Build and use GPU-enabled FFmpeg in doc CI (#3045)

Summary:
This commit add the step to build FFmpeg with GPU decoder in build_doc job so that we can use GPU decoder/encoder in documentations.

Pull Request resolved: https://github.com/pytorch/audio/pull/3045

Reviewed By: nateanl

Differential Revision: D45965739

Pulled By: mthrok

fbshipit-source-id: c167eb3ef347860a51efa906068fa2daa556f017

.github/workflows/build_docs.yml

@@ -37,6 +37,8 @@ jobs:
         export PYTHON_VERSION="${{ matrix.python_version }}"
         export CU_VERSION="${{ matrix.cuda_arch_version }}"
         export CUDATOOLKIT="pytorch-cuda=${CU_VERSION}"
+        export CODEC_HEADER_VERSION=n11.1.5.2
+        export COMPUTE_CAPABILITY=86
 
         # Set CHANNEL
         if [[(${GITHUB_EVENT_NAME} = 'pull_request' && (${GITHUB_BASE_REF} = 'release'*)) || (${GITHUB_REF} = 'refs/heads/release'*) ]]; then
@@ -49,6 +51,50 @@ jobs:
         conda create --quiet -y --prefix ci_env python="${PYTHON_VERSION}"
         conda activate ./ci_env
 
+        # Install FFmpeg-dependencies
+        conda install --quiet -y -c conda-forge yasm x264 gnutls pkg-config lame libopus libvpx openh264 openssl x264
+
+        # Build FFmpeg with NVIDIA Video Codec SDK
+        # TODO cache this
+        (
+            git clone --quiet https://git.videolan.org/git/ffmpeg/nv-codec-headers.git
+            cd nv-codec-headers
+            git checkout ${CODEC_HEADER_VERSION}
+            make PREFIX=${CONDA_PREFIX} install
+        )
+
+        (
+            wget -q https://github.com/FFmpeg/FFmpeg/archive/refs/tags/n5.1.3.tar.gz
+            tar -xf n5.1.3.tar.gz
+            cd ./FFmpeg-n5.1.3
+            export revision=5.1.3
+            ./configure \
+                --prefix=${CONDA_PREFIX} \
+                --extra-cflags="-I${CONDA_PREFIX}/include" \
+                --extra-ldflags="-L${CONDA_PREFIX}/lib" \
+                --nvccflags="-gencode arch=compute_${COMPUTE_CAPABILITY},code=sm_${COMPUTE_CAPABILITY} -O2" \
+                --disable-doc \
+                --enable-rpath \
+                --disable-static \
+                --enable-protocol=https \
+                --enable-gnutls \
+                --enable-shared \
+                --enable-gpl \
+                --enable-nonfree \
+                --enable-libmp3lame \
+                --enable-libx264 \
+                --enable-cuda-nvcc \
+                --enable-nvenc \
+                --enable-cuvid \
+                --enable-nvdec
+            make clean
+            make -j > /dev/null 2>&1
+            make install
+            # test
+            src="https://download.pytorch.org/torchaudio/tutorial-assets/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4_small.mp4"
+            ffmpeg -y -vsync 0 -hwaccel cuvid -hwaccel_output_format cuda -c:v h264_cuvid -resize 360x240 -i "${src}" -c:a copy -c:v h264_nvenc -b:v 5M test.mp4
+        )
+
         # Install PyTorch
         set -ex
         set +u  # don't know why
@@ -60,14 +106,12 @@ jobs:
           "${CUDATOOLKIT}"
 
         # Install torchaudio
-        # TODO: Enable NVDec/NVEnc
-        conda install --quiet -y -c conda-forge 'ffmpeg>=5.0,<6' pkg-config
         pip --quiet install cmake>=3.18.0 ninja
         cd packaging
         . ./pkg_helpers.bash
         setup_build_version
         cd ../
-        USE_FFMPEG=1 pip install --progress-bar off -v -e . --no-use-pep517
+        FFMPEG_ROOT=${CONDA_PREFIX} USE_FFMPEG=1 USE_CUDA=1 pip install --progress-bar off -v -e . --no-use-pep517
 
         # Install runtime dependencies
         pip --quiet install git+https://github.com/kpu/kenlm/ flashlight-text

docs/source/index.rst

@@ -44,7 +44,6 @@ model implementations and application components.
    tutorials/streamreader_advanced_tutorial
    tutorials/streamwriter_basic_tutorial
    tutorials/streamwriter_advanced
-   hw_acceleration_tutorial
 
    tutorials/effector_tutorial
    tutorials/audio_resampling_tutorial