9 files changed, 212 insertions, 81 deletions

diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 0932fadc2..2f986097f 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -223,6 +223,9 @@ void Maxwell3D::ProcessMacro(u32 method, const u32* base_start, u32 amount, bool
 }
 
 void Maxwell3D::RefreshParametersImpl() {
+    if (!Settings::IsGPULevelHigh()) {
+        return;
+    }
     size_t current_index = 0;
     for (auto& segment : macro_segments) {
         if (segment.first == 0) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 90e35e307..4993d4709 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -1287,8 +1287,7 @@ bool AccelerateDMA::DmaBufferImageCopy(const Tegra::DMA::ImageCopy& copy_info,
     }
     const u32 buffer_size = static_cast<u32>(buffer_operand.pitch * buffer_operand.height);
     static constexpr auto sync_info = VideoCommon::ObtainBufferSynchronize::FullSynchronize;
-    const auto post_op = IS_IMAGE_UPLOAD ? VideoCommon::ObtainBufferOperation::DoNothing
-                                         : VideoCommon::ObtainBufferOperation::MarkAsWritten;
+    const auto post_op = VideoCommon::ObtainBufferOperation::DoNothing;
     const auto [buffer, offset] =
         buffer_cache.ObtainBuffer(buffer_operand.address, buffer_size, sync_info, post_op);
 
@@ -1299,7 +1298,8 @@ bool AccelerateDMA::DmaBufferImageCopy(const Tegra::DMA::ImageCopy& copy_info,
     if constexpr (IS_IMAGE_UPLOAD) {
         image->UploadMemory(buffer->Handle(), offset, copy_span);
     } else {
-        image->DownloadMemory(buffer->Handle(), offset, copy_span);
+        texture_cache.DownloadImageIntoBuffer(image, buffer->Handle(), offset, copy_span,
+                                              buffer_operand.address, buffer_size);
     }
     return true;
 }
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 0b9c4a904..032a8ebc5 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -803,30 +803,40 @@ void Image::UploadMemory(const ImageBufferMap& map,
 
 void Image::DownloadMemory(GLuint buffer_handle, size_t buffer_offset,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
+    std::array buffer_handles{buffer_handle};
+    std::array buffer_offsets{buffer_offset};
+    DownloadMemory(buffer_handles, buffer_offsets, copies);
+}
+
+void Image::DownloadMemory(std::span<GLuint> buffer_handles, std::span<size_t> buffer_offsets,
+                           std::span<const VideoCommon::BufferImageCopy> copies) {
     const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
     if (is_rescaled) {
         ScaleDown();
     }
     glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT); // TODO: Move this to its own API
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, buffer_handle);
-    glPixelStorei(GL_PACK_ALIGNMENT, 1);
+    for (size_t i = 0; i < buffer_handles.size(); i++) {
+        auto& buffer_handle = buffer_handles[i];
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, buffer_handle);
+        glPixelStorei(GL_PACK_ALIGNMENT, 1);
 
-    u32 current_row_length = std::numeric_limits<u32>::max();
-    u32 current_image_height = std::numeric_limits<u32>::max();
+        u32 current_row_length = std::numeric_limits<u32>::max();
+        u32 current_image_height = std::numeric_limits<u32>::max();
 
-    for (const VideoCommon::BufferImageCopy& copy : copies) {
-        if (copy.image_subresource.base_level >= gl_num_levels) {
-            continue;
-        }
-        if (current_row_length != copy.buffer_row_length) {
-            current_row_length = copy.buffer_row_length;
-            glPixelStorei(GL_PACK_ROW_LENGTH, current_row_length);
-        }
-        if (current_image_height != copy.buffer_image_height) {
-            current_image_height = copy.buffer_image_height;
-            glPixelStorei(GL_PACK_IMAGE_HEIGHT, current_image_height);
+        for (const VideoCommon::BufferImageCopy& copy : copies) {
+            if (copy.image_subresource.base_level >= gl_num_levels) {
+                continue;
+            }
+            if (current_row_length != copy.buffer_row_length) {
+                current_row_length = copy.buffer_row_length;
+                glPixelStorei(GL_PACK_ROW_LENGTH, current_row_length);
+            }
+            if (current_image_height != copy.buffer_image_height) {
+                current_image_height = copy.buffer_image_height;
+                glPixelStorei(GL_PACK_IMAGE_HEIGHT, current_image_height);
+            }
+            CopyImageToBuffer(copy, buffer_offsets[i]);
         }
-        CopyImageToBuffer(copy, buffer_offset);
     }
     if (is_rescaled) {
         ScaleUp(true);
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.h b/src/video_core/renderer_opengl/gl_texture_cache.h
index 911e4607a..0dd039ed2 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.h
+++ b/src/video_core/renderer_opengl/gl_texture_cache.h
@@ -215,6 +215,9 @@ public:
     void DownloadMemory(GLuint buffer_handle, size_t buffer_offset,
                         std::span<const VideoCommon::BufferImageCopy> copies);
 
+    void DownloadMemory(std::span<GLuint> buffer_handle, std::span<size_t> buffer_offset,
+                        std::span<const VideoCommon::BufferImageCopy> copies);
+
     void DownloadMemory(ImageBufferMap& map, std::span<const VideoCommon::BufferImageCopy> copies);
 
     GLuint StorageHandle() noexcept;
@@ -376,6 +379,7 @@ struct TextureCacheParams {
     using Sampler = OpenGL::Sampler;
     using Framebuffer = OpenGL::Framebuffer;
     using AsyncBuffer = u32;
+    using BufferType = GLuint;
 };
 
 using TextureCache = VideoCommon::TextureCache<TextureCacheParams>;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 673ab478e..2559a3aa7 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -781,8 +781,7 @@ bool AccelerateDMA::DmaBufferImageCopy(const Tegra::DMA::ImageCopy& copy_info,
     }
     const u32 buffer_size = static_cast<u32>(buffer_operand.pitch * buffer_operand.height);
     static constexpr auto sync_info = VideoCommon::ObtainBufferSynchronize::FullSynchronize;
-    const auto post_op = IS_IMAGE_UPLOAD ? VideoCommon::ObtainBufferOperation::DoNothing
-                                         : VideoCommon::ObtainBufferOperation::MarkAsWritten;
+    const auto post_op = VideoCommon::ObtainBufferOperation::DoNothing;
     const auto [buffer, offset] =
         buffer_cache.ObtainBuffer(buffer_operand.address, buffer_size, sync_info, post_op);
 
@@ -793,7 +792,8 @@ bool AccelerateDMA::DmaBufferImageCopy(const Tegra::DMA::ImageCopy& copy_info,
     if constexpr (IS_IMAGE_UPLOAD) {
         image->UploadMemory(buffer->Handle(), offset, copy_span);
     } else {
-        image->DownloadMemory(buffer->Handle(), offset, copy_span);
+        texture_cache.DownloadImageIntoBuffer(image, buffer->Handle(), offset, copy_span,
+                                              buffer_operand.address, buffer_size);
     }
     return true;
 }
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index ae15f6976..d0a7d8f35 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -1,10 +1,11 @@
 // SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
+// SPDX-License-Identifier: GPL-3.0-or-later
 
 #include <algorithm>
 #include <array>
 #include <span>
 #include <vector>
+#include <boost/container/small_vector.hpp>
 
 #include "common/bit_cast.h"
 #include "common/bit_util.h"
@@ -1343,14 +1344,31 @@ void Image::UploadMemory(const StagingBufferRef& map, std::span<const BufferImag
 
 void Image::DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
+    std::array buffer_handles{
+        buffer,
+    };
+    std::array buffer_offsets{
+        offset,
+    };
+    DownloadMemory(buffer_handles, buffer_offsets, copies);
+}
+
+void Image::DownloadMemory(std::span<VkBuffer> buffers_span, std::span<VkDeviceSize> offsets_span,
+                           std::span<const VideoCommon::BufferImageCopy> copies) {
     const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
     if (is_rescaled) {
         ScaleDown();
     }
-    std::vector vk_copies = TransformBufferImageCopies(copies, offset, aspect_mask);
+    boost::container::small_vector<VkBuffer, 1> buffers_vector{};
+    boost::container::small_vector<std::vector<VkBufferImageCopy>, 1> vk_copies;
+    for (size_t index = 0; index < buffers_span.size(); index++) {
+        buffers_vector.emplace_back(buffers_span[index]);
+        vk_copies.emplace_back(
+            TransformBufferImageCopies(copies, offsets_span[index], aspect_mask));
+    }
     scheduler->RequestOutsideRenderPassOperationContext();
-    scheduler->Record([buffer, image = *original_image, aspect_mask = aspect_mask,
-                       vk_copies](vk::CommandBuffer cmdbuf) {
+    scheduler->Record([buffers = std::move(buffers_vector), image = *original_image,
+                       aspect_mask = aspect_mask, vk_copies](vk::CommandBuffer cmdbuf) {
         const VkImageMemoryBarrier read_barrier{
             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
             .pNext = nullptr,
@@ -1369,6 +1387,20 @@ void Image::DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
                 .layerCount = VK_REMAINING_ARRAY_LAYERS,
             },
         };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, read_barrier);
+
+        for (size_t index = 0; index < buffers.size(); index++) {
+            cmdbuf.CopyImageToBuffer(image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, buffers[index],
+                                     vk_copies[index]);
+        }
+
+        const VkMemoryBarrier memory_write_barrier{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+        };
         const VkImageMemoryBarrier image_write_barrier{
             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
             .pNext = nullptr,
@@ -1387,15 +1419,6 @@ void Image::DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
                 .layerCount = VK_REMAINING_ARRAY_LAYERS,
             },
         };
-        const VkMemoryBarrier memory_write_barrier{
-            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
-            .pNext = nullptr,
-            .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
-            .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
-        };
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                               0, read_barrier);
-        cmdbuf.CopyImageToBuffer(image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, buffer, vk_copies);
         cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
                                0, memory_write_barrier, nullptr, image_write_barrier);
     });
@@ -1405,7 +1428,13 @@ void Image::DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
 }
 
 void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferImageCopy> copies) {
-    DownloadMemory(map.buffer, map.offset, copies);
+    std::array buffers{
+        map.buffer,
+    };
+    std::array offsets{
+        map.offset,
+    };
+    DownloadMemory(buffers, offsets, copies);
 }
 
 bool Image::IsRescaled() const noexcept {
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h
index d5ee23f8d..c656c5386 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.h
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.h
@@ -1,5 +1,5 @@
 // SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
+// SPDX-License-Identifier: GPL-3.0-or-later
 
 #pragma once
 
@@ -141,6 +141,9 @@ public:
     void DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
                         std::span<const VideoCommon::BufferImageCopy> copies);
 
+    void DownloadMemory(std::span<VkBuffer> buffers, std::span<VkDeviceSize> offsets,
+                        std::span<const VideoCommon::BufferImageCopy> copies);
+
     void DownloadMemory(const StagingBufferRef& map,
                         std::span<const VideoCommon::BufferImageCopy> copies);
 
@@ -371,6 +374,7 @@ struct TextureCacheParams {
     using Sampler = Vulkan::Sampler;
     using Framebuffer = Vulkan::Framebuffer;
     using AsyncBuffer = Vulkan::StagingBufferRef;
+    using BufferType = VkBuffer;
 };
 
 using TextureCache = VideoCommon::TextureCache<TextureCacheParams>;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index ed5c768d8..e601f8446 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -1,9 +1,10 @@
-// SPDX-FileCopyrightText: 2021 yuzu Emulator Project
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 #pragma once
 
 #include <unordered_set>
+#include <boost/container/small_vector.hpp>
 
 #include "common/alignment.h"
 #include "common/settings.h"
@@ -17,15 +18,10 @@
 
 namespace VideoCommon {
 
-using Tegra::Texture::SwizzleSource;
-using Tegra::Texture::TextureType;
 using Tegra::Texture::TICEntry;
 using Tegra::Texture::TSCEntry;
 using VideoCore::Surface::GetFormatType;
-using VideoCore::Surface::IsCopyCompatible;
 using VideoCore::Surface::PixelFormat;
-using VideoCore::Surface::PixelFormatFromDepthFormat;
-using VideoCore::Surface::PixelFormatFromRenderTargetFormat;
 using VideoCore::Surface::SurfaceType;
 using namespace Common::Literals;
 
@@ -143,6 +139,13 @@ void TextureCache<P>::TickFrame() {
     runtime.TickFrame();
     critical_gc = 0;
     ++frame_tick;
+
+    if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
+        for (auto& buffer : async_buffers_death_ring) {
+            runtime.FreeDeferredStagingBuffer(buffer);
+        }
+        async_buffers_death_ring.clear();
+    }
 }
 
 template <class P>
@@ -661,25 +664,39 @@ template <class P>
 void TextureCache<P>::CommitAsyncFlushes() {
     // This is intentionally passing the value by copy
     if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
-        const std::span<const ImageId> download_ids = uncommitted_downloads;
+        auto& download_ids = uncommitted_downloads;
         if (download_ids.empty()) {
             committed_downloads.emplace_back(std::move(uncommitted_downloads));
             uncommitted_downloads.clear();
-            async_buffers.emplace_back(std::optional<AsyncBuffer>{});
+            async_buffers.emplace_back(std::move(uncommitted_async_buffers));
+            uncommitted_async_buffers.clear();
             return;
         }
         size_t total_size_bytes = 0;
-        for (const ImageId image_id : download_ids) {
-            total_size_bytes += slot_images[image_id].unswizzled_size_bytes;
+        size_t last_async_buffer_id = uncommitted_async_buffers.size();
+        bool any_none_dma = false;
+        for (PendingDownload& download_info : download_ids) {
+            if (download_info.is_swizzle) {
+                total_size_bytes +=
+                    Common::AlignUp(slot_images[download_info.object_id].unswizzled_size_bytes, 64);
+                any_none_dma = true;
+                download_info.async_buffer_id = last_async_buffer_id;
+            }
         }
-        auto download_map = runtime.DownloadStagingBuffer(total_size_bytes, true);
-        for (const ImageId image_id : download_ids) {
-            Image& image = slot_images[image_id];
-            const auto copies = FullDownloadCopies(image.info);
-            image.DownloadMemory(download_map, copies);
-            download_map.offset += Common::AlignUp(image.unswizzled_size_bytes, 64);
+        if (any_none_dma) {
+            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];
+                    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);
         }
-        async_buffers.emplace_back(download_map);
+        async_buffers.emplace_back(std::move(uncommitted_async_buffers));
+        uncommitted_async_buffers.clear();
     }
     committed_downloads.emplace_back(std::move(uncommitted_downloads));
     uncommitted_downloads.clear();
@@ -691,39 +708,57 @@ void TextureCache<P>::PopAsyncFlushes() {
         return;
     }
     if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
-        const std::span<const ImageId> download_ids = committed_downloads.front();
+        const auto& download_ids = committed_downloads.front();
         if (download_ids.empty()) {
             committed_downloads.pop_front();
             async_buffers.pop_front();
             return;
         }
-        auto download_map = *async_buffers.front();
-        std::span<u8> download_span = download_map.mapped_span;
+        auto download_map = std::move(async_buffers.front());
         for (size_t i = download_ids.size(); i > 0; i--) {
-            const ImageBase& image = slot_images[download_ids[i - 1]];
-            const auto copies = FullDownloadCopies(image.info);
-            download_map.offset -= Common::AlignUp(image.unswizzled_size_bytes, 64);
-            std::span<u8> download_span_alt = download_span.subspan(download_map.offset);
-            SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span_alt,
-                         swizzle_data_buffer);
+            auto& download_info = download_ids[i - 1];
+            auto& download_buffer = download_map[download_info.async_buffer_id];
+            if (download_info.is_swizzle) {
+                const ImageBase& image = slot_images[download_info.object_id];
+                const auto copies = FullDownloadCopies(image.info);
+                download_buffer.offset -= Common::AlignUp(image.unswizzled_size_bytes, 64);
+                std::span<u8> download_span =
+                    download_buffer.mapped_span.subspan(download_buffer.offset);
+                SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span,
+                             swizzle_data_buffer);
+            } else {
+                const BufferDownload& buffer_info = slot_buffer_downloads[download_info.object_id];
+                std::span<u8> download_span =
+                    download_buffer.mapped_span.subspan(download_buffer.offset);
+                gpu_memory->WriteBlockUnsafe(buffer_info.address, download_span.data(),
+                                             buffer_info.size);
+                slot_buffer_downloads.erase(download_info.object_id);
+            }
+        }
+        for (auto& download_buffer : download_map) {
+            async_buffers_death_ring.emplace_back(download_buffer);
         }
-        runtime.FreeDeferredStagingBuffer(download_map);
         committed_downloads.pop_front();
         async_buffers.pop_front();
     } else {
-        const std::span<const ImageId> download_ids = committed_downloads.front();
+        const auto& download_ids = committed_downloads.front();
         if (download_ids.empty()) {
             committed_downloads.pop_front();
             return;
         }
         size_t total_size_bytes = 0;
-        for (const ImageId image_id : download_ids) {
-            total_size_bytes += slot_images[image_id].unswizzled_size_bytes;
+        for (const PendingDownload& download_info : download_ids) {
+            if (download_info.is_swizzle) {
+                total_size_bytes += slot_images[download_info.object_id].unswizzled_size_bytes;
+            }
         }
         auto download_map = runtime.DownloadStagingBuffer(total_size_bytes);
         const size_t original_offset = download_map.offset;
-        for (const ImageId image_id : download_ids) {
-            Image& image = slot_images[image_id];
+        for (const PendingDownload& download_info : download_ids) {
+            if (!download_info.is_swizzle) {
+                continue;
+            }
+            Image& image = slot_images[download_info.object_id];
             const auto copies = FullDownloadCopies(image.info);
             image.DownloadMemory(download_map, copies);
             download_map.offset += image.unswizzled_size_bytes;
@@ -732,8 +767,11 @@ void TextureCache<P>::PopAsyncFlushes() {
         runtime.Finish();
         download_map.offset = original_offset;
         std::span<u8> download_span = download_map.mapped_span;
-        for (const ImageId image_id : download_ids) {
-            const ImageBase& image = slot_images[image_id];
+        for (const PendingDownload& download_info : download_ids) {
+            if (!download_info.is_swizzle) {
+                continue;
+            }
+            const ImageBase& image = slot_images[download_info.object_id];
             const auto copies = FullDownloadCopies(image.info);
             SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span,
                          swizzle_data_buffer);
@@ -834,6 +872,33 @@ std::pair<typename TextureCache<P>::Image*, BufferImageCopy> TextureCache<P>::Dm
 }
 
 template <class P>
+void TextureCache<P>::DownloadImageIntoBuffer(typename TextureCache<P>::Image* image,
+                                              typename TextureCache<P>::BufferType buffer,
+                                              size_t buffer_offset,
+                                              std::span<const VideoCommon::BufferImageCopy> copies,
+                                              GPUVAddr address, size_t size) {
+    if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
+        const BufferDownload new_buffer_download{address, size};
+        auto slot = slot_buffer_downloads.insert(new_buffer_download);
+        const PendingDownload new_download{false, uncommitted_async_buffers.size(), slot};
+        uncommitted_downloads.emplace_back(new_download);
+        auto download_map = runtime.DownloadStagingBuffer(size, true);
+        uncommitted_async_buffers.emplace_back(download_map);
+        std::array buffers{
+            buffer,
+            download_map.buffer,
+        };
+        std::array buffer_offsets{
+            buffer_offset,
+            download_map.offset,
+        };
+        image->DownloadMemory(buffers, buffer_offsets, copies);
+    } else {
+        image->DownloadMemory(buffer, buffer_offset, copies);
+    }
+}
+
+template <class P>
 void TextureCache<P>::RefreshContents(Image& image, ImageId image_id) {
     if (False(image.flags & ImageFlagBits::CpuModified)) {
         // Only upload modified images
@@ -2209,7 +2274,8 @@ void TextureCache<P>::BindRenderTarget(ImageViewId* old_id, ImageViewId new_id)
     if (new_id) {
         const ImageViewBase& old_view = slot_image_views[new_id];
         if (True(old_view.flags & ImageViewFlagBits::PreemtiveDownload)) {
-            uncommitted_downloads.push_back(old_view.image_id);
+            const PendingDownload new_download{true, 0, old_view.image_id};
+            uncommitted_downloads.emplace_back(new_download);
         }
     }
     *old_id = new_id;
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index 5a5b4179c..758b7e212 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/src/video_core/texture_cache/texture_cache_base.h
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2021 yuzu Emulator Project
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 #pragma once
@@ -40,14 +40,9 @@ struct ChannelState;
 
 namespace VideoCommon {
 
-using Tegra::Texture::SwizzleSource;
 using Tegra::Texture::TICEntry;
 using Tegra::Texture::TSCEntry;
-using VideoCore::Surface::GetFormatType;
-using VideoCore::Surface::IsCopyCompatible;
 using VideoCore::Surface::PixelFormat;
-using VideoCore::Surface::PixelFormatFromDepthFormat;
-using VideoCore::Surface::PixelFormatFromRenderTargetFormat;
 using namespace Common::Literals;
 
 struct ImageViewInOut {
@@ -119,6 +114,7 @@ class TextureCache : public VideoCommon::ChannelSetupCaches<TextureCacheChannelI
     using Sampler = typename P::Sampler;
     using Framebuffer = typename P::Framebuffer;
     using AsyncBuffer = typename P::AsyncBuffer;
+    using BufferType = typename P::BufferType;
 
     struct BlitImages {
         ImageId dst_id;
@@ -215,6 +211,10 @@ public:
         const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& buffer_operand,
         const Tegra::DMA::ImageOperand& image_operand, ImageId image_id, bool modifies_image);
 
+    void DownloadImageIntoBuffer(Image* image, BufferType buffer, size_t buffer_offset,
+                                 std::span<const VideoCommon::BufferImageCopy> copies,
+                                 GPUVAddr address = 0, size_t size = 0);
+
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
@@ -424,17 +424,32 @@ private:
     u64 critical_memory;
     size_t critical_gc;
 
+    struct BufferDownload {
+        GPUVAddr address;
+        size_t size;
+    };
+
+    struct PendingDownload {
+        bool is_swizzle;
+        size_t async_buffer_id;
+        SlotId object_id;
+    };
+
     SlotVector<Image> slot_images;
     SlotVector<ImageMapView> slot_map_views;
     SlotVector<ImageView> slot_image_views;
     SlotVector<ImageAlloc> slot_image_allocs;
     SlotVector<Sampler> slot_samplers;
     SlotVector<Framebuffer> slot_framebuffers;
+    SlotVector<BufferDownload> slot_buffer_downloads;
 
     // TODO: This data structure is not optimal and it should be reworked
-    std::vector<ImageId> uncommitted_downloads;
-    std::deque<std::vector<ImageId>> committed_downloads;
-    std::deque<std::optional<AsyncBuffer>> async_buffers;
+
+    std::vector<PendingDownload> uncommitted_downloads;
+    std::deque<std::vector<PendingDownload>> committed_downloads;
+    std::vector<AsyncBuffer> uncommitted_async_buffers;
+    std::deque<std::vector<AsyncBuffer>> async_buffers;
+    std::deque<AsyncBuffer> async_buffers_death_ring;
 
     struct LRUItemParams {
         using ObjectType = ImageId;