renderer_vulkan: Introduce separate cmd buffer for uploads

src/common/settings.h

@@ -358,6 +358,8 @@ struct Values {
                                           Category::RendererDebug};
     // TODO: remove this once AMDVLK supports VK_EXT_depth_bias_control
     bool renderer_amdvlk_depth_bias_workaround{};
+    Setting<bool> disable_buffer_reorder{linkage, false, "disable_buffer_reorder",
+                                         Category::RendererDebug};
 
     // System
     SwitchableSetting<Language, true> language_index{linkage,

src/video_core/CMakeLists.txt

@@ -15,6 +15,7 @@ add_library(video_core STATIC
     buffer_cache/buffer_cache.cpp
     buffer_cache/buffer_cache.h
     buffer_cache/memory_tracker_base.h
+    buffer_cache/usage_tracker.h
     buffer_cache/word_manager.h
     cache_types.h
     cdma_pusher.cpp

src/video_core/buffer_cache/buffer_cache.h

@@ -67,6 +67,7 @@ void BufferCache<P>::TickFrame() {
     if (!channel_state) {
         return;
     }
+    runtime.TickFrame(slot_buffers);
 
     // Calculate hits and shots and move hit bits to the right
     const u32 hits = std::reduce(channel_state->uniform_cache_hits.begin(),
@@ -230,7 +231,10 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     for (const IntervalType& add_interval : tmp_intervals) {
         common_ranges.add(add_interval);
     }
-    runtime.CopyBuffer(dest_buffer, src_buffer, copies);
+    const auto& copy = copies[0];
+    src_buffer.MarkUsage(copy.src_offset, copy.size);
+    dest_buffer.MarkUsage(copy.dst_offset, copy.size);
+    runtime.CopyBuffer(dest_buffer, src_buffer, copies, true);
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
@@ -258,9 +262,10 @@ bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
     common_ranges.subtract(subtract_interval);
 
     const BufferId buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size));
-    auto& dest_buffer = slot_buffers[buffer];
+    Buffer& dest_buffer = slot_buffers[buffer];
     const u32 offset = dest_buffer.Offset(*cpu_dst_address);
     runtime.ClearBuffer(dest_buffer, offset, size, value);
+    dest_buffer.MarkUsage(offset, size);
     return true;
 }
 
@@ -603,6 +608,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
             VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset);
             const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size};
             async_downloads += std::make_pair(base_interval, 1);
+            buffer.MarkUsage(copy.src_offset, copy.size);
             runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
             normalized_copies.push_back(second_copy);
         }
@@ -621,8 +627,9 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
                     // Have in mind the staging buffer offset for the copy
                     copy.dst_offset += download_staging.offset;
                     const std::array copies{copy};
-                    runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies,
-                                       false);
+                    Buffer& buffer = slot_buffers[buffer_id];
+                    buffer.MarkUsage(copy.src_offset, copy.size);
+                    runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
                 }
                 runtime.PostCopyBarrier();
                 runtime.Finish();
@@ -742,7 +749,7 @@ void BufferCache<P>::BindHostIndexBuffer() {
                 {BufferCopy{.src_offset = upload_staging.offset, .dst_offset = 0, .size = size}}};
             std::memcpy(upload_staging.mapped_span.data(),
                         draw_state.inline_index_draw_indexes.data(), size);
-            runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
+            runtime.CopyBuffer(buffer, upload_staging.buffer, copies, true);
         } else {
             buffer.ImmediateUpload(0, draw_state.inline_index_draw_indexes);
         }
@@ -754,6 +761,7 @@ void BufferCache<P>::BindHostIndexBuffer() {
             offset + draw_state.index_buffer.first * draw_state.index_buffer.FormatSizeInBytes();
         runtime.BindIndexBuffer(buffer, new_offset, size);
     } else {
+        buffer.MarkUsage(offset, size);
         runtime.BindIndexBuffer(draw_state.topology, draw_state.index_buffer.format,
                                 draw_state.index_buffer.first, draw_state.index_buffer.count,
                                 buffer, offset, size);
@@ -790,6 +798,7 @@ void BufferCache<P>::BindHostVertexBuffers() {
 
             const u32 stride = maxwell3d->regs.vertex_streams[index].stride;
             const u32 offset = buffer.Offset(binding.cpu_addr);
+            buffer.MarkUsage(offset, binding.size);
 
             host_bindings.buffers.push_back(&buffer);
             host_bindings.offsets.push_back(offset);
@@ -895,6 +904,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
         channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
     }
+    buffer.MarkUsage(offset, size);
     if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
         runtime.BindUniformBuffer(stage, binding_index, buffer, offset, size);
     } else {
@@ -913,6 +923,7 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         const bool is_written = ((channel_state->written_storage_buffers[stage] >> index) & 1) != 0;
 
         if (is_written) {
@@ -943,6 +954,7 @@ void BufferCache<P>::BindHostGraphicsTextureBuffers(size_t stage) {
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
         const PixelFormat format = binding.format;
+        buffer.MarkUsage(offset, size);
         if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) {
             if (((channel_state->image_texture_buffers[stage] >> index) & 1) != 0) {
                 runtime.BindImageBuffer(buffer, offset, size, format);
@@ -975,9 +987,10 @@ void BufferCache<P>::BindHostTransformFeedbackBuffers() {
         MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         host_bindings.buffers.push_back(&buffer);
         host_bindings.offsets.push_back(offset);
-        host_bindings.sizes.push_back(binding.size);
+        host_bindings.sizes.push_back(size);
     }
     if (host_bindings.buffers.size() > 0) {
         runtime.BindTransformFeedbackBuffers(host_bindings);
@@ -1001,6 +1014,7 @@ void BufferCache<P>::BindHostComputeUniformBuffers() {
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
             runtime.BindComputeUniformBuffer(binding_index, buffer, offset, size);
             ++binding_index;
@@ -1021,6 +1035,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         const bool is_written =
             ((channel_state->written_compute_storage_buffers >> index) & 1) != 0;
 
@@ -1053,6 +1068,7 @@ void BufferCache<P>::BindHostComputeTextureBuffers() {
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
         const PixelFormat format = binding.format;
+        buffer.MarkUsage(offset, size);
         if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) {
             if (((channel_state->image_compute_texture_buffers >> index) & 1) != 0) {
                 runtime.BindImageBuffer(buffer, offset, size, format);
@@ -1172,10 +1188,11 @@ void BufferCache<P>::UpdateVertexBuffer(u32 index) {
     if (!gpu_memory->IsWithinGPUAddressRange(gpu_addr_end)) {
         size = static_cast<u32>(gpu_memory->MaxContinuousRange(gpu_addr_begin, size));
     }
+    const BufferId buffer_id = FindBuffer(*cpu_addr, size);
     channel_state->vertex_buffers[index] = Binding{
         .cpu_addr = *cpu_addr,
         .size = size,
-        .buffer_id = FindBuffer(*cpu_addr, size),
+        .buffer_id = buffer_id,
     };
 }
 
@@ -1406,7 +1423,8 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
         .dst_offset = dst_base_offset,
         .size = overlap.SizeBytes(),
     });
-    runtime.CopyBuffer(new_buffer, overlap, copies);
+    new_buffer.MarkUsage(copies[0].dst_offset, copies[0].size);
+    runtime.CopyBuffer(new_buffer, overlap, copies, true);
     DeleteBuffer(overlap_id, true);
 }
 
@@ -1419,7 +1437,9 @@ BufferId BufferCache<P>::CreateBuffer(VAddr cpu_addr, u32 wanted_size) {
     const u32 size = static_cast<u32>(overlap.end - overlap.begin);
     const BufferId new_buffer_id = slot_buffers.insert(runtime, rasterizer, overlap.begin, size);
     auto& new_buffer = slot_buffers[new_buffer_id];
-    runtime.ClearBuffer(new_buffer, 0, new_buffer.SizeBytes(), 0);
+    const size_t size_bytes = new_buffer.SizeBytes();
+    runtime.ClearBuffer(new_buffer, 0, size_bytes, 0);
+    new_buffer.MarkUsage(0, size_bytes);
     for (const BufferId overlap_id : overlap.ids) {
         JoinOverlap(new_buffer_id, overlap_id, !overlap.has_stream_leap);
     }
@@ -1472,11 +1492,6 @@ void BufferCache<P>::TouchBuffer(Buffer& buffer, BufferId buffer_id) noexcept {
 
 template <class P>
 bool BufferCache<P>::SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size) {
-    return SynchronizeBufferImpl(buffer, cpu_addr, size);
-}
-
-template <class P>
-bool BufferCache<P>::SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size) {
     boost::container::small_vector<BufferCopy, 4> copies;
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
@@ -1498,51 +1513,6 @@ bool BufferCache<P>::SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 s
     return false;
 }
 
-template <class P>
-bool BufferCache<P>::SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr, u32 size) {
-    boost::container::small_vector<BufferCopy, 4> copies;
-    u64 total_size_bytes = 0;
-    u64 largest_copy = 0;
-    IntervalSet found_sets{};
-    auto make_copies = [&] {
-        for (auto& interval : found_sets) {
-            const std::size_t sub_size = interval.upper() - interval.lower();
-            const VAddr cpu_addr_ = interval.lower();
-            copies.push_back(BufferCopy{
-                .src_offset = total_size_bytes,
-                .dst_offset = cpu_addr_ - buffer.CpuAddr(),
-                .size = sub_size,
-            });
-            total_size_bytes += sub_size;
-            largest_copy = std::max<u64>(largest_copy, sub_size);
-        }
-        const std::span<BufferCopy> copies_span(copies.data(), copies.size());
-        UploadMemory(buffer, total_size_bytes, largest_copy, copies_span);
-    };
-    memory_tracker.ForEachUploadRange(cpu_addr, size, [&](u64 cpu_addr_out, u64 range_size) {
-        const VAddr base_adr = cpu_addr_out;
-        const VAddr end_adr = base_adr + range_size;
-        const IntervalType add_interval{base_adr, end_adr};
-        found_sets.add(add_interval);
-    });
-    if (found_sets.empty()) {
-        return true;
-    }
-    const IntervalType search_interval{cpu_addr, cpu_addr + size};
-    auto it = common_ranges.lower_bound(search_interval);
-    auto it_end = common_ranges.upper_bound(search_interval);
-    if (it == common_ranges.end()) {
-        make_copies();
-        return false;
-    }
-    while (it != it_end) {
-        found_sets.subtract(*it);
-        it++;
-    }
-    make_copies();
-    return false;
-}
-
 template <class P>
 void BufferCache<P>::UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                                   std::span<BufferCopy> copies) {
@@ -1591,7 +1561,8 @@ void BufferCache<P>::MappedUploadMemory([[maybe_unused]] Buffer& buffer,
             // Apply the staging offset
             copy.src_offset += upload_staging.offset;
         }
-        runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
+        const bool can_reorder = runtime.CanReorderUpload(buffer, copies);
+        runtime.CopyBuffer(buffer, upload_staging.buffer, copies, true, can_reorder);
     }
 }
 
@@ -1633,7 +1604,8 @@ void BufferCache<P>::InlineMemoryImplementation(VAddr dest_address, size_t copy_
         }};
         u8* const src_pointer = upload_staging.mapped_span.data();
         std::memcpy(src_pointer, inlined_buffer.data(), copy_size);
-        runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
+        const bool can_reorder = runtime.CanReorderUpload(buffer, copies);
+        runtime.CopyBuffer(buffer, upload_staging.buffer, copies, true, can_reorder);
     } else {
         buffer.ImmediateUpload(buffer.Offset(dest_address), inlined_buffer.first(copy_size));
     }
@@ -1686,8 +1658,9 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
         for (BufferCopy& copy : copies) {
             // Modify copies to have the staging offset in mind
             copy.dst_offset += download_staging.offset;
+            buffer.MarkUsage(copy.src_offset, copy.size);
         }
-        runtime.CopyBuffer(download_staging.buffer, buffer, copies_span);
+        runtime.CopyBuffer(download_staging.buffer, buffer, copies_span, true);
         runtime.Finish();
         for (const BufferCopy& copy : copies) {
             const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;

src/video_core/buffer_cache/buffer_cache_base.h

@@ -529,10 +529,6 @@ private:
 
     bool SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size);
 
-    bool SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size);
-
-    bool SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr, u32 size);
-
     void UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                       std::span<BufferCopy> copies);
 

src/video_core/buffer_cache/usage_tracker.h

@@ -0,0 +1,79 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "common/alignment.h"
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+class UsageTracker {
+    static constexpr size_t BYTES_PER_BIT_SHIFT = 6;
+    static constexpr size_t PAGE_SHIFT = 6 + BYTES_PER_BIT_SHIFT;
+    static constexpr size_t PAGE_BYTES = 1 << PAGE_SHIFT;
+
+public:
+    explicit UsageTracker(size_t size) {
+        const size_t num_pages = (size >> PAGE_SHIFT) + 1;
+        pages.resize(num_pages, 0ULL);
+    }
+
+    void Reset() noexcept {
+        std::ranges::fill(pages, 0ULL);
+    }
+
+    void Track(u64 offset, u64 size) noexcept {
+        const size_t page = offset >> PAGE_SHIFT;
+        const size_t page_end = (offset + size) >> PAGE_SHIFT;
+        TrackPage(page, offset, size);
+        if (page == page_end) {
+            return;
+        }
+        for (size_t i = page + 1; i < page_end; i++) {
+            pages[i] = ~u64{0};
+        }
+        const size_t offset_end = offset + size;
+        const size_t offset_end_page_aligned = Common::AlignDown(offset_end, PAGE_BYTES);
+        TrackPage(page_end, offset_end_page_aligned, offset_end - offset_end_page_aligned);
+    }
+
+    [[nodiscard]] bool IsUsed(u64 offset, u64 size) const noexcept {
+        const size_t page = offset >> PAGE_SHIFT;
+        const size_t page_end = (offset + size) >> PAGE_SHIFT;
+        if (IsPageUsed(page, offset, size)) {
+            return true;
+        }
+        for (size_t i = page + 1; i < page_end; i++) {
+            if (pages[i] != 0) {
+                return true;
+            }
+        }
+        const size_t offset_end = offset + size;
+        const size_t offset_end_page_aligned = Common::AlignDown(offset_end, PAGE_BYTES);
+        return IsPageUsed(page_end, offset_end_page_aligned, offset_end - offset_end_page_aligned);
+    }
+
+private:
+    void TrackPage(u64 page, u64 offset, u64 size) noexcept {
+        const size_t offset_in_page = offset % PAGE_BYTES;
+        const size_t first_bit = offset_in_page >> BYTES_PER_BIT_SHIFT;
+        const size_t num_bits = std::min(size, PAGE_BYTES) >> BYTES_PER_BIT_SHIFT;
+        const size_t mask = ~u64{0} >> (64 - num_bits);
+        pages[page] |= (~u64{0} & mask) << first_bit;
+    }
+
+    bool IsPageUsed(u64 page, u64 offset, u64 size) const noexcept {
+        const size_t offset_in_page = offset % PAGE_BYTES;
+        const size_t first_bit = offset_in_page >> BYTES_PER_BIT_SHIFT;
+        const size_t num_bits = std::min(size, PAGE_BYTES) >> BYTES_PER_BIT_SHIFT;
+        const size_t mask = ~u64{0} >> (64 - num_bits);
+        const size_t mask2 = (~u64{0} & mask) << first_bit;
+        return (pages[page] & mask2) != 0;
+    }
+
+private:
+    std::vector<u64> pages;
+};
+
+} // namespace VideoCommon

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -178,13 +178,14 @@ void BufferCacheRuntime::CopyBuffer(GLuint dst_buffer, Buffer& src_buffer,
 }
 
 void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, GLuint src_buffer,
-                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier) {
+                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                                    bool) {
     CopyBuffer(dst_buffer.Handle(), src_buffer, copies, barrier);
 }
 
 void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
-                                    std::span<const VideoCommon::BufferCopy> copies) {
-    CopyBuffer(dst_buffer.Handle(), src_buffer.Handle(), copies);
+                                    std::span<const VideoCommon::BufferCopy> copies, bool) {
+    CopyBuffer(dst_buffer.Handle(), src_buffer.Handle(), copies, true);
 }
 
 void BufferCacheRuntime::PreCopyBarrier() {

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -30,6 +30,8 @@ public:
 
     void MakeResident(GLenum access) noexcept;
 
+    void MarkUsage(u64 offset, u64 size) {}
+
     [[nodiscard]] GLuint View(u32 offset, u32 size, VideoCore::Surface::PixelFormat format);
 
     [[nodiscard]] GLuint64EXT HostGpuAddr() const noexcept {
@@ -66,22 +68,29 @@ public:
 
     [[nodiscard]] StagingBufferMap DownloadStagingBuffer(size_t size);
 
+    bool CanReorderUpload(const Buffer&, std::span<const VideoCommon::BufferCopy>) {
+        return false;
+    }
+
     void CopyBuffer(GLuint dst_buffer, GLuint src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier);
 
     void CopyBuffer(GLuint dst_buffer, Buffer& src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier);
 
     void CopyBuffer(Buffer& dst_buffer, GLuint src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                    bool can_reorder_upload = false);
 
     void CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies);
+                    std::span<const VideoCommon::BufferCopy> copies, bool);
 
     void PreCopyBarrier();
     void PostCopyBarrier();
     void Finish();
 
+    void TickFrame(VideoCommon::SlotVector<Buffer>&) noexcept {}
+
     void ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value);
 
     void BindIndexBuffer(Buffer& buffer, u32 offset, u32 size);

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -79,13 +79,13 @@ vk::Buffer CreateBuffer(const Device& device, const MemoryAllocator& memory_allo
 } // Anonymous namespace
 
 Buffer::Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params)
-    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(null_params) {}
+    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(null_params), tracker{4096} {}
 
 Buffer::Buffer(BufferCacheRuntime& runtime, VideoCore::RasterizerInterface& rasterizer_,
                VAddr cpu_addr_, u64 size_bytes_)
     : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(rasterizer_, cpu_addr_, size_bytes_),
-      device{&runtime.device}, buffer{
-                                   CreateBuffer(*device, runtime.memory_allocator, SizeBytes())} {
+      device{&runtime.device}, buffer{CreateBuffer(*device, runtime.memory_allocator, SizeBytes())},
+      tracker{SizeBytes()} {
     if (runtime.device.HasDebuggingToolAttached()) {
         buffer.SetObjectNameEXT(fmt::format("Buffer 0x{:x}", CpuAddr()).c_str());
     }
@@ -355,12 +355,31 @@ bool BufferCacheRuntime::CanReportMemoryUsage() const {
     return device.CanReportMemoryUsage();
 }
 
+void BufferCacheRuntime::TickFrame(VideoCommon::SlotVector<Buffer>& slot_buffers) noexcept {
+    for (auto it = slot_buffers.begin(); it != slot_buffers.end(); it++) {
+        it->ResetUsageTracking();
+    }
+}
+
 void BufferCacheRuntime::Finish() {
     scheduler.Finish();
 }
 
+bool BufferCacheRuntime::CanReorderUpload(const Buffer& buffer,
+                                          std::span<const VideoCommon::BufferCopy> copies) {
+    if (Settings::values.disable_buffer_reorder) {
+        return false;
+    }
+    const bool can_use_upload_cmdbuf =
+        std::ranges::all_of(copies, [&](const VideoCommon::BufferCopy& copy) {
+            return !buffer.IsRegionUsed(copy.dst_offset, copy.size);
+        });
+    return can_use_upload_cmdbuf;
+}
+
 void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
-                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier) {
+                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                                    bool can_reorder_upload) {
     if (dst_buffer == VK_NULL_HANDLE || src_buffer == VK_NULL_HANDLE) {
         return;
     }
@@ -376,9 +395,18 @@ void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
         .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
         .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
     };
+
     // Measuring a popular game, this number never exceeds the specified size once data is warmed up
     boost::container::small_vector<VkBufferCopy, 8> vk_copies(copies.size());
     std::ranges::transform(copies, vk_copies.begin(), MakeBufferCopy);
+    if (src_buffer == staging_pool.StreamBuf() && can_reorder_upload) {
+        scheduler.RecordWithUploadBuffer([src_buffer, dst_buffer, vk_copies](
+                                             vk::CommandBuffer, vk::CommandBuffer upload_cmdbuf) {
+            upload_cmdbuf.CopyBuffer(src_buffer, dst_buffer, vk_copies);
+        });
+        return;
+    }
+
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([src_buffer, dst_buffer, vk_copies, barrier](vk::CommandBuffer cmdbuf) {
         if (barrier) {

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -5,6 +5,7 @@
 
 #include "video_core/buffer_cache/buffer_cache_base.h"
 #include "video_core/buffer_cache/memory_tracker_base.h"
+#include "video_core/buffer_cache/usage_tracker.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
@@ -34,6 +35,18 @@ public:
         return *buffer;
     }
 
+    [[nodiscard]] bool IsRegionUsed(u64 offset, u64 size) const noexcept {
+        return tracker.IsUsed(offset, size);
+    }
+
+    void MarkUsage(u64 offset, u64 size) noexcept {
+        tracker.Track(offset, size);
+    }
+
+    void ResetUsageTracking() noexcept {
+        tracker.Reset();
+    }
+
     operator VkBuffer() const noexcept {
         return *buffer;
     }
@@ -49,6 +62,7 @@ private:
     const Device* device{};
     vk::Buffer buffer;
     std::vector<BufferView> views;
+    VideoCommon::UsageTracker tracker;
 };
 
 class QuadArrayIndexBuffer;
@@ -67,6 +81,8 @@ public:
                                 ComputePassDescriptorQueue& compute_pass_descriptor_queue,
                                 DescriptorPool& descriptor_pool);
 
+    void TickFrame(VideoCommon::SlotVector<Buffer>& slot_buffers) noexcept;
+
     void Finish();
 
     u64 GetDeviceLocalMemory() const;
@@ -79,12 +95,15 @@ public:
 
     [[nodiscard]] StagingBufferRef DownloadStagingBuffer(size_t size, bool deferred = false);
 
+    bool CanReorderUpload(const Buffer& buffer, std::span<const VideoCommon::BufferCopy> copies);
+
     void FreeDeferredStagingBuffer(StagingBufferRef& ref);
 
     void PreCopyBarrier();
 
     void CopyBuffer(VkBuffer src_buffer, VkBuffer dst_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                    bool can_reorder_upload = false);
 
     void PostCopyBarrier();
 

src/video_core/renderer_vulkan/vk_master_semaphore.cpp

@@ -100,12 +100,14 @@ void MasterSemaphore::Wait(u64 tick) {
     Refresh();
 }
 
-VkResult MasterSemaphore::SubmitQueue(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                                      VkSemaphore wait_semaphore, u64 host_tick) {
+VkResult MasterSemaphore::SubmitQueue(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                                      VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                                      u64 host_tick) {
     if (semaphore) {
-        return SubmitQueueTimeline(cmdbuf, signal_semaphore, wait_semaphore, host_tick);
+        return SubmitQueueTimeline(cmdbuf, upload_cmdbuf, signal_semaphore, wait_semaphore,
+                                   host_tick);
     } else {
-        return SubmitQueueFence(cmdbuf, signal_semaphore, wait_semaphore, host_tick);
+        return SubmitQueueFence(cmdbuf, upload_cmdbuf, signal_semaphore, wait_semaphore, host_tick);
     }
 }
 
@@ -115,6 +117,7 @@ static constexpr std::array<VkPipelineStageFlags, 2> wait_stage_masks{
 };
 
 VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
+                                              vk::CommandBuffer& upload_cmdbuf,
                                               VkSemaphore signal_semaphore,
                                               VkSemaphore wait_semaphore, u64 host_tick) {
     const VkSemaphore timeline_semaphore = *semaphore;
@@ -123,6 +126,8 @@ VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
     const std::array signal_values{host_tick, u64(0)};
     const std::array signal_semaphores{timeline_semaphore, signal_semaphore};
 
+    const std::array cmdbuffers{*upload_cmdbuf, *cmdbuf};
+
     const u32 num_wait_semaphores = wait_semaphore ? 1 : 0;
     const VkTimelineSemaphoreSubmitInfo timeline_si{
         .sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO,
@@ -138,8 +143,8 @@ VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
         .waitSemaphoreCount = num_wait_semaphores,
         .pWaitSemaphores = &wait_semaphore,
         .pWaitDstStageMask = wait_stage_masks.data(),
-        .commandBufferCount = 1,
-        .pCommandBuffers = cmdbuf.address(),
+        .commandBufferCount = static_cast<u32>(cmdbuffers.size()),
+        .pCommandBuffers = cmdbuffers.data(),
         .signalSemaphoreCount = num_signal_semaphores,
         .pSignalSemaphores = signal_semaphores.data(),
     };
@@ -147,19 +152,23 @@ VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
     return device.GetGraphicsQueue().Submit(submit_info);
 }
 
-VkResult MasterSemaphore::SubmitQueueFence(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                                           VkSemaphore wait_semaphore, u64 host_tick) {
+VkResult MasterSemaphore::SubmitQueueFence(vk::CommandBuffer& cmdbuf,
+                                           vk::CommandBuffer& upload_cmdbuf,
+                                           VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                                           u64 host_tick) {
     const u32 num_signal_semaphores = signal_semaphore ? 1 : 0;
     const u32 num_wait_semaphores = wait_semaphore ? 1 : 0;
 
+    const std::array cmdbuffers{*upload_cmdbuf, *cmdbuf};
+
     const VkSubmitInfo submit_info{
         .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
         .pNext = nullptr,
         .waitSemaphoreCount = num_wait_semaphores,
         .pWaitSemaphores = &wait_semaphore,
         .pWaitDstStageMask = wait_stage_masks.data(),
-        .commandBufferCount = 1,
-        .pCommandBuffers = cmdbuf.address(),
+        .commandBufferCount = static_cast<u32>(cmdbuffers.size()),
+        .pCommandBuffers = cmdbuffers.data(),
         .signalSemaphoreCount = num_signal_semaphores,
         .pSignalSemaphores = &signal_semaphore,
     };

src/video_core/renderer_vulkan/vk_master_semaphore.h

@@ -52,14 +52,16 @@ public:
     void Wait(u64 tick);
 
     /// Submits the device graphics queue, updating the tick as necessary
-    VkResult SubmitQueue(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                         VkSemaphore wait_semaphore, u64 host_tick);
+    VkResult SubmitQueue(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                         VkSemaphore signal_semaphore, VkSemaphore wait_semaphore, u64 host_tick);
 
 private:
-    VkResult SubmitQueueTimeline(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                                 VkSemaphore wait_semaphore, u64 host_tick);
-    VkResult SubmitQueueFence(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                              VkSemaphore wait_semaphore, u64 host_tick);
+    VkResult SubmitQueueTimeline(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                                 VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                                 u64 host_tick);
+    VkResult SubmitQueueFence(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                              VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                              u64 host_tick);
 
     void WaitThread(std::stop_token token);
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -22,11 +22,12 @@ namespace Vulkan {
 
 MICROPROFILE_DECLARE(Vulkan_WaitForWorker);
 
-void Scheduler::CommandChunk::ExecuteAll(vk::CommandBuffer cmdbuf) {
+void Scheduler::CommandChunk::ExecuteAll(vk::CommandBuffer cmdbuf,
+                                         vk::CommandBuffer upload_cmdbuf) {
     auto command = first;
     while (command != nullptr) {
         auto next = command->GetNext();
-        command->Execute(cmdbuf);
+        command->Execute(cmdbuf, upload_cmdbuf);
         command->~Command();
         command = next;
     }
@@ -180,7 +181,7 @@ void Scheduler::WorkerThread(std::stop_token stop_token) {
             // Perform the work, tracking whether the chunk was a submission
             // before executing.
             const bool has_submit = work->HasSubmit();
-            work->ExecuteAll(current_cmdbuf);
+            work->ExecuteAll(current_cmdbuf, current_upload_cmdbuf);
 
             // If the chunk was a submission, reallocate the command buffer.
             if (has_submit) {
@@ -205,6 +206,13 @@ void Scheduler::AllocateWorkerCommandBuffer() {
         .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
         .pInheritanceInfo = nullptr,
     });
+    current_upload_cmdbuf = vk::CommandBuffer(command_pool->Commit(), device.GetDispatchLoader());
+    current_upload_cmdbuf.Begin({
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+        .pNext = nullptr,
+        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+        .pInheritanceInfo = nullptr,
+    });
 }
 
 u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
@@ -212,7 +220,9 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
     InvalidateState();
 
     const u64 signal_value = master_semaphore->NextTick();
-    Record([signal_semaphore, wait_semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
+    RecordWithUploadBuffer([signal_semaphore, wait_semaphore, signal_value,
+                            this](vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) {
+        upload_cmdbuf.End();
         cmdbuf.End();
 
         if (on_submit) {
@@ -221,7 +231,7 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
 
         std::scoped_lock lock{submit_mutex};
         switch (const VkResult result = master_semaphore->SubmitQueue(
-                    cmdbuf, signal_semaphore, wait_semaphore, signal_value)) {
+                    cmdbuf, upload_cmdbuf, signal_semaphore, wait_semaphore, signal_value)) {
         case VK_SUCCESS:
             break;
         case VK_ERROR_DEVICE_LOST:

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -80,7 +80,8 @@ public:
 
     /// Send work to a separate thread.
     template <typename T>
-    void Record(T&& command) {
+        requires std::is_invocable_v<T, vk::CommandBuffer, vk::CommandBuffer>
+    void RecordWithUploadBuffer(T&& command) {
         if (chunk->Record(command)) {
             return;
         }
@@ -88,6 +89,15 @@ public:
         (void)chunk->Record(command);
     }
 
+    template <typename T>
+        requires std::is_invocable_v<T, vk::CommandBuffer>
+    void Record(T&& c) {
+        this->RecordWithUploadBuffer(
+            [command = std::move(c)](vk::CommandBuffer cmdbuf, vk::CommandBuffer) {
+                command(cmdbuf);
+            });
+    }
+
     /// Returns the current command buffer tick.
     [[nodiscard]] u64 CurrentTick() const noexcept {
         return master_semaphore->CurrentTick();
@@ -119,7 +129,7 @@ private:
     public:
         virtual ~Command() = default;
 
-        virtual void Execute(vk::CommandBuffer cmdbuf) const = 0;
+        virtual void Execute(vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) const = 0;
 
         Command* GetNext() const {
             return next;
@@ -142,8 +152,8 @@ private:
         TypedCommand(TypedCommand&&) = delete;
         TypedCommand& operator=(TypedCommand&&) = delete;
 
-        void Execute(vk::CommandBuffer cmdbuf) const override {
-            command(cmdbuf);
+        void Execute(vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) const override {
+            command(cmdbuf, upload_cmdbuf);
         }
 
     private:
@@ -152,7 +162,7 @@ private:
 
     class CommandChunk final {
     public:
-        void ExecuteAll(vk::CommandBuffer cmdbuf);
+        void ExecuteAll(vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf);
 
         template <typename T>
         bool Record(T& command) {
@@ -228,6 +238,7 @@ private:
     VideoCommon::QueryCacheBase<QueryCacheParams>* query_cache = nullptr;
 
     vk::CommandBuffer current_cmdbuf;
+    vk::CommandBuffer current_upload_cmdbuf;
 
     std::unique_ptr<CommandChunk> chunk;
     std::function<void()> on_submit;

src/video_core/renderer_vulkan/vk_smaa.cpp

@@ -672,7 +672,7 @@ void SMAA::UploadImages(Scheduler& scheduler) {
     UploadImage(m_device, m_allocator, scheduler, m_static_images[Search], search_extent,
                 VK_FORMAT_R8_UNORM, ARRAY_TO_SPAN(searchTexBytes));
 
-    scheduler.Record([&](vk::CommandBuffer& cmdbuf) {
+    scheduler.Record([&](vk::CommandBuffer cmdbuf) {
         for (auto& images : m_dynamic_images) {
             for (size_t i = 0; i < MaxDynamicImage; i++) {
                 ClearColorImage(cmdbuf, *images.images[i]);
@@ -707,7 +707,7 @@ VkImageView SMAA::Draw(Scheduler& scheduler, size_t image_index, VkImage source_
     UpdateDescriptorSets(source_image_view, image_index);
 
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([=, this](vk::CommandBuffer& cmdbuf) {
+    scheduler.Record([=, this](vk::CommandBuffer cmdbuf) {
         TransitionImageLayout(cmdbuf, source_image, VK_IMAGE_LAYOUT_GENERAL);
         TransitionImageLayout(cmdbuf, edges_image, VK_IMAGE_LAYOUT_GENERAL);
         BeginRenderPass(cmdbuf, m_renderpasses[EdgeDetection], edge_detection_framebuffer,

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -36,6 +36,10 @@ public:
     StagingBufferRef Request(size_t size, MemoryUsage usage, bool deferred = false);
     void FreeDeferred(StagingBufferRef& ref);
 
+    [[nodiscard]] VkBuffer StreamBuf() const noexcept {
+        return *stream_buffer;
+    }
+
     void TickFrame();
 
 private:

src/video_core/texture_cache/slot_vector.h

@@ -138,6 +138,10 @@ public:
         return Iterator(this, SlotId{SlotId::INVALID_INDEX});
     }
 
+    [[nodiscard]] size_t size() const noexcept {
+        return values_capacity - free_list.size();
+    }
+
 private:
     struct NonTrivialDummy {
         NonTrivialDummy() noexcept {}

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -1101,6 +1101,10 @@ public:
         return &handle;
     }
 
+    VkCommandBuffer operator*() const noexcept {
+        return handle;
+    }
+
     void Begin(const VkCommandBufferBeginInfo& begin_info) const {
         Check(dld->vkBeginCommandBuffer(handle, &begin_info));
     }

src/yuzu/configuration/configure_debug.cpp

@@ -51,6 +51,8 @@ void ConfigureDebug::SetConfiguration() {
     ui->enable_all_controllers->setChecked(Settings::values.enable_all_controllers.GetValue());
     ui->enable_renderdoc_hotkey->setEnabled(runtime_lock);
     ui->enable_renderdoc_hotkey->setChecked(Settings::values.enable_renderdoc_hotkey.GetValue());
+    ui->disable_buffer_reorder->setEnabled(runtime_lock);
+    ui->disable_buffer_reorder->setChecked(Settings::values.disable_buffer_reorder.GetValue());
     ui->enable_graphics_debugging->setEnabled(runtime_lock);
     ui->enable_graphics_debugging->setChecked(Settings::values.renderer_debug.GetValue());
     ui->enable_shader_feedback->setEnabled(runtime_lock);
@@ -96,6 +98,7 @@ void ConfigureDebug::ApplyConfiguration() {
     Settings::values.enable_all_controllers = ui->enable_all_controllers->isChecked();
     Settings::values.renderer_debug = ui->enable_graphics_debugging->isChecked();
     Settings::values.enable_renderdoc_hotkey = ui->enable_renderdoc_hotkey->isChecked();
+    Settings::values.disable_buffer_reorder = ui->disable_buffer_reorder->isChecked();
     Settings::values.renderer_shader_feedback = ui->enable_shader_feedback->isChecked();
     Settings::values.cpu_debug_mode = ui->enable_cpu_debugging->isChecked();
     Settings::values.enable_nsight_aftermath = ui->enable_nsight_aftermath->isChecked();

src/yuzu/configuration/configure_debug.ui

@@ -271,19 +271,6 @@
           </widget>
          </item>
          <item row="8" column="0">
-          <widget class="QCheckBox" name="disable_macro_hle">
-           <property name="enabled">
-            <bool>true</bool>
-           </property>
-           <property name="toolTip">
-            <string>When checked, it disables the macro HLE functions. Enabling this makes games run slower</string>
-           </property>
-           <property name="text">
-            <string>Disable Macro HLE</string>
-           </property>
-          </widget>
-         </item>
-         <item row="7" column="0">
           <widget class="QCheckBox" name="dump_macros">
            <property name="enabled">
             <bool>true</bool>
@@ -306,17 +293,27 @@
            </property>
           </widget>
          </item>
-         <item row="2" column="0">
-          <widget class="QCheckBox" name="enable_shader_feedback">
+         <item row="6" column="0">
+          <widget class="QCheckBox" name="dump_shaders">
+           <property name="enabled">
+            <bool>true</bool>
+           </property>
            <property name="toolTip">
-            <string>When checked, yuzu will log statistics about the compiled pipeline cache</string>
+            <string>When checked, it will dump all the original assembler shaders from the disk shader cache or game as found</string>
            </property>
            <property name="text">
-            <string>Enable Shader Feedback</string>
+            <string>Dump Game Shaders</string>
            </property>
           </widget>
          </item>
-         <item row="6" column="0">
+         <item row="1" column="0">
+          <widget class="QCheckBox" name="enable_renderdoc_hotkey">
+           <property name="text">
+            <string>Enable Renderdoc Hotkey</string>
+           </property>
+          </widget>
+         </item>
+         <item row="7" column="0">
           <widget class="QCheckBox" name="disable_macro_jit">
            <property name="enabled">
             <bool>true</bool>
@@ -330,20 +327,17 @@
           </widget>
          </item>
          <item row="9" column="0">
-          <spacer name="verticalSpacer_5">
-           <property name="orientation">
-            <enum>Qt::Vertical</enum>
+          <widget class="QCheckBox" name="disable_macro_hle">
+           <property name="enabled">
+            <bool>true</bool>
            </property>
-           <property name="sizeType">
-            <enum>QSizePolicy::Preferred</enum>
+           <property name="toolTip">
+            <string>When checked, it disables the macro HLE functions. Enabling this makes games run slower</string>
            </property>
-           <property name="sizeHint" stdset="0">
-            <size>
-             <width>20</width>
-             <height>0</height>
-            </size>
+           <property name="text">
+            <string>Disable Macro HLE</string>
            </property>
-          </spacer>
+          </widget>
          </item>
          <item row="0" column="0">
           <widget class="QCheckBox" name="enable_graphics_debugging">
@@ -358,23 +352,39 @@
            </property>
           </widget>
          </item>
-         <item row="5" column="0">
-          <widget class="QCheckBox" name="dump_shaders">
-           <property name="enabled">
-            <bool>true</bool>
+         <item row="10" column="0">
+          <spacer name="verticalSpacer_5">
+           <property name="orientation">
+            <enum>Qt::Vertical</enum>
+           </property>
+           <property name="sizeType">
+            <enum>QSizePolicy::Preferred</enum>
            </property>
+           <property name="sizeHint" stdset="0">
+            <size>
+             <width>20</width>
+             <height>0</height>
+            </size>
+           </property>
+          </spacer>
+         </item>
+         <item row="2" column="0">
+          <widget class="QCheckBox" name="enable_shader_feedback">
            <property name="toolTip">
-            <string>When checked, it will dump all the original assembler shaders from the disk shader cache or game as found</string>
+            <string>When checked, yuzu will log statistics about the compiled pipeline cache</string>
            </property>
            <property name="text">
-            <string>Dump Game Shaders</string>
+            <string>Enable Shader Feedback</string>
            </property>
           </widget>
          </item>
-         <item row="1" column="0">
-          <widget class="QCheckBox" name="enable_renderdoc_hotkey">
+         <item row="5" column="0">
+          <widget class="QCheckBox" name="disable_buffer_reorder">
+           <property name="toolTip">
+            <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;When checked, disables reording of mapped memory uploads which allows to associate uploads with specific draws. May reduce performance in some cases.&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+           </property>
            <property name="text">
-            <string>Enable Renderdoc Hotkey</string>
+            <string>Disable Buffer Reorder</string>
            </property>
           </widget>
          </item>