Merge pull request #7138 from ameerj/vic-fmt

vic: Implement RGBX8 video frame format
This commit is contained in:
Mai M 2021-10-08 19:19:20 -04:00 committed by GitHub
commit 39cd6306e6
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 167 additions and 138 deletions

View File

@ -16,6 +16,7 @@ extern "C" {
} }
#include "common/assert.h" #include "common/assert.h"
#include "common/bit_field.h"
#include "common/logging/log.h" #include "common/logging/log.h"
#include "video_core/command_classes/nvdec.h" #include "video_core/command_classes/nvdec.h"
@ -26,6 +27,25 @@ extern "C" {
#include "video_core/textures/decoders.h" #include "video_core/textures/decoders.h"
namespace Tegra { namespace Tegra {
namespace {
enum class VideoPixelFormat : u64_le {
RGBA8 = 0x1f,
BGRA8 = 0x20,
RGBX8 = 0x23,
Yuv420 = 0x44,
};
} // Anonymous namespace
union VicConfig {
u64_le raw{};
BitField<0, 7, VideoPixelFormat> pixel_format;
BitField<7, 2, u64_le> chroma_loc_horiz;
BitField<9, 2, u64_le> chroma_loc_vert;
BitField<11, 4, u64_le> block_linear_kind;
BitField<15, 4, u64_le> block_linear_height_log2;
BitField<32, 14, u64_le> surface_width_minus1;
BitField<46, 14, u64_le> surface_height_minus1;
};
Vic::Vic(GPU& gpu_, std::shared_ptr<Nvdec> nvdec_processor_) Vic::Vic(GPU& gpu_, std::shared_ptr<Nvdec> nvdec_processor_)
: gpu(gpu_), : gpu(gpu_),
@ -65,134 +85,155 @@ void Vic::Execute() {
if (!frame) { if (!frame) {
return; return;
} }
const auto pixel_format = static_cast<VideoPixelFormat>(config.pixel_format.Value()); const u64 surface_width = config.surface_width_minus1 + 1;
switch (pixel_format) { const u64 surface_height = config.surface_height_minus1 + 1;
if (static_cast<u64>(frame->width) != surface_width ||
static_cast<u64>(frame->height) != surface_height) {
// TODO: Properly support multiple video streams with differing frame dimensions
LOG_WARNING(Debug, "Frame dimensions {}x{} do not match expected surface dimensions {}x{}",
frame->width, frame->height, surface_width, surface_height);
return;
}
switch (config.pixel_format) {
case VideoPixelFormat::RGBA8:
case VideoPixelFormat::BGRA8: case VideoPixelFormat::BGRA8:
case VideoPixelFormat::RGBA8: { case VideoPixelFormat::RGBX8:
LOG_TRACE(Service_NVDRV, "Writing RGB Frame"); WriteRGBFrame(frame, config);
if (scaler_ctx == nullptr || frame->width != scaler_width ||
frame->height != scaler_height) {
const AVPixelFormat target_format =
(pixel_format == VideoPixelFormat::RGBA8) ? AV_PIX_FMT_RGBA : AV_PIX_FMT_BGRA;
sws_freeContext(scaler_ctx);
scaler_ctx = 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 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) {
converted_frame_buffer = AVMallocPtr{static_cast<u8*>(av_malloc(linear_size)), av_free};
}
const std::array<int, 4> converted_stride{frame->width * 4, frame->height * 4, 0, 0};
u8* const converted_frame_buf_addr{converted_frame_buffer.get()};
sws_scale(scaler_ctx, frame->data, frame->linesize, 0, frame->height,
&converted_frame_buf_addr, converted_stride.data());
const u32 blk_kind = static_cast<u32>(config.block_linear_kind);
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, width, height, 1, block_height, 0);
luma_buffer.resize(size);
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);
} else {
// send pitch linear frame
gpu.MemoryManager().WriteBlock(output_surface_luma_address, converted_frame_buf_addr,
linear_size);
}
break; break;
} case VideoPixelFormat::Yuv420:
case VideoPixelFormat::Yuv420: { WriteYUVFrame(frame, config);
LOG_TRACE(Service_NVDRV, "Writing YUV420 Frame");
const std::size_t surface_width = config.surface_width_minus1 + 1;
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 aligned_width = (surface_width + 0xff) & ~0xffUL;
const auto stride = static_cast<size_t>(frame->linesize[0]);
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_src[src + x];
}
}
gpu.MemoryManager().WriteBlock(output_surface_luma_address, luma_buffer.data(),
luma_buffer.size());
// Chroma
const std::size_t half_height = frame_height / 2;
const auto half_stride = static_cast<size_t>(frame->linesize[1]);
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_address, chroma_buffer.data(),
chroma_buffer.size());
break; break;
}
default: default:
UNIMPLEMENTED_MSG("Unknown video pixel format {}", config.pixel_format.Value()); UNIMPLEMENTED_MSG("Unknown video pixel format {:X}", config.pixel_format.Value());
break; break;
} }
} }
void Vic::WriteRGBFrame(const AVFrame* frame, const VicConfig& config) {
LOG_TRACE(Service_NVDRV, "Writing RGB Frame");
if (!scaler_ctx || frame->width != scaler_width || frame->height != scaler_height) {
const AVPixelFormat target_format = [pixel_format = config.pixel_format]() {
switch (pixel_format) {
case VideoPixelFormat::RGBA8:
return AV_PIX_FMT_RGBA;
case VideoPixelFormat::BGRA8:
return AV_PIX_FMT_BGRA;
case VideoPixelFormat::RGBX8:
return AV_PIX_FMT_RGB0;
default:
return AV_PIX_FMT_RGBA;
}
}();
sws_freeContext(scaler_ctx);
// Frames are decoded into either YUV420 or NV12 formats. Convert to desired RGB 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;
converted_frame_buffer.reset();
}
// Get Converted frame
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) {
converted_frame_buffer = AVMallocPtr{static_cast<u8*>(av_malloc(linear_size)), av_free};
}
const std::array<int, 4> converted_stride{frame->width * 4, frame->height * 4, 0, 0};
u8* const converted_frame_buf_addr{converted_frame_buffer.get()};
sws_scale(scaler_ctx, frame->data, frame->linesize, 0, frame->height, &converted_frame_buf_addr,
converted_stride.data());
const u32 blk_kind = static_cast<u32>(config.block_linear_kind);
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 = Texture::CalculateSize(true, 4, width, height, 1, block_height, 0);
luma_buffer.resize(size);
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);
} else {
// send pitch linear frame
gpu.MemoryManager().WriteBlock(output_surface_luma_address, converted_frame_buf_addr,
linear_size);
}
}
void Vic::WriteYUVFrame(const AVFrame* frame, const VicConfig& config) {
LOG_TRACE(Service_NVDRV, "Writing YUV420 Frame");
const std::size_t surface_width = config.surface_width_minus1 + 1;
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 aligned_width = (surface_width + 0xff) & ~0xffUL;
const auto stride = static_cast<size_t>(frame->linesize[0]);
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_src[src + x];
}
}
gpu.MemoryManager().WriteBlock(output_surface_luma_address, luma_buffer.data(),
luma_buffer.size());
// Chroma
const std::size_t half_height = frame_height / 2;
const auto half_stride = static_cast<size_t>(frame->linesize[1]);
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_address, chroma_buffer.data(),
chroma_buffer.size());
}
} // namespace Tegra } // namespace Tegra

View File

@ -6,7 +6,6 @@
#include <memory> #include <memory>
#include <vector> #include <vector>
#include "common/bit_field.h"
#include "common/common_types.h" #include "common/common_types.h"
struct SwsContext; struct SwsContext;
@ -14,6 +13,7 @@ struct SwsContext;
namespace Tegra { namespace Tegra {
class GPU; class GPU;
class Nvdec; class Nvdec;
union VicConfig;
class Vic { class Vic {
public: public:
@ -27,6 +27,7 @@ public:
}; };
explicit Vic(GPU& gpu, std::shared_ptr<Nvdec> nvdec_processor); explicit Vic(GPU& gpu, std::shared_ptr<Nvdec> nvdec_processor);
~Vic(); ~Vic();
/// Write to the device state. /// Write to the device state.
@ -35,22 +36,9 @@ public:
private: private:
void Execute(); void Execute();
enum class VideoPixelFormat : u64_le { void WriteRGBFrame(const AVFrame* frame, const VicConfig& config);
RGBA8 = 0x1f,
BGRA8 = 0x20,
Yuv420 = 0x44,
};
union VicConfig { void WriteYUVFrame(const AVFrame* frame, const VicConfig& config);
u64_le raw{};
BitField<0, 7, u64_le> pixel_format;
BitField<7, 2, u64_le> chroma_loc_horiz;
BitField<9, 2, u64_le> chroma_loc_vert;
BitField<11, 4, u64_le> block_linear_kind;
BitField<15, 4, u64_le> block_linear_height_log2;
BitField<32, 14, u64_le> surface_width_minus1;
BitField<46, 14, u64_le> surface_height_minus1;
};
GPU& gpu; GPU& gpu;
std::shared_ptr<Tegra::Nvdec> nvdec_processor; std::shared_ptr<Tegra::Nvdec> nvdec_processor;