early-access version 3552

src/video_core/texture_cache/image_info.h

@@ -39,6 +39,8 @@ struct ImageInfo {
     u32 tile_width_spacing = 0;
     bool rescaleable = false;
     bool downscaleable = false;
+    bool forced_flushed = false;
+    bool dma_downloaded = false;
 };
 
 } // namespace VideoCommon

src/video_core/texture_cache/image_view_base.cpp

@@ -26,8 +26,9 @@ ImageViewBase::ImageViewBase(const ImageViewInfo& info, const ImageInfo& image_i
     ASSERT_MSG(VideoCore::Surface::IsViewCompatible(image_info.format, info.format, false, true),
                "Image view format {} is incompatible with image format {}", info.format,
                image_info.format);
-    const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
-    if (image_info.type == ImageType::Linear && is_async) {
+    const bool preemptive =
+        !Settings::values.use_reactive_flushing.GetValue() && image_info.type == ImageType::Linear;
+    if (image_info.forced_flushed || preemptive) {
         flags |= ImageViewFlagBits::PreemtiveDownload;
     }
     if (image_info.type == ImageType::e3D && info.type != ImageViewType::e3D) {

src/video_core/texture_cache/texture_cache.h

@@ -490,6 +490,32 @@ void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
     }
 }
 
+template <class P>
+std::optional<VideoCore::RasterizerDownloadArea> TextureCache<P>::GetFlushArea(VAddr cpu_addr,
+                                                                               u64 size) {
+    std::optional<VideoCore::RasterizerDownloadArea> area{};
+    ForEachImageInRegion(cpu_addr, size, [&](ImageId, ImageBase& image) {
+        if (False(image.flags & ImageFlagBits::GpuModified)) {
+            return;
+        }
+        if (!area) {
+            area.emplace();
+            area->start_address = cpu_addr;
+            area->end_address = cpu_addr + size;
+            area->preemtive = true;
+        }
+        area->start_address = std::min(area->start_address, image.cpu_addr);
+        area->end_address = std::max(area->end_address, image.cpu_addr_end);
+        for (auto image_view_id : image.image_view_ids) {
+            auto& image_view = slot_image_views[image_view_id];
+            image_view.flags |= ImageViewFlagBits::PreemtiveDownload;
+        }
+        area->preemtive &= image.info.forced_flushed;
+        image.info.forced_flushed = true;
+    });
+    return area;
+}
+
 template <class P>
 void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
     std::vector<ImageId> deleted_images;
@@ -683,18 +709,43 @@ void TextureCache<P>::CommitAsyncFlushes() {
                 download_info.async_buffer_id = last_async_buffer_id;
             }
         }
+
         if (any_none_dma) {
+            bool all_pre_sync = true;
             auto download_map = runtime.DownloadStagingBuffer(total_size_bytes, true);
             for (const PendingDownload& download_info : download_ids) {
                 if (download_info.is_swizzle) {
                     Image& image = slot_images[download_info.object_id];
+                    all_pre_sync &= image.info.dma_downloaded;
+                    image.info.dma_downloaded = true;
                     const auto copies = FullDownloadCopies(image.info);
                     image.DownloadMemory(download_map, copies);
                     download_map.offset += Common::AlignUp(image.unswizzled_size_bytes, 64);
                 }
             }
-            uncommitted_async_buffers.emplace_back(download_map);
+            if (!all_pre_sync) {
+                runtime.Finish();
+                auto it = download_ids.begin();
+                while (it != download_ids.end()) {
+                    const PendingDownload& download_info = *it;
+                    if (download_info.is_swizzle) {
+                        const ImageBase& image = slot_images[download_info.object_id];
+                        const auto copies = FullDownloadCopies(image.info);
+                        download_map.offset -= Common::AlignUp(image.unswizzled_size_bytes, 64);
+                        std::span<u8> download_span =
+                            download_map.mapped_span.subspan(download_map.offset);
+                        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span,
+                                     swizzle_data_buffer);
+                        it = download_ids.erase(it);
+                    } else {
+                        it++;
+                    }
+                }
+            } else {
+                uncommitted_async_buffers.emplace_back(download_map);
+            }
         }
+
         async_buffers.emplace_back(std::move(uncommitted_async_buffers));
         uncommitted_async_buffers.clear();
     }
@@ -789,11 +840,16 @@ ImageId TextureCache<P>::DmaImageId(const Tegra::DMA::ImageOperand& operand) {
     if (!dst_id) {
         return NULL_IMAGE_ID;
     }
-    const auto& image = slot_images[dst_id];
+    auto& image = slot_images[dst_id];
     if (False(image.flags & ImageFlagBits::GpuModified)) {
         // No need to waste time on an image that's synced with guest
         return NULL_IMAGE_ID;
     }
+    if (!image.info.dma_downloaded) {
+        // Force a full sync.
+        image.info.dma_downloaded = true;
+        return NULL_IMAGE_ID;
+    }
     const auto base = image.TryFindBase(operand.address);
     if (!base) {
         return NULL_IMAGE_ID;

src/video_core/texture_cache/texture_cache_base.h

@@ -179,6 +179,8 @@ public:
     /// Download contents of host images to guest memory in a region
     void DownloadMemory(VAddr cpu_addr, size_t size);
 
+    std::optional<VideoCore::RasterizerDownloadArea> GetFlushArea(VAddr cpu_addr, u64 size);
+
     /// Remove images in a region
     void UnmapMemory(VAddr cpu_addr, size_t size);