torchcodec 0.10.0__cp312-cp312-manylinux_2_28_x86_64.whl

torchcodec/_core/CudaDeviceInterface.cpp

@@ -0,0 +1,602 @@
+#include <ATen/cuda/CUDAEvent.h>
+#include <c10/cuda/CUDAStream.h>
+#include <torch/types.h>
+#include <mutex>
+
+#include "Cache.h"
+#include "CudaDeviceInterface.h"
+#include "FFMPEGCommon.h"
+#include "ValidationUtils.h"
+
+extern "C" {
+#include <libavutil/hwcontext_cuda.h>
+#include <libavutil/pixdesc.h>
+}
+
+namespace facebook::torchcodec {
+namespace {
+
+static bool g_cuda = registerDeviceInterface(
+    DeviceInterfaceKey(torch::kCUDA),
+    [](const torch::Device& device) {
+      return new CudaDeviceInterface(device);
+    });
+
+// 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.
+
+// 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);
+
+int getFlagsAVHardwareDeviceContextCreate() {
+// 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.
+#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 26, 100)
+  return AV_CUDA_USE_CURRENT_CONTEXT;
+#else
+  return 0;
+#endif
+}
+
+UniqueAVBufferRef getHardwareDeviceContext(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");
+  int deviceIndex = getDeviceIndex(device);
+
+  UniqueAVBufferRef hardwareDeviceCtx = g_cached_hw_device_ctxs.get(device);
+  if (hardwareDeviceCtx) {
+    return hardwareDeviceCtx;
+  }
+
+  // Create hardware device context
+  c10::cuda::CUDAGuard deviceGuard(device);
+  // We set the device because we may be called from a different thread than
+  // the one that initialized the cuda context.
+  TORCH_CHECK(
+      cudaSetDevice(deviceIndex) == cudaSuccess, "Failed to set CUDA device");
+  AVBufferRef* hardwareDeviceCtxRaw = nullptr;
+  std::string deviceOrdinal = std::to_string(deviceIndex);
+
+  int err = av_hwdevice_ctx_create(
+      &hardwareDeviceCtxRaw,
+      type,
+      deviceOrdinal.c_str(),
+      nullptr,
+      getFlagsAVHardwareDeviceContextCreate());
+
+  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 UniqueAVBufferRef(hardwareDeviceCtxRaw);
+}
+
+} // 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());
+
+  initializeCudaContextWithPytorch(device_);
+
+  hardwareDeviceCtx_ = getHardwareDeviceContext(device_);
+  nppCtx_ = getNppStreamContext(device_);
+}
+
+CudaDeviceInterface::~CudaDeviceInterface() {
+  if (hardwareDeviceCtx_) {
+    g_cached_hw_device_ctxs.addIfCacheHasCapacity(
+        device_, std::move(hardwareDeviceCtx_));
+  }
+  returnNppStreamContextToCache(device_, std::move(nppCtx_));
+}
+
+void CudaDeviceInterface::initialize(
+    const AVStream* avStream,
+    const UniqueDecodingAVFormatContext& avFormatCtx,
+    const SharedAVCodecContext& codecContext) {
+  TORCH_CHECK(avStream != nullptr, "avStream is null");
+  codecContext_ = codecContext;
+  timeBase_ = avStream->time_base;
+
+  // TODO: Ideally, we should keep all interface implementations independent.
+  cpuInterface_ = createDeviceInterface(torch::kCPU);
+  TORCH_CHECK(
+      cpuInterface_ != nullptr, "Failed to create CPU device interface");
+  cpuInterface_->initialize(avStream, avFormatCtx, codecContext);
+  cpuInterface_->initializeVideo(
+      VideoStreamOptions(),
+      {},
+      /*resizedOutputDims=*/std::nullopt);
+}
+
+void CudaDeviceInterface::initializeVideo(
+    const VideoStreamOptions& videoStreamOptions,
+    [[maybe_unused]] const std::vector<std::unique_ptr<Transform>>& transforms,
+    [[maybe_unused]] const std::optional<FrameDims>& resizedOutputDims) {
+  videoStreamOptions_ = videoStreamOptions;
+}
+
+void CudaDeviceInterface::registerHardwareDeviceWithCodec(
+    AVCodecContext* codecContext) {
+  TORCH_CHECK(
+      hardwareDeviceCtx_, "Hardware device context has not been initialized");
+  TORCH_CHECK(codecContext != nullptr, "codecContext is null");
+  codecContext->hw_device_ctx = av_buffer_ref(hardwareDeviceCtx_.get());
+}
+
+UniqueAVFrame CudaDeviceInterface::maybeConvertAVFrameToNV12OrRGB24(
+    UniqueAVFrame& avFrame) {
+  // We need FFmpeg filters to handle those conversion cases which are not
+  // directly implemented in CUDA or CPU device interface (in case of a
+  // fallback).
+
+  // Input frame is on CPU, we will just pass it to CPU device interface, so
+  // skipping filters context as CPU device interface will handle everything for
+  // us.
+  if (avFrame->format != AV_PIX_FMT_CUDA) {
+    return std::move(avFrame);
+  }
+
+  auto hwFramesCtx =
+      reinterpret_cast<AVHWFramesContext*>(avFrame->hw_frames_ctx->data);
+  TORCH_CHECK(
+      hwFramesCtx != nullptr,
+      "The AVFrame does not have a hw_frames_ctx. "
+      "That's unexpected, please report this to the TorchCodec repo.");
+
+  AVPixelFormat actualFormat = hwFramesCtx->sw_format;
+
+  // If the frame is already in NV12 format, we don't need to do anything.
+  if (actualFormat == AV_PIX_FMT_NV12) {
+    return std::move(avFrame);
+  }
+
+  AVPixelFormat outputFormat;
+  std::stringstream filters;
+
+  unsigned version_int = avfilter_version();
+  if (version_int < AV_VERSION_INT(8, 0, 103)) {
+    // Color conversion support ('format=' option) was added to scale_cuda from
+    // n5.0. With the earlier version of ffmpeg we have no choice but use CPU
+    // filters. See:
+    // https://github.com/FFmpeg/FFmpeg/commit/62dc5df941f5e196164c151691e4274195523e95
+    outputFormat = AV_PIX_FMT_RGB24;
+
+    auto actualFormatName = av_get_pix_fmt_name(actualFormat);
+    TORCH_CHECK(
+        actualFormatName != nullptr,
+        "The actual format of a frame is unknown to FFmpeg. "
+        "That's unexpected, please report this to the TorchCodec repo.");
+
+    filters << "hwdownload,format=" << actualFormatName;
+  } else {
+    // Actual output color format will be set via filter options
+    outputFormat = AV_PIX_FMT_CUDA;
+
+    filters << "scale_cuda=format=nv12:interp_algo=bilinear";
+  }
+
+  enum AVPixelFormat frameFormat =
+      static_cast<enum AVPixelFormat>(avFrame->format);
+
+  auto newContext = std::make_unique<FiltersContext>(
+      avFrame->width,
+      avFrame->height,
+      frameFormat,
+      avFrame->sample_aspect_ratio,
+      avFrame->width,
+      avFrame->height,
+      outputFormat,
+      filters.str(),
+      timeBase_,
+      av_buffer_ref(avFrame->hw_frames_ctx));
+
+  if (!nv12Conversion_ || *nv12ConversionContext_ != *newContext) {
+    nv12Conversion_ =
+        std::make_unique<FilterGraph>(*newContext, videoStreamOptions_);
+    nv12ConversionContext_ = std::move(newContext);
+  }
+  auto filteredAVFrame = nv12Conversion_->convert(avFrame);
+
+  // If this check fails it means the frame wasn't
+  // reshaped to its expected dimensions by filtergraph.
+  TORCH_CHECK(
+      (filteredAVFrame->width == nv12ConversionContext_->outputWidth) &&
+          (filteredAVFrame->height == nv12ConversionContext_->outputHeight),
+      "Expected frame from filter graph of ",
+      nv12ConversionContext_->outputWidth,
+      "x",
+      nv12ConversionContext_->outputHeight,
+      ", got ",
+      filteredAVFrame->width,
+      "x",
+      filteredAVFrame->height);
+
+  return filteredAVFrame;
+}
+
+void CudaDeviceInterface::convertAVFrameToFrameOutput(
+    UniqueAVFrame& avFrame,
+    FrameOutput& frameOutput,
+    std::optional<torch::Tensor> preAllocatedOutputTensor) {
+  validatePreAllocatedTensorShape(preAllocatedOutputTensor, avFrame);
+
+  hasDecodedFrame_ = true;
+
+  // All of our CUDA decoding assumes NV12 format. We handle non-NV12 formats by
+  // converting them to NV12.
+  avFrame = maybeConvertAVFrameToNV12OrRGB24(avFrame);
+
+  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. There are two possible
+    // reasons:
+    //
+    //   1. During maybeConvertAVFrameToNV12OrRGB24(), we had a non-NV12 format
+    //      frame and we're on FFmpeg 4.4 or earlier. In such cases, we had to
+    //      use CPU filters and we just converted the frame to RGB24.
+    //   2. This is what NVDEC gave 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.
+    //
+    // In both cases, we have a frame on the CPU. We send the frame back to the
+    // CUDA device when we're done.
+
+    enum AVPixelFormat frameFormat =
+        static_cast<enum AVPixelFormat>(avFrame->format);
+
+    FrameOutput cpuFrameOutput;
+    if (frameFormat == AV_PIX_FMT_RGB24) {
+      // Reason 1 above. The frame is already in RGB24, we just need to convert
+      // it to a tensor.
+      cpuFrameOutput.data = rgbAVFrameToTensor(avFrame);
+    } else {
+      // Reason 2 above. We need to do a full conversion which requires an
+      // actual CPU device.
+      cpuInterface_->convertAVFrameToFrameOutput(avFrame, cpuFrameOutput);
+    }
+
+    // Finally, we need to send the frame back to the GPU. Note that the
+    // pre-allocated tensor is on the GPU, so we can't send that to the CPU
+    // device interface. We copy it over here.
+    if (preAllocatedOutputTensor.has_value()) {
+      preAllocatedOutputTensor.value().copy_(cpuFrameOutput.data);
+      frameOutput.data = preAllocatedOutputTensor.value();
+    } else {
+      frameOutput.data = cpuFrameOutput.data.to(device_);
+    }
+
+    usingCPUFallback_ = true;
+    return;
+  }
+
+  usingCPUFallback_ = false;
+
+  // 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. This SHOULD
+  // be enforced by our call to maybeConvertAVFrameToNV12OrRGB24() above.
+  TORCH_CHECK(
+      avFrame->hw_frames_ctx != nullptr,
+      "The AVFrame does not have a hw_frames_ctx. This should never happen");
+  AVHWFramesContext* hwFramesCtx =
+      reinterpret_cast<AVHWFramesContext*>(avFrame->hw_frames_ctx->data);
+  TORCH_CHECK(
+      hwFramesCtx != nullptr,
+      "The AVFrame does not have a valid hw_frames_ctx. This should never happen");
+
+  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. "
+      "That's unexpected, please report this to the TorchCodec repo.");
+
+  // Figure out the NVDEC stream from the avFrame's hardware context.
+  // In reality, we know that this stream is 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);
+  TORCH_CHECK(cudaDeviceCtx != nullptr, "The hardware context is null");
+  at::cuda::CUDAStream nvdecStream = // That's always the default stream. Sad.
+      c10::cuda::getStreamFromExternal(cudaDeviceCtx->stream, device_.index());
+
+  frameOutput.data = convertNV12FrameToRGB(
+      avFrame, device_, nppCtx_, nvdecStream, preAllocatedOutputTensor);
+}
+
+// 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,
+    bool isDecoder) {
+  void* i = nullptr;
+  const AVCodec* codec = nullptr;
+  while ((codec = av_codec_iterate(&i)) != nullptr) {
+    TORCH_CHECK(
+        codec != nullptr,
+        "codec returned by av_codec_iterate should not be null");
+    if (isDecoder) {
+      if (codec->id != codecId || !av_codec_is_decoder(codec)) {
+        continue;
+      }
+    } else {
+      if (codec->id != codecId || !av_codec_is_encoder(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;
+}
+
+std::string CudaDeviceInterface::getDetails() {
+  // Note: for this interface specifically the fallback is only known after a
+  // frame has been decoded, not before: that's when FFmpeg decides to fallback,
+  // so we can't know earlier.
+  if (!hasDecodedFrame_) {
+    return std::string(
+        "FFmpeg CUDA Device Interface. Fallback status unknown (no frames decoded).");
+  }
+  return std::string("FFmpeg CUDA Device Interface. Using ") +
+      (usingCPUFallback_ ? "CPU fallback." : "NVDEC.");
+}
+
+// --------------------------------------------------------------------------
+// Below are methods exclusive to video encoding:
+// --------------------------------------------------------------------------
+namespace {
+// Note: [RGB -> YUV Color Conversion, limited color range]
+//
+// For context on this subject, first read the note:
+// [YUV -> RGB Color Conversion, color space and color range]
+// https://github.com/meta-pytorch/torchcodec/blob/main/src/torchcodec/_core/CUDACommon.cpp#L63-L65
+//
+// Lets encode RGB -> YUV in the limited color range for BT.601 color space.
+// In limited range, the [0, 255] range is mapped into [16-235] for Y, and into
+// [16-240] for U,V.
+// To implement, we get the full range conversion matrix as before, then scale:
+// - Y channel: scale by (235-16)/255 = 219/255
+// - U,V channels: scale by (240-16)/255 = 224/255
+// https://en.wikipedia.org/wiki/YCbCr#Y%E2%80%B2PbPr_to_Y%E2%80%B2CbCr
+//
+// ```py
+// import torch
+// kr, kg, kb = 0.299, 0.587, 0.114  # BT.601 luma coefficients
+// u_scale = 2 * (1 - kb)
+// v_scale = 2 * (1 - kr)
+//
+// rgb_to_yuv_full = 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]
+// ])
+//
+// full_to_limited_y_scale = 219.0 / 255.0
+// full_to_limited_uv_scale = 224.0 / 255.0
+//
+// rgb_to_yuv_limited = rgb_to_yuv_full * torch.tensor([
+//     [full_to_limited_y_scale],
+//     [full_to_limited_uv_scale],
+//     [full_to_limited_uv_scale]
+// ])
+//
+// print("RGB->YUV matrix (Limited Range BT.601):")
+// print(rgb_to_yuv_limited)
+// ```
+//
+// This yields:
+// tensor([[ 0.2568,  0.5041,  0.0979],
+//         [-0.1482, -0.2910,  0.4392],
+//         [ 0.4392, -0.3678, -0.0714]])
+//
+// Which matches https://fourcc.org/fccyvrgb.php
+//
+// To perform color conversion in NPP, we are required to provide these color
+// conversion matrices to ColorTwist functions, for example,
+// `nppiRGBToNV12_8u_ColorTwist32f_C3P2R_Ctx`.
+// https://docs.nvidia.com/cuda/npp/image_color_conversion.html
+//
+// These offsets are added in the 4th column of each conversion matrix below.
+// - In limited range, Y is offset by 16 to add the lower margin.
+// - In both color ranges, U,V are offset by 128 to be centered around 0.
+//
+// RGB to YUV conversion matrices to use in NPP color conversion functions
+struct ColorConversionMatrices {
+  static constexpr Npp32f BT601_LIMITED[3][4] = {
+      {0.2568f, 0.5041f, 0.0979f, 16.0f},
+      {-0.1482f, -0.2910f, 0.4392f, 128.0f},
+      {0.4392f, -0.3678f, -0.0714f, 128.0f}};
+
+  static constexpr Npp32f BT601_FULL[3][4] = {
+      {0.2990f, 0.5870f, 0.1140f, 0.0f},
+      {-0.1687f, -0.3313f, 0.5000f, 128.0f},
+      {0.5000f, -0.4187f, -0.0813f, 128.0f}};
+
+  static constexpr Npp32f BT709_LIMITED[3][4] = {
+      {0.1826f, 0.6142f, 0.0620f, 16.0f},
+      {-0.1006f, -0.3386f, 0.4392f, 128.0f},
+      {0.4392f, -0.3989f, -0.0403f, 128.0f}};
+
+  static constexpr Npp32f BT709_FULL[3][4] = {
+      {0.2126f, 0.7152f, 0.0722f, 0.0f},
+      {-0.1146f, -0.3854f, 0.5000f, 128.0f},
+      {0.5000f, -0.4542f, -0.0458f, 128.0f}};
+
+  static constexpr Npp32f BT2020_LIMITED[3][4] = {
+      {0.2256f, 0.5823f, 0.0509f, 16.0f},
+      {-0.1227f, -0.3166f, 0.4392f, 128.0f},
+      {0.4392f, -0.4039f, -0.0353f, 128.0f}};
+
+  static constexpr Npp32f BT2020_FULL[3][4] = {
+      {0.2627f, 0.6780f, 0.0593f, 0.0f},
+      {-0.139630f, -0.360370f, 0.5000f, 128.0f},
+      {0.5000f, -0.459786f, -0.040214f, 128.0f}};
+};
+
+// Returns conversion matrix based on codec context color space and range
+const Npp32f (*getConversionMatrix(AVCodecContext* codecContext))[4] {
+  if (codecContext->color_range == AVCOL_RANGE_MPEG || // limited range
+      codecContext->color_range == AVCOL_RANGE_UNSPECIFIED) {
+    if (codecContext->colorspace == AVCOL_SPC_BT470BG) {
+      return ColorConversionMatrices::BT601_LIMITED;
+    } else if (codecContext->colorspace == AVCOL_SPC_BT709) {
+      return ColorConversionMatrices::BT709_LIMITED;
+    } else if (codecContext->colorspace == AVCOL_SPC_BT2020_NCL) {
+      return ColorConversionMatrices::BT2020_LIMITED;
+    } else { // default to BT.601
+      return ColorConversionMatrices::BT601_LIMITED;
+    }
+  } else if (codecContext->color_range == AVCOL_RANGE_JPEG) { // full range
+    if (codecContext->colorspace == AVCOL_SPC_BT470BG) {
+      return ColorConversionMatrices::BT601_FULL;
+    } else if (codecContext->colorspace == AVCOL_SPC_BT709) {
+      return ColorConversionMatrices::BT709_FULL;
+    } else if (codecContext->colorspace == AVCOL_SPC_BT2020_NCL) {
+      return ColorConversionMatrices::BT2020_FULL;
+    } else { // default to BT.601
+      return ColorConversionMatrices::BT601_FULL;
+    }
+  }
+  return ColorConversionMatrices::BT601_LIMITED;
+}
+} // namespace
+
+UniqueAVFrame CudaDeviceInterface::convertCUDATensorToAVFrameForEncoding(
+    const torch::Tensor& tensor,
+    int frameIndex,
+    AVCodecContext* codecContext) {
+  TORCH_CHECK(
+      tensor.dim() == 3 && tensor.size(0) == 3,
+      "Expected 3D RGB tensor (CHW format), got shape: ",
+      tensor.sizes());
+  TORCH_CHECK(
+      tensor.device().type() == torch::kCUDA,
+      "Expected tensor on CUDA device, got: ",
+      tensor.device().str());
+
+  UniqueAVFrame avFrame(av_frame_alloc());
+  TORCH_CHECK(avFrame != nullptr, "Failed to allocate AVFrame");
+  int height = static_cast<int>(tensor.size(1));
+  int width = static_cast<int>(tensor.size(2));
+
+  // TODO-VideoEncoder: Unify AVFrame creation with CPU version of this method
+  avFrame->format = AV_PIX_FMT_CUDA;
+  avFrame->height = height;
+  avFrame->width = width;
+  avFrame->pts = frameIndex;
+
+  // FFmpeg's av_hwframe_get_buffer is used to allocate memory on CUDA device.
+  // TODO-VideoEncoder: Consider using pytorch to allocate CUDA memory for
+  // efficiency
+  int ret =
+      av_hwframe_get_buffer(codecContext->hw_frames_ctx, avFrame.get(), 0);
+  TORCH_CHECK(
+      ret >= 0,
+      "Failed to allocate hardware frame: ",
+      getFFMPEGErrorStringFromErrorCode(ret));
+
+  TORCH_CHECK(
+      avFrame != nullptr && avFrame->data[0] != nullptr,
+      "avFrame must be pre-allocated with CUDA memory");
+
+  // TODO VideoEncoder: Investigate ways to avoid this copy
+  torch::Tensor hwcFrame = tensor.permute({1, 2, 0}).contiguous();
+
+  NppiSize oSizeROI = {width, height};
+  NppStatus status;
+  // Convert to NV12, as CUDA_ENCODING_PIXEL_FORMAT is always NV12 currently
+  status = nppiRGBToNV12_8u_ColorTwist32f_C3P2R_Ctx(
+      static_cast<const Npp8u*>(hwcFrame.data_ptr()),
+      hwcFrame.stride(0) * hwcFrame.element_size(),
+      avFrame->data,
+      avFrame->linesize,
+      oSizeROI,
+      getConversionMatrix(codecContext),
+      *nppCtx_);
+
+  TORCH_CHECK(
+      status == NPP_SUCCESS,
+      "Failed to convert RGB to ",
+      av_get_pix_fmt_name(DeviceInterface::CUDA_ENCODING_PIXEL_FORMAT),
+      ": NPP error code ",
+      status);
+
+  avFrame->colorspace = codecContext->colorspace;
+  avFrame->color_range = codecContext->color_range;
+  return avFrame;
+}
+
+// Allocates and initializes AVHWFramesContext, and sets pixel format fields
+// to enable encoding with CUDA device. The hw_frames_ctx field is needed by
+// FFmpeg to allocate frames on GPU's memory.
+void CudaDeviceInterface::setupHardwareFrameContextForEncoding(
+    AVCodecContext* codecContext) {
+  TORCH_CHECK(codecContext != nullptr, "codecContext is null");
+  TORCH_CHECK(
+      hardwareDeviceCtx_, "Hardware device context has not been initialized");
+
+  AVBufferRef* hwFramesCtxRef = av_hwframe_ctx_alloc(hardwareDeviceCtx_.get());
+  TORCH_CHECK(
+      hwFramesCtxRef != nullptr,
+      "Failed to allocate hardware frames context for codec");
+
+  codecContext->sw_pix_fmt = DeviceInterface::CUDA_ENCODING_PIXEL_FORMAT;
+  // Always set pixel format to support CUDA encoding.
+  codecContext->pix_fmt = AV_PIX_FMT_CUDA;
+
+  AVHWFramesContext* hwFramesCtx =
+      reinterpret_cast<AVHWFramesContext*>(hwFramesCtxRef->data);
+  hwFramesCtx->format = codecContext->pix_fmt;
+  hwFramesCtx->sw_format = codecContext->sw_pix_fmt;
+  hwFramesCtx->width = codecContext->width;
+  hwFramesCtx->height = codecContext->height;
+
+  int ret = av_hwframe_ctx_init(hwFramesCtxRef);
+  if (ret < 0) {
+    av_buffer_unref(&hwFramesCtxRef);
+    TORCH_CHECK(
+        false,
+        "Failed to initialize CUDA frames context for codec: ",
+        getFFMPEGErrorStringFromErrorCode(ret));
+  }
+  codecContext->hw_frames_ctx = hwFramesCtxRef;
+}
+} // namespace facebook::torchcodec

torchcodec/_core/CudaDeviceInterface.h

@@ -0,0 +1,79 @@
+// 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 "CUDACommon.h"
+#include "DeviceInterface.h"
+#include "FilterGraph.h"
+
+namespace facebook::torchcodec {
+
+class CudaDeviceInterface : public DeviceInterface {
+ public:
+  CudaDeviceInterface(const torch::Device& device);
+
+  virtual ~CudaDeviceInterface();
+
+  std::optional<const AVCodec*> findCodec(
+      const AVCodecID& codecId,
+      bool isDecoder = true) override;
+
+  void initialize(
+      const AVStream* avStream,
+      const UniqueDecodingAVFormatContext& avFormatCtx,
+      const SharedAVCodecContext& codecContext) override;
+
+  void initializeVideo(
+      const VideoStreamOptions& videoStreamOptions,
+      [[maybe_unused]] const std::vector<std::unique_ptr<Transform>>&
+          transforms,
+      [[maybe_unused]] const std::optional<FrameDims>& resizedOutputDims)
+      override;
+
+  void registerHardwareDeviceWithCodec(AVCodecContext* codecContext) override;
+
+  void convertAVFrameToFrameOutput(
+      UniqueAVFrame& avFrame,
+      FrameOutput& frameOutput,
+      std::optional<torch::Tensor> preAllocatedOutputTensor) override;
+
+  std::string getDetails() override;
+
+  UniqueAVFrame convertCUDATensorToAVFrameForEncoding(
+      const torch::Tensor& tensor,
+      int frameIndex,
+      AVCodecContext* codecContext) override;
+
+  void setupHardwareFrameContextForEncoding(
+      AVCodecContext* codecContext) override;
+
+ private:
+  // Our CUDA decoding code assumes NV12 format. In order to handle other
+  // kinds of input, we need to convert them to NV12. Our current implementation
+  // does this using filtergraph.
+  UniqueAVFrame maybeConvertAVFrameToNV12OrRGB24(UniqueAVFrame& avFrame);
+
+  // We sometimes encounter frames that cannot be decoded on the CUDA device.
+  // Rather than erroring out, we decode them on the CPU.
+  std::unique_ptr<DeviceInterface> cpuInterface_;
+
+  VideoStreamOptions videoStreamOptions_;
+  AVRational timeBase_;
+
+  UniqueAVBufferRef hardwareDeviceCtx_;
+  UniqueNppContext nppCtx_;
+
+  // This filtergraph instance is only used for NV12 format conversion in
+  // maybeConvertAVFrameToNV12().
+  std::unique_ptr<FiltersContext> nv12ConversionContext_;
+  std::unique_ptr<FilterGraph> nv12Conversion_;
+
+  bool usingCPUFallback_ = false;
+  bool hasDecodedFrame_ = false;
+};
+
+} // namespace facebook::torchcodec