2 files changed, 114 insertions, 180 deletions

diff --git a/src/video_core/host1x/ffmpeg/ffmpeg.cpp b/src/video_core/host1x/ffmpeg/ffmpeg.cpp
index 1003cd38d..d603bad8b 100644
--- a/src/video_core/host1x/ffmpeg/ffmpeg.cpp
+++ b/src/video_core/host1x/ffmpeg/ffmpeg.cpp
@@ -5,7 +5,9 @@
 #include "common/logging/log.h"
 #include "common/scope_exit.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/host1x/ffmpeg/ffmpeg.h"
+#include "video_core/memory_manager.h"
 
 extern "C" {
 #ifdef LIBVA_FOUND
@@ -149,6 +151,7 @@ bool HardwareContext::InitializeForDecoder(DecoderContext& decoder_context,
         }
     }
 
+    LOG_INFO(HW_GPU, "Hardware decoding is disabled due to implementation issues, using CPU.");
     return false;
 }
 
@@ -183,8 +186,8 @@ bool HardwareContext::InitializeWithType(AVHWDeviceType type) {
     return true;
 }
 
-DecoderContext::DecoderContext(const Decoder& decoder) {
-    m_codec_context = avcodec_alloc_context3(decoder.GetCodec());
+DecoderContext::DecoderContext(const Decoder& decoder) : m_decoder{decoder} {
+    m_codec_context = avcodec_alloc_context3(m_decoder.GetCodec());
     av_opt_set(m_codec_context->priv_data, "tune", "zerolatency", 0);
     m_codec_context->thread_count = 0;
     m_codec_context->thread_type &= ~FF_THREAD_FRAME;
@@ -216,6 +219,25 @@ bool DecoderContext::OpenContext(const Decoder& decoder) {
 }
 
 bool DecoderContext::SendPacket(const Packet& packet) {
+    m_temp_frame = std::make_shared<Frame>();
+    m_got_frame = 0;
+
+// Android can randomly crash when calling decode directly, so skip.
+// TODO update ffmpeg and hope that fixes it.
+#ifndef ANDROID
+    if (!m_codec_context->hw_device_ctx && m_codec_context->codec_id == AV_CODEC_ID_H264) {
+        m_decode_order = true;
+        auto* codec{ffcodec(m_decoder.GetCodec())};
+        if (const int ret = codec->cb.decode(m_codec_context, m_temp_frame->GetFrame(),
+                                             &m_got_frame, packet.GetPacket());
+            ret < 0) {
+            LOG_DEBUG(Service_NVDRV, "avcodec_send_packet error {}", AVError(ret));
+            return false;
+        }
+        return true;
+    }
+#endif
+
     if (const int ret = avcodec_send_packet(m_codec_context, packet.GetPacket()); ret < 0) {
         LOG_ERROR(HW_GPU, "avcodec_send_packet error: {}", AVError(ret));
         return false;
@@ -224,139 +246,73 @@ bool DecoderContext::SendPacket(const Packet& packet) {
     return true;
 }
 
-std::unique_ptr<Frame> DecoderContext::ReceiveFrame(bool* out_is_interlaced) {
-    auto dst_frame = std::make_unique<Frame>();
-
-    const auto ReceiveImpl = [&](AVFrame* frame) {
-        if (const int ret = avcodec_receive_frame(m_codec_context, frame); ret < 0) {
-            LOG_ERROR(HW_GPU, "avcodec_receive_frame error: {}", AVError(ret));
-            return false;
+std::shared_ptr<Frame> DecoderContext::ReceiveFrame() {
+    // Android can randomly crash when calling decode directly, so skip.
+    // TODO update ffmpeg and hope that fixes it.
+#ifndef ANDROID
+    if (!m_codec_context->hw_device_ctx && m_codec_context->codec_id == AV_CODEC_ID_H264) {
+        m_decode_order = true;
+        auto* codec{ffcodec(m_decoder.GetCodec())};
+        int ret{0};
+
+        if (m_got_frame == 0) {
+            Packet packet{{}};
+            auto* pkt = packet.GetPacket();
+            pkt->data = nullptr;
+            pkt->size = 0;
+            ret = codec->cb.decode(m_codec_context, m_temp_frame->GetFrame(), &m_got_frame, pkt);
+            m_codec_context->has_b_frames = 0;
         }
 
-        *out_is_interlaced =
-#if defined(FF_API_INTERLACED_FRAME) || LIBAVUTIL_VERSION_MAJOR >= 59
-            (frame->flags & AV_FRAME_FLAG_INTERLACED) != 0;
-#else
-            frame->interlaced_frame != 0;
-#endif
-        return true;
-    };
-
-    if (m_codec_context->hw_device_ctx) {
-        // If we have a hardware context, make a separate frame here to receive the
-        // hardware result before sending it to the output.
-        Frame intermediate_frame;
-
-        if (!ReceiveImpl(intermediate_frame.GetFrame())) {
+        if (m_got_frame == 0 || ret < 0) {
+            LOG_ERROR(Service_NVDRV, "Failed to receive a frame! error {}", ret);
             return {};
         }
+    } else
+#endif
+    {
 
-        dst_frame->SetFormat(PreferredGpuFormat);
-        if (const int ret =
-                av_hwframe_transfer_data(dst_frame->GetFrame(), intermediate_frame.GetFrame(), 0);
-            ret < 0) {
-            LOG_ERROR(HW_GPU, "av_hwframe_transfer_data error: {}", AVError(ret));
-            return {};
-        }
-    } else {
-        // Otherwise, decode the frame as normal.
-        if (!ReceiveImpl(dst_frame->GetFrame())) {
-            return {};
-        }
-    }
-
-    return dst_frame;
-}
-
-DeinterlaceFilter::DeinterlaceFilter(const Frame& frame) {
-    const AVFilter* buffer_src = avfilter_get_by_name("buffer");
-    const AVFilter* buffer_sink = avfilter_get_by_name("buffersink");
-    AVFilterInOut* inputs = avfilter_inout_alloc();
-    AVFilterInOut* outputs = avfilter_inout_alloc();
-    SCOPE_EXIT {
-        avfilter_inout_free(&inputs);
-        avfilter_inout_free(&outputs);
-    };
-
-    // Don't know how to get the accurate time_base but it doesn't matter for yadif filter
-    // so just use 1/1 to make buffer filter happy
-    std::string args = fmt::format("video_size={}x{}:pix_fmt={}:time_base=1/1", frame.GetWidth(),
-                                   frame.GetHeight(), static_cast<int>(frame.GetPixelFormat()));
-
-    m_filter_graph = avfilter_graph_alloc();
-    int ret = avfilter_graph_create_filter(&m_source_context, buffer_src, "in", args.c_str(),
-                                           nullptr, m_filter_graph);
-    if (ret < 0) {
-        LOG_ERROR(HW_GPU, "avfilter_graph_create_filter source error: {}", AVError(ret));
-        return;
-    }
-
-    ret = avfilter_graph_create_filter(&m_sink_context, buffer_sink, "out", nullptr, nullptr,
-                                       m_filter_graph);
-    if (ret < 0) {
-        LOG_ERROR(HW_GPU, "avfilter_graph_create_filter sink error: {}", AVError(ret));
-        return;
-    }
-
-    inputs->name = av_strdup("out");
-    inputs->filter_ctx = m_sink_context;
-    inputs->pad_idx = 0;
-    inputs->next = nullptr;
-
-    outputs->name = av_strdup("in");
-    outputs->filter_ctx = m_source_context;
-    outputs->pad_idx = 0;
-    outputs->next = nullptr;
-
-    const char* description = "yadif=1:-1:0";
-    ret = avfilter_graph_parse_ptr(m_filter_graph, description, &inputs, &outputs, nullptr);
-    if (ret < 0) {
-        LOG_ERROR(HW_GPU, "avfilter_graph_parse_ptr error: {}", AVError(ret));
-        return;
-    }
-
-    ret = avfilter_graph_config(m_filter_graph, nullptr);
-    if (ret < 0) {
-        LOG_ERROR(HW_GPU, "avfilter_graph_config error: {}", AVError(ret));
-        return;
-    }
-
-    m_initialized = true;
-}
-
-bool DeinterlaceFilter::AddSourceFrame(const Frame& frame) {
-    if (const int ret = av_buffersrc_add_frame_flags(m_source_context, frame.GetFrame(),
-                                                     AV_BUFFERSRC_FLAG_KEEP_REF);
-        ret < 0) {
-        LOG_ERROR(HW_GPU, "av_buffersrc_add_frame_flags error: {}", AVError(ret));
-        return false;
-    }
+        const auto ReceiveImpl = [&](AVFrame* frame) {
+            if (const int ret = avcodec_receive_frame(m_codec_context, frame); ret < 0) {
+                LOG_ERROR(HW_GPU, "avcodec_receive_frame error: {}", AVError(ret));
+                return false;
+            }
 
-    return true;
-}
+            return true;
+        };
 
-std::unique_ptr<Frame> DeinterlaceFilter::DrainSinkFrame() {
-    auto dst_frame = std::make_unique<Frame>();
-    const int ret = av_buffersink_get_frame(m_sink_context, dst_frame->GetFrame());
+        if (m_codec_context->hw_device_ctx) {
+            // If we have a hardware context, make a separate frame here to receive the
+            // hardware result before sending it to the output.
+            Frame intermediate_frame;
 
-    if (ret == AVERROR(EAGAIN) || ret == AVERROR(AVERROR_EOF)) {
-        return {};
-    }
+            if (!ReceiveImpl(intermediate_frame.GetFrame())) {
+                return {};
+            }
 
-    if (ret < 0) {
-        LOG_ERROR(HW_GPU, "av_buffersink_get_frame error: {}", AVError(ret));
-        return {};
+            m_temp_frame->SetFormat(PreferredGpuFormat);
+            if (const int ret = av_hwframe_transfer_data(m_temp_frame->GetFrame(),
+                                                         intermediate_frame.GetFrame(), 0);
+                ret < 0) {
+                LOG_ERROR(HW_GPU, "av_hwframe_transfer_data error: {}", AVError(ret));
+                return {};
+            }
+        } else {
+            // Otherwise, decode the frame as normal.
+            if (!ReceiveImpl(m_temp_frame->GetFrame())) {
+                return {};
+            }
+        }
     }
 
-    return dst_frame;
-}
-
-DeinterlaceFilter::~DeinterlaceFilter() {
-    avfilter_graph_free(&m_filter_graph);
+#if defined(FF_API_INTERLACED_FRAME) || LIBAVUTIL_VERSION_MAJOR >= 59
+    m_temp_frame->GetFrame()->interlaced_frame =
+        (m_temp_frame->GetFrame()->flags & AV_FRAME_FLAG_INTERLACED) != 0;
+#endif
+    return std::move(m_temp_frame);
 }
 
 void DecodeApi::Reset() {
-    m_deinterlace_filter.reset();
     m_hardware_context.reset();
     m_decoder_context.reset();
     m_decoder.reset();
@@ -382,43 +338,14 @@ bool DecodeApi::Initialize(Tegra::Host1x::NvdecCommon::VideoCodec codec) {
     return true;
 }
 
-bool DecodeApi::SendPacket(std::span<const u8> packet_data, size_t configuration_size) {
+bool DecodeApi::SendPacket(std::span<const u8> packet_data) {
     FFmpeg::Packet packet(packet_data);
     return m_decoder_context->SendPacket(packet);
 }
 
-void DecodeApi::ReceiveFrames(std::queue<std::unique_ptr<Frame>>& frame_queue) {
+std::shared_ptr<Frame> DecodeApi::ReceiveFrame() {
     // Receive raw frame from decoder.
-    bool is_interlaced;
-    auto frame = m_decoder_context->ReceiveFrame(&is_interlaced);
-    if (!frame) {
-        return;
-    }
-
-    if (!is_interlaced) {
-        // If the frame is not interlaced, we can pend it now.
-        frame_queue.push(std::move(frame));
-    } else {
-        // Create the deinterlacer if needed.
-        if (!m_deinterlace_filter) {
-            m_deinterlace_filter.emplace(*frame);
-        }
-
-        // Add the frame we just received.
-        if (!m_deinterlace_filter->AddSourceFrame(*frame)) {
-            return;
-        }
-
-        // Pend output fields.
-        while (true) {
-            auto filter_frame = m_deinterlace_filter->DrainSinkFrame();
-            if (!filter_frame) {
-                break;
-            }
-
-            frame_queue.push(std::move(filter_frame));
-        }
-    }
+    return m_decoder_context->ReceiveFrame();
 }
 
 } // namespace FFmpeg
diff --git a/src/video_core/host1x/ffmpeg/ffmpeg.h b/src/video_core/host1x/ffmpeg/ffmpeg.h
index 1de0bbd83..a74fcba80 100644
--- a/src/video_core/host1x/ffmpeg/ffmpeg.h
+++ b/src/video_core/host1x/ffmpeg/ffmpeg.h
@@ -20,17 +20,20 @@ extern "C" {
 #endif
 
 #include <libavcodec/avcodec.h>
-#include <libavfilter/avfilter.h>
-#include <libavfilter/buffersink.h>
-#include <libavfilter/buffersrc.h>
-#include <libavutil/avutil.h>
 #include <libavutil/opt.h>
+#ifndef ANDROID
+#include <libavcodec/codec_internal.h>
+#endif
 
 #if defined(__GNUC__) || defined(__clang__)
 #pragma GCC diagnostic pop
 #endif
 }
 
+namespace Tegra {
+class MemoryManager;
+}
+
 namespace FFmpeg {
 
 class Packet;
@@ -90,6 +93,10 @@ public:
         return m_frame->data[plane];
     }
 
+    const u8* GetPlane(int plane) const {
+        return m_frame->data[plane];
+    }
+
     u8** GetPlanes() const {
         return m_frame->data;
     }
@@ -98,6 +105,14 @@ public:
         m_frame->format = format;
     }
 
+    bool IsInterlaced() const {
+        return m_frame->interlaced_frame != 0;
+    }
+
+    bool IsHardwareDecoded() const {
+        return m_frame->hw_frames_ctx != nullptr;
+    }
+
     AVFrame* GetFrame() const {
         return m_frame;
     }
@@ -160,33 +175,22 @@ public:
     void InitializeHardwareDecoder(const HardwareContext& context, AVPixelFormat hw_pix_fmt);
     bool OpenContext(const Decoder& decoder);
     bool SendPacket(const Packet& packet);
-    std::unique_ptr<Frame> ReceiveFrame(bool* out_is_interlaced);
+    std::shared_ptr<Frame> ReceiveFrame();
 
     AVCodecContext* GetCodecContext() const {
         return m_codec_context;
     }
 
-private:
-    AVCodecContext* m_codec_context{};
-};
-
-// Wraps an AVFilterGraph.
-class DeinterlaceFilter {
-public:
-    YUZU_NON_COPYABLE(DeinterlaceFilter);
-    YUZU_NON_MOVEABLE(DeinterlaceFilter);
-
-    explicit DeinterlaceFilter(const Frame& frame);
-    ~DeinterlaceFilter();
-
-    bool AddSourceFrame(const Frame& frame);
-    std::unique_ptr<Frame> DrainSinkFrame();
+    bool UsingDecodeOrder() const {
+        return m_decode_order;
+    }
 
 private:
-    AVFilterGraph* m_filter_graph{};
-    AVFilterContext* m_source_context{};
-    AVFilterContext* m_sink_context{};
-    bool m_initialized{};
+    const Decoder& m_decoder;
+    AVCodecContext* m_codec_context{};
+    s32 m_got_frame{};
+    std::shared_ptr<Frame> m_temp_frame{};
+    bool m_decode_order{};
 };
 
 class DecodeApi {
@@ -200,14 +204,17 @@ public:
     bool Initialize(Tegra::Host1x::NvdecCommon::VideoCodec codec);
     void Reset();
 
-    bool SendPacket(std::span<const u8> packet_data, size_t configuration_size);
-    void ReceiveFrames(std::queue<std::unique_ptr<Frame>>& frame_queue);
+    bool UsingDecodeOrder() const {
+        return m_decoder_context->UsingDecodeOrder();
+    }
+
+    bool SendPacket(std::span<const u8> packet_data);
+    std::shared_ptr<Frame> ReceiveFrame();
 
 private:
     std::optional<FFmpeg::Decoder> m_decoder;
     std::optional<FFmpeg::DecoderContext> m_decoder_context;
     std::optional<FFmpeg::HardwareContext> m_hardware_context;
-    std::optional<FFmpeg::DeinterlaceFilter> m_deinterlace_filter;
 };
 
 } // namespace FFmpeg