texture_cache: Async download of GPU modified linear images

src/video_core/amdgpu/liverpool.cpp

@@ -124,6 +124,11 @@ void Liverpool::Process(std::stop_token stoken) {
             if (task.done()) {
                 task.destroy();
 
+                if (rasterizer) {
+                    rasterizer->EndCommandList();
+                    rasterizer->Flush();
+                }
+
                 std::scoped_lock lock{queue.m_access};
                 queue.submits.pop();
 
@@ -135,11 +140,6 @@ void Liverpool::Process(std::stop_token stoken) {
 
         if (submit_done) {
             VideoCore::EndCapture();
-
-            if (rasterizer) {
-                rasterizer->ProcessFaults();
-                rasterizer->Flush();
-            }
             submit_done = false;
         }
 

src/video_core/buffer_cache/buffer_cache.h

@@ -112,7 +112,7 @@ public:
     /// Invalidates any buffer in the logical page range.
     void InvalidateMemory(VAddr device_addr, u64 size);
 
-    /// Waits on pending downloads in the logical page range.
+    /// Flushes any GPU modified buffer in the logical page range back to CPU memory.
     void ReadMemory(VAddr device_addr, u64 size, bool is_write = false);
 
     /// Binds host vertex buffers for the current draw.

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -272,6 +272,8 @@ void Rasterizer::EliminateFastClear() {
 void Rasterizer::Draw(bool is_indexed, u32 index_offset) {
     RENDERER_TRACE;
 
+    scheduler.PopPendingOperations();
+
     if (!FilterDraw()) {
         return;
     }
@@ -317,6 +319,8 @@ void Rasterizer::DrawIndirect(bool is_indexed, VAddr arg_address, u32 offset, u3
                               u32 max_count, VAddr count_address) {
     RENDERER_TRACE;
 
+    scheduler.PopPendingOperations();
+
     if (!FilterDraw()) {
         return;
     }
@@ -380,6 +384,8 @@ void Rasterizer::DrawIndirect(bool is_indexed, VAddr arg_address, u32 offset, u3
 void Rasterizer::DispatchDirect() {
     RENDERER_TRACE;
 
+    scheduler.PopPendingOperations();
+
     const auto& cs_program = liverpool->GetCsRegs();
     const ComputePipeline* pipeline = pipeline_cache.GetComputePipeline();
     if (!pipeline) {
@@ -407,6 +413,8 @@ void Rasterizer::DispatchDirect() {
 void Rasterizer::DispatchIndirect(VAddr address, u32 offset, u32 size) {
     RENDERER_TRACE;
 
+    scheduler.PopPendingOperations();
+
     const auto& cs_program = liverpool->GetCsRegs();
     const ComputePipeline* pipeline = pipeline_cache.GetComputePipeline();
     if (!pipeline) {
@@ -439,11 +447,12 @@ void Rasterizer::Finish() {
     scheduler.Finish();
 }
 
-void Rasterizer::ProcessFaults() {
+void Rasterizer::EndCommandList() {
     if (fault_process_pending) {
         fault_process_pending = false;
         buffer_cache.ProcessFaultBuffer();
     }
+    texture_cache.ProcessDownloadImages();
 }
 
 bool Rasterizer::BindResources(const Pipeline* pipeline) {
@@ -649,8 +658,7 @@ void Rasterizer::BindTextures(const Shader::Info& stage, Shader::Backend::Bindin
             if (instance.IsNullDescriptorSupported()) {
                 image_infos.emplace_back(VK_NULL_HANDLE, VK_NULL_HANDLE, vk::ImageLayout::eGeneral);
             } else {
-                auto& null_image_view =
-                    texture_cache.FindTexture(VideoCore::NULL_IMAGE_ID, desc.view_info);
+                auto& null_image_view = texture_cache.FindTexture(VideoCore::NULL_IMAGE_ID, desc);
                 image_infos.emplace_back(VK_NULL_HANDLE, *null_image_view.image_view,
                                          vk::ImageLayout::eGeneral);
             }
@@ -664,7 +672,7 @@ void Rasterizer::BindTextures(const Shader::Info& stage, Shader::Backend::Bindin
             bound_images.emplace_back(image_id);
 
             auto& image = texture_cache.GetImage(image_id);
-            auto& image_view = texture_cache.FindTexture(image_id, desc.view_info);
+            auto& image_view = texture_cache.FindTexture(image_id, desc);
 
             if (image.binding.force_general || image.binding.is_target) {
                 image.Transit(vk::ImageLayout::eGeneral,

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -68,7 +68,7 @@ public:
     void CpSync();
     u64 Flush();
     void Finish();
-    void ProcessFaults();
+    void EndCommandList();
 
     PipelineCache& GetPipelineCache() {
         return pipeline_cache;

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -101,6 +101,14 @@ void Scheduler::Wait(u64 tick) {
     }
 }
 
+void Scheduler::PopPendingOperations() {
+    master_semaphore.Refresh();
+    while (!pending_ops.empty() && master_semaphore.IsFree(pending_ops.front().gpu_tick)) {
+        pending_ops.front().callback();
+        pending_ops.pop();
+    }
+}
+
 void Scheduler::AllocateWorkerCommandBuffers() {
     const vk::CommandBufferBeginInfo begin_info = {
         .flags = vk::CommandBufferUsageFlagBits::eOneTimeSubmit,
@@ -175,10 +183,7 @@ void Scheduler::SubmitExecution(SubmitInfo& info) {
     AllocateWorkerCommandBuffers();
 
     // Apply pending operations
-    while (!pending_ops.empty() && IsFree(pending_ops.front().gpu_tick)) {
-        pending_ops.front().callback();
-        pending_ops.pop();
-    }
+    PopPendingOperations();
 }
 
 void DynamicState::Commit(const Instance& instance, const vk::CommandBuffer& cmdbuf) {

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -317,6 +317,9 @@ public:
     /// Waits for the given tick to trigger on the GPU.
     void Wait(u64 tick);
 
+    /// Attempts to execute operations whose tick the GPU has caught up with.
+    void PopPendingOperations();
+
     /// Starts a new rendering scope with provided state.
     void BeginRendering(const RenderState& new_state);
 

src/video_core/texture_cache/texture_cache.cpp

@@ -6,6 +6,7 @@
 
 #include "common/assert.h"
 #include "common/debug.h"
+#include "core/memory.h"
 #include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/page_manager.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
@@ -58,6 +59,50 @@ ImageId TextureCache::GetNullImage(const vk::Format format) {
     return null_id;
 }
 
+void TextureCache::ProcessDownloadImages() {
+    for (const ImageId image_id : download_images) {
+        DownloadImageMemory(image_id);
+    }
+    download_images.clear();
+}
+
+void TextureCache::DownloadImageMemory(ImageId image_id) {
+    Image& image = slot_images[image_id];
+    if (False(image.flags & ImageFlagBits::GpuModified)) {
+        return;
+    }
+    auto& download_buffer = buffer_cache.GetUtilityBuffer(MemoryUsage::Download);
+    const u32 download_size = image.info.size.width * image.info.size.height *
+                              image.info.resources.layers * (image.info.num_bits / 8);
+    // ASSERT(download_size == image.info.guest_size);
+    const auto [download, offset] = download_buffer.Map(download_size);
+    download_buffer.Commit();
+    const vk::BufferImageCopy image_download = {
+        .bufferOffset = offset,
+        .bufferRowLength = 0,
+        .bufferImageHeight = 0,
+        .imageSubresource =
+            {
+                .aspectMask = image.info.IsDepthStencil() ? vk::ImageAspectFlagBits::eDepth
+                                                          : vk::ImageAspectFlagBits::eColor,
+                .mipLevel = 0,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        .imageOffset = {0, 0, 0},
+        .imageExtent = {image.info.size.width, image.info.size.height, 1},
+    };
+    scheduler.EndRendering();
+    const auto cmdbuf = scheduler.CommandBuffer();
+    image.Transit(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead, {});
+    cmdbuf.copyImageToBuffer(image.image, vk::ImageLayout::eTransferSrcOptimal,
+                             download_buffer.Handle(), image_download);
+    scheduler.DeferOperation([device_addr = image.info.guest_address, download, download_size] {
+        auto* memory = Core::Memory::Instance();
+        memory->TryWriteBacking(std::bit_cast<u8*>(device_addr), download, download_size);
+    });
+}
+
 void TextureCache::MarkAsMaybeDirty(ImageId image_id, Image& image) {
     if (image.hash == 0) {
         // Initialize hash
@@ -437,16 +482,25 @@ ImageView& TextureCache::RegisterImageView(ImageId image_id, const ImageViewInfo
     return slot_image_views[view_id];
 }
 
-ImageView& TextureCache::FindTexture(ImageId image_id, const ImageViewInfo& view_info) {
+ImageView& TextureCache::FindTexture(ImageId image_id, const BaseDesc& desc) {
     Image& image = slot_images[image_id];
+    if (desc.type == BindingType::Storage) {
+        image.flags |= ImageFlagBits::GpuModified;
+        if (image.info.tiling_mode == AmdGpu::TilingMode::Display_Linear) {
+            download_images.emplace(image_id);
+        }
+    }
     UpdateImage(image_id);
-    return RegisterImageView(image_id, view_info);
+    return RegisterImageView(image_id, desc.view_info);
 }
 
 ImageView& TextureCache::FindRenderTarget(BaseDesc& desc) {
     const ImageId image_id = FindImage(desc);
     Image& image = slot_images[image_id];
     image.flags |= ImageFlagBits::GpuModified;
+    if (image.info.tiling_mode == AmdGpu::TilingMode::Display_Linear) {
+        download_images.emplace(image_id);
+    }
     image.usage.render_target = 1u;
     UpdateImage(image_id);
 
@@ -474,6 +528,9 @@ ImageView& TextureCache::FindDepthTarget(BaseDesc& desc) {
     const ImageId image_id = FindImage(desc);
     Image& image = slot_images[image_id];
     image.flags |= ImageFlagBits::GpuModified;
+    if (image.info.tiling_mode == AmdGpu::TilingMode::Display_Linear) {
+        download_images.emplace(image_id);
+    }
     image.usage.depth_target = 1u;
     image.usage.stencil = image.info.HasStencil();
     UpdateImage(image_id);

src/video_core/texture_cache/texture_cache.h

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include <unordered_set>
 #include <boost/container/small_vector.hpp>
 #include <tsl/robin_map.h>
 
@@ -105,11 +106,14 @@ public:
     /// Evicts any images that overlap the unmapped range.
     void UnmapMemory(VAddr cpu_addr, size_t size);
 
+    /// Schedules a copy of pending images for download back to CPU memory.
+    void ProcessDownloadImages();
+
     /// Retrieves the image handle of the image with the provided attributes.
     [[nodiscard]] ImageId FindImage(BaseDesc& desc, FindFlags flags = {});
 
     /// Retrieves an image view with the properties of the specified image id.
-    [[nodiscard]] ImageView& FindTexture(ImageId image_id, const ImageViewInfo& view_info);
+    [[nodiscard]] ImageView& FindTexture(ImageId image_id, const BaseDesc& desc);
 
     /// Retrieves the render target with specified properties
     [[nodiscard]] ImageView& FindRenderTarget(BaseDesc& desc);
@@ -252,6 +256,9 @@ private:
     /// Gets or creates a null image for a particular format.
     ImageId GetNullImage(vk::Format format);
 
+    /// Copies image memory back to CPU.
+    void DownloadImageMemory(ImageId image_id);
+
     /// Create an image from the given parameters
     [[nodiscard]] ImageId InsertImage(const ImageInfo& info, VAddr cpu_addr);
 
@@ -293,6 +300,7 @@ private:
     Common::SlotVector<ImageView> slot_image_views;
     tsl::robin_map<u64, Sampler> samplers;
     tsl::robin_map<vk::Format, ImageId> null_images;
+    std::unordered_set<ImageId> download_images;
     PageTable page_table;
     std::mutex mutex;