shader: Implement SULD and SUST

8 files changed, 135 insertions, 65 deletions

diff --git a/src/video_core/renderer_vulkan/blit_image.cpp b/src/video_core/renderer_vulkan/blit_image.cpp
index 6c0d5c7f4..39fe9289b 100644
--- a/src/video_core/renderer_vulkan/blit_image.cpp
+++ b/src/video_core/renderer_vulkan/blit_image.cpp
@@ -361,7 +361,7 @@ void BlitImageHelper::BlitColor(const Framebuffer* dst_framebuffer, const ImageV
         .operation = operation,
     };
     const VkPipelineLayout layout = *one_texture_pipeline_layout;
-    const VkImageView src_view = src_image_view.Handle(ImageViewType::e2D);
+    const VkImageView src_view = src_image_view.Handle(Shader::TextureType::Color2D);
     const VkSampler sampler = is_linear ? *linear_sampler : *nearest_sampler;
     const VkPipeline pipeline = FindOrEmplacePipeline(key);
     const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
@@ -435,7 +435,7 @@ void BlitImageHelper::ConvertR16ToD16(const Framebuffer* dst_framebuffer,
 void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
                               const ImageView& src_image_view) {
     const VkPipelineLayout layout = *one_texture_pipeline_layout;
-    const VkImageView src_view = src_image_view.Handle(ImageViewType::e2D);
+    const VkImageView src_view = src_image_view.Handle(Shader::TextureType::Color2D);
     const VkSampler sampler = *nearest_sampler;
     const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
     const VkExtent2D extent{
diff --git a/src/video_core/renderer_vulkan/pipeline_helper.h b/src/video_core/renderer_vulkan/pipeline_helper.h
index cff93cc60..d2c3f11c1 100644
--- a/src/video_core/renderer_vulkan/pipeline_helper.h
+++ b/src/video_core/renderer_vulkan/pipeline_helper.h
@@ -97,6 +97,9 @@ public:
         for ([[maybe_unused]] const auto& desc : info.texture_descriptors) {
             Add(VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, stage);
         }
+        for (const auto& desc : info.image_descriptors) {
+            Add(VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, stage);
+        }
     }
 
 private:
@@ -127,36 +130,6 @@ private:
     size_t offset{};
 };
 
-inline VideoCommon::ImageViewType CastType(Shader::TextureType type) {
-    switch (type) {
-    case Shader::TextureType::Color1D:
-    case Shader::TextureType::Shadow1D:
-        return VideoCommon::ImageViewType::e1D;
-    case Shader::TextureType::ColorArray1D:
-    case Shader::TextureType::ShadowArray1D:
-        return VideoCommon::ImageViewType::e1DArray;
-    case Shader::TextureType::Color2D:
-    case Shader::TextureType::Shadow2D:
-        return VideoCommon::ImageViewType::e2D;
-    case Shader::TextureType::ColorArray2D:
-    case Shader::TextureType::ShadowArray2D:
-        return VideoCommon::ImageViewType::e2DArray;
-    case Shader::TextureType::Color3D:
-    case Shader::TextureType::Shadow3D:
-        return VideoCommon::ImageViewType::e3D;
-    case Shader::TextureType::ColorCube:
-    case Shader::TextureType::ShadowCube:
-        return VideoCommon::ImageViewType::Cube;
-    case Shader::TextureType::ColorArrayCube:
-    case Shader::TextureType::ShadowArrayCube:
-        return VideoCommon::ImageViewType::CubeArray;
-    case Shader::TextureType::Buffer:
-        break;
-    }
-    UNREACHABLE_MSG("Invalid texture type {}", type);
-    return {};
-}
-
 inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& samplers,
                                  const ImageId*& image_view_ids, TextureCache& texture_cache,
                                  VKUpdateDescriptorQueue& update_descriptor_queue) {
@@ -164,9 +137,17 @@ inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& sam
     for (const auto& desc : info.texture_descriptors) {
         const VkSampler sampler{*(samplers++)};
         ImageView& image_view{texture_cache.GetImageView(*(image_view_ids++))};
-        const VkImageView vk_image_view{image_view.Handle(CastType(desc.type))};
+        const VkImageView vk_image_view{image_view.Handle(desc.type)};
         update_descriptor_queue.AddSampledImage(vk_image_view, sampler);
     }
+    for (const auto& desc : info.image_descriptors) {
+        ImageView& image_view{texture_cache.GetImageView(*(image_view_ids++))};
+        if (desc.is_written) {
+            texture_cache.MarkModification(image_view.image_id);
+        }
+        const VkImageView vk_image_view{image_view.StorageView(desc.type, desc.format)};
+        update_descriptor_queue.AddImage(vk_image_view);
+    }
 }
 
 } // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp b/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
index ac47b1f3c..3d690f335 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
@@ -108,6 +108,10 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
         Sampler* const sampler = texture_cache.GetComputeSampler(handle.sampler);
         samplers.push_back(sampler->Handle());
     }
+    for (const auto& desc : info.image_descriptors) {
+        const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
+        image_view_indices.push_back(handle.image);
+    }
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
     texture_cache.FillComputeImageViews(indices_span, image_view_ids);
 
diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
index 57e2d569c..23c01f24e 100644
--- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
+++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
@@ -186,6 +186,10 @@ void GraphicsPipeline::Configure(bool is_indexed) {
             Sampler* const sampler{texture_cache.GetGraphicsSampler(handle.sampler)};
             samplers.push_back(sampler->Handle());
         }
+        for (const auto& desc : info.image_descriptors) {
+            const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
+            image_view_indices.push_back(handle.image);
+        }
     }
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
     texture_cache.FillGraphicsImageViews(indices_span, image_view_ids);
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 0292a1b94..2ba44330f 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -494,7 +494,7 @@ bool RasterizerVulkan::AccelerateDisplay(const Tegra::FramebufferConfig& config,
     if (!image_view) {
         return false;
     }
-    screen_info.image_view = image_view->Handle(VideoCommon::ImageViewType::e2D);
+    screen_info.image_view = image_view->Handle(Shader::TextureType::Color2D);
     screen_info.width = image_view->size.width;
     screen_info.height = image_view->size.height;
     screen_info.is_srgb = VideoCore::Surface::IsPixelFormatSRGB(image_view->format);
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index 70328680d..8e029bcb3 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -215,6 +215,30 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     return VK_COMPONENT_SWIZZLE_ZERO;
 }
 
+[[nodiscard]] VkImageViewType ImageViewType(Shader::TextureType type) {
+    switch (type) {
+    case Shader::TextureType::Color1D:
+        return VK_IMAGE_VIEW_TYPE_1D;
+    case Shader::TextureType::Color2D:
+        return VK_IMAGE_VIEW_TYPE_2D;
+    case Shader::TextureType::ColorCube:
+        return VK_IMAGE_VIEW_TYPE_CUBE;
+    case Shader::TextureType::Color3D:
+        return VK_IMAGE_VIEW_TYPE_3D;
+    case Shader::TextureType::ColorArray1D:
+        return VK_IMAGE_VIEW_TYPE_1D_ARRAY;
+    case Shader::TextureType::ColorArray2D:
+        return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
+    case Shader::TextureType::ColorArrayCube:
+        return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
+    case Shader::TextureType::Buffer:
+        UNREACHABLE_MSG("Texture buffers can't be image views");
+        return VK_IMAGE_VIEW_TYPE_1D;
+    }
+    UNREACHABLE_MSG("Invalid image view type={}", type);
+    return VK_IMAGE_VIEW_TYPE_2D;
+}
+
 [[nodiscard]] VkImageViewType ImageViewType(VideoCommon::ImageViewType type) {
     switch (type) {
     case VideoCommon::ImageViewType::e1D:
@@ -232,7 +256,7 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     case VideoCommon::ImageViewType::CubeArray:
         return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
     case VideoCommon::ImageViewType::Rect:
-        LOG_WARNING(Render_Vulkan, "Unnormalized image view type not supported");
+        UNIMPLEMENTED_MSG("Rect image view");
         return VK_IMAGE_VIEW_TYPE_2D;
     case VideoCommon::ImageViewType::Buffer:
         UNREACHABLE_MSG("Texture buffers can't be image views");
@@ -539,6 +563,28 @@ struct RangedBarrierRange {
     }
 };
 
+[[nodiscard]] VkFormat Format(Shader::ImageFormat format) {
+    switch (format) {
+    case Shader::ImageFormat::Typeless:
+        break;
+    case Shader::ImageFormat::R8_SINT:
+        return VK_FORMAT_R8_SINT;
+    case Shader::ImageFormat::R8_UINT:
+        return VK_FORMAT_R8_UINT;
+    case Shader::ImageFormat::R16_UINT:
+        return VK_FORMAT_R16_UINT;
+    case Shader::ImageFormat::R16_SINT:
+        return VK_FORMAT_R16_SINT;
+    case Shader::ImageFormat::R32_UINT:
+        return VK_FORMAT_R32_UINT;
+    case Shader::ImageFormat::R32G32_UINT:
+        return VK_FORMAT_R32G32_UINT;
+    case Shader::ImageFormat::R32G32B32A32_UINT:
+        return VK_FORMAT_R32G32B32A32_UINT;
+    }
+    UNREACHABLE_MSG("Invalid image format={}", format);
+    return VK_FORMAT_R32_UINT;
+}
 } // Anonymous namespace
 
 void TextureCacheRuntime::Finish() {
@@ -577,7 +623,7 @@ void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst
             return;
         }
     }
-    ASSERT(src.ImageFormat() == dst.ImageFormat());
+    ASSERT(src.format == dst.format);
     ASSERT(!(is_dst_msaa && !is_src_msaa));
     ASSERT(operation == Fermi2D::Operation::SrcCopy);
 
@@ -915,8 +961,9 @@ void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferIm
 ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewInfo& info,
                      ImageId image_id_, Image& image)
     : VideoCommon::ImageViewBase{info, image.info, image_id_}, device{&runtime.device},
-      image_handle{image.Handle()}, image_format{image.info.format}, samples{ConvertSampleCount(
-                                                                         image.info.num_samples)} {
+      image_handle{image.Handle()}, samples{ConvertSampleCount(image.info.num_samples)} {
+    using Shader::TextureType;
+
     const VkImageAspectFlags aspect_mask = ImageViewAspectMask(info);
     std::array<SwizzleSource, 4> swizzle{
         SwizzleSource::R,
@@ -954,39 +1001,39 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
         },
         .subresourceRange = MakeSubresourceRange(aspect_mask, info.range),
     };
-    const auto create = [&](VideoCommon::ImageViewType view_type, std::optional<u32> num_layers) {
+    const auto create = [&](TextureType tex_type, std::optional<u32> num_layers) {
         VkImageViewCreateInfo ci{create_info};
-        ci.viewType = ImageViewType(view_type);
+        ci.viewType = ImageViewType(tex_type);
         if (num_layers) {
             ci.subresourceRange.layerCount = *num_layers;
         }
         vk::ImageView handle = device->GetLogical().CreateImageView(ci);
         if (device->HasDebuggingToolAttached()) {
-            handle.SetObjectNameEXT(VideoCommon::Name(*this, view_type).c_str());
+            handle.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
         }
-        image_views[static_cast<size_t>(view_type)] = std::move(handle);
+        image_views[static_cast<size_t>(tex_type)] = std::move(handle);
     };
     switch (info.type) {
     case VideoCommon::ImageViewType::e1D:
     case VideoCommon::ImageViewType::e1DArray:
-        create(VideoCommon::ImageViewType::e1D, 1);
-        create(VideoCommon::ImageViewType::e1DArray, std::nullopt);
-        render_target = Handle(VideoCommon::ImageViewType::e1DArray);
+        create(TextureType::Color1D, 1);
+        create(TextureType::ColorArray1D, std::nullopt);
+        render_target = Handle(TextureType::ColorArray1D);
         break;
     case VideoCommon::ImageViewType::e2D:
     case VideoCommon::ImageViewType::e2DArray:
-        create(VideoCommon::ImageViewType::e2D, 1);
-        create(VideoCommon::ImageViewType::e2DArray, std::nullopt);
-        render_target = Handle(VideoCommon::ImageViewType::e2DArray);
+        create(TextureType::Color2D, 1);
+        create(TextureType::ColorArray2D, std::nullopt);
+        render_target = Handle(Shader::TextureType::ColorArray2D);
         break;
     case VideoCommon::ImageViewType::e3D:
-        create(VideoCommon::ImageViewType::e3D, std::nullopt);
-        render_target = Handle(VideoCommon::ImageViewType::e3D);
+        create(TextureType::Color3D, std::nullopt);
+        render_target = Handle(Shader::TextureType::Color3D);
         break;
     case VideoCommon::ImageViewType::Cube:
     case VideoCommon::ImageViewType::CubeArray:
-        create(VideoCommon::ImageViewType::Cube, 6);
-        create(VideoCommon::ImageViewType::CubeArray, std::nullopt);
+        create(TextureType::ColorCube, 6);
+        create(TextureType::ColorArrayCube, std::nullopt);
         break;
     case VideoCommon::ImageViewType::Rect:
         UNIMPLEMENTED();
@@ -1009,7 +1056,8 @@ VkImageView ImageView::DepthView() {
     if (depth_view) {
         return *depth_view;
     }
-    depth_view = MakeDepthStencilView(VK_IMAGE_ASPECT_DEPTH_BIT);
+    const auto& info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
+    depth_view = MakeView(info.format, VK_IMAGE_ASPECT_DEPTH_BIT);
     return *depth_view;
 }
 
@@ -1017,18 +1065,38 @@ VkImageView ImageView::StencilView() {
     if (stencil_view) {
         return *stencil_view;
     }
-    stencil_view = MakeDepthStencilView(VK_IMAGE_ASPECT_STENCIL_BIT);
+    const auto& info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
+    stencil_view = MakeView(info.format, VK_IMAGE_ASPECT_STENCIL_BIT);
     return *stencil_view;
 }
 
-vk::ImageView ImageView::MakeDepthStencilView(VkImageAspectFlags aspect_mask) {
+VkImageView ImageView::StorageView(Shader::TextureType texture_type,
+                                   Shader::ImageFormat image_format) {
+    if (image_format == Shader::ImageFormat::Typeless) {
+        return Handle(texture_type);
+    }
+    const bool is_signed{image_format == Shader::ImageFormat::R8_SINT ||
+                         image_format == Shader::ImageFormat::R16_SINT};
+    if (!storage_views) {
+        storage_views = std::make_unique<StorageViews>();
+    }
+    auto& views{is_signed ? storage_views->signeds : storage_views->unsigneds};
+    auto& view{views[static_cast<size_t>(texture_type)]};
+    if (view) {
+        return *view;
+    }
+    view = MakeView(Format(image_format), VK_IMAGE_ASPECT_COLOR_BIT);
+    return *view;
+}
+
+vk::ImageView ImageView::MakeView(VkFormat vk_format, VkImageAspectFlags aspect_mask) {
     return device->GetLogical().CreateImageView({
         .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
         .image = image_handle,
         .viewType = ImageViewType(type),
-        .format = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format).format,
+        .format = vk_format,
         .components{
             .r = VK_COMPONENT_SWIZZLE_IDENTITY,
             .g = VK_COMPONENT_SWIZZLE_IDENTITY,
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h
index 498e76a1c..0b73d55f8 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.h
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.h
@@ -7,6 +7,7 @@
 #include <compare>
 #include <span>
 
+#include "shader_recompiler/shader_info.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
 #include "video_core/texture_cache/texture_cache.h"
 #include "video_core/vulkan_common/vulkan_memory_allocator.h"
@@ -133,8 +134,11 @@ public:
 
     [[nodiscard]] VkImageView StencilView();
 
-    [[nodiscard]] VkImageView Handle(VideoCommon::ImageViewType query_type) const noexcept {
-        return *image_views[static_cast<size_t>(query_type)];
+    [[nodiscard]] VkImageView StorageView(Shader::TextureType texture_type,
+                                          Shader::ImageFormat image_format);
+
+    [[nodiscard]] VkImageView Handle(Shader::TextureType texture_type) const noexcept {
+        return *image_views[static_cast<size_t>(texture_type)];
     }
 
     [[nodiscard]] VkImage ImageHandle() const noexcept {
@@ -145,10 +149,6 @@ public:
         return render_target;
     }
 
-    [[nodiscard]] PixelFormat ImageFormat() const noexcept {
-        return image_format;
-    }
-
     [[nodiscard]] VkSampleCountFlagBits Samples() const noexcept {
         return samples;
     }
@@ -162,15 +162,20 @@ public:
     }
 
 private:
-    [[nodiscard]] vk::ImageView MakeDepthStencilView(VkImageAspectFlags aspect_mask);
+    struct StorageViews {
+        std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> signeds;
+        std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> unsigneds;
+    };
+
+    [[nodiscard]] vk::ImageView MakeView(VkFormat vk_format, VkImageAspectFlags aspect_mask);
 
     const Device* device = nullptr;
-    std::array<vk::ImageView, VideoCommon::NUM_IMAGE_VIEW_TYPES> image_views;
+    std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> image_views;
+    std::unique_ptr<StorageViews> storage_views;
     vk::ImageView depth_view;
     vk::ImageView stencil_view;
     VkImage image_handle = VK_NULL_HANDLE;
     VkImageView render_target = VK_NULL_HANDLE;
-    PixelFormat image_format = PixelFormat::Invalid;
     VkSampleCountFlagBits samples = VK_SAMPLE_COUNT_1_BIT;
     GPUVAddr gpu_addr = 0;
     u32 buffer_size = 0;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 5e8d99482..255b07cf8 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -117,6 +117,9 @@ public:
     /// Return a reference to the given image view id
     [[nodiscard]] ImageView& GetImageView(ImageViewId id) noexcept;
 
+    /// Mark an image as modified from the GPU
+    void MarkModification(ImageId id) noexcept;
+
     /// Fill image_view_ids with the graphics images in indices
     void FillGraphicsImageViews(std::span<const u32> indices,
                                 std::span<ImageViewId> image_view_ids);
@@ -527,6 +530,11 @@ typename P::ImageView& TextureCache<P>::GetImageView(ImageViewId id) noexcept {
 }
 
 template <class P>
+void TextureCache<P>::MarkModification(ImageId id) noexcept {
+    MarkModification(slot_images[id]);
+}
+
+template <class P>
 void TextureCache<P>::FillGraphicsImageViews(std::span<const u32> indices,
                                              std::span<ImageViewId> image_view_ids) {
     FillImageViews(graphics_image_table, graphics_image_view_ids, indices, image_view_ids);