Merge branch 'master' into ssbo-align

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_opengl/gl_graphics_pipeline.cpp

@@ -559,7 +559,9 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
 }
 
 void GraphicsPipeline::ConfigureTransformFeedbackImpl() const {
-    glTransformFeedbackAttribsNV(num_xfb_attribs, xfb_attribs.data(), GL_SEPARATE_ATTRIBS);
+    const GLenum buffer_mode =
+        num_xfb_buffers_active == 1 ? GL_INTERLEAVED_ATTRIBS : GL_SEPARATE_ATTRIBS;
+    glTransformFeedbackAttribsNV(num_xfb_attribs, xfb_attribs.data(), buffer_mode);
 }
 
 void GraphicsPipeline::GenerateTransformFeedbackState() {
@@ -567,12 +569,14 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
     // when this is required.
     GLint* cursor{xfb_attribs.data()};
 
+    num_xfb_buffers_active = 0;
     for (size_t feedback = 0; feedback < Maxwell::NumTransformFeedbackBuffers; ++feedback) {
         const auto& layout = key.xfb_state.layouts[feedback];
         UNIMPLEMENTED_IF_MSG(layout.stride != layout.varying_count * 4, "Stride padding");
         if (layout.varying_count == 0) {
             continue;
         }
+        num_xfb_buffers_active++;
 
         const auto& locations = key.xfb_state.varyings[feedback];
         std::optional<u32> current_index;

src/video_core/renderer_opengl/gl_graphics_pipeline.h

@@ -154,6 +154,7 @@ private:
 
     static constexpr std::size_t XFB_ENTRY_STRIDE = 3;
     GLsizei num_xfb_attribs{};
+    u32 num_xfb_buffers_active{};
     std::array<GLint, 128 * XFB_ENTRY_STRIDE * Maxwell::NumTransformFeedbackBuffers> xfb_attribs{};
 
     std::mutex built_mutex;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -555,7 +555,7 @@ void RasterizerOpenGL::OnCacheInvalidation(VAddr addr, u64 size) {
     }
     {
         std::scoped_lock lock{buffer_cache.mutex};
-        buffer_cache.CachedWriteMemory(addr, size);
+        buffer_cache.WriteMemory(addr, size);
     }
     shader_cache.InvalidateRegion(addr, size);
 }