nullequal/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

util_shaders: Unify BGRA copy passes

Browse source
This commit is contained in:
ameerj 2021-09-15 01:02:28 -04:00
parent 081ccc6441
commit 7f737b022a
5 changed files with 36 additions and 82 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/video_core/renderer_opengl/gl_texture_cache.cpp
View file

@ -461,7 +461,7 @@ bool TextureCacheRuntime::CanImageBeCopied(const Image& dst, const Image& src) {
if (dst.info.type == ImageType::e3D && dst.info.format == PixelFormat::BC4_UNORM) { if (dst.info.type == ImageType::e3D && dst.info.format == PixelFormat::BC4_UNORM) {
return false; return false;
} }
if (IsPixelFormatBGR(dst.info.format) || IsPixelFormatBGR(src.info.format)) { if (IsPixelFormatBGR(dst.info.format) != IsPixelFormatBGR(src.info.format)) {
return false; return false;
} }
return true; return true;

8
src/video_core/renderer_opengl/gl_texture_cache.h
View file

@ -162,6 +162,14 @@ public:
return texture.handle; return texture.handle;
} }
GLuint GlFormat() const noexcept {
return gl_format;
}
GLuint GlType() const noexcept {
return gl_type;
}
private: private:
void CopyBufferToImage(const VideoCommon::BufferImageCopy& copy, size_t buffer_offset); void CopyBufferToImage(const VideoCommon::BufferImageCopy& copy, size_t buffer_offset);

4
src/video_core/renderer_opengl/maxwell_to_gl.h
View file

@ -52,7 +52,7 @@ constexpr std::array<FormatTuple, VideoCore::Surface::MaxPixelFormat> FORMAT_TAB
{GL_COMPRESSED_RGB_BPTC_UNSIGNED_FLOAT}, // BC6H_UFLOAT {GL_COMPRESSED_RGB_BPTC_UNSIGNED_FLOAT}, // BC6H_UFLOAT
{GL_COMPRESSED_RGB_BPTC_SIGNED_FLOAT}, // BC6H_SFLOAT {GL_COMPRESSED_RGB_BPTC_SIGNED_FLOAT}, // BC6H_SFLOAT
{GL_COMPRESSED_RGBA_ASTC_4x4_KHR}, // ASTC_2D_4X4_UNORM {GL_COMPRESSED_RGBA_ASTC_4x4_KHR}, // ASTC_2D_4X4_UNORM
{GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE}, // B8G8R8A8_UNORM {GL_RGBA8, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV}, // B8G8R8A8_UNORM
{GL_RGBA32F, GL_RGBA, GL_FLOAT}, // R32G32B32A32_FLOAT {GL_RGBA32F, GL_RGBA, GL_FLOAT}, // R32G32B32A32_FLOAT
{GL_RGBA32I, GL_RGBA_INTEGER, GL_INT}, // R32G32B32A32_SINT {GL_RGBA32I, GL_RGBA_INTEGER, GL_INT}, // R32G32B32A32_SINT
{GL_RG32F, GL_RG, GL_FLOAT}, // R32G32_FLOAT {GL_RG32F, GL_RG, GL_FLOAT}, // R32G32_FLOAT
@ -81,7 +81,7 @@ constexpr std::array<FormatTuple, VideoCore::Surface::MaxPixelFormat> FORMAT_TAB
{GL_COMPRESSED_RGBA_ASTC_8x8_KHR}, // ASTC_2D_8X8_UNORM {GL_COMPRESSED_RGBA_ASTC_8x8_KHR}, // ASTC_2D_8X8_UNORM
{GL_COMPRESSED_RGBA_ASTC_8x5_KHR}, // ASTC_2D_8X5_UNORM {GL_COMPRESSED_RGBA_ASTC_8x5_KHR}, // ASTC_2D_8X5_UNORM
{GL_COMPRESSED_RGBA_ASTC_5x4_KHR}, // ASTC_2D_5X4_UNORM {GL_COMPRESSED_RGBA_ASTC_5x4_KHR}, // ASTC_2D_5X4_UNORM
{GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE}, // B8G8R8A8_SRGB {GL_SRGB8_ALPHA8, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV}, // B8G8R8A8_SRGB
{GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT1_EXT}, // BC1_RGBA_SRGB {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT1_EXT}, // BC1_RGBA_SRGB
{GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT3_EXT}, // BC2_SRGB {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT3_EXT}, // BC2_SRGB
{GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT5_EXT}, // BC3_SRGB {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT5_EXT}, // BC3_SRGB

85
src/video_core/renderer_opengl/util_shaders.cpp
View file

@ -257,38 +257,31 @@ void UtilShaders::CopyBC4(Image& dst_image, Image& src_image, std::span<const Im
void UtilShaders::CopyBGR(Image& dst_image, Image& src_image, void UtilShaders::CopyBGR(Image& dst_image, Image& src_image,
std::span<const VideoCommon::ImageCopy> copies) { std::span<const VideoCommon::ImageCopy> copies) {
static constexpr GLuint BINDING_INPUT_IMAGE = 0;
static constexpr GLuint BINDING_OUTPUT_IMAGE = 1;
static constexpr VideoCommon::Offset3D zero_offset{0, 0, 0}; static constexpr VideoCommon::Offset3D zero_offset{0, 0, 0};
const u32 bytes_per_block = BytesPerBlock(dst_image.info.format); for (const ImageCopy& copy : copies) {
switch (bytes_per_block) { ASSERT(copy.src_offset == zero_offset);
case 2: ASSERT(copy.dst_offset == zero_offset);
// BGR565 copy
for (const ImageCopy& copy : copies) { if (bgr_pbo_size < dst_image.unswizzled_size_bytes) {
ASSERT(copy.src_offset == zero_offset); bgr_pbo.Create();
ASSERT(copy.dst_offset == zero_offset); bgr_pbo_size = dst_image.unswizzled_size_bytes;
bgr_copy_pass.Execute(dst_image, src_image, copy); glNamedBufferData(bgr_pbo.handle, bgr_pbo_size, nullptr, GL_STREAM_COPY);
} }
break; // Copy from source to PBO
case 4: { glPixelStorei(GL_PACK_ALIGNMENT, 1);
// BGRA8 copy glPixelStorei(GL_PACK_ROW_LENGTH, copy.extent.width);
program_manager.BindComputeProgram(copy_bgra_program.handle); glBindBuffer(GL_PIXEL_PACK_BUFFER, bgr_pbo.handle);
constexpr GLenum FORMAT = GL_RGBA8; glGetTextureSubImage(src_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height,
for (const ImageCopy& copy : copies) { copy.src_subresource.num_layers, src_image.GlFormat(),
ASSERT(copy.src_offset == zero_offset); src_image.GlType(), static_cast<GLsizei>(bgr_pbo_size), nullptr);
ASSERT(copy.dst_offset == zero_offset);
glBindImageTexture(BINDING_INPUT_IMAGE, src_image.StorageHandle(), // Copy from PBO to destination in reverse order
copy.src_subresource.base_level, GL_FALSE, 0, GL_READ_ONLY, FORMAT); glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
glBindImageTexture(BINDING_OUTPUT_IMAGE, dst_image.StorageHandle(), glPixelStorei(GL_UNPACK_ROW_LENGTH, copy.extent.width);
copy.dst_subresource.base_level, GL_FALSE, 0, GL_WRITE_ONLY, FORMAT); glBindBuffer(GL_PIXEL_UNPACK_BUFFER, bgr_pbo.handle);
glDispatchCompute(copy.extent.width, copy.extent.height, copy.extent.depth); glTextureSubImage3D(dst_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height,
} copy.dst_subresource.num_layers, dst_image.GlFormat(),
program_manager.RestoreGuestCompute(); dst_image.GlType(), nullptr);
break;
}
default:
UNREACHABLE();
break;
} }
} }
@ -309,36 +302,4 @@ GLenum StoreFormat(u32 bytes_per_block) {
return GL_R8UI; return GL_R8UI;
} }
void Bgr565CopyPass::Execute(const Image& dst_image, const Image& src_image,
const ImageCopy& copy) {
if (CopyBufferCreationNeeded(copy)) {
CreateNewCopyBuffer(copy, GL_TEXTURE_2D_ARRAY, GL_RGB565);
}
// Copy from source to PBO
glPixelStorei(GL_PACK_ALIGNMENT, 1);
glPixelStorei(GL_PACK_ROW_LENGTH, copy.extent.width);
glBindBuffer(GL_PIXEL_PACK_BUFFER, bgr16_pbo.handle);
glGetTextureSubImage(src_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height,
copy.src_subresource.num_layers, GL_RGB, GL_UNSIGNED_SHORT_5_6_5,
static_cast<GLsizei>(bgr16_pbo_size), nullptr);
// Copy from PBO to destination in reverse order
glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
glPixelStorei(GL_UNPACK_ROW_LENGTH, copy.extent.width);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, bgr16_pbo.handle);
glTextureSubImage3D(dst_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height,
copy.dst_subresource.num_layers, GL_RGB, GL_UNSIGNED_SHORT_5_6_5_REV,
nullptr);
}
bool Bgr565CopyPass::CopyBufferCreationNeeded(const ImageCopy& copy) {
return bgr16_pbo_size < NumPixelsInCopy(copy) * sizeof(u16);
}
void Bgr565CopyPass::CreateNewCopyBuffer(const ImageCopy& copy, GLenum target, GLuint format) {
bgr16_pbo.Create();
bgr16_pbo_size = NumPixelsInCopy(copy) * sizeof(u16);
glNamedBufferData(bgr16_pbo.handle, bgr16_pbo_size, nullptr, GL_STREAM_COPY);
}
} // namespace OpenGL } // namespace OpenGL

19
src/video_core/renderer_opengl/util_shaders.h
View file

@ -19,22 +19,6 @@ class ProgramManager;
struct ImageBufferMap; struct ImageBufferMap;
class Bgr565CopyPass {
public:
Bgr565CopyPass() = default;
~Bgr565CopyPass() = default;
void Execute(const Image& dst_image, const Image& src_image,
const VideoCommon::ImageCopy& copy);
private:
[[nodiscard]] bool CopyBufferCreationNeeded(const VideoCommon::ImageCopy& copy);
void CreateNewCopyBuffer(const VideoCommon::ImageCopy& copy, GLenum target, GLuint format);
OGLBuffer bgr16_pbo;
size_t bgr16_pbo_size{};
};
class UtilShaders { class UtilShaders {
public: public:
explicit UtilShaders(ProgramManager& program_manager); explicit UtilShaders(ProgramManager& program_manager);
@ -70,7 +54,8 @@ private:
OGLProgram copy_bgra_program; OGLProgram copy_bgra_program;
OGLProgram copy_bc4_program; OGLProgram copy_bc4_program;
Bgr565CopyPass bgr_copy_pass; OGLBuffer bgr_pbo;
size_t bgr_pbo_size{};
}; };
GLenum StoreFormat(u32 bytes_per_block); GLenum StoreFormat(u32 bytes_per_block);