Mirrors/yuzu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

textures: add BC1 and BC3 compressors and recompression setting

Browse source
This commit is contained in:
Liam 2023-05-20 17:15:36 -04:00
parent f82efe9f65
commit 415c78b87c
23 changed files with 1150 additions and 27 deletions
Download patch file Download diff file Expand all files Collapse all files

6
src/video_core/CMakeLists.txt
View file

@ -246,10 +246,14 @@ add_library(video_core STATIC
texture_cache/util.h
textures/astc.h
textures/astc.cpp
textures/bcn.cpp
textures/bcn.h
textures/decoders.cpp
textures/decoders.h
textures/texture.cpp
textures/texture.h
textures/workers.cpp
textures/workers.h
transform_feedback.cpp
transform_feedback.h
video_core.cpp
@ -275,7 +279,7 @@ add_library(video_core STATIC
create_target_directory_groups(video_core)
target_link_libraries(video_core PUBLIC common core)
target_link_libraries(video_core PUBLIC glad shader_recompiler)
target_link_libraries(video_core PUBLIC glad shader_recompiler stb)
if (YUZU_USE_BUNDLED_FFMPEG AND NOT WIN32)
add_dependencies(video_core ffmpeg-build)

2
src/video_core/buffer_cache/buffer_cache.h
View file

@ -1664,7 +1664,7 @@ typename BufferCache<P>::Binding BufferCache<P>::StorageBufferBinding(GPUVAddr s
// cbufs, which do not store the sizes adjacent to the addresses, so use the fully
// mapped buffer size for now.
const u32 memory_layout_size = static_cast<u32>(gpu_memory->GetMemoryLayoutSize(gpu_addr));
return memory_layout_size;
return std::min(memory_layout_size, static_cast<u32>(8_MiB));
}();
const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
if (!cpu_addr || size == 0) {

31
src/video_core/renderer_opengl/gl_texture_cache.cpp
View file

@ -233,6 +233,8 @@ void ApplySwizzle(GLuint handle, PixelFormat format, std::array<SwizzleSource, 4
const VideoCommon::ImageInfo& info) {
if (IsPixelFormatASTC(info.format) && info.size.depth == 1 && !runtime.HasNativeASTC()) {
return Settings::values.accelerate_astc.GetValue() &&
Settings::values.astc_recompression.GetValue() ==
Settings::AstcRecompression::Uncompressed &&
!Settings::values.async_astc.GetValue();
}
// Disable other accelerated uploads for now as they don't implement swizzled uploads
@ -437,6 +439,19 @@ OGLTexture MakeImage(const VideoCommon::ImageInfo& info, GLenum gl_internal_form
return GL_R32UI;
}
[[nodiscard]] GLenum SelectAstcFormat(PixelFormat format, bool is_srgb) {
switch (Settings::values.astc_recompression.GetValue()) {
case Settings::AstcRecompression::Bc1:
return is_srgb ? GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT1_EXT : GL_COMPRESSED_RGBA_S3TC_DXT1_EXT;
break;
case Settings::AstcRecompression::Bc3:
return is_srgb ? GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT5_EXT : GL_COMPRESSED_RGBA_S3TC_DXT5_EXT;
break;
default:
return is_srgb ? GL_SRGB8_ALPHA8 : GL_RGBA8;
}
}
} // Anonymous namespace
ImageBufferMap::~ImageBufferMap() {
@ -739,9 +754,16 @@ Image::Image(TextureCacheRuntime& runtime_, const VideoCommon::ImageInfo& info_,
if (IsConverted(runtime->device, info.format, info.type)) {
flags |= ImageFlagBits::Converted;
flags |= ImageFlagBits::CostlyLoad;
gl_internal_format = IsPixelFormatSRGB(info.format) ? GL_SRGB8_ALPHA8 : GL_RGBA8;
const bool is_srgb = IsPixelFormatSRGB(info.format);
gl_internal_format = is_srgb ? GL_SRGB8_ALPHA8 : GL_RGBA8;
gl_format = GL_RGBA;
gl_type = GL_UNSIGNED_INT_8_8_8_8_REV;
if (IsPixelFormatASTC(info.format)) {
gl_internal_format = SelectAstcFormat(info.format, is_srgb);
gl_format = GL_NONE;
}
} else {
const auto& tuple = MaxwellToGL::GetFormatTuple(info.format);
gl_internal_format = tuple.internal_format;
@ -1130,7 +1152,12 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
views{runtime.null_image_views} {
const Device& device = runtime.device;
if (True(image.flags & ImageFlagBits::Converted)) {
internal_format = IsPixelFormatSRGB(info.format) ? GL_SRGB8_ALPHA8 : GL_RGBA8;
const bool is_srgb = IsPixelFormatSRGB(info.format);
internal_format = is_srgb ? GL_SRGB8_ALPHA8 : GL_RGBA8;
if (IsPixelFormatASTC(info.format)) {
internal_format = SelectAstcFormat(info.format, is_srgb);
}
} else {
internal_format = MaxwellToGL::GetFormatTuple(format).internal_format;
}

24
src/video_core/renderer_vulkan/maxwell_to_vk.cpp
View file

@ -6,6 +6,7 @@
#include "common/assert.h"
#include "common/common_types.h"
#include "common/logging/log.h"
#include "common/settings.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/renderer_vulkan/maxwell_to_vk.h"
#include "video_core/surface.h"
@ -237,14 +238,25 @@ FormatInfo SurfaceFormat(const Device& device, FormatType format_type, bool with
PixelFormat pixel_format) {
ASSERT(static_cast<size_t>(pixel_format) < std::size(tex_format_tuples));
FormatTuple tuple = tex_format_tuples[static_cast<size_t>(pixel_format)];
// Use A8B8G8R8_UNORM on hardware that doesn't support ASTC natively
// Transcode on hardware that doesn't support ASTC natively
if (!device.IsOptimalAstcSupported() && VideoCore::Surface::IsPixelFormatASTC(pixel_format)) {
const bool is_srgb = with_srgb && VideoCore::Surface::IsPixelFormatSRGB(pixel_format);
if (is_srgb) {
tuple.format = VK_FORMAT_A8B8G8R8_SRGB_PACK32;
} else {
tuple.format = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
tuple.usage |= Storage;
switch (Settings::values.astc_recompression.GetValue()) {
case Settings::AstcRecompression::Uncompressed:
if (is_srgb) {
tuple.format = VK_FORMAT_A8B8G8R8_SRGB_PACK32;
} else {
tuple.format = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
tuple.usage |= Storage;
}
break;
case Settings::AstcRecompression::Bc1:
tuple.format = is_srgb ? VK_FORMAT_BC1_RGBA_SRGB_BLOCK : VK_FORMAT_BC1_RGBA_UNORM_BLOCK;
break;
case Settings::AstcRecompression::Bc3:
tuple.format = is_srgb ? VK_FORMAT_BC3_SRGB_BLOCK : VK_FORMAT_BC3_UNORM_BLOCK;
break;
}
}
const bool attachable = (tuple.usage & Attachable) != 0;

8
src/video_core/renderer_vulkan/vk_texture_cache.cpp
View file

@ -1268,7 +1268,9 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
if (IsPixelFormatASTC(info.format) && !runtime->device.IsOptimalAstcSupported()) {
if (Settings::values.async_astc.GetValue()) {
flags |= VideoCommon::ImageFlagBits::AsynchronousDecode;
} else if (Settings::values.accelerate_astc.GetValue() && info.size.depth == 1) {
} else if (Settings::values.astc_recompression.GetValue() ==
Settings::AstcRecompression::Uncompressed &&
Settings::values.accelerate_astc.GetValue() && info.size.depth == 1) {
flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
}
flags |= VideoCommon::ImageFlagBits::Converted;
@ -1283,7 +1285,9 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
.usage = VK_IMAGE_USAGE_STORAGE_BIT,
};
current_image = *original_image;
if (IsPixelFormatASTC(info.format) && !runtime->device.IsOptimalAstcSupported()) {
if (IsPixelFormatASTC(info.format) && !runtime->device.IsOptimalAstcSupported() &&
Settings::values.astc_recompression.GetValue() ==
Settings::AstcRecompression::Uncompressed) {
const auto& device = runtime->device.GetLogical();
storage_image_views.reserve(info.resources.levels);
for (s32 level = 0; level < info.resources.levels; ++level) {

79
src/video_core/texture_cache/util.cpp
View file

@ -18,6 +18,8 @@
#include "common/bit_util.h"
#include "common/common_types.h"
#include "common/div_ceil.h"
#include "common/scratch_buffer.h"
#include "common/settings.h"
#include "video_core/compatible_formats.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/memory_manager.h"
@ -28,6 +30,7 @@
#include "video_core/texture_cache/samples_helper.h"
#include "video_core/texture_cache/util.h"
#include "video_core/textures/astc.h"
#include "video_core/textures/bcn.h"
#include "video_core/textures/decoders.h"
namespace VideoCommon {
@ -585,6 +588,21 @@ u32 CalculateConvertedSizeBytes(const ImageInfo& info) noexcept {
return info.size.width * BytesPerBlock(info.format);
}
static constexpr Extent2D TILE_SIZE{1, 1};
if (IsPixelFormatASTC(info.format) && Settings::values.astc_recompression.GetValue() !=
Settings::AstcRecompression::Uncompressed) {
const u32 bpp_div =
Settings::values.astc_recompression.GetValue() == Settings::AstcRecompression::Bc1 ? 2
: 1;
// NumBlocksPerLayer doesn't account for this correctly, so we have to do it manually.
u32 output_size = 0;
for (s32 i = 0; i < info.resources.levels; i++) {
const auto mip_size = AdjustMipSize(info.size, i);
const u32 plane_dim =
Common::AlignUp(mip_size.width, 4U) * Common::AlignUp(mip_size.height, 4U);
output_size += (plane_dim * info.size.depth * info.resources.layers) / bpp_div;
}
return output_size;
}
return NumBlocksPerLayer(info, TILE_SIZE) * info.resources.layers * CONVERTED_BYTES_PER_BLOCK;
}
@ -885,6 +903,7 @@ BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
std::span<BufferImageCopy> copies) {
u32 output_offset = 0;
Common::ScratchBuffer<u8> decode_scratch;
const Extent2D tile_size = DefaultBlockSize(info.format);
for (BufferImageCopy& copy : copies) {
@ -895,22 +914,58 @@ void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8
ASSERT(copy.image_extent == mip_size);
ASSERT(copy.buffer_row_length == Common::AlignUp(mip_size.width, tile_size.width));
ASSERT(copy.buffer_image_height == Common::AlignUp(mip_size.height, tile_size.height));
if (IsPixelFormatASTC(info.format)) {
Tegra::Texture::ASTC::Decompress(
input.subspan(copy.buffer_offset), copy.image_extent.width,
copy.image_extent.height,
copy.image_subresource.num_layers * copy.image_extent.depth, tile_size.width,
tile_size.height, output.subspan(output_offset));
} else {
DecompressBC4(input.subspan(copy.buffer_offset), copy.image_extent,
output.subspan(output_offset));
}
const auto input_offset = input.subspan(copy.buffer_offset);
copy.buffer_offset = output_offset;
copy.buffer_row_length = mip_size.width;
copy.buffer_image_height = mip_size.height;
output_offset += copy.image_extent.width * copy.image_extent.height *
copy.image_subresource.num_layers * CONVERTED_BYTES_PER_BLOCK;
const auto recompression_setting = Settings::values.astc_recompression.GetValue();
const bool astc = IsPixelFormatASTC(info.format);
if (astc && recompression_setting == Settings::AstcRecompression::Uncompressed) {
Tegra::Texture::ASTC::Decompress(
input_offset, copy.image_extent.width, copy.image_extent.height,
copy.image_subresource.num_layers * copy.image_extent.depth, tile_size.width,
tile_size.height, output.subspan(output_offset));
output_offset += copy.image_extent.width * copy.image_extent.height *
copy.image_subresource.num_layers * CONVERTED_BYTES_PER_BLOCK;
} else if (astc) {
// BC1 uses 0.5 bytes per texel
// BC3 uses 1 byte per texel
const auto compress = recompression_setting == Settings::AstcRecompression::Bc1
? Tegra::Texture::BCN::CompressBC1
: Tegra::Texture::BCN::CompressBC3;
const auto bpp_div = recompression_setting == Settings::AstcRecompression::Bc1 ? 2 : 1;
const u32 plane_dim = copy.image_extent.width * copy.image_extent.height;
const u32 level_size = plane_dim * copy.image_extent.depth *
copy.image_subresource.num_layers * CONVERTED_BYTES_PER_BLOCK;
decode_scratch.resize_destructive(level_size);
Tegra::Texture::ASTC::Decompress(
input_offset, copy.image_extent.width, copy.image_extent.height,
copy.image_subresource.num_layers * copy.image_extent.depth, tile_size.width,
tile_size.height, decode_scratch);
compress(decode_scratch, copy.image_extent.width, copy.image_extent.height,
copy.image_subresource.num_layers * copy.image_extent.depth,
output.subspan(output_offset));
const u32 aligned_plane_dim = Common::AlignUp(copy.image_extent.width, 4) *
Common::AlignUp(copy.image_extent.height, 4);
copy.buffer_size =
(aligned_plane_dim * copy.image_extent.depth * copy.image_subresource.num_layers) /
bpp_div;
output_offset += static_cast<u32>(copy.buffer_size);
} else {
DecompressBC4(input_offset, copy.image_extent, output.subspan(output_offset));
output_offset += copy.image_extent.width * copy.image_extent.height *
copy.image_subresource.num_layers * CONVERTED_BYTES_PER_BLOCK;
}
}
}

5
src/video_core/textures/astc.cpp
View file

@ -16,8 +16,8 @@
#include "common/alignment.h"
#include "common/common_types.h"
#include "common/polyfill_ranges.h"
#include "common/thread_worker.h"
#include "video_core/textures/astc.h"
#include "video_core/textures/workers.h"
class InputBitStream {
public:
@ -1656,8 +1656,7 @@ void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height,
const u32 rows = Common::DivideUp(height, block_height);
const u32 cols = Common::DivideUp(width, block_width);
static Common::ThreadWorker workers{std::max(std::thread::hardware_concurrency(), 2U) / 2,
"ASTCDecompress"};
Common::ThreadWorker& workers{GetThreadWorkers()};
for (u32 z = 0; z < depth; ++z) {
const u32 depth_offset = z * height * width * 4;

87
src/video_core/textures/bcn.cpp Normal file
View file

@ -0,0 +1,87 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <stb_dxt.h>
#include <string.h>
#include "common/alignment.h"
#include "video_core/textures/bcn.h"
#include "video_core/textures/workers.h"
namespace Tegra::Texture::BCN {
using BCNCompressor = void(u8* block_output, const u8* block_input, bool any_alpha);
template <u32 BytesPerBlock, bool ThresholdAlpha = false>
void CompressBCN(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
std::span<uint8_t> output, BCNCompressor f) {
constexpr u8 alpha_threshold = 128;
constexpr u32 bytes_per_px = 4;
const u32 plane_dim = width * height;
Common::ThreadWorker& workers{GetThreadWorkers()};
for (u32 z = 0; z < depth; z++) {
for (u32 y = 0; y < height; y += 4) {
auto compress_row = [z, y, width, height, plane_dim, f, data, output]() {
for (u32 x = 0; x < width; x += 4) {
// Gather 4x4 block of RGBA texels
u8 input_colors[4][4][4];
bool any_alpha = false;
for (u32 j = 0; j < 4; j++) {
for (u32 i = 0; i < 4; i++) {
const size_t coord =
(z * plane_dim + (y + j) * width + (x + i)) * bytes_per_px;
if ((x + i < width) && (y + j < height)) {
if constexpr (ThresholdAlpha) {
if (data[coord + 3] >= alpha_threshold) {
input_colors[j][i][0] = data[coord + 0];
input_colors[j][i][1] = data[coord + 1];
input_colors[j][i][2] = data[coord + 2];
input_colors[j][i][3] = 255;
} else {
any_alpha = true;
memset(input_colors[j][i], 0, bytes_per_px);
}
} else {
memcpy(input_colors[j][i], &data[coord], bytes_per_px);
}
} else {
memset(input_colors[j][i], 0, bytes_per_px);
}
}
}
const u32 bytes_per_row = BytesPerBlock * Common::DivideUp(width, 4U);
const u32 bytes_per_plane = bytes_per_row * Common::DivideUp(height, 4U);
f(output.data() + z * bytes_per_plane + (y / 4) * bytes_per_row +
(x / 4) * BytesPerBlock,
reinterpret_cast<u8*>(input_colors), any_alpha);
}
};
workers.QueueWork(std::move(compress_row));
}
workers.WaitForRequests();
}
}
void CompressBC1(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
std::span<uint8_t> output) {
CompressBCN<8, true>(data, width, height, depth, output,
[](u8* block_output, const u8* block_input, bool any_alpha) {
stb_compress_bc1_block(block_output, block_input, any_alpha,
STB_DXT_NORMAL);
});
}
void CompressBC3(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
std::span<uint8_t> output) {
CompressBCN<16, false>(data, width, height, depth, output,
[](u8* block_output, const u8* block_input, bool any_alpha) {
stb_compress_bc3_block(block_output, block_input, STB_DXT_NORMAL);
});
}
} // namespace Tegra::Texture::BCN

17
src/video_core/textures/bcn.h Normal file
View file

@ -0,0 +1,17 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include <span>
#include <stdint.h>
namespace Tegra::Texture::BCN {
void CompressBC1(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
std::span<uint8_t> output);
void CompressBC3(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
std::span<uint8_t> output);
} // namespace Tegra::Texture::BCN

15
src/video_core/textures/workers.cpp Normal file
View file

@ -0,0 +1,15 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include "video_core/textures/workers.h"
namespace Tegra::Texture {
Common::ThreadWorker& GetThreadWorkers() {
static Common::ThreadWorker workers{std::max(std::thread::hardware_concurrency(), 2U) / 2,
"ImageTranscode"};
return workers;
}
} // namespace Tegra::Texture

12
src/video_core/textures/workers.h Normal file
View file

@ -0,0 +1,12 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include "common/thread_worker.h"
namespace Tegra::Texture {
Common::ThreadWorker& GetThreadWorkers();
}

6
src/video_core/vulkan_common/vulkan_device.cpp
View file

@ -1001,6 +1001,11 @@ u64 Device::GetDeviceMemoryUsage() const {
}
void Device::CollectPhysicalMemoryInfo() {
// Account for resolution scaling in memory limits
const size_t normal_memory = 6_GiB;
const size_t scaler_memory = 1_GiB * Settings::values.resolution_info.ScaleUp(1);
// Calculate limits using memory budget
VkPhysicalDeviceMemoryBudgetPropertiesEXT budget{};
budget.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT;
const auto mem_info =
@ -1030,6 +1035,7 @@ void Device::CollectPhysicalMemoryInfo() {
if (!is_integrated) {
const u64 reserve_memory = std::min<u64>(device_access_memory / 8, 1_GiB);
device_access_memory -= reserve_memory;
device_access_memory = std::min<u64>(device_access_memory, normal_memory + scaler_memory);
return;
}
const s64 available_memory = static_cast<s64>(device_access_memory - device_initial_usage);