Merge pull request #7396 from FernandoS27/blit-this-mf

TextureCache: Eliminate format deduction as full depth conversion has been supported.

src/video_core/host_shaders/CMakeLists.txt

@@ -11,13 +11,9 @@ set(SHADER_FILES
     block_linear_unswizzle_2d.comp
     block_linear_unswizzle_3d.comp
     convert_abgr8_to_d24s8.frag
-    convert_b10g11r11_to_d24s8.frag
     convert_d24s8_to_abgr8.frag
-    convert_d24s8_to_b10g11r11.frag
-    convert_d24s8_to_r16g16.frag
     convert_depth_to_float.frag
     convert_float_to_depth.frag
-    convert_r16g16_to_d24s8.frag
     full_screen_triangle.vert
     fxaa.frag
     fxaa.vert

src/video_core/host_shaders/convert_abgr8_to_d24s8.frag

@@ -9,9 +9,10 @@ layout(binding = 0) uniform sampler2D color_texture;
 
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
-    uvec4 color = uvec4(texelFetch(color_texture, coord, 0).rgba * (exp2(8) - 1.0f));
-    uint depth_unorm = (color.r << 16) | (color.g << 8) | color.b;
+    uvec4 color = uvec4(texelFetch(color_texture, coord, 0).abgr * (exp2(8) - 1.0f));
+    uvec4 bytes = color << uvec4(24, 16, 8, 0);
+    uint depth_stencil_unorm = bytes.x | bytes.y | bytes.z | bytes.w;
 
-    gl_FragDepth = float(depth_unorm) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(color.a);
+    gl_FragDepth = float(depth_stencil_unorm & 0x00FFFFFFu) / (exp2(24.0) - 1.0f);
+    gl_FragStencilRefARB = int(depth_stencil_unorm >> 24);
 }

src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag

@@ -1,19 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-#extension GL_ARB_shader_stencil_export : require
-
-layout(binding = 0) uniform sampler2D color_texture;
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    vec4 color = texelFetch(color_texture, coord, 0).rgba;
-    uint depth_stencil_unorm = (uint(color.b * (exp2(10) - 1.0f)) << 22)
-                      | (uint(color.g * (exp2(11) - 1.0f)) << 11)
-                      | (uint(color.r * (exp2(11) - 1.0f)));
-
-    gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
-}

src/video_core/host_shaders/convert_d24s8_to_abgr8.frag

@@ -14,8 +14,10 @@ void main() {
     uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
     uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
 
-    color.r = float(depth >> 16) / (exp2(8) - 1.0);
-    color.g = float((depth >> 8) & 0x00FF) / (exp2(8) - 1.0);
-    color.b = float(depth & 0x00FF) / (exp2(8) - 1.0);
-    color.a = float(stencil) / (exp2(8) - 1.0);
+    highp uint depth_val =
+        uint(textureLod(depth_tex, coord, 0).r * (exp2(32.0) - 1.0));
+    lowp uint stencil_val = textureLod(stencil_tex, coord, 0).r;
+    highp uvec4 components =
+        uvec4(stencil_val, (uvec3(depth_val) >> uvec3(24u, 16u, 8u)) & 0x000000FFu);
+    color.abgr = vec4(components) / (exp2(8.0) - 1.0);
 }

src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag

@@ -1,21 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-
-layout(binding = 0) uniform sampler2D depth_tex;
-layout(binding = 1) uniform isampler2D stencil_tex;
-
-layout(location = 0) out vec4 color;
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
-    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
-
-    color.b = float(depth >> 22) / (exp2(10) - 1.0);
-    color.g = float((depth >> 11) & 0x00FF) / (exp2(11) - 1.0);
-    color.r = float(depth & 0x00FF) / (exp2(11) - 1.0);
-    color.a = 1.0f;
-}

src/video_core/host_shaders/convert_d24s8_to_r16g16.frag

@@ -1,21 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-
-layout(binding = 0) uniform sampler2D depth_tex;
-layout(binding = 1) uniform isampler2D stencil_tex;
-
-layout(location = 0) out vec4 color;
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
-    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
-
-    color.r = float(depth >> 16) / (exp2(16) - 1.0);
-    color.g = float((depth >> 16) & 0x00FF) / (exp2(16) - 1.0);
-    color.b = 0.0f;
-    color.a = 1.0f;
-}

src/video_core/host_shaders/convert_r16g16_to_d24s8.frag

@@ -1,18 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-#extension GL_ARB_shader_stencil_export : require
-
-layout(binding = 0) uniform sampler2D color_texture;
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    vec4 color = texelFetch(color_texture, coord, 0).rgba;
-    uint depth_stencil_unorm = (uint(color.r * (exp2(16) - 1.0f)) << 16)
-                      | (uint(color.g * (exp2(16) - 1.0f)) << 16);
-
-    gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
-}

src/video_core/renderer_vulkan/blit_image.cpp

@@ -5,13 +5,9 @@
 #include <algorithm>
 
 #include "video_core/host_shaders/convert_abgr8_to_d24s8_frag_spv.h"
-#include "video_core/host_shaders/convert_b10g11r11_to_d24s8_frag_spv.h"
 #include "video_core/host_shaders/convert_d24s8_to_abgr8_frag_spv.h"
-#include "video_core/host_shaders/convert_d24s8_to_b10g11r11_frag_spv.h"
-#include "video_core/host_shaders/convert_d24s8_to_r16g16_frag_spv.h"
 #include "video_core/host_shaders/convert_depth_to_float_frag_spv.h"
 #include "video_core/host_shaders/convert_float_to_depth_frag_spv.h"
-#include "video_core/host_shaders/convert_r16g16_to_d24s8_frag_spv.h"
 #include "video_core/host_shaders/full_screen_triangle_vert_spv.h"
 #include "video_core/host_shaders/vulkan_blit_color_float_frag_spv.h"
 #include "video_core/host_shaders/vulkan_blit_depth_stencil_frag_spv.h"
@@ -361,11 +357,7 @@ BlitImageHelper::BlitImageHelper(const Device& device_, VKScheduler& scheduler_,
       convert_depth_to_float_frag(BuildShader(device, CONVERT_DEPTH_TO_FLOAT_FRAG_SPV)),
       convert_float_to_depth_frag(BuildShader(device, CONVERT_FLOAT_TO_DEPTH_FRAG_SPV)),
       convert_abgr8_to_d24s8_frag(BuildShader(device, CONVERT_ABGR8_TO_D24S8_FRAG_SPV)),
-      convert_b10g11r11_to_d24s8_frag(BuildShader(device, CONVERT_B10G11R11_TO_D24S8_FRAG_SPV)),
-      convert_r16g16_to_d24s8_frag(BuildShader(device, CONVERT_R16G16_TO_D24S8_FRAG_SPV)),
       convert_d24s8_to_abgr8_frag(BuildShader(device, CONVERT_D24S8_TO_ABGR8_FRAG_SPV)),
-      convert_d24s8_to_b10g11r11_frag(BuildShader(device, CONVERT_D24S8_TO_B10G11R11_FRAG_SPV)),
-      convert_d24s8_to_r16g16_frag(BuildShader(device, CONVERT_D24S8_TO_R16G16_FRAG_SPV)),
       linear_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_LINEAR>)),
       nearest_sampler(device.GetLogical().CreateSampler(SAMPLER_CREATE_INFO<VK_FILTER_NEAREST>)) {
     if (device.IsExtShaderStencilExportSupported()) {
@@ -461,29 +453,10 @@ void BlitImageHelper::ConvertR16ToD16(const Framebuffer* dst_framebuffer,
 }
 
 void BlitImageHelper::ConvertABGR8ToD24S8(const Framebuffer* dst_framebuffer,
-                                          const ImageView& src_image_view, u32 up_scale,
-                                          u32 down_shift) {
+                                          ImageView& src_image_view, u32 up_scale, u32 down_shift) {
     ConvertPipelineDepthTargetEx(convert_abgr8_to_d24s8_pipeline, dst_framebuffer->RenderPass(),
                                  convert_abgr8_to_d24s8_frag, true);
-    Convert(*convert_abgr8_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
-            down_shift);
-}
-
-void BlitImageHelper::ConvertB10G11R11ToD24S8(const Framebuffer* dst_framebuffer,
-                                              const ImageView& src_image_view, u32 up_scale,
-                                              u32 down_shift) {
-    ConvertPipelineDepthTargetEx(convert_b10g11r11_to_d24s8_pipeline, dst_framebuffer->RenderPass(),
-                                 convert_b10g11r11_to_d24s8_frag, true);
-    Convert(*convert_b10g11r11_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
-            down_shift);
-}
-
-void BlitImageHelper::ConvertR16G16ToD24S8(const Framebuffer* dst_framebuffer,
-                                           const ImageView& src_image_view, u32 up_scale,
-                                           u32 down_shift) {
-    ConvertPipelineDepthTargetEx(convert_r16g16_to_d24s8_pipeline, dst_framebuffer->RenderPass(),
-                                 convert_r16g16_to_d24s8_frag, true);
-    Convert(*convert_r16g16_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
+    ConvertColor(*convert_abgr8_to_d24s8_pipeline, dst_framebuffer, src_image_view, up_scale,
                  down_shift);
 }
 
@@ -495,24 +468,6 @@ void BlitImageHelper::ConvertD24S8ToABGR8(const Framebuffer* dst_framebuffer,
                         down_shift);
 }
 
-void BlitImageHelper::ConvertD24S8ToB10G11R11(const Framebuffer* dst_framebuffer,
-                                              ImageView& src_image_view, u32 up_scale,
-                                              u32 down_shift) {
-    ConvertPipelineColorTargetEx(convert_d24s8_to_b10g11r11_pipeline, dst_framebuffer->RenderPass(),
-                                 convert_d24s8_to_b10g11r11_frag, false);
-    ConvertDepthStencil(*convert_d24s8_to_b10g11r11_pipeline, dst_framebuffer, src_image_view,
-                        up_scale, down_shift);
-}
-
-void BlitImageHelper::ConvertD24S8ToR16G16(const Framebuffer* dst_framebuffer,
-                                           ImageView& src_image_view, u32 up_scale,
-                                           u32 down_shift) {
-    ConvertPipelineColorTargetEx(convert_d24s8_to_r16g16_pipeline, dst_framebuffer->RenderPass(),
-                                 convert_d24s8_to_r16g16_frag, false);
-    ConvertDepthStencil(*convert_d24s8_to_r16g16_pipeline, dst_framebuffer, src_image_view,
-                        up_scale, down_shift);
-}
-
 void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
                               const ImageView& src_image_view, u32 up_scale, u32 down_shift) {
     const VkPipelineLayout layout = *one_texture_pipeline_layout;
@@ -560,6 +515,53 @@ void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_frameb
     scheduler.InvalidateState();
 }
 
+void BlitImageHelper::ConvertColor(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
+                                   ImageView& src_image_view, u32 up_scale, u32 down_shift) {
+    const VkPipelineLayout layout = *one_texture_pipeline_layout;
+    const VkImageView src_view = src_image_view.ColorView();
+    const VkSampler sampler = *nearest_sampler;
+    const VkExtent2D extent{
+        .width = std::max((src_image_view.size.width * up_scale) >> down_shift, 1U),
+        .height = std::max((src_image_view.size.height * up_scale) >> down_shift, 1U),
+    };
+    scheduler.RequestRenderpass(dst_framebuffer);
+    scheduler.Record([pipeline, layout, sampler, src_view, extent, up_scale, down_shift,
+                      this](vk::CommandBuffer cmdbuf) {
+        const VkOffset2D offset{
+            .x = 0,
+            .y = 0,
+        };
+        const VkViewport viewport{
+            .x = 0.0f,
+            .y = 0.0f,
+            .width = static_cast<float>(extent.width),
+            .height = static_cast<float>(extent.height),
+            .minDepth = 0.0f,
+            .maxDepth = 0.0f,
+        };
+        const VkRect2D scissor{
+            .offset = offset,
+            .extent = extent,
+        };
+        const PushConstants push_constants{
+            .tex_scale = {viewport.width, viewport.height},
+            .tex_offset = {0.0f, 0.0f},
+        };
+        const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
+        UpdateOneTextureDescriptorSet(device, descriptor_set, sampler, src_view);
+
+        // TODO: Barriers
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+        cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, layout, 0, descriptor_set,
+                                  nullptr);
+        cmdbuf.SetViewport(0, viewport);
+        cmdbuf.SetScissor(0, scissor);
+        cmdbuf.PushConstants(layout, VK_SHADER_STAGE_VERTEX_BIT, push_constants);
+        cmdbuf.Draw(3, 1, 0, 0);
+    });
+    scheduler.InvalidateState();
+}
+
 void BlitImageHelper::ConvertDepthStencil(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
                                           ImageView& src_image_view, u32 up_scale, u32 down_shift) {
     const VkPipelineLayout layout = *two_textures_pipeline_layout;

src/video_core/renderer_vulkan/blit_image.h

@@ -56,28 +56,19 @@ public:
     void ConvertR16ToD16(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
                          u32 up_scale, u32 down_shift);
 
-    void ConvertABGR8ToD24S8(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
-                             u32 up_scale, u32 down_shift);
-
-    void ConvertB10G11R11ToD24S8(const Framebuffer* dst_framebuffer,
-                                 const ImageView& src_image_view, u32 up_scale, u32 down_shift);
-
-    void ConvertR16G16ToD24S8(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
+    void ConvertABGR8ToD24S8(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
                              u32 up_scale, u32 down_shift);
 
     void ConvertD24S8ToABGR8(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
                              u32 up_scale, u32 down_shift);
 
-    void ConvertD24S8ToB10G11R11(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
-                                 u32 up_scale, u32 down_shift);
-
-    void ConvertD24S8ToR16G16(const Framebuffer* dst_framebuffer, ImageView& src_image_view,
-                              u32 up_scale, u32 down_shift);
-
 private:
     void Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
                  const ImageView& src_image_view, u32 up_scale, u32 down_shift);
 
+    void ConvertColor(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
+                      ImageView& src_image_view, u32 up_scale, u32 down_shift);
+
     void ConvertDepthStencil(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
                              ImageView& src_image_view, u32 up_scale, u32 down_shift);
 
@@ -114,11 +105,7 @@ private:
     vk::ShaderModule convert_depth_to_float_frag;
     vk::ShaderModule convert_float_to_depth_frag;
     vk::ShaderModule convert_abgr8_to_d24s8_frag;
-    vk::ShaderModule convert_b10g11r11_to_d24s8_frag;
-    vk::ShaderModule convert_r16g16_to_d24s8_frag;
     vk::ShaderModule convert_d24s8_to_abgr8_frag;
-    vk::ShaderModule convert_d24s8_to_b10g11r11_frag;
-    vk::ShaderModule convert_d24s8_to_r16g16_frag;
     vk::Sampler linear_sampler;
     vk::Sampler nearest_sampler;
 
@@ -131,11 +118,7 @@ private:
     vk::Pipeline convert_d16_to_r16_pipeline;
     vk::Pipeline convert_r16_to_d16_pipeline;
     vk::Pipeline convert_abgr8_to_d24s8_pipeline;
-    vk::Pipeline convert_b10g11r11_to_d24s8_pipeline;
-    vk::Pipeline convert_r16g16_to_d24s8_pipeline;
     vk::Pipeline convert_d24s8_to_abgr8_pipeline;
-    vk::Pipeline convert_d24s8_to_b10g11r11_pipeline;
-    vk::Pipeline convert_d24s8_to_r16g16_pipeline;
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -775,8 +775,18 @@ StagingBufferRef TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
 
 bool TextureCacheRuntime::ShouldReinterpret(Image& dst, Image& src) {
     if (VideoCore::Surface::GetFormatType(dst.info.format) ==
-        VideoCore::Surface::SurfaceType::DepthStencil) {
-        return !device.IsExtShaderStencilExportSupported();
+            VideoCore::Surface::SurfaceType::DepthStencil &&
+        !device.IsExtShaderStencilExportSupported()) {
+        return true;
+    }
+    if (VideoCore::Surface::GetFormatType(src.info.format) ==
+            VideoCore::Surface::SurfaceType::DepthStencil &&
+        !device.IsExtShaderStencilExportSupported()) {
+        return true;
+    }
+    if (dst.info.format == PixelFormat::D32_FLOAT_S8_UINT ||
+        src.info.format == PixelFormat::D32_FLOAT_S8_UINT) {
+        return true;
     }
     return false;
 }
@@ -1058,21 +1068,10 @@ void TextureCacheRuntime::ConvertImage(Framebuffer* dst, ImageView& dst_view, Im
         }
         break;
     case PixelFormat::A8B8G8R8_UNORM:
-    case PixelFormat::B8G8R8A8_UNORM:
         if (src_view.format == PixelFormat::S8_UINT_D24_UNORM) {
             return blit_image_helper.ConvertD24S8ToABGR8(dst, src_view, up_scale, down_shift);
         }
         break;
-    case PixelFormat::B10G11R11_FLOAT:
-        if (src_view.format == PixelFormat::S8_UINT_D24_UNORM) {
-            return blit_image_helper.ConvertD24S8ToB10G11R11(dst, src_view, up_scale, down_shift);
-        }
-        break;
-    case PixelFormat::R16G16_UNORM:
-        if (src_view.format == PixelFormat::S8_UINT_D24_UNORM) {
-            return blit_image_helper.ConvertD24S8ToR16G16(dst, src_view, up_scale, down_shift);
-        }
-        break;
     case PixelFormat::R32_FLOAT:
         if (src_view.format == PixelFormat::D32_FLOAT) {
             return blit_image_helper.ConvertD32ToR32(dst, src_view, up_scale, down_shift);
@@ -1084,16 +1083,7 @@ void TextureCacheRuntime::ConvertImage(Framebuffer* dst, ImageView& dst_view, Im
         }
         break;
     case PixelFormat::S8_UINT_D24_UNORM:
-        if (src_view.format == PixelFormat::A8B8G8R8_UNORM ||
-            src_view.format == PixelFormat::B8G8R8A8_UNORM) {
         return blit_image_helper.ConvertABGR8ToD24S8(dst, src_view, up_scale, down_shift);
-        }
-        if (src_view.format == PixelFormat::B10G11R11_FLOAT) {
-            return blit_image_helper.ConvertB10G11R11ToD24S8(dst, src_view, up_scale, down_shift);
-        }
-        if (src_view.format == PixelFormat::R16G16_UNORM) {
-            return blit_image_helper.ConvertR16G16ToD24S8(dst, src_view, up_scale, down_shift);
-        }
         break;
     case PixelFormat::D32_FLOAT:
         if (src_view.format == PixelFormat::R32_FLOAT) {
@@ -1590,6 +1580,14 @@ VkImageView ImageView::StencilView() {
     return *stencil_view;
 }
 
+VkImageView ImageView::ColorView() {
+    if (color_view) {
+        return *color_view;
+    }
+    color_view = MakeView(VK_FORMAT_R8G8B8A8_UNORM, VK_IMAGE_ASPECT_COLOR_BIT);
+    return *color_view;
+}
+
 VkImageView ImageView::StorageView(Shader::TextureType texture_type,
                                    Shader::ImageFormat image_format) {
     if (image_format == Shader::ImageFormat::Typeless) {

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -184,6 +184,8 @@ public:
 
     [[nodiscard]] VkImageView StencilView();
 
+    [[nodiscard]] VkImageView ColorView();
+
     [[nodiscard]] VkImageView StorageView(Shader::TextureType texture_type,
                                           Shader::ImageFormat image_format);
 
@@ -224,6 +226,7 @@ private:
     std::unique_ptr<StorageViews> storage_views;
     vk::ImageView depth_view;
     vk::ImageView stencil_view;
+    vk::ImageView color_view;
     VkImage image_handle = VK_NULL_HANDLE;
     VkImageView render_target = VK_NULL_HANDLE;
     VkSampleCountFlagBits samples = VK_SAMPLE_COUNT_1_BIT;

src/video_core/texture_cache/texture_cache.h

@@ -472,7 +472,7 @@ template <class P>
 void TextureCache<P>::BlitImage(const Tegra::Engines::Fermi2D::Surface& dst,
                                 const Tegra::Engines::Fermi2D::Surface& src,
                                 const Tegra::Engines::Fermi2D::Config& copy) {
-    const BlitImages images = GetBlitImages(dst, src);
+    const BlitImages images = GetBlitImages(dst, src, copy);
     const ImageId dst_id = images.dst_id;
     const ImageId src_id = images.src_id;
 
@@ -762,12 +762,15 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
     const bool broken_views =
         runtime.HasBrokenTextureViewFormats() || True(options & RelaxedOptions::ForceBrokenViews);
     const bool native_bgr = runtime.HasNativeBgr();
-    ImageId image_id;
+    const bool flexible_formats = True(options & RelaxedOptions::Format);
+    ImageId image_id{};
+    boost::container::small_vector<ImageId, 1> image_ids;
     const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
         if (True(existing_image.flags & ImageFlagBits::Remapped)) {
             return false;
         }
-        if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear) {
+        if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear)
+            [[unlikely]] {
             const bool strict_size = False(options & RelaxedOptions::Size) &&
                                      True(existing_image.flags & ImageFlagBits::Strong);
             const ImageInfo& existing = existing_image.info;
@@ -776,18 +779,28 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
                 IsPitchLinearSameSize(existing, info, strict_size) &&
                 IsViewCompatible(existing.format, info.format, broken_views, native_bgr)) {
                 image_id = existing_image_id;
-                return true;
+                image_ids.push_back(existing_image_id);
+                return !flexible_formats && existing.format == info.format;
             }
         } else if (IsSubresource(info, existing_image, gpu_addr, options, broken_views,
                                  native_bgr)) {
             image_id = existing_image_id;
-            return true;
+            image_ids.push_back(existing_image_id);
+            return !flexible_formats && existing_image.info.format == info.format;
         }
         return false;
     };
     ForEachImageInRegion(*cpu_addr, CalculateGuestSizeInBytes(info), lambda);
+    if (image_ids.size() <= 1) [[likely]] {
         return image_id;
     }
+    auto image_ids_compare = [this](ImageId a, ImageId b) {
+        auto& image_a = slot_images[a];
+        auto& image_b = slot_images[b];
+        return image_a.modification_tick < image_b.modification_tick;
+    };
+    return *std::ranges::max_element(image_ids, image_ids_compare);
+}
 
 template <class P>
 bool TextureCache<P>::ImageCanRescale(ImageBase& image) {
@@ -1078,32 +1091,58 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
 
 template <class P>
 typename TextureCache<P>::BlitImages TextureCache<P>::GetBlitImages(
-    const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src) {
-    static constexpr auto FIND_OPTIONS = RelaxedOptions::Format | RelaxedOptions::Samples;
+    const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src,
+    const Tegra::Engines::Fermi2D::Config& copy) {
+
+    static constexpr auto FIND_OPTIONS = RelaxedOptions::Samples;
     const GPUVAddr dst_addr = dst.Address();
     const GPUVAddr src_addr = src.Address();
     ImageInfo dst_info(dst);
     ImageInfo src_info(src);
+    const bool can_be_depth_blit =
+        dst_info.format == src_info.format && copy.filter == Tegra::Engines::Fermi2D::Filter::Point;
     ImageId dst_id;
     ImageId src_id;
+    RelaxedOptions try_options = FIND_OPTIONS;
+    if (can_be_depth_blit) {
+        try_options |= RelaxedOptions::Format;
+    }
     do {
         has_deleted_images = false;
-        dst_id = FindImage(dst_info, dst_addr, FIND_OPTIONS);
-        src_id = FindImage(src_info, src_addr, FIND_OPTIONS);
-        const ImageBase* const dst_image = dst_id ? &slot_images[dst_id] : nullptr;
+        src_id = FindImage(src_info, src_addr, try_options);
+        dst_id = FindImage(dst_info, dst_addr, try_options);
         const ImageBase* const src_image = src_id ? &slot_images[src_id] : nullptr;
+        if (src_image && src_image->info.num_samples > 1) {
+            RelaxedOptions find_options{FIND_OPTIONS | RelaxedOptions::ForceBrokenViews};
+            src_id = FindOrInsertImage(src_info, src_addr, find_options);
+            dst_id = FindOrInsertImage(dst_info, dst_addr, find_options);
+            if (has_deleted_images) {
+                continue;
+            }
+            break;
+        }
+        if (can_be_depth_blit) {
+            const ImageBase* const dst_image = src_id ? &slot_images[src_id] : nullptr;
             DeduceBlitImages(dst_info, src_info, dst_image, src_image);
             if (GetFormatType(dst_info.format) != GetFormatType(src_info.format)) {
                 continue;
             }
-        RelaxedOptions find_options{};
-        if (src_info.num_samples > 1) {
-            // it's a resolve, we must enforce the same format.
-            find_options = RelaxedOptions::ForceBrokenViews;
         }
-        src_id = FindOrInsertImage(src_info, src_addr, find_options);
-        dst_id = FindOrInsertImage(dst_info, dst_addr, find_options);
+        if (!src_id) {
+            src_id = InsertImage(src_info, src_addr, RelaxedOptions{});
+        }
+        if (!dst_id) {
+            dst_id = InsertImage(dst_info, dst_addr, RelaxedOptions{});
+        }
     } while (has_deleted_images);
+    if (GetFormatType(dst_info.format) != SurfaceType::ColorTexture) {
+        // Make sure the images are depth and/or stencil textures.
+        do {
+            has_deleted_images = false;
+            src_id = FindOrInsertImage(src_info, src_addr, RelaxedOptions{});
+            dst_id = FindOrInsertImage(dst_info, dst_addr, RelaxedOptions{});
+        } while (has_deleted_images);
+    }
     return BlitImages{
         .dst_id = dst_id,
         .src_id = src_id,
@@ -1160,7 +1199,14 @@ template <class P>
 ImageViewId TextureCache<P>::FindRenderTargetView(const ImageInfo& info, GPUVAddr gpu_addr,
                                                   bool is_clear) {
     const auto options = is_clear ? RelaxedOptions::Samples : RelaxedOptions{};
-    const ImageId image_id = FindOrInsertImage(info, gpu_addr, options);
+    ImageId image_id{};
+    bool delete_state = has_deleted_images;
+    do {
+        has_deleted_images = false;
+        image_id = FindOrInsertImage(info, gpu_addr, options);
+        delete_state |= has_deleted_images;
+    } while (has_deleted_images);
+    has_deleted_images = delete_state;
     if (!image_id) {
         return NULL_IMAGE_VIEW_ID;
     }
@@ -1783,7 +1829,13 @@ void TextureCache<P>::CopyImage(ImageId dst_id, ImageId src_id, std::vector<Imag
         const SubresourceExtent src_extent{.levels = 1, .layers = 1};
         const SubresourceRange dst_range{.base = dst_base, .extent = dst_extent};
         const SubresourceRange src_range{.base = src_base, .extent = src_extent};
-        const ImageViewInfo dst_view_info(ImageViewType::e2D, dst.info.format, dst_range);
+        PixelFormat dst_format = dst.info.format;
+        if (GetFormatType(src.info.format) == SurfaceType::DepthStencil &&
+            GetFormatType(dst_format) == SurfaceType::ColorTexture &&
+            BytesPerBlock(dst_format) == 4) {
+            dst_format = PixelFormat::A8B8G8R8_UNORM;
+        }
+        const ImageViewInfo dst_view_info(ImageViewType::e2D, dst_format, dst_range);
         const ImageViewInfo src_view_info(ImageViewType::e2D, src.info.format, src_range);
         const auto [dst_framebuffer_id, dst_view_id] = RenderTargetFromImage(dst_id, dst_view_info);
         Framebuffer* const dst_framebuffer = &slot_framebuffers[dst_framebuffer_id];

src/video_core/texture_cache/texture_cache_base.h

@@ -252,7 +252,8 @@ private:
 
     /// Return a blit image pair from the given guest blit parameters
     [[nodiscard]] BlitImages GetBlitImages(const Tegra::Engines::Fermi2D::Surface& dst,
-                                           const Tegra::Engines::Fermi2D::Surface& src);
+                                           const Tegra::Engines::Fermi2D::Surface& src,
+                                           const Tegra::Engines::Fermi2D::Config& copy);
 
     /// Find or create a sampler from a guest descriptor sampler
     [[nodiscard]] SamplerId FindSampler(const TSCEntry& config);

src/video_core/texture_cache/util.cpp

@@ -1151,29 +1151,16 @@ bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr
 
 void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
                       const ImageBase* src) {
-    bool is_resolve = false;
-    const auto original_src_format = src_info.format;
     const auto original_dst_format = dst_info.format;
-    if (src) {
-        if (GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
+    if (src && GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
         src_info.format = src->info.format;
     }
-        is_resolve = src->info.num_samples > 1;
-        src_info.num_samples = src->info.num_samples;
-        src_info.size = src->info.size;
-    }
     if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) {
         dst_info.format = dst->info.format;
     }
     if (src && GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
-        if (dst) {
-            if (GetFormatType(dst->info.format) == SurfaceType::ColorTexture) {
-                src_info.format = original_src_format;
-            }
-        } else {
         dst_info.format = src->info.format;
     }
-    }
     if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) {
         if (src) {
             if (GetFormatType(src->info.format) == SurfaceType::ColorTexture) {
@@ -1183,7 +1170,6 @@ void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase*
             src_info.format = dst->info.format;
         }
     }
-    ASSERT(!is_resolve || dst_info.format == src_info.format);
 }
 
 u32 MapSizeBytes(const ImageBase& image) {