nvdec: Implement VA-API hardware video acceleration (#6713)

* nvdec: VA-API

* Verify formatting

* Forgot a semicolon for Windows

* Clarify comment about AV_PIX_FMT_NV12

* Fix assert log spam from missing negation

* vic: Remove forgotten debug code

* Address lioncash's review

* Mention VA-API is Intel/AMD

* Address v1993's review

* Hopefully fix CMakeLists style this time

* vic: Improve cache locality

* vic: Fix off-by-one error

* codec: Async

* codec: Forgot the GetValue()

* nvdec: Address ameerj's review

* codec: Fallback to CPU without VA-API support

* cmake: Address lat9nq's review

* cmake: Make VA-API optional

* vaapi: Multiple GPU

* Apply suggestions from code review

Co-authored-by: Ameer J <52414509+ameerj@users.noreply.github.com>

* nvdec: Address ameerj's review

* codec: Use anonymous instead of static

* nvdec: Remove enum and fix memory leak

* nvdec: Address ameerj's review

* codec: Remove preparation for threading

Co-authored-by: Ameer J <52414509+ameerj@users.noreply.github.com>

5 files changed, 175 insertions, 72 deletions

diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 333f6f35f..1eb67c051 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -1,5 +1,10 @@
 add_subdirectory(host_shaders)
 
+if(LIBVA_FOUND)
+    set_source_files_properties(command_classes/codecs/codec.cpp
+        PROPERTIES COMPILE_DEFINITIONS LIBVA_FOUND=1)
+endif()
+
 add_library(video_core STATIC
     buffer_cache/buffer_base.h
     buffer_cache/buffer_cache.cpp
diff --git a/src/video_core/command_classes/codecs/codec.cpp b/src/video_core/command_classes/codecs/codec.cpp
index 1b4bbc8ac..f798a0053 100644
--- a/src/video_core/command_classes/codecs/codec.cpp
+++ b/src/video_core/command_classes/codecs/codec.cpp
@@ -2,7 +2,6 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include <cstring>
 #include <fstream>
 #include <vector>
 #include "common/assert.h"
@@ -17,10 +16,47 @@ 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) {
+    for (const AVPixelFormat* p = pix_fmts; *p != AV_PIX_FMT_NONE; ++p) {
+        if (*p == AV_PIX_FMT_VAAPI) {
+            return AV_PIX_FMT_VAAPI;
+        }
+    }
+    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;
+}
+} // namespace
+#endif
 
 void AVFrameDeleter(AVFrame* ptr) {
-    av_frame_unref(ptr);
-    av_free(ptr);
+    av_frame_free(&ptr);
 }
 
 Codec::Codec(GPU& gpu_, const NvdecCommon::NvdecRegisters& regs)
@@ -32,19 +68,31 @@ Codec::~Codec() {
         return;
     }
     // Free libav memory
-    AVFrame* av_frame{nullptr};
     avcodec_send_packet(av_codec_ctx, nullptr);
-    av_frame = av_frame_alloc();
+    AVFrame* av_frame = av_frame_alloc();
     avcodec_receive_frame(av_codec_ctx, av_frame);
     avcodec_flush_buffers(av_codec_ctx);
-
-    av_frame_unref(av_frame);
-    av_free(av_frame);
+    av_frame_free(&av_frame);
     avcodec_close(av_codec_ctx);
+    av_buffer_unref(&av_hw_device);
+}
+
+void Codec::InitializeHwdec() {
+    // Prioritize integrated GPU to mitigate bandwidth bottlenecks
+#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;
+    }
+#endif
+    // TODO more GPU accelerated decoders
 }
 
 void Codec::Initialize() {
-    AVCodecID codec{AV_CODEC_ID_NONE};
+    AVCodecID codec;
     switch (current_codec) {
     case NvdecCommon::VideoCodec::H264:
         codec = AV_CODEC_ID_H264;
@@ -53,22 +101,24 @@ void Codec::Initialize() {
         codec = AV_CODEC_ID_VP9;
         break;
     default:
+        UNIMPLEMENTED_MSG("Unknown codec {}", current_codec);
         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);
-
-    // TODO(ameerj): libavcodec gpu hw acceleration
-
+    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;
-    return;
 }
 
 void Codec::SetTargetCodec(NvdecCommon::VideoCodec codec) {
@@ -80,36 +130,64 @@ void Codec::SetTargetCodec(NvdecCommon::VideoCodec codec) {
 
 void Codec::Decode() {
     const bool is_first_frame = !initialized;
-    if (!initialized) {
+    if (is_first_frame) {
         Initialize();
     }
-
     bool vp9_hidden_frame = false;
-    AVPacket packet{};
-    av_init_packet(&packet);
     std::vector<u8> frame_data;
-
     if (current_codec == NvdecCommon::VideoCodec::H264) {
         frame_data = h264_decoder->ComposeFrameHeader(state, is_first_frame);
     } else if (current_codec == NvdecCommon::VideoCodec::Vp9) {
         frame_data = vp9_decoder->ComposeFrameHeader(state);
         vp9_hidden_frame = vp9_decoder->WasFrameHidden();
     }
-
+    AVPacket packet{};
+    av_init_packet(&packet);
     packet.data = frame_data.data();
     packet.size = static_cast<s32>(frame_data.size());
-
-    avcodec_send_packet(av_codec_ctx, &packet);
-
-    if (!vp9_hidden_frame) {
-        // Only receive/store visible frames
-        AVFramePtr frame = AVFramePtr{av_frame_alloc(), AVFrameDeleter};
-        avcodec_receive_frame(av_codec_ctx, frame.get());
-        av_frames.push(std::move(frame));
-        // Limit queue to 10 frames. Workaround for ZLA decode and queue spam
-        if (av_frames.size() > 10) {
-            av_frames.pop();
-        }
+    if (const int ret = avcodec_send_packet(av_codec_ctx, &packet); ret) {
+        LOG_DEBUG(Service_NVDRV, "avcodec_send_packet error {}", ret);
+        return;
+    }
+    // Only receive/store visible frames
+    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) {
+        LOG_DEBUG(Service_NVDRV, "avcodec_receive_frame error {}", ret);
+        av_frame_free(&hw_frame);
+        return;
+    }
+    if (!hw_frame->width || !hw_frame->height) {
+        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");
+        // 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);
+        return;
+    }
+    av_frames.push(AVFramePtr{sw_frame, AVFrameDeleter});
+    if (av_frames.size() > 10) {
+        LOG_TRACE(Service_NVDRV, "av_frames.push overflow dropped frame");
+        av_frames.pop();
     }
 }
 
@@ -119,7 +197,6 @@ AVFramePtr Codec::GetCurrentFrame() {
     if (av_frames.empty()) {
         return AVFramePtr{nullptr, AVFrameDeleter};
     }
-
     AVFramePtr frame = std::move(av_frames.front());
     av_frames.pop();
     return frame;
@@ -144,6 +221,5 @@ std::string_view Codec::GetCurrentCodecName() const {
     default:
         return "Unknown";
     }
-};
-
+}
 } // namespace Tegra
diff --git a/src/video_core/command_classes/codecs/codec.h b/src/video_core/command_classes/codecs/codec.h
index 96c823c76..71936203f 100644
--- a/src/video_core/command_classes/codecs/codec.h
+++ b/src/video_core/command_classes/codecs/codec.h
@@ -22,7 +22,6 @@ extern "C" {
 
 namespace Tegra {
 class GPU;
-struct VicRegisters;
 
 void AVFrameDeleter(AVFrame* ptr);
 using AVFramePtr = std::unique_ptr<AVFrame, decltype(&AVFrameDeleter)>;
@@ -55,10 +54,13 @@ public:
     [[nodiscard]] std::string_view GetCurrentCodecName() const;
 
 private:
+    void InitializeHwdec();
+
     bool initialized{};
     NvdecCommon::VideoCodec current_codec{NvdecCommon::VideoCodec::None};
 
     AVCodec* av_codec{nullptr};
+    AVBufferRef* av_hw_device{nullptr};
     AVCodecContext* av_codec_ctx{nullptr};
 
     GPU& gpu;
diff --git a/src/video_core/command_classes/vic.cpp b/src/video_core/command_classes/vic.cpp
index ffb7c82a1..d5e77941c 100644
--- a/src/video_core/command_classes/vic.cpp
+++ b/src/video_core/command_classes/vic.cpp
@@ -46,11 +46,8 @@ void Vic::ProcessMethod(Method method, u32 argument) {
     case Method::SetOutputSurfaceLumaOffset:
         output_surface_luma_address = arg;
         break;
-    case Method::SetOutputSurfaceChromaUOffset:
-        output_surface_chroma_u_address = arg;
-        break;
-    case Method::SetOutputSurfaceChromaVOffset:
-        output_surface_chroma_v_address = arg;
+    case Method::SetOutputSurfaceChromaOffset:
+        output_surface_chroma_address = arg;
         break;
     default:
         break;
@@ -65,11 +62,10 @@ void Vic::Execute() {
     const VicConfig config{gpu.MemoryManager().Read<u64>(config_struct_address + 0x20)};
     const AVFramePtr frame_ptr = nvdec_processor->GetFrame();
     const auto* frame = frame_ptr.get();
-    if (!frame || frame->width == 0 || frame->height == 0) {
+    if (!frame) {
         return;
     }
-    const VideoPixelFormat pixel_format =
-        static_cast<VideoPixelFormat>(config.pixel_format.Value());
+    const auto pixel_format = static_cast<VideoPixelFormat>(config.pixel_format.Value());
     switch (pixel_format) {
     case VideoPixelFormat::BGRA8:
     case VideoPixelFormat::RGBA8: {
@@ -83,16 +79,18 @@ void Vic::Execute() {
             sws_freeContext(scaler_ctx);
             scaler_ctx = nullptr;
 
-            // FFmpeg returns all frames in YUV420, convert it into expected format
-            scaler_ctx =
-                sws_getContext(frame->width, frame->height, AV_PIX_FMT_YUV420P, frame->width,
-                               frame->height, target_format, 0, nullptr, nullptr, nullptr);
+            // Frames are decoded into either YUV420 or NV12 formats. Convert to desired format
+            scaler_ctx = sws_getContext(frame->width, frame->height,
+                                        static_cast<AVPixelFormat>(frame->format), frame->width,
+                                        frame->height, target_format, 0, nullptr, nullptr, nullptr);
 
             scaler_width = frame->width;
             scaler_height = frame->height;
         }
         // Get Converted frame
-        const std::size_t linear_size = frame->width * frame->height * 4;
+        const u32 width = static_cast<u32>(frame->width);
+        const u32 height = static_cast<u32>(frame->height);
+        const std::size_t linear_size = width * height * 4;
 
         // Only allocate frame_buffer once per stream, as the size is not expected to change
         if (!converted_frame_buffer) {
@@ -109,11 +107,10 @@ void Vic::Execute() {
         if (blk_kind != 0) {
             // swizzle pitch linear to block linear
             const u32 block_height = static_cast<u32>(config.block_linear_height_log2);
-            const auto size = Tegra::Texture::CalculateSize(true, 4, frame->width, frame->height, 1,
-                                                            block_height, 0);
+            const auto size =
+                Tegra::Texture::CalculateSize(true, 4, width, height, 1, block_height, 0);
             luma_buffer.resize(size);
-            Tegra::Texture::SwizzleSubrect(frame->width, frame->height, frame->width * 4,
-                                           frame->width, 4, luma_buffer.data(),
+            Tegra::Texture::SwizzleSubrect(width, height, width * 4, width, 4, luma_buffer.data(),
                                            converted_frame_buffer.get(), block_height, 0, 0);
 
             gpu.MemoryManager().WriteBlock(output_surface_luma_address, luma_buffer.data(), size);
@@ -131,41 +128,65 @@ void Vic::Execute() {
         const std::size_t surface_height = config.surface_height_minus1 + 1;
         const auto frame_width = std::min(surface_width, static_cast<size_t>(frame->width));
         const auto frame_height = std::min(surface_height, static_cast<size_t>(frame->height));
-        const std::size_t half_width = frame_width / 2;
-        const std::size_t half_height = frame_height / 2;
-        const std::size_t aligned_width = (surface_width + 0xff) & ~0xff;
+        const std::size_t aligned_width = (surface_width + 0xff) & ~0xffUL;
 
-        const auto* luma_ptr = frame->data[0];
-        const auto* chroma_b_ptr = frame->data[1];
-        const auto* chroma_r_ptr = frame->data[2];
         const auto stride = static_cast<size_t>(frame->linesize[0]);
-        const auto half_stride = static_cast<size_t>(frame->linesize[1]);
 
         luma_buffer.resize(aligned_width * surface_height);
         chroma_buffer.resize(aligned_width * surface_height / 2);
 
         // Populate luma buffer
+        const u8* luma_src = frame->data[0];
         for (std::size_t y = 0; y < frame_height; ++y) {
             const std::size_t src = y * stride;
             const std::size_t dst = y * aligned_width;
             for (std::size_t x = 0; x < frame_width; ++x) {
-                luma_buffer[dst + x] = luma_ptr[src + x];
+                luma_buffer[dst + x] = luma_src[src + x];
             }
         }
         gpu.MemoryManager().WriteBlock(output_surface_luma_address, luma_buffer.data(),
                                        luma_buffer.size());
 
-        // Populate chroma buffer from both channels with interleaving.
-        for (std::size_t y = 0; y < half_height; ++y) {
-            const std::size_t src = y * half_stride;
-            const std::size_t dst = y * aligned_width;
+        // Chroma
+        const std::size_t half_height = frame_height / 2;
+        const auto half_stride = static_cast<size_t>(frame->linesize[1]);
 
-            for (std::size_t x = 0; x < half_width; ++x) {
-                chroma_buffer[dst + x * 2] = chroma_b_ptr[src + x];
-                chroma_buffer[dst + x * 2 + 1] = chroma_r_ptr[src + x];
+        switch (frame->format) {
+        case AV_PIX_FMT_YUV420P: {
+            // Frame from FFmpeg software
+            // Populate chroma buffer from both channels with interleaving.
+            const std::size_t half_width = frame_width / 2;
+            const u8* chroma_b_src = frame->data[1];
+            const u8* chroma_r_src = frame->data[2];
+            for (std::size_t y = 0; y < half_height; ++y) {
+                const std::size_t src = y * half_stride;
+                const std::size_t dst = y * aligned_width;
+
+                for (std::size_t x = 0; x < half_width; ++x) {
+                    chroma_buffer[dst + x * 2] = chroma_b_src[src + x];
+                    chroma_buffer[dst + x * 2 + 1] = chroma_r_src[src + x];
+                }
+            }
+            break;
+        }
+        case AV_PIX_FMT_NV12: {
+            // Frame from VA-API hardware
+            // This is already interleaved so just copy
+            const u8* chroma_src = frame->data[1];
+            for (std::size_t y = 0; y < half_height; ++y) {
+                const std::size_t src = y * stride;
+                const std::size_t dst = y * aligned_width;
+                for (std::size_t x = 0; x < frame_width; ++x) {
+                    chroma_buffer[dst + x] = chroma_src[src + x];
+                }
             }
+            break;
+        }
+        default:
+            UNREACHABLE();
+            break;
         }
-        gpu.MemoryManager().WriteBlock(output_surface_chroma_u_address, chroma_buffer.data(),
+        gpu.MemoryManager().WriteBlock(output_surface_chroma_address, chroma_buffer.data(),
                                        chroma_buffer.size());
         break;
     }
diff --git a/src/video_core/command_classes/vic.h b/src/video_core/command_classes/vic.h
index f5a2ed100..74246e08c 100644
--- a/src/video_core/command_classes/vic.h
+++ b/src/video_core/command_classes/vic.h
@@ -22,8 +22,8 @@ public:
         SetControlParams = 0x1c1,
         SetConfigStructOffset = 0x1c2,
         SetOutputSurfaceLumaOffset = 0x1c8,
-        SetOutputSurfaceChromaUOffset = 0x1c9,
-        SetOutputSurfaceChromaVOffset = 0x1ca
+        SetOutputSurfaceChromaOffset = 0x1c9,
+        SetOutputSurfaceChromaUnusedOffset = 0x1ca
     };
 
     explicit Vic(GPU& gpu, std::shared_ptr<Nvdec> nvdec_processor);
@@ -64,8 +64,7 @@ private:
 
     GPUVAddr config_struct_address{};
     GPUVAddr output_surface_luma_address{};
-    GPUVAddr output_surface_chroma_u_address{};
-    GPUVAddr output_surface_chroma_v_address{};
+    GPUVAddr output_surface_chroma_address{};
 
     SwsContext* scaler_ctx{};
     s32 scaler_width{};