nvdec: Implement GPU accelerated decoding for all platforms

Supplements the VAAPI intel gpu decoder by implementing the D3D11VA decoder for Windows, and CUVID/VDPAU for Nvidia and AMD on drivers linux respectively.

src/video_core/command_classes/codecs/codec.cpp

@@ -16,44 +16,17 @@ 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 = {
-    "i915",
-    "amdgpu",
-};
-
-AVPixelFormat GetHwFormat(AVCodecContext*, const AVPixelFormat* pix_fmts) {
+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) {
-            return AV_PIX_FMT_VAAPI;
+        if (*p == av_codec_ctx->pix_fmt) {
+            return av_codec_ctx->pix_fmt;
         }
     }
     LOG_INFO(Service_NVDRV, "Could not find compatible GPU AV format, falling back to CPU");
-    return *pix_fmts;
-}
-
-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;
+    return AV_PIX_FMT_NONE;
 }
 } // namespace
-#endif
 
 void AVFrameDeleter(AVFrame* ptr) {
     av_frame_free(&ptr);
@@ -69,26 +42,75 @@ Codec::~Codec() {
     }
     // Free libav memory
     avcodec_send_packet(av_codec_ctx, nullptr);
-    AVFrame* av_frame = av_frame_alloc();
-    avcodec_receive_frame(av_codec_ctx, av_frame);
+    AVFramePtr av_frame{av_frame_alloc(), AVFrameDeleter};
+    avcodec_receive_frame(av_codec_ctx, av_frame.get());
     avcodec_flush_buffers(av_codec_ctx);
-    av_frame_free(&av_frame);
     avcodec_close(av_codec_ctx);
-    av_buffer_unref(&av_hw_device);
+    av_buffer_unref(&av_gpu_decoder);
 }
 
-void Codec::InitializeHwdec() {
-    // Prioritize integrated GPU to mitigate bandwidth bottlenecks
+bool Codec::CreateGpuAvDevice() {
 #if defined(LIBVA_FOUND)
-    if (CreateVaapiHwdevice(&av_hw_device)) {
-        const auto hw_device_ctx = av_buffer_ref(av_hw_device);
-        ASSERT_MSG(hw_device_ctx, "av_buffer_ref failed");
-        av_codec_ctx->hw_device_ctx = hw_device_ctx;
-        av_codec_ctx->get_format = GetHwFormat;
-        return;
+    static constexpr std::array<const char*, 3> VAAPI_DRIVERS = {
+        "i915",
+        "iHD",
+        "amdgpu",
+    };
+    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
-    // TODO more GPU accelerated decoders
+    };
+    for (const auto& type : GPU_DECODER_TYPES) {
+        av_hwdevice_ctx_create(&av_gpu_decoder, type, nullptr, nullptr, 0);
+        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::InitializeGpuDecoder() {
+    if (!CreateGpuAvDevice()) {
+        av_buffer_unref(&av_gpu_decoder);
+        return;
+    }
+    auto* hw_device_ctx = av_buffer_ref(av_gpu_decoder);
+    ASSERT_MSG(hw_device_ctx, "av_buffer_ref failed");
+    av_codec_ctx->hw_device_ctx = hw_device_ctx;
+    av_codec_ctx->get_format = GetGpuFormat;
+    using_gpu_decode = true;
 }
 
 void Codec::Initialize() {
@@ -107,7 +129,8 @@ void Codec::Initialize() {
     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();
+
+    InitializeGpuDecoder();
     if (!av_codec_ctx->hw_device_ctx) {
         LOG_INFO(Service_NVDRV, "Using FFmpeg software decoding");
     }
@@ -115,7 +138,7 @@ void Codec::Initialize() {
     if (av_error < 0) {
         LOG_ERROR(Service_NVDRV, "avcodec_open2() Failed.");
         avcodec_close(av_codec_ctx);
-        av_buffer_unref(&av_hw_device);
+        av_buffer_unref(&av_gpu_decoder);
         return;
     }
     initialized = true;
@@ -153,38 +176,33 @@ void Codec::Decode() {
     if (vp9_hidden_frame) {
         return;
     }
-    AVFrame* hw_frame = av_frame_alloc();
-    AVFrame* sw_frame = hw_frame;
-    ASSERT_MSG(hw_frame, "av_frame_alloc hw_frame failed");
-    if (const int ret = avcodec_receive_frame(av_codec_ctx, hw_frame); ret) {
+    AVFramePtr initial_frame{av_frame_alloc(), AVFrameDeleter};
+    AVFramePtr final_frame{nullptr, AVFrameDeleter};
+    ASSERT_MSG(initial_frame, "av_frame_alloc initial_frame failed");
+    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)
-    // Hardware acceleration code from FFmpeg/doc/examples/hw_decode.c under MIT license
-    if (hw_frame->format == AV_PIX_FMT_VAAPI) {
-        sw_frame = av_frame_alloc();
-        ASSERT_MSG(sw_frame, "av_frame_alloc sw_frame failed");
+    if (using_gpu_decode) {
+        final_frame = AVFramePtr{av_frame_alloc(), AVFrameDeleter};
+        ASSERT_MSG(final_frame, "av_frame_alloc final_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;
-        const int transfer_data_ret = av_hwframe_transfer_data(sw_frame, hw_frame, 0);
-        ASSERT_MSG(!transfer_data_ret, "av_hwframe_transfer_data error {}", transfer_data_ret);
-        av_frame_free(&hw_frame);
-    }
-#endif
-    if (sw_frame->format != AV_PIX_FMT_YUV420P && sw_frame->format != AV_PIX_FMT_NV12) {
-        UNIMPLEMENTED_MSG("Unexpected video format from host graphics: {}", sw_frame->format);
-        av_frame_free(&sw_frame);
+        final_frame->format = AV_PIX_FMT_NV12;
+        const int ret = av_hwframe_transfer_data(final_frame.get(), initial_frame.get(), 0);
+        ASSERT_MSG(!ret, "av_hwframe_transfer_data error {}", ret);
+    } else {
+        final_frame = std::move(initial_frame);
+    }
+    if (final_frame->format != AV_PIX_FMT_YUV420P && final_frame->format != AV_PIX_FMT_NV12) {
+        UNIMPLEMENTED_MSG("Unexpected video format: {}", final_frame->format);
         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 +50,22 @@ 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 InitializeGpuDecoder();
+
+    bool CreateGpuAvDevice();
 
     bool initialized{};
+    bool using_gpu_decode{};
     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;