torchcodec 0.7.0__cp312-cp312-win_amd64.whl

torchcodec/_core/CMakeLists.txt

@@ -0,0 +1,292 @@
+cmake_minimum_required(VERSION 3.18)
+project(TorchCodec)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+set(PYBIND11_FINDPYTHON ON)
+find_package(pybind11 REQUIRED)
+find_package(Torch REQUIRED)
+find_package(Python3 ${PYTHON_VERSION} EXACT COMPONENTS Development)
+
+if(DEFINED TORCHCODEC_DISABLE_COMPILE_WARNING_AS_ERROR AND TORCHCODEC_DISABLE_COMPILE_WARNING_AS_ERROR)
+    set(TORCHCODEC_WERROR_OPTION "")
+else()
+    if (WIN32)
+        # TODO set warnings as errors on Windows as well.
+        # set(TORCHCODEC_WERROR_OPTION "/WX")
+    else()
+        set(TORCHCODEC_WERROR_OPTION "-Werror")
+    endif()
+endif()
+
+if (WIN32)
+  # Avoid warnings about non-ASCII characters in source files.
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4819")
+  # Important for when we add Windows CUDA: exporting all symbols is limited to
+  # 65535 symbols, which (apparently) will not work for CUDA.
+  # https://github.com/pytorch/pytorch/pull/3650
+  set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+endif()
+
+if (WIN32)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /W4 ${TORCHCODEC_WERROR_OPTION} ${TORCH_CXX_FLAGS}")
+else()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -pedantic ${TORCHCODEC_WERROR_OPTION} ${TORCH_CXX_FLAGS}")
+endif()
+
+
+function(make_torchcodec_sublibrary
+    library_name
+    type
+    sources
+    library_dependencies)
+
+    add_library(${library_name} ${type} ${sources})
+    set_target_properties(${library_name} PROPERTIES CXX_STANDARD 17)
+    target_include_directories(${library_name}
+        PRIVATE
+        ./../../../
+        "${TORCH_INSTALL_PREFIX}/include"
+        ${Python3_INCLUDE_DIRS}
+    )
+
+    # Avoid adding the "lib" prefix which we already add explicitly.
+    set_target_properties(${library_name} PROPERTIES PREFIX "")
+
+    target_link_libraries(
+        ${library_name}
+        PUBLIC
+        ${library_dependencies}
+    )
+
+endfunction()
+
+function(make_torchcodec_libraries
+    ffmpeg_major_version
+    ffmpeg_target)
+
+    # We create three shared libraries per version of FFmpeg, where the version
+    # is denoted by N:
+    #
+    # 1. libtorchcodec_coreN.{ext}: Base library which contains the
+    #    implementation of VideoDecoder and everything VideoDecoder needs. On
+    #    Linux, {ext} is so. On Mac, it is dylib. On Windows it's dll.
+    #
+    # 2. libtorchcodec_custom_opsN.{ext}: Implementation of the PyTorch custom
+    #    ops. Depends on libtorchcodec_coreN.{ext}. On Linux, {ext} is so.
+    #    On Mac, it is dylib. On Windows it's dll.
+    #
+    # 3. libtorchcodec_pybind_opsN.{ext}: Implementation of the pybind11 ops. We
+    #    keep these separate from the PyTorch custom ops because we have to
+    #    load these libraries separately on the Python side. Depends on
+    #    libtorchcodec_coreN.{ext}. On BOTH Linux and Mac {ext} is so. On
+    #    Windows, it's pyd.
+
+    # 1. Create libtorchcodec_coreN.{ext}.
+    set(core_library_name "libtorchcodec_core${ffmpeg_major_version}")
+    set(core_sources
+        AVIOContextHolder.cpp
+        AVIOTensorContext.cpp
+        FFMPEGCommon.cpp
+        FilterGraph.cpp
+        Frame.cpp
+        DeviceInterface.cpp
+        CpuDeviceInterface.cpp
+        SingleStreamDecoder.cpp
+        Encoder.cpp
+        ValidationUtils.cpp
+    )
+
+    if(ENABLE_CUDA)
+	    list(APPEND core_sources CudaDeviceInterface.cpp)
+    endif()
+
+    set(core_library_dependencies
+        ${ffmpeg_target}
+        ${TORCH_LIBRARIES}
+    )
+
+    if(ENABLE_CUDA)
+        list(APPEND core_library_dependencies
+            ${CUDA_nppi_LIBRARY}
+            ${CUDA_nppicc_LIBRARY}
+        )
+    endif()
+
+    make_torchcodec_sublibrary(
+        "${core_library_name}"
+        SHARED
+        "${core_sources}"
+        "${core_library_dependencies}"
+    )
+
+    # 2. Create libtorchcodec_custom_opsN.{ext}.
+    set(custom_ops_library_name "libtorchcodec_custom_ops${ffmpeg_major_version}")
+    set(custom_ops_sources
+        AVIOTensorContext.cpp
+        custom_ops.cpp
+    )
+    set(custom_ops_dependencies
+        ${core_library_name}
+        ${Python3_LIBRARIES}
+    )
+    make_torchcodec_sublibrary(
+        "${custom_ops_library_name}"
+        SHARED
+        "${custom_ops_sources}"
+        "${custom_ops_dependencies}"
+    )
+
+    # 3. Create libtorchcodec_pybind_opsN.so.
+    set(pybind_ops_library_name "libtorchcodec_pybind_ops${ffmpeg_major_version}")
+    set(pybind_ops_sources
+        AVIOFileLikeContext.cpp
+        pybind_ops.cpp
+    )
+    set(pybind_ops_dependencies
+       ${core_library_name}
+       pybind11::module # This library dependency makes sure we have the right
+                        # Python libraries included as well as all of the right
+                        # settings so that we can successfully load the shared
+                        # library as a Python module on Mac. If we instead use
+                        # ${Python3_LIBRARIES}, it works on Linux but not on
+                        # Mac.
+    )
+    make_torchcodec_sublibrary(
+        "${pybind_ops_library_name}"
+        MODULE # Note that this not SHARED; otherwise we build the wrong kind
+               # of library on Mac. On Mac, SHARED becomes .dylib and MODULE becomes
+               # a .so. We want pybind11 libraries to become .so. If this is
+               # changed to SHARED, we will be able to succesfully compile a
+               # .dylib, but we will not be able to succesfully import that as
+               # a Python module on Mac.
+        "${pybind_ops_sources}"
+        "${pybind_ops_dependencies}"
+    )
+
+    if(WIN32)
+      # On Windows, we need to set the suffix to .pyd so that Python can
+      # import the shared library as a module. Just setting the MODULE type
+      # isn't enough.
+      set_target_properties(${pybind_ops_library_name} PROPERTIES SUFFIX ".pyd")
+    endif()
+
+    # pybind11 limits the visibility of symbols in the shared library to prevent
+    # stray initialization of py::objects. The rest of the object code must
+    # match. See:
+    #   https://pybind11.readthedocs.io/en/stable/faq.html#someclass-declared-with-greater-visibility-than-the-type-of-its-field-someclass-member-wattributes
+    if(NOT WIN32)
+        target_compile_options(
+            ${pybind_ops_library_name}
+            PUBLIC
+            "-fvisibility=hidden"
+        )
+    endif()
+
+    # The value we use here must match the value we return from
+    # _get_pybind_ops_module_name() on the Python side. If the values do not
+    # match, then we will be unable to import the C++ shared library as a
+    # Python module at runtime.
+    target_compile_definitions(
+        ${pybind_ops_library_name}
+        PRIVATE
+        PYBIND_OPS_MODULE_NAME=core_pybind_ops
+    )
+
+    if(APPLE)
+        # If we don't make sure this flag is set, we run into segfauls at import
+        # time on Mac. See:
+        # https://github.com/pybind/pybind11/issues/3907#issuecomment-1170412764
+        target_link_options(
+            ${pybind_ops_library_name}
+            PUBLIC
+            "LINKER:-undefined,dynamic_lookup"
+        )
+    endif()
+
+    # Install all libraries.
+    set(
+        all_libraries
+        ${core_library_name}
+        ${custom_ops_library_name}
+        ${pybind_ops_library_name}
+    )
+
+    # The install step is invoked within CMakeBuild.build_library() in
+    # setup.py and just copies the built files from the temp
+    # cmake/setuptools build folder into the CMAKE_INSTALL_PREFIX folder. We
+    # still need to manually pass "DESTINATION ..." for cmake to copy those
+    # files in CMAKE_INSTALL_PREFIX instead of CMAKE_INSTALL_PREFIX/lib.
+    install(
+        TARGETS ${all_libraries}
+        LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX}
+        RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}  # For Windows
+    )
+
+endfunction()
+
+if(DEFINED ENV{BUILD_AGAINST_ALL_FFMPEG_FROM_S3})
+    message(
+        STATUS
+        "Building and dynamically linking libtorchcodec against our pre-built
+        non-GPL FFmpeg libraries. These libraries are only used at build time,
+        you still need a different FFmpeg to be installed for run time!"
+    )
+
+    # This will expose the ffmpeg4, ffmpeg5, ffmpeg6, and ffmpeg7 targets
+    include(
+        ${CMAKE_CURRENT_SOURCE_DIR}/fetch_and_expose_non_gpl_ffmpeg_libs.cmake
+    )
+
+    make_torchcodec_libraries(7 ffmpeg7)
+    make_torchcodec_libraries(6 ffmpeg6)
+    make_torchcodec_libraries(4 ffmpeg4)
+    make_torchcodec_libraries(5 ffmpeg5)
+else()
+    message(
+        STATUS
+        "Building and dynamically linking libtorchcodec against the installed
+        FFmpeg libraries. This require pkg-config to be installed. If you have
+        installed FFmpeg from conda, make sure pkg-config is installed from
+        conda as well."
+    )
+    find_package(PkgConfig REQUIRED)
+    pkg_check_modules(LIBAV REQUIRED IMPORTED_TARGET
+        libavdevice
+        libavfilter
+        libavformat
+        libavcodec
+        libavutil
+        libswresample
+        libswscale
+    )
+
+    # Split libavcodec's version string by '.' and convert it to a list
+    string(REPLACE "." ";" libavcodec_version_list ${LIBAV_libavcodec_VERSION})
+    # Get the first element of the list, which is the major version
+    list(GET libavcodec_version_list 0 libavcodec_major_version)
+
+    if (${libavcodec_major_version} STREQUAL "58")
+        set(ffmpeg_major_version "4")
+    elseif (${libavcodec_major_version} STREQUAL "59")
+        set(ffmpeg_major_version "5")
+    elseif (${libavcodec_major_version} STREQUAL "60")
+        set(ffmpeg_major_version "6")
+    elseif (${libavcodec_major_version} STREQUAL "61")
+        set(ffmpeg_major_version "7")
+    else()
+        message(
+            FATAL_ERROR
+            "Unsupported libavcodec version: ${libavcodec_major_version}"
+        )
+    endif()
+
+    make_torchcodec_libraries(${ffmpeg_major_version} PkgConfig::LIBAV)
+
+    # Expose these values updwards so that the test compilation does not need
+    # to re-figure it out. FIXME: it's not great that we just copy-paste the
+    # library names.
+    set(libtorchcodec_library_name "libtorchcodec_core${ffmpeg_major_version}" PARENT_SCOPE)
+    set(libtorchcodec_custom_ops_name "libtorchcodec_custom_ops${ffmpeg_major_version}" PARENT_SCOPE)
+    set(libav_include_dirs ${LIBAV_INCLUDE_DIRS} PARENT_SCOPE)
+endif()

torchcodec/_core/Cache.h

@@ -0,0 +1,138 @@
+// 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 <torch/types.h>
+#include <memory>
+#include <mutex>
+
+namespace facebook::torchcodec {
+
+// This header defines simple cache class primitives to store reusable objects
+// across TorchCodec stream instances. Intended usage is to store hardware
+// contexts creation of which is expensive. The cache mechanism is as follows:
+// 1. 'PerGpuCache' provides a dynamic cache with the specified maximum capacity
+//    for the given number of GPUs.
+// 2. When stream object (e.g. SingleStreamDecoder) is destoyed cachable object
+//    must be released to the cache. Cache will accept the object if it is not
+//    full.
+// 3. When stream object (e.g. SingleStreamDecoder) is created cachable object
+//    must be first queried from the cache. If the cache is empty then new
+//    object must be created.
+
+template <typename T, typename D = std::default_delete<T>>
+class Cache {
+ public:
+  using element_type = std::unique_ptr<T, D>;
+
+  explicit Cache(int capacity) : capacity_(capacity) {}
+
+  // Adds an object to the cache if the cache has capacity. Returns true
+  // if object was added and false otherwise.
+  bool addIfCacheHasCapacity(element_type&& obj);
+
+  // Returns an object from the cache. Cache does not hold a reference
+  // to the object after this call.
+  element_type get();
+
+ private:
+  int capacity_;
+  std::mutex mutex_;
+  std::vector<element_type> cache_;
+};
+
+template <typename T, typename D>
+bool Cache<T, D>::addIfCacheHasCapacity(element_type&& obj) {
+  std::scoped_lock lock(mutex_);
+  if (capacity_ >= 0 && cache_.size() >= static_cast<size_t>(capacity_)) {
+    return false;
+  }
+  cache_.push_back(std::move(obj));
+  return true;
+}
+
+template <typename T, typename D>
+typename Cache<T, D>::element_type Cache<T, D>::get() {
+  std::scoped_lock lock(mutex_);
+  if (cache_.empty()) {
+    return nullptr;
+  }
+
+  element_type obj = std::move(cache_.back());
+  cache_.pop_back();
+  return obj;
+}
+
+template <typename T, typename D = std::default_delete<T>>
+class PerGpuCache {
+ public:
+  using element_type = typename Cache<T, D>::element_type;
+
+  // Initializes 'maxGpus' number of caches. Each cache can hold no
+  // more than 'capacity' items. If 'capacity' <0 cache size is unlimited.
+  PerGpuCache(int maxGpus, int capacity) {
+    TORCH_CHECK(maxGpus > 0, "maxGpus for PerGpuCache must be >0");
+    for (int i = 0; i < maxGpus; ++i) {
+      cache_.emplace_back(std::make_unique<Cache<T, D>>(capacity));
+    }
+  }
+
+  // Adds an object to the specified device cache if the cache has
+  // capacity. Returns true if object was added and false otherwise.
+  bool addIfCacheHasCapacity(const torch::Device& device, element_type&& obj);
+
+  // Returns an object from the cache of the specified device. Cache
+  // does not hold a reference to the object after this call.
+  element_type get(const torch::Device& device);
+
+ private:
+  // 'Cache' class implementation contains mutex which makes it non-movable
+  // and non-copyable, so we need to wrap it in std::unique_ptr.
+  std::vector<std::unique_ptr<Cache<T, D>>> cache_;
+};
+
+// Note: this function is inline for convenience, not performance. Because the
+// rest of this file is template functions, they must all be defined in this
+// header. This function is not a template function, and should, in principle,
+// be defined in a .cpp file to preserve the One Definition Rule. That's
+// annoying for such a small amount of code, so we just inline it. If this file
+// grows, and there are more such functions, we should break them out into a
+// .cpp file.
+inline torch::DeviceIndex getNonNegativeDeviceIndex(
+    const torch::Device& device) {
+  torch::DeviceIndex deviceIndex = device.index();
+  // For single GPU machines libtorch returns -1 for the device index. So for
+  // that case we set the device index to 0. That's used in per-gpu cache
+  // implementation and during initialization of CUDA and FFmpeg contexts
+  // which require non negative indices.
+  deviceIndex = std::max<at::DeviceIndex>(deviceIndex, 0);
+  TORCH_CHECK(deviceIndex >= 0, "Device index out of range");
+  return deviceIndex;
+}
+
+template <typename T, typename D>
+bool PerGpuCache<T, D>::addIfCacheHasCapacity(
+    const torch::Device& device,
+    element_type&& obj) {
+  torch::DeviceIndex deviceIndex = getNonNegativeDeviceIndex(device);
+  TORCH_CHECK(
+      static_cast<size_t>(deviceIndex) < cache_.size(),
+      "Device index out of range");
+  return cache_[deviceIndex]->addIfCacheHasCapacity(std::move(obj));
+}
+
+template <typename T, typename D>
+typename PerGpuCache<T, D>::element_type PerGpuCache<T, D>::get(
+    const torch::Device& device) {
+  torch::DeviceIndex deviceIndex = getNonNegativeDeviceIndex(device);
+  TORCH_CHECK(
+      static_cast<size_t>(deviceIndex) < cache_.size(),
+      "Device index out of range");
+  return cache_[deviceIndex]->get();
+}
+
+} // namespace facebook::torchcodec

torchcodec/_core/CpuDeviceInterface.cpp

@@ -0,0 +1,266 @@
+// 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.
+
+#include "src/torchcodec/_core/CpuDeviceInterface.h"
+
+namespace facebook::torchcodec {
+namespace {
+
+static bool g_cpu = registerDeviceInterface(
+    torch::kCPU,
+    [](const torch::Device& device) { return new CpuDeviceInterface(device); });
+
+} // namespace
+
+bool CpuDeviceInterface::SwsFrameContext::operator==(
+    const CpuDeviceInterface::SwsFrameContext& other) const {
+  return inputWidth == other.inputWidth && inputHeight == other.inputHeight &&
+      inputFormat == other.inputFormat && outputWidth == other.outputWidth &&
+      outputHeight == other.outputHeight;
+}
+
+bool CpuDeviceInterface::SwsFrameContext::operator!=(
+    const CpuDeviceInterface::SwsFrameContext& other) const {
+  return !(*this == other);
+}
+
+CpuDeviceInterface::CpuDeviceInterface(const torch::Device& device)
+    : DeviceInterface(device) {
+  TORCH_CHECK(g_cpu, "CpuDeviceInterface was not registered!");
+  TORCH_CHECK(
+      device_.type() == torch::kCPU, "Unsupported device: ", device_.str());
+}
+
+// Note [preAllocatedOutputTensor with swscale and filtergraph]:
+// Callers may pass a pre-allocated tensor, where the output.data tensor will
+// be stored. This parameter is honored in any case, but it only leads to a
+// speed-up when swscale is used. With swscale, we can tell ffmpeg to place the
+// decoded frame directly into `preAllocatedtensor.data_ptr()`. We haven't yet
+// found a way to do that with filtegraph.
+// TODO: Figure out whether that's possible!
+// Dimension order of the preAllocatedOutputTensor must be HWC, regardless of
+// `dimension_order` parameter. It's up to callers to re-shape it if needed.
+void CpuDeviceInterface::convertAVFrameToFrameOutput(
+    const VideoStreamOptions& videoStreamOptions,
+    const AVRational& timeBase,
+    UniqueAVFrame& avFrame,
+    FrameOutput& frameOutput,
+    std::optional<torch::Tensor> preAllocatedOutputTensor) {
+  auto frameDims =
+      getHeightAndWidthFromOptionsOrAVFrame(videoStreamOptions, avFrame);
+  int expectedOutputHeight = frameDims.height;
+  int expectedOutputWidth = frameDims.width;
+
+  if (preAllocatedOutputTensor.has_value()) {
+    auto shape = preAllocatedOutputTensor.value().sizes();
+    TORCH_CHECK(
+        (shape.size() == 3) && (shape[0] == expectedOutputHeight) &&
+            (shape[1] == expectedOutputWidth) && (shape[2] == 3),
+        "Expected pre-allocated tensor of shape ",
+        expectedOutputHeight,
+        "x",
+        expectedOutputWidth,
+        "x3, got ",
+        shape);
+  }
+
+  torch::Tensor outputTensor;
+  enum AVPixelFormat frameFormat =
+      static_cast<enum AVPixelFormat>(avFrame->format);
+
+  // By default, we want to use swscale for color conversion because it is
+  // faster. However, it has width requirements, so we may need to fall back
+  // to filtergraph. We also need to respect what was requested from the
+  // options; we respect the options unconditionally, so it's possible for
+  // swscale's width requirements to be violated. We don't expose the ability to
+  // choose color conversion library publicly; we only use this ability
+  // internally.
+
+  // swscale requires widths to be multiples of 32:
+  // https://stackoverflow.com/questions/74351955/turn-off-sw-scale-conversion-to-planar-yuv-32-byte-alignment-requirements
+  // so we fall back to filtergraph if the width is not a multiple of 32.
+  auto defaultLibrary = (expectedOutputWidth % 32 == 0)
+      ? ColorConversionLibrary::SWSCALE
+      : ColorConversionLibrary::FILTERGRAPH;
+
+  ColorConversionLibrary colorConversionLibrary =
+      videoStreamOptions.colorConversionLibrary.value_or(defaultLibrary);
+
+  if (colorConversionLibrary == ColorConversionLibrary::SWSCALE) {
+    // We need to compare the current frame context with our previous frame
+    // context. If they are different, then we need to re-create our colorspace
+    // conversion objects. We create our colorspace conversion objects late so
+    // that we don't have to depend on the unreliable metadata in the header.
+    // And we sometimes re-create them because it's possible for frame
+    // resolution to change mid-stream. Finally, we want to reuse the colorspace
+    // conversion objects as much as possible for performance reasons.
+    SwsFrameContext swsFrameContext;
+
+    swsFrameContext.inputWidth = avFrame->width;
+    swsFrameContext.inputHeight = avFrame->height;
+    swsFrameContext.inputFormat = frameFormat;
+    swsFrameContext.outputWidth = expectedOutputWidth;
+    swsFrameContext.outputHeight = expectedOutputHeight;
+
+    outputTensor = preAllocatedOutputTensor.value_or(allocateEmptyHWCTensor(
+        expectedOutputHeight, expectedOutputWidth, torch::kCPU));
+
+    if (!swsContext_ || prevSwsFrameContext_ != swsFrameContext) {
+      createSwsContext(swsFrameContext, avFrame->colorspace);
+      prevSwsFrameContext_ = swsFrameContext;
+    }
+    int resultHeight =
+        convertAVFrameToTensorUsingSwsScale(avFrame, outputTensor);
+    // If this check failed, it would mean that the frame wasn't reshaped to
+    // the expected height.
+    // TODO: Can we do the same check for width?
+    TORCH_CHECK(
+        resultHeight == expectedOutputHeight,
+        "resultHeight != expectedOutputHeight: ",
+        resultHeight,
+        " != ",
+        expectedOutputHeight);
+
+    frameOutput.data = outputTensor;
+  } else if (colorConversionLibrary == ColorConversionLibrary::FILTERGRAPH) {
+    // See comment above in swscale branch about the filterGraphContext_
+    // creation. creation
+    FiltersContext filtersContext;
+
+    filtersContext.inputWidth = avFrame->width;
+    filtersContext.inputHeight = avFrame->height;
+    filtersContext.inputFormat = frameFormat;
+    filtersContext.inputAspectRatio = avFrame->sample_aspect_ratio;
+    filtersContext.outputWidth = expectedOutputWidth;
+    filtersContext.outputHeight = expectedOutputHeight;
+    filtersContext.outputFormat = AV_PIX_FMT_RGB24;
+    filtersContext.timeBase = timeBase;
+
+    std::stringstream filters;
+    filters << "scale=" << expectedOutputWidth << ":" << expectedOutputHeight;
+    filters << ":sws_flags=bilinear";
+
+    filtersContext.filtergraphStr = filters.str();
+
+    if (!filterGraphContext_ || prevFiltersContext_ != filtersContext) {
+      filterGraphContext_ =
+          std::make_unique<FilterGraph>(filtersContext, videoStreamOptions);
+      prevFiltersContext_ = std::move(filtersContext);
+    }
+    outputTensor = convertAVFrameToTensorUsingFilterGraph(avFrame);
+
+    // Similarly to above, if this check fails it means the frame wasn't
+    // reshaped to its expected dimensions by filtergraph.
+    auto shape = outputTensor.sizes();
+    TORCH_CHECK(
+        (shape.size() == 3) && (shape[0] == expectedOutputHeight) &&
+            (shape[1] == expectedOutputWidth) && (shape[2] == 3),
+        "Expected output tensor of shape ",
+        expectedOutputHeight,
+        "x",
+        expectedOutputWidth,
+        "x3, got ",
+        shape);
+
+    if (preAllocatedOutputTensor.has_value()) {
+      // We have already validated that preAllocatedOutputTensor and
+      // outputTensor have the same shape.
+      preAllocatedOutputTensor.value().copy_(outputTensor);
+      frameOutput.data = preAllocatedOutputTensor.value();
+    } else {
+      frameOutput.data = outputTensor;
+    }
+  } else {
+    TORCH_CHECK(
+        false,
+        "Invalid color conversion library: ",
+        static_cast<int>(colorConversionLibrary));
+  }
+}
+
+int CpuDeviceInterface::convertAVFrameToTensorUsingSwsScale(
+    const UniqueAVFrame& avFrame,
+    torch::Tensor& outputTensor) {
+  uint8_t* pointers[4] = {
+      outputTensor.data_ptr<uint8_t>(), nullptr, nullptr, nullptr};
+  int expectedOutputWidth = outputTensor.sizes()[1];
+  int linesizes[4] = {expectedOutputWidth * 3, 0, 0, 0};
+  int resultHeight = sws_scale(
+      swsContext_.get(),
+      avFrame->data,
+      avFrame->linesize,
+      0,
+      avFrame->height,
+      pointers,
+      linesizes);
+  return resultHeight;
+}
+
+torch::Tensor CpuDeviceInterface::convertAVFrameToTensorUsingFilterGraph(
+    const UniqueAVFrame& avFrame) {
+  UniqueAVFrame filteredAVFrame = filterGraphContext_->convert(avFrame);
+
+  TORCH_CHECK_EQ(filteredAVFrame->format, AV_PIX_FMT_RGB24);
+
+  auto frameDims = getHeightAndWidthFromResizedAVFrame(*filteredAVFrame.get());
+  int height = frameDims.height;
+  int width = frameDims.width;
+  std::vector<int64_t> shape = {height, width, 3};
+  std::vector<int64_t> strides = {filteredAVFrame->linesize[0], 3, 1};
+  AVFrame* filteredAVFramePtr = filteredAVFrame.release();
+  auto deleter = [filteredAVFramePtr](void*) {
+    UniqueAVFrame avFrameToDelete(filteredAVFramePtr);
+  };
+  return torch::from_blob(
+      filteredAVFramePtr->data[0], shape, strides, deleter, {torch::kUInt8});
+}
+
+void CpuDeviceInterface::createSwsContext(
+    const SwsFrameContext& swsFrameContext,
+    const enum AVColorSpace colorspace) {
+  SwsContext* swsContext = sws_getContext(
+      swsFrameContext.inputWidth,
+      swsFrameContext.inputHeight,
+      swsFrameContext.inputFormat,
+      swsFrameContext.outputWidth,
+      swsFrameContext.outputHeight,
+      AV_PIX_FMT_RGB24,
+      SWS_BILINEAR,
+      nullptr,
+      nullptr,
+      nullptr);
+  TORCH_CHECK(swsContext, "sws_getContext() returned nullptr");
+
+  int* invTable = nullptr;
+  int* table = nullptr;
+  int srcRange, dstRange, brightness, contrast, saturation;
+  int ret = sws_getColorspaceDetails(
+      swsContext,
+      &invTable,
+      &srcRange,
+      &table,
+      &dstRange,
+      &brightness,
+      &contrast,
+      &saturation);
+  TORCH_CHECK(ret != -1, "sws_getColorspaceDetails returned -1");
+
+  const int* colorspaceTable = sws_getCoefficients(colorspace);
+  ret = sws_setColorspaceDetails(
+      swsContext,
+      colorspaceTable,
+      srcRange,
+      colorspaceTable,
+      dstRange,
+      brightness,
+      contrast,
+      saturation);
+  TORCH_CHECK(ret != -1, "sws_setColorspaceDetails returned -1");
+
+  swsContext_.reset(swsContext);
+}
+
+} // namespace facebook::torchcodec