texture_cache: Implement color to multisampled depth blit pass (#3103)

CMakeLists.txt

@@ -952,6 +952,10 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
                src/video_core/renderer_vulkan/host_passes/fsr_pass.h
                src/video_core/renderer_vulkan/host_passes/pp_pass.cpp
                src/video_core/renderer_vulkan/host_passes/pp_pass.h
+               src/video_core/texture_cache/blit_helper.cpp
+               src/video_core/texture_cache/blit_helper.h
+               src/video_core/texture_cache/host_compatibility.cpp
+               src/video_core/texture_cache/host_compatibility.h
                src/video_core/texture_cache/image.cpp
                src/video_core/texture_cache/image.h
                src/video_core/texture_cache/image_info.cpp
@@ -965,8 +969,6 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
                src/video_core/texture_cache/tile_manager.cpp
                src/video_core/texture_cache/tile_manager.h
                src/video_core/texture_cache/types.h
-               src/video_core/texture_cache/host_compatibility.cpp
-               src/video_core/texture_cache/host_compatibility.h
                src/video_core/page_manager.cpp
                src/video_core/page_manager.h
                src/video_core/multi_level_page_table.h

src/video_core/host_shaders/CMakeLists.txt

@@ -11,6 +11,7 @@ set(SHADER_FILES
     detilers/micro_32bpp.comp
     detilers/micro_64bpp.comp
     detilers/micro_8bpp.comp
+    color_to_ms_depth.frag
     fault_buffer_process.comp
     fs_tri.vert
     fsr.comp

src/video_core/host_shaders/color_to_ms_depth.frag

@@ -0,0 +1,15 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#version 450 core
+#extension GL_EXT_samplerless_texture_functions : require
+
+layout (binding = 0, set = 0) uniform texture2D color;
+
+layout (location = 0) in vec2 uv;
+
+void main()
+{
+    ivec2 coord = ivec2(uv * vec2(textureSize(color, 0).xy));
+    gl_FragDepth = texelFetch(color, coord, 0)[gl_SampleID];
+}

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -328,6 +328,7 @@ public:
         return render_state;
     }
 
+    /// Returns the current pipeline dynamic state tracking.
     DynamicState& GetDynamicState() {
         return dynamic_state;
     }

src/video_core/texture_cache/blit_helper.cpp

@@ -0,0 +1,256 @@
+// SPDX-FileCopyrightText: Copyright 2025 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "video_core/renderer_vulkan/vk_instance.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_shader_util.h"
+#include "video_core/texture_cache/blit_helper.h"
+#include "video_core/texture_cache/image.h"
+
+#include "video_core/host_shaders/color_to_ms_depth_frag.h"
+#include "video_core/host_shaders/fs_tri_vert.h"
+
+namespace VideoCore {
+
+static vk::SampleCountFlagBits ToSampleCount(u32 num_samples) {
+    switch (num_samples) {
+    case 1:
+        return vk::SampleCountFlagBits::e1;
+    case 2:
+        return vk::SampleCountFlagBits::e2;
+    case 4:
+        return vk::SampleCountFlagBits::e4;
+    case 8:
+        return vk::SampleCountFlagBits::e8;
+    case 16:
+        return vk::SampleCountFlagBits::e16;
+    default:
+        UNREACHABLE_MSG("Unknown samples count = {}", num_samples);
+    }
+}
+
+BlitHelper::BlitHelper(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_)
+    : instance{instance_}, scheduler{scheduler_} {
+    CreateShaders();
+    CreatePipelineLayouts();
+}
+
+BlitHelper::~BlitHelper() = default;
+
+void BlitHelper::BlitColorToMsDepth(Image& source, Image& dest) {
+    source.Transit(vk::ImageLayout::eShaderReadOnlyOptimal, vk::AccessFlagBits2::eShaderRead, {});
+    dest.Transit(vk::ImageLayout::eDepthAttachmentOptimal,
+                 vk::AccessFlagBits2::eDepthStencilAttachmentWrite, {});
+
+    const vk::ImageViewUsageCreateInfo color_usage_ci{.usage = vk::ImageUsageFlagBits::eSampled};
+    const vk::ImageViewCreateInfo color_view_ci = {
+        .pNext = &color_usage_ci,
+        .image = source.image,
+        .viewType = vk::ImageViewType::e2D,
+        .format = source.info.pixel_format,
+        .subresourceRange{
+            .aspectMask = vk::ImageAspectFlagBits::eColor,
+            .baseMipLevel = 0U,
+            .levelCount = 1U,
+            .baseArrayLayer = 0U,
+            .layerCount = 1U,
+        },
+    };
+    const auto [color_view_result, color_view] =
+        instance.GetDevice().createImageView(color_view_ci);
+    ASSERT_MSG(color_view_result == vk::Result::eSuccess, "Failed to create image view: {}",
+               vk::to_string(color_view_result));
+    const vk::ImageViewUsageCreateInfo depth_usage_ci{
+        .usage = vk::ImageUsageFlagBits::eDepthStencilAttachment};
+    const vk::ImageViewCreateInfo depth_view_ci = {
+        .pNext = &depth_usage_ci,
+        .image = dest.image,
+        .viewType = vk::ImageViewType::e2D,
+        .format = dest.info.pixel_format,
+        .subresourceRange{
+            .aspectMask = vk::ImageAspectFlagBits::eDepth,
+            .baseMipLevel = 0U,
+            .levelCount = 1U,
+            .baseArrayLayer = 0U,
+            .layerCount = 1U,
+        },
+    };
+    const auto [depth_view_result, depth_view] =
+        instance.GetDevice().createImageView(depth_view_ci);
+    ASSERT_MSG(depth_view_result == vk::Result::eSuccess, "Failed to create image view: {}",
+               vk::to_string(depth_view_result));
+    scheduler.DeferOperation([device = instance.GetDevice(), color_view, depth_view] {
+        device.destroyImageView(color_view);
+        device.destroyImageView(depth_view);
+    });
+
+    Vulkan::RenderState state{};
+    state.has_depth = true;
+    state.width = dest.info.size.width;
+    state.height = dest.info.size.height;
+    state.depth_attachment = vk::RenderingAttachmentInfo{
+        .imageView = depth_view,
+        .imageLayout = vk::ImageLayout::eDepthAttachmentOptimal,
+        .loadOp = vk::AttachmentLoadOp::eDontCare,
+        .storeOp = vk::AttachmentStoreOp::eStore,
+        .clearValue = vk::ClearValue{.depthStencil = {.depth = 0.f}},
+    };
+    scheduler.BeginRendering(state);
+
+    const auto cmdbuf = scheduler.CommandBuffer();
+    const vk::DescriptorImageInfo image_info = {
+        .sampler = VK_NULL_HANDLE,
+        .imageView = color_view,
+        .imageLayout = vk::ImageLayout::eShaderReadOnlyOptimal,
+    };
+    const vk::WriteDescriptorSet texture_write = {
+        .dstSet = VK_NULL_HANDLE,
+        .dstBinding = 0U,
+        .dstArrayElement = 0U,
+        .descriptorCount = 1U,
+        .descriptorType = vk::DescriptorType::eSampledImage,
+        .pImageInfo = &image_info,
+    };
+    cmdbuf.pushDescriptorSetKHR(vk::PipelineBindPoint::eGraphics, *single_texture_pl_layout, 0U,
+                                texture_write);
+
+    const DepthPipelineKey key{dest.info.num_samples, dest.info.pixel_format};
+    const vk::Pipeline depth_pipeline = GetDepthToMsPipeline(key);
+    cmdbuf.bindPipeline(vk::PipelineBindPoint::eGraphics, depth_pipeline);
+
+    const vk::Viewport viewport = {
+        .x = 0,
+        .y = 0,
+        .width = float(state.width),
+        .height = float(state.height),
+        .minDepth = 0.f,
+        .maxDepth = 1.f,
+    };
+    cmdbuf.setViewport(0, viewport);
+
+    const vk::Rect2D scissor = {
+        .offset = {0, 0},
+        .extent = {state.width, state.height},
+    };
+    cmdbuf.setScissor(0, scissor);
+
+    cmdbuf.draw(3, 1, 0, 0);
+
+    scheduler.GetDynamicState().Invalidate();
+}
+
+vk::Pipeline BlitHelper::GetDepthToMsPipeline(const DepthPipelineKey& key) {
+    auto it = std::ranges::find(color_to_ms_depth_pl, key, &DepthPipeline::first);
+    if (it != color_to_ms_depth_pl.end()) {
+        return *it->second;
+    }
+    CreateColorToMSDepthPipeline(key);
+    return *color_to_ms_depth_pl.back().second;
+}
+
+void BlitHelper::CreateShaders() {
+    fs_tri_vertex = Vulkan::Compile(HostShaders::FS_TRI_VERT, vk::ShaderStageFlagBits::eVertex,
+                                    instance.GetDevice());
+    color_to_ms_depth_frag =
+        Vulkan::Compile(HostShaders::COLOR_TO_MS_DEPTH_FRAG, vk::ShaderStageFlagBits::eFragment,
+                        instance.GetDevice());
+}
+
+void BlitHelper::CreatePipelineLayouts() {
+    const vk::DescriptorSetLayoutBinding texture_binding = {
+        .binding = 0,
+        .descriptorType = vk::DescriptorType::eSampledImage,
+        .descriptorCount = 1,
+        .stageFlags = vk::ShaderStageFlagBits::eFragment,
+    };
+    const vk::DescriptorSetLayoutCreateInfo desc_layout_ci = {
+        .flags = vk::DescriptorSetLayoutCreateFlagBits::ePushDescriptorKHR,
+        .bindingCount = 1U,
+        .pBindings = &texture_binding,
+    };
+    auto [desc_layout_result, desc_layout] =
+        instance.GetDevice().createDescriptorSetLayoutUnique(desc_layout_ci);
+    single_texture_descriptor_set_layout = std::move(desc_layout);
+    const vk::DescriptorSetLayout set_layout = *single_texture_descriptor_set_layout;
+    const vk::PipelineLayoutCreateInfo layout_info = {
+        .setLayoutCount = 1U,
+        .pSetLayouts = &set_layout,
+        .pushConstantRangeCount = 0U,
+        .pPushConstantRanges = nullptr,
+    };
+    auto [layout_result, pipeline_layout] =
+        instance.GetDevice().createPipelineLayoutUnique(layout_info);
+    ASSERT_MSG(layout_result == vk::Result::eSuccess,
+               "Failed to create graphics pipeline layout: {}", vk::to_string(layout_result));
+    Vulkan::SetObjectName(instance.GetDevice(), *pipeline_layout, "Single texture pipeline layout");
+    single_texture_pl_layout = std::move(pipeline_layout);
+}
+
+void BlitHelper::CreateColorToMSDepthPipeline(const DepthPipelineKey& key) {
+    const vk::PipelineInputAssemblyStateCreateInfo input_assembly = {
+        .topology = vk::PrimitiveTopology::eTriangleList,
+    };
+    const vk::PipelineMultisampleStateCreateInfo multisampling = {
+        .rasterizationSamples = ToSampleCount(key.num_samples),
+    };
+    const vk::PipelineDepthStencilStateCreateInfo depth_state = {
+        .depthTestEnable = true,
+        .depthWriteEnable = true,
+        .depthCompareOp = vk::CompareOp::eAlways,
+    };
+    const std::array dynamic_states = {vk::DynamicState::eViewportWithCount,
+                                       vk::DynamicState::eScissorWithCount};
+    const vk::PipelineDynamicStateCreateInfo dynamic_info = {
+        .dynamicStateCount = static_cast<u32>(dynamic_states.size()),
+        .pDynamicStates = dynamic_states.data(),
+    };
+
+    std::array<vk::PipelineShaderStageCreateInfo, 2> shader_stages;
+    shader_stages[0] = {
+        .stage = vk::ShaderStageFlagBits::eVertex,
+        .module = fs_tri_vertex,
+        .pName = "main",
+    };
+    shader_stages[1] = {
+        .stage = vk::ShaderStageFlagBits::eFragment,
+        .module = color_to_ms_depth_frag,
+        .pName = "main",
+    };
+
+    const vk::PipelineRenderingCreateInfo pipeline_rendering_ci = {
+        .colorAttachmentCount = 0U,
+        .pColorAttachmentFormats = nullptr,
+        .depthAttachmentFormat = key.depth_format,
+        .stencilAttachmentFormat = vk::Format::eUndefined,
+    };
+
+    const vk::PipelineColorBlendStateCreateInfo color_blending{};
+    const vk::PipelineViewportStateCreateInfo viewport_info{};
+    const vk::PipelineVertexInputStateCreateInfo vertex_input_info{};
+    const vk::PipelineRasterizationStateCreateInfo raster_state{.lineWidth = 1.f};
+
+    const vk::GraphicsPipelineCreateInfo pipeline_info = {
+        .pNext = &pipeline_rendering_ci,
+        .stageCount = static_cast<u32>(shader_stages.size()),
+        .pStages = shader_stages.data(),
+        .pVertexInputState = &vertex_input_info,
+        .pInputAssemblyState = &input_assembly,
+        .pViewportState = &viewport_info,
+        .pRasterizationState = &raster_state,
+        .pMultisampleState = &multisampling,
+        .pDepthStencilState = &depth_state,
+        .pColorBlendState = &color_blending,
+        .pDynamicState = &dynamic_info,
+        .layout = *single_texture_pl_layout,
+    };
+
+    auto [pipeline_result, pipeline] =
+        instance.GetDevice().createGraphicsPipelineUnique(VK_NULL_HANDLE, pipeline_info);
+    ASSERT_MSG(pipeline_result == vk::Result::eSuccess, "Failed to create graphics pipeline: {}",
+               vk::to_string(pipeline_result));
+    Vulkan::SetObjectName(instance.GetDevice(), *pipeline, "Color to MS Depth {}", key.num_samples);
+
+    color_to_ms_depth_pl.emplace_back(key, std::move(pipeline));
+}
+
+} // namespace VideoCore

src/video_core/texture_cache/blit_helper.h

@@ -0,0 +1,55 @@
+// SPDX-FileCopyrightText: Copyright 2025 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <tsl/robin_map.h>
+
+#include "common/types.h"
+#include "video_core/renderer_vulkan/vk_common.h"
+
+namespace Vulkan {
+class Instance;
+class Scheduler;
+} // namespace Vulkan
+
+namespace VideoCore {
+
+class Image;
+class ImageView;
+
+class BlitHelper {
+    static constexpr size_t MaxMsPipelines = 6;
+
+public:
+    explicit BlitHelper(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler);
+    ~BlitHelper();
+
+    void BlitColorToMsDepth(Image& source, Image& dest);
+
+private:
+    void CreateShaders();
+    void CreatePipelineLayouts();
+
+    struct DepthPipelineKey {
+        u32 num_samples;
+        vk::Format depth_format;
+
+        auto operator<=>(const DepthPipelineKey&) const noexcept = default;
+    };
+    vk::Pipeline GetDepthToMsPipeline(const DepthPipelineKey& key);
+    void CreateColorToMSDepthPipeline(const DepthPipelineKey& key);
+
+private:
+    const Vulkan::Instance& instance;
+    Vulkan::Scheduler& scheduler;
+    vk::UniqueDescriptorSetLayout single_texture_descriptor_set_layout;
+    vk::UniquePipelineLayout single_texture_pl_layout;
+    vk::ShaderModule fs_tri_vertex;
+    vk::ShaderModule color_to_ms_depth_frag;
+
+    using DepthPipeline = std::pair<DepthPipelineKey, vk::UniquePipeline>;
+    std::vector<DepthPipeline> color_to_ms_depth_pl{};
+};
+
+} // namespace VideoCore

src/video_core/texture_cache/image_info.h

@@ -47,6 +47,7 @@ struct ImageInfo {
         VAddr cmask_addr;
         VAddr fmask_addr;
         VAddr htile_addr;
+        u32 htile_clear_mask{u32(-1)};
     } meta_info{};
 
     struct {

src/video_core/texture_cache/texture_cache.cpp

@@ -22,7 +22,7 @@ static constexpr u64 NumFramesBeforeRemoval = 32;
 TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
                            BufferCache& buffer_cache_, PageManager& tracker_)
     : instance{instance_}, scheduler{scheduler_}, buffer_cache{buffer_cache_}, tracker{tracker_},
-      tile_manager{instance, scheduler} {
+      blit_helper{instance, scheduler}, tile_manager{instance, scheduler} {
     // Create basic null image at fixed image ID.
     const auto null_id = GetNullImage(vk::Format::eR8G8B8A8Unorm);
     ASSERT(null_id.index == NULL_IMAGE_ID.index);
@@ -177,10 +177,20 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested_info, Bindi
         auto& new_image = slot_images[new_image_id];
         new_image.usage = cache_image.usage;
         new_image.flags &= ~ImageFlagBits::Dirty;
+        // When creating a depth buffer through overlap resolution don't clear it on first use.
+        new_image.info.meta_info.htile_clear_mask = 0;
 
+        if (cache_image.info.num_samples == 1 && new_info.num_samples == 1) {
             // Perform depth<->color copy using the intermediate copy buffer.
             const auto& copy_buffer = buffer_cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
             new_image.CopyImageWithBuffer(cache_image, copy_buffer.Handle(), 0);
+        } else if (cache_image.info.num_samples == 1 && new_info.IsDepthStencil() &&
+                   new_info.num_samples > 1) {
+            // Perform a rendering pass to transfer the channels of source as samples in dest.
+            blit_helper.BlitColorToMsDepth(cache_image, new_image);
+        } else {
+            LOG_WARNING(Render_Vulkan, "Unimplemented depth overlap copy");
+        }
 
         // Free the cache image.
         FreeImage(cache_image_id);
@@ -202,7 +212,8 @@ std::tuple<ImageId, int, int> TextureCache::ResolveOverlap(const ImageInfo& imag
 
     if (image_info.guest_address == tex_cache_image.info.guest_address) { // Equal address
         if (image_info.BlockDim() != tex_cache_image.info.BlockDim() ||
-            image_info.num_bits != tex_cache_image.info.num_bits) {
+            image_info.num_bits * image_info.num_samples !=
+                tex_cache_image.info.num_bits * tex_cache_image.info.num_samples) {
             // Very likely this kind of overlap is caused by allocation from a pool.
             if (safe_to_delete) {
                 FreeImage(cache_image_id);
@@ -470,8 +481,10 @@ ImageView& TextureCache::FindDepthTarget(BaseDesc& desc) {
     // Register meta data for this depth buffer
     if (!(image.flags & ImageFlagBits::MetaRegistered)) {
         if (desc.info.meta_info.htile_addr) {
-            surface_metas.emplace(desc.info.meta_info.htile_addr,
-                                  MetaDataInfo{.type = MetaDataInfo::Type::HTile});
+            surface_metas.emplace(
+                desc.info.meta_info.htile_addr,
+                MetaDataInfo{.type = MetaDataInfo::Type::HTile,
+                             .clear_mask = image.info.meta_info.htile_clear_mask});
             image.info.meta_info.htile_addr = desc.info.meta_info.htile_addr;
             image.flags |= ImageFlagBits::MetaRegistered;
         }

src/video_core/texture_cache/texture_cache.h

@@ -9,6 +9,7 @@
 #include "common/slot_vector.h"
 #include "video_core/amdgpu/resource.h"
 #include "video_core/multi_level_page_table.h"
+#include "video_core/texture_cache/blit_helper.h"
 #include "video_core/texture_cache/image.h"
 #include "video_core/texture_cache/image_view.h"
 #include "video_core/texture_cache/sampler.h"
@@ -286,6 +287,7 @@ private:
     Vulkan::Scheduler& scheduler;
     BufferCache& buffer_cache;
     PageManager& tracker;
+    BlitHelper blit_helper;
     TileManager tile_manager;
     Common::SlotVector<Image> slot_images;
     Common::SlotVector<ImageView> slot_image_views;