gl_buffer_cache: Enable async downloads

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -146,8 +146,12 @@ StagingBufferMap BufferCacheRuntime::UploadStagingBuffer(size_t size) {
     return staging_buffer_pool.RequestUploadBuffer(size);
 }
 
-StagingBufferMap BufferCacheRuntime::DownloadStagingBuffer(size_t size) {
-    return staging_buffer_pool.RequestDownloadBuffer(size);
+StagingBufferMap BufferCacheRuntime::DownloadStagingBuffer(size_t size, bool deferred) {
+    return staging_buffer_pool.RequestDownloadBuffer(size, deferred);
+}
+
+void BufferCacheRuntime::FreeDeferredStagingBuffer(StagingBufferMap& buffer) {
+    staging_buffer_pool.FreeDeferredStagingBuffer(buffer.index);
 }
 
 u64 BufferCacheRuntime::GetDeviceMemoryUsage() const {

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -64,7 +64,9 @@ public:
 
     [[nodiscard]] StagingBufferMap UploadStagingBuffer(size_t size);
 
-    [[nodiscard]] StagingBufferMap DownloadStagingBuffer(size_t size);
+    [[nodiscard]] StagingBufferMap DownloadStagingBuffer(size_t size, bool deferred = false);
+
+    void FreeDeferredStagingBuffer(StagingBufferMap& buffer);
 
     void CopyBuffer(GLuint dst_buffer, GLuint src_buffer,
                     std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
@@ -233,7 +235,7 @@ struct BufferCacheParams {
     static constexpr bool NEEDS_BIND_STORAGE_INDEX = true;
     static constexpr bool USE_MEMORY_MAPS = true;
     static constexpr bool SEPARATE_IMAGE_BUFFER_BINDINGS = true;
-    static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = false;
+    static constexpr bool IMPLEMENTS_ASYNC_DOWNLOADS = true;
 
     // TODO: Investigate why OpenGL seems to perform worse with persistently mapped buffer uploads
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = false;

src/video_core/renderer_opengl/gl_staging_buffer_pool.cpp

@@ -28,19 +28,26 @@ StagingBuffers::StagingBuffers(GLenum storage_flags_, GLenum map_flags_)
 
 StagingBuffers::~StagingBuffers() = default;
 
-StagingBufferMap StagingBuffers::RequestMap(size_t requested_size, bool insert_fence) {
+StagingBufferMap StagingBuffers::RequestMap(size_t requested_size, bool insert_fence,
+                                            bool deferred) {
     MICROPROFILE_SCOPE(OpenGL_BufferRequest);
 
     const size_t index = RequestBuffer(requested_size);
     OGLSync* const sync = insert_fence ? &allocs[index].sync : nullptr;
     allocs[index].sync_index = insert_fence ? ++current_sync_index : 0;
+    allocs[index].deferred = deferred;
     return StagingBufferMap{
         .mapped_span = std::span(allocs[index].map, requested_size),
         .sync = sync,
         .buffer = allocs[index].buffer.handle,
+        .index = index,
     };
 }
 
+void StagingBuffers::FreeDeferredStagingBuffer(size_t index) {
+    allocs[index].deferred = false;
+}
+
 size_t StagingBuffers::RequestBuffer(size_t requested_size) {
     if (const std::optional<size_t> index = FindBuffer(requested_size); index) {
         return *index;
@@ -68,6 +75,9 @@ std::optional<size_t> StagingBuffers::FindBuffer(size_t requested_size) {
         if (buffer_size < requested_size || buffer_size >= smallest_buffer) {
             continue;
         }
+        if (alloc.deferred) {
+            continue;
+        }
         if (alloc.sync.handle != 0) {
             if (alloc.sync_index >= known_unsignaled_index) {
                 // This fence is later than a fence that is known to not be signaled
@@ -138,8 +148,12 @@ StagingBufferMap StagingBufferPool::RequestUploadBuffer(size_t size) {
     return upload_buffers.RequestMap(size, true);
 }
 
-StagingBufferMap StagingBufferPool::RequestDownloadBuffer(size_t size) {
-    return download_buffers.RequestMap(size, false);
+StagingBufferMap StagingBufferPool::RequestDownloadBuffer(size_t size, bool deferred) {
+    return download_buffers.RequestMap(size, false, deferred);
+}
+
+void StagingBufferPool::FreeDeferredStagingBuffer(size_t index) {
+    download_buffers.FreeDeferredStagingBuffer(index);
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_staging_buffer_pool.h

@@ -26,13 +26,16 @@ struct StagingBufferMap {
     size_t offset = 0;
     OGLSync* sync;
     GLuint buffer;
+    size_t index;
 };
 
 struct StagingBuffers {
     explicit StagingBuffers(GLenum storage_flags_, GLenum map_flags_);
     ~StagingBuffers();
 
-    StagingBufferMap RequestMap(size_t requested_size, bool insert_fence);
+    StagingBufferMap RequestMap(size_t requested_size, bool insert_fence, bool deferred = false);
+
+    void FreeDeferredStagingBuffer(size_t index);
 
     size_t RequestBuffer(size_t requested_size);
 
@@ -44,6 +47,7 @@ struct StagingBuffers {
         u8* map;
         size_t size;
         size_t sync_index;
+        bool deferred;
     };
     std::vector<StagingBufferAlloc> allocs;
     GLenum storage_flags;
@@ -88,7 +92,8 @@ public:
     ~StagingBufferPool() = default;
 
     StagingBufferMap RequestUploadBuffer(size_t size);
-    StagingBufferMap RequestDownloadBuffer(size_t size);
+    StagingBufferMap RequestDownloadBuffer(size_t size, bool deferred = false);
+    void FreeDeferredStagingBuffer(size_t index);
 
 private:
     StagingBuffers upload_buffers{GL_MAP_WRITE_BIT, GL_MAP_WRITE_BIT | GL_MAP_FLUSH_EXPLICIT_BIT};