Merge pull request #10155 from FernandoS27/reactive-flushing-new

Y.F.C. bring back Reactive Flushing

23 files changed, 302 insertions, 92 deletions

diff --git a/src/video_core/buffer_cache/buffer_base.h b/src/video_core/buffer_cache/buffer_base.h
index 9cbd95c4b..0bb3bf8ae 100644
--- a/src/video_core/buffer_cache/buffer_base.h
+++ b/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;
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index e534e1e9c..6624919a4 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -23,8 +23,6 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
     common_ranges.clear();
     inline_buffer_id = NULL_BUFFER_ID;
 
-    active_async_buffers = !Settings::IsGPULevelHigh();
-
     if (!runtime.CanReportMemoryUsage()) {
         minimum_memory = DEFAULT_EXPECTED_MEMORY;
         critical_memory = DEFAULT_CRITICAL_MEMORY;
@@ -75,8 +73,6 @@ void BufferCache<P>::TickFrame() {
     uniform_cache_hits[0] = 0;
     uniform_cache_shots[0] = 0;
 
-    active_async_buffers = !Settings::IsGPULevelHigh();
-
     const bool skip_preferred = hits * 256 < shots * 251;
     uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
 
@@ -111,9 +107,25 @@ 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>
@@ -205,7 +217,7 @@ 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 +453,7 @@ 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>
@@ -474,9 +484,8 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
 
     if (committed_ranges.empty()) {
         if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
-            if (active_async_buffers) {
-                async_buffers.emplace_back(std::optional<Async_Buffer>{});
-            }
+
+            async_buffers.emplace_back(std::optional<Async_Buffer>{});
         }
         return;
     }
@@ -537,64 +546,65 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
     committed_ranges.clear();
     if (downloads.empty()) {
         if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
-            if (active_async_buffers) {
-                async_buffers.emplace_back(std::optional<Async_Buffer>{});
-            }
+
+            async_buffers.emplace_back(std::optional<Async_Buffer>{});
         }
         return;
     }
-    if (active_async_buffers) {
-        if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
-            auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
-            boost::container::small_vector<BufferCopy, 4> normalized_copies;
-            IntervalSet new_async_range{};
-            runtime.PreCopyBarrier();
-            for (auto& [copy, buffer_id] : downloads) {
-                copy.dst_offset += download_staging.offset;
-                const std::array copies{copy};
-                BufferCopy second_copy{copy};
-                Buffer& buffer = slot_buffers[buffer_id];
-                second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
-                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);
-                runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
-                normalized_copies.push_back(second_copy);
-            }
-            runtime.PostCopyBarrier();
-            pending_downloads.emplace_back(std::move(normalized_copies));
-            async_buffers.emplace_back(download_staging);
-        } else {
-            committed_ranges.clear();
-            uncommitted_ranges.clear();
+    if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
+        auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
+        boost::container::small_vector<BufferCopy, 4> normalized_copies;
+        IntervalSet new_async_range{};
+        runtime.PreCopyBarrier();
+        for (auto& [copy, buffer_id] : downloads) {
+            copy.dst_offset += download_staging.offset;
+            const std::array copies{copy};
+            BufferCopy second_copy{copy};
+            Buffer& buffer = slot_buffers[buffer_id];
+            second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
+            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);
+            runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
+            normalized_copies.push_back(second_copy);
         }
+        runtime.PostCopyBarrier();
+        pending_downloads.emplace_back(std::move(normalized_copies));
+        async_buffers.emplace_back(download_staging);
     } else {
-        if constexpr (USE_MEMORY_MAPS) {
-            auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
-            runtime.PreCopyBarrier();
-            for (auto& [copy, buffer_id] : downloads) {
-                // 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);
-            }
-            runtime.PostCopyBarrier();
-            runtime.Finish();
-            for (const auto& [copy, buffer_id] : downloads) {
-                const Buffer& buffer = slot_buffers[buffer_id];
-                const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
-                // Undo the modified offset
-                const u64 dst_offset = copy.dst_offset - download_staging.offset;
-                const u8* read_mapped_memory = download_staging.mapped_span.data() + dst_offset;
-                cpu_memory.WriteBlockUnsafe(cpu_addr, read_mapped_memory, copy.size);
-            }
+        if (!Settings::IsGPULevelHigh()) {
+            committed_ranges.clear();
+            uncommitted_ranges.clear();
         } else {
-            const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
-            for (const auto& [copy, buffer_id] : downloads) {
-                Buffer& buffer = slot_buffers[buffer_id];
-                buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
-                const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
-                cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
+            if constexpr (USE_MEMORY_MAPS) {
+                auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
+                runtime.PreCopyBarrier();
+                for (auto& [copy, buffer_id] : downloads) {
+                    // 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);
+                }
+                runtime.PostCopyBarrier();
+                runtime.Finish();
+                for (const auto& [copy, buffer_id] : downloads) {
+                    const Buffer& buffer = slot_buffers[buffer_id];
+                    const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
+                    // Undo the modified offset
+                    const u64 dst_offset = copy.dst_offset - download_staging.offset;
+                    const u8* read_mapped_memory = download_staging.mapped_span.data() + dst_offset;
+                    cpu_memory.WriteBlockUnsafe(cpu_addr, read_mapped_memory, copy.size);
+                }
+            } else {
+                const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
+                for (const auto& [copy, buffer_id] : downloads) {
+                    Buffer& buffer = slot_buffers[buffer_id];
+                    buffer.ImmediateDownload(copy.src_offset,
+                                             immediate_buffer.subspan(0, copy.size));
+                    const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
+                    cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
+                }
             }
         }
     }
@@ -1629,7 +1639,6 @@ 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) {
diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h
index 656baa550..0445ec47f 100644
--- a/src/video_core/buffer_cache/buffer_cache_base.h
+++ b/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;
@@ -572,9 +572,8 @@ private:
     u64 critical_memory = 0;
     BufferId inline_buffer_id;
 
-    bool active_async_buffers = false;
-
     std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table;
+    std::vector<u8> tmp_buffer;
 };
 
 } // namespace VideoCommon
diff --git a/src/video_core/buffer_cache/memory_tracker_base.h b/src/video_core/buffer_cache/memory_tracker_base.h
index dc4ebfcaa..6036b21c9 100644
--- a/src/video_core/buffer_cache/memory_tracker_base.h
+++ b/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) {
diff --git a/src/video_core/buffer_cache/word_manager.h b/src/video_core/buffer_cache/word_manager.h
index a42455045..a336bde41 100644
--- a/src/video_core/buffer_cache/word_manager.h
+++ b/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>
@@ -292,6 +302,9 @@ public:
                  (pending_pointer - pending_offset) * BYTES_PER_PAGE);
         };
         IterateWords(offset, size, [&](size_t index, u64 mask) {
+            if constexpr (type == Type::GPU) {
+                mask &= ~untracked_words[index];
+            }
             const u64 word = state_words[index] & mask;
             if constexpr (clear) {
                 if constexpr (type == Type::CPU || type == Type::CachedCPU) {
@@ -340,8 +353,13 @@ public:
         static_assert(type != Type::Untracked);
 
         const std::span<const u64> state_words = words.template Span<type>();
+        [[maybe_unused]] const std::span<const u64> untracked_words =
+            words.template Span<Type::Untracked>();
         bool result = false;
         IterateWords(offset, size, [&](size_t index, u64 mask) {
+            if constexpr (type == Type::GPU) {
+                mask &= ~untracked_words[index];
+            }
             const u64 word = state_words[index] & mask;
             if (word != 0) {
                 result = true;
@@ -362,9 +380,14 @@ public:
     [[nodiscard]] std::pair<u64, u64> ModifiedRegion(u64 offset, u64 size) const noexcept {
         static_assert(type != Type::Untracked);
         const std::span<const u64> state_words = words.template Span<type>();
+        [[maybe_unused]] const std::span<const u64> untracked_words =
+            words.template Span<Type::Untracked>();
         u64 begin = std::numeric_limits<u64>::max();
         u64 end = 0;
         IterateWords(offset, size, [&](size_t index, u64 mask) {
+            if constexpr (type == Type::GPU) {
+                mask &= ~untracked_words[index];
+            }
             const u64 word = state_words[index] & mask;
             if (word == 0) {
                 return;
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index e68850dc5..ebe5536de 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/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.ReadBlockUnsafe(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,
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 3b2f6aab6..35d699bbf 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -59,6 +59,11 @@ public:
         buffer_cache.AccumulateFlushes();
     }
 
+    void SignalReference() {
+        std::function<void()> do_nothing([] {});
+        SignalFence(std::move(do_nothing));
+    }
+
     void SyncOperation(std::function<void()>&& func) {
         uncommitted_operations.emplace_back(std::move(func));
     }
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 2e7f9c5ed..295a416a8 100644
--- a/src/video_core/gpu.cpp
+++ b/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);
 }
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 8a871593a..e49c40cf2 100644
--- a/src/video_core/gpu.h
+++ b/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;
@@ -241,6 +242,9 @@ public:
     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);
 
     /// Notify rasterizer that any caches of the specified region should be invalidated
diff --git a/src/video_core/query_cache.h b/src/video_core/query_cache.h
index 941de95c1..1528cc1dd 100644
--- a/src/video_core/query_cache.h
+++ b/src/video_core/query_cache.h
@@ -255,7 +255,6 @@ private:
                 if (!in_range(query)) {
                     continue;
                 }
-                rasterizer.UpdatePagesCachedCount(query.GetCpuAddr(), query.SizeInBytes(), -1);
                 AsyncJobId async_job_id = query.GetAsyncJob();
                 auto flush_result = query.Flush(async);
                 if (async_job_id == NULL_ASYNC_JOB_ID) {
@@ -273,7 +272,6 @@ private:
 
     /// Registers the passed parameters as cached and returns a pointer to the stored cached query.
     CachedQuery* Register(VideoCore::QueryType type, VAddr cpu_addr, u8* host_ptr, bool timestamp) {
-        rasterizer.UpdatePagesCachedCount(cpu_addr, CachedQuery::SizeInBytes(timestamp), 1);
         const u64 page = static_cast<u64>(cpu_addr) >> YUZU_PAGEBITS;
         return &cached_queries[page].emplace_back(static_cast<QueryCache&>(*this), type, cpu_addr,
                                                   host_ptr);
diff --git a/src/video_core/rasterizer_download_area.h b/src/video_core/rasterizer_download_area.h
new file mode 100644
index 000000000..2d7425c79
--- /dev/null
+++ b/src/video_core/rasterizer_download_area.h
@@ -0,0 +1,16 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "common/common_types.h"
+
+namespace VideoCore {
+
+struct RasterizerDownloadArea {
+    VAddr start_address;
+    VAddr end_address;
+    bool preemtive;
+};
+
+} // namespace VideoCore
\ No newline at end of file
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 33e2610bc..7566a8c4e 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/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;
diff --git a/src/video_core/renderer_null/null_rasterizer.cpp b/src/video_core/renderer_null/null_rasterizer.cpp
index 2b5c7defa..bf2ce4c49 100644
--- a/src/video_core/renderer_null/null_rasterizer.cpp
+++ b/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) {}
diff --git a/src/video_core/renderer_null/null_rasterizer.h b/src/video_core/renderer_null/null_rasterizer.h
index 0c59e6a1f..a8d35d2c1 100644
--- a/src/video_core/renderer_null/null_rasterizer.h
+++ b/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;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 0089b4b27..f5baa0f3c 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/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) {
@@ -1281,7 +1304,7 @@ bool AccelerateDMA::DmaBufferImageCopy(const Tegra::DMA::ImageCopy& copy_info,
                                        const Tegra::DMA::BufferOperand& buffer_operand,
                                        const Tegra::DMA::ImageOperand& image_operand) {
     std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
-    const auto image_id = texture_cache.DmaImageId(image_operand);
+    const auto image_id = texture_cache.DmaImageId(image_operand, IS_IMAGE_UPLOAD);
     if (image_id == VideoCommon::NULL_IMAGE_ID) {
         return false;
     }
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index ad6978bd0..410d8ffc5 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/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;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index d1489fc95..628e1376f 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -502,6 +502,22 @@ 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;
+        }
+    }
+    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;
@@ -598,7 +614,7 @@ void RasterizerVulkan::SignalSyncPoint(u32 value) {
 }
 
 void RasterizerVulkan::SignalReference() {
-    fence_manager.SignalOrdering();
+    fence_manager.SignalReference();
 }
 
 void RasterizerVulkan::ReleaseFences() {
@@ -631,7 +647,7 @@ void RasterizerVulkan::WaitForIdle() {
         cmdbuf.SetEvent(event, flags);
         cmdbuf.WaitEvents(event, flags, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, {}, {}, {});
     });
-    SignalReference();
+    fence_manager.SignalOrdering();
 }
 
 void RasterizerVulkan::FragmentBarrier() {
@@ -777,7 +793,7 @@ bool AccelerateDMA::DmaBufferImageCopy(const Tegra::DMA::ImageCopy& copy_info,
                                        const Tegra::DMA::BufferOperand& buffer_operand,
                                        const Tegra::DMA::ImageOperand& image_operand) {
     std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
-    const auto image_id = texture_cache.DmaImageId(image_operand);
+    const auto image_id = texture_cache.DmaImageId(image_operand, IS_IMAGE_UPLOAD);
     if (image_id == VideoCommon::NULL_IMAGE_ID) {
         return false;
     }
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 1659fbc13..9bd422850 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/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;
diff --git a/src/video_core/texture_cache/image_info.cpp b/src/video_core/texture_cache/image_info.cpp
index 11f3f78a1..e8ddde691 100644
--- a/src/video_core/texture_cache/image_info.cpp
+++ b/src/video_core/texture_cache/image_info.cpp
@@ -4,6 +4,7 @@
 #include <fmt/format.h>
 
 #include "common/assert.h"
+#include "common/settings.h"
 #include "video_core/surface.h"
 #include "video_core/texture_cache/format_lookup_table.h"
 #include "video_core/texture_cache/image_info.h"
@@ -22,6 +23,8 @@ using VideoCore::Surface::PixelFormat;
 using VideoCore::Surface::SurfaceType;
 
 ImageInfo::ImageInfo(const TICEntry& config) noexcept {
+    forced_flushed = config.IsPitchLinear() && !Settings::values.use_reactive_flushing.GetValue();
+    dma_downloaded = forced_flushed;
     format = PixelFormatFromTextureInfo(config.format, config.r_type, config.g_type, config.b_type,
                                         config.a_type, config.srgb_conversion);
     num_samples = NumSamples(config.msaa_mode);
@@ -117,6 +120,9 @@ ImageInfo::ImageInfo(const TICEntry& config) noexcept {
 
 ImageInfo::ImageInfo(const Maxwell3D::Regs::RenderTargetConfig& ct,
                      Tegra::Texture::MsaaMode msaa_mode) noexcept {
+    forced_flushed =
+        ct.tile_mode.is_pitch_linear && !Settings::values.use_reactive_flushing.GetValue();
+    dma_downloaded = forced_flushed;
     format = VideoCore::Surface::PixelFormatFromRenderTargetFormat(ct.format);
     rescaleable = false;
     if (ct.tile_mode.is_pitch_linear) {
@@ -155,6 +161,9 @@ ImageInfo::ImageInfo(const Maxwell3D::Regs::RenderTargetConfig& ct,
 
 ImageInfo::ImageInfo(const Maxwell3D::Regs::Zeta& zt, const Maxwell3D::Regs::ZetaSize& zt_size,
                      Tegra::Texture::MsaaMode msaa_mode) noexcept {
+    forced_flushed =
+        zt.tile_mode.is_pitch_linear && !Settings::values.use_reactive_flushing.GetValue();
+    dma_downloaded = forced_flushed;
     format = VideoCore::Surface::PixelFormatFromDepthFormat(zt.format);
     size.width = zt_size.width;
     size.height = zt_size.height;
@@ -195,6 +204,9 @@ ImageInfo::ImageInfo(const Maxwell3D::Regs::Zeta& zt, const Maxwell3D::Regs::Zet
 
 ImageInfo::ImageInfo(const Fermi2D::Surface& config) noexcept {
     UNIMPLEMENTED_IF_MSG(config.layer != 0, "Surface layer is not zero");
+    forced_flushed = config.linear == Fermi2D::MemoryLayout::Pitch &&
+                     !Settings::values.use_reactive_flushing.GetValue();
+    dma_downloaded = forced_flushed;
     format = VideoCore::Surface::PixelFormatFromRenderTargetFormat(config.format);
     rescaleable = false;
     if (config.linear == Fermi2D::MemoryLayout::Pitch) {
diff --git a/src/video_core/texture_cache/image_info.h b/src/video_core/texture_cache/image_info.h
index 4b7dfa315..8a4cb0cbd 100644
--- a/src/video_core/texture_cache/image_info.h
+++ b/src/video_core/texture_cache/image_info.h
@@ -39,6 +39,8 @@ struct ImageInfo {
     u32 tile_width_spacing = 0;
     bool rescaleable = false;
     bool downscaleable = false;
+    bool forced_flushed = false;
+    bool dma_downloaded = false;
 };
 
 } // namespace VideoCommon
diff --git a/src/video_core/texture_cache/image_view_base.cpp b/src/video_core/texture_cache/image_view_base.cpp
index bcad40353..d134b6738 100644
--- a/src/video_core/texture_cache/image_view_base.cpp
+++ b/src/video_core/texture_cache/image_view_base.cpp
@@ -4,7 +4,6 @@
 #include <algorithm>
 
 #include "common/assert.h"
-#include "common/settings.h"
 #include "video_core/compatible_formats.h"
 #include "video_core/surface.h"
 #include "video_core/texture_cache/formatter.h"
@@ -26,8 +25,7 @@ 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) {
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index b5297e76b..e1198dcf8 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -491,6 +491,32 @@ 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;
     ForEachImageInRegion(cpu_addr, size, [&](ImageId id, Image&) { deleted_images.push_back(id); });
@@ -683,6 +709,7 @@ void TextureCache<P>::CommitAsyncFlushes() {
                 download_info.async_buffer_id = last_async_buffer_id;
             }
         }
+
         if (any_none_dma) {
             auto download_map = runtime.DownloadStagingBuffer(total_size_bytes, true);
             for (const PendingDownload& download_info : download_ids) {
@@ -695,6 +722,7 @@ void TextureCache<P>::CommitAsyncFlushes() {
             }
             uncommitted_async_buffers.emplace_back(download_map);
         }
+
         async_buffers.emplace_back(std::move(uncommitted_async_buffers));
         uncommitted_async_buffers.clear();
     }
@@ -783,17 +811,22 @@ void TextureCache<P>::PopAsyncFlushes() {
 }
 
 template <class P>
-ImageId TextureCache<P>::DmaImageId(const Tegra::DMA::ImageOperand& operand) {
+ImageId TextureCache<P>::DmaImageId(const Tegra::DMA::ImageOperand& operand, bool is_upload) {
     const ImageInfo dst_info(operand);
     const ImageId dst_id = FindDMAImage(dst_info, operand.address);
     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 (!is_upload && !image.info.dma_downloaded) {
+        // Force a full sync.
+        image.info.dma_downloaded = true;
+        return NULL_IMAGE_ID;
+    }
     const auto base = image.TryFindBase(operand.address);
     if (!base) {
         return NULL_IMAGE_ID;
@@ -1290,7 +1323,6 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
             all_siblings.push_back(overlap_id);
         } else {
             bad_overlap_ids.push_back(overlap_id);
-            overlap.flags |= ImageFlagBits::BadOverlap;
         }
     };
     ForEachImageInRegion(cpu_addr, size_bytes, region_check);
@@ -1401,7 +1433,12 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
         ImageBase& aliased = slot_images[aliased_id];
         aliased.overlapping_images.push_back(new_image_id);
         new_image.overlapping_images.push_back(aliased_id);
-        new_image.flags |= ImageFlagBits::BadOverlap;
+        if (aliased.info.resources.levels == 1 && aliased.overlapping_images.size() > 1) {
+            aliased.flags |= ImageFlagBits::BadOverlap;
+        }
+        if (new_image.info.resources.levels == 1 && new_image.overlapping_images.size() > 1) {
+            new_image.flags |= ImageFlagBits::BadOverlap;
+        }
     }
     RegisterImage(new_image_id);
     return new_image_id;
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index 758b7e212..0720494e5 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/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);
 
@@ -205,7 +207,7 @@ public:
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
 
-    [[nodiscard]] ImageId DmaImageId(const Tegra::DMA::ImageOperand& operand);
+    [[nodiscard]] ImageId DmaImageId(const Tegra::DMA::ImageOperand& operand, bool is_upload);
 
     [[nodiscard]] std::pair<Image*, BufferImageCopy> DmaBufferImageCopy(
         const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& buffer_operand,