summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--src/common/bit_util.h7
-rw-r--r--src/video_core/host_shaders/CMakeLists.txt6
-rw-r--r--src/video_core/host_shaders/convert_abgr8_to_d24s8.frag17
-rw-r--r--src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag19
-rw-r--r--src/video_core/host_shaders/convert_d24s8_to_abgr8.frag21
-rw-r--r--src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag21
-rw-r--r--src/video_core/host_shaders/convert_d24s8_to_r16g16.frag21
-rw-r--r--src/video_core/host_shaders/convert_r16g16_to_d24s8.frag18
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.cpp10
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.h7
-rw-r--r--src/video_core/renderer_vulkan/blit_image.cpp171
-rw-r--r--src/video_core/renderer_vulkan/blit_image.h39
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.cpp199
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.h11
-rw-r--r--src/video_core/texture_cache/texture_cache.h19
-rw-r--r--src/video_core/texture_cache/texture_cache_base.h2
-rw-r--r--src/video_core/texture_cache/types.h1
-rw-r--r--src/video_core/texture_cache/util.cpp29
18 files changed, 595 insertions, 23 deletions
diff --git a/src/common/bit_util.h b/src/common/bit_util.h
index 64520ca4e..eef8c1c5a 100644
--- a/src/common/bit_util.h
+++ b/src/common/bit_util.h
@@ -7,6 +7,7 @@
7#include <bit> 7#include <bit>
8#include <climits> 8#include <climits>
9#include <cstddef> 9#include <cstddef>
10#include <type_traits>
10 11
11#include "common/common_types.h" 12#include "common/common_types.h"
12 13
@@ -44,4 +45,10 @@ template <typename T>
44 return static_cast<u32>(log2_f + static_cast<u64>((value ^ (1ULL << log2_f)) != 0ULL)); 45 return static_cast<u32>(log2_f + static_cast<u64>((value ^ (1ULL << log2_f)) != 0ULL));
45} 46}
46 47
48template <typename T>
49requires std::is_integral_v<T>
50[[nodiscard]] T NextPow2(T value) {
51 return static_cast<T>(1ULL << ((8U * sizeof(T)) - std::countl_zero(value - 1U)));
52}
53
47} // namespace Common 54} // namespace Common
diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index d779a967a..1c91999d7 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -10,8 +10,14 @@ set(SHADER_FILES
10 astc_decoder.comp 10 astc_decoder.comp
11 block_linear_unswizzle_2d.comp 11 block_linear_unswizzle_2d.comp
12 block_linear_unswizzle_3d.comp 12 block_linear_unswizzle_3d.comp
13 convert_abgr8_to_d24s8.frag
14 convert_b10g11r11_to_d24s8.frag
15 convert_d24s8_to_abgr8.frag
16 convert_d24s8_to_b10g11r11.frag
17 convert_d24s8_to_r16g16.frag
13 convert_depth_to_float.frag 18 convert_depth_to_float.frag
14 convert_float_to_depth.frag 19 convert_float_to_depth.frag
20 convert_r16g16_to_d24s8.frag
15 full_screen_triangle.vert 21 full_screen_triangle.vert
16 fxaa.frag 22 fxaa.frag
17 fxaa.vert 23 fxaa.vert
diff --git a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
new file mode 100644
index 000000000..4e4ab6a26
--- /dev/null
+++ b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
@@ -0,0 +1,17 @@
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 450
6#extension GL_ARB_shader_stencil_export : require
7
8layout(binding = 0) uniform sampler2D color_texture;
9
10void main() {
11 ivec2 coord = ivec2(gl_FragCoord.xy);
12 uvec4 color = uvec4(texelFetch(color_texture, coord, 0).rgba * (exp2(8) - 1.0f));
13 uint depth_unorm = (color.r << 16) | (color.g << 8) | color.b;
14
15 gl_FragDepth = float(depth_unorm) / (exp2(24.0) - 1.0f);
16 gl_FragStencilRefARB = int(color.a);
17}
diff --git a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
new file mode 100644
index 000000000..2999a84cf
--- /dev/null
+++ b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
@@ -0,0 +1,19 @@
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 450
6#extension GL_ARB_shader_stencil_export : require
7
8layout(binding = 0) uniform sampler2D color_texture;
9
10void main() {
11 ivec2 coord = ivec2(gl_FragCoord.xy);
12 vec4 color = texelFetch(color_texture, coord, 0).rgba;
13 uint depth_stencil_unorm = (uint(color.b * (exp2(10) - 1.0f)) << 22)
14 | (uint(color.g * (exp2(11) - 1.0f)) << 11)
15 | (uint(color.r * (exp2(11) - 1.0f)));
16
17 gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
18 gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
19}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
new file mode 100644
index 000000000..ff3bf8209
--- /dev/null
+++ b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
@@ -0,0 +1,21 @@
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 450
6
7layout(binding = 0) uniform sampler2D depth_tex;
8layout(binding = 1) uniform isampler2D stencil_tex;
9
10layout(location = 0) out vec4 color;
11
12void main() {
13 ivec2 coord = ivec2(gl_FragCoord.xy);
14 uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
15 uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
16
17 color.r = float(depth >> 16) / (exp2(8) - 1.0);
18 color.g = float((depth >> 8) & 0x00FF) / (exp2(8) - 1.0);
19 color.b = float(depth & 0x00FF) / (exp2(8) - 1.0);
20 color.a = float(stencil) / (exp2(8) - 1.0);
21}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
new file mode 100644
index 000000000..c743d3a13
--- /dev/null
+++ b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
@@ -0,0 +1,21 @@
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 450
6
7layout(binding = 0) uniform sampler2D depth_tex;
8layout(binding = 1) uniform isampler2D stencil_tex;
9
10layout(location = 0) out vec4 color;
11
12void main() {
13 ivec2 coord = ivec2(gl_FragCoord.xy);
14 uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
15 uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
16
17 color.b = float(depth >> 22) / (exp2(10) - 1.0);
18 color.g = float((depth >> 11) & 0x00FF) / (exp2(11) - 1.0);
19 color.r = float(depth & 0x00FF) / (exp2(11) - 1.0);
20 color.a = 1.0f;
21}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
new file mode 100644
index 000000000..2a9443d3d
--- /dev/null
+++ b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
@@ -0,0 +1,21 @@
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 450
6
7layout(binding = 0) uniform sampler2D depth_tex;
8layout(binding = 1) uniform isampler2D stencil_tex;
9
10layout(location = 0) out vec4 color;
11
12void main() {
13 ivec2 coord = ivec2(gl_FragCoord.xy);
14 uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
15 uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
16
17 color.r = float(depth >> 16) / (exp2(16) - 1.0);
18 color.g = float((depth >> 16) & 0x00FF) / (exp2(16) - 1.0);
19 color.b = 0.0f;
20 color.a = 1.0f;
21}
diff --git a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
new file mode 100644
index 000000000..3df70575e
--- /dev/null
+++ b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
@@ -0,0 +1,18 @@
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 450
6#extension GL_ARB_shader_stencil_export : require
7
8layout(binding = 0) uniform sampler2D color_texture;
9
10void main() {
11 ivec2 coord = ivec2(gl_FragCoord.xy);
12 vec4 color = texelFetch(color_texture, coord, 0).rgba;
13 uint depth_stencil_unorm = (uint(color.r * (exp2(16) - 1.0f)) << 16)
14 | (uint(color.g * (exp2(16) - 1.0f)) << 16);
15
16 gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
17 gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
18}
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 3dbdcc2c4..14e6522f2 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -9,6 +9,7 @@
9 9
10#include <glad/glad.h> 10#include <glad/glad.h>
11 11
12#include "common/bit_util.h"
12#include "common/literals.h" 13#include "common/literals.h"
13#include "common/settings.h" 14#include "common/settings.h"
14#include "video_core/renderer_opengl/gl_device.h" 15#include "video_core/renderer_opengl/gl_device.h"
@@ -398,9 +399,6 @@ OGLTexture MakeImage(const VideoCommon::ImageInfo& info, GLenum gl_internal_form
398 return GL_R32UI; 399 return GL_R32UI;
399} 400}
400 401
401[[nodiscard]] u32 NextPow2(u32 value) {
402 return 1U << (32U - std::countl_zero(value - 1U));
403}
404} // Anonymous namespace 402} // Anonymous namespace
405 403
406ImageBufferMap::~ImageBufferMap() { 404ImageBufferMap::~ImageBufferMap() {
@@ -527,8 +525,8 @@ void TextureCacheRuntime::CopyImage(Image& dst_image, Image& src_image,
527 } 525 }
528} 526}
529 527
530void TextureCacheRuntime::ConvertImage(Image& dst, Image& src, 528void TextureCacheRuntime::ReinterpretImage(Image& dst, Image& src,
531 std::span<const VideoCommon::ImageCopy> copies) { 529 std::span<const VideoCommon::ImageCopy> copies) {
532 LOG_DEBUG(Render_OpenGL, "Converting {} to {}", src.info.format, dst.info.format); 530 LOG_DEBUG(Render_OpenGL, "Converting {} to {}", src.info.format, dst.info.format);
533 format_conversion_pass.ConvertImage(dst, src, copies); 531 format_conversion_pass.ConvertImage(dst, src, copies);
534} 532}
@@ -1333,7 +1331,7 @@ void FormatConversionPass::ConvertImage(Image& dst_image, Image& src_image,
1333 const u32 copy_size = region.width * region.height * region.depth * img_bpp; 1331 const u32 copy_size = region.width * region.height * region.depth * img_bpp;
1334 if (pbo_size < copy_size) { 1332 if (pbo_size < copy_size) {
1335 intermediate_pbo.Create(); 1333 intermediate_pbo.Create();
1336 pbo_size = NextPow2(copy_size); 1334 pbo_size = Common::NextPow2(copy_size);
1337 glNamedBufferData(intermediate_pbo.handle, pbo_size, nullptr, GL_STREAM_COPY); 1335 glNamedBufferData(intermediate_pbo.handle, pbo_size, nullptr, GL_STREAM_COPY);
1338 } 1336 }
1339 // Copy from source to PBO 1337 // Copy from source to PBO
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.h b/src/video_core/renderer_opengl/gl_texture_cache.h
index c0534b1f1..37d5e6a6b 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.h
+++ b/src/video_core/renderer_opengl/gl_texture_cache.h
@@ -84,9 +84,13 @@ public:
84 84
85 u64 GetDeviceLocalMemory() const; 85 u64 GetDeviceLocalMemory() const;
86 86
87 bool ShouldReinterpret([[maybe_unused]] Image& dst, [[maybe_unused]] Image& src) {
88 return true;
89 }
90
87 void CopyImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies); 91 void CopyImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies);
88 92
89 void ConvertImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies); 93 void ReinterpretImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies);
90 94
91 void ConvertImage(Framebuffer* dst, ImageView& dst_view, ImageView& src_view, bool rescaled) { 95 void ConvertImage(Framebuffer* dst, ImageView& dst_view, ImageView& src_view, bool rescaled) {
92 UNIMPLEMENTED(); 96 UNIMPLEMENTED();
@@ -339,7 +343,6 @@ struct TextureCacheParams {
339 static constexpr bool FRAMEBUFFER_BLITS = true; 343 static constexpr bool FRAMEBUFFER_BLITS = true;
340 static constexpr bool HAS_EMULATED_COPIES = true; 344 static constexpr bool HAS_EMULATED_COPIES = true;
341 static constexpr bool HAS_DEVICE_MEMORY_INFO = true; 345 static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
342 static constexpr bool HAS_PIXEL_FORMAT_CONVERSIONS = true;
343 346
344 using Runtime = OpenGL::TextureCacheRuntime; 347 using Runtime = OpenGL::TextureCacheRuntime;
345 using Image = OpenGL::Image; 348 using Image = OpenGL::Image;
diff --git a/src/video_core/renderer_vulkan/blit_image.cpp b/src/video_core/renderer_vulkan/blit_image.cpp
index b3884a4f5..28b631f73 100644
--- a/src/video_core/renderer_vulkan/blit_image.cpp
+++ b/src/video_core/renderer_vulkan/blit_image.cpp
@@ -4,8 +4,14 @@
4 4
5#include <algorithm> 5#include <algorithm>
6 6
7#include "video_core/host_shaders/convert_abgr8_to_d24s8_frag_spv.h"
8#include "video_core/host_shaders/convert_b10g11r11_to_d24s8_frag_spv.h"
9#include "video_core/host_shaders/convert_d24s8_to_abgr8_frag_spv.h"
10#include "video_core/host_shaders/convert_d24s8_to_b10g11r11_frag_spv.h"
11#include "video_core/host_shaders/convert_d24s8_to_r16g16_frag_spv.h"
7#include "video_core/host_shaders/convert_depth_to_float_frag_spv.h" 12#include "video_core/host_shaders/convert_depth_to_float_frag_spv.h"
8#include "video_core/host_shaders/convert_float_to_depth_frag_spv.h" 13#include "video_core/host_shaders/convert_float_to_depth_frag_spv.h"
14#include "video_core/host_shaders/convert_r16g16_to_d24s8_frag_spv.h"
9#include "video_core/host_shaders/full_screen_triangle_vert_spv.h" 15#include "video_core/host_shaders/full_screen_triangle_vert_spv.h"
10#include "video_core/host_shaders/vulkan_blit_color_float_frag_spv.h" 16#include "video_core/host_shaders/vulkan_blit_color_float_frag_spv.h"
11#include "video_core/host_shaders/vulkan_blit_depth_stencil_frag_spv.h" 17#include "video_core/host_shaders/vulkan_blit_depth_stencil_frag_spv.h"
@@ -354,6 +360,12 @@ BlitImageHelper::BlitImageHelper(const Device& device_, VKScheduler& scheduler_,
354 blit_color_to_color_frag(BuildShader(device, VULKAN_BLIT_COLOR_FLOAT_FRAG_SPV)), 360 blit_color_to_color_frag(BuildShader(device, VULKAN_BLIT_COLOR_FLOAT_FRAG_SPV)),
355 convert_depth_to_float_frag(BuildShader(device, CONVERT_DEPTH_TO_FLOAT_FRAG_SPV)), 361 convert_depth_to_float_frag(BuildShader(device, CONVERT_DEPTH_TO_FLOAT_FRAG_SPV)),
356 convert_float_to_depth_frag(BuildShader(device, CONVERT_FLOAT_TO_DEPTH_FRAG_SPV)), 362 convert_float_to_depth_frag(BuildShader(device, CONVERT_FLOAT_TO_DEPTH_FRAG_SPV)),
363 convert_abgr8_to_d24s8_frag(BuildShader(device, CONVERT_ABGR8_TO_D24S8_FRAG_SPV)),
364 convert_b10g11r11_to_d24s8_frag(BuildShader(device, CONVERT_B10G11R11_TO_D24S8_FRAG_SPV)),
365 convert_r16g16_to_d24s8_frag(BuildShader(device, CONVERT_R16G16_TO_D24S8_FRAG_SPV)),
366 convert_d24s8_to_abgr8_frag(BuildShader(device, CONVERT_D24S8_TO_ABGR8_FRAG_SPV)),
367 convert_d24s8_to_b10g11r11_frag(BuildShader(device, CONVERT_D24S8_TO_B10G11R11_FRAG_SPV)),
368 convert_d24s8_to_r16g16_frag(BuildShader(device, CONVERT_D24S8_TO_R16G16_FRAG_SPV)),
357 linear_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_LINEAR>)), 369 linear_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_LINEAR>)),
358 nearest_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_NEAREST>)) { 370 nearest_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_NEAREST>)) {
359 if (device.IsExtShaderStencilExportSupported()) { 371 if (device.IsExtShaderStencilExportSupported()) {
@@ -448,6 +460,59 @@ void BlitImageHelper::ConvertR16ToD16(const Framebuffer* dst_framebuffer,
448 Convert(*convert_r16_to_d16_pipeline, dst_framebuffer, src_image_view, up_scale, down_shift); 460 Convert(*convert_r16_to_d16_pipeline, dst_framebuffer, src_image_view, up_scale, down_shift);
449} 461}
450 462
463void BlitImageHelper::ConvertABGR8ToD24S8(const Framebuffer* dst_framebuffer,
464 const ImageView& src_image_view, u32 up_scale,
465 u32 down_shift) {
466 ConvertPipelineDepthTargetEx(convert_abgr8_to_d24s8_pipeline, dst_framebuffer->RenderPass(),
467 convert_abgr8_to_d24s8_frag, true);
468 Convert(*convert_abgr8_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
469 down_shift);
470}
471
472void BlitImageHelper::ConvertB10G11R11ToD24S8(const Framebuffer* dst_framebuffer,
473 const ImageView& src_image_view, u32 up_scale,
474 u32 down_shift) {
475 ConvertPipelineDepthTargetEx(convert_b10g11r11_to_d24s8_pipeline, dst_framebuffer->RenderPass(),
476 convert_b10g11r11_to_d24s8_frag, true);
477 Convert(*convert_b10g11r11_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
478 down_shift);
479}
480
481void BlitImageHelper::ConvertR16G16ToD24S8(const Framebuffer* dst_framebuffer,
482 const ImageView& src_image_view, u32 up_scale,
483 u32 down_shift) {
484 ConvertPipelineDepthTargetEx(convert_r16g16_to_d24s8_pipeline, dst_framebuffer->RenderPass(),
485 convert_r16g16_to_d24s8_frag, true);
486 Convert(*convert_r16g16_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
487 down_shift);
488}
489
490void BlitImageHelper::ConvertD24S8ToABGR8(const Framebuffer* dst_framebuffer,
491 ImageView& src_image_view, u32 up_scale, u32 down_shift) {
492 ConvertPipelineColorTargetEx(convert_d24s8_to_abgr8_pipeline, dst_framebuffer->RenderPass(),
493 convert_d24s8_to_abgr8_frag, false);
494 ConvertDepthStencil(*convert_d24s8_to_abgr8_pipeline, dst_framebuffer, src_image_view, up_scale,
495 down_shift);
496}
497
498void BlitImageHelper::ConvertD24S8ToB10G11R11(const Framebuffer* dst_framebuffer,
499 ImageView& src_image_view, u32 up_scale,
500 u32 down_shift) {
501 ConvertPipelineColorTargetEx(convert_d24s8_to_b10g11r11_pipeline, dst_framebuffer->RenderPass(),
502 convert_d24s8_to_b10g11r11_frag, false);
503 ConvertDepthStencil(*convert_d24s8_to_b10g11r11_pipeline, dst_framebuffer, src_image_view,
504 up_scale, down_shift);
505}
506
507void BlitImageHelper::ConvertD24S8ToR16G16(const Framebuffer* dst_framebuffer,
508 ImageView& src_image_view, u32 up_scale,
509 u32 down_shift) {
510 ConvertPipelineColorTargetEx(convert_d24s8_to_r16g16_pipeline, dst_framebuffer->RenderPass(),
511 convert_d24s8_to_r16g16_frag, false);
512 ConvertDepthStencil(*convert_d24s8_to_r16g16_pipeline, dst_framebuffer, src_image_view,
513 up_scale, down_shift);
514}
515
451void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer, 516void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
452 const ImageView& src_image_view, u32 up_scale, u32 down_shift) { 517 const ImageView& src_image_view, u32 up_scale, u32 down_shift) {
453 const VkPipelineLayout layout = *one_texture_pipeline_layout; 518 const VkPipelineLayout layout = *one_texture_pipeline_layout;
@@ -495,6 +560,54 @@ void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_frameb
495 scheduler.InvalidateState(); 560 scheduler.InvalidateState();
496} 561}
497 562
563void BlitImageHelper::ConvertDepthStencil(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
564 ImageView& src_image_view, u32 up_scale, u32 down_shift) {
565 const VkPipelineLayout layout = *two_textures_pipeline_layout;
566 const VkImageView src_depth_view = src_image_view.DepthView();
567 const VkImageView src_stencil_view = src_image_view.StencilView();
568 const VkSampler sampler = *nearest_sampler;
569 const VkExtent2D extent{
570 .width = std::max((src_image_view.size.width * up_scale) >> down_shift, 1U),
571 .height = std::max((src_image_view.size.height * up_scale) >> down_shift, 1U),
572 };
573 scheduler.RequestRenderpass(dst_framebuffer);
574 scheduler.Record([pipeline, layout, sampler, src_depth_view, src_stencil_view, extent, up_scale,
575 down_shift, this](vk::CommandBuffer cmdbuf) {
576 const VkOffset2D offset{
577 .x = 0,
578 .y = 0,
579 };
580 const VkViewport viewport{
581 .x = 0.0f,
582 .y = 0.0f,
583 .width = static_cast<float>(extent.width),
584 .height = static_cast<float>(extent.height),
585 .minDepth = 0.0f,
586 .maxDepth = 0.0f,
587 };
588 const VkRect2D scissor{
589 .offset = offset,
590 .extent = extent,
591 };
592 const PushConstants push_constants{
593 .tex_scale = {viewport.width, viewport.height},
594 .tex_offset = {0.0f, 0.0f},
595 };
596 const VkDescriptorSet descriptor_set = two_textures_descriptor_allocator.Commit();
597 UpdateTwoTexturesDescriptorSet(device, descriptor_set, sampler, src_depth_view,
598 src_stencil_view);
599 // TODO: Barriers
600 cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
601 cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, layout, 0, descriptor_set,
602 nullptr);
603 cmdbuf.SetViewport(0, viewport);
604 cmdbuf.SetScissor(0, scissor);
605 cmdbuf.PushConstants(layout, VK_SHADER_STAGE_VERTEX_BIT, push_constants);
606 cmdbuf.Draw(3, 1, 0, 0);
607 });
608 scheduler.InvalidateState();
609}
610
498VkPipeline BlitImageHelper::FindOrEmplaceColorPipeline(const BlitImagePipelineKey& key) { 611VkPipeline BlitImageHelper::FindOrEmplaceColorPipeline(const BlitImagePipelineKey& key) {
499 const auto it = std::ranges::find(blit_color_keys, key); 612 const auto it = std::ranges::find(blit_color_keys, key);
500 if (it != blit_color_keys.end()) { 613 if (it != blit_color_keys.end()) {
@@ -636,4 +749,62 @@ void BlitImageHelper::ConvertColorToDepthPipeline(vk::Pipeline& pipeline, VkRend
636 }); 749 });
637} 750}
638 751
752void BlitImageHelper::ConvertPipelineColorTargetEx(vk::Pipeline& pipeline, VkRenderPass renderpass,
753 vk::ShaderModule& module, bool single_texture) {
754 if (pipeline) {
755 return;
756 }
757 const std::array stages = MakeStages(*full_screen_vert, *module);
758 pipeline = device.GetLogical().CreateGraphicsPipeline({
759 .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
760 .pNext = nullptr,
761 .flags = 0,
762 .stageCount = static_cast<u32>(stages.size()),
763 .pStages = stages.data(),
764 .pVertexInputState = &PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
765 .pInputAssemblyState = &PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
766 .pTessellationState = nullptr,
767 .pViewportState = &PIPELINE_VIEWPORT_STATE_CREATE_INFO,
768 .pRasterizationState = &PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
769 .pMultisampleState = &PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
770 .pDepthStencilState = nullptr,
771 .pColorBlendState = &PIPELINE_COLOR_BLEND_STATE_GENERIC_CREATE_INFO,
772 .pDynamicState = &PIPELINE_DYNAMIC_STATE_CREATE_INFO,
773 .layout = single_texture ? *one_texture_pipeline_layout : *two_textures_pipeline_layout,
774 .renderPass = renderpass,
775 .subpass = 0,
776 .basePipelineHandle = VK_NULL_HANDLE,
777 .basePipelineIndex = 0,
778 });
779}
780
781void BlitImageHelper::ConvertPipelineDepthTargetEx(vk::Pipeline& pipeline, VkRenderPass renderpass,
782 vk::ShaderModule& module, bool single_texture) {
783 if (pipeline) {
784 return;
785 }
786 const std::array stages = MakeStages(*full_screen_vert, *module);
787 pipeline = device.GetLogical().CreateGraphicsPipeline({
788 .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
789 .pNext = nullptr,
790 .flags = 0,
791 .stageCount = static_cast<u32>(stages.size()),
792 .pStages = stages.data(),
793 .pVertexInputState = &PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
794 .pInputAssemblyState = &PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
795 .pTessellationState = nullptr,
796 .pViewportState = &PIPELINE_VIEWPORT_STATE_CREATE_INFO,
797 .pRasterizationState = &PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
798 .pMultisampleState = &PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
799 .pDepthStencilState = &PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
800 .pColorBlendState = &PIPELINE_COLOR_BLEND_STATE_EMPTY_CREATE_INFO,
801 .pDynamicState = &PIPELINE_DYNAMIC_STATE_CREATE_INFO,
802 .layout = single_texture ? *one_texture_pipeline_layout : *two_textures_pipeline_layout,
803 .renderPass = renderpass,
804 .subpass = 0,
805 .basePipelineHandle = VK_NULL_HANDLE,
806 .basePipelineIndex = 0,
807 });
808}
809
639} // namespace Vulkan 810} // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/blit_image.h b/src/video_core/renderer_vulkan/blit_image.h
index d77f76678..cec095341 100644
--- a/src/video_core/renderer_vulkan/blit_image.h
+++ b/src/video_core/renderer_vulkan/blit_image.h
@@ -56,10 +56,31 @@ public:
56 void ConvertR16ToD16(const Framebuffer* dst_framebuffer, const ImageView& src_image_view, 56 void ConvertR16ToD16(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
57 u32 up_scale, u32 down_shift); 57 u32 up_scale, u32 down_shift);
58 58
59 void ConvertABGR8ToD24S8(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
60 u32 up_scale, u32 down_shift);
61
62 void ConvertB10G11R11ToD24S8(const Framebuffer* dst_framebuffer,
63 const ImageView& src_image_view, u32 up_scale, u32 down_shift);
64
65 void ConvertR16G16ToD24S8(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
66 u32 up_scale, u32 down_shift);
67
68 void ConvertD24S8ToABGR8(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
69 u32 up_scale, u32 down_shift);
70
71 void ConvertD24S8ToB10G11R11(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
72 u32 up_scale, u32 down_shift);
73
74 void ConvertD24S8ToR16G16(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
75 u32 up_scale, u32 down_shift);
76
59private: 77private:
60 void Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer, 78 void Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
61 const ImageView& src_image_view, u32 up_scale, u32 down_shift); 79 const ImageView& src_image_view, u32 up_scale, u32 down_shift);
62 80
81 void ConvertDepthStencil(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
82 ImageView& src_image_view, u32 up_scale, u32 down_shift);
83
63 [[nodiscard]] VkPipeline FindOrEmplaceColorPipeline(const BlitImagePipelineKey& key); 84 [[nodiscard]] VkPipeline FindOrEmplaceColorPipeline(const BlitImagePipelineKey& key);
64 85
65 [[nodiscard]] VkPipeline FindOrEmplaceDepthStencilPipeline(const BlitImagePipelineKey& key); 86 [[nodiscard]] VkPipeline FindOrEmplaceDepthStencilPipeline(const BlitImagePipelineKey& key);
@@ -68,6 +89,12 @@ private:
68 89
69 void ConvertColorToDepthPipeline(vk::Pipeline& pipeline, VkRenderPass renderpass); 90 void ConvertColorToDepthPipeline(vk::Pipeline& pipeline, VkRenderPass renderpass);
70 91
92 void ConvertPipelineColorTargetEx(vk::Pipeline& pipeline, VkRenderPass renderpass,
93 vk::ShaderModule& module, bool single_texture);
94
95 void ConvertPipelineDepthTargetEx(vk::Pipeline& pipeline, VkRenderPass renderpass,
96 vk::ShaderModule& module, bool single_texture);
97
71 const Device& device; 98 const Device& device;
72 VKScheduler& scheduler; 99 VKScheduler& scheduler;
73 StateTracker& state_tracker; 100 StateTracker& state_tracker;
@@ -83,6 +110,12 @@ private:
83 vk::ShaderModule blit_depth_stencil_frag; 110 vk::ShaderModule blit_depth_stencil_frag;
84 vk::ShaderModule convert_depth_to_float_frag; 111 vk::ShaderModule convert_depth_to_float_frag;
85 vk::ShaderModule convert_float_to_depth_frag; 112 vk::ShaderModule convert_float_to_depth_frag;
113 vk::ShaderModule convert_abgr8_to_d24s8_frag;
114 vk::ShaderModule convert_b10g11r11_to_d24s8_frag;
115 vk::ShaderModule convert_r16g16_to_d24s8_frag;
116 vk::ShaderModule convert_d24s8_to_abgr8_frag;
117 vk::ShaderModule convert_d24s8_to_b10g11r11_frag;
118 vk::ShaderModule convert_d24s8_to_r16g16_frag;
86 vk::Sampler linear_sampler; 119 vk::Sampler linear_sampler;
87 vk::Sampler nearest_sampler; 120 vk::Sampler nearest_sampler;
88 121
@@ -94,6 +127,12 @@ private:
94 vk::Pipeline convert_r32_to_d32_pipeline; 127 vk::Pipeline convert_r32_to_d32_pipeline;
95 vk::Pipeline convert_d16_to_r16_pipeline; 128 vk::Pipeline convert_d16_to_r16_pipeline;
96 vk::Pipeline convert_r16_to_d16_pipeline; 129 vk::Pipeline convert_r16_to_d16_pipeline;
130 vk::Pipeline convert_abgr8_to_d24s8_pipeline;
131 vk::Pipeline convert_b10g11r11_to_d24s8_pipeline;
132 vk::Pipeline convert_r16g16_to_d24s8_pipeline;
133 vk::Pipeline convert_d24s8_to_abgr8_pipeline;
134 vk::Pipeline convert_d24s8_to_b10g11r11_pipeline;
135 vk::Pipeline convert_d24s8_to_r16g16_pipeline;
97}; 136};
98 137
99} // namespace Vulkan 138} // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index 9bc846b94..3964424af 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -8,6 +8,7 @@
8#include <vector> 8#include <vector>
9 9
10#include "common/bit_cast.h" 10#include "common/bit_cast.h"
11#include "common/bit_util.h"
11#include "common/settings.h" 12#include "common/settings.h"
12 13
13#include "video_core/engines/fermi_2d.h" 14#include "video_core/engines/fermi_2d.h"
@@ -313,6 +314,19 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
313 }; 314 };
314} 315}
315 316
317[[nodiscard]] VkBufferImageCopy MakeBufferImageCopy(const VideoCommon::ImageCopy& copy, bool is_src,
318 VkImageAspectFlags aspect_mask) noexcept {
319 return VkBufferImageCopy{
320 .bufferOffset = 0,
321 .bufferRowLength = 0,
322 .bufferImageHeight = 0,
323 .imageSubresource = MakeImageSubresourceLayers(
324 is_src ? copy.src_subresource : copy.dst_subresource, aspect_mask),
325 .imageOffset = MakeOffset3D(is_src ? copy.src_offset : copy.dst_offset),
326 .imageExtent = MakeExtent3D(copy.extent),
327 };
328}
329
316[[maybe_unused]] [[nodiscard]] std::vector<VkBufferCopy> TransformBufferCopies( 330[[maybe_unused]] [[nodiscard]] std::vector<VkBufferCopy> TransformBufferCopies(
317 std::span<const VideoCommon::BufferCopy> copies, size_t buffer_offset) { 331 std::span<const VideoCommon::BufferCopy> copies, size_t buffer_offset) {
318 std::vector<VkBufferCopy> result(copies.size()); 332 std::vector<VkBufferCopy> result(copies.size());
@@ -759,6 +773,163 @@ StagingBufferRef TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
759 return staging_buffer_pool.Request(size, MemoryUsage::Download); 773 return staging_buffer_pool.Request(size, MemoryUsage::Download);
760} 774}
761 775
776bool TextureCacheRuntime::ShouldReinterpret(Image& dst, Image& src) {
777 if (VideoCore::Surface::GetFormatType(dst.info.format) ==
778 VideoCore::Surface::SurfaceType::DepthStencil) {
779 return !device.IsExtShaderStencilExportSupported();
780 }
781 return false;
782}
783
784VkBuffer TextureCacheRuntime::GetTemporaryBuffer(size_t needed_size) {
785 const auto level = (8 * sizeof(size_t)) - std::countl_zero(needed_size - 1ULL);
786 if (buffer_commits[level]) {
787 return *buffers[level];
788 }
789 const auto new_size = Common::NextPow2(needed_size);
790 VkBufferUsageFlags flags = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
791 VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
792 VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT;
793 buffers[level] = device.GetLogical().CreateBuffer({
794 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
795 .pNext = nullptr,
796 .flags = 0,
797 .size = new_size,
798 .usage = flags,
799 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
800 .queueFamilyIndexCount = 0,
801 .pQueueFamilyIndices = nullptr,
802 });
803 buffer_commits[level] = std::make_unique<MemoryCommit>(
804 memory_allocator.Commit(buffers[level], MemoryUsage::DeviceLocal));
805 return *buffers[level];
806}
807
808void TextureCacheRuntime::ReinterpretImage(Image& dst, Image& src,
809 std::span<const VideoCommon::ImageCopy> copies) {
810 std::vector<VkBufferImageCopy> vk_in_copies(copies.size());
811 std::vector<VkBufferImageCopy> vk_out_copies(copies.size());
812 const VkImageAspectFlags src_aspect_mask = src.AspectMask();
813 const VkImageAspectFlags dst_aspect_mask = dst.AspectMask();
814
815 std::ranges::transform(copies, vk_in_copies.begin(), [src_aspect_mask](const auto& copy) {
816 return MakeBufferImageCopy(copy, true, src_aspect_mask);
817 });
818 std::ranges::transform(copies, vk_out_copies.begin(), [dst_aspect_mask](const auto& copy) {
819 return MakeBufferImageCopy(copy, false, dst_aspect_mask);
820 });
821 const u32 img_bpp = BytesPerBlock(src.info.format);
822 size_t total_size = 0;
823 for (const auto& copy : copies) {
824 total_size += copy.extent.width * copy.extent.height * copy.extent.depth * img_bpp;
825 }
826 const VkBuffer copy_buffer = GetTemporaryBuffer(total_size);
827 const VkImage dst_image = dst.Handle();
828 const VkImage src_image = src.Handle();
829 scheduler.RequestOutsideRenderPassOperationContext();
830 scheduler.Record([dst_image, src_image, copy_buffer, src_aspect_mask, dst_aspect_mask,
831 vk_in_copies, vk_out_copies](vk::CommandBuffer cmdbuf) {
832 RangedBarrierRange dst_range;
833 RangedBarrierRange src_range;
834 for (const VkBufferImageCopy& copy : vk_in_copies) {
835 src_range.AddLayers(copy.imageSubresource);
836 }
837 for (const VkBufferImageCopy& copy : vk_out_copies) {
838 dst_range.AddLayers(copy.imageSubresource);
839 }
840 static constexpr VkMemoryBarrier READ_BARRIER{
841 .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
842 .pNext = nullptr,
843 .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
844 .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
845 };
846 static constexpr VkMemoryBarrier WRITE_BARRIER{
847 .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
848 .pNext = nullptr,
849 .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
850 .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
851 };
852 const std::array pre_barriers{
853 VkImageMemoryBarrier{
854 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
855 .pNext = nullptr,
856 .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
857 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
858 VK_ACCESS_TRANSFER_WRITE_BIT,
859 .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
860 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
861 .newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
862 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
863 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
864 .image = src_image,
865 .subresourceRange = src_range.SubresourceRange(src_aspect_mask),
866 },
867 };
868 const std::array middle_in_barrier{
869 VkImageMemoryBarrier{
870 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
871 .pNext = nullptr,
872 .srcAccessMask = 0,
873 .dstAccessMask = 0,
874 .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
875 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
876 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
877 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
878 .image = src_image,
879 .subresourceRange = src_range.SubresourceRange(src_aspect_mask),
880 },
881 };
882 const std::array middle_out_barrier{
883 VkImageMemoryBarrier{
884 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
885 .pNext = nullptr,
886 .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
887 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
888 VK_ACCESS_TRANSFER_WRITE_BIT,
889 .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
890 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
891 .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
892 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
893 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
894 .image = dst_image,
895 .subresourceRange = dst_range.SubresourceRange(dst_aspect_mask),
896 },
897 };
898 const std::array post_barriers{
899 VkImageMemoryBarrier{
900 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
901 .pNext = nullptr,
902 .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
903 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT |
904 VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
905 VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
906 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
907 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
908 VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
909 .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
910 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
911 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
912 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
913 .image = dst_image,
914 .subresourceRange = dst_range.SubresourceRange(dst_aspect_mask),
915 },
916 };
917 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
918 0, {}, {}, pre_barriers);
919
920 cmdbuf.CopyImageToBuffer(src_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, copy_buffer,
921 vk_in_copies);
922 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
923 0, WRITE_BARRIER, nullptr, middle_in_barrier);
924
925 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
926 0, READ_BARRIER, {}, middle_out_barrier);
927 cmdbuf.CopyBufferToImage(copy_buffer, dst_image, VK_IMAGE_LAYOUT_GENERAL, vk_out_copies);
928 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
929 0, {}, {}, post_barriers);
930 });
931}
932
762void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src, 933void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src,
763 const Region2D& dst_region, const Region2D& src_region, 934 const Region2D& dst_region, const Region2D& src_region,
764 Tegra::Engines::Fermi2D::Filter filter, 935 Tegra::Engines::Fermi2D::Filter filter,
@@ -886,6 +1057,22 @@ void TextureCacheRuntime::ConvertImage(Framebuffer* dst, ImageView& dst_view, Im
886 return blit_image_helper.ConvertD16ToR16(dst, src_view, up_scale, down_shift); 1057 return blit_image_helper.ConvertD16ToR16(dst, src_view, up_scale, down_shift);
887 } 1058 }
888 break; 1059 break;
1060 case PixelFormat::A8B8G8R8_UNORM:
1061 case PixelFormat::B8G8R8A8_UNORM:
1062 if (src_view.format == PixelFormat::S8_UINT_D24_UNORM) {
1063 return blit_image_helper.ConvertD24S8ToABGR8(dst, src_view, up_scale, down_shift);
1064 }
1065 break;
1066 case PixelFormat::B10G11R11_FLOAT:
1067 if (src_view.format == PixelFormat::S8_UINT_D24_UNORM) {
1068 return blit_image_helper.ConvertD24S8ToB10G11R11(dst, src_view, up_scale, down_shift);
1069 }
1070 break;
1071 case PixelFormat::R16G16_UNORM:
1072 if (src_view.format == PixelFormat::S8_UINT_D24_UNORM) {
1073 return blit_image_helper.ConvertD24S8ToR16G16(dst, src_view, up_scale, down_shift);
1074 }
1075 break;
889 case PixelFormat::R32_FLOAT: 1076 case PixelFormat::R32_FLOAT:
890 if (src_view.format == PixelFormat::D32_FLOAT) { 1077 if (src_view.format == PixelFormat::D32_FLOAT) {
891 return blit_image_helper.ConvertD32ToR32(dst, src_view, up_scale, down_shift); 1078 return blit_image_helper.ConvertD32ToR32(dst, src_view, up_scale, down_shift);
@@ -896,6 +1083,18 @@ void TextureCacheRuntime::ConvertImage(Framebuffer* dst, ImageView& dst_view, Im
896 return blit_image_helper.ConvertR16ToD16(dst, src_view, up_scale, down_shift); 1083 return blit_image_helper.ConvertR16ToD16(dst, src_view, up_scale, down_shift);
897 } 1084 }
898 break; 1085 break;
1086 case PixelFormat::S8_UINT_D24_UNORM:
1087 if (src_view.format == PixelFormat::A8B8G8R8_UNORM ||
1088 src_view.format == PixelFormat::B8G8R8A8_UNORM) {
1089 return blit_image_helper.ConvertABGR8ToD24S8(dst, src_view, up_scale, down_shift);
1090 }
1091 if (src_view.format == PixelFormat::B10G11R11_FLOAT) {
1092 return blit_image_helper.ConvertB10G11R11ToD24S8(dst, src_view, up_scale, down_shift);
1093 }
1094 if (src_view.format == PixelFormat::R16G16_UNORM) {
1095 return blit_image_helper.ConvertR16G16ToD24S8(dst, src_view, up_scale, down_shift);
1096 }
1097 break;
899 case PixelFormat::D32_FLOAT: 1098 case PixelFormat::D32_FLOAT:
900 if (src_view.format == PixelFormat::R32_FLOAT) { 1099 if (src_view.format == PixelFormat::R32_FLOAT) {
901 return blit_image_helper.ConvertR32ToD32(dst, src_view, up_scale, down_shift); 1100 return blit_image_helper.ConvertR32ToD32(dst, src_view, up_scale, down_shift);
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h
index f5f8f9a74..44e9dcee4 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.h
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.h
@@ -61,6 +61,10 @@ public:
61 61
62 void CopyImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies); 62 void CopyImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies);
63 63
64 bool ShouldReinterpret(Image& dst, Image& src);
65
66 void ReinterpretImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies);
67
64 void ConvertImage(Framebuffer* dst, ImageView& dst_view, ImageView& src_view, bool rescaled); 68 void ConvertImage(Framebuffer* dst, ImageView& dst_view, ImageView& src_view, bool rescaled);
65 69
66 bool CanAccelerateImageUpload(Image&) const noexcept { 70 bool CanAccelerateImageUpload(Image&) const noexcept {
@@ -82,6 +86,8 @@ public:
82 return true; 86 return true;
83 } 87 }
84 88
89 [[nodiscard]] VkBuffer GetTemporaryBuffer(size_t needed_size);
90
85 const Device& device; 91 const Device& device;
86 VKScheduler& scheduler; 92 VKScheduler& scheduler;
87 MemoryAllocator& memory_allocator; 93 MemoryAllocator& memory_allocator;
@@ -90,6 +96,10 @@ public:
90 ASTCDecoderPass& astc_decoder_pass; 96 ASTCDecoderPass& astc_decoder_pass;
91 RenderPassCache& render_pass_cache; 97 RenderPassCache& render_pass_cache;
92 const Settings::ResolutionScalingInfo& resolution; 98 const Settings::ResolutionScalingInfo& resolution;
99
100 constexpr static size_t indexing_slots = 8 * sizeof(size_t);
101 std::array<vk::Buffer, indexing_slots> buffers{};
102 std::array<std::unique_ptr<MemoryCommit>, indexing_slots> buffer_commits{};
93}; 103};
94 104
95class Image : public VideoCommon::ImageBase { 105class Image : public VideoCommon::ImageBase {
@@ -316,7 +326,6 @@ struct TextureCacheParams {
316 static constexpr bool FRAMEBUFFER_BLITS = false; 326 static constexpr bool FRAMEBUFFER_BLITS = false;
317 static constexpr bool HAS_EMULATED_COPIES = false; 327 static constexpr bool HAS_EMULATED_COPIES = false;
318 static constexpr bool HAS_DEVICE_MEMORY_INFO = true; 328 static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
319 static constexpr bool HAS_PIXEL_FORMAT_CONVERSIONS = false;
320 329
321 using Runtime = Vulkan::TextureCacheRuntime; 330 using Runtime = Vulkan::TextureCacheRuntime;
322 using Image = Vulkan::Image; 331 using Image = Vulkan::Image;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 241f71a91..44a0d42ba 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -475,6 +475,7 @@ void TextureCache<P>::BlitImage(const Tegra::Engines::Fermi2D::Surface& dst,
475 const BlitImages images = GetBlitImages(dst, src); 475 const BlitImages images = GetBlitImages(dst, src);
476 const ImageId dst_id = images.dst_id; 476 const ImageId dst_id = images.dst_id;
477 const ImageId src_id = images.src_id; 477 const ImageId src_id = images.src_id;
478
478 PrepareImage(src_id, false, false); 479 PrepareImage(src_id, false, false);
479 PrepareImage(dst_id, true, false); 480 PrepareImage(dst_id, true, false);
480 481
@@ -758,7 +759,8 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
758 return ImageId{}; 759 return ImageId{};
759 } 760 }
760 } 761 }
761 const bool broken_views = runtime.HasBrokenTextureViewFormats(); 762 const bool broken_views =
763 runtime.HasBrokenTextureViewFormats() || True(options & RelaxedOptions::ForceBrokenViews);
762 const bool native_bgr = runtime.HasNativeBgr(); 764 const bool native_bgr = runtime.HasNativeBgr();
763 ImageId image_id; 765 ImageId image_id;
764 const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) { 766 const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
@@ -1094,12 +1096,13 @@ typename TextureCache<P>::BlitImages TextureCache<P>::GetBlitImages(
1094 if (GetFormatType(dst_info.format) != GetFormatType(src_info.format)) { 1096 if (GetFormatType(dst_info.format) != GetFormatType(src_info.format)) {
1095 continue; 1097 continue;
1096 } 1098 }
1097 if (!dst_id) { 1099 RelaxedOptions find_options{};
1098 dst_id = InsertImage(dst_info, dst_addr, RelaxedOptions{}); 1100 if (src_info.num_samples > 1) {
1099 } 1101 // it's a resolve, we must enforce the same format.
1100 if (!src_id) { 1102 find_options = RelaxedOptions::ForceBrokenViews;
1101 src_id = InsertImage(src_info, src_addr, RelaxedOptions{});
1102 } 1103 }
1104 src_id = FindOrInsertImage(src_info, src_addr, find_options);
1105 dst_id = FindOrInsertImage(dst_info, dst_addr, find_options);
1103 } while (has_deleted_images); 1106 } while (has_deleted_images);
1104 return BlitImages{ 1107 return BlitImages{
1105 .dst_id = dst_id, 1108 .dst_id = dst_id,
@@ -1759,8 +1762,8 @@ void TextureCache<P>::CopyImage(ImageId dst_id, ImageId src_id, std::vector<Imag
1759 } 1762 }
1760 UNIMPLEMENTED_IF(dst.info.type != ImageType::e2D); 1763 UNIMPLEMENTED_IF(dst.info.type != ImageType::e2D);
1761 UNIMPLEMENTED_IF(src.info.type != ImageType::e2D); 1764 UNIMPLEMENTED_IF(src.info.type != ImageType::e2D);
1762 if constexpr (HAS_PIXEL_FORMAT_CONVERSIONS) { 1765 if (runtime.ShouldReinterpret(dst, src)) {
1763 return runtime.ConvertImage(dst, src, copies); 1766 return runtime.ReinterpretImage(dst, src, copies);
1764 } 1767 }
1765 for (const ImageCopy& copy : copies) { 1768 for (const ImageCopy& copy : copies) {
1766 UNIMPLEMENTED_IF(copy.dst_subresource.num_layers != 1); 1769 UNIMPLEMENTED_IF(copy.dst_subresource.num_layers != 1);
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index a9504c0e8..643ad811c 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/src/video_core/texture_cache/texture_cache_base.h
@@ -59,8 +59,6 @@ class TextureCache {
59 static constexpr bool HAS_EMULATED_COPIES = P::HAS_EMULATED_COPIES; 59 static constexpr bool HAS_EMULATED_COPIES = P::HAS_EMULATED_COPIES;
60 /// True when the API can provide info about the memory of the device. 60 /// True when the API can provide info about the memory of the device.
61 static constexpr bool HAS_DEVICE_MEMORY_INFO = P::HAS_DEVICE_MEMORY_INFO; 61 static constexpr bool HAS_DEVICE_MEMORY_INFO = P::HAS_DEVICE_MEMORY_INFO;
62 /// True when the API provides utilities for pixel format conversions.
63 static constexpr bool HAS_PIXEL_FORMAT_CONVERSIONS = P::HAS_PIXEL_FORMAT_CONVERSIONS;
64 62
65 static constexpr u64 DEFAULT_EXPECTED_MEMORY = 1_GiB; 63 static constexpr u64 DEFAULT_EXPECTED_MEMORY = 1_GiB;
66 static constexpr u64 DEFAULT_CRITICAL_MEMORY = 2_GiB; 64 static constexpr u64 DEFAULT_CRITICAL_MEMORY = 2_GiB;
diff --git a/src/video_core/texture_cache/types.h b/src/video_core/texture_cache/types.h
index 5c274abdf..5ac27b3a7 100644
--- a/src/video_core/texture_cache/types.h
+++ b/src/video_core/texture_cache/types.h
@@ -54,6 +54,7 @@ enum class RelaxedOptions : u32 {
54 Size = 1 << 0, 54 Size = 1 << 0,
55 Format = 1 << 1, 55 Format = 1 << 1,
56 Samples = 1 << 2, 56 Samples = 1 << 2,
57 ForceBrokenViews = 1 << 3,
57}; 58};
58DECLARE_ENUM_FLAG_OPERATORS(RelaxedOptions) 59DECLARE_ENUM_FLAG_OPERATORS(RelaxedOptions)
59 60
diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp
index ddc9fb13a..e4d82631e 100644
--- a/src/video_core/texture_cache/util.cpp
+++ b/src/video_core/texture_cache/util.cpp
@@ -1151,18 +1151,39 @@ bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr
1151 1151
1152void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst, 1152void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
1153 const ImageBase* src) { 1153 const ImageBase* src) {
1154 if (src && GetFormatType(src->info.format) != SurfaceType::ColorTexture) { 1154 bool is_resolve = false;
1155 src_info.format = src->info.format; 1155 const auto original_src_format = src_info.format;
1156 const auto original_dst_format = dst_info.format;
1157 if (src) {
1158 if (GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
1159 src_info.format = src->info.format;
1160 }
1161 is_resolve = src->info.num_samples > 1;
1162 src_info.num_samples = src->info.num_samples;
1163 src_info.size = src->info.size;
1156 } 1164 }
1157 if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) { 1165 if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) {
1158 dst_info.format = dst->info.format; 1166 dst_info.format = dst->info.format;
1159 } 1167 }
1160 if (src && GetFormatType(src->info.format) != SurfaceType::ColorTexture) { 1168 if (src && GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
1161 dst_info.format = src->info.format; 1169 if (dst) {
1170 if (GetFormatType(dst->info.format) == SurfaceType::ColorTexture) {
1171 src_info.format = original_src_format;
1172 }
1173 } else {
1174 dst_info.format = src->info.format;
1175 }
1162 } 1176 }
1163 if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) { 1177 if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) {
1164 src_info.format = dst->info.format; 1178 if (src) {
1179 if (GetFormatType(src->info.format) == SurfaceType::ColorTexture) {
1180 dst_info.format = original_dst_format;
1181 }
1182 } else {
1183 src_info.format = dst->info.format;
1184 }
1165 } 1185 }
1186 ASSERT(!is_resolve || dst_info.format == src_info.format);
1166} 1187}
1167 1188
1168u32 MapSizeBytes(const ImageBase& image) { 1189u32 MapSizeBytes(const ImageBase& image) {