From 7f737b022aaead6cb85504d738b769ebcd557018 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Wed, 15 Sep 2021 01:02:28 -0400 Subject: util_shaders: Unify BGRA copy passes --- .../renderer_opengl/gl_texture_cache.cpp | 2 +- src/video_core/renderer_opengl/gl_texture_cache.h | 8 ++ src/video_core/renderer_opengl/maxwell_to_gl.h | 4 +- src/video_core/renderer_opengl/util_shaders.cpp | 85 ++++++---------------- src/video_core/renderer_opengl/util_shaders.h | 19 +---- 5 files changed, 36 insertions(+), 82 deletions(-) diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp index b0aee6cc1..3ec78d866 100644 --- a/src/video_core/renderer_opengl/gl_texture_cache.cpp +++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp @@ -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) { return false; } - if (IsPixelFormatBGR(dst.info.format) || IsPixelFormatBGR(src.info.format)) { + if (IsPixelFormatBGR(dst.info.format) != IsPixelFormatBGR(src.info.format)) { return false; } return true; diff --git a/src/video_core/renderer_opengl/gl_texture_cache.h b/src/video_core/renderer_opengl/gl_texture_cache.h index 4a4f6301c..990a8ddcb 100644 --- a/src/video_core/renderer_opengl/gl_texture_cache.h +++ b/src/video_core/renderer_opengl/gl_texture_cache.h @@ -162,6 +162,14 @@ public: return texture.handle; } + GLuint GlFormat() const noexcept { + return gl_format; + } + + GLuint GlType() const noexcept { + return gl_type; + } + private: void CopyBufferToImage(const VideoCommon::BufferImageCopy& copy, size_t buffer_offset); diff --git a/src/video_core/renderer_opengl/maxwell_to_gl.h b/src/video_core/renderer_opengl/maxwell_to_gl.h index 672f94bfc..39158aa3e 100644 --- a/src/video_core/renderer_opengl/maxwell_to_gl.h +++ b/src/video_core/renderer_opengl/maxwell_to_gl.h @@ -52,7 +52,7 @@ constexpr std::array FORMAT_TAB {GL_COMPRESSED_RGB_BPTC_UNSIGNED_FLOAT}, // BC6H_UFLOAT {GL_COMPRESSED_RGB_BPTC_SIGNED_FLOAT}, // BC6H_SFLOAT {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_RGBA32I, GL_RGBA_INTEGER, GL_INT}, // R32G32B32A32_SINT {GL_RG32F, GL_RG, GL_FLOAT}, // R32G32_FLOAT @@ -81,7 +81,7 @@ constexpr std::array FORMAT_TAB {GL_COMPRESSED_RGBA_ASTC_8x8_KHR}, // ASTC_2D_8X8_UNORM {GL_COMPRESSED_RGBA_ASTC_8x5_KHR}, // ASTC_2D_8X5_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_DXT3_EXT}, // BC2_SRGB {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT5_EXT}, // BC3_SRGB diff --git a/src/video_core/renderer_opengl/util_shaders.cpp b/src/video_core/renderer_opengl/util_shaders.cpp index 333f35a1c..2cb6e19b3 100644 --- a/src/video_core/renderer_opengl/util_shaders.cpp +++ b/src/video_core/renderer_opengl/util_shaders.cpp @@ -257,38 +257,31 @@ void UtilShaders::CopyBC4(Image& dst_image, Image& src_image, std::span copies) { - static constexpr GLuint BINDING_INPUT_IMAGE = 0; - static constexpr GLuint BINDING_OUTPUT_IMAGE = 1; static constexpr VideoCommon::Offset3D zero_offset{0, 0, 0}; - const u32 bytes_per_block = BytesPerBlock(dst_image.info.format); - switch (bytes_per_block) { - case 2: - // BGR565 copy - for (const ImageCopy& copy : copies) { - ASSERT(copy.src_offset == zero_offset); - ASSERT(copy.dst_offset == zero_offset); - bgr_copy_pass.Execute(dst_image, src_image, copy); - } - break; - case 4: { - // BGRA8 copy - program_manager.BindComputeProgram(copy_bgra_program.handle); - constexpr GLenum FORMAT = GL_RGBA8; - for (const ImageCopy& copy : copies) { - ASSERT(copy.src_offset == zero_offset); - ASSERT(copy.dst_offset == zero_offset); - glBindImageTexture(BINDING_INPUT_IMAGE, src_image.StorageHandle(), - copy.src_subresource.base_level, GL_FALSE, 0, GL_READ_ONLY, FORMAT); - glBindImageTexture(BINDING_OUTPUT_IMAGE, dst_image.StorageHandle(), - copy.dst_subresource.base_level, GL_FALSE, 0, GL_WRITE_ONLY, FORMAT); - glDispatchCompute(copy.extent.width, copy.extent.height, copy.extent.depth); + for (const ImageCopy& copy : copies) { + ASSERT(copy.src_offset == zero_offset); + ASSERT(copy.dst_offset == zero_offset); + + if (bgr_pbo_size < dst_image.unswizzled_size_bytes) { + bgr_pbo.Create(); + bgr_pbo_size = dst_image.unswizzled_size_bytes; + glNamedBufferData(bgr_pbo.handle, bgr_pbo_size, nullptr, GL_STREAM_COPY); } - program_manager.RestoreGuestCompute(); - break; - } - default: - UNREACHABLE(); - break; + // Copy from source to PBO + glPixelStorei(GL_PACK_ALIGNMENT, 1); + glPixelStorei(GL_PACK_ROW_LENGTH, copy.extent.width); + glBindBuffer(GL_PIXEL_PACK_BUFFER, bgr_pbo.handle); + glGetTextureSubImage(src_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height, + copy.src_subresource.num_layers, src_image.GlFormat(), + src_image.GlType(), static_cast(bgr_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, bgr_pbo.handle); + glTextureSubImage3D(dst_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height, + copy.dst_subresource.num_layers, dst_image.GlFormat(), + dst_image.GlType(), nullptr); } } @@ -309,36 +302,4 @@ GLenum StoreFormat(u32 bytes_per_block) { 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(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 diff --git a/src/video_core/renderer_opengl/util_shaders.h b/src/video_core/renderer_opengl/util_shaders.h index ef881e35f..b474480cb 100644 --- a/src/video_core/renderer_opengl/util_shaders.h +++ b/src/video_core/renderer_opengl/util_shaders.h @@ -19,22 +19,6 @@ class ProgramManager; 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 { public: explicit UtilShaders(ProgramManager& program_manager); @@ -70,7 +54,8 @@ private: OGLProgram copy_bgra_program; OGLProgram copy_bc4_program; - Bgr565CopyPass bgr_copy_pass; + OGLBuffer bgr_pbo; + size_t bgr_pbo_size{}; }; GLenum StoreFormat(u32 bytes_per_block); -- cgit v1.2.3