GPU: Add Reactive flushing

src/core/core.cpp

@@ -612,6 +612,10 @@ void System::PrepareReschedule(const u32 core_index) {
     impl->kernel.PrepareReschedule(core_index);
 }
 
+size_t System::GetCurrentHostThreadID() const {
+    return impl->kernel.GetCurrentHostThreadID();
+}
+
 PerfStatsResults System::GetAndResetPerfStats() {
     return impl->GetAndResetPerfStats();
 }

src/core/core.h

@@ -222,6 +222,8 @@ public:
     /// Prepare the core emulation for a reschedule
     void PrepareReschedule(u32 core_index);
 
+    [[nodiscard]] size_t GetCurrentHostThreadID() const;
+
     /// Gets and resets core performance statistics
     [[nodiscard]] PerfStatsResults GetAndResetPerfStats();
 

src/core/memory.cpp

@@ -13,10 +13,13 @@
 #include "common/swap.h"
 #include "core/core.h"
 #include "core/device_memory.h"
+#include "core/hardware_properties.h"
 #include "core/hle/kernel/k_page_table.h"
 #include "core/hle/kernel/k_process.h"
 #include "core/memory.h"
 #include "video_core/gpu.h"
+#include "video_core/rasterizer_download_area.h"
+
 
 namespace Core::Memory {
 
@@ -243,7 +246,7 @@ struct Memory::Impl {
             [&](const Common::ProcessAddress current_vaddr, const std::size_t copy_amount,
                 const u8* const host_ptr) {
                 if constexpr (!UNSAFE) {
-                    system.GPU().FlushRegion(GetInteger(current_vaddr), copy_amount);
+                    HandleRasterizerDownload(GetInteger(current_vaddr), copy_amount);
                 }
                 std::memcpy(dest_buffer, host_ptr, copy_amount);
             },
@@ -334,7 +337,7 @@ struct Memory::Impl {
             },
             [&](const Common::ProcessAddress current_vaddr, const std::size_t copy_amount,
                 u8* const host_ptr) {
-                system.GPU().FlushRegion(GetInteger(current_vaddr), copy_amount);
+                HandleRasterizerDownload(GetInteger(current_vaddr), copy_amount);
                 WriteBlockImpl<false>(process, dest_addr, host_ptr, copy_amount);
             },
             [&](const std::size_t copy_amount) {
@@ -373,7 +376,7 @@ struct Memory::Impl {
                                  const std::size_t block_size) {
             // dc ivac: Invalidate to point of coherency
             // GPU flush -> CPU invalidate
-            system.GPU().FlushRegion(GetInteger(current_vaddr), block_size);
+            HandleRasterizerDownload(GetInteger(current_vaddr), block_size);
         };
         return PerformCacheOperation(process, dest_addr, size, on_rasterizer);
     }
@@ -462,8 +465,7 @@ struct Memory::Impl {
         }
 
         if (Settings::IsFastmemEnabled()) {
-            const bool is_read_enable = !Settings::IsGPULevelExtreme() || !cached;
-            system.DeviceMemory().buffer.Protect(vaddr, size, is_read_enable, !cached);
+            system.DeviceMemory().buffer.Protect(vaddr, size, !cached, !cached);
         }
 
         // Iterate over a contiguous CPU address space, which corresponds to the specified GPU
@@ -651,7 +653,9 @@ struct Memory::Impl {
                 LOG_ERROR(HW_Memory, "Unmapped Read{} @ 0x{:016X}", sizeof(T) * 8,
                           GetInteger(vaddr));
             },
-            [&]() { system.GPU().FlushRegion(GetInteger(vaddr), sizeof(T)); });
+            [&]() {
+                HandleRasterizerDownload(GetInteger(vaddr), sizeof(T));
+            });
         if (ptr) {
             std::memcpy(&result, ptr, sizeof(T));
         }
@@ -712,7 +716,18 @@ struct Memory::Impl {
         return true;
     }
 
+    void HandleRasterizerDownload(VAddr address, size_t size) {
+        const size_t core = system.GetCurrentHostThreadID();
+        auto& current_area = rasterizer_areas[core];
+        const VAddr end_address = address + size;
+        if (current_area.start_address <= address && end_address <= current_area.end_address) [[likely]] {
+            return;
+        }
+        current_area = system.GPU().OnCPURead(address, size);
+    }
+
     Common::PageTable* current_page_table = nullptr;
+    std::array<VideoCore::RasterizerDownloadArea, Core::Hardware::NUM_CPU_CORES> rasterizer_areas{};
     Core::System& system;
 };
 

src/video_core/buffer_cache/buffer_base.h

@@ -18,6 +18,7 @@ namespace VideoCommon {
 enum class BufferFlagBits {
     Picked = 1 << 0,
     CachedWrites = 1 << 1,
+    PreemtiveDownload = 1 << 2,
 };
 DECLARE_ENUM_FLAG_OPERATORS(BufferFlagBits)
 
@@ -54,6 +55,10 @@ public:
         flags |= BufferFlagBits::Picked;
     }
 
+    void MarkPreemtiveDownload() noexcept {
+        flags |= BufferFlagBits::PreemtiveDownload;
+    }
+
     /// Unmark buffer as picked
     void Unpick() noexcept {
         flags &= ~BufferFlagBits::Picked;
@@ -84,6 +89,10 @@ public:
         return True(flags & BufferFlagBits::CachedWrites);
     }
 
+    bool IsPreemtiveDownload() const noexcept {
+        return True(flags & BufferFlagBits::PreemtiveDownload);
+    }
+
     /// Returns the base CPU address of the buffer
     [[nodiscard]] VAddr CpuAddr() const noexcept {
         return cpu_addr;

src/video_core/buffer_cache/buffer_cache.h

@@ -111,9 +111,24 @@ void BufferCache<P>::WriteMemory(VAddr cpu_addr, u64 size) {
 template <class P>
 void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
     memory_tracker.CachedCpuWrite(cpu_addr, size);
-    const IntervalType add_interval{Common::AlignDown(cpu_addr, YUZU_PAGESIZE),
-                                    Common::AlignUp(cpu_addr + size, YUZU_PAGESIZE)};
-    cached_ranges.add(add_interval);
+}
+
+template <class P>
+std::optional<VideoCore::RasterizerDownloadArea> BufferCache<P>::GetFlushArea(VAddr cpu_addr,
+                                                                              u64 size) {
+    std::optional<VideoCore::RasterizerDownloadArea> area{};
+    area.emplace();
+    VAddr cpu_addr_start_aligned = Common::AlignDown(cpu_addr, Core::Memory::YUZU_PAGESIZE);
+    VAddr cpu_addr_end_aligned = Common::AlignUp(cpu_addr + size, Core::Memory::YUZU_PAGESIZE);
+    area->start_address = cpu_addr_start_aligned;
+    area->end_address = cpu_addr_end_aligned;
+    if (memory_tracker.IsRegionPreflushable(cpu_addr, size)) {
+        area->preemtive = true;
+        return area;
+    };
+    memory_tracker.MarkRegionAsPreflushable(cpu_addr_start_aligned, cpu_addr_end_aligned - cpu_addr_start_aligned);
+    area->preemtive = !IsRegionGpuModified(cpu_addr, size);
+    return area;
 }
 
 template <class P>
@@ -191,8 +206,10 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
         const VAddr new_base_address = *cpu_dest_address + diff;
         const IntervalType add_interval{new_base_address, new_base_address + size};
         tmp_intervals.push_back(add_interval);
-        uncommitted_ranges.add(add_interval);
-        pending_ranges.add(add_interval);
+        if (memory_tracker.IsRegionPreflushable(new_base_address, new_base_address + size)) {
+            uncommitted_ranges.add(add_interval);
+            pending_ranges.add(add_interval);
+        }
     };
     ForEachInRangeSet(common_ranges, *cpu_src_address, amount, mirror);
     // This subtraction in this order is important for overlapping copies.
@@ -205,7 +222,7 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
-    std::vector<u8> tmp_buffer(amount);
+    tmp_buffer.resize(amount);
     cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount);
     cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount);
     return true;
@@ -441,9 +458,7 @@ void BufferCache<P>::BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_add
 
 template <class P>
 void BufferCache<P>::FlushCachedWrites() {
-    cached_write_buffer_ids.clear();
     memory_tracker.FlushCachedWrites();
-    cached_ranges.clear();
 }
 
 template <class P>
@@ -1221,6 +1236,9 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s
 
     const IntervalType base_interval{cpu_addr, cpu_addr + size};
     common_ranges.add(base_interval);
+    if (!memory_tracker.IsRegionPreflushable(cpu_addr, cpu_addr + size)) {
+        return;
+    }
     uncommitted_ranges.add(base_interval);
     pending_ranges.add(base_interval);
 }
@@ -1629,7 +1647,6 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
     replace(transform_feedback_buffers);
     replace(compute_uniform_buffers);
     replace(compute_storage_buffers);
-    std::erase(cached_write_buffer_ids, buffer_id);
 
     // Mark the whole buffer as CPU written to stop tracking CPU writes
     if (!do_not_mark) {

src/video_core/buffer_cache/buffer_cache_base.h

@@ -188,6 +188,8 @@ public:
 
     void DownloadMemory(VAddr cpu_addr, u64 size);
 
+    std::optional<VideoCore::RasterizerDownloadArea> GetFlushArea(VAddr cpu_addr, u64 size);
+
     bool InlineMemory(VAddr dest_address, size_t copy_size, std::span<const u8> inlined_buffer);
 
     void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size);
@@ -541,8 +543,6 @@ private:
                        std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>, Empty>
         uniform_buffer_binding_sizes{};
 
-    std::vector<BufferId> cached_write_buffer_ids;
-
     MemoryTracker memory_tracker;
     IntervalSet uncommitted_ranges;
     IntervalSet common_ranges;
@@ -575,6 +575,7 @@ private:
     bool active_async_buffers = false;
 
     std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table;
+    std::vector<u8> tmp_buffer;
 };
 
 } // namespace VideoCommon

src/video_core/buffer_cache/memory_tracker_base.h

@@ -66,6 +66,14 @@ public:
             });
     }
 
+    /// Returns true if a region has been marked as Preflushable
+    [[nodiscard]] bool IsRegionPreflushable(VAddr query_cpu_addr, u64 query_size) noexcept {
+        return IteratePages<false>(
+            query_cpu_addr, query_size, [](Manager* manager, u64 offset, size_t size) {
+                return manager->template IsRegionModified<Type::Preflushable>(offset, size);
+            });
+    }
+
     /// Mark region as CPU modified, notifying the rasterizer about this change
     void MarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 query_size) {
         IteratePages<true>(dirty_cpu_addr, query_size,
@@ -93,6 +101,15 @@ public:
                            });
     }
 
+    /// Mark region as modified from the host GPU
+    void MarkRegionAsPreflushable(VAddr dirty_cpu_addr, u64 query_size) noexcept {
+        IteratePages<true>(dirty_cpu_addr, query_size,
+                           [](Manager* manager, u64 offset, size_t size) {
+                               manager->template ChangeRegionState<Type::Preflushable, true>(
+                                   manager->GetCpuAddr() + offset, size);
+                           });
+    }
+
     /// Unmark region as modified from the host GPU
     void UnmarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 query_size) noexcept {
         IteratePages<true>(dirty_cpu_addr, query_size,
@@ -102,6 +119,15 @@ public:
                            });
     }
 
+    /// Unmark region as modified from the host GPU
+    void UnmarkRegionAsPreflushable(VAddr dirty_cpu_addr, u64 query_size) noexcept {
+        IteratePages<true>(dirty_cpu_addr, query_size,
+                           [](Manager* manager, u64 offset, size_t size) {
+                               manager->template ChangeRegionState<Type::Preflushable, false>(
+                                   manager->GetCpuAddr() + offset, size);
+                           });
+    }
+
     /// Mark region as modified from the CPU
     /// but don't mark it as modified until FlusHCachedWrites is called.
     void CachedCpuWrite(VAddr dirty_cpu_addr, u64 query_size) {

src/video_core/buffer_cache/word_manager.h

@@ -26,6 +26,7 @@ enum class Type {
     GPU,
     CachedCPU,
     Untracked,
+    Preflushable,
 };
 
 /// Vector tracking modified pages tightly packed with small vector optimization
@@ -55,17 +56,20 @@ struct Words {
             gpu.stack.fill(0);
             cached_cpu.stack.fill(0);
             untracked.stack.fill(~u64{0});
+            preflushable.stack.fill(0);
         } else {
             // Share allocation between CPU and GPU pages and set their default values
-            u64* const alloc = new u64[num_words * 4];
+            u64* const alloc = new u64[num_words * 5];
             cpu.heap = alloc;
             gpu.heap = alloc + num_words;
             cached_cpu.heap = alloc + num_words * 2;
             untracked.heap = alloc + num_words * 3;
+            preflushable.heap = alloc + num_words * 4;
             std::fill_n(cpu.heap, num_words, ~u64{0});
             std::fill_n(gpu.heap, num_words, 0);
             std::fill_n(cached_cpu.heap, num_words, 0);
             std::fill_n(untracked.heap, num_words, ~u64{0});
+            std::fill_n(preflushable.heap, num_words, 0);
         }
         // Clean up tailing bits
         const u64 last_word_size = size_bytes % BYTES_PER_WORD;
@@ -88,13 +92,14 @@ struct Words {
         gpu = rhs.gpu;
         cached_cpu = rhs.cached_cpu;
         untracked = rhs.untracked;
+        preflushable = rhs.preflushable;
         rhs.cpu.heap = nullptr;
         return *this;
     }
 
     Words(Words&& rhs) noexcept
         : size_bytes{rhs.size_bytes}, num_words{rhs.num_words}, cpu{rhs.cpu}, gpu{rhs.gpu},
-          cached_cpu{rhs.cached_cpu}, untracked{rhs.untracked} {
+          cached_cpu{rhs.cached_cpu}, untracked{rhs.untracked}, preflushable{rhs.preflushable} {
         rhs.cpu.heap = nullptr;
     }
 
@@ -129,6 +134,8 @@ struct Words {
             return std::span<u64>(cached_cpu.Pointer(IsShort()), num_words);
         } else if constexpr (type == Type::Untracked) {
             return std::span<u64>(untracked.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::Preflushable) {
+            return std::span<u64>(preflushable.Pointer(IsShort()), num_words);
         }
     }
 
@@ -142,6 +149,8 @@ struct Words {
             return std::span<const u64>(cached_cpu.Pointer(IsShort()), num_words);
         } else if constexpr (type == Type::Untracked) {
             return std::span<const u64>(untracked.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::Preflushable) {
+            return std::span<const u64>(preflushable.Pointer(IsShort()), num_words);
         }
     }
 
@@ -151,6 +160,7 @@ struct Words {
     WordsArray<stack_words> gpu;
     WordsArray<stack_words> cached_cpu;
     WordsArray<stack_words> untracked;
+    WordsArray<stack_words> preflushable;
 };
 
 template <class RasterizerInterface, size_t stack_words = 1>

src/video_core/engines/maxwell_dma.cpp

@@ -223,7 +223,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
     write_buffer.resize_destructive(dst_size);
 
     memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
-    memory_manager.ReadBlockUnsafe(dst_operand.address, write_buffer.data(), dst_size);
+    memory_manager.ReadBlock(dst_operand.address, write_buffer.data(), dst_size);
 
     UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
                      src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
@@ -288,11 +288,7 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
     write_buffer.resize_destructive(dst_size);
 
     memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
-    if (Settings::IsGPULevelExtreme()) {
-        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
-    } else {
-        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
-    }
+    memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
 
     // If the input is linear and the output is tiled, swizzle the input and copy it over.
     SwizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,

src/video_core/fence_manager.h

@@ -55,8 +55,8 @@ public:
 
     // Unlike other fences, this one doesn't
     void SignalOrdering() {
-        std::scoped_lock lock{buffer_cache.mutex};
-        buffer_cache.AccumulateFlushes();
+        std::function<void()> do_nothing([]{});
+        SignalFence(std::move(do_nothing));
     }
 
     void SyncOperation(std::function<void()>&& func) {

src/video_core/gpu.cpp

@@ -283,6 +283,21 @@ struct GPU::Impl {
         gpu_thread.FlushRegion(addr, size);
     }
 
+    VideoCore::RasterizerDownloadArea OnCPURead(VAddr addr, u64 size) {
+        auto raster_area = rasterizer->GetFlushArea(addr, size);
+        if (raster_area.preemtive) {
+            return raster_area;
+        }
+        raster_area.preemtive = true;
+        const u64 fence = RequestSyncOperation([this, &raster_area]() {
+            rasterizer->FlushRegion(raster_area.start_address,
+                                    raster_area.end_address - raster_area.start_address);
+        });
+        gpu_thread.TickGPU();
+        WaitForSyncOperation(fence);
+        return raster_area;
+    }
+
     /// Notify rasterizer that any caches of the specified region should be invalidated
     void InvalidateRegion(VAddr addr, u64 size) {
         gpu_thread.InvalidateRegion(addr, size);
@@ -538,6 +553,10 @@ void GPU::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
     impl->SwapBuffers(framebuffer);
 }
 
+VideoCore::RasterizerDownloadArea GPU::OnCPURead(VAddr addr, u64 size) {
+    return impl->OnCPURead(addr, size);
+}
+
 void GPU::FlushRegion(VAddr addr, u64 size) {
     impl->FlushRegion(addr, size);
 }

src/video_core/gpu.h

@@ -10,6 +10,7 @@
 #include "core/hle/service/nvdrv/nvdata.h"
 #include "video_core/cdma_pusher.h"
 #include "video_core/framebuffer_config.h"
+#include "video_core/rasterizer_download_area.h"
 
 namespace Core {
 class System;
@@ -240,6 +241,9 @@ public:
     /// Swap buffers (render frame)
     void SwapBuffers(const Tegra::FramebufferConfig* framebuffer);
 
+    /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
+    [[nodiscard]] VideoCore::RasterizerDownloadArea OnCPURead(VAddr addr, u64 size);
+
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     void FlushRegion(VAddr addr, u64 size);
 

src/video_core/rasterizer_download_area.h

@@ -0,0 +1,13 @@
+#pragma once
+
+#include "common/common_types.h"
+
+namespace VideoCore {
+
+struct RasterizerDownloadArea {
+    VAddr start_address;
+    VAddr end_address;
+    bool preemtive;
+};
+
+} // namespace VideoCore

src/video_core/rasterizer_interface.h

@@ -12,6 +12,7 @@
 #include "video_core/cache_types.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/gpu.h"
+#include "video_core/rasterizer_download_area.h"
 
 namespace Tegra {
 class MemoryManager;
@@ -95,6 +96,8 @@ public:
     virtual bool MustFlushRegion(VAddr addr, u64 size,
                                  VideoCommon::CacheType which = VideoCommon::CacheType::All) = 0;
 
+    virtual RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) = 0;
+
     /// Notify rasterizer that any caches of the specified region should be invalidated
     virtual void InvalidateRegion(VAddr addr, u64 size,
                                   VideoCommon::CacheType which = VideoCommon::CacheType::All) = 0;

src/video_core/renderer_null/null_rasterizer.cpp

@@ -1,6 +1,8 @@
 // SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#include "common/alignment.h"
+#include "core/memory.h"
 #include "video_core/host1x/host1x.h"
 #include "video_core/memory_manager.h"
 #include "video_core/renderer_null/null_rasterizer.h"
@@ -46,6 +48,14 @@ bool RasterizerNull::MustFlushRegion(VAddr addr, u64 size, VideoCommon::CacheTyp
 }
 void RasterizerNull::InvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType) {}
 void RasterizerNull::OnCPUWrite(VAddr addr, u64 size) {}
+VideoCore::RasterizerDownloadArea RasterizerNull::GetFlushArea(VAddr addr, u64 size) {
+    VideoCore::RasterizerDownloadArea new_area{
+        .start_address = Common::AlignDown(addr, Core::Memory::YUZU_PAGESIZE),
+        .end_address = Common::AlignUp(addr + size, Core::Memory::YUZU_PAGESIZE),
+        .preemtive = true,
+    };
+    return new_area;
+}
 void RasterizerNull::InvalidateGPUCache() {}
 void RasterizerNull::UnmapMemory(VAddr addr, u64 size) {}
 void RasterizerNull::ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) {}

src/video_core/renderer_null/null_rasterizer.h

@@ -54,6 +54,7 @@ public:
     void InvalidateRegion(VAddr addr, u64 size,
                           VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void OnCPUWrite(VAddr addr, u64 size) override;
+    VideoCore::RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) override;
     void InvalidateGPUCache() override;
     void UnmapMemory(VAddr addr, u64 size) override;
     void ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) override;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -433,6 +433,29 @@ bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size, VideoCommon::CacheT
     return false;
 }
 
+VideoCore::RasterizerDownloadArea RasterizerOpenGL::GetFlushArea(VAddr addr, u64 size) {
+    {
+        std::scoped_lock lock{texture_cache.mutex};
+        auto area = texture_cache.GetFlushArea(addr, size);
+        if (area) {
+            return *area;
+        }
+    }
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        auto area = buffer_cache.GetFlushArea(addr, size);
+        if (area) {
+            return *area;
+        }
+    }
+    VideoCore::RasterizerDownloadArea new_area{
+        .start_address = Common::AlignDown(addr, Core::Memory::YUZU_PAGESIZE),
+        .end_address = Common::AlignUp(addr + size, Core::Memory::YUZU_PAGESIZE),
+        .preemtive = true,
+    };
+    return new_area;
+}
+
 void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType which) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
     if (addr == 0 || size == 0) {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -95,6 +95,7 @@ public:
                      VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     bool MustFlushRegion(VAddr addr, u64 size,
                          VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
+    VideoCore::RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size,
                           VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void OnCPUWrite(VAddr addr, u64 size) override;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -502,6 +502,29 @@ bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size, VideoCommon::CacheT
     return false;
 }
 
+VideoCore::RasterizerDownloadArea RasterizerVulkan::GetFlushArea(VAddr addr, u64 size) {
+    {
+        std::scoped_lock lock{texture_cache.mutex};
+        auto area = texture_cache.GetFlushArea(addr, size);
+        if (area) {
+            return *area;
+        }
+    }
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        auto area = buffer_cache.GetFlushArea(addr, size);
+        if (area) {
+            return *area;
+        }
+    }
+    VideoCore::RasterizerDownloadArea new_area{
+        .start_address = Common::AlignDown(addr, Core::Memory::YUZU_PAGESIZE),
+        .end_address = Common::AlignUp(addr + size, Core::Memory::YUZU_PAGESIZE),
+        .preemtive = true,
+    };
+    return new_area;
+}
+
 void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType which) {
     if (addr == 0 || size == 0) {
         return;

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -92,6 +92,7 @@ public:
                      VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     bool MustFlushRegion(VAddr addr, u64 size,
                          VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
+    VideoCore::RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size,
                           VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void InnerInvalidation(std::span<const std::pair<VAddr, std::size_t>> sequences) override;

src/video_core/texture_cache/image_info.h

@@ -39,6 +39,7 @@ struct ImageInfo {
     u32 tile_width_spacing = 0;
     bool rescaleable = false;
     bool downscaleable = false;
+    bool forced_flushed = false;
 };
 
 } // namespace VideoCommon

src/video_core/texture_cache/image_view_base.cpp

@@ -26,8 +26,7 @@ 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) {
+    if (image_info.forced_flushed) {
         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;
@@ -789,11 +815,15 @@ 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.forced_flushed) {
+        image.info.forced_flushed = 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);