summaryrefslogtreecommitdiffstats
path: root/src/video_core/renderer_opengl
diff options
context:
space:
mode:
authorameerj <52414509+ameerj@users.noreply.github.com>2021-03-25 21:53:51 +0100
committerameerj <52414509+ameerj@users.noreply.github.com>2021-03-25 21:53:51 +0100
commit2f83d9a61bca42d9ef24074beb2b11b19bd4cecd (patch)
tree514e40eb750280c2e3025f9301befb6f8c9b46e9 /src/video_core/renderer_opengl
parentastc_decoder: Reimplement Layers (diff)
downloadyuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.tar
yuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.tar.gz
yuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.tar.bz2
yuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.tar.lz
yuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.tar.xz
yuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.tar.zst
yuzu-2f83d9a61bca42d9ef24074beb2b11b19bd4cecd.zip
Diffstat (limited to 'src/video_core/renderer_opengl')
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.cpp11
-rw-r--r--src/video_core/renderer_opengl/util_shaders.cpp96
-rw-r--r--src/video_core/renderer_opengl/util_shaders.h8
3 files changed, 46 insertions, 69 deletions
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 29105ecad..623b43d8a 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -307,7 +307,7 @@ void ApplySwizzle(GLuint handle, PixelFormat format, std::array<SwizzleSource, 4
[[nodiscard]] bool CanBeAccelerated(const TextureCacheRuntime& runtime,
const VideoCommon::ImageInfo& info) {
- return (!runtime.HasNativeASTC() && IsPixelFormatASTC(info.format));
+ return !runtime.HasNativeASTC() && IsPixelFormatASTC(info.format);
// Disable other accelerated uploads for now as they don't implement swizzled uploads
return false;
switch (info.type) {
@@ -568,12 +568,13 @@ void TextureCacheRuntime::BlitFramebuffer(Framebuffer* dst, Framebuffer* src,
void TextureCacheRuntime::AccelerateImageUpload(Image& image, const ImageBufferMap& map,
std::span<const SwizzleParameters> swizzles) {
- if (IsPixelFormatASTC(image.info.format)) {
- return util_shaders.ASTCDecode(image, map, swizzles);
- }
switch (image.info.type) {
case ImageType::e2D:
- return util_shaders.BlockLinearUpload2D(image, map, swizzles);
+ if (IsPixelFormatASTC(image.info.format)) {
+ return util_shaders.ASTCDecode(image, map, swizzles);
+ } else {
+ return util_shaders.BlockLinearUpload2D(image, map, swizzles);
+ }
case ImageType::e3D:
return util_shaders.BlockLinearUpload3D(image, map, swizzles);
case ImageType::Linear:
diff --git a/src/video_core/renderer_opengl/util_shaders.cpp b/src/video_core/renderer_opengl/util_shaders.cpp
index 85722c54a..47fddcb6e 100644
--- a/src/video_core/renderer_opengl/util_shaders.cpp
+++ b/src/video_core/renderer_opengl/util_shaders.cpp
@@ -2,11 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
-#include <bit>
-#include <fstream>
#include <span>
-#include <streambuf>
-#include <string>
#include <string_view>
#include <glad/glad.h>
@@ -24,7 +20,6 @@
#include "video_core/renderer_opengl/gl_shader_manager.h"
#include "video_core/renderer_opengl/gl_texture_cache.h"
#include "video_core/renderer_opengl/util_shaders.h"
-#include "video_core/surface.h"
#include "video_core/texture_cache/accelerated_swizzle.h"
#include "video_core/texture_cache/types.h"
#include "video_core/texture_cache/util.h"
@@ -36,6 +31,7 @@ namespace OpenGL {
using namespace HostShaders;
using namespace Tegra::Texture::ASTC;
+using VideoCommon::Extent2D;
using VideoCommon::Extent3D;
using VideoCommon::ImageCopy;
using VideoCommon::ImageType;
@@ -69,33 +65,15 @@ UtilShaders::UtilShaders(ProgramManager& program_manager_)
pitch_unswizzle_program(MakeProgram(PITCH_UNSWIZZLE_COMP)),
copy_bgra_program(MakeProgram(OPENGL_COPY_BGRA_COMP)),
copy_bc4_program(MakeProgram(OPENGL_COPY_BC4_COMP)) {
- MakeBuffers();
-}
-
-UtilShaders::~UtilShaders() = default;
-
-void UtilShaders::MakeBuffers() {
const auto swizzle_table = Tegra::Texture::MakeSwizzleTable();
swizzle_table_buffer.Create();
+ astc_buffer.Create();
glNamedBufferStorage(swizzle_table_buffer.handle, sizeof(swizzle_table), &swizzle_table, 0);
-
- astc_encodings_buffer.Create();
- glNamedBufferStorage(astc_encodings_buffer.handle, sizeof(EncodingsValues), &EncodingsValues,
- 0);
- replicate_6_to_8_buffer.Create();
- glNamedBufferStorage(replicate_6_to_8_buffer.handle, sizeof(REPLICATE_6_BIT_TO_8_TABLE),
- &REPLICATE_6_BIT_TO_8_TABLE, 0);
- replicate_7_to_8_buffer.Create();
- glNamedBufferStorage(replicate_7_to_8_buffer.handle, sizeof(REPLICATE_7_BIT_TO_8_TABLE),
- &REPLICATE_7_BIT_TO_8_TABLE, 0);
- replicate_8_to_8_buffer.Create();
- glNamedBufferStorage(replicate_8_to_8_buffer.handle, sizeof(REPLICATE_8_BIT_TO_8_TABLE),
- &REPLICATE_8_BIT_TO_8_TABLE, 0);
- replicate_byte_to_16_buffer.Create();
- glNamedBufferStorage(replicate_byte_to_16_buffer.handle, sizeof(REPLICATE_BYTE_TO_16_TABLE),
- &REPLICATE_BYTE_TO_16_TABLE, 0);
+ glNamedBufferStorage(astc_buffer.handle, sizeof(ASTC_BUFFER_DATA), &ASTC_BUFFER_DATA, 0);
}
+UtilShaders::~UtilShaders() = default;
+
void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
std::span<const VideoCommon::SwizzleParameters> swizzles) {
static constexpr GLuint BINDING_SWIZZLE_BUFFER = 0;
@@ -108,47 +86,51 @@ void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
static constexpr GLuint BINDING_BYTE_TO_16_BUFFER = 6;
static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
- static constexpr GLuint LOC_NUM_IMAGE_BLOCKS = 0;
- static constexpr GLuint LOC_BLOCK_DIMS = 1;
- const Extent3D tile_size = {
- VideoCore::Surface::DefaultBlockWidth(image.info.format),
- VideoCore::Surface::DefaultBlockHeight(image.info.format),
+ const Extent2D tile_size{
+ .width = VideoCore::Surface::DefaultBlockWidth(image.info.format),
+ .height = VideoCore::Surface::DefaultBlockHeight(image.info.format),
};
program_manager.BindHostCompute(astc_decoder_program.handle);
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_SWIZZLE_BUFFER, swizzle_table_buffer.handle);
- glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_ENC_BUFFER, astc_encodings_buffer.handle);
- glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_6_TO_8_BUFFER,
- replicate_6_to_8_buffer.handle);
- glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_7_TO_8_BUFFER,
- replicate_7_to_8_buffer.handle);
- glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_8_TO_8_BUFFER,
- replicate_8_to_8_buffer.handle);
- glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_BYTE_TO_16_BUFFER,
- replicate_byte_to_16_buffer.handle);
+ glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_ENC_BUFFER, astc_buffer.handle,
+ offsetof(AstcBufferData, encoding_values),
+ sizeof(AstcBufferData::encoding_values));
+ glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_6_TO_8_BUFFER, astc_buffer.handle,
+ offsetof(AstcBufferData, replicate_6_to_8),
+ sizeof(AstcBufferData::replicate_6_to_8));
+ glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_7_TO_8_BUFFER, astc_buffer.handle,
+ offsetof(AstcBufferData, replicate_7_to_8),
+ sizeof(AstcBufferData::replicate_7_to_8));
+ glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_8_TO_8_BUFFER, astc_buffer.handle,
+ offsetof(AstcBufferData, replicate_8_to_8),
+ sizeof(AstcBufferData::replicate_8_to_8));
+ glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_BYTE_TO_16_BUFFER, astc_buffer.handle,
+ offsetof(AstcBufferData, replicate_byte_to_16),
+ sizeof(AstcBufferData::replicate_byte_to_16));
glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
- glUniform2ui(LOC_BLOCK_DIMS, tile_size.width, tile_size.height);
+ glUniform2ui(1, tile_size.width, tile_size.height);
+ // Ensure buffer data is valid before dispatching
+ glFlush();
for (const SwizzleParameters& swizzle : swizzles) {
- glBindImageTexture(BINDING_OUTPUT_IMAGE, image.StorageHandle(), swizzle.level, GL_TRUE, 0,
- GL_WRITE_ONLY, GL_RGBA8);
const size_t input_offset = swizzle.buffer_offset + map.offset;
- const auto num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
- const auto num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
-
- glUniform2ui(LOC_NUM_IMAGE_BLOCKS, swizzle.num_tiles.width, swizzle.num_tiles.height);
+ const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
+ const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
- // To unswizzle the ASTC data
const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
- glUniform3uiv(2, 1, params.origin.data());
- glUniform3iv(3, 1, params.destination.data());
- glUniform1ui(4, params.bytes_per_block_log2);
- glUniform1ui(5, params.layer_stride);
- glUniform1ui(6, params.block_size);
- glUniform1ui(7, params.x_shift);
- glUniform1ui(8, params.block_height);
- glUniform1ui(9, params.block_height_mask);
+ ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0}));
+ ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0}));
+ glUniform1ui(2, params.bytes_per_block_log2);
+ glUniform1ui(3, params.layer_stride);
+ glUniform1ui(4, params.block_size);
+ glUniform1ui(5, params.x_shift);
+ glUniform1ui(6, params.block_height);
+ glUniform1ui(7, params.block_height_mask);
+
+ glBindImageTexture(BINDING_OUTPUT_IMAGE, image.StorageHandle(), swizzle.level, GL_TRUE, 0,
+ GL_WRITE_ONLY, GL_RGBA8);
// ASTC texture data
glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,
image.guest_size_bytes - swizzle.buffer_offset);
diff --git a/src/video_core/renderer_opengl/util_shaders.h b/src/video_core/renderer_opengl/util_shaders.h
index 08a1cb9b2..53d65f368 100644
--- a/src/video_core/renderer_opengl/util_shaders.h
+++ b/src/video_core/renderer_opengl/util_shaders.h
@@ -40,8 +40,6 @@ public:
explicit UtilShaders(ProgramManager& program_manager);
~UtilShaders();
- void MakeBuffers();
-
void ASTCDecode(Image& image, const ImageBufferMap& map,
std::span<const VideoCommon::SwizzleParameters> swizzles);
@@ -64,11 +62,7 @@ private:
ProgramManager& program_manager;
OGLBuffer swizzle_table_buffer;
- OGLBuffer astc_encodings_buffer;
- OGLBuffer replicate_6_to_8_buffer;
- OGLBuffer replicate_7_to_8_buffer;
- OGLBuffer replicate_8_to_8_buffer;
- OGLBuffer replicate_byte_to_16_buffer;
+ OGLBuffer astc_buffer;
OGLProgram astc_decoder_program;
OGLProgram block_linear_unswizzle_2d_program;