diff --git a/07_nlm_denoise/blackbook537/.gitattributes b/07_nlm_denoise/blackbook537/.gitattributes
new file mode 100644
index 00000000..d3a1da34
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/.gitattributes
@@ -0,0 +1,14 @@
+* text=auto eol=lf
+*.sh text eol=lf
+*.py text eol=lf
+*.cpp text eol=lf
+*.h text eol=lf
+*.cu text eol=lf
+*.mu text eol=lf
+*.maca text eol=lf
+Makefile text eol=lf
+*.png binary
+*.pdf binary
+# Historical evidence hashes cover original line endings and tool-output spaces.
+# Do not normalize or whitespace-check these archived records as source code.
+assets/results/** -text -whitespace
diff --git a/07_nlm_denoise/blackbook537/.gitignore b/07_nlm_denoise/blackbook537/.gitignore
new file mode 100644
index 00000000..8c17b850
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/.gitignore
@@ -0,0 +1,13 @@
+/build/
+/output/
+__pycache__/
+*.pyc
+*.o
+*.obj
+*.nsys-rep
+*.ncu-rep
+*.qdrep
+.vscode/
+.idea/
+*.swp
+.DS_Store
diff --git a/07_nlm_denoise/blackbook537/LICENSE b/07_nlm_denoise/blackbook537/LICENSE
new file mode 100644
index 00000000..52c5e5f5
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 InfiniTensor
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/07_nlm_denoise/blackbook537/Makefile b/07_nlm_denoise/blackbook537/Makefile
new file mode 100644
index 00000000..52c158ef
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/Makefile
@@ -0,0 +1,150 @@
+# NLM denoise: make build / run / validate / test / bench / clean
+.DEFAULT_GOAL := build
+SHELL := /bin/bash
+.DELETE_ON_ERROR:
+
+PLATFORM    ?= nvidia
+WITH_OPENCV ?= 0
+FAST_EXP    ?= 0
+
+# 可选覆盖变量（默认即训练营服务器环境，无需设置）：
+#   NVCC     nvcc 路径（默认取 PATH 中的 nvcc）
+#   CXX_HOST nvcc 的主机编译器（本地 gcc 版本过新时指定，如 conda gcc13）
+#   CUDA_RT  非系统位置的 cudart 运行库目录（需含 include/ 与 lib/）
+#   ARCH     目标 GPU 架构（如 sm_80/sm_86/sm_90；留空用 nvcc 默认全架构，
+#            会产生 pre-sm_75 弃用警告并增加编译时间）
+NVCC        ?= nvcc
+CXX_HOST    ?=
+CUDA_RT     ?=
+ARCH        ?=
+COREX_HOME  ?= /usr/local/corex
+COREX_CXX   ?= $(COREX_HOME)/bin/clang++
+MUSA_HOME   ?= /usr/local/musa
+MCC         ?= $(MUSA_HOME)/bin/mcc
+MACA_HOME   ?= /opt/maca
+MACA_CUDA   ?= $(MACA_HOME)/tools/cu-bridge
+MXCC        ?= $(MACA_HOME)/mxgpu_llvm/bin/mxcc
+
+SUFFIX          := cu
+PLATFORM_DEFINE := -DPLATFORM_NVIDIA
+CFLAGS          := -std=c++17 -O3
+EXTRA_LIBS      :=
+COMP_ONLY_FLAGS :=
+
+ifeq ($(PLATFORM),nvidia)
+    CC              := $(NVCC)
+    PLATFORM_DEFINE := -DPLATFORM_NVIDIA
+    ifneq ($(CXX_HOST),)
+        CFLAGS += -ccbin $(CXX_HOST)
+    endif
+    ifneq ($(CUDA_RT),)
+        CFLAGS     += -I$(CUDA_RT)/include
+        EXTRA_LIBS += -L$(CUDA_RT)/lib -lcudart_static -lpthread -ldl -lrt
+    endif
+    ifneq ($(ARCH),)
+        CFLAGS += -arch=$(ARCH)
+    endif
+else ifeq ($(PLATFORM),iluvatar)
+    CC              := $(COREX_CXX)
+    CFLAGS          := -std=c++17 -O3 -I$(COREX_HOME)/include -fPIC
+    COMP_ONLY_FLAGS := -x ivcore --cuda-path=$(COREX_HOME)
+    PLATFORM_DEFINE := -DPLATFORM_ILUVATAR
+    EXTRA_LIBS      := -L$(COREX_HOME)/lib64 -Wl,-rpath,$(COREX_HOME)/lib64 -lcudart
+else ifeq ($(PLATFORM),metax)
+    CC              := $(MXCC)
+    CFLAGS          := -std=c++17 -O3 -I$(MACA_CUDA)/include
+    SUFFIX          := maca
+    PLATFORM_DEFINE := -DPLATFORM_METAX
+    # MACA 3.x 的 CUDA 兼容头位于 cu-bridge；将设备对象单独链接时还需
+    # 显式链接 wcuda 符号实现。rpath 使产物无需额外设置 LD_LIBRARY_PATH。
+    EXTRA_LIBS      := -L$(MACA_HOME)/lib -Wl,-rpath,$(MACA_HOME)/lib \
+                       -lruntime_cu -lsymbol_cu
+else ifeq ($(PLATFORM),moore)
+    CC              := $(MCC)
+    CFLAGS          := -std=c++11 -O3
+    SUFFIX          := mu
+    PLATFORM_DEFINE := -DPLATFORM_MOORE
+    # rpath 让构建产物在未手工设置 LD_LIBRARY_PATH 时也能找到 libmusart；
+    # 避免绑定发行版特定的 GCC 目录，mcc 会自行选择兼容的 Host 工具链。
+    EXTRA_LIBS      := -I$(MUSA_HOME)/include -L$(MUSA_HOME)/lib \
+                       -Wl,-rpath,$(MUSA_HOME)/lib -lmusart
+else
+    $(error Unsupported PLATFORM '$(PLATFORM)' (expected: nvidia, iluvatar, metax, moore))
+endif
+
+ifeq ($(FAST_EXP),1)
+    PLATFORM_DEFINE += -DNLM_FAST_EXP
+endif
+
+ifeq ($(WITH_OPENCV),1)
+    PLATFORM_DEFINE += -DWITH_OPENCV
+    CFLAGS    += $(shell pkg-config --cflags opencv4)
+    EXTRA_LIBS += $(shell pkg-config --libs opencv4)
+endif
+
+
+TARGET          := build/nlm_denoise
+PLATFORM_TARGET := build/$(PLATFORM)/nlm_denoise
+OBJ_DIR         := build/obj/$(PLATFORM)
+INCLUDES        := -Iinclude
+HOST_SRCS       := $(wildcard src/*.cpp)
+HOST_OBJS       := $(patsubst src/%.cpp,$(OBJ_DIR)/%.o,$(HOST_SRCS))
+KER_SRC         := src/kernels.$(SUFFIX)
+KER_OBJ         := $(OBJ_DIR)/kernels.o
+OBJS            := $(HOST_OBJS) $(KER_OBJ)
+HEADERS         := $(wildcard include/*.h)
+
+INPUT       ?= test_images/noisy_256x256_3ch_sigma25.png
+OUTPUT      ?= output/images/denoised.png
+PARAMS      ?= params.txt
+MODE        ?= $(if $(filter metax,$(PLATFORM)),1,2)
+RUN_LOG     ?= output/logs/nlm_perf.csv
+SIZES       ?= 1920x1080,3840x2160
+CHANNELS    ?= 1,3
+PARAM_SETS  ?= all
+WARMUP      ?= 3
+REPEAT      ?= 10
+BENCH_LOG   ?= output/results/benchmark.csv
+
+.PHONY: all build run validate test bench clean FORCE
+all: build
+build: $(PLATFORM_TARGET)
+
+# Keep the documented CLI path while all recipes use their own platform binary.
+$(TARGET): $(PLATFORM_TARGET) FORCE
+	@cp -f $(PLATFORM_TARGET) $(TARGET)
+build: $(TARGET)
+
+run: build
+	$(PLATFORM_TARGET) run -i "$(INPUT)" -o "$(OUTPUT)" -p "$(PARAMS)" --kernel $(MODE) --log "$(RUN_LOG)"
+
+validate: build
+	$(PLATFORM_TARGET) validate -i "$(INPUT)" -o "$(OUTPUT)" -p "$(PARAMS)"
+
+test: build
+	$(PLATFORM_TARGET) units
+	$(PLATFORM_TARGET) validate -i "$(INPUT)" -p "$(PARAMS)"
+
+bench: build
+	$(PLATFORM_TARGET) bench --sizes "$(SIZES)" --channels "$(CHANNELS)" --param-sets "$(PARAM_SETS)" --warmup $(WARMUP) --repeat $(REPEAT) --log "$(BENCH_LOG)"
+
+clean:
+	rm -rf build
+
+# Track compiler and flag changes without touching unchanged objects.
+$(OBJ_DIR)/build_flags.txt: FORCE
+	@mkdir -p $(OBJ_DIR)
+	@printf '%s\n' '$(CC) $(CFLAGS) $(COMP_ONLY_FLAGS) $(PLATFORM_DEFINE) $(INCLUDES) $(EXTRA_LIBS)' > $@.tmp
+	@cmp -s $@.tmp $@ && rm -f $@.tmp || mv -f $@.tmp $@
+
+$(PLATFORM_TARGET): $(OBJS)
+	@mkdir -p $(dir $@)
+	$(CC) $(CFLAGS) $(PLATFORM_DEFINE) -o $@ $(OBJS) $(EXTRA_LIBS)
+
+$(OBJ_DIR)/%.o: src/%.cpp $(HEADERS) $(OBJ_DIR)/build_flags.txt
+	@mkdir -p $(dir $@)
+	$(CC) $(CFLAGS) $(PLATFORM_DEFINE) $(INCLUDES) -c $< -o $@
+
+$(KER_OBJ): $(KER_SRC) src/kernels.cu $(HEADERS) $(OBJ_DIR)/build_flags.txt
+	@mkdir -p $(dir $@)
+	$(CC) $(CFLAGS) $(COMP_ONLY_FLAGS) $(PLATFORM_DEFINE) $(INCLUDES) -c $(KER_SRC) -o $@
diff --git a/07_nlm_denoise/blackbook537/README.md b/07_nlm_denoise/blackbook537/README.md
new file mode 100644
index 00000000..5063eade
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/README.md
@@ -0,0 +1,142 @@
+# CUDA 非局部均值图像降噪
+
+2026 夏季训练营 · 选题 07 · blackbook537。支持灰度/RGB PNG、JPG，包含 CPU 参考、三个 GPU 版本、质量验证和性能基准。
+
+目录按参考作业的 `src / include / scripts / assets / test_images` 组织。算法保留本项目实现，NVIDIA 与 Iluvatar 共用 CUDA 源文件，MACA/MUSA 使用各自编译入口。
+
+## 项目结构
+
+```text
+blackbook537/
+├── src/
+│   ├── main.cpp              # 命令行入口
+│   ├── kernels.cu           # V0/V1/V2 核心实现，NVIDIA / Iluvatar
+│   ├── kernels.maca         # MetaX 编译入口
+│   ├── kernels.mu           # Moore Threads 编译入口
+│   ├── pipeline.cpp         # 显存、传输、转换与 kernel 调度
+│   ├── nlm_cpu_ref.cpp      # 同语义 CPU 参考
+│   ├── image_io.cpp         # PNG/JPG 读写
+│   ├── params.cpp           # 参数解析
+│   ├── benchmark.cpp        # 性能统计
+│   ├── metrics.cpp          # MAE / PSNR
+│   ├── validate.cpp         # CPU/GPU 对比
+│   └── self_check.cpp       # 专项要求保留的功能自检
+├── include/                 # 扁平头文件与 stb 依赖
+├── scripts/
+│   ├── reproduce.sh         # 四平台统一复现入口
+│   ├── run_quality_sweep.sh # 配对图像质量扫描
+│   ├── collect_environment.sh
+│   ├── nsys_profile.sh
+│   ├── ncu_profile.sh
+│   ├── generate_report_assets.py
+│   ├── export_report.py     # 从 Markdown 导出 PDF
+│   └── params/              # 非默认参数组
+├── assets/                  # 报告图表、results/ 原始证据
+├── test_images/             # 示例输入、干净参考和历史降噪示例
+├── params.txt               # 默认 pr=3, sr=10, h=10, sigma=25
+├── lab_report.md            # 可编辑实验报告
+├── lab_report.pdf           # 阅读版实验报告
+├── SUBMISSION.md            # 补充文件建议与提交前检查
+├── requirements-report.txt # 仅报告工具使用的 Python 依赖
+├── Makefile
+├── LICENSE
+├── .gitattributes
+├── .gitignore
+└── README.md
+```
+
+`build/` 和 `output/` 在运行时生成并被忽略。仓库外层选题目录遵循本季度提交要求。
+
+## 构建与运行
+
+使用 Linux 或 WSL、GNU Make，以及对应平台 SDK。NVIDIA/MetaX/Iluvatar 使用 C++17，Moore 使用 C++11。普通构建不依赖 Python 或 OpenCV。
+
+```bash
+cd 07_nlm_denoise/blackbook537
+make build PLATFORM=nvidia ARCH=sm_86   # 示例：RTX 3060；4090 D 使用 sm_89
+make run PLATFORM=nvidia ARCH=sm_86    # 随附 256×256 RGB 图片，可快速检查
+
+# 1080p 降噪，选择 V2，结果位于 output/images/
+make run PLATFORM=nvidia ARCH=sm_86 \
+  INPUT=test_images/noisy_1920x1080_3ch_sigma25.png MODE=2
+
+# 自定义输入和参数；JPG 输入请将 OUTPUT 也指定为 JPG
+make run INPUT=your_image.jpg OUTPUT=output/images/denoised.jpg PARAMS=params.txt MODE=1
+```
+
+各平台构建选项：
+
+| 平台 | 命令 | 主要可覆盖变量 |
+|---|---|---|
+| NVIDIA | `make build PLATFORM=nvidia ARCH=sm_89` | `NVCC`、`CXX_HOST`、`CUDA_RT`、`ARCH` |
+| Iluvatar | `make build PLATFORM=iluvatar` | `COREX_HOME`、`COREX_CXX` |
+| MetaX | `make build PLATFORM=metax` | `MACA_HOME`、`MACA_CUDA`、`MXCC` |
+| Moore | `make build PLATFORM=moore` | `MUSA_HOME`、`MCC` |
+
+运行时使用相同的 `PLATFORM` 和构建选项。对象及可执行文件按平台隔离；公共入口 `build/nlm_denoise` 指向最后选择的平台副本。脚本使用 `build/<platform>/nlm_denoise`。修改编译器、编译选项或头文件后，Make 会重新编译相应对象。
+
+| 命令 | 行为 |
+|---|---|
+| `make` / `make build` | 只构建 |
+| `make run` | 对图片降噪 |
+| `make validate` | 三个 GPU 版本对 CPU 参考计算 MAE/PSNR |
+| `make test` | Host 自检和三版本 GPU 验证 |
+| `make bench` | 1080p/4K × 灰度/RGB × 四参数组 × 三版本统计 |
+| `make clean` | 删除构建产物；保留实验输出 |
+
+## 版本与参数
+
+| GPU 版本 | 实现 | 使用说明 |
+|---|---|---|
+| V0 (`MODE=0`) | 每线程一像素，全局内存直接访问 | 优化对照 |
+| V1 (`MODE=1`) | 带 halo 的 shared memory tile | MetaX 的 Make 默认版本 |
+| V2 (`MODE=2`) | V1 + patch 半径模板展开 | 其他平台的 Make 默认版本；pr≥4 回退 V1 |
+
+CPU 参考不占用 GPU 版本编号。直接运行 CLI 时 `--kernel` 默认仍为 2。
+`patch_radius` 支持 1–8，`search_radius` 支持 1–32，参数合法范围由解析器检查。默认精确 `expf`；`FAST_EXP=1` 开启近似指数；`WITH_OPENCV=1` 加入可选 OpenCV 对比（需要 `pkg-config opencv4`）。
+
+**算法口径需要关注：** 当前采用 patch 距离求和，噪声减项为 `2*sigma²*N_patch`。题面写作 `2*sigma²`，两者不能直接宣称等价；详见[实验报告](lab_report.md)。生成器的 `--sigma` 表示均匀噪声幅度，不能当作高斯噪声标准差。
+
+## 验证与复现
+
+```bash
+make test PLATFORM=nvidia ARCH=sm_86
+make validate INPUT=test_images/noisy_1920x1080_3ch_sigma25.png
+
+# 相同输入、相同参数的 GPU 与 CPU 计时
+./build/nlm_denoise bench --sizes 1920x1080 --channels 3 --param-sets base \
+  --warmup 3 --repeat 10 --with-cpu --cpu-repeat 3 --log output/results/cpu_baseline.csv
+
+# 完整复现；其他平台改 PLATFORM，NVIDIA 指定实际架构
+PLATFORM=nvidia ARCH=sm_89 bash scripts/reproduce.sh
+PLATFORM=metax bash scripts/reproduce.sh
+
+# 单独复现质量扫描
+PLATFORM=nvidia ARCH=sm_86 bash scripts/run_quality_sweep.sh
+```
+
+完整矩阵和 CPU 参考耗时较长。快速检查可使用 `SIZES=64x48 VALIDATE_SIZE=64x48 QUALITY_SIZE=64x48 WARMUP=0 REPEAT=1 CPU_REPEAT=1 bash scripts/reproduce.sh`；这种小图检查不替代正式性能验收。
+
+CLI 还提供 `gen`（合成图）、`metrics`（两图指标）和 `units`（Host 自检）。无参数运行可查看帮助。退出码：0 成功、1 验证不通过、2 参数或运行错误。
+
+## 实验报告与证据
+
+[实验报告 PDF](lab_report.pdf) · [Markdown 源稿](lab_report.md) · [原始结果索引](assets/results/README.md) · [补充文件建议](SUBMISSION.md)
+
+历史 1080p RGB/base GPU kernel 数据：4090 D V2 **57.73 ms（最小值）**；S4000 V2 **467.985 ms（均值）**；C500 V1 **418.397 ms（均值）**；MR-V100 V2 **716.595 ms（均值）**。设备与统计口径不同，不能把这些数字直接作为统一排名。
+
+![历史 NVIDIA 性能图](assets/performance_4090.png)
+
+![配对合成图质量示例](assets/quality_triptych.png)
+
+`e2e_ms` 为内存图像处理流程计时，不包含文件解码、编码；`Mpx/s` 当前按 kernel 时间计算。历史材料不能证明本次重构已在所有平台重跑。当前核对记录见 [SUBMISSION.md](SUBMISSION.md)。
+
+## 报告工具与依赖
+
+```bash
+python3 -m pip install -r requirements-report.txt
+python3 scripts/export_report.py
+python3 scripts/generate_report_assets.py --performance-only
+```
+
+stb_image v2.30 / stb_image_write v1.16 用于图像编解码；第三方许可证完整保留在对应头文件末尾，项目许可证见 [LICENSE](LICENSE)。所有示例图来源和噪声解释见 [test_images/README.md](test_images/README.md)。
diff --git a/07_nlm_denoise/blackbook537/SUBMISSION.md b/07_nlm_denoise/blackbook537/SUBMISSION.md
new file mode 100644
index 00000000..5b15d0ca
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/SUBMISSION.md
@@ -0,0 +1,80 @@
+# 提交整理与补充建议
+
+## 已完成的结构调整
+
+以参考作业的五个主要目录 `src/`、`include/`、`scripts/`、`assets/`、`test_images/` 为格式，重构本人的 NLM 提交。仓库外层的季度说明、选题目录和 LICENSE 保持原有内容。
+
+| 原位置/内容 | 新位置/处理 |
+|---|---|
+| 多层 `include/{core,nlm,pal,kernels,tester}` | 扁平 `include/` |
+| 四平台 `kernels/` 与共享 `.inl` | `src/kernels.cu/.maca/.mu`；CUDA 和 CoreX 共用 `.cu` |
+| `tester/` | `src/benchmark.cpp`、`metrics.cpp`、`validate.cpp`、`self_check.cpp` |
+| `third_party/stb/` | `include/`，保留头文件内完整许可证 |
+| `params/default.txt` | 根目录 `params.txt` |
+| 多份相同默认参数 | 合并到 `params.txt`；其余位于 `scripts/params/` |
+| `docs/` 多篇重叠说明 | `README.md` 与 `lab_report.md/.pdf` |
+| `docs/assets/` | `assets/` |
+| `data/clean/`、`data/noisy/` | `test_images/`，新增来源说明 |
+| 四份平台复现脚本 | 统一 `scripts/reproduce.sh` |
+| 个人路径 WSL 脚本、过程记录、重复输出图片 | 移出提交；保存在仓库外完整备份中 |
+| `experiments/results/` | 精选 CSV、环境、验证与 trace 位于 `assets/results/` |
+| 所有运行输出 | `output/`，由 `.gitignore` 排除 |
+
+`make run` 现执行降噪；`make test` 保留自检与 GPU 验证；默认 `make` 只构建。编译对象按平台隔离，头文件及编译选项变动会触发重编。脚本使用 LF 换行。
+
+## 建议补充的文件（按优先级）
+
+| 优先级 | 建议文件 | 应提供的实质内容 | 当前状态 |
+|---|---|---|---|
+| P0 | `assets/review_response.md` | PR 链接、逐条评语、对应修改和验证证据 | 原 PR #68 已关闭且未合并；尚未取得具体评语 |
+| P0 | `assets/algorithm_contract.md` | 题面 dist 定义、2*sigma² 与 2*sigma²*N_patch 的关系、h 标定、RGB/边界处理、教师认可口径或严格题面验证结果 | 已在报告揭示差异，尚需解决 |
+| P0 | `assets/results/<device>/environment.txt` 与新 CSV | 最终 commit、工作树状态、编译选项；NVIDIA 1080p/4K mean/min/stddev 与同条件 CPU 重复数据 | 旧 NVIDIA 主要为 legacy 口径 |
+| P1 | `test_images/dataset_manifest.csv` | 自然图来源、授权、分辨率、通道、SHA256、clean/noisy 对应、噪声生成参数 | 当前只有合成示例说明 |
+| P1 | 独立正确性样例与验证结果 | 小尺寸手算例、边界/纹理图、标准差受控的噪声；与题意独立对照，避免只证明 CPU/GPU 相互一致 | 现有 12 项自检与 CPU/GPU 验证保留 |
+| P1 | `assets/nsight/` | 原生 nsys 时间线和 ncu 的吞吐、占用率、寄存器/内存指标截图，附命令和环境 | 有 nsys 文本与图表，缺 ncu 计数器 |
+| P2 | `assets/demo.mp4` | 参数选择、输入/输出、可重复的运行演示 | 尚未制作，不影响代码构建 |
+
+不要创建空白占位文件来代替证据。自然图的使用授权和结果需真实提供；新的时间线截图应来自本人实现，不能使用参考作业截图。
+
+## 优先解决的内容问题
+
+1. 当前噪声生成器使用均匀噪声幅度，题面 sigma 是标准差估计。报告已说明区别；若改为高斯噪声或标准差标定，应更新文件名/元数据并重跑质量实验，不能直接沿用旧指标。
+2. 当前公式减项按 patch 大小和通道数缩放。CPU 与 GPU 一致不能单独证明满足题意。先确定算法口径，再进行正式重跑。
+3. `e2e_ms` 不包含文件 I/O 和函数退出时的显存释放，吞吐量按 kernel 计算。不要再将其称为完整应用耗时或直接声称达到完整 4K 交互目标。
+4. 旧 NVIDIA 的 kernel min 与其他平台的 kernel mean 不可混排；缺少采样值时不补造标准差。
+5. 主核耗时占比高只说明优化对象，不能据此宣称 compute-bound；ncu 数据仍需补采。
+
+## 本次验证记录
+
+2026-09-16，在 WSL Ubuntu、RTX 3060 Laptop、CUDA 12.9、sm_86 上实际执行：
+
+| 检查 | 结果 |
+|---|---|
+| 全部迁移源码干净构建 | PASS |
+| Host 源码 C++11 语法兼容检查 | PASS，`g++ -std=c++11 -Iinclude -fsyntax-only src/*.cpp` |
+| `make test`：12 项 Host 自检 + 256×256 RGB 三版本对 CPU | PASS，V0/V1/V2 MAE=0、PSNR=inf |
+| `make run` 默认图像处理 | PASS，写入 output/images/ 与 output/logs/ |
+| 相同配置再次构建 | PASS，kernel 对象时间戳不变，无重复编译 |
+| NVIDIA / Iluvatar / MetaX / Moore 构建命令检查 | PASS；后三平台仅检查构建依赖和命令生成 |
+| 所有保留 Bash 脚本语法检查 | PASS |
+| 64×48 全流程复现：构建、验证、24 组合、CPU、质量、证据哈希 | PASS；本机 profiler 缺失，明确标记 SKIPPED |
+| 1080p clean/noisy 图重新生成 | 与保留样例逐字节一致 |
+| 1080p V2/base 输出 | 与保留的历史 MetaX 输出逐字节一致 |
+| 31×23 灰度图（非整块边界）三版本对 CPU | PASS，MAE=0、PSNR=inf |
+
+本地日志保留在被忽略的 `output/refactor_validation/`，可按需审阅；小图完整复现仅验证脚本链路，不替代 1080p/4K 正式性能验收。国产平台需在对应设备上重跑，历史日志不替代本次重构验证。
+
+## 提交前命令
+
+```bash
+make build PLATFORM=nvidia ARCH=sm_89
+make test PLATFORM=nvidia ARCH=sm_89
+PLATFORM=nvidia ARCH=sm_89 bash scripts/reproduce.sh
+python3 scripts/export_report.py
+git diff --check
+git status --short
+```
+
+确认 PR 目标分支为 `2026-summer-project`，变更范围仅包含 `07_nlm_denoise/blackbook537/`。
+
+原提交为 [PR #68](https://github.com/InfiniTensor/Learning-CUDA/pull/68)，目前已关闭且未合并。此次更新沿用 `blackbook537/Learning-CUDA-2026` 的 `submit/nlm-denoise-blackbook537` 分支。重新提交时，应说明本次结构整理与验证范围，保留上述尚需补充的内容，不将历史跨平台数据写成本次重新实测。
diff --git a/07_nlm_denoise/blackbook537/assets/README.md b/07_nlm_denoise/blackbook537/assets/README.md
new file mode 100644
index 00000000..2bab0846
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/README.md
@@ -0,0 +1,18 @@
+# 报告图表
+
+图表来自本项目历史结果。`results/` 保存相应 CSV、环境、验证日志与 MetaX 原生 trace。性能图中的 kernel min / mean 以图下注释和原始 CSV 为准。
+
+| 图表 | 数据 |
+|---|---|
+| [RTX 4090 D 性能](performance_4090.png) | rtx4090d/benchmark_legacy.csv；kernel min，e2e mean |
+| [MTT S4000 性能](performance_moore_s4000.png) | moore_s4000_musa5.1/benchmark_1080p_rgb_base.csv、benchmark_4k_rgb_base.csv |
+| [MetaX C500 性能](performance_metax_c500.png) | metax_c500_maca3.0/benchmark_base.csv |
+| [Iluvatar MR-V100 性能](performance_iluvatar_mrv100.png) | iluvatar_mrv100_corex4.4.0/benchmark_base.csv |
+| [RTX 3060 质量与延迟](quality_latency_tradeoff.png) | rtx3060_laptop/quality_tradeoff.csv、benchmark_quality.csv |
+| [MetaX 质量与延迟](quality_latency_metax_c500.png) | metax_c500_maca3.0/quality_tradeoff.csv、benchmark_base.csv、benchmark_matrix.csv；V2 |
+| [Iluvatar 质量与延迟](quality_latency_iluvatar_mrv100.png) | iluvatar_mrv100_corex4.4.0/quality_tradeoff.csv、benchmark_base.csv、benchmark_matrix.csv；V2 |
+| [合成配对图](quality_triptych.png) | test_images/ 中的 1080p clean/noisy/V2 输出 |
+| [nsys 摘要](nsys_kernel_summary.png) | rtx4090d/nsys_summary_legacy.csv |
+| [MetaX 时间线](mctracer_timeline_metax_c500.png) | metax_c500_maca3.0/mctracer/nlm_1080_v1-3213.json |
+
+重复的跨平台三联图及旧验收卡已从提交中移除；对应原始版本保留在仓库外备份。图表不用于证明所有平台在当前重构后均重新运行成功。
diff --git a/07_nlm_denoise/blackbook537/assets/mctracer_timeline_metax_c500.png b/07_nlm_denoise/blackbook537/assets/mctracer_timeline_metax_c500.png
new file mode 100644
index 00000000..84c84fa4
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/mctracer_timeline_metax_c500.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/nsys_kernel_summary.png b/07_nlm_denoise/blackbook537/assets/nsys_kernel_summary.png
new file mode 100644
index 00000000..dacd649f
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/nsys_kernel_summary.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/performance_4090.png b/07_nlm_denoise/blackbook537/assets/performance_4090.png
new file mode 100644
index 00000000..3b3f1bee
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/performance_4090.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/performance_iluvatar_mrv100.png b/07_nlm_denoise/blackbook537/assets/performance_iluvatar_mrv100.png
new file mode 100644
index 00000000..9542b1bd
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/performance_iluvatar_mrv100.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/performance_metax_c500.png b/07_nlm_denoise/blackbook537/assets/performance_metax_c500.png
new file mode 100644
index 00000000..491931a8
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/performance_metax_c500.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/performance_moore_s4000.png b/07_nlm_denoise/blackbook537/assets/performance_moore_s4000.png
new file mode 100644
index 00000000..d7186d31
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/performance_moore_s4000.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/quality_latency_iluvatar_mrv100.png b/07_nlm_denoise/blackbook537/assets/quality_latency_iluvatar_mrv100.png
new file mode 100644
index 00000000..38f08d18
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/quality_latency_iluvatar_mrv100.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/quality_latency_metax_c500.png b/07_nlm_denoise/blackbook537/assets/quality_latency_metax_c500.png
new file mode 100644
index 00000000..13e2b54e
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/quality_latency_metax_c500.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/quality_latency_tradeoff.png b/07_nlm_denoise/blackbook537/assets/quality_latency_tradeoff.png
new file mode 100644
index 00000000..f0fab150
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/quality_latency_tradeoff.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/quality_triptych.png b/07_nlm_denoise/blackbook537/assets/quality_triptych.png
new file mode 100644
index 00000000..f2004284
Binary files /dev/null and b/07_nlm_denoise/blackbook537/assets/quality_triptych.png differ
diff --git a/07_nlm_denoise/blackbook537/assets/results/README.md b/07_nlm_denoise/blackbook537/assets/results/README.md
new file mode 100644
index 00000000..3c86e6ac
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/README.md
@@ -0,0 +1,22 @@
+# 历史实验结果索引
+
+这些文件来自结构重构前的已有实验。整理时只改变文件位置，原始 CSV、环境、日志和 JSON 内容保持原样；其中旧路径与旧命令是历史记录。删除的中间构建日志、重复图像和旧校验清单保存在仓库外备份。
+
+| 目录 | 数据与说明 |
+|---|---|
+| [rtx4090d](rtx4090d/) | 2026-09-07，CUDA 12.8；`benchmark_legacy.csv` 为 kernel min / e2e mean，缺 stddev；另含 CPU 对照、质量、nsys 摘要及原始综合日志 |
+| [rtx3060_laptop](rtx3060_laptop/) | legacy 基准，以及 2026-09-15 的配对质量补充实验；`benchmark_quality.csv` 为 warmup=3/repeat=5；环境为当时未提交工作树 |
+| [moore_s4000_musa5.1](moore_s4000_musa5.1/) | 2026-09-15，MUSA 5.1；base 1080p/4K 与 RGB all 为 repeat=10；完整矩阵 repeat=3；CPU base 仅一次，不能写成三次 |
+| [metax_c500_maca3.0](metax_c500_maca3.0/) | 2026-09-15，MACA 3.0，C500 25% sGPU；base repeat=10、矩阵 repeat=3、CPU repeat=3；含原生 mcTracer JSON |
+| [iluvatar_mrv100_corex4.4.0](iluvatar_mrv100_corex4.4.0/) | 2026-09-15/16，CoreX 4.4.0；base repeat=10、矩阵 repeat=3、CPU repeat=3；上传工作树快照的来源记录仍保留 |
+
+`SHA256SUMS.txt` 以本目录为起点覆盖当前保留的原始证据文件。这个清单只验证整理后的文件完整性，不代表重新运行了实验；旧归档哈希仍属于当时的原始归档。
+
+Linux 下复核：
+
+```bash
+cd assets/results
+sha256sum -c SHA256SUMS.txt
+```
+
+新实验默认写入 `output/results/<platform>/`，待检查环境、参数、结果与最终 commit 一致后，再选择需要的证据纳入本目录。完整复现命令见 [README](../../README.md)。
diff --git a/07_nlm_denoise/blackbook537/assets/results/SHA256SUMS.txt b/07_nlm_denoise/blackbook537/assets/results/SHA256SUMS.txt
new file mode 100644
index 00000000..e182c2f0
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/SHA256SUMS.txt
@@ -0,0 +1,37 @@
+f205f9aa4956ce8e5489bfba167ce77e98b568a615faeaa5fd24083d2a08c538  iluvatar_mrv100_corex4.4.0/benchmark_base.csv
+9ae3b424fa96deebd9742f699f203169589c93bddc8fde5c81d82ccc4b198914  iluvatar_mrv100_corex4.4.0/benchmark_matrix.csv
+4365d3d58b07339ec4a07e4cc15756df3a379e2ccf610460677b2f1668e8e998  iluvatar_mrv100_corex4.4.0/cpu_baseline_repeat3.csv
+4ad87ea2cdd30046aa1a1ab5f24e9dbe3fd92060c8cd70e757f0353187aa8699  iluvatar_mrv100_corex4.4.0/environment.txt
+cfa64f393705f363fce7a158053c3105c8a6ba7014515b93aebee6503d9c93c9  iluvatar_mrv100_corex4.4.0/logs/validate_1920x1080_rgb.log
+f38fa3bf4b4c1a6eae7a89befc68dcd4e8058cd1453a39e91d83cc658dd6f8cc  iluvatar_mrv100_corex4.4.0/quality_sigma.csv
+10ff2f0f0eddbee3e04a53e716634eef4a4a1077871bf5e57f3b278a45d9ce07  iluvatar_mrv100_corex4.4.0/quality_tradeoff.csv
+29c50bd29d97749fdd0075b7212a12862cf225f92a439e6dcbd7ea990529d7bc  iluvatar_mrv100_corex4.4.0/source_snapshot.txt
+93a071eaef3c8e8de61eaa0a0188bb65ed31d87d2f9e385aa823dcf3da1680bb  metax_c500_maca3.0/benchmark_base.csv
+fc28b85e9ccde1903e844233cfa2c4023eda3e745f9a2b8641b3445aa18c85c4  metax_c500_maca3.0/benchmark_base_gray.csv
+dfc2a7478d4c43a7d5a6cc15b7a3478c7363329be4233c4f623c3116f4d987fa  metax_c500_maca3.0/benchmark_matrix.csv
+d9c09a3394279895b2998a9ead3b8f74dcd8f833a9af98331a26966781f2d056  metax_c500_maca3.0/cpu_baseline_repeat3.csv
+49ce065e1743e5eca16c8bc6831c5e39a2339297974cecafe8934d75dce08d86  metax_c500_maca3.0/environment.txt
+88363202761d37ab37c59954c40ad62de5b945c45040491c25ed4b073794af37  metax_c500_maca3.0/mctracer/nlm_1080_v1-3213.json
+c2fb08ee3a3db697314dd3b6cbf5783f3ef99faadd0901fe4b2942ed50360aad  metax_c500_maca3.0/quality_sigma.csv
+b2ee0cfad4014338de0c15023f002bbd1332ceae370580766fae333c855f0ba1  metax_c500_maca3.0/quality_tradeoff.csv
+dc85f1178fdcc42723e1fb06019d996f3e34120c4e3434f139ccb78041ab5cf8  metax_c500_maca3.0/validate_1920x1080_rgb.log
+caafa1da2e73a75c8550090483d3e91c64e14b1829970c06642722c04743956e  moore_s4000_musa5.1/benchmark_1080p_rgb_all.csv
+184cc396ae05a343a772a8da5fe1f0bca7e9c5e3da1abe0a13e512f9c857c268  moore_s4000_musa5.1/benchmark_1080p_rgb_base.csv
+d7eae800f04749588cc8f4b5d61023035fe5cf668c1b161d8abece8d4a009c5a  moore_s4000_musa5.1/benchmark_4k_rgb_base.csv
+d5fc45c92d760cf10558bb1ba6810e3506d66d048d4b8af98794e7443834aef7  moore_s4000_musa5.1/benchmark_full_matrix_repeat3.csv
+94dd3f4e6de2605417af71d99b5b061a136b8e5b71cff09d16813febbbe13b65  moore_s4000_musa5.1/environment_collector_verified.txt
+4d5cdefc38e1fb3950612f892b61c149117687b7ada62d539e68b44334ef5f19  moore_s4000_musa5.1/quality_sigma.csv
+a90e63e0f55148ca80bae8b15c1ca6b8bfe5f2bba4411f0e45d45cf869f29e86  moore_s4000_musa5.1/quality_tradeoff.csv
+192feeefb416de501dd514a2628d621188bb08ca51fbca8ee13422117d7ab191  moore_s4000_musa5.1/validate_1920x1080_rgb.log
+fcf711e4147e65ef886ced1ea2db5880cce30655e38fe52942e05f1cf42dd373  rtx3060_laptop/benchmark_legacy.csv
+58304cf46e149ea19c274bf54e5bdd1d234c0cf2bb19355b13c2e0e290f9dad5  rtx3060_laptop/benchmark_quality.csv
+2c4a7cc2959fdab97b3c8c5ba29006319e4fa335402541d6525d7b5e98e68889  rtx3060_laptop/environment_quality.txt
+b968ad72878ac8696bcaaaa4ec04e8ffb2f1da9427164e7c1777a4027cc70546  rtx3060_laptop/quality_sigma.csv
+403ef89d444fd8c6c5795f8fd579a97dcfe6dae0f9ed6ad76b5a285b4765d379  rtx3060_laptop/quality_tradeoff.csv
+0db2e2d915f2ff866e36ad36cf0de7261361d340c8e1159f0ba3d6d2584cd050  rtx3060_laptop/run_all_legacy.txt
+7049b96bfca333b967b34e881f8e90e18ba32ad0828b8f3f36b984a9787d48fd  rtx4090d/benchmark_legacy.csv
+c01b04cdeb457c660e959b4f469918e29e9cf97c4c3ea25a940a07b7869744a3  rtx4090d/environment_legacy.txt
+870987ff40256722cee7f32c363d913d72df7feeeb7a168aedda4a688b631416  rtx4090d/nsys_summary_legacy.csv
+9042cc72d90062f0c023b68483f21ff1c73e5aec56ca3eb686c1f06c3163d99a  rtx4090d/quality_legacy.csv
+81b59daa55dc6a30b1e7236ef666bbf4007a7a1c75e3bb04c7bc023db8eb2850  rtx4090d/run_all_legacy.txt
+f020002cdaa65c6e57cde77780037c1b848168d98f2dc6e7b6e7f06293d4df22  rtx4090d/validation_legacy.csv
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/benchmark_base.csv b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/benchmark_base.csv
new file mode 100644
index 00000000..b47c4d14
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/benchmark_base.csv
@@ -0,0 +1,7 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+base,1920x1080,3,3,10,10.000000,25.000000,0,3,10,1561.032275,1556.602905,1.771660,1565.984094,1561.638428,1.811760,1.328352,3,232814.573362,232737.430104,63.254730,148.669820
+base,1920x1080,3,3,10,10.000000,25.000000,1,3,10,779.245551,767.521729,7.909994,783.766174,771.945557,8.029191,2.661035,3,232814.573362,232737.430104,63.254730,297.045957
+base,1920x1080,3,3,10,10.000000,25.000000,2,3,10,716.594922,705.615356,9.801939,721.346332,710.146057,9.947215,2.893685,3,232814.573362,232737.430104,63.254730,322.750062
+base,3840x2160,3,3,10,10.000000,25.000000,0,3,10,6157.024951,6149.317383,6.053048,6176.088037,6167.896973,6.139685,1.347144,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,1,3,10,3100.126782,3062.085693,20.516572,3118.679126,3080.606689,20.576561,2.675503,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,2,3,10,2869.991064,2823.746338,42.442398,2888.718970,2843.283447,42.347571,2.890044,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/benchmark_matrix.csv b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/benchmark_matrix.csv
new file mode 100644
index 00000000..a48c32b0
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/benchmark_matrix.csv
@@ -0,0 +1,49 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+small,1920x1080,1,2,7,10.000000,25.000000,0,1,3,110.075401,109.807373,0.203227,113.023483,111.548820,1.281020,18.837996,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,1,2,7,10.000000,25.000000,1,1,3,48.315779,48.225536,0.063975,50.079185,50.001888,0.080786,42.917656,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,1,2,7,10.000000,25.000000,2,1,3,48.153309,47.902267,0.177689,49.923332,49.908306,0.015528,43.062461,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,0,1,3,412.247742,411.911560,0.293113,414.234792,413.839905,0.520222,5.029985,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,1,1,3,178.838618,178.829086,0.006773,180.647334,180.593277,0.039465,11.594811,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,2,1,3,178.384206,178.356384,0.035730,180.499201,180.448990,0.037843,11.624347,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,0,1,3,411.884654,410.574799,0.985028,414.237559,413.042664,0.915347,5.034419,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,1,1,3,178.830358,178.824203,0.004377,180.610469,180.500824,0.138085,11.595347,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,2,1,3,178.352158,178.349945,0.003118,180.530909,180.481705,0.037993,11.626436,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,0,1,3,1261.722778,1258.562256,2.235274,1263.971151,1260.607788,2.403835,1.643467,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,1,1,3,614.759135,614.263611,0.356902,616.952271,616.330139,0.561860,3.373028,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,2,1,3,614.330627,613.909180,0.501169,616.509847,615.886047,0.484838,3.375381,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,0,1,3,407.255351,406.792450,0.328558,413.692769,411.352966,3.048501,5.091646,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,1,1,3,151.200958,150.836227,0.367675,155.520564,155.134949,0.519185,13.714199,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,2,1,3,144.948583,144.930161,0.018506,149.895325,149.814972,0.096745,14.305762,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,0,1,3,1559.592651,1558.821899,0.830700,1564.879272,1563.836792,1.020669,1.329578,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,1,1,3,787.504435,781.217957,4.581173,792.165568,785.335754,4.962077,2.633128,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,2,1,3,727.279460,712.251099,11.368889,732.490682,716.997437,11.708834,2.851174,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,0,1,3,1561.503215,1560.492310,1.289256,1569.216024,1565.276611,3.779472,1.327951,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,1,1,3,780.457438,773.256958,5.091541,785.600016,778.657166,4.911380,2.656903,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,2,1,3,724.909058,722.450439,3.413207,730.018636,727.441895,3.541073,2.860497,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,0,1,3,4888.372884,4879.338379,6.595452,4893.556152,4884.440918,6.615305,0.424190,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,1,1,3,3756.251790,3755.744873,0.431653,3761.770264,3761.220703,0.427554,0.552040,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,2,1,3,3756.514648,3754.959229,1.162039,3762.661214,3761.162354,1.088797,0.552001,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,0,1,3,442.792104,441.429047,1.032451,452.033285,447.258667,4.430310,18.732041,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,1,1,3,190.896907,190.875732,0.018400,196.865921,196.729996,0.106085,43.449630,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,2,1,3,190.784963,190.715744,0.051305,196.747615,196.683945,0.079528,43.475124,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,0,1,3,1681.082642,1676.939575,5.605624,1687.454102,1683.334961,5.641976,4.933963,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,1,1,3,710.423726,710.411377,0.012537,716.396118,715.987061,0.384389,11.675286,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,2,1,3,709.839233,709.789062,0.047672,715.694194,715.556519,0.140165,11.684899,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,0,1,3,1683.890259,1671.197021,15.248118,1690.323934,1677.393799,15.323472,4.925737,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,1,1,3,710.400309,710.378479,0.017140,717.004903,716.504028,0.354187,11.675671,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,2,1,3,709.984456,709.878357,0.075492,715.909139,715.730896,0.216671,11.682509,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,0,1,3,5219.344076,5218.896973,0.425018,5226.740885,5226.617188,0.156331,1.589165,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,1,1,3,2437.269450,2434.812988,1.736991,2443.931152,2442.047119,1.334822,3.403153,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,2,1,3,2431.300700,2427.561523,2.715137,2438.108643,2434.489990,2.704408,3.411507,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,0,1,3,1593.497355,1591.744019,2.063988,1621.089966,1615.066040,6.659392,5.205155,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,1,1,3,600.433980,599.606018,0.657078,618.827861,618.012634,0.619045,13.814008,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,2,1,3,574.664368,574.621216,0.032348,593.176941,592.515747,0.472129,14.433468,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,0,1,3,6158.861654,6157.038086,1.838196,6178.013184,6175.827637,2.118879,1.346742,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,1,1,3,3103.822266,3062.217285,32.499737,3122.170085,3081.743164,31.455323,2.672318,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,2,1,3,2806.920736,2731.674072,66.356029,2825.245687,2750.477539,65.298297,2.954982,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,0,1,3,6161.010905,6155.158203,8.175672,6180.084798,6174.314941,7.587828,1.346273,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,1,1,3,3071.419189,3060.039062,16.064246,3089.492188,3076.310303,16.311514,2.700511,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,2,1,3,2861.546794,2796.239502,46.179360,2880.702148,2818.173340,44.221587,2.898572,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,0,1,3,19406.244141,19377.421875,33.134622,19426.278646,19396.865234,33.693630,0.427409,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,1,1,3,14899.965169,14893.223633,5.363872,14920.315755,14913.674805,5.284395,0.556672,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,2,1,3,14896.569010,14887.253906,7.416163,14921.927734,14914.188477,7.629688,0.556799,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/cpu_baseline_repeat3.csv b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/cpu_baseline_repeat3.csv
new file mode 100644
index 00000000..08100d43
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/cpu_baseline_repeat3.csv
@@ -0,0 +1,2 @@
+size,channels,config,pr,sr,h,sigma,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms
+1920x1080,3,base,3,10,10.000000,25.000000,3,232814.573362,232737.430104,63.254730
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/environment.txt b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/environment.txt
new file mode 100644
index 00000000..ad75a57b
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/environment.txt
@@ -0,0 +1,69 @@
+collected_at_utc=2026-09-15T16:25:21Z
+hostname=5f5ee6a151dd
+os=Linux 5f5ee6a151dd 5.4.0-148-generic #165-Ubuntu SMP Tue Apr 18 08:53:12 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux
+NAME="Ubuntu"
+VERSION_ID="24.04"
+VERSION="24.04.4 LTS (Noble Numbat)"
+
+[gpu]
+Timestamp    Wed Sep 16 00:25:21 2026
++--------------------------------------------------------------------------------+
+|  IX-ML: 4.4.0            Driver Version: 4.4.0          CUDA Version: 10.2     |
+|----------------------------------+----------------------+----------------------|
+| GPU  Name                        | Bus-Id               | Clock-SM  Clock-Mem  |
+| Fan  Temp  Perf    Pwr:Usage/Cap |      Memory-Usage    | GPU-Util  Compute M. |
+|==================================+======================+======================|
+| 0    Iluvatar MR-V100            | 00000000:15:00.0     | 1500MHz   1600MHz    |
+| N/A  35C   P0       38W / 150W   | 68MiB / 32768MiB     | 0%        Default    |
++----------------------------------+----------------------+----------------------+
+
++--------------------------------------------------------------------------------+
+| Processes:                                                          GPU Memory |
+|  GPU        PID      Process name                                   Usage(MiB) |
+|================================================================================|
+|  No running processes found                                                    |
++--------------------------------------------------------------------------------+
+
+[cpu]
+nproc=112
+Architecture:                    x86_64
+CPU(s):                          112
+Model name:                      Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz
+Thread(s) per core:              2
+Core(s) per socket:              28
+Socket(s):                       2
+
+[accelerator-toolkit]
+corex-clang++=/usr/local/corex/bin/clang++
+clang version 18.1.8 (4.4.0 862f87fe94c5eaa2928bda965ed6abb85a25eb7c)
+Target: x86_64-unknown-linux-gnu
+Thread model: posix
+InstalledDir: /usr/local/corex/bin
+corex-nvcc=/usr/local/corex/bin/nvcc
+nvcc: NVIDIA (R) Cuda compiler driver
+Copyright (c) 2005-2019 NVIDIA Corporation
+Built on Wed_Oct_23_19:24:38_PDT_2019
+Cuda compilation tools, release 10.2, V10.2.89
+COREX_HOME=/usr/local/corex
+COREX_HOME_REALPATH=/usr/local/corex-4.4.0
+
+[compiler]
+g++ (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
+GNU Make 4.3
+
+[profilers]
+nsys=not-found
+ncu=not-found
+muprof=not-found
+mxprof=not-found
+ixprof=not-found
+mcTracer=not-found
+
+[build-contract]
+PLATFORM=iluvatar
+COREX_HOME=/usr/local/corex
+COREX_CXX=/usr/local/corex/bin/clang++
+FAST_EXP=0
+WITH_OPENCV=0
+git_commit=unavailable
+git_worktree=unavailable (uploaded source snapshot)
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/logs/validate_1920x1080_rgb.log b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/logs/validate_1920x1080_rgb.log
new file mode 100644
index 00000000..09fe83a7
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/logs/validate_1920x1080_rgb.log
@@ -0,0 +1,8 @@
+[validate] 图像 1920x1080x3，参数 pr=3 sr=10 h=10.0 sigma=25.0
+[validate] CPU 参考耗时: 232556.7 ms
+ver    kernel(ms)   e2e(ms)      MAE(vsCPU)   PSNR(dB)   判定
+0      1561.551     1601.035     0.0000       inf        PASS
+1      792.049      796.391      0.0000       inf        PASS
+2      712.314      716.229      0.0000       inf        PASS
+[validate] 已保存 GPU(v2) 输出: experiments/results/current/iluvatar_mrv100_corex4.4.0/denoised_1920x1080_base_v2.png
+[validate] 总体判定: PASS
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/quality_sigma.csv b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/quality_sigma.csv
new file mode 100644
index 00000000..2979cfd5
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/quality_sigma.csv
@@ -0,0 +1,7 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma10,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/noisy_1920x1080_3ch_sigma10.png,1920,1080,3,4.997221,32.883942
+denoised-sigma10,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/denoised_base_1920x1080_3ch_sigma10.png,1920,1080,3,0.304271,53.265898
+noisy-sigma25,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/noisy_1920x1080_3ch_sigma25.png,1920,1080,3,12.485684,24.947823
+denoised-sigma25,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/denoised_base_1920x1080_3ch_sigma25.png,1920,1080,3,0.597227,49.414889
+noisy-sigma50,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/noisy_1920x1080_3ch_sigma50.png,1920,1080,3,24.883010,18.960290
+denoised-sigma50,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/denoised_base_1920x1080_3ch_sigma50.png,1920,1080,3,1.190769,43.964257
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/quality_tradeoff.csv b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/quality_tradeoff.csv
new file mode 100644
index 00000000..1d13aca3
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/quality_tradeoff.csv
@@ -0,0 +1,5 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma25,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/noisy_1920x1080_3ch_sigma25.png,1920,1080,3,12.485684,24.947823
+small,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/denoised_small_1920x1080_3ch_sigma25.png,1920,1080,3,0.803424,47.358490
+base,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/denoised_base_1920x1080_3ch_sigma25.png,1920,1080,3,0.597227,49.414889
+large,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/clean_1920x1080_3ch.png,experiments/results/current/iluvatar_mrv100_corex4.4.0/quality_images/denoised_large_1920x1080_3ch_sigma25.png,1920,1080,3,0.463451,50.948416
diff --git a/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/source_snapshot.txt b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/source_snapshot.txt
new file mode 100644
index 00000000..2bf7d91b
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/iluvatar_mrv100_corex4.4.0/source_snapshot.txt
@@ -0,0 +1,5 @@
+source snapshot and applied patch archives:
+7272d32769d40a63a965f322b4d429a4965d7f931adb136ac69fb7107269e5b6  ../source.tar.gz
+1b8ea9dab37c01aea009b7eb4a0819e2630668531ee44243424bdcd7c80578b7  ../patch.tar.gz
+ec5e2ecfe7c45a0eeef0668aa1e99da1c12ec1eff19c59874cefca23ae48c9bf  ../patch_v2.tar.gz
+tested source control note: uploaded snapshot of local uncommitted worktree; git metadata intentionally excluded
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_base.csv b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_base.csv
new file mode 100644
index 00000000..4501fc69
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_base.csv
@@ -0,0 +1,7 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+base,1920x1080,3,3,10,10.000000,25.000000,0,3,10,795.647772,794.322937,1.090644,798.188031,797.002625,1.025455,2.606178,1,179756.446891,179756.446891,0.000000,225.205641
+base,1920x1080,3,3,10,10.000000,25.000000,1,3,10,418.420218,418.399475,0.012818,420.728488,420.557281,0.122575,4.955783,1,179756.446891,179756.446891,0.000000,427.250476
+base,1920x1080,3,3,10,10.000000,25.000000,2,3,10,1486.953772,1485.741821,0.654279,1489.554797,1488.348511,0.650486,1.394529,1,179756.446891,179756.446891,0.000000,120.677968
+base,3840x2160,3,3,10,10.000000,25.000000,0,3,10,3114.038403,3112.780273,0.895396,3123.607959,3122.292236,0.814412,2.663551,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,1,3,10,1645.635645,1645.616333,0.017891,1655.068079,1654.860596,0.106856,5.040241,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,2,3,10,5894.478613,5893.784180,0.499588,5904.009033,5903.217773,0.595635,1.407147,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_base_gray.csv b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_base_gray.csv
new file mode 100644
index 00000000..407d4a47
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_base_gray.csv
@@ -0,0 +1,7 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+base,1920x1080,1,3,10,10.000000,25.000000,0,1,3,428.420349,428.408844,0.011594,429.623301,429.192902,0.583016,4.840106,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,1,1,3,159.979950,159.977219,0.002423,161.236781,160.783890,0.325449,12.961624,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,2,1,3,1365.514852,1364.965576,0.643406,1367.177572,1366.687988,0.643214,1.518548,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,0,1,3,1677.981689,1677.955322,0.024405,1682.484701,1680.829590,1.182699,4.943081,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,1,1,3,624.214966,624.213013,0.001795,626.646016,626.615173,0.028304,13.287730,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,2,1,3,5422.634928,5421.552734,0.962776,5427.650716,5424.878906,2.103560,1.529588,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_matrix.csv b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_matrix.csv
new file mode 100644
index 00000000..87d301eb
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/benchmark_matrix.csv
@@ -0,0 +1,37 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+small,1920x1080,1,2,7,10.000000,25.000000,0,1,3,124.490667,124.484863,0.004152,125.600766,125.335663,0.366963,16.656670,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,1,2,7,10.000000,25.000000,1,1,3,45.472341,45.470463,0.001975,46.286012,46.278595,0.005457,45.601347,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,1,2,7,10.000000,25.000000,2,1,3,64.742401,64.704002,0.028946,65.553449,65.522598,0.031297,32.028469,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,0,1,3,428.705963,428.697357,0.011466,430.039174,429.638428,0.283388,4.836882,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,1,1,3,159.982249,159.981049,0.001190,160.805796,160.770279,0.034008,12.961438,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,2,1,3,1366.574056,1365.123291,1.322678,1368.271525,1366.753784,1.383191,1.517371,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,0,1,3,1270.217896,1270.018311,0.141966,1271.494670,1271.118774,0.361480,1.632476,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,1,1,3,482.296570,482.288391,0.005783,483.634755,483.200775,0.306876,4.299429,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,2,1,3,482.289500,482.283020,0.005438,483.187195,483.185394,0.001470,4.299492,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,0,1,3,208.371114,207.775742,0.424419,211.430145,210.323837,1.106714,9.951475,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,1,1,3,71.794261,71.793152,0.000791,74.213326,74.195648,0.023532,28.882532,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,2,1,3,105.359019,105.285889,0.061302,107.776677,107.711121,0.055193,19.681277,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,0,1,3,795.318441,794.650879,0.842124,798.024984,797.294739,0.878254,2.607258,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,1,1,3,418.403066,418.388489,0.011509,420.740336,420.633484,0.075584,4.955987,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,2,1,3,1487.943766,1487.868408,0.077818,1491.377360,1490.647095,0.796123,1.393601,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,0,1,3,2505.419189,2502.314209,2.701461,2509.017415,2506.995850,2.233248,0.827646,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,1,1,3,2356.978597,2356.744629,0.294875,2359.896077,2359.692139,0.262277,0.879770,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,2,1,3,2356.567790,2355.624512,0.835182,2359.574870,2358.506836,0.990491,0.879924,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,0,1,3,487.528697,487.515381,0.012591,491.373088,490.302063,1.348502,17.013152,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,1,1,3,177.408340,177.406204,0.001542,180.040700,179.944458,0.110239,46.753157,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,2,1,3,254.696452,254.535431,0.146169,257.309550,257.130157,0.152726,32.565825,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,0,1,3,1678.633464,1678.570190,0.046173,1682.997192,1682.088867,1.035282,4.941162,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,1,1,3,624.225281,624.218628,0.005901,627.466817,627.442078,0.025129,13.287511,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,2,1,3,5422.730794,5422.113281,0.585132,5427.406738,5425.670898,1.293725,1.529561,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,0,1,3,4973.012044,4972.897461,0.085565,4976.503581,4976.351562,0.121676,1.667883,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,1,1,3,1881.727743,1881.723877,0.003140,1885.173462,1884.999268,0.190840,4.407864,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,2,1,3,1881.728841,1881.723877,0.004547,1885.123332,1885.072021,0.052918,4.407861,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,0,1,3,811.838704,811.763428,0.053278,822.923177,820.263062,3.746618,10.216808,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,1,1,3,280.384257,280.364288,0.024357,288.247416,288.217041,0.023882,29.582260,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,2,1,3,414.482351,414.349823,0.104524,422.416331,422.272125,0.138796,20.011467,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,0,1,3,3112.634115,3111.530762,0.784562,3121.310872,3120.247803,0.763634,2.664753,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,1,1,3,1645.637044,1645.609497,0.030196,1656.506795,1653.906372,2.924078,5.040237,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,2,1,3,5898.279460,5897.552734,0.649538,5907.088704,5906.401367,0.622719,1.406241,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,0,1,3,9806.705404,9803.552734,2.484112,9819.252279,9818.011719,0.945210,0.845789,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,1,1,3,9310.156901,9309.583008,0.502887,9319.997721,9319.424805,0.509252,0.890898,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,2,1,3,9310.460286,9309.819336,0.460961,9320.696615,9319.684570,0.805056,0.890869,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/cpu_baseline_repeat3.csv b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/cpu_baseline_repeat3.csv
new file mode 100644
index 00000000..03f7909f
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/cpu_baseline_repeat3.csv
@@ -0,0 +1,4 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+base,1920x1080,3,3,10,10.000000,25.000000,0,3,10,795.539520,793.667358,1.128572,797.893823,795.942688,1.161652,2.606533,3,179598.340789,179467.824297,100.088485,225.090527
+base,1920x1080,3,3,10,10.000000,25.000000,1,3,10,418.396954,418.379517,0.013650,420.732855,420.653534,0.049251,4.956059,3,179598.340789,179467.824297,100.088485,426.870254
+base,1920x1080,3,3,10,10.000000,25.000000,2,3,10,1488.151355,1486.244629,1.160960,1490.697253,1489.130981,1.160322,1.393407,3,179598.340789,179467.824297,100.088485,120.479420
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/environment.txt b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/environment.txt
new file mode 100644
index 00000000..e3e6a732
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/environment.txt
@@ -0,0 +1,77 @@
+collected_at_utc=2026-09-15T15:06:40Z
+hostname=5bd71ded5feb
+os=Linux 5bd71ded5feb 5.15.0-58-generic #64~20.04.1-Ubuntu SMP Fri Jan 6 16:42:31 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux
+NAME="Ubuntu"
+VERSION_ID="22.04"
+VERSION="22.04.3 LTS (Jammy Jellyfish)"
+
+[gpu]
+
+mx-smi  version: 2.2.6
+
+mx-smi  version: 2.2.6
+
+=================== MetaX System Management Interface Log ===================
+Timestamp                                         : Tue Sep 15 23:06:40 2026
+
+Attached GPUs                                     : 1
++---------------------------------------------------------------------------------+
+| MX-SMI 2.2.6                        Kernel Mode Driver Version: 3.8.30          |
+| MACA Version: 3.0.0.8               BIOS Version: 1.33.5.0                      |
+|------------------------------------+---------------------+----------------------+
+| GPU     NAME         Persistence-M | Bus-id              | GPU-Util      sGPU-M |
+| Temp    Pwr:Usage/Cap         Perf | Memory-Usage        | GPU-State            |
+|====================================+=====================+======================|
+| 0       MetaX C500             Off | 0000:0f:00.0        | 0%             *sGPU |
+| 35C     45W / 350W              P0 | 826/65536 MiB       | Available            |
++------------------------------------+---------------------+----------------------+
+
++---------------------------------------------------------------------------------+
+| Sliced GPU                                                                      |
+|------------------------------------+---------------------+----------------------+
+| Minor    GPU    sGPU-Id    Compute | Vram Quota          | sGPU-Util            |
+|====================================+=====================+======================|
+| 005      0      2          25%     | 0/16000 MiB         | 0%                   |
++------------------------------------+---------------------+----------------------+
+
++---------------------------------------------------------------------------------+
+| Process:                                                                        |
+|  GPU                    PID         Process Name                 GPU Memory     |
+|                                                                  Usage(MiB)     |
+|=================================================================================|
+|  no process found                                                               |
++---------------------------------------------------------------------------------+
+
+End of Log
+
+[accelerator-toolkit]
+mxcc=/opt/maca/mxgpu_llvm/bin/mxcc
+mxcc version 1.0.0 (df29922f9c)
+Target: x86_64-unknown-linux-gnu
+Thread model: posix
+InstalledDir: /opt/maca/mxgpu_llvm/bin
+MACA_HOME=/opt/maca
+MACA_CUDA=/opt/maca/tools/cu-bridge
+MACA_HOME_REALPATH=/opt/maca-3.0.0
+
+[compiler]
+g++ (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
+GNU Make 4.3
+
+[profilers]
+nsys=not-found
+ncu=not-found
+muprof=not-found
+mxprof=not-found
+mcTracer=/opt/maca/bin/mcTracer
+(Tracer): 23:06:40 [INFO ] ***** Version: 
+
+[build-contract]
+PLATFORM=metax
+MACA_HOME=/opt/maca
+MACA_CUDA=/opt/maca/tools/cu-bridge
+MXCC=/opt/maca/mxgpu_llvm/bin/mxcc
+FAST_EXP=0
+WITH_OPENCV=0
+git_commit=unavailable
+git_worktree=unavailable (uploaded source snapshot)
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/mctracer/nlm_1080_v1-3213.json b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/mctracer/nlm_1080_v1-3213.json
new file mode 100644
index 00000000..92e9050b
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/mctracer/nlm_1080_v1-3213.json
@@ -0,0 +1,617 @@
+{
+"traceEvents": [
+{
+	"args":	{
+		"name":	"C500",
+		"version":	"0.0.1"
+	},
+	"cat":	"__metadata",
+	"ph":	"M",
+	"name":	"process_name",
+	"pid":	2
+},{
+	"args":	{
+		"co_id":	1,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645737715365,
+	"dur":	42682,
+	"name":	"mcMalloc",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	2,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645737769142,
+	"dur":	12079,
+	"name":	"mcMalloc",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	3,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645737781848,
+	"dur":	11709,
+	"name":	"mcMalloc",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	4,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645737793992,
+	"dur":	10262,
+	"name":	"mcMalloc",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"bytes":	6220800,
+		"chnl_id":	0,
+		"chnl_type":	0,
+		"co_id":	5,
+		"copy_kind":	1,
+		"device_id":	0,
+		"dst_kind":	3,
+		"flags":	1,
+		"kind":	1,
+		"src_kind":	2,
+		"stream_id":	0
+	},
+	"ph":	"X",
+	"ts":	1789484645747685632,
+	"dur":	6656,
+	"name":	"memcpy HTOD(pinned,device)",
+	"pid":	3,
+	"cat":	"0",
+	"tid":	0
+}
+,{
+	"ph":	"f",
+	"id":	16777221,
+	"ts":	1789484645747685632,
+	"name":	"Launch",
+	"pid":	3,
+	"bp":	"e",
+	"cat":	"Launch",
+	"tid":	0
+}
+,{
+	"args":	{
+		"co_id":	5,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645737823818,
+	"dur":	9942713,
+	"name":	"mcMemcpy",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"ph":	"s",
+	"id":	16777221,
+	"ts":	1789484645737823818,
+	"name":	"Launch",
+	"pid":	1,
+	"cat":	"Launch",
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"block":	{
+			"x":	256,
+			"y":	1,
+			"z":	1
+		},
+		"co_id":	6,
+		"complete_ts":	1789484645749576387,
+		"device_id":	0,
+		"dispatch_id":	0,
+		"grid":	{
+			"x":	24300,
+			"y":	1,
+			"z":	1
+		},
+		"hw_queue_id":	[25335152],
+		"is_dynamic_parallel":	false,
+		"is_recompiled":	false,
+		"launch_type":	0,
+		"max_block_size":	512,
+		"mem":	{
+			"dynamic_shared":	0,
+			"private_per_thread":	0,
+			"private_total":	0,
+			"registers_per_thread":	8,
+			"static_shared":	0
+		},
+		"mtreg_occupancy(%)":	1,
+		"name":	"_ZN3nlm22CvtU8ToF32PlanarKernelEPKhPfiii",
+		"queue_id":	0,
+		"queue_ts":	1789484645749212113,
+		"shared_memeory_occupancy(%)":	0,
+		"submit_ts":	1789484645749221012
+	},
+	"ph":	"X",
+	"ts":	1789484645749271040,
+	"dur":	235520,
+	"name":	"nlm::CvtU8ToF32PlanarKernel(unsigned char const*, float*, int, int, int)",
+	"pid":	2,
+	"cat":	"0",
+	"tid":	1024
+}
+,{
+	"ph":	"f",
+	"id":	16777222,
+	"ts":	1789484645749271040,
+	"name":	"Launch",
+	"pid":	2,
+	"bp":	"e",
+	"cat":	"Launch",
+	"tid":	1024
+}
+,{
+	"args":	{
+		"co_id":	6,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645747782481,
+	"dur":	1434368,
+	"name":	"mcLaunchKernel",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"ph":	"s",
+	"id":	16777222,
+	"ts":	1789484645747782481,
+	"name":	"Launch",
+	"pid":	1,
+	"cat":	"Launch",
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	7,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749224312,
+	"dur":	461,
+	"name":	"mcGetDevice",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	8,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749229400,
+	"dur":	798,
+	"name":	"mcGetDeviceProperties",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	9,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749240043,
+	"dur":	413,
+	"name":	"mcGetDeviceCount",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	10,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749244723,
+	"dur":	2243,
+	"name":	"mcEventCreate",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	11,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749247557,
+	"dur":	403,
+	"name":	"mcEventCreate",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	12,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749258960,
+	"dur":	3763,
+	"name":	"mcEventRecord",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"block":	{
+			"x":	16,
+			"y":	16,
+			"z":	1
+		},
+		"co_id":	13,
+		"complete_ts":	1789484646171908751,
+		"device_id":	0,
+		"dispatch_id":	0,
+		"grid":	{
+			"x":	120,
+			"y":	68,
+			"z":	1
+		},
+		"hw_queue_id":	[25335152],
+		"is_dynamic_parallel":	false,
+		"is_recompiled":	false,
+		"launch_type":	0,
+		"max_block_size":	512,
+		"mem":	{
+			"dynamic_shared":	21672,
+			"private_per_thread":	0,
+			"private_total":	0,
+			"registers_per_thread":	83,
+			"static_shared":	0
+		},
+		"mtreg_occupancy(%)":	16,
+		"name":	"_ZN3nlm13NlmSmemKernelEPKfPfiii12NlmParamsDev",
+		"queue_id":	0,
+		"queue_ts":	1789484645749294841,
+		"shared_memeory_occupancy(%)":	33,
+		"submit_ts":	1789484645749582668
+	},
+	"ph":	"X",
+	"ts":	1789484645749598208,
+	"dur":	422226432,
+	"name":	"nlm::NlmSmemKernel(float const*, float*, int, int, int, NlmParamsDev)",
+	"pid":	2,
+	"cat":	"0",
+	"tid":	1024
+}
+,{
+	"ph":	"f",
+	"id":	16777229,
+	"ts":	1789484645749598208,
+	"name":	"Launch",
+	"pid":	2,
+	"bp":	"e",
+	"cat":	"Launch",
+	"tid":	1024
+}
+,{
+	"args":	{
+		"co_id":	13,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749263929,
+	"dur":	31687,
+	"name":	"mcLaunchKernel",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"ph":	"s",
+	"id":	16777229,
+	"ts":	1789484645749263929,
+	"name":	"Launch",
+	"pid":	1,
+	"cat":	"Launch",
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	14,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749296224,
+	"dur":	1070,
+	"name":	"mcEventRecord",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	15,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484645749306277,
+	"dur":	422608911,
+	"name":	"mcEventSynchronize",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	16,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646171923468,
+	"dur":	2466,
+	"name":	"mcEventElapsedTime",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	17,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646171939093,
+	"dur":	2566,
+	"name":	"mcEventDestroy",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	18,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646171942440,
+	"dur":	484,
+	"name":	"mcEventDestroy",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"block":	{
+			"x":	256,
+			"y":	1,
+			"z":	1
+		},
+		"co_id":	19,
+		"complete_ts":	1789484646175530794,
+		"device_id":	0,
+		"dispatch_id":	0,
+		"grid":	{
+			"x":	24300,
+			"y":	1,
+			"z":	1
+		},
+		"hw_queue_id":	[25335152],
+		"is_dynamic_parallel":	false,
+		"is_recompiled":	false,
+		"launch_type":	0,
+		"max_block_size":	512,
+		"mem":	{
+			"dynamic_shared":	0,
+			"private_per_thread":	0,
+			"private_total":	0,
+			"registers_per_thread":	8,
+			"static_shared":	0
+		},
+		"mtreg_occupancy(%)":	1,
+		"name":	"_ZN3nlm27CvtF32ToU8InterleavedKernelEPKfPhiii",
+		"queue_id":	0,
+		"queue_ts":	1789484646171996452,
+		"shared_memeory_occupancy(%)":	0,
+		"submit_ts":	1789484646171998553
+	},
+	"ph":	"X",
+	"ts":	1789484646172017152,
+	"dur":	549376,
+	"name":	"nlm::CvtF32ToU8InterleavedKernel(float const*, unsigned char*, int, int, int)",
+	"pid":	2,
+	"cat":	"0",
+	"tid":	1024
+}
+,{
+	"ph":	"f",
+	"id":	16777235,
+	"ts":	1789484646172017152,
+	"name":	"Launch",
+	"pid":	2,
+	"bp":	"e",
+	"cat":	"Launch",
+	"tid":	1024
+}
+,{
+	"args":	{
+		"co_id":	19,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646171944111,
+	"dur":	54228,
+	"name":	"mcLaunchKernel",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"ph":	"s",
+	"id":	16777235,
+	"ts":	1789484646171944111,
+	"name":	"Launch",
+	"pid":	1,
+	"cat":	"Launch",
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"bytes":	6220800,
+		"chnl_id":	0,
+		"chnl_type":	0,
+		"co_id":	20,
+		"copy_kind":	2,
+		"device_id":	0,
+		"dst_kind":	1,
+		"flags":	1,
+		"kind":	1,
+		"src_kind":	3,
+		"stream_id":	0
+	},
+	"ph":	"X",
+	"ts":	1789484646174692096,
+	"dur":	833024,
+	"name":	"memcpy DTOH(device,pageable)",
+	"pid":	3,
+	"cat":	"0",
+	"tid":	0
+}
+,{
+	"ph":	"f",
+	"id":	16777236,
+	"ts":	1789484646174692096,
+	"name":	"Launch",
+	"pid":	3,
+	"bp":	"e",
+	"cat":	"Launch",
+	"tid":	0
+}
+,{
+	"args":	{
+		"co_id":	20,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646174665434,
+	"dur":	880394,
+	"name":	"mcMemcpy",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"ph":	"s",
+	"id":	16777236,
+	"ts":	1789484646174665434,
+	"name":	"Launch",
+	"pid":	1,
+	"cat":	"Launch",
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	21,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646175554686,
+	"dur":	416555,
+	"name":	"mcFree",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	22,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646175972661,
+	"dur":	11640,
+	"name":	"mcFree",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	23,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646175985221,
+	"dur":	8773,
+	"name":	"mcFree",
+	"pid":	1,
+	"tid":	1068141760
+}
+,{
+	"args":	{
+		"co_id":	24,
+		"process_id":	3213,
+		"ret":	0,
+		"thread_id":	1068141760
+	},
+	"ph":	"X",
+	"ts":	1789484646175994579,
+	"dur":	8322,
+	"name":	"mcFree",
+	"pid":	1,
+	"tid":	1068141760
+}
+]
+}
\ No newline at end of file
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/quality_sigma.csv b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/quality_sigma.csv
new file mode 100644
index 00000000..249ff177
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/quality_sigma.csv
@@ -0,0 +1,7 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma10,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/noisy_1920x1080_3ch_sigma10.png,1920,1080,3,4.997221,32.883942
+denoised-sigma10,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/denoised_base_1920x1080_3ch_sigma10.png,1920,1080,3,0.304271,53.265898
+noisy-sigma25,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/noisy_1920x1080_3ch_sigma25.png,1920,1080,3,12.485684,24.947823
+denoised-sigma25,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/denoised_base_1920x1080_3ch_sigma25.png,1920,1080,3,0.597227,49.414889
+noisy-sigma50,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/noisy_1920x1080_3ch_sigma50.png,1920,1080,3,24.883010,18.960290
+denoised-sigma50,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/denoised_base_1920x1080_3ch_sigma50.png,1920,1080,3,1.190769,43.964257
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/quality_tradeoff.csv b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/quality_tradeoff.csv
new file mode 100644
index 00000000..fe8fbbdb
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/quality_tradeoff.csv
@@ -0,0 +1,5 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma25,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/noisy_1920x1080_3ch_sigma25.png,1920,1080,3,12.485684,24.947823
+small,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/denoised_small_1920x1080_3ch_sigma25.png,1920,1080,3,0.803424,47.358490
+base,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/denoised_base_1920x1080_3ch_sigma25.png,1920,1080,3,0.597227,49.414889
+large,experiments/results/current/metax_c500/quality_images/clean_1920x1080_3ch.png,experiments/results/current/metax_c500/quality_images/denoised_large_1920x1080_3ch_sigma25.png,1920,1080,3,0.463451,50.948416
diff --git a/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/validate_1920x1080_rgb.log b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/validate_1920x1080_rgb.log
new file mode 100644
index 00000000..c6200fa3
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/metax_c500_maca3.0/validate_1920x1080_rgb.log
@@ -0,0 +1,8 @@
+[validate] 图像 1920x1080x3，参数 pr=3 sr=10 h=10.0 sigma=25.0
+[validate] CPU 参考耗时: 179382.8 ms
+ver    kernel(ms)   e2e(ms)      MAE(vsCPU)   PSNR(dB)   判定
+0      798.987      824.309      0.0000       inf        PASS
+1      418.401      422.132      0.0000       inf        PASS
+2      1487.083     1489.690     0.0000       inf        PASS
+[validate] 已保存 GPU(v2) 输出: experiments/results/current/metax_c500/denoised_1920x1080_base_v2.png
+[validate] 总体判定: PASS
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_1080p_rgb_all.csv b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_1080p_rgb_all.csv
new file mode 100644
index 00000000..eb6a2364
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_1080p_rgb_all.csv
@@ -0,0 +1,13 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+small,1920x1080,3,2,7,10.000000,25.000000,0,3,10,274.253793,274.236572,0.009154,277.307324,276.820526,0.722822,7.560880,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,1,3,10,130.402187,130.399399,0.001606,133.138817,132.948471,0.529844,15.901574,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,2,3,10,106.782271,106.780212,0.001414,109.351600,109.337646,0.007513,19.418954,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,0,3,10,1030.038464,1029.910889,0.068615,1033.298572,1032.700928,0.763538,2.013129,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,1,3,10,569.418866,569.048340,0.123897,572.389532,571.735962,0.588812,3.641607,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,2,3,10,467.984213,467.950806,0.021188,471.184372,470.823730,0.250381,4.430919,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,0,3,10,1030.034717,1029.933105,0.081052,1037.431775,1035.752441,1.034724,2.013136,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,1,3,10,569.467346,569.449097,0.011830,571.794067,571.674683,0.055476,3.641297,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,2,3,10,467.943237,467.886871,0.032439,470.513162,470.088501,0.546349,4.431307,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,0,3,10,3217.569775,3217.068115,0.340089,3220.866943,3220.053711,0.671239,0.644462,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,1,3,10,2428.735010,2416.576172,14.748511,2434.028906,2419.438477,13.871443,0.853778,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,2,3,10,2422.753589,2416.565186,12.014790,2426.481226,2419.429932,12.715772,0.855886,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_1080p_rgb_base.csv b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_1080p_rgb_base.csv
new file mode 100644
index 00000000..deec8ebd
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_1080p_rgb_base.csv
@@ -0,0 +1,4 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+base,1920x1080,3,3,10,10.000000,25.000000,0,3,10,1030.027124,1029.960938,0.044574,1033.800415,1032.837036,0.908436,2.013151,1,186736.456402,186736.456402,0.000000,180.631052
+base,1920x1080,3,3,10,10.000000,25.000000,1,3,10,567.906073,566.320374,1.575765,571.938690,569.702942,1.890596,3.651308,1,186736.456402,186736.456402,0.000000,326.497332
+base,1920x1080,3,3,10,10.000000,25.000000,2,3,10,467.984882,467.899536,0.065601,471.323572,470.700806,0.598145,4.430912,1,186736.456402,186736.456402,0.000000,396.195878
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_4k_rgb_base.csv b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_4k_rgb_base.csv
new file mode 100644
index 00000000..e244430a
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_4k_rgb_base.csv
@@ -0,0 +1,4 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+base,3840x2160,3,3,10,10.000000,25.000000,0,3,10,4101.873877,4101.651367,0.129775,4130.489404,4117.537598,7.840829,2.022100,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,1,3,10,2237.455176,2237.411377,0.037148,2261.089990,2250.583740,10.076719,3.707069,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,2,3,10,1842.981885,1842.171265,1.425139,1873.326904,1860.179565,8.183939,4.500533,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_full_matrix_repeat3.csv b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_full_matrix_repeat3.csv
new file mode 100644
index 00000000..658ae99d
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/benchmark_full_matrix_repeat3.csv
@@ -0,0 +1,49 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+small,1920x1080,1,2,7,10.000000,25.000000,0,1,3,129.346898,129.345535,0.001068,130.821121,130.555420,0.368692,16.031308,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,1,2,7,10.000000,25.000000,1,1,3,56.231148,56.229008,0.001803,57.325453,57.272289,0.058644,36.876359,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,1,2,7,10.000000,25.000000,2,1,3,39.690538,39.687019,0.002873,40.768103,40.680580,0.070208,52.244190,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,0,1,3,479.323741,479.317322,0.004867,480.609172,480.468292,0.161676,4.326095,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,1,1,3,204.874705,204.872925,0.001306,206.120575,205.868927,0.353545,10.121308,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,1,3,10,10.000000,25.000000,2,1,3,146.523112,146.517349,0.004285,147.580027,147.495392,0.101640,14.152034,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,0,1,3,479.319773,479.307251,0.009350,480.863241,480.452393,0.404624,4.326131,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,1,1,3,204.875992,204.871429,0.004923,205.959315,205.948380,0.010152,10.121244,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,1,3,10,15.000000,25.000000,2,1,3,146.527629,146.525482,0.001552,147.991857,147.609726,0.378727,14.151597,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,0,1,3,1484.131999,1484.116211,0.012308,1485.840332,1485.494873,0.381111,1.397180,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,1,1,3,631.279907,631.271118,0.009531,632.626750,632.376099,0.351888,3.284755,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,1,4,14,10.000000,25.000000,2,1,3,631.272949,631.271362,0.001604,632.382446,632.355042,0.019809,3.284791,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,0,1,3,274.241364,274.239380,0.002635,277.480967,276.761993,1.000706,7.561223,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,1,1,3,130.403687,130.402573,0.001198,132.897771,132.893692,0.003812,15.901391,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,2,1,3,106.784383,106.781555,0.003556,109.278903,109.264809,0.011890,19.418570,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,0,1,3,1030.051025,1029.957764,0.071263,1035.736450,1032.769897,2.971621,2.013104,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,1,1,3,569.454915,569.449280,0.007042,576.555664,576.150940,0.294630,3.641377,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,2,1,3,467.985341,467.969666,0.013104,480.872843,475.156403,7.719309,4.430908,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,0,1,3,1030.052043,1030.040771,0.008428,1049.653727,1046.366211,4.214041,2.013102,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,1,1,3,569.459208,569.437622,0.016384,576.800822,572.657959,5.783659,3.641349,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,1920x1080,3,3,10,15.000000,25.000000,2,1,3,468.007009,467.978058,0.027955,471.209442,471.098419,0.084458,4.430703,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,0,1,3,3218.469482,3217.895020,0.812068,3222.596191,3221.941895,0.465941,0.644281,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,1,1,3,2416.803467,2416.689941,0.104611,2419.873128,2419.746094,0.162862,0.857993,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,2,1,3,2436.756348,2416.510498,14.317428,2440.094401,2419.832520,14.334395,0.850967,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,0,1,3,514.850505,514.848511,0.002524,519.099813,518.074036,1.373081,16.110308,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,1,1,3,222.907445,222.905411,0.001809,226.185954,226.114120,0.063163,37.210063,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,1,2,7,10.000000,25.000000,2,1,3,157.077784,157.075668,0.001751,160.261276,160.239197,0.027813,52.804412,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,0,1,3,1908.620524,1908.615112,0.003899,1912.280192,1912.180908,0.096140,4.345756,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,1,1,3,812.816162,812.812012,0.003041,816.151245,816.078064,0.052870,10.204521,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,1,3,10,10.000000,25.000000,2,1,3,580.606445,580.603088,0.004619,583.813497,583.795166,0.013774,14.285753,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,0,1,3,1908.627686,1908.621460,0.004665,1912.238647,1912.064209,0.179364,4.345740,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,1,1,3,812.820862,812.819519,0.001647,816.189351,816.158936,0.023715,10.204462,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,1,3,10,15.000000,25.000000,2,1,3,580.607992,580.605713,0.001756,584.958374,583.913940,0.961233,14.285714,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,0,1,3,5910.332357,5910.329102,0.003618,5932.037923,5928.812988,2.434471,1.403373,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,1,1,3,2498.901204,2498.893066,0.006579,2517.193278,2510.129395,5.091973,3.319219,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,1,4,14,10.000000,25.000000,2,1,3,2498.902425,2498.896973,0.005475,2517.673503,2512.231689,5.483564,3.319217,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,0,1,3,1091.943604,1091.927124,0.018526,1139.071777,1120.375854,16.656299,7.595997,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,1,1,3,516.066447,516.064453,0.001671,543.061869,527.822388,18.487320,16.072349,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,3840x2160,3,2,7,10.000000,25.000000,2,1,3,422.355794,422.349182,0.004681,448.068197,439.588593,9.663223,19.638419,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,0,1,3,4102.078288,4101.984375,0.072424,4152.101237,4150.945801,1.183154,2.021999,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,1,1,3,2237.445312,2237.418213,0.032793,2271.912435,2269.194336,1.922216,3.707085,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,3840x2160,3,3,10,10.000000,25.000000,2,1,3,1842.180216,1841.973267,0.146479,1876.934977,1873.409790,2.813918,4.502491,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,0,1,3,4101.948730,4101.866699,0.061871,4151.048503,4128.649414,16.427898,2.022063,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,1,1,3,2237.446452,2237.425781,0.024398,2274.024007,2270.760986,2.387083,3.707083,0,-1.000000,-1.000000,-1.000000,-1.000000
+strong-h,3840x2160,3,3,10,15.000000,25.000000,2,1,3,1842.395549,1842.203247,0.140065,1869.796672,1858.244629,8.257476,4.501965,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,0,1,3,12813.651693,12813.555664,0.119616,12831.928385,12829.979492,2.112187,0.647310,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,1,1,3,9593.831706,9593.754883,0.057084,9613.700521,9612.013672,2.234694,0.864556,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,3840x2160,3,4,14,10.000000,25.000000,2,1,3,9593.986654,9593.896484,0.081418,9611.934896,9607.855469,4.753878,0.864542,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/environment_collector_verified.txt b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/environment_collector_verified.txt
new file mode 100644
index 00000000..d1b69f67
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/environment_collector_verified.txt
@@ -0,0 +1,223 @@
+collected_at_utc=2026-09-15T13:58:05Z
+hostname=afaaa9126062
+os=Linux afaaa9126062 5.15.0-105-generic #115~20.04.1-Ubuntu SMP Mon Apr 15 17:33:04 UTC 2024 x86_64 x86_64 x86_64 GNU/Linux
+NAME="Ubuntu"
+VERSION_ID="22.04"
+VERSION="22.04.5 LTS (Jammy Jellyfish)"
+
+[gpu]
+compiler: mcc
+--------------------------------------------------------------------------------
+device#                          0
+Name:                            MTT S4000
+Driver Version:                  5.1
+Runtime Version:                 5.1
+compute capability:              2.2
+pciBusID:                        0x9
+pciDeviceID:                     0x0
+pciDomainID:                     0x0
+multiProcessorCount:             64
+maxThreadsPerMultiProcessor:     6144
+isMultiGpuBoard:                 1
+clockRate:                       1791 Mhz
+memoryClockRate:                 8006 Mhz
+memoryBusWidth:                  384
+totalGlobalMem:                  47.91 GB
+sharedMemPerMultiprocessor:      72.00 KB
+totalConstMem:                   8192
+sharedMemPerBlock:               72.00 KB
+canMapHostMemory:                1
+regsPerBlock:                    131072
+warpSize:                        128
+l2CacheSize:                     25165824
+computeMode:                     0
+maxThreadsPerBlock:              1024
+maxThreadsDim.x:                 1024
+maxThreadsDim.y:                 1024
+maxThreadsDim.z:                 1024
+maxGridSize.x:                   2147483647
+maxGridSize.y:                   2147483647
+maxGridSize.z:                   2147483647
+concurrentKernels:               1
+cooperativeLaunch:               1
+cooperativeMultiDeviceLaunch:    0
+isIntegrated:                    0
+maxTexture1D:                    32768
+maxTexture2D.width:              32768
+maxTexture2D.height:             32768
+maxTexture3D.width:              16384
+maxTexture3D.height:             16384
+maxTexture3D.depth:              2048
+peers:                           
+non-peers:                       device#0 
+memInfo.total:                   47.91 GB
+memInfo.free:                    45.81 GB (96%)
+
+[accelerator-toolkit]
+mcc=/usr/local/musa/bin/mcc
+clang version 14.0.0 (git@sh-code.mthreads.com:sw/mtcc.git 61c54cd8e5c72ef146656c7beee1889a837ed7bc)
+mcc version 5.1.0
+Target: x86_64-unknown-linux-gnu
+Thread model: posix
+InstalledDir: /usr/local/musa/bin
+musa_toolkits:
+{
+	"version":	"5.1.0",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"AIMODULE_1.4.0",
+	"commit id":	"e1481406739bd545c45c4b852450a81ac5feee6c",
+	"commit date":	"2026-03-06 20:38:51 +0800"
+}
+CUB:
+{
+	"version":	"1.17.2",
+	"git branch":	"develop-1.17",
+	"git tag":	"No tag",
+	"commit id":	"2aa7e2b6a2e82ca339ca132d345377d6faef6024",
+	"commit date":	"2026-04-22 21:48:03 +0800"
+}
+MUPTI:
+{
+	"version":	"1.2.0",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"453615f9e030b4c0bd101ce44d60346a1ffee11b",
+	"commit date":	"2026-04-28 12:40:08 +0800"
+}
+Thrust:
+{
+	"version":	"1.17.2",
+	"git branch":	"develop-1.17",
+	"git tag":	"No tag",
+	"commit id":	"bffd03bde9af75266cc1a51d46cde48f4bcd9280",
+	"commit date":	"2026-04-22 21:51:23 +0800"
+}
+mcc:
+{
+	"version":	"5.1.0",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"61c54cd8e5c72ef146656c7beee1889a837ed7bc",
+	"commit date":	"2026-04-28 15:32:37 +0800"
+}
+mccl:
+{
+	"version":	"2.1.3",
+	"build archs":	"22",
+	"git branch":	"release_musa_5.1.0_s4000",
+	"git tag":	"No tag",
+	"commit id":	"f13ff3006e6f7fa7e70064fc989364661b3ad9d2",
+	"commit date":	"2026-04-15 10:49:18 +0800"
+}
+mtjpeg:
+{
+	"version":	"1.0.5",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"070f3ff97088cfdef12da04b09af22bfc862874c",
+	"commit date":	"2026-03-23 13:40:41 +0800"
+}
+muPP:
+{
+	"version":	"1.13.0",
+	"build archs":	"22;31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"AIMODULE_1.4.0",
+	"commit id":	"0a8b4f2b1c8f9928130fa882a97705be2fc693e2",
+	"commit date":	"2026-03-12 14:48:07 +0800"
+}
+mublasLt:
+{
+	"version":	"1.12.0",
+	"build archs":	"22,31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"AIMODULE_1.4.0",
+	"commit id":	"b99c7919a587a92ec1f24b75a633d18da754377b",
+	"commit date":	"2026-04-07 14:08:36 +0800"
+}
+mublas:
+{
+	"version":	"1.12.0",
+	"build archs":	"22,31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"4bced69bfb393863c43c23e9d21fbce6e2ff9ffd",
+	"commit date":	"2026-04-23 19:16:27 +0800"
+}
+mudnn:
+{
+	"version":	"3.3.0",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"bd105d42d4413d10f51cf91785cf4e06acee30ac",
+	"commit date":	"2026-04-26 17:03:03 +0800"
+}
+mufft:
+{
+	"version":	"1.12.0",
+	"build archs":	"22;31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"AIMODULE_1.4.0",
+	"commit id":	"0a2e9ac103e2d54c384947561076c682c6811c1e",
+	"commit date":	"2026-03-11 19:24:16 +0800"
+}
+murand:
+{
+	"version":	"1.3.0",
+	"build archs":	"22;31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"AIMODULE_1.4.0",
+	"commit id":	"525224ed521b4cb0af25ada59b84583241e57761",
+	"commit date":	"2026-03-27 19:58:21 +0800"
+}
+musify:
+{
+	"version":	"1.2.0",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"a5f79eb9e507dd7bfa1b5bd8b6b833f671c6b8c1",
+	"commit date":	"2026-04-08 14:58:40 +0800"
+}
+musolver:
+{
+	"version":	"1.6.0",
+	"build archs":	"22,31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"f6062eb2da90ad3b5091a91d3348423649275d12",
+	"commit date":	"2026-04-25 19:06:04 +0800"
+}
+musparse:
+{
+	"version":	"1.7.0",
+	"build archs":	"22,31",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"3b5df2ca75c504c98821bc2cd71a9160ddf4eb55",
+	"commit date":	"2026-04-27 17:08:43 +0800"
+}
+musa_runtime:
+{
+	"version":	"5.1.0",
+	"git branch":	"release_musa_5.1.0",
+	"git tag":	"No tag",
+	"commit id":	"73921cd20f278bd905e50dbd920c976d93cc18b4",
+	"commit date":	"2026-04-24 13:42:11 +0800"
+}
+
+[compiler]
+g++ (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0
+GNU Make 4.3
+
+[profilers]
+nsys=not-found
+ncu=not-found
+muprof=not-found
+
+[build-contract]
+PLATFORM=moore
+MUSA_HOME=/usr/local/musa
+FAST_EXP=0
+WITH_OPENCV=0
+git_commit=unavailable
+git_worktree=clean
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/quality_sigma.csv b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/quality_sigma.csv
new file mode 100644
index 00000000..56eca177
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/quality_sigma.csv
@@ -0,0 +1,7 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma10,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/noisy_1920x1080_3ch_sigma10_seed7.png,1920,1080,3,4.997221,32.883942
+denoised-sigma10,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/denoised_base_1920x1080_3ch_sigma10_seed7.png,1920,1080,3,0.304271,53.265898
+noisy-sigma25,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/noisy_1920x1080_3ch_sigma25_seed7.png,1920,1080,3,12.485684,24.947823
+denoised-sigma25,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/denoised_base_1920x1080_3ch_sigma25_seed7.png,1920,1080,3,0.597227,49.414889
+noisy-sigma50,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/noisy_1920x1080_3ch_sigma50_seed7.png,1920,1080,3,24.883010,18.960290
+denoised-sigma50,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/denoised_base_1920x1080_3ch_sigma50_seed7.png,1920,1080,3,1.190769,43.964257
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/quality_tradeoff.csv b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/quality_tradeoff.csv
new file mode 100644
index 00000000..44237c02
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/quality_tradeoff.csv
@@ -0,0 +1,5 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma25,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/noisy_1920x1080_3ch_sigma25_seed7.png,1920,1080,3,12.485684,24.947823
+small,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/denoised_small_1920x1080_3ch_sigma25_seed7.png,1920,1080,3,0.803424,47.358490
+base,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/denoised_base_1920x1080_3ch_sigma25_seed7.png,1920,1080,3,0.597227,49.414889
+large,experiments/results/moore_s4000_musa5.1/quality_images/clean_1920x1080_3ch_seed7.png,experiments/results/moore_s4000_musa5.1/quality_images/denoised_large_1920x1080_3ch_sigma25_seed7.png,1920,1080,3,0.463451,50.948416
diff --git a/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/validate_1920x1080_rgb.log b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/validate_1920x1080_rgb.log
new file mode 100644
index 00000000..f5bdcde3
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/moore_s4000_musa5.1/validate_1920x1080_rgb.log
@@ -0,0 +1,8 @@
+[validate] 图像 1920x1080x3，参数 pr=3 sr=10 h=10.0 sigma=25.0
+[validate] CPU 参考耗时: 187922.0 ms
+ver    kernel(ms)   e2e(ms)      MAE(vsCPU)   PSNR(dB)   判定
+0      1030.048     1109.572     0.0000       inf        PASS
+1      569.469      573.384      0.0000       inf        PASS
+2      468.039      471.062      0.0000       inf        PASS
+[validate] 已保存 GPU(v2) 输出: experiments/results/moore_s4000_musa5.1/denoised_1920x1080_base_v2.png
+[validate] 总体判定: PASS
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/benchmark_legacy.csv b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/benchmark_legacy.csv
new file mode 100644
index 00000000..e6a1f6ea
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/benchmark_legacy.csv
@@ -0,0 +1,49 @@
+size,channels,pr,sr,h,sigma,kernel_ver,kernel_ms,e2e_ms,mpx_s,cpu_ms,speedup
+1920x1080,1,3,10,10,25,0,422.151,426.491,4.91198,-1,-1
+1920x1080,1,3,10,10,25,1,197.33,201.464,10.5083,-1,-1
+1920x1080,1,3,10,10,25,2,152.199,155.809,13.6243,-1,-1
+1920x1080,1,3,10,15,25,0,404.96,421.488,5.1205,-1,-1
+1920x1080,1,3,10,15,25,1,187.304,199.081,11.0708,-1,-1
+1920x1080,1,3,10,15,25,2,148.226,156.576,13.9894,-1,-1
+1920x1080,1,2,7,10,15,0,111.487,119.248,18.5995,-1,-1
+1920x1080,1,2,7,10,15,1,49.8289,57.7302,41.6144,-1,-1
+1920x1080,1,2,7,10,15,2,38.5659,41.934,53.7677,-1,-1
+1920x1080,1,4,14,12,35,0,1286.27,1346.18,1.6121,-1,-1
+1920x1080,1,4,14,12,35,1,598.219,614.698,3.46629,-1,-1
+1920x1080,1,4,14,12,35,2,587.015,627.243,3.53245,-1,-1
+1920x1080,3,3,10,10,25,0,870.892,893.541,2.38101,-1,-1
+1920x1080,3,3,10,10,25,1,365.143,392.385,5.67887,-1,-1
+1920x1080,3,3,10,10,25,2,353.994,374.229,5.85773,-1,-1
+1920x1080,3,3,10,15,25,0,844.082,905.64,2.45663,-1,-1
+1920x1080,3,3,10,15,25,1,397.056,412.219,5.22244,-1,-1
+1920x1080,3,3,10,15,25,2,380.475,397.698,5.45002,-1,-1
+1920x1080,3,2,7,10,15,0,234.539,248.302,8.84117,-1,-1
+1920x1080,3,2,7,10,15,1,104.366,116.672,19.8685,-1,-1
+1920x1080,3,2,7,10,15,2,97.3568,108.617,21.299,-1,-1
+1920x1080,3,4,14,12,35,0,2684.31,2857.94,0.77249,-1,-1
+1920x1080,3,4,14,12,35,1,1504.75,1535.02,1.37804,-1,-1
+1920x1080,3,4,14,12,35,2,1538.15,1549.59,1.34811,-1,-1
+3840x2160,1,3,10,10,25,0,2017.37,2039.27,4.1115,-1,-1
+3840x2160,1,3,10,10,25,1,878.583,900.289,9.44066,-1,-1
+3840x2160,1,3,10,10,25,2,693.923,709.221,11.9529,-1,-1
+3840x2160,1,3,10,15,25,0,1991.97,2011.27,4.16392,-1,-1
+3840x2160,1,3,10,15,25,1,875.186,897.331,9.4773,-1,-1
+3840x2160,1,3,10,15,25,2,715.99,730.313,11.5845,-1,-1
+3840x2160,1,2,7,10,15,0,509.793,565.186,16.2701,-1,-1
+3840x2160,1,2,7,10,15,1,224.844,246.884,36.8896,-1,-1
+3840x2160,1,2,7,10,15,2,165.908,175.714,49.9938,-1,-1
+3840x2160,1,4,14,12,35,0,6134.59,6300.13,1.35207,-1,-1
+3840x2160,1,4,14,12,35,1,2766.1,2799.31,2.99859,-1,-1
+3840x2160,1,4,14,12,35,2,2692.87,2748.51,3.08014,-1,-1
+3840x2160,3,3,10,10,25,0,3771.64,3868.83,2.19915,-1,-1
+3840x2160,3,3,10,10,25,1,1706.93,1730.46,4.85926,-1,-1
+3840x2160,3,3,10,10,25,2,1659.49,1691.04,4.99815,-1,-1
+3840x2160,3,3,10,15,25,0,3807.33,3845.76,2.17854,-1,-1
+3840x2160,3,3,10,15,25,1,1725.52,1746,4.8069,-1,-1
+3840x2160,3,3,10,15,25,2,1670.13,1693.09,4.96631,-1,-1
+3840x2160,3,2,7,10,15,0,1027.49,1055.27,8.07247,-1,-1
+3840x2160,3,2,7,10,15,1,473.802,495.377,17.5061,-1,-1
+3840x2160,3,2,7,10,15,2,441.068,461.316,18.8053,-1,-1
+3840x2160,3,4,14,12,35,0,11574,11860.4,0.716639,-1,-1
+3840x2160,3,4,14,12,35,1,6285.2,6323.47,1.31967,-1,-1
+3840x2160,3,4,14,12,35,2,6316.24,6512.54,1.31319,-1,-1
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/benchmark_quality.csv b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/benchmark_quality.csv
new file mode 100644
index 00000000..0da71a87
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/benchmark_quality.csv
@@ -0,0 +1,10 @@
+config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e
+small,1920x1080,3,2,7,10.000000,25.000000,0,3,5,206.358325,205.456390,1.112540,211.635452,210.481689,1.121096,10.048541,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,1,3,5,96.905011,95.821823,0.958317,102.151357,100.839897,0.868854,21.398274,0,-1.000000,-1.000000,-1.000000,-1.000000
+small,1920x1080,3,2,7,10.000000,25.000000,2,3,5,90.727629,88.535042,1.105170,95.985898,94.068398,0.976344,22.855221,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,0,3,5,773.808130,766.097412,6.025820,778.856470,771.509155,5.892213,2.679734,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,1,3,5,373.257837,372.642822,0.698041,378.362042,377.567810,0.780827,5.555409,0,-1.000000,-1.000000,-1.000000,-1.000000
+base,1920x1080,3,3,10,10.000000,25.000000,2,3,5,352.407758,352.066559,0.409771,357.572998,356.940491,0.604280,5.884093,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,0,3,5,2548.151709,2491.449463,43.053494,2553.959375,2496.928955,43.024549,0.813766,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,1,3,5,1427.287671,1423.611938,4.364124,1432.517676,1428.449463,4.570228,1.452826,0,-1.000000,-1.000000,-1.000000,-1.000000
+large,1920x1080,3,4,14,10.000000,25.000000,2,3,5,1452.325684,1450.172363,2.479933,1457.433716,1455.473755,2.397685,1.427779,0,-1.000000,-1.000000,-1.000000,-1.000000
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/environment_quality.txt b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/environment_quality.txt
new file mode 100644
index 00000000..30de9d0a
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/environment_quality.txt
@@ -0,0 +1,33 @@
+collected_at_utc=2026-09-15T05:56:53Z
+hostname=DESKTOP-9OMGKMK
+os=Linux DESKTOP-9OMGKMK 6.18.33.2-microsoft-standard-WSL2 #1 SMP PREEMPT_DYNAMIC Thu Jun 18 21:54:43 UTC 2026 x86_64 GNU/Linux
+NAME="Ubuntu"
+VERSION_ID="26.04"
+VERSION="26.04 LTS (Resolute Raccoon)"
+
+[gpu]
+NVIDIA GeForce RTX 3060 Laptop GPU, GPU-3a2c7f60-59bc-affa-0466-f486df5eb745, 610.88, 6144 MiB, 8.6
+
+[cuda]
+nvcc=/home/zhoufei/cuda/cuda_nvcc-linux-x86_64-12.9.86-archive/bin/nvcc
+nvcc: NVIDIA (R) Cuda compiler driver
+Copyright (c) 2005-2025 NVIDIA Corporation
+Built on Tue_May_27_02:21:03_PDT_2025
+Cuda compilation tools, release 12.9, V12.9.86
+Build cuda_12.9.r12.9/compiler.36037853_0
+
+[compiler]
+x86_64-conda-linux-gnu-g++ (conda-forge gcc 13.4.0-20) 13.4.0
+GNU Make 4.4.1
+
+[profilers]
+nsys=not-found
+ncu=not-found
+
+[build-contract]
+PLATFORM=nvidia
+ARCH=sm_86
+FAST_EXP=0
+WITH_OPENCV=0
+git_commit=549f45ffa93c1864c88e5b218fd2c50a3d192437
+git_worktree=dirty
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/quality_sigma.csv b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/quality_sigma.csv
new file mode 100644
index 00000000..f710f5be
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/quality_sigma.csv
@@ -0,0 +1,7 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma10,experiments/work/clean_1920x1080_3ch.png,experiments/work/noisy_1920x1080_3ch_sigma10.png,1920,1080,3,4.997221,32.883942
+denoised-sigma10,experiments/work/clean_1920x1080_3ch.png,experiments/work/denoised_base_1920x1080_3ch_sigma10.png,1920,1080,3,0.304271,53.265898
+noisy-sigma25,experiments/work/clean_1920x1080_3ch.png,experiments/work/noisy_1920x1080_3ch_sigma25.png,1920,1080,3,12.485684,24.947823
+denoised-sigma25,experiments/work/clean_1920x1080_3ch.png,experiments/work/denoised_base_1920x1080_3ch_sigma25.png,1920,1080,3,0.597227,49.414889
+noisy-sigma50,experiments/work/clean_1920x1080_3ch.png,experiments/work/noisy_1920x1080_3ch_sigma50.png,1920,1080,3,24.883010,18.960290
+denoised-sigma50,experiments/work/clean_1920x1080_3ch.png,experiments/work/denoised_base_1920x1080_3ch_sigma50.png,1920,1080,3,1.190769,43.964257
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/quality_tradeoff.csv b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/quality_tradeoff.csv
new file mode 100644
index 00000000..d3a73dd7
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/quality_tradeoff.csv
@@ -0,0 +1,5 @@
+label,reference,test,width,height,channels,mae,psnr_db
+noisy-sigma25,experiments/work/clean_1920x1080_3ch.png,experiments/work/noisy_1920x1080_3ch_sigma25.png,1920,1080,3,12.485684,24.947823
+small,experiments/work/clean_1920x1080_3ch.png,experiments/work/denoised_small_1920x1080_3ch_sigma25.png,1920,1080,3,0.803424,47.358490
+base,experiments/work/clean_1920x1080_3ch.png,experiments/work/denoised_base_1920x1080_3ch_sigma25.png,1920,1080,3,0.597227,49.414889
+large,experiments/work/clean_1920x1080_3ch.png,experiments/work/denoised_large_1920x1080_3ch_sigma25.png,1920,1080,3,0.463451,50.948416
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/run_all_legacy.txt b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/run_all_legacy.txt
new file mode 100644
index 00000000..d1aae610
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx3060_laptop/run_all_legacy.txt
@@ -0,0 +1,101 @@
+===============================================================
+ nlm_denoise —— NVIDIA 平台实测结果（nvidia_result.txt）
+ 生成方式: bash scripts/gen_result_wsl.sh（可复现）
+===============================================================
+
+[环境]
+GPU     : NVIDIA GeForce RTX 3060 Laptop GPU
+CUDA    : release 12.9
+OS      : Linux 6.18.33.2-microsoft-standard-WSL2
+参数基线: pr=3 sr=10 h=10.0 sigma=25.0
+
+===============================================================
+[1] make run VERBOSE=true（单元测试 + GPU 端到端校验 256x256）
+===============================================================
+=== Running tests (output from build/nlm_denoise) ===
+=== Verbose mode: Enabled (using '--verbose') ===
+build/nlm_denoise units
+========== NLM 单元测试 ==========
+[units] 参数解析
+  [PASS] 正常解析（含注释/缺省）
+  [PASS] 未知 key 拒绝
+  [PASS] 越界取值拒绝
+[units] 质量指标
+  [PASS] MAE(相同图像)=0
+  [PASS] PSNR(相同图像)=inf
+  [PASS] MAE(恒差10)=10
+  [PASS] PSNR(恒差10)≈28.13dB
+[units] CPU 参考实现
+  [PASS] 常量灰度图恒等
+  [PASS] 常量 RGB 图恒等
+  [PASS] 确定性（两次运行一致）
+  [PASS] 极小图（5x7, sr=10）不崩溃且尺寸正确
+    降噪前后对干净图 MAE: 12.48 -> 1.57
+  [PASS] 降噪有效性（MAE 下降）
+========== 全部通过（失败 0 项）==========
+=== Verbose extra: GPU 端到端正确性校验（256x256 RGB）===
+[gen] 已生成 256x256x3 sigma=25.0 -> /tmp/nlm_make_run.png
+[validate] 图像 256x256x3，参数 pr=3 sr=10 h=10.0 sigma=25.0
+[validate] CPU 参考耗时: 5077.8 ms
+ver    kernel(ms)   e2e(ms)      MAE(vsCPU)   PSNR(dB)   判定
+0      29.435       248.460      0.0000       inf        PASS
+1      13.586       14.708       0.0000       inf        PASS
+2      13.318       14.908       0.0000       inf        PASS
+[validate] 总体判定: PASS
+
+===============================================================
+[2] 正确性校验 validate（512x512 RGB，三版本 vs CPU 参考）
+===============================================================
+[gen] 已生成 512x512x3 sigma=25.0 -> /tmp/nlm_res_512.png
+[validate] 图像 512x512x3，参数 pr=3 sr=10 h=10.0 sigma=25.0
+[validate] CPU 参考耗时: 19981.0 ms
+ver    kernel(ms)   e2e(ms)      MAE(vsCPU)   PSNR(dB)   判定
+0      110.222      335.570      0.0000       inf        PASS
+1      51.163       53.033       0.0000       inf        PASS
+2      44.296       46.678       0.0000       inf        PASS
+[validate] 总体判定: PASS
+
+===============================================================
+[3] 性能基准抽样（640x360，完整 48 组数据见 bench.csv）
+===============================================================
+[bench] warmup=1 repeat=3，输出 /tmp/nlm_res_bench.csv
+[bench] 640x360x1 base     v0: kernel=47.460 ms  e2e=49.395 ms  4.9 Mpx/s
+[bench] 640x360x1 base     v1: kernel=22.047 ms  e2e=23.698 ms  10.5 Mpx/s
+[bench] 640x360x1 base     v2: kernel=17.622 ms  e2e=19.207 ms  13.1 Mpx/s
+[bench] 640x360x1 strong-h v0: kernel=43.867 ms  e2e=48.350 ms  5.3 Mpx/s
+[bench] 640x360x1 strong-h v1: kernel=21.043 ms  e2e=24.679 ms  10.9 Mpx/s
+[bench] 640x360x1 strong-h v2: kernel=17.576 ms  e2e=19.119 ms  13.1 Mpx/s
+[bench] 640x360x1 small    v0: kernel=13.418 ms  e2e=14.850 ms  17.2 Mpx/s
+[bench] 640x360x1 small    v1: kernel=5.364 ms  e2e=6.948 ms  43.0 Mpx/s
+[bench] 640x360x1 small    v2: kernel=3.986 ms  e2e=5.369 ms  57.8 Mpx/s
+[bench] 640x360x1 large    v0: kernel=140.188 ms  e2e=148.394 ms  1.6 Mpx/s
+[bench] 640x360x1 large    v1: kernel=68.777 ms  e2e=70.522 ms  3.3 Mpx/s
+[bench] 640x360x1 large    v2: kernel=65.242 ms  e2e=67.839 ms  3.5 Mpx/s
+[bench] 640x360x3 base     v0: kernel=87.725 ms  e2e=89.905 ms  2.6 Mpx/s
+[bench] 640x360x3 base     v1: kernel=37.180 ms  e2e=42.286 ms  6.2 Mpx/s
+[bench] 640x360x3 base     v2: kernel=38.963 ms  e2e=41.604 ms  5.9 Mpx/s
+[bench] 640x360x3 strong-h v0: kernel=87.507 ms  e2e=89.743 ms  2.6 Mpx/s
+[bench] 640x360x3 strong-h v1: kernel=40.482 ms  e2e=42.578 ms  5.7 Mpx/s
+[bench] 640x360x3 strong-h v2: kernel=37.172 ms  e2e=40.020 ms  6.2 Mpx/s
+[bench] 640x360x3 small    v0: kernel=20.902 ms  e2e=23.873 ms  11.0 Mpx/s
+[bench] 640x360x3 small    v1: kernel=11.015 ms  e2e=12.831 ms  20.9 Mpx/s
+[bench] 640x360x3 small    v2: kernel=10.006 ms  e2e=11.924 ms  23.0 Mpx/s
+[bench] 640x360x3 large    v0: kernel=258.631 ms  e2e=269.029 ms  0.9 Mpx/s
+[bench] 640x360x3 large    v1: kernel=144.836 ms  e2e=153.271 ms  1.6 Mpx/s
+[bench] 640x360x3 large    v2: kernel=153.670 ms  e2e=156.464 ms  1.5 Mpx/s
+[bench] 完成，CSV 已写入 /tmp/nlm_res_bench.csv
+
+----- bench.csv 摘要（RTX 3060 Laptop 实测，warmup=3 repeat=10）-----
+size,channels,pr,sr,h,sigma,kernel_ver,kernel_ms,e2e_ms,mpx_s,cpu_ms,speedup
+1920x1080,1,3,10,10,25,0,422.151,426.491,4.91198,-1,-1
+1920x1080,1,3,10,10,25,1,197.33,201.464,10.5083,-1,-1
+1920x1080,1,3,10,10,25,2,152.199,155.809,13.6243,-1,-1
+1920x1080,3,3,10,10,25,0,870.892,893.541,2.38101,-1,-1
+1920x1080,3,3,10,10,25,1,365.143,392.385,5.67887,-1,-1
+1920x1080,3,3,10,10,25,2,353.994,374.229,5.85773,-1,-1
+3840x2160,1,3,10,10,25,0,2017.37,2039.27,4.1115,-1,-1
+3840x2160,1,3,10,10,25,1,878.583,900.289,9.44066,-1,-1
+3840x2160,1,3,10,10,25,2,693.923,709.221,11.9529,-1,-1
+3840x2160,3,3,10,10,25,0,3771.64,3868.83,2.19915,-1,-1
+3840x2160,3,3,10,10,25,1,1706.93,1730.46,4.85926,-1,-1
+3840x2160,3,3,10,10,25,2,1659.49,1691.04,4.99815,-1,-1
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx4090d/benchmark_legacy.csv b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/benchmark_legacy.csv
new file mode 100644
index 00000000..ceb9eb6e
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/benchmark_legacy.csv
@@ -0,0 +1,49 @@
+size,channels,pr,sr,h,sigma,kernel_ver,kernel_ms,e2e_ms,mpx_s,cpu_ms,speedup
+1920x1080,1,3,10,10,25,0,72.021,73.3564,28.7916,-1,-1
+1920x1080,1,3,10,10,25,1,31.399,32.9606,66.0402,-1,-1
+1920x1080,1,3,10,10,25,2,26.9916,28.4374,76.8239,-1,-1
+1920x1080,1,3,10,15,25,0,72.0691,73.4943,28.7724,-1,-1
+1920x1080,1,3,10,15,25,1,31.4204,32.9142,65.9953,-1,-1
+1920x1080,1,3,10,15,25,2,26.9978,28.5832,76.8064,-1,-1
+1920x1080,1,2,7,10,15,0,19.9782,21.4234,103.793,-1,-1
+1920x1080,1,2,7,10,15,1,8.79104,10.1764,235.877,-1,-1
+1920x1080,1,2,7,10,15,2,6.6857,7.97942,310.155,-1,-1
+1920x1080,1,4,14,12,35,0,219.77,221.643,9.43533,-1,-1
+1920x1080,1,4,14,12,35,1,95.5075,96.5187,21.7114,-1,-1
+1920x1080,1,4,14,12,35,2,95.5034,96.4983,21.7123,-1,-1
+1920x1080,3,3,10,10,25,0,125.909,129.137,16.469,-1,-1
+1920x1080,3,3,10,10,25,1,58.8974,61.9797,35.207,-1,-1
+1920x1080,3,3,10,10,25,2,57.728,60.6587,35.9202,-1,-1
+1920x1080,3,3,10,15,25,0,125.947,130.057,16.4641,-1,-1
+1920x1080,3,3,10,15,25,1,59.1022,61.7263,35.085,-1,-1
+1920x1080,3,3,10,15,25,2,57.8929,60.4233,35.8179,-1,-1
+1920x1080,3,2,7,10,15,0,34.26,37.781,60.5255,-1,-1
+1920x1080,3,2,7,10,15,1,15.8116,19.3021,131.144,-1,-1
+1920x1080,3,2,7,10,15,2,14.9924,18.4552,138.31,-1,-1
+1920x1080,3,4,14,12,35,0,389.667,393.404,5.32147,-1,-1
+1920x1080,3,4,14,12,35,1,233.681,237.497,8.87364,-1,-1
+1920x1080,3,4,14,12,35,2,233.679,237.292,8.87372,-1,-1
+3840x2160,1,3,10,10,25,0,286.953,291.7,28.905,-1,-1
+3840x2160,1,3,10,10,25,1,124.657,127.953,66.538,-1,-1
+3840x2160,1,3,10,10,25,2,106.371,109.509,77.9761,-1,-1
+3840x2160,1,3,10,15,25,0,286.96,291.46,28.9044,-1,-1
+3840x2160,1,3,10,15,25,1,124.658,129.147,66.5374,-1,-1
+3840x2160,1,3,10,15,25,2,106.386,110.804,77.9648,-1,-1
+3840x2160,1,2,7,10,15,0,79.5464,84.0145,104.271,-1,-1
+3840x2160,1,2,7,10,15,1,34.8774,39.3418,237.816,-1,-1
+3840x2160,1,2,7,10,15,2,26.4448,30.8649,313.65,-1,-1
+3840x2160,1,4,14,12,35,0,875.202,880.017,9.47713,-1,-1
+3840x2160,1,4,14,12,35,1,378.942,384.487,21.8883,-1,-1
+3840x2160,1,4,14,12,35,2,378.941,382.351,21.8883,-1,-1
+3840x2160,3,3,10,10,25,0,502.12,516.696,16.5187,-1,-1
+3840x2160,3,3,10,10,25,1,235.926,249.267,35.1569,-1,-1
+3840x2160,3,3,10,10,25,2,229.223,242.733,36.1848,-1,-1
+3840x2160,3,3,10,15,25,0,503.897,518.236,16.4605,-1,-1
+3840x2160,3,3,10,15,25,1,236.172,249.428,35.1201,-1,-1
+3840x2160,3,3,10,15,25,2,229.225,242.209,36.1845,-1,-1
+3840x2160,3,2,7,10,15,0,136.475,148.992,60.7761,-1,-1
+3840x2160,3,2,7,10,15,1,63.2095,75.9547,131.221,-1,-1
+3840x2160,3,2,7,10,15,2,59.5139,71.9586,139.369,-1,-1
+3840x2160,3,4,14,12,35,0,1569.88,1584.02,5.28347,-1,-1
+3840x2160,3,4,14,12,35,1,925.448,939.821,8.96258,-1,-1
+3840x2160,3,4,14,12,35,2,925.396,938.748,8.96308,-1,-1
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx4090d/environment_legacy.txt b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/environment_legacy.txt
new file mode 100644
index 00000000..7745d94c
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/environment_legacy.txt
@@ -0,0 +1,12 @@
+collected_at=2026-09-07
+gpu=NVIDIA GeForce RTX 4090 D
+compute_capability=8.9
+architecture=sm_89
+memory_mib=24564
+driver=570.124.06
+cuda_toolkit=12.8.61
+cuda_symlink=/usr/local/cuda -> cuda-12.8
+host=Docker container, 128 vCPU
+nsight_systems=2024.6.2.225
+nsight_compute=2025.1.0.0 (ERR_NVGPUCTRPERM: host counters restricted)
+build_command=make PLATFORM=nvidia ARCH=sm_89
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx4090d/nsys_summary_legacy.csv b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/nsys_summary_legacy.csv
new file mode 100644
index 00000000..eaf2524d
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/nsys_summary_legacy.csv
@@ -0,0 +1,4 @@
+implementation,kernel_name,gpu_ms,gpu_share_percent
+V0,NlmNaiveKernel,133.217,100.0
+V1,NlmSmemKernel,62.126,99.9
+V2,NlmSmemUnrollKernel<3>,61.036,99.9
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx4090d/quality_legacy.csv b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/quality_legacy.csv
new file mode 100644
index 00000000..b8a31720
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/quality_legacy.csv
@@ -0,0 +1,3 @@
+label,mae,psnr_db,source
+noisy-sigma25,12.4857,24.95,paired synthetic images seed=7
+base-v2,0.5972,49.41,paired synthetic images seed=7
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx4090d/run_all_legacy.txt b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/run_all_legacy.txt
new file mode 100644
index 00000000..9e81d3e6
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/run_all_legacy.txt
@@ -0,0 +1,198 @@
+===============================================================
+ nlm_denoise —— NVIDIA RTX 4090 平台全程验证实测结果
+ 验证目的：确认项目在英伟达平台可跑完全程（构建→单测→校验→基准→质量评估→profiling）
+ 验证时间：2026-09-07
+ 结论：全部通过（PASS）
+===============================================================
+
+[环境]
+GPU      : NVIDIA GeForce RTX 4090 D（compute capability 8.9 / sm_89），24564 MiB
+驱动     : 570.124.06
+CUDA     : release 12.8, V12.8.61（/usr/local/cuda -> cuda-12.8）
+主机     : Docker 容器（/.dockerenv 存在），128 vCPU
+构建命令 : make PLATFORM=nvidia ARCH=sm_89（ARCH 必须显式传入；Makefile 默认留空）
+Profiler : Nsight Systems 2024.6.2.225（可用）
+           Nsight Compute 2025.1.0.0（已安装，但计数器权限被宿主限制，见 [6]）
+
+===============================================================
+[1] 构建（PLATFORM=nvidia, -arch=sm_89）
+===============================================================
+9 个编译单元 + kernels/nvidia/kernels.cu 全部编译通过，链接生成 build/nlm_denoise。
+无 warning、无 error。
+注：首次尝试因 nvcc 不在 PATH 失败（make: nvcc: No such file or directory, Error 127），
+    将 /usr/local/cuda/bin 加入 PATH 后正常——属环境配置问题，非代码问题。
+
+===============================================================
+[2] 单元测试（units，12 项）
+===============================================================
+[units] 参数解析
+  [PASS] 正常解析（含注释/缺省）
+  [PASS] 未知 key 拒绝
+  [PASS] 越界取值拒绝
+[units] 质量指标
+  [PASS] MAE(相同图像)=0
+  [PASS] PSNR(相同图像)=inf
+  [PASS] MAE(恒差10)=10
+  [PASS] PSNR(恒差10)≈28.13dB
+[units] CPU 参考实现
+  [PASS] 常量灰度图恒等
+  [PASS] 常量 RGB 图恒等
+  [PASS] 确定性（两次运行一致）
+  [PASS] 极小图（5x7, sr=10）不崩溃且尺寸正确
+    降噪前后对干净图 MAE: 12.48 -> 1.57
+  [PASS] 降噪有效性（MAE 下降）
+========== 全部通过（失败 0 项）==========
+
+===============================================================
+[3] 正确性校验（validate，1080p RGB 合成含噪图，任务基线参数 pr=3 sr=10 h=10 σ=25）
+===============================================================
+[validate] CPU 参考耗时: 174535.3 ms
+ver    kernel(ms)   e2e(ms)      MAE(vsCPU)   PSNR(dB)   判定
+0      132.935      305.017      0.0000       inf        PASS
+1      58.876       61.969       0.0000       inf        PASS
+2      57.614       61.489       0.0000       inf        PASS
+[validate] 总体判定: PASS
+
+三个 kernel 版本与 CPU 参考 bit 级一致（MAE=0、PSNR=inf）。
+对单线程 CPU 参考加速比 ≈ 174535 / 61.5 ≈ 2838×（端到端）。
+
+===============================================================
+[4] 性能基准（bench，48 组 = 2 分辨率 × 2 通道 × 4 参数 × 3 版本）
+     warmup=3, repeat=10, kernel_ms 取最小值；完整数据见 benchmark_legacy.csv
+===============================================================
+kernel 耗时（ms）与吞吐量（Mpx/s）：
+
+| 配置                      | V0 naive | V1 smem | V2 unroll | V2 吞吐   | V2 vs V0 |
+|---------------------------|----------|---------|-----------|-----------|----------|
+| 1080p 灰度 base           |   72.02  |  31.40  |   26.99   |  76.8     |  2.67×   |
+| 1080p 灰度 strong-h       |   72.07  |  31.42  |   27.00   |  76.8     |  2.67×   |
+| 1080p 灰度 small(pr2,sr7) |   19.98  |   8.79  |    6.69   | 310.2     |  2.99×   |
+| 1080p 灰度 large(pr4,sr14)|  219.77  |  95.51  |   95.50 * |  21.7     |  2.30×   |
+| 1080p RGB  base           |  125.91  |  58.90  |   57.73   |  35.9     |  2.18×   |
+| 1080p RGB  strong-h       |  125.95  |  59.10  |   57.89   |  35.8     |  2.18×   |
+| 1080p RGB  small          |   34.26  |  15.81  |   14.99   | 138.3     |  2.29×   |
+| 1080p RGB  large          |  389.67  | 233.68  |  233.68 * |   8.9     |  1.67×   |
+| 4K 灰度 base              |  286.95  | 124.66  |  106.37   |  78.0     |  2.70×   |
+| 4K 灰度 small             |   79.55  |  34.88  |   26.44   | 313.7     |  3.01×   |
+| 4K 灰度 large             |  875.20  | 378.94  |  378.94 * |  21.9     |  2.31×   |
+| 4K RGB  base              |  502.12  | 235.93  |  229.22   |  36.2     |  2.19×   |
+| 4K RGB  small             |  136.48  |  63.21  |   59.51   | 139.4     |  2.29×   |
+| 4K RGB  large             | 1569.88  | 925.45  |  925.40 * |   9.0     |  1.70×   |
+
+* V2 在 pr≥4 自动回退 V1 通用循环（模板全展开会导致寄存器溢出），故两者持平，
+  差异 <0.01 ms 属测量噪声——回退修正逻辑在 4090 上同样生效且行为正确。
+
+跨平台对比（V2，kernel ms）：
+| 配置            | RTX 3060 Laptop | RTX 4090 D | 4090 加速 |
+|-----------------|-----------------|------------|-----------|
+| 1080p 灰度 base |      152.2      |    26.99   |   5.64×   |
+| 1080p RGB base  |      354.0      |    57.73   |   6.13×   |
+| 4K 灰度 base    |      693.9      |   106.37   |   6.52×   |
+| 4K RGB base     |     1659.5      |   229.22   |   7.24×   |
+
+4K 交互式（进阶目标）在 4090 上的达成情况：
+| 配置          | kernel   | 帧率     | 是否交互（≥10 fps / ≤100 ms） |
+|---------------|----------|----------|-------------------------------|
+| 4K 灰度 small |  26.44ms | 37.8 fps | 达成                          |
+| 4K RGB  small |  59.51ms | 16.8 fps | 达成                          |
+| 4K 灰度 base  | 106.37ms |  9.4 fps | 接近（略超 100 ms）           |
+| 4K RGB  base  | 229.22ms |  4.4 fps | 未达成                        |
+=> 结论：4090 上以 small 参数（pr=2, sr=7）可实现 4K 交互式处理；
+   任务基线参数（pr=3, sr=10）下 4K RGB 仍需算法级近似（积分图）才能交互。
+
+===============================================================
+[5] 图像质量指标（与 ground truth 对比，1080p RGB，seed=7 严格配对）
+===============================================================
+数据对：gen --sigma 0 --seed 7（干净原图）与 gen --sigma 25 --seed 7（含噪图），
+        降噪输出由 V2 + 任务基线参数产生。
+
+noisy vs clean               MAE=12.4857  PSNR=24.95 dB
+denoised vs clean            MAE=0.5972   PSNR=49.41 dB
+
+改善：MAE ↓95.2%，PSNR +24.46 dB。
+与 RTX 3060 Laptop 上测得的数值**完全一致**（12.4857/24.95 与 0.5972/49.41），
+印证算法输出 bit 级确定、跨 GPU 架构可复现。
+
+===============================================================
+[6] Profiling 分析
+===============================================================
+--- nsys（Nsight Systems 2024.6.2，可用，已实测）---
+1080p RGB 基线，cuda_gpu_kern_sum（单次 run，--trace=cuda）：
+
+| kernel                          | GPU 时间    | 占比   |
+|---------------------------------|-------------|--------|
+| nlm::NlmNaiveKernel        (V0) | 133.217 ms  | 100.0% |
+| nlm::NlmSmemKernel         (V1) |  62.126 ms  |  99.9% |
+| nlm::NlmSmemUnrollKernel<3>(V2) |  61.036 ms  |  99.9% |
+| nlm::CvtU8ToF32PlanarKernel     |  ~0.021 ms  |   0.0% |
+| nlm::CvtF32ToU8InterleavedKernel|  ~0.021 ms  |   0.0% |
+
+cuda_api_sum（V2）：
+| API                 | 总时间    | 调用数 | 备注                                   |
+|---------------------|-----------|--------|----------------------------------------|
+| cudaMalloc          |  96.18 ms |   4    | 单次最大 95.98 ms（首次分配页映射开销）|
+| cudaEventSynchronize|  61.04 ms |   1    | 即等待 kernel 完成                     |
+| cudaMemcpy          |   1.24 ms |   2    | H2D + D2H 合计                         |
+| cudaFree            |   0.96 ms |   4    |                                        |
+| cudaLaunchKernel    |   0.70 ms |   3    | 启动开销可忽略                         |
+
+关键结论：
+1. **NLM kernel 占 GPU 时间 99.9%**，u8↔f32 转换 kernel 各约 21 µs（合计 <0.07%），
+   设备侧数据通路无优化空间剩余——瓶颈完全在 NLM 计算本身；
+2. **H2D+D2H 仅 1.24 ms**（占 kernel 时间 2%），PCIe 传输不是瓶颈，
+   pinned memory / stream 分片对单帧延迟收益有限（主要为多帧流水铺路）；
+3. **首次 cudaMalloc 约 96 ms** 是 host 侧最大单项开销，机载常驻进程应预分配缓冲池消除；
+4. **PNG 编解码主导单图 e2e**：run（含 PNG I/O）e2e=219.4 ms / kernel=60.9 ms
+   → host 侧 I/O 约 154.6 ms；而 bench（内存合成图，无 PNG I/O）e2e=60.66 ms /
+   kernel=57.73 ms → 通路开销仅 2.93 ms。故实时机载管线应绕开 PNG，
+   直接消费传感器缓冲或用 nvJPEG 硬解。
+
+--- ncu（Nsight Compute 2025.1.0.0，已安装但不可用）---
+现象：ERR_NVGPUCTRPERM - The user does not have permission to access NVIDIA GPU
+      Performance Counters on the target device 0
+根因：/proc/driver/nvidia/params 中 RmProfilingAdminOnly: 1，且当前为 Docker 容器
+      （/.dockerenv 存在）。容器内即使是 root 也无法修改宿主内核模块参数。
+解除方式（需宿主侧操作，容器内不可行）：
+      宿主机设置 NVreg_RestrictProfilingToAdminUsers=0 后重载 nvidia 驱动/重启，
+      或以特权容器（--cap-add=SYS_ADMIN）运行。
+替代方案：以 nsys kernel/API 级时间分解 + 设备属性静态推算完成瓶颈论证（见下）。
+
+--- 算力利用率推算（基于 cudaGetDeviceProperties 实测属性）---
+设备属性实测：
+| GPU               | SM 数 | FP32 cores | clockRate | FP32 峰值      | smem/block optin |
+|-------------------|-------|------------|-----------|----------------|------------------|
+| RTX 4090 D        |  114  |   14592    | 2520 MHz  | 73.5 TFLOPS    | 99.0 KB          |
+| RTX 3060 Laptop   |   30  |    3840    | 1702 MHz  | 13.1 TFLOPS    | 99.0 KB          |
+（峰值 = cores × 2 FLOP/FMA × clock；4090 D = 36.8 TFMA/s，3060 Laptop = 6.5 TFMA/s）
+
+计算量（1080p RGB，pr=3 sr=10）：
+  每输出像素 FMA = (2sr+1)² × (2pr+1)² × C = 441 × 49 × 3 = 64,827 FMA
+  单帧合计     = 2,073,600 px × 64,827 = 1.344×10¹¹ FMA（= 2.69×10¹¹ FLOP）
+
+| GPU             | kernel(V2) | 实测吞吐     | 峰值        | 利用率 |
+|-----------------|------------|--------------|-------------|--------|
+| RTX 4090 D      |  57.73 ms  | 2.33 TFMA/s  | 36.8 TFMA/s |  6.3%  |
+| RTX 3060 Laptop | 354.0 ms   | 0.38 TFMA/s  |  6.5 TFMA/s |  5.8%  |
+
+**关键结论**：两卡利用率同为约 6%（4090 6.3%、3060 Laptop 5.8%），实测加速比 6.13×
+与峰值算力比 5.62×（73.5/13.1）基本吻合——说明 kernel 瓶颈**不在硬件算力档位，
+而在指令级效率**：内层循环为 smem load → 减法 → FMA 的短依赖链，
+地址计算与 smem 访问指令稀释了 FMA 密度，换更强的卡只能等比提速。
+占用率静态推算（1080p RGB，256 线程/block，smem 约 20.7 KB/block）：
+  受 smem 限制 99/20.7 ≈ 4 block/SM = 1024 线程/SM，对 maxThreadsPerSM=1536 → 约 67% 占用率，
+  足以隐藏访存延迟，故进一步提速需提高每指令有效负载（多像素复用/寄存器 tile）
+  或降低计算量（积分图近似），而非单纯提高占用率。
+
+===============================================================
+[7] 全程验证结论
+===============================================================
+构建 → 单元测试(12/12) → 正确性校验(MAE=0, PASS) → 性能基准(48 组) →
+质量评估(PSNR 49.41 dB) → nsys profiling 分析：**全部通过，无失败项**。
+
+项目在 NVIDIA 平台（RTX 4090 D, CUDA 12.8, sm_89）可完整跑完全程，
+结果与 RTX 3060 Laptop（CUDA 12.9, sm_86）bit 级一致，跨架构可复现。
+
+已知限制（均如实记录，非本次验证引入）：
+- ncu 计数器受容器/宿主权限限制不可用（解除方式见 [6]）；
+- 4K RGB 在任务基线参数下 229 ms/帧，未达交互式；需 V3 积分图近似（见总结报告 §6.3）；
+- 国产平台（沐曦/摩尔/天数）代码就绪但未在对应硬件实测。
diff --git a/07_nlm_denoise/blackbook537/assets/results/rtx4090d/validation_legacy.csv b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/validation_legacy.csv
new file mode 100644
index 00000000..3b43dc51
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/assets/results/rtx4090d/validation_legacy.csv
@@ -0,0 +1,4 @@
+implementation,kernel_ms,e2e_ms,mae_vs_cpu,psnr_vs_cpu,cpu_ms,status
+V0,132.935,305.017,0.0000,inf,174535.3,PASS
+V1,58.876,61.969,0.0000,inf,174535.3,PASS
+V2,57.614,61.489,0.0000,inf,174535.3,PASS
diff --git a/07_nlm_denoise/blackbook537/include/benchmark.h b/07_nlm_denoise/blackbook537/include/benchmark.h
new file mode 100644
index 00000000..e4b4cf80
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/benchmark.h
@@ -0,0 +1,74 @@
+#ifndef NLM_BENCHMARK_H
+#define NLM_BENCHMARK_H
+
+#include <cmath>
+#include <cstdint>
+#include <string>
+
+#include "image_io.h"
+
+// ============================================================================
+// 质量指标、基准与验证共享接口
+// ============================================================================
+
+// 平均绝对误差（要求两图同尺寸同通道）
+inline double ComputeMAE(const ImageU8& a, const ImageU8& b) {
+    if (a.data.size() != b.data.size() || a.data.empty()) return -1.0;
+    double acc = 0.0;
+    for (size_t i = 0; i < a.data.size(); ++i)
+        acc += std::fabs((double)a.data[i] - (double)b.data[i]);
+    return acc / (double)a.data.size();
+}
+
+// 峰值信噪比（dB）；两图完全一致时返回 +inf
+inline double ComputePSNR(const ImageU8& a, const ImageU8& b) {
+    if (a.data.size() != b.data.size() || a.data.empty()) return -1.0;
+    double acc = 0.0;
+    for (size_t i = 0; i < a.data.size(); ++i) {
+        const double d = (double)a.data[i] - (double)b.data[i];
+        acc += d * d;
+    }
+    const double mse = acc / (double)a.data.size();
+    if (mse == 0.0) return HUGE_VAL;
+    return 20.0 * std::log10(255.0 / std::sqrt(mse));
+}
+
+// 合成测试图：平滑渐变背景 + LCG 均匀噪声（用于无外部图像时的端到端测试与基准）
+inline ImageU8 MakeSyntheticImage(int w, int h, int c, float sigma, uint32_t seed) {
+    ImageU8 img;
+    img.width = w;
+    img.height = h;
+    img.channels = c;
+    img.data.resize((size_t)w * h * c);
+    uint32_t state = seed ? seed : 1u;
+    for (int y = 0; y < h; ++y) {
+        for (int x = 0; x < w; ++x) {
+            for (int ch = 0; ch < c; ++ch) {
+                // 背景：x/y 方向渐变 + 通道偏移，制造纹理差异
+                float base = (float)((x * 255 / (w > 1 ? w - 1 : 1) +
+                                      y * 255 / (h > 1 ? h - 1 : 1)) / 2);
+                base = base * 0.8f + 20.0f + 15.0f * ch;
+                // LCG -> [0,1) -> 均匀噪声 [-sigma, +sigma]
+                state = state * 1664525u + 1013904223u;
+                const float u = (float)(state >> 8) / 16777216.0f;
+                float v = base + (u - 0.5f) * 2.0f * sigma;
+                int q = (int)(v + 0.5f);
+                q = q < 0 ? 0 : (q > 255 ? 255 : q);
+                img.data[((size_t)y * w + x) * c + ch] = (uint8_t)q;
+            }
+        }
+    }
+    return img;
+}
+
+// ---------------- 基准与验证子命令入口（main.cpp 调度）----------------
+int RunValidate(const std::string& input, const std::string& output,
+                const std::string& params_path);
+int RunBenchmark(const std::string& sizes_spec, const std::string& channels_spec,
+                 const std::string& param_sets_spec, const std::string& csv_path,
+                 int warmup, int repeat, bool with_cpu, int cpu_repeat);
+int RunMetrics(const std::string& reference_path, const std::string& test_path,
+               const std::string& label, const std::string& csv_path);
+int RunUnitTests();
+
+#endif // NLM_BENCHMARK_H
diff --git a/07_nlm_denoise/blackbook537/include/image_io.h b/07_nlm_denoise/blackbook537/include/image_io.h
new file mode 100644
index 00000000..c8c161f6
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/image_io.h
@@ -0,0 +1,34 @@
+#ifndef NLM_IMAGE_IO_H
+#define NLM_IMAGE_IO_H
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+// ============================================================================
+// 图像对象（Host 侧，interleaved 布局，与文件格式一致）
+//   channels == 1 : 灰度图，data[y*w + x]
+//   channels == 3 : RGB 图， data[(y*w + x)*3 + ch]
+// 像素值范围 [0,255]（8 位整数），处理时由转换 kernel 转为浮点。
+// ============================================================================
+struct ImageU8 {
+    int width;
+    int height;
+    int channels;                 // ∈ {1, 3}
+    std::vector<uint8_t> data;
+
+    ImageU8() : width(0), height(0), channels(0) {}
+};
+
+// 解码 PNG/JPG/BMP/TGA；仅接受灰度或 RGB（带 alpha 的文件自动丢弃 alpha 通道）
+bool LoadImage(const char* path, ImageU8* out, std::string* err);
+
+// 按输出扩展名编码（.png/.jpg/.jpeg/.bmp/.tga，缺省 PNG），与输入同格式语义
+// 保存前自动逐级创建父目录（如 output/images/ 不存在时创建），便于集中管理数据
+bool SaveImage(const char* path, const ImageU8& img, std::string* err);
+
+// 确保路径的父目录存在（逐级创建，已存在则跳过；相对/无目录路径直接成功）
+// 数据集中管理配套：日志等非图像输出路径同样可用（POSIX mkdir，Linux 部署）
+void EnsureParentDir(const char* path);
+
+#endif // NLM_IMAGE_IO_H
diff --git a/07_nlm_denoise/blackbook537/include/kernels.h b/07_nlm_denoise/blackbook537/include/kernels.h
new file mode 100644
index 00000000..ac42bd67
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/kernels.h
@@ -0,0 +1,53 @@
+#ifndef NLM_KERNELS_H
+#define NLM_KERNELS_H
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+#include "params.h"
+
+// ============================================================================
+// 设备侧参数（Host 预计算，避免 kernel 内重复计算）
+// 权重公式：w(p,q) = exp(-max(dist(P_p,P_q) - 2*sigma^2*N_patch, 0) / h^2)
+//   dist 为两 patch 的 L2 距离之和（RGB 时跨通道求和，三通道共享权重）
+//   N_patch = (2*pr+1)^2 * channels
+// ============================================================================
+struct NlmParamsDev {
+    int   pr;             // patch 半径
+    int   sr;             // 搜索窗半径
+    float inv_h2;         // 1 / h^2
+    float two_sigma2_n;   // 2 * sigma^2 * N_patch
+};
+
+inline NlmParamsDev MakeDevParams(const NlmParams& p, int channels) {
+    NlmParamsDev d;
+    d.pr = p.patch_radius;
+    d.sr = p.search_radius;
+    d.inv_h2 = 1.0f / (p.h * p.h);
+    const float n_patch =
+        (float)((2 * p.patch_radius + 1) * (2 * p.patch_radius + 1) * channels);
+    d.two_sigma2_n = 2.0f * p.sigma * p.sigma * n_patch;
+    return d;
+}
+
+// ---------------- 设备资源管理（pipeline 经此持有设备内存，无需包含 GPU 头文件）----------------
+bool GpuAlloc(void** ptr, size_t bytes, std::string* err);
+bool GpuFree(void* ptr, std::string* err);
+bool GpuCopyH2D(void* dst, const void* src, size_t bytes, std::string* err);
+bool GpuCopyD2H(void* dst, const void* src, size_t bytes, std::string* err);
+
+// ---------------- 设备端 kernel 调度 ----------------
+// 数据布局约定：planar f32，第 ch 通道位于 d_ptr + ch*w*h
+bool GpuCvtU8ToF32Planar(const uint8_t* d_in, float* d_out, int w, int h, int c,
+                         std::string* err);
+bool GpuCvtF32ToU8Interleaved(const float* d_in, uint8_t* d_out, int w, int h, int c,
+                              std::string* err);
+
+// NLM 主体。version: 0=naive, 1=smem, 2=smem+模板展开（3 为预留进阶版本）。
+// kernel_ms 输出 kernel 纯计算时间（GPU event 计时）。
+// smem 需求超出设备上限时自动回退 naive 路径（保证功能正确）。
+bool GpuNlm(const float* d_src, float* d_dst, int w, int h, int c,
+            const NlmParamsDev& p, int version, float* kernel_ms, std::string* err);
+
+#endif // NLM_KERNELS_H
diff --git a/07_nlm_denoise/blackbook537/include/nlm_cpu_ref.h b/07_nlm_denoise/blackbook537/include/nlm_cpu_ref.h
new file mode 100644
index 00000000..ea176803
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/nlm_cpu_ref.h
@@ -0,0 +1,17 @@
+#ifndef NLM_CPU_REF_H
+#define NLM_CPU_REF_H
+
+#include "image_io.h"
+#include "params.h"
+
+// ============================================================================
+// CPU 参考实现（正确性 ground truth + 加速比基线）
+// 与 GPU kernel 语义严格一致：
+//   - planar 浮点计算，RGB 跨通道联合 patch 距离、共享权重
+//   - 边界：每次最终访问坐标独立 clamp（等价 OpenCV BORDER_REPLICATE）
+//   - 权重：w = exp(-max(dist - 2*sigma^2*N_patch, 0) / h^2)
+// 单线程、无第三方依赖。
+// ============================================================================
+void NlmDenoiseCpuRef(const ImageU8& src, ImageU8* dst, const NlmParams& params);
+
+#endif // NLM_CPU_REF_H
diff --git a/07_nlm_denoise/blackbook537/include/params.h b/07_nlm_denoise/blackbook537/include/params.h
new file mode 100644
index 00000000..896412ac
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/params.h
@@ -0,0 +1,28 @@
+#ifndef NLM_PARAMS_H
+#define NLM_PARAMS_H
+
+#include <string>
+
+// ============================================================================
+// NLM 滤波参数（对应任务文档参数文件定义）
+//   patch_radius  : patch 半径，实际 patch 大小 (2*patch_radius+1)^2
+//   search_radius : 搜索窗口半径，实际窗口 (2*search_radius+1)^2
+//   h             : 滤波强度
+//   sigma         : 噪声标准差估计（8 位图像域 0-255）
+// 任务基线：patch_radius=3, search_radius=10, h=10.0, sigma=25.0
+// ============================================================================
+struct NlmParams {
+    int   patch_radius;   // >= 1，任务要求至少支持到 3
+    int   search_radius;  // >= 1，任务要求至少支持到 10
+    float h;              // > 0
+    float sigma;          // >= 0
+
+    NlmParams() : patch_radius(3), search_radius(10), h(10.0f), sigma(25.0f) {}
+};
+
+// 解析文本参数文件（"key = value" 格式，'#' 之后为注释）。
+// 缺失字段使用默认值；未知 key / 非法取值 / 越界取值均返回 false 并填写 err。
+// 取值范围约束：pr ∈ [1,8]，sr ∈ [1,32]，h ∈ (0,1000]，sigma ∈ [0,1000]。
+bool ParseParams(const char* path, NlmParams* out, std::string* err);
+
+#endif // NLM_PARAMS_H
diff --git a/07_nlm_denoise/blackbook537/include/pipeline.h b/07_nlm_denoise/blackbook537/include/pipeline.h
new file mode 100644
index 00000000..4a137b6c
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/pipeline.h
@@ -0,0 +1,42 @@
+#ifndef NLM_PIPELINE_H
+#define NLM_PIPELINE_H
+
+#include <chrono>
+#include <string>
+
+#include "image_io.h"
+#include "params.h"
+
+// Host 侧计时器（chrono，供端到端/H2D/D2H 计时）
+struct HostTimer {
+    std::chrono::steady_clock::time_point t0;
+    void Start() { t0 = std::chrono::steady_clock::now(); }
+    double Ms() const {
+        return std::chrono::duration<double, std::milli>(
+                   std::chrono::steady_clock::now() - t0).count();
+    }
+};
+
+// 性能报告（计时口径见架构设计文档 §5.3）
+struct NlmPerfReport {
+    float kernel_ms;          // NLM kernel 纯计算时间（GPU event）
+    float e2e_ms;             // 端到端 GPU 时间（含 H2D/D2H 与转换 kernel，Host chrono）
+    float h2d_ms;
+    float d2h_ms;
+    float throughput_mpx_s;   // 按 kernel_ms 计算的吞吐量（百万像素/秒）
+    float cpu_ref_ms;         // CPU 参考耗时（未测量为 -1）
+    float speedup_vs_cpu;     // cpu_ref_ms / e2e_ms（未测量为 -1）
+
+    NlmPerfReport()
+        : kernel_ms(0), e2e_ms(0), h2d_ms(0), d2h_ms(0),
+          throughput_mpx_s(0), cpu_ref_ms(-1.0f), speedup_vs_cpu(-1.0f) {}
+};
+
+// GPU 降噪全流程：u8 interleaved -> planar f32 -> NLM -> 量化回 u8。
+// 无状态纯函数接口（外层可直接包帧循环用于交互式场景）。
+// kernel_version: 0=naive, 1=smem, 2=smem+模板展开。
+bool NlmDenoiseGpu(const ImageU8& src, ImageU8* dst,
+                   const NlmParams& params, int kernel_version,
+                   NlmPerfReport* perf, std::string* err);
+
+#endif // NLM_PIPELINE_H
diff --git a/07_nlm_denoise/blackbook537/include/platform_api.h b/07_nlm_denoise/blackbook537/include/platform_api.h
new file mode 100644
index 00000000..9d4cde62
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/platform_api.h
@@ -0,0 +1,75 @@
+#ifndef NLM_PLATFORM_API_H
+#define NLM_PLATFORM_API_H
+
+// ============================================================================
+// 平台抽象层（PAL, Platform Abstraction Layer）
+//
+// 统一封装设备运行时 API，算法核心代码 100% 共享，平台差异全部收敛到本文件。
+// 编译期平台宏由 Makefile 注入：
+//   -DPLATFORM_NVIDIA / -DPLATFORM_METAX / -DPLATFORM_MOORE / -DPLATFORM_ILUVATAR
+//
+// 说明：NVIDIA、沐曦 MACA、天数 CoreX 均提供 CUDA 兼容运行时（API 同名），
+//       摩尔线程 MUSA 使用 musa 前缀，单独适配。
+// ============================================================================
+
+#define NLM_STR_IMPL_(x) #x
+#define NLM_STR(x) NLM_STR_IMPL_(x)
+
+#if defined(PLATFORM_MOORE)
+// ---------------- 摩尔线程 MUSA ----------------
+#include <musa_runtime.h>
+#include <musa_runtime_api.h>
+typedef musaError_t gpuError_t;
+typedef musaEvent_t gpuEvent_t;
+#define GPU_SUCCESS                musaSuccess
+#define GPU_MEMCPY_H2D             musaMemcpyHostToDevice
+#define GPU_MEMCPY_D2H             musaMemcpyDeviceToHost
+#define GPU_MALLOC(p, b)           musaMalloc((p), (b))
+#define GPU_FREE(p)                musaFree(p)
+#define GPU_MEMCPY(d, s, b, k)     musaMemcpy((d), (s), (b), (k))
+#define GPU_EVENT_CREATE(e)        musaEventCreate(e)
+#define GPU_EVENT_DESTROY(e)       musaEventDestroy(e)
+#define GPU_EVENT_RECORD(e)        musaEventRecord(e)
+#define GPU_EVENT_SYNC(e)          musaEventSynchronize(e)
+#define GPU_EVENT_ELAPSED(ms,a,b)  musaEventElapsedTime((ms), (a), (b))
+#define GPU_GET_LAST_ERROR()       musaGetLastError()
+#define GPU_GET_ERROR_STRING(e)    musaGetErrorString(e)
+#define GPU_GET_DEVICE(d)          musaGetDevice(d)
+#define GPU_FUNC_SET_MAX_SMEM(f,b) musaFuncSetAttribute((const void*)(f), musaFuncAttributeMaxDynamicSharedMemorySize, (int)(b))
+#define GPU_GET_SMEM_OPTIN(d,v)    musaDeviceGetAttribute((v), musaDevAttrMaxSharedMemoryPerBlockOptin, (d))
+
+#else
+// ---------------- NVIDIA CUDA / 沐曦 MACA / 天数 CoreX ----------------
+#include <cuda_runtime.h>
+typedef cudaError_t gpuError_t;
+typedef cudaEvent_t gpuEvent_t;
+#define GPU_SUCCESS                cudaSuccess
+#define GPU_MEMCPY_H2D             cudaMemcpyHostToDevice
+#define GPU_MEMCPY_D2H             cudaMemcpyDeviceToHost
+#define GPU_MALLOC(p, b)           cudaMalloc((p), (b))
+#define GPU_FREE(p)                cudaFree(p)
+#define GPU_MEMCPY(d, s, b, k)     cudaMemcpy((d), (s), (b), (k))
+#define GPU_EVENT_CREATE(e)        cudaEventCreate(e)
+#define GPU_EVENT_DESTROY(e)       cudaEventDestroy(e)
+#define GPU_EVENT_RECORD(e)        cudaEventRecord(e)
+#define GPU_EVENT_SYNC(e)          cudaEventSynchronize(e)
+#define GPU_EVENT_ELAPSED(ms,a,b)  cudaEventElapsedTime((ms), (a), (b))
+#define GPU_GET_LAST_ERROR()       cudaGetLastError()
+#define GPU_GET_ERROR_STRING(e)    cudaGetErrorString(e)
+#define GPU_GET_DEVICE(d)          cudaGetDevice(d)
+#define GPU_FUNC_SET_MAX_SMEM(f,b) cudaFuncSetAttribute((const void*)(f), cudaFuncAttributeMaxDynamicSharedMemorySize, (int)(b))
+#define GPU_GET_SMEM_OPTIN(d,v)    cudaDeviceGetAttribute((v), cudaDevAttrMaxSharedMemoryPerBlockOptin, (d))
+#endif
+
+// 统一错误检查宏：任一 GPU API 失败即将错误信息写入 err 并返回 false
+#define GPU_CHECK(call, err)                                                          \
+    do {                                                                              \
+        gpuError_t e_ = (call);                                                       \
+        if (e_ != GPU_SUCCESS) {                                                      \
+            if (err) *err = std::string("[GPU] ") + GPU_GET_ERROR_STRING(e_) +        \
+                            " @ " __FILE__ ":" NLM_STR(__LINE__);                     \
+            return false;                                                             \
+        }                                                                             \
+    } while (0)
+
+#endif // NLM_PLATFORM_API_H
diff --git a/07_nlm_denoise/blackbook537/include/stb_image.h b/07_nlm_denoise/blackbook537/include/stb_image.h
new file mode 100644
index 00000000..9eedabed
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/stb_image.h
@@ -0,0 +1,7988 @@
+/* stb_image - v2.30 - public domain image loader - http://nothings.org/stb
+                                  no warranty implied; use at your own risk
+
+   Do this:
+      #define STB_IMAGE_IMPLEMENTATION
+   before you include this file in *one* C or C++ file to create the implementation.
+
+   // i.e. it should look like this:
+   #include ...
+   #include ...
+   #include ...
+   #define STB_IMAGE_IMPLEMENTATION
+   #include "stb_image.h"
+
+   You can #define STBI_ASSERT(x) before the #include to avoid using assert.h.
+   And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free
+
+
+   QUICK NOTES:
+      Primarily of interest to game developers and other people who can
+          avoid problematic images and only need the trivial interface
+
+      JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib)
+      PNG 1/2/4/8/16-bit-per-channel
+
+      TGA (not sure what subset, if a subset)
+      BMP non-1bpp, non-RLE
+      PSD (composited view only, no extra channels, 8/16 bit-per-channel)
+
+      GIF (*comp always reports as 4-channel)
+      HDR (radiance rgbE format)
+      PIC (Softimage PIC)
+      PNM (PPM and PGM binary only)
+
+      Animated GIF still needs a proper API, but here's one way to do it:
+          http://gist.github.com/urraka/685d9a6340b26b830d49
+
+      - decode from memory or through FILE (define STBI_NO_STDIO to remove code)
+      - decode from arbitrary I/O callbacks
+      - SIMD acceleration on x86/x64 (SSE2) and ARM (NEON)
+
+   Full documentation under "DOCUMENTATION" below.
+
+
+LICENSE
+
+  See end of file for license information.
+
+RECENT REVISION HISTORY:
+
+      2.30  (2024-05-31) avoid erroneous gcc warning
+      2.29  (2023-05-xx) optimizations
+      2.28  (2023-01-29) many error fixes, security errors, just tons of stuff
+      2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
+      2.26  (2020-07-13) many minor fixes
+      2.25  (2020-02-02) fix warnings
+      2.24  (2020-02-02) fix warnings; thread-local failure_reason and flip_vertically
+      2.23  (2019-08-11) fix clang static analysis warning
+      2.22  (2019-03-04) gif fixes, fix warnings
+      2.21  (2019-02-25) fix typo in comment
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) bugfix, 1-bit BMP, 16-bitness query, fix warnings
+      2.16  (2017-07-23) all functions have 16-bit variants; optimizations; bugfixes
+      2.15  (2017-03-18) fix png-1,2,4; all Imagenet JPGs; no runtime SSE detection on GCC
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-12-04) experimental 16-bit API, only for PNG so far; fixes
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) 16-bit PNGS; enable SSE2 in non-gcc x64
+                         RGB-format JPEG; remove white matting in PSD;
+                         allocate large structures on the stack;
+                         correct channel count for PNG & BMP
+      2.10  (2016-01-22) avoid warning introduced in 2.09
+      2.09  (2016-01-16) 16-bit TGA; comments in PNM files; STBI_REALLOC_SIZED
+
+   See end of file for full revision history.
+
+
+ ============================    Contributors    =========================
+
+ Image formats                          Extensions, features
+    Sean Barrett (jpeg, png, bmp)          Jetro Lauha (stbi_info)
+    Nicolas Schulz (hdr, psd)              Martin "SpartanJ" Golini (stbi_info)
+    Jonathan Dummer (tga)                  James "moose2000" Brown (iPhone PNG)
+    Jean-Marc Lienher (gif)                Ben "Disch" Wenger (io callbacks)
+    Tom Seddon (pic)                       Omar Cornut (1/2/4-bit PNG)
+    Thatcher Ulrich (psd)                  Nicolas Guillemot (vertical flip)
+    Ken Miller (pgm, ppm)                  Richard Mitton (16-bit PSD)
+    github:urraka (animated gif)           Junggon Kim (PNM comments)
+    Christopher Forseth (animated gif)     Daniel Gibson (16-bit TGA)
+                                           socks-the-fox (16-bit PNG)
+                                           Jeremy Sawicki (handle all ImageNet JPGs)
+ Optimizations & bugfixes                  Mikhail Morozov (1-bit BMP)
+    Fabian "ryg" Giesen                    Anael Seghezzi (is-16-bit query)
+    Arseny Kapoulkine                      Simon Breuss (16-bit PNM)
+    John-Mark Allen
+    Carmelo J Fdez-Aguera
+
+ Bug & warning fixes
+    Marc LeBlanc            David Woo          Guillaume George     Martins Mozeiko
+    Christpher Lloyd        Jerry Jansson      Joseph Thomson       Blazej Dariusz Roszkowski
+    Phil Jordan                                Dave Moore           Roy Eltham
+    Hayaki Saito            Nathan Reed        Won Chun
+    Luke Graham             Johan Duparc       Nick Verigakis       the Horde3D community
+    Thomas Ruf              Ronny Chevalier                         github:rlyeh
+    Janez Zemva             John Bartholomew   Michal Cichon        github:romigrou
+    Jonathan Blow           Ken Hamada         Tero Hanninen        github:svdijk
+    Eugene Golushkov        Laurent Gomila     Cort Stratton        github:snagar
+    Aruelien Pocheville     Sergio Gonzalez    Thibault Reuille     github:Zelex
+    Cass Everitt            Ryamond Barbiero                        github:grim210
+    Paul Du Bois            Engin Manap        Aldo Culquicondor    github:sammyhw
+    Philipp Wiesemann       Dale Weiler        Oriol Ferrer Mesia   github:phprus
+    Josh Tobin              Neil Bickford      Matthew Gregan       github:poppolopoppo
+    Julian Raschke          Gregory Mullen     Christian Floisand   github:darealshinji
+    Baldur Karlsson         Kevin Schmidt      JR Smith             github:Michaelangel007
+                            Brad Weinberger    Matvey Cherevko      github:mosra
+    Luca Sas                Alexander Veselov  Zack Middleton       [reserved]
+    Ryan C. Gordon          [reserved]                              [reserved]
+                     DO NOT ADD YOUR NAME HERE
+
+                     Jacko Dirks
+
+  To add your name to the credits, pick a random blank space in the middle and fill it.
+  80% of merge conflicts on stb PRs are due to people adding their name at the end
+  of the credits.
+*/
+
+#ifndef STBI_INCLUDE_STB_IMAGE_H
+#define STBI_INCLUDE_STB_IMAGE_H
+
+// DOCUMENTATION
+//
+// Limitations:
+//    - no 12-bit-per-channel JPEG
+//    - no JPEGs with arithmetic coding
+//    - GIF always returns *comp=4
+//
+// Basic usage (see HDR discussion below for HDR usage):
+//    int x,y,n;
+//    unsigned char *data = stbi_load(filename, &x, &y, &n, 0);
+//    // ... process data if not NULL ...
+//    // ... x = width, y = height, n = # 8-bit components per pixel ...
+//    // ... replace '0' with '1'..'4' to force that many components per pixel
+//    // ... but 'n' will always be the number that it would have been if you said 0
+//    stbi_image_free(data);
+//
+// Standard parameters:
+//    int *x                 -- outputs image width in pixels
+//    int *y                 -- outputs image height in pixels
+//    int *channels_in_file  -- outputs # of image components in image file
+//    int desired_channels   -- if non-zero, # of image components requested in result
+//
+// The return value from an image loader is an 'unsigned char *' which points
+// to the pixel data, or NULL on an allocation failure or if the image is
+// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels,
+// with each pixel consisting of N interleaved 8-bit components; the first
+// pixel pointed to is top-left-most in the image. There is no padding between
+// image scanlines or between pixels, regardless of format. The number of
+// components N is 'desired_channels' if desired_channels is non-zero, or
+// *channels_in_file otherwise. If desired_channels is non-zero,
+// *channels_in_file has the number of components that _would_ have been
+// output otherwise. E.g. if you set desired_channels to 4, you will always
+// get RGBA output, but you can check *channels_in_file to see if it's trivially
+// opaque because e.g. there were only 3 channels in the source image.
+//
+// An output image with N components has the following components interleaved
+// in this order in each pixel:
+//
+//     N=#comp     components
+//       1           grey
+//       2           grey, alpha
+//       3           red, green, blue
+//       4           red, green, blue, alpha
+//
+// If image loading fails for any reason, the return value will be NULL,
+// and *x, *y, *channels_in_file will be unchanged. The function
+// stbi_failure_reason() can be queried for an extremely brief, end-user
+// unfriendly explanation of why the load failed. Define STBI_NO_FAILURE_STRINGS
+// to avoid compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly
+// more user-friendly ones.
+//
+// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
+//
+// To query the width, height and component count of an image without having to
+// decode the full file, you can use the stbi_info family of functions:
+//
+//   int x,y,n,ok;
+//   ok = stbi_info(filename, &x, &y, &n);
+//   // returns ok=1 and sets x, y, n if image is a supported format,
+//   // 0 otherwise.
+//
+// Note that stb_image pervasively uses ints in its public API for sizes,
+// including sizes of memory buffers. This is now part of the API and thus
+// hard to change without causing breakage. As a result, the various image
+// loaders all have certain limits on image size; these differ somewhat
+// by format but generally boil down to either just under 2GB or just under
+// 1GB. When the decoded image would be larger than this, stb_image decoding
+// will fail.
+//
+// Additionally, stb_image will reject image files that have any of their
+// dimensions set to a larger value than the configurable STBI_MAX_DIMENSIONS,
+// which defaults to 2**24 = 16777216 pixels. Due to the above memory limit,
+// the only way to have an image with such dimensions load correctly
+// is for it to have a rather extreme aspect ratio. Either way, the
+// assumption here is that such larger images are likely to be malformed
+// or malicious. If you do need to load an image with individual dimensions
+// larger than that, and it still fits in the overall size limit, you can
+// #define STBI_MAX_DIMENSIONS on your own to be something larger.
+//
+// ===========================================================================
+//
+// UNICODE:
+//
+//   If compiling for Windows and you wish to use Unicode filenames, compile
+//   with
+//       #define STBI_WINDOWS_UTF8
+//   and pass utf8-encoded filenames. Call stbi_convert_wchar_to_utf8 to convert
+//   Windows wchar_t filenames to utf8.
+//
+// ===========================================================================
+//
+// Philosophy
+//
+// stb libraries are designed with the following priorities:
+//
+//    1. easy to use
+//    2. easy to maintain
+//    3. good performance
+//
+// Sometimes I let "good performance" creep up in priority over "easy to maintain",
+// and for best performance I may provide less-easy-to-use APIs that give higher
+// performance, in addition to the easy-to-use ones. Nevertheless, it's important
+// to keep in mind that from the standpoint of you, a client of this library,
+// all you care about is #1 and #3, and stb libraries DO NOT emphasize #3 above all.
+//
+// Some secondary priorities arise directly from the first two, some of which
+// provide more explicit reasons why performance can't be emphasized.
+//
+//    - Portable ("ease of use")
+//    - Small source code footprint ("easy to maintain")
+//    - No dependencies ("ease of use")
+//
+// ===========================================================================
+//
+// I/O callbacks
+//
+// I/O callbacks allow you to read from arbitrary sources, like packaged
+// files or some other source. Data read from callbacks are processed
+// through a small internal buffer (currently 128 bytes) to try to reduce
+// overhead.
+//
+// The three functions you must define are "read" (reads some bytes of data),
+// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end).
+//
+// ===========================================================================
+//
+// SIMD support
+//
+// The JPEG decoder will try to automatically use SIMD kernels on x86 when
+// supported by the compiler. For ARM Neon support, you must explicitly
+// request it.
+//
+// (The old do-it-yourself SIMD API is no longer supported in the current
+// code.)
+//
+// On x86, SSE2 will automatically be used when available based on a run-time
+// test; if not, the generic C versions are used as a fall-back. On ARM targets,
+// the typical path is to have separate builds for NEON and non-NEON devices
+// (at least this is true for iOS and Android). Therefore, the NEON support is
+// toggled by a build flag: define STBI_NEON to get NEON loops.
+//
+// If for some reason you do not want to use any of SIMD code, or if
+// you have issues compiling it, you can disable it entirely by
+// defining STBI_NO_SIMD.
+//
+// ===========================================================================
+//
+// HDR image support   (disable by defining STBI_NO_HDR)
+//
+// stb_image supports loading HDR images in general, and currently the Radiance
+// .HDR file format specifically. You can still load any file through the existing
+// interface; if you attempt to load an HDR file, it will be automatically remapped
+// to LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1;
+// both of these constants can be reconfigured through this interface:
+//
+//     stbi_hdr_to_ldr_gamma(2.2f);
+//     stbi_hdr_to_ldr_scale(1.0f);
+//
+// (note, do not use _inverse_ constants; stbi_image will invert them
+// appropriately).
+//
+// Additionally, there is a new, parallel interface for loading files as
+// (linear) floats to preserve the full dynamic range:
+//
+//    float *data = stbi_loadf(filename, &x, &y, &n, 0);
+//
+// If you load LDR images through this interface, those images will
+// be promoted to floating point values, run through the inverse of
+// constants corresponding to the above:
+//
+//     stbi_ldr_to_hdr_scale(1.0f);
+//     stbi_ldr_to_hdr_gamma(2.2f);
+//
+// Finally, given a filename (or an open file or memory block--see header
+// file for details) containing image data, you can query for the "most
+// appropriate" interface to use (that is, whether the image is HDR or
+// not), using:
+//
+//     stbi_is_hdr(char *filename);
+//
+// ===========================================================================
+//
+// iPhone PNG support:
+//
+// We optionally support converting iPhone-formatted PNGs (which store
+// premultiplied BGRA) back to RGB, even though they're internally encoded
+// differently. To enable this conversion, call
+// stbi_convert_iphone_png_to_rgb(1).
+//
+// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
+// pixel to remove any premultiplied alpha *only* if the image file explicitly
+// says there's premultiplied data (currently only happens in iPhone images,
+// and only if iPhone convert-to-rgb processing is on).
+//
+// ===========================================================================
+//
+// ADDITIONAL CONFIGURATION
+//
+//  - You can suppress implementation of any of the decoders to reduce
+//    your code footprint by #defining one or more of the following
+//    symbols before creating the implementation.
+//
+//        STBI_NO_JPEG
+//        STBI_NO_PNG
+//        STBI_NO_BMP
+//        STBI_NO_PSD
+//        STBI_NO_TGA
+//        STBI_NO_GIF
+//        STBI_NO_HDR
+//        STBI_NO_PIC
+//        STBI_NO_PNM   (.ppm and .pgm)
+//
+//  - You can request *only* certain decoders and suppress all other ones
+//    (this will be more forward-compatible, as addition of new decoders
+//    doesn't require you to disable them explicitly):
+//
+//        STBI_ONLY_JPEG
+//        STBI_ONLY_PNG
+//        STBI_ONLY_BMP
+//        STBI_ONLY_PSD
+//        STBI_ONLY_TGA
+//        STBI_ONLY_GIF
+//        STBI_ONLY_HDR
+//        STBI_ONLY_PIC
+//        STBI_ONLY_PNM   (.ppm and .pgm)
+//
+//   - If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still
+//     want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB
+//
+//  - If you define STBI_MAX_DIMENSIONS, stb_image will reject images greater
+//    than that size (in either width or height) without further processing.
+//    This is to let programs in the wild set an upper bound to prevent
+//    denial-of-service attacks on untrusted data, as one could generate a
+//    valid image of gigantic dimensions and force stb_image to allocate a
+//    huge block of memory and spend disproportionate time decoding it. By
+//    default this is set to (1 << 24), which is 16777216, but that's still
+//    very big.
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif // STBI_NO_STDIO
+
+#define STBI_VERSION 1
+
+enum
+{
+   STBI_default = 0, // only used for desired_channels
+
+   STBI_grey       = 1,
+   STBI_grey_alpha = 2,
+   STBI_rgb        = 3,
+   STBI_rgb_alpha  = 4
+};
+
+#include <stdlib.h>
+typedef unsigned char stbi_uc;
+typedef unsigned short stbi_us;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef STBIDEF
+#ifdef STB_IMAGE_STATIC
+#define STBIDEF static
+#else
+#define STBIDEF extern
+#endif
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PRIMARY API - works on images of any type
+//
+
+//
+// load image by filename, open file, or memory buffer
+//
+
+typedef struct
+{
+   int      (*read)  (void *user,char *data,int size);   // fill 'data' with 'size' bytes.  return number of bytes actually read
+   void     (*skip)  (void *user,int n);                 // skip the next 'n' bytes, or 'unget' the last -n bytes if negative
+   int      (*eof)   (void *user);                       // returns nonzero if we are at end of file/data
+} stbi_io_callbacks;
+
+////////////////////////////////////
+//
+// 8-bits-per-channel interface
+//
+
+STBIDEF stbi_uc *stbi_load_from_memory   (stbi_uc           const *buffer, int len   , int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk  , void *user, int *x, int *y, int *channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_uc *stbi_load            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_uc *stbi_load_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+// for stbi_load_from_file, file pointer is left pointing immediately after image
+#endif
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
+#endif
+
+#ifdef STBI_WINDOWS_UTF8
+STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
+#endif
+
+////////////////////////////////////
+//
+// 16-bits-per-channel interface
+//
+
+STBIDEF stbi_us *stbi_load_16_from_memory   (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_us *stbi_load_16          (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+#endif
+
+////////////////////////////////////
+//
+// float-per-channel interface
+//
+#ifndef STBI_NO_LINEAR
+   STBIDEF float *stbi_loadf_from_memory     (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
+   STBIDEF float *stbi_loadf_from_callbacks  (stbi_io_callbacks const *clbk, void *user, int *x, int *y,  int *channels_in_file, int desired_channels);
+
+   #ifndef STBI_NO_STDIO
+   STBIDEF float *stbi_loadf            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+   STBIDEF float *stbi_loadf_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+   #endif
+#endif
+
+#ifndef STBI_NO_HDR
+   STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma);
+   STBIDEF void   stbi_hdr_to_ldr_scale(float scale);
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_LINEAR
+   STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma);
+   STBIDEF void   stbi_ldr_to_hdr_scale(float scale);
+#endif // STBI_NO_LINEAR
+
+// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR
+STBIDEF int    stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user);
+STBIDEF int    stbi_is_hdr_from_memory(stbi_uc const *buffer, int len);
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_is_hdr          (char const *filename);
+STBIDEF int      stbi_is_hdr_from_file(FILE *f);
+#endif // STBI_NO_STDIO
+
+
+// get a VERY brief reason for failure
+// on most compilers (and ALL modern mainstream compilers) this is threadsafe
+STBIDEF const char *stbi_failure_reason  (void);
+
+// free the loaded image -- this is just free()
+STBIDEF void     stbi_image_free      (void *retval_from_stbi_load);
+
+// get image dimensions & components without fully decoding
+STBIDEF int      stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp);
+STBIDEF int      stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp);
+STBIDEF int      stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len);
+STBIDEF int      stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *clbk, void *user);
+
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_info               (char const *filename,     int *x, int *y, int *comp);
+STBIDEF int      stbi_info_from_file     (FILE *f,                  int *x, int *y, int *comp);
+STBIDEF int      stbi_is_16_bit          (char const *filename);
+STBIDEF int      stbi_is_16_bit_from_file(FILE *f);
+#endif
+
+
+
+// for image formats that explicitly notate that they have premultiplied alpha,
+// we just return the colors as stored in the file. set this flag to force
+// unpremultiplication. results are undefined if the unpremultiply overflow.
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply);
+
+// indicate whether we should process iphone images back to canonical format,
+// or just pass them through "as-is"
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert);
+
+// flip the image vertically, so the first pixel in the output array is the bottom left
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
+
+// as above, but only applies to images loaded on the thread that calls the function
+// this function is only available if your compiler supports thread-local variables;
+// calling it will fail to link if your compiler doesn't
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply);
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert);
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip);
+
+// ZLIB client - used by PNG, available for other purposes
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen);
+STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header);
+STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen);
+STBIDEF int   stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
+
+STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen);
+STBIDEF int   stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+//
+//
+////   end header file   /////////////////////////////////////////////////////
+#endif // STBI_INCLUDE_STB_IMAGE_H
+
+#ifdef STB_IMAGE_IMPLEMENTATION
+
+#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \
+  || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \
+  || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \
+  || defined(STBI_ONLY_ZLIB)
+   #ifndef STBI_ONLY_JPEG
+   #define STBI_NO_JPEG
+   #endif
+   #ifndef STBI_ONLY_PNG
+   #define STBI_NO_PNG
+   #endif
+   #ifndef STBI_ONLY_BMP
+   #define STBI_NO_BMP
+   #endif
+   #ifndef STBI_ONLY_PSD
+   #define STBI_NO_PSD
+   #endif
+   #ifndef STBI_ONLY_TGA
+   #define STBI_NO_TGA
+   #endif
+   #ifndef STBI_ONLY_GIF
+   #define STBI_NO_GIF
+   #endif
+   #ifndef STBI_ONLY_HDR
+   #define STBI_NO_HDR
+   #endif
+   #ifndef STBI_ONLY_PIC
+   #define STBI_NO_PIC
+   #endif
+   #ifndef STBI_ONLY_PNM
+   #define STBI_NO_PNM
+   #endif
+#endif
+
+#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB)
+#define STBI_NO_ZLIB
+#endif
+
+
+#include <stdarg.h>
+#include <stddef.h> // ptrdiff_t on osx
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
+#include <math.h>  // ldexp, pow
+#endif
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif
+
+#ifndef STBI_ASSERT
+#include <assert.h>
+#define STBI_ASSERT(x) assert(x)
+#endif
+
+#ifdef __cplusplus
+#define STBI_EXTERN extern "C"
+#else
+#define STBI_EXTERN extern
+#endif
+
+
+#ifndef _MSC_VER
+   #ifdef __cplusplus
+   #define stbi_inline inline
+   #else
+   #define stbi_inline
+   #endif
+#else
+   #define stbi_inline __forceinline
+#endif
+
+#ifndef STBI_NO_THREAD_LOCALS
+   #if defined(__cplusplus) &&  __cplusplus >= 201103L
+      #define STBI_THREAD_LOCAL       thread_local
+   #elif defined(__GNUC__) && __GNUC__ < 5
+      #define STBI_THREAD_LOCAL       __thread
+   #elif defined(_MSC_VER)
+      #define STBI_THREAD_LOCAL       __declspec(thread)
+   #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
+      #define STBI_THREAD_LOCAL       _Thread_local
+   #endif
+
+   #ifndef STBI_THREAD_LOCAL
+      #if defined(__GNUC__)
+        #define STBI_THREAD_LOCAL       __thread
+      #endif
+   #endif
+#endif
+
+#if defined(_MSC_VER) || defined(__SYMBIAN32__)
+typedef unsigned short stbi__uint16;
+typedef   signed short stbi__int16;
+typedef unsigned int   stbi__uint32;
+typedef   signed int   stbi__int32;
+#else
+#include <stdint.h>
+typedef uint16_t stbi__uint16;
+typedef int16_t  stbi__int16;
+typedef uint32_t stbi__uint32;
+typedef int32_t  stbi__int32;
+#endif
+
+// should produce compiler error if size is wrong
+typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1];
+
+#ifdef _MSC_VER
+#define STBI_NOTUSED(v)  (void)(v)
+#else
+#define STBI_NOTUSED(v)  (void)sizeof(v)
+#endif
+
+#ifdef _MSC_VER
+#define STBI_HAS_LROTL
+#endif
+
+#ifdef STBI_HAS_LROTL
+   #define stbi_lrot(x,y)  _lrotl(x,y)
+#else
+   #define stbi_lrot(x,y)  (((x) << (y)) | ((x) >> (-(y) & 31)))
+#endif
+
+#if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED))
+// ok
+#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC) && !defined(STBI_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC (or STBI_REALLOC_SIZED)."
+#endif
+
+#ifndef STBI_MALLOC
+#define STBI_MALLOC(sz)           malloc(sz)
+#define STBI_REALLOC(p,newsz)     realloc(p,newsz)
+#define STBI_FREE(p)              free(p)
+#endif
+
+#ifndef STBI_REALLOC_SIZED
+#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz)
+#endif
+
+// x86/x64 detection
+#if defined(__x86_64__) || defined(_M_X64)
+#define STBI__X64_TARGET
+#elif defined(__i386) || defined(_M_IX86)
+#define STBI__X86_TARGET
+#endif
+
+#if defined(__GNUC__) && defined(STBI__X86_TARGET) && !defined(__SSE2__) && !defined(STBI_NO_SIMD)
+// gcc doesn't support sse2 intrinsics unless you compile with -msse2,
+// which in turn means it gets to use SSE2 everywhere. This is unfortunate,
+// but previous attempts to provide the SSE2 functions with runtime
+// detection caused numerous issues. The way architecture extensions are
+// exposed in GCC/Clang is, sadly, not really suited for one-file libs.
+// New behavior: if compiled with -msse2, we use SSE2 without any
+// detection; if not, we don't use it at all.
+#define STBI_NO_SIMD
+#endif
+
+#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD)
+// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET
+//
+// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the
+// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant.
+// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not
+// simultaneously enabling "-mstackrealign".
+//
+// See https://github.com/nothings/stb/issues/81 for more information.
+//
+// So default to no SSE2 on 32-bit MinGW. If you've read this far and added
+// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2.
+#define STBI_NO_SIMD
+#endif
+
+#if !defined(STBI_NO_SIMD) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET))
+#define STBI_SSE2
+#include <emmintrin.h>
+
+#ifdef _MSC_VER
+
+#if _MSC_VER >= 1400  // not VC6
+#include <intrin.h> // __cpuid
+static int stbi__cpuid3(void)
+{
+   int info[4];
+   __cpuid(info,1);
+   return info[3];
+}
+#else
+static int stbi__cpuid3(void)
+{
+   int res;
+   __asm {
+      mov  eax,1
+      cpuid
+      mov  res,edx
+   }
+   return res;
+}
+#endif
+
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void)
+{
+   int info3 = stbi__cpuid3();
+   return ((info3 >> 26) & 1) != 0;
+}
+#endif
+
+#else // assume GCC-style if not VC++
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void)
+{
+   // If we're even attempting to compile this on GCC/Clang, that means
+   // -msse2 is on, which means the compiler is allowed to use SSE2
+   // instructions at will, and so are we.
+   return 1;
+}
+#endif
+
+#endif
+#endif
+
+// ARM NEON
+#if defined(STBI_NO_SIMD) && defined(STBI_NEON)
+#undef STBI_NEON
+#endif
+
+#ifdef STBI_NEON
+#include <arm_neon.h>
+#ifdef _MSC_VER
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+#else
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+#endif
+#endif
+
+#ifndef STBI_SIMD_ALIGN
+#define STBI_SIMD_ALIGN(type, name) type name
+#endif
+
+#ifndef STBI_MAX_DIMENSIONS
+#define STBI_MAX_DIMENSIONS (1 << 24)
+#endif
+
+///////////////////////////////////////////////
+//
+//  stbi__context struct and start_xxx functions
+
+// stbi__context structure is our basic context used by all images, so it
+// contains all the IO context, plus some basic image information
+typedef struct
+{
+   stbi__uint32 img_x, img_y;
+   int img_n, img_out_n;
+
+   stbi_io_callbacks io;
+   void *io_user_data;
+
+   int read_from_callbacks;
+   int buflen;
+   stbi_uc buffer_start[128];
+   int callback_already_read;
+
+   stbi_uc *img_buffer, *img_buffer_end;
+   stbi_uc *img_buffer_original, *img_buffer_original_end;
+} stbi__context;
+
+
+static void stbi__refill_buffer(stbi__context *s);
+
+// initialize a memory-decode context
+static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len)
+{
+   s->io.read = NULL;
+   s->read_from_callbacks = 0;
+   s->callback_already_read = 0;
+   s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer;
+   s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len;
+}
+
+// initialize a callback-based context
+static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user)
+{
+   s->io = *c;
+   s->io_user_data = user;
+   s->buflen = sizeof(s->buffer_start);
+   s->read_from_callbacks = 1;
+   s->callback_already_read = 0;
+   s->img_buffer = s->img_buffer_original = s->buffer_start;
+   stbi__refill_buffer(s);
+   s->img_buffer_original_end = s->img_buffer_end;
+}
+
+#ifndef STBI_NO_STDIO
+
+static int stbi__stdio_read(void *user, char *data, int size)
+{
+   return (int) fread(data,1,size,(FILE*) user);
+}
+
+static void stbi__stdio_skip(void *user, int n)
+{
+   int ch;
+   fseek((FILE*) user, n, SEEK_CUR);
+   ch = fgetc((FILE*) user);  /* have to read a byte to reset feof()'s flag */
+   if (ch != EOF) {
+      ungetc(ch, (FILE *) user);  /* push byte back onto stream if valid. */
+   }
+}
+
+static int stbi__stdio_eof(void *user)
+{
+   return feof((FILE*) user) || ferror((FILE *) user);
+}
+
+static stbi_io_callbacks stbi__stdio_callbacks =
+{
+   stbi__stdio_read,
+   stbi__stdio_skip,
+   stbi__stdio_eof,
+};
+
+static void stbi__start_file(stbi__context *s, FILE *f)
+{
+   stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f);
+}
+
+//static void stop_file(stbi__context *s) { }
+
+#endif // !STBI_NO_STDIO
+
+static void stbi__rewind(stbi__context *s)
+{
+   // conceptually rewind SHOULD rewind to the beginning of the stream,
+   // but we just rewind to the beginning of the initial buffer, because
+   // we only use it after doing 'test', which only ever looks at at most 92 bytes
+   s->img_buffer = s->img_buffer_original;
+   s->img_buffer_end = s->img_buffer_original_end;
+}
+
+enum
+{
+   STBI_ORDER_RGB,
+   STBI_ORDER_BGR
+};
+
+typedef struct
+{
+   int bits_per_channel;
+   int num_channels;
+   int channel_order;
+} stbi__result_info;
+
+#ifndef STBI_NO_JPEG
+static int      stbi__jpeg_test(stbi__context *s);
+static void    *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PNG
+static int      stbi__png_test(stbi__context *s);
+static void    *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__png_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__png_is16(stbi__context *s);
+#endif
+
+#ifndef STBI_NO_BMP
+static int      stbi__bmp_test(stbi__context *s);
+static void    *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_TGA
+static int      stbi__tga_test(stbi__context *s);
+static void    *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__tga_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PSD
+static int      stbi__psd_test(stbi__context *s);
+static void    *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc);
+static int      stbi__psd_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__psd_is16(stbi__context *s);
+#endif
+
+#ifndef STBI_NO_HDR
+static int      stbi__hdr_test(stbi__context *s);
+static float   *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PIC
+static int      stbi__pic_test(stbi__context *s);
+static void    *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__pic_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_GIF
+static int      stbi__gif_test(stbi__context *s);
+static void    *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static void    *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
+static int      stbi__gif_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PNM
+static int      stbi__pnm_test(stbi__context *s);
+static void    *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__pnm_is16(stbi__context *s);
+#endif
+
+static
+#ifdef STBI_THREAD_LOCAL
+STBI_THREAD_LOCAL
+#endif
+const char *stbi__g_failure_reason;
+
+STBIDEF const char *stbi_failure_reason(void)
+{
+   return stbi__g_failure_reason;
+}
+
+#ifndef STBI_NO_FAILURE_STRINGS
+static int stbi__err(const char *str)
+{
+   stbi__g_failure_reason = str;
+   return 0;
+}
+#endif
+
+static void *stbi__malloc(size_t size)
+{
+    return STBI_MALLOC(size);
+}
+
+// stb_image uses ints pervasively, including for offset calculations.
+// therefore the largest decoded image size we can support with the
+// current code, even on 64-bit targets, is INT_MAX. this is not a
+// significant limitation for the intended use case.
+//
+// we do, however, need to make sure our size calculations don't
+// overflow. hence a few helper functions for size calculations that
+// multiply integers together, making sure that they're non-negative
+// and no overflow occurs.
+
+// return 1 if the sum is valid, 0 on overflow.
+// negative terms are considered invalid.
+static int stbi__addsizes_valid(int a, int b)
+{
+   if (b < 0) return 0;
+   // now 0 <= b <= INT_MAX, hence also
+   // 0 <= INT_MAX - b <= INTMAX.
+   // And "a + b <= INT_MAX" (which might overflow) is the
+   // same as a <= INT_MAX - b (no overflow)
+   return a <= INT_MAX - b;
+}
+
+// returns 1 if the product is valid, 0 on overflow.
+// negative factors are considered invalid.
+static int stbi__mul2sizes_valid(int a, int b)
+{
+   if (a < 0 || b < 0) return 0;
+   if (b == 0) return 1; // mul-by-0 is always safe
+   // portable way to check for no overflows in a*b
+   return a <= INT_MAX/b;
+}
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad2sizes_valid(int a, int b, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add);
+}
+#endif
+
+// returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad3sizes_valid(int a, int b, int c, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
+      stbi__addsizes_valid(a*b*c, add);
+}
+
+// returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
+      stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add);
+}
+#endif
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// mallocs with size overflow checking
+static void *stbi__malloc_mad2(int a, int b, int add)
+{
+   if (!stbi__mad2sizes_valid(a, b, add)) return NULL;
+   return stbi__malloc(a*b + add);
+}
+#endif
+
+static void *stbi__malloc_mad3(int a, int b, int c, int add)
+{
+   if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL;
+   return stbi__malloc(a*b*c + add);
+}
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static void *stbi__malloc_mad4(int a, int b, int c, int d, int add)
+{
+   if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL;
+   return stbi__malloc(a*b*c*d + add);
+}
+#endif
+
+// returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow.
+static int stbi__addints_valid(int a, int b)
+{
+   if ((a >= 0) != (b >= 0)) return 1; // a and b have different signs, so no overflow
+   if (a < 0 && b < 0) return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0.
+   return a <= INT_MAX - b;
+}
+
+// returns 1 if the product of two ints fits in a signed short, 0 on overflow.
+static int stbi__mul2shorts_valid(int a, int b)
+{
+   if (b == 0 || b == -1) return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow
+   if ((a >= 0) == (b >= 0)) return a <= SHRT_MAX/b; // product is positive, so similar to mul2sizes_valid
+   if (b < 0) return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN
+   return a >= SHRT_MIN / b;
+}
+
+// stbi__err - error
+// stbi__errpf - error returning pointer to float
+// stbi__errpuc - error returning pointer to unsigned char
+
+#ifdef STBI_NO_FAILURE_STRINGS
+   #define stbi__err(x,y)  0
+#elif defined(STBI_FAILURE_USERMSG)
+   #define stbi__err(x,y)  stbi__err(y)
+#else
+   #define stbi__err(x,y)  stbi__err(x)
+#endif
+
+#define stbi__errpf(x,y)   ((float *)(size_t) (stbi__err(x,y)?NULL:NULL))
+#define stbi__errpuc(x,y)  ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL))
+
+STBIDEF void stbi_image_free(void *retval_from_stbi_load)
+{
+   STBI_FREE(retval_from_stbi_load);
+}
+
+#ifndef STBI_NO_LINEAR
+static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp);
+#endif
+
+#ifndef STBI_NO_HDR
+static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp);
+#endif
+
+static int stbi__vertically_flip_on_load_global = 0;
+
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip)
+{
+   stbi__vertically_flip_on_load_global = flag_true_if_should_flip;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__vertically_flip_on_load  stbi__vertically_flip_on_load_global
+#else
+static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set;
+
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip)
+{
+   stbi__vertically_flip_on_load_local = flag_true_if_should_flip;
+   stbi__vertically_flip_on_load_set = 1;
+}
+
+#define stbi__vertically_flip_on_load  (stbi__vertically_flip_on_load_set       \
+                                         ? stbi__vertically_flip_on_load_local  \
+                                         : stbi__vertically_flip_on_load_global)
+#endif // STBI_THREAD_LOCAL
+
+static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
+{
+   memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields
+   ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed
+   ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order
+   ri->num_channels = 0;
+
+   // test the formats with a very explicit header first (at least a FOURCC
+   // or distinctive magic number first)
+   #ifndef STBI_NO_PNG
+   if (stbi__png_test(s))  return stbi__png_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_BMP
+   if (stbi__bmp_test(s))  return stbi__bmp_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_GIF
+   if (stbi__gif_test(s))  return stbi__gif_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_test(s))  return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc);
+   #else
+   STBI_NOTUSED(bpc);
+   #endif
+   #ifndef STBI_NO_PIC
+   if (stbi__pic_test(s))  return stbi__pic_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   // then the formats that can end up attempting to load with just 1 or 2
+   // bytes matching expectations; these are prone to false positives, so
+   // try them later
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_test(s))  return stbi__pnm_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_test(s)) {
+      float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri);
+      return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp);
+   }
+   #endif
+
+   #ifndef STBI_NO_TGA
+   // test tga last because it's a crappy test!
+   if (stbi__tga_test(s))
+      return stbi__tga_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels)
+{
+   int i;
+   int img_len = w * h * channels;
+   stbi_uc *reduced;
+
+   reduced = (stbi_uc *) stbi__malloc(img_len);
+   if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory");
+
+   for (i = 0; i < img_len; ++i)
+      reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling
+
+   STBI_FREE(orig);
+   return reduced;
+}
+
+static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels)
+{
+   int i;
+   int img_len = w * h * channels;
+   stbi__uint16 *enlarged;
+
+   enlarged = (stbi__uint16 *) stbi__malloc(img_len*2);
+   if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
+
+   for (i = 0; i < img_len; ++i)
+      enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff
+
+   STBI_FREE(orig);
+   return enlarged;
+}
+
+static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel)
+{
+   int row;
+   size_t bytes_per_row = (size_t)w * bytes_per_pixel;
+   stbi_uc temp[2048];
+   stbi_uc *bytes = (stbi_uc *)image;
+
+   for (row = 0; row < (h>>1); row++) {
+      stbi_uc *row0 = bytes + row*bytes_per_row;
+      stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row;
+      // swap row0 with row1
+      size_t bytes_left = bytes_per_row;
+      while (bytes_left) {
+         size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp);
+         memcpy(temp, row0, bytes_copy);
+         memcpy(row0, row1, bytes_copy);
+         memcpy(row1, temp, bytes_copy);
+         row0 += bytes_copy;
+         row1 += bytes_copy;
+         bytes_left -= bytes_copy;
+      }
+   }
+}
+
+#ifndef STBI_NO_GIF
+static void stbi__vertical_flip_slices(void *image, int w, int h, int z, int bytes_per_pixel)
+{
+   int slice;
+   int slice_size = w * h * bytes_per_pixel;
+
+   stbi_uc *bytes = (stbi_uc *)image;
+   for (slice = 0; slice < z; ++slice) {
+      stbi__vertical_flip(bytes, w, h, bytes_per_pixel);
+      bytes += slice_size;
+   }
+}
+#endif
+
+static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__result_info ri;
+   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8);
+
+   if (result == NULL)
+      return NULL;
+
+   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+   if (ri.bits_per_channel != 8) {
+      result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
+      ri.bits_per_channel = 8;
+   }
+
+   // @TODO: move stbi__convert_format to here
+
+   if (stbi__vertically_flip_on_load) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc));
+   }
+
+   return (unsigned char *) result;
+}
+
+static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__result_info ri;
+   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16);
+
+   if (result == NULL)
+      return NULL;
+
+   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+   if (ri.bits_per_channel != 16) {
+      result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
+      ri.bits_per_channel = 16;
+   }
+
+   // @TODO: move stbi__convert_format16 to here
+   // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision
+
+   if (stbi__vertically_flip_on_load) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16));
+   }
+
+   return (stbi__uint16 *) result;
+}
+
+#if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR)
+static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp)
+{
+   if (stbi__vertically_flip_on_load && result != NULL) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(float));
+   }
+}
+#endif
+
+#ifndef STBI_NO_STDIO
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
+STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
+#endif
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
+{
+	return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE *stbi__fopen(char const *filename, char const *mode)
+{
+   FILE *f;
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+   wchar_t wMode[64];
+   wchar_t wFilename[1024];
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
+      return 0;
+
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
+      return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+	if (0 != _wfopen_s(&f, wFilename, wMode))
+		f = 0;
+#else
+   f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != fopen_s(&f, filename, mode))
+      f=0;
+#else
+   f = fopen(filename, mode);
+#endif
+   return f;
+}
+
+
+STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   unsigned char *result;
+   if (!f) return stbi__errpuc("can't fopen", "Unable to open file");
+   result = stbi_load_from_file(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   unsigned char *result;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+   if (result) {
+      // need to 'unget' all the characters in the IO buffer
+      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
+   }
+   return result;
+}
+
+STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__uint16 *result;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp);
+   if (result) {
+      // need to 'unget' all the characters in the IO buffer
+      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
+   }
+   return result;
+}
+
+STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   stbi__uint16 *result;
+   if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file");
+   result = stbi_load_from_file_16(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+
+#endif //!STBI_NO_STDIO
+
+STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
+}
+
+STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
+}
+
+STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+}
+
+STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+}
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
+{
+   unsigned char *result;
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+
+   result = (unsigned char*) stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp);
+   if (stbi__vertically_flip_on_load) {
+      stbi__vertical_flip_slices( result, *x, *y, *z, *comp );
+   }
+
+   return result;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   unsigned char *data;
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_test(s)) {
+      stbi__result_info ri;
+      float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri);
+      if (hdr_data)
+         stbi__float_postprocess(hdr_data,x,y,comp,req_comp);
+      return hdr_data;
+   }
+   #endif
+   data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp);
+   if (data)
+      return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp);
+   return stbi__errpf("unknown image type", "Image not of any known type, or corrupt");
+}
+
+STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+
+STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   float *result;
+   FILE *f = stbi__fopen(filename, "rb");
+   if (!f) return stbi__errpf("can't fopen", "Unable to open file");
+   result = stbi_loadf_from_file(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_file(&s,f);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+#endif // !STBI_NO_STDIO
+
+#endif // !STBI_NO_LINEAR
+
+// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is
+// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always
+// reports false!
+
+STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len)
+{
+   #ifndef STBI_NO_HDR
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__hdr_test(&s);
+   #else
+   STBI_NOTUSED(buffer);
+   STBI_NOTUSED(len);
+   return 0;
+   #endif
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_is_hdr          (char const *filename)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   int result=0;
+   if (f) {
+      result = stbi_is_hdr_from_file(f);
+      fclose(f);
+   }
+   return result;
+}
+
+STBIDEF int stbi_is_hdr_from_file(FILE *f)
+{
+   #ifndef STBI_NO_HDR
+   long pos = ftell(f);
+   int res;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   res = stbi__hdr_test(&s);
+   fseek(f, pos, SEEK_SET);
+   return res;
+   #else
+   STBI_NOTUSED(f);
+   return 0;
+   #endif
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int      stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user)
+{
+   #ifndef STBI_NO_HDR
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__hdr_test(&s);
+   #else
+   STBI_NOTUSED(clbk);
+   STBI_NOTUSED(user);
+   return 0;
+   #endif
+}
+
+#ifndef STBI_NO_LINEAR
+static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f;
+
+STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; }
+STBIDEF void   stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; }
+#endif
+
+static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f;
+
+STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; }
+STBIDEF void   stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; }
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// Common code used by all image loaders
+//
+
+enum
+{
+   STBI__SCAN_load=0,
+   STBI__SCAN_type,
+   STBI__SCAN_header
+};
+
+static void stbi__refill_buffer(stbi__context *s)
+{
+   int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen);
+   s->callback_already_read += (int) (s->img_buffer - s->img_buffer_original);
+   if (n == 0) {
+      // at end of file, treat same as if from memory, but need to handle case
+      // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file
+      s->read_from_callbacks = 0;
+      s->img_buffer = s->buffer_start;
+      s->img_buffer_end = s->buffer_start+1;
+      *s->img_buffer = 0;
+   } else {
+      s->img_buffer = s->buffer_start;
+      s->img_buffer_end = s->buffer_start + n;
+   }
+}
+
+stbi_inline static stbi_uc stbi__get8(stbi__context *s)
+{
+   if (s->img_buffer < s->img_buffer_end)
+      return *s->img_buffer++;
+   if (s->read_from_callbacks) {
+      stbi__refill_buffer(s);
+      return *s->img_buffer++;
+   }
+   return 0;
+}
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+stbi_inline static int stbi__at_eof(stbi__context *s)
+{
+   if (s->io.read) {
+      if (!(s->io.eof)(s->io_user_data)) return 0;
+      // if feof() is true, check if buffer = end
+      // special case: we've only got the special 0 character at the end
+      if (s->read_from_callbacks == 0) return 1;
+   }
+
+   return s->img_buffer >= s->img_buffer_end;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC)
+// nothing
+#else
+static void stbi__skip(stbi__context *s, int n)
+{
+   if (n == 0) return;  // already there!
+   if (n < 0) {
+      s->img_buffer = s->img_buffer_end;
+      return;
+   }
+   if (s->io.read) {
+      int blen = (int) (s->img_buffer_end - s->img_buffer);
+      if (blen < n) {
+         s->img_buffer = s->img_buffer_end;
+         (s->io.skip)(s->io_user_data, n - blen);
+         return;
+      }
+   }
+   s->img_buffer += n;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM)
+// nothing
+#else
+static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n)
+{
+   if (s->io.read) {
+      int blen = (int) (s->img_buffer_end - s->img_buffer);
+      if (blen < n) {
+         int res, count;
+
+         memcpy(buffer, s->img_buffer, blen);
+
+         count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen);
+         res = (count == (n-blen));
+         s->img_buffer = s->img_buffer_end;
+         return res;
+      }
+   }
+
+   if (s->img_buffer+n <= s->img_buffer_end) {
+      memcpy(buffer, s->img_buffer, n);
+      s->img_buffer += n;
+      return 1;
+   } else
+      return 0;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static int stbi__get16be(stbi__context *s)
+{
+   int z = stbi__get8(s);
+   return (z << 8) + stbi__get8(s);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static stbi__uint32 stbi__get32be(stbi__context *s)
+{
+   stbi__uint32 z = stbi__get16be(s);
+   return (z << 16) + stbi__get16be(s);
+}
+#endif
+
+#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF)
+// nothing
+#else
+static int stbi__get16le(stbi__context *s)
+{
+   int z = stbi__get8(s);
+   return z + (stbi__get8(s) << 8);
+}
+#endif
+
+#ifndef STBI_NO_BMP
+static stbi__uint32 stbi__get32le(stbi__context *s)
+{
+   stbi__uint32 z = stbi__get16le(s);
+   z += (stbi__uint32)stbi__get16le(s) << 16;
+   return z;
+}
+#endif
+
+#define STBI__BYTECAST(x)  ((stbi_uc) ((x) & 255))  // truncate int to byte without warnings
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+//////////////////////////////////////////////////////////////////////////////
+//
+//  generic converter from built-in img_n to req_comp
+//    individual types do this automatically as much as possible (e.g. jpeg
+//    does all cases internally since it needs to colorspace convert anyway,
+//    and it never has alpha, so very few cases ). png can automatically
+//    interleave an alpha=255 channel, but falls back to this for other cases
+//
+//  assume data buffer is malloced, so malloc a new one and free that one
+//  only failure mode is malloc failing
+
+static stbi_uc stbi__compute_y(int r, int g, int b)
+{
+   return (stbi_uc) (((r*77) + (g*150) +  (29*b)) >> 8);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y)
+{
+   int i,j;
+   unsigned char *good;
+
+   if (req_comp == img_n) return data;
+   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+   good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0);
+   if (good == NULL) {
+      STBI_FREE(data);
+      return stbi__errpuc("outofmem", "Out of memory");
+   }
+
+   for (j=0; j < (int) y; ++j) {
+      unsigned char *src  = data + j * x * img_n   ;
+      unsigned char *dest = good + j * x * req_comp;
+
+      #define STBI__COMBO(a,b)  ((a)*8+(b))
+      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
+      // convert source image with img_n components to one with req_comp components;
+      // avoid switch per pixel, so use switch per scanline and massive macros
+      switch (STBI__COMBO(img_n, req_comp)) {
+         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=255;                                     } break;
+         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
+         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=255;                     } break;
+         STBI__CASE(2,1) { dest[0]=src[0];                                                  } break;
+         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
+         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                  } break;
+         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=255;        } break;
+         STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = 255;    } break;
+         STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = src[3]; } break;
+         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                    } break;
+         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return stbi__errpuc("unsupported", "Unsupported format conversion");
+      }
+      #undef STBI__CASE
+   }
+
+   STBI_FREE(data);
+   return good;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 stbi__compute_y_16(int r, int g, int b)
+{
+   return (stbi__uint16) (((r*77) + (g*150) +  (29*b)) >> 8);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y)
+{
+   int i,j;
+   stbi__uint16 *good;
+
+   if (req_comp == img_n) return data;
+   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+   good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2);
+   if (good == NULL) {
+      STBI_FREE(data);
+      return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
+   }
+
+   for (j=0; j < (int) y; ++j) {
+      stbi__uint16 *src  = data + j * x * img_n   ;
+      stbi__uint16 *dest = good + j * x * req_comp;
+
+      #define STBI__COMBO(a,b)  ((a)*8+(b))
+      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
+      // convert source image with img_n components to one with req_comp components;
+      // avoid switch per pixel, so use switch per scanline and massive macros
+      switch (STBI__COMBO(img_n, req_comp)) {
+         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=0xffff;                                     } break;
+         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
+         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=0xffff;                     } break;
+         STBI__CASE(2,1) { dest[0]=src[0];                                                     } break;
+         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
+         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                     } break;
+         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=0xffff;        } break;
+         STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = 0xffff; } break;
+         STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = src[3]; } break;
+         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                       } break;
+         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return (stbi__uint16*) stbi__errpuc("unsupported", "Unsupported format conversion");
+      }
+      #undef STBI__CASE
+   }
+
+   STBI_FREE(data);
+   return good;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp)
+{
+   int i,k,n;
+   float *output;
+   if (!data) return NULL;
+   output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0);
+   if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); }
+   // compute number of non-alpha components
+   if (comp & 1) n = comp; else n = comp-1;
+   for (i=0; i < x*y; ++i) {
+      for (k=0; k < n; ++k) {
+         output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale);
+      }
+   }
+   if (n < comp) {
+      for (i=0; i < x*y; ++i) {
+         output[i*comp + n] = data[i*comp + n]/255.0f;
+      }
+   }
+   STBI_FREE(data);
+   return output;
+}
+#endif
+
+#ifndef STBI_NO_HDR
+#define stbi__float2int(x)   ((int) (x))
+static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp)
+{
+   int i,k,n;
+   stbi_uc *output;
+   if (!data) return NULL;
+   output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0);
+   if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); }
+   // compute number of non-alpha components
+   if (comp & 1) n = comp; else n = comp-1;
+   for (i=0; i < x*y; ++i) {
+      for (k=0; k < n; ++k) {
+         float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f;
+         if (z < 0) z = 0;
+         if (z > 255) z = 255;
+         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
+      }
+      if (k < comp) {
+         float z = data[i*comp+k] * 255 + 0.5f;
+         if (z < 0) z = 0;
+         if (z > 255) z = 255;
+         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
+      }
+   }
+   STBI_FREE(data);
+   return output;
+}
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//  "baseline" JPEG/JFIF decoder
+//
+//    simple implementation
+//      - doesn't support delayed output of y-dimension
+//      - simple interface (only one output format: 8-bit interleaved RGB)
+//      - doesn't try to recover corrupt jpegs
+//      - doesn't allow partial loading, loading multiple at once
+//      - still fast on x86 (copying globals into locals doesn't help x86)
+//      - allocates lots of intermediate memory (full size of all components)
+//        - non-interleaved case requires this anyway
+//        - allows good upsampling (see next)
+//    high-quality
+//      - upsampled channels are bilinearly interpolated, even across blocks
+//      - quality integer IDCT derived from IJG's 'slow'
+//    performance
+//      - fast huffman; reasonable integer IDCT
+//      - some SIMD kernels for common paths on targets with SSE2/NEON
+//      - uses a lot of intermediate memory, could cache poorly
+
+#ifndef STBI_NO_JPEG
+
+// huffman decoding acceleration
+#define FAST_BITS   9  // larger handles more cases; smaller stomps less cache
+
+typedef struct
+{
+   stbi_uc  fast[1 << FAST_BITS];
+   // weirdly, repacking this into AoS is a 10% speed loss, instead of a win
+   stbi__uint16 code[256];
+   stbi_uc  values[256];
+   stbi_uc  size[257];
+   unsigned int maxcode[18];
+   int    delta[17];   // old 'firstsymbol' - old 'firstcode'
+} stbi__huffman;
+
+typedef struct
+{
+   stbi__context *s;
+   stbi__huffman huff_dc[4];
+   stbi__huffman huff_ac[4];
+   stbi__uint16 dequant[4][64];
+   stbi__int16 fast_ac[4][1 << FAST_BITS];
+
+// sizes for components, interleaved MCUs
+   int img_h_max, img_v_max;
+   int img_mcu_x, img_mcu_y;
+   int img_mcu_w, img_mcu_h;
+
+// definition of jpeg image component
+   struct
+   {
+      int id;
+      int h,v;
+      int tq;
+      int hd,ha;
+      int dc_pred;
+
+      int x,y,w2,h2;
+      stbi_uc *data;
+      void *raw_data, *raw_coeff;
+      stbi_uc *linebuf;
+      short   *coeff;   // progressive only
+      int      coeff_w, coeff_h; // number of 8x8 coefficient blocks
+   } img_comp[4];
+
+   stbi__uint32   code_buffer; // jpeg entropy-coded buffer
+   int            code_bits;   // number of valid bits
+   unsigned char  marker;      // marker seen while filling entropy buffer
+   int            nomore;      // flag if we saw a marker so must stop
+
+   int            progressive;
+   int            spec_start;
+   int            spec_end;
+   int            succ_high;
+   int            succ_low;
+   int            eob_run;
+   int            jfif;
+   int            app14_color_transform; // Adobe APP14 tag
+   int            rgb;
+
+   int scan_n, order[4];
+   int restart_interval, todo;
+
+// kernels
+   void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]);
+   void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step);
+   stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs);
+} stbi__jpeg;
+
+static int stbi__build_huffman(stbi__huffman *h, int *count)
+{
+   int i,j,k=0;
+   unsigned int code;
+   // build size list for each symbol (from JPEG spec)
+   for (i=0; i < 16; ++i) {
+      for (j=0; j < count[i]; ++j) {
+         h->size[k++] = (stbi_uc) (i+1);
+         if(k >= 257) return stbi__err("bad size list","Corrupt JPEG");
+      }
+   }
+   h->size[k] = 0;
+
+   // compute actual symbols (from jpeg spec)
+   code = 0;
+   k = 0;
+   for(j=1; j <= 16; ++j) {
+      // compute delta to add to code to compute symbol id
+      h->delta[j] = k - code;
+      if (h->size[k] == j) {
+         while (h->size[k] == j)
+            h->code[k++] = (stbi__uint16) (code++);
+         if (code-1 >= (1u << j)) return stbi__err("bad code lengths","Corrupt JPEG");
+      }
+      // compute largest code + 1 for this size, preshifted as needed later
+      h->maxcode[j] = code << (16-j);
+      code <<= 1;
+   }
+   h->maxcode[j] = 0xffffffff;
+
+   // build non-spec acceleration table; 255 is flag for not-accelerated
+   memset(h->fast, 255, 1 << FAST_BITS);
+   for (i=0; i < k; ++i) {
+      int s = h->size[i];
+      if (s <= FAST_BITS) {
+         int c = h->code[i] << (FAST_BITS-s);
+         int m = 1 << (FAST_BITS-s);
+         for (j=0; j < m; ++j) {
+            h->fast[c+j] = (stbi_uc) i;
+         }
+      }
+   }
+   return 1;
+}
+
+// build a table that decodes both magnitude and value of small ACs in
+// one go.
+static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h)
+{
+   int i;
+   for (i=0; i < (1 << FAST_BITS); ++i) {
+      stbi_uc fast = h->fast[i];
+      fast_ac[i] = 0;
+      if (fast < 255) {
+         int rs = h->values[fast];
+         int run = (rs >> 4) & 15;
+         int magbits = rs & 15;
+         int len = h->size[fast];
+
+         if (magbits && len + magbits <= FAST_BITS) {
+            // magnitude code followed by receive_extend code
+            int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits);
+            int m = 1 << (magbits - 1);
+            if (k < m) k += (~0U << magbits) + 1;
+            // if the result is small enough, we can fit it in fast_ac table
+            if (k >= -128 && k <= 127)
+               fast_ac[i] = (stbi__int16) ((k * 256) + (run * 16) + (len + magbits));
+         }
+      }
+   }
+}
+
+static void stbi__grow_buffer_unsafe(stbi__jpeg *j)
+{
+   do {
+      unsigned int b = j->nomore ? 0 : stbi__get8(j->s);
+      if (b == 0xff) {
+         int c = stbi__get8(j->s);
+         while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes
+         if (c != 0) {
+            j->marker = (unsigned char) c;
+            j->nomore = 1;
+            return;
+         }
+      }
+      j->code_buffer |= b << (24 - j->code_bits);
+      j->code_bits += 8;
+   } while (j->code_bits <= 24);
+}
+
+// (1 << n) - 1
+static const stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535};
+
+// decode a jpeg huffman value from the bitstream
+stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h)
+{
+   unsigned int temp;
+   int c,k;
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+
+   // look at the top FAST_BITS and determine what symbol ID it is,
+   // if the code is <= FAST_BITS
+   c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+   k = h->fast[c];
+   if (k < 255) {
+      int s = h->size[k];
+      if (s > j->code_bits)
+         return -1;
+      j->code_buffer <<= s;
+      j->code_bits -= s;
+      return h->values[k];
+   }
+
+   // naive test is to shift the code_buffer down so k bits are
+   // valid, then test against maxcode. To speed this up, we've
+   // preshifted maxcode left so that it has (16-k) 0s at the
+   // end; in other words, regardless of the number of bits, it
+   // wants to be compared against something shifted to have 16;
+   // that way we don't need to shift inside the loop.
+   temp = j->code_buffer >> 16;
+   for (k=FAST_BITS+1 ; ; ++k)
+      if (temp < h->maxcode[k])
+         break;
+   if (k == 17) {
+      // error! code not found
+      j->code_bits -= 16;
+      return -1;
+   }
+
+   if (k > j->code_bits)
+      return -1;
+
+   // convert the huffman code to the symbol id
+   c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k];
+   if(c < 0 || c >= 256) // symbol id out of bounds!
+       return -1;
+   STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]);
+
+   // convert the id to a symbol
+   j->code_bits -= k;
+   j->code_buffer <<= k;
+   return h->values[c];
+}
+
+// bias[n] = (-1<<n) + 1
+static const int stbi__jbias[16] = {0,-1,-3,-7,-15,-31,-63,-127,-255,-511,-1023,-2047,-4095,-8191,-16383,-32767};
+
+// combined JPEG 'receive' and JPEG 'extend', since baseline
+// always extends everything it receives.
+stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n)
+{
+   unsigned int k;
+   int sgn;
+   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
+
+   sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative)
+   k = stbi_lrot(j->code_buffer, n);
+   j->code_buffer = k & ~stbi__bmask[n];
+   k &= stbi__bmask[n];
+   j->code_bits -= n;
+   return k + (stbi__jbias[n] & (sgn - 1));
+}
+
+// get some unsigned bits
+stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n)
+{
+   unsigned int k;
+   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
+   k = stbi_lrot(j->code_buffer, n);
+   j->code_buffer = k & ~stbi__bmask[n];
+   k &= stbi__bmask[n];
+   j->code_bits -= n;
+   return k;
+}
+
+stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j)
+{
+   unsigned int k;
+   if (j->code_bits < 1) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < 1) return 0; // ran out of bits from stream, return 0s intead of continuing
+   k = j->code_buffer;
+   j->code_buffer <<= 1;
+   --j->code_bits;
+   return k & 0x80000000;
+}
+
+// given a value that's at position X in the zigzag stream,
+// where does it appear in the 8x8 matrix coded as row-major?
+static const stbi_uc stbi__jpeg_dezigzag[64+15] =
+{
+    0,  1,  8, 16,  9,  2,  3, 10,
+   17, 24, 32, 25, 18, 11,  4,  5,
+   12, 19, 26, 33, 40, 48, 41, 34,
+   27, 20, 13,  6,  7, 14, 21, 28,
+   35, 42, 49, 56, 57, 50, 43, 36,
+   29, 22, 15, 23, 30, 37, 44, 51,
+   58, 59, 52, 45, 38, 31, 39, 46,
+   53, 60, 61, 54, 47, 55, 62, 63,
+   // let corrupt input sample past end
+   63, 63, 63, 63, 63, 63, 63, 63,
+   63, 63, 63, 63, 63, 63, 63
+};
+
+// decode one 64-entry block--
+static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant)
+{
+   int diff,dc,k;
+   int t;
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+   t = stbi__jpeg_huff_decode(j, hdc);
+   if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG");
+
+   // 0 all the ac values now so we can do it 32-bits at a time
+   memset(data,0,64*sizeof(data[0]));
+
+   diff = t ? stbi__extend_receive(j, t) : 0;
+   if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta","Corrupt JPEG");
+   dc = j->img_comp[b].dc_pred + diff;
+   j->img_comp[b].dc_pred = dc;
+   if (!stbi__mul2shorts_valid(dc, dequant[0])) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+   data[0] = (short) (dc * dequant[0]);
+
+   // decode AC components, see JPEG spec
+   k = 1;
+   do {
+      unsigned int zig;
+      int c,r,s;
+      if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+      c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+      r = fac[c];
+      if (r) { // fast-AC path
+         k += (r >> 4) & 15; // run
+         s = r & 15; // combined length
+         if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
+         j->code_buffer <<= s;
+         j->code_bits -= s;
+         // decode into unzigzag'd location
+         zig = stbi__jpeg_dezigzag[k++];
+         data[zig] = (short) ((r >> 8) * dequant[zig]);
+      } else {
+         int rs = stbi__jpeg_huff_decode(j, hac);
+         if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+         s = rs & 15;
+         r = rs >> 4;
+         if (s == 0) {
+            if (rs != 0xf0) break; // end block
+            k += 16;
+         } else {
+            k += r;
+            // decode into unzigzag'd location
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]);
+         }
+      }
+   } while (k < 64);
+   return 1;
+}
+
+static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b)
+{
+   int diff,dc;
+   int t;
+   if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+
+   if (j->succ_high == 0) {
+      // first scan for DC coefficient, must be first
+      memset(data,0,64*sizeof(data[0])); // 0 all the ac values now
+      t = stbi__jpeg_huff_decode(j, hdc);
+      if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      diff = t ? stbi__extend_receive(j, t) : 0;
+
+      if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta", "Corrupt JPEG");
+      dc = j->img_comp[b].dc_pred + diff;
+      j->img_comp[b].dc_pred = dc;
+      if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      data[0] = (short) (dc * (1 << j->succ_low));
+   } else {
+      // refinement scan for DC coefficient
+      if (stbi__jpeg_get_bit(j))
+         data[0] += (short) (1 << j->succ_low);
+   }
+   return 1;
+}
+
+// @OPTIMIZE: store non-zigzagged during the decode passes,
+// and only de-zigzag when dequantizing
+static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac)
+{
+   int k;
+   if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+   if (j->succ_high == 0) {
+      int shift = j->succ_low;
+
+      if (j->eob_run) {
+         --j->eob_run;
+         return 1;
+      }
+
+      k = j->spec_start;
+      do {
+         unsigned int zig;
+         int c,r,s;
+         if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+         c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+         r = fac[c];
+         if (r) { // fast-AC path
+            k += (r >> 4) & 15; // run
+            s = r & 15; // combined length
+            if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
+            j->code_buffer <<= s;
+            j->code_bits -= s;
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short) ((r >> 8) * (1 << shift));
+         } else {
+            int rs = stbi__jpeg_huff_decode(j, hac);
+            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+               if (r < 15) {
+                  j->eob_run = (1 << r);
+                  if (r)
+                     j->eob_run += stbi__jpeg_get_bits(j, r);
+                  --j->eob_run;
+                  break;
+               }
+               k += 16;
+            } else {
+               k += r;
+               zig = stbi__jpeg_dezigzag[k++];
+               data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift));
+            }
+         }
+      } while (k <= j->spec_end);
+   } else {
+      // refinement scan for these AC coefficients
+
+      short bit = (short) (1 << j->succ_low);
+
+      if (j->eob_run) {
+         --j->eob_run;
+         for (k = j->spec_start; k <= j->spec_end; ++k) {
+            short *p = &data[stbi__jpeg_dezigzag[k]];
+            if (*p != 0)
+               if (stbi__jpeg_get_bit(j))
+                  if ((*p & bit)==0) {
+                     if (*p > 0)
+                        *p += bit;
+                     else
+                        *p -= bit;
+                  }
+         }
+      } else {
+         k = j->spec_start;
+         do {
+            int r,s;
+            int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh
+            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+               if (r < 15) {
+                  j->eob_run = (1 << r) - 1;
+                  if (r)
+                     j->eob_run += stbi__jpeg_get_bits(j, r);
+                  r = 64; // force end of block
+               } else {
+                  // r=15 s=0 should write 16 0s, so we just do
+                  // a run of 15 0s and then write s (which is 0),
+                  // so we don't have to do anything special here
+               }
+            } else {
+               if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG");
+               // sign bit
+               if (stbi__jpeg_get_bit(j))
+                  s = bit;
+               else
+                  s = -bit;
+            }
+
+            // advance by r
+            while (k <= j->spec_end) {
+               short *p = &data[stbi__jpeg_dezigzag[k++]];
+               if (*p != 0) {
+                  if (stbi__jpeg_get_bit(j))
+                     if ((*p & bit)==0) {
+                        if (*p > 0)
+                           *p += bit;
+                        else
+                           *p -= bit;
+                     }
+               } else {
+                  if (r == 0) {
+                     *p = (short) s;
+                     break;
+                  }
+                  --r;
+               }
+            }
+         } while (k <= j->spec_end);
+      }
+   }
+   return 1;
+}
+
+// take a -128..127 value and stbi__clamp it and convert to 0..255
+stbi_inline static stbi_uc stbi__clamp(int x)
+{
+   // trick to use a single test to catch both cases
+   if ((unsigned int) x > 255) {
+      if (x < 0) return 0;
+      if (x > 255) return 255;
+   }
+   return (stbi_uc) x;
+}
+
+#define stbi__f2f(x)  ((int) (((x) * 4096 + 0.5)))
+#define stbi__fsh(x)  ((x) * 4096)
+
+// derived from jidctint -- DCT_ISLOW
+#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \
+   int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \
+   p2 = s2;                                    \
+   p3 = s6;                                    \
+   p1 = (p2+p3) * stbi__f2f(0.5411961f);       \
+   t2 = p1 + p3*stbi__f2f(-1.847759065f);      \
+   t3 = p1 + p2*stbi__f2f( 0.765366865f);      \
+   p2 = s0;                                    \
+   p3 = s4;                                    \
+   t0 = stbi__fsh(p2+p3);                      \
+   t1 = stbi__fsh(p2-p3);                      \
+   x0 = t0+t3;                                 \
+   x3 = t0-t3;                                 \
+   x1 = t1+t2;                                 \
+   x2 = t1-t2;                                 \
+   t0 = s7;                                    \
+   t1 = s5;                                    \
+   t2 = s3;                                    \
+   t3 = s1;                                    \
+   p3 = t0+t2;                                 \
+   p4 = t1+t3;                                 \
+   p1 = t0+t3;                                 \
+   p2 = t1+t2;                                 \
+   p5 = (p3+p4)*stbi__f2f( 1.175875602f);      \
+   t0 = t0*stbi__f2f( 0.298631336f);           \
+   t1 = t1*stbi__f2f( 2.053119869f);           \
+   t2 = t2*stbi__f2f( 3.072711026f);           \
+   t3 = t3*stbi__f2f( 1.501321110f);           \
+   p1 = p5 + p1*stbi__f2f(-0.899976223f);      \
+   p2 = p5 + p2*stbi__f2f(-2.562915447f);      \
+   p3 = p3*stbi__f2f(-1.961570560f);           \
+   p4 = p4*stbi__f2f(-0.390180644f);           \
+   t3 += p1+p4;                                \
+   t2 += p2+p3;                                \
+   t1 += p2+p4;                                \
+   t0 += p1+p3;
+
+static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64])
+{
+   int i,val[64],*v=val;
+   stbi_uc *o;
+   short *d = data;
+
+   // columns
+   for (i=0; i < 8; ++i,++d, ++v) {
+      // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing
+      if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0
+           && d[40]==0 && d[48]==0 && d[56]==0) {
+         //    no shortcut                 0     seconds
+         //    (1|2|3|4|5|6|7)==0          0     seconds
+         //    all separate               -0.047 seconds
+         //    1 && 2|3 && 4|5 && 6|7:    -0.047 seconds
+         int dcterm = d[0]*4;
+         v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm;
+      } else {
+         STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56])
+         // constants scaled things up by 1<<12; let's bring them back
+         // down, but keep 2 extra bits of precision
+         x0 += 512; x1 += 512; x2 += 512; x3 += 512;
+         v[ 0] = (x0+t3) >> 10;
+         v[56] = (x0-t3) >> 10;
+         v[ 8] = (x1+t2) >> 10;
+         v[48] = (x1-t2) >> 10;
+         v[16] = (x2+t1) >> 10;
+         v[40] = (x2-t1) >> 10;
+         v[24] = (x3+t0) >> 10;
+         v[32] = (x3-t0) >> 10;
+      }
+   }
+
+   for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) {
+      // no fast case since the first 1D IDCT spread components out
+      STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7])
+      // constants scaled things up by 1<<12, plus we had 1<<2 from first
+      // loop, plus horizontal and vertical each scale by sqrt(8) so together
+      // we've got an extra 1<<3, so 1<<17 total we need to remove.
+      // so we want to round that, which means adding 0.5 * 1<<17,
+      // aka 65536. Also, we'll end up with -128 to 127 that we want
+      // to encode as 0..255 by adding 128, so we'll add that before the shift
+      x0 += 65536 + (128<<17);
+      x1 += 65536 + (128<<17);
+      x2 += 65536 + (128<<17);
+      x3 += 65536 + (128<<17);
+      // tried computing the shifts into temps, or'ing the temps to see
+      // if any were out of range, but that was slower
+      o[0] = stbi__clamp((x0+t3) >> 17);
+      o[7] = stbi__clamp((x0-t3) >> 17);
+      o[1] = stbi__clamp((x1+t2) >> 17);
+      o[6] = stbi__clamp((x1-t2) >> 17);
+      o[2] = stbi__clamp((x2+t1) >> 17);
+      o[5] = stbi__clamp((x2-t1) >> 17);
+      o[3] = stbi__clamp((x3+t0) >> 17);
+      o[4] = stbi__clamp((x3-t0) >> 17);
+   }
+}
+
+#ifdef STBI_SSE2
+// sse2 integer IDCT. not the fastest possible implementation but it
+// produces bit-identical results to the generic C version so it's
+// fully "transparent".
+static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
+{
+   // This is constructed to match our regular (generic) integer IDCT exactly.
+   __m128i row0, row1, row2, row3, row4, row5, row6, row7;
+   __m128i tmp;
+
+   // dot product constant: even elems=x, odd elems=y
+   #define dct_const(x,y)  _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y))
+
+   // out(0) = c0[even]*x + c0[odd]*y   (c0, x, y 16-bit, out 32-bit)
+   // out(1) = c1[even]*x + c1[odd]*y
+   #define dct_rot(out0,out1, x,y,c0,c1) \
+      __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \
+      __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \
+      __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \
+      __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \
+      __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \
+      __m128i out1##_h = _mm_madd_epi16(c0##hi, c1)
+
+   // out = in << 12  (in 16-bit, out 32-bit)
+   #define dct_widen(out, in) \
+      __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \
+      __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4)
+
+   // wide add
+   #define dct_wadd(out, a, b) \
+      __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \
+      __m128i out##_h = _mm_add_epi32(a##_h, b##_h)
+
+   // wide sub
+   #define dct_wsub(out, a, b) \
+      __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \
+      __m128i out##_h = _mm_sub_epi32(a##_h, b##_h)
+
+   // butterfly a/b, add bias, then shift by "s" and pack
+   #define dct_bfly32o(out0, out1, a,b,bias,s) \
+      { \
+         __m128i abiased_l = _mm_add_epi32(a##_l, bias); \
+         __m128i abiased_h = _mm_add_epi32(a##_h, bias); \
+         dct_wadd(sum, abiased, b); \
+         dct_wsub(dif, abiased, b); \
+         out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \
+         out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \
+      }
+
+   // 8-bit interleave step (for transposes)
+   #define dct_interleave8(a, b) \
+      tmp = a; \
+      a = _mm_unpacklo_epi8(a, b); \
+      b = _mm_unpackhi_epi8(tmp, b)
+
+   // 16-bit interleave step (for transposes)
+   #define dct_interleave16(a, b) \
+      tmp = a; \
+      a = _mm_unpacklo_epi16(a, b); \
+      b = _mm_unpackhi_epi16(tmp, b)
+
+   #define dct_pass(bias,shift) \
+      { \
+         /* even part */ \
+         dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \
+         __m128i sum04 = _mm_add_epi16(row0, row4); \
+         __m128i dif04 = _mm_sub_epi16(row0, row4); \
+         dct_widen(t0e, sum04); \
+         dct_widen(t1e, dif04); \
+         dct_wadd(x0, t0e, t3e); \
+         dct_wsub(x3, t0e, t3e); \
+         dct_wadd(x1, t1e, t2e); \
+         dct_wsub(x2, t1e, t2e); \
+         /* odd part */ \
+         dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \
+         dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \
+         __m128i sum17 = _mm_add_epi16(row1, row7); \
+         __m128i sum35 = _mm_add_epi16(row3, row5); \
+         dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \
+         dct_wadd(x4, y0o, y4o); \
+         dct_wadd(x5, y1o, y5o); \
+         dct_wadd(x6, y2o, y5o); \
+         dct_wadd(x7, y3o, y4o); \
+         dct_bfly32o(row0,row7, x0,x7,bias,shift); \
+         dct_bfly32o(row1,row6, x1,x6,bias,shift); \
+         dct_bfly32o(row2,row5, x2,x5,bias,shift); \
+         dct_bfly32o(row3,row4, x3,x4,bias,shift); \
+      }
+
+   __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f));
+   __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f));
+   __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f));
+   __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f));
+   __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f));
+   __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f));
+   __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f));
+   __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f));
+
+   // rounding biases in column/row passes, see stbi__idct_block for explanation.
+   __m128i bias_0 = _mm_set1_epi32(512);
+   __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17));
+
+   // load
+   row0 = _mm_load_si128((const __m128i *) (data + 0*8));
+   row1 = _mm_load_si128((const __m128i *) (data + 1*8));
+   row2 = _mm_load_si128((const __m128i *) (data + 2*8));
+   row3 = _mm_load_si128((const __m128i *) (data + 3*8));
+   row4 = _mm_load_si128((const __m128i *) (data + 4*8));
+   row5 = _mm_load_si128((const __m128i *) (data + 5*8));
+   row6 = _mm_load_si128((const __m128i *) (data + 6*8));
+   row7 = _mm_load_si128((const __m128i *) (data + 7*8));
+
+   // column pass
+   dct_pass(bias_0, 10);
+
+   {
+      // 16bit 8x8 transpose pass 1
+      dct_interleave16(row0, row4);
+      dct_interleave16(row1, row5);
+      dct_interleave16(row2, row6);
+      dct_interleave16(row3, row7);
+
+      // transpose pass 2
+      dct_interleave16(row0, row2);
+      dct_interleave16(row1, row3);
+      dct_interleave16(row4, row6);
+      dct_interleave16(row5, row7);
+
+      // transpose pass 3
+      dct_interleave16(row0, row1);
+      dct_interleave16(row2, row3);
+      dct_interleave16(row4, row5);
+      dct_interleave16(row6, row7);
+   }
+
+   // row pass
+   dct_pass(bias_1, 17);
+
+   {
+      // pack
+      __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7
+      __m128i p1 = _mm_packus_epi16(row2, row3);
+      __m128i p2 = _mm_packus_epi16(row4, row5);
+      __m128i p3 = _mm_packus_epi16(row6, row7);
+
+      // 8bit 8x8 transpose pass 1
+      dct_interleave8(p0, p2); // a0e0a1e1...
+      dct_interleave8(p1, p3); // c0g0c1g1...
+
+      // transpose pass 2
+      dct_interleave8(p0, p1); // a0c0e0g0...
+      dct_interleave8(p2, p3); // b0d0f0h0...
+
+      // transpose pass 3
+      dct_interleave8(p0, p2); // a0b0c0d0...
+      dct_interleave8(p1, p3); // a4b4c4d4...
+
+      // store
+      _mm_storel_epi64((__m128i *) out, p0); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p2); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p1); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p3); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e));
+   }
+
+#undef dct_const
+#undef dct_rot
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_interleave8
+#undef dct_interleave16
+#undef dct_pass
+}
+
+#endif // STBI_SSE2
+
+#ifdef STBI_NEON
+
+// NEON integer IDCT. should produce bit-identical
+// results to the generic C version.
+static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
+{
+   int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;
+
+   int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f));
+   int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f));
+   int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f));
+   int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f));
+   int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f));
+   int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f));
+   int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f));
+   int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f));
+   int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f));
+   int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f));
+   int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f));
+   int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f));
+
+#define dct_long_mul(out, inq, coeff) \
+   int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \
+   int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff)
+
+#define dct_long_mac(out, acc, inq, coeff) \
+   int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \
+   int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff)
+
+#define dct_widen(out, inq) \
+   int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \
+   int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12)
+
+// wide add
+#define dct_wadd(out, a, b) \
+   int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \
+   int32x4_t out##_h = vaddq_s32(a##_h, b##_h)
+
+// wide sub
+#define dct_wsub(out, a, b) \
+   int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \
+   int32x4_t out##_h = vsubq_s32(a##_h, b##_h)
+
+// butterfly a/b, then shift using "shiftop" by "s" and pack
+#define dct_bfly32o(out0,out1, a,b,shiftop,s) \
+   { \
+      dct_wadd(sum, a, b); \
+      dct_wsub(dif, a, b); \
+      out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \
+      out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \
+   }
+
+#define dct_pass(shiftop, shift) \
+   { \
+      /* even part */ \
+      int16x8_t sum26 = vaddq_s16(row2, row6); \
+      dct_long_mul(p1e, sum26, rot0_0); \
+      dct_long_mac(t2e, p1e, row6, rot0_1); \
+      dct_long_mac(t3e, p1e, row2, rot0_2); \
+      int16x8_t sum04 = vaddq_s16(row0, row4); \
+      int16x8_t dif04 = vsubq_s16(row0, row4); \
+      dct_widen(t0e, sum04); \
+      dct_widen(t1e, dif04); \
+      dct_wadd(x0, t0e, t3e); \
+      dct_wsub(x3, t0e, t3e); \
+      dct_wadd(x1, t1e, t2e); \
+      dct_wsub(x2, t1e, t2e); \
+      /* odd part */ \
+      int16x8_t sum15 = vaddq_s16(row1, row5); \
+      int16x8_t sum17 = vaddq_s16(row1, row7); \
+      int16x8_t sum35 = vaddq_s16(row3, row5); \
+      int16x8_t sum37 = vaddq_s16(row3, row7); \
+      int16x8_t sumodd = vaddq_s16(sum17, sum35); \
+      dct_long_mul(p5o, sumodd, rot1_0); \
+      dct_long_mac(p1o, p5o, sum17, rot1_1); \
+      dct_long_mac(p2o, p5o, sum35, rot1_2); \
+      dct_long_mul(p3o, sum37, rot2_0); \
+      dct_long_mul(p4o, sum15, rot2_1); \
+      dct_wadd(sump13o, p1o, p3o); \
+      dct_wadd(sump24o, p2o, p4o); \
+      dct_wadd(sump23o, p2o, p3o); \
+      dct_wadd(sump14o, p1o, p4o); \
+      dct_long_mac(x4, sump13o, row7, rot3_0); \
+      dct_long_mac(x5, sump24o, row5, rot3_1); \
+      dct_long_mac(x6, sump23o, row3, rot3_2); \
+      dct_long_mac(x7, sump14o, row1, rot3_3); \
+      dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \
+      dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \
+      dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \
+      dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \
+   }
+
+   // load
+   row0 = vld1q_s16(data + 0*8);
+   row1 = vld1q_s16(data + 1*8);
+   row2 = vld1q_s16(data + 2*8);
+   row3 = vld1q_s16(data + 3*8);
+   row4 = vld1q_s16(data + 4*8);
+   row5 = vld1q_s16(data + 5*8);
+   row6 = vld1q_s16(data + 6*8);
+   row7 = vld1q_s16(data + 7*8);
+
+   // add DC bias
+   row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0));
+
+   // column pass
+   dct_pass(vrshrn_n_s32, 10);
+
+   // 16bit 8x8 transpose
+   {
+// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively.
+// whether compilers actually get this is another story, sadly.
+#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; }
+#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); }
+#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); }
+
+      // pass 1
+      dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6
+      dct_trn16(row2, row3);
+      dct_trn16(row4, row5);
+      dct_trn16(row6, row7);
+
+      // pass 2
+      dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4
+      dct_trn32(row1, row3);
+      dct_trn32(row4, row6);
+      dct_trn32(row5, row7);
+
+      // pass 3
+      dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0
+      dct_trn64(row1, row5);
+      dct_trn64(row2, row6);
+      dct_trn64(row3, row7);
+
+#undef dct_trn16
+#undef dct_trn32
+#undef dct_trn64
+   }
+
+   // row pass
+   // vrshrn_n_s32 only supports shifts up to 16, we need
+   // 17. so do a non-rounding shift of 16 first then follow
+   // up with a rounding shift by 1.
+   dct_pass(vshrn_n_s32, 16);
+
+   {
+      // pack and round
+      uint8x8_t p0 = vqrshrun_n_s16(row0, 1);
+      uint8x8_t p1 = vqrshrun_n_s16(row1, 1);
+      uint8x8_t p2 = vqrshrun_n_s16(row2, 1);
+      uint8x8_t p3 = vqrshrun_n_s16(row3, 1);
+      uint8x8_t p4 = vqrshrun_n_s16(row4, 1);
+      uint8x8_t p5 = vqrshrun_n_s16(row5, 1);
+      uint8x8_t p6 = vqrshrun_n_s16(row6, 1);
+      uint8x8_t p7 = vqrshrun_n_s16(row7, 1);
+
+      // again, these can translate into one instruction, but often don't.
+#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; }
+#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); }
+#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); }
+
+      // sadly can't use interleaved stores here since we only write
+      // 8 bytes to each scan line!
+
+      // 8x8 8-bit transpose pass 1
+      dct_trn8_8(p0, p1);
+      dct_trn8_8(p2, p3);
+      dct_trn8_8(p4, p5);
+      dct_trn8_8(p6, p7);
+
+      // pass 2
+      dct_trn8_16(p0, p2);
+      dct_trn8_16(p1, p3);
+      dct_trn8_16(p4, p6);
+      dct_trn8_16(p5, p7);
+
+      // pass 3
+      dct_trn8_32(p0, p4);
+      dct_trn8_32(p1, p5);
+      dct_trn8_32(p2, p6);
+      dct_trn8_32(p3, p7);
+
+      // store
+      vst1_u8(out, p0); out += out_stride;
+      vst1_u8(out, p1); out += out_stride;
+      vst1_u8(out, p2); out += out_stride;
+      vst1_u8(out, p3); out += out_stride;
+      vst1_u8(out, p4); out += out_stride;
+      vst1_u8(out, p5); out += out_stride;
+      vst1_u8(out, p6); out += out_stride;
+      vst1_u8(out, p7);
+
+#undef dct_trn8_8
+#undef dct_trn8_16
+#undef dct_trn8_32
+   }
+
+#undef dct_long_mul
+#undef dct_long_mac
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_pass
+}
+
+#endif // STBI_NEON
+
+#define STBI__MARKER_none  0xff
+// if there's a pending marker from the entropy stream, return that
+// otherwise, fetch from the stream and get a marker. if there's no
+// marker, return 0xff, which is never a valid marker value
+static stbi_uc stbi__get_marker(stbi__jpeg *j)
+{
+   stbi_uc x;
+   if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; }
+   x = stbi__get8(j->s);
+   if (x != 0xff) return STBI__MARKER_none;
+   while (x == 0xff)
+      x = stbi__get8(j->s); // consume repeated 0xff fill bytes
+   return x;
+}
+
+// in each scan, we'll have scan_n components, and the order
+// of the components is specified by order[]
+#define STBI__RESTART(x)     ((x) >= 0xd0 && (x) <= 0xd7)
+
+// after a restart interval, stbi__jpeg_reset the entropy decoder and
+// the dc prediction
+static void stbi__jpeg_reset(stbi__jpeg *j)
+{
+   j->code_bits = 0;
+   j->code_buffer = 0;
+   j->nomore = 0;
+   j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0;
+   j->marker = STBI__MARKER_none;
+   j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff;
+   j->eob_run = 0;
+   // no more than 1<<31 MCUs if no restart_interal? that's plenty safe,
+   // since we don't even allow 1<<30 pixels
+}
+
+static int stbi__parse_entropy_coded_data(stbi__jpeg *z)
+{
+   stbi__jpeg_reset(z);
+   if (!z->progressive) {
+      if (z->scan_n == 1) {
+         int i,j;
+         STBI_SIMD_ALIGN(short, data[64]);
+         int n = z->order[0];
+         // non-interleaved data, we just need to process one block at a time,
+         // in trivial scanline order
+         // number of blocks to do just depends on how many actual "pixels" this
+         // component has, independent of interleaved MCU blocking and such
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               int ha = z->img_comp[n].ha;
+               if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
+               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
+               // every data block is an MCU, so countdown the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  // if it's NOT a restart, then just bail, so we get corrupt data
+                  // rather than no data
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      } else { // interleaved
+         int i,j,k,x,y;
+         STBI_SIMD_ALIGN(short, data[64]);
+         for (j=0; j < z->img_mcu_y; ++j) {
+            for (i=0; i < z->img_mcu_x; ++i) {
+               // scan an interleaved mcu... process scan_n components in order
+               for (k=0; k < z->scan_n; ++k) {
+                  int n = z->order[k];
+                  // scan out an mcu's worth of this component; that's just determined
+                  // by the basic H and V specified for the component
+                  for (y=0; y < z->img_comp[n].v; ++y) {
+                     for (x=0; x < z->img_comp[n].h; ++x) {
+                        int x2 = (i*z->img_comp[n].h + x)*8;
+                        int y2 = (j*z->img_comp[n].v + y)*8;
+                        int ha = z->img_comp[n].ha;
+                        if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
+                        z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data);
+                     }
+                  }
+               }
+               // after all interleaved components, that's an interleaved MCU,
+               // so now count down the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      }
+   } else {
+      if (z->scan_n == 1) {
+         int i,j;
+         int n = z->order[0];
+         // non-interleaved data, we just need to process one block at a time,
+         // in trivial scanline order
+         // number of blocks to do just depends on how many actual "pixels" this
+         // component has, independent of interleaved MCU blocking and such
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+               if (z->spec_start == 0) {
+                  if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                     return 0;
+               } else {
+                  int ha = z->img_comp[n].ha;
+                  if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha]))
+                     return 0;
+               }
+               // every data block is an MCU, so countdown the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      } else { // interleaved
+         int i,j,k,x,y;
+         for (j=0; j < z->img_mcu_y; ++j) {
+            for (i=0; i < z->img_mcu_x; ++i) {
+               // scan an interleaved mcu... process scan_n components in order
+               for (k=0; k < z->scan_n; ++k) {
+                  int n = z->order[k];
+                  // scan out an mcu's worth of this component; that's just determined
+                  // by the basic H and V specified for the component
+                  for (y=0; y < z->img_comp[n].v; ++y) {
+                     for (x=0; x < z->img_comp[n].h; ++x) {
+                        int x2 = (i*z->img_comp[n].h + x);
+                        int y2 = (j*z->img_comp[n].v + y);
+                        short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w);
+                        if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                           return 0;
+                     }
+                  }
+               }
+               // after all interleaved components, that's an interleaved MCU,
+               // so now count down the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      }
+   }
+}
+
+static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant)
+{
+   int i;
+   for (i=0; i < 64; ++i)
+      data[i] *= dequant[i];
+}
+
+static void stbi__jpeg_finish(stbi__jpeg *z)
+{
+   if (z->progressive) {
+      // dequantize and idct the data
+      int i,j,n;
+      for (n=0; n < z->s->img_n; ++n) {
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+               stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]);
+               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
+            }
+         }
+      }
+   }
+}
+
+static int stbi__process_marker(stbi__jpeg *z, int m)
+{
+   int L;
+   switch (m) {
+      case STBI__MARKER_none: // no marker found
+         return stbi__err("expected marker","Corrupt JPEG");
+
+      case 0xDD: // DRI - specify restart interval
+         if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG");
+         z->restart_interval = stbi__get16be(z->s);
+         return 1;
+
+      case 0xDB: // DQT - define quantization table
+         L = stbi__get16be(z->s)-2;
+         while (L > 0) {
+            int q = stbi__get8(z->s);
+            int p = q >> 4, sixteen = (p != 0);
+            int t = q & 15,i;
+            if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG");
+            if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG");
+
+            for (i=0; i < 64; ++i)
+               z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s));
+            L -= (sixteen ? 129 : 65);
+         }
+         return L==0;
+
+      case 0xC4: // DHT - define huffman table
+         L = stbi__get16be(z->s)-2;
+         while (L > 0) {
+            stbi_uc *v;
+            int sizes[16],i,n=0;
+            int q = stbi__get8(z->s);
+            int tc = q >> 4;
+            int th = q & 15;
+            if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG");
+            for (i=0; i < 16; ++i) {
+               sizes[i] = stbi__get8(z->s);
+               n += sizes[i];
+            }
+            if(n > 256) return stbi__err("bad DHT header","Corrupt JPEG"); // Loop over i < n would write past end of values!
+            L -= 17;
+            if (tc == 0) {
+               if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0;
+               v = z->huff_dc[th].values;
+            } else {
+               if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0;
+               v = z->huff_ac[th].values;
+            }
+            for (i=0; i < n; ++i)
+               v[i] = stbi__get8(z->s);
+            if (tc != 0)
+               stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th);
+            L -= n;
+         }
+         return L==0;
+   }
+
+   // check for comment block or APP blocks
+   if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) {
+      L = stbi__get16be(z->s);
+      if (L < 2) {
+         if (m == 0xFE)
+            return stbi__err("bad COM len","Corrupt JPEG");
+         else
+            return stbi__err("bad APP len","Corrupt JPEG");
+      }
+      L -= 2;
+
+      if (m == 0xE0 && L >= 5) { // JFIF APP0 segment
+         static const unsigned char tag[5] = {'J','F','I','F','\0'};
+         int ok = 1;
+         int i;
+         for (i=0; i < 5; ++i)
+            if (stbi__get8(z->s) != tag[i])
+               ok = 0;
+         L -= 5;
+         if (ok)
+            z->jfif = 1;
+      } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment
+         static const unsigned char tag[6] = {'A','d','o','b','e','\0'};
+         int ok = 1;
+         int i;
+         for (i=0; i < 6; ++i)
+            if (stbi__get8(z->s) != tag[i])
+               ok = 0;
+         L -= 6;
+         if (ok) {
+            stbi__get8(z->s); // version
+            stbi__get16be(z->s); // flags0
+            stbi__get16be(z->s); // flags1
+            z->app14_color_transform = stbi__get8(z->s); // color transform
+            L -= 6;
+         }
+      }
+
+      stbi__skip(z->s, L);
+      return 1;
+   }
+
+   return stbi__err("unknown marker","Corrupt JPEG");
+}
+
+// after we see SOS
+static int stbi__process_scan_header(stbi__jpeg *z)
+{
+   int i;
+   int Ls = stbi__get16be(z->s);
+   z->scan_n = stbi__get8(z->s);
+   if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG");
+   if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG");
+   for (i=0; i < z->scan_n; ++i) {
+      int id = stbi__get8(z->s), which;
+      int q = stbi__get8(z->s);
+      for (which = 0; which < z->s->img_n; ++which)
+         if (z->img_comp[which].id == id)
+            break;
+      if (which == z->s->img_n) return 0; // no match
+      z->img_comp[which].hd = q >> 4;   if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG");
+      z->img_comp[which].ha = q & 15;   if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG");
+      z->order[i] = which;
+   }
+
+   {
+      int aa;
+      z->spec_start = stbi__get8(z->s);
+      z->spec_end   = stbi__get8(z->s); // should be 63, but might be 0
+      aa = stbi__get8(z->s);
+      z->succ_high = (aa >> 4);
+      z->succ_low  = (aa & 15);
+      if (z->progressive) {
+         if (z->spec_start > 63 || z->spec_end > 63  || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13)
+            return stbi__err("bad SOS", "Corrupt JPEG");
+      } else {
+         if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG");
+         if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG");
+         z->spec_end = 63;
+      }
+   }
+
+   return 1;
+}
+
+static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why)
+{
+   int i;
+   for (i=0; i < ncomp; ++i) {
+      if (z->img_comp[i].raw_data) {
+         STBI_FREE(z->img_comp[i].raw_data);
+         z->img_comp[i].raw_data = NULL;
+         z->img_comp[i].data = NULL;
+      }
+      if (z->img_comp[i].raw_coeff) {
+         STBI_FREE(z->img_comp[i].raw_coeff);
+         z->img_comp[i].raw_coeff = 0;
+         z->img_comp[i].coeff = 0;
+      }
+      if (z->img_comp[i].linebuf) {
+         STBI_FREE(z->img_comp[i].linebuf);
+         z->img_comp[i].linebuf = NULL;
+      }
+   }
+   return why;
+}
+
+static int stbi__process_frame_header(stbi__jpeg *z, int scan)
+{
+   stbi__context *s = z->s;
+   int Lf,p,i,q, h_max=1,v_max=1,c;
+   Lf = stbi__get16be(s);         if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG
+   p  = stbi__get8(s);            if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline
+   s->img_y = stbi__get16be(s);   if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG
+   s->img_x = stbi__get16be(s);   if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   c = stbi__get8(s);
+   if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG");
+   s->img_n = c;
+   for (i=0; i < c; ++i) {
+      z->img_comp[i].data = NULL;
+      z->img_comp[i].linebuf = NULL;
+   }
+
+   if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG");
+
+   z->rgb = 0;
+   for (i=0; i < s->img_n; ++i) {
+      static const unsigned char rgb[3] = { 'R', 'G', 'B' };
+      z->img_comp[i].id = stbi__get8(s);
+      if (s->img_n == 3 && z->img_comp[i].id == rgb[i])
+         ++z->rgb;
+      q = stbi__get8(s);
+      z->img_comp[i].h = (q >> 4);  if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG");
+      z->img_comp[i].v = q & 15;    if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG");
+      z->img_comp[i].tq = stbi__get8(s);  if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG");
+   }
+
+   if (scan != STBI__SCAN_load) return 1;
+
+   if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode");
+
+   for (i=0; i < s->img_n; ++i) {
+      if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h;
+      if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v;
+   }
+
+   // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios
+   // and I've never seen a non-corrupted JPEG file actually use them
+   for (i=0; i < s->img_n; ++i) {
+      if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG");
+      if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG");
+   }
+
+   // compute interleaved mcu info
+   z->img_h_max = h_max;
+   z->img_v_max = v_max;
+   z->img_mcu_w = h_max * 8;
+   z->img_mcu_h = v_max * 8;
+   // these sizes can't be more than 17 bits
+   z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w;
+   z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h;
+
+   for (i=0; i < s->img_n; ++i) {
+      // number of effective pixels (e.g. for non-interleaved MCU)
+      z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max;
+      z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max;
+      // to simplify generation, we'll allocate enough memory to decode
+      // the bogus oversized data from using interleaved MCUs and their
+      // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't
+      // discard the extra data until colorspace conversion
+      //
+      // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier)
+      // so these muls can't overflow with 32-bit ints (which we require)
+      z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8;
+      z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8;
+      z->img_comp[i].coeff = 0;
+      z->img_comp[i].raw_coeff = 0;
+      z->img_comp[i].linebuf = NULL;
+      z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15);
+      if (z->img_comp[i].raw_data == NULL)
+         return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
+      // align blocks for idct using mmx/sse
+      z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15);
+      if (z->progressive) {
+         // w2, h2 are multiples of 8 (see above)
+         z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8;
+         z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8;
+         z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15);
+         if (z->img_comp[i].raw_coeff == NULL)
+            return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
+         z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15);
+      }
+   }
+
+   return 1;
+}
+
+// use comparisons since in some cases we handle more than one case (e.g. SOF)
+#define stbi__DNL(x)         ((x) == 0xdc)
+#define stbi__SOI(x)         ((x) == 0xd8)
+#define stbi__EOI(x)         ((x) == 0xd9)
+#define stbi__SOF(x)         ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2)
+#define stbi__SOS(x)         ((x) == 0xda)
+
+#define stbi__SOF_progressive(x)   ((x) == 0xc2)
+
+static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan)
+{
+   int m;
+   z->jfif = 0;
+   z->app14_color_transform = -1; // valid values are 0,1,2
+   z->marker = STBI__MARKER_none; // initialize cached marker to empty
+   m = stbi__get_marker(z);
+   if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG");
+   if (scan == STBI__SCAN_type) return 1;
+   m = stbi__get_marker(z);
+   while (!stbi__SOF(m)) {
+      if (!stbi__process_marker(z,m)) return 0;
+      m = stbi__get_marker(z);
+      while (m == STBI__MARKER_none) {
+         // some files have extra padding after their blocks, so ok, we'll scan
+         if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG");
+         m = stbi__get_marker(z);
+      }
+   }
+   z->progressive = stbi__SOF_progressive(m);
+   if (!stbi__process_frame_header(z, scan)) return 0;
+   return 1;
+}
+
+static stbi_uc stbi__skip_jpeg_junk_at_end(stbi__jpeg *j)
+{
+   // some JPEGs have junk at end, skip over it but if we find what looks
+   // like a valid marker, resume there
+   while (!stbi__at_eof(j->s)) {
+      stbi_uc x = stbi__get8(j->s);
+      while (x == 0xff) { // might be a marker
+         if (stbi__at_eof(j->s)) return STBI__MARKER_none;
+         x = stbi__get8(j->s);
+         if (x != 0x00 && x != 0xff) {
+            // not a stuffed zero or lead-in to another marker, looks
+            // like an actual marker, return it
+            return x;
+         }
+         // stuffed zero has x=0 now which ends the loop, meaning we go
+         // back to regular scan loop.
+         // repeated 0xff keeps trying to read the next byte of the marker.
+      }
+   }
+   return STBI__MARKER_none;
+}
+
+// decode image to YCbCr format
+static int stbi__decode_jpeg_image(stbi__jpeg *j)
+{
+   int m;
+   for (m = 0; m < 4; m++) {
+      j->img_comp[m].raw_data = NULL;
+      j->img_comp[m].raw_coeff = NULL;
+   }
+   j->restart_interval = 0;
+   if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0;
+   m = stbi__get_marker(j);
+   while (!stbi__EOI(m)) {
+      if (stbi__SOS(m)) {
+         if (!stbi__process_scan_header(j)) return 0;
+         if (!stbi__parse_entropy_coded_data(j)) return 0;
+         if (j->marker == STBI__MARKER_none ) {
+         j->marker = stbi__skip_jpeg_junk_at_end(j);
+            // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0
+         }
+         m = stbi__get_marker(j);
+         if (STBI__RESTART(m))
+            m = stbi__get_marker(j);
+      } else if (stbi__DNL(m)) {
+         int Ld = stbi__get16be(j->s);
+         stbi__uint32 NL = stbi__get16be(j->s);
+         if (Ld != 4) return stbi__err("bad DNL len", "Corrupt JPEG");
+         if (NL != j->s->img_y) return stbi__err("bad DNL height", "Corrupt JPEG");
+         m = stbi__get_marker(j);
+      } else {
+         if (!stbi__process_marker(j, m)) return 1;
+         m = stbi__get_marker(j);
+      }
+   }
+   if (j->progressive)
+      stbi__jpeg_finish(j);
+   return 1;
+}
+
+// static jfif-centered resampling (across block boundaries)
+
+typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1,
+                                    int w, int hs);
+
+#define stbi__div4(x) ((stbi_uc) ((x) >> 2))
+
+static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   STBI_NOTUSED(out);
+   STBI_NOTUSED(in_far);
+   STBI_NOTUSED(w);
+   STBI_NOTUSED(hs);
+   return in_near;
+}
+
+static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate two samples vertically for every one in input
+   int i;
+   STBI_NOTUSED(hs);
+   for (i=0; i < w; ++i)
+      out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2);
+   return out;
+}
+
+static stbi_uc*  stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate two samples horizontally for every one in input
+   int i;
+   stbi_uc *input = in_near;
+
+   if (w == 1) {
+      // if only one sample, can't do any interpolation
+      out[0] = out[1] = input[0];
+      return out;
+   }
+
+   out[0] = input[0];
+   out[1] = stbi__div4(input[0]*3 + input[1] + 2);
+   for (i=1; i < w-1; ++i) {
+      int n = 3*input[i]+2;
+      out[i*2+0] = stbi__div4(n+input[i-1]);
+      out[i*2+1] = stbi__div4(n+input[i+1]);
+   }
+   out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2);
+   out[i*2+1] = input[w-1];
+
+   STBI_NOTUSED(in_far);
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+
+#define stbi__div16(x) ((stbi_uc) ((x) >> 4))
+
+static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate 2x2 samples for every one in input
+   int i,t0,t1;
+   if (w == 1) {
+      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
+      return out;
+   }
+
+   t1 = 3*in_near[0] + in_far[0];
+   out[0] = stbi__div4(t1+2);
+   for (i=1; i < w; ++i) {
+      t0 = t1;
+      t1 = 3*in_near[i]+in_far[i];
+      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
+      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
+   }
+   out[w*2-1] = stbi__div4(t1+2);
+
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate 2x2 samples for every one in input
+   int i=0,t0,t1;
+
+   if (w == 1) {
+      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
+      return out;
+   }
+
+   t1 = 3*in_near[0] + in_far[0];
+   // process groups of 8 pixels for as long as we can.
+   // note we can't handle the last pixel in a row in this loop
+   // because we need to handle the filter boundary conditions.
+   for (; i < ((w-1) & ~7); i += 8) {
+#if defined(STBI_SSE2)
+      // load and perform the vertical filtering pass
+      // this uses 3*x + y = 4*x + (y - x)
+      __m128i zero  = _mm_setzero_si128();
+      __m128i farb  = _mm_loadl_epi64((__m128i *) (in_far + i));
+      __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i));
+      __m128i farw  = _mm_unpacklo_epi8(farb, zero);
+      __m128i nearw = _mm_unpacklo_epi8(nearb, zero);
+      __m128i diff  = _mm_sub_epi16(farw, nearw);
+      __m128i nears = _mm_slli_epi16(nearw, 2);
+      __m128i curr  = _mm_add_epi16(nears, diff); // current row
+
+      // horizontal filter works the same based on shifted vers of current
+      // row. "prev" is current row shifted right by 1 pixel; we need to
+      // insert the previous pixel value (from t1).
+      // "next" is current row shifted left by 1 pixel, with first pixel
+      // of next block of 8 pixels added in.
+      __m128i prv0 = _mm_slli_si128(curr, 2);
+      __m128i nxt0 = _mm_srli_si128(curr, 2);
+      __m128i prev = _mm_insert_epi16(prv0, t1, 0);
+      __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7);
+
+      // horizontal filter, polyphase implementation since it's convenient:
+      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+      // note the shared term.
+      __m128i bias  = _mm_set1_epi16(8);
+      __m128i curs = _mm_slli_epi16(curr, 2);
+      __m128i prvd = _mm_sub_epi16(prev, curr);
+      __m128i nxtd = _mm_sub_epi16(next, curr);
+      __m128i curb = _mm_add_epi16(curs, bias);
+      __m128i even = _mm_add_epi16(prvd, curb);
+      __m128i odd  = _mm_add_epi16(nxtd, curb);
+
+      // interleave even and odd pixels, then undo scaling.
+      __m128i int0 = _mm_unpacklo_epi16(even, odd);
+      __m128i int1 = _mm_unpackhi_epi16(even, odd);
+      __m128i de0  = _mm_srli_epi16(int0, 4);
+      __m128i de1  = _mm_srli_epi16(int1, 4);
+
+      // pack and write output
+      __m128i outv = _mm_packus_epi16(de0, de1);
+      _mm_storeu_si128((__m128i *) (out + i*2), outv);
+#elif defined(STBI_NEON)
+      // load and perform the vertical filtering pass
+      // this uses 3*x + y = 4*x + (y - x)
+      uint8x8_t farb  = vld1_u8(in_far + i);
+      uint8x8_t nearb = vld1_u8(in_near + i);
+      int16x8_t diff  = vreinterpretq_s16_u16(vsubl_u8(farb, nearb));
+      int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2));
+      int16x8_t curr  = vaddq_s16(nears, diff); // current row
+
+      // horizontal filter works the same based on shifted vers of current
+      // row. "prev" is current row shifted right by 1 pixel; we need to
+      // insert the previous pixel value (from t1).
+      // "next" is current row shifted left by 1 pixel, with first pixel
+      // of next block of 8 pixels added in.
+      int16x8_t prv0 = vextq_s16(curr, curr, 7);
+      int16x8_t nxt0 = vextq_s16(curr, curr, 1);
+      int16x8_t prev = vsetq_lane_s16(t1, prv0, 0);
+      int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7);
+
+      // horizontal filter, polyphase implementation since it's convenient:
+      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+      // note the shared term.
+      int16x8_t curs = vshlq_n_s16(curr, 2);
+      int16x8_t prvd = vsubq_s16(prev, curr);
+      int16x8_t nxtd = vsubq_s16(next, curr);
+      int16x8_t even = vaddq_s16(curs, prvd);
+      int16x8_t odd  = vaddq_s16(curs, nxtd);
+
+      // undo scaling and round, then store with even/odd phases interleaved
+      uint8x8x2_t o;
+      o.val[0] = vqrshrun_n_s16(even, 4);
+      o.val[1] = vqrshrun_n_s16(odd,  4);
+      vst2_u8(out + i*2, o);
+#endif
+
+      // "previous" value for next iter
+      t1 = 3*in_near[i+7] + in_far[i+7];
+   }
+
+   t0 = t1;
+   t1 = 3*in_near[i] + in_far[i];
+   out[i*2] = stbi__div16(3*t1 + t0 + 8);
+
+   for (++i; i < w; ++i) {
+      t0 = t1;
+      t1 = 3*in_near[i]+in_far[i];
+      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
+      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
+   }
+   out[w*2-1] = stbi__div4(t1+2);
+
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+#endif
+
+static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // resample with nearest-neighbor
+   int i,j;
+   STBI_NOTUSED(in_far);
+   for (i=0; i < w; ++i)
+      for (j=0; j < hs; ++j)
+         out[i*hs+j] = in_near[i];
+   return out;
+}
+
+// this is a reduced-precision calculation of YCbCr-to-RGB introduced
+// to make sure the code produces the same results in both SIMD and scalar
+#define stbi__float2fixed(x)  (((int) ((x) * 4096.0f + 0.5f)) << 8)
+static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step)
+{
+   int i;
+   for (i=0; i < count; ++i) {
+      int y_fixed = (y[i] << 20) + (1<<19); // rounding
+      int r,g,b;
+      int cr = pcr[i] - 128;
+      int cb = pcb[i] - 128;
+      r = y_fixed +  cr* stbi__float2fixed(1.40200f);
+      g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
+      b = y_fixed                                     +   cb* stbi__float2fixed(1.77200f);
+      r >>= 20;
+      g >>= 20;
+      b >>= 20;
+      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
+      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
+      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
+      out[0] = (stbi_uc)r;
+      out[1] = (stbi_uc)g;
+      out[2] = (stbi_uc)b;
+      out[3] = 255;
+      out += step;
+   }
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step)
+{
+   int i = 0;
+
+#ifdef STBI_SSE2
+   // step == 3 is pretty ugly on the final interleave, and i'm not convinced
+   // it's useful in practice (you wouldn't use it for textures, for example).
+   // so just accelerate step == 4 case.
+   if (step == 4) {
+      // this is a fairly straightforward implementation and not super-optimized.
+      __m128i signflip  = _mm_set1_epi8(-0x80);
+      __m128i cr_const0 = _mm_set1_epi16(   (short) ( 1.40200f*4096.0f+0.5f));
+      __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f));
+      __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f));
+      __m128i cb_const1 = _mm_set1_epi16(   (short) ( 1.77200f*4096.0f+0.5f));
+      __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128);
+      __m128i xw = _mm_set1_epi16(255); // alpha channel
+
+      for (; i+7 < count; i += 8) {
+         // load
+         __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i));
+         __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i));
+         __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i));
+         __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128
+         __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128
+
+         // unpack to short (and left-shift cr, cb by 8)
+         __m128i yw  = _mm_unpacklo_epi8(y_bias, y_bytes);
+         __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased);
+         __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased);
+
+         // color transform
+         __m128i yws = _mm_srli_epi16(yw, 4);
+         __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw);
+         __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw);
+         __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1);
+         __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1);
+         __m128i rws = _mm_add_epi16(cr0, yws);
+         __m128i gwt = _mm_add_epi16(cb0, yws);
+         __m128i bws = _mm_add_epi16(yws, cb1);
+         __m128i gws = _mm_add_epi16(gwt, cr1);
+
+         // descale
+         __m128i rw = _mm_srai_epi16(rws, 4);
+         __m128i bw = _mm_srai_epi16(bws, 4);
+         __m128i gw = _mm_srai_epi16(gws, 4);
+
+         // back to byte, set up for transpose
+         __m128i brb = _mm_packus_epi16(rw, bw);
+         __m128i gxb = _mm_packus_epi16(gw, xw);
+
+         // transpose to interleave channels
+         __m128i t0 = _mm_unpacklo_epi8(brb, gxb);
+         __m128i t1 = _mm_unpackhi_epi8(brb, gxb);
+         __m128i o0 = _mm_unpacklo_epi16(t0, t1);
+         __m128i o1 = _mm_unpackhi_epi16(t0, t1);
+
+         // store
+         _mm_storeu_si128((__m128i *) (out + 0), o0);
+         _mm_storeu_si128((__m128i *) (out + 16), o1);
+         out += 32;
+      }
+   }
+#endif
+
+#ifdef STBI_NEON
+   // in this version, step=3 support would be easy to add. but is there demand?
+   if (step == 4) {
+      // this is a fairly straightforward implementation and not super-optimized.
+      uint8x8_t signflip = vdup_n_u8(0x80);
+      int16x8_t cr_const0 = vdupq_n_s16(   (short) ( 1.40200f*4096.0f+0.5f));
+      int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f));
+      int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f));
+      int16x8_t cb_const1 = vdupq_n_s16(   (short) ( 1.77200f*4096.0f+0.5f));
+
+      for (; i+7 < count; i += 8) {
+         // load
+         uint8x8_t y_bytes  = vld1_u8(y + i);
+         uint8x8_t cr_bytes = vld1_u8(pcr + i);
+         uint8x8_t cb_bytes = vld1_u8(pcb + i);
+         int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip));
+         int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip));
+
+         // expand to s16
+         int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4));
+         int16x8_t crw = vshll_n_s8(cr_biased, 7);
+         int16x8_t cbw = vshll_n_s8(cb_biased, 7);
+
+         // color transform
+         int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0);
+         int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0);
+         int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1);
+         int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1);
+         int16x8_t rws = vaddq_s16(yws, cr0);
+         int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1);
+         int16x8_t bws = vaddq_s16(yws, cb1);
+
+         // undo scaling, round, convert to byte
+         uint8x8x4_t o;
+         o.val[0] = vqrshrun_n_s16(rws, 4);
+         o.val[1] = vqrshrun_n_s16(gws, 4);
+         o.val[2] = vqrshrun_n_s16(bws, 4);
+         o.val[3] = vdup_n_u8(255);
+
+         // store, interleaving r/g/b/a
+         vst4_u8(out, o);
+         out += 8*4;
+      }
+   }
+#endif
+
+   for (; i < count; ++i) {
+      int y_fixed = (y[i] << 20) + (1<<19); // rounding
+      int r,g,b;
+      int cr = pcr[i] - 128;
+      int cb = pcb[i] - 128;
+      r = y_fixed + cr* stbi__float2fixed(1.40200f);
+      g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
+      b = y_fixed                                   +   cb* stbi__float2fixed(1.77200f);
+      r >>= 20;
+      g >>= 20;
+      b >>= 20;
+      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
+      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
+      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
+      out[0] = (stbi_uc)r;
+      out[1] = (stbi_uc)g;
+      out[2] = (stbi_uc)b;
+      out[3] = 255;
+      out += step;
+   }
+}
+#endif
+
+// set up the kernels
+static void stbi__setup_jpeg(stbi__jpeg *j)
+{
+   j->idct_block_kernel = stbi__idct_block;
+   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row;
+   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2;
+
+#ifdef STBI_SSE2
+   if (stbi__sse2_available()) {
+      j->idct_block_kernel = stbi__idct_simd;
+      j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+      j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+   }
+#endif
+
+#ifdef STBI_NEON
+   j->idct_block_kernel = stbi__idct_simd;
+   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+#endif
+}
+
+// clean up the temporary component buffers
+static void stbi__cleanup_jpeg(stbi__jpeg *j)
+{
+   stbi__free_jpeg_components(j, j->s->img_n, 0);
+}
+
+typedef struct
+{
+   resample_row_func resample;
+   stbi_uc *line0,*line1;
+   int hs,vs;   // expansion factor in each axis
+   int w_lores; // horizontal pixels pre-expansion
+   int ystep;   // how far through vertical expansion we are
+   int ypos;    // which pre-expansion row we're on
+} stbi__resample;
+
+// fast 0..255 * 0..255 => 0..255 rounded multiplication
+static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y)
+{
+   unsigned int t = x*y + 128;
+   return (stbi_uc) ((t + (t >>8)) >> 8);
+}
+
+static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp)
+{
+   int n, decode_n, is_rgb;
+   z->s->img_n = 0; // make stbi__cleanup_jpeg safe
+
+   // validate req_comp
+   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
+
+   // load a jpeg image from whichever source, but leave in YCbCr format
+   if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; }
+
+   // determine actual number of components to generate
+   n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1;
+
+   is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif));
+
+   if (z->s->img_n == 3 && n < 3 && !is_rgb)
+      decode_n = 1;
+   else
+      decode_n = z->s->img_n;
+
+   // nothing to do if no components requested; check this now to avoid
+   // accessing uninitialized coutput[0] later
+   if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; }
+
+   // resample and color-convert
+   {
+      int k;
+      unsigned int i,j;
+      stbi_uc *output;
+      stbi_uc *coutput[4] = { NULL, NULL, NULL, NULL };
+
+      stbi__resample res_comp[4];
+
+      for (k=0; k < decode_n; ++k) {
+         stbi__resample *r = &res_comp[k];
+
+         // allocate line buffer big enough for upsampling off the edges
+         // with upsample factor of 4
+         z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3);
+         if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
+
+         r->hs      = z->img_h_max / z->img_comp[k].h;
+         r->vs      = z->img_v_max / z->img_comp[k].v;
+         r->ystep   = r->vs >> 1;
+         r->w_lores = (z->s->img_x + r->hs-1) / r->hs;
+         r->ypos    = 0;
+         r->line0   = r->line1 = z->img_comp[k].data;
+
+         if      (r->hs == 1 && r->vs == 1) r->resample = resample_row_1;
+         else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2;
+         else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2;
+         else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel;
+         else                               r->resample = stbi__resample_row_generic;
+      }
+
+      // can't error after this so, this is safe
+      output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1);
+      if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
+
+      // now go ahead and resample
+      for (j=0; j < z->s->img_y; ++j) {
+         stbi_uc *out = output + n * z->s->img_x * j;
+         for (k=0; k < decode_n; ++k) {
+            stbi__resample *r = &res_comp[k];
+            int y_bot = r->ystep >= (r->vs >> 1);
+            coutput[k] = r->resample(z->img_comp[k].linebuf,
+                                     y_bot ? r->line1 : r->line0,
+                                     y_bot ? r->line0 : r->line1,
+                                     r->w_lores, r->hs);
+            if (++r->ystep >= r->vs) {
+               r->ystep = 0;
+               r->line0 = r->line1;
+               if (++r->ypos < z->img_comp[k].y)
+                  r->line1 += z->img_comp[k].w2;
+            }
+         }
+         if (n >= 3) {
+            stbi_uc *y = coutput[0];
+            if (z->s->img_n == 3) {
+               if (is_rgb) {
+                  for (i=0; i < z->s->img_x; ++i) {
+                     out[0] = y[i];
+                     out[1] = coutput[1][i];
+                     out[2] = coutput[2][i];
+                     out[3] = 255;
+                     out += n;
+                  }
+               } else {
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+               }
+            } else if (z->s->img_n == 4) {
+               if (z->app14_color_transform == 0) { // CMYK
+                  for (i=0; i < z->s->img_x; ++i) {
+                     stbi_uc m = coutput[3][i];
+                     out[0] = stbi__blinn_8x8(coutput[0][i], m);
+                     out[1] = stbi__blinn_8x8(coutput[1][i], m);
+                     out[2] = stbi__blinn_8x8(coutput[2][i], m);
+                     out[3] = 255;
+                     out += n;
+                  }
+               } else if (z->app14_color_transform == 2) { // YCCK
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+                  for (i=0; i < z->s->img_x; ++i) {
+                     stbi_uc m = coutput[3][i];
+                     out[0] = stbi__blinn_8x8(255 - out[0], m);
+                     out[1] = stbi__blinn_8x8(255 - out[1], m);
+                     out[2] = stbi__blinn_8x8(255 - out[2], m);
+                     out += n;
+                  }
+               } else { // YCbCr + alpha?  Ignore the fourth channel for now
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+               }
+            } else
+               for (i=0; i < z->s->img_x; ++i) {
+                  out[0] = out[1] = out[2] = y[i];
+                  out[3] = 255; // not used if n==3
+                  out += n;
+               }
+         } else {
+            if (is_rgb) {
+               if (n == 1)
+                  for (i=0; i < z->s->img_x; ++i)
+                     *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+               else {
+                  for (i=0; i < z->s->img_x; ++i, out += 2) {
+                     out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+                     out[1] = 255;
+                  }
+               }
+            } else if (z->s->img_n == 4 && z->app14_color_transform == 0) {
+               for (i=0; i < z->s->img_x; ++i) {
+                  stbi_uc m = coutput[3][i];
+                  stbi_uc r = stbi__blinn_8x8(coutput[0][i], m);
+                  stbi_uc g = stbi__blinn_8x8(coutput[1][i], m);
+                  stbi_uc b = stbi__blinn_8x8(coutput[2][i], m);
+                  out[0] = stbi__compute_y(r, g, b);
+                  out[1] = 255;
+                  out += n;
+               }
+            } else if (z->s->img_n == 4 && z->app14_color_transform == 2) {
+               for (i=0; i < z->s->img_x; ++i) {
+                  out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]);
+                  out[1] = 255;
+                  out += n;
+               }
+            } else {
+               stbi_uc *y = coutput[0];
+               if (n == 1)
+                  for (i=0; i < z->s->img_x; ++i) out[i] = y[i];
+               else
+                  for (i=0; i < z->s->img_x; ++i) { *out++ = y[i]; *out++ = 255; }
+            }
+         }
+      }
+      stbi__cleanup_jpeg(z);
+      *out_x = z->s->img_x;
+      *out_y = z->s->img_y;
+      if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output
+      return output;
+   }
+}
+
+static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   unsigned char* result;
+   stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__errpuc("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   STBI_NOTUSED(ri);
+   j->s = s;
+   stbi__setup_jpeg(j);
+   result = load_jpeg_image(j, x,y,comp,req_comp);
+   STBI_FREE(j);
+   return result;
+}
+
+static int stbi__jpeg_test(stbi__context *s)
+{
+   int r;
+   stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__err("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   j->s = s;
+   stbi__setup_jpeg(j);
+   r = stbi__decode_jpeg_header(j, STBI__SCAN_type);
+   stbi__rewind(s);
+   STBI_FREE(j);
+   return r;
+}
+
+static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp)
+{
+   if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) {
+      stbi__rewind( j->s );
+      return 0;
+   }
+   if (x) *x = j->s->img_x;
+   if (y) *y = j->s->img_y;
+   if (comp) *comp = j->s->img_n >= 3 ? 3 : 1;
+   return 1;
+}
+
+static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int result;
+   stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg)));
+   if (!j) return stbi__err("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   j->s = s;
+   result = stbi__jpeg_info_raw(j, x, y, comp);
+   STBI_FREE(j);
+   return result;
+}
+#endif
+
+// public domain zlib decode    v0.2  Sean Barrett 2006-11-18
+//    simple implementation
+//      - all input must be provided in an upfront buffer
+//      - all output is written to a single output buffer (can malloc/realloc)
+//    performance
+//      - fast huffman
+
+#ifndef STBI_NO_ZLIB
+
+// fast-way is faster to check than jpeg huffman, but slow way is slower
+#define STBI__ZFAST_BITS  9 // accelerate all cases in default tables
+#define STBI__ZFAST_MASK  ((1 << STBI__ZFAST_BITS) - 1)
+#define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet
+
+// zlib-style huffman encoding
+// (jpegs packs from left, zlib from right, so can't share code)
+typedef struct
+{
+   stbi__uint16 fast[1 << STBI__ZFAST_BITS];
+   stbi__uint16 firstcode[16];
+   int maxcode[17];
+   stbi__uint16 firstsymbol[16];
+   stbi_uc  size[STBI__ZNSYMS];
+   stbi__uint16 value[STBI__ZNSYMS];
+} stbi__zhuffman;
+
+stbi_inline static int stbi__bitreverse16(int n)
+{
+  n = ((n & 0xAAAA) >>  1) | ((n & 0x5555) << 1);
+  n = ((n & 0xCCCC) >>  2) | ((n & 0x3333) << 2);
+  n = ((n & 0xF0F0) >>  4) | ((n & 0x0F0F) << 4);
+  n = ((n & 0xFF00) >>  8) | ((n & 0x00FF) << 8);
+  return n;
+}
+
+stbi_inline static int stbi__bit_reverse(int v, int bits)
+{
+   STBI_ASSERT(bits <= 16);
+   // to bit reverse n bits, reverse 16 and shift
+   // e.g. 11 bits, bit reverse and shift away 5
+   return stbi__bitreverse16(v) >> (16-bits);
+}
+
+static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num)
+{
+   int i,k=0;
+   int code, next_code[16], sizes[17];
+
+   // DEFLATE spec for generating codes
+   memset(sizes, 0, sizeof(sizes));
+   memset(z->fast, 0, sizeof(z->fast));
+   for (i=0; i < num; ++i)
+      ++sizes[sizelist[i]];
+   sizes[0] = 0;
+   for (i=1; i < 16; ++i)
+      if (sizes[i] > (1 << i))
+         return stbi__err("bad sizes", "Corrupt PNG");
+   code = 0;
+   for (i=1; i < 16; ++i) {
+      next_code[i] = code;
+      z->firstcode[i] = (stbi__uint16) code;
+      z->firstsymbol[i] = (stbi__uint16) k;
+      code = (code + sizes[i]);
+      if (sizes[i])
+         if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG");
+      z->maxcode[i] = code << (16-i); // preshift for inner loop
+      code <<= 1;
+      k += sizes[i];
+   }
+   z->maxcode[16] = 0x10000; // sentinel
+   for (i=0; i < num; ++i) {
+      int s = sizelist[i];
+      if (s) {
+         int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s];
+         stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i);
+         z->size [c] = (stbi_uc     ) s;
+         z->value[c] = (stbi__uint16) i;
+         if (s <= STBI__ZFAST_BITS) {
+            int j = stbi__bit_reverse(next_code[s],s);
+            while (j < (1 << STBI__ZFAST_BITS)) {
+               z->fast[j] = fastv;
+               j += (1 << s);
+            }
+         }
+         ++next_code[s];
+      }
+   }
+   return 1;
+}
+
+// zlib-from-memory implementation for PNG reading
+//    because PNG allows splitting the zlib stream arbitrarily,
+//    and it's annoying structurally to have PNG call ZLIB call PNG,
+//    we require PNG read all the IDATs and combine them into a single
+//    memory buffer
+
+typedef struct
+{
+   stbi_uc *zbuffer, *zbuffer_end;
+   int num_bits;
+   int hit_zeof_once;
+   stbi__uint32 code_buffer;
+
+   char *zout;
+   char *zout_start;
+   char *zout_end;
+   int   z_expandable;
+
+   stbi__zhuffman z_length, z_distance;
+} stbi__zbuf;
+
+stbi_inline static int stbi__zeof(stbi__zbuf *z)
+{
+   return (z->zbuffer >= z->zbuffer_end);
+}
+
+stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z)
+{
+   return stbi__zeof(z) ? 0 : *z->zbuffer++;
+}
+
+static void stbi__fill_bits(stbi__zbuf *z)
+{
+   do {
+      if (z->code_buffer >= (1U << z->num_bits)) {
+        z->zbuffer = z->zbuffer_end;  /* treat this as EOF so we fail. */
+        return;
+      }
+      z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits;
+      z->num_bits += 8;
+   } while (z->num_bits <= 24);
+}
+
+stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n)
+{
+   unsigned int k;
+   if (z->num_bits < n) stbi__fill_bits(z);
+   k = z->code_buffer & ((1 << n) - 1);
+   z->code_buffer >>= n;
+   z->num_bits -= n;
+   return k;
+}
+
+static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z)
+{
+   int b,s,k;
+   // not resolved by fast table, so compute it the slow way
+   // use jpeg approach, which requires MSbits at top
+   k = stbi__bit_reverse(a->code_buffer, 16);
+   for (s=STBI__ZFAST_BITS+1; ; ++s)
+      if (k < z->maxcode[s])
+         break;
+   if (s >= 16) return -1; // invalid code!
+   // code size is s, so:
+   b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s];
+   if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere!
+   if (z->size[b] != s) return -1;  // was originally an assert, but report failure instead.
+   a->code_buffer >>= s;
+   a->num_bits -= s;
+   return z->value[b];
+}
+
+stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z)
+{
+   int b,s;
+   if (a->num_bits < 16) {
+      if (stbi__zeof(a)) {
+         if (!a->hit_zeof_once) {
+            // This is the first time we hit eof, insert 16 extra padding btis
+            // to allow us to keep going; if we actually consume any of them
+            // though, that is invalid data. This is caught later.
+            a->hit_zeof_once = 1;
+            a->num_bits += 16; // add 16 implicit zero bits
+         } else {
+            // We already inserted our extra 16 padding bits and are again
+            // out, this stream is actually prematurely terminated.
+            return -1;
+         }
+      } else {
+         stbi__fill_bits(a);
+      }
+   }
+   b = z->fast[a->code_buffer & STBI__ZFAST_MASK];
+   if (b) {
+      s = b >> 9;
+      a->code_buffer >>= s;
+      a->num_bits -= s;
+      return b & 511;
+   }
+   return stbi__zhuffman_decode_slowpath(a, z);
+}
+
+static int stbi__zexpand(stbi__zbuf *z, char *zout, int n)  // need to make room for n bytes
+{
+   char *q;
+   unsigned int cur, limit, old_limit;
+   z->zout = zout;
+   if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG");
+   cur   = (unsigned int) (z->zout - z->zout_start);
+   limit = old_limit = (unsigned) (z->zout_end - z->zout_start);
+   if (UINT_MAX - cur < (unsigned) n) return stbi__err("outofmem", "Out of memory");
+   while (cur + n > limit) {
+      if(limit > UINT_MAX / 2) return stbi__err("outofmem", "Out of memory");
+      limit *= 2;
+   }
+   q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit);
+   STBI_NOTUSED(old_limit);
+   if (q == NULL) return stbi__err("outofmem", "Out of memory");
+   z->zout_start = q;
+   z->zout       = q + cur;
+   z->zout_end   = q + limit;
+   return 1;
+}
+
+static const int stbi__zlength_base[31] = {
+   3,4,5,6,7,8,9,10,11,13,
+   15,17,19,23,27,31,35,43,51,59,
+   67,83,99,115,131,163,195,227,258,0,0 };
+
+static const int stbi__zlength_extra[31]=
+{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 };
+
+static const int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,
+257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0};
+
+static const int stbi__zdist_extra[32] =
+{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13};
+
+static int stbi__parse_huffman_block(stbi__zbuf *a)
+{
+   char *zout = a->zout;
+   for(;;) {
+      int z = stbi__zhuffman_decode(a, &a->z_length);
+      if (z < 256) {
+         if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes
+         if (zout >= a->zout_end) {
+            if (!stbi__zexpand(a, zout, 1)) return 0;
+            zout = a->zout;
+         }
+         *zout++ = (char) z;
+      } else {
+         stbi_uc *p;
+         int len,dist;
+         if (z == 256) {
+            a->zout = zout;
+            if (a->hit_zeof_once && a->num_bits < 16) {
+               // The first time we hit zeof, we inserted 16 extra zero bits into our bit
+               // buffer so the decoder can just do its speculative decoding. But if we
+               // actually consumed any of those bits (which is the case when num_bits < 16),
+               // the stream actually read past the end so it is malformed.
+               return stbi__err("unexpected end","Corrupt PNG");
+            }
+            return 1;
+         }
+         if (z >= 286) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data
+         z -= 257;
+         len = stbi__zlength_base[z];
+         if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]);
+         z = stbi__zhuffman_decode(a, &a->z_distance);
+         if (z < 0 || z >= 30) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data
+         dist = stbi__zdist_base[z];
+         if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]);
+         if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG");
+         if (len > a->zout_end - zout) {
+            if (!stbi__zexpand(a, zout, len)) return 0;
+            zout = a->zout;
+         }
+         p = (stbi_uc *) (zout - dist);
+         if (dist == 1) { // run of one byte; common in images.
+            stbi_uc v = *p;
+            if (len) { do *zout++ = v; while (--len); }
+         } else {
+            if (len) { do *zout++ = *p++; while (--len); }
+         }
+      }
+   }
+}
+
+static int stbi__compute_huffman_codes(stbi__zbuf *a)
+{
+   static const stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 };
+   stbi__zhuffman z_codelength;
+   stbi_uc lencodes[286+32+137];//padding for maximum single op
+   stbi_uc codelength_sizes[19];
+   int i,n;
+
+   int hlit  = stbi__zreceive(a,5) + 257;
+   int hdist = stbi__zreceive(a,5) + 1;
+   int hclen = stbi__zreceive(a,4) + 4;
+   int ntot  = hlit + hdist;
+
+   memset(codelength_sizes, 0, sizeof(codelength_sizes));
+   for (i=0; i < hclen; ++i) {
+      int s = stbi__zreceive(a,3);
+      codelength_sizes[length_dezigzag[i]] = (stbi_uc) s;
+   }
+   if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0;
+
+   n = 0;
+   while (n < ntot) {
+      int c = stbi__zhuffman_decode(a, &z_codelength);
+      if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG");
+      if (c < 16)
+         lencodes[n++] = (stbi_uc) c;
+      else {
+         stbi_uc fill = 0;
+         if (c == 16) {
+            c = stbi__zreceive(a,2)+3;
+            if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG");
+            fill = lencodes[n-1];
+         } else if (c == 17) {
+            c = stbi__zreceive(a,3)+3;
+         } else if (c == 18) {
+            c = stbi__zreceive(a,7)+11;
+         } else {
+            return stbi__err("bad codelengths", "Corrupt PNG");
+         }
+         if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG");
+         memset(lencodes+n, fill, c);
+         n += c;
+      }
+   }
+   if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG");
+   if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0;
+   if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0;
+   return 1;
+}
+
+static int stbi__parse_uncompressed_block(stbi__zbuf *a)
+{
+   stbi_uc header[4];
+   int len,nlen,k;
+   if (a->num_bits & 7)
+      stbi__zreceive(a, a->num_bits & 7); // discard
+   // drain the bit-packed data into header
+   k = 0;
+   while (a->num_bits > 0) {
+      header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check
+      a->code_buffer >>= 8;
+      a->num_bits -= 8;
+   }
+   if (a->num_bits < 0) return stbi__err("zlib corrupt","Corrupt PNG");
+   // now fill header the normal way
+   while (k < 4)
+      header[k++] = stbi__zget8(a);
+   len  = header[1] * 256 + header[0];
+   nlen = header[3] * 256 + header[2];
+   if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG");
+   if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG");
+   if (a->zout + len > a->zout_end)
+      if (!stbi__zexpand(a, a->zout, len)) return 0;
+   memcpy(a->zout, a->zbuffer, len);
+   a->zbuffer += len;
+   a->zout += len;
+   return 1;
+}
+
+static int stbi__parse_zlib_header(stbi__zbuf *a)
+{
+   int cmf   = stbi__zget8(a);
+   int cm    = cmf & 15;
+   /* int cinfo = cmf >> 4; */
+   int flg   = stbi__zget8(a);
+   if (stbi__zeof(a)) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
+   if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
+   if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png
+   if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png
+   // window = 1 << (8 + cinfo)... but who cares, we fully buffer output
+   return 1;
+}
+
+static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] =
+{
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8
+};
+static const stbi_uc stbi__zdefault_distance[32] =
+{
+   5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
+};
+/*
+Init algorithm:
+{
+   int i;   // use <= to match clearly with spec
+   for (i=0; i <= 143; ++i)     stbi__zdefault_length[i]   = 8;
+   for (   ; i <= 255; ++i)     stbi__zdefault_length[i]   = 9;
+   for (   ; i <= 279; ++i)     stbi__zdefault_length[i]   = 7;
+   for (   ; i <= 287; ++i)     stbi__zdefault_length[i]   = 8;
+
+   for (i=0; i <=  31; ++i)     stbi__zdefault_distance[i] = 5;
+}
+*/
+
+static int stbi__parse_zlib(stbi__zbuf *a, int parse_header)
+{
+   int final, type;
+   if (parse_header)
+      if (!stbi__parse_zlib_header(a)) return 0;
+   a->num_bits = 0;
+   a->code_buffer = 0;
+   a->hit_zeof_once = 0;
+   do {
+      final = stbi__zreceive(a,1);
+      type = stbi__zreceive(a,2);
+      if (type == 0) {
+         if (!stbi__parse_uncompressed_block(a)) return 0;
+      } else if (type == 3) {
+         return 0;
+      } else {
+         if (type == 1) {
+            // use fixed code lengths
+            if (!stbi__zbuild_huffman(&a->z_length  , stbi__zdefault_length  , STBI__ZNSYMS)) return 0;
+            if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance,  32)) return 0;
+         } else {
+            if (!stbi__compute_huffman_codes(a)) return 0;
+         }
+         if (!stbi__parse_huffman_block(a)) return 0;
+      }
+   } while (!final);
+   return 1;
+}
+
+static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header)
+{
+   a->zout_start = obuf;
+   a->zout       = obuf;
+   a->zout_end   = obuf + olen;
+   a->z_expandable = exp;
+
+   return stbi__parse_zlib(a, parse_header);
+}
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(initial_size);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer + len;
+   if (stbi__do_zlib(&a, p, initial_size, 1, 1)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen)
+{
+   return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen);
+}
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(initial_size);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer + len;
+   if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen)
+{
+   stbi__zbuf a;
+   a.zbuffer = (stbi_uc *) ibuffer;
+   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
+   if (stbi__do_zlib(&a, obuffer, olen, 0, 1))
+      return (int) (a.zout - a.zout_start);
+   else
+      return -1;
+}
+
+STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(16384);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer+len;
+   if (stbi__do_zlib(&a, p, 16384, 1, 0)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen)
+{
+   stbi__zbuf a;
+   a.zbuffer = (stbi_uc *) ibuffer;
+   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
+   if (stbi__do_zlib(&a, obuffer, olen, 0, 0))
+      return (int) (a.zout - a.zout_start);
+   else
+      return -1;
+}
+#endif
+
+// public domain "baseline" PNG decoder   v0.10  Sean Barrett 2006-11-18
+//    simple implementation
+//      - only 8-bit samples
+//      - no CRC checking
+//      - allocates lots of intermediate memory
+//        - avoids problem of streaming data between subsystems
+//        - avoids explicit window management
+//    performance
+//      - uses stb_zlib, a PD zlib implementation with fast huffman decoding
+
+#ifndef STBI_NO_PNG
+typedef struct
+{
+   stbi__uint32 length;
+   stbi__uint32 type;
+} stbi__pngchunk;
+
+static stbi__pngchunk stbi__get_chunk_header(stbi__context *s)
+{
+   stbi__pngchunk c;
+   c.length = stbi__get32be(s);
+   c.type   = stbi__get32be(s);
+   return c;
+}
+
+static int stbi__check_png_header(stbi__context *s)
+{
+   static const stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 };
+   int i;
+   for (i=0; i < 8; ++i)
+      if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG");
+   return 1;
+}
+
+typedef struct
+{
+   stbi__context *s;
+   stbi_uc *idata, *expanded, *out;
+   int depth;
+} stbi__png;
+
+
+enum {
+   STBI__F_none=0,
+   STBI__F_sub=1,
+   STBI__F_up=2,
+   STBI__F_avg=3,
+   STBI__F_paeth=4,
+   // synthetic filter used for first scanline to avoid needing a dummy row of 0s
+   STBI__F_avg_first
+};
+
+static stbi_uc first_row_filter[5] =
+{
+   STBI__F_none,
+   STBI__F_sub,
+   STBI__F_none,
+   STBI__F_avg_first,
+   STBI__F_sub // Paeth with b=c=0 turns out to be equivalent to sub
+};
+
+static int stbi__paeth(int a, int b, int c)
+{
+   // This formulation looks very different from the reference in the PNG spec, but is
+   // actually equivalent and has favorable data dependencies and admits straightforward
+   // generation of branch-free code, which helps performance significantly.
+   int thresh = c*3 - (a + b);
+   int lo = a < b ? a : b;
+   int hi = a < b ? b : a;
+   int t0 = (hi <= thresh) ? lo : c;
+   int t1 = (thresh <= lo) ? hi : t0;
+   return t1;
+}
+
+static const stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 };
+
+// adds an extra all-255 alpha channel
+// dest == src is legal
+// img_n must be 1 or 3
+static void stbi__create_png_alpha_expand8(stbi_uc *dest, stbi_uc *src, stbi__uint32 x, int img_n)
+{
+   int i;
+   // must process data backwards since we allow dest==src
+   if (img_n == 1) {
+      for (i=x-1; i >= 0; --i) {
+         dest[i*2+1] = 255;
+         dest[i*2+0] = src[i];
+      }
+   } else {
+      STBI_ASSERT(img_n == 3);
+      for (i=x-1; i >= 0; --i) {
+         dest[i*4+3] = 255;
+         dest[i*4+2] = src[i*3+2];
+         dest[i*4+1] = src[i*3+1];
+         dest[i*4+0] = src[i*3+0];
+      }
+   }
+}
+
+// create the png data from post-deflated data
+static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color)
+{
+   int bytes = (depth == 16 ? 2 : 1);
+   stbi__context *s = a->s;
+   stbi__uint32 i,j,stride = x*out_n*bytes;
+   stbi__uint32 img_len, img_width_bytes;
+   stbi_uc *filter_buf;
+   int all_ok = 1;
+   int k;
+   int img_n = s->img_n; // copy it into a local for later
+
+   int output_bytes = out_n*bytes;
+   int filter_bytes = img_n*bytes;
+   int width = x;
+
+   STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1);
+   a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into
+   if (!a->out) return stbi__err("outofmem", "Out of memory");
+
+   // note: error exits here don't need to clean up a->out individually,
+   // stbi__do_png always does on error.
+   if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) return stbi__err("too large", "Corrupt PNG");
+   img_width_bytes = (((img_n * x * depth) + 7) >> 3);
+   if (!stbi__mad2sizes_valid(img_width_bytes, y, img_width_bytes)) return stbi__err("too large", "Corrupt PNG");
+   img_len = (img_width_bytes + 1) * y;
+
+   // we used to check for exact match between raw_len and img_len on non-interlaced PNGs,
+   // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros),
+   // so just check for raw_len < img_len always.
+   if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG");
+
+   // Allocate two scan lines worth of filter workspace buffer.
+   filter_buf = (stbi_uc *) stbi__malloc_mad2(img_width_bytes, 2, 0);
+   if (!filter_buf) return stbi__err("outofmem", "Out of memory");
+
+   // Filtering for low-bit-depth images
+   if (depth < 8) {
+      filter_bytes = 1;
+      width = img_width_bytes;
+   }
+
+   for (j=0; j < y; ++j) {
+      // cur/prior filter buffers alternate
+      stbi_uc *cur = filter_buf + (j & 1)*img_width_bytes;
+      stbi_uc *prior = filter_buf + (~j & 1)*img_width_bytes;
+      stbi_uc *dest = a->out + stride*j;
+      int nk = width * filter_bytes;
+      int filter = *raw++;
+
+      // check filter type
+      if (filter > 4) {
+         all_ok = stbi__err("invalid filter","Corrupt PNG");
+         break;
+      }
+
+      // if first row, use special filter that doesn't sample previous row
+      if (j == 0) filter = first_row_filter[filter];
+
+      // perform actual filtering
+      switch (filter) {
+      case STBI__F_none:
+         memcpy(cur, raw, nk);
+         break;
+      case STBI__F_sub:
+         memcpy(cur, raw, filter_bytes);
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]);
+         break;
+      case STBI__F_up:
+         for (k = 0; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + prior[k]);
+         break;
+      case STBI__F_avg:
+         for (k = 0; k < filter_bytes; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1));
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1));
+         break;
+      case STBI__F_paeth:
+         for (k = 0; k < filter_bytes; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + prior[k]); // prior[k] == stbi__paeth(0,prior[k],0)
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes], prior[k], prior[k-filter_bytes]));
+         break;
+      case STBI__F_avg_first:
+         memcpy(cur, raw, filter_bytes);
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1));
+         break;
+      }
+
+      raw += nk;
+
+      // expand decoded bits in cur to dest, also adding an extra alpha channel if desired
+      if (depth < 8) {
+         stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range
+         stbi_uc *in = cur;
+         stbi_uc *out = dest;
+         stbi_uc inb = 0;
+         stbi__uint32 nsmp = x*img_n;
+
+         // expand bits to bytes first
+         if (depth == 4) {
+            for (i=0; i < nsmp; ++i) {
+               if ((i & 1) == 0) inb = *in++;
+               *out++ = scale * (inb >> 4);
+               inb <<= 4;
+            }
+         } else if (depth == 2) {
+            for (i=0; i < nsmp; ++i) {
+               if ((i & 3) == 0) inb = *in++;
+               *out++ = scale * (inb >> 6);
+               inb <<= 2;
+            }
+         } else {
+            STBI_ASSERT(depth == 1);
+            for (i=0; i < nsmp; ++i) {
+               if ((i & 7) == 0) inb = *in++;
+               *out++ = scale * (inb >> 7);
+               inb <<= 1;
+            }
+         }
+
+         // insert alpha=255 values if desired
+         if (img_n != out_n)
+            stbi__create_png_alpha_expand8(dest, dest, x, img_n);
+      } else if (depth == 8) {
+         if (img_n == out_n)
+            memcpy(dest, cur, x*img_n);
+         else
+            stbi__create_png_alpha_expand8(dest, cur, x, img_n);
+      } else if (depth == 16) {
+         // convert the image data from big-endian to platform-native
+         stbi__uint16 *dest16 = (stbi__uint16*)dest;
+         stbi__uint32 nsmp = x*img_n;
+
+         if (img_n == out_n) {
+            for (i = 0; i < nsmp; ++i, ++dest16, cur += 2)
+               *dest16 = (cur[0] << 8) | cur[1];
+         } else {
+            STBI_ASSERT(img_n+1 == out_n);
+            if (img_n == 1) {
+               for (i = 0; i < x; ++i, dest16 += 2, cur += 2) {
+                  dest16[0] = (cur[0] << 8) | cur[1];
+                  dest16[1] = 0xffff;
+               }
+            } else {
+               STBI_ASSERT(img_n == 3);
+               for (i = 0; i < x; ++i, dest16 += 4, cur += 6) {
+                  dest16[0] = (cur[0] << 8) | cur[1];
+                  dest16[1] = (cur[2] << 8) | cur[3];
+                  dest16[2] = (cur[4] << 8) | cur[5];
+                  dest16[3] = 0xffff;
+               }
+            }
+         }
+      }
+   }
+
+   STBI_FREE(filter_buf);
+   if (!all_ok) return 0;
+
+   return 1;
+}
+
+static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced)
+{
+   int bytes = (depth == 16 ? 2 : 1);
+   int out_bytes = out_n * bytes;
+   stbi_uc *final;
+   int p;
+   if (!interlaced)
+      return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color);
+
+   // de-interlacing
+   final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0);
+   if (!final) return stbi__err("outofmem", "Out of memory");
+   for (p=0; p < 7; ++p) {
+      int xorig[] = { 0,4,0,2,0,1,0 };
+      int yorig[] = { 0,0,4,0,2,0,1 };
+      int xspc[]  = { 8,8,4,4,2,2,1 };
+      int yspc[]  = { 8,8,8,4,4,2,2 };
+      int i,j,x,y;
+      // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1
+      x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p];
+      y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p];
+      if (x && y) {
+         stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y;
+         if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) {
+            STBI_FREE(final);
+            return 0;
+         }
+         for (j=0; j < y; ++j) {
+            for (i=0; i < x; ++i) {
+               int out_y = j*yspc[p]+yorig[p];
+               int out_x = i*xspc[p]+xorig[p];
+               memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes,
+                      a->out + (j*x+i)*out_bytes, out_bytes);
+            }
+         }
+         STBI_FREE(a->out);
+         image_data += img_len;
+         image_data_len -= img_len;
+      }
+   }
+   a->out = final;
+
+   return 1;
+}
+
+static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi_uc *p = z->out;
+
+   // compute color-based transparency, assuming we've
+   // already got 255 as the alpha value in the output
+   STBI_ASSERT(out_n == 2 || out_n == 4);
+
+   if (out_n == 2) {
+      for (i=0; i < pixel_count; ++i) {
+         p[1] = (p[0] == tc[0] ? 0 : 255);
+         p += 2;
+      }
+   } else {
+      for (i=0; i < pixel_count; ++i) {
+         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+            p[3] = 0;
+         p += 4;
+      }
+   }
+   return 1;
+}
+
+static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi__uint16 *p = (stbi__uint16*) z->out;
+
+   // compute color-based transparency, assuming we've
+   // already got 65535 as the alpha value in the output
+   STBI_ASSERT(out_n == 2 || out_n == 4);
+
+   if (out_n == 2) {
+      for (i = 0; i < pixel_count; ++i) {
+         p[1] = (p[0] == tc[0] ? 0 : 65535);
+         p += 2;
+      }
+   } else {
+      for (i = 0; i < pixel_count; ++i) {
+         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+            p[3] = 0;
+         p += 4;
+      }
+   }
+   return 1;
+}
+
+static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n)
+{
+   stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y;
+   stbi_uc *p, *temp_out, *orig = a->out;
+
+   p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0);
+   if (p == NULL) return stbi__err("outofmem", "Out of memory");
+
+   // between here and free(out) below, exitting would leak
+   temp_out = p;
+
+   if (pal_img_n == 3) {
+      for (i=0; i < pixel_count; ++i) {
+         int n = orig[i]*4;
+         p[0] = palette[n  ];
+         p[1] = palette[n+1];
+         p[2] = palette[n+2];
+         p += 3;
+      }
+   } else {
+      for (i=0; i < pixel_count; ++i) {
+         int n = orig[i]*4;
+         p[0] = palette[n  ];
+         p[1] = palette[n+1];
+         p[2] = palette[n+2];
+         p[3] = palette[n+3];
+         p += 4;
+      }
+   }
+   STBI_FREE(a->out);
+   a->out = temp_out;
+
+   STBI_NOTUSED(len);
+
+   return 1;
+}
+
+static int stbi__unpremultiply_on_load_global = 0;
+static int stbi__de_iphone_flag_global = 0;
+
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_global = flag_true_if_should_convert;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__unpremultiply_on_load  stbi__unpremultiply_on_load_global
+#define stbi__de_iphone_flag  stbi__de_iphone_flag_global
+#else
+static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set;
+static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set;
+
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply;
+   stbi__unpremultiply_on_load_set = 1;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_local = flag_true_if_should_convert;
+   stbi__de_iphone_flag_set = 1;
+}
+
+#define stbi__unpremultiply_on_load  (stbi__unpremultiply_on_load_set           \
+                                       ? stbi__unpremultiply_on_load_local      \
+                                       : stbi__unpremultiply_on_load_global)
+#define stbi__de_iphone_flag  (stbi__de_iphone_flag_set                         \
+                                ? stbi__de_iphone_flag_local                    \
+                                : stbi__de_iphone_flag_global)
+#endif // STBI_THREAD_LOCAL
+
+static void stbi__de_iphone(stbi__png *z)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi_uc *p = z->out;
+
+   if (s->img_out_n == 3) {  // convert bgr to rgb
+      for (i=0; i < pixel_count; ++i) {
+         stbi_uc t = p[0];
+         p[0] = p[2];
+         p[2] = t;
+         p += 3;
+      }
+   } else {
+      STBI_ASSERT(s->img_out_n == 4);
+      if (stbi__unpremultiply_on_load) {
+         // convert bgr to rgb and unpremultiply
+         for (i=0; i < pixel_count; ++i) {
+            stbi_uc a = p[3];
+            stbi_uc t = p[0];
+            if (a) {
+               stbi_uc half = a / 2;
+               p[0] = (p[2] * 255 + half) / a;
+               p[1] = (p[1] * 255 + half) / a;
+               p[2] = ( t   * 255 + half) / a;
+            } else {
+               p[0] = p[2];
+               p[2] = t;
+            }
+            p += 4;
+         }
+      } else {
+         // convert bgr to rgb
+         for (i=0; i < pixel_count; ++i) {
+            stbi_uc t = p[0];
+            p[0] = p[2];
+            p[2] = t;
+            p += 4;
+         }
+      }
+   }
+}
+
+#define STBI__PNG_TYPE(a,b,c,d)  (((unsigned) (a) << 24) + ((unsigned) (b) << 16) + ((unsigned) (c) << 8) + (unsigned) (d))
+
+static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp)
+{
+   stbi_uc palette[1024], pal_img_n=0;
+   stbi_uc has_trans=0, tc[3]={0};
+   stbi__uint16 tc16[3];
+   stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0;
+   int first=1,k,interlace=0, color=0, is_iphone=0;
+   stbi__context *s = z->s;
+
+   z->expanded = NULL;
+   z->idata = NULL;
+   z->out = NULL;
+
+   if (!stbi__check_png_header(s)) return 0;
+
+   if (scan == STBI__SCAN_type) return 1;
+
+   for (;;) {
+      stbi__pngchunk c = stbi__get_chunk_header(s);
+      switch (c.type) {
+         case STBI__PNG_TYPE('C','g','B','I'):
+            is_iphone = 1;
+            stbi__skip(s, c.length);
+            break;
+         case STBI__PNG_TYPE('I','H','D','R'): {
+            int comp,filter;
+            if (!first) return stbi__err("multiple IHDR","Corrupt PNG");
+            first = 0;
+            if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG");
+            s->img_x = stbi__get32be(s);
+            s->img_y = stbi__get32be(s);
+            if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+            if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+            z->depth = stbi__get8(s);  if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16)  return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only");
+            color = stbi__get8(s);  if (color > 6)         return stbi__err("bad ctype","Corrupt PNG");
+            if (color == 3 && z->depth == 16)                  return stbi__err("bad ctype","Corrupt PNG");
+            if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG");
+            comp  = stbi__get8(s);  if (comp) return stbi__err("bad comp method","Corrupt PNG");
+            filter= stbi__get8(s);  if (filter) return stbi__err("bad filter method","Corrupt PNG");
+            interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG");
+            if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG");
+            if (!pal_img_n) {
+               s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0);
+               if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode");
+            } else {
+               // if paletted, then pal_n is our final components, and
+               // img_n is # components to decompress/filter.
+               s->img_n = 1;
+               if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG");
+            }
+            // even with SCAN_header, have to scan to see if we have a tRNS
+            break;
+         }
+
+         case STBI__PNG_TYPE('P','L','T','E'):  {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG");
+            pal_len = c.length / 3;
+            if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG");
+            for (i=0; i < pal_len; ++i) {
+               palette[i*4+0] = stbi__get8(s);
+               palette[i*4+1] = stbi__get8(s);
+               palette[i*4+2] = stbi__get8(s);
+               palette[i*4+3] = 255;
+            }
+            break;
+         }
+
+         case STBI__PNG_TYPE('t','R','N','S'): {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG");
+            if (pal_img_n) {
+               if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; }
+               if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG");
+               if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG");
+               pal_img_n = 4;
+               for (i=0; i < c.length; ++i)
+                  palette[i*4+3] = stbi__get8(s);
+            } else {
+               if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG");
+               if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG");
+               has_trans = 1;
+               // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now.
+               if (scan == STBI__SCAN_header) { ++s->img_n; return 1; }
+               if (z->depth == 16) {
+                  for (k = 0; k < s->img_n && k < 3; ++k) // extra loop test to suppress false GCC warning
+                     tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
+               } else {
+                  for (k = 0; k < s->img_n && k < 3; ++k)
+                     tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
+               }
+            }
+            break;
+         }
+
+         case STBI__PNG_TYPE('I','D','A','T'): {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG");
+            if (scan == STBI__SCAN_header) {
+               // header scan definitely stops at first IDAT
+               if (pal_img_n)
+                  s->img_n = pal_img_n;
+               return 1;
+            }
+            if (c.length > (1u << 30)) return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes");
+            if ((int)(ioff + c.length) < (int)ioff) return 0;
+            if (ioff + c.length > idata_limit) {
+               stbi__uint32 idata_limit_old = idata_limit;
+               stbi_uc *p;
+               if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096;
+               while (ioff + c.length > idata_limit)
+                  idata_limit *= 2;
+               STBI_NOTUSED(idata_limit_old);
+               p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory");
+               z->idata = p;
+            }
+            if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG");
+            ioff += c.length;
+            break;
+         }
+
+         case STBI__PNG_TYPE('I','E','N','D'): {
+            stbi__uint32 raw_len, bpl;
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (scan != STBI__SCAN_load) return 1;
+            if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG");
+            // initial guess for decoded data size to avoid unnecessary reallocs
+            bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component
+            raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */;
+            z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone);
+            if (z->expanded == NULL) return 0; // zlib should set error
+            STBI_FREE(z->idata); z->idata = NULL;
+            if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans)
+               s->img_out_n = s->img_n+1;
+            else
+               s->img_out_n = s->img_n;
+            if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0;
+            if (has_trans) {
+               if (z->depth == 16) {
+                  if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0;
+               } else {
+                  if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0;
+               }
+            }
+            if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2)
+               stbi__de_iphone(z);
+            if (pal_img_n) {
+               // pal_img_n == 3 or 4
+               s->img_n = pal_img_n; // record the actual colors we had
+               s->img_out_n = pal_img_n;
+               if (req_comp >= 3) s->img_out_n = req_comp;
+               if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n))
+                  return 0;
+            } else if (has_trans) {
+               // non-paletted image with tRNS -> source image has (constant) alpha
+               ++s->img_n;
+            }
+            STBI_FREE(z->expanded); z->expanded = NULL;
+            // end of PNG chunk, read and skip CRC
+            stbi__get32be(s);
+            return 1;
+         }
+
+         default:
+            // if critical, fail
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if ((c.type & (1 << 29)) == 0) {
+               #ifndef STBI_NO_FAILURE_STRINGS
+               // not threadsafe
+               static char invalid_chunk[] = "XXXX PNG chunk not known";
+               invalid_chunk[0] = STBI__BYTECAST(c.type >> 24);
+               invalid_chunk[1] = STBI__BYTECAST(c.type >> 16);
+               invalid_chunk[2] = STBI__BYTECAST(c.type >>  8);
+               invalid_chunk[3] = STBI__BYTECAST(c.type >>  0);
+               #endif
+               return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type");
+            }
+            stbi__skip(s, c.length);
+            break;
+      }
+      // end of PNG chunk, read and skip CRC
+      stbi__get32be(s);
+   }
+}
+
+static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri)
+{
+   void *result=NULL;
+   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
+   if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) {
+      if (p->depth <= 8)
+         ri->bits_per_channel = 8;
+      else if (p->depth == 16)
+         ri->bits_per_channel = 16;
+      else
+         return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth");
+      result = p->out;
+      p->out = NULL;
+      if (req_comp && req_comp != p->s->img_out_n) {
+         if (ri->bits_per_channel == 8)
+            result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+         else
+            result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+         p->s->img_out_n = req_comp;
+         if (result == NULL) return result;
+      }
+      *x = p->s->img_x;
+      *y = p->s->img_y;
+      if (n) *n = p->s->img_n;
+   }
+   STBI_FREE(p->out);      p->out      = NULL;
+   STBI_FREE(p->expanded); p->expanded = NULL;
+   STBI_FREE(p->idata);    p->idata    = NULL;
+
+   return result;
+}
+
+static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi__png p;
+   p.s = s;
+   return stbi__do_png(&p, x,y,comp,req_comp, ri);
+}
+
+static int stbi__png_test(stbi__context *s)
+{
+   int r;
+   r = stbi__check_png_header(s);
+   stbi__rewind(s);
+   return r;
+}
+
+static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp)
+{
+   if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) {
+      stbi__rewind( p->s );
+      return 0;
+   }
+   if (x) *x = p->s->img_x;
+   if (y) *y = p->s->img_y;
+   if (comp) *comp = p->s->img_n;
+   return 1;
+}
+
+static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   stbi__png p;
+   p.s = s;
+   return stbi__png_info_raw(&p, x, y, comp);
+}
+
+static int stbi__png_is16(stbi__context *s)
+{
+   stbi__png p;
+   p.s = s;
+   if (!stbi__png_info_raw(&p, NULL, NULL, NULL))
+	   return 0;
+   if (p.depth != 16) {
+      stbi__rewind(p.s);
+      return 0;
+   }
+   return 1;
+}
+#endif
+
+// Microsoft/Windows BMP image
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_test_raw(stbi__context *s)
+{
+   int r;
+   int sz;
+   if (stbi__get8(s) != 'B') return 0;
+   if (stbi__get8(s) != 'M') return 0;
+   stbi__get32le(s); // discard filesize
+   stbi__get16le(s); // discard reserved
+   stbi__get16le(s); // discard reserved
+   stbi__get32le(s); // discard data offset
+   sz = stbi__get32le(s);
+   r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124);
+   return r;
+}
+
+static int stbi__bmp_test(stbi__context *s)
+{
+   int r = stbi__bmp_test_raw(s);
+   stbi__rewind(s);
+   return r;
+}
+
+
+// returns 0..31 for the highest set bit
+static int stbi__high_bit(unsigned int z)
+{
+   int n=0;
+   if (z == 0) return -1;
+   if (z >= 0x10000) { n += 16; z >>= 16; }
+   if (z >= 0x00100) { n +=  8; z >>=  8; }
+   if (z >= 0x00010) { n +=  4; z >>=  4; }
+   if (z >= 0x00004) { n +=  2; z >>=  2; }
+   if (z >= 0x00002) { n +=  1;/* >>=  1;*/ }
+   return n;
+}
+
+static int stbi__bitcount(unsigned int a)
+{
+   a = (a & 0x55555555) + ((a >>  1) & 0x55555555); // max 2
+   a = (a & 0x33333333) + ((a >>  2) & 0x33333333); // max 4
+   a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits
+   a = (a + (a >> 8)); // max 16 per 8 bits
+   a = (a + (a >> 16)); // max 32 per 8 bits
+   return a & 0xff;
+}
+
+// extract an arbitrarily-aligned N-bit value (N=bits)
+// from v, and then make it 8-bits long and fractionally
+// extend it to full full range.
+static int stbi__shiftsigned(unsigned int v, int shift, int bits)
+{
+   static unsigned int mul_table[9] = {
+      0,
+      0xff/*0b11111111*/, 0x55/*0b01010101*/, 0x49/*0b01001001*/, 0x11/*0b00010001*/,
+      0x21/*0b00100001*/, 0x41/*0b01000001*/, 0x81/*0b10000001*/, 0x01/*0b00000001*/,
+   };
+   static unsigned int shift_table[9] = {
+      0, 0,0,1,0,2,4,6,0,
+   };
+   if (shift < 0)
+      v <<= -shift;
+   else
+      v >>= shift;
+   STBI_ASSERT(v < 256);
+   v >>= (8-bits);
+   STBI_ASSERT(bits >= 0 && bits <= 8);
+   return (int) ((unsigned) v * mul_table[bits]) >> shift_table[bits];
+}
+
+typedef struct
+{
+   int bpp, offset, hsz;
+   unsigned int mr,mg,mb,ma, all_a;
+   int extra_read;
+} stbi__bmp_data;
+
+static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress)
+{
+   // BI_BITFIELDS specifies masks explicitly, don't override
+   if (compress == 3)
+      return 1;
+
+   if (compress == 0) {
+      if (info->bpp == 16) {
+         info->mr = 31u << 10;
+         info->mg = 31u <<  5;
+         info->mb = 31u <<  0;
+      } else if (info->bpp == 32) {
+         info->mr = 0xffu << 16;
+         info->mg = 0xffu <<  8;
+         info->mb = 0xffu <<  0;
+         info->ma = 0xffu << 24;
+         info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
+      } else {
+         // otherwise, use defaults, which is all-0
+         info->mr = info->mg = info->mb = info->ma = 0;
+      }
+      return 1;
+   }
+   return 0; // error
+}
+
+static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
+{
+   int hsz;
+   if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP");
+   stbi__get32le(s); // discard filesize
+   stbi__get16le(s); // discard reserved
+   stbi__get16le(s); // discard reserved
+   info->offset = stbi__get32le(s);
+   info->hsz = hsz = stbi__get32le(s);
+   info->mr = info->mg = info->mb = info->ma = 0;
+   info->extra_read = 14;
+
+   if (info->offset < 0) return stbi__errpuc("bad BMP", "bad BMP");
+
+   if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown");
+   if (hsz == 12) {
+      s->img_x = stbi__get16le(s);
+      s->img_y = stbi__get16le(s);
+   } else {
+      s->img_x = stbi__get32le(s);
+      s->img_y = stbi__get32le(s);
+   }
+   if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP");
+   info->bpp = stbi__get16le(s);
+   if (hsz != 12) {
+      int compress = stbi__get32le(s);
+      if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE");
+      if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes
+      if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel
+      stbi__get32le(s); // discard sizeof
+      stbi__get32le(s); // discard hres
+      stbi__get32le(s); // discard vres
+      stbi__get32le(s); // discard colorsused
+      stbi__get32le(s); // discard max important
+      if (hsz == 40 || hsz == 56) {
+         if (hsz == 56) {
+            stbi__get32le(s);
+            stbi__get32le(s);
+            stbi__get32le(s);
+            stbi__get32le(s);
+         }
+         if (info->bpp == 16 || info->bpp == 32) {
+            if (compress == 0) {
+               stbi__bmp_set_mask_defaults(info, compress);
+            } else if (compress == 3) {
+               info->mr = stbi__get32le(s);
+               info->mg = stbi__get32le(s);
+               info->mb = stbi__get32le(s);
+               info->extra_read += 12;
+               // not documented, but generated by photoshop and handled by mspaint
+               if (info->mr == info->mg && info->mg == info->mb) {
+                  // ?!?!?
+                  return stbi__errpuc("bad BMP", "bad BMP");
+               }
+            } else
+               return stbi__errpuc("bad BMP", "bad BMP");
+         }
+      } else {
+         // V4/V5 header
+         int i;
+         if (hsz != 108 && hsz != 124)
+            return stbi__errpuc("bad BMP", "bad BMP");
+         info->mr = stbi__get32le(s);
+         info->mg = stbi__get32le(s);
+         info->mb = stbi__get32le(s);
+         info->ma = stbi__get32le(s);
+         if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs
+            stbi__bmp_set_mask_defaults(info, compress);
+         stbi__get32le(s); // discard color space
+         for (i=0; i < 12; ++i)
+            stbi__get32le(s); // discard color space parameters
+         if (hsz == 124) {
+            stbi__get32le(s); // discard rendering intent
+            stbi__get32le(s); // discard offset of profile data
+            stbi__get32le(s); // discard size of profile data
+            stbi__get32le(s); // discard reserved
+         }
+      }
+   }
+   return (void *) 1;
+}
+
+
+static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *out;
+   unsigned int mr=0,mg=0,mb=0,ma=0, all_a;
+   stbi_uc pal[256][4];
+   int psize=0,i,j,width;
+   int flip_vertically, pad, target;
+   stbi__bmp_data info;
+   STBI_NOTUSED(ri);
+
+   info.all_a = 255;
+   if (stbi__bmp_parse_header(s, &info) == NULL)
+      return NULL; // error code already set
+
+   flip_vertically = ((int) s->img_y) > 0;
+   s->img_y = abs((int) s->img_y);
+
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   mr = info.mr;
+   mg = info.mg;
+   mb = info.mb;
+   ma = info.ma;
+   all_a = info.all_a;
+
+   if (info.hsz == 12) {
+      if (info.bpp < 24)
+         psize = (info.offset - info.extra_read - 24) / 3;
+   } else {
+      if (info.bpp < 16)
+         psize = (info.offset - info.extra_read - info.hsz) >> 2;
+   }
+   if (psize == 0) {
+      // accept some number of extra bytes after the header, but if the offset points either to before
+      // the header ends or implies a large amount of extra data, reject the file as malformed
+      int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original);
+      int header_limit = 1024; // max we actually read is below 256 bytes currently.
+      int extra_data_limit = 256*4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size.
+      if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) {
+         return stbi__errpuc("bad header", "Corrupt BMP");
+      }
+      // we established that bytes_read_so_far is positive and sensible.
+      // the first half of this test rejects offsets that are either too small positives, or
+      // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn
+      // ensures the number computed in the second half of the test can't overflow.
+      if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) {
+         return stbi__errpuc("bad offset", "Corrupt BMP");
+      } else {
+         stbi__skip(s, info.offset - bytes_read_so_far);
+      }
+   }
+
+   if (info.bpp == 24 && ma == 0xff000000)
+      s->img_n = 3;
+   else
+      s->img_n = ma ? 4 : 3;
+   if (req_comp && req_comp >= 3) // we can directly decode 3 or 4
+      target = req_comp;
+   else
+      target = s->img_n; // if they want monochrome, we'll post-convert
+
+   // sanity-check size
+   if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0))
+      return stbi__errpuc("too large", "Corrupt BMP");
+
+   out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0);
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   if (info.bpp < 16) {
+      int z=0;
+      if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); }
+      for (i=0; i < psize; ++i) {
+         pal[i][2] = stbi__get8(s);
+         pal[i][1] = stbi__get8(s);
+         pal[i][0] = stbi__get8(s);
+         if (info.hsz != 12) stbi__get8(s);
+         pal[i][3] = 255;
+      }
+      stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4));
+      if (info.bpp == 1) width = (s->img_x + 7) >> 3;
+      else if (info.bpp == 4) width = (s->img_x + 1) >> 1;
+      else if (info.bpp == 8) width = s->img_x;
+      else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); }
+      pad = (-width)&3;
+      if (info.bpp == 1) {
+         for (j=0; j < (int) s->img_y; ++j) {
+            int bit_offset = 7, v = stbi__get8(s);
+            for (i=0; i < (int) s->img_x; ++i) {
+               int color = (v>>bit_offset)&0x1;
+               out[z++] = pal[color][0];
+               out[z++] = pal[color][1];
+               out[z++] = pal[color][2];
+               if (target == 4) out[z++] = 255;
+               if (i+1 == (int) s->img_x) break;
+               if((--bit_offset) < 0) {
+                  bit_offset = 7;
+                  v = stbi__get8(s);
+               }
+            }
+            stbi__skip(s, pad);
+         }
+      } else {
+         for (j=0; j < (int) s->img_y; ++j) {
+            for (i=0; i < (int) s->img_x; i += 2) {
+               int v=stbi__get8(s),v2=0;
+               if (info.bpp == 4) {
+                  v2 = v & 15;
+                  v >>= 4;
+               }
+               out[z++] = pal[v][0];
+               out[z++] = pal[v][1];
+               out[z++] = pal[v][2];
+               if (target == 4) out[z++] = 255;
+               if (i+1 == (int) s->img_x) break;
+               v = (info.bpp == 8) ? stbi__get8(s) : v2;
+               out[z++] = pal[v][0];
+               out[z++] = pal[v][1];
+               out[z++] = pal[v][2];
+               if (target == 4) out[z++] = 255;
+            }
+            stbi__skip(s, pad);
+         }
+      }
+   } else {
+      int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0;
+      int z = 0;
+      int easy=0;
+      stbi__skip(s, info.offset - info.extra_read - info.hsz);
+      if (info.bpp == 24) width = 3 * s->img_x;
+      else if (info.bpp == 16) width = 2*s->img_x;
+      else /* bpp = 32 and pad = 0 */ width=0;
+      pad = (-width) & 3;
+      if (info.bpp == 24) {
+         easy = 1;
+      } else if (info.bpp == 32) {
+         if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000)
+            easy = 2;
+      }
+      if (!easy) {
+         if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
+         // right shift amt to put high bit in position #7
+         rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr);
+         gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg);
+         bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb);
+         ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma);
+         if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
+      }
+      for (j=0; j < (int) s->img_y; ++j) {
+         if (easy) {
+            for (i=0; i < (int) s->img_x; ++i) {
+               unsigned char a;
+               out[z+2] = stbi__get8(s);
+               out[z+1] = stbi__get8(s);
+               out[z+0] = stbi__get8(s);
+               z += 3;
+               a = (easy == 2 ? stbi__get8(s) : 255);
+               all_a |= a;
+               if (target == 4) out[z++] = a;
+            }
+         } else {
+            int bpp = info.bpp;
+            for (i=0; i < (int) s->img_x; ++i) {
+               stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s));
+               unsigned int a;
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount));
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount));
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount));
+               a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255);
+               all_a |= a;
+               if (target == 4) out[z++] = STBI__BYTECAST(a);
+            }
+         }
+         stbi__skip(s, pad);
+      }
+   }
+
+   // if alpha channel is all 0s, replace with all 255s
+   if (target == 4 && all_a == 0)
+      for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4)
+         out[i] = 255;
+
+   if (flip_vertically) {
+      stbi_uc t;
+      for (j=0; j < (int) s->img_y>>1; ++j) {
+         stbi_uc *p1 = out +      j     *s->img_x*target;
+         stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target;
+         for (i=0; i < (int) s->img_x*target; ++i) {
+            t = p1[i]; p1[i] = p2[i]; p2[i] = t;
+         }
+      }
+   }
+
+   if (req_comp && req_comp != target) {
+      out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y);
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+
+   *x = s->img_x;
+   *y = s->img_y;
+   if (comp) *comp = s->img_n;
+   return out;
+}
+#endif
+
+// Targa Truevision - TGA
+// by Jonathan Dummer
+#ifndef STBI_NO_TGA
+// returns STBI_rgb or whatever, 0 on error
+static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16)
+{
+   // only RGB or RGBA (incl. 16bit) or grey allowed
+   if (is_rgb16) *is_rgb16 = 0;
+   switch(bits_per_pixel) {
+      case 8:  return STBI_grey;
+      case 16: if(is_grey) return STBI_grey_alpha;
+               // fallthrough
+      case 15: if(is_rgb16) *is_rgb16 = 1;
+               return STBI_rgb;
+      case 24: // fallthrough
+      case 32: return bits_per_pixel/8;
+      default: return 0;
+   }
+}
+
+static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp)
+{
+    int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp;
+    int sz, tga_colormap_type;
+    stbi__get8(s);                   // discard Offset
+    tga_colormap_type = stbi__get8(s); // colormap type
+    if( tga_colormap_type > 1 ) {
+        stbi__rewind(s);
+        return 0;      // only RGB or indexed allowed
+    }
+    tga_image_type = stbi__get8(s); // image type
+    if ( tga_colormap_type == 1 ) { // colormapped (paletted) image
+        if (tga_image_type != 1 && tga_image_type != 9) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s,4);       // skip index of first colormap entry and number of entries
+        sz = stbi__get8(s);    //   check bits per palette color entry
+        if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s,4);       // skip image x and y origin
+        tga_colormap_bpp = sz;
+    } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE
+        if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) {
+            stbi__rewind(s);
+            return 0; // only RGB or grey allowed, +/- RLE
+        }
+        stbi__skip(s,9); // skip colormap specification and image x/y origin
+        tga_colormap_bpp = 0;
+    }
+    tga_w = stbi__get16le(s);
+    if( tga_w < 1 ) {
+        stbi__rewind(s);
+        return 0;   // test width
+    }
+    tga_h = stbi__get16le(s);
+    if( tga_h < 1 ) {
+        stbi__rewind(s);
+        return 0;   // test height
+    }
+    tga_bits_per_pixel = stbi__get8(s); // bits per pixel
+    stbi__get8(s); // ignore alpha bits
+    if (tga_colormap_bpp != 0) {
+        if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) {
+            // when using a colormap, tga_bits_per_pixel is the size of the indexes
+            // I don't think anything but 8 or 16bit indexes makes sense
+            stbi__rewind(s);
+            return 0;
+        }
+        tga_comp = stbi__tga_get_comp(tga_colormap_bpp, 0, NULL);
+    } else {
+        tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL);
+    }
+    if(!tga_comp) {
+      stbi__rewind(s);
+      return 0;
+    }
+    if (x) *x = tga_w;
+    if (y) *y = tga_h;
+    if (comp) *comp = tga_comp;
+    return 1;                   // seems to have passed everything
+}
+
+static int stbi__tga_test(stbi__context *s)
+{
+   int res = 0;
+   int sz, tga_color_type;
+   stbi__get8(s);      //   discard Offset
+   tga_color_type = stbi__get8(s);   //   color type
+   if ( tga_color_type > 1 ) goto errorEnd;   //   only RGB or indexed allowed
+   sz = stbi__get8(s);   //   image type
+   if ( tga_color_type == 1 ) { // colormapped (paletted) image
+      if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9
+      stbi__skip(s,4);       // skip index of first colormap entry and number of entries
+      sz = stbi__get8(s);    //   check bits per palette color entry
+      if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
+      stbi__skip(s,4);       // skip image x and y origin
+   } else { // "normal" image w/o colormap
+      if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE
+      stbi__skip(s,9); // skip colormap specification and image x/y origin
+   }
+   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test width
+   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test height
+   sz = stbi__get8(s);   //   bits per pixel
+   if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index
+   if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
+
+   res = 1; // if we got this far, everything's good and we can return 1 instead of 0
+
+errorEnd:
+   stbi__rewind(s);
+   return res;
+}
+
+// read 16bit value and convert to 24bit RGB
+static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out)
+{
+   stbi__uint16 px = (stbi__uint16)stbi__get16le(s);
+   stbi__uint16 fiveBitMask = 31;
+   // we have 3 channels with 5bits each
+   int r = (px >> 10) & fiveBitMask;
+   int g = (px >> 5) & fiveBitMask;
+   int b = px & fiveBitMask;
+   // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later
+   out[0] = (stbi_uc)((r * 255)/31);
+   out[1] = (stbi_uc)((g * 255)/31);
+   out[2] = (stbi_uc)((b * 255)/31);
+
+   // some people claim that the most significant bit might be used for alpha
+   // (possibly if an alpha-bit is set in the "image descriptor byte")
+   // but that only made 16bit test images completely translucent..
+   // so let's treat all 15 and 16bit TGAs as RGB with no alpha.
+}
+
+static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   //   read in the TGA header stuff
+   int tga_offset = stbi__get8(s);
+   int tga_indexed = stbi__get8(s);
+   int tga_image_type = stbi__get8(s);
+   int tga_is_RLE = 0;
+   int tga_palette_start = stbi__get16le(s);
+   int tga_palette_len = stbi__get16le(s);
+   int tga_palette_bits = stbi__get8(s);
+   int tga_x_origin = stbi__get16le(s);
+   int tga_y_origin = stbi__get16le(s);
+   int tga_width = stbi__get16le(s);
+   int tga_height = stbi__get16le(s);
+   int tga_bits_per_pixel = stbi__get8(s);
+   int tga_comp, tga_rgb16=0;
+   int tga_inverted = stbi__get8(s);
+   // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?)
+   //   image data
+   unsigned char *tga_data;
+   unsigned char *tga_palette = NULL;
+   int i, j;
+   unsigned char raw_data[4] = {0};
+   int RLE_count = 0;
+   int RLE_repeating = 0;
+   int read_next_pixel = 1;
+   STBI_NOTUSED(ri);
+   STBI_NOTUSED(tga_x_origin); // @TODO
+   STBI_NOTUSED(tga_y_origin); // @TODO
+
+   if (tga_height > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (tga_width > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   //   do a tiny bit of precessing
+   if ( tga_image_type >= 8 )
+   {
+      tga_image_type -= 8;
+      tga_is_RLE = 1;
+   }
+   tga_inverted = 1 - ((tga_inverted >> 5) & 1);
+
+   //   If I'm paletted, then I'll use the number of bits from the palette
+   if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16);
+   else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16);
+
+   if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency
+      return stbi__errpuc("bad format", "Can't find out TGA pixelformat");
+
+   //   tga info
+   *x = tga_width;
+   *y = tga_height;
+   if (comp) *comp = tga_comp;
+
+   if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0))
+      return stbi__errpuc("too large", "Corrupt TGA");
+
+   tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0);
+   if (!tga_data) return stbi__errpuc("outofmem", "Out of memory");
+
+   // skip to the data's starting position (offset usually = 0)
+   stbi__skip(s, tga_offset );
+
+   if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) {
+      for (i=0; i < tga_height; ++i) {
+         int row = tga_inverted ? tga_height -i - 1 : i;
+         stbi_uc *tga_row = tga_data + row*tga_width*tga_comp;
+         stbi__getn(s, tga_row, tga_width * tga_comp);
+      }
+   } else  {
+      //   do I need to load a palette?
+      if ( tga_indexed)
+      {
+         if (tga_palette_len == 0) {  /* you have to have at least one entry! */
+            STBI_FREE(tga_data);
+            return stbi__errpuc("bad palette", "Corrupt TGA");
+         }
+
+         //   any data to skip? (offset usually = 0)
+         stbi__skip(s, tga_palette_start );
+         //   load the palette
+         tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0);
+         if (!tga_palette) {
+            STBI_FREE(tga_data);
+            return stbi__errpuc("outofmem", "Out of memory");
+         }
+         if (tga_rgb16) {
+            stbi_uc *pal_entry = tga_palette;
+            STBI_ASSERT(tga_comp == STBI_rgb);
+            for (i=0; i < tga_palette_len; ++i) {
+               stbi__tga_read_rgb16(s, pal_entry);
+               pal_entry += tga_comp;
+            }
+         } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) {
+               STBI_FREE(tga_data);
+               STBI_FREE(tga_palette);
+               return stbi__errpuc("bad palette", "Corrupt TGA");
+         }
+      }
+      //   load the data
+      for (i=0; i < tga_width * tga_height; ++i)
+      {
+         //   if I'm in RLE mode, do I need to get a RLE stbi__pngchunk?
+         if ( tga_is_RLE )
+         {
+            if ( RLE_count == 0 )
+            {
+               //   yep, get the next byte as a RLE command
+               int RLE_cmd = stbi__get8(s);
+               RLE_count = 1 + (RLE_cmd & 127);
+               RLE_repeating = RLE_cmd >> 7;
+               read_next_pixel = 1;
+            } else if ( !RLE_repeating )
+            {
+               read_next_pixel = 1;
+            }
+         } else
+         {
+            read_next_pixel = 1;
+         }
+         //   OK, if I need to read a pixel, do it now
+         if ( read_next_pixel )
+         {
+            //   load however much data we did have
+            if ( tga_indexed )
+            {
+               // read in index, then perform the lookup
+               int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s);
+               if ( pal_idx >= tga_palette_len ) {
+                  // invalid index
+                  pal_idx = 0;
+               }
+               pal_idx *= tga_comp;
+               for (j = 0; j < tga_comp; ++j) {
+                  raw_data[j] = tga_palette[pal_idx+j];
+               }
+            } else if(tga_rgb16) {
+               STBI_ASSERT(tga_comp == STBI_rgb);
+               stbi__tga_read_rgb16(s, raw_data);
+            } else {
+               //   read in the data raw
+               for (j = 0; j < tga_comp; ++j) {
+                  raw_data[j] = stbi__get8(s);
+               }
+            }
+            //   clear the reading flag for the next pixel
+            read_next_pixel = 0;
+         } // end of reading a pixel
+
+         // copy data
+         for (j = 0; j < tga_comp; ++j)
+           tga_data[i*tga_comp+j] = raw_data[j];
+
+         //   in case we're in RLE mode, keep counting down
+         --RLE_count;
+      }
+      //   do I need to invert the image?
+      if ( tga_inverted )
+      {
+         for (j = 0; j*2 < tga_height; ++j)
+         {
+            int index1 = j * tga_width * tga_comp;
+            int index2 = (tga_height - 1 - j) * tga_width * tga_comp;
+            for (i = tga_width * tga_comp; i > 0; --i)
+            {
+               unsigned char temp = tga_data[index1];
+               tga_data[index1] = tga_data[index2];
+               tga_data[index2] = temp;
+               ++index1;
+               ++index2;
+            }
+         }
+      }
+      //   clear my palette, if I had one
+      if ( tga_palette != NULL )
+      {
+         STBI_FREE( tga_palette );
+      }
+   }
+
+   // swap RGB - if the source data was RGB16, it already is in the right order
+   if (tga_comp >= 3 && !tga_rgb16)
+   {
+      unsigned char* tga_pixel = tga_data;
+      for (i=0; i < tga_width * tga_height; ++i)
+      {
+         unsigned char temp = tga_pixel[0];
+         tga_pixel[0] = tga_pixel[2];
+         tga_pixel[2] = temp;
+         tga_pixel += tga_comp;
+      }
+   }
+
+   // convert to target component count
+   if (req_comp && req_comp != tga_comp)
+      tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height);
+
+   //   the things I do to get rid of an error message, and yet keep
+   //   Microsoft's C compilers happy... [8^(
+   tga_palette_start = tga_palette_len = tga_palette_bits =
+         tga_x_origin = tga_y_origin = 0;
+   STBI_NOTUSED(tga_palette_start);
+   //   OK, done
+   return tga_data;
+}
+#endif
+
+// *************************************************************************************************
+// Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_test(stbi__context *s)
+{
+   int r = (stbi__get32be(s) == 0x38425053);
+   stbi__rewind(s);
+   return r;
+}
+
+static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount)
+{
+   int count, nleft, len;
+
+   count = 0;
+   while ((nleft = pixelCount - count) > 0) {
+      len = stbi__get8(s);
+      if (len == 128) {
+         // No-op.
+      } else if (len < 128) {
+         // Copy next len+1 bytes literally.
+         len++;
+         if (len > nleft) return 0; // corrupt data
+         count += len;
+         while (len) {
+            *p = stbi__get8(s);
+            p += 4;
+            len--;
+         }
+      } else if (len > 128) {
+         stbi_uc   val;
+         // Next -len+1 bytes in the dest are replicated from next source byte.
+         // (Interpret len as a negative 8-bit int.)
+         len = 257 - len;
+         if (len > nleft) return 0; // corrupt data
+         val = stbi__get8(s);
+         count += len;
+         while (len) {
+            *p = val;
+            p += 4;
+            len--;
+         }
+      }
+   }
+
+   return 1;
+}
+
+static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
+{
+   int pixelCount;
+   int channelCount, compression;
+   int channel, i;
+   int bitdepth;
+   int w,h;
+   stbi_uc *out;
+   STBI_NOTUSED(ri);
+
+   // Check identifier
+   if (stbi__get32be(s) != 0x38425053)   // "8BPS"
+      return stbi__errpuc("not PSD", "Corrupt PSD image");
+
+   // Check file type version.
+   if (stbi__get16be(s) != 1)
+      return stbi__errpuc("wrong version", "Unsupported version of PSD image");
+
+   // Skip 6 reserved bytes.
+   stbi__skip(s, 6 );
+
+   // Read the number of channels (R, G, B, A, etc).
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16)
+      return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image");
+
+   // Read the rows and columns of the image.
+   h = stbi__get32be(s);
+   w = stbi__get32be(s);
+
+   if (h > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (w > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   // Make sure the depth is 8 bits.
+   bitdepth = stbi__get16be(s);
+   if (bitdepth != 8 && bitdepth != 16)
+      return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit");
+
+   // Make sure the color mode is RGB.
+   // Valid options are:
+   //   0: Bitmap
+   //   1: Grayscale
+   //   2: Indexed color
+   //   3: RGB color
+   //   4: CMYK color
+   //   7: Multichannel
+   //   8: Duotone
+   //   9: Lab color
+   if (stbi__get16be(s) != 3)
+      return stbi__errpuc("wrong color format", "PSD is not in RGB color format");
+
+   // Skip the Mode Data.  (It's the palette for indexed color; other info for other modes.)
+   stbi__skip(s,stbi__get32be(s) );
+
+   // Skip the image resources.  (resolution, pen tool paths, etc)
+   stbi__skip(s, stbi__get32be(s) );
+
+   // Skip the reserved data.
+   stbi__skip(s, stbi__get32be(s) );
+
+   // Find out if the data is compressed.
+   // Known values:
+   //   0: no compression
+   //   1: RLE compressed
+   compression = stbi__get16be(s);
+   if (compression > 1)
+      return stbi__errpuc("bad compression", "PSD has an unknown compression format");
+
+   // Check size
+   if (!stbi__mad3sizes_valid(4, w, h, 0))
+      return stbi__errpuc("too large", "Corrupt PSD");
+
+   // Create the destination image.
+
+   if (!compression && bitdepth == 16 && bpc == 16) {
+      out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0);
+      ri->bits_per_channel = 16;
+   } else
+      out = (stbi_uc *) stbi__malloc(4 * w*h);
+
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   pixelCount = w*h;
+
+   // Initialize the data to zero.
+   //memset( out, 0, pixelCount * 4 );
+
+   // Finally, the image data.
+   if (compression) {
+      // RLE as used by .PSD and .TIFF
+      // Loop until you get the number of unpacked bytes you are expecting:
+      //     Read the next source byte into n.
+      //     If n is between 0 and 127 inclusive, copy the next n+1 bytes literally.
+      //     Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times.
+      //     Else if n is 128, noop.
+      // Endloop
+
+      // The RLE-compressed data is preceded by a 2-byte data count for each row in the data,
+      // which we're going to just skip.
+      stbi__skip(s, h * channelCount * 2 );
+
+      // Read the RLE data by channel.
+      for (channel = 0; channel < 4; channel++) {
+         stbi_uc *p;
+
+         p = out+channel;
+         if (channel >= channelCount) {
+            // Fill this channel with default data.
+            for (i = 0; i < pixelCount; i++, p += 4)
+               *p = (channel == 3 ? 255 : 0);
+         } else {
+            // Read the RLE data.
+            if (!stbi__psd_decode_rle(s, p, pixelCount)) {
+               STBI_FREE(out);
+               return stbi__errpuc("corrupt", "bad RLE data");
+            }
+         }
+      }
+
+   } else {
+      // We're at the raw image data.  It's each channel in order (Red, Green, Blue, Alpha, ...)
+      // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image.
+
+      // Read the data by channel.
+      for (channel = 0; channel < 4; channel++) {
+         if (channel >= channelCount) {
+            // Fill this channel with default data.
+            if (bitdepth == 16 && bpc == 16) {
+               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
+               stbi__uint16 val = channel == 3 ? 65535 : 0;
+               for (i = 0; i < pixelCount; i++, q += 4)
+                  *q = val;
+            } else {
+               stbi_uc *p = out+channel;
+               stbi_uc val = channel == 3 ? 255 : 0;
+               for (i = 0; i < pixelCount; i++, p += 4)
+                  *p = val;
+            }
+         } else {
+            if (ri->bits_per_channel == 16) {    // output bpc
+               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
+               for (i = 0; i < pixelCount; i++, q += 4)
+                  *q = (stbi__uint16) stbi__get16be(s);
+            } else {
+               stbi_uc *p = out+channel;
+               if (bitdepth == 16) {  // input bpc
+                  for (i = 0; i < pixelCount; i++, p += 4)
+                     *p = (stbi_uc) (stbi__get16be(s) >> 8);
+               } else {
+                  for (i = 0; i < pixelCount; i++, p += 4)
+                     *p = stbi__get8(s);
+               }
+            }
+         }
+      }
+   }
+
+   // remove weird white matte from PSD
+   if (channelCount >= 4) {
+      if (ri->bits_per_channel == 16) {
+         for (i=0; i < w*h; ++i) {
+            stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i;
+            if (pixel[3] != 0 && pixel[3] != 65535) {
+               float a = pixel[3] / 65535.0f;
+               float ra = 1.0f / a;
+               float inv_a = 65535.0f * (1 - ra);
+               pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a);
+               pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a);
+               pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a);
+            }
+         }
+      } else {
+         for (i=0; i < w*h; ++i) {
+            unsigned char *pixel = out + 4*i;
+            if (pixel[3] != 0 && pixel[3] != 255) {
+               float a = pixel[3] / 255.0f;
+               float ra = 1.0f / a;
+               float inv_a = 255.0f * (1 - ra);
+               pixel[0] = (unsigned char) (pixel[0]*ra + inv_a);
+               pixel[1] = (unsigned char) (pixel[1]*ra + inv_a);
+               pixel[2] = (unsigned char) (pixel[2]*ra + inv_a);
+            }
+         }
+      }
+   }
+
+   // convert to desired output format
+   if (req_comp && req_comp != 4) {
+      if (ri->bits_per_channel == 16)
+         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h);
+      else
+         out = stbi__convert_format(out, 4, req_comp, w, h);
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+
+   if (comp) *comp = 4;
+   *y = h;
+   *x = w;
+
+   return out;
+}
+#endif
+
+// *************************************************************************************************
+// Softimage PIC loader
+// by Tom Seddon
+//
+// See http://softimage.wiki.softimage.com/index.php/INFO:_PIC_file_format
+// See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_is4(stbi__context *s,const char *str)
+{
+   int i;
+   for (i=0; i<4; ++i)
+      if (stbi__get8(s) != (stbi_uc)str[i])
+         return 0;
+
+   return 1;
+}
+
+static int stbi__pic_test_core(stbi__context *s)
+{
+   int i;
+
+   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34"))
+      return 0;
+
+   for(i=0;i<84;++i)
+      stbi__get8(s);
+
+   if (!stbi__pic_is4(s,"PICT"))
+      return 0;
+
+   return 1;
+}
+
+typedef struct
+{
+   stbi_uc size,type,channel;
+} stbi__pic_packet;
+
+static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest)
+{
+   int mask=0x80, i;
+
+   for (i=0; i<4; ++i, mask>>=1) {
+      if (channel & mask) {
+         if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short");
+         dest[i]=stbi__get8(s);
+      }
+   }
+
+   return dest;
+}
+
+static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src)
+{
+   int mask=0x80,i;
+
+   for (i=0;i<4; ++i, mask>>=1)
+      if (channel&mask)
+         dest[i]=src[i];
+}
+
+static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result)
+{
+   int act_comp=0,num_packets=0,y,chained;
+   stbi__pic_packet packets[10];
+
+   // this will (should...) cater for even some bizarre stuff like having data
+    // for the same channel in multiple packets.
+   do {
+      stbi__pic_packet *packet;
+
+      if (num_packets==sizeof(packets)/sizeof(packets[0]))
+         return stbi__errpuc("bad format","too many packets");
+
+      packet = &packets[num_packets++];
+
+      chained = stbi__get8(s);
+      packet->size    = stbi__get8(s);
+      packet->type    = stbi__get8(s);
+      packet->channel = stbi__get8(s);
+
+      act_comp |= packet->channel;
+
+      if (stbi__at_eof(s))          return stbi__errpuc("bad file","file too short (reading packets)");
+      if (packet->size != 8)  return stbi__errpuc("bad format","packet isn't 8bpp");
+   } while (chained);
+
+   *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel?
+
+   for(y=0; y<height; ++y) {
+      int packet_idx;
+
+      for(packet_idx=0; packet_idx < num_packets; ++packet_idx) {
+         stbi__pic_packet *packet = &packets[packet_idx];
+         stbi_uc *dest = result+y*width*4;
+
+         switch (packet->type) {
+            default:
+               return stbi__errpuc("bad format","packet has bad compression type");
+
+            case 0: {//uncompressed
+               int x;
+
+               for(x=0;x<width;++x, dest+=4)
+                  if (!stbi__readval(s,packet->channel,dest))
+                     return 0;
+               break;
+            }
+
+            case 1://Pure RLE
+               {
+                  int left=width, i;
+
+                  while (left>0) {
+                     stbi_uc count,value[4];
+
+                     count=stbi__get8(s);
+                     if (stbi__at_eof(s))   return stbi__errpuc("bad file","file too short (pure read count)");
+
+                     if (count > left)
+                        count = (stbi_uc) left;
+
+                     if (!stbi__readval(s,packet->channel,value))  return 0;
+
+                     for(i=0; i<count; ++i,dest+=4)
+                        stbi__copyval(packet->channel,dest,value);
+                     left -= count;
+                  }
+               }
+               break;
+
+            case 2: {//Mixed RLE
+               int left=width;
+               while (left>0) {
+                  int count = stbi__get8(s), i;
+                  if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (mixed read count)");
+
+                  if (count >= 128) { // Repeated
+                     stbi_uc value[4];
+
+                     if (count==128)
+                        count = stbi__get16be(s);
+                     else
+                        count -= 127;
+                     if (count > left)
+                        return stbi__errpuc("bad file","scanline overrun");
+
+                     if (!stbi__readval(s,packet->channel,value))
+                        return 0;
+
+                     for(i=0;i<count;++i, dest += 4)
+                        stbi__copyval(packet->channel,dest,value);
+                  } else { // Raw
+                     ++count;
+                     if (count>left) return stbi__errpuc("bad file","scanline overrun");
+
+                     for(i=0;i<count;++i, dest+=4)
+                        if (!stbi__readval(s,packet->channel,dest))
+                           return 0;
+                  }
+                  left-=count;
+               }
+               break;
+            }
+         }
+      }
+   }
+
+   return result;
+}
+
+static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *result;
+   int i, x,y, internal_comp;
+   STBI_NOTUSED(ri);
+
+   if (!comp) comp = &internal_comp;
+
+   for (i=0; i<92; ++i)
+      stbi__get8(s);
+
+   x = stbi__get16be(s);
+   y = stbi__get16be(s);
+
+   if (y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (pic header)");
+   if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode");
+
+   stbi__get32be(s); //skip `ratio'
+   stbi__get16be(s); //skip `fields'
+   stbi__get16be(s); //skip `pad'
+
+   // intermediate buffer is RGBA
+   result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0);
+   if (!result) return stbi__errpuc("outofmem", "Out of memory");
+   memset(result, 0xff, x*y*4);
+
+   if (!stbi__pic_load_core(s,x,y,comp, result)) {
+      STBI_FREE(result);
+      result=0;
+   }
+   *px = x;
+   *py = y;
+   if (req_comp == 0) req_comp = *comp;
+   result=stbi__convert_format(result,4,req_comp,x,y);
+
+   return result;
+}
+
+static int stbi__pic_test(stbi__context *s)
+{
+   int r = stbi__pic_test_core(s);
+   stbi__rewind(s);
+   return r;
+}
+#endif
+
+// *************************************************************************************************
+// GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb
+
+#ifndef STBI_NO_GIF
+typedef struct
+{
+   stbi__int16 prefix;
+   stbi_uc first;
+   stbi_uc suffix;
+} stbi__gif_lzw;
+
+typedef struct
+{
+   int w,h;
+   stbi_uc *out;                 // output buffer (always 4 components)
+   stbi_uc *background;          // The current "background" as far as a gif is concerned
+   stbi_uc *history;
+   int flags, bgindex, ratio, transparent, eflags;
+   stbi_uc  pal[256][4];
+   stbi_uc lpal[256][4];
+   stbi__gif_lzw codes[8192];
+   stbi_uc *color_table;
+   int parse, step;
+   int lflags;
+   int start_x, start_y;
+   int max_x, max_y;
+   int cur_x, cur_y;
+   int line_size;
+   int delay;
+} stbi__gif;
+
+static int stbi__gif_test_raw(stbi__context *s)
+{
+   int sz;
+   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0;
+   sz = stbi__get8(s);
+   if (sz != '9' && sz != '7') return 0;
+   if (stbi__get8(s) != 'a') return 0;
+   return 1;
+}
+
+static int stbi__gif_test(stbi__context *s)
+{
+   int r = stbi__gif_test_raw(s);
+   stbi__rewind(s);
+   return r;
+}
+
+static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp)
+{
+   int i;
+   for (i=0; i < num_entries; ++i) {
+      pal[i][2] = stbi__get8(s);
+      pal[i][1] = stbi__get8(s);
+      pal[i][0] = stbi__get8(s);
+      pal[i][3] = transp == i ? 0 : 255;
+   }
+}
+
+static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info)
+{
+   stbi_uc version;
+   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8')
+      return stbi__err("not GIF", "Corrupt GIF");
+
+   version = stbi__get8(s);
+   if (version != '7' && version != '9')    return stbi__err("not GIF", "Corrupt GIF");
+   if (stbi__get8(s) != 'a')                return stbi__err("not GIF", "Corrupt GIF");
+
+   stbi__g_failure_reason = "";
+   g->w = stbi__get16le(s);
+   g->h = stbi__get16le(s);
+   g->flags = stbi__get8(s);
+   g->bgindex = stbi__get8(s);
+   g->ratio = stbi__get8(s);
+   g->transparent = -1;
+
+   if (g->w > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   if (g->h > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+
+   if (comp != 0) *comp = 4;  // can't actually tell whether it's 3 or 4 until we parse the comments
+
+   if (is_info) return 1;
+
+   if (g->flags & 0x80)
+      stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1);
+
+   return 1;
+}
+
+static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp)
+{
+   stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif));
+   if (!g) return stbi__err("outofmem", "Out of memory");
+   if (!stbi__gif_header(s, g, comp, 1)) {
+      STBI_FREE(g);
+      stbi__rewind( s );
+      return 0;
+   }
+   if (x) *x = g->w;
+   if (y) *y = g->h;
+   STBI_FREE(g);
+   return 1;
+}
+
+static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code)
+{
+   stbi_uc *p, *c;
+   int idx;
+
+   // recurse to decode the prefixes, since the linked-list is backwards,
+   // and working backwards through an interleaved image would be nasty
+   if (g->codes[code].prefix >= 0)
+      stbi__out_gif_code(g, g->codes[code].prefix);
+
+   if (g->cur_y >= g->max_y) return;
+
+   idx = g->cur_x + g->cur_y;
+   p = &g->out[idx];
+   g->history[idx / 4] = 1;
+
+   c = &g->color_table[g->codes[code].suffix * 4];
+   if (c[3] > 128) { // don't render transparent pixels;
+      p[0] = c[2];
+      p[1] = c[1];
+      p[2] = c[0];
+      p[3] = c[3];
+   }
+   g->cur_x += 4;
+
+   if (g->cur_x >= g->max_x) {
+      g->cur_x = g->start_x;
+      g->cur_y += g->step;
+
+      while (g->cur_y >= g->max_y && g->parse > 0) {
+         g->step = (1 << g->parse) * g->line_size;
+         g->cur_y = g->start_y + (g->step >> 1);
+         --g->parse;
+      }
+   }
+}
+
+static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g)
+{
+   stbi_uc lzw_cs;
+   stbi__int32 len, init_code;
+   stbi__uint32 first;
+   stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear;
+   stbi__gif_lzw *p;
+
+   lzw_cs = stbi__get8(s);
+   if (lzw_cs > 12) return NULL;
+   clear = 1 << lzw_cs;
+   first = 1;
+   codesize = lzw_cs + 1;
+   codemask = (1 << codesize) - 1;
+   bits = 0;
+   valid_bits = 0;
+   for (init_code = 0; init_code < clear; init_code++) {
+      g->codes[init_code].prefix = -1;
+      g->codes[init_code].first = (stbi_uc) init_code;
+      g->codes[init_code].suffix = (stbi_uc) init_code;
+   }
+
+   // support no starting clear code
+   avail = clear+2;
+   oldcode = -1;
+
+   len = 0;
+   for(;;) {
+      if (valid_bits < codesize) {
+         if (len == 0) {
+            len = stbi__get8(s); // start new block
+            if (len == 0)
+               return g->out;
+         }
+         --len;
+         bits |= (stbi__int32) stbi__get8(s) << valid_bits;
+         valid_bits += 8;
+      } else {
+         stbi__int32 code = bits & codemask;
+         bits >>= codesize;
+         valid_bits -= codesize;
+         // @OPTIMIZE: is there some way we can accelerate the non-clear path?
+         if (code == clear) {  // clear code
+            codesize = lzw_cs + 1;
+            codemask = (1 << codesize) - 1;
+            avail = clear + 2;
+            oldcode = -1;
+            first = 0;
+         } else if (code == clear + 1) { // end of stream code
+            stbi__skip(s, len);
+            while ((len = stbi__get8(s)) > 0)
+               stbi__skip(s,len);
+            return g->out;
+         } else if (code <= avail) {
+            if (first) {
+               return stbi__errpuc("no clear code", "Corrupt GIF");
+            }
+
+            if (oldcode >= 0) {
+               p = &g->codes[avail++];
+               if (avail > 8192) {
+                  return stbi__errpuc("too many codes", "Corrupt GIF");
+               }
+
+               p->prefix = (stbi__int16) oldcode;
+               p->first = g->codes[oldcode].first;
+               p->suffix = (code == avail) ? p->first : g->codes[code].first;
+            } else if (code == avail)
+               return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+
+            stbi__out_gif_code(g, (stbi__uint16) code);
+
+            if ((avail & codemask) == 0 && avail <= 0x0FFF) {
+               codesize++;
+               codemask = (1 << codesize) - 1;
+            }
+
+            oldcode = code;
+         } else {
+            return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+         }
+      }
+   }
+}
+
+// this function is designed to support animated gifs, although stb_image doesn't support it
+// two back is the image from two frames ago, used for a very specific disposal format
+static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp, stbi_uc *two_back)
+{
+   int dispose;
+   int first_frame;
+   int pi;
+   int pcount;
+   STBI_NOTUSED(req_comp);
+
+   // on first frame, any non-written pixels get the background colour (non-transparent)
+   first_frame = 0;
+   if (g->out == 0) {
+      if (!stbi__gif_header(s, g, comp,0)) return 0; // stbi__g_failure_reason set by stbi__gif_header
+      if (!stbi__mad3sizes_valid(4, g->w, g->h, 0))
+         return stbi__errpuc("too large", "GIF image is too large");
+      pcount = g->w * g->h;
+      g->out = (stbi_uc *) stbi__malloc(4 * pcount);
+      g->background = (stbi_uc *) stbi__malloc(4 * pcount);
+      g->history = (stbi_uc *) stbi__malloc(pcount);
+      if (!g->out || !g->background || !g->history)
+         return stbi__errpuc("outofmem", "Out of memory");
+
+      // image is treated as "transparent" at the start - ie, nothing overwrites the current background;
+      // background colour is only used for pixels that are not rendered first frame, after that "background"
+      // color refers to the color that was there the previous frame.
+      memset(g->out, 0x00, 4 * pcount);
+      memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent)
+      memset(g->history, 0x00, pcount);        // pixels that were affected previous frame
+      first_frame = 1;
+   } else {
+      // second frame - how do we dispose of the previous one?
+      dispose = (g->eflags & 0x1C) >> 2;
+      pcount = g->w * g->h;
+
+      if ((dispose == 3) && (two_back == 0)) {
+         dispose = 2; // if I don't have an image to revert back to, default to the old background
+      }
+
+      if (dispose == 3) { // use previous graphic
+         for (pi = 0; pi < pcount; ++pi) {
+            if (g->history[pi]) {
+               memcpy( &g->out[pi * 4], &two_back[pi * 4], 4 );
+            }
+         }
+      } else if (dispose == 2) {
+         // restore what was changed last frame to background before that frame;
+         for (pi = 0; pi < pcount; ++pi) {
+            if (g->history[pi]) {
+               memcpy( &g->out[pi * 4], &g->background[pi * 4], 4 );
+            }
+         }
+      } else {
+         // This is a non-disposal case eithe way, so just
+         // leave the pixels as is, and they will become the new background
+         // 1: do not dispose
+         // 0:  not specified.
+      }
+
+      // background is what out is after the undoing of the previou frame;
+      memcpy( g->background, g->out, 4 * g->w * g->h );
+   }
+
+   // clear my history;
+   memset( g->history, 0x00, g->w * g->h );        // pixels that were affected previous frame
+
+   for (;;) {
+      int tag = stbi__get8(s);
+      switch (tag) {
+         case 0x2C: /* Image Descriptor */
+         {
+            stbi__int32 x, y, w, h;
+            stbi_uc *o;
+
+            x = stbi__get16le(s);
+            y = stbi__get16le(s);
+            w = stbi__get16le(s);
+            h = stbi__get16le(s);
+            if (((x + w) > (g->w)) || ((y + h) > (g->h)))
+               return stbi__errpuc("bad Image Descriptor", "Corrupt GIF");
+
+            g->line_size = g->w * 4;
+            g->start_x = x * 4;
+            g->start_y = y * g->line_size;
+            g->max_x   = g->start_x + w * 4;
+            g->max_y   = g->start_y + h * g->line_size;
+            g->cur_x   = g->start_x;
+            g->cur_y   = g->start_y;
+
+            // if the width of the specified rectangle is 0, that means
+            // we may not see *any* pixels or the image is malformed;
+            // to make sure this is caught, move the current y down to
+            // max_y (which is what out_gif_code checks).
+            if (w == 0)
+               g->cur_y = g->max_y;
+
+            g->lflags = stbi__get8(s);
+
+            if (g->lflags & 0x40) {
+               g->step = 8 * g->line_size; // first interlaced spacing
+               g->parse = 3;
+            } else {
+               g->step = g->line_size;
+               g->parse = 0;
+            }
+
+            if (g->lflags & 0x80) {
+               stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1);
+               g->color_table = (stbi_uc *) g->lpal;
+            } else if (g->flags & 0x80) {
+               g->color_table = (stbi_uc *) g->pal;
+            } else
+               return stbi__errpuc("missing color table", "Corrupt GIF");
+
+            o = stbi__process_gif_raster(s, g);
+            if (!o) return NULL;
+
+            // if this was the first frame,
+            pcount = g->w * g->h;
+            if (first_frame && (g->bgindex > 0)) {
+               // if first frame, any pixel not drawn to gets the background color
+               for (pi = 0; pi < pcount; ++pi) {
+                  if (g->history[pi] == 0) {
+                     g->pal[g->bgindex][3] = 255; // just in case it was made transparent, undo that; It will be reset next frame if need be;
+                     memcpy( &g->out[pi * 4], &g->pal[g->bgindex], 4 );
+                  }
+               }
+            }
+
+            return o;
+         }
+
+         case 0x21: // Comment Extension.
+         {
+            int len;
+            int ext = stbi__get8(s);
+            if (ext == 0xF9) { // Graphic Control Extension.
+               len = stbi__get8(s);
+               if (len == 4) {
+                  g->eflags = stbi__get8(s);
+                  g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths.
+
+                  // unset old transparent
+                  if (g->transparent >= 0) {
+                     g->pal[g->transparent][3] = 255;
+                  }
+                  if (g->eflags & 0x01) {
+                     g->transparent = stbi__get8(s);
+                     if (g->transparent >= 0) {
+                        g->pal[g->transparent][3] = 0;
+                     }
+                  } else {
+                     // don't need transparent
+                     stbi__skip(s, 1);
+                     g->transparent = -1;
+                  }
+               } else {
+                  stbi__skip(s, len);
+                  break;
+               }
+            }
+            while ((len = stbi__get8(s)) != 0) {
+               stbi__skip(s, len);
+            }
+            break;
+         }
+
+         case 0x3B: // gif stream termination code
+            return (stbi_uc *) s; // using '1' causes warning on some compilers
+
+         default:
+            return stbi__errpuc("unknown code", "Corrupt GIF");
+      }
+   }
+}
+
+static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays)
+{
+   STBI_FREE(g->out);
+   STBI_FREE(g->history);
+   STBI_FREE(g->background);
+
+   if (out) STBI_FREE(out);
+   if (delays && *delays) STBI_FREE(*delays);
+   return stbi__errpuc("outofmem", "Out of memory");
+}
+
+static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
+{
+   if (stbi__gif_test(s)) {
+      int layers = 0;
+      stbi_uc *u = 0;
+      stbi_uc *out = 0;
+      stbi_uc *two_back = 0;
+      stbi__gif g;
+      int stride;
+      int out_size = 0;
+      int delays_size = 0;
+
+      STBI_NOTUSED(out_size);
+      STBI_NOTUSED(delays_size);
+
+      memset(&g, 0, sizeof(g));
+      if (delays) {
+         *delays = 0;
+      }
+
+      do {
+         u = stbi__gif_load_next(s, &g, comp, req_comp, two_back);
+         if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
+
+         if (u) {
+            *x = g.w;
+            *y = g.h;
+            ++layers;
+            stride = g.w * g.h * 4;
+
+            if (out) {
+               void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride );
+               if (!tmp)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
+               else {
+                   out = (stbi_uc*) tmp;
+                   out_size = layers * stride;
+               }
+
+               if (delays) {
+                  int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers );
+                  if (!new_delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  *delays = new_delays;
+                  delays_size = layers * sizeof(int);
+               }
+            } else {
+               out = (stbi_uc*)stbi__malloc( layers * stride );
+               if (!out)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
+               out_size = layers * stride;
+               if (delays) {
+                  *delays = (int*) stbi__malloc( layers * sizeof(int) );
+                  if (!*delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  delays_size = layers * sizeof(int);
+               }
+            }
+            memcpy( out + ((layers - 1) * stride), u, stride );
+            if (layers >= 2) {
+               two_back = out - 2 * stride;
+            }
+
+            if (delays) {
+               (*delays)[layers - 1U] = g.delay;
+            }
+         }
+      } while (u != 0);
+
+      // free temp buffer;
+      STBI_FREE(g.out);
+      STBI_FREE(g.history);
+      STBI_FREE(g.background);
+
+      // do the final conversion after loading everything;
+      if (req_comp && req_comp != 4)
+         out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h);
+
+      *z = layers;
+      return out;
+   } else {
+      return stbi__errpuc("not GIF", "Image was not as a gif type.");
+   }
+}
+
+static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *u = 0;
+   stbi__gif g;
+   memset(&g, 0, sizeof(g));
+   STBI_NOTUSED(ri);
+
+   u = stbi__gif_load_next(s, &g, comp, req_comp, 0);
+   if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
+   if (u) {
+      *x = g.w;
+      *y = g.h;
+
+      // moved conversion to after successful load so that the same
+      // can be done for multiple frames.
+      if (req_comp && req_comp != 4)
+         u = stbi__convert_format(u, 4, req_comp, g.w, g.h);
+   } else if (g.out) {
+      // if there was an error and we allocated an image buffer, free it!
+      STBI_FREE(g.out);
+   }
+
+   // free buffers needed for multiple frame loading;
+   STBI_FREE(g.history);
+   STBI_FREE(g.background);
+
+   return u;
+}
+
+static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   return stbi__gif_info_raw(s,x,y,comp);
+}
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR loader
+// originally by Nicolas Schulz
+#ifndef STBI_NO_HDR
+static int stbi__hdr_test_core(stbi__context *s, const char *signature)
+{
+   int i;
+   for (i=0; signature[i]; ++i)
+      if (stbi__get8(s) != signature[i])
+          return 0;
+   stbi__rewind(s);
+   return 1;
+}
+
+static int stbi__hdr_test(stbi__context* s)
+{
+   int r = stbi__hdr_test_core(s, "#?RADIANCE\n");
+   stbi__rewind(s);
+   if(!r) {
+       r = stbi__hdr_test_core(s, "#?RGBE\n");
+       stbi__rewind(s);
+   }
+   return r;
+}
+
+#define STBI__HDR_BUFLEN  1024
+static char *stbi__hdr_gettoken(stbi__context *z, char *buffer)
+{
+   int len=0;
+   char c = '\0';
+
+   c = (char) stbi__get8(z);
+
+   while (!stbi__at_eof(z) && c != '\n') {
+      buffer[len++] = c;
+      if (len == STBI__HDR_BUFLEN-1) {
+         // flush to end of line
+         while (!stbi__at_eof(z) && stbi__get8(z) != '\n')
+            ;
+         break;
+      }
+      c = (char) stbi__get8(z);
+   }
+
+   buffer[len] = 0;
+   return buffer;
+}
+
+static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp)
+{
+   if ( input[3] != 0 ) {
+      float f1;
+      // Exponent
+      f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8));
+      if (req_comp <= 2)
+         output[0] = (input[0] + input[1] + input[2]) * f1 / 3;
+      else {
+         output[0] = input[0] * f1;
+         output[1] = input[1] * f1;
+         output[2] = input[2] * f1;
+      }
+      if (req_comp == 2) output[1] = 1;
+      if (req_comp == 4) output[3] = 1;
+   } else {
+      switch (req_comp) {
+         case 4: output[3] = 1; /* fallthrough */
+         case 3: output[0] = output[1] = output[2] = 0;
+                 break;
+         case 2: output[1] = 1; /* fallthrough */
+         case 1: output[0] = 0;
+                 break;
+      }
+   }
+}
+
+static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   char buffer[STBI__HDR_BUFLEN];
+   char *token;
+   int valid = 0;
+   int width, height;
+   stbi_uc *scanline;
+   float *hdr_data;
+   int len;
+   unsigned char count, value;
+   int i, j, k, c1,c2, z;
+   const char *headerToken;
+   STBI_NOTUSED(ri);
+
+   // Check identifier
+   headerToken = stbi__hdr_gettoken(s,buffer);
+   if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0)
+      return stbi__errpf("not HDR", "Corrupt HDR image");
+
+   // Parse header
+   for(;;) {
+      token = stbi__hdr_gettoken(s,buffer);
+      if (token[0] == 0) break;
+      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
+   }
+
+   if (!valid)    return stbi__errpf("unsupported format", "Unsupported HDR format");
+
+   // Parse width and height
+   // can't use sscanf() if we're not using stdio!
+   token = stbi__hdr_gettoken(s,buffer);
+   if (strncmp(token, "-Y ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+   token += 3;
+   height = (int) strtol(token, &token, 10);
+   while (*token == ' ') ++token;
+   if (strncmp(token, "+X ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+   token += 3;
+   width = (int) strtol(token, NULL, 10);
+
+   if (height > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
+   if (width > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
+
+   *x = width;
+   *y = height;
+
+   if (comp) *comp = 3;
+   if (req_comp == 0) req_comp = 3;
+
+   if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0))
+      return stbi__errpf("too large", "HDR image is too large");
+
+   // Read data
+   hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0);
+   if (!hdr_data)
+      return stbi__errpf("outofmem", "Out of memory");
+
+   // Load image data
+   // image data is stored as some number of sca
+   if ( width < 8 || width >= 32768) {
+      // Read flat data
+      for (j=0; j < height; ++j) {
+         for (i=0; i < width; ++i) {
+            stbi_uc rgbe[4];
+           main_decode_loop:
+            stbi__getn(s, rgbe, 4);
+            stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp);
+         }
+      }
+   } else {
+      // Read RLE-encoded data
+      scanline = NULL;
+
+      for (j = 0; j < height; ++j) {
+         c1 = stbi__get8(s);
+         c2 = stbi__get8(s);
+         len = stbi__get8(s);
+         if (c1 != 2 || c2 != 2 || (len & 0x80)) {
+            // not run-length encoded, so we have to actually use THIS data as a decoded
+            // pixel (note this can't be a valid pixel--one of RGB must be >= 128)
+            stbi_uc rgbe[4];
+            rgbe[0] = (stbi_uc) c1;
+            rgbe[1] = (stbi_uc) c2;
+            rgbe[2] = (stbi_uc) len;
+            rgbe[3] = (stbi_uc) stbi__get8(s);
+            stbi__hdr_convert(hdr_data, rgbe, req_comp);
+            i = 1;
+            j = 0;
+            STBI_FREE(scanline);
+            goto main_decode_loop; // yes, this makes no sense
+         }
+         len <<= 8;
+         len |= stbi__get8(s);
+         if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); }
+         if (scanline == NULL) {
+            scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0);
+            if (!scanline) {
+               STBI_FREE(hdr_data);
+               return stbi__errpf("outofmem", "Out of memory");
+            }
+         }
+
+         for (k = 0; k < 4; ++k) {
+            int nleft;
+            i = 0;
+            while ((nleft = width - i) > 0) {
+               count = stbi__get8(s);
+               if (count > 128) {
+                  // Run
+                  value = stbi__get8(s);
+                  count -= 128;
+                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
+                  for (z = 0; z < count; ++z)
+                     scanline[i++ * 4 + k] = value;
+               } else {
+                  // Dump
+                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
+                  for (z = 0; z < count; ++z)
+                     scanline[i++ * 4 + k] = stbi__get8(s);
+               }
+            }
+         }
+         for (i=0; i < width; ++i)
+            stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp);
+      }
+      if (scanline)
+         STBI_FREE(scanline);
+   }
+
+   return hdr_data;
+}
+
+static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   char buffer[STBI__HDR_BUFLEN];
+   char *token;
+   int valid = 0;
+   int dummy;
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   if (stbi__hdr_test(s) == 0) {
+       stbi__rewind( s );
+       return 0;
+   }
+
+   for(;;) {
+      token = stbi__hdr_gettoken(s,buffer);
+      if (token[0] == 0) break;
+      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
+   }
+
+   if (!valid) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token = stbi__hdr_gettoken(s,buffer);
+   if (strncmp(token, "-Y ", 3)) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token += 3;
+   *y = (int) strtol(token, &token, 10);
+   while (*token == ' ') ++token;
+   if (strncmp(token, "+X ", 3)) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token += 3;
+   *x = (int) strtol(token, NULL, 10);
+   *comp = 3;
+   return 1;
+}
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   void *p;
+   stbi__bmp_data info;
+
+   info.all_a = 255;
+   p = stbi__bmp_parse_header(s, &info);
+   if (p == NULL) {
+      stbi__rewind( s );
+      return 0;
+   }
+   if (x) *x = s->img_x;
+   if (y) *y = s->img_y;
+   if (comp) {
+      if (info.bpp == 24 && info.ma == 0xff000000)
+         *comp = 3;
+      else
+         *comp = info.ma ? 4 : 3;
+   }
+   return 1;
+}
+#endif
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int channelCount, dummy, depth;
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+   if (stbi__get32be(s) != 0x38425053) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 1) {
+       stbi__rewind( s );
+       return 0;
+   }
+   stbi__skip(s, 6);
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   *y = stbi__get32be(s);
+   *x = stbi__get32be(s);
+   depth = stbi__get16be(s);
+   if (depth != 8 && depth != 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 3) {
+       stbi__rewind( s );
+       return 0;
+   }
+   *comp = 4;
+   return 1;
+}
+
+static int stbi__psd_is16(stbi__context *s)
+{
+   int channelCount, depth;
+   if (stbi__get32be(s) != 0x38425053) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 1) {
+       stbi__rewind( s );
+       return 0;
+   }
+   stbi__skip(s, 6);
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   STBI_NOTUSED(stbi__get32be(s));
+   STBI_NOTUSED(stbi__get32be(s));
+   depth = stbi__get16be(s);
+   if (depth != 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   return 1;
+}
+#endif
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int act_comp=0,num_packets=0,chained,dummy;
+   stbi__pic_packet packets[10];
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) {
+      stbi__rewind(s);
+      return 0;
+   }
+
+   stbi__skip(s, 88);
+
+   *x = stbi__get16be(s);
+   *y = stbi__get16be(s);
+   if (stbi__at_eof(s)) {
+      stbi__rewind( s);
+      return 0;
+   }
+   if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) {
+      stbi__rewind( s );
+      return 0;
+   }
+
+   stbi__skip(s, 8);
+
+   do {
+      stbi__pic_packet *packet;
+
+      if (num_packets==sizeof(packets)/sizeof(packets[0]))
+         return 0;
+
+      packet = &packets[num_packets++];
+      chained = stbi__get8(s);
+      packet->size    = stbi__get8(s);
+      packet->type    = stbi__get8(s);
+      packet->channel = stbi__get8(s);
+      act_comp |= packet->channel;
+
+      if (stbi__at_eof(s)) {
+          stbi__rewind( s );
+          return 0;
+      }
+      if (packet->size != 8) {
+          stbi__rewind( s );
+          return 0;
+      }
+   } while (chained);
+
+   *comp = (act_comp & 0x10 ? 4 : 3);
+
+   return 1;
+}
+#endif
+
+// *************************************************************************************************
+// Portable Gray Map and Portable Pixel Map loader
+// by Ken Miller
+//
+// PGM: http://netpbm.sourceforge.net/doc/pgm.html
+// PPM: http://netpbm.sourceforge.net/doc/ppm.html
+//
+// Known limitations:
+//    Does not support comments in the header section
+//    Does not support ASCII image data (formats P2 and P3)
+
+#ifndef STBI_NO_PNM
+
+static int      stbi__pnm_test(stbi__context *s)
+{
+   char p, t;
+   p = (char) stbi__get8(s);
+   t = (char) stbi__get8(s);
+   if (p != 'P' || (t != '5' && t != '6')) {
+       stbi__rewind( s );
+       return 0;
+   }
+   return 1;
+}
+
+static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *out;
+   STBI_NOTUSED(ri);
+
+   ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n);
+   if (ri->bits_per_channel == 0)
+      return 0;
+
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   *x = s->img_x;
+   *y = s->img_y;
+   if (comp) *comp = s->img_n;
+
+   if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0))
+      return stbi__errpuc("too large", "PNM too large");
+
+   out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0);
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) {
+      STBI_FREE(out);
+      return stbi__errpuc("bad PNM", "PNM file truncated");
+   }
+
+   if (req_comp && req_comp != s->img_n) {
+      if (ri->bits_per_channel == 16) {
+         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, s->img_n, req_comp, s->img_x, s->img_y);
+      } else {
+         out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y);
+      }
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+   return out;
+}
+
+static int      stbi__pnm_isspace(char c)
+{
+   return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r';
+}
+
+static void     stbi__pnm_skip_whitespace(stbi__context *s, char *c)
+{
+   for (;;) {
+      while (!stbi__at_eof(s) && stbi__pnm_isspace(*c))
+         *c = (char) stbi__get8(s);
+
+      if (stbi__at_eof(s) || *c != '#')
+         break;
+
+      while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' )
+         *c = (char) stbi__get8(s);
+   }
+}
+
+static int      stbi__pnm_isdigit(char c)
+{
+   return c >= '0' && c <= '9';
+}
+
+static int      stbi__pnm_getinteger(stbi__context *s, char *c)
+{
+   int value = 0;
+
+   while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) {
+      value = value*10 + (*c - '0');
+      *c = (char) stbi__get8(s);
+      if((value > 214748364) || (value == 214748364 && *c > '7'))
+          return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int");
+   }
+
+   return value;
+}
+
+static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int maxv, dummy;
+   char c, p, t;
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   stbi__rewind(s);
+
+   // Get identifier
+   p = (char) stbi__get8(s);
+   t = (char) stbi__get8(s);
+   if (p != 'P' || (t != '5' && t != '6')) {
+       stbi__rewind(s);
+       return 0;
+   }
+
+   *comp = (t == '6') ? 3 : 1;  // '5' is 1-component .pgm; '6' is 3-component .ppm
+
+   c = (char) stbi__get8(s);
+   stbi__pnm_skip_whitespace(s, &c);
+
+   *x = stbi__pnm_getinteger(s, &c); // read width
+   if(*x == 0)
+       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+   stbi__pnm_skip_whitespace(s, &c);
+
+   *y = stbi__pnm_getinteger(s, &c); // read height
+   if (*y == 0)
+       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+   stbi__pnm_skip_whitespace(s, &c);
+
+   maxv = stbi__pnm_getinteger(s, &c);  // read max value
+   if (maxv > 65535)
+      return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images");
+   else if (maxv > 255)
+      return 16;
+   else
+      return 8;
+}
+
+static int stbi__pnm_is16(stbi__context *s)
+{
+   if (stbi__pnm_info(s, NULL, NULL, NULL) == 16)
+	   return 1;
+   return 0;
+}
+#endif
+
+static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp)
+{
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_info(s, x, y, comp)) return 1;
+   #endif
+
+   #ifndef STBI_NO_PNG
+   if (stbi__png_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_GIF
+   if (stbi__gif_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_BMP
+   if (stbi__bmp_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PIC
+   if (stbi__pic_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_info(s, x, y, comp))  return 1;
+   #endif
+
+   // test tga last because it's a crappy test!
+   #ifndef STBI_NO_TGA
+   if (stbi__tga_info(s, x, y, comp))
+       return 1;
+   #endif
+   return stbi__err("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static int stbi__is_16_main(stbi__context *s)
+{
+   #ifndef STBI_NO_PNG
+   if (stbi__png_is16(s))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_is16(s))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_is16(s))  return 1;
+   #endif
+   return 0;
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp)
+{
+    FILE *f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f) return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_info_from_file(f, x, y, comp);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp)
+{
+   int r;
+   stbi__context s;
+   long pos = ftell(f);
+   stbi__start_file(&s, f);
+   r = stbi__info_main(&s,x,y,comp);
+   fseek(f,pos,SEEK_SET);
+   return r;
+}
+
+STBIDEF int stbi_is_16_bit(char const *filename)
+{
+    FILE *f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f) return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_is_16_bit_from_file(f);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_is_16_bit_from_file(FILE *f)
+{
+   int r;
+   stbi__context s;
+   long pos = ftell(f);
+   stbi__start_file(&s, f);
+   r = stbi__is_16_main(&s);
+   fseek(f,pos,SEEK_SET);
+   return r;
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__info_main(&s,x,y,comp);
+}
+
+STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
+   return stbi__info_main(&s,x,y,comp);
+}
+
+STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__is_16_main(&s);
+}
+
+STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *c, void *user)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
+   return stbi__is_16_main(&s);
+}
+
+#endif // STB_IMAGE_IMPLEMENTATION
+
+/*
+   revision history:
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) change sbti__shiftsigned to avoid clang -O2 bug
+                         1-bit BMP
+                         *_is_16_bit api
+                         avoid warnings
+      2.16  (2017-07-23) all functions have 16-bit variants;
+                         STBI_NO_STDIO works again;
+                         compilation fixes;
+                         fix rounding in unpremultiply;
+                         optimize vertical flip;
+                         disable raw_len validation;
+                         documentation fixes
+      2.15  (2017-03-18) fix png-1,2,4 bug; now all Imagenet JPGs decode;
+                         warning fixes; disable run-time SSE detection on gcc;
+                         uniform handling of optional "return" values;
+                         thread-safe initialization of zlib tables
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-11-29) add 16-bit API, only supported for PNG right now
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) allocate large structures on the stack
+                         remove white matting for transparent PSD
+                         fix reported channel count for PNG & BMP
+                         re-enable SSE2 in non-gcc 64-bit
+                         support RGB-formatted JPEG
+                         read 16-bit PNGs (only as 8-bit)
+      2.10  (2016-01-22) avoid warning introduced in 2.09 by STBI_REALLOC_SIZED
+      2.09  (2016-01-16) allow comments in PNM files
+                         16-bit-per-pixel TGA (not bit-per-component)
+                         info() for TGA could break due to .hdr handling
+                         info() for BMP to shares code instead of sloppy parse
+                         can use STBI_REALLOC_SIZED if allocator doesn't support realloc
+                         code cleanup
+      2.08  (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA
+      2.07  (2015-09-13) fix compiler warnings
+                         partial animated GIF support
+                         limited 16-bpc PSD support
+                         #ifdef unused functions
+                         bug with < 92 byte PIC,PNM,HDR,TGA
+      2.06  (2015-04-19) fix bug where PSD returns wrong '*comp' value
+      2.05  (2015-04-19) fix bug in progressive JPEG handling, fix warning
+      2.04  (2015-04-15) try to re-enable SIMD on MinGW 64-bit
+      2.03  (2015-04-12) extra corruption checking (mmozeiko)
+                         stbi_set_flip_vertically_on_load (nguillemot)
+                         fix NEON support; fix mingw support
+      2.02  (2015-01-19) fix incorrect assert, fix warning
+      2.01  (2015-01-17) fix various warnings; suppress SIMD on gcc 32-bit without -msse2
+      2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG
+      2.00  (2014-12-25) optimize JPG, including x86 SSE2 & NEON SIMD (ryg)
+                         progressive JPEG (stb)
+                         PGM/PPM support (Ken Miller)
+                         STBI_MALLOC,STBI_REALLOC,STBI_FREE
+                         GIF bugfix -- seemingly never worked
+                         STBI_NO_*, STBI_ONLY_*
+      1.48  (2014-12-14) fix incorrectly-named assert()
+      1.47  (2014-12-14) 1/2/4-bit PNG support, both direct and paletted (Omar Cornut & stb)
+                         optimize PNG (ryg)
+                         fix bug in interlaced PNG with user-specified channel count (stb)
+      1.46  (2014-08-26)
+              fix broken tRNS chunk (colorkey-style transparency) in non-paletted PNG
+      1.45  (2014-08-16)
+              fix MSVC-ARM internal compiler error by wrapping malloc
+      1.44  (2014-08-07)
+              various warning fixes from Ronny Chevalier
+      1.43  (2014-07-15)
+              fix MSVC-only compiler problem in code changed in 1.42
+      1.42  (2014-07-09)
+              don't define _CRT_SECURE_NO_WARNINGS (affects user code)
+              fixes to stbi__cleanup_jpeg path
+              added STBI_ASSERT to avoid requiring assert.h
+      1.41  (2014-06-25)
+              fix search&replace from 1.36 that messed up comments/error messages
+      1.40  (2014-06-22)
+              fix gcc struct-initialization warning
+      1.39  (2014-06-15)
+              fix to TGA optimization when req_comp != number of components in TGA;
+              fix to GIF loading because BMP wasn't rewinding (whoops, no GIFs in my test suite)
+              add support for BMP version 5 (more ignored fields)
+      1.38  (2014-06-06)
+              suppress MSVC warnings on integer casts truncating values
+              fix accidental rename of 'skip' field of I/O
+      1.37  (2014-06-04)
+              remove duplicate typedef
+      1.36  (2014-06-03)
+              convert to header file single-file library
+              if de-iphone isn't set, load iphone images color-swapped instead of returning NULL
+      1.35  (2014-05-27)
+              various warnings
+              fix broken STBI_SIMD path
+              fix bug where stbi_load_from_file no longer left file pointer in correct place
+              fix broken non-easy path for 32-bit BMP (possibly never used)
+              TGA optimization by Arseny Kapoulkine
+      1.34  (unknown)
+              use STBI_NOTUSED in stbi__resample_row_generic(), fix one more leak in tga failure case
+      1.33  (2011-07-14)
+              make stbi_is_hdr work in STBI_NO_HDR (as specified), minor compiler-friendly improvements
+      1.32  (2011-07-13)
+              support for "info" function for all supported filetypes (SpartanJ)
+      1.31  (2011-06-20)
+              a few more leak fixes, bug in PNG handling (SpartanJ)
+      1.30  (2011-06-11)
+              added ability to load files via callbacks to accomidate custom input streams (Ben Wenger)
+              removed deprecated format-specific test/load functions
+              removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway
+              error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha)
+              fix inefficiency in decoding 32-bit BMP (David Woo)
+      1.29  (2010-08-16)
+              various warning fixes from Aurelien Pocheville
+      1.28  (2010-08-01)
+              fix bug in GIF palette transparency (SpartanJ)
+      1.27  (2010-08-01)
+              cast-to-stbi_uc to fix warnings
+      1.26  (2010-07-24)
+              fix bug in file buffering for PNG reported by SpartanJ
+      1.25  (2010-07-17)
+              refix trans_data warning (Won Chun)
+      1.24  (2010-07-12)
+              perf improvements reading from files on platforms with lock-heavy fgetc()
+              minor perf improvements for jpeg
+              deprecated type-specific functions so we'll get feedback if they're needed
+              attempt to fix trans_data warning (Won Chun)
+      1.23    fixed bug in iPhone support
+      1.22  (2010-07-10)
+              removed image *writing* support
+              stbi_info support from Jetro Lauha
+              GIF support from Jean-Marc Lienher
+              iPhone PNG-extensions from James Brown
+              warning-fixes from Nicolas Schulz and Janez Zemva (i.stbi__err. Janez (U+017D)emva)
+      1.21    fix use of 'stbi_uc' in header (reported by jon blow)
+      1.20    added support for Softimage PIC, by Tom Seddon
+      1.19    bug in interlaced PNG corruption check (found by ryg)
+      1.18  (2008-08-02)
+              fix a threading bug (local mutable static)
+      1.17    support interlaced PNG
+      1.16    major bugfix - stbi__convert_format converted one too many pixels
+      1.15    initialize some fields for thread safety
+      1.14    fix threadsafe conversion bug
+              header-file-only version (#define STBI_HEADER_FILE_ONLY before including)
+      1.13    threadsafe
+      1.12    const qualifiers in the API
+      1.11    Support installable IDCT, colorspace conversion routines
+      1.10    Fixes for 64-bit (don't use "unsigned long")
+              optimized upsampling by Fabian "ryg" Giesen
+      1.09    Fix format-conversion for PSD code (bad global variables!)
+      1.08    Thatcher Ulrich's PSD code integrated by Nicolas Schulz
+      1.07    attempt to fix C++ warning/errors again
+      1.06    attempt to fix C++ warning/errors again
+      1.05    fix TGA loading to return correct *comp and use good luminance calc
+      1.04    default float alpha is 1, not 255; use 'void *' for stbi_image_free
+      1.03    bugfixes to STBI_NO_STDIO, STBI_NO_HDR
+      1.02    support for (subset of) HDR files, float interface for preferred access to them
+      1.01    fix bug: possible bug in handling right-side up bmps... not sure
+              fix bug: the stbi__bmp_load() and stbi__tga_load() functions didn't work at all
+      1.00    interface to zlib that skips zlib header
+      0.99    correct handling of alpha in palette
+      0.98    TGA loader by lonesock; dynamically add loaders (untested)
+      0.97    jpeg errors on too large a file; also catch another malloc failure
+      0.96    fix detection of invalid v value - particleman@mollyrocket forum
+      0.95    during header scan, seek to markers in case of padding
+      0.94    STBI_NO_STDIO to disable stdio usage; rename all #defines the same
+      0.93    handle jpegtran output; verbose errors
+      0.92    read 4,8,16,24,32-bit BMP files of several formats
+      0.91    output 24-bit Windows 3.0 BMP files
+      0.90    fix a few more warnings; bump version number to approach 1.0
+      0.61    bugfixes due to Marc LeBlanc, Christopher Lloyd
+      0.60    fix compiling as c++
+      0.59    fix warnings: merge Dave Moore's -Wall fixes
+      0.58    fix bug: zlib uncompressed mode len/nlen was wrong endian
+      0.57    fix bug: jpg last huffman symbol before marker was >9 bits but less than 16 available
+      0.56    fix bug: zlib uncompressed mode len vs. nlen
+      0.55    fix bug: restart_interval not initialized to 0
+      0.54    allow NULL for 'int *comp'
+      0.53    fix bug in png 3->4; speedup png decoding
+      0.52    png handles req_comp=3,4 directly; minor cleanup; jpeg comments
+      0.51    obey req_comp requests, 1-component jpegs return as 1-component,
+              on 'test' only check type, not whether we support this variant
+      0.50  (2006-11-19)
+              first released version
+*/
+
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/
diff --git a/07_nlm_denoise/blackbook537/include/stb_image_write.h b/07_nlm_denoise/blackbook537/include/stb_image_write.h
new file mode 100644
index 00000000..e4b32ed1
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/include/stb_image_write.h
@@ -0,0 +1,1724 @@
+/* stb_image_write - v1.16 - public domain - http://nothings.org/stb
+   writes out PNG/BMP/TGA/JPEG/HDR images to C stdio - Sean Barrett 2010-2015
+                                     no warranty implied; use at your own risk
+
+   Before #including,
+
+       #define STB_IMAGE_WRITE_IMPLEMENTATION
+
+   in the file that you want to have the implementation.
+
+   Will probably not work correctly with strict-aliasing optimizations.
+
+ABOUT:
+
+   This header file is a library for writing images to C stdio or a callback.
+
+   The PNG output is not optimal; it is 20-50% larger than the file
+   written by a decent optimizing implementation; though providing a custom
+   zlib compress function (see STBIW_ZLIB_COMPRESS) can mitigate that.
+   This library is designed for source code compactness and simplicity,
+   not optimal image file size or run-time performance.
+
+BUILDING:
+
+   You can #define STBIW_ASSERT(x) before the #include to avoid using assert.h.
+   You can #define STBIW_MALLOC(), STBIW_REALLOC(), and STBIW_FREE() to replace
+   malloc,realloc,free.
+   You can #define STBIW_MEMMOVE() to replace memmove()
+   You can #define STBIW_ZLIB_COMPRESS to use a custom zlib-style compress function
+   for PNG compression (instead of the builtin one), it must have the following signature:
+   unsigned char * my_compress(unsigned char *data, int data_len, int *out_len, int quality);
+   The returned data will be freed with STBIW_FREE() (free() by default),
+   so it must be heap allocated with STBIW_MALLOC() (malloc() by default),
+
+UNICODE:
+
+   If compiling for Windows and you wish to use Unicode filenames, compile
+   with
+       #define STBIW_WINDOWS_UTF8
+   and pass utf8-encoded filenames. Call stbiw_convert_wchar_to_utf8 to convert
+   Windows wchar_t filenames to utf8.
+
+USAGE:
+
+   There are five functions, one for each image file format:
+
+     int stbi_write_png(char const *filename, int w, int h, int comp, const void *data, int stride_in_bytes);
+     int stbi_write_bmp(char const *filename, int w, int h, int comp, const void *data);
+     int stbi_write_tga(char const *filename, int w, int h, int comp, const void *data);
+     int stbi_write_jpg(char const *filename, int w, int h, int comp, const void *data, int quality);
+     int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
+
+     void stbi_flip_vertically_on_write(int flag); // flag is non-zero to flip data vertically
+
+   There are also five equivalent functions that use an arbitrary write function. You are
+   expected to open/close your file-equivalent before and after calling these:
+
+     int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
+     int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+     int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+     int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
+     int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality);
+
+   where the callback is:
+      void stbi_write_func(void *context, void *data, int size);
+
+   You can configure it with these global variables:
+      int stbi_write_tga_with_rle;             // defaults to true; set to 0 to disable RLE
+      int stbi_write_png_compression_level;    // defaults to 8; set to higher for more compression
+      int stbi_write_force_png_filter;         // defaults to -1; set to 0..5 to force a filter mode
+
+
+   You can define STBI_WRITE_NO_STDIO to disable the file variant of these
+   functions, so the library will not use stdio.h at all. However, this will
+   also disable HDR writing, because it requires stdio for formatted output.
+
+   Each function returns 0 on failure and non-0 on success.
+
+   The functions create an image file defined by the parameters. The image
+   is a rectangle of pixels stored from left-to-right, top-to-bottom.
+   Each pixel contains 'comp' channels of data stored interleaved with 8-bits
+   per channel, in the following order: 1=Y, 2=YA, 3=RGB, 4=RGBA. (Y is
+   monochrome color.) The rectangle is 'w' pixels wide and 'h' pixels tall.
+   The *data pointer points to the first byte of the top-left-most pixel.
+   For PNG, "stride_in_bytes" is the distance in bytes from the first byte of
+   a row of pixels to the first byte of the next row of pixels.
+
+   PNG creates output files with the same number of components as the input.
+   The BMP format expands Y to RGB in the file format and does not
+   output alpha.
+
+   PNG supports writing rectangles of data even when the bytes storing rows of
+   data are not consecutive in memory (e.g. sub-rectangles of a larger image),
+   by supplying the stride between the beginning of adjacent rows. The other
+   formats do not. (Thus you cannot write a native-format BMP through the BMP
+   writer, both because it is in BGR order and because it may have padding
+   at the end of the line.)
+
+   PNG allows you to set the deflate compression level by setting the global
+   variable 'stbi_write_png_compression_level' (it defaults to 8).
+
+   HDR expects linear float data. Since the format is always 32-bit rgb(e)
+   data, alpha (if provided) is discarded, and for monochrome data it is
+   replicated across all three channels.
+
+   TGA supports RLE or non-RLE compressed data. To use non-RLE-compressed
+   data, set the global variable 'stbi_write_tga_with_rle' to 0.
+
+   JPEG does ignore alpha channels in input data; quality is between 1 and 100.
+   Higher quality looks better but results in a bigger image.
+   JPEG baseline (no JPEG progressive).
+
+CREDITS:
+
+
+   Sean Barrett           -    PNG/BMP/TGA
+   Baldur Karlsson        -    HDR
+   Jean-Sebastien Guay    -    TGA monochrome
+   Tim Kelsey             -    misc enhancements
+   Alan Hickman           -    TGA RLE
+   Emmanuel Julien        -    initial file IO callback implementation
+   Jon Olick              -    original jo_jpeg.cpp code
+   Daniel Gibson          -    integrate JPEG, allow external zlib
+   Aarni Koskela          -    allow choosing PNG filter
+
+   bugfixes:
+      github:Chribba
+      Guillaume Chereau
+      github:jry2
+      github:romigrou
+      Sergio Gonzalez
+      Jonas Karlsson
+      Filip Wasil
+      Thatcher Ulrich
+      github:poppolopoppo
+      Patrick Boettcher
+      github:xeekworx
+      Cap Petschulat
+      Simon Rodriguez
+      Ivan Tikhonov
+      github:ignotion
+      Adam Schackart
+      Andrew Kensler
+
+LICENSE
+
+  See end of file for license information.
+
+*/
+
+#ifndef INCLUDE_STB_IMAGE_WRITE_H
+#define INCLUDE_STB_IMAGE_WRITE_H
+
+#include <stdlib.h>
+
+// if STB_IMAGE_WRITE_STATIC causes problems, try defining STBIWDEF to 'inline' or 'static inline'
+#ifndef STBIWDEF
+#ifdef STB_IMAGE_WRITE_STATIC
+#define STBIWDEF  static
+#else
+#ifdef __cplusplus
+#define STBIWDEF  extern "C"
+#else
+#define STBIWDEF  extern
+#endif
+#endif
+#endif
+
+#ifndef STB_IMAGE_WRITE_STATIC  // C++ forbids static forward declarations
+STBIWDEF int stbi_write_tga_with_rle;
+STBIWDEF int stbi_write_png_compression_level;
+STBIWDEF int stbi_write_force_png_filter;
+#endif
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_png(char const *filename, int w, int h, int comp, const void  *data, int stride_in_bytes);
+STBIWDEF int stbi_write_bmp(char const *filename, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_tga(char const *filename, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void  *data, int quality);
+
+#ifdef STBIW_WINDOWS_UTF8
+STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
+#endif
+#endif
+
+typedef void stbi_write_func(void *context, void *data, int size);
+
+STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
+STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
+STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void  *data, int quality);
+
+STBIWDEF void stbi_flip_vertically_on_write(int flip_boolean);
+
+#endif//INCLUDE_STB_IMAGE_WRITE_H
+
+#ifdef STB_IMAGE_WRITE_IMPLEMENTATION
+
+#ifdef _WIN32
+   #ifndef _CRT_SECURE_NO_WARNINGS
+   #define _CRT_SECURE_NO_WARNINGS
+   #endif
+   #ifndef _CRT_NONSTDC_NO_DEPRECATE
+   #define _CRT_NONSTDC_NO_DEPRECATE
+   #endif
+#endif
+
+#ifndef STBI_WRITE_NO_STDIO
+#include <stdio.h>
+#endif // STBI_WRITE_NO_STDIO
+
+#include <stdarg.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#if defined(STBIW_MALLOC) && defined(STBIW_FREE) && (defined(STBIW_REALLOC) || defined(STBIW_REALLOC_SIZED))
+// ok
+#elif !defined(STBIW_MALLOC) && !defined(STBIW_FREE) && !defined(STBIW_REALLOC) && !defined(STBIW_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBIW_MALLOC, STBIW_FREE, and STBIW_REALLOC (or STBIW_REALLOC_SIZED)."
+#endif
+
+#ifndef STBIW_MALLOC
+#define STBIW_MALLOC(sz)        malloc(sz)
+#define STBIW_REALLOC(p,newsz)  realloc(p,newsz)
+#define STBIW_FREE(p)           free(p)
+#endif
+
+#ifndef STBIW_REALLOC_SIZED
+#define STBIW_REALLOC_SIZED(p,oldsz,newsz) STBIW_REALLOC(p,newsz)
+#endif
+
+
+#ifndef STBIW_MEMMOVE
+#define STBIW_MEMMOVE(a,b,sz) memmove(a,b,sz)
+#endif
+
+
+#ifndef STBIW_ASSERT
+#include <assert.h>
+#define STBIW_ASSERT(x) assert(x)
+#endif
+
+#define STBIW_UCHAR(x) (unsigned char) ((x) & 0xff)
+
+#ifdef STB_IMAGE_WRITE_STATIC
+static int stbi_write_png_compression_level = 8;
+static int stbi_write_tga_with_rle = 1;
+static int stbi_write_force_png_filter = -1;
+#else
+int stbi_write_png_compression_level = 8;
+int stbi_write_tga_with_rle = 1;
+int stbi_write_force_png_filter = -1;
+#endif
+
+static int stbi__flip_vertically_on_write = 0;
+
+STBIWDEF void stbi_flip_vertically_on_write(int flag)
+{
+   stbi__flip_vertically_on_write = flag;
+}
+
+typedef struct
+{
+   stbi_write_func *func;
+   void *context;
+   unsigned char buffer[64];
+   int buf_used;
+} stbi__write_context;
+
+// initialize a callback-based context
+static void stbi__start_write_callbacks(stbi__write_context *s, stbi_write_func *c, void *context)
+{
+   s->func    = c;
+   s->context = context;
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+
+static void stbi__stdio_write(void *context, void *data, int size)
+{
+   fwrite(data,1,size,(FILE*) context);
+}
+
+#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
+#ifdef __cplusplus
+#define STBIW_EXTERN extern "C"
+#else
+#define STBIW_EXTERN extern
+#endif
+STBIW_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
+STBIW_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
+
+STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
+{
+   return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE *stbiw__fopen(char const *filename, char const *mode)
+{
+   FILE *f;
+#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
+   wchar_t wMode[64];
+   wchar_t wFilename[1024];
+   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
+      return 0;
+
+   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
+      return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != _wfopen_s(&f, wFilename, wMode))
+      f = 0;
+#else
+   f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != fopen_s(&f, filename, mode))
+      f=0;
+#else
+   f = fopen(filename, mode);
+#endif
+   return f;
+}
+
+static int stbi__start_write_file(stbi__write_context *s, const char *filename)
+{
+   FILE *f = stbiw__fopen(filename, "wb");
+   stbi__start_write_callbacks(s, stbi__stdio_write, (void *) f);
+   return f != NULL;
+}
+
+static void stbi__end_write_file(stbi__write_context *s)
+{
+   fclose((FILE *)s->context);
+}
+
+#endif // !STBI_WRITE_NO_STDIO
+
+typedef unsigned int stbiw_uint32;
+typedef int stb_image_write_test[sizeof(stbiw_uint32)==4 ? 1 : -1];
+
+static void stbiw__writefv(stbi__write_context *s, const char *fmt, va_list v)
+{
+   while (*fmt) {
+      switch (*fmt++) {
+         case ' ': break;
+         case '1': { unsigned char x = STBIW_UCHAR(va_arg(v, int));
+                     s->func(s->context,&x,1);
+                     break; }
+         case '2': { int x = va_arg(v,int);
+                     unsigned char b[2];
+                     b[0] = STBIW_UCHAR(x);
+                     b[1] = STBIW_UCHAR(x>>8);
+                     s->func(s->context,b,2);
+                     break; }
+         case '4': { stbiw_uint32 x = va_arg(v,int);
+                     unsigned char b[4];
+                     b[0]=STBIW_UCHAR(x);
+                     b[1]=STBIW_UCHAR(x>>8);
+                     b[2]=STBIW_UCHAR(x>>16);
+                     b[3]=STBIW_UCHAR(x>>24);
+                     s->func(s->context,b,4);
+                     break; }
+         default:
+            STBIW_ASSERT(0);
+            return;
+      }
+   }
+}
+
+static void stbiw__writef(stbi__write_context *s, const char *fmt, ...)
+{
+   va_list v;
+   va_start(v, fmt);
+   stbiw__writefv(s, fmt, v);
+   va_end(v);
+}
+
+static void stbiw__write_flush(stbi__write_context *s)
+{
+   if (s->buf_used) {
+      s->func(s->context, &s->buffer, s->buf_used);
+      s->buf_used = 0;
+   }
+}
+
+static void stbiw__putc(stbi__write_context *s, unsigned char c)
+{
+   s->func(s->context, &c, 1);
+}
+
+static void stbiw__write1(stbi__write_context *s, unsigned char a)
+{
+   if ((size_t)s->buf_used + 1 > sizeof(s->buffer))
+      stbiw__write_flush(s);
+   s->buffer[s->buf_used++] = a;
+}
+
+static void stbiw__write3(stbi__write_context *s, unsigned char a, unsigned char b, unsigned char c)
+{
+   int n;
+   if ((size_t)s->buf_used + 3 > sizeof(s->buffer))
+      stbiw__write_flush(s);
+   n = s->buf_used;
+   s->buf_used = n+3;
+   s->buffer[n+0] = a;
+   s->buffer[n+1] = b;
+   s->buffer[n+2] = c;
+}
+
+static void stbiw__write_pixel(stbi__write_context *s, int rgb_dir, int comp, int write_alpha, int expand_mono, unsigned char *d)
+{
+   unsigned char bg[3] = { 255, 0, 255}, px[3];
+   int k;
+
+   if (write_alpha < 0)
+      stbiw__write1(s, d[comp - 1]);
+
+   switch (comp) {
+      case 2: // 2 pixels = mono + alpha, alpha is written separately, so same as 1-channel case
+      case 1:
+         if (expand_mono)
+            stbiw__write3(s, d[0], d[0], d[0]); // monochrome bmp
+         else
+            stbiw__write1(s, d[0]);  // monochrome TGA
+         break;
+      case 4:
+         if (!write_alpha) {
+            // composite against pink background
+            for (k = 0; k < 3; ++k)
+               px[k] = bg[k] + ((d[k] - bg[k]) * d[3]) / 255;
+            stbiw__write3(s, px[1 - rgb_dir], px[1], px[1 + rgb_dir]);
+            break;
+         }
+         /* FALLTHROUGH */
+      case 3:
+         stbiw__write3(s, d[1 - rgb_dir], d[1], d[1 + rgb_dir]);
+         break;
+   }
+   if (write_alpha > 0)
+      stbiw__write1(s, d[comp - 1]);
+}
+
+static void stbiw__write_pixels(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, void *data, int write_alpha, int scanline_pad, int expand_mono)
+{
+   stbiw_uint32 zero = 0;
+   int i,j, j_end;
+
+   if (y <= 0)
+      return;
+
+   if (stbi__flip_vertically_on_write)
+      vdir *= -1;
+
+   if (vdir < 0) {
+      j_end = -1; j = y-1;
+   } else {
+      j_end =  y; j = 0;
+   }
+
+   for (; j != j_end; j += vdir) {
+      for (i=0; i < x; ++i) {
+         unsigned char *d = (unsigned char *) data + (j*x+i)*comp;
+         stbiw__write_pixel(s, rgb_dir, comp, write_alpha, expand_mono, d);
+      }
+      stbiw__write_flush(s);
+      s->func(s->context, &zero, scanline_pad);
+   }
+}
+
+static int stbiw__outfile(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, int expand_mono, void *data, int alpha, int pad, const char *fmt, ...)
+{
+   if (y < 0 || x < 0) {
+      return 0;
+   } else {
+      va_list v;
+      va_start(v, fmt);
+      stbiw__writefv(s, fmt, v);
+      va_end(v);
+      stbiw__write_pixels(s,rgb_dir,vdir,x,y,comp,data,alpha,pad, expand_mono);
+      return 1;
+   }
+}
+
+static int stbi_write_bmp_core(stbi__write_context *s, int x, int y, int comp, const void *data)
+{
+   if (comp != 4) {
+      // write RGB bitmap
+      int pad = (-x*3) & 3;
+      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *) data,0,pad,
+              "11 4 22 4" "4 44 22 444444",
+              'B', 'M', 14+40+(x*3+pad)*y, 0,0, 14+40,  // file header
+               40, x,y, 1,24, 0,0,0,0,0,0);             // bitmap header
+   } else {
+      // RGBA bitmaps need a v4 header
+      // use BI_BITFIELDS mode with 32bpp and alpha mask
+      // (straight BI_RGB with alpha mask doesn't work in most readers)
+      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *)data,1,0,
+         "11 4 22 4" "4 44 22 444444 4444 4 444 444 444 444",
+         'B', 'M', 14+108+x*y*4, 0, 0, 14+108, // file header
+         108, x,y, 1,32, 3,0,0,0,0,0, 0xff0000,0xff00,0xff,0xff000000u, 0, 0,0,0, 0,0,0, 0,0,0, 0,0,0); // bitmap V4 header
+   }
+}
+
+STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_bmp_core(&s, x, y, comp, data);
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_bmp(char const *filename, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_bmp_core(&s, x, y, comp, data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif //!STBI_WRITE_NO_STDIO
+
+static int stbi_write_tga_core(stbi__write_context *s, int x, int y, int comp, void *data)
+{
+   int has_alpha = (comp == 2 || comp == 4);
+   int colorbytes = has_alpha ? comp-1 : comp;
+   int format = colorbytes < 2 ? 3 : 2; // 3 color channels (RGB/RGBA) = 2, 1 color channel (Y/YA) = 3
+
+   if (y < 0 || x < 0)
+      return 0;
+
+   if (!stbi_write_tga_with_rle) {
+      return stbiw__outfile(s, -1, -1, x, y, comp, 0, (void *) data, has_alpha, 0,
+         "111 221 2222 11", 0, 0, format, 0, 0, 0, 0, 0, x, y, (colorbytes + has_alpha) * 8, has_alpha * 8);
+   } else {
+      int i,j,k;
+      int jend, jdir;
+
+      stbiw__writef(s, "111 221 2222 11", 0,0,format+8, 0,0,0, 0,0,x,y, (colorbytes + has_alpha) * 8, has_alpha * 8);
+
+      if (stbi__flip_vertically_on_write) {
+         j = 0;
+         jend = y;
+         jdir = 1;
+      } else {
+         j = y-1;
+         jend = -1;
+         jdir = -1;
+      }
+      for (; j != jend; j += jdir) {
+         unsigned char *row = (unsigned char *) data + j * x * comp;
+         int len;
+
+         for (i = 0; i < x; i += len) {
+            unsigned char *begin = row + i * comp;
+            int diff = 1;
+            len = 1;
+
+            if (i < x - 1) {
+               ++len;
+               diff = memcmp(begin, row + (i + 1) * comp, comp);
+               if (diff) {
+                  const unsigned char *prev = begin;
+                  for (k = i + 2; k < x && len < 128; ++k) {
+                     if (memcmp(prev, row + k * comp, comp)) {
+                        prev += comp;
+                        ++len;
+                     } else {
+                        --len;
+                        break;
+                     }
+                  }
+               } else {
+                  for (k = i + 2; k < x && len < 128; ++k) {
+                     if (!memcmp(begin, row + k * comp, comp)) {
+                        ++len;
+                     } else {
+                        break;
+                     }
+                  }
+               }
+            }
+
+            if (diff) {
+               unsigned char header = STBIW_UCHAR(len - 1);
+               stbiw__write1(s, header);
+               for (k = 0; k < len; ++k) {
+                  stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin + k * comp);
+               }
+            } else {
+               unsigned char header = STBIW_UCHAR(len - 129);
+               stbiw__write1(s, header);
+               stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin);
+            }
+         }
+      }
+      stbiw__write_flush(s);
+   }
+   return 1;
+}
+
+STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_tga_core(&s, x, y, comp, (void *) data);
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_tga(char const *filename, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_tga_core(&s, x, y, comp, (void *) data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR writer
+// by Baldur Karlsson
+
+#define stbiw__max(a, b)  ((a) > (b) ? (a) : (b))
+
+#ifndef STBI_WRITE_NO_STDIO
+
+static void stbiw__linear_to_rgbe(unsigned char *rgbe, float *linear)
+{
+   int exponent;
+   float maxcomp = stbiw__max(linear[0], stbiw__max(linear[1], linear[2]));
+
+   if (maxcomp < 1e-32f) {
+      rgbe[0] = rgbe[1] = rgbe[2] = rgbe[3] = 0;
+   } else {
+      float normalize = (float) frexp(maxcomp, &exponent) * 256.0f/maxcomp;
+
+      rgbe[0] = (unsigned char)(linear[0] * normalize);
+      rgbe[1] = (unsigned char)(linear[1] * normalize);
+      rgbe[2] = (unsigned char)(linear[2] * normalize);
+      rgbe[3] = (unsigned char)(exponent + 128);
+   }
+}
+
+static void stbiw__write_run_data(stbi__write_context *s, int length, unsigned char databyte)
+{
+   unsigned char lengthbyte = STBIW_UCHAR(length+128);
+   STBIW_ASSERT(length+128 <= 255);
+   s->func(s->context, &lengthbyte, 1);
+   s->func(s->context, &databyte, 1);
+}
+
+static void stbiw__write_dump_data(stbi__write_context *s, int length, unsigned char *data)
+{
+   unsigned char lengthbyte = STBIW_UCHAR(length);
+   STBIW_ASSERT(length <= 128); // inconsistent with spec but consistent with official code
+   s->func(s->context, &lengthbyte, 1);
+   s->func(s->context, data, length);
+}
+
+static void stbiw__write_hdr_scanline(stbi__write_context *s, int width, int ncomp, unsigned char *scratch, float *scanline)
+{
+   unsigned char scanlineheader[4] = { 2, 2, 0, 0 };
+   unsigned char rgbe[4];
+   float linear[3];
+   int x;
+
+   scanlineheader[2] = (width&0xff00)>>8;
+   scanlineheader[3] = (width&0x00ff);
+
+   /* skip RLE for images too small or large */
+   if (width < 8 || width >= 32768) {
+      for (x=0; x < width; x++) {
+         switch (ncomp) {
+            case 4: /* fallthrough */
+            case 3: linear[2] = scanline[x*ncomp + 2];
+                    linear[1] = scanline[x*ncomp + 1];
+                    linear[0] = scanline[x*ncomp + 0];
+                    break;
+            default:
+                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
+                    break;
+         }
+         stbiw__linear_to_rgbe(rgbe, linear);
+         s->func(s->context, rgbe, 4);
+      }
+   } else {
+      int c,r;
+      /* encode into scratch buffer */
+      for (x=0; x < width; x++) {
+         switch(ncomp) {
+            case 4: /* fallthrough */
+            case 3: linear[2] = scanline[x*ncomp + 2];
+                    linear[1] = scanline[x*ncomp + 1];
+                    linear[0] = scanline[x*ncomp + 0];
+                    break;
+            default:
+                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
+                    break;
+         }
+         stbiw__linear_to_rgbe(rgbe, linear);
+         scratch[x + width*0] = rgbe[0];
+         scratch[x + width*1] = rgbe[1];
+         scratch[x + width*2] = rgbe[2];
+         scratch[x + width*3] = rgbe[3];
+      }
+
+      s->func(s->context, scanlineheader, 4);
+
+      /* RLE each component separately */
+      for (c=0; c < 4; c++) {
+         unsigned char *comp = &scratch[width*c];
+
+         x = 0;
+         while (x < width) {
+            // find first run
+            r = x;
+            while (r+2 < width) {
+               if (comp[r] == comp[r+1] && comp[r] == comp[r+2])
+                  break;
+               ++r;
+            }
+            if (r+2 >= width)
+               r = width;
+            // dump up to first run
+            while (x < r) {
+               int len = r-x;
+               if (len > 128) len = 128;
+               stbiw__write_dump_data(s, len, &comp[x]);
+               x += len;
+            }
+            // if there's a run, output it
+            if (r+2 < width) { // same test as what we break out of in search loop, so only true if we break'd
+               // find next byte after run
+               while (r < width && comp[r] == comp[x])
+                  ++r;
+               // output run up to r
+               while (x < r) {
+                  int len = r-x;
+                  if (len > 127) len = 127;
+                  stbiw__write_run_data(s, len, comp[x]);
+                  x += len;
+               }
+            }
+         }
+      }
+   }
+}
+
+static int stbi_write_hdr_core(stbi__write_context *s, int x, int y, int comp, float *data)
+{
+   if (y <= 0 || x <= 0 || data == NULL)
+      return 0;
+   else {
+      // Each component is stored separately. Allocate scratch space for full output scanline.
+      unsigned char *scratch = (unsigned char *) STBIW_MALLOC(x*4);
+      int i, len;
+      char buffer[128];
+      char header[] = "#?RADIANCE\n# Written by stb_image_write.h\nFORMAT=32-bit_rle_rgbe\n";
+      s->func(s->context, header, sizeof(header)-1);
+
+#ifdef __STDC_LIB_EXT1__
+      len = sprintf_s(buffer, sizeof(buffer), "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
+#else
+      len = sprintf(buffer, "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
+#endif
+      s->func(s->context, buffer, len);
+
+      for(i=0; i < y; i++)
+         stbiw__write_hdr_scanline(s, x, comp, scratch, data + comp*x*(stbi__flip_vertically_on_write ? y-1-i : i));
+      STBIW_FREE(scratch);
+      return 1;
+   }
+}
+
+STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const float *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_hdr_core(&s, x, y, comp, (float *) data);
+}
+
+STBIWDEF int stbi_write_hdr(char const *filename, int x, int y, int comp, const float *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_hdr_core(&s, x, y, comp, (float *) data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif // STBI_WRITE_NO_STDIO
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PNG writer
+//
+
+#ifndef STBIW_ZLIB_COMPRESS
+// stretchy buffer; stbiw__sbpush() == vector<>::push_back() -- stbiw__sbcount() == vector<>::size()
+#define stbiw__sbraw(a) ((int *) (void *) (a) - 2)
+#define stbiw__sbm(a)   stbiw__sbraw(a)[0]
+#define stbiw__sbn(a)   stbiw__sbraw(a)[1]
+
+#define stbiw__sbneedgrow(a,n)  ((a)==0 || stbiw__sbn(a)+n >= stbiw__sbm(a))
+#define stbiw__sbmaybegrow(a,n) (stbiw__sbneedgrow(a,(n)) ? stbiw__sbgrow(a,n) : 0)
+#define stbiw__sbgrow(a,n)  stbiw__sbgrowf((void **) &(a), (n), sizeof(*(a)))
+
+#define stbiw__sbpush(a, v)      (stbiw__sbmaybegrow(a,1), (a)[stbiw__sbn(a)++] = (v))
+#define stbiw__sbcount(a)        ((a) ? stbiw__sbn(a) : 0)
+#define stbiw__sbfree(a)         ((a) ? STBIW_FREE(stbiw__sbraw(a)),0 : 0)
+
+static void *stbiw__sbgrowf(void **arr, int increment, int itemsize)
+{
+   int m = *arr ? 2*stbiw__sbm(*arr)+increment : increment+1;
+   void *p = STBIW_REALLOC_SIZED(*arr ? stbiw__sbraw(*arr) : 0, *arr ? (stbiw__sbm(*arr)*itemsize + sizeof(int)*2) : 0, itemsize * m + sizeof(int)*2);
+   STBIW_ASSERT(p);
+   if (p) {
+      if (!*arr) ((int *) p)[1] = 0;
+      *arr = (void *) ((int *) p + 2);
+      stbiw__sbm(*arr) = m;
+   }
+   return *arr;
+}
+
+static unsigned char *stbiw__zlib_flushf(unsigned char *data, unsigned int *bitbuffer, int *bitcount)
+{
+   while (*bitcount >= 8) {
+      stbiw__sbpush(data, STBIW_UCHAR(*bitbuffer));
+      *bitbuffer >>= 8;
+      *bitcount -= 8;
+   }
+   return data;
+}
+
+static int stbiw__zlib_bitrev(int code, int codebits)
+{
+   int res=0;
+   while (codebits--) {
+      res = (res << 1) | (code & 1);
+      code >>= 1;
+   }
+   return res;
+}
+
+static unsigned int stbiw__zlib_countm(unsigned char *a, unsigned char *b, int limit)
+{
+   int i;
+   for (i=0; i < limit && i < 258; ++i)
+      if (a[i] != b[i]) break;
+   return i;
+}
+
+static unsigned int stbiw__zhash(unsigned char *data)
+{
+   stbiw_uint32 hash = data[0] + (data[1] << 8) + (data[2] << 16);
+   hash ^= hash << 3;
+   hash += hash >> 5;
+   hash ^= hash << 4;
+   hash += hash >> 17;
+   hash ^= hash << 25;
+   hash += hash >> 6;
+   return hash;
+}
+
+#define stbiw__zlib_flush() (out = stbiw__zlib_flushf(out, &bitbuf, &bitcount))
+#define stbiw__zlib_add(code,codebits) \
+      (bitbuf |= (code) << bitcount, bitcount += (codebits), stbiw__zlib_flush())
+#define stbiw__zlib_huffa(b,c)  stbiw__zlib_add(stbiw__zlib_bitrev(b,c),c)
+// default huffman tables
+#define stbiw__zlib_huff1(n)  stbiw__zlib_huffa(0x30 + (n), 8)
+#define stbiw__zlib_huff2(n)  stbiw__zlib_huffa(0x190 + (n)-144, 9)
+#define stbiw__zlib_huff3(n)  stbiw__zlib_huffa(0 + (n)-256,7)
+#define stbiw__zlib_huff4(n)  stbiw__zlib_huffa(0xc0 + (n)-280,8)
+#define stbiw__zlib_huff(n)  ((n) <= 143 ? stbiw__zlib_huff1(n) : (n) <= 255 ? stbiw__zlib_huff2(n) : (n) <= 279 ? stbiw__zlib_huff3(n) : stbiw__zlib_huff4(n))
+#define stbiw__zlib_huffb(n) ((n) <= 143 ? stbiw__zlib_huff1(n) : stbiw__zlib_huff2(n))
+
+#define stbiw__ZHASH   16384
+
+#endif // STBIW_ZLIB_COMPRESS
+
+STBIWDEF unsigned char * stbi_zlib_compress(unsigned char *data, int data_len, int *out_len, int quality)
+{
+#ifdef STBIW_ZLIB_COMPRESS
+   // user provided a zlib compress implementation, use that
+   return STBIW_ZLIB_COMPRESS(data, data_len, out_len, quality);
+#else // use builtin
+   static unsigned short lengthc[] = { 3,4,5,6,7,8,9,10,11,13,15,17,19,23,27,31,35,43,51,59,67,83,99,115,131,163,195,227,258, 259 };
+   static unsigned char  lengtheb[]= { 0,0,0,0,0,0,0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,  4,  5,  5,  5,  5,  0 };
+   static unsigned short distc[]   = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577, 32768 };
+   static unsigned char  disteb[]  = { 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13 };
+   unsigned int bitbuf=0;
+   int i,j, bitcount=0;
+   unsigned char *out = NULL;
+   unsigned char ***hash_table = (unsigned char***) STBIW_MALLOC(stbiw__ZHASH * sizeof(unsigned char**));
+   if (hash_table == NULL)
+      return NULL;
+   if (quality < 5) quality = 5;
+
+   stbiw__sbpush(out, 0x78);   // DEFLATE 32K window
+   stbiw__sbpush(out, 0x5e);   // FLEVEL = 1
+   stbiw__zlib_add(1,1);  // BFINAL = 1
+   stbiw__zlib_add(1,2);  // BTYPE = 1 -- fixed huffman
+
+   for (i=0; i < stbiw__ZHASH; ++i)
+      hash_table[i] = NULL;
+
+   i=0;
+   while (i < data_len-3) {
+      // hash next 3 bytes of data to be compressed
+      int h = stbiw__zhash(data+i)&(stbiw__ZHASH-1), best=3;
+      unsigned char *bestloc = 0;
+      unsigned char **hlist = hash_table[h];
+      int n = stbiw__sbcount(hlist);
+      for (j=0; j < n; ++j) {
+         if (hlist[j]-data > i-32768) { // if entry lies within window
+            int d = stbiw__zlib_countm(hlist[j], data+i, data_len-i);
+            if (d >= best) { best=d; bestloc=hlist[j]; }
+         }
+      }
+      // when hash table entry is too long, delete half the entries
+      if (hash_table[h] && stbiw__sbn(hash_table[h]) == 2*quality) {
+         STBIW_MEMMOVE(hash_table[h], hash_table[h]+quality, sizeof(hash_table[h][0])*quality);
+         stbiw__sbn(hash_table[h]) = quality;
+      }
+      stbiw__sbpush(hash_table[h],data+i);
+
+      if (bestloc) {
+         // "lazy matching" - check match at *next* byte, and if it's better, do cur byte as literal
+         h = stbiw__zhash(data+i+1)&(stbiw__ZHASH-1);
+         hlist = hash_table[h];
+         n = stbiw__sbcount(hlist);
+         for (j=0; j < n; ++j) {
+            if (hlist[j]-data > i-32767) {
+               int e = stbiw__zlib_countm(hlist[j], data+i+1, data_len-i-1);
+               if (e > best) { // if next match is better, bail on current match
+                  bestloc = NULL;
+                  break;
+               }
+            }
+         }
+      }
+
+      if (bestloc) {
+         int d = (int) (data+i - bestloc); // distance back
+         STBIW_ASSERT(d <= 32767 && best <= 258);
+         for (j=0; best > lengthc[j+1]-1; ++j);
+         stbiw__zlib_huff(j+257);
+         if (lengtheb[j]) stbiw__zlib_add(best - lengthc[j], lengtheb[j]);
+         for (j=0; d > distc[j+1]-1; ++j);
+         stbiw__zlib_add(stbiw__zlib_bitrev(j,5),5);
+         if (disteb[j]) stbiw__zlib_add(d - distc[j], disteb[j]);
+         i += best;
+      } else {
+         stbiw__zlib_huffb(data[i]);
+         ++i;
+      }
+   }
+   // write out final bytes
+   for (;i < data_len; ++i)
+      stbiw__zlib_huffb(data[i]);
+   stbiw__zlib_huff(256); // end of block
+   // pad with 0 bits to byte boundary
+   while (bitcount)
+      stbiw__zlib_add(0,1);
+
+   for (i=0; i < stbiw__ZHASH; ++i)
+      (void) stbiw__sbfree(hash_table[i]);
+   STBIW_FREE(hash_table);
+
+   // store uncompressed instead if compression was worse
+   if (stbiw__sbn(out) > data_len + 2 + ((data_len+32766)/32767)*5) {
+      stbiw__sbn(out) = 2;  // truncate to DEFLATE 32K window and FLEVEL = 1
+      for (j = 0; j < data_len;) {
+         int blocklen = data_len - j;
+         if (blocklen > 32767) blocklen = 32767;
+         stbiw__sbpush(out, data_len - j == blocklen); // BFINAL = ?, BTYPE = 0 -- no compression
+         stbiw__sbpush(out, STBIW_UCHAR(blocklen)); // LEN
+         stbiw__sbpush(out, STBIW_UCHAR(blocklen >> 8));
+         stbiw__sbpush(out, STBIW_UCHAR(~blocklen)); // NLEN
+         stbiw__sbpush(out, STBIW_UCHAR(~blocklen >> 8));
+         memcpy(out+stbiw__sbn(out), data+j, blocklen);
+         stbiw__sbn(out) += blocklen;
+         j += blocklen;
+      }
+   }
+
+   {
+      // compute adler32 on input
+      unsigned int s1=1, s2=0;
+      int blocklen = (int) (data_len % 5552);
+      j=0;
+      while (j < data_len) {
+         for (i=0; i < blocklen; ++i) { s1 += data[j+i]; s2 += s1; }
+         s1 %= 65521; s2 %= 65521;
+         j += blocklen;
+         blocklen = 5552;
+      }
+      stbiw__sbpush(out, STBIW_UCHAR(s2 >> 8));
+      stbiw__sbpush(out, STBIW_UCHAR(s2));
+      stbiw__sbpush(out, STBIW_UCHAR(s1 >> 8));
+      stbiw__sbpush(out, STBIW_UCHAR(s1));
+   }
+   *out_len = stbiw__sbn(out);
+   // make returned pointer freeable
+   STBIW_MEMMOVE(stbiw__sbraw(out), out, *out_len);
+   return (unsigned char *) stbiw__sbraw(out);
+#endif // STBIW_ZLIB_COMPRESS
+}
+
+static unsigned int stbiw__crc32(unsigned char *buffer, int len)
+{
+#ifdef STBIW_CRC32
+    return STBIW_CRC32(buffer, len);
+#else
+   static unsigned int crc_table[256] =
+   {
+      0x00000000, 0x77073096, 0xEE0E612C, 0x990951BA, 0x076DC419, 0x706AF48F, 0xE963A535, 0x9E6495A3,
+      0x0eDB8832, 0x79DCB8A4, 0xE0D5E91E, 0x97D2D988, 0x09B64C2B, 0x7EB17CBD, 0xE7B82D07, 0x90BF1D91,
+      0x1DB71064, 0x6AB020F2, 0xF3B97148, 0x84BE41DE, 0x1ADAD47D, 0x6DDDE4EB, 0xF4D4B551, 0x83D385C7,
+      0x136C9856, 0x646BA8C0, 0xFD62F97A, 0x8A65C9EC, 0x14015C4F, 0x63066CD9, 0xFA0F3D63, 0x8D080DF5,
+      0x3B6E20C8, 0x4C69105E, 0xD56041E4, 0xA2677172, 0x3C03E4D1, 0x4B04D447, 0xD20D85FD, 0xA50AB56B,
+      0x35B5A8FA, 0x42B2986C, 0xDBBBC9D6, 0xACBCF940, 0x32D86CE3, 0x45DF5C75, 0xDCD60DCF, 0xABD13D59,
+      0x26D930AC, 0x51DE003A, 0xC8D75180, 0xBFD06116, 0x21B4F4B5, 0x56B3C423, 0xCFBA9599, 0xB8BDA50F,
+      0x2802B89E, 0x5F058808, 0xC60CD9B2, 0xB10BE924, 0x2F6F7C87, 0x58684C11, 0xC1611DAB, 0xB6662D3D,
+      0x76DC4190, 0x01DB7106, 0x98D220BC, 0xEFD5102A, 0x71B18589, 0x06B6B51F, 0x9FBFE4A5, 0xE8B8D433,
+      0x7807C9A2, 0x0F00F934, 0x9609A88E, 0xE10E9818, 0x7F6A0DBB, 0x086D3D2D, 0x91646C97, 0xE6635C01,
+      0x6B6B51F4, 0x1C6C6162, 0x856530D8, 0xF262004E, 0x6C0695ED, 0x1B01A57B, 0x8208F4C1, 0xF50FC457,
+      0x65B0D9C6, 0x12B7E950, 0x8BBEB8EA, 0xFCB9887C, 0x62DD1DDF, 0x15DA2D49, 0x8CD37CF3, 0xFBD44C65,
+      0x4DB26158, 0x3AB551CE, 0xA3BC0074, 0xD4BB30E2, 0x4ADFA541, 0x3DD895D7, 0xA4D1C46D, 0xD3D6F4FB,
+      0x4369E96A, 0x346ED9FC, 0xAD678846, 0xDA60B8D0, 0x44042D73, 0x33031DE5, 0xAA0A4C5F, 0xDD0D7CC9,
+      0x5005713C, 0x270241AA, 0xBE0B1010, 0xC90C2086, 0x5768B525, 0x206F85B3, 0xB966D409, 0xCE61E49F,
+      0x5EDEF90E, 0x29D9C998, 0xB0D09822, 0xC7D7A8B4, 0x59B33D17, 0x2EB40D81, 0xB7BD5C3B, 0xC0BA6CAD,
+      0xEDB88320, 0x9ABFB3B6, 0x03B6E20C, 0x74B1D29A, 0xEAD54739, 0x9DD277AF, 0x04DB2615, 0x73DC1683,
+      0xE3630B12, 0x94643B84, 0x0D6D6A3E, 0x7A6A5AA8, 0xE40ECF0B, 0x9309FF9D, 0x0A00AE27, 0x7D079EB1,
+      0xF00F9344, 0x8708A3D2, 0x1E01F268, 0x6906C2FE, 0xF762575D, 0x806567CB, 0x196C3671, 0x6E6B06E7,
+      0xFED41B76, 0x89D32BE0, 0x10DA7A5A, 0x67DD4ACC, 0xF9B9DF6F, 0x8EBEEFF9, 0x17B7BE43, 0x60B08ED5,
+      0xD6D6A3E8, 0xA1D1937E, 0x38D8C2C4, 0x4FDFF252, 0xD1BB67F1, 0xA6BC5767, 0x3FB506DD, 0x48B2364B,
+      0xD80D2BDA, 0xAF0A1B4C, 0x36034AF6, 0x41047A60, 0xDF60EFC3, 0xA867DF55, 0x316E8EEF, 0x4669BE79,
+      0xCB61B38C, 0xBC66831A, 0x256FD2A0, 0x5268E236, 0xCC0C7795, 0xBB0B4703, 0x220216B9, 0x5505262F,
+      0xC5BA3BBE, 0xB2BD0B28, 0x2BB45A92, 0x5CB36A04, 0xC2D7FFA7, 0xB5D0CF31, 0x2CD99E8B, 0x5BDEAE1D,
+      0x9B64C2B0, 0xEC63F226, 0x756AA39C, 0x026D930A, 0x9C0906A9, 0xEB0E363F, 0x72076785, 0x05005713,
+      0x95BF4A82, 0xE2B87A14, 0x7BB12BAE, 0x0CB61B38, 0x92D28E9B, 0xE5D5BE0D, 0x7CDCEFB7, 0x0BDBDF21,
+      0x86D3D2D4, 0xF1D4E242, 0x68DDB3F8, 0x1FDA836E, 0x81BE16CD, 0xF6B9265B, 0x6FB077E1, 0x18B74777,
+      0x88085AE6, 0xFF0F6A70, 0x66063BCA, 0x11010B5C, 0x8F659EFF, 0xF862AE69, 0x616BFFD3, 0x166CCF45,
+      0xA00AE278, 0xD70DD2EE, 0x4E048354, 0x3903B3C2, 0xA7672661, 0xD06016F7, 0x4969474D, 0x3E6E77DB,
+      0xAED16A4A, 0xD9D65ADC, 0x40DF0B66, 0x37D83BF0, 0xA9BCAE53, 0xDEBB9EC5, 0x47B2CF7F, 0x30B5FFE9,
+      0xBDBDF21C, 0xCABAC28A, 0x53B39330, 0x24B4A3A6, 0xBAD03605, 0xCDD70693, 0x54DE5729, 0x23D967BF,
+      0xB3667A2E, 0xC4614AB8, 0x5D681B02, 0x2A6F2B94, 0xB40BBE37, 0xC30C8EA1, 0x5A05DF1B, 0x2D02EF8D
+   };
+
+   unsigned int crc = ~0u;
+   int i;
+   for (i=0; i < len; ++i)
+      crc = (crc >> 8) ^ crc_table[buffer[i] ^ (crc & 0xff)];
+   return ~crc;
+#endif
+}
+
+#define stbiw__wpng4(o,a,b,c,d) ((o)[0]=STBIW_UCHAR(a),(o)[1]=STBIW_UCHAR(b),(o)[2]=STBIW_UCHAR(c),(o)[3]=STBIW_UCHAR(d),(o)+=4)
+#define stbiw__wp32(data,v) stbiw__wpng4(data, (v)>>24,(v)>>16,(v)>>8,(v));
+#define stbiw__wptag(data,s) stbiw__wpng4(data, s[0],s[1],s[2],s[3])
+
+static void stbiw__wpcrc(unsigned char **data, int len)
+{
+   unsigned int crc = stbiw__crc32(*data - len - 4, len+4);
+   stbiw__wp32(*data, crc);
+}
+
+static unsigned char stbiw__paeth(int a, int b, int c)
+{
+   int p = a + b - c, pa = abs(p-a), pb = abs(p-b), pc = abs(p-c);
+   if (pa <= pb && pa <= pc) return STBIW_UCHAR(a);
+   if (pb <= pc) return STBIW_UCHAR(b);
+   return STBIW_UCHAR(c);
+}
+
+// @OPTIMIZE: provide an option that always forces left-predict or paeth predict
+static void stbiw__encode_png_line(unsigned char *pixels, int stride_bytes, int width, int height, int y, int n, int filter_type, signed char *line_buffer)
+{
+   static int mapping[] = { 0,1,2,3,4 };
+   static int firstmap[] = { 0,1,0,5,6 };
+   int *mymap = (y != 0) ? mapping : firstmap;
+   int i;
+   int type = mymap[filter_type];
+   unsigned char *z = pixels + stride_bytes * (stbi__flip_vertically_on_write ? height-1-y : y);
+   int signed_stride = stbi__flip_vertically_on_write ? -stride_bytes : stride_bytes;
+
+   if (type==0) {
+      memcpy(line_buffer, z, width*n);
+      return;
+   }
+
+   // first loop isn't optimized since it's just one pixel
+   for (i = 0; i < n; ++i) {
+      switch (type) {
+         case 1: line_buffer[i] = z[i]; break;
+         case 2: line_buffer[i] = z[i] - z[i-signed_stride]; break;
+         case 3: line_buffer[i] = z[i] - (z[i-signed_stride]>>1); break;
+         case 4: line_buffer[i] = (signed char) (z[i] - stbiw__paeth(0,z[i-signed_stride],0)); break;
+         case 5: line_buffer[i] = z[i]; break;
+         case 6: line_buffer[i] = z[i]; break;
+      }
+   }
+   switch (type) {
+      case 1: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-n]; break;
+      case 2: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-signed_stride]; break;
+      case 3: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - ((z[i-n] + z[i-signed_stride])>>1); break;
+      case 4: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], z[i-signed_stride], z[i-signed_stride-n]); break;
+      case 5: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - (z[i-n]>>1); break;
+      case 6: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], 0,0); break;
+   }
+}
+
+STBIWDEF unsigned char *stbi_write_png_to_mem(const unsigned char *pixels, int stride_bytes, int x, int y, int n, int *out_len)
+{
+   int force_filter = stbi_write_force_png_filter;
+   int ctype[5] = { -1, 0, 4, 2, 6 };
+   unsigned char sig[8] = { 137,80,78,71,13,10,26,10 };
+   unsigned char *out,*o, *filt, *zlib;
+   signed char *line_buffer;
+   int j,zlen;
+
+   if (stride_bytes == 0)
+      stride_bytes = x * n;
+
+   if (force_filter >= 5) {
+      force_filter = -1;
+   }
+
+   filt = (unsigned char *) STBIW_MALLOC((x*n+1) * y); if (!filt) return 0;
+   line_buffer = (signed char *) STBIW_MALLOC(x * n); if (!line_buffer) { STBIW_FREE(filt); return 0; }
+   for (j=0; j < y; ++j) {
+      int filter_type;
+      if (force_filter > -1) {
+         filter_type = force_filter;
+         stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, force_filter, line_buffer);
+      } else { // Estimate the best filter by running through all of them:
+         int best_filter = 0, best_filter_val = 0x7fffffff, est, i;
+         for (filter_type = 0; filter_type < 5; filter_type++) {
+            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, filter_type, line_buffer);
+
+            // Estimate the entropy of the line using this filter; the less, the better.
+            est = 0;
+            for (i = 0; i < x*n; ++i) {
+               est += abs((signed char) line_buffer[i]);
+            }
+            if (est < best_filter_val) {
+               best_filter_val = est;
+               best_filter = filter_type;
+            }
+         }
+         if (filter_type != best_filter) {  // If the last iteration already got us the best filter, don't redo it
+            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, best_filter, line_buffer);
+            filter_type = best_filter;
+         }
+      }
+      // when we get here, filter_type contains the filter type, and line_buffer contains the data
+      filt[j*(x*n+1)] = (unsigned char) filter_type;
+      STBIW_MEMMOVE(filt+j*(x*n+1)+1, line_buffer, x*n);
+   }
+   STBIW_FREE(line_buffer);
+   zlib = stbi_zlib_compress(filt, y*( x*n+1), &zlen, stbi_write_png_compression_level);
+   STBIW_FREE(filt);
+   if (!zlib) return 0;
+
+   // each tag requires 12 bytes of overhead
+   out = (unsigned char *) STBIW_MALLOC(8 + 12+13 + 12+zlen + 12);
+   if (!out) return 0;
+   *out_len = 8 + 12+13 + 12+zlen + 12;
+
+   o=out;
+   STBIW_MEMMOVE(o,sig,8); o+= 8;
+   stbiw__wp32(o, 13); // header length
+   stbiw__wptag(o, "IHDR");
+   stbiw__wp32(o, x);
+   stbiw__wp32(o, y);
+   *o++ = 8;
+   *o++ = STBIW_UCHAR(ctype[n]);
+   *o++ = 0;
+   *o++ = 0;
+   *o++ = 0;
+   stbiw__wpcrc(&o,13);
+
+   stbiw__wp32(o, zlen);
+   stbiw__wptag(o, "IDAT");
+   STBIW_MEMMOVE(o, zlib, zlen);
+   o += zlen;
+   STBIW_FREE(zlib);
+   stbiw__wpcrc(&o, zlen);
+
+   stbiw__wp32(o,0);
+   stbiw__wptag(o, "IEND");
+   stbiw__wpcrc(&o,0);
+
+   STBIW_ASSERT(o == out + *out_len);
+
+   return out;
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_png(char const *filename, int x, int y, int comp, const void *data, int stride_bytes)
+{
+   FILE *f;
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len);
+   if (png == NULL) return 0;
+
+   f = stbiw__fopen(filename, "wb");
+   if (!f) { STBIW_FREE(png); return 0; }
+   fwrite(png, 1, len, f);
+   fclose(f);
+   STBIW_FREE(png);
+   return 1;
+}
+#endif
+
+STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int stride_bytes)
+{
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len);
+   if (png == NULL) return 0;
+   func(context, png, len);
+   STBIW_FREE(png);
+   return 1;
+}
+
+
+/* ***************************************************************************
+ *
+ * JPEG writer
+ *
+ * This is based on Jon Olick's jo_jpeg.cpp:
+ * public domain Simple, Minimalistic JPEG writer - http://www.jonolick.com/code.html
+ */
+
+static const unsigned char stbiw__jpg_ZigZag[] = { 0,1,5,6,14,15,27,28,2,4,7,13,16,26,29,42,3,8,12,17,25,30,41,43,9,11,18,
+      24,31,40,44,53,10,19,23,32,39,45,52,54,20,22,33,38,46,51,55,60,21,34,37,47,50,56,59,61,35,36,48,49,57,58,62,63 };
+
+static void stbiw__jpg_writeBits(stbi__write_context *s, int *bitBufP, int *bitCntP, const unsigned short *bs) {
+   int bitBuf = *bitBufP, bitCnt = *bitCntP;
+   bitCnt += bs[1];
+   bitBuf |= bs[0] << (24 - bitCnt);
+   while(bitCnt >= 8) {
+      unsigned char c = (bitBuf >> 16) & 255;
+      stbiw__putc(s, c);
+      if(c == 255) {
+         stbiw__putc(s, 0);
+      }
+      bitBuf <<= 8;
+      bitCnt -= 8;
+   }
+   *bitBufP = bitBuf;
+   *bitCntP = bitCnt;
+}
+
+static void stbiw__jpg_DCT(float *d0p, float *d1p, float *d2p, float *d3p, float *d4p, float *d5p, float *d6p, float *d7p) {
+   float d0 = *d0p, d1 = *d1p, d2 = *d2p, d3 = *d3p, d4 = *d4p, d5 = *d5p, d6 = *d6p, d7 = *d7p;
+   float z1, z2, z3, z4, z5, z11, z13;
+
+   float tmp0 = d0 + d7;
+   float tmp7 = d0 - d7;
+   float tmp1 = d1 + d6;
+   float tmp6 = d1 - d6;
+   float tmp2 = d2 + d5;
+   float tmp5 = d2 - d5;
+   float tmp3 = d3 + d4;
+   float tmp4 = d3 - d4;
+
+   // Even part
+   float tmp10 = tmp0 + tmp3;   // phase 2
+   float tmp13 = tmp0 - tmp3;
+   float tmp11 = tmp1 + tmp2;
+   float tmp12 = tmp1 - tmp2;
+
+   d0 = tmp10 + tmp11;       // phase 3
+   d4 = tmp10 - tmp11;
+
+   z1 = (tmp12 + tmp13) * 0.707106781f; // c4
+   d2 = tmp13 + z1;       // phase 5
+   d6 = tmp13 - z1;
+
+   // Odd part
+   tmp10 = tmp4 + tmp5;       // phase 2
+   tmp11 = tmp5 + tmp6;
+   tmp12 = tmp6 + tmp7;
+
+   // The rotator is modified from fig 4-8 to avoid extra negations.
+   z5 = (tmp10 - tmp12) * 0.382683433f; // c6
+   z2 = tmp10 * 0.541196100f + z5; // c2-c6
+   z4 = tmp12 * 1.306562965f + z5; // c2+c6
+   z3 = tmp11 * 0.707106781f; // c4
+
+   z11 = tmp7 + z3;      // phase 5
+   z13 = tmp7 - z3;
+
+   *d5p = z13 + z2;         // phase 6
+   *d3p = z13 - z2;
+   *d1p = z11 + z4;
+   *d7p = z11 - z4;
+
+   *d0p = d0;  *d2p = d2;  *d4p = d4;  *d6p = d6;
+}
+
+static void stbiw__jpg_calcBits(int val, unsigned short bits[2]) {
+   int tmp1 = val < 0 ? -val : val;
+   val = val < 0 ? val-1 : val;
+   bits[1] = 1;
+   while(tmp1 >>= 1) {
+      ++bits[1];
+   }
+   bits[0] = val & ((1<<bits[1])-1);
+}
+
+static int stbiw__jpg_processDU(stbi__write_context *s, int *bitBuf, int *bitCnt, float *CDU, int du_stride, float *fdtbl, int DC, const unsigned short HTDC[256][2], const unsigned short HTAC[256][2]) {
+   const unsigned short EOB[2] = { HTAC[0x00][0], HTAC[0x00][1] };
+   const unsigned short M16zeroes[2] = { HTAC[0xF0][0], HTAC[0xF0][1] };
+   int dataOff, i, j, n, diff, end0pos, x, y;
+   int DU[64];
+
+   // DCT rows
+   for(dataOff=0, n=du_stride*8; dataOff<n; dataOff+=du_stride) {
+      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+1], &CDU[dataOff+2], &CDU[dataOff+3], &CDU[dataOff+4], &CDU[dataOff+5], &CDU[dataOff+6], &CDU[dataOff+7]);
+   }
+   // DCT columns
+   for(dataOff=0; dataOff<8; ++dataOff) {
+      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+du_stride], &CDU[dataOff+du_stride*2], &CDU[dataOff+du_stride*3], &CDU[dataOff+du_stride*4],
+                     &CDU[dataOff+du_stride*5], &CDU[dataOff+du_stride*6], &CDU[dataOff+du_stride*7]);
+   }
+   // Quantize/descale/zigzag the coefficients
+   for(y = 0, j=0; y < 8; ++y) {
+      for(x = 0; x < 8; ++x,++j) {
+         float v;
+         i = y*du_stride+x;
+         v = CDU[i]*fdtbl[j];
+         // DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? ceilf(v - 0.5f) : floorf(v + 0.5f));
+         // ceilf() and floorf() are C99, not C89, but I /think/ they're not needed here anyway?
+         DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? v - 0.5f : v + 0.5f);
+      }
+   }
+
+   // Encode DC
+   diff = DU[0] - DC;
+   if (diff == 0) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[0]);
+   } else {
+      unsigned short bits[2];
+      stbiw__jpg_calcBits(diff, bits);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[bits[1]]);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
+   }
+   // Encode ACs
+   end0pos = 63;
+   for(; (end0pos>0)&&(DU[end0pos]==0); --end0pos) {
+   }
+   // end0pos = first element in reverse order !=0
+   if(end0pos == 0) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
+      return DU[0];
+   }
+   for(i = 1; i <= end0pos; ++i) {
+      int startpos = i;
+      int nrzeroes;
+      unsigned short bits[2];
+      for (; DU[i]==0 && i<=end0pos; ++i) {
+      }
+      nrzeroes = i-startpos;
+      if ( nrzeroes >= 16 ) {
+         int lng = nrzeroes>>4;
+         int nrmarker;
+         for (nrmarker=1; nrmarker <= lng; ++nrmarker)
+            stbiw__jpg_writeBits(s, bitBuf, bitCnt, M16zeroes);
+         nrzeroes &= 15;
+      }
+      stbiw__jpg_calcBits(DU[i], bits);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTAC[(nrzeroes<<4)+bits[1]]);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
+   }
+   if(end0pos != 63) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
+   }
+   return DU[0];
+}
+
+static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality) {
+   // Constants that don't pollute global namespace
+   static const unsigned char std_dc_luminance_nrcodes[] = {0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0};
+   static const unsigned char std_dc_luminance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
+   static const unsigned char std_ac_luminance_nrcodes[] = {0,0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d};
+   static const unsigned char std_ac_luminance_values[] = {
+      0x01,0x02,0x03,0x00,0x04,0x11,0x05,0x12,0x21,0x31,0x41,0x06,0x13,0x51,0x61,0x07,0x22,0x71,0x14,0x32,0x81,0x91,0xa1,0x08,
+      0x23,0x42,0xb1,0xc1,0x15,0x52,0xd1,0xf0,0x24,0x33,0x62,0x72,0x82,0x09,0x0a,0x16,0x17,0x18,0x19,0x1a,0x25,0x26,0x27,0x28,
+      0x29,0x2a,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59,
+      0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x83,0x84,0x85,0x86,0x87,0x88,0x89,
+      0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6,
+      0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe1,0xe2,
+      0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
+   };
+   static const unsigned char std_dc_chrominance_nrcodes[] = {0,0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0};
+   static const unsigned char std_dc_chrominance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
+   static const unsigned char std_ac_chrominance_nrcodes[] = {0,0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77};
+   static const unsigned char std_ac_chrominance_values[] = {
+      0x00,0x01,0x02,0x03,0x11,0x04,0x05,0x21,0x31,0x06,0x12,0x41,0x51,0x07,0x61,0x71,0x13,0x22,0x32,0x81,0x08,0x14,0x42,0x91,
+      0xa1,0xb1,0xc1,0x09,0x23,0x33,0x52,0xf0,0x15,0x62,0x72,0xd1,0x0a,0x16,0x24,0x34,0xe1,0x25,0xf1,0x17,0x18,0x19,0x1a,0x26,
+      0x27,0x28,0x29,0x2a,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,
+      0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x82,0x83,0x84,0x85,0x86,0x87,
+      0x88,0x89,0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,
+      0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,
+      0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
+   };
+   // Huffman tables
+   static const unsigned short YDC_HT[256][2] = { {0,2},{2,3},{3,3},{4,3},{5,3},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9}};
+   static const unsigned short UVDC_HT[256][2] = { {0,2},{1,2},{2,2},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9},{1022,10},{2046,11}};
+   static const unsigned short YAC_HT[256][2] = {
+      {10,4},{0,2},{1,2},{4,3},{11,4},{26,5},{120,7},{248,8},{1014,10},{65410,16},{65411,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {12,4},{27,5},{121,7},{502,9},{2038,11},{65412,16},{65413,16},{65414,16},{65415,16},{65416,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {28,5},{249,8},{1015,10},{4084,12},{65417,16},{65418,16},{65419,16},{65420,16},{65421,16},{65422,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {58,6},{503,9},{4085,12},{65423,16},{65424,16},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {59,6},{1016,10},{65430,16},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {122,7},{2039,11},{65438,16},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {123,7},{4086,12},{65446,16},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {250,8},{4087,12},{65454,16},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {504,9},{32704,15},{65462,16},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {505,9},{65470,16},{65471,16},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {506,9},{65479,16},{65480,16},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1017,10},{65488,16},{65489,16},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1018,10},{65497,16},{65498,16},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2040,11},{65506,16},{65507,16},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {65515,16},{65516,16},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2041,11},{65525,16},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
+   };
+   static const unsigned short UVAC_HT[256][2] = {
+      {0,2},{1,2},{4,3},{10,4},{24,5},{25,5},{56,6},{120,7},{500,9},{1014,10},{4084,12},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {11,4},{57,6},{246,8},{501,9},{2038,11},{4085,12},{65416,16},{65417,16},{65418,16},{65419,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {26,5},{247,8},{1015,10},{4086,12},{32706,15},{65420,16},{65421,16},{65422,16},{65423,16},{65424,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {27,5},{248,8},{1016,10},{4087,12},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{65430,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {58,6},{502,9},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{65438,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {59,6},{1017,10},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{65446,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {121,7},{2039,11},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{65454,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {122,7},{2040,11},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{65462,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {249,8},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{65470,16},{65471,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {503,9},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{65479,16},{65480,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {504,9},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{65488,16},{65489,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {505,9},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{65497,16},{65498,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {506,9},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{65506,16},{65507,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2041,11},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{65515,16},{65516,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {16352,14},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{65525,16},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1018,10},{32707,15},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
+   };
+   static const int YQT[] = {16,11,10,16,24,40,51,61,12,12,14,19,26,58,60,55,14,13,16,24,40,57,69,56,14,17,22,29,51,87,80,62,18,22,
+                             37,56,68,109,103,77,24,35,55,64,81,104,113,92,49,64,78,87,103,121,120,101,72,92,95,98,112,100,103,99};
+   static const int UVQT[] = {17,18,24,47,99,99,99,99,18,21,26,66,99,99,99,99,24,26,56,99,99,99,99,99,47,66,99,99,99,99,99,99,
+                              99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99};
+   static const float aasf[] = { 1.0f * 2.828427125f, 1.387039845f * 2.828427125f, 1.306562965f * 2.828427125f, 1.175875602f * 2.828427125f,
+                                 1.0f * 2.828427125f, 0.785694958f * 2.828427125f, 0.541196100f * 2.828427125f, 0.275899379f * 2.828427125f };
+
+   int row, col, i, k, subsample;
+   float fdtbl_Y[64], fdtbl_UV[64];
+   unsigned char YTable[64], UVTable[64];
+
+   if(!data || !width || !height || comp > 4 || comp < 1) {
+      return 0;
+   }
+
+   quality = quality ? quality : 90;
+   subsample = quality <= 90 ? 1 : 0;
+   quality = quality < 1 ? 1 : quality > 100 ? 100 : quality;
+   quality = quality < 50 ? 5000 / quality : 200 - quality * 2;
+
+   for(i = 0; i < 64; ++i) {
+      int uvti, yti = (YQT[i]*quality+50)/100;
+      YTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (yti < 1 ? 1 : yti > 255 ? 255 : yti);
+      uvti = (UVQT[i]*quality+50)/100;
+      UVTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (uvti < 1 ? 1 : uvti > 255 ? 255 : uvti);
+   }
+
+   for(row = 0, k = 0; row < 8; ++row) {
+      for(col = 0; col < 8; ++col, ++k) {
+         fdtbl_Y[k]  = 1 / (YTable [stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
+         fdtbl_UV[k] = 1 / (UVTable[stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
+      }
+   }
+
+   // Write Headers
+   {
+      static const unsigned char head0[] = { 0xFF,0xD8,0xFF,0xE0,0,0x10,'J','F','I','F',0,1,1,0,0,1,0,1,0,0,0xFF,0xDB,0,0x84,0 };
+      static const unsigned char head2[] = { 0xFF,0xDA,0,0xC,3,1,0,2,0x11,3,0x11,0,0x3F,0 };
+      const unsigned char head1[] = { 0xFF,0xC0,0,0x11,8,(unsigned char)(height>>8),STBIW_UCHAR(height),(unsigned char)(width>>8),STBIW_UCHAR(width),
+                                      3,1,(unsigned char)(subsample?0x22:0x11),0,2,0x11,1,3,0x11,1,0xFF,0xC4,0x01,0xA2,0 };
+      s->func(s->context, (void*)head0, sizeof(head0));
+      s->func(s->context, (void*)YTable, sizeof(YTable));
+      stbiw__putc(s, 1);
+      s->func(s->context, UVTable, sizeof(UVTable));
+      s->func(s->context, (void*)head1, sizeof(head1));
+      s->func(s->context, (void*)(std_dc_luminance_nrcodes+1), sizeof(std_dc_luminance_nrcodes)-1);
+      s->func(s->context, (void*)std_dc_luminance_values, sizeof(std_dc_luminance_values));
+      stbiw__putc(s, 0x10); // HTYACinfo
+      s->func(s->context, (void*)(std_ac_luminance_nrcodes+1), sizeof(std_ac_luminance_nrcodes)-1);
+      s->func(s->context, (void*)std_ac_luminance_values, sizeof(std_ac_luminance_values));
+      stbiw__putc(s, 1); // HTUDCinfo
+      s->func(s->context, (void*)(std_dc_chrominance_nrcodes+1), sizeof(std_dc_chrominance_nrcodes)-1);
+      s->func(s->context, (void*)std_dc_chrominance_values, sizeof(std_dc_chrominance_values));
+      stbiw__putc(s, 0x11); // HTUACinfo
+      s->func(s->context, (void*)(std_ac_chrominance_nrcodes+1), sizeof(std_ac_chrominance_nrcodes)-1);
+      s->func(s->context, (void*)std_ac_chrominance_values, sizeof(std_ac_chrominance_values));
+      s->func(s->context, (void*)head2, sizeof(head2));
+   }
+
+   // Encode 8x8 macroblocks
+   {
+      static const unsigned short fillBits[] = {0x7F, 7};
+      int DCY=0, DCU=0, DCV=0;
+      int bitBuf=0, bitCnt=0;
+      // comp == 2 is grey+alpha (alpha is ignored)
+      int ofsG = comp > 2 ? 1 : 0, ofsB = comp > 2 ? 2 : 0;
+      const unsigned char *dataR = (const unsigned char *)data;
+      const unsigned char *dataG = dataR + ofsG;
+      const unsigned char *dataB = dataR + ofsB;
+      int x, y, pos;
+      if(subsample) {
+         for(y = 0; y < height; y += 16) {
+            for(x = 0; x < width; x += 16) {
+               float Y[256], U[256], V[256];
+               for(row = y, pos = 0; row < y+16; ++row) {
+                  // row >= height => use last input row
+                  int clamped_row = (row < height) ? row : height - 1;
+                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
+                  for(col = x; col < x+16; ++col, ++pos) {
+                     // if col >= width => use pixel from last input column
+                     int p = base_p + ((col < width) ? col : (width-1))*comp;
+                     float r = dataR[p], g = dataG[p], b = dataB[p];
+                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
+                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
+                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
+                  }
+               }
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+0,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+8,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+128, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+136, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+
+               // subsample U,V
+               {
+                  float subU[64], subV[64];
+                  int yy, xx;
+                  for(yy = 0, pos = 0; yy < 8; ++yy) {
+                     for(xx = 0; xx < 8; ++xx, ++pos) {
+                        int j = yy*32+xx*2;
+                        subU[pos] = (U[j+0] + U[j+1] + U[j+16] + U[j+17]) * 0.25f;
+                        subV[pos] = (V[j+0] + V[j+1] + V[j+16] + V[j+17]) * 0.25f;
+                     }
+                  }
+                  DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subU, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
+                  DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subV, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
+               }
+            }
+         }
+      } else {
+         for(y = 0; y < height; y += 8) {
+            for(x = 0; x < width; x += 8) {
+               float Y[64], U[64], V[64];
+               for(row = y, pos = 0; row < y+8; ++row) {
+                  // row >= height => use last input row
+                  int clamped_row = (row < height) ? row : height - 1;
+                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
+                  for(col = x; col < x+8; ++col, ++pos) {
+                     // if col >= width => use pixel from last input column
+                     int p = base_p + ((col < width) ? col : (width-1))*comp;
+                     float r = dataR[p], g = dataG[p], b = dataB[p];
+                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
+                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
+                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
+                  }
+               }
+
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y, 8, fdtbl_Y,  DCY, YDC_HT, YAC_HT);
+               DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, U, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
+               DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, V, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
+            }
+         }
+      }
+
+      // Do the bit alignment of the EOI marker
+      stbiw__jpg_writeBits(s, &bitBuf, &bitCnt, fillBits);
+   }
+
+   // EOI
+   stbiw__putc(s, 0xFF);
+   stbiw__putc(s, 0xD9);
+
+   return 1;
+}
+
+STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality);
+}
+
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_jpg_core(&s, x, y, comp, data, quality);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif
+
+#endif // STB_IMAGE_WRITE_IMPLEMENTATION
+
+/* Revision history
+      1.16  (2021-07-11)
+             make Deflate code emit uncompressed blocks when it would otherwise expand
+             support writing BMPs with alpha channel
+      1.15  (2020-07-13) unknown
+      1.14  (2020-02-02) updated JPEG writer to downsample chroma channels
+      1.13
+      1.12
+      1.11  (2019-08-11)
+
+      1.10  (2019-02-07)
+             support utf8 filenames in Windows; fix warnings and platform ifdefs
+      1.09  (2018-02-11)
+             fix typo in zlib quality API, improve STB_I_W_STATIC in C++
+      1.08  (2018-01-29)
+             add stbi__flip_vertically_on_write, external zlib, zlib quality, choose PNG filter
+      1.07  (2017-07-24)
+             doc fix
+      1.06 (2017-07-23)
+             writing JPEG (using Jon Olick's code)
+      1.05   ???
+      1.04 (2017-03-03)
+             monochrome BMP expansion
+      1.03   ???
+      1.02 (2016-04-02)
+             avoid allocating large structures on the stack
+      1.01 (2016-01-16)
+             STBIW_REALLOC_SIZED: support allocators with no realloc support
+             avoid race-condition in crc initialization
+             minor compile issues
+      1.00 (2015-09-14)
+             installable file IO function
+      0.99 (2015-09-13)
+             warning fixes; TGA rle support
+      0.98 (2015-04-08)
+             added STBIW_MALLOC, STBIW_ASSERT etc
+      0.97 (2015-01-18)
+             fixed HDR asserts, rewrote HDR rle logic
+      0.96 (2015-01-17)
+             add HDR output
+             fix monochrome BMP
+      0.95 (2014-08-17)
+             add monochrome TGA output
+      0.94 (2014-05-31)
+             rename private functions to avoid conflicts with stb_image.h
+      0.93 (2014-05-27)
+             warning fixes
+      0.92 (2010-08-01)
+             casts to unsigned char to fix warnings
+      0.91 (2010-07-17)
+             first public release
+      0.90   first internal release
+*/
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/
diff --git a/07_nlm_denoise/blackbook537/lab_report.md b/07_nlm_denoise/blackbook537/lab_report.md
new file mode 100644
index 00000000..54c8a44d
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/lab_report.md
@@ -0,0 +1,139 @@
+# CUDA 非局部均值降噪实验报告
+
+2026 夏季训练营 · 选题 07 · blackbook537
+
+本报告整理已有实验数据与实现，并按提交阅读顺序重写。跨平台性能来自重构前的存档，不代表本次结构调整后的全平台复测。每项数值均对应 assets/results/ 中的文件。
+
+## 1. 实现与算法约定
+
+程序读取灰度/RGB PNG 或 JPG，在 GPU 上完成 u8 interleaved 到 planar f32 的转换、NLM 加权计算和 u8 量化，再保存图像。输出格式由扩展名决定，使用时应与输入保持一致。主要计算由本项目 kernel 实现，stb 只承担图像编解码。
+
+默认参数 pr=3、sr=10、h=10、sigma=25；支持 pr=1–8、sr=1–32。RGB 的三个通道共享由跨通道 patch 距离计算的权重；边界访问采用 clamp，等价于复制边界。
+
+```text
+N_patch = (2*pr+1)^2 * channels
+dist = sum((P_p - P_q)^2)
+w(p,q) = exp(-max(dist - 2*sigma^2*N_patch, 0) / h^2)
+output(p) = sum(w(p,q)*input(q)) / sum(w(p,q))
+```
+
+题面示意公式的减项为 2*sigma²，而现有代码为 2*sigma²*N_patch。两种形式不能在固定 h 下直接视为等价。此次结构重构保留现有计算语义；需要按教师认可的 dist 定义进一步说明或修正。CPU/GPU 自洽验证只能证明双方采用相同计算，不能代替题意确认。
+
+## 2. 优化实现
+
+| 版本 | 实现 | 作用和限制 |
+|---|---|---|
+| V0 | 每线程处理一个像素 | 直接访问全局内存，作为基准 |
+| V1 | shared memory + halo tile | 复用相邻 patch 数据；行 stride 加 padding |
+| V2 | V1 + patch 模板展开 | pr≤3 使用展开；pr≥4 回退 V1 |
+
+线程块协作加载复制边界的 halo，再在 tile 内计算距离。共享内存容量不足时回退 V0。V2 的速度收益与平台有关：历史 C500 数据中，base 参数 V2 比 V1 慢约 3.56 倍，因此 Make 在 MetaX 上默认选择 V1。
+
+平台差异集中在 include/platform_api.h。src/kernels.cu 保存算法；MACA/MUSA 编译入口包含同一实现；Host 代码只保留一份。默认使用 expf，FAST_EXP=1 可选近似路径，新的近似误差结论需单独复测。
+
+<!-- pagebreak -->
+
+## 3. 正确性与图像质量
+
+src/self_check.cpp 保留参数解析、指标函数、CPU 常量图不变性、确定性等 12 项自检。validate 对 V0/V1/V2 分别与 CPU 参考计算 MAE/PSNR，验证失败返回非零退出码。本季度 NLM 专项要求程序包含测试，因此迁移后保留这些入口。
+
+历史 1080p RGB/base 验证记录中，三个 GPU 版本的 MAE=0、PSNR=inf。该结论仅限已测图像与配置；不能推广为任意输入或所有编译器均逐位一致。
+
+合成图使用平滑渐变背景和确定性的 LCG 均匀噪声。生成器参数 sigma 为噪声幅度 a，未量化、未裁剪时标准差约为 a/sqrt(3)。因此以下结果不能表述为“高斯标准差 25 的自然图像降噪性能”。
+
+| 图像/配置 | MAE 相对 clean | PSNR (dB) |
+|---|---:|---:|
+| 含噪，幅度 25 | 12.485684 | 24.947823 |
+| small，pr2/sr7 | 0.803424 | 47.358490 |
+| base，pr3/sr10 | 0.597227 | 49.414889 |
+| large，pr4/sr14 | 0.463451 | 50.948416 |
+
+来源：assets/results/rtx3060_laptop/quality_tradeoff.csv。三个滤波配置固定同一输入、h=10、sigma=25；只改变 patch 和搜索窗口。base 相对 noisy 提高约 24.47 dB。由于背景较平滑，不能据此证明复杂纹理和真实低照度场景下的细节保持能力。
+
+![历史配对合成图：clean / noisy / denoised](assets/quality_triptych.png)
+
+<!-- pagebreak -->
+
+## 4. 性能结果与统计口径
+
+新 benchmark 每配置预热后重复采样，输出 kernel/e2e 的 mean、min、stddev。CPU 可对相同输入和参数单独重复。历史 NVIDIA legacy 文件只保存 kernel min 与 e2e mean，缺少原始样本，不能补算标准差。
+
+e2e_ms 在 NlmDenoiseGpu 内计时：包含显存分配、H2D、转换、NLM、量化和 D2H；不包含文件读写，也不包含函数退出时的资源释放。H2D/D2H 使用 Host 计时，NLM kernel 使用设备 Event。Mpx/s 由像素数除以 kernel 时间计算。
+
+| GPU / 配置 | 版本 | kernel (ms) | e2e (ms) | 统计口径 |
+|---|---|---:|---:|---|
+| 4090 D / 1080p RGB | V2 | 57.728 | 60.659 | min / mean |
+| 4090 D / 4K RGB | V2 | 229.223 | 242.733 | min / mean |
+| S4000 / 1080p RGB | V2 | 467.985 | 471.324 | mean，10 次 |
+| S4000 / 4K RGB | V2 | 1842.982 | 1873.327 | mean，10 次 |
+| C500 / 1080p RGB | V1 | 418.397 | 420.733 | mean，10 次 |
+| C500 / 4K RGB | V1 | 1645.636 | 1655.068 | mean，10 次 |
+| MR-V100 / 1080p RGB | V2 | 716.595 | 721.346 | mean，10 次 |
+| MR-V100 / 4K RGB | V2 | 2869.991 | 2888.719 | mean，10 次 |
+
+均为 base 参数；C500 为 25% sGPU 配额。表格用于展示本项目各环境的历史行为，设备配额、编译器和统计口径不同，不能用作跨 GPU 公平排名。原始文件、环境与 CPU 重复次数见 assets/results/README.md。
+
+![RTX 4090 D 历史性能](assets/performance_4090.png)
+
+<!-- pagebreak -->
+
+## 5. 质量与延迟权衡
+
+下表来自 RTX 3060 Laptop 的配对质量与性能补充实验，V2、1080p RGB、同一幅度 25 输入、h=10、sigma=25。延迟为 5 次采样的均值与总体标准差。
+
+| 配置 | PSNR (dB) | e2e mean ± stddev (ms) |
+|---|---:|---:|
+| small | 47.358 | 95.986 ± 0.976 |
+| base | 49.415 | 357.573 ± 0.604 |
+| large | 50.948 | 1457.434 ± 2.398 |
+
+small 相对 base 约快 3.7 倍，PSNR 降低 2.06 dB；large 延迟约为 base 的 4.1 倍，PSNR 仅增加 1.53 dB。扩大窗口的质量收益在此合成场景中递减。来源：rtx3060_laptop/benchmark_quality.csv 和 quality_tradeoff.csv。
+
+![质量与延迟](assets/quality_latency_tradeoff.png)
+
+历史 4090 D 的 4K RGB/small kernel 最小耗时为 59.51 ms，而 base 为 229.22 ms。倒数帧率只描述 kernel 的理想处理能力，不能作为包含图像 I/O、资源释放和持续帧输入的完整应用帧率。small 还使用更小窗口，不能代替 base 条件的进阶目标。
+
+<!-- pagebreak -->
+
+## 6. 性能分析与证据边界
+
+历史 RTX 4090 D nsys 摘要中，1080p RGB 的 V2 NLM kernel 为 61.036 ms，约占该设备区间的 99.9%。C500 的 mcTracer 1080p/V1 记录为 422.226 ms，约占求和设备事件时长的 99.617%。这些结果说明 NLM 主核占据主要设备时间。
+
+![MetaX mcTracer 时间线](assets/mctracer_timeline_metax_c500.png)
+
+时间占比本身不能区分算术吞吐、访存、指令调度或寄存器压力造成的瓶颈。现有材料缺少 ncu 硬件计数器；因此不再将“计算受限”“消除了 bank conflict”或“达到某占用率”写作已测结论。共享内存复用和模板展开的解释属于实现分析，需由计数器补充验证。
+
+历史 NVIDIA 记录中 ncu 受硬件计数器访问限制；Moore/Iluvatar 环境未提供相应 profiler CLI。保留 nsys_profile.sh、ncu_profile.sh 和统一脚本中的 mcTracer 入口，工具缺失时明确记录跳过。
+
+## 7. 后续工作
+
+第一优先级是明确题面 dist 定义、sigma 标定及验收口径，并以独立样例验证；调整算法后须重采全部受影响结果。随后补充自然图像与受控噪声的来源和指标，再在正式提交 commit 上重跑 NVIDIA 1080p/4K 矩阵。
+
+性能方向包括常驻缓冲池、减少重复 patch 距离计算、积分图或其他近似、寄存器复用。实际收益与精度代价需要实验评估，现有材料不足以预报具体加速倍数。
+
+<!-- pagebreak -->
+
+## 8. 复现与提交阅读指南
+
+```bash
+make build PLATFORM=nvidia ARCH=sm_89
+make test PLATFORM=nvidia ARCH=sm_89
+make run PLATFORM=nvidia ARCH=sm_89 \
+  INPUT=test_images/noisy_1920x1080_3ch_sigma25.png
+PLATFORM=nvidia ARCH=sm_89 bash scripts/reproduce.sh
+```
+
+其他设备分别设置 PLATFORM=iluvatar、metax、moore，SDK 路径通过 Make 变量或环境变量指定。脚本默认将新结果写入 output/results/<platform>/，不会自动覆盖历史证据。profiler 缺失时单独记录状态；小图 smoke run 不能作为性能验收。
+
+| 材料 | 用途 |
+|---|---|
+| README.md | 构建、运行、版本、平台选择 |
+| lab_report.md / lab_report.pdf | 实现、质量、性能与局限 |
+| assets/results/README.md | 历史数据来源与重复次数 |
+| assets/results/SHA256SUMS.txt | 整理后证据文件完整性 |
+| test_images/README.md | 示例来源、配对关系、噪声定义 |
+| SUBMISSION.md | 补充文件建议与本次结构验证记录 |
+
+本次以参考作业的目录布局组织本项目，没有采用其双边滤波算法代码。多层 include、独立 tester/kernels 和重复实验文档已合并，原始冗余材料在本地仓库外备份。必要测试依据 NLM 专项的提交要求保留。
+
+报告内容依据本地《2026夏季训练营 CUDA 方向项目》NLM 专项（第 13–15 页）、本项目源码和历史证据整理。没有收到具体 PR 评审意见，因此本报告不推断 PR 未通过的确定原因，也不声称目录调整即可保证通过。
diff --git a/07_nlm_denoise/blackbook537/lab_report.pdf b/07_nlm_denoise/blackbook537/lab_report.pdf
new file mode 100644
index 00000000..d30a6377
Binary files /dev/null and b/07_nlm_denoise/blackbook537/lab_report.pdf differ
diff --git a/07_nlm_denoise/blackbook537/params.txt b/07_nlm_denoise/blackbook537/params.txt
new file mode 100644
index 00000000..26f1c117
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/params.txt
@@ -0,0 +1,4 @@
+patch_radius = 3          # patch 半径，实际 patch 大小为 (2*patch_radius+1)^2
+search_radius = 10        # 搜索窗口半径
+h = 10.0                  # 滤波强度
+sigma = 25.0              # 噪声标准差估计，对于8位图像范围0-255
diff --git a/07_nlm_denoise/blackbook537/requirements-report.txt b/07_nlm_denoise/blackbook537/requirements-report.txt
new file mode 100644
index 00000000..59696df2
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/requirements-report.txt
@@ -0,0 +1,2 @@
+Pillow>=10.0,<12
+reportlab>=4.0,<5
diff --git a/07_nlm_denoise/blackbook537/scripts/collect_environment.sh b/07_nlm_denoise/blackbook537/scripts/collect_environment.sh
new file mode 100644
index 00000000..e53a0519
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/collect_environment.sh
@@ -0,0 +1,153 @@
+#!/usr/bin/env bash
+# Collect hardware/software versions without installing or changing the environment.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+OUT="${1:-output/results/environment.txt}"
+PLATFORM="${PLATFORM:-nvidia}"
+COREX_HOME="${COREX_HOME:-/usr/local/corex}"
+MUSA_HOME="${MUSA_HOME:-/usr/local/musa}"
+MACA_HOME="${MACA_HOME:-/opt/maca}"
+MACA_CUDA="${MACA_CUDA:-${MACA_HOME}/tools/cu-bridge}"
+mkdir -p "$(dirname "${OUT}")"
+
+{
+    echo "collected_at_utc=$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+    echo "hostname=$(hostname)"
+    echo "os=$(uname -a)"
+    if test -f /etc/os-release; then
+        grep -E '^(NAME|VERSION|VERSION_ID)=' /etc/os-release
+    fi
+    echo
+    echo "[gpu]"
+    if [[ "${PLATFORM}" == "iluvatar" ]]; then
+        if command -v ixsmi >/dev/null 2>&1; then
+            ixsmi 2>&1 || true
+        else
+            echo "ixsmi=not-found"
+        fi
+    elif [[ "${PLATFORM}" == "moore" ]]; then
+        MUSA_INFO="${MUSA_HOME}/bin/musaInfo"
+        if test -x "${MUSA_INFO}"; then
+            LD_LIBRARY_PATH="${MUSA_HOME}/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}" \
+                "${MUSA_INFO}"
+        else
+            echo "musaInfo=not-found (${MUSA_INFO})"
+        fi
+    elif [[ "${PLATFORM}" == "metax" ]]; then
+        if command -v mx-smi >/dev/null 2>&1; then
+            mx-smi --version 2>&1 || true
+            mx-smi 2>&1 || true
+        else
+            echo "mx-smi=not-found"
+        fi
+    elif command -v nvidia-smi >/dev/null 2>&1; then
+        nvidia-smi --query-gpu=name,uuid,driver_version,memory.total,compute_cap \
+            --format=csv,noheader
+    else
+        echo "gpu-query=not-found for PLATFORM=${PLATFORM}"
+    fi
+    echo
+    echo "[cpu]"
+    echo "nproc=$(nproc)"
+    if command -v lscpu >/dev/null 2>&1; then
+        lscpu | grep -E '^(Architecture|CPU\(s\)|Model name|Thread\(s\) per core|Core\(s\) per socket|Socket\(s\)):' || true
+    fi
+    echo
+    echo "[accelerator-toolkit]"
+    if [[ "${PLATFORM}" == "iluvatar" ]]; then
+        COREX_COMPILER="${COREX_CXX:-${COREX_HOME}/bin/clang++}"
+        if test -x "${COREX_COMPILER}"; then
+            echo "corex-clang++=${COREX_COMPILER}"
+            "${COREX_COMPILER}" --version
+        else
+            echo "corex-clang++=not-found (${COREX_COMPILER})"
+        fi
+        COREX_NVCC="${COREX_NVCC:-${COREX_HOME}/bin/nvcc}"
+        if test -x "${COREX_NVCC}"; then
+            echo "corex-nvcc=${COREX_NVCC}"
+            "${COREX_NVCC}" --version
+        else
+            echo "corex-nvcc=not-found (${COREX_NVCC})"
+        fi
+        echo "COREX_HOME=${COREX_HOME}"
+        test -L "${COREX_HOME}" && echo "COREX_HOME_REALPATH=$(readlink -f "${COREX_HOME}")"
+    elif [[ "${PLATFORM}" == "moore" ]]; then
+        MUSA_COMPILER="${MCC:-${MUSA_HOME}/bin/mcc}"
+        if test -x "${MUSA_COMPILER}"; then
+            echo "mcc=${MUSA_COMPILER}"
+            "${MUSA_COMPILER}" --version
+            if test -x "${MUSA_HOME}/bin/musa_version_query"; then
+                "${MUSA_HOME}/bin/musa_version_query"
+            fi
+        else
+            echo "mcc=not-found (${MUSA_COMPILER})"
+        fi
+    elif [[ "${PLATFORM}" == "metax" ]]; then
+        MACA_COMPILER="${MXCC:-${MACA_HOME}/mxgpu_llvm/bin/mxcc}"
+        if test -x "${MACA_COMPILER}"; then
+            echo "mxcc=${MACA_COMPILER}"
+            "${MACA_COMPILER}" --version
+        else
+            echo "mxcc=not-found (${MACA_COMPILER})"
+        fi
+        echo "MACA_HOME=${MACA_HOME}"
+        echo "MACA_CUDA=${MACA_CUDA}"
+        test -L "${MACA_HOME}" && echo "MACA_HOME_REALPATH=$(readlink -f "${MACA_HOME}")"
+    else
+        CUDA_COMPILER="${NVCC:-nvcc}"
+        if command -v "${CUDA_COMPILER}" >/dev/null 2>&1 || test -x "${CUDA_COMPILER}"; then
+            echo "nvcc=${CUDA_COMPILER}"
+            "${CUDA_COMPILER}" --version
+        else
+            echo "nvcc=not-found"
+        fi
+    fi
+    echo
+    echo "[compiler]"
+    "${CXX_HOST:-g++}" --version | head -n 1
+    make --version | head -n 1
+    echo
+    echo "[profilers]"
+    if command -v nsys >/dev/null 2>&1; then nsys --version; else echo "nsys=not-found"; fi
+    if command -v ncu >/dev/null 2>&1; then ncu --version; else echo "ncu=not-found"; fi
+    if command -v muprof >/dev/null 2>&1; then muprof --version; else echo "muprof=not-found"; fi
+    if command -v mxprof >/dev/null 2>&1; then mxprof --version; else echo "mxprof=not-found"; fi
+    if command -v ixprof >/dev/null 2>&1; then ixprof --version; else echo "ixprof=not-found"; fi
+    if test -x "${MACA_HOME}/bin/mcTracer"; then
+        echo "mcTracer=${MACA_HOME}/bin/mcTracer"
+        "${MACA_HOME}/bin/mcTracer" --help 2>&1 | grep -A 1 -m 1 'Version:' || true
+    else
+        echo "mcTracer=not-found"
+    fi
+    echo
+    echo "[build-contract]"
+    echo "PLATFORM=${PLATFORM}"
+    if [[ "${PLATFORM}" == "nvidia" ]]; then
+        echo "ARCH=${ARCH:-compiler-default}"
+    elif [[ "${PLATFORM}" == "iluvatar" ]]; then
+        echo "COREX_HOME=${COREX_HOME}"
+        echo "COREX_CXX=${COREX_CXX:-${COREX_HOME}/bin/clang++}"
+    elif [[ "${PLATFORM}" == "metax" ]]; then
+        echo "MACA_HOME=${MACA_HOME}"
+        echo "MACA_CUDA=${MACA_CUDA}"
+        echo "MXCC=${MXCC:-${MACA_HOME}/mxgpu_llvm/bin/mxcc}"
+    else
+        echo "MUSA_HOME=${MUSA_HOME}"
+    fi
+    echo "FAST_EXP=${FAST_EXP:-0}"
+    echo "WITH_OPENCV=${WITH_OPENCV:-0}"
+    if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+        echo "git_commit=$(git rev-parse HEAD)"
+        if test -n "$(git status --porcelain)"; then
+            echo "git_worktree=dirty"
+        else
+            echo "git_worktree=clean"
+        fi
+    else
+        echo "git_commit=unavailable"
+        echo "git_worktree=unavailable (uploaded source snapshot)"
+    fi
+} | tee "${OUT}"
+
+echo "Environment manifest: ${OUT}"
diff --git a/07_nlm_denoise/blackbook537/scripts/export_report.py b/07_nlm_denoise/blackbook537/scripts/export_report.py
new file mode 100644
index 00000000..42e2b8f4
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/export_report.py
@@ -0,0 +1,133 @@
+#!/usr/bin/env python3
+"""Export the checked-in Markdown report; no performance values are generated."""
+from __future__ import annotations
+
+import argparse
+import html
+import os
+from pathlib import Path
+import re
+
+from reportlab.lib import colors
+from reportlab.lib.enums import TA_LEFT
+from reportlab.lib.pagesizes import A4
+from reportlab.lib.styles import ParagraphStyle
+from reportlab.pdfbase import pdfmetrics
+from reportlab.pdfbase.cidfonts import UnicodeCIDFont
+from reportlab.pdfbase.ttfonts import TTFont
+from reportlab.platypus import (
+    Image, KeepTogether, PageBreak, Paragraph, SimpleDocTemplate, Spacer, Table, TableStyle,
+)
+
+ROOT = Path(__file__).resolve().parent.parent
+WIDTH = A4[0] - 80
+
+
+def register_font() -> str:
+    candidates = [os.environ.get('PDF_CJK_FONT', ''), 'C:/Windows/Fonts/simsun.ttc',
+                  '/usr/share/fonts/truetype/arphic/uming.ttc']
+    for candidate in candidates:
+        if candidate and Path(candidate).is_file():
+            pdfmetrics.registerFont(TTFont('ReportCJK', candidate, subfontIndex=0))
+            return 'ReportCJK'
+    pdfmetrics.registerFont(UnicodeCIDFont('STSong-Light'))
+    return 'STSong-Light'
+
+
+def inline(text: str) -> str:
+    text = html.escape(text)
+    text = re.sub(r'`([^`]+)`', r'<font color="#1d4ed8">\1</font>', text)
+    return re.sub(r'\[([^\]]+)\]\(([^)]+)\)', r'\1', text)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--input', type=Path, default=ROOT / 'lab_report.md')
+    parser.add_argument('--output', type=Path, default=ROOT / 'lab_report.pdf')
+    args = parser.parse_args()
+    font = register_font()
+    base = dict(fontName=font, textColor=colors.HexColor('#253047'), wordWrap='CJK')
+    body = ParagraphStyle('Body', fontSize=10, leading=16, spaceAfter=9, **base)
+    title = ParagraphStyle('Title', fontSize=23, leading=32, spaceAfter=14, **base)
+    heading = ParagraphStyle('Heading', fontSize=15, leading=23, spaceBefore=9,
+                             spaceAfter=10, keepWithNext=True, **base)
+    cell = ParagraphStyle('Cell', fontSize=9, leading=13, **base)
+    code = ParagraphStyle('Code', fontSize=8.3, leading=12.5, spaceAfter=8,
+                          leftIndent=8, rightIndent=8, **base)
+    caption = ParagraphStyle('Caption', fontSize=8.5, leading=12, spaceAfter=9, **base)
+    story = []
+    lines = args.input.read_text(encoding='utf-8').splitlines()
+    i = 0
+    while i < len(lines):
+        line = lines[i].strip()
+        if not line:
+            i += 1
+            continue
+        if line == '<!-- pagebreak -->':
+            story.append(PageBreak())
+        elif line.startswith('```'):
+            block = []
+            i += 1
+            while i < len(lines) and not lines[i].startswith('```'):
+                block.append(html.escape(lines[i]).replace(' ', '&#160;'))
+                i += 1
+            story.append(Paragraph('<br/>'.join(block), code))
+        elif line.startswith('|'):
+            rows = []
+            while i < len(lines) and lines[i].strip().startswith('|'):
+                cells = [v.strip() for v in lines[i].strip().strip('|').split('|')]
+                if not all(re.fullmatch(r':?-+:?', v) for v in cells):
+                    rows.append(cells)
+                i += 1
+            ncols = len(rows[0])
+            weights = [min(30, max(10, max(len(row[j]) for row in rows))) for j in range(ncols)]
+            widths = [WIDTH * w / sum(weights) for w in weights]
+            table = Table([[Paragraph(inline(v), cell) for v in row] for row in rows],
+                          colWidths=widths, repeatRows=1, hAlign='LEFT')
+            table.setStyle(TableStyle([
+                ('BACKGROUND', (0, 0), (-1, 0), colors.HexColor('#e8eef9')),
+                ('ROWBACKGROUNDS', (0, 1), (-1, -1), [colors.white, colors.HexColor('#f7f9fc')]),
+                ('LINEBELOW', (0, 0), (-1, 0), 0.7, colors.HexColor('#b9c7df')),
+                ('VALIGN', (0, 0), (-1, -1), 'TOP'),
+                ('TOPPADDING', (0, 0), (-1, -1), 7),
+                ('BOTTOMPADDING', (0, 0), (-1, -1), 7),
+                ('LEFTPADDING', (0, 0), (-1, -1), 7),
+                ('RIGHTPADDING', (0, 0), (-1, -1), 7),
+            ]))
+            story.extend([table, Spacer(1, 10)])
+            continue
+        elif line.startswith('!['):
+            match = re.fullmatch(r'!\[([^\]]*)\]\(([^)]+)\)', line)
+            if not match:
+                raise ValueError(f'Invalid image: {line}')
+            asset = args.input.parent / match[2]
+            image = Image(str(asset))
+            scale = min(WIDTH / image.imageWidth, 265 / image.imageHeight)
+            image.drawWidth = image.imageWidth * scale
+            image.drawHeight = image.imageHeight * scale
+            story.append(KeepTogether([image, Spacer(1, 5), Paragraph(inline(match[1]), caption)]))
+        elif line.startswith('# '):
+            story.append(Paragraph(inline(line[2:]), title))
+        elif line.startswith('## '):
+            story.append(Paragraph(inline(line[3:]), heading))
+        else:
+            story.append(Paragraph(inline(line), body))
+        i += 1
+
+    def footer(canvas, doc):
+        canvas.setStrokeColor(colors.HexColor('#cbd5e1'))
+        canvas.line(40, 34, A4[0]-40, 34)
+        canvas.setFont(font, 8)
+        canvas.setFillColor(colors.HexColor('#64748b'))
+        canvas.drawString(40, 21, 'CUDA NLM / blackbook537 / 2026')
+        canvas.drawRightString(A4[0]-40, 21, str(doc.page))
+
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    doc = SimpleDocTemplate(str(args.output), pagesize=A4, leftMargin=40, rightMargin=40,
+                            topMargin=36, bottomMargin=47, title='CUDA NLM 实验报告', author='blackbook537')
+    doc.build(story, onFirstPage=footer, onLaterPages=footer)
+    print(args.output)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/07_nlm_denoise/blackbook537/scripts/generate_report_assets.py b/07_nlm_denoise/blackbook537/scripts/generate_report_assets.py
new file mode 100644
index 00000000..46253e4a
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/generate_report_assets.py
@@ -0,0 +1,397 @@
+#!/usr/bin/env python3
+"""Generate reviewable PNG assets from experiment CSVs and output images.
+
+Only Pillow is required. The script never invents missing values: a chart is skipped
+with an explicit message when its required input is absent or incomplete.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+from pathlib import Path
+
+from PIL import Image, ImageDraw, ImageFont
+
+
+BG = "#f7f9fc"
+INK = "#172033"
+MUTED = "#5f6b7a"
+BLUE = "#2563eb"
+TEAL = "#0f9d8a"
+ORANGE = "#f59e0b"
+RED = "#dc2626"
+GRID = "#d8dee9"
+
+
+def font(size: int, bold: bool = False) -> ImageFont.FreeTypeFont | ImageFont.ImageFont:
+    candidates = [
+        Path("C:/Windows/Fonts/arialbd.ttf" if bold else "C:/Windows/Fonts/arial.ttf"),
+        Path("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf" if bold else
+             "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf"),
+    ]
+    for path in candidates:
+        if path.exists():
+            return ImageFont.truetype(str(path), size=size)
+    return ImageFont.load_default()
+
+
+def canvas(title: str, subtitle: str = "") -> tuple[Image.Image, ImageDraw.ImageDraw]:
+    image = Image.new("RGB", (1600, 900), BG)
+    draw = ImageDraw.Draw(image)
+    draw.text((70, 45), title, fill=INK, font=font(42, True))
+    if subtitle:
+        draw.text((72, 100), subtitle, fill=MUTED, font=font(22))
+    return image, draw
+
+
+def read_csv(path: Path) -> list[dict[str, str]]:
+    with path.open("r", encoding="utf-8-sig", newline="") as handle:
+        return list(csv.DictReader(handle))
+
+
+def value(row: dict[str, str], *keys: str) -> float:
+    for key in keys:
+        raw = row.get(key)
+        if raw not in (None, ""):
+            return float(raw)
+    raise KeyError(keys)
+
+
+def infer_config(row: dict[str, str]) -> str:
+    if row.get("config"):
+        return row["config"]
+    pr = int(float(row["pr"]))
+    sr = int(float(row["sr"]))
+    h = float(row["h"])
+    if pr == 2 and sr == 7:
+        return "small"
+    if pr == 4 and sr == 14:
+        return "large"
+    if pr == 3 and sr == 10 and h == 15:
+        return "strong-h"
+    if pr == 3 and sr == 10:
+        return "base"
+    return f"pr{pr}-sr{sr}"
+
+
+def draw_axes(draw: ImageDraw.ImageDraw, box: tuple[int, int, int, int], ymax: float,
+              ylabel: str) -> None:
+    left, top, right, bottom = box
+    draw.line((left, top, left, bottom), fill=INK, width=3)
+    draw.line((left, bottom, right, bottom), fill=INK, width=3)
+    for i in range(6):
+        y = bottom - (bottom - top) * i / 5
+        val = ymax * i / 5
+        draw.line((left, int(y), right, int(y)), fill=GRID, width=1)
+        draw.text((left - 95, int(y) - 12), f"{val:.0f}", fill=MUTED, font=font(18))
+    draw.text((left, top - 42), ylabel, fill=MUTED, font=font(20))
+
+
+def performance_chart(csv_paths: list[Path], output: Path, device_label: str) -> bool:
+    rows: list[dict[str, str]] = []
+    for csv_path in csv_paths:
+        rows.extend(read_csv(csv_path))
+    chosen: dict[tuple[str, int], dict[str, str]] = {}
+    for row in rows:
+        if row.get("channels") != "3" or infer_config(row) != "base":
+            continue
+        size = row.get("size", "")
+        if size in ("1920x1080", "3840x2160"):
+            chosen[(size, int(row["kernel_ver"]))] = row
+    if len(chosen) != 6:
+        print(f"SKIP performance chart: expected 6 base RGB rows in {csv_paths}")
+        return False
+
+    legacy = any("kernel_mean_ms" not in row for row in chosen.values())
+    statistic = "kernel min / pipeline e2e mean (legacy)" if legacy else "kernel mean / pipeline e2e mean"
+    image, draw = canvas(f"{device_label}: V0 / V1 / V2 latency",
+                         f"Archived experiment: {statistic}; excludes image I/O")
+    box = (150, 190, 1510, 760)
+    vals = [value(row, "kernel_mean_ms", "kernel_ms") for row in chosen.values()]
+    vals += [value(row, "e2e_mean_ms", "e2e_ms") for row in chosen.values()]
+    ymax = math.ceil(max(vals) / 100.0) * 100.0
+    draw_axes(draw, box, ymax, "milliseconds")
+    group_centers = [430, 1180]
+    colors = [BLUE, TEAL, ORANGE]
+    left, top, _, bottom = box
+    for group_index, size in enumerate(("1920x1080", "3840x2160")):
+        center = group_centers[group_index]
+        for ver in range(3):
+            row = chosen[(size, ver)]
+            kernel_ms = value(row, "kernel_mean_ms", "kernel_ms")
+            e2e_ms = value(row, "e2e_mean_ms", "e2e_ms")
+            x = center + (ver - 1) * 150
+            for offset, number, shade in ((-34, kernel_ms, colors[ver]),
+                                           (34, e2e_ms, "#94a3b8")):
+                height = (bottom - top) * number / ymax
+                draw.rounded_rectangle((x + offset - 26, bottom - height,
+                                        x + offset + 26, bottom), radius=6, fill=shade)
+                draw.text((x + offset - 38, bottom - height - 30), f"{number:.1f}",
+                          fill=INK, font=font(17))
+            draw.text((x - 20, bottom + 16), f"V{ver}", fill=INK, font=font(20, True))
+        draw.text((center - 75, bottom + 55), size, fill=INK, font=font(24, True))
+    for idx, shade in enumerate(colors):
+        draw.rectangle((960 + idx * 24, 115, 980 + idx * 24, 140), fill=shade)
+    draw.text((1042, 112), "kernel", fill=INK, font=font(20))
+    draw.rectangle((1170, 115, 1200, 140), fill="#94a3b8")
+    draw.text((1210, 112), "GPU e2e", fill=INK, font=font(20))
+    output.parent.mkdir(parents=True, exist_ok=True)
+    image.save(output, optimize=True)
+    return True
+
+
+def triptych(clean_path: Path, noisy_path: Path, denoised_path: Path, output: Path) -> bool:
+    paths = [clean_path, noisy_path, denoised_path]
+    if not all(path.exists() for path in paths):
+        print(f"SKIP triptych: missing one of {paths}")
+        return False
+    images = [Image.open(path).convert("RGB") for path in paths]
+    if len({im.size for im in images}) != 1:
+        print("SKIP triptych: image dimensions do not match")
+        return False
+    image = Image.new("RGB", (1800, 1050), BG)
+    draw = ImageDraw.Draw(image)
+    draw.text((65, 35), "NLM visual quality: full frame and enlarged ROI", fill=INK,
+              font=font(40, True))
+    names = ["Clean reference", "Uniform noise (amplitude=25)", "Denoised V2 (base)"]
+    full_size = (540, 304)
+    source_w, source_h = images[0].size
+    roi_w, roi_h = 240, 136
+    roi_left = max(0, source_w // 2 - roi_w // 2)
+    roi_top = max(0, source_h // 2 - roi_h // 2)
+    for idx, (name, source) in enumerate(zip(names, images)):
+        x = 40 + idx * 590
+        y = 115
+        full = source.resize(full_size, Image.Resampling.LANCZOS)
+        image.paste(full, (x, y))
+        scale_x = full_size[0] / source_w
+        scale_y = full_size[1] / source_h
+        draw.rectangle((x + roi_left * scale_x, y + roi_top * scale_y,
+                        x + (roi_left + roi_w) * scale_x,
+                        y + (roi_top + roi_h) * scale_y), outline=RED, width=4)
+        draw.text((x, y + full_size[1] + 12), name, fill=INK, font=font(24, True))
+        crop = source.crop((roi_left, roi_top, roi_left + roi_w, roi_top + roi_h))
+        crop = crop.resize((540, 306), Image.Resampling.NEAREST)
+        image.paste(crop, (x, 520))
+        draw.rectangle((x, 520, x + 540, 826), outline=RED, width=3)
+        draw.text((x, 840), "Center ROI, 2.25x nearest-neighbor zoom", fill=MUTED,
+                  font=font(19))
+    draw.text((65, 980), "Paired deterministic synthetic images, seed=7; no post-processing.",
+              fill=MUTED, font=font(21))
+    output.parent.mkdir(parents=True, exist_ok=True)
+    image.save(output, optimize=True)
+    return True
+
+
+def scatter_chart(quality_csv: Path, benchmark_csvs: list[Path], output: Path,
+                  device_label: str) -> bool:
+    if not quality_csv.exists() or not benchmark_csvs or \
+            not all(path.exists() for path in benchmark_csvs):
+        print("SKIP quality-latency chart: CSV missing")
+        return False
+    quality = {row["label"]: float(row["psnr_db"]) for row in read_csv(quality_csv)
+               if row.get("label") in ("small", "base", "large")}
+    latency: dict[str, float] = {}
+    for benchmark_csv in benchmark_csvs:
+        for row in read_csv(benchmark_csv):
+            if row.get("size") == "1920x1080" and row.get("channels") == "3" and \
+                    row.get("kernel_ver") == "2":
+                name = infer_config(row)
+                if name in ("small", "base", "large"):
+                    latency[name] = value(row, "e2e_mean_ms", "e2e_ms")
+    if len(quality) != 3 or len(latency) != 3:
+        print("SKIP quality-latency chart: small/base/large rows incomplete")
+        return False
+
+    image, draw = canvas(f"Quality / latency trade-off ({device_label})",
+                         "Same uniform-noise amplitude=25 input; V2; higher PSNR and lower latency are better")
+    box = (170, 190, 1510, 770)
+    xs = list(latency.values())
+    ys = list(quality.values())
+    xmin, xmax = 0.0, max(xs) * 1.15
+    ymin, ymax = min(ys) - 1.0, max(ys) + 1.0
+    left, top, right, bottom = box
+    draw.line((left, top, left, bottom), fill=INK, width=3)
+    draw.line((left, bottom, right, bottom), fill=INK, width=3)
+    for i in range(6):
+        x = left + (right - left) * i / 5
+        draw.line((int(x), top, int(x), bottom), fill=GRID, width=1)
+        draw.text((int(x) - 20, bottom + 15), f"{xmax * i / 5:.0f}", fill=MUTED,
+                  font=font(18))
+        y = bottom - (bottom - top) * i / 5
+        draw.line((left, int(y), right, int(y)), fill=GRID, width=1)
+        draw.text((left - 75, int(y) - 12), f"{ymin + (ymax-ymin)*i/5:.1f}",
+                  fill=MUTED, font=font(18))
+    draw.text((650, 830), "GPU e2e mean (ms)", fill=INK, font=font(22))
+    draw.text((25, 155), "PSNR (dB)", fill=INK, font=font(22))
+    colors = {"small": TEAL, "base": BLUE, "large": ORANGE}
+    for name in ("small", "base", "large"):
+        px = left + (right - left) * (latency[name] - xmin) / (xmax - xmin)
+        py = bottom - (bottom - top) * (quality[name] - ymin) / (ymax - ymin)
+        draw.ellipse((px - 14, py - 14, px + 14, py + 14), fill=colors[name])
+        label_x = px + 20 if px < right - 330 else px - 330
+        draw.text((label_x, py - 35), f"{name}: {quality[name]:.2f} dB, "
+                  f"{latency[name]:.1f} ms", fill=INK, font=font(20, True))
+    output.parent.mkdir(parents=True, exist_ok=True)
+    image.save(output, optimize=True)
+    return True
+
+
+def mctracer_timeline(trace_path: Path, output: Path, device_label: str) -> bool:
+    if not trace_path.exists():
+        print(f"SKIP mcTracer timeline: missing {trace_path}")
+        return False
+    with trace_path.open("r", encoding="utf-8") as handle:
+        events = json.load(handle).get("traceEvents", [])
+    events = [event for event in events
+              if event.get("ph") == "X" and event.get("cat") == "0" and
+              float(event.get("dur", 0)) > 0]
+    if not events:
+        print(f"SKIP mcTracer timeline: no device duration events in {trace_path}")
+        return False
+
+    def label(name: str) -> str:
+        if "HTOD" in name:
+            return "H2D"
+        if "DTOH" in name:
+            return "D2H"
+        if "CvtU8" in name:
+            return "U8 to F32"
+        if "CvtF32" in name:
+            return "F32 to U8"
+        if "Nlm" in name:
+            return "NLM V1"
+        return name.split("(", 1)[0]
+
+    events.sort(key=lambda event: float(event.get("ts", 0)))
+    start = min(float(event["ts"]) for event in events)
+    finish = max(float(event["ts"]) + float(event["dur"]) for event in events)
+    span = max(finish - start, 1.0)
+    device_total = sum(float(event["dur"]) for event in events)
+
+    image, draw = canvas(f"{device_label}: MACA device timeline",
+                         "mcTracer raw JSON; 1080p RGB base, V1")
+    left, right = 285, 1500
+    top, row_height = 190, 82
+    colors = {"H2D": TEAL, "D2H": RED, "U8 to F32": BLUE,
+              "F32 to U8": ORANGE, "NLM V1": "#7c3aed"}
+    for index, event in enumerate(events):
+        name = label(str(event.get("name", "event")))
+        duration = float(event["dur"])
+        x0 = left + (right - left) * (float(event["ts"]) - start) / span
+        x1 = left + (right - left) * (float(event["ts"]) + duration - start) / span
+        x1 = max(x1, x0 + 6)
+        y = top + index * row_height
+        draw.text((65, y + 8), name, fill=INK, font=font(21, True))
+        draw.line((left, y + 25, right, y + 25), fill=GRID, width=2)
+        draw.rounded_rectangle((x0, y + 9, x1, y + 41), radius=5,
+                               fill=colors.get(name, MUTED))
+        draw.text((left, y + 48),
+                  f"{duration / 1e6:.6f} ms; {100.0 * duration / device_total:.3f}%",
+                  fill=MUTED, font=font(17))
+
+    bottom = top + len(events) * row_height + 20
+    draw.text((65, bottom),
+              f"Device-event span: {span / 1e6:.3f} ms | "
+              f"summed device duration: {device_total / 1e6:.3f} ms",
+              fill=INK, font=font(22, True))
+    draw.text((65, bottom + 42),
+              "The NLM kernel occupies 99.6% of traced device time; transfer bars "
+              "are widened to remain visible.",
+              fill=MUTED, font=font(20))
+    output.parent.mkdir(parents=True, exist_ok=True)
+    image.save(output, optimize=True)
+    return True
+
+
+def nsys_chart(csv_path: Path, output: Path) -> bool:
+    if not csv_path.exists():
+        print("SKIP nsys chart: CSV missing")
+        return False
+    rows = read_csv(csv_path)
+    if not rows:
+        return False
+    image, draw = canvas("Nsight Systems: NLM kernel time",
+                         "RTX 4090 D verified profile; conversion kernels are below 0.1%")
+    box = (220, 220, 1460, 750)
+    ymax = math.ceil(max(float(row["gpu_ms"]) for row in rows) / 20.0) * 20.0
+    draw_axes(draw, box, ymax, "GPU milliseconds")
+    left, top, _, bottom = box
+    for idx, row in enumerate(rows):
+        x = 470 + idx * 350
+        number = float(row["gpu_ms"])
+        height = (bottom - top) * number / ymax
+        draw.rounded_rectangle((x - 85, bottom - height, x + 85, bottom),
+                               radius=10, fill=(BLUE, TEAL, ORANGE)[idx])
+        draw.text((x - 52, bottom - height - 42), f"{number:.3f} ms", fill=INK,
+                  font=font(22, True))
+        draw.text((x - 25, bottom + 18), row["implementation"], fill=INK,
+                  font=font(25, True))
+        draw.text((x - 125, bottom + 57), f"GPU share {row['gpu_share_percent']}%",
+                  fill=MUTED, font=font(19))
+    output.parent.mkdir(parents=True, exist_ok=True)
+    image.save(output, optimize=True)
+    return True
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--assets", type=Path, default=Path("assets"))
+    parser.add_argument("--benchmark", type=Path, action="append", default=[])
+    parser.add_argument("--performance-device", default="RTX 4090 D")
+    parser.add_argument("--performance-output", default="performance_4090.png")
+    parser.add_argument("--performance-only", action="store_true")
+    parser.add_argument("--quality", type=Path,
+                        default=Path("assets/results/rtx3060_laptop/quality_tradeoff.csv"))
+    parser.add_argument("--quality-benchmark", type=Path, action="append", default=[])
+    parser.add_argument("--quality-device", default="RTX 3060 Laptop (archived)")
+    parser.add_argument("--clean", type=Path, default=Path("test_images/clean_1920x1080_3ch.png"))
+    parser.add_argument("--noisy", type=Path,
+                        default=Path("test_images/noisy_1920x1080_3ch_sigma25.png"))
+    parser.add_argument("--denoised", type=Path,
+                        default=Path("test_images/denoised_1920x1080_base_v2.png"))
+    parser.add_argument("--triptych-output", default="quality_triptych.png")
+    parser.add_argument("--quality-output", default="quality_latency_tradeoff.png")
+    parser.add_argument("--trace", type=Path)
+    parser.add_argument("--trace-output", default="mctracer_timeline.png")
+    parser.add_argument("--experiment-only", action="store_true")
+    args = parser.parse_args()
+
+    benchmark_paths = args.benchmark or [
+        Path("assets/results/rtx4090d/benchmark_legacy.csv")
+    ]
+    performance_generated = performance_chart(
+        benchmark_paths, args.assets / args.performance_output,
+        args.performance_device)
+    if args.performance_only:
+        print(f"Generated {int(performance_generated)}/1 assets in {args.assets}")
+        return 0 if performance_generated else 1
+
+    generated = int(performance_generated)
+    generated += triptych(args.clean, args.noisy, args.denoised,
+                          args.assets / args.triptych_output)
+    quality_benchmarks = args.quality_benchmark or [
+        Path("assets/results/rtx3060_laptop/benchmark_quality.csv")
+    ]
+    generated += scatter_chart(args.quality, quality_benchmarks,
+                               args.assets / args.quality_output,
+                               args.quality_device)
+    if args.trace:
+        generated += mctracer_timeline(args.trace, args.assets / args.trace_output,
+                                       args.performance_device)
+    if args.experiment_only:
+        expected = 4 if args.trace else 3
+        print(f"Generated {generated}/{expected} assets in {args.assets}")
+        return 0 if generated == expected else 1
+    generated += nsys_chart(Path("assets/results/rtx4090d/nsys_summary_legacy.csv"),
+                            args.assets / "nsys_kernel_summary.png")
+    expected = 5 if args.trace else 4
+    print(f"Generated {generated}/{expected} assets in {args.assets}")
+    return 0 if generated == expected else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/07_nlm_denoise/blackbook537/scripts/ncu_profile.sh b/07_nlm_denoise/blackbook537/scripts/ncu_profile.sh
new file mode 100644
index 00000000..640b9497
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/ncu_profile.sh
@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+# Nsight Compute 指标采集：对 V0/V1/V2 三个 kernel 版本分别采样，
+# 输出 SpeedOfLight / MemoryWorkloadAnalysis / Occupancy 三组关键指标，
+# 用于报告中的"瓶颈 -> 优化 -> 指标变化"闭环分析。
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+mkdir -p output/work output/results
+IMG=output/work/nlm_ncu.png
+./build/nlm_denoise gen -o "${IMG}" --size 1920x1080 --channels 3 --sigma 25
+
+for V in 0 1 2; do
+    echo "===== ncu kernel v${V} ====="
+    ncu --set speedoflight \
+        --section MemoryWorkloadAnalysis \
+        --section Occupancy \
+        --kernel-name-base demangled \
+        -o "output/work/ncu_v${V}" -f \
+        ./build/nlm_denoise run -i "${IMG}" -o output/work/nlm_ncu_out.png \
+            -p params.txt --kernel "${V}" --log /dev/null
+done
+echo "报告文件: output/work/ncu_v0.ncu-rep / ncu_v1.ncu-rep / ncu_v2.ncu-rep"
diff --git a/07_nlm_denoise/blackbook537/scripts/nsys_profile.sh b/07_nlm_denoise/blackbook537/scripts/nsys_profile.sh
new file mode 100644
index 00000000..f6350eb2
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/nsys_profile.sh
@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+# Nsight Systems 时间线采集：确认 NLM kernel 占端到端时间比例、
+# 检查 H2D/D2H 气泡，支撑 4K 交互式场景的 stream 分片优化验证。
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+mkdir -p output/work output/results
+IMG=output/work/nlm_nsys.png
+./build/nlm_denoise gen -o "${IMG}" --size 3840x2160 --channels 3 --sigma 25
+
+nsys profile -o output/work/nsys_4k -f true \
+    ./build/nlm_denoise run -i "${IMG}" -o output/work/nlm_nsys_out.png \
+        -p params.txt --kernel 2 --log /dev/null
+
+nsys stats output/work/nsys_4k.nsys-rep \
+    --report cuda_gpu_kern_sum --report cuda_gpu_mem_time_sum \
+    | tee output/results/nsys_stats_4k.txt
+echo "报告文件: output/work/nsys_4k.nsys-rep"
diff --git a/07_nlm_denoise/blackbook537/scripts/params/large.txt b/07_nlm_denoise/blackbook537/scripts/params/large.txt
new file mode 100644
index 00000000..e45df86e
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/params/large.txt
@@ -0,0 +1,5 @@
+# 质量/压力测试：扩大 patch/search window；噪声标定仍为 25。
+patch_radius = 4
+search_radius = 14
+h = 10
+sigma = 25
diff --git a/07_nlm_denoise/blackbook537/scripts/params/sigma10.txt b/07_nlm_denoise/blackbook537/scripts/params/sigma10.txt
new file mode 100644
index 00000000..6d445976
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/params/sigma10.txt
@@ -0,0 +1,4 @@
+patch_radius = 3
+search_radius = 10
+h = 10
+sigma = 10
diff --git a/07_nlm_denoise/blackbook537/scripts/params/sigma50.txt b/07_nlm_denoise/blackbook537/scripts/params/sigma50.txt
new file mode 100644
index 00000000..553bc6f5
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/params/sigma50.txt
@@ -0,0 +1,4 @@
+patch_radius = 3
+search_radius = 10
+h = 10
+sigma = 50
diff --git a/07_nlm_denoise/blackbook537/scripts/params/small.txt b/07_nlm_denoise/blackbook537/scripts/params/small.txt
new file mode 100644
index 00000000..3635eee5
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/params/small.txt
@@ -0,0 +1,5 @@
+# 速度优先：缩小 patch/search window；与 base 使用相同噪声标定，便于公平比较。
+patch_radius = 2
+search_radius = 7
+h = 10
+sigma = 25
diff --git a/07_nlm_denoise/blackbook537/scripts/params/strong_h.txt b/07_nlm_denoise/blackbook537/scripts/params/strong_h.txt
new file mode 100644
index 00000000..9e1d2093
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/params/strong_h.txt
@@ -0,0 +1,5 @@
+# 与 base 计算量相同，提高 h 以增强平滑。
+patch_radius = 3
+search_radius = 10
+h = 15
+sigma = 25
diff --git a/07_nlm_denoise/blackbook537/scripts/reproduce.sh b/07_nlm_denoise/blackbook537/scripts/reproduce.sh
new file mode 100644
index 00000000..ee57daab
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/reproduce.sh
@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+# One entry for all platforms; outputs are excluded from the submission tree.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+export PLATFORM="${PLATFORM:-nvidia}"
+export RESULT_DIR="${RESULT_DIR:-output/results/${PLATFORM}}"
+export WORK_DIR="${WORK_DIR:-output/work/${PLATFORM}}"
+export ARCH="${ARCH:-}"
+SIZES="${SIZES:-1920x1080,3840x2160}"
+CHANNELS="${CHANNELS:-1,3}"
+WARMUP="${WARMUP:-3}"
+REPEAT="${REPEAT:-10}"
+CPU_REPEAT="${CPU_REPEAT:-3}"
+VALIDATE_SIZE="${VALIDATE_SIZE:-1920x1080}"
+export SIZE="${QUALITY_SIZE:-1920x1080}"
+
+case "${PLATFORM}" in
+    nvidia)
+        if ! command -v "${NVCC:-nvcc}" >/dev/null 2>&1 && test -x /usr/local/cuda/bin/nvcc; then
+            export PATH="/usr/local/cuda/bin:${PATH}"
+        fi ;;
+    iluvatar)
+        export COREX_HOME="${COREX_HOME:-/usr/local/corex}"
+        export PATH="${COREX_HOME}/bin:${PATH}" ;;
+    metax)
+        export MACA_HOME="${MACA_HOME:-/opt/maca}"
+        export PATH="${MACA_HOME}/bin:${PATH}" ;;
+    moore) export MUSA_HOME="${MUSA_HOME:-/usr/local/musa}" ;;
+    *) echo "Unsupported PLATFORM: ${PLATFORM}" >&2; exit 2 ;;
+esac
+mkdir -p "${RESULT_DIR}" "${WORK_DIR}"
+BIN="./build/${PLATFORM}/nlm_denoise"
+
+echo '[1/7] Environment and build'
+bash scripts/collect_environment.sh "${RESULT_DIR}/environment.txt"
+make build -j"$(nproc)" 2>&1 | tee "${RESULT_DIR}/build.log"
+
+echo '[2/7] CPU self checks and three-version GPU validation'
+"${BIN}" units 2>&1 | tee "${RESULT_DIR}/self_check.log"
+"${BIN}" gen -o "${WORK_DIR}/validation.png" --size "${VALIDATE_SIZE}" --channels 3 --sigma 25 --seed 7
+"${BIN}" validate -i "${WORK_DIR}/validation.png" -p params.txt \
+    -o "${WORK_DIR}/validated.png" 2>&1 | tee "${RESULT_DIR}/validation.log"
+
+echo '[3/7] GPU matrix'
+"${BIN}" bench --sizes "${SIZES}" --channels "${CHANNELS}" --param-sets all \
+    --warmup "${WARMUP}" --repeat "${REPEAT}" --log "${RESULT_DIR}/benchmark.csv" \
+    2>&1 | tee "${RESULT_DIR}/benchmark.log"
+
+echo '[4/7] CPU baseline on the same input and parameters'
+"${BIN}" bench --sizes "${VALIDATE_SIZE}" --channels 3 --param-sets base \
+    --warmup "${WARMUP}" --repeat "${REPEAT}" --with-cpu --cpu-repeat "${CPU_REPEAT}" \
+    --log "${RESULT_DIR}/cpu_baseline.csv" 2>&1 | tee "${RESULT_DIR}/cpu_baseline.log"
+
+echo '[5/7] Paired-image quality sweeps (V2 on every platform)'
+CHANNELS=3 bash scripts/run_quality_sweep.sh 2>&1 | tee "${RESULT_DIR}/quality.log"
+
+echo '[6/7] Optional device profiler'
+if [[ "${PLATFORM}" == "nvidia" ]] && command -v nsys >/dev/null 2>&1; then
+    nsys profile -f true -o "${WORK_DIR}/nsys_v2" "${BIN}" run \
+        -i "${WORK_DIR}/validation.png" -o "${WORK_DIR}/profiled.png" -p params.txt --kernel 2 --log /dev/null
+    nsys stats "${WORK_DIR}/nsys_v2.nsys-rep" --report cuda_gpu_kern_sum \
+        --report cuda_gpu_mem_time_sum > "${RESULT_DIR}/profiler.txt"
+elif [[ "${PLATFORM}" == "metax" ]] && test -x "${MACA_HOME}/bin/mcTracer"; then
+    mkdir -p "${RESULT_DIR}/mctracer"
+    "${MACA_HOME}/bin/mcTracer" --mctx --odname "${RESULT_DIR}/mctracer" --name nlm_v1 \
+        "${BIN}" run -i "${WORK_DIR}/validation.png" -o "${WORK_DIR}/profiled.png" \
+        -p params.txt --kernel 1 --log /dev/null 2>&1 | tee "${RESULT_DIR}/profiler.txt"
+else
+    echo 'SKIPPED: no supported profiler CLI found; this run provides no new hardware profiling evidence.' \
+        | tee "${RESULT_DIR}/profiler.txt"
+fi
+
+echo '[7/7] Hash evidence'
+(cd "${RESULT_DIR}"; find . -type f ! -name SHA256SUMS.txt -print0 | sort -z | xargs -0 sha256sum > SHA256SUMS.txt)
+echo "Core validation and benchmark complete: ${RESULT_DIR}; profiler status: ${RESULT_DIR}/profiler.txt"
diff --git a/07_nlm_denoise/blackbook537/scripts/run_quality_sweep.sh b/07_nlm_denoise/blackbook537/scripts/run_quality_sweep.sh
new file mode 100644
index 00000000..e5e18acc
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/scripts/run_quality_sweep.sh
@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+# Reproduce quality experiments on paired deterministic synthetic images.
+# The clean/noisy pair always uses the same seed; only noise amplitude changes.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+RESULT_DIR="${RESULT_DIR:-output/results}"
+WORK_DIR="${WORK_DIR:-output/work}"
+PLATFORM="${PLATFORM:-nvidia}"
+ARCH="${ARCH:-}"
+SEED="${SEED:-7}"
+SIZE="${SIZE:-1920x1080}"
+CHANNELS="${CHANNELS:-3}"
+
+mkdir -p "${RESULT_DIR}" "${WORK_DIR}"
+rm -f "${RESULT_DIR}/quality_tradeoff.csv" "${RESULT_DIR}/quality_sigma.csv"
+
+if [[ "${PLATFORM}" == "nvidia" ]]; then
+    make build PLATFORM="${PLATFORM}" ARCH="${ARCH}" -j"$(nproc)"
+else
+    make build PLATFORM="${PLATFORM}" -j"$(nproc)"
+fi
+BIN="./build/${PLATFORM}/nlm_denoise"
+
+CLEAN="${WORK_DIR}/clean_${SIZE}_${CHANNELS}ch.png"
+"${BIN}" gen -o "${CLEAN}" --size "${SIZE}" \
+    --channels "${CHANNELS}" --sigma 0 --seed "${SEED}"
+
+# A fair small/base/large comparison: same sigma=25 input, h=10, only pr/sr change.
+NOISY25="${WORK_DIR}/noisy_${SIZE}_${CHANNELS}ch_sigma25.png"
+"${BIN}" gen -o "${NOISY25}" --size "${SIZE}" \
+    --channels "${CHANNELS}" --sigma 25 --seed "${SEED}"
+"${BIN}" metrics --reference "${CLEAN}" --test "${NOISY25}" \
+    --label noisy-sigma25 --log "${RESULT_DIR}/quality_tradeoff.csv"
+
+for NAME in small base large; do
+    OUTPUT="${WORK_DIR}/denoised_${NAME}_${SIZE}_${CHANNELS}ch_sigma25.png"
+    PARAM_FILE="scripts/params/${NAME}.txt"
+    if [[ "${NAME}" == "base" ]]; then PARAM_FILE=params.txt; fi
+    "${BIN}" run -i "${NOISY25}" -o "${OUTPUT}" \
+        -p "${PARAM_FILE}" --kernel 2 --log /dev/null
+    "${BIN}" metrics --reference "${CLEAN}" --test "${OUTPUT}" \
+        --label "${NAME}" --log "${RESULT_DIR}/quality_tradeoff.csv"
+done
+
+# Noise-strength sweep: parameter sigma matches the generated input amplitude.
+for SIGMA in 10 25 50; do
+    NOISY="${WORK_DIR}/noisy_${SIZE}_${CHANNELS}ch_sigma${SIGMA}.png"
+    OUTPUT="${WORK_DIR}/denoised_base_${SIZE}_${CHANNELS}ch_sigma${SIGMA}.png"
+    "${BIN}" gen -o "${NOISY}" --size "${SIZE}" \
+        --channels "${CHANNELS}" --sigma "${SIGMA}" --seed "${SEED}"
+    "${BIN}" metrics --reference "${CLEAN}" --test "${NOISY}" \
+        --label "noisy-sigma${SIGMA}" --log "${RESULT_DIR}/quality_sigma.csv"
+    PARAM_FILE="scripts/params/sigma${SIGMA}.txt"
+    if [[ "${SIGMA}" == "25" ]]; then PARAM_FILE=params.txt; fi
+    "${BIN}" run -i "${NOISY}" -o "${OUTPUT}" \
+        -p "${PARAM_FILE}" --kernel 2 --log /dev/null
+    "${BIN}" metrics --reference "${CLEAN}" --test "${OUTPUT}" \
+        --label "denoised-sigma${SIGMA}" --log "${RESULT_DIR}/quality_sigma.csv"
+done
+
+echo "Quality results: ${RESULT_DIR}/quality_tradeoff.csv, ${RESULT_DIR}/quality_sigma.csv"
+echo "Generated images: ${WORK_DIR}/"
diff --git a/07_nlm_denoise/blackbook537/src/benchmark.cpp b/07_nlm_denoise/blackbook537/src/benchmark.cpp
new file mode 100644
index 00000000..60b855bd
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/benchmark.cpp
@@ -0,0 +1,243 @@
+// ============================================================================
+// bench 子命令：性能基准测试
+//   扫描 分辨率 x 通道数 x 参数组合 x kernel 版本，输出可复核的统计 CSV。
+//   GPU 每配置执行 warmup 次预热与 repeat 次采样，报告 mean/min/stddev；
+//   可选 CPU 基线对同一输入、同一参数独立采样，避免旧版跨参数错配。
+// ============================================================================
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <fstream>
+#include <iomanip>
+#include <string>
+#include <vector>
+
+#include "nlm_cpu_ref.h"
+#include "pipeline.h"
+#include "benchmark.h"
+
+namespace {
+
+struct SizeSpec {
+    int w;
+    int h;
+};
+
+struct ParamSet {
+    std::string name;
+    NlmParams p;
+};
+
+struct Stats {
+    double mean;
+    double min;
+    double stddev;
+};
+
+Stats Summarize(const std::vector<double>& samples) {
+    Stats out = {-1.0, -1.0, -1.0};
+    if (samples.empty()) return out;
+    out.min = *std::min_element(samples.begin(), samples.end());
+    double sum = 0.0;
+    for (size_t i = 0; i < samples.size(); ++i) sum += samples[i];
+    out.mean = sum / static_cast<double>(samples.size());
+    double squared = 0.0;
+    for (size_t i = 0; i < samples.size(); ++i) {
+        const double delta = samples[i] - out.mean;
+        squared += delta * delta;
+    }
+    out.stddev = std::sqrt(squared / static_cast<double>(samples.size()));
+    return out;
+}
+
+std::vector<SizeSpec> ParseSizes(const std::string& spec) {
+    std::vector<SizeSpec> out;
+    size_t pos = 0;
+    while (pos < spec.size()) {
+        const size_t comma = spec.find(',', pos);
+        const std::string tok = spec.substr(
+            pos, comma == std::string::npos ? comma : comma - pos);
+        int w = 0;
+        int h = 0;
+        if (std::sscanf(tok.c_str(), "%dx%d", &w, &h) == 2 && w > 0 && h > 0) {
+            out.push_back(SizeSpec{w, h});
+        }
+        pos = comma == std::string::npos ? spec.size() : comma + 1;
+    }
+    return out;
+}
+
+std::vector<int> ParseChannels(const std::string& spec) {
+    std::vector<int> out;
+    size_t pos = 0;
+    while (pos < spec.size()) {
+        const size_t comma = spec.find(',', pos);
+        const int c = std::atoi(spec.substr(
+            pos, comma == std::string::npos ? comma : comma - pos).c_str());
+        if (c == 1 || c == 3) out.push_back(c);
+        pos = comma == std::string::npos ? spec.size() : comma + 1;
+    }
+    return out;
+}
+
+bool ListContains(const std::string& spec, const std::string& name) {
+    if (spec.empty() || spec == "all") return true;
+    size_t pos = 0;
+    while (pos < spec.size()) {
+        const size_t comma = spec.find(',', pos);
+        const std::string token = spec.substr(
+            pos, comma == std::string::npos ? comma : comma - pos);
+        if (token == name) return true;
+        pos = comma == std::string::npos ? spec.size() : comma + 1;
+    }
+    return false;
+}
+
+std::vector<ParamSet> SelectParamSets(const std::string& spec) {
+    std::vector<ParamSet> all;
+
+    NlmParams small;
+    small.patch_radius = 2;
+    small.search_radius = 7;
+    small.h = 10.0f;
+    small.sigma = 25.0f;
+    all.push_back(ParamSet{"small", small});
+
+    NlmParams base;
+    all.push_back(ParamSet{"base", base});
+
+    NlmParams strong_h;
+    strong_h.h = 15.0f;
+    all.push_back(ParamSet{"strong-h", strong_h});
+
+    NlmParams large;
+    large.patch_radius = 4;
+    large.search_radius = 14;
+    large.h = 10.0f;
+    large.sigma = 25.0f;
+    all.push_back(ParamSet{"large", large});
+
+    std::vector<ParamSet> selected;
+    for (size_t i = 0; i < all.size(); ++i) {
+        if (ListContains(spec, all[i].name)) selected.push_back(all[i]);
+    }
+    return selected;
+}
+
+}  // namespace
+
+int RunBenchmark(const std::string& sizes_spec, const std::string& channels_spec,
+                 const std::string& param_sets_spec, const std::string& csv_path,
+                 int warmup, int repeat, bool with_cpu, int cpu_repeat) {
+    const std::vector<SizeSpec> sizes = ParseSizes(sizes_spec);
+    const std::vector<int> channels = ParseChannels(channels_spec);
+    const std::vector<ParamSet> sets = SelectParamSets(param_sets_spec);
+    if (sizes.empty() || channels.empty() || sets.empty()) {
+        std::fprintf(stderr, "[bench] 无效的 --sizes、--channels 或 --param-sets\n");
+        return 2;
+    }
+    if (warmup < 0) warmup = 0;
+    if (repeat < 1) repeat = 1;
+    if (cpu_repeat < 1) cpu_repeat = 1;
+
+    EnsureParentDir(csv_path.c_str());
+    std::ofstream csv(csv_path.c_str());
+    if (!csv) {
+        std::fprintf(stderr, "[bench] 无法写入 CSV: %s\n", csv_path.c_str());
+        return 2;
+    }
+    csv << "config,size,channels,pr,sr,h,sigma,kernel_ver,warmup,repeat,"
+           "kernel_mean_ms,kernel_min_ms,kernel_stddev_ms,"
+           "e2e_mean_ms,e2e_min_ms,e2e_stddev_ms,mpx_s,"
+           "cpu_repeat,cpu_mean_ms,cpu_min_ms,cpu_stddev_ms,speedup_e2e\n";
+    csv << std::fixed << std::setprecision(6);
+
+    std::printf("[bench] param_sets=%s warmup=%d repeat=%d cpu_repeat=%d -> %s\n",
+                param_sets_spec.c_str(), warmup, repeat, cpu_repeat, csv_path.c_str());
+    const int kernel_versions[] = {0, 1, 2};
+
+    for (size_t si = 0; si < sizes.size(); ++si) {
+        const int w = sizes[si].w;
+        const int h = sizes[si].h;
+        for (size_t ci = 0; ci < channels.size(); ++ci) {
+            const int c = channels[ci];
+            for (size_t pi = 0; pi < sets.size(); ++pi) {
+                const ParamSet& set = sets[pi];
+                const NlmParams& prm = set.p;
+                const unsigned seed = 777u + static_cast<unsigned>(c);
+                const ImageU8 src = MakeSyntheticImage(w, h, c, prm.sigma, seed);
+
+                Stats cpu_stats = {-1.0, -1.0, -1.0};
+                int measured_cpu_repeat = 0;
+                if (with_cpu && static_cast<size_t>(w) * h <= 1920u * 1080u) {
+                    std::vector<double> cpu_samples;
+                    for (int it = 0; it < cpu_repeat; ++it) {
+                        ImageU8 dst_cpu;
+                        HostTimer timer;
+                        timer.Start();
+                        NlmDenoiseCpuRef(src, &dst_cpu, prm);
+                        cpu_samples.push_back(timer.Ms());
+                    }
+                    cpu_stats = Summarize(cpu_samples);
+                    measured_cpu_repeat = cpu_repeat;
+                    std::printf("[bench] CPU %dx%dx%d %-8s: mean=%.1f min=%.1f std=%.1f ms\n",
+                                w, h, c, set.name.c_str(), cpu_stats.mean,
+                                cpu_stats.min, cpu_stats.stddev);
+                } else if (with_cpu) {
+                    std::printf("[bench] CPU %dx%dx%d %-8s: 跳过（超过 1080p 安全上限）\n",
+                                w, h, c, set.name.c_str());
+                }
+
+                for (size_t ki = 0;
+                     ki < sizeof(kernel_versions) / sizeof(kernel_versions[0]); ++ki) {
+                    const int ver = kernel_versions[ki];
+                    std::vector<double> kernel_samples;
+                    std::vector<double> e2e_samples;
+                    std::string err;
+                    for (int it = 0; it < warmup + repeat; ++it) {
+                        ImageU8 dst;
+                        NlmPerfReport perf;
+                        if (!NlmDenoiseGpu(src, &dst, prm, ver, &perf, &err)) {
+                            std::fprintf(stderr, "[bench] v%d 失败: %s\n", ver, err.c_str());
+                            return 2;
+                        }
+                        if (it >= warmup) {
+                            kernel_samples.push_back(perf.kernel_ms);
+                            e2e_samples.push_back(perf.e2e_ms);
+                        }
+                    }
+                    const Stats kernel_stats = Summarize(kernel_samples);
+                    const Stats e2e_stats = Summarize(e2e_samples);
+                    const double mpx_s = kernel_stats.mean > 0.0
+                                             ? static_cast<double>(w) * h /
+                                                   (kernel_stats.mean * 1e3)
+                                             : -1.0;
+                    const double speedup = cpu_stats.mean > 0.0
+                                               ? cpu_stats.mean / e2e_stats.mean
+                                               : -1.0;
+
+                    csv << set.name << ',' << w << 'x' << h << ',' << c << ','
+                        << prm.patch_radius << ',' << prm.search_radius << ',' << prm.h << ','
+                        << prm.sigma << ',' << ver << ',' << warmup << ',' << repeat << ','
+                        << kernel_stats.mean << ',' << kernel_stats.min << ','
+                        << kernel_stats.stddev << ',' << e2e_stats.mean << ','
+                        << e2e_stats.min << ',' << e2e_stats.stddev << ',' << mpx_s << ','
+                        << measured_cpu_repeat << ',' << cpu_stats.mean << ',' << cpu_stats.min
+                        << ',' << cpu_stats.stddev << ',' << speedup << '\n';
+
+                    std::printf(
+                        "[bench] %dx%dx%d %-8s v%d: kernel %.3f±%.3f ms "
+                        "e2e %.3f±%.3f ms %.1f Mpx/s%s\n",
+                        w, h, c, set.name.c_str(), ver, kernel_stats.mean,
+                        kernel_stats.stddev, e2e_stats.mean, e2e_stats.stddev, mpx_s,
+                        speedup > 0.0
+                            ? (std::string(" speedup=") + std::to_string(speedup) + "x").c_str()
+                            : "");
+                }
+            }
+        }
+    }
+    std::printf("[bench] 完成，CSV 已写入 %s\n", csv_path.c_str());
+    return 0;
+}
diff --git a/07_nlm_denoise/blackbook537/src/image_io.cpp b/07_nlm_denoise/blackbook537/src/image_io.cpp
new file mode 100644
index 00000000..67f74f0e
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/image_io.cpp
@@ -0,0 +1,109 @@
+#include "image_io.h"
+
+#include <cstring>
+#include <sys/stat.h>
+#include <sys/types.h>
+
+// stb_image / stb_image_write：单头文件图像编解码库（零外部依赖，便于多平台移植）
+#define STB_IMAGE_IMPLEMENTATION
+#define STBI_FAILURE_USERMSG
+#include "stb_image.h"
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#include "stb_image_write.h"
+
+namespace {
+
+std::string LowerExt(const char* path) {
+    std::string p(path);
+    size_t dot = p.find_last_of('.');
+    if (dot == std::string::npos) return "";
+    std::string ext = p.substr(dot + 1);
+    for (size_t i = 0; i < ext.size(); ++i)
+        if (ext[i] >= 'A' && ext[i] <= 'Z') ext[i] = (char)(ext[i] - 'A' + 'a');
+    return ext;
+}
+
+} // namespace
+
+void EnsureParentDir(const char* path) {
+    if (!path) return;
+    const std::string p(path);
+    const size_t slash = p.find_last_of('/');
+    if (slash == std::string::npos || slash == 0) return;  // 无目录或根目录
+    // 逐级创建（已存在时 mkdir 报 EEXIST，直接忽略）
+    for (size_t i = 1; i < slash; ++i) {
+        if (p[i] == '/') {
+            std::string part = p.substr(0, i);
+            mkdir(part.c_str(), 0755);
+        }
+    }
+    mkdir(p.substr(0, slash).c_str(), 0755);
+}
+
+bool LoadImage(const char* path, ImageU8* out, std::string* err) {
+    if (!path || !out) {
+        if (err) *err = "LoadImage: 空指针参数";
+        return false;
+    }
+    int w = 0, h = 0, c = 0;
+    unsigned char* raw = stbi_load(path, &w, &h, &c, 0);
+    if (!raw) {
+        if (err) *err = std::string("图像解码失败: ") + path + " (" + stbi_failure_reason() + ")";
+        return false;
+    }
+    // 灰度+alpha / RGBA：重新解码并丢弃 alpha，保证通道语义 ∈ {1,3}
+    if (c == 2 || c == 4) {
+        stbi_image_free(raw);
+        const int want = (c == 2) ? 1 : 3;
+        raw = stbi_load(path, &w, &h, &c, want);
+        if (!raw) {
+            if (err) *err = std::string("图像解码失败(重试): ") + path;
+            return false;
+        }
+        c = want;
+    }
+    if (c != 1 && c != 3) {
+        stbi_image_free(raw);
+        if (err) *err = "仅支持灰度或 RGB 图像，实际通道数: " + std::to_string(c);
+        return false;
+    }
+
+    out->width = w;
+    out->height = h;
+    out->channels = c;
+    out->data.assign(raw, raw + (size_t)w * h * c);
+    stbi_image_free(raw);
+    return true;
+}
+
+bool SaveImage(const char* path, const ImageU8& img, std::string* err) {
+    if (!path) {
+        if (err) *err = "SaveImage: 空路径";
+        return false;
+    }
+    if (img.width <= 0 || img.height <= 0 ||
+        (img.channels != 1 && img.channels != 3) ||
+        img.data.size() != (size_t)img.width * img.height * img.channels) {
+        if (err) *err = "SaveImage: 非法图像对象";
+        return false;
+    }
+    EnsureParentDir(path);  // 数据集中管理：输出目录不存在时自动创建
+    const std::string ext = LowerExt(path);
+    int ok = 0;
+    if (ext == "jpg" || ext == "jpeg") {
+        ok = stbi_write_jpg(path, img.width, img.height, img.channels, img.data.data(), 95);
+    } else if (ext == "bmp") {
+        ok = stbi_write_bmp(path, img.width, img.height, img.channels, img.data.data());
+    } else if (ext == "tga") {
+        ok = stbi_write_tga(path, img.width, img.height, img.channels, img.data.data());
+    } else {
+        // 缺省 / .png
+        ok = stbi_write_png(path, img.width, img.height, img.channels,
+                            img.data.data(), img.width * img.channels);
+    }
+    if (!ok) {
+        if (err) *err = std::string("图像编码失败: ") + path;
+        return false;
+    }
+    return true;
+}
diff --git a/07_nlm_denoise/blackbook537/src/kernels.cu b/07_nlm_denoise/blackbook537/src/kernels.cu
new file mode 100644
index 00000000..898f1d9a
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/kernels.cu
@@ -0,0 +1,376 @@
+// ============================================================================
+// NLM 全部设备端实现与 Host 端调度（算法核心，多平台 100% 共享）
+// NVIDIA / Iluvatar 直接编译本文件，MACA / MUSA 入口包含本文件，
+// 平台差异仅通过 platform_api.h 中的 PAL 宏适配。
+// ============================================================================
+#include "kernels.h"
+#include "platform_api.h"
+
+namespace nlm {
+
+// ---------------------------------------------------------------------------
+// 设备端小工具
+// ---------------------------------------------------------------------------
+__device__ __forceinline__ int ClampI(int v, int lo, int hi) {
+    return v < lo ? lo : (v > hi ? hi : v);
+}
+
+// NLM 权重：w = exp(-max(dist - 2*sigma^2*N_patch, 0) / h^2)
+// NLM_FAST_EXP（编译期开关）：使用 __expf 快速指数，误差写入报告误差来源分析
+__device__ __forceinline__ float NlmWeight(float dist, float ts2n, float inv_h2) {
+    const float d = fmaxf(dist - ts2n, 0.0f);
+#ifdef NLM_FAST_EXP
+    return __expf(-d * inv_h2);
+#else
+    return expf(-d * inv_h2);
+#endif
+}
+
+// ---------------------------------------------------------------------------
+// 数据转换 kernel：interleaved u8 <-> planar f32
+// ---------------------------------------------------------------------------
+__global__ void CvtU8ToF32PlanarKernel(const uint8_t* __restrict__ in,
+                                       float* __restrict__ out,
+                                       int w, int h, int c) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int n = w * h * c;
+    if (idx >= n) return;
+    const int ch = idx / (w * h);          // planar 通道
+    const int rem = idx - ch * (w * h);    // 通道内像素偏移
+    out[(size_t)ch * w * h + rem] = (float)in[(size_t)rem * c + ch];
+}
+
+__global__ void CvtF32ToU8InterleavedKernel(const float* __restrict__ in,
+                                            uint8_t* __restrict__ out,
+                                            int w, int h, int c) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int n = w * h * c;
+    if (idx >= n) return;
+    const int ch = idx / (w * h);
+    const int rem = idx - ch * (w * h);
+    const float v = in[(size_t)ch * w * h + rem];
+    int q = (int)(v + 0.5f);               // round
+    q = q < 0 ? 0 : (q > 255 ? 255 : q);   // clamp [0,255]
+    out[(size_t)rem * c + ch] = (uint8_t)q;
+}
+
+// ---------------------------------------------------------------------------
+// V0 naive kernel：1 thread = 1 output pixel，全部走全局内存
+// 边界语义：对每次最终访问坐标独立 clamp（等价 OpenCV BORDER_REPLICATE），
+// 与 CPU 参考实现严格一致。
+// ---------------------------------------------------------------------------
+__global__ void NlmNaiveKernel(const float* __restrict__ src,
+                               float* __restrict__ dst,
+                               int w, int h, int c, NlmParamsDev p) {
+    const int x = blockIdx.x * blockDim.x + threadIdx.x;
+    const int y = blockIdx.y * blockDim.y + threadIdx.y;
+    if (x >= w || y >= h) return;
+
+    const int pr = p.pr, sr = p.sr;
+    const size_t plane = (size_t)w * h;
+    float sum_w = 0.0f;
+    float sum[3] = {0.0f, 0.0f, 0.0f};
+
+    for (int dy = -sr; dy <= sr; ++dy) {
+        for (int dx = -sr; dx <= sr; ++dx) {
+            const int qx = x + dx, qy = y + dy;
+            // ---- patch L2 距离（跨通道求和，共享权重）----
+            float dist = 0.0f;
+            for (int py = -pr; py <= pr; ++py) {
+                const int ay = ClampI(y + py, 0, h - 1);
+                const int by = ClampI(qy + py, 0, h - 1);
+                for (int px = -pr; px <= pr; ++px) {
+                    const int ax = ClampI(x + px, 0, w - 1);
+                    const int bx = ClampI(qx + px, 0, w - 1);
+                    for (int ch = 0; ch < c; ++ch) {
+                        const float d = src[ch * plane + (size_t)ay * w + ax] -
+                                        src[ch * plane + (size_t)by * w + bx];
+                        dist += d * d;
+                    }
+                }
+            }
+            const float wgt = NlmWeight(dist, p.two_sigma2_n, p.inv_h2);
+            sum_w += wgt;
+            const int cqx = ClampI(qx, 0, w - 1);
+            const int cqy = ClampI(qy, 0, h - 1);
+            for (int ch = 0; ch < c; ++ch)
+                sum[ch] += wgt * src[ch * plane + (size_t)cqy * w + cqx];
+        }
+    }
+
+    // sum_w 恒 >= 1（自身权重 w(p,p)=exp(0)=1），无除零风险
+    const float inv = 1.0f / sum_w;
+    for (int ch = 0; ch < c; ++ch)
+        dst[ch * plane + (size_t)y * w + x] = sum[ch] * inv;
+}
+
+// ---------------------------------------------------------------------------
+// V1 smem kernel：block 计算 16x16 输出 tile，halo 区域（含搜索窗+patch 半径）
+// 协同加载进 shared memory，patch 距离全部命中 smem。
+//
+// 关键性质（正确性论证）：smem[l] = src[clamp(b0 - radius + l)]，
+// 而线程中心局部坐标 lc 满足 b0 - radius + lc == 像素全局坐标，
+// 因此 smem[lc + 偏移] == src[clamp(全局坐标 + 偏移)]，
+// 与 CPU 参考的"最终坐标独立 clamp"语义逐位一致，内层循环无需任何 clamp。
+// ---------------------------------------------------------------------------
+__global__ void NlmSmemKernel(const float* __restrict__ src,
+                              float* __restrict__ dst,
+                              int w, int h, int c, NlmParamsDev p) {
+    extern __shared__ float smem[];
+    const int pr = p.pr, sr = p.sr;
+    const int radius = sr + pr;
+    const int tileW = blockDim.x + 2 * radius;
+    const int tileH = blockDim.y + 2 * radius;
+    const int stride = tileW + 1;            // padding 消除 bank conflict
+    const int slab = tileH * stride;         // 每通道 smem 元素数
+    const int bx0 = blockIdx.x * blockDim.x;
+    const int by0 = blockIdx.y * blockDim.y;
+    const size_t plane = (size_t)w * h;
+
+    // 协同加载 halo tile（复制边界）
+    const int tid = threadIdx.y * blockDim.x + threadIdx.x;
+    const int nthreads = blockDim.x * blockDim.y;
+    for (int ch = 0; ch < c; ++ch) {
+        for (int idx = tid; idx < tileW * tileH; idx += nthreads) {
+            const int lx = idx % tileW;
+            const int ly = idx / tileW;
+            const int gx = ClampI(bx0 - radius + lx, 0, w - 1);
+            const int gy = ClampI(by0 - radius + ly, 0, h - 1);
+            smem[ch * slab + ly * stride + lx] = src[ch * plane + (size_t)gy * w + gx];
+        }
+    }
+    __syncthreads();
+
+    const int x = bx0 + threadIdx.x;
+    const int y = by0 + threadIdx.y;
+    if (x >= w || y >= h) return;
+
+    const int lcx = threadIdx.x + radius;
+    const int lcy = threadIdx.y + radius;
+    float sum_w = 0.0f;
+    float sum[3] = {0.0f, 0.0f, 0.0f};
+
+    for (int dy = -sr; dy <= sr; ++dy) {
+        const int qly = lcy + dy;
+        for (int dx = -sr; dx <= sr; ++dx) {
+            const int qlx = lcx + dx;
+            float dist = 0.0f;
+            for (int py = -pr; py <= pr; ++py) {
+                const int rowA = (lcy + py) * stride + lcx;
+                const int rowB = (qly + py) * stride + qlx;
+                for (int px = -pr; px <= pr; ++px) {
+                    for (int ch = 0; ch < c; ++ch) {
+                        const float d = smem[ch * slab + rowA + px] -
+                                        smem[ch * slab + rowB + px];
+                        dist += d * d;
+                    }
+                }
+            }
+            const float wgt = NlmWeight(dist, p.two_sigma2_n, p.inv_h2);
+            sum_w += wgt;
+            for (int ch = 0; ch < c; ++ch)
+                sum[ch] += wgt * smem[ch * slab + qly * stride + qlx];
+        }
+    }
+
+    const float inv = 1.0f / sum_w;
+    for (int ch = 0; ch < c; ++ch)
+        dst[ch * plane + (size_t)y * w + x] = sum[ch] * inv;
+}
+
+// ---------------------------------------------------------------------------
+// V2 kernel：V1 + patch 半径模板化（#pragma unroll 完全展开内层循环）
+// 与 V1 语义逐位一致，仅通过编译期常量消除循环开销、提升指令级并行。
+// ---------------------------------------------------------------------------
+template <int PR>
+__global__ __launch_bounds__(256) void NlmSmemUnrollKernel(
+        const float* __restrict__ src, float* __restrict__ dst,
+        int w, int h, int c, NlmParamsDev p) {
+    extern __shared__ float smem[];
+    const int pr = PR;
+    const int sr = p.sr;
+    const int radius = sr + pr;
+    const int tileW = blockDim.x + 2 * radius;
+    const int tileH = blockDim.y + 2 * radius;
+    const int stride = tileW + 1;
+    const int slab = tileH * stride;
+    const int bx0 = blockIdx.x * blockDim.x;
+    const int by0 = blockIdx.y * blockDim.y;
+    const size_t plane = (size_t)w * h;
+
+    const int tid = threadIdx.y * blockDim.x + threadIdx.x;
+    const int nthreads = blockDim.x * blockDim.y;
+    for (int ch = 0; ch < c; ++ch) {
+        for (int idx = tid; idx < tileW * tileH; idx += nthreads) {
+            const int lx = idx % tileW;
+            const int ly = idx / tileW;
+            const int gx = ClampI(bx0 - radius + lx, 0, w - 1);
+            const int gy = ClampI(by0 - radius + ly, 0, h - 1);
+            smem[ch * slab + ly * stride + lx] = src[ch * plane + (size_t)gy * w + gx];
+        }
+    }
+    __syncthreads();
+
+    const int x = bx0 + threadIdx.x;
+    const int y = by0 + threadIdx.y;
+    if (x >= w || y >= h) return;
+
+    const int lcx = threadIdx.x + radius;
+    const int lcy = threadIdx.y + radius;
+    float sum_w = 0.0f;
+    float sum[3] = {0.0f, 0.0f, 0.0f};
+
+    for (int dy = -sr; dy <= sr; ++dy) {
+        const int qly = lcy + dy;
+        for (int dx = -sr; dx <= sr; ++dx) {
+            const int qlx = lcx + dx;
+            float dist = 0.0f;
+#pragma unroll
+            for (int py = -PR; py <= PR; ++py) {
+                const int rowA = (lcy + py) * stride + lcx;
+                const int rowB = (qly + py) * stride + qlx;
+#pragma unroll
+                for (int px = -PR; px <= PR; ++px) {
+                    for (int ch = 0; ch < c; ++ch) {
+                        const float d = smem[ch * slab + rowA + px] -
+                                        smem[ch * slab + rowB + px];
+                        dist += d * d;
+                    }
+                }
+            }
+            const float wgt = NlmWeight(dist, p.two_sigma2_n, p.inv_h2);
+            sum_w += wgt;
+            for (int ch = 0; ch < c; ++ch)
+                sum[ch] += wgt * smem[ch * slab + qly * stride + qlx];
+        }
+    }
+
+    const float inv = 1.0f / sum_w;
+    for (int ch = 0; ch < c; ++ch)
+        dst[ch * plane + (size_t)y * w + x] = sum[ch] * inv;
+}
+
+} // namespace nlm
+
+// ===========================================================================
+// Host 端调度接口（对 pipeline 暴露）
+// ===========================================================================
+bool GpuAlloc(void** ptr, size_t bytes, std::string* err) {
+    if (!ptr || bytes == 0) {
+        if (err) *err = "GpuAlloc: 非法参数";
+        return false;
+    }
+    GPU_CHECK(GPU_MALLOC(ptr, bytes), err);
+    return true;
+}
+
+bool GpuFree(void* ptr, std::string* err) {
+    if (!ptr) return true;
+    GPU_CHECK(GPU_FREE(ptr), err);
+    return true;
+}
+
+bool GpuCopyH2D(void* dst, const void* src, size_t bytes, std::string* err) {
+    GPU_CHECK(GPU_MEMCPY(dst, src, bytes, GPU_MEMCPY_H2D), err);
+    return true;
+}
+
+bool GpuCopyD2H(void* dst, const void* src, size_t bytes, std::string* err) {
+    GPU_CHECK(GPU_MEMCPY(dst, src, bytes, GPU_MEMCPY_D2H), err);
+    return true;
+}
+
+bool GpuCvtU8ToF32Planar(const uint8_t* d_in, float* d_out, int w, int h, int c,
+                         std::string* err) {
+    const int n = w * h * c;
+    const int block = 256;
+    const int grid = (n + block - 1) / block;
+    nlm::CvtU8ToF32PlanarKernel<<<grid, block>>>(d_in, d_out, w, h, c);
+    GPU_CHECK(GPU_GET_LAST_ERROR(), err);
+    return true;
+}
+
+bool GpuCvtF32ToU8Interleaved(const float* d_in, uint8_t* d_out, int w, int h, int c,
+                              std::string* err) {
+    const int n = w * h * c;
+    const int block = 256;
+    const int grid = (n + block - 1) / block;
+    nlm::CvtF32ToU8InterleavedKernel<<<grid, block>>>(d_in, d_out, w, h, c);
+    GPU_CHECK(GPU_GET_LAST_ERROR(), err);
+    return true;
+}
+
+bool GpuNlm(const float* d_src, float* d_dst, int w, int h, int c,
+            const NlmParamsDev& p, int version, float* kernel_ms, std::string* err) {
+    typedef void (*NlmKernelFn)(const float*, float*, int, int, int, NlmParamsDev);
+
+    const dim3 block(16, 16);
+    const dim3 grid((w + (int)block.x - 1) / (int)block.x,
+                    (h + (int)block.y - 1) / (int)block.y, 1);
+
+    // smem 需求：c 个通道，每通道 (tileH) x (tileW+1) 个 float
+    const int radius = p.sr + p.pr;
+    const int tileW = (int)block.x + 2 * radius;
+    const int tileH = (int)block.y + 2 * radius;
+    const size_t smem_bytes = (size_t)c * tileH * (tileW + 1) * sizeof(float);
+
+    NlmKernelFn fn = NULL;
+    bool use_smem = false;
+
+    if (version == 0) {
+        fn = nlm::NlmNaiveKernel;
+    } else if (version == 1 || version == 2) {
+        // 查询设备动态 smem 上限，不足则回退 naive（保证功能正确）
+        int dev = 0, smem_optin = 48 * 1024;
+        GPU_CHECK(GPU_GET_DEVICE(&dev), err);
+        GPU_CHECK(GPU_GET_SMEM_OPTIN(dev, &smem_optin), err);
+        if (smem_bytes <= (size_t)smem_optin) {
+            use_smem = true;
+            if (version == 1) {
+                fn = nlm::NlmSmemKernel;
+            } else {
+                // V2：按 patch 半径模板分派。
+                // 实测（RTX 3060, bench.csv）：pr<=3 时完全展开有 5%~30% 收益；
+                // pr=4 时 9x9x3 内层全展开导致寄存器溢出、占用率塌陷，
+                // 性能反降 ~70%（如 4K RGB large 参数 5540ms -> 9472ms），
+                // 因此 pr>=4 一律退回 V1 通用循环路径。
+                switch (p.pr) {
+                    case 1: fn = nlm::NlmSmemUnrollKernel<1>; break;
+                    case 2: fn = nlm::NlmSmemUnrollKernel<2>; break;
+                    case 3: fn = nlm::NlmSmemUnrollKernel<3>; break;
+                    default: fn = nlm::NlmSmemKernel; break;
+                }
+            }
+            if (smem_bytes > 48 * 1024) {
+                GPU_CHECK(GPU_FUNC_SET_MAX_SMEM(fn, smem_bytes), err);
+            }
+        } else {
+            fn = nlm::NlmNaiveKernel;  // smem 不足，回退
+        }
+    } else {
+        if (err) *err = "kernel version=" + std::to_string(version) +
+                        " 未实现（V3 为预留进阶版本）";
+        return false;
+    }
+
+    gpuEvent_t e0, e1;
+    GPU_CHECK(GPU_EVENT_CREATE(&e0), err);
+    GPU_CHECK(GPU_EVENT_CREATE(&e1), err);
+
+    GPU_CHECK(GPU_EVENT_RECORD(e0), err);
+    if (use_smem) {
+        fn<<<grid, block, smem_bytes>>>(d_src, d_dst, w, h, c, p);
+    } else {
+        fn<<<grid, block>>>(d_src, d_dst, w, h, c, p);
+    }
+    GPU_CHECK(GPU_GET_LAST_ERROR(), err);
+    GPU_CHECK(GPU_EVENT_RECORD(e1), err);
+    GPU_CHECK(GPU_EVENT_SYNC(e1), err);
+
+    float ms = 0.0f;
+    GPU_CHECK(GPU_EVENT_ELAPSED(&ms, e0, e1), err);
+    GPU_EVENT_DESTROY(e0);
+    GPU_EVENT_DESTROY(e1);
+    if (kernel_ms) *kernel_ms = ms;
+    return true;
+}
diff --git a/07_nlm_denoise/blackbook537/src/kernels.maca b/07_nlm_denoise/blackbook537/src/kernels.maca
new file mode 100644
index 00000000..6a66535a
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/kernels.maca
@@ -0,0 +1,2 @@
+// MetaX MACA 编译入口；共享算法，平台 API 由编译宏选择。
+#include "kernels.cu"
diff --git a/07_nlm_denoise/blackbook537/src/kernels.mu b/07_nlm_denoise/blackbook537/src/kernels.mu
new file mode 100644
index 00000000..ef800a9c
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/kernels.mu
@@ -0,0 +1,2 @@
+// Moore Threads MUSA 编译入口；共享算法，保持 C++11 兼容。
+#include "kernels.cu"
diff --git a/07_nlm_denoise/blackbook537/src/main.cpp b/07_nlm_denoise/blackbook537/src/main.cpp
new file mode 100644
index 00000000..838c2884
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/main.cpp
@@ -0,0 +1,231 @@
+// ============================================================================
+// nlm_denoise —— 实时图像非局部均值降噪（CUDA）CLI 入口
+//
+// 子命令：
+//   run      -i input.png -o output.png -p params.txt [--kernel v] [--log f]
+//   validate -i input.png [-o output.png] -p params.txt
+//   bench    [--sizes 1920x1080,3840x2160] [--channels 1,3]
+//            [--param-sets all|small,base,strong-h,large] [--log results.csv]
+//            [--warmup 3] [--repeat 10] [--with-cpu] [--cpu-repeat 1]
+//   metrics  --reference clean.png --test output.png [--label name] [--log quality.csv]
+//   gen      -o noisy.png [--size 1920x1080] [--channels 3] [--sigma 25] [--seed 1]
+//   units    （Host 侧单元测试，无需 GPU）
+// ============================================================================
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+
+#include "image_io.h"
+#include "nlm_cpu_ref.h"
+#include "params.h"
+#include "pipeline.h"
+#include "benchmark.h"
+
+namespace {
+
+void PrintUsage() {
+    std::printf(
+        "用法: nlm_denoise <子命令> [选项]\n"
+        "\n"
+        "子命令:\n"
+        "  run       对单张图像执行 GPU NLM 降噪并输出性能日志\n"
+        "  validate  与 CPU 参考（及可选 OpenCV）对比，输出 MAE/PSNR 校验\n"
+        "  bench     分辨率 x 通道 x 参数组合 x kernel 版本性能扫描（CSV）\n"
+        "  metrics   计算测试图相对参考图的 MAE/PSNR，可追加写入 CSV\n"
+        "  gen       生成合成含噪测试图\n"
+        "  units     运行 Host 侧单元测试（无需 GPU）\n"
+        "\n"
+        "run 选项:\n"
+        "  -i <file>     输入图像（PNG/JPG，灰度或 RGB）\n"
+        "  -o <file>     输出图像（建议 output/images/，目录不存在时自动创建）\n"
+        "  -p <file>     参数文件（默认 params.txt）\n"
+        "  --kernel <v>  kernel 版本 0=naive 1=smem 2=smem+unroll（默认 2）\n"
+        "  --log <file>  性能日志文件（默认 output/logs/nlm_perf.log，追加写）\n"
+        "  --with-cpu    同时运行 CPU 参考以计算加速比\n"
+        "\n"
+        "bench 选项:\n"
+        "  --sizes <list>      尺寸列表（默认 1920x1080）\n"
+        "  --channels <list>   1,3（默认 1,3）\n"
+        "  --cpu-repeat <n> CPU 重复次数（默认 1）\n"
+        "  --param-sets <list> all 或 small,base,strong-h,large（默认 all）\n"
+        "  --with-cpu          CPU 统计（仅支持不超过 1080p）\n"
+        "\n"
+        "metrics 选项:\n"
+        "  --reference <file>  干净参考图（必填）\n"
+        "  --test <file>       待评估图（必填）\n"
+        "  --label <name>      CSV 行标签（默认 unnamed）\n"
+        "  --log <file>        可选 CSV 输出\n"
+        "\n"
+        "gen 选项:\n"
+        "  -o <file>     输出图像（缺省按规范自动命名：output/generated/noisy_<尺寸>_<c>ch_sigma<σ>.png）\n");
+}
+
+// 解析 "-i xx / --key value / --flag" 到 map（flag 型参数值为 "1"）
+std::map<std::string, std::string> ParseArgs(int argc, char** argv, int start) {
+    std::map<std::string, std::string> m;
+    for (int i = start; i < argc; ++i) {
+        std::string a = argv[i];
+        if (a.size() > 0 && a[0] == '-') {
+            if (i + 1 < argc && argv[i + 1][0] != '-') {
+                m[a] = argv[++i];
+            } else {
+                m[a] = "1";  // flag
+            }
+        }
+    }
+    return m;
+}
+
+std::string GetArg(const std::map<std::string, std::string>& m,
+                   const char* key, const char* dflt) {
+    std::map<std::string, std::string>::const_iterator it = m.find(key);
+    return it == m.end() ? std::string(dflt) : it->second;
+}
+
+int CmdRun(const std::map<std::string, std::string>& args) {
+    const std::string input = GetArg(args, "-i", "");
+    const std::string output = GetArg(args, "-o", "");
+    const std::string params_path = GetArg(args, "-p", "params.txt");
+    const std::string log_path = GetArg(args, "--log", "output/logs/nlm_perf.log");
+    const int kernel_ver = std::atoi(GetArg(args, "--kernel", "2").c_str());
+    const bool with_cpu = args.count("--with-cpu") > 0;
+
+    if (input.empty() || output.empty()) {
+        std::fprintf(stderr, "[run] 必须指定 -i 与 -o\n");
+        return 2;
+    }
+
+    std::string err;
+    ImageU8 src;
+    if (!LoadImage(input.c_str(), &src, &err)) {
+        std::fprintf(stderr, "[run] %s\n", err.c_str());
+        return 2;
+    }
+    NlmParams params;
+    if (!ParseParams(params_path.c_str(), &params, &err)) {
+        std::fprintf(stderr, "[run] %s\n", err.c_str());
+        return 2;
+    }
+
+    ImageU8 dst;
+    NlmPerfReport perf;
+    if (!NlmDenoiseGpu(src, &dst, params, kernel_ver, &perf, &err)) {
+        std::fprintf(stderr, "[run] %s\n", err.c_str());
+        return 2;
+    }
+    if (!SaveImage(output.c_str(), dst, &err)) {
+        std::fprintf(stderr, "[run] %s\n", err.c_str());
+        return 2;
+    }
+
+    if (with_cpu) {
+        ImageU8 cpu;
+        HostTimer t;
+        t.Start();
+        NlmDenoiseCpuRef(src, &cpu, params);
+        perf.cpu_ref_ms = (float)t.Ms();
+        perf.speedup_vs_cpu = perf.cpu_ref_ms / perf.e2e_ms;
+        std::printf("[run] 对 CPU 参考: MAE=%.4f  PSNR=%.2f dB\n",
+                    ComputeMAE(dst, cpu), ComputePSNR(dst, cpu));
+    }
+
+    std::printf("[run] %dx%dx%d  kernel v%d  pr=%d sr=%d h=%.1f sigma=%.1f\n",
+                src.width, src.height, src.channels, kernel_ver,
+                params.patch_radius, params.search_radius, params.h, params.sigma);
+    std::printf("[run] kernel=%.3f ms  e2e=%.3f ms (H2D %.3f / D2H %.3f)  %.2f Mpx/s\n",
+                perf.kernel_ms, perf.e2e_ms, perf.h2d_ms, perf.d2h_ms,
+                perf.throughput_mpx_s);
+    if (perf.cpu_ref_ms > 0.0f)
+        std::printf("[run] CPU=%.1f ms  加速比=%.1fx\n", perf.cpu_ref_ms, perf.speedup_vs_cpu);
+
+    // 性能日志（CSV 行，追加写）；目录不存在时自动创建（数据集中管理）
+    EnsureParentDir(log_path.c_str());
+    std::ofstream log(log_path.c_str(), std::ios::app);
+    if (log) {
+        log << src.width << "x" << src.height << ',' << src.channels << ','
+            << params.patch_radius << ',' << params.search_radius << ','
+            << params.h << ',' << params.sigma << ','
+            << kernel_ver << ',' << perf.kernel_ms << ',' << perf.e2e_ms << ','
+            << perf.throughput_mpx_s << ','
+            << perf.cpu_ref_ms << ',' << perf.speedup_vs_cpu << '\n';
+    }
+    return 0;
+}
+
+int CmdGen(const std::map<std::string, std::string>& args) {
+    const std::string size = GetArg(args, "--size", "1920x1080");
+    const int channels = std::atoi(GetArg(args, "--channels", "3").c_str());
+    const float sigma = (float)std::atof(GetArg(args, "--sigma", "25").c_str());
+    const unsigned seed = (unsigned)std::atoi(GetArg(args, "--seed", "1").c_str());
+    int w = 0, h = 0;
+    if (std::sscanf(size.c_str(), "%dx%d", &w, &h) != 2 || w <= 0 || h <= 0 ||
+        (channels != 1 && channels != 3)) {
+        std::fprintf(stderr, "[gen] 非法 --size/--channels\n");
+        return 2;
+    }
+    // 缺省输出：按命名规范自动生成 output/generated/noisy_<尺寸>_<c>ch_sigma<σ>.png
+    char auto_name[256];
+    std::snprintf(auto_name, sizeof(auto_name),
+                  "output/generated/noisy_%dx%d_%dch_sigma%d.png", w, h, channels,
+                  (int)(sigma + 0.5f));
+    const std::string output = GetArg(args, "-o", auto_name);
+    ImageU8 img = MakeSyntheticImage(w, h, channels, sigma, seed);
+    std::string err;
+    if (!SaveImage(output.c_str(), img, &err)) {
+        std::fprintf(stderr, "[gen] %s\n", err.c_str());
+        return 2;
+    }
+    std::printf("[gen] 已生成 %dx%dx%d sigma=%.1f -> %s\n", w, h, channels, sigma, output.c_str());
+    return 0;
+}
+
+} // namespace
+
+int main(int argc, char** argv) {
+    if (argc < 2) {
+        PrintUsage();
+        return 2;
+    }
+    const std::string cmd = argv[1];
+    const std::map<std::string, std::string> args = ParseArgs(argc, argv, 2);
+
+    if (cmd == "run") {
+        return CmdRun(args);
+    } else if (cmd == "validate") {
+        const std::string input = GetArg(args, "-i", "");
+        if (input.empty()) {
+            std::fprintf(stderr, "[validate] 必须指定 -i\n");
+            return 2;
+        }
+        return RunValidate(input, GetArg(args, "-o", ""),
+                           GetArg(args, "-p", "params.txt"));
+    } else if (cmd == "bench") {
+        return RunBenchmark(GetArg(args, "--sizes", "1920x1080"),
+                            GetArg(args, "--channels", "1,3"),
+                            GetArg(args, "--param-sets", "all"),
+                            GetArg(args, "--log", "output/results/benchmark.csv"),
+                            std::atoi(GetArg(args, "--warmup", "3").c_str()),
+                            std::atoi(GetArg(args, "--repeat", "10").c_str()),
+                            args.count("--with-cpu") > 0,
+                            std::atoi(GetArg(args, "--cpu-repeat", "1").c_str()));
+    } else if (cmd == "metrics") {
+        const std::string reference = GetArg(args, "--reference", "");
+        const std::string test = GetArg(args, "--test", "");
+        if (reference.empty() || test.empty()) {
+            std::fprintf(stderr, "[metrics] 必须指定 --reference 与 --test\n");
+            return 2;
+        }
+        return RunMetrics(reference, test, GetArg(args, "--label", "unnamed"),
+                          GetArg(args, "--log", ""));
+    } else if (cmd == "gen") {
+        return CmdGen(args);
+    } else if (cmd == "units") {
+        return RunUnitTests();
+    }
+
+    std::fprintf(stderr, "未知子命令: %s\n\n", cmd.c_str());
+    PrintUsage();
+    return 2;
+}
diff --git a/07_nlm_denoise/blackbook537/src/metrics.cpp b/07_nlm_denoise/blackbook537/src/metrics.cpp
new file mode 100644
index 00000000..66b63227
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/metrics.cpp
@@ -0,0 +1,66 @@
+// ============================================================================
+// metrics 子命令：对同尺寸图像计算 MAE/PSNR，并可追加写入结构化 CSV。
+// 该路径只读取图像，不依赖 GPU，便于在服务器与开发机之间复核质量结果。
+// ============================================================================
+#include <cstdio>
+#include <fstream>
+#include <iomanip>
+#include <string>
+
+#include "image_io.h"
+#include "benchmark.h"
+
+namespace {
+
+bool FileHasContent(const std::string& path) {
+    std::ifstream in(path.c_str(), std::ios::binary | std::ios::ate);
+    return in && in.tellg() > 0;
+}
+
+}  // namespace
+
+int RunMetrics(const std::string& reference_path, const std::string& test_path,
+               const std::string& label, const std::string& csv_path) {
+    std::string err;
+    ImageU8 reference;
+    ImageU8 test;
+    if (!LoadImage(reference_path.c_str(), &reference, &err)) {
+        std::fprintf(stderr, "[metrics] reference: %s\n", err.c_str());
+        return 2;
+    }
+    if (!LoadImage(test_path.c_str(), &test, &err)) {
+        std::fprintf(stderr, "[metrics] test: %s\n", err.c_str());
+        return 2;
+    }
+    if (reference.width != test.width || reference.height != test.height ||
+        reference.channels != test.channels) {
+        std::fprintf(stderr,
+                     "[metrics] 尺寸/通道不一致: reference=%dx%dx%d test=%dx%dx%d\n",
+                     reference.width, reference.height, reference.channels,
+                     test.width, test.height, test.channels);
+        return 2;
+    }
+
+    const double mae = ComputeMAE(reference, test);
+    const double psnr = ComputePSNR(reference, test);
+    std::printf("[metrics] %s  %dx%dx%d  MAE=%.6f  PSNR=%.6f dB\n",
+                label.c_str(), reference.width, reference.height, reference.channels,
+                mae, psnr);
+
+    if (!csv_path.empty()) {
+        const bool has_content = FileHasContent(csv_path);
+        EnsureParentDir(csv_path.c_str());
+        std::ofstream csv(csv_path.c_str(), std::ios::app);
+        if (!csv) {
+            std::fprintf(stderr, "[metrics] 无法写入 CSV: %s\n", csv_path.c_str());
+            return 2;
+        }
+        if (!has_content) {
+            csv << "label,reference,test,width,height,channels,mae,psnr_db\n";
+        }
+        csv << label << ',' << reference_path << ',' << test_path << ','
+            << reference.width << ',' << reference.height << ',' << reference.channels << ','
+            << std::fixed << std::setprecision(6) << mae << ',' << psnr << '\n';
+    }
+    return 0;
+}
diff --git a/07_nlm_denoise/blackbook537/src/nlm_cpu_ref.cpp b/07_nlm_denoise/blackbook537/src/nlm_cpu_ref.cpp
new file mode 100644
index 00000000..337f8e1b
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/nlm_cpu_ref.cpp
@@ -0,0 +1,83 @@
+#include "nlm_cpu_ref.h"
+
+#include <cmath>
+#include <vector>
+
+namespace {
+
+inline int ClampI(int v, int lo, int hi) {
+    return v < lo ? lo : (v > hi ? hi : v);
+}
+
+} // namespace
+
+void NlmDenoiseCpuRef(const ImageU8& src, ImageU8* dst, const NlmParams& params) {
+    const int w = src.width, h = src.height, c = src.channels;
+    const size_t plane = (size_t)w * h;
+    const int pr = params.patch_radius, sr = params.search_radius;
+
+    // interleaved u8 -> planar f32
+    std::vector<float> in(plane * c);
+    for (size_t i = 0; i < plane; ++i)
+        for (int ch = 0; ch < c; ++ch)
+            in[(size_t)ch * plane + i] = (float)src.data[i * c + ch];
+
+    std::vector<float> out(plane * c);
+
+    const float inv_h2 = 1.0f / (params.h * params.h);
+    const float n_patch = (float)((2 * pr + 1) * (2 * pr + 1) * c);
+    const float two_sigma2_n = 2.0f * params.sigma * params.sigma * n_patch;
+
+    for (int y = 0; y < h; ++y) {
+        for (int x = 0; x < w; ++x) {
+            float sum_w = 0.0f;
+            float sum[3] = {0.0f, 0.0f, 0.0f};
+
+            for (int dy = -sr; dy <= sr; ++dy) {
+                for (int dx = -sr; dx <= sr; ++dx) {
+                    const int qx = x + dx, qy = y + dy;
+                    // patch L2 距离（跨通道求和）
+                    float dist = 0.0f;
+                    for (int py = -pr; py <= pr; ++py) {
+                        const int ay = ClampI(y + py, 0, h - 1);
+                        const int by = ClampI(qy + py, 0, h - 1);
+                        for (int px = -pr; px <= pr; ++px) {
+                            const int ax = ClampI(x + px, 0, w - 1);
+                            const int bx = ClampI(qx + px, 0, w - 1);
+                            for (int ch = 0; ch < c; ++ch) {
+                                const float d = in[(size_t)ch * plane + (size_t)ay * w + ax] -
+                                                in[(size_t)ch * plane + (size_t)by * w + bx];
+                                dist += d * d;
+                            }
+                        }
+                    }
+                    const float dsub = dist - two_sigma2_n;
+                    const float wgt = std::exp(-(dsub > 0.0f ? dsub : 0.0f) * inv_h2);
+                    sum_w += wgt;
+                    const int cqx = ClampI(qx, 0, w - 1);
+                    const int cqy = ClampI(qy, 0, h - 1);
+                    for (int ch = 0; ch < c; ++ch)
+                        sum[ch] += wgt * in[(size_t)ch * plane + (size_t)cqy * w + cqx];
+                }
+            }
+
+            const float inv = 1.0f / sum_w;  // sum_w >= 1，无除零
+            for (int ch = 0; ch < c; ++ch)
+                out[(size_t)ch * plane + (size_t)y * w + x] = sum[ch] * inv;
+        }
+    }
+
+    // planar f32 -> interleaved u8（round + clamp）
+    dst->width = w;
+    dst->height = h;
+    dst->channels = c;
+    dst->data.resize(plane * c);
+    for (size_t i = 0; i < plane; ++i) {
+        for (int ch = 0; ch < c; ++ch) {
+            const float v = out[(size_t)ch * plane + i];
+            int q = (int)(v + 0.5f);
+            q = q < 0 ? 0 : (q > 255 ? 255 : q);
+            dst->data[i * c + ch] = (uint8_t)q;
+        }
+    }
+}
diff --git a/07_nlm_denoise/blackbook537/src/params.cpp b/07_nlm_denoise/blackbook537/src/params.cpp
new file mode 100644
index 00000000..b2ac5ca6
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/params.cpp
@@ -0,0 +1,119 @@
+#include "params.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+namespace {
+
+// 去掉首尾空白
+std::string Trim(const std::string& s) {
+    size_t b = s.find_first_not_of(" \t\r\n");
+    if (b == std::string::npos) return "";
+    size_t e = s.find_last_not_of(" \t\r\n");
+    return s.substr(b, e - b + 1);
+}
+
+bool ParseInt(const std::string& s, int* out) {
+    char* end = NULL;
+    long v = std::strtol(s.c_str(), &end, 10);
+    if (end == s.c_str() || *end != '\0') return false;
+    *out = (int)v;
+    return true;
+}
+
+bool ParseFloat(const std::string& s, float* out) {
+    char* end = NULL;
+    float v = std::strtof(s.c_str(), &end);
+    if (end == s.c_str() || *end != '\0') return false;
+    *out = v;
+    return true;
+}
+
+} // namespace
+
+bool ParseParams(const char* path, NlmParams* out, std::string* err) {
+    if (!path || !out) {
+        if (err) *err = "ParseParams: 空指针参数";
+        return false;
+    }
+    FILE* fp = std::fopen(path, "r");
+    if (!fp) {
+        if (err) *err = std::string("无法打开参数文件: ") + path;
+        return false;
+    }
+
+    NlmParams p;  // 默认值与任务文档示例一致
+    char line[512];
+    int lineno = 0;
+    while (std::fgets(line, sizeof(line), fp)) {
+        ++lineno;
+        std::string s(line);
+        // 去掉注释
+        size_t sharp = s.find('#');
+        if (sharp != std::string::npos) s = s.substr(0, sharp);
+        s = Trim(s);
+        if (s.empty()) continue;
+
+        size_t eq = s.find('=');
+        if (eq == std::string::npos) {
+            std::fclose(fp);
+            if (err) *err = "参数文件第 " + std::to_string(lineno) + " 行缺少 '='";
+            return false;
+        }
+        std::string key = Trim(s.substr(0, eq));
+        std::string val = Trim(s.substr(eq + 1));
+
+        if (key == "patch_radius") {
+            if (!ParseInt(val, &p.patch_radius)) {
+                std::fclose(fp);
+                if (err) *err = "patch_radius 非法取值: " + val;
+                return false;
+            }
+        } else if (key == "search_radius") {
+            if (!ParseInt(val, &p.search_radius)) {
+                std::fclose(fp);
+                if (err) *err = "search_radius 非法取值: " + val;
+                return false;
+            }
+        } else if (key == "h") {
+            if (!ParseFloat(val, &p.h)) {
+                std::fclose(fp);
+                if (err) *err = "h 非法取值: " + val;
+                return false;
+            }
+        } else if (key == "sigma") {
+            if (!ParseFloat(val, &p.sigma)) {
+                std::fclose(fp);
+                if (err) *err = "sigma 非法取值: " + val;
+                return false;
+            }
+        } else {
+            std::fclose(fp);
+            if (err) *err = "未知参数 key: " + key;
+            return false;
+        }
+    }
+    std::fclose(fp);
+
+    // 取值范围校验（覆盖任务基线 pr=3 / sr=10，上限为防御性约束）
+    if (p.patch_radius < 1 || p.patch_radius > 8) {
+        if (err) *err = "patch_radius 超出允许范围 [1,8]";
+        return false;
+    }
+    if (p.search_radius < 1 || p.search_radius > 32) {
+        if (err) *err = "search_radius 超出允许范围 [1,32]";
+        return false;
+    }
+    if (!(p.h > 0.0f) || p.h > 1000.0f) {
+        if (err) *err = "h 超出允许范围 (0,1000]";
+        return false;
+    }
+    if (p.sigma < 0.0f || p.sigma > 1000.0f) {
+        if (err) *err = "sigma 超出允许范围 [0,1000]";
+        return false;
+    }
+
+    *out = p;
+    return true;
+}
diff --git a/07_nlm_denoise/blackbook537/src/pipeline.cpp b/07_nlm_denoise/blackbook537/src/pipeline.cpp
new file mode 100644
index 00000000..2bf4ff1e
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/pipeline.cpp
@@ -0,0 +1,82 @@
+#include "pipeline.h"
+
+#include "kernels.h"
+
+bool NlmDenoiseGpu(const ImageU8& src, ImageU8* dst,
+                   const NlmParams& params, int kernel_version,
+                   NlmPerfReport* perf, std::string* err) {
+    // ---- 输入校验（系统边界防御）----
+    if (!dst || !err) return false;
+    if (src.width <= 0 || src.height <= 0 ||
+        (src.channels != 1 && src.channels != 3) ||
+        src.data.size() != (size_t)src.width * src.height * src.channels) {
+        *err = "NlmDenoiseGpu: 非法输入图像";
+        return false;
+    }
+
+    const int w = src.width, h = src.height, c = src.channels;
+    const size_t n_px = (size_t)w * h;
+    const size_t u8_bytes = n_px * c * sizeof(uint8_t);
+    const size_t f32_bytes = n_px * c * sizeof(float);
+
+    HostTimer t_total, t_stage;
+    t_total.Start();
+
+    // ---- 设备资源申请 ----
+    uint8_t* d_src_u8 = NULL;
+    uint8_t* d_dst_u8 = NULL;
+    float* d_src_f32 = NULL;
+    float* d_dst_f32 = NULL;
+    bool ok = false;
+
+    // 统一资源释放（失败回滚 + 正常退出共用）
+    struct Cleanup {
+        uint8_t** a; uint8_t** b; float** x; float** y;
+        ~Cleanup() {
+            std::string ignore;
+            if (a && *a) GpuFree(*a, &ignore);
+            if (b && *b) GpuFree(*b, &ignore);
+            if (x && *x) GpuFree(*x, &ignore);
+            if (y && *y) GpuFree(*y, &ignore);
+        }
+    } cleanup{&d_src_u8, &d_dst_u8, &d_src_f32, &d_dst_f32};
+
+    if (!GpuAlloc((void**)&d_src_u8, u8_bytes, err)) return false;
+    if (!GpuAlloc((void**)&d_dst_u8, u8_bytes, err)) return false;
+    if (!GpuAlloc((void**)&d_src_f32, f32_bytes, err)) return false;
+    if (!GpuAlloc((void**)&d_dst_f32, f32_bytes, err)) return false;
+
+    // ---- H2D + 转换为 planar f32 ----
+    t_stage.Start();
+    if (!GpuCopyH2D(d_src_u8, src.data.data(), u8_bytes, err)) return false;
+    if (perf) perf->h2d_ms = (float)t_stage.Ms();
+
+    if (!GpuCvtU8ToF32Planar(d_src_u8, d_src_f32, w, h, c, err)) return false;
+
+    // ---- NLM 主计算（kernel 纯计算时间由 GPU event 测量）----
+    const NlmParamsDev pdev = MakeDevParams(params, c);
+    float kernel_ms = 0.0f;
+    if (!GpuNlm(d_src_f32, d_dst_f32, w, h, c, pdev, kernel_version,
+                &kernel_ms, err)) return false;
+
+    // ---- 量化回 u8 + D2H ----
+    if (!GpuCvtF32ToU8Interleaved(d_dst_f32, d_dst_u8, w, h, c, err)) return false;
+
+    t_stage.Start();
+    dst->width = w;
+    dst->height = h;
+    dst->channels = c;
+    dst->data.resize(n_px * c);
+    if (!GpuCopyD2H(dst->data.data(), d_dst_u8, u8_bytes, err)) return false;
+    if (perf) perf->d2h_ms = (float)t_stage.Ms();
+
+    ok = true;
+    if (perf) {
+        perf->kernel_ms = kernel_ms;
+        perf->e2e_ms = (float)t_total.Ms();
+        perf->throughput_mpx_s =
+            kernel_ms > 0.0f ? (float)((double)n_px / ((double)kernel_ms * 1e3)) : 0.0f;
+    }
+    (void)ok;
+    return true;
+}
diff --git a/07_nlm_denoise/blackbook537/src/self_check.cpp b/07_nlm_denoise/blackbook537/src/self_check.cpp
new file mode 100644
index 00000000..011aad7e
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/self_check.cpp
@@ -0,0 +1,142 @@
+// ============================================================================
+// units 子命令：Host 侧单元测试（不依赖 GPU，可在任意构建环境运行）
+// 覆盖：参数解析、质量指标、CPU 参考实现的关键不变式
+// ============================================================================
+#include <cmath>
+#include <cstdio>
+#include <string>
+
+#include "nlm_cpu_ref.h"
+#include "params.h"
+#include "benchmark.h"
+
+namespace {
+
+int g_failures = 0;
+
+void Check(bool cond, const char* name) {
+    std::printf("  [%s] %s\n", cond ? "PASS" : "FAIL", name);
+    if (!cond) ++g_failures;
+}
+
+// ---- 参数解析 ----
+void TestParams() {
+    std::printf("[units] 参数解析\n");
+    const char* path = "output/work/units_params_tmp.txt";
+    EnsureParentDir(path);
+    {
+        FILE* fp = std::fopen(path, "w");
+        std::fprintf(fp,
+                     "# 注释行\n"
+                     "patch_radius = 3\n"
+                     "search_radius = 10   # 行内注释\n"
+                     "h = 12.5\n"
+                     "sigma = 30.0\n");
+        std::fclose(fp);
+    }
+    NlmParams p;
+    std::string err;
+    Check(ParseParams(path, &p, &err) &&
+          p.patch_radius == 3 && p.search_radius == 10 &&
+          p.h == 12.5f && p.sigma == 30.0f,
+          "正常解析（含注释/缺省）");
+
+    {  // 未知 key 必须报错
+        FILE* fp = std::fopen(path, "w");
+        std::fprintf(fp, "unknown_key = 1\n");
+        std::fclose(fp);
+        Check(!ParseParams(path, &p, &err), "未知 key 拒绝");
+    }
+    {  // 越界取值必须报错
+        FILE* fp = std::fopen(path, "w");
+        std::fprintf(fp, "patch_radius = 0\n");
+        std::fclose(fp);
+        Check(!ParseParams(path, &p, &err), "越界取值拒绝");
+    }
+    std::remove(path);
+}
+
+// ---- 质量指标 ----
+void TestMetrics() {
+    std::printf("[units] 质量指标\n");
+    ImageU8 a = MakeSyntheticImage(16, 16, 1, 0.0f, 42u);
+    ImageU8 b = a;
+    Check(ComputeMAE(a, b) == 0.0, "MAE(相同图像)=0");
+    Check(std::isinf(ComputePSNR(a, b)), "PSNR(相同图像)=inf");
+    for (size_t i = 0; i < b.data.size(); ++i) b.data[i] = (uint8_t)(a.data[i] + 10);
+    Check(std::fabs(ComputeMAE(a, b) - 10.0) < 1e-9, "MAE(恒差10)=10");
+    // PSNR = 20*log10(255/10) ≈ 28.13 dB
+    Check(std::fabs(ComputePSNR(a, b) - 28.1308) < 0.01, "PSNR(恒差10)≈28.13dB");
+}
+
+// ---- CPU 参考实现不变式 ----
+void TestCpuRef() {
+    std::printf("[units] CPU 参考实现\n");
+    NlmParams p;
+    p.patch_radius = 2;
+    p.search_radius = 3;
+
+    {  // 常量图：所有 patch 距离为 0，权重全 1，输出必须逐位等于输入
+        ImageU8 img;
+        img.width = 48; img.height = 48; img.channels = 1;
+        img.data.assign(48 * 48, 128);
+        ImageU8 out;
+        NlmDenoiseCpuRef(img, &out, p);
+        Check(out.data == img.data, "常量灰度图恒等");
+    }
+    {  // RGB 常量图
+        ImageU8 img;
+        img.width = 32; img.height = 32; img.channels = 3;
+        img.data.resize(32 * 32 * 3);
+        for (size_t i = 0; i < img.data.size(); i += 3) {
+            img.data[i] = 100; img.data[i + 1] = 150; img.data[i + 2] = 200;
+        }
+        ImageU8 out;
+        NlmDenoiseCpuRef(img, &out, p);
+        Check(out.data == img.data, "常量 RGB 图恒等");
+    }
+    {  // 确定性：同一输入两次运行结果 bit 级一致
+        ImageU8 img = MakeSyntheticImage(64, 64, 1, 25.0f, 7u);
+        ImageU8 o1, o2;
+        NlmDenoiseCpuRef(img, &o1, p);
+        NlmDenoiseCpuRef(img, &o2, p);
+        Check(o1.data == o2.data, "确定性（两次运行一致）");
+    }
+    {  // 极小图（图像小于搜索窗）：clamp 边界不越界、输出范围合法
+        ImageU8 img = MakeSyntheticImage(5, 7, 1, 25.0f, 9u);
+        NlmParams big;
+        big.patch_radius = 3;
+        big.search_radius = 10;
+        ImageU8 out;
+        NlmDenoiseCpuRef(img, &out, big);
+        bool range_ok = (out.data.size() == img.data.size());
+        Check(range_ok, "极小图（5x7, sr=10）不崩溃且尺寸正确");
+    }
+    {  // 降噪有效性：含噪渐变图降噪后 MAE(输出, 干净图) 应小于 MAE(含噪图, 干净图)
+        ImageU8 clean = MakeSyntheticImage(64, 64, 1, 0.0f, 5u);
+        ImageU8 noisy = MakeSyntheticImage(64, 64, 1, 25.0f, 5u);
+        NlmParams q;  // 基线参数
+        ImageU8 out;
+        NlmDenoiseCpuRef(noisy, &out, q);
+        const double mae_before = ComputeMAE(noisy, clean);
+        const double mae_after = ComputeMAE(out, clean);
+        std::printf("    降噪前后对干净图 MAE: %.2f -> %.2f\n", mae_before, mae_after);
+        Check(mae_after < mae_before, "降噪有效性（MAE 下降）");
+    }
+}
+
+} // namespace
+
+int RunUnitTests() {
+    std::printf("========== NLM 单元测试 ==========" "\n");
+    TestParams();
+    TestMetrics();
+    TestCpuRef();
+    std::printf("========== %s（失败 %d 项）==========\n",
+                g_failures == 0 ? "全部通过" : "存在失败", g_failures);
+    return g_failures;
+}
+
+#ifdef NLM_UNITS_STANDALONE
+int main() { return RunUnitTests(); }
+#endif
diff --git a/07_nlm_denoise/blackbook537/src/validate.cpp b/07_nlm_denoise/blackbook537/src/validate.cpp
new file mode 100644
index 00000000..275d2caf
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/src/validate.cpp
@@ -0,0 +1,104 @@
+// ============================================================================
+// validate 子命令：正确性验证
+//   主基线：自研 CPU 参考实现（语义与 GPU kernel 严格一致），逐版本对比 MAE/PSNR
+//   辅基线（可选，-DWITH_OPENCV）：OpenCV fastNlMeansDenoising(Colored) 交叉验证
+// 通过阈值（架构设计 §10.1）：vs CPU 参考 V0/V1 MAE<=0.5，V2 MAE<=1.0，PSNR>=30dB
+// ============================================================================
+#include <cstdio>
+#include <string>
+
+#include "nlm_cpu_ref.h"
+#include "pipeline.h"
+#include "benchmark.h"
+
+#ifdef WITH_OPENCV
+#include <opencv2/opencv.hpp>
+#include <opencv2/photo.hpp>
+#endif
+
+int RunValidate(const std::string& input, const std::string& output,
+                const std::string& params_path) {
+    std::string err;
+    ImageU8 src;
+    if (!LoadImage(input.c_str(), &src, &err)) {
+        std::fprintf(stderr, "[validate] %s\n", err.c_str());
+        return 2;
+    }
+    NlmParams params;
+    if (!ParseParams(params_path.c_str(), &params, &err)) {
+        std::fprintf(stderr, "[validate] %s\n", err.c_str());
+        return 2;
+    }
+
+    std::printf("[validate] 图像 %dx%dx%d，参数 pr=%d sr=%d h=%.1f sigma=%.1f\n",
+                src.width, src.height, src.channels,
+                params.patch_radius, params.search_radius, params.h, params.sigma);
+
+    // ---- CPU 参考（主基线）----
+    ImageU8 cpu;
+    HostTimer t;
+    t.Start();
+    NlmDenoiseCpuRef(src, &cpu, params);
+    const double cpu_ms = t.Ms();
+    std::printf("[validate] CPU 参考耗时: %.1f ms\n", cpu_ms);
+
+    // ---- 各 GPU kernel 版本对比 ----
+    std::printf("%-6s %-12s %-12s %-12s %-10s %s\n",
+                "ver", "kernel(ms)", "e2e(ms)", "MAE(vsCPU)", "PSNR(dB)", "判定");
+    bool pass = true;
+    ImageU8 last_gpu;
+    for (int v = 0; v <= 2; ++v) {
+        ImageU8 gpu;
+        NlmPerfReport perf;
+        if (!NlmDenoiseGpu(src, &gpu, params, v, &perf, &err)) {
+            std::fprintf(stderr, "[validate] kernel v%d 运行失败: %s\n", v, err.c_str());
+            return 2;
+        }
+        const double mae = ComputeMAE(gpu, cpu);
+        const double psnr = ComputePSNR(gpu, cpu);
+        const double mae_thr = (v == 2) ? 1.0 : 0.5;
+        const bool ok = (mae <= mae_thr) && (psnr >= 30.0);
+        std::printf("%-6d %-12.3f %-12.3f %-12.4f %-10.2f %s\n",
+                    v, perf.kernel_ms, perf.e2e_ms, mae, psnr, ok ? "PASS" : "FAIL");
+        pass = pass && ok;
+        if (v == 2) last_gpu = gpu;
+    }
+
+#ifdef WITH_OPENCV
+    // ---- OpenCV 参考（辅基线，信息性输出，不作为 PASS/FAIL 门槛）----
+    {
+        cv::Mat cv_src(src.height, src.width,
+                       src.channels == 1 ? CV_8UC1 : CV_8UC3, src.data.data());
+        cv::Mat cv_dst;
+        if (src.channels == 1) {
+            cv::fastNlMeansDenoising(cv_src, cv_dst, params.h,
+                                     2 * params.patch_radius + 1,
+                                     2 * params.search_radius + 1);
+        } else {
+            cv::fastNlMeansDenoisingColored(cv_src, cv_dst, params.h, params.h,
+                                            2 * params.patch_radius + 1,
+                                            2 * params.search_radius + 1);
+        }
+        ImageU8 ocv;
+        ocv.width = src.width;
+        ocv.height = src.height;
+        ocv.channels = src.channels;
+        ocv.data.assign(cv_dst.data, cv_dst.data + src.data.size());
+        std::printf("[validate] OpenCV 参考对比（信息性）: GPU(v2) vs OpenCV  MAE=%.4f  PSNR=%.2f dB\n",
+                    ComputeMAE(last_gpu, ocv), ComputePSNR(last_gpu, ocv));
+        std::printf("[validate] 注：OpenCV 的 h 标定与任务公式存在实现差异，"
+                    "仅作交叉验证，验收以 CPU 参考为准。\n");
+    }
+#endif
+
+    if (!output.empty()) {
+        if (!SaveImage(output.c_str(), last_gpu, &err)) {
+            std::fprintf(stderr, "[validate] %s\n", err.c_str());
+            return 2;
+        }
+        std::printf("[validate] 已保存 GPU(v2) 输出: %s\n", output.c_str());
+    }
+
+    std::printf("[validate] 总体判定: %s\n", pass ? "PASS" : "FAIL");
+    return pass ? 0 : 1;
+}
diff --git a/07_nlm_denoise/blackbook537/test_images/README.md b/07_nlm_denoise/blackbook537/test_images/README.md
new file mode 100644
index 00000000..0fa58dec
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/test_images/README.md
@@ -0,0 +1,22 @@
+# 示例图像来源
+
+图片由本项目生成器产生，没有采用参考作业的图片或输出。所有文件为 8 位 RGB PNG。
+
+| 文件 | 用途/来源 |
+|---|---|
+| `clean_1920x1080_3ch.png` | 平滑渐变参考图，生成器噪声幅度 0 |
+| `noisy_1920x1080_3ch_sigma25.png` | 对应 1080p 含噪图，seed=7、均匀噪声幅度 25 |
+| `noisy_256x256_3ch_sigma25.png` | 默认快速运行输入；历史文件，不用于正式质量表 |
+| `denoised_1920x1080_base_v2.png` | 从历史 MetaX 证据中保留的 V2/base 输出，仅作演示 |
+
+生成器使用 LCG 确定性序列，将均匀噪声加到渐变图后量化、裁剪到 0–255。文件名中的 `sigma25` 是旧命名，含义为均匀噪声幅度，并非高斯标准差 25。灰度、复杂纹理、自然图像与标准差受控噪声的补充建议见 [SUBMISSION.md](../SUBMISSION.md)。
+
+复现 1080p 配对图（输出到被忽略的运行目录）：
+
+```bash
+./build/nlm_denoise gen --size 1920x1080 --channels 3 --sigma 0 --seed 7 -o output/work/clean.png
+./build/nlm_denoise gen --size 1920x1080 --channels 3 --sigma 25 --seed 7 -o output/work/noisy.png
+./build/nlm_denoise run -i output/work/noisy.png -o output/images/denoised.png -p params.txt --kernel 2
+```
+
+`SHA256SUMS.txt` 记录当前四张图的哈希。可在本目录执行 `sha256sum -c SHA256SUMS.txt`。
diff --git a/07_nlm_denoise/blackbook537/test_images/SHA256SUMS.txt b/07_nlm_denoise/blackbook537/test_images/SHA256SUMS.txt
new file mode 100644
index 00000000..8661abfd
--- /dev/null
+++ b/07_nlm_denoise/blackbook537/test_images/SHA256SUMS.txt
@@ -0,0 +1,4 @@
+061b764c95d5788d850c8121802290343e739107104061bbac1a83dbf463f844  clean_1920x1080_3ch.png
+2f8e77c56606ef55111bf1ab2720ab6e3405ea1ff598868d8b4e2c6f737d5758  denoised_1920x1080_base_v2.png
+ea2fb82017ac764aa8b649ae3d77242cdd9f6520dee7bb956955a81baa1ba859  noisy_1920x1080_3ch_sigma25.png
+d857a577bba448c8ab828a14a51ba89c14c29aa2bf1dc1dc51af654a2d744a8d  noisy_256x256_3ch_sigma25.png
diff --git a/07_nlm_denoise/blackbook537/test_images/clean_1920x1080_3ch.png b/07_nlm_denoise/blackbook537/test_images/clean_1920x1080_3ch.png
new file mode 100644
index 00000000..3cc1c4bd
Binary files /dev/null and b/07_nlm_denoise/blackbook537/test_images/clean_1920x1080_3ch.png differ
diff --git a/07_nlm_denoise/blackbook537/test_images/denoised_1920x1080_base_v2.png b/07_nlm_denoise/blackbook537/test_images/denoised_1920x1080_base_v2.png
new file mode 100644
index 00000000..9f4514d1
Binary files /dev/null and b/07_nlm_denoise/blackbook537/test_images/denoised_1920x1080_base_v2.png differ
diff --git a/07_nlm_denoise/blackbook537/test_images/noisy_1920x1080_3ch_sigma25.png b/07_nlm_denoise/blackbook537/test_images/noisy_1920x1080_3ch_sigma25.png
new file mode 100644
index 00000000..690c4e33
Binary files /dev/null and b/07_nlm_denoise/blackbook537/test_images/noisy_1920x1080_3ch_sigma25.png differ
diff --git a/07_nlm_denoise/blackbook537/test_images/noisy_256x256_3ch_sigma25.png b/07_nlm_denoise/blackbook537/test_images/noisy_256x256_3ch_sigma25.png
new file mode 100644
index 00000000..d5d7ecb0
Binary files /dev/null and b/07_nlm_denoise/blackbook537/test_images/noisy_256x256_3ch_sigma25.png differ