fix: Solve CUDA AV1 decoding

CONTRIBUTING.md

@@ -42,11 +42,12 @@ git clone [email protected]:pytorch/torchcodec.git
 cd torchcodec
 
 pip install -e ".[dev]" --no-build-isolation -vv
+# Or, for cuda support: ENABLE_CUDA=1 pip install -e ".[dev]" --no-build-isolation -vv
 ```
 
 ### Running unit tests
 
-To run python tests run:
+To run python tests run (please make sure `torchvision` is installed):
 
 ```bash
 pytest test -vvv

src/torchcodec/decoders/_core/CPUOnlyDevice.cpp

@@ -35,4 +35,11 @@ void releaseContextOnCuda(
   throwUnsupportedDeviceError(device);
 }
 
+void forceCudaCodec(
+    const torch::Device& device,
+    AVCodecPtr* codec,
+    const AVCodecID& codecId) {
+  throwUnsupportedDeviceError(device);
+}
+
 } // namespace facebook::torchcodec

src/torchcodec/decoders/_core/CudaDevice.cpp

@@ -256,4 +256,36 @@ void convertAVFrameToDecodedOutputOnCuda(
           << " took: " << duration.count() << "us" << std::endl;
 }
 
+// inspired by https://github.com/FFmpeg/FFmpeg/commit/ad67ea9
+void forceCudaCodec(
+    const torch::Device& device,
+    AVCodecPtr* codec,
+    const AVCodecID& codecId) {
+  if (device.type() != torch::kCUDA) {
+    return;
+  }
+
+  const AVCodec* c;
+  void* i = NULL;
+  bool found = false;
+
+  while (!found && (c = av_codec_iterate(&i))) {
+    const AVCodecHWConfig* config;
+
+    if (c->id != codecId || !av_codec_is_decoder(c)) {
+      continue;
+    }
+
+    for (int j = 0; config = avcodec_get_hw_config(c, j); j++) {
+      if (config->device_type == AV_HWDEVICE_TYPE_CUDA) {
+        found = true;
+      }
+    }
+  }
+
+  if (found) {
+    *codec = c;
+  }
+}
+
 } // namespace facebook::torchcodec

src/torchcodec/decoders/_core/DeviceInterface.h

@@ -10,6 +10,7 @@
 #include <memory>
 #include <stdexcept>
 #include <string>
+#include "FFMPEGCommon.h"
 #include "src/torchcodec/decoders/_core/VideoDecoder.h"
 
 extern "C" {
@@ -43,4 +44,9 @@ void releaseContextOnCuda(
     const torch::Device& device,
     AVCodecContext* codecContext);
 
+void forceCudaCodec(
+    const torch::Device& device,
+    AVCodecPtr* codec,
+    const AVCodecID& codecId);
+
 } // namespace facebook::torchcodec

src/torchcodec/decoders/_core/VideoDecoder.cpp

@@ -456,6 +456,12 @@ void VideoDecoder::addVideoStreamDecoder(
         "Stream with index " + std::to_string(streamNumber) +
         " is not a video stream.");
   }
+
+  if (options.device.type() == torch::kCUDA) {
+    forceCudaCodec(
+        options.device, &codec, streamInfo.stream->codecpar->codec_id);
+  }
+
   AVCodecContext* codecContext = avcodec_alloc_context3(codec);
   codecContext->thread_count = options.ffmpegThreadCount.value_or(0);
   TORCH_CHECK(codecContext != nullptr);

test/decoders/test_video_decoder.py

@@ -11,7 +11,14 @@
 
 from torchcodec.decoders import _core, VideoDecoder
 
-from ..utils import assert_frames_equal, cpu_and_cuda, H265_VIDEO, in_fbcode, NASA_VIDEO
+from ..utils import (
+    assert_frames_equal,
+    AV1_VIDEO,
+    cpu_and_cuda,
+    H265_VIDEO,
+    in_fbcode,
+    NASA_VIDEO,
+)
 
 
 class TestVideoDecoder:
@@ -409,6 +416,17 @@ def test_get_frames_at_fails(self, device):
         with pytest.raises(RuntimeError, match="Expected a value of type"):
             decoder.get_frames_at([0.3])
 
+    def test_get_frame_at_av1(self):
+        # We don't parametrize with CUDA because the current GPUs on CI do not
+        # support AV1:
+        decoder = VideoDecoder(AV1_VIDEO.path, device="cpu")
+        ref_frame11 = AV1_VIDEO.get_frame_data_by_index(10)
+        ref_frame_info11 = AV1_VIDEO.get_frame_info(10)
+        decoded_frame11 = decoder.get_frame_at(10)
+        assert decoded_frame11.duration_seconds == ref_frame_info11.duration_seconds
+        assert decoded_frame11.pts_seconds == ref_frame_info11.pts_seconds
+        assert_frames_equal(decoded_frame11.data, ref_frame11.to(device="cpu"))
+
     @pytest.mark.parametrize("device", cpu_and_cuda())
     def test_get_frame_played_at(self, device):
         decoder = VideoDecoder(NASA_VIDEO.path, device=device)

test/generate_reference_resources.sh

@@ -61,3 +61,19 @@ do
   python3 "$TORCHCODEC_PATH/test/convert_image_to_tensor.py" "$bmp"
   rm -f "$bmp"
 done
+
+# This video was generated by running the following:
+# ffmpeg -f lavfi -i testsrc=duration=5:size=640x360:rate=25,format=yuv420p -c:v libaom-av1 -crf 30 -colorspace bt709 -color_primaries bt709 -color_trc bt709 av1_video.mkv
+# Note that this video only has 1 stream, at index 0.
+VIDEO_PATH=$RESOURCES_DIR/av1_video.mkv
+FRAMES=(10)
+for frame in "${FRAMES[@]}"; do
+  frame_name=$(printf "%06d" "$frame")
+  ffmpeg -y -i "$VIDEO_PATH" -vf select="eq(n\,$frame)" -vsync vfr -q:v 2 "$VIDEO_PATH.stream0.frame$frame_name.bmp"
+done
+
+for bmp in "$RESOURCES_DIR"/*.bmp
+do
+  python3 "$TORCHCODEC_PATH/test/convert_image_to_tensor.py" "$bmp"
+  rm -f "$bmp"
+done

test/utils.py

@@ -312,3 +312,18 @@ def get_empty_chw_tensor(self, *, stream_index: int) -> torch.Tensor:
         },
     },
 )
+
+AV1_VIDEO = TestVideo(
+    filename="av1_video.mkv",
+    default_stream_index=0,
+    # This metadata is extracted manually.
+    #  $ ffprobe -v error -hide_banner -select_streams v:0 -show_frames -of json test/resources/av1_video.mkv > out.json
+    stream_infos={
+        0: TestVideoStreamInfo(width=640, height=360, num_color_channels=3),
+    },
+    frames={
+        0: {
+            10: TestFrameInfo(pts_seconds=0.400000, duration_seconds=0.040000),
+        },
+    },
+)