Fence Manager: Force ordering on WFI.

src/video_core/buffer_cache/buffer_cache.h

@@ -152,6 +152,8 @@ public:
     /// Return true when there are uncommitted buffers to be downloaded
     [[nodiscard]] bool HasUncommittedFlushes() const noexcept;
 
+    void AccumulateFlushes();
+
     /// Return true when the caller should wait for async downloads
     [[nodiscard]] bool ShouldWaitAsyncFlushes() const noexcept;
 
@@ -334,6 +336,7 @@ private:
     std::vector<BufferId> cached_write_buffer_ids;
 
     IntervalSet uncommitted_ranges;
+    std::deque<IntervalSet> committed_ranges;
 
     size_t immediate_buffer_capacity = 0;
     std::unique_ptr<u8[]> immediate_buffer_alloc;
@@ -551,7 +554,19 @@ void BufferCache<P>::FlushCachedWrites() {
 
 template <class P>
 bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
-    return !uncommitted_ranges.empty();
+    return !uncommitted_ranges.empty() || !committed_ranges.empty();
+}
+
+template <class P>
+void BufferCache<P>::AccumulateFlushes() {
+    if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) {
+        uncommitted_ranges.clear();
+        return;
+    }
+    if (uncommitted_ranges.empty()) {
+        return;
+    }
+    committed_ranges.emplace_back(std::move(uncommitted_ranges));
 }
 
 template <class P>
@@ -561,8 +576,8 @@ bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
 
 template <class P>
 void BufferCache<P>::CommitAsyncFlushesHigh() {
-    const IntervalSet& intervals = uncommitted_ranges;
+    AccumulateFlushes();
-    if (intervals.empty()) {
+    if (committed_ranges.empty()) {
         return;
     }
     MICROPROFILE_SCOPE(GPU_DownloadMemory);
@@ -570,6 +585,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
     boost::container::small_vector<std::pair<BufferCopy, BufferId>, 1> downloads;
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
+    for (const IntervalSet& intervals : committed_ranges) {
         for (auto& interval : intervals) {
             const std::size_t size = interval.upper() - interval.lower();
             const VAddr cpu_addr = interval.lower();
@@ -607,6 +623,8 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
                     });
             });
         }
+    }
+    committed_ranges.clear();
     if (downloads.empty()) {
         return;
     }
@@ -644,6 +662,7 @@ void BufferCache<P>::CommitAsyncFlushes() {
         CommitAsyncFlushesHigh();
     } else {
         uncommitted_ranges.clear();
+        committed_ranges.clear();
     }
 }
 

src/video_core/fence_manager.h

@@ -54,6 +54,12 @@ public:
         delayed_destruction_ring.Tick();
     }
 
+    // Unlike other fences, this one doesn't
+    void SignalOrdering() {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.AccumulateFlushes();
+    }
+
     void SignalReference() {
         // Only sync references on High
         if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) {

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -657,6 +657,10 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
 
 void RasterizerOpenGL::WaitForIdle() {
     glMemoryBarrier(GL_ALL_BARRIER_BITS);
+    if (!gpu.IsAsync()) {
+        return;
+    }
+    fence_manager.SignalOrdering();
 }
 
 void RasterizerOpenGL::FragmentBarrier() {

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -619,6 +619,10 @@ void RasterizerVulkan::WaitForIdle() {
         cmdbuf.SetEvent(event, flags);
         cmdbuf.WaitEvents(event, flags, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, {}, {}, {});
     });
+    if (!gpu.IsAsync()) {
+        return;
+    }
+    fence_manager.SignalOrdering();
 }
 
 void RasterizerVulkan::FragmentBarrier() {