├── source ├── config.h.in ├── common.h ├── source.cpp └── kernel.cu ├── README.md ├── CMakeLists.txt ├── .github └── workflows │ ├── linux.yml │ └── windows.yml └── LICENSE /source/config.h.in: -------------------------------------------------------------------------------- 1 | #define VERSION "@VCS_TAG@" -------------------------------------------------------------------------------- /source/common.h: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | enum struct ChannelMode { Y, UV, YUV, RGB }; 4 | 5 | extern cudaError_t nlmeans( 6 | void * d_dst, 7 | void * d_src, 8 | float * d_buffer, 9 | float * d_buffer_bwd, 10 | float * d_buffer_fwd, 11 | float * d_wdst, 12 | float * d_weight, 13 | float * d_max_weight, 14 | bool is_float, 15 | int bits_per_sample, 16 | int width, int height, int image_stride, int buffer_stride, 17 | int radius, int spatial_radius, int block_radius, float h2_inv_norm, 18 | ChannelMode channels, int wmode, float wref, bool has_ref, 19 | cudaStream_t stream 20 | ); 21 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # vs-nlm-cuda 2 | Non-local means denoise filter in CUDA, drop-in replacement of the KNLMeansCL for VapourSynth 3 | 4 | ## Usage 5 | Prototype: 6 | 7 | `core.nlm_cuda.NLMeans(clip clip[, int d = 1, int a = 2, int s = 4, float h = 1.2, string channels = "AUTO", int wmode = 0, float wref = 1.0, clip rclip = None, int device_id = 0, int num_streams = 1])` 8 | 9 | ## Compilation 10 | ```bash 11 | cmake -S . -B build -D CMAKE_BUILD_TYPE=Release \ 12 | -D CMAKE_CUDA_FLAGS="--use_fast_math" \ 13 | -D CMAKE_CUDA_ARCHITECTURES="50;61-real;70-virtual;75-real;86-real;89-real" 14 | 15 | cmake --build build 16 | 17 | cmake --install build 18 | ``` 19 | 20 | -------------------------------------------------------------------------------- /CMakeLists.txt: -------------------------------------------------------------------------------- 1 | cmake_minimum_required(VERSION 3.22.0) 2 | 3 | project(vs-nlm-cuda VERSION 0.1 LANGUAGES CUDA CXX) 4 | 5 | find_package(CUDAToolkit REQUIRED) 6 | 7 | add_library(vsnlm_cuda SHARED 8 | source/source.cpp 9 | source/kernel.cu 10 | ) 11 | 12 | set_target_properties(vsnlm_cuda PROPERTIES 13 | CUDA_EXTENSIONS OFF 14 | CUDA_STANDARD 17 15 | CUDA_STANDARD_REQUIRED ON 16 | CXX_EXTENSIONS OFF 17 | CXX_STANDARD 20 18 | CXX_STANDARD_REQUIRED ON 19 | ) 20 | 21 | target_link_libraries(vsnlm_cuda PRIVATE CUDA::cudart_static) 22 | 23 | find_package(PkgConfig QUIET MODULE) 24 | 25 | if(PKG_CONFIG_FOUND) 26 | pkg_search_module(VS vapoursynth) 27 | 28 | if(VS_FOUND) 29 | message(STATUS "Found VapourSynth r${VS_VERSION}") 30 | 31 | cmake_path(APPEND install_dir ${VS_LIBDIR} vapoursynth) 32 | target_include_directories(vsnlm_cuda PRIVATE ${VS_INCLUDE_DIRS}) 33 | 34 | install(TARGETS vsnlm_cuda LIBRARY DESTINATION ${install_dir}) 35 | endif() 36 | endif() 37 | 38 | if(NOT VS_FOUND) 39 | set(VS_INCLUDE_DIR "" CACHE PATH "Path to VapourSynth headers") 40 | 41 | if(VS_INCLUDE_DIR STREQUAL "") 42 | message(WARNING "VapourSynth not found") 43 | endif() 44 | 45 | target_include_directories(vsnlm_cuda PRIVATE ${VS_INCLUDE_DIR}) 46 | 47 | install(TARGETS vsnlm_cuda LIBRARY RUNTIME) 48 | endif() 49 | 50 | find_package(Git QUIET) 51 | 52 | if(GIT_FOUND) 53 | execute_process( 54 | COMMAND ${GIT_EXECUTABLE} describe --tags --long --always 55 | WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}" 56 | OUTPUT_VARIABLE VCS_TAG 57 | ) 58 | if(VCS_TAG) 59 | string(STRIP ${VCS_TAG} VCS_TAG) 60 | endif() 61 | endif() 62 | 63 | if(VCS_TAG) 64 | message(STATUS "vs-nlm-cuda ${VCS_TAG}") 65 | else() 66 | message(WARNING "unknown plugin version") 67 | set(VCS_TAG "unknown") 68 | endif() 69 | 70 | configure_file(source/config.h.in config.h) 71 | 72 | target_include_directories(vsnlm_cuda PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) 73 | -------------------------------------------------------------------------------- /.github/workflows/linux.yml: -------------------------------------------------------------------------------- 1 | name: Build (Linux) 2 | 3 | on: 4 | push: 5 | paths: 6 | - 'CMakeLists.txt' 7 | - 'source/*' 8 | - '.github/workflows/linux.yml' 9 | workflow_dispatch: 10 | 11 | jobs: 12 | build-linux: 13 | runs-on: ubuntu-24.04 14 | steps: 15 | - name: Checkout repo 16 | uses: actions/checkout@v4 17 | with: 18 | submodules: true 19 | 20 | - name: Setup Ninja 21 | run: pip3 install ninja 22 | 23 | - name: Setup CUDA 24 | run: | 25 | wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb 26 | sudo dpkg -i cuda-keyring_1.1-1_all.deb 27 | sudo apt-get update 28 | sudo apt-get install -y cuda-nvcc-13-0 cuda-cudart-dev-13-0 cuda-nvrtc-dev-13-0 29 | echo "PATH=/usr/local/cuda/bin${PATH:+:${PATH}}" >> $GITHUB_ENV 30 | echo "CUDA_PATH=/usr/local/cuda" >> $GITHUB_ENV 31 | echo "LD_LIBRARY_PATH=/usr/local/cuda/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}" >> $GITHUB_ENV 32 | 33 | - name: Download VapourSynth headers 34 | run: | 35 | wget -q -O vs.zip https://github.com/vapoursynth/vapoursynth/archive/refs/tags/R57.zip 36 | unzip -q vs.zip 37 | mv vapoursynth*/ vapoursynth 38 | 39 | - name: Configure 40 | run: cmake -S . -B build -G Ninja 41 | -D VS_INCLUDE_DIR="`pwd`/vapoursynth/include" 42 | -D CMAKE_BUILD_TYPE=Release 43 | -D CMAKE_CXX_COMPILER=g++-12 44 | -D CMAKE_CXX_FLAGS="-Wall -ffast-math" 45 | -D CMAKE_CUDA_FLAGS="--threads 0 --use_fast_math --resource-usage -Wno-deprecated-gpu-targets" 46 | -D CMAKE_CUDA_ARCHITECTURES="75;86-real;89-real;120-real" 47 | 48 | - name: Build 49 | run: cmake --build build --config Release --verbose 50 | 51 | - name: Install 52 | run: cmake --install build --prefix artifact 53 | 54 | - name: Upload 55 | uses: actions/upload-artifact@v4 56 | with: 57 | name: vs-nlm-cuda-linux 58 | path: artifact 59 | -------------------------------------------------------------------------------- /.github/workflows/windows.yml: -------------------------------------------------------------------------------- 1 | name: Build (Windows) 2 | 3 | on: 4 | push: 5 | paths: 6 | - 'CMakeLists.txt' 7 | - 'source/*' 8 | - '.github/workflows/windows.yml' 9 | workflow_dispatch: 10 | inputs: 11 | tag: 12 | description: 'which tag to upload to' 13 | default: '' 14 | 15 | jobs: 16 | build-windows: 17 | runs-on: windows-2025 18 | 19 | defaults: 20 | run: 21 | shell: cmd 22 | 23 | steps: 24 | - name: Checkout repo 25 | uses: actions/checkout@v4 26 | with: 27 | submodules: true 28 | 29 | - name: Setup MSVC 30 | uses: ilammy/msvc-dev-cmd@v1 31 | 32 | - name: Setup Ninja 33 | run: pip install ninja 34 | 35 | - name: Cache CUDA 36 | id: cache-cuda 37 | uses: actions/cache@v4 38 | with: 39 | path: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA 40 | key: ${{ runner.os }}-cuda-12.9.1 41 | 42 | - name: Setup CUDA 43 | if: steps.cache-cuda.outputs.cache-hit != 'true' 44 | run: | 45 | curl -s -o cuda_installer.exe -L https://developer.download.nvidia.com/compute/cuda/13.0.1/network_installers/cuda_13.0.1_windows_network.exe 46 | cuda_installer.exe -s nvcc_13.0 cudart_13.0 nvrtc_dev_13.0 crt_13.0 nvptxcompiler_13.0 nvvm_13.0 47 | 48 | - name: Download VapourSynth headers 49 | run: | 50 | curl -s -o vs.zip -L https://github.com/vapoursynth/vapoursynth/archive/refs/tags/R57.zip 51 | unzip -q vs.zip 52 | mv vapoursynth-*/ vapoursynth/ 53 | 54 | - name: Configure 55 | run: cmake -S . -B build_cuda -G Ninja -LA 56 | -D CMAKE_BUILD_TYPE=Release 57 | -D VS_INCLUDE_DIR="%cd%\vapoursynth\include" 58 | -D CMAKE_CXX_FLAGS="/fp:fast /EHsc" 59 | -D CMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded 60 | -D CMAKE_CUDA_FLAGS="--threads 0 --use_fast_math --resource-usage -Wno-deprecated-gpu-targets" 61 | -D CMAKE_CUDA_ARCHITECTURES="75;86-real;89-real;120-real" 62 | env: 63 | CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0 64 | 65 | - name: Build 66 | run: cmake --build build_cuda --verbose 67 | 68 | - name: Install 69 | run: cmake --install build_cuda --prefix install 70 | 71 | - name: Prepare for upload 72 | run: | 73 | mkdir artifact 74 | copy install\bin\*.dll artifact 75 | 76 | - name: Upload 77 | uses: actions/upload-artifact@v4 78 | with: 79 | name: vs-nlm-cuda-windows 80 | path: artifact 81 | 82 | - name: Compress artifact for release 83 | if: github.event_name == 'workflow_dispatch' && github.event.inputs.tag != '' 84 | run: | 85 | cd artifact 86 | 87 | mkdir vs-nlm-cuda-windows-${{ github.event.inputs.tag }} 88 | xcopy vsnlm_cuda.dll vs-nlm-cuda-windows-${{ github.event.inputs.tag }} /f 89 | 7z a -t7z -mx=9 ../vs-nlm-cuda-windows-${{ github.event.inputs.tag }}.7z vs-nlm-cuda-windows-${{ github.event.inputs.tag }} 90 | 91 | - name: Release 92 | uses: softprops/action-gh-release@v1 93 | if: github.event_name == 'workflow_dispatch' && github.event.inputs.tag != '' 94 | with: 95 | tag_name: ${{ github.event.inputs.tag }} 96 | files: vs-nlm-cuda-windows-${{ github.event.inputs.tag }}.7z 97 | fail_on_unmatched_files: true 98 | generate_release_notes: false 99 | prerelease: true 100 | 101 | -------------------------------------------------------------------------------- /source/source.cpp: -------------------------------------------------------------------------------- 1 | // based on KNLMeansCL by Khanattila 2 | 3 | #include 4 | #include 5 | #include 6 | #include 7 | #include 8 | #include 9 | #include 10 | #include 11 | #include 12 | #include 13 | #include 14 | #include 15 | #include 16 | #include 17 | #include 18 | #include 19 | #include 20 | 21 | #include 22 | #include 23 | 24 | #include "common.h" 25 | #include // generated by cmake and git 26 | 27 | static bool success(cudaError_t result) { 28 | return result == cudaSuccess; 29 | } 30 | 31 | static const char * get_error(cudaError_t error) { 32 | return cudaGetErrorString(error); 33 | } 34 | 35 | #define showError(expr) show_error_impl(expr, # expr, __LINE__) 36 | template 37 | static void show_error_impl(T result, const char * source, int line_no) { 38 | if (!success(result)) [[unlikely]] { 39 | std::fprintf(stderr, "[%d] %s failed: %s\n", line_no, source, get_error(result)); 40 | } 41 | } 42 | 43 | #define checkError(expr) do { \ 44 | if (auto result = expr; !success(result)) [[unlikely]] { \ 45 | std::ostringstream error; \ 46 | error << '[' << __LINE__ << "] '" # expr "' failed: " << get_error(result); \ 47 | return set_error(error.str().c_str()); \ 48 | } \ 49 | } while (0) 50 | 51 | static void cudaStreamDestroyCustom(cudaStream_t stream) { 52 | showError(cudaStreamDestroy(stream)); 53 | } 54 | 55 | static void cudaEventDestroyCustom(cudaEvent_t event) { 56 | showError(cudaEventDestroy(event)); 57 | } 58 | 59 | static void cudaFreeCustom(void * p) { 60 | showError(cudaFree(p)); 61 | } 62 | 63 | static void cudaFreeHostCustom(void * p) { 64 | showError(cudaFreeHost(p)); 65 | } 66 | 67 | 68 | struct node_freer { 69 | const VSAPI * & vsapi; 70 | VSNodeRef * node {}; 71 | void release() { 72 | node = nullptr; 73 | } 74 | ~node_freer() { 75 | if (node) { 76 | vsapi->freeNode(node); 77 | } 78 | } 79 | }; 80 | 81 | 82 | template 83 | requires 84 | std::default_initializable && 85 | std::is_trivially_copy_assignable_v && 86 | std::convertible_to && 87 | std::invocable && 88 | (std::is_pointer_v || unsafe) // e.g. CUdeviceptr is not a pointer 89 | struct Resource { 90 | T data; 91 | 92 | [[nodiscard]] constexpr Resource() noexcept = default; 93 | 94 | [[nodiscard]] constexpr Resource(T x) noexcept : data(x) {} 95 | 96 | [[nodiscard]] constexpr Resource(Resource&& other) noexcept 97 | : data(std::exchange(other.data, T{})) 98 | { } 99 | 100 | Resource& operator=(Resource&& other) noexcept { 101 | if (this == &other) return *this; 102 | deleter_(data); 103 | data = std::exchange(other.data, T{}); 104 | return *this; 105 | } 106 | 107 | Resource operator=(Resource other) = delete; 108 | 109 | Resource(const Resource& other) = delete; 110 | 111 | constexpr operator T() const noexcept { 112 | return data; 113 | } 114 | 115 | constexpr auto deleter_(T x) noexcept { 116 | if (x) { 117 | deleter(x); 118 | x = T{}; 119 | } 120 | } 121 | 122 | Resource& operator=(T x) noexcept { 123 | deleter_(data); 124 | data = x; 125 | return *this; 126 | } 127 | 128 | constexpr ~Resource() noexcept { 129 | deleter_(data); 130 | } 131 | }; 132 | 133 | 134 | template 135 | static T square(const T & x) { 136 | return x * x; 137 | } 138 | 139 | 140 | struct ticket_semaphore { 141 | std::atomic ticket {}; 142 | std::atomic current {}; 143 | 144 | void acquire() noexcept { 145 | intptr_t tk { ticket.fetch_add(1, std::memory_order::acquire) }; 146 | while (true) { 147 | intptr_t curr { current.load(std::memory_order::acquire) }; 148 | if (tk <= curr) { 149 | return; 150 | } 151 | current.wait(curr, std::memory_order::relaxed); 152 | } 153 | } 154 | 155 | void release() noexcept { 156 | current.fetch_add(1, std::memory_order::release); 157 | current.notify_all(); 158 | } 159 | }; 160 | 161 | 162 | // [num_planes, width, height] 163 | static std::array get_filter_shape(ChannelMode channels, const VSVideoInfo * vi) { 164 | switch (channels) { 165 | case ChannelMode::Y: 166 | return { 1, vi->width, vi->height }; 167 | case ChannelMode::UV: 168 | return { 2, vi->width >> vi->format->subSamplingW, vi->height >> vi->format->subSamplingH }; 169 | case ChannelMode::YUV: 170 | case ChannelMode::RGB: 171 | return { 3, vi->width, vi->height }; 172 | } 173 | 174 | return {}; 175 | } 176 | 177 | 178 | struct NLMeansStreamData { 179 | Resource stream; 180 | 181 | Resource event; 182 | 183 | int buffer_stride; 184 | int image_stride; 185 | Resource h_buffer; 186 | Resource d_src_ref; 187 | Resource d_buffer; 188 | Resource d_buffer_bwd; 189 | Resource d_buffer_fwd; // may be empty 190 | Resource d_wdst; 191 | Resource d_weight; 192 | Resource d_max_weight; 193 | Resource d_dst; 194 | }; 195 | 196 | 197 | struct NLMeansData { 198 | VSNodeRef * node; 199 | int radius; // d 200 | int spatial_radius; // a 201 | int block_radius; // s 202 | float h2_inv_norm; // h 203 | ChannelMode channels; 204 | int wmode; 205 | float wref; 206 | VSNodeRef * ref_node; // rclip 207 | int device_id; 208 | 209 | ticket_semaphore semaphore; 210 | std::vector stream_data; 211 | std::vector ticket; 212 | std::mutex ticket_lock; 213 | }; 214 | 215 | static void VS_CC NLMeansInit( 216 | VSMap *in, VSMap *out, void **instanceData, VSNode *node, 217 | VSCore *core, const VSAPI *vsapi 218 | ) noexcept { 219 | 220 | auto d = static_cast(*instanceData); 221 | 222 | auto vi = vsapi->getVideoInfo(d->node); 223 | vsapi->setVideoInfo(vi, 1, node); 224 | } 225 | 226 | static const VSFrameRef *VS_CC NLMeansGetFrame( 227 | int n, int activationReason, void **instanceData, void **frameData, 228 | VSFrameContext *frameCtx, VSCore *core, const VSAPI *vsapi 229 | ) noexcept { 230 | 231 | auto d = static_cast(*instanceData); 232 | 233 | if (activationReason == arInitial) { 234 | int start = std::max(n - d->radius, 0); 235 | auto vi = vsapi->getVideoInfo(d->node); 236 | int end = std::min(n + d->radius, vi->numFrames - 1); 237 | for (int i = start; i <= end; i++) { 238 | vsapi->requestFrameFilter(i, d->node, frameCtx); 239 | if (d->ref_node) { 240 | vsapi->requestFrameFilter(i, d->ref_node, frameCtx); 241 | } 242 | } 243 | return nullptr; 244 | } else if (activationReason != arAllFramesReady) { 245 | return nullptr; 246 | } 247 | 248 | auto set_error = [vsapi, frameCtx](const char * error_message) -> std::nullptr_t { 249 | vsapi->setFilterError(error_message, frameCtx); 250 | return nullptr; 251 | }; 252 | 253 | checkError(cudaSetDevice(d->device_id)); 254 | 255 | auto vi = vsapi->getVideoInfo(d->node); 256 | 257 | std::vectorfreeFrame)>> src_frames; 258 | src_frames.reserve((d->ref_node ? 2 : 1) * (2 * d->radius + 1)); 259 | for (int i = n - d->radius; i <= n + d->radius; i++) { 260 | src_frames.emplace_back( 261 | vsapi->getFrameFilter(std::clamp(i, 0, vi->numFrames - 1), d->node, frameCtx), 262 | vsapi->freeFrame 263 | ); 264 | } 265 | if (d->ref_node) { 266 | for (int i = n - d->radius; i <= n + d->radius; i++) { 267 | src_frames.emplace_back( 268 | vsapi->getFrameFilter(std::clamp(i, 0, vi->numFrames - 1), d->ref_node, frameCtx), 269 | vsapi->freeFrame 270 | ); 271 | } 272 | } 273 | 274 | auto & src_center_frame = src_frames[d->radius]; 275 | auto format = vsapi->getFrameFormat(src_center_frame.get()); 276 | 277 | const VSFrameRef * fr[] { 278 | (d->channels == ChannelMode::UV) ? src_center_frame.get() : nullptr, 279 | (d->channels == ChannelMode::Y) ? src_center_frame.get() : nullptr, 280 | (d->channels == ChannelMode::Y) ? src_center_frame.get() : nullptr 281 | }; 282 | const int pl[] { 0, 1, 2 }; 283 | std::unique_ptrfreeFrame)> dst_frame { 284 | vsapi->newVideoFrame2(format, vi->width, vi->height, fr, pl, src_center_frame.get(), core), 285 | vsapi->freeFrame 286 | }; 287 | 288 | d->semaphore.acquire(); 289 | 290 | int ticket; 291 | { 292 | std::lock_guard lock { d->ticket_lock }; 293 | ticket = d->ticket.back(); 294 | d->ticket.pop_back(); 295 | } 296 | 297 | auto & stream_data = d->stream_data[ticket]; 298 | 299 | auto [filter_num_planes, filter_width, filter_height] = get_filter_shape(d->channels, vi); 300 | 301 | { 302 | auto h_buffer = stream_data.h_buffer.data; 303 | 304 | for (const auto & src_frame : src_frames) { 305 | for (int plane = 0; plane < 3; plane++) { 306 | if (fr[plane] != nullptr) { 307 | continue; 308 | } 309 | 310 | auto srcp = vsapi->getReadPtr(src_frame.get(), plane); 311 | int pitch = vsapi->getStride(src_frame.get(), plane); 312 | 313 | vs_bitblt( 314 | h_buffer, stream_data.image_stride * vi->format->bytesPerSample, 315 | srcp, pitch, 316 | filter_width * vi->format->bytesPerSample, filter_height 317 | ); 318 | 319 | h_buffer += filter_height * stream_data.image_stride * vi->format->bytesPerSample; 320 | } 321 | } 322 | } 323 | 324 | { 325 | checkError(cudaMemcpyAsync( 326 | stream_data.d_src_ref, 327 | stream_data.h_buffer, 328 | (d->ref_node ? 2 : 1) * (2 * d->radius + 1) * filter_num_planes * filter_height * stream_data.image_stride * vi->format->bytesPerSample, 329 | cudaMemcpyHostToDevice, 330 | stream_data.stream 331 | )); 332 | 333 | checkError(nlmeans( 334 | stream_data.d_dst, 335 | stream_data.d_src_ref, 336 | stream_data.d_buffer, 337 | stream_data.d_buffer_bwd, 338 | stream_data.d_buffer_fwd, 339 | stream_data.d_wdst, 340 | stream_data.d_weight, 341 | stream_data.d_max_weight, 342 | vi->format->sampleType == stFloat, 343 | vi->format->bitsPerSample, 344 | filter_width, 345 | filter_height, 346 | stream_data.image_stride, 347 | stream_data.buffer_stride, 348 | d->radius, 349 | d->spatial_radius, 350 | d->block_radius, 351 | d->h2_inv_norm, 352 | d->channels, 353 | d->wmode, 354 | d->wref, 355 | d->ref_node != nullptr, 356 | stream_data.stream 357 | )); 358 | 359 | checkError(cudaMemcpyAsync( 360 | stream_data.h_buffer, 361 | stream_data.d_dst, 362 | filter_num_planes * filter_height * stream_data.image_stride * vi->format->bytesPerSample, 363 | cudaMemcpyDeviceToHost, 364 | stream_data.stream 365 | )); 366 | 367 | checkError(cudaEventRecord(stream_data.event, stream_data.stream)); 368 | checkError(cudaEventSynchronize(stream_data.event)); 369 | } 370 | 371 | { 372 | auto h_buffer = stream_data.h_buffer.data; 373 | 374 | for (int plane = 0; plane < 3; plane++) { 375 | if (fr[plane] != nullptr) { 376 | continue; 377 | } 378 | 379 | auto dstp = vsapi->getWritePtr(dst_frame.get(), plane); 380 | int pitch = vsapi->getStride(dst_frame.get(), plane); 381 | 382 | vs_bitblt( 383 | dstp, pitch, 384 | h_buffer, stream_data.image_stride * vi->format->bytesPerSample, 385 | filter_width * vi->format->bytesPerSample, filter_height 386 | ); 387 | 388 | h_buffer += filter_height * stream_data.image_stride * vi->format->bytesPerSample; 389 | } 390 | } 391 | 392 | { 393 | std::lock_guard lock { d->ticket_lock }; 394 | d->ticket.emplace_back(ticket); 395 | } 396 | d->semaphore.release(); 397 | 398 | return dst_frame.release(); 399 | } 400 | 401 | static void VS_CC NLMeansFree( 402 | void *instanceData, VSCore *core, const VSAPI *vsapi 403 | ) noexcept { 404 | 405 | auto d = static_cast(instanceData); 406 | 407 | vsapi->freeNode(d->node); 408 | if (d->ref_node) { 409 | vsapi->freeNode(d->ref_node); 410 | } 411 | 412 | showError(cudaSetDevice(d->device_id)); 413 | 414 | delete d; 415 | } 416 | 417 | static void VS_CC NLMeansCreate( 418 | const VSMap *in, VSMap *out, void *userData, 419 | VSCore *core, const VSAPI *vsapi 420 | ) noexcept { 421 | 422 | auto d = std::make_unique(); 423 | 424 | d->node = vsapi->propGetNode(in, "clip", 0, nullptr); 425 | node_freer node_freer_node { vsapi, d->node }; 426 | 427 | auto set_error = [vsapi, out](const char * error_message) -> void { 428 | vsapi->setError(out, error_message); 429 | return ; 430 | }; 431 | 432 | auto vi = vsapi->getVideoInfo(d->node); 433 | 434 | if (!isConstantFormat(vi) || 435 | (vi->format->sampleType == stInteger && vi->format->bitsPerSample > 16) || 436 | (vi->format->sampleType == stFloat && vi->format->bitsPerSample != 32) 437 | ) { 438 | return set_error("only constant format 1-16 bit integer or 32-bit float input is supported"); 439 | } 440 | 441 | int error; 442 | 443 | d->radius = int64ToIntS(vsapi->propGetInt(in, "d", 0, &error)); 444 | if (error) { 445 | d->radius = 1; 446 | } 447 | if (d->radius < 0) { 448 | return set_error("\"d\" must be non-negative"); 449 | } 450 | 451 | d->spatial_radius = int64ToIntS(vsapi->propGetInt(in, "a", 0, &error)); 452 | if (error) { 453 | d->spatial_radius = 2; 454 | } 455 | if (d->spatial_radius <= 0) { 456 | return set_error("\"a\" must be positive"); 457 | } 458 | 459 | d->block_radius = int64ToIntS(vsapi->propGetInt(in, "s", 0, &error)); 460 | if (error) { 461 | d->block_radius = 4; 462 | } 463 | if (d->block_radius < 0) { 464 | return set_error("\"s\" must be non-negative"); 465 | } 466 | 467 | auto h = vsapi->propGetFloat(in, "h", 0, &error); 468 | if (error) { 469 | h = 1.2; 470 | } 471 | if (h <= 0.0) { 472 | return set_error("\"h\" must be positive"); 473 | } 474 | d->h2_inv_norm = static_cast(square(255) / (3 * square(2 * d->block_radius + 1) * square(h))); 475 | 476 | auto channels = vsapi->propGetData(in, "channels", 0, &error); 477 | if (error) { 478 | channels = "AUTO"; 479 | } 480 | auto channels_len = strlen(channels); 481 | if (channels_len == 1 && *channels == 'Y') { 482 | d->channels = ChannelMode::Y; 483 | } else if (channels_len == 2 && strncmp(channels, "UV", 2) == 0) { 484 | d->channels = ChannelMode::UV; 485 | } else if (channels_len == 3 && strncmp(channels, "YUV", 3) == 0) { 486 | d->channels = ChannelMode::YUV; 487 | } else if (channels_len == 3 && strncmp(channels, "RGB", 3) == 0) { 488 | d->channels = ChannelMode::RGB; 489 | } else if (channels_len == 4 && strncmp(channels, "AUTO", 4) == 0) { 490 | if (vi->format->colorFamily == cmRGB) { 491 | d->channels = ChannelMode::RGB; 492 | } else { 493 | d->channels = ChannelMode::Y; 494 | } 495 | } else { 496 | return set_error("\"channels\" must be \"Y\", \"UV\', \"YUV\", \"RGB\" or \"AUTO\""); 497 | } 498 | 499 | if (d->channels == ChannelMode::Y) { 500 | if (vi->format->colorFamily != cmGray && vi->format->colorFamily != cmYUV) { 501 | return set_error("color family must be Gray or YUV for \"channels\" == \"Y\""); 502 | } 503 | } else if (d->channels == ChannelMode::UV) { 504 | if (vi->format->colorFamily != cmYUV) { 505 | return set_error("color family must be YUV for \"channels\" == \"UV\""); 506 | } 507 | } else if (d->channels == ChannelMode::YUV) { 508 | if (vi->format->colorFamily != cmYUV || vi->format->subSamplingW || vi->format->subSamplingH) { 509 | return set_error("color family must be YUV444 for \"channels\" == \"YUV\""); 510 | } 511 | } else if (d->channels == ChannelMode::RGB) { 512 | if (vi->format->colorFamily != cmRGB) { 513 | return set_error("color family must be RGB for \"channels\" == \"RGB\""); 514 | } 515 | } 516 | 517 | d->wmode = static_cast(vsapi->propGetInt(in, "wmode", 0, &error)); 518 | if (error) { 519 | d->wmode = 0; 520 | } 521 | if (d->wmode < 0 || d->wmode > 3) { 522 | return set_error("\"wmode\" must be 0, 1, 2 or 3"); 523 | } 524 | 525 | d->wref = static_cast(vsapi->propGetFloat(in, "wref", 0, &error)); 526 | if (error) { 527 | d->wref = 1.0f; 528 | } 529 | 530 | d->ref_node = vsapi->propGetNode(in, "rclip", 0, &error); 531 | if (error) { 532 | d->ref_node = nullptr; 533 | } 534 | if (d->ref_node) { 535 | const auto ref_vi = vsapi->getVideoInfo(d->ref_node); 536 | if (!isSameFormat(vi, ref_vi) || vi->numFrames != ref_vi->numFrames) { 537 | vsapi->freeNode(d->ref_node); 538 | return set_error("\"rclip\" must be of the same format as \"clip\""); 539 | } 540 | } 541 | 542 | node_freer node_freer_ref { vsapi, d->ref_node }; 543 | 544 | d->device_id = int64ToIntS(vsapi->propGetInt(in, "device_id", 0, &error)); 545 | if (error) { 546 | d->device_id = 0; 547 | } 548 | 549 | int num_streams = int64ToIntS(vsapi->propGetInt(in, "num_streams", 0, &error)); 550 | if (error) { 551 | num_streams = 1; 552 | } 553 | if (num_streams <= 0) { 554 | return set_error("\"num_streams\" must be positive"); 555 | } 556 | d->semaphore.current.store(num_streams - 1, std::memory_order::relaxed); 557 | d->ticket.reserve(num_streams); 558 | for (int i = 0; i < num_streams; i++) { 559 | d->ticket.emplace_back(i); 560 | } 561 | 562 | checkError(cudaSetDevice(d->device_id)); 563 | 564 | auto [filter_num_planes, filter_width, filter_height] = get_filter_shape(d->channels, vi); 565 | 566 | d->stream_data.resize(num_streams); 567 | for (int i = 0; i < num_streams; i++) { 568 | auto & stream_data = d->stream_data[i]; 569 | 570 | checkError(cudaStreamCreateWithFlags(&stream_data.stream.data, cudaStreamNonBlocking)); 571 | 572 | checkError(cudaEventCreateWithFlags( 573 | &stream_data.event.data, 574 | cudaEventBlockingSync | cudaEventDisableTiming 575 | )); 576 | 577 | size_t image_pitch; 578 | checkError(cudaMallocPitch( 579 | &stream_data.d_src_ref.data, 580 | &image_pitch, 581 | filter_width * vi->format->bytesPerSample, 582 | (d->ref_node ? 2 : 1) * (2 * d->radius + 1) * filter_num_planes * filter_height 583 | )); 584 | stream_data.image_stride = static_cast(image_pitch / vi->format->bytesPerSample); 585 | 586 | checkError(cudaMalloc(&stream_data.d_dst.data, filter_num_planes * filter_height * image_pitch)); 587 | 588 | size_t buffer_pitch; 589 | checkError(cudaMallocPitch(&stream_data.d_max_weight.data, &buffer_pitch, filter_width * sizeof(float), filter_height)); 590 | stream_data.buffer_stride = static_cast(buffer_pitch / sizeof(float)); 591 | 592 | checkError(cudaMalloc(&stream_data.d_buffer.data, filter_height * buffer_pitch)); 593 | checkError(cudaMalloc(&stream_data.d_buffer_bwd.data, filter_height * buffer_pitch)); 594 | if (d->radius > 0) { 595 | checkError(cudaMalloc(&stream_data.d_buffer_fwd.data, filter_height * buffer_pitch)); 596 | } 597 | checkError(cudaMalloc(&stream_data.d_wdst.data, filter_num_planes * filter_height * buffer_pitch)); 598 | checkError(cudaMalloc(&stream_data.d_weight.data, filter_height * buffer_pitch)); 599 | 600 | auto h_buffer_size = std::max( 601 | (d->ref_node ? 2 : 1) * (2 * d->radius + 1) * filter_num_planes * filter_height * image_pitch, 602 | filter_num_planes * filter_height * buffer_pitch 603 | ); 604 | checkError(cudaMallocHost(&stream_data.h_buffer.data, h_buffer_size)); 605 | } 606 | 607 | vsapi->createFilter( 608 | in, out, "NLMeans", 609 | NLMeansInit, NLMeansGetFrame, NLMeansFree, 610 | fmParallel, 0, d.release(), core 611 | ); 612 | 613 | node_freer_node.release(); 614 | node_freer_ref.release(); 615 | } 616 | 617 | VS_EXTERNAL_API(void) 618 | VapourSynthPluginInit(VSConfigPlugin configFunc, VSRegisterFunction registerFunc, VSPlugin *plugin) { 619 | configFunc( 620 | "io.github.amusementclub.nlm_cuda", 621 | "nlm_cuda", 622 | "Non-local means denoise filter implemented in CUDA", 623 | VAPOURSYNTH_API_VERSION, 1, plugin 624 | ); 625 | 626 | registerFunc( 627 | "NLMeans", 628 | "clip:clip;" 629 | "d:int:opt;" 630 | "a:int:opt;" 631 | "s:int:opt;" 632 | "h:float:opt;" 633 | "channels:data:opt;" 634 | "wmode:int:opt;" 635 | "wref:float:opt;" 636 | "rclip:clip:opt;" 637 | "device_id:int:opt;" 638 | "num_streams:int:opt;", 639 | NLMeansCreate, nullptr, plugin 640 | ); 641 | 642 | auto getVersion = [](const VSMap *, VSMap * out, void *, VSCore *, const VSAPI *vsapi) { 643 | vsapi->propSetData(out, "version", VERSION, -1, paReplace); 644 | vsapi->propSetInt(out, "cuda_version", CUDART_VERSION, paReplace); 645 | }; 646 | registerFunc("Version", "", getVersion, nullptr, plugin); 647 | } 648 | -------------------------------------------------------------------------------- /source/kernel.cu: -------------------------------------------------------------------------------- 1 | // based on KNLMeansCL by Khanattila 2 | 3 | #include "common.h" 4 | 5 | #include 6 | #include 7 | #include 8 | 9 | template 10 | __host__ __device__ 11 | static inline T square(T x) { 12 | return x * x; 13 | } 14 | 15 | template 16 | __global__ 17 | __launch_bounds__(128) 18 | static void distance_horizontal( 19 | float * __restrict__ buffer, 20 | const T * __restrict__ src, 21 | const T * __restrict__ neighbor_src, 22 | int width, int height, int image_stride, int buffer_stride, 23 | int offset_x, int offset_y, 24 | int block_radius, ChannelMode channels, 25 | float inv_divisor 26 | ) { 27 | 28 | int y = blockIdx.y * blockDim.y + threadIdx.y; 29 | 30 | if (y >= height) { 31 | return ; 32 | } 33 | 34 | auto scale = [inv_divisor](float x) -> float { 35 | if constexpr (!std::is_floating_point_v) { 36 | x *= square(inv_divisor); 37 | } 38 | return x; 39 | }; 40 | 41 | extern __shared__ float local_buffer[]; // shape: (block_h, results_per_thread * block_w + 2 * nlm_s) 42 | // ignores bank conflicts here for higher occupancy 43 | // at most (32 / blockDim.x) smem requests generated per instruction 44 | int local_buffer_stride = results_per_thread * blockDim.x + 2 * block_radius; 45 | 46 | auto clamp_x = [width](int coord) { return min(max(coord, 0), width - 1); }; 47 | auto clamp_y = [height](int coord) { return min(max(coord, 0), height - 1); }; 48 | 49 | for (int i = threadIdx.x; i < results_per_thread * blockDim.x + 2 * block_radius; i += blockDim.x) { 50 | int x = -block_radius + blockIdx.x * results_per_thread * blockDim.x; 51 | 52 | switch (channels) { 53 | case ChannelMode::Y: { 54 | auto center = static_cast(src[y * image_stride + clamp_x(x + i)]); 55 | auto neighbor = static_cast(neighbor_src[clamp_y(y + offset_y) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 56 | auto diff = center - neighbor; 57 | local_buffer[threadIdx.y * local_buffer_stride + i] = scale(square(diff) * 3); 58 | break; 59 | } 60 | case ChannelMode::UV: { 61 | auto center_u = static_cast(src[y * image_stride + clamp_x(x + i)]); 62 | auto neighbor_u = static_cast(neighbor_src[(clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 63 | auto diff_u = center_u - neighbor_u; 64 | auto center_v = static_cast(src[(height + y) * image_stride + clamp_x(x + i)]); 65 | auto neighbor_v = static_cast(neighbor_src[(height + clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 66 | auto diff_v = center_v - neighbor_v; 67 | local_buffer[threadIdx.y * local_buffer_stride + i] = scale((square(diff_u) + square(diff_v)) * 1.5f); 68 | break; 69 | } 70 | case ChannelMode::YUV: { 71 | auto center_y = static_cast(src[y * image_stride + clamp_x(x + i)]); 72 | auto neighbor_y = static_cast(neighbor_src[(clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 73 | auto diff_y = center_y - neighbor_y; 74 | auto center_u = static_cast(src[(height + y) * image_stride + clamp_x(x + i)]); 75 | auto neighbor_u = static_cast(neighbor_src[(height + clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 76 | auto diff_u = center_u - neighbor_u; 77 | auto center_v = static_cast(src[(2 * height + y) * image_stride + clamp_x(x + i)]); 78 | auto neighbor_v = static_cast(neighbor_src[(2 * height + clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 79 | auto diff_v = center_v - neighbor_v; 80 | local_buffer[threadIdx.y * local_buffer_stride + i] = scale(square(diff_y) + square(diff_u) + square(diff_v)); 81 | break; 82 | } 83 | case ChannelMode::RGB: { 84 | auto center_r = static_cast(src[y * image_stride + clamp_x(x + i)]); 85 | auto neighbor_r = static_cast(neighbor_src[(clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 86 | auto diff_r = center_r - neighbor_r; 87 | auto center_g = static_cast(src[(height + y) * image_stride + clamp_x(x + i)]); 88 | auto neighbor_g = static_cast(neighbor_src[(height + clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 89 | auto diff_g = center_g - neighbor_g; 90 | auto center_b = static_cast(src[(2 * height + y) * image_stride + clamp_x(x + i)]); 91 | auto neighbor_b = static_cast(neighbor_src[(2 * height + clamp_y(y + offset_y)) * image_stride + clamp_x(clamp_x(x + i) + offset_x)]); 92 | auto diff_b = center_b - neighbor_b; 93 | auto weight = (center_r + neighbor_r) / 6 * inv_divisor; 94 | local_buffer[threadIdx.y * local_buffer_stride + i] = scale( 95 | (2.0f / 3.0f + weight) * square(diff_r) + 96 | (4.0f / 3.0f) * square(diff_g) + 97 | (1.0f - weight) * square(diff_b) 98 | ); 99 | break; 100 | } 101 | default: 102 | __builtin_unreachable(); 103 | } 104 | } 105 | 106 | __syncwarp(); 107 | 108 | int x = blockIdx.x * results_per_thread * blockDim.x + threadIdx.x; 109 | int local_x = threadIdx.x + block_radius; 110 | for (int loop = 0; loop < results_per_thread; loop++, x += blockDim.x, local_x += blockDim.x) { 111 | if (x >= width) { 112 | return ; 113 | } 114 | 115 | float sum = 0.0f; 116 | for (int i = -block_radius; i <= block_radius; i++) { 117 | sum += local_buffer[threadIdx.y * local_buffer_stride + (local_x + i)]; 118 | } 119 | 120 | buffer[y * buffer_stride + x] = sum; 121 | } 122 | } 123 | 124 | template 125 | static void distance_horizontal_dispatch( 126 | float * buffer, 127 | const void * src, int src_offset, int neighbor_src_offset, 128 | int width, int height, int image_stride, int buffer_stride, 129 | int offset_x, int offset_y, 130 | int block_radius, ChannelMode channels, 131 | dim3 grid, dim3 block, size_t dyn_smem_size, cudaStream_t stream, 132 | bool is_float, int bits_per_sample 133 | ) { 134 | if (is_float) { 135 | if (bits_per_sample == 32) { 136 | distance_horizontal<<>>( 137 | buffer, 138 | static_cast(src) + src_offset, 139 | static_cast(src) + neighbor_src_offset, 140 | width, height, image_stride, buffer_stride, 141 | offset_x, offset_y, 142 | block_radius, channels, 0.0f 143 | ); 144 | } 145 | } else { 146 | float inv_divisor = 1.0f / ((1 << bits_per_sample) - 1); 147 | 148 | if (bits_per_sample <= 8) { 149 | distance_horizontal<<>>( 150 | buffer, 151 | static_cast(src) + src_offset, 152 | static_cast(src) + neighbor_src_offset, 153 | width, height, image_stride, buffer_stride, 154 | offset_x, offset_y, 155 | block_radius, channels, inv_divisor 156 | ); 157 | } else if (bits_per_sample <= 16) { 158 | distance_horizontal<<>>( 159 | buffer, 160 | static_cast(src) + src_offset, 161 | static_cast(src) + neighbor_src_offset, 162 | width, height, image_stride, buffer_stride, 163 | offset_x, offset_y, 164 | block_radius, channels, inv_divisor 165 | ); 166 | } 167 | } 168 | } 169 | 170 | template 171 | __global__ 172 | __launch_bounds__(128) 173 | static void vertical( 174 | float * __restrict__ buffer_out, 175 | const float * __restrict__ buffer, 176 | int width, int height, int buffer_stride, 177 | int block_radius, 178 | float h2_inv_norm, 179 | int wmode 180 | ) { 181 | 182 | int x = blockIdx.x * blockDim.x + threadIdx.x; 183 | 184 | if (x >= width) { 185 | return ; 186 | } 187 | 188 | extern __shared__ float local_buffer[]; // shape: (results_per_thread * block_h + 2 * nlm_s, block_w) 189 | 190 | auto clamp_y = [height](int coord) { return min(max(coord, 0), height - 1); }; 191 | 192 | for (int i = threadIdx.y; i < results_per_thread * blockDim.y + 2 * block_radius; i += blockDim.y) { 193 | int y = -block_radius + blockIdx.y * results_per_thread * blockDim.y; 194 | local_buffer[i * blockDim.x + threadIdx.x] = buffer[clamp_y(y + i) * buffer_stride + x]; 195 | } 196 | 197 | __syncthreads(); 198 | 199 | int y = blockIdx.y * results_per_thread * blockDim.y + threadIdx.y; 200 | int local_y = threadIdx.y + block_radius; 201 | for (int loop = 0; loop < results_per_thread; loop++, y += blockDim.y, local_y += blockDim.y) { 202 | if (y >= height) { 203 | return ; 204 | } 205 | 206 | float sum = 0.0f; 207 | for (int i = -block_radius; i <= block_radius; i++) { 208 | sum += local_buffer[(local_y + i) * blockDim.x + threadIdx.x]; 209 | } 210 | 211 | float val; 212 | if (wmode == 0) { 213 | val = expf(-sum * h2_inv_norm); 214 | } else { 215 | val = fdimf(1.0f, sum * h2_inv_norm); 216 | if (wmode >= 2) { 217 | val *= val; 218 | } 219 | if (wmode == 3) { 220 | val *= val; 221 | val *= val; 222 | } 223 | } 224 | 225 | buffer_out[y * buffer_stride + x] = val; 226 | } 227 | } 228 | 229 | template 230 | __global__ 231 | __launch_bounds__(128) 232 | static void accumulation( 233 | float * __restrict__ wdst, 234 | float * __restrict__ weight, 235 | float * __restrict__ max_weight, 236 | const T * __restrict__ src, 237 | const float * __restrict__ buffer_bwd, 238 | const float * __restrict__ buffer_fwd, 239 | int width, int height, int image_stride, int buffer_stride, 240 | int offset_x, int offset_y, int offset_t, 241 | int num_planes 242 | ) { 243 | 244 | int x = blockIdx.x * blockDim.x + threadIdx.x; 245 | int y = blockIdx.y * blockDim.y + threadIdx.y; 246 | 247 | if (x >= width || y >= height) { 248 | return ; 249 | } 250 | 251 | auto clamp_x = [width](int coord) { return min(max(coord, 0), width - 1); }; 252 | auto clamp_y = [height](int coord) { return min(max(coord, 0), height - 1); }; 253 | 254 | auto center_weight = buffer_bwd[y * buffer_stride + x]; 255 | auto mirror_weight = buffer_fwd[clamp_y(y - offset_y) * buffer_stride + clamp_x(x - offset_x)]; 256 | max_weight[y * buffer_stride + x] = fmaxf(max_weight[y * buffer_stride + x], fmaxf(center_weight, mirror_weight)); 257 | weight[y * buffer_stride + x] += center_weight + mirror_weight; 258 | 259 | for (int plane = 0; plane < num_planes; plane++) { 260 | auto north_west = static_cast(src[((offset_t * num_planes + plane) * height + clamp_y(y + offset_y)) * image_stride + clamp_x(x + offset_x)]); 261 | auto south_east = static_cast(src[((-offset_t * num_planes + plane) * height + clamp_y(y - offset_y)) * image_stride + clamp_x(x - offset_x)]); 262 | wdst[(plane * height + y) * buffer_stride + x] += center_weight * north_west + mirror_weight * south_east; 263 | } 264 | } 265 | 266 | static void accumulation_dispatch( 267 | float * wdst, 268 | float * weight, 269 | float * max_weight, 270 | const void * src, int src_offset, 271 | const float * buffer_bwd, 272 | const float * buffer_fwd, 273 | int width, int height, int image_stride, int buffer_stride, 274 | int offset_x, int offset_y, int offset_t, 275 | int num_planes, 276 | dim3 grid, dim3 block, size_t dyn_smem_size, cudaStream_t stream, 277 | bool is_float, int bits_per_sample 278 | ) { 279 | if (is_float) { 280 | if (bits_per_sample == 32) { 281 | accumulation<<>>( 282 | wdst, 283 | weight, 284 | max_weight, 285 | static_cast(src) + src_offset, 286 | buffer_bwd, 287 | buffer_fwd, 288 | width, height, image_stride, buffer_stride, 289 | offset_x, offset_y, offset_t, 290 | num_planes 291 | ); 292 | } 293 | } else { 294 | if (bits_per_sample <= 8) { 295 | accumulation<<>>( 296 | wdst, 297 | weight, 298 | max_weight, 299 | static_cast(src) + src_offset, 300 | buffer_bwd, 301 | buffer_fwd, 302 | width, height, image_stride, buffer_stride, 303 | offset_x, offset_y, offset_t, 304 | num_planes 305 | ); 306 | } else if (bits_per_sample <= 16) { 307 | accumulation<<>>( 308 | wdst, 309 | weight, 310 | max_weight, 311 | static_cast(src) + src_offset, 312 | buffer_bwd, 313 | buffer_fwd, 314 | width, height, image_stride, buffer_stride, 315 | offset_x, offset_y, offset_t, 316 | num_planes 317 | ); 318 | } 319 | } 320 | } 321 | 322 | template 323 | __global__ 324 | __launch_bounds__(128) 325 | static void finish( 326 | T * __restrict__ dst, 327 | const T * __restrict__ src, 328 | const float * __restrict__ wdst, 329 | const float * __restrict__ weight, 330 | const float * __restrict__ max_weight, 331 | int width, int height, int image_stride, int buffer_stride, 332 | int num_planes, float wref, int peak 333 | ) { 334 | 335 | int x = blockIdx.x * blockDim.x + threadIdx.x; 336 | int y = blockIdx.y * blockDim.y + threadIdx.y; 337 | 338 | if (x >= width || y >= height) { 339 | return ; 340 | } 341 | 342 | auto from_float = [peak](float x) -> T { 343 | if constexpr (std::is_floating_point_v) { 344 | return x; 345 | } else { 346 | return static_cast(max(0, min(__float2int_rn(x), peak))); 347 | } 348 | }; 349 | 350 | // max_weight is initialized to 0 351 | auto center_weight = wref * fmaxf(FLT_EPSILON, max_weight[y * buffer_stride + x]); 352 | auto denominator = center_weight + weight[y * buffer_stride + x]; 353 | 354 | for (int plane = 0; plane < num_planes; plane++) { 355 | auto numerator = center_weight * static_cast(src[(plane * height + y) * image_stride + x]) + wdst[(plane * height + y) * buffer_stride + x]; 356 | dst[(plane * height + y) * image_stride + x] = from_float(numerator / denominator); 357 | } 358 | } 359 | 360 | static void finish_dispatch( 361 | void * dst, 362 | const void * src, int src_offset, 363 | const float * wdst, 364 | const float * weight, 365 | const float * max_weight, 366 | int width, int height, int image_stride, int buffer_stride, 367 | int num_planes, float wref, 368 | dim3 grid, dim3 block, size_t dyn_smem_size, cudaStream_t stream, 369 | bool is_float, int bits_per_sample 370 | ) { 371 | if (is_float) { 372 | if (bits_per_sample == 32) { 373 | finish<<>>( 374 | static_cast(dst), 375 | static_cast(src) + src_offset, 376 | wdst, 377 | weight, 378 | max_weight, 379 | width, height, image_stride, buffer_stride, 380 | num_planes, wref, 0 381 | ); 382 | } 383 | } else { 384 | int peak = (1 << bits_per_sample) - 1; 385 | 386 | if (bits_per_sample <= 8) { 387 | finish<<>>( 388 | static_cast(dst), 389 | static_cast(src) + src_offset, 390 | wdst, 391 | weight, 392 | max_weight, 393 | width, height, image_stride, buffer_stride, 394 | num_planes, wref, peak 395 | ); 396 | } else if (bits_per_sample <= 16) { 397 | finish<<>>( 398 | static_cast(dst), 399 | static_cast(src) + src_offset, 400 | wdst, 401 | weight, 402 | max_weight, 403 | width, height, image_stride, buffer_stride, 404 | num_planes, wref, peak 405 | ); 406 | } 407 | } 408 | } 409 | 410 | cudaError_t nlmeans( 411 | void * d_dst, 412 | void * d_src, 413 | float * d_buffer, 414 | float * d_buffer_bwd, 415 | float * d_buffer_fwd, 416 | float * d_wdst, 417 | float * d_weight, 418 | float * d_max_weight, 419 | bool is_float, 420 | int bits_per_sample, 421 | int width, int height, int image_stride, int buffer_stride, 422 | int radius, int spatial_radius, int block_radius, float h2_inv_norm, 423 | ChannelMode channels, int wmode, float wref, bool has_ref, 424 | cudaStream_t stream 425 | ) { 426 | 427 | int num_planes {}; 428 | if (channels == ChannelMode::Y) { 429 | num_planes = 1; 430 | } else if (channels == ChannelMode::UV) { 431 | num_planes = 2; 432 | } else if (channels == ChannelMode::YUV || channels == ChannelMode::RGB) { 433 | num_planes = 3; 434 | } 435 | 436 | if (auto error = cudaMemsetAsync(d_wdst, 0, num_planes * height * buffer_stride * sizeof(float), stream); error != cudaSuccess) { 437 | return error; 438 | } 439 | if (auto error = cudaMemsetAsync(d_weight, 0, height * buffer_stride * sizeof(float), stream); error != cudaSuccess) { 440 | return error; 441 | } 442 | if (auto error = cudaMemsetAsync(d_max_weight, 0, height * buffer_stride * sizeof(float), stream); error != cudaSuccess) { 443 | return error; 444 | } 445 | 446 | dim3 block { 16, 8, 1 }; 447 | // assert(block.x * block.y == 128); 448 | 449 | dim3 grid { (width + block.x - 1) / block.x, (height + block.y - 1) / block.y, 1 }; 450 | 451 | constexpr int hrz_result = 3; 452 | dim3 hrz_grid { (width + hrz_result * block.x - 1) / (hrz_result * block.x), (height + block.y - 1) / block.y, 1 }; 453 | int hrz_smem_stride = hrz_result * block.x + 2 * block_radius; 454 | auto hrz_smem = block.y * hrz_smem_stride * sizeof(float); 455 | 456 | constexpr int vrt_result = 3; 457 | dim3 vrt_grid { (width + block.x - 1) / block.x, (height + vrt_result * block.y - 1) / (vrt_result * block.y), 1 }; 458 | auto vrt_smem = (vrt_result * block.y + 2 * block_radius) * block.x * sizeof(float); 459 | 460 | for (int offset_t = -radius; offset_t <= 0; offset_t++) { 461 | for (int offset_y = -spatial_radius; offset_y <= spatial_radius; offset_y++) { 462 | for (int offset_x = -spatial_radius; offset_x <= spatial_radius; offset_x++) { 463 | if ((offset_t * (2 * spatial_radius + 1) + offset_y) * (2 * spatial_radius + 1) + offset_x >= 0) { 464 | continue; 465 | } 466 | 467 | distance_horizontal_dispatch( 468 | d_buffer, 469 | d_src, 470 | (has_ref * (2 * radius + 1) + radius) * num_planes * height * image_stride, 471 | ((has_ref * (2 * radius + 1) + radius) + offset_t) * num_planes * height * image_stride, 472 | width, height, image_stride, buffer_stride, 473 | offset_x, offset_y, 474 | block_radius, channels, 475 | hrz_grid, block, hrz_smem, stream, 476 | is_float, bits_per_sample 477 | ); 478 | 479 | vertical<<>>( 480 | d_buffer_bwd, 481 | d_buffer, 482 | width, height, buffer_stride, 483 | block_radius, h2_inv_norm, wmode 484 | ); 485 | 486 | if (offset_t == 0) { 487 | accumulation_dispatch( 488 | d_wdst, d_weight, d_max_weight, 489 | d_src, radius * num_planes * height * image_stride, 490 | d_buffer_bwd, d_buffer_bwd, 491 | width, height, image_stride, buffer_stride, 492 | offset_x, offset_y, offset_t, 493 | num_planes, 494 | grid, block, 0, stream, 495 | is_float, bits_per_sample 496 | ); 497 | continue; 498 | } 499 | 500 | distance_horizontal_dispatch( 501 | d_buffer, 502 | d_src, 503 | ((has_ref * (2 * radius + 1) + radius) - offset_t) * num_planes * height * image_stride, 504 | (has_ref * (2 * radius + 1) + radius) * num_planes * height * image_stride, 505 | width, height, image_stride, buffer_stride, 506 | offset_x, offset_y, 507 | block_radius, channels, 508 | hrz_grid, block, hrz_smem, stream, 509 | is_float, bits_per_sample 510 | ); 511 | 512 | vertical<<>>( 513 | d_buffer_fwd, 514 | d_buffer, 515 | width, height, buffer_stride, 516 | block_radius, h2_inv_norm, wmode 517 | ); 518 | 519 | accumulation_dispatch( 520 | d_wdst, d_weight, d_max_weight, 521 | d_src, radius * num_planes * height * image_stride, 522 | d_buffer_bwd, d_buffer_fwd, 523 | width, height, image_stride, buffer_stride, 524 | offset_x, offset_y, offset_t, 525 | num_planes, 526 | grid, block, 0, stream, 527 | is_float, bits_per_sample 528 | ); 529 | } 530 | } 531 | } 532 | 533 | finish_dispatch( 534 | d_dst, 535 | d_src, radius * num_planes * height * image_stride, 536 | d_wdst, d_weight, d_max_weight, 537 | width, height, image_stride, buffer_stride, 538 | num_planes, wref, 539 | grid, block, 0, stream, 540 | is_float, bits_per_sample 541 | ); 542 | 543 | return cudaSuccess; 544 | } 545 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | --------------------------------------------------------------------------------