Temporary revert GPU support

CMakeLists.txt

@@ -1,9 +1,6 @@
 cmake_minimum_required(VERSION 3.18)
 project(TorchCodec)
 
-option(ENABLE_CUDA "Enable CUDA decoding using NVDEC" OFF)
-option(ENABLE_NVTX "Enable NVTX annotations for profiling" OFF)
-
 add_subdirectory(src/torchcodec/decoders/_core)
 
 

README.md

@@ -140,7 +140,3 @@ guide](CONTRIBUTING.md) for more details.
 ## License
 
 TorchCodec is released under the [BSD 3 license](./LICENSE).
-
-
-If you are building with ENABLE_CUDA and/or ENABLE_NVTX please review
-[Nvidia licenses](https://docs.nvidia.com/cuda/eula/index.html).

benchmarks/decoders/BenchmarkDecodersMain.cpp

@@ -63,7 +63,7 @@ void runNDecodeIterations(
     decoder->addVideoStreamDecoder(-1);
     for (double pts : ptsList) {
       decoder->setCursorPtsInSeconds(pts);
-      torch::Tensor tensor = decoder->getNextDecodedOutputNoDemux().frame;
+      torch::Tensor tensor = decoder->getNextDecodedOutput().frame;
     }
     if (i + 1 == warmupIterations) {
       start = std::chrono::high_resolution_clock::now();
@@ -95,7 +95,7 @@ void runNdecodeIterationsGrabbingConsecutiveFrames(
         VideoDecoder::createFromFilePath(videoPath);
     decoder->addVideoStreamDecoder(-1);
     for (int j = 0; j < consecutiveFrameCount; ++j) {
-      torch::Tensor tensor = decoder->getNextDecodedOutputNoDemux().frame;
+      torch::Tensor tensor = decoder->getNextDecodedOutput().frame;
     }
     if (i + 1 == warmupIterations) {
       start = std::chrono::high_resolution_clock::now();
@@ -145,8 +145,7 @@ void runNDecodeIterationsWithCustomOps(
         /*height=*/std::nullopt,
         /*thread_count=*/std::nullopt,
         /*dimension_order=*/std::nullopt,
-        /*stream_index=*/std::nullopt,
-        /*device_string=*/std::nullopt);
+        /*stream_index=*/std::nullopt);
 
     for (double pts : ptsList) {
       seekFrameOp.call(decoderTensor, pts);

examples/basic_example.py

@@ -171,14 +171,3 @@ def plot(frames: torch.Tensor, title : Optional[str] = None):
 # %%
 plot(frame_at_2_seconds.data, "Frame displayed at 2 seconds")
 plot(first_two_seconds.data, "Frames displayed during [0, 2) seconds")
-
-# %%
-# Using a CUDA GPU to accelerate decoding
-# ---------------------------------------
-#
-# If you have a CUDA GPU that has NVDEC, you can decode on the GPU.
-if torch.cuda.is_available():
-    cuda_decoder = SimpleVideoDecoder(raw_video_bytes, device="cuda:0")
-    cuda_frame = cuda_decoder.get_frame_displayed_at(seconds=2)
-    print(cuda_frame.data.device)  # should be cuda:0
-    plot(cuda_frame.data.to("cpu"), "Frame displayed at 2 seconds on CUDA")

setup.py

@@ -112,16 +112,12 @@ def _build_all_extensions_with_cmake(self):
         torch_dir = Path(torch.utils.cmake_prefix_path) / "Torch"
         cmake_build_type = os.environ.get("CMAKE_BUILD_TYPE", "Release")
         python_version = sys.version_info
-        enable_cuda = os.environ.get("ENABLE_CUDA", "")
-        enable_nvtx = os.environ.get("ENABLE_NVTX", "")
         cmake_args = [
             f"-DCMAKE_INSTALL_PREFIX={self._install_prefix}",
             f"-DTorch_DIR={torch_dir}",
             "-DCMAKE_VERBOSE_MAKEFILE=ON",
             f"-DCMAKE_BUILD_TYPE={cmake_build_type}",
             f"-DPYTHON_VERSION={python_version.major}.{python_version.minor}",
-            f"-DENABLE_CUDA={enable_cuda}",
-            f"-DENABLE_NVTX={enable_nvtx}",
         ]
 
         Path(self.build_temp).mkdir(parents=True, exist_ok=True)

src/torchcodec/_samplers/video_clip_sampler.py

@@ -31,7 +31,6 @@ class VideoTooShortException(Exception):
 @dataclass
 class DecoderArgs:
     num_threads: int = 0
-    device: torch.device = torch.device("cpu")
 
 
 @dataclass
@@ -164,7 +163,6 @@ def forward(self, video_data: Tensor) -> Union[List[Any]]:
             width=target_width,
             height=target_height,
             num_threads=self.decoder_args.num_threads,
-            device_string=str(self.decoder_args.device),
         )
 
         clips: List[Any] = []

src/torchcodec/decoders/_core/CMakeLists.txt

@@ -4,28 +4,6 @@ set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 find_package(Torch REQUIRED)
-
-if(ENABLE_CUDA)
-    find_package(CUDA REQUIRED)
-
-    if(ENABLE_NVTX)
-        # We only need CPM for NVTX:
-        # https://github.com/NVIDIA/NVTX#cmake
-        file(
-        DOWNLOAD
-        https://github.com/cpm-cmake/CPM.cmake/releases/download/v0.38.3/CPM.cmake
-        ${CMAKE_CURRENT_BINARY_DIR}/cmake/CPM.cmake
-        EXPECTED_HASH SHA256=cc155ce02e7945e7b8967ddfaff0b050e958a723ef7aad3766d368940cb15494
-        )
-        include(${CMAKE_CURRENT_BINARY_DIR}/cmake/CPM.cmake)
-        CPMAddPackage(
-            NAME NVTX
-            GITHUB_REPOSITORY NVIDIA/NVTX
-            GIT_TAG v3.1.0-c-cpp
-            GIT_SHALLOW TRUE)
-    endif()
-endif()
-
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")
 find_package(Python3 ${PYTHON_VERSION} EXACT COMPONENTS Development)
 
@@ -41,12 +19,6 @@ function(make_torchcodec_library library_name ffmpeg_target)
     )
     add_library(${library_name} SHARED ${sources})
     set_property(TARGET ${library_name} PROPERTY CXX_STANDARD 17)
-    if(ENABLE_CUDA)
-        target_compile_definitions(${library_name} PRIVATE ENABLE_CUDA=1)
-    endif()
-    if(ENABLE_NVTX)
-        target_compile_definitions(${library_name} PRIVATE ENABLE_NVTX=1)
-    endif()
 
     target_include_directories(
         ${library_name}
@@ -56,17 +28,12 @@ function(make_torchcodec_library library_name ffmpeg_target)
         ${Python3_INCLUDE_DIRS}
     )
 
-    set(NEEDED_LIBRARIES ${ffmpeg_target} ${TORCH_LIBRARIES} ${Python3_LIBRARIES})
-    if(ENABLE_CUDA)
-        list(APPEND NEEDED_LIBRARIES ${CUDA_CUDA_LIBRARY} ${CUDA_nppi_LIBRARY} ${CUDA_nppicc_LIBRARY} )
-    endif()
-    if(ENABLE_NVTX)
-        list(APPEND NEEDED_LIBRARIES nvtx3-cpp)
-    endif()
     target_link_libraries(
         ${library_name}
         PUBLIC
-        ${NEEDED_LIBRARIES}
+        ${ffmpeg_target}
+        ${TORCH_LIBRARIES}
+        ${Python3_LIBRARIES}
     )
 
     # We already set the library_name to be libtorchcodecN, so we don't want

src/torchcodec/decoders/_core/FFMPEGCommon.h

@@ -57,8 +57,6 @@ using UniqueAVFilterInOut = std::unique_ptr<
     Deleterp<AVFilterInOut, void, avfilter_inout_free>>;
 using UniqueAVIOContext = std::
     unique_ptr<AVIOContext, Deleterp<AVIOContext, void, avio_context_free>>;
-using UniqueAVBufferRef =
-    std::unique_ptr<AVBufferRef, Deleterp<AVBufferRef, void, av_buffer_unref>>;
 
 // av_find_best_stream is not const-correct before commit:
 // https://github.com/FFmpeg/FFmpeg/commit/46dac8cf3d250184ab4247809bc03f60e14f4c0c