Merge pull request #11556 from GPUCode/msaa-image-vk

renderer_vulkan: Implement MSAA image copies

src/video_core/host_shaders/convert_msaa_to_non_msaa.comp

@@ -15,11 +15,14 @@ void main() {
 
     // TODO: Specialization constants for num_samples?
     const int num_samples = imageSamples(msaa_in);
+    const ivec3 msaa_size = imageSize(msaa_in);
+    const ivec3 out_size = imageSize(output_img);
+    const ivec3 scale = out_size / msaa_size;
     for (int curr_sample = 0; curr_sample < num_samples; ++curr_sample) {
         const vec4 pixel = imageLoad(msaa_in, coords, curr_sample);
 
-        const int single_sample_x = 2 * coords.x + (curr_sample & 1);
-        const int single_sample_y = 2 * coords.y + ((curr_sample / 2) & 1);
+        const int single_sample_x = scale.x * coords.x + (curr_sample & 1);
+        const int single_sample_y = scale.y * coords.y + ((curr_sample / 2) & 1);
         const ivec3 dest_coords = ivec3(single_sample_x, single_sample_y, coords.z);
 
         if (any(greaterThanEqual(dest_coords, imageSize(output_img)))) {

src/video_core/host_shaders/convert_non_msaa_to_msaa.comp

@@ -15,9 +15,12 @@ void main() {
 
     // TODO: Specialization constants for num_samples?
     const int num_samples = imageSamples(output_msaa);
+    const ivec3 msaa_size = imageSize(output_msaa);
+    const ivec3 out_size = imageSize(img_in);
+    const ivec3 scale = out_size / msaa_size;
     for (int curr_sample = 0; curr_sample < num_samples; ++curr_sample) {
-        const int single_sample_x = 2 * coords.x + (curr_sample & 1);
-        const int single_sample_y = 2 * coords.y + ((curr_sample / 2) & 1);
+        const int single_sample_x = scale.x * coords.x + (curr_sample & 1);
+        const int single_sample_y = scale.y * coords.y + ((curr_sample / 2) & 1);
         const ivec3 single_coords = ivec3(single_sample_x, single_sample_y, coords.z);
 
         if (any(greaterThanEqual(single_coords, imageSize(img_in)))) {

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -3,6 +3,7 @@
 
 #include <array>
 #include <memory>
+#include <numeric>
 #include <optional>
 #include <utility>
 
@@ -11,7 +12,10 @@
 #include "common/assert.h"
 #include "common/common_types.h"
 #include "common/div_ceil.h"
+#include "common/vector_math.h"
 #include "video_core/host_shaders/astc_decoder_comp_spv.h"
+#include "video_core/host_shaders/convert_msaa_to_non_msaa_comp_spv.h"
+#include "video_core/host_shaders/convert_non_msaa_to_msaa_comp_spv.h"
 #include "video_core/host_shaders/queries_prefix_scan_sum_comp_spv.h"
 #include "video_core/host_shaders/queries_prefix_scan_sum_nosubgroups_comp_spv.h"
 #include "video_core/host_shaders/resolve_conditional_render_comp_spv.h"
@@ -131,6 +135,33 @@ constexpr DescriptorBankInfo ASTC_BANK_INFO{
     .score = 2,
 };
 
+constexpr std::array<VkDescriptorSetLayoutBinding, ASTC_NUM_BINDINGS> MSAA_DESCRIPTOR_SET_BINDINGS{{
+    {
+        .binding = 0,
+        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+        .descriptorCount = 1,
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .pImmutableSamplers = nullptr,
+    },
+    {
+        .binding = 1,
+        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+        .descriptorCount = 1,
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .pImmutableSamplers = nullptr,
+    },
+}};
+
+constexpr DescriptorBankInfo MSAA_BANK_INFO{
+    .uniform_buffers = 0,
+    .storage_buffers = 0,
+    .texture_buffers = 0,
+    .image_buffers = 0,
+    .textures = 0,
+    .images = 2,
+    .score = 2,
+};
+
 constexpr VkDescriptorUpdateTemplateEntry INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE{
     .dstBinding = 0,
     .dstArrayElement = 0,
@@ -149,6 +180,15 @@ constexpr VkDescriptorUpdateTemplateEntry QUERIES_SCAN_DESCRIPTOR_UPDATE_TEMPLAT
     .stride = sizeof(DescriptorUpdateEntry),
 };
 
+constexpr VkDescriptorUpdateTemplateEntry MSAA_DESCRIPTOR_UPDATE_TEMPLATE{
+    .dstBinding = 0,
+    .dstArrayElement = 0,
+    .descriptorCount = 2,
+    .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+    .offset = 0,
+    .stride = sizeof(DescriptorUpdateEntry),
+};
+
 constexpr std::array<VkDescriptorUpdateTemplateEntry, ASTC_NUM_BINDINGS>
     ASTC_PASS_DESCRIPTOR_UPDATE_TEMPLATE_ENTRY{{
         {
@@ -224,6 +264,9 @@ ComputePass::ComputePass(const Device& device_, DescriptorPool& descriptor_pool,
         });
         descriptor_allocator = descriptor_pool.Allocator(*descriptor_set_layout, bank_info);
     }
+    if (code.empty()) {
+        return;
+    }
     module = device.GetLogical().CreateShaderModule({
         .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
         .pNext = nullptr,
@@ -590,4 +633,100 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
     scheduler.Finish();
 }
 
+MSAACopyPass::MSAACopyPass(const Device& device_, Scheduler& scheduler_,
+                           DescriptorPool& descriptor_pool_,
+                           StagingBufferPool& staging_buffer_pool_,
+                           ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, descriptor_pool_, MSAA_DESCRIPTOR_SET_BINDINGS,
+                  MSAA_DESCRIPTOR_UPDATE_TEMPLATE, MSAA_BANK_INFO, {},
+                  CONVERT_NON_MSAA_TO_MSAA_COMP_SPV),
+      scheduler{scheduler_}, staging_buffer_pool{staging_buffer_pool_},
+      compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {
+    const auto make_msaa_pipeline = [this](size_t i, std::span<const u32> code) {
+        modules[i] = device.GetLogical().CreateShaderModule({
+            .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .codeSize = static_cast<u32>(code.size_bytes()),
+            .pCode = code.data(),
+        });
+        pipelines[i] = device.GetLogical().CreateComputePipeline({
+            .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .stage{
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+                .pNext = nullptr,
+                .flags = 0,
+                .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+                .module = *modules[i],
+                .pName = "main",
+                .pSpecializationInfo = nullptr,
+            },
+            .layout = *layout,
+            .basePipelineHandle = nullptr,
+            .basePipelineIndex = 0,
+        });
+    };
+    make_msaa_pipeline(0, CONVERT_NON_MSAA_TO_MSAA_COMP_SPV);
+    make_msaa_pipeline(1, CONVERT_MSAA_TO_NON_MSAA_COMP_SPV);
+}
+
+MSAACopyPass::~MSAACopyPass() = default;
+
+void MSAACopyPass::CopyImage(Image& dst_image, Image& src_image,
+                             std::span<const VideoCommon::ImageCopy> copies,
+                             bool msaa_to_non_msaa) {
+    const VkPipeline msaa_pipeline = *pipelines[msaa_to_non_msaa ? 1 : 0];
+    scheduler.RequestOutsideRenderPassOperationContext();
+    for (const VideoCommon::ImageCopy& copy : copies) {
+        ASSERT(copy.src_subresource.base_layer == 0);
+        ASSERT(copy.src_subresource.num_layers == 1);
+        ASSERT(copy.dst_subresource.base_layer == 0);
+        ASSERT(copy.dst_subresource.num_layers == 1);
+
+        compute_pass_descriptor_queue.Acquire();
+        compute_pass_descriptor_queue.AddImage(
+            src_image.StorageImageView(copy.src_subresource.base_level));
+        compute_pass_descriptor_queue.AddImage(
+            dst_image.StorageImageView(copy.dst_subresource.base_level));
+        const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+
+        const Common::Vec3<u32> num_dispatches = {
+            Common::DivCeil(copy.extent.width, 8U),
+            Common::DivCeil(copy.extent.height, 8U),
+            copy.extent.depth,
+        };
+
+        scheduler.Record([this, dst = dst_image.Handle(), msaa_pipeline, num_dispatches,
+                          descriptor_data](vk::CommandBuffer cmdbuf) {
+            const VkDescriptorSet set = descriptor_allocator.Commit();
+            device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
+            cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, msaa_pipeline);
+            cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
+            cmdbuf.Dispatch(num_dispatches.x, num_dispatches.y, num_dispatches.z);
+            const VkImageMemoryBarrier write_barrier{
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
+                .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+                .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = dst,
+                .subresourceRange{
+                    .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+                    .baseMipLevel = 0,
+                    .levelCount = VK_REMAINING_MIP_LEVELS,
+                    .baseArrayLayer = 0,
+                    .layerCount = VK_REMAINING_ARRAY_LAYERS,
+                },
+            };
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                                   VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, write_barrier);
+        });
+    }
+}
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -11,6 +11,7 @@
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_update_descriptor.h"
+#include "video_core/texture_cache/types.h"
 #include "video_core/vulkan_common/vulkan_memory_allocator.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
@@ -130,4 +131,22 @@ private:
     MemoryAllocator& memory_allocator;
 };
 
+class MSAACopyPass final : public ComputePass {
+public:
+    explicit MSAACopyPass(const Device& device_, Scheduler& scheduler_,
+                          DescriptorPool& descriptor_pool_, StagingBufferPool& staging_buffer_pool_,
+                          ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~MSAACopyPass();
+
+    void CopyImage(Image& dst_image, Image& src_image,
+                   std::span<const VideoCommon::ImageCopy> copies, bool msaa_to_non_msaa);
+
+private:
+    Scheduler& scheduler;
+    StagingBufferPool& staging_buffer_pool;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+    std::array<vk::ShaderModule, 2> modules;
+    std::array<vk::Pipeline, 2> pipelines;
+};
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -176,6 +176,36 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     return allocator.CreateImage(image_ci);
 }
 
+[[nodiscard]] vk::ImageView MakeStorageView(const vk::Device& device, u32 level, VkImage image,
+                                            VkFormat format) {
+    static constexpr VkImageViewUsageCreateInfo storage_image_view_usage_create_info{
+        .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_USAGE_CREATE_INFO,
+        .pNext = nullptr,
+        .usage = VK_IMAGE_USAGE_STORAGE_BIT,
+    };
+    return device.CreateImageView(VkImageViewCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+        .pNext = &storage_image_view_usage_create_info,
+        .flags = 0,
+        .image = image,
+        .viewType = VK_IMAGE_VIEW_TYPE_2D_ARRAY,
+        .format = format,
+        .components{
+            .r = VK_COMPONENT_SWIZZLE_IDENTITY,
+            .g = VK_COMPONENT_SWIZZLE_IDENTITY,
+            .b = VK_COMPONENT_SWIZZLE_IDENTITY,
+            .a = VK_COMPONENT_SWIZZLE_IDENTITY,
+        },
+        .subresourceRange{
+            .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            .baseMipLevel = level,
+            .levelCount = 1,
+            .baseArrayLayer = 0,
+            .layerCount = VK_REMAINING_ARRAY_LAYERS,
+        },
+    });
+}
+
 [[nodiscard]] VkImageAspectFlags ImageAspectMask(PixelFormat format) {
     switch (VideoCore::Surface::GetFormatType(format)) {
     case VideoCore::Surface::SurfaceType::ColorTexture:
@@ -817,6 +847,10 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, Scheduler& sched
         astc_decoder_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
                                   compute_pass_descriptor_queue, memory_allocator);
     }
+    if (device.IsStorageImageMultisampleSupported()) {
+        msaa_copy_pass = std::make_unique<MSAACopyPass>(
+            device, scheduler, descriptor_pool, staging_buffer_pool, compute_pass_descriptor_queue);
+    }
     if (!device.IsKhrImageFormatListSupported()) {
         return;
     }
@@ -1285,7 +1319,11 @@ void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
 
 void TextureCacheRuntime::CopyImageMSAA(Image& dst, Image& src,
                                         std::span<const VideoCommon::ImageCopy> copies) {
-    UNIMPLEMENTED_MSG("Copying images with different samples is not implemented in Vulkan.");
+    const bool msaa_to_non_msaa = src.info.num_samples > 1 && dst.info.num_samples == 1;
+    if (msaa_copy_pass) {
+        return msaa_copy_pass->CopyImage(dst, src, copies, msaa_to_non_msaa);
+    }
+    UNIMPLEMENTED_MSG("Copying images with different samples is not supported.");
 }
 
 u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
@@ -1333,39 +1371,15 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
     if (runtime->device.HasDebuggingToolAttached()) {
         original_image.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
     }
-    static constexpr VkImageViewUsageCreateInfo storage_image_view_usage_create_info{
-        .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_USAGE_CREATE_INFO,
-        .pNext = nullptr,
-        .usage = VK_IMAGE_USAGE_STORAGE_BIT,
-    };
     current_image = *original_image;
+    storage_image_views.resize(info.resources.levels);
     if (IsPixelFormatASTC(info.format) && !runtime->device.IsOptimalAstcSupported() &&
         Settings::values.astc_recompression.GetValue() ==
             Settings::AstcRecompression::Uncompressed) {
         const auto& device = runtime->device.GetLogical();
-        storage_image_views.reserve(info.resources.levels);
         for (s32 level = 0; level < info.resources.levels; ++level) {
-            storage_image_views.push_back(device.CreateImageView(VkImageViewCreateInfo{
-                .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
-                .pNext = &storage_image_view_usage_create_info,
-                .flags = 0,
-                .image = *original_image,
-                .viewType = VK_IMAGE_VIEW_TYPE_2D_ARRAY,
-                .format = VK_FORMAT_A8B8G8R8_UNORM_PACK32,
-                .components{
-                    .r = VK_COMPONENT_SWIZZLE_IDENTITY,
-                    .g = VK_COMPONENT_SWIZZLE_IDENTITY,
-                    .b = VK_COMPONENT_SWIZZLE_IDENTITY,
-                    .a = VK_COMPONENT_SWIZZLE_IDENTITY,
-                },
-                .subresourceRange{
-                    .aspectMask = aspect_mask,
-                    .baseMipLevel = static_cast<u32>(level),
-                    .levelCount = 1,
-                    .baseArrayLayer = 0,
-                    .layerCount = VK_REMAINING_ARRAY_LAYERS,
-                },
-            }));
+            storage_image_views[level] =
+                MakeStorageView(device, level, *original_image, VK_FORMAT_A8B8G8R8_UNORM_PACK32);
         }
     }
 }
@@ -1496,6 +1510,17 @@ void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferIm
     DownloadMemory(buffers, offsets, copies);
 }
 
+VkImageView Image::StorageImageView(s32 level) noexcept {
+    auto& view = storage_image_views[level];
+    if (!view) {
+        const auto format_info =
+            MaxwellToVK::SurfaceFormat(runtime->device, FormatType::Optimal, true, info.format);
+        view =
+            MakeStorageView(runtime->device.GetLogical(), level, current_image, format_info.format);
+    }
+    return *view;
+}
+
 bool Image::IsRescaled() const noexcept {
     return True(flags & ImageFlagBits::Rescaled);
 }

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -117,6 +117,7 @@ public:
     BlitImageHelper& blit_image_helper;
     RenderPassCache& render_pass_cache;
     std::optional<ASTCDecoderPass> astc_decoder_pass;
+    std::unique_ptr<MSAACopyPass> msaa_copy_pass;
     const Settings::ResolutionScalingInfo& resolution;
     std::array<std::vector<VkFormat>, VideoCore::Surface::MaxPixelFormat> view_formats;
 
@@ -161,15 +162,13 @@ public:
         return aspect_mask;
     }
 
-    [[nodiscard]] VkImageView StorageImageView(s32 level) const noexcept {
-        return *storage_image_views[level];
-    }
-
     /// Returns true when the image is already initialized and mark it as initialized
     [[nodiscard]] bool ExchangeInitialization() noexcept {
         return std::exchange(initialized, true);
     }
 
+    VkImageView StorageImageView(s32 level) noexcept;
+
     bool IsRescaled() const noexcept;
 
     bool ScaleUp(bool ignore = false);

src/video_core/vulkan_common/vulkan_device.h

@@ -324,6 +324,11 @@ public:
         return features.shader_float16_int8.shaderInt8;
     }
 
+    /// Returns true if the device supports binding multisample images as storage images.
+    bool IsStorageImageMultisampleSupported() const {
+        return features.features.shaderStorageImageMultisample;
+    }
+
     /// Returns true if the device warp size can potentially be bigger than guest's warp size.
     bool IsWarpSizePotentiallyBiggerThanGuest() const {
         return is_warp_potentially_bigger;