Upload folder using huggingface_hub

stable-diffusion.cpp/.dockerignore

@@ -0,0 +1,6 @@
+build*/
+test/
+
+.cache/
+*.swp
+models/

stable-diffusion.cpp/.github/workflows/build.yml

@@ -0,0 +1,201 @@
+name: CI
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+      - ci
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+
+jobs:
+  ubuntu-latest-cmake:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake ..
+          cmake --build . --config Release
+
+      #- name: Test
+        #id: cmake_test
+        #run: |
+          #cd build
+          #ctest --verbose --timeout 900
+
+  macOS-latest-cmake:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake ..
+          cmake --build . --config Release
+
+      #- name: Test
+        #id: cmake_test
+        #run: |
+          #cd build
+          #ctest --verbose --timeout 900
+
+  windows-latest-cmake:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        include:
+          - build: 'noavx'
+            defines: '-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
+          - build: 'avx2'
+            defines: '-DGGML_AVX2=ON'
+          - build: 'avx'
+            defines: '-DGGML_AVX2=OFF'
+          - build: 'avx512'
+            defines: '-DGGML_AVX512=ON'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. ${{ matrix.defines }}
+          cmake --build . --config Release
+
+      - name: Check AVX512F support
+        id: check_avx512f
+        if: ${{ matrix.build == 'avx512' }}
+        continue-on-error: true
+        run: |
+          cd build
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
+          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
+          & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
+          .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
+
+      #- name: Test
+        #id: cmake_test
+        #run: |
+          #cd build
+          #ctest -C Release --verbose --timeout 900
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          Copy-Item ggml/LICENSE .\build\bin\Release\ggml.txt
+          Copy-Item LICENSE .\build\bin\Release\stable-diffusion.cpp.txt
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
+
+  release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+
+    runs-on: ubuntu-latest
+
+    needs:
+      - ubuntu-latest-cmake
+      - macOS-latest-cmake
+      - windows-latest-cmake
+
+    steps:
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v3
+
+      - name: Get commit hash
+        id: commit
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Create release
+        id: create_release
+        uses: anzz1/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+
+      - name: Upload release
+        id: upload_release
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/${file}`)
+                });
+              }
+            }

stable-diffusion.cpp/.gitignore

@@ -0,0 +1,5 @@
+build*/
+test/
+
+.cache/
+*.swp

stable-diffusion.cpp/.gitmodules

@@ -0,0 +1,3 @@
+[submodule "ggml"]
+	path = ggml
+	url = https://github.com/leejet/ggml.git

stable-diffusion.cpp/CMakeLists.txt

@@ -0,0 +1,45 @@
+cmake_minimum_required(VERSION 3.12)
+project("stable-diffusion")
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(SD_STANDALONE ON)
+else()
+    set(SD_STANDALONE OFF)
+endif()
+
+#
+# Option list
+#
+
+# general
+#option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
+option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
+option(BUILD_SHARED_LIBS             "sd: build shared libs" OFF)
+#option(SD_BUILD_SERVER               "sd: build server example"                           ON)
+
+
+# deps
+add_subdirectory(ggml)
+
+set(SD_LIB stable-diffusion)
+
+add_library(${SD_LIB} stable-diffusion.h stable-diffusion.cpp)
+target_link_libraries(${SD_LIB} PUBLIC ggml)
+target_include_directories(${SD_LIB} PUBLIC .)
+target_compile_features(${SD_LIB} PUBLIC cxx_std_11)
+
+
+if (SD_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif()
+

stable-diffusion.cpp/Dockerfile

@@ -0,0 +1,17 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && apt-get install -y build-essential git cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

stable-diffusion.cpp/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 leejet
+
+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.

stable-diffusion.cpp/README.md

@@ -0,0 +1,198 @@
+<p align="center">
+  <img src="./assets/a%20lovely%20cat.png" width="256x">
+</p>
+
+# stable-diffusion.cpp
+
+Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in pure C/C++
+
+## Features
+
+- Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
+- 16-bit, 32-bit float support
+- 4-bit, 5-bit and 8-bit integer quantization support
+- Accelerated memory-efficient CPU inference
+    - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image
+- AVX, AVX2 and AVX512 support for x86 architectures
+- SD1.x and SD2.x support
+- Original `txt2img` and `img2img` mode
+- Negative prompt
+- [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
+- Sampling method
+    - `Euler A`
+    - `Euler`
+    - `Heun`
+    - `DPM2`
+    - `DPM++ 2M`
+    - [`DPM++ 2M v2`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457)
+    - `DPM++ 2S a`
+- Cross-platform reproducibility (`--rng cuda`, consistent with the `stable-diffusion-webui GPU RNG`)
+- Embedds generation parameters into png output as webui-compatible text string
+- Supported platforms
+    - Linux
+    - Mac OS
+    - Windows
+    - Android (via Termux)
+
+### TODO
+
+- [ ] More sampling methods
+- [ ] GPU support
+- [ ] Make inference faster
+    - The current implementation of ggml_conv_2d is slow and has high memory usage
+- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
+- [ ] LoRA support
+- [ ] k-quants support
+
+## Usage
+
+### Get the Code
+
+```
+git clone --recursive https://github.com/leejet/stable-diffusion.cpp
+cd stable-diffusion.cpp
+```
+
+- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
+
+```
+cd stable-diffusion.cpp
+git pull origin master
+git submodule init
+git submodule update
+```
+
+### Convert weights
+
+- download original weights(.ckpt or .safetensors). For example
+    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
+    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
+    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
+
+    ```shell
+    curl -L -O https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
+    # curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
+    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-2-1/blob/main/v2-1_768-nonema-pruned.safetensors
+    ```
+
+- convert weights to ggml model format
+
+    ```shell
+    cd models
+    pip install -r requirements.txt
+    python convert.py [path to weights] --out_type [output precision]
+    # For example, python convert.py sd-v1-4.ckpt --out_type f16
+    ```
+
+### Quantization
+
+You can specify the output model format using the --out_type parameter
+
+- `f16` for 16-bit floating-point
+- `f32` for 32-bit floating-point
+- `q8_0` for 8-bit integer quantization 
+- `q5_0` or `q5_1` for 5-bit integer quantization 
+- `q4_0` or `q4_1` for 4-bit integer quantization
+
+### Build
+
+#### Build from scratch
+
+```shell
+mkdir build
+cd build
+cmake ..
+cmake --build . --config Release
+```
+
+##### Using OpenBLAS
+
+```
+cmake .. -DGGML_OPENBLAS=ON
+cmake --build . --config Release
+```
+
+### Run
+
+```
+usage: ./bin/sd [arguments]
+
+arguments:
+  -h, --help                         show this help message and exit
+  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)
+  -t, --threads N                    number of threads to use during computation (default: -1).
+                                     If threads <= 0, then threads will be set to the number of CPU physical cores
+  -m, --model [MODEL]                path to model
+  -i, --init-img [IMAGE]             path to the input image, required by img2img
+  -o, --output OUTPUT                path to write result image to (default: .\output.png)
+  -p, --prompt [PROMPT]              the prompt to render
+  -n, --negative-prompt PROMPT       the negative prompt (default: "")
+  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
+  --strength STRENGTH                strength for noising/unnoising (default: 0.75)
+                                     1.0 corresponds to full destruction of information in init image
+  -H, --height H                     image height, in pixel space (default: 512)
+  -W, --width W                      image width, in pixel space (default: 512)
+  --sampling-method {euler, euler_a, heun, dpm++2m, dpm++2mv2}
+                                     sampling method (default: "euler_a")
+  --steps  STEPS                     number of sample steps (default: 20)
+  --rng {std_default, cuda}          RNG (default: cuda)
+  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
+  -v, --verbose                      print extra info
+```
+
+#### txt2img example
+
+```
+./bin/sd -m ../models/sd-v1-4-ggml-model-f16.bin -p "a lovely cat"
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----  |----  |----  |----  |----  |----  |----  |
+| ![](./assets/f32.png) |![](./assets/f16.png) |![](./assets/q8_0.png) |![](./assets/q5_0.png) |![](./assets/q5_1.png) |![](./assets/q4_0.png) |![](./assets/q4_1.png) |
+
+#### img2img example
+
+- `./output.png` is the image generated from the above txt2img pipeline
+
+
+```
+./bin/sd --mode img2img -m ../models/sd-v1-4-ggml-model-f16.bin -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+```
+
+<p align="center">
+  <img src="./assets/img2img_output.png" width="256x">
+</p>
+
+### Docker
+
+#### Building using Docker
+
+```shell
+docker build -t sd .
+```
+
+#### Run
+
+```shell
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
+# For example
+# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4-ggml-model-f16.bin -p "a lovely cat" -v -o /output/output.png
+```
+
+## Memory/Disk Requirements
+
+| precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----         | ----  |----  |----  |----  |----  |----  |----  |
+|  **Disk**        | 2.7G | 2.0G | 1.7G | 1.6G | 1.6G | 1.5G | 1.5G |
+|  **Memory**(txt2img - 512 x 512) | ~2.8G | ~2.3G | ~2.1G | ~2.0G | ~2.0G | ~2.0G | ~2.0G |
+
+
+## References
+
+- [ggml](https://github.com/ggerganov/ggml)
+- [stable-diffusion](https://github.com/CompVis/stable-diffusion)
+- [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
+- [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui)
+- [k-diffusion](https://github.com/crowsonkb/k-diffusion)

stable-diffusion.cpp/examples/CMakeLists.txt

@@ -0,0 +1,8 @@
+# TODO: move into its own subdirectoy
+# TODO: make stb libs a target (maybe common)
+set(SD_TARGET sd)
+
+add_executable(${SD_TARGET} main.cpp stb_image.h stb_image_write.h)
+install(TARGETS ${SD_TARGET} RUNTIME)
+target_link_libraries(${SD_TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${SD_TARGET} PUBLIC cxx_std_11)

stable-diffusion.cpp/examples/main.cpp

@@ -0,0 +1,473 @@
+#include <stdio.h>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <random>
+#include <string>
+#include <thread>
+#include <unordered_set>
+
+#include "stable-diffusion.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb_image.h"
+
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#define STB_IMAGE_WRITE_STATIC
+#include "stb_image_write.h"
+
+#if defined(__APPLE__) && defined(__MACH__)
+#include <sys/sysctl.h>
+#include <sys/types.h>
+#endif
+
+#if !defined(_WIN32)
+#include <sys/ioctl.h>
+#include <unistd.h>
+#endif
+
+#define TXT2IMG "txt2img"
+#define IMG2IMG "img2img"
+
+// get_num_physical_cores is copy from
+// https://github.com/ggerganov/llama.cpp/blob/master/examples/common.cpp
+// LICENSE: https://github.com/ggerganov/llama.cpp/blob/master/LICENSE
+int32_t get_num_physical_cores() {
+#ifdef __linux__
+    // enumerate the set of thread siblings, num entries is num cores
+    std::unordered_set<std::string> siblings;
+    for (uint32_t cpu = 0; cpu < UINT32_MAX; ++cpu) {
+        std::ifstream thread_siblings("/sys/devices/system/cpu" + std::to_string(cpu) + "/topology/thread_siblings");
+        if (!thread_siblings.is_open()) {
+            break;  // no more cpus
+        }
+        std::string line;
+        if (std::getline(thread_siblings, line)) {
+            siblings.insert(line);
+        }
+    }
+    if (siblings.size() > 0) {
+        return static_cast<int32_t>(siblings.size());
+    }
+#elif defined(__APPLE__) && defined(__MACH__)
+    int32_t num_physical_cores;
+    size_t len = sizeof(num_physical_cores);
+    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+#elif defined(_WIN32)
+    // TODO: Implement
+#endif
+    unsigned int n_threads = std::thread::hardware_concurrency();
+    return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
+}
+
+const char* rng_type_to_str[] = {
+    "std_default",
+    "cuda",
+};
+
+// Names of the sampler method, same order as enum SampleMethod in stable-diffusion.h
+const char* sample_method_str[] = {
+    "euler_a",
+    "euler",
+    "heun",
+    "dpm2",
+    "dpm++2s_a",
+    "dpm++2m",
+    "dpm++2mv2"};
+
+// Names of the sigma schedule overrides, same order as Schedule in stable-diffusion.h
+const char* schedule_str[] = {
+    "default",
+    "discrete",
+    "karras"};
+
+struct Option {
+    int n_threads = -1;
+    std::string mode = TXT2IMG;
+    std::string model_path;
+    std::string output_path = "output.png";
+    std::string init_img;
+    std::string prompt;
+    std::string negative_prompt;
+    float cfg_scale = 7.0f;
+    int w = 512;
+    int h = 512;
+    SampleMethod sample_method = EULER_A;
+    Schedule schedule = DEFAULT;
+    int sample_steps = 20;
+    float strength = 0.75f;
+    RNGType rng_type = CUDA_RNG;
+    int64_t seed = 42;
+    bool verbose = false;
+
+    void print() {
+        printf("Option: \n");
+        printf("    n_threads:       %d\n", n_threads);
+        printf("    mode:            %s\n", mode.c_str());
+        printf("    model_path:      %s\n", model_path.c_str());
+        printf("    output_path:     %s\n", output_path.c_str());
+        printf("    init_img:        %s\n", init_img.c_str());
+        printf("    prompt:          %s\n", prompt.c_str());
+        printf("    negative_prompt: %s\n", negative_prompt.c_str());
+        printf("    cfg_scale:       %.2f\n", cfg_scale);
+        printf("    width:           %d\n", w);
+        printf("    height:          %d\n", h);
+        printf("    sample_method:   %s\n", sample_method_str[sample_method]);
+        printf("    schedule:        %s\n", schedule_str[schedule]);
+        printf("    sample_steps:    %d\n", sample_steps);
+        printf("    strength:        %.2f\n", strength);
+        printf("    rng:             %s\n", rng_type_to_str[rng_type]);
+        printf("    seed:            %ld\n", seed);
+    }
+};
+
+void print_usage(int argc, const char* argv[]) {
+    printf("usage: %s [arguments]\n", argv[0]);
+    printf("\n");
+    printf("arguments:\n");
+    printf("  -h, --help                         show this help message and exit\n");
+    printf("  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)\n");
+    printf("  -t, --threads N                    number of threads to use during computation (default: -1).\n");
+    printf("                                     If threads <= 0, then threads will be set to the number of CPU physical cores\n");
+    printf("  -m, --model [MODEL]                path to model\n");
+    printf("  -i, --init-img [IMAGE]             path to the input image, required by img2img\n");
+    printf("  -o, --output OUTPUT                path to write result image to (default: .\\output.png)\n");
+    printf("  -p, --prompt [PROMPT]              the prompt to render\n");
+    printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
+    printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
+    printf("  --strength STRENGTH                strength for noising/unnoising (default: 0.75)\n");
+    printf("                                     1.0 corresponds to full destruction of information in init image\n");
+    printf("  -H, --height H                     image height, in pixel space (default: 512)\n");
+    printf("  -W, --width W                      image width, in pixel space (default: 512)\n");
+    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2}\n");
+    printf("                                     sampling method (default: \"euler_a\")\n");
+    printf("  --steps  STEPS                     number of sample steps (default: 20)\n");
+    printf("  --rng {std_default, cuda}          RNG (default: cuda)\n");
+    printf("  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)\n");
+    printf("  --schedule {discrete, karras}      Denoiser sigma schedule (default: discrete)\n");
+    printf("  -v, --verbose                      print extra info\n");
+}
+
+void parse_args(int argc, const char* argv[], Option* opt) {
+    bool invalid_arg = false;
+
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->n_threads = std::stoi(argv[i]);
+        } else if (arg == "-M" || arg == "--mode") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->mode = argv[i];
+
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->model_path = argv[i];
+        } else if (arg == "-i" || arg == "--init-img") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->init_img = argv[i];
+        } else if (arg == "-o" || arg == "--output") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->output_path = argv[i];
+        } else if (arg == "-p" || arg == "--prompt") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->prompt = argv[i];
+        } else if (arg == "-n" || arg == "--negative-prompt") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->negative_prompt = argv[i];
+        } else if (arg == "--cfg-scale") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->cfg_scale = std::stof(argv[i]);
+        } else if (arg == "--strength") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->strength = std::stof(argv[i]);
+        } else if (arg == "-H" || arg == "--height") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->h = std::stoi(argv[i]);
+        } else if (arg == "-W" || arg == "--width") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->w = std::stoi(argv[i]);
+        } else if (arg == "--steps") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->sample_steps = std::stoi(argv[i]);
+        } else if (arg == "--rng") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            std::string rng_type_str = argv[i];
+            if (rng_type_str == "std_default") {
+                opt->rng_type = STD_DEFAULT_RNG;
+            } else if (rng_type_str == "cuda") {
+                opt->rng_type = CUDA_RNG;
+            } else {
+                invalid_arg = true;
+                break;
+            }
+        } else if (arg == "--schedule") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            const char* schedule_selected = argv[i];
+            int schedule_found = -1;
+            for (int d = 0; d < N_SCHEDULES; d++) {
+                if (!strcmp(schedule_selected, schedule_str[d])) {
+                    schedule_found = d;
+                }
+            }
+            if (schedule_found == -1) {
+                invalid_arg = true;
+                break;
+            }
+            opt->schedule = (Schedule)schedule_found;
+        } else if (arg == "-s" || arg == "--seed") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->seed = std::stoll(argv[i]);
+        } else if (arg == "--sampling-method") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            const char* sample_method_selected = argv[i];
+            int sample_method_found = -1;
+            for (int m = 0; m < N_SAMPLE_METHODS; m++) {
+                if (!strcmp(sample_method_selected, sample_method_str[m])) {
+                    sample_method_found = m;
+                }
+            }
+            if (sample_method_found == -1) {
+                invalid_arg = true;
+                break;
+            }
+            opt->sample_method = (SampleMethod)sample_method_found;
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv);
+            exit(0);
+        } else if (arg == "-v" || arg == "--verbose") {
+            opt->verbose = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv);
+            exit(1);
+        }
+        if (invalid_arg) {
+            fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+            print_usage(argc, argv);
+            exit(1);
+        }
+    }
+
+    if (opt->n_threads <= 0) {
+        opt->n_threads = get_num_physical_cores();
+    }
+
+    if (opt->mode != TXT2IMG && opt->mode != IMG2IMG) {
+        fprintf(stderr, "error: invalid mode %s, must be one of ['%s', '%s']\n",
+                opt->mode.c_str(), TXT2IMG, IMG2IMG);
+        exit(1);
+    }
+
+    if (opt->prompt.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: prompt\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->model_path.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: model_path\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->mode == IMG2IMG && opt->init_img.length() == 0) {
+        fprintf(stderr, "error: when using the img2img mode, the following arguments are required: init-img\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->output_path.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: output_path\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->w <= 0 || opt->w % 64 != 0) {
+        fprintf(stderr, "error: the width must be a multiple of 64\n");
+        exit(1);
+    }
+
+    if (opt->h <= 0 || opt->h % 64 != 0) {
+        fprintf(stderr, "error: the height must be a multiple of 64\n");
+        exit(1);
+    }
+
+    if (opt->sample_steps <= 0) {
+        fprintf(stderr, "error: the sample_steps must be greater than 0\n");
+        exit(1);
+    }
+
+    if (opt->strength < 0.f || opt->strength > 1.f) {
+        fprintf(stderr, "error: can only work with strength in [0.0, 1.0]\n");
+        exit(1);
+    }
+
+    if (opt->seed < 0) {
+        srand((int)time(NULL));
+        opt->seed = rand();
+    }
+}
+
+std::string basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
+}
+
+int main(int argc, const char* argv[]) {
+    Option opt;
+    parse_args(argc, argv, &opt);
+
+    if (opt.verbose) {
+        opt.print();
+        printf("%s", sd_get_system_info().c_str());
+        set_sd_log_level(SDLogLevel::DEBUG);
+    }
+
+    bool vae_decode_only = true;
+    std::vector<uint8_t> init_img;
+    if (opt.mode == IMG2IMG) {
+        vae_decode_only = false;
+
+        int c = 0;
+        unsigned char* img_data = stbi_load(opt.init_img.c_str(), &opt.w, &opt.h, &c, 3);
+        if (img_data == NULL) {
+            fprintf(stderr, "load image from '%s' failed\n", opt.init_img.c_str());
+            return 1;
+        }
+        if (c != 3) {
+            fprintf(stderr, "input image must be a 3 channels RGB image, but got %d channels\n", c);
+            free(img_data);
+            return 1;
+        }
+        if (opt.w <= 0 || opt.w % 64 != 0) {
+            fprintf(stderr, "error: the width of image must be a multiple of 64\n");
+            free(img_data);
+            return 1;
+        }
+        if (opt.h <= 0 || opt.h % 64 != 0) {
+            fprintf(stderr, "error: the height of image must be a multiple of 64\n");
+            free(img_data);
+            return 1;
+        }
+        init_img.assign(img_data, img_data + (opt.w * opt.h * c));
+    }
+
+    StableDiffusion sd(opt.n_threads, vae_decode_only, true, opt.rng_type);
+    if (!sd.load_from_file(opt.model_path, opt.schedule)) {
+        return 1;
+    }
+
+    std::vector<uint8_t> img;
+    if (opt.mode == TXT2IMG) {
+        img = sd.txt2img(opt.prompt,
+                         opt.negative_prompt,
+                         opt.cfg_scale,
+                         opt.w,
+                         opt.h,
+                         opt.sample_method,
+                         opt.sample_steps,
+                         opt.seed);
+    } else {
+        img = sd.img2img(init_img,
+                         opt.prompt,
+                         opt.negative_prompt,
+                         opt.cfg_scale,
+                         opt.w,
+                         opt.h,
+                         opt.sample_method,
+                         opt.sample_steps,
+                         opt.strength,
+                         opt.seed);
+    }
+
+    if (img.size() == 0) {
+        fprintf(stderr, "generate failed\n");
+        return 1;
+    }
+
+    std::string parameter_string = opt.prompt + "\n";
+    if (opt.negative_prompt.size() != 0) {
+        parameter_string += "Negative prompt: " + opt.negative_prompt + "\n";
+    }
+    parameter_string += "Steps: " + std::to_string(opt.sample_steps) + ", ";
+    parameter_string += "CFG scale: " + std::to_string(opt.cfg_scale) + ", ";
+    parameter_string += "Seed: " + std::to_string(opt.seed) + ", ";
+    parameter_string += "Size: " + std::to_string(opt.w) + "x" + std::to_string(opt.h) + ", ";
+    parameter_string += "Model: " + basename(opt.model_path) + ", ";
+    parameter_string += "RNG: " + std::string(rng_type_to_str[opt.rng_type]) + ", ";
+    parameter_string += "Sampler: " + std::string(sample_method_str[opt.sample_method]);
+    if (opt.schedule == KARRAS) {
+        parameter_string += " karras";
+    }
+    parameter_string += ", ";
+    parameter_string += "Version: stable-diffusion.cpp";
+
+    stbi_write_png(opt.output_path.c_str(), opt.w, opt.h, 3, img.data(), 0, parameter_string.c_str());
+    printf("save result image to '%s'\n", opt.output_path.c_str());
+
+    return 0;
+}

stable-diffusion.cpp/examples/stb_image_write.h

@@ -0,0 +1,1741 @@
+/* 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, const char* parameters = NULL);
+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, const char* parameters)
+{
+   int force_filter = stbi_write_force_png_filter;
+   int param_length = 0;
+   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;
+
+   if(parameters != NULL) {
+      param_length = strlen(parameters);
+      param_length += strlen("parameters") + 1; // For the name and the null-byte
+   }
+
+   // each tag requires 12 bytes of overhead
+   out = (unsigned char *) STBIW_MALLOC(8 + 12+13 + 12+zlen + 12 + ((parameters)?(param_length+12):0));
+   if (!out) return 0;
+   *out_len = 8 + 12+13 + 12+zlen + 12 + ((parameters)?(param_length+12):0);
+
+   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);
+
+   if(parameters != NULL) {
+      stbiw__wp32(o, param_length);
+      stbiw__wptag(o, "tEXt");
+      STBIW_MEMMOVE(o, "parameters", strlen("parameters"));
+      o+=strlen("parameters");
+      *o++ = 0; // Null pyte separator
+      STBIW_MEMMOVE(o, parameters, strlen(parameters));
+      o+=strlen(parameters);
+      stbiw__wpcrc(&o, param_length);
+   }
+
+   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, const char* parameters)
+{
+   FILE *f;
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len, parameters);
+   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, NULL);
+   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.
+------------------------------------------------------------------------------
+*/

stable-diffusion.cpp/models/.gitignore

@@ -0,0 +1,5 @@
+*.bin
+*.ckpt
+*.safetensor
+*.safetensors
+*.log

stable-diffusion.cpp/models/README.md

@@ -0,0 +1,26 @@
+# Model Convert Script
+
+## Requirements
+
+- vocab.json, from https://huggingface.co/openai/clip-vit-large-patch14/raw/main/vocab.json
+
+
+```shell
+pip install -r requirements.txt
+```
+
+## Usage
+```
+usage: convert.py [-h] [--out_type {f32,f16,q4_0,q4_1,q5_0,q5_1,q8_0}] [--out_file OUT_FILE] model_path
+
+Convert Stable Diffuison model to GGML compatible file format
+
+positional arguments:
+  model_path            model file path (*.pth, *.pt, *.ckpt, *.safetensors)
+
+options:
+  -h, --help            show this help message and exit
+  --out_type {f32,f16,q4_0,q4_1,q5_0,q5_1,q8_0}
+                        output format (default: based on input)
+  --out_file OUT_FILE   path to write to; default: based on input and current working directory
+```

stable-diffusion.cpp/models/convert.py

@@ -0,0 +1,385 @@
+import struct
+import json
+import os
+
+import numpy as np
+import torch
+import safetensors.torch
+
+this_file_dir = os.path.dirname(__file__)
+vocab_dir = this_file_dir
+
+SD1 = 0
+SD2 = 1
+
+ggml_ftype_str_to_int = {
+    "f32": 0,
+    "f16": 1,
+    "q4_0": 2,
+    "q4_1": 3,
+    "q5_0": 8,
+    "q5_1": 9,
+    "q8_0": 7
+}
+
+ggml_ttype_str_to_int = {
+    "f32": 0,
+    "f16": 1,
+    "q4_0": 2,
+    "q4_1": 3,
+    "q5_0": 6,
+    "q5_1": 7,
+    "q8_0": 8
+}
+
+QK4_0 = 32
+def quantize_q4_0(x):
+    assert x.shape[-1] % QK4_0 == 0 and x.shape[-1] > QK4_0
+    x = x.reshape(-1, QK4_0)
+    max = np.take_along_axis(x, np.argmax(np.abs(x), axis=-1)[:, np.newaxis], axis=-1)
+    d = max / -8
+    qs = ((x / d) + 8).round().clip(min=0, max=15).astype(np.int8)
+    half = QK4_0 // 2
+    qs = qs[:, :half] | (qs[:, half:] << 4)
+    d = d.astype(np.float16).view(np.int8)
+    y = np.concatenate((d, qs), axis=-1)
+    return y
+
+QK4_1 = 32
+def quantize_q4_1(x):
+    assert x.shape[-1] % QK4_1 == 0 and x.shape[-1] > QK4_1
+    x = x.reshape(-1, QK4_1)
+    min = np.min(x, axis=-1, keepdims=True)
+    max = np.max(x, axis=-1, keepdims=True)
+    d = (max - min) / ((1 << 4) - 1)
+    qs = ((x - min) / d).round().clip(min=0, max=15).astype(np.int8)
+    half = QK4_1 // 2
+    qs = qs[:, :half] | (qs[:, half:] << 4)
+    d = d.astype(np.float16).view(np.int8)
+    m = min.astype(np.float16).view(np.int8)
+    y = np.concatenate((d, m, qs), axis=-1)
+    return y
+
+QK5_0 = 32
+def quantize_q5_0(x):
+    assert x.shape[-1] % QK5_0 == 0 and x.shape[-1] > QK5_0
+    x = x.reshape(-1, QK5_0)
+    max = np.take_along_axis(x, np.argmax(np.abs(x), axis=-1)[:, np.newaxis], axis=-1)
+    d = max / -16
+    xi = ((x / d) + 16).round().clip(min=0, max=31).astype(np.int8)
+    half = QK5_0 // 2
+    qs = (xi[:, :half] & 0x0F) | (xi[:, half:] << 4)
+    qh = np.zeros(qs.shape[:-1], dtype=np.int32)
+    for i in range(QK5_0):
+        qh |= ((xi[:, i] & 0x10) >> 4).astype(np.int32) << i
+    d = d.astype(np.float16).view(np.int8)
+    qh = qh[..., np.newaxis].view(np.int8)
+    y = np.concatenate((d, qh, qs), axis=-1)
+    return y
+
+QK5_1 = 32
+def quantize_q5_1(x):
+    assert x.shape[-1] % QK5_1 == 0 and x.shape[-1] > QK5_1
+    x = x.reshape(-1, QK5_1)
+    min = np.min(x, axis=-1, keepdims=True)
+    max = np.max(x, axis=-1, keepdims=True)
+    d = (max - min) / ((1 << 5) - 1)
+    xi = ((x - min) / d).round().clip(min=0, max=31).astype(np.int8)
+    half = QK5_1//2
+    qs = (xi[:, :half] & 0x0F) | (xi[:, half:] << 4)
+    qh = np.zeros(xi.shape[:-1], dtype=np.int32)
+    for i in range(QK5_1):
+        qh |= ((xi[:, i] & 0x10) >> 4).astype(np.int32) << i
+    d = d.astype(np.float16).view(np.int8)
+    m = min.astype(np.float16).view(np.int8)
+    qh = qh[..., np.newaxis].view(np.int8)
+    ndarray = np.concatenate((d, m, qh, qs), axis=-1)
+    return ndarray
+
+QK8_0 = 32
+def quantize_q8_0(x):
+    assert x.shape[-1] % QK8_0 == 0 and x.shape[-1] > QK8_0
+    x = x.reshape(-1, QK8_0)
+    amax = np.max(np.abs(x), axis=-1, keepdims=True) 
+    d = amax / ((1 << 7) - 1)
+    qs = (x / d).round().clip(min=-128, max=127).astype(np.int8)
+    d = d.astype(np.float16).view(np.int8)
+    y = np.concatenate((d, qs), axis=-1)
+    return y
+
+# copy from https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py#L16
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+def load_model_from_file(model_path):
+    print("loading model from {}".format(model_path))
+    if model_path.lower().endswith(".safetensors"):
+        pl_sd = safetensors.torch.load_file(model_path, device="cpu")
+    else:
+        pl_sd = torch.load(model_path, map_location="cpu")
+    state_dict = pl_sd["state_dict"] if "state_dict" in pl_sd else pl_sd
+    print("loading model from {} completed".format(model_path))
+    return state_dict
+
+def get_alpha_comprod(linear_start=0.00085, linear_end=0.0120, timesteps=1000):
+    betas = torch.linspace(linear_start ** 0.5, linear_end ** 0.5, timesteps, dtype=torch.float32) ** 2
+    alphas = 1. - betas
+    alphas_cumprod = np.cumprod(alphas.numpy(), axis=0)
+    return torch.tensor(alphas_cumprod)
+
+unused_tensors = [
+    "betas",
+    "alphas_cumprod_prev",
+    "sqrt_alphas_cumprod",
+    "sqrt_one_minus_alphas_cumprod",
+    "log_one_minus_alphas_cumprod",
+    "sqrt_recip_alphas_cumprod",
+    "sqrt_recipm1_alphas_cumprod",
+    "posterior_variance",
+    "posterior_log_variance_clipped",
+    "posterior_mean_coef1",
+    "posterior_mean_coef2",
+    "cond_stage_model.transformer.text_model.embeddings.position_ids",
+    "cond_stage_model.model.logit_scale",
+    "cond_stage_model.model.text_projection",
+    "model_ema.decay",
+    "model_ema.num_updates",
+    "control_model",
+    "lora_te_text_model",
+    "embedding_manager"
+]
+
+
+def preprocess(state_dict):
+    alphas_cumprod = state_dict.get("alphas_cumprod")
+    if alphas_cumprod != None:
+        # print((np.abs(get_alpha_comprod().numpy() - alphas_cumprod.numpy()) < 0.000001).all())
+        pass
+    else:
+        print("no alphas_cumprod in file, generate new one")
+        alphas_cumprod = get_alpha_comprod()
+        state_dict["alphas_cumprod"] = alphas_cumprod
+    
+    new_state_dict = {}
+    for name, w in state_dict.items():
+        # ignore unused tensors
+        if not isinstance(w, torch.Tensor):
+            continue
+        skip = False
+        for unused_tensor in unused_tensors:
+            if name.startswith(unused_tensor):
+                skip = True
+                break
+        if skip:
+            continue
+
+        # # convert BF16 to FP16
+        if w.dtype == torch.bfloat16:
+            w = w.to(torch.float16)
+
+        # convert open_clip to hf CLIPTextModel (for SD2.x)
+        open_clip_to_hf_clip_model = {
+            "cond_stage_model.model.ln_final.bias": "cond_stage_model.transformer.text_model.final_layer_norm.bias",
+            "cond_stage_model.model.ln_final.weight": "cond_stage_model.transformer.text_model.final_layer_norm.weight",
+            "cond_stage_model.model.positional_embedding": "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight",
+            "cond_stage_model.model.token_embedding.weight": "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight",
+            "first_stage_model.decoder.mid.attn_1.to_k.bias": "first_stage_model.decoder.mid.attn_1.k.bias",
+            "first_stage_model.decoder.mid.attn_1.to_k.weight": "first_stage_model.decoder.mid.attn_1.k.weight",
+            "first_stage_model.decoder.mid.attn_1.to_out.0.bias": "first_stage_model.decoder.mid.attn_1.proj_out.bias",
+            "first_stage_model.decoder.mid.attn_1.to_out.0.weight": "first_stage_model.decoder.mid.attn_1.proj_out.weight",
+            "first_stage_model.decoder.mid.attn_1.to_q.bias": "first_stage_model.decoder.mid.attn_1.q.bias",
+            "first_stage_model.decoder.mid.attn_1.to_q.weight": "first_stage_model.decoder.mid.attn_1.q.weight",
+            "first_stage_model.decoder.mid.attn_1.to_v.bias": "first_stage_model.decoder.mid.attn_1.v.bias",
+            "first_stage_model.decoder.mid.attn_1.to_v.weight": "first_stage_model.decoder.mid.attn_1.v.weight",
+        }
+        open_clip_to_hk_clip_resblock = {
+            "attn.out_proj.bias": "self_attn.out_proj.bias",
+            "attn.out_proj.weight": "self_attn.out_proj.weight",
+            "ln_1.bias": "layer_norm1.bias",
+            "ln_1.weight": "layer_norm1.weight",
+            "ln_2.bias": "layer_norm2.bias",
+            "ln_2.weight": "layer_norm2.weight",
+            "mlp.c_fc.bias": "mlp.fc1.bias",
+            "mlp.c_fc.weight": "mlp.fc1.weight",
+            "mlp.c_proj.bias": "mlp.fc2.bias",
+            "mlp.c_proj.weight": "mlp.fc2.weight",
+        }
+        open_clip_resblock_prefix = "cond_stage_model.model.transformer.resblocks."
+        hf_clip_resblock_prefix = "cond_stage_model.transformer.text_model.encoder.layers."
+        if name in open_clip_to_hf_clip_model:
+            new_name = open_clip_to_hf_clip_model[name]
+            print(f"preprocess {name} => {new_name}")
+            name = new_name
+        if name.startswith(open_clip_resblock_prefix):
+            remain = name[len(open_clip_resblock_prefix):]
+            idx = remain.split(".")[0]
+            suffix = remain[len(idx)+1:]
+            if suffix == "attn.in_proj_weight":
+                w_q, w_k, w_v = w.chunk(3)
+                for new_suffix, new_w in zip(["self_attn.q_proj.weight", "self_attn.k_proj.weight", "self_attn.v_proj.weight"], [w_q, w_k, w_v]):
+                    new_name = hf_clip_resblock_prefix + idx + "." + new_suffix
+                    new_state_dict[new_name] = new_w
+                    print(f"preprocess {name}{w.size()} => {new_name}{new_w.size()}")
+            elif suffix == "attn.in_proj_bias":
+                w_q, w_k, w_v = w.chunk(3)
+                for new_suffix, new_w in zip(["self_attn.q_proj.bias", "self_attn.k_proj.bias", "self_attn.v_proj.bias"], [w_q, w_k, w_v]):
+                    new_name = hf_clip_resblock_prefix + idx + "." + new_suffix
+                    new_state_dict[new_name] = new_w
+                    print(f"preprocess {name}{w.size()} => {new_name}{new_w.size()}")
+            else:
+                new_suffix = open_clip_to_hk_clip_resblock[suffix]
+                new_name = hf_clip_resblock_prefix + idx + "." + new_suffix
+                new_state_dict[new_name] = w
+                print(f"preprocess {name} => {new_name}")
+            continue
+        
+        # convert unet transformer linear to conv2d 1x1
+        if name.startswith("model.diffusion_model.") and (name.endswith("proj_in.weight") or name.endswith("proj_out.weight")):
+            if len(w.shape) == 2:
+                new_w = w.unsqueeze(2).unsqueeze(3)
+                new_state_dict[name] = new_w
+                print(f"preprocess {name} {w.size()} => {name} {new_w.size()}")
+                continue
+
+        # convert vae attn block linear to conv2d 1x1
+        if name.startswith("first_stage_model.") and "attn_1" in name:
+            if len(w.shape) == 2:
+                new_w = w.unsqueeze(2).unsqueeze(3)
+                new_state_dict[name] = new_w
+                print(f"preprocess {name} {w.size()} => {name} {new_w.size()}")
+                continue
+
+        new_state_dict[name] = w
+    return new_state_dict
+
+def convert(model_path, out_type = None, out_file=None):
+    # load model
+    with open(os.path.join(vocab_dir, "vocab.json"), encoding="utf-8") as f:
+        clip_vocab = json.load(f)
+    
+    state_dict = load_model_from_file(model_path)
+    model_type = SD1
+    if "cond_stage_model.model.token_embedding.weight" in state_dict.keys():
+        model_type = SD2
+        print("Stable diffuison 2.x")
+    else:
+        print("Stable diffuison 1.x")
+    state_dict = preprocess(state_dict)
+
+    # output option
+    if out_type == None:
+        weight = state_dict["model.diffusion_model.input_blocks.0.0.weight"].numpy()
+        if weight.dtype == np.float32:
+            out_type = "f32"
+        elif weight.dtype == np.float16:
+            out_type = "f16"
+        elif weight.dtype == np.float64:
+            out_type = "f32"
+        else:
+            raise Exception("unsupported weight type %s" % weight.dtype)
+    if out_file == None:
+        out_file = os.path.splitext(os.path.basename(model_path))[0] + f"-ggml-model-{out_type}.bin"
+        out_file = os.path.join(os.getcwd(), out_file)
+    print(f"Saving GGML compatible file to {out_file}")
+
+    # convert and save
+    with open(out_file, "wb") as file:
+        # magic: ggml in hex
+        file.write(struct.pack("i", 0x67676D6C))
+        # model & file type
+        ftype = (model_type << 16) | ggml_ftype_str_to_int[out_type]
+        file.write(struct.pack("i", ftype))
+
+        # vocab
+        byte_encoder = bytes_to_unicode()
+        byte_decoder = {v: k for k, v in byte_encoder.items()}
+        file.write(struct.pack("i", len(clip_vocab)))
+        for key in clip_vocab:
+            text = bytearray([byte_decoder[c] for c in key])
+            file.write(struct.pack("i", len(text)))
+            file.write(text)
+        
+        # weights
+        for name in state_dict.keys():
+            if not isinstance(state_dict[name], torch.Tensor):
+                continue
+            skip = False
+            for unused_tensor in unused_tensors:
+                if name.startswith(unused_tensor):
+                    skip = True
+                    break
+            if skip:
+                continue
+            if name in unused_tensors:
+                continue
+            data = state_dict[name].numpy()
+
+            n_dims = len(data.shape)
+            shape = data.shape
+            old_type = data.dtype
+
+            ttype = "f32"
+            if n_dims == 4:
+                data = data.astype(np.float16)
+                ttype = "f16"
+            elif n_dims == 2 and name[-7:] == ".weight":
+                if out_type == "f32":
+                    data = data.astype(np.float32)
+                elif out_type == "f16":
+                    data = data.astype(np.float16)
+                elif out_type == "q4_0":
+                    data = quantize_q4_0(data)
+                elif out_type == "q4_1":
+                    data = quantize_q4_1(data)
+                elif out_type == "q5_0":
+                    data = quantize_q5_0(data)
+                elif out_type == "q5_1":
+                    data = quantize_q5_1(data)
+                elif out_type == "q8_0":
+                    data = quantize_q8_0(data)
+                else:
+                    raise Exception("invalid out_type {}".format(out_type))
+                ttype = out_type
+            else:
+                data = data.astype(np.float32)
+                ttype = "f32"
+            
+            print("Processing tensor: {} with shape {}, {} -> {}".format(name, data.shape, old_type, ttype))
+
+            # header
+            name_bytes = name.encode("utf-8")
+            file.write(struct.pack("iii", n_dims, len(name_bytes), ggml_ttype_str_to_int[ttype]))
+            for i in range(n_dims):
+                file.write(struct.pack("i", shape[n_dims - 1 - i]))
+            file.write(name_bytes)
+            # data
+            data.tofile(file)
+        print("Convert done")
+        print(f"Saved GGML compatible file to {out_file}")
+
+if __name__ == "__main__":
+    import argparse
+    parser = argparse.ArgumentParser(description="Convert Stable Diffuison model to GGML compatible file format")
+    parser.add_argument("--out_type", choices=["f32", "f16", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0"], help="output format (default: based on input)")
+    parser.add_argument("--out_file", help="path to write to; default: based on input and current working directory")
+    parser.add_argument("model_path", help="model file path (*.pth, *.pt, *.ckpt, *.safetensors)")
+    args = parser.parse_args()
+    convert(args.model_path, args.out_type, args.out_file)

stable-diffusion.cpp/models/requirements.txt

@@ -0,0 +1,4 @@
+numpy
+torch
+safetensors
+pytorch_lightning

stable-diffusion.cpp/rng.h

@@ -0,0 +1,35 @@
+#ifndef __RNG_H__
+#define __RNG_H__
+
+#include <random>
+#include <vector>
+
+class RNG {
+   public:
+    virtual void manual_seed(uint64_t seed) = 0;
+    virtual std::vector<float> randn(uint32_t n) = 0;
+};
+
+class STDDefaultRNG : public RNG {
+   private:
+    std::default_random_engine generator;
+
+   public:
+    void manual_seed(uint64_t seed) {
+        generator.seed(seed);
+    }
+
+    std::vector<float> randn(uint32_t n) {
+        std::vector<float> result;
+        float mean = 0.0;
+        float stddev = 1.0;
+        std::normal_distribution<float> distribution(mean, stddev);
+        for (int i = 0; i < n; i++) {
+            float random_number = distribution(generator);
+            result.push_back(random_number);
+        }
+        return result;
+    }
+};
+
+#endif  // __RNG_H__

stable-diffusion.cpp/rng_philox.h

@@ -0,0 +1,125 @@
+#ifndef __RNG_PHILOX_H__
+#define __RNG_PHILOX_H__
+
+#include <cmath>
+#include <vector>
+
+#include "rng.h"
+
+// RNG imitiating torch cuda randn on CPU.
+// Port from: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/5ef669de080814067961f28357256e8fe27544f4/modules/rng_philox.py
+class PhiloxRNG : public RNG {
+   private:
+    uint64_t seed;
+    uint32_t offset;
+
+   private:
+    std::vector<uint32_t> philox_m = {0xD2511F53, 0xCD9E8D57};
+    std::vector<uint32_t> philox_w = {0x9E3779B9, 0xBB67AE85};
+    float two_pow32_inv = 2.3283064e-10;
+    float two_pow32_inv_2pi = 2.3283064e-10 * 6.2831855;
+
+    std::vector<uint32_t> uint32(uint64_t x) {
+        std::vector<uint32_t> result(2);
+        result[0] = static_cast<uint32_t>(x & 0xFFFFFFFF);
+        result[1] = static_cast<uint32_t>(x >> 32);
+        return result;
+    }
+
+    std::vector<std::vector<uint32_t>> uint32(const std::vector<uint64_t>& x) {
+        int N = x.size();
+        std::vector<std::vector<uint32_t>> result(2, std::vector<uint32_t>(N));
+
+        for (int i = 0; i < N; ++i) {
+            result[0][i] = static_cast<uint32_t>(x[i] & 0xFFFFFFFF);
+            result[1][i] = static_cast<uint32_t>(x[i] >> 32);
+        }
+
+        return result;
+    }
+
+    //  A single round of the Philox 4x32 random number generator.
+    void philox4_round(std::vector<std::vector<uint32_t>>& counter,
+                       const std::vector<std::vector<uint32_t>>& key) {
+        uint32_t N = counter[0].size();
+        for (uint32_t i = 0; i < N; i++) {
+            std::vector<uint32_t> v1 = uint32(static_cast<uint64_t>(counter[0][i]) * static_cast<uint64_t>(philox_m[0]));
+            std::vector<uint32_t> v2 = uint32(static_cast<uint64_t>(counter[2][i]) * static_cast<uint64_t>(philox_m[1]));
+
+            counter[0][i] = v2[1] ^ counter[1][i] ^ key[0][i];
+            counter[1][i] = v2[0];
+            counter[2][i] = v1[1] ^ counter[3][i] ^ key[1][i];
+            counter[3][i] = v1[0];
+        }
+    }
+
+    // Generates 32-bit random numbers using the Philox 4x32 random number generator.
+    // Parameters:
+    //     counter : A 4xN array of 32-bit integers representing the counter values (offset into generation).
+    //     key : A 2xN array of 32-bit integers representing the key values (seed).
+    //     rounds : The number of rounds to perform.
+    // Returns:
+    //     std::vector<std::vector<uint32_t>>: A 4xN array of 32-bit integers containing the generated random numbers.
+    std::vector<std::vector<uint32_t>> philox4_32(std::vector<std::vector<uint32_t>>& counter,
+                                                  std::vector<std::vector<uint32_t>>& key,
+                                                  int rounds = 10) {
+        uint32_t N = counter[0].size();
+        for (int i = 0; i < rounds - 1; ++i) {
+            philox4_round(counter, key);
+
+            for (uint32_t j = 0; j < N; ++j) {
+                key[0][j] += philox_w[0];
+                key[1][j] += philox_w[1];
+            }
+        }
+
+        philox4_round(counter, key);
+        return counter;
+    }
+
+    float box_muller(float x, float y) {
+        float u = x * two_pow32_inv + two_pow32_inv / 2;
+        float v = y * two_pow32_inv_2pi + two_pow32_inv_2pi / 2;
+
+        float s = sqrt(-2.0 * log(u));
+
+        float r1 = s * sin(v);
+        return r1;
+    }
+
+   public:
+    PhiloxRNG(uint64_t seed = 0) {
+        this->seed = seed;
+        this->offset = 0;
+    }
+
+    void manual_seed(uint64_t seed) {
+        this->seed = seed;
+        this->offset = 0;
+    }
+
+    std::vector<float> randn(uint32_t n) {
+        std::vector<std::vector<uint32_t>> counter(4, std::vector<uint32_t>(n, 0));
+        for (uint32_t i = 0; i < n; i++) {
+            counter[0][i] = this->offset;
+        }
+
+        for (uint32_t i = 0; i < n; i++) {
+            counter[2][i] = i;
+        }
+        this->offset += 1;
+
+        std::vector<uint64_t> key(n, this->seed);
+        std::vector<std::vector<uint32_t>> key_uint32 = uint32(key);
+
+        std::vector<std::vector<uint32_t>> g = philox4_32(counter, key_uint32);
+
+        std::vector<float> result;
+        for (int i = 0; i < n; ++i) {
+            result.push_back(box_muller(g[0][i], g[1][i]));
+        }
+        return result;
+    }
+};
+
+#endif  // __RNG_PHILOX_H__

stable-diffusion.cpp/stable-diffusion.h

@@ -0,0 +1,74 @@
+#ifndef __STABLE_DIFFUSION_H__
+#define __STABLE_DIFFUSION_H__
+
+#include <memory>
+#include <vector>
+
+enum SDLogLevel {
+    DEBUG,
+    INFO,
+    WARN,
+    ERROR
+};
+
+enum RNGType {
+    STD_DEFAULT_RNG,
+    CUDA_RNG
+};
+
+enum SampleMethod {
+    EULER_A,
+    EULER,
+    HEUN,
+    DPM2,
+    DPMPP2S_A,
+    DPMPP2M,
+    DPMPP2Mv2,
+    N_SAMPLE_METHODS
+};
+
+enum Schedule {
+    DEFAULT,
+    DISCRETE,
+    KARRAS,
+    N_SCHEDULES
+};
+
+class StableDiffusionGGML;
+
+class StableDiffusion {
+   private:
+    std::shared_ptr<StableDiffusionGGML> sd;
+
+   public:
+    StableDiffusion(int n_threads = -1,
+                    bool vae_decode_only = false,
+                    bool free_params_immediately = false,
+                    RNGType rng_type = STD_DEFAULT_RNG);
+    bool load_from_file(const std::string& file_path, Schedule d = DEFAULT);
+    std::vector<uint8_t> txt2img(
+        const std::string& prompt,
+        const std::string& negative_prompt,
+        float cfg_scale,
+        int width,
+        int height,
+        SampleMethod sample_method,
+        int sample_steps,
+        int64_t seed);
+    std::vector<uint8_t> img2img(
+        const std::vector<uint8_t>& init_img,
+        const std::string& prompt,
+        const std::string& negative_prompt,
+        float cfg_scale,
+        int width,
+        int height,
+        SampleMethod sample_method,
+        int sample_steps,
+        float strength,
+        int64_t seed);
+};
+
+void set_sd_log_level(SDLogLevel level);
+std::string sd_get_system_info();
+
+#endif  // __STABLE_DIFFUSION_H__