Merge pull request #6846 from ameerj/nvdec-gpu-decode

nvdec: Add GPU video decoding for all capable drivers and platforms

CMakeLists.txt

@@ -518,6 +518,10 @@ set(FFmpeg_COMPONENTS
     avutil
     swscale)
 
+if (${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
+    Include(FindPkgConfig REQUIRED)
+    pkg_check_modules(LIBVA libva)
+endif()
 if (NOT YUZU_USE_BUNDLED_FFMPEG)
     # Use system installed FFmpeg
     find_package(FFmpeg QUIET COMPONENTS ${FFmpeg_COMPONENTS})
@@ -540,6 +544,9 @@ endif()
 
 if (YUZU_USE_BUNDLED_FFMPEG)
     if (NOT WIN32)
+        # TODO(lat9nq): Move this to externals/ffmpeg/CMakeLists.txt (and move externals/ffmpeg to
+        # externals/ffmpeg/ffmpeg)
+
         # Build FFmpeg from externals
         message(STATUS "Using FFmpeg from externals")
 
@@ -579,20 +586,23 @@ if (YUZU_USE_BUNDLED_FFMPEG)
                 CACHE PATH "Paths to FFmpeg libraries" FORCE)
         endforeach()
 
-        set(FFmpeg_INCLUDE_DIR
-            "${FFmpeg_PREFIX};${FFmpeg_BUILD_DIR}"
-            CACHE PATH "Path to FFmpeg headers" FORCE)
-
-        if (${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
         Include(FindPkgConfig REQUIRED)
         pkg_check_modules(LIBVA libva)
-        endif()
+        pkg_check_modules(CUDA cuda)
+        pkg_check_modules(FFNVCODEC ffnvcodec)
+        pkg_check_modules(VDPAU vdpau)
+
+        set(FFmpeg_HWACCEL_LIBRARIES)
+        set(FFmpeg_HWACCEL_FLAGS)
+        set(FFmpeg_HWACCEL_INCLUDE_DIRS)
+        set(FFmpeg_HWACCEL_LDFLAGS)
+
         if(LIBVA_FOUND)
             pkg_check_modules(LIBDRM libdrm REQUIRED)
             find_package(X11 REQUIRED)
             pkg_check_modules(LIBVA-DRM libva-drm REQUIRED)
             pkg_check_modules(LIBVA-X11 libva-x11 REQUIRED)
-            set(FFmpeg_LIBVA_LIBRARIES
+            list(APPEND FFmpeg_HWACCEL_LIBRARIES
                 ${LIBDRM_LIBRARIES}
                 ${X11_LIBRARIES}
                 ${LIBVA-DRM_LIBRARIES}
@@ -602,11 +612,56 @@ if (YUZU_USE_BUNDLED_FFMPEG)
                 --enable-hwaccel=h264_vaapi
                 --enable-hwaccel=vp9_vaapi
                 --enable-libdrm)
+            list(APPEND FFmpeg_HWACCEL_INCLUDE_DIRS
+                ${LIBDRM_INCLUDE_DIRS}
+                ${X11_INCLUDE_DIRS}
+                ${LIBVA-DRM_INCLUDE_DIRS}
+                ${LIBVA-X11_INCLUDE_DIRS}
+                ${LIBVA_INCLUDE_DIRS}
+            )
             message(STATUS "VA-API found")
         else()
             set(FFmpeg_HWACCEL_FLAGS --disable-vaapi)
         endif()
 
+        if (FFNVCODEC_FOUND AND CUDA_FOUND)
+            list(APPEND FFmpeg_HWACCEL_FLAGS
+                --enable-cuvid
+                --enable-ffnvcodec
+                --enable-nvdec
+                --enable-hwaccel=h264_nvdec
+                --enable-hwaccel=vp9_nvdec
+                --extra-cflags=-I${CUDA_INCLUDE_DIRS}
+            )
+            list(APPEND FFmpeg_HWACCEL_LIBRARIES
+                ${FFNVCODEC_LIBRARIES}
+                ${CUDA_LIBRARIES}
+            )
+            list(APPEND FFmpeg_HWACCEL_INCLUDE_DIRS
+                ${FFNVCODEC_INCLUDE_DIRS}
+                ${CUDA_INCLUDE_DIRS}
+            )
+            list(APPEND FFmpeg_HWACCEL_LDFLAGS
+                ${FFNVCODEC_LDFLAGS}
+                ${CUDA_LDFLAGS}
+            )
+            message(STATUS "ffnvcodec libraries version ${FFNVCODEC_VERSION} found")
+        endif()
+
+        if (VDPAU_FOUND)
+            list(APPEND FFmpeg_HWACCEL_FLAGS
+                --enable-vdpau
+                --enable-hwaccel=h264_vdpau
+                --enable-hwaccel=vp9_vdpau
+            )
+            list(APPEND FFmpeg_HWACCEL_LIBRARIES ${VDPAU_LIBRARIES})
+            list(APPEND FFmpeg_HWACCEL_INCLUDE_DIRS ${VDPAU_INCLUDE_DIRS})
+            list(APPEND FFmpeg_HWACCEL_LDFLAGS ${VDPAU_LDFLAGS})
+            message(STATUS "vdpau libraries version ${VDPAU_VERSION} found")
+        else()
+            list(APPEND FFmpeg_HWACCEL_FLAGS --disable-vdpau)
+        endif()
+
         # `configure` parameters builds only exactly what yuzu needs from FFmpeg
         # `--disable-vdpau` is needed to avoid linking issues
         add_custom_command(
@@ -624,7 +679,6 @@ if (YUZU_USE_BUNDLED_FFMPEG)
                     --disable-network
                     --disable-postproc
                     --disable-swresample
-                    --disable-vdpau
                     --enable-decoder=h264
                     --enable-decoder=vp9
                     --cc="${CMAKE_C_COMPILER}"
@@ -653,15 +707,26 @@ if (YUZU_USE_BUNDLED_FFMPEG)
                 ${FFmpeg_BUILD_DIR}
         )
 
+        set(FFmpeg_INCLUDE_DIR
+            "${FFmpeg_PREFIX};${FFmpeg_BUILD_DIR};${FFmpeg_HWACCEL_INCLUDE_DIRS}"
+            CACHE PATH "Path to FFmpeg headers" FORCE)
+
+        set(FFmpeg_LDFLAGS
+            "${FFmpeg_HWACCEL_LDFLAGS}"
+            CACHE STRING "FFmpeg linker flags" FORCE)
+
         # ALL makes this custom target build every time
         # but it won't actually build if the DEPENDS parameter is up to date
         add_custom_target(ffmpeg-configure ALL DEPENDS ${FFmpeg_MAKEFILE})
         add_custom_target(ffmpeg-build ALL DEPENDS ${FFmpeg_BUILD_LIBRARIES} ffmpeg-configure)
         link_libraries(${FFmpeg_LIBVA_LIBRARIES})
-        set(FFmpeg_LIBRARIES ${FFmpeg_LIBVA_LIBRARIES} ${FFmpeg_BUILD_LIBRARIES}
+        set(FFmpeg_LIBRARIES ${FFmpeg_BUILD_LIBRARIES} ${FFmpeg_HWACCEL_LIBRARIES}
             CACHE PATH "Paths to FFmpeg libraries" FORCE)
         unset(FFmpeg_BUILD_LIBRARIES)
-        unset(FFmpeg_LIBVA_LIBRARIES)
+        unset(FFmpeg_HWACCEL_FLAGS)
+        unset(FFmpeg_HWACCEL_INCLUDE_DIRS)
+        unset(FFmpeg_HWACCEL_LDFLAGS)
+        unset(FFmpeg_HWACCEL_LIBRARIES)
 
         if (FFmpeg_FOUND)
             message(STATUS "Found FFmpeg version ${FFmpeg_VERSION}")
@@ -670,12 +735,13 @@ if (YUZU_USE_BUNDLED_FFMPEG)
         endif()
     else() # WIN32
         # Use yuzu FFmpeg binaries
-        set(FFmpeg_EXT_NAME "ffmpeg-4.3.1")
+        set(FFmpeg_EXT_NAME "ffmpeg-4.4")
         set(FFmpeg_PATH "${CMAKE_BINARY_DIR}/externals/${FFmpeg_EXT_NAME}")
         download_bundled_external("ffmpeg/" ${FFmpeg_EXT_NAME} "")
         set(FFmpeg_FOUND YES)
         set(FFmpeg_INCLUDE_DIR "${FFmpeg_PATH}/include" CACHE PATH "Path to FFmpeg headers" FORCE)
         set(FFmpeg_LIBRARY_DIR "${FFmpeg_PATH}/bin" CACHE PATH "Path to FFmpeg library directory" FORCE)
+        set(FFmpeg_LDFLAGS "" CACHE STRING "FFmpeg linker flags" FORCE)
         set(FFmpeg_DLL_DIR "${FFmpeg_PATH}/bin" CACHE PATH "Path to FFmpeg dll's" FORCE)
         set(FFmpeg_LIBRARIES
             ${FFmpeg_LIBRARY_DIR}/swscale.lib

src/common/settings.cpp

@@ -54,7 +54,7 @@ void LogSettings() {
     log_setting("Renderer_GPUAccuracyLevel", values.gpu_accuracy.GetValue());
     log_setting("Renderer_UseAsynchronousGpuEmulation",
                 values.use_asynchronous_gpu_emulation.GetValue());
-    log_setting("Renderer_UseNvdecEmulation", values.use_nvdec_emulation.GetValue());
+    log_setting("Renderer_NvdecEmulation", values.nvdec_emulation.GetValue());
     log_setting("Renderer_AccelerateASTC", values.accelerate_astc.GetValue());
     log_setting("Renderer_UseVsync", values.use_vsync.GetValue());
     log_setting("Renderer_ShaderBackend", values.shader_backend.GetValue());
@@ -136,7 +136,7 @@ void RestoreGlobalState(bool is_powered_on) {
     values.use_disk_shader_cache.SetGlobal(true);
     values.gpu_accuracy.SetGlobal(true);
     values.use_asynchronous_gpu_emulation.SetGlobal(true);
-    values.use_nvdec_emulation.SetGlobal(true);
+    values.nvdec_emulation.SetGlobal(true);
     values.accelerate_astc.SetGlobal(true);
     values.use_vsync.SetGlobal(true);
     values.shader_backend.SetGlobal(true);

src/common/settings.h

@@ -48,6 +48,12 @@ enum class FullscreenMode : u32 {
     Exclusive = 1,
 };
 
+enum class NvdecEmulation : u32 {
+    Off = 0,
+    CPU = 1,
+    GPU = 2,
+};
+
 /** The BasicSetting class is a simple resource manager. It defines a label and default value
  * alongside the actual value of the setting for simpler and less-error prone use with frontend
  * configurations. Setting a default value and label is required, though subclasses may deviate from
@@ -466,7 +472,7 @@ struct Values {
     RangedSetting<GPUAccuracy> gpu_accuracy{GPUAccuracy::High, GPUAccuracy::Normal,
                                             GPUAccuracy::Extreme, "gpu_accuracy"};
     Setting<bool> use_asynchronous_gpu_emulation{true, "use_asynchronous_gpu_emulation"};
-    Setting<bool> use_nvdec_emulation{true, "use_nvdec_emulation"};
+    Setting<NvdecEmulation> nvdec_emulation{NvdecEmulation::GPU, "nvdec_emulation"};
     Setting<bool> accelerate_astc{true, "accelerate_astc"};
     Setting<bool> use_vsync{true, "use_vsync"};
     BasicRangedSetting<u16> fps_cap{1000, 1, 1000, "fps_cap"};

src/core/telemetry_session.cpp

@@ -72,6 +72,18 @@ static const char* TranslateGPUAccuracyLevel(Settings::GPUAccuracy backend) {
     return "Unknown";
 }
 
+static const char* TranslateNvdecEmulation(Settings::NvdecEmulation backend) {
+    switch (backend) {
+    case Settings::NvdecEmulation::Off:
+        return "Off";
+    case Settings::NvdecEmulation::CPU:
+        return "CPU";
+    case Settings::NvdecEmulation::GPU:
+        return "GPU";
+    }
+    return "Unknown";
+}
+
 u64 GetTelemetryId() {
     u64 telemetry_id{};
     const auto filename = Common::FS::GetYuzuPath(Common::FS::YuzuPath::ConfigDir) / "telemetry_id";
@@ -229,8 +241,8 @@ void TelemetrySession::AddInitialInfo(Loader::AppLoader& app_loader,
              TranslateGPUAccuracyLevel(Settings::values.gpu_accuracy.GetValue()));
     AddField(field_type, "Renderer_UseAsynchronousGpuEmulation",
              Settings::values.use_asynchronous_gpu_emulation.GetValue());
-    AddField(field_type, "Renderer_UseNvdecEmulation",
+    AddField(field_type, "Renderer_NvdecEmulation",
-             Settings::values.use_nvdec_emulation.GetValue());
+             TranslateNvdecEmulation(Settings::values.nvdec_emulation.GetValue()));
     AddField(field_type, "Renderer_AccelerateASTC", Settings::values.accelerate_astc.GetValue());
     AddField(field_type, "Renderer_UseVsync", Settings::values.use_vsync.GetValue());
     AddField(field_type, "Renderer_ShaderBackend",

src/video_core/CMakeLists.txt

@@ -231,6 +231,7 @@ endif()
 
 target_include_directories(video_core PRIVATE ${FFmpeg_INCLUDE_DIR})
 target_link_libraries(video_core PRIVATE ${FFmpeg_LIBRARIES})
+target_link_options(video_core PRIVATE ${FFmpeg_LDFLAGS})
 
 add_dependencies(video_core host_shaders)
 target_include_directories(video_core PRIVATE ${HOST_SHADERS_INCLUDE})

src/video_core/command_classes/codecs/codec.cpp

@@ -5,6 +5,7 @@
 #include <fstream>
 #include <vector>
 #include "common/assert.h"
+#include "common/settings.h"
 #include "video_core/command_classes/codecs/codec.h"
 #include "video_core/command_classes/codecs/h264.h"
 #include "video_core/command_classes/codecs/vp9.h"
@@ -16,44 +17,28 @@ extern "C" {
 }
 
 namespace Tegra {
-#if defined(LIBVA_FOUND)
-// Hardware acceleration code from FFmpeg/doc/examples/hw_decode.c originally under MIT license
 namespace {
-constexpr std::array<const char*, 2> VAAPI_DRIVERS = {
+constexpr AVPixelFormat PREFERRED_GPU_FMT = AV_PIX_FMT_NV12;
-    "i915",
+constexpr AVPixelFormat PREFERRED_CPU_FMT = AV_PIX_FMT_YUV420P;
-    "amdgpu",
-};
 
-AVPixelFormat GetHwFormat(AVCodecContext*, const AVPixelFormat* pix_fmts) {
+void AVPacketDeleter(AVPacket* ptr) {
+    av_packet_free(&ptr);
+}
+
+using AVPacketPtr = std::unique_ptr<AVPacket, decltype(&AVPacketDeleter)>;
+
+AVPixelFormat GetGpuFormat(AVCodecContext* av_codec_ctx, const AVPixelFormat* pix_fmts) {
     for (const AVPixelFormat* p = pix_fmts; *p != AV_PIX_FMT_NONE; ++p) {
-        if (*p == AV_PIX_FMT_VAAPI) {
+        if (*p == av_codec_ctx->pix_fmt) {
-            return AV_PIX_FMT_VAAPI;
+            return av_codec_ctx->pix_fmt;
         }
     }
     LOG_INFO(Service_NVDRV, "Could not find compatible GPU AV format, falling back to CPU");
-    return *pix_fmts;
+    av_buffer_unref(&av_codec_ctx->hw_device_ctx);
-}
+    av_codec_ctx->pix_fmt = PREFERRED_CPU_FMT;
-
+    return PREFERRED_CPU_FMT;
-bool CreateVaapiHwdevice(AVBufferRef** av_hw_device) {
-    AVDictionary* hwdevice_options = nullptr;
-    av_dict_set(&hwdevice_options, "connection_type", "drm", 0);
-    for (const auto& driver : VAAPI_DRIVERS) {
-        av_dict_set(&hwdevice_options, "kernel_driver", driver, 0);
-        const int hwdevice_error = av_hwdevice_ctx_create(av_hw_device, AV_HWDEVICE_TYPE_VAAPI,
-                                                          nullptr, hwdevice_options, 0);
-        if (hwdevice_error >= 0) {
-            LOG_INFO(Service_NVDRV, "Using VA-API with {}", driver);
-            av_dict_free(&hwdevice_options);
-            return true;
-        }
-        LOG_DEBUG(Service_NVDRV, "VA-API av_hwdevice_ctx_create failed {}", hwdevice_error);
-    }
-    LOG_DEBUG(Service_NVDRV, "VA-API av_hwdevice_ctx_create failed for all drivers");
-    av_dict_free(&hwdevice_options);
-    return false;
 }
 } // namespace
-#endif
 
 void AVFrameDeleter(AVFrame* ptr) {
     av_frame_free(&ptr);
@@ -68,56 +53,110 @@ Codec::~Codec() {
         return;
     }
     // Free libav memory
-    avcodec_send_packet(av_codec_ctx, nullptr);
+    avcodec_free_context(&av_codec_ctx);
-    AVFrame* av_frame = av_frame_alloc();
+    av_buffer_unref(&av_gpu_decoder);
-    avcodec_receive_frame(av_codec_ctx, av_frame);
-    avcodec_flush_buffers(av_codec_ctx);
-    av_frame_free(&av_frame);
-    avcodec_close(av_codec_ctx);
-    av_buffer_unref(&av_hw_device);
 }
 
-void Codec::InitializeHwdec() {
+bool Codec::CreateGpuAvDevice() {
-    // Prioritize integrated GPU to mitigate bandwidth bottlenecks
 #if defined(LIBVA_FOUND)
-    if (CreateVaapiHwdevice(&av_hw_device)) {
+    static constexpr std::array<const char*, 3> VAAPI_DRIVERS = {
-        const auto hw_device_ctx = av_buffer_ref(av_hw_device);
+        "i915",
-        ASSERT_MSG(hw_device_ctx, "av_buffer_ref failed");
+        "iHD",
-        av_codec_ctx->hw_device_ctx = hw_device_ctx;
+        "amdgpu",
-        av_codec_ctx->get_format = GetHwFormat;
+    };
+    AVDictionary* hwdevice_options = nullptr;
+    av_dict_set(&hwdevice_options, "connection_type", "drm", 0);
+    for (const auto& driver : VAAPI_DRIVERS) {
+        av_dict_set(&hwdevice_options, "kernel_driver", driver, 0);
+        const int hwdevice_error = av_hwdevice_ctx_create(&av_gpu_decoder, AV_HWDEVICE_TYPE_VAAPI,
+                                                          nullptr, hwdevice_options, 0);
+        if (hwdevice_error >= 0) {
+            LOG_INFO(Service_NVDRV, "Using VA-API with {}", driver);
+            av_dict_free(&hwdevice_options);
+            av_codec_ctx->pix_fmt = AV_PIX_FMT_VAAPI;
+            return true;
+        }
+        LOG_DEBUG(Service_NVDRV, "VA-API av_hwdevice_ctx_create failed {}", hwdevice_error);
+    }
+    LOG_DEBUG(Service_NVDRV, "VA-API av_hwdevice_ctx_create failed for all drivers");
+    av_dict_free(&hwdevice_options);
+#endif
+    static constexpr auto HW_CONFIG_METHOD = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX;
+    static constexpr std::array GPU_DECODER_TYPES{
+        AV_HWDEVICE_TYPE_CUDA,
+#ifdef _WIN32
+        AV_HWDEVICE_TYPE_D3D11VA,
+#else
+        AV_HWDEVICE_TYPE_VDPAU,
+#endif
+    };
+    for (const auto& type : GPU_DECODER_TYPES) {
+        const int hwdevice_res = av_hwdevice_ctx_create(&av_gpu_decoder, type, nullptr, nullptr, 0);
+        if (hwdevice_res < 0) {
+            LOG_DEBUG(Service_NVDRV, "{} av_hwdevice_ctx_create failed {}",
+                      av_hwdevice_get_type_name(type), hwdevice_res);
+            continue;
+        }
+        for (int i = 0;; i++) {
+            const AVCodecHWConfig* config = avcodec_get_hw_config(av_codec, i);
+            if (!config) {
+                LOG_DEBUG(Service_NVDRV, "{} decoder does not support device type {}.",
+                          av_codec->name, av_hwdevice_get_type_name(type));
+                break;
+            }
+            if (config->methods & HW_CONFIG_METHOD && config->device_type == type) {
+                av_codec_ctx->pix_fmt = config->pix_fmt;
+                LOG_INFO(Service_NVDRV, "Using {} GPU decoder", av_hwdevice_get_type_name(type));
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+void Codec::InitializeAvCodecContext() {
+    av_codec_ctx = avcodec_alloc_context3(av_codec);
+    av_opt_set(av_codec_ctx->priv_data, "tune", "zerolatency", 0);
+}
+
+void Codec::InitializeGpuDecoder() {
+    if (!CreateGpuAvDevice()) {
+        av_buffer_unref(&av_gpu_decoder);
         return;
     }
-#endif
+    auto* hw_device_ctx = av_buffer_ref(av_gpu_decoder);
-    // TODO more GPU accelerated decoders
+    ASSERT_MSG(hw_device_ctx, "av_buffer_ref failed");
+    av_codec_ctx->hw_device_ctx = hw_device_ctx;
+    av_codec_ctx->get_format = GetGpuFormat;
 }
 
 void Codec::Initialize() {
-    AVCodecID codec;
+    const AVCodecID codec = [&] {
         switch (current_codec) {
         case NvdecCommon::VideoCodec::H264:
-        codec = AV_CODEC_ID_H264;
+            return AV_CODEC_ID_H264;
-        break;
         case NvdecCommon::VideoCodec::Vp9:
-        codec = AV_CODEC_ID_VP9;
+            return AV_CODEC_ID_VP9;
-        break;
         default:
             UNIMPLEMENTED_MSG("Unknown codec {}", current_codec);
+            return AV_CODEC_ID_NONE;
+        }
+    }();
+    av_codec = avcodec_find_decoder(codec);
+
+    InitializeAvCodecContext();
+    if (Settings::values.nvdec_emulation.GetValue() == Settings::NvdecEmulation::GPU) {
+        InitializeGpuDecoder();
+    }
+    if (const int res = avcodec_open2(av_codec_ctx, av_codec, nullptr); res < 0) {
+        LOG_ERROR(Service_NVDRV, "avcodec_open2() Failed with result {}", res);
+        avcodec_free_context(&av_codec_ctx);
+        av_buffer_unref(&av_gpu_decoder);
         return;
     }
-    av_codec = avcodec_find_decoder(codec);
-    av_codec_ctx = avcodec_alloc_context3(av_codec);
-    av_opt_set(av_codec_ctx->priv_data, "tune", "zerolatency", 0);
-    InitializeHwdec();
     if (!av_codec_ctx->hw_device_ctx) {
         LOG_INFO(Service_NVDRV, "Using FFmpeg software decoding");
     }
-    const auto av_error = avcodec_open2(av_codec_ctx, av_codec, nullptr);
-    if (av_error < 0) {
-        LOG_ERROR(Service_NVDRV, "avcodec_open2() Failed.");
-        avcodec_close(av_codec_ctx);
-        av_buffer_unref(&av_hw_device);
-        return;
-    }
     initialized = true;
 }
 
@@ -133,6 +172,9 @@ void Codec::Decode() {
     if (is_first_frame) {
         Initialize();
     }
+    if (!initialized) {
+        return;
+    }
     bool vp9_hidden_frame = false;
     std::vector<u8> frame_data;
     if (current_codec == NvdecCommon::VideoCodec::H264) {
@@ -141,50 +183,48 @@ void Codec::Decode() {
         frame_data = vp9_decoder->ComposeFrameHeader(state);
         vp9_hidden_frame = vp9_decoder->WasFrameHidden();
     }
-    AVPacket packet{};
+    AVPacketPtr packet{av_packet_alloc(), AVPacketDeleter};
-    av_init_packet(&packet);
+    if (!packet) {
-    packet.data = frame_data.data();
+        LOG_ERROR(Service_NVDRV, "av_packet_alloc failed");
-    packet.size = static_cast<s32>(frame_data.size());
+        return;
-    if (const int ret = avcodec_send_packet(av_codec_ctx, &packet); ret) {
+    }
-        LOG_DEBUG(Service_NVDRV, "avcodec_send_packet error {}", ret);
+    packet->data = frame_data.data();
+    packet->size = static_cast<s32>(frame_data.size());
+    if (const int res = avcodec_send_packet(av_codec_ctx, packet.get()); res != 0) {
+        LOG_DEBUG(Service_NVDRV, "avcodec_send_packet error {}", res);
         return;
     }
     // Only receive/store visible frames
     if (vp9_hidden_frame) {
         return;
     }
-    AVFrame* hw_frame = av_frame_alloc();
+    AVFramePtr initial_frame{av_frame_alloc(), AVFrameDeleter};
-    AVFrame* sw_frame = hw_frame;
+    AVFramePtr final_frame{nullptr, AVFrameDeleter};
-    ASSERT_MSG(hw_frame, "av_frame_alloc hw_frame failed");
+    ASSERT_MSG(initial_frame, "av_frame_alloc initial_frame failed");
-    if (const int ret = avcodec_receive_frame(av_codec_ctx, hw_frame); ret) {
+    if (const int ret = avcodec_receive_frame(av_codec_ctx, initial_frame.get()); ret) {
         LOG_DEBUG(Service_NVDRV, "avcodec_receive_frame error {}", ret);
-        av_frame_free(&hw_frame);
         return;
     }
-    if (!hw_frame->width || !hw_frame->height) {
+    if (initial_frame->width == 0 || initial_frame->height == 0) {
         LOG_WARNING(Service_NVDRV, "Zero width or height in frame");
-        av_frame_free(&hw_frame);
         return;
     }
-#if defined(LIBVA_FOUND)
+    if (av_codec_ctx->hw_device_ctx) {
-    // Hardware acceleration code from FFmpeg/doc/examples/hw_decode.c under MIT license
+        final_frame = AVFramePtr{av_frame_alloc(), AVFrameDeleter};
-    if (hw_frame->format == AV_PIX_FMT_VAAPI) {
+        ASSERT_MSG(final_frame, "av_frame_alloc final_frame failed");
-        sw_frame = av_frame_alloc();
-        ASSERT_MSG(sw_frame, "av_frame_alloc sw_frame failed");
         // Can't use AV_PIX_FMT_YUV420P and share code with software decoding in vic.cpp
         // because Intel drivers crash unless using AV_PIX_FMT_NV12
-        sw_frame->format = AV_PIX_FMT_NV12;
+        final_frame->format = PREFERRED_GPU_FMT;
-        const int transfer_data_ret = av_hwframe_transfer_data(sw_frame, hw_frame, 0);
+        const int ret = av_hwframe_transfer_data(final_frame.get(), initial_frame.get(), 0);
-        ASSERT_MSG(!transfer_data_ret, "av_hwframe_transfer_data error {}", transfer_data_ret);
+        ASSERT_MSG(!ret, "av_hwframe_transfer_data error {}", ret);
-        av_frame_free(&hw_frame);
+    } else {
+        final_frame = std::move(initial_frame);
     }
-#endif
+    if (final_frame->format != PREFERRED_CPU_FMT && final_frame->format != PREFERRED_GPU_FMT) {
-    if (sw_frame->format != AV_PIX_FMT_YUV420P && sw_frame->format != AV_PIX_FMT_NV12) {
+        UNIMPLEMENTED_MSG("Unexpected video format: {}", final_frame->format);
-        UNIMPLEMENTED_MSG("Unexpected video format from host graphics: {}", sw_frame->format);
-        av_frame_free(&sw_frame);
         return;
     }
-    av_frames.push(AVFramePtr{sw_frame, AVFrameDeleter});
+    av_frames.push(std::move(final_frame));
     if (av_frames.size() > 10) {
         LOG_TRACE(Service_NVDRV, "av_frames.push overflow dropped frame");
         av_frames.pop();

src/video_core/command_classes/codecs/codec.h

@@ -50,18 +50,23 @@ public:
 
     /// Returns the value of current_codec
     [[nodiscard]] NvdecCommon::VideoCodec GetCurrentCodec() const;
+
     /// Return name of the current codec
     [[nodiscard]] std::string_view GetCurrentCodecName() const;
 
 private:
-    void InitializeHwdec();
+    void InitializeAvCodecContext();
+
+    void InitializeGpuDecoder();
+
+    bool CreateGpuAvDevice();
 
     bool initialized{};
     NvdecCommon::VideoCodec current_codec{NvdecCommon::VideoCodec::None};
 
     AVCodec* av_codec{nullptr};
-    AVBufferRef* av_hw_device{nullptr};
     AVCodecContext* av_codec_ctx{nullptr};
+    AVBufferRef* av_gpu_decoder{nullptr};
 
     GPU& gpu;
     const NvdecCommon::NvdecRegisters& state;

src/video_core/command_classes/codecs/h264.cpp

@@ -95,7 +95,8 @@ const std::vector<u8>& H264::ComposeFrameHeader(const NvdecCommon::NvdecRegister
     const s32 pic_height = context.h264_parameter_set.frame_height_in_map_units /
                            (context.h264_parameter_set.frame_mbs_only_flag ? 1 : 2);
 
-    writer.WriteUe(16);
+    // TODO (ameerj): Where do we get this number, it seems to be particular for each stream
+    writer.WriteUe(6); // Max number of reference frames
     writer.WriteBit(false);
     writer.WriteUe(context.h264_parameter_set.pic_width_in_mbs - 1);
     writer.WriteUe(pic_height - 1);

src/video_core/video_core.cpp

@@ -37,7 +37,8 @@ std::unique_ptr<VideoCore::RendererBase> CreateRenderer(
 namespace VideoCore {
 
 std::unique_ptr<Tegra::GPU> CreateGPU(Core::Frontend::EmuWindow& emu_window, Core::System& system) {
-    const bool use_nvdec = Settings::values.use_nvdec_emulation.GetValue();
+    const auto nvdec_value = Settings::values.nvdec_emulation.GetValue();
+    const bool use_nvdec = nvdec_value != Settings::NvdecEmulation::Off;
     const bool use_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
     auto gpu = std::make_unique<Tegra::GPU>(system, use_async, use_nvdec);
     auto context = emu_window.CreateSharedContext();

src/yuzu/configuration/config.cpp

@@ -812,7 +812,7 @@ void Config::ReadRendererValues() {
     ReadGlobalSetting(Settings::values.use_disk_shader_cache);
     ReadGlobalSetting(Settings::values.gpu_accuracy);
     ReadGlobalSetting(Settings::values.use_asynchronous_gpu_emulation);
-    ReadGlobalSetting(Settings::values.use_nvdec_emulation);
+    ReadGlobalSetting(Settings::values.nvdec_emulation);
     ReadGlobalSetting(Settings::values.accelerate_astc);
     ReadGlobalSetting(Settings::values.use_vsync);
     ReadGlobalSetting(Settings::values.shader_backend);
@@ -1349,7 +1349,10 @@ void Config::SaveRendererValues() {
                  static_cast<u32>(Settings::values.gpu_accuracy.GetDefault()),
                  Settings::values.gpu_accuracy.UsingGlobal());
     WriteGlobalSetting(Settings::values.use_asynchronous_gpu_emulation);
-    WriteGlobalSetting(Settings::values.use_nvdec_emulation);
+    WriteSetting(QString::fromStdString(Settings::values.nvdec_emulation.GetLabel()),
+                 static_cast<u32>(Settings::values.nvdec_emulation.GetValue(global)),
+                 static_cast<u32>(Settings::values.nvdec_emulation.GetDefault()),
+                 Settings::values.nvdec_emulation.UsingGlobal());
     WriteGlobalSetting(Settings::values.accelerate_astc);
     WriteGlobalSetting(Settings::values.use_vsync);
     WriteSetting(QString::fromStdString(Settings::values.shader_backend.GetLabel()),

src/yuzu/configuration/config.h

@@ -182,5 +182,6 @@ private:
 Q_DECLARE_METATYPE(Settings::CPUAccuracy);
 Q_DECLARE_METATYPE(Settings::GPUAccuracy);
 Q_DECLARE_METATYPE(Settings::FullscreenMode);
+Q_DECLARE_METATYPE(Settings::NvdecEmulation);
 Q_DECLARE_METATYPE(Settings::RendererBackend);
 Q_DECLARE_METATYPE(Settings::ShaderBackend);

src/yuzu/configuration/configure_graphics.cpp

@@ -88,24 +88,30 @@ void ConfigureGraphics::SetConfiguration() {
     ui->api_widget->setEnabled(runtime_lock);
     ui->use_asynchronous_gpu_emulation->setEnabled(runtime_lock);
     ui->use_disk_shader_cache->setEnabled(runtime_lock);
-    ui->use_nvdec_emulation->setEnabled(runtime_lock);
+    ui->nvdec_emulation_widget->setEnabled(runtime_lock);
     ui->accelerate_astc->setEnabled(runtime_lock);
     ui->use_disk_shader_cache->setChecked(Settings::values.use_disk_shader_cache.GetValue());
     ui->use_asynchronous_gpu_emulation->setChecked(
         Settings::values.use_asynchronous_gpu_emulation.GetValue());
-    ui->use_nvdec_emulation->setChecked(Settings::values.use_nvdec_emulation.GetValue());
     ui->accelerate_astc->setChecked(Settings::values.accelerate_astc.GetValue());
 
     if (Settings::IsConfiguringGlobal()) {
         ui->api->setCurrentIndex(static_cast<int>(Settings::values.renderer_backend.GetValue()));
         ui->fullscreen_mode_combobox->setCurrentIndex(
             static_cast<int>(Settings::values.fullscreen_mode.GetValue()));
+        ui->nvdec_emulation->setCurrentIndex(
+            static_cast<int>(Settings::values.nvdec_emulation.GetValue()));
         ui->aspect_ratio_combobox->setCurrentIndex(Settings::values.aspect_ratio.GetValue());
     } else {
         ConfigurationShared::SetPerGameSetting(ui->api, &Settings::values.renderer_backend);
         ConfigurationShared::SetHighlight(ui->api_widget,
                                           !Settings::values.renderer_backend.UsingGlobal());
 
+        ConfigurationShared::SetPerGameSetting(ui->nvdec_emulation,
+                                               &Settings::values.nvdec_emulation);
+        ConfigurationShared::SetHighlight(ui->nvdec_emulation_widget,
+                                          !Settings::values.nvdec_emulation.UsingGlobal());
+
         ConfigurationShared::SetPerGameSetting(ui->fullscreen_mode_combobox,
                                                &Settings::values.fullscreen_mode);
         ConfigurationShared::SetHighlight(ui->fullscreen_mode_label,
@@ -137,8 +143,6 @@ void ConfigureGraphics::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_asynchronous_gpu_emulation,
                                              ui->use_asynchronous_gpu_emulation,
                                              use_asynchronous_gpu_emulation);
-    ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_nvdec_emulation,
-                                             ui->use_nvdec_emulation, use_nvdec_emulation);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.accelerate_astc, ui->accelerate_astc,
                                              accelerate_astc);
 
@@ -147,6 +151,9 @@ void ConfigureGraphics::ApplyConfiguration() {
         if (Settings::values.renderer_backend.UsingGlobal()) {
             Settings::values.renderer_backend.SetValue(GetCurrentGraphicsBackend());
         }
+        if (Settings::values.nvdec_emulation.UsingGlobal()) {
+            Settings::values.nvdec_emulation.SetValue(GetCurrentNvdecEmulation());
+        }
         if (Settings::values.shader_backend.UsingGlobal()) {
             Settings::values.shader_backend.SetValue(shader_backend);
         }
@@ -180,6 +187,13 @@ void ConfigureGraphics::ApplyConfiguration() {
             }
         }
 
+        if (ui->nvdec_emulation->currentIndex() == ConfigurationShared::USE_GLOBAL_INDEX) {
+            Settings::values.nvdec_emulation.SetGlobal(true);
+        } else {
+            Settings::values.nvdec_emulation.SetGlobal(false);
+            Settings::values.nvdec_emulation.SetValue(GetCurrentNvdecEmulation());
+        }
+
         if (ui->bg_combobox->currentIndex() == ConfigurationShared::USE_GLOBAL_INDEX) {
             Settings::values.bg_red.SetGlobal(true);
             Settings::values.bg_green.SetGlobal(true);
@@ -278,6 +292,20 @@ Settings::RendererBackend ConfigureGraphics::GetCurrentGraphicsBackend() const {
                                                   ConfigurationShared::USE_GLOBAL_OFFSET);
 }
 
+Settings::NvdecEmulation ConfigureGraphics::GetCurrentNvdecEmulation() const {
+    if (Settings::IsConfiguringGlobal()) {
+        return static_cast<Settings::NvdecEmulation>(ui->nvdec_emulation->currentIndex());
+    }
+
+    if (ui->nvdec_emulation->currentIndex() == ConfigurationShared::USE_GLOBAL_INDEX) {
+        Settings::values.nvdec_emulation.SetGlobal(true);
+        return Settings::values.nvdec_emulation.GetValue();
+    }
+    Settings::values.nvdec_emulation.SetGlobal(false);
+    return static_cast<Settings::NvdecEmulation>(ui->nvdec_emulation->currentIndex() -
+                                                 ConfigurationShared::USE_GLOBAL_OFFSET);
+}
+
 void ConfigureGraphics::SetupPerGameUI() {
     if (Settings::IsConfiguringGlobal()) {
         ui->api->setEnabled(Settings::values.renderer_backend.UsingGlobal());
@@ -286,7 +314,7 @@ void ConfigureGraphics::SetupPerGameUI() {
         ui->aspect_ratio_combobox->setEnabled(Settings::values.aspect_ratio.UsingGlobal());
         ui->use_asynchronous_gpu_emulation->setEnabled(
             Settings::values.use_asynchronous_gpu_emulation.UsingGlobal());
-        ui->use_nvdec_emulation->setEnabled(Settings::values.use_nvdec_emulation.UsingGlobal());
+        ui->nvdec_emulation->setEnabled(Settings::values.nvdec_emulation.UsingGlobal());
         ui->accelerate_astc->setEnabled(Settings::values.accelerate_astc.UsingGlobal());
         ui->use_disk_shader_cache->setEnabled(Settings::values.use_disk_shader_cache.UsingGlobal());
         ui->bg_button->setEnabled(Settings::values.bg_red.UsingGlobal());
@@ -301,8 +329,6 @@ void ConfigureGraphics::SetupPerGameUI() {
 
     ConfigurationShared::SetColoredTristate(
         ui->use_disk_shader_cache, Settings::values.use_disk_shader_cache, use_disk_shader_cache);
-    ConfigurationShared::SetColoredTristate(
-        ui->use_nvdec_emulation, Settings::values.use_nvdec_emulation, use_nvdec_emulation);
     ConfigurationShared::SetColoredTristate(ui->accelerate_astc, Settings::values.accelerate_astc,
                                             accelerate_astc);
     ConfigurationShared::SetColoredTristate(ui->use_asynchronous_gpu_emulation,
@@ -316,4 +342,6 @@ void ConfigureGraphics::SetupPerGameUI() {
         static_cast<int>(Settings::values.fullscreen_mode.GetValue(true)));
     ConfigurationShared::InsertGlobalItem(
         ui->api, static_cast<int>(Settings::values.renderer_backend.GetValue(true)));
+    ConfigurationShared::InsertGlobalItem(
+        ui->nvdec_emulation, static_cast<int>(Settings::values.nvdec_emulation.GetValue(true)));
 }

src/yuzu/configuration/configure_graphics.h

@@ -43,6 +43,7 @@ private:
     void SetupPerGameUI();
 
     Settings::RendererBackend GetCurrentGraphicsBackend() const;
+    Settings::NvdecEmulation GetCurrentNvdecEmulation() const;
 
     std::unique_ptr<Ui::ConfigureGraphics> ui;
     QColor bg_color;

src/yuzu/configuration/configure_graphics.ui

@@ -167,13 +167,6 @@
           </property>
          </widget>
         </item>
-        <item>
-         <widget class="QCheckBox" name="use_nvdec_emulation">
-          <property name="text">
-           <string>Use NVDEC emulation</string>
-          </property>
-         </widget>
-        </item>
         <item>
           <widget class="QCheckBox" name="accelerate_astc">
             <property name="text">
@@ -181,6 +174,50 @@
             </property>
           </widget>
         </item>
+        <item>
+         <widget class="QWidget" name="nvdec_emulation_widget" native="true">
+          <layout class="QHBoxLayout" name="nvdec_emulation_layout">
+           <property name="leftMargin">
+            <number>0</number>
+           </property>
+           <property name="topMargin">
+            <number>0</number>
+           </property>
+           <property name="rightMargin">
+            <number>0</number>
+           </property>
+           <property name="bottomMargin">
+            <number>0</number>
+           </property>
+           <item>
+            <widget class="QLabel" name="nvdec_emulation_label">
+             <property name="text">
+              <string>NVDEC emulation:</string>
+             </property>
+            </widget>
+           </item>
+           <item>
+            <widget class="QComboBox" name="nvdec_emulation">
+             <item>
+              <property name="text">
+               <string>Disabled</string>
+              </property>
+             </item>
+             <item>
+              <property name="text">
+               <string>CPU Decoding</string>
+              </property>
+             </item>
+             <item>
+              <property name="text">
+               <string>GPU Decoding</string>
+              </property>
+             </item>
+            </widget>
+           </item>
+          </layout>
+         </widget>
+        </item>
         <item>
          <widget class="QWidget" name="fullscreen_mode_layout" native="true">
           <layout class="QHBoxLayout" name="horizontalLayout_1">

src/yuzu_cmd/config.cpp

@@ -465,7 +465,7 @@ void Config::ReadValues() {
     ReadSetting("Renderer", Settings::values.disable_fps_limit);
     ReadSetting("Renderer", Settings::values.shader_backend);
     ReadSetting("Renderer", Settings::values.use_asynchronous_shaders);
-    ReadSetting("Renderer", Settings::values.use_nvdec_emulation);
+    ReadSetting("Renderer", Settings::values.nvdec_emulation);
     ReadSetting("Renderer", Settings::values.accelerate_astc);
     ReadSetting("Renderer", Settings::values.use_fast_gpu_time);
 

src/yuzu_cmd/default_ini.h

@@ -261,9 +261,9 @@ shader_backend =
 # 0 (default): Off, 1: On
 use_asynchronous_shaders =
 
-# Enable NVDEC emulation.
+# NVDEC emulation.
-# 0: Off, 1 (default): On
+# 0: Disabled, 1: CPU Decoding, 2 (default): GPU Decoding
-use_nvdec_emulation =
+nvdec_emulation =
 
 # Accelerate ASTC texture decoding.
 # 0: Off, 1 (default): On