torchcodec 0.7.0__cp39-cp39-win_amd64.whl

torchcodec/_core/CpuDeviceInterface.h

@@ -0,0 +1,70 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#pragma once
+
+#include "src/torchcodec/_core/DeviceInterface.h"
+#include "src/torchcodec/_core/FFMPEGCommon.h"
+#include "src/torchcodec/_core/FilterGraph.h"
+
+namespace facebook::torchcodec {
+
+class CpuDeviceInterface : public DeviceInterface {
+ public:
+  CpuDeviceInterface(const torch::Device& device);
+
+  virtual ~CpuDeviceInterface() {}
+
+  std::optional<const AVCodec*> findCodec(
+      [[maybe_unused]] const AVCodecID& codecId) override {
+    return std::nullopt;
+  }
+
+  void initializeContext(
+      [[maybe_unused]] AVCodecContext* codecContext) override {}
+
+  void convertAVFrameToFrameOutput(
+      const VideoStreamOptions& videoStreamOptions,
+      const AVRational& timeBase,
+      UniqueAVFrame& avFrame,
+      FrameOutput& frameOutput,
+      std::optional<torch::Tensor> preAllocatedOutputTensor =
+          std::nullopt) override;
+
+ private:
+  int convertAVFrameToTensorUsingSwsScale(
+      const UniqueAVFrame& avFrame,
+      torch::Tensor& outputTensor);
+
+  torch::Tensor convertAVFrameToTensorUsingFilterGraph(
+      const UniqueAVFrame& avFrame);
+
+  struct SwsFrameContext {
+    int inputWidth;
+    int inputHeight;
+    AVPixelFormat inputFormat;
+    int outputWidth;
+    int outputHeight;
+    bool operator==(const SwsFrameContext&) const;
+    bool operator!=(const SwsFrameContext&) const;
+  };
+
+  void createSwsContext(
+      const SwsFrameContext& swsFrameContext,
+      const enum AVColorSpace colorspace);
+
+  // color-conversion fields. Only one of FilterGraphContext and
+  // UniqueSwsContext should be non-null.
+  std::unique_ptr<FilterGraph> filterGraphContext_;
+  UniqueSwsContext swsContext_;
+
+  // Used to know whether a new FilterGraphContext or UniqueSwsContext should
+  // be created before decoding a new frame.
+  SwsFrameContext prevSwsFrameContext_;
+  FiltersContext prevFiltersContext_;
+};
+
+} // namespace facebook::torchcodec

torchcodec/_core/CudaDeviceInterface.cpp

@@ -0,0 +1,514 @@
+#include <ATen/cuda/CUDAEvent.h>
+#include <c10/cuda/CUDAStream.h>
+#include <npp.h>
+#include <torch/types.h>
+#include <mutex>
+
+#include "src/torchcodec/_core/Cache.h"
+#include "src/torchcodec/_core/CudaDeviceInterface.h"
+#include "src/torchcodec/_core/FFMPEGCommon.h"
+
+extern "C" {
+#include <libavutil/hwcontext_cuda.h>
+#include <libavutil/pixdesc.h>
+}
+
+namespace facebook::torchcodec {
+namespace {
+
+static bool g_cuda =
+    registerDeviceInterface(torch::kCUDA, [](const torch::Device& device) {
+      return new CudaDeviceInterface(device);
+    });
+
+// BT.709 full range color conversion matrix for YUV to RGB conversion.
+// See Note [YUV -> RGB Color Conversion, color space and color range] below.
+constexpr Npp32f bt709FullRangeColorTwist[3][4] = {
+    {1.0f, 0.0f, 1.5748f, 0.0f},
+    {1.0f, -0.187324273f, -0.468124273f, -128.0f},
+    {1.0f, 1.8556f, 0.0f, -128.0f}};
+
+// We reuse cuda contexts across VideoDeoder instances. This is because
+// creating a cuda context is expensive. The cache mechanism is as follows:
+// 1. There is a cache of size MAX_CONTEXTS_PER_GPU_IN_CACHE cuda contexts for
+//    each GPU.
+// 2. When we destroy a SingleStreamDecoder instance we release the cuda context
+// to
+//    the cache if the cache is not full.
+// 3. When we create a SingleStreamDecoder instance we try to get a cuda context
+// from
+//    the cache. If the cache is empty we create a new cuda context.
+
+// Pytorch can only handle up to 128 GPUs.
+// https://github.com/pytorch/pytorch/blob/e30c55ee527b40d67555464b9e402b4b7ce03737/c10/cuda/CUDAMacros.h#L44
+const int MAX_CUDA_GPUS = 128;
+// Set to -1 to have an infinitely sized cache. Set it to 0 to disable caching.
+// Set to a positive number to have a cache of that size.
+const int MAX_CONTEXTS_PER_GPU_IN_CACHE = -1;
+PerGpuCache<AVBufferRef, Deleterp<AVBufferRef, void, av_buffer_unref>>
+    g_cached_hw_device_ctxs(MAX_CUDA_GPUS, MAX_CONTEXTS_PER_GPU_IN_CACHE);
+PerGpuCache<NppStreamContext> g_cached_npp_ctxs(
+    MAX_CUDA_GPUS,
+    MAX_CONTEXTS_PER_GPU_IN_CACHE);
+
+#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 26, 100)
+
+AVBufferRef* getFFMPEGContextFromExistingCudaContext(
+    const torch::Device& device,
+    torch::DeviceIndex nonNegativeDeviceIndex,
+    enum AVHWDeviceType type) {
+  c10::cuda::CUDAGuard deviceGuard(device);
+  // Valid values for the argument to cudaSetDevice are 0 to maxDevices - 1:
+  // https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__DEVICE.html#group__CUDART__DEVICE_1g159587909ffa0791bbe4b40187a4c6bb
+  // So we ensure the deviceIndex is not negative.
+  // We set the device because we may be called from a different thread than
+  // the one that initialized the cuda context.
+  cudaSetDevice(nonNegativeDeviceIndex);
+  AVBufferRef* hw_device_ctx = nullptr;
+  std::string deviceOrdinal = std::to_string(nonNegativeDeviceIndex);
+  int err = av_hwdevice_ctx_create(
+      &hw_device_ctx,
+      type,
+      deviceOrdinal.c_str(),
+      nullptr,
+      AV_CUDA_USE_CURRENT_CONTEXT);
+  if (err < 0) {
+    /* clang-format off */
+    TORCH_CHECK(
+        false,
+        "Failed to create specified HW device. This typically happens when ",
+        "your installed FFmpeg doesn't support CUDA (see ",
+        "https://github.com/pytorch/torchcodec#installing-cuda-enabled-torchcodec",
+        "). FFmpeg error: ", getFFMPEGErrorStringFromErrorCode(err));
+    /* clang-format on */
+  }
+  return hw_device_ctx;
+}
+
+#else
+
+AVBufferRef* getFFMPEGContextFromNewCudaContext(
+    [[maybe_unused]] const torch::Device& device,
+    torch::DeviceIndex nonNegativeDeviceIndex,
+    enum AVHWDeviceType type) {
+  AVBufferRef* hw_device_ctx = nullptr;
+  std::string deviceOrdinal = std::to_string(nonNegativeDeviceIndex);
+  int err = av_hwdevice_ctx_create(
+      &hw_device_ctx, type, deviceOrdinal.c_str(), nullptr, 0);
+  if (err < 0) {
+    TORCH_CHECK(
+        false,
+        "Failed to create specified HW device",
+        getFFMPEGErrorStringFromErrorCode(err));
+  }
+  return hw_device_ctx;
+}
+
+#endif
+
+UniqueAVBufferRef getCudaContext(const torch::Device& device) {
+  enum AVHWDeviceType type = av_hwdevice_find_type_by_name("cuda");
+  TORCH_CHECK(type != AV_HWDEVICE_TYPE_NONE, "Failed to find cuda device");
+  torch::DeviceIndex nonNegativeDeviceIndex = getNonNegativeDeviceIndex(device);
+
+  UniqueAVBufferRef hw_device_ctx = g_cached_hw_device_ctxs.get(device);
+  if (hw_device_ctx) {
+    return hw_device_ctx;
+  }
+
+  // 58.26.100 introduced the concept of reusing the existing cuda context
+  // which is much faster and lower memory than creating a new cuda context.
+  // So we try to use that if it is available.
+  // FFMPEG 6.1.2 appears to be the earliest release that contains version
+  // 58.26.100 of avutil.
+  // https://github.com/FFmpeg/FFmpeg/blob/4acb9b7d1046944345ae506165fb55883d04d8a6/doc/APIchanges#L265
+#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 26, 100)
+  return UniqueAVBufferRef(getFFMPEGContextFromExistingCudaContext(
+      device, nonNegativeDeviceIndex, type));
+#else
+  return UniqueAVBufferRef(
+      getFFMPEGContextFromNewCudaContext(device, nonNegativeDeviceIndex, type));
+#endif
+}
+
+std::unique_ptr<NppStreamContext> getNppStreamContext(
+    const torch::Device& device) {
+  torch::DeviceIndex nonNegativeDeviceIndex = getNonNegativeDeviceIndex(device);
+
+  std::unique_ptr<NppStreamContext> nppCtx = g_cached_npp_ctxs.get(device);
+  if (nppCtx) {
+    return nppCtx;
+  }
+
+  // From 12.9, NPP recommends using a user-created NppStreamContext and using
+  // the `_Ctx()` calls:
+  // https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#npp-release-12-9-update-1
+  // And the nppGetStreamContext() helper is deprecated. We are explicitly
+  // supposed to create the NppStreamContext manually from the CUDA device
+  // properties:
+  // https://github.com/NVIDIA/CUDALibrarySamples/blob/d97803a40fab83c058bb3d68b6c38bd6eebfff43/NPP/README.md?plain=1#L54-L72
+
+  nppCtx = std::make_unique<NppStreamContext>();
+  cudaDeviceProp prop{};
+  cudaError_t err = cudaGetDeviceProperties(&prop, nonNegativeDeviceIndex);
+  TORCH_CHECK(
+      err == cudaSuccess,
+      "cudaGetDeviceProperties failed: ",
+      cudaGetErrorString(err));
+
+  nppCtx->nCudaDeviceId = nonNegativeDeviceIndex;
+  nppCtx->nMultiProcessorCount = prop.multiProcessorCount;
+  nppCtx->nMaxThreadsPerMultiProcessor = prop.maxThreadsPerMultiProcessor;
+  nppCtx->nMaxThreadsPerBlock = prop.maxThreadsPerBlock;
+  nppCtx->nSharedMemPerBlock = prop.sharedMemPerBlock;
+  nppCtx->nCudaDevAttrComputeCapabilityMajor = prop.major;
+  nppCtx->nCudaDevAttrComputeCapabilityMinor = prop.minor;
+
+  return nppCtx;
+}
+
+} // namespace
+
+CudaDeviceInterface::CudaDeviceInterface(const torch::Device& device)
+    : DeviceInterface(device) {
+  TORCH_CHECK(g_cuda, "CudaDeviceInterface was not registered!");
+  TORCH_CHECK(
+      device_.type() == torch::kCUDA, "Unsupported device: ", device_.str());
+}
+
+CudaDeviceInterface::~CudaDeviceInterface() {
+  if (ctx_) {
+    g_cached_hw_device_ctxs.addIfCacheHasCapacity(device_, std::move(ctx_));
+  }
+  if (nppCtx_) {
+    g_cached_npp_ctxs.addIfCacheHasCapacity(device_, std::move(nppCtx_));
+  }
+}
+
+void CudaDeviceInterface::initializeContext(AVCodecContext* codecContext) {
+  TORCH_CHECK(!ctx_, "FFmpeg HW device context already initialized");
+
+  // It is important for pytorch itself to create the cuda context. If ffmpeg
+  // creates the context it may not be compatible with pytorch.
+  // This is a dummy tensor to initialize the cuda context.
+  torch::Tensor dummyTensorForCudaInitialization = torch::empty(
+      {1}, torch::TensorOptions().dtype(torch::kUInt8).device(device_));
+  ctx_ = getCudaContext(device_);
+  nppCtx_ = getNppStreamContext(device_);
+  codecContext->hw_device_ctx = av_buffer_ref(ctx_.get());
+  return;
+}
+
+void CudaDeviceInterface::convertAVFrameToFrameOutput(
+    const VideoStreamOptions& videoStreamOptions,
+    [[maybe_unused]] const AVRational& timeBase,
+    UniqueAVFrame& avFrame,
+    FrameOutput& frameOutput,
+    std::optional<torch::Tensor> preAllocatedOutputTensor) {
+  if (avFrame->format != AV_PIX_FMT_CUDA) {
+    // The frame's format is AV_PIX_FMT_CUDA if and only if its content is on
+    // the GPU. In this branch, the frame is on the CPU: this is what NVDEC
+    // gives us if it wasn't able to decode a frame, for whatever reason.
+    // Typically that happens if the video's encoder isn't supported by NVDEC.
+    // Below, we choose to convert the frame's color-space using the CPU
+    // codepath, and send it back to the GPU at the very end.
+    // TODO: A possibly better solution would be to send the frame to the GPU
+    // first, and do the color conversion there.
+    auto cpuDevice = torch::Device(torch::kCPU);
+    auto cpuInterface = createDeviceInterface(cpuDevice);
+
+    FrameOutput cpuFrameOutput;
+    cpuInterface->convertAVFrameToFrameOutput(
+        videoStreamOptions,
+        timeBase,
+        avFrame,
+        cpuFrameOutput,
+        preAllocatedOutputTensor);
+
+    frameOutput.data = cpuFrameOutput.data.to(device_);
+    return;
+  }
+
+  // Above we checked that the AVFrame was on GPU, but that's not enough, we
+  // also need to check that the AVFrame is in AV_PIX_FMT_NV12 format (8 bits),
+  // because this is what the NPP color conversion routines expect.
+  // TODO: we should investigate how to can perform color conversion for
+  // non-8bit videos. This is supported on CPU.
+  TORCH_CHECK(
+      avFrame->hw_frames_ctx != nullptr,
+      "The AVFrame does not have a hw_frames_ctx. "
+      "That's unexpected, please report this to the TorchCodec repo.");
+
+  auto hwFramesCtx =
+      reinterpret_cast<AVHWFramesContext*>(avFrame->hw_frames_ctx->data);
+  AVPixelFormat actualFormat = hwFramesCtx->sw_format;
+  TORCH_CHECK(
+      actualFormat == AV_PIX_FMT_NV12,
+      "The AVFrame is ",
+      (av_get_pix_fmt_name(actualFormat) ? av_get_pix_fmt_name(actualFormat)
+                                         : "unknown"),
+      ", but we expected AV_PIX_FMT_NV12. This typically happens when "
+      "the video isn't 8bit, which is not supported on CUDA at the moment. "
+      "Try using the CPU device instead. "
+      "If the video is 10bit, we are tracking 10bit support in "
+      "https://github.com/pytorch/torchcodec/issues/776");
+
+  auto frameDims =
+      getHeightAndWidthFromOptionsOrAVFrame(videoStreamOptions, avFrame);
+  int height = frameDims.height;
+  int width = frameDims.width;
+  torch::Tensor& dst = frameOutput.data;
+  if (preAllocatedOutputTensor.has_value()) {
+    dst = preAllocatedOutputTensor.value();
+    auto shape = dst.sizes();
+    TORCH_CHECK(
+        (shape.size() == 3) && (shape[0] == height) && (shape[1] == width) &&
+            (shape[2] == 3),
+        "Expected tensor of shape ",
+        height,
+        "x",
+        width,
+        "x3, got ",
+        shape);
+  } else {
+    dst = allocateEmptyHWCTensor(height, width, device_);
+  }
+
+  torch::DeviceIndex deviceIndex = getNonNegativeDeviceIndex(device_);
+
+  // Create a CUDA event and attach it to the AVFrame's CUDA stream. That's the
+  // NVDEC stream, i.e. the CUDA stream that the frame was decoded on.
+  // We will be waiting for this event to complete before calling the NPP
+  // functions, to ensure NVDEC has finished decoding the frame before running
+  // the NPP color-conversion.
+  // Note that our code is generic and assumes that the NVDEC's stream can be
+  // arbitrary, but unfortunately we know it's hardcoded to be the default
+  // stream by FFmpeg:
+  // https://github.com/FFmpeg/FFmpeg/blob/66e40840d15b514f275ce3ce2a4bf72ec68c7311/libavutil/hwcontext_cuda.c#L387-L388
+  TORCH_CHECK(
+      hwFramesCtx->device_ctx != nullptr,
+      "The AVFrame's hw_frames_ctx does not have a device_ctx. ");
+  auto cudaDeviceCtx =
+      static_cast<AVCUDADeviceContext*>(hwFramesCtx->device_ctx->hwctx);
+  at::cuda::CUDAEvent nvdecDoneEvent;
+  at::cuda::CUDAStream nvdecStream = // That's always the default stream. Sad.
+      c10::cuda::getStreamFromExternal(cudaDeviceCtx->stream, deviceIndex);
+  nvdecDoneEvent.record(nvdecStream);
+
+  // Don't start NPP work before NVDEC is done decoding the frame!
+  at::cuda::CUDAStream nppStream = at::cuda::getCurrentCUDAStream(deviceIndex);
+  nvdecDoneEvent.block(nppStream);
+
+  // Create the NPP context if we haven't yet.
+  nppCtx_->hStream = nppStream.stream();
+  cudaError_t err =
+      cudaStreamGetFlags(nppCtx_->hStream, &nppCtx_->nStreamFlags);
+  TORCH_CHECK(
+      err == cudaSuccess,
+      "cudaStreamGetFlags failed: ",
+      cudaGetErrorString(err));
+
+  NppiSize oSizeROI = {width, height};
+  Npp8u* yuvData[2] = {avFrame->data[0], avFrame->data[1]};
+
+  NppStatus status;
+
+  // For background, see
+  // Note [YUV -> RGB Color Conversion, color space and color range]
+  if (avFrame->colorspace == AVColorSpace::AVCOL_SPC_BT709) {
+    if (avFrame->color_range == AVColorRange::AVCOL_RANGE_JPEG) {
+      // NPP provides a pre-defined color conversion function for BT.709 full
+      // range: nppiNV12ToRGB_709HDTV_8u_P2C3R_Ctx. But it's not closely
+      // matching the results we have on CPU. So we're using a custom color
+      // conversion matrix, which provides more accurate results. See the note
+      // mentioned above for details, and headaches.
+
+      int srcStep[2] = {avFrame->linesize[0], avFrame->linesize[1]};
+
+      status = nppiNV12ToRGB_8u_ColorTwist32f_P2C3R_Ctx(
+          yuvData,
+          srcStep,
+          static_cast<Npp8u*>(dst.data_ptr()),
+          dst.stride(0),
+          oSizeROI,
+          bt709FullRangeColorTwist,
+          *nppCtx_);
+    } else {
+      // If not full range, we assume studio limited range.
+      // The color conversion matrix for BT.709 limited range should be:
+      // static const Npp32f bt709LimitedRangeColorTwist[3][4] = {
+      //   {1.16438356f, 0.0f, 1.79274107f, -16.0f},
+      //   {1.16438356f, -0.213248614f, -0.5329093290f, -128.0f},
+      //   {1.16438356f, 2.11240179f, 0.0f, -128.0f}
+      // };
+      // We get very close results to CPU with that, but using the pre-defined
+      // nppiNV12ToRGB_709CSC_8u_P2C3R_Ctx seems to be even more accurate.
+      status = nppiNV12ToRGB_709CSC_8u_P2C3R_Ctx(
+          yuvData,
+          avFrame->linesize[0],
+          static_cast<Npp8u*>(dst.data_ptr()),
+          dst.stride(0),
+          oSizeROI,
+          *nppCtx_);
+    }
+  } else {
+    // TODO we're assuming BT.601 color space (and probably limited range) by
+    // calling nppiNV12ToRGB_8u_P2C3R_Ctx. We should handle BT.601 full range,
+    // and other color-spaces like 2020.
+    status = nppiNV12ToRGB_8u_P2C3R_Ctx(
+        yuvData,
+        avFrame->linesize[0],
+        static_cast<Npp8u*>(dst.data_ptr()),
+        dst.stride(0),
+        oSizeROI,
+        *nppCtx_);
+  }
+  TORCH_CHECK(status == NPP_SUCCESS, "Failed to convert NV12 frame.");
+}
+
+// inspired by https://github.com/FFmpeg/FFmpeg/commit/ad67ea9
+// we have to do this because of an FFmpeg bug where hardware decoding is not
+// appropriately set, so we just go off and find the matching codec for the CUDA
+// device
+std::optional<const AVCodec*> CudaDeviceInterface::findCodec(
+    const AVCodecID& codecId) {
+  void* i = nullptr;
+  const AVCodec* codec = nullptr;
+  while ((codec = av_codec_iterate(&i)) != nullptr) {
+    if (codec->id != codecId || !av_codec_is_decoder(codec)) {
+      continue;
+    }
+
+    const AVCodecHWConfig* config = nullptr;
+    for (int j = 0; (config = avcodec_get_hw_config(codec, j)) != nullptr;
+         ++j) {
+      if (config->device_type == AV_HWDEVICE_TYPE_CUDA) {
+        return codec;
+      }
+    }
+  }
+
+  return std::nullopt;
+}
+
+} // namespace facebook::torchcodec
+
+/* clang-format off */
+// Note: [YUV -> RGB Color Conversion, color space and color range]
+//
+// The frames we get from the decoder (FFmpeg decoder, or NVCUVID) are in YUV
+// format. We need to convert them to RGB. This note attempts to describe this
+// process. There may be some inaccuracies and approximations that experts will
+// notice, but our goal is only to provide a good enough understanding of the
+// process for torchcodec developers to implement and maintain it.
+// On CPU, filtergraph and swscale handle everything for us. With CUDA, we have
+// to do a lot of the heavy lifting ourselves.
+//
+// Color space and color range
+// ---------------------------
+// Two main characteristics of a frame will affect the conversion process:
+// 1. Color space: This basically defines what YUV values correspond to which
+//    physical wavelength. No need to go into details here,the point is that
+//    videos can come in different color spaces, the most common ones being
+//    BT.601 and BT.709, but there are others.
+//    In FFmpeg this is represented with AVColorSpace:
+//    https://ffmpeg.org/doxygen/4.0/pixfmt_8h.html#aff71a069509a1ad3ff54d53a1c894c85
+// 2. Color range: This defines the range of YUV values. There is:
+//    - full range, also called PC range: AVCOL_RANGE_JPEG
+//    - and the "limited" range, also called studio or TV range: AVCOL_RANGE_MPEG
+//    https://ffmpeg.org/doxygen/4.0/pixfmt_8h.html#a3da0bf691418bc22c4bcbe6583ad589a
+//
+// Color space and color range are independent concepts, so we can have a BT.709
+// with full range, and another one with limited range. Same for BT.601.
+//
+// In the first version of this note we'll focus on the full color range. It
+// will later be updated to account for the limited range.
+//
+// Color conversion matrix
+// -----------------------
+// YUV -> RGB conversion is defined as the reverse process of the RGB -> YUV,
+// So this is where we'll start.
+// At the core of a RGB -> YUV conversion are the "luma coefficients", which are
+// specific to a given color space and defined by the color space standard. In
+// FFmpeg they can be found here:
+// https://github.com/FFmpeg/FFmpeg/blob/7d606ef0ccf2946a4a21ab1ec23486cadc21864b/libavutil/csp.c#L46-L56
+//
+// For example, the BT.709 coefficients are: kr=0.2126, kg=0.7152, kb=0.0722
+// Coefficients must sum to 1.
+//
+// Conventionally Y is in [0, 1] range, and U and V are in [-0.5, 0.5] range
+// (that's mathematically, in practice they are represented in integer range).
+// The conversion is defined as:
+// https://en.wikipedia.org/wiki/YCbCr#R'G'B'_to_Y%E2%80%B2PbPr
+// Y = kr*R + kg*G + kb*B
+// U = (B - Y) * 0.5 / (1 - kb) = (B - Y) / u_scale where u_scale = 2 * (1 - kb)
+// V = (R - Y) * 0.5 / (1 - kr) = (R - Y) / v_scale where v_scale = 2 * (1 - kr)
+//
+// Putting all this into matrix form, we get:
+// [Y]   = [kr               kg            kb            ]  [R]
+// [U]     [-kr/u_scale      -kg/u_scale   (1-kb)/u_scale]  [G]
+// [V]     [(1-kr)/v_scale   -kg/v_scale   -kb)/v_scale  ]  [B]
+//
+//
+// Now, to convert YUV to RGB, we just need to invert this matrix:
+// ```py
+// import torch
+// kr, kg, kb = 0.2126, 0.7152, 0.0722  # BT.709  luma coefficients
+// u_scale = 2 * (1 - kb)
+// v_scale = 2 * (1 - kr)
+//
+// rgb_to_yuv = torch.tensor([
+//     [kr, kg, kb],
+//     [-kr/u_scale, -kg/u_scale, (1-kb)/u_scale],
+//     [(1-kr)/v_scale, -kg/v_scale, -kb/v_scale]
+// ])
+//
+// yuv_to_rgb_full = torch.linalg.inv(rgb_to_yuv)
+// print("YUV->RGB matrix (Full Range):")
+// print(yuv_to_rgb_full)
+// ```
+// And we get:
+// tensor([[ 1.0000e+00, -3.3142e-09,  1.5748e+00],
+//         [ 1.0000e+00, -1.8732e-01, -4.6812e-01],
+//         [ 1.0000e+00,  1.8556e+00,  4.6231e-09]])
+//
+// Which matches https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.709_conversion
+//
+// Color conversion in NPP
+// -----------------------
+// https://docs.nvidia.com/cuda/npp/image_color_conversion.html.
+//
+// NPP provides different ways to convert YUV to RGB:
+// - pre-defined color conversion functions like
+//   nppiNV12ToRGB_709CSC_8u_P2C3R_Ctx and nppiNV12ToRGB_709HDTV_8u_P2C3R_Ctx
+//   which are for BT.709 limited and full range, respectively.
+// - generic color conversion functions that accept a custom color conversion
+//   matrix, called ColorTwist, like nppiNV12ToRGB_8u_ColorTwist32f_P2C3R_Ctx
+//
+// We use the pre-defined functions or the color twist functions depending on
+// which one we find to be closer to the CPU results.
+//
+// The color twist functionality is *partially* described in a section named
+// "YUVToRGBColorTwist". Importantly:
+//
+// - The `nppiNV12ToRGB_8u_ColorTwist32f_P2C3R_Ctx` function takes the YUV data
+//   and the color-conversion matrix as input. The function itself and the
+//   matrix assume different ranges for YUV values:
+// - The **matrix coefficient** must assume that Y is in [0, 1] and U,V are in
+//   [-0.5, 0.5]. That's how we defined our matrix above.
+// - The function `nppiNV12ToRGB_8u_ColorTwist32f_P2C3R_Ctx` however expects all
+//   of the input Y, U, V to be in [0, 255]. That's how the data comes out of
+//   the decoder.
+// - But *internally*, `nppiNV12ToRGB_8u_ColorTwist32f_P2C3R_Ctx` needs U and V to
+//   be centered around 0, i.e. in [-128, 127]. So we need to apply a -128
+//   offset to U and V. Y doesn't need to be offset. The offset can be applied
+//   by adding a 4th column to the matrix.
+//
+//
+// So our conversion matrix becomes the following, with new offset column:
+// tensor([[ 1.0000e+00, -3.3142e-09,  1.5748e+00,     0]
+//         [ 1.0000e+00, -1.8732e-01, -4.6812e-01,     -128]
+//         [ 1.0000e+00,  1.8556e+00,  4.6231e-09 ,    -128]])
+//
+// And that's what we need to pass for BT701, full range.
+/* clang-format on */

torchcodec/_core/CudaDeviceInterface.h

@@ -0,0 +1,37 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#pragma once
+
+#include <npp.h>
+#include "src/torchcodec/_core/DeviceInterface.h"
+
+namespace facebook::torchcodec {
+
+class CudaDeviceInterface : public DeviceInterface {
+ public:
+  CudaDeviceInterface(const torch::Device& device);
+
+  virtual ~CudaDeviceInterface();
+
+  std::optional<const AVCodec*> findCodec(const AVCodecID& codecId) override;
+
+  void initializeContext(AVCodecContext* codecContext) override;
+
+  void convertAVFrameToFrameOutput(
+      const VideoStreamOptions& videoStreamOptions,
+      const AVRational& timeBase,
+      UniqueAVFrame& avFrame,
+      FrameOutput& frameOutput,
+      std::optional<torch::Tensor> preAllocatedOutputTensor =
+          std::nullopt) override;
+
+ private:
+  UniqueAVBufferRef ctx_;
+  std::unique_ptr<NppStreamContext> nppCtx_;
+};
+
+} // namespace facebook::torchcodec