spirv: Implement image buffers

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -54,28 +54,30 @@ Id ImageType(EmitContext& ctx, const TextureDescriptor& desc) {
     throw InvalidArgument("Invalid texture type {}", desc.type);
 }
 
+spv::ImageFormat GetImageFormat(ImageFormat format) {
+    switch (format) {
+    case ImageFormat::Typeless:
+        return spv::ImageFormat::Unknown;
+    case ImageFormat::R8_UINT:
+        return spv::ImageFormat::R8ui;
+    case ImageFormat::R8_SINT:
+        return spv::ImageFormat::R8i;
+    case ImageFormat::R16_UINT:
+        return spv::ImageFormat::R16ui;
+    case ImageFormat::R16_SINT:
+        return spv::ImageFormat::R16i;
+    case ImageFormat::R32_UINT:
+        return spv::ImageFormat::R32ui;
+    case ImageFormat::R32G32_UINT:
+        return spv::ImageFormat::Rg32ui;
+    case ImageFormat::R32G32B32A32_UINT:
+        return spv::ImageFormat::Rgba32ui;
+    }
+    throw InvalidArgument("Invalid image format {}", format);
+}
+
 Id ImageType(EmitContext& ctx, const ImageDescriptor& desc) {
-    const spv::ImageFormat format{[&] {
-        switch (desc.format) {
-        case ImageFormat::Typeless:
-            return spv::ImageFormat::Unknown;
-        case ImageFormat::R8_UINT:
-            return spv::ImageFormat::R8ui;
-        case ImageFormat::R8_SINT:
-            return spv::ImageFormat::R8i;
-        case ImageFormat::R16_UINT:
-            return spv::ImageFormat::R16ui;
-        case ImageFormat::R16_SINT:
-            return spv::ImageFormat::R16i;
-        case ImageFormat::R32_UINT:
-            return spv::ImageFormat::R32ui;
-        case ImageFormat::R32G32_UINT:
-            return spv::ImageFormat::Rg32ui;
-        case ImageFormat::R32G32B32A32_UINT:
-            return spv::ImageFormat::Rgba32ui;
-        }
-        throw InvalidArgument("Invalid image format {}", desc.format);
-    }()};
+    const spv::ImageFormat format{GetImageFormat(desc.format)};
     const Id type{ctx.U32[1]};
     switch (desc.type) {
     case TextureType::Color1D:
@@ -388,6 +390,7 @@ EmitContext::EmitContext(const Profile& profile_, IR::Program& program, u32& bin
     DefineConstantBuffers(program.info, binding);
     DefineStorageBuffers(program.info, binding);
     DefineTextureBuffers(program.info, binding);
+    DefineImageBuffers(program.info, binding);
     DefineTextures(program.info, binding);
     DefineImages(program.info, binding);
     DefineAttributeMemAccess(program.info);
@@ -883,6 +886,31 @@ void EmitContext::DefineTextureBuffers(const Info& info, u32& binding) {
     }
 }
 
+void EmitContext::DefineImageBuffers(const Info& info, u32& binding) {
+    image_buffers.reserve(info.image_buffer_descriptors.size());
+    for (const ImageBufferDescriptor& desc : info.image_buffer_descriptors) {
+        if (desc.count != 1) {
+            throw NotImplementedException("Array of image buffers");
+        }
+        const spv::ImageFormat format{GetImageFormat(desc.format)};
+        const Id image_type{TypeImage(U32[4], spv::Dim::Buffer, false, false, false, 2, format)};
+        const Id pointer_type{TypePointer(spv::StorageClass::UniformConstant, image_type)};
+        const Id id{AddGlobalVariable(pointer_type, spv::StorageClass::UniformConstant)};
+        Decorate(id, spv::Decoration::Binding, binding);
+        Decorate(id, spv::Decoration::DescriptorSet, 0U);
+        Name(id, fmt::format("imgbuf{}_{:02x}", desc.cbuf_index, desc.cbuf_offset));
+        const ImageBufferDefinition def{
+            .id = id,
+            .image_type = image_type,
+        };
+        image_buffers.insert(image_buffers.end(), desc.count, def);
+        if (profile.supported_spirv >= 0x00010400) {
+            interfaces.push_back(id);
+        }
+        binding += desc.count;
+    }
+}
+
 void EmitContext::DefineTextures(const Info& info, u32& binding) {
     textures.reserve(info.texture_descriptors.size());
     for (const TextureDescriptor& desc : info.texture_descriptors) {

src/shader_recompiler/backend/spirv/emit_context.h

@@ -35,6 +35,11 @@ struct TextureDefinition {
     Id image_type;
 };
 
+struct ImageBufferDefinition {
+    Id id;
+    Id image_type;
+};
+
 struct ImageDefinition {
     Id id;
     Id image_type;
@@ -136,6 +141,7 @@ public:
     std::array<UniformDefinitions, Info::MAX_CBUFS> cbufs{};
     std::array<StorageDefinitions, Info::MAX_SSBOS> ssbos{};
     std::vector<Id> texture_buffers;
+    std::vector<ImageBufferDefinition> image_buffers;
     std::vector<TextureDefinition> textures;
     std::vector<ImageDefinition> images;
 
@@ -213,6 +219,7 @@ private:
     void DefineConstantBuffers(const Info& info, u32& binding);
     void DefineStorageBuffers(const Info& info, u32& binding);
     void DefineTextureBuffers(const Info& info, u32& binding);
+    void DefineImageBuffers(const Info& info, u32& binding);
     void DefineTextures(const Info& info, u32& binding);
     void DefineImages(const Info& info, u32& binding);
     void DefineAttributeMemAccess(const Info& info);

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -149,7 +149,8 @@ Id Image(EmitContext& ctx, const IR::Value& index, IR::TextureInstInfo info) {
         throw NotImplementedException("Indirect image indexing");
     }
     if (info.type == TextureType::Buffer) {
-        throw NotImplementedException("Image buffer");
+        const ImageBufferDefinition def{ctx.image_buffers.at(index.U32())};
+        return ctx.OpLoad(def.image_type, def.id);
     } else {
         const ImageDefinition def{ctx.images.at(index.U32())};
         return ctx.OpLoad(def.image_type, def.id);

src/shader_recompiler/ir_opt/texture_pass.cpp

@@ -158,9 +158,11 @@ TextureInst MakeInst(Environment& env, IR::Block* block, IR::Inst& inst) {
 class Descriptors {
 public:
     explicit Descriptors(TextureBufferDescriptors& texture_buffer_descriptors_,
+                         ImageBufferDescriptors& image_buffer_descriptors_,
                          TextureDescriptors& texture_descriptors_,
                          ImageDescriptors& image_descriptors_)
         : texture_buffer_descriptors{texture_buffer_descriptors_},
+          image_buffer_descriptors{image_buffer_descriptors_},
           texture_descriptors{texture_descriptors_}, image_descriptors{image_descriptors_} {}
 
     u32 Add(const TextureBufferDescriptor& desc) {
@@ -170,6 +172,13 @@ public:
         });
     }
 
+    u32 Add(const ImageBufferDescriptor& desc) {
+        return Add(image_buffer_descriptors, desc, [&desc](const auto& existing) {
+            return desc.format == existing.format && desc.cbuf_index == existing.cbuf_index &&
+                   desc.cbuf_offset == existing.cbuf_offset;
+        });
+    }
+
     u32 Add(const TextureDescriptor& desc) {
         return Add(texture_descriptors, desc, [&desc](const auto& existing) {
             return desc.cbuf_index == existing.cbuf_index &&
@@ -200,6 +209,7 @@ private:
     }
 
     TextureBufferDescriptors& texture_buffer_descriptors;
+    ImageBufferDescriptors& image_buffer_descriptors;
     TextureDescriptors& texture_descriptors;
     ImageDescriptors& image_descriptors;
 };
@@ -224,6 +234,7 @@ void TexturePass(Environment& env, IR::Program& program) {
     });
     Descriptors descriptors{
         program.info.texture_buffer_descriptors,
+        program.info.image_buffer_descriptors,
         program.info.texture_descriptors,
         program.info.image_descriptors,
     };
@@ -261,7 +272,13 @@ void TexturePass(Environment& env, IR::Program& program) {
         case IR::Opcode::ImageWrite: {
             const bool is_written{inst->GetOpcode() == IR::Opcode::ImageWrite};
             if (flags.type == TextureType::Buffer) {
-                throw NotImplementedException("Image buffer");
+                index = descriptors.Add(ImageBufferDescriptor{
+                    .format = flags.image_format,
+                    .is_written = is_written,
+                    .cbuf_index = cbuf.index,
+                    .cbuf_offset = cbuf.offset,
+                    .count = 1,
+                });
             } else {
                 index = descriptors.Add(ImageDescriptor{
                     .type = flags.type,

src/shader_recompiler/shader_info.h

@@ -67,6 +67,15 @@ struct TextureBufferDescriptor {
 };
 using TextureBufferDescriptors = boost::container::small_vector<TextureBufferDescriptor, 6>;
 
+struct ImageBufferDescriptor {
+    ImageFormat format;
+    bool is_written;
+    u32 cbuf_index;
+    u32 cbuf_offset;
+    u32 count;
+};
+using ImageBufferDescriptors = boost::container::small_vector<ImageBufferDescriptor, 2>;
+
 struct TextureDescriptor {
     TextureType type;
     bool is_depth;
@@ -153,6 +162,7 @@ struct Info {
         constant_buffer_descriptors;
     boost::container::static_vector<StorageBufferDescriptor, MAX_SSBOS> storage_buffers_descriptors;
     TextureBufferDescriptors texture_buffer_descriptors;
+    ImageBufferDescriptors image_buffer_descriptors;
     TextureDescriptors texture_descriptors;
     ImageDescriptors image_descriptors;
 };

src/video_core/buffer_cache/buffer_cache.h

@@ -154,7 +154,7 @@ public:
     void UnbindGraphicsTextureBuffers(size_t stage);
 
     void BindGraphicsTextureBuffer(size_t stage, size_t tbo_index, GPUVAddr gpu_addr, u32 size,
-                                   PixelFormat format);
+                                   PixelFormat format, bool is_written);
 
     void UnbindComputeStorageBuffers();
 
@@ -163,8 +163,8 @@ public:
 
     void UnbindComputeTextureBuffers();
 
-    void BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_addr, u32 size,
-                                  PixelFormat format);
+    void BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_addr, u32 size, PixelFormat format,
+                                  bool is_written);
 
     void FlushCachedWrites();
 
@@ -393,7 +393,9 @@ private:
     u32 written_compute_storage_buffers = 0;
 
     std::array<u32, NUM_STAGES> enabled_texture_buffers{};
+    std::array<u32, NUM_STAGES> written_texture_buffers{};
     u32 enabled_compute_texture_buffers = 0;
+    u32 written_compute_texture_buffers = 0;
 
     std::array<u32, NUM_STAGES> fast_bound_uniform_buffers{};
 
@@ -700,12 +702,14 @@ void BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index,
 template <class P>
 void BufferCache<P>::UnbindGraphicsTextureBuffers(size_t stage) {
     enabled_texture_buffers[stage] = 0;
+    written_texture_buffers[stage] = 0;
 }
 
 template <class P>
 void BufferCache<P>::BindGraphicsTextureBuffer(size_t stage, size_t tbo_index, GPUVAddr gpu_addr,
-                                               u32 size, PixelFormat format) {
+                                               u32 size, PixelFormat format, bool is_written) {
     enabled_texture_buffers[stage] |= 1U << tbo_index;
+    written_texture_buffers[stage] |= (is_written ? 1U : 0U) << tbo_index;
     texture_buffers[stage][tbo_index] = GetTextureBufferBinding(gpu_addr, size, format);
 }
 
@@ -732,12 +736,14 @@ void BufferCache<P>::BindComputeStorageBuffer(size_t ssbo_index, u32 cbuf_index,
 template <class P>
 void BufferCache<P>::UnbindComputeTextureBuffers() {
     enabled_compute_texture_buffers = 0;
+    written_compute_texture_buffers = 0;
 }
 
 template <class P>
 void BufferCache<P>::BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_addr, u32 size,
-                                              PixelFormat format) {
+                                              PixelFormat format, bool is_written) {
     enabled_compute_texture_buffers |= 1U << tbo_index;
+    written_compute_texture_buffers |= (is_written ? 1U : 0U) << tbo_index;
     compute_texture_buffers[tbo_index] = GetTextureBufferBinding(gpu_addr, size, format);
 }
 
@@ -1274,6 +1280,10 @@ void BufferCache<P>::UpdateTextureBuffers(size_t stage) {
     ForEachEnabledBit(enabled_texture_buffers[stage], [&](u32 index) {
         Binding& binding = texture_buffers[stage][index];
         binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size);
+        // Mark buffer as written if needed
+        if (((written_texture_buffers[stage] >> index) & 1) != 0) {
+            MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, binding.size);
+        }
     });
 }
 
@@ -1343,6 +1353,10 @@ void BufferCache<P>::UpdateComputeTextureBuffers() {
     ForEachEnabledBit(enabled_compute_texture_buffers, [&](u32 index) {
         Binding& binding = compute_texture_buffers[index];
         binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size);
+        // Mark as written if needed
+        if (((written_compute_texture_buffers >> index) & 1) != 0) {
+            MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, binding.size);
+        }
     });
 }
 

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -88,6 +88,7 @@ public:
         Add(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, stage, info.constant_buffer_descriptors.size());
         Add(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, stage, info.storage_buffers_descriptors.size());
         Add(VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER, stage, info.texture_buffer_descriptors.size());
+        Add(VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER, stage, info.image_buffer_descriptors.size());
         Add(VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, stage, info.texture_descriptors.size());
         Add(VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, stage, info.image_descriptors.size());
     }
@@ -126,6 +127,7 @@ inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& sam
                                  const ImageId*& image_view_ids, TextureCache& texture_cache,
                                  VKUpdateDescriptorQueue& update_descriptor_queue) {
     image_view_ids += info.texture_buffer_descriptors.size();
+    image_view_ids += info.image_buffer_descriptors.size();
     for (const auto& desc : info.texture_descriptors) {
         const VkSampler sampler{*(samplers++)};
         ImageView& image_view{texture_cache.GetImageView(*(image_view_ids++))};

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -97,10 +97,12 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
         const u32 raw_handle{gpu_memory.Read<u32>(addr)};
         return TextureHandle(raw_handle, via_header_index);
     }};
-    for (const auto& desc : info.texture_buffer_descriptors) {
+    const auto add_image{[&](const auto& desc) {
         const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
         image_view_indices.push_back(handle.image);
-    }
+    }};
+    std::ranges::for_each(info.texture_buffer_descriptors, add_image);
+    std::ranges::for_each(info.image_buffer_descriptors, add_image);
     for (const auto& desc : info.texture_descriptors) {
         const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
         image_view_indices.push_back(handle.image);
@@ -108,24 +110,29 @@ 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);
-    }
+    std::ranges::for_each(info.image_descriptors, add_image);
+
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
     texture_cache.FillComputeImageViews(indices_span, image_view_ids);
 
     buffer_cache.UnbindComputeTextureBuffers();
     ImageId* texture_buffer_ids{image_view_ids.data()};
     size_t index{};
-    for (const auto& desc : info.texture_buffer_descriptors) {
+    const auto add_buffer{[&](const auto& desc) {
         ASSERT(desc.count == 1);
+        bool is_written{false};
+        if constexpr (std::is_same_v<decltype(desc), const Shader::ImageBufferDescriptor&>) {
+            is_written = desc.is_written;
+        }
         ImageView& image_view = texture_cache.GetImageView(*texture_buffer_ids);
         buffer_cache.BindComputeTextureBuffer(index, image_view.GpuAddr(), image_view.BufferSize(),
-                                              image_view.format);
+                                              image_view.format, is_written);
         ++texture_buffer_ids;
         ++index;
-    }
+    }};
+    std::ranges::for_each(info.texture_buffer_descriptors, add_buffer);
+    std::ranges::for_each(info.image_buffer_descriptors, add_buffer);
+
     buffer_cache.UpdateComputeBuffers();
     buffer_cache.BindHostComputeBuffers();
 

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -175,10 +175,12 @@ void GraphicsPipeline::Configure(bool is_indexed) {
             const u32 raw_handle{gpu_memory.Read<u32>(addr)};
             return TextureHandle(raw_handle, via_header_index);
         }};
-        for (const auto& desc : info.texture_buffer_descriptors) {
+        const auto add_image{[&](const auto& desc) {
             const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
             image_view_indices.push_back(handle.image);
-        }
+        }};
+        std::ranges::for_each(info.texture_buffer_descriptors, add_image);
+        std::ranges::for_each(info.image_buffer_descriptors, add_image);
         for (const auto& desc : info.texture_descriptors) {
             const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
             image_view_indices.push_back(handle.image);
@@ -186,28 +188,33 @@ 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);
-        }
+        std::ranges::for_each(info.image_descriptors, add_image);
     }
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
     texture_cache.FillGraphicsImageViews(indices_span, image_view_ids);
 
     ImageId* texture_buffer_index{image_view_ids.data()};
     for (size_t stage = 0; stage < Maxwell::MaxShaderStage; ++stage) {
-        const Shader::Info& info{stage_infos[stage]};
-        buffer_cache.UnbindGraphicsTextureBuffers(stage);
         size_t index{};
-        for (const auto& desc : info.texture_buffer_descriptors) {
+        const auto add_buffer{[&](const auto& desc) {
             ASSERT(desc.count == 1);
-            ImageView& image_view = texture_cache.GetImageView(*texture_buffer_index);
+            bool is_written{false};
+            if constexpr (std::is_same_v<decltype(desc), const Shader::ImageBufferDescriptor&>) {
+                is_written = desc.is_written;
+            }
+            ImageView& image_view{texture_cache.GetImageView(*texture_buffer_index)};
             buffer_cache.BindGraphicsTextureBuffer(stage, index, image_view.GpuAddr(),
-                                                   image_view.BufferSize(), image_view.format);
+                                                   image_view.BufferSize(), image_view.format,
+                                                   is_written);
             ++index;
             ++texture_buffer_index;
-        }
+        }};
+        const Shader::Info& info{stage_infos[stage]};
+        buffer_cache.UnbindGraphicsTextureBuffers(stage);
+        std::ranges::for_each(info.texture_buffer_descriptors, add_buffer);
+        std::ranges::for_each(info.image_buffer_descriptors, add_buffer);
         texture_buffer_index += info.texture_descriptors.size();
+        texture_buffer_index += info.image_descriptors.size();
     }
     buffer_cache.UpdateGraphicsBuffers(is_indexed);