update

Author: mthrok

.circleci/config.yml

@@ -654,6 +654,7 @@ jobs:
           command: .circleci/unittest/linux/scripts/run_test.sh
           environment:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CUDA: true
+              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_AUDIO_OUT_DEVICE: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MACOS: true
@@ -689,11 +690,12 @@ jobs:
       - run:
           name: Run tests
           environment:
+              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_CUDA_SMALL_MEMORY: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MACOS: true
           command: |
-            docker run -t --gpus all -v $PWD:$PWD -w $PWD -e "CI=${CI}" -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310 -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_CUDA_SMALL_MEMORY -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MACOS "${image_name}" .circleci/unittest/linux/scripts/run_test.sh
+            docker run -t --gpus all -v $PWD:$PWD -w $PWD -e "CI=${CI}" -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310 -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_CUDA_SMALL_MEMORY -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MACOS -e TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL "${image_name}" .circleci/unittest/linux/scripts/run_test.sh
       - store_test_results:
           path: test-results
       - store_artifacts:
@@ -726,6 +728,7 @@ jobs:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_SOX: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CUDA: true
+              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_KALDI: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310: true
@@ -814,6 +817,7 @@ jobs:
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CUDA: true
+              TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_QUANTIZATION: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310: true
               TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_sentencepiece: true

.circleci/config.yml.in

@@ -72,6 +72,7 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_CUDA_SMALL_MEMORY=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310=true
 
         declare -a args=(

.github/workflows/unittest-linux-gpu.yml

@@ -268,6 +268,17 @@ def skipIfNoModule(module, display_name=None):
     reason="This feature is only available for MacOS.",
     key="NO_MACOS",
 )
+def skipIfNoHWAccel(name):
+    key = "NO_HW_ACCEL"
+    if not is_ffmpeg_available():
+        return _skipIf(True, reason="ffmpeg features are not available.", key=key)
+    if not torch.cuda.is_available():
+        return _skipIf(True, reason="CUDA is not available.", key=key)
+    if torchaudio._extension._check_cuda_version() is None:
+        return _skipIf(True, "Torchaudio is not compiled with CUDA.", key=key)
+    if name not in get_video_decoders() and name not in get_video_encoders():
+        return _skipIf(True, f"{name} is not in the list of available decoders or encoders", key=key)
+    return _pass
 
 
 def zip_equal(*iterables):

test/torchaudio_unittest/common_utils/case_utils.py

@@ -1054,10 +1054,10 @@ def test_png_yuv_read_out(self):
             self.assertEqual(chunks[10], bgra, atol=0, rtol=0)
 
 
-@skipIfNoFFmpeg
 @skipIfNoHWAccel("h264_cuvid")
 class CuvidHWAccelInterfaceTest(TorchaudioTestCase):
     def test_dup_hw_acel(self):
+        """Specifying the same source stream with and without HW accel should fail (instead of segfault later)"""
         src = get_asset_path("nasa_13013.mp4")
         r = StreamReader(src)
         r.add_video_stream(-1, decoder="h264_cuvid")
@@ -1070,55 +1070,87 @@ def test_dup_hw_acel(self):
             r.add_video_stream(-1, decoder="h264_cuvid")
 
 
-@skipIfNoFFmpeg
-@skipIfNoHWAccel("h264_cuvid")
 @_media_source
-class H264CuvidDecoderTest(_MediaSourceMixin, TempDirMixin, TorchaudioTestCase):
+class CudaDecoderTest(_MediaSourceMixin, TempDirMixin, TorchaudioTestCase):
+    @skipIfNoHWAccel("h264_cuvid")
     def test_h264_cuvid(self):
+        """GPU decoder works for H264"""
         src = self.get_src(get_asset_path("nasa_13013.mp4"))
         r = StreamReader(src)
-        r.add_video_stream(-1, decoder="h264_cuvid")
-        r.process_all_packets()
-        (chunk,) = r.pop_chunks()
+        r.add_video_stream(10, decoder="h264_cuvid")
 
-        self.assertEqual(chunk.device, torch.device("cpu"))
-        self.assertEqual(chunk.dtype, torch.uint8)
-        self.assertEqual(chunk.shape, torch.Size([390, 3, 270, 480]))
+        num_frames = 0
+        for (chunk,) in r.stream():
+            self.assertEqual(chunk.device, torch.device("cpu"))
+            self.assertEqual(chunk.dtype, torch.uint8)
+            self.assertEqual(chunk.shape, torch.Size([10, 3, 270, 480]))
+            num_frames += chunk.size(0)
+        assert num_frames == 390
 
+    @skipIfNoHWAccel("h264_cuvid")
     def test_h264_cuvid_hw_accel(self):
+        """GPU decoder works for H264 with HW acceleration, and put the frames on CUDA tensor"""
         src = self.get_src(get_asset_path("nasa_13013.mp4"))
         r = StreamReader(src)
-        r.add_video_stream(-1, decoder="h264_cuvid", hw_accel="cuda")
-        r.process_all_packets()
-        (chunk,) = r.pop_chunks()
-
-        self.assertEqual(chunk.device, torch.device("cuda:0"))
-        self.assertEqual(chunk.dtype, torch.uint8)
-        self.assertEqual(chunk.shape, torch.Size([390, 3, 270, 480]))
+        r.add_video_stream(10, decoder="h264_cuvid", hw_accel="cuda")
 
+        num_frames = 0
+        for (chunk,) in r.stream():
+            self.assertEqual(chunk.device, torch.device("cuda:0"))
+            self.assertEqual(chunk.dtype, torch.uint8)
+            self.assertEqual(chunk.shape, torch.Size([10, 3, 270, 480]))
+            num_frames += chunk.size(0)
+        assert num_frames == 390
 
-@skipIfNoFFmpeg
-@skipIfNoHWAccel("hevc_cuvid")
-@_media_source
-class HevcCuvidDecoderTest(_MediaSourceMixin, TempDirMixin, TorchaudioTestCase):
+    @skipIfNoHWAccel("hevc_cuvid")
     def test_hevc_cuvid(self):
+        """GPU decoder works for H265/HEVC"""
         src = self.get_src(get_asset_path("testsrc.hevc"))
         r = StreamReader(src)
-        r.add_video_stream(-1, decoder="hevc_cuvid")
-        r.process_all_packets()
-        (chunk,) = r.pop_chunks()
+        r.add_video_stream(10, decoder="hevc_cuvid")
 
-        self.assertEqual(chunk.device, torch.device("cpu"))
-        self.assertEqual(chunk.dtype, torch.uint8)
-        self.assertEqual(chunk.shape, torch.Size([300, 3, 144, 256]))
+        num_frames = 0
+        for (chunk,) in r.stream():
+            self.assertEqual(chunk.device, torch.device("cpu"))
+            self.assertEqual(chunk.dtype, torch.uint8)
+            self.assertEqual(chunk.shape, torch.Size([10, 3, 144, 256]))
+            num_frames += chunk.size(0)
+        assert num_frames == 300
 
+    @skipIfNoHWAccel("hevc_cuvid")
     def test_hevc_cuvid_hw_accel(self):
+        """GPU decoder works for H265/HEVC with HW acceleration, and put the frames on CUDA tensor"""
         src = self.get_src(get_asset_path("testsrc.hevc"))
         r = StreamReader(src)
-        r.add_video_stream(-1, decoder="hevc_cuvid", hw_accel="cuda")
-        r.process_all_packets()
-        (chunk,) = r.pop_chunks()
+        r.add_video_stream(10, decoder="hevc_cuvid", hw_accel="cuda")
 
-        self.assertEqual(chunk.device, torch.device("cuda:0"))
-        self.assertEqual(chunk.dtype, torch.int16)
-        self.assertEqual(chunk.shape, torch.Size([300, 3, 144, 256]))
+        num_frames = 0
+        for (chunk,) in r.stream():
+            self.assertEqual(chunk.device, torch.device("cuda:0"))
+            self.assertEqual(chunk.dtype, torch.int16)
+            self.assertEqual(chunk.shape, torch.Size([10, 3, 144, 256]))
+            num_frames += chunk.size(0)
+        assert num_frames == 300
+
+
+@skipIfNoHWAccel("h264_cuvid")
+class FilterGraphWithCudaAccel(TorchaudioTestCase):
+    def test_sclae_cuda_change_size(self):
+        """scale_cuda filter can be used when HW accel is on"""
+        src = get_asset_path("nasa_13013.mp4")
+        r = StreamReader(src)
+        r.add_video_stream(10, decoder="h264_cuvid", hw_accel="cuda", filter_desc="scale_cuda=iw/2:ih/2")
+        num_frames = 0
+        for (chunk,) in r.stream():
+            self.assertEqual(chunk.device, torch.device("cuda:0"))
+            self.assertEqual(chunk.dtype, torch.uint8)
+            self.assertEqual(chunk.shape, torch.Size([10, 3, 135, 240]))
+            num_frames += chunk.size(0)
+        assert num_frames == 390
+
+    def test_scale_cuda_format(self):
+        """yuv444p format conversion does not work (yet)"""
+        src = get_asset_path("nasa_13013.mp4")
+        r = StreamReader(src)
+        with self.assertRaises(RuntimeError):
+            r.add_video_stream(10, decoder="h264_cuvid", hw_accel="cuda", filter_desc="scale_cuda=format=yuv444p")