vp8: Implement header composition

Enables frame decoding with FFmpeg

4 files changed, 90 insertions, 6 deletions

diff --git a/src/video_core/command_classes/codecs/codec.cpp b/src/video_core/command_classes/codecs/codec.cpp
index fc3b8db99..cf5ede204 100644
--- a/src/video_core/command_classes/codecs/codec.cpp
+++ b/src/video_core/command_classes/codecs/codec.cpp
@@ -185,7 +185,7 @@ void Codec::Decode() {
         case Tegra::NvdecCommon::VideoCodec::H264:
             return h264_decoder->ComposeFrameHeader(state, is_first_frame);
         case Tegra::NvdecCommon::VideoCodec::VP8:
-            return vp8_decoder->ComposeFrameHeader(state, is_first_frame);
+            return vp8_decoder->ComposeFrameHeader(state);
         case Tegra::NvdecCommon::VideoCodec::VP9:
             vp9_decoder->ComposeFrameHeader(state);
             vp9_hidden_frame = vp9_decoder->WasFrameHidden();
diff --git a/src/video_core/command_classes/codecs/vp8.cpp b/src/video_core/command_classes/codecs/vp8.cpp
index 976e9f9b7..3ee269948 100644
--- a/src/video_core/command_classes/codecs/vp8.cpp
+++ b/src/video_core/command_classes/codecs/vp8.cpp
@@ -6,15 +6,50 @@
 #include <vector>
 
 #include "video_core/command_classes/codecs/vp8.h"
+#include "video_core/gpu.h"
+#include "video_core/memory_manager.h"
 
 namespace Tegra::Decoder {
 VP8::VP8(GPU& gpu_) : gpu(gpu_) {}
 
 VP8::~VP8() = default;
 
-const std::vector<u8>& VP8::ComposeFrameHeader(const NvdecCommon::NvdecRegisters& state,
-                                               bool is_first_frame) {
-    return {};
+const std::vector<u8>& VP8::ComposeFrameHeader(const NvdecCommon::NvdecRegisters& state) {
+    VP8PictureInfo info;
+    gpu.MemoryManager().ReadBlock(state.picture_info_offset, &info, sizeof(VP8PictureInfo));
+
+    const bool is_key_frame = info.key_frame == 1u;
+    const auto bitstream_size = static_cast<size_t>(info.vld_buffer_size);
+    const size_t header_size = is_key_frame ? 10u : 3u;
+    frame.resize(header_size + bitstream_size);
+
+    // Based on page 30 of the VP8 specification.
+    // https://datatracker.ietf.org/doc/rfc6386/
+    frame[0] = is_key_frame ? 0u : 1u; // 1-bit frame type (0: keyframe, 1: interframes).
+    frame[0] |= static_cast<u8>((info.version & 7u) << 1u); // 3-bit version number
+    frame[0] |= static_cast<u8>(1u << 4u);                  // 1-bit show_frame flag
+
+    // The next 19-bits are the first partition size
+    frame[0] |= static_cast<u8>((info.first_part_size & 7u) << 5u);
+    frame[1] = static_cast<u8>((info.first_part_size & 0x7f8u) >> 3u);
+    frame[2] = static_cast<u8>((info.first_part_size & 0x7f800u) >> 11u);
+
+    if (is_key_frame) {
+        frame[3] = 0x9du;
+        frame[4] = 0x01u;
+        frame[5] = 0x2au;
+        // TODO(ameerj): Horizontal/Vertical Scale
+        // 16 bits: (2 bits Horizontal Scale << 14) | Width (14 bits)
+        frame[6] = static_cast<u8>(info.frame_width & 0xff);
+        frame[7] = static_cast<u8>(((info.frame_width >> 8) & 0x3f));
+        // 16 bits:(2 bits Vertical Scale << 14) | Height (14 bits)
+        frame[8] = static_cast<u8>(info.frame_height & 0xff);
+        frame[9] = static_cast<u8>(((info.frame_height >> 8) & 0x3f));
+    }
+    const u64 bitstream_offset = state.frame_bitstream_offset;
+    gpu.MemoryManager().ReadBlock(bitstream_offset, frame.data() + header_size, bitstream_size);
+
+    return frame;
 }
 
 } // namespace Tegra::Decoder
diff --git a/src/video_core/command_classes/codecs/vp8.h b/src/video_core/command_classes/codecs/vp8.h
index 70c75f414..d71917596 100644
--- a/src/video_core/command_classes/codecs/vp8.h
+++ b/src/video_core/command_classes/codecs/vp8.h
@@ -4,8 +4,10 @@
 
 #pragma once
 
+#include <array>
 #include <vector>
 
+#include "common/common_funcs.h"
 #include "common/common_types.h"
 #include "video_core/command_classes/nvdec_common.h"
 
@@ -20,11 +22,53 @@ public:
 
     /// Compose the VP8 header of the frame for FFmpeg decoding
     [[nodiscard]] const std::vector<u8>& ComposeFrameHeader(
-        const NvdecCommon::NvdecRegisters& state, bool is_first_frame = false);
+        const NvdecCommon::NvdecRegisters& state);
 
 private:
     std::vector<u8> frame;
     GPU& gpu;
+
+    struct VP8PictureInfo {
+        INSERT_PADDING_WORDS_NOINIT(14);
+        u16 frame_width;  // actual frame width
+        u16 frame_height; // actual frame height
+        u8 key_frame;
+        u8 version;
+        union {
+            u8 raw;
+            BitField<0, 2, u8> tile_format;
+            BitField<2, 3, u8> gob_height;
+            BitField<5, 3, u8> reserverd_surface_format;
+        };
+        u8 error_conceal_on;  // 1: error conceal on; 0: off
+        u32 first_part_size;  // the size of first partition(frame header and mb header partition)
+        u32 hist_buffer_size; // in units of 256
+        u32 vld_buffer_size;  // in units of 1
+        // Current frame buffers
+        std::array<u32, 2> frame_stride; // [y_c]
+        u32 luma_top_offset;             // offset of luma top field in units of 256
+        u32 luma_bot_offset;             // offset of luma bottom field in units of 256
+        u32 luma_frame_offset;           // offset of luma frame in units of 256
+        u32 chroma_top_offset;           // offset of chroma top field in units of 256
+        u32 chroma_bot_offset;           // offset of chroma bottom field in units of 256
+        u32 chroma_frame_offset;         // offset of chroma frame in units of 256
+
+        INSERT_PADDING_BYTES_NOINIT(0x1c); // NvdecDisplayParams
+
+        // Decode picture buffer related
+        s8 current_output_memory_layout;
+        // output NV12/NV24 setting. index 0: golden; 1: altref; 2: last
+        std::array<s8, 3> output_memory_layout;
+
+        u8 segmentation_feature_data_update;
+        INSERT_PADDING_BYTES_NOINIT(3);
+
+        // ucode return result
+        u32 result_value;
+        std::array<u32, 8> partition_offset;
+        INSERT_PADDING_WORDS_NOINIT(3);
+    };
+    static_assert(sizeof(VP8PictureInfo) == 0xc0, "PictureInfo is an invalid size");
 };
 
 } // namespace Decoder
diff --git a/src/video_core/command_classes/nvdec_common.h b/src/video_core/command_classes/nvdec_common.h
index 26c974b00..8a35c44a1 100644
--- a/src/video_core/command_classes/nvdec_common.h
+++ b/src/video_core/command_classes/nvdec_common.h
@@ -50,7 +50,10 @@ struct NvdecRegisters {
             u64 h264_last_surface_chroma_offset;       ///< 0x0858
             std::array<u64, 17> surface_luma_offset;   ///< 0x0860
             std::array<u64, 17> surface_chroma_offset; ///< 0x08E8
-            INSERT_PADDING_WORDS_NOINIT(132);          ///< 0x0970
+            INSERT_PADDING_WORDS_NOINIT(68);           ///< 0x0970
+            u64 vp8_prob_data_offset;                  ///< 0x0A80
+            u64 vp8_header_partition_buf_offset;       ///< 0x0A88
+            INSERT_PADDING_WORDS_NOINIT(60);           ///< 0x0A90
             u64 vp9_entropy_probs_offset;              ///< 0x0B80
             u64 vp9_backward_updates_offset;           ///< 0x0B88
             u64 vp9_last_frame_segmap_offset;          ///< 0x0B90
@@ -81,6 +84,8 @@ ASSERT_REG_POSITION(h264_last_surface_luma_offset, 0x10A);
 ASSERT_REG_POSITION(h264_last_surface_chroma_offset, 0x10B);
 ASSERT_REG_POSITION(surface_luma_offset, 0x10C);
 ASSERT_REG_POSITION(surface_chroma_offset, 0x11D);
+ASSERT_REG_POSITION(vp8_prob_data_offset, 0x150);
+ASSERT_REG_POSITION(vp8_header_partition_buf_offset, 0x151);
 ASSERT_REG_POSITION(vp9_entropy_probs_offset, 0x170);
 ASSERT_REG_POSITION(vp9_backward_updates_offset, 0x171);
 ASSERT_REG_POSITION(vp9_last_frame_segmap_offset, 0x172);