summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat (limited to '')
-rw-r--r--src/video_core/host_shaders/CMakeLists.txt1
-rw-r--r--src/video_core/host_shaders/opengl_copy_bgra.comp15
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.cpp20
-rw-r--r--src/video_core/renderer_opengl/util_shaders.cpp76
-rw-r--r--src/video_core/renderer_opengl/util_shaders.h22
5 files changed, 132 insertions, 2 deletions
diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index 970120acc..3494318ca 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -5,6 +5,7 @@ set(SHADER_FILES
5 convert_float_to_depth.frag 5 convert_float_to_depth.frag
6 full_screen_triangle.vert 6 full_screen_triangle.vert
7 opengl_copy_bc4.comp 7 opengl_copy_bc4.comp
8 opengl_copy_bgra.comp
8 opengl_present.frag 9 opengl_present.frag
9 opengl_present.vert 10 opengl_present.vert
10 pitch_unswizzle.comp 11 pitch_unswizzle.comp
diff --git a/src/video_core/host_shaders/opengl_copy_bgra.comp b/src/video_core/host_shaders/opengl_copy_bgra.comp
new file mode 100644
index 000000000..2571a4abf
--- /dev/null
+++ b/src/video_core/host_shaders/opengl_copy_bgra.comp
@@ -0,0 +1,15 @@
1// Copyright 2021 yuzu Emulator Project
2// Licensed under GPLv2 or any later version
3// Refer to the license.txt file included.
4
5#version 430 core
6
7layout (local_size_x = 4, local_size_y = 4) in;
8
9layout(binding = 0, rgba8) readonly uniform image2DArray bgr_input;
10layout(binding = 1, rgba8) writeonly uniform image2DArray bgr_output;
11
12void main() {
13 vec4 color = imageLoad(bgr_input, ivec3(gl_GlobalInvocationID));
14 imageStore(bgr_output, ivec3(gl_GlobalInvocationID), color.bgra);
15}
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 12434db67..e028677e9 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -96,7 +96,7 @@ constexpr std::array<FormatTuple, MaxPixelFormat> FORMAT_TABLE = {{
96 {GL_COMPRESSED_RGB_BPTC_UNSIGNED_FLOAT}, // BC6H_UFLOAT 96 {GL_COMPRESSED_RGB_BPTC_UNSIGNED_FLOAT}, // BC6H_UFLOAT
97 {GL_COMPRESSED_RGB_BPTC_SIGNED_FLOAT}, // BC6H_SFLOAT 97 {GL_COMPRESSED_RGB_BPTC_SIGNED_FLOAT}, // BC6H_SFLOAT
98 {GL_COMPRESSED_RGBA_ASTC_4x4_KHR}, // ASTC_2D_4X4_UNORM 98 {GL_COMPRESSED_RGBA_ASTC_4x4_KHR}, // ASTC_2D_4X4_UNORM
99 {GL_RGBA8, GL_BGRA, GL_UNSIGNED_BYTE}, // B8G8R8A8_UNORM 99 {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE}, // B8G8R8A8_UNORM
100 {GL_RGBA32F, GL_RGBA, GL_FLOAT}, // R32G32B32A32_FLOAT 100 {GL_RGBA32F, GL_RGBA, GL_FLOAT}, // R32G32B32A32_FLOAT
101 {GL_RGBA32I, GL_RGBA_INTEGER, GL_INT}, // R32G32B32A32_SINT 101 {GL_RGBA32I, GL_RGBA_INTEGER, GL_INT}, // R32G32B32A32_SINT
102 {GL_RG32F, GL_RG, GL_FLOAT}, // R32G32_FLOAT 102 {GL_RG32F, GL_RG, GL_FLOAT}, // R32G32_FLOAT
@@ -125,7 +125,7 @@ constexpr std::array<FormatTuple, MaxPixelFormat> FORMAT_TABLE = {{
125 {GL_COMPRESSED_RGBA_ASTC_8x8_KHR}, // ASTC_2D_8X8_UNORM 125 {GL_COMPRESSED_RGBA_ASTC_8x8_KHR}, // ASTC_2D_8X8_UNORM
126 {GL_COMPRESSED_RGBA_ASTC_8x5_KHR}, // ASTC_2D_8X5_UNORM 126 {GL_COMPRESSED_RGBA_ASTC_8x5_KHR}, // ASTC_2D_8X5_UNORM
127 {GL_COMPRESSED_RGBA_ASTC_5x4_KHR}, // ASTC_2D_5X4_UNORM 127 {GL_COMPRESSED_RGBA_ASTC_5x4_KHR}, // ASTC_2D_5X4_UNORM
128 {GL_SRGB8_ALPHA8, GL_BGRA, GL_UNSIGNED_BYTE}, // B8G8R8A8_UNORM 128 {GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE}, // B8G8R8A8_SRGB
129 {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT1_EXT}, // BC1_RGBA_SRGB 129 {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT1_EXT}, // BC1_RGBA_SRGB
130 {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT3_EXT}, // BC2_SRGB 130 {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT3_EXT}, // BC2_SRGB
131 {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT5_EXT}, // BC3_SRGB 131 {GL_COMPRESSED_SRGB_ALPHA_S3TC_DXT5_EXT}, // BC3_SRGB
@@ -396,6 +396,17 @@ void AttachTexture(GLuint fbo, GLenum attachment, const ImageView* image_view) {
396 } 396 }
397} 397}
398 398
399[[nodiscard]] bool IsPixelFormatBGR(PixelFormat format) {
400 switch (format) {
401 case PixelFormat::B5G6R5_UNORM:
402 case PixelFormat::B8G8R8A8_UNORM:
403 case PixelFormat::B8G8R8A8_SRGB:
404 return true;
405 default:
406 return false;
407 }
408}
409
399} // Anonymous namespace 410} // Anonymous namespace
400 411
401ImageBufferMap::~ImageBufferMap() { 412ImageBufferMap::~ImageBufferMap() {
@@ -512,6 +523,9 @@ bool TextureCacheRuntime::CanImageBeCopied(const Image& dst, const Image& src) {
512 if (dst.info.type == ImageType::e3D && dst.info.format == PixelFormat::BC4_UNORM) { 523 if (dst.info.type == ImageType::e3D && dst.info.format == PixelFormat::BC4_UNORM) {
513 return false; 524 return false;
514 } 525 }
526 if (IsPixelFormatBGR(dst.info.format) || IsPixelFormatBGR(src.info.format)) {
527 return false;
528 }
515 return true; 529 return true;
516} 530}
517 531
@@ -520,6 +534,8 @@ void TextureCacheRuntime::EmulateCopyImage(Image& dst, Image& src,
520 if (dst.info.type == ImageType::e3D && dst.info.format == PixelFormat::BC4_UNORM) { 534 if (dst.info.type == ImageType::e3D && dst.info.format == PixelFormat::BC4_UNORM) {
521 ASSERT(src.info.type == ImageType::e3D); 535 ASSERT(src.info.type == ImageType::e3D);
522 util_shaders.CopyBC4(dst, src, copies); 536 util_shaders.CopyBC4(dst, src, copies);
537 } else if (IsPixelFormatBGR(dst.info.format) || IsPixelFormatBGR(src.info.format)) {
538 util_shaders.CopyBGR(dst, src, copies);
523 } else { 539 } else {
524 UNREACHABLE(); 540 UNREACHABLE();
525 } 541 }
diff --git a/src/video_core/renderer_opengl/util_shaders.cpp b/src/video_core/renderer_opengl/util_shaders.cpp
index 31ec68505..2fe4799bc 100644
--- a/src/video_core/renderer_opengl/util_shaders.cpp
+++ b/src/video_core/renderer_opengl/util_shaders.cpp
@@ -14,6 +14,7 @@
14#include "video_core/host_shaders/block_linear_unswizzle_2d_comp.h" 14#include "video_core/host_shaders/block_linear_unswizzle_2d_comp.h"
15#include "video_core/host_shaders/block_linear_unswizzle_3d_comp.h" 15#include "video_core/host_shaders/block_linear_unswizzle_3d_comp.h"
16#include "video_core/host_shaders/opengl_copy_bc4_comp.h" 16#include "video_core/host_shaders/opengl_copy_bc4_comp.h"
17#include "video_core/host_shaders/opengl_copy_bgra_comp.h"
17#include "video_core/host_shaders/pitch_unswizzle_comp.h" 18#include "video_core/host_shaders/pitch_unswizzle_comp.h"
18#include "video_core/renderer_opengl/gl_resource_manager.h" 19#include "video_core/renderer_opengl/gl_resource_manager.h"
19#include "video_core/renderer_opengl/gl_shader_manager.h" 20#include "video_core/renderer_opengl/gl_shader_manager.h"
@@ -48,6 +49,11 @@ OGLProgram MakeProgram(std::string_view source) {
48 return program; 49 return program;
49} 50}
50 51
52size_t NumPixelsInCopy(const VideoCommon::ImageCopy& copy) {
53 return static_cast<size_t>(copy.extent.width * copy.extent.height *
54 copy.src_subresource.num_layers);
55}
56
51} // Anonymous namespace 57} // Anonymous namespace
52 58
53UtilShaders::UtilShaders(ProgramManager& program_manager_) 59UtilShaders::UtilShaders(ProgramManager& program_manager_)
@@ -55,6 +61,7 @@ UtilShaders::UtilShaders(ProgramManager& program_manager_)
55 block_linear_unswizzle_2d_program(MakeProgram(BLOCK_LINEAR_UNSWIZZLE_2D_COMP)), 61 block_linear_unswizzle_2d_program(MakeProgram(BLOCK_LINEAR_UNSWIZZLE_2D_COMP)),
56 block_linear_unswizzle_3d_program(MakeProgram(BLOCK_LINEAR_UNSWIZZLE_3D_COMP)), 62 block_linear_unswizzle_3d_program(MakeProgram(BLOCK_LINEAR_UNSWIZZLE_3D_COMP)),
57 pitch_unswizzle_program(MakeProgram(PITCH_UNSWIZZLE_COMP)), 63 pitch_unswizzle_program(MakeProgram(PITCH_UNSWIZZLE_COMP)),
64 copy_bgra_program(MakeProgram(OPENGL_COPY_BGRA_COMP)),
58 copy_bc4_program(MakeProgram(OPENGL_COPY_BC4_COMP)) { 65 copy_bc4_program(MakeProgram(OPENGL_COPY_BC4_COMP)) {
59 const auto swizzle_table = Tegra::Texture::MakeSwizzleTable(); 66 const auto swizzle_table = Tegra::Texture::MakeSwizzleTable();
60 swizzle_table_buffer.Create(); 67 swizzle_table_buffer.Create();
@@ -205,6 +212,43 @@ void UtilShaders::CopyBC4(Image& dst_image, Image& src_image, std::span<const Im
205 program_manager.RestoreGuestCompute(); 212 program_manager.RestoreGuestCompute();
206} 213}
207 214
215void UtilShaders::CopyBGR(Image& dst_image, Image& src_image,
216 std::span<const VideoCommon::ImageCopy> copies) {
217 static constexpr GLuint BINDING_INPUT_IMAGE = 0;
218 static constexpr GLuint BINDING_OUTPUT_IMAGE = 1;
219 static constexpr VideoCommon::Offset3D zero_offset{0, 0, 0};
220 const u32 bytes_per_block = BytesPerBlock(dst_image.info.format);
221 switch (bytes_per_block) {
222 case 2:
223 // BGR565 copy
224 for (const ImageCopy& copy : copies) {
225 ASSERT(copy.src_offset == zero_offset);
226 ASSERT(copy.dst_offset == zero_offset);
227 bgr_copy_pass.Execute(dst_image, src_image, copy);
228 }
229 break;
230 case 4: {
231 // BGRA8 copy
232 program_manager.BindHostCompute(copy_bgra_program.handle);
233 constexpr GLenum FORMAT = GL_RGBA8;
234 for (const ImageCopy& copy : copies) {
235 ASSERT(copy.src_offset == zero_offset);
236 ASSERT(copy.dst_offset == zero_offset);
237 glBindImageTexture(BINDING_INPUT_IMAGE, src_image.StorageHandle(),
238 copy.src_subresource.base_level, GL_FALSE, 0, GL_READ_ONLY, FORMAT);
239 glBindImageTexture(BINDING_OUTPUT_IMAGE, dst_image.StorageHandle(),
240 copy.dst_subresource.base_level, GL_FALSE, 0, GL_WRITE_ONLY, FORMAT);
241 glDispatchCompute(copy.extent.width, copy.extent.height, copy.extent.depth);
242 }
243 program_manager.RestoreGuestCompute();
244 break;
245 }
246 default:
247 UNREACHABLE();
248 break;
249 }
250}
251
208GLenum StoreFormat(u32 bytes_per_block) { 252GLenum StoreFormat(u32 bytes_per_block) {
209 switch (bytes_per_block) { 253 switch (bytes_per_block) {
210 case 1: 254 case 1:
@@ -222,4 +266,36 @@ GLenum StoreFormat(u32 bytes_per_block) {
222 return GL_R8UI; 266 return GL_R8UI;
223} 267}
224 268
269void Bgr565CopyPass::Execute(const Image& dst_image, const Image& src_image,
270 const ImageCopy& copy) {
271 if (CopyBufferCreationNeeded(copy)) {
272 CreateNewCopyBuffer(copy, GL_TEXTURE_2D_ARRAY, GL_RGB565);
273 }
274 // Copy from source to PBO
275 glPixelStorei(GL_PACK_ALIGNMENT, 1);
276 glPixelStorei(GL_PACK_ROW_LENGTH, copy.extent.width);
277 glBindBuffer(GL_PIXEL_PACK_BUFFER, bgr16_pbo.handle);
278 glGetTextureSubImage(src_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height,
279 copy.src_subresource.num_layers, GL_RGB, GL_UNSIGNED_SHORT_5_6_5,
280 static_cast<GLsizei>(bgr16_pbo_size), nullptr);
281
282 // Copy from PBO to destination in reverse order
283 glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
284 glPixelStorei(GL_UNPACK_ROW_LENGTH, copy.extent.width);
285 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, bgr16_pbo.handle);
286 glTextureSubImage3D(dst_image.Handle(), 0, 0, 0, 0, copy.extent.width, copy.extent.height,
287 copy.dst_subresource.num_layers, GL_RGB, GL_UNSIGNED_SHORT_5_6_5_REV,
288 nullptr);
289}
290
291bool Bgr565CopyPass::CopyBufferCreationNeeded(const ImageCopy& copy) {
292 return bgr16_pbo_size < NumPixelsInCopy(copy) * sizeof(u16);
293}
294
295void Bgr565CopyPass::CreateNewCopyBuffer(const ImageCopy& copy, GLenum target, GLuint format) {
296 bgr16_pbo.Create();
297 bgr16_pbo_size = NumPixelsInCopy(copy) * sizeof(u16);
298 glNamedBufferData(bgr16_pbo.handle, bgr16_pbo_size, nullptr, GL_STREAM_COPY);
299}
300
225} // namespace OpenGL 301} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/util_shaders.h b/src/video_core/renderer_opengl/util_shaders.h
index 7b1d16b09..93b009743 100644
--- a/src/video_core/renderer_opengl/util_shaders.h
+++ b/src/video_core/renderer_opengl/util_shaders.h
@@ -19,6 +19,22 @@ class ProgramManager;
19 19
20struct ImageBufferMap; 20struct ImageBufferMap;
21 21
22class Bgr565CopyPass {
23public:
24 Bgr565CopyPass() = default;
25 ~Bgr565CopyPass() = default;
26
27 void Execute(const Image& dst_image, const Image& src_image,
28 const VideoCommon::ImageCopy& copy);
29
30private:
31 [[nodiscard]] bool CopyBufferCreationNeeded(const VideoCommon::ImageCopy& copy);
32 void CreateNewCopyBuffer(const VideoCommon::ImageCopy& copy, GLenum target, GLuint format);
33
34 OGLBuffer bgr16_pbo;
35 size_t bgr16_pbo_size{};
36};
37
22class UtilShaders { 38class UtilShaders {
23public: 39public:
24 explicit UtilShaders(ProgramManager& program_manager); 40 explicit UtilShaders(ProgramManager& program_manager);
@@ -36,6 +52,9 @@ public:
36 void CopyBC4(Image& dst_image, Image& src_image, 52 void CopyBC4(Image& dst_image, Image& src_image,
37 std::span<const VideoCommon::ImageCopy> copies); 53 std::span<const VideoCommon::ImageCopy> copies);
38 54
55 void CopyBGR(Image& dst_image, Image& src_image,
56 std::span<const VideoCommon::ImageCopy> copies);
57
39private: 58private:
40 ProgramManager& program_manager; 59 ProgramManager& program_manager;
41 60
@@ -44,7 +63,10 @@ private:
44 OGLProgram block_linear_unswizzle_2d_program; 63 OGLProgram block_linear_unswizzle_2d_program;
45 OGLProgram block_linear_unswizzle_3d_program; 64 OGLProgram block_linear_unswizzle_3d_program;
46 OGLProgram pitch_unswizzle_program; 65 OGLProgram pitch_unswizzle_program;
66 OGLProgram copy_bgra_program;
47 OGLProgram copy_bc4_program; 67 OGLProgram copy_bc4_program;
68
69 Bgr565CopyPass bgr_copy_pass;
48}; 70};
49 71
50GLenum StoreFormat(u32 bytes_per_block); 72GLenum StoreFormat(u32 bytes_per_block);