do vertex restride on the GPU & don't over-sync

cemu-project · Sep 23, 2024 · 3cf831d · 3cf831d
1 parent 8b68df0
commit 3cf831d
Show file tree

Hide file tree

Showing 3 changed files with 14 additions and 12 deletions.
diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.cpp b/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.cpp
@@ -9,7 +9,7 @@ MetalVertexBufferCache::~MetalVertexBufferCache()
 {
 }
 
-MetalRestridedBufferRange MetalVertexBufferCache::RestrideBufferIfNeeded(MTL::Buffer* bufferCache, uint32 bufferIndex, size_t stride)
+MetalRestridedBufferRange MetalVertexBufferCache::RestrideBufferIfNeeded(MTL::Buffer* bufferCache, uint32 bufferIndex, size_t stride, std::vector<MTL::Resource*>& barrierBuffers)
 {
     auto vertexBufferRange = m_bufferRanges[bufferIndex];
     auto& restrideInfo = *vertexBufferRange.restrideInfo;
@@ -28,14 +28,14 @@ MetalRestridedBufferRange MetalVertexBufferCache::RestrideBufferIfNeeded(MTL::Bu
         restrideInfo.allocation = m_bufferAllocator.GetBufferAllocation(newSize);
         buffer = m_bufferAllocator.GetBuffer(restrideInfo.allocation.bufferIndex);
 
-        // HACK: the restriding is done on the CPU, since doing it on the GPU was causing over-synchronization
+        /*
         uint8* oldPtr = (uint8*)bufferCache->contents() + vertexBufferRange.offset;
         uint8* newPtr = (uint8*)buffer->contents() + restrideInfo.allocation.offset;
 
         for (size_t elem = 0; elem < vertexBufferRange.size / stride; elem++)
     		memcpy(newPtr + elem * newStride, oldPtr + elem * stride, stride);
+        */
 
-        /*
         if (m_mtlr->GetEncoderType() == MetalEncoderType::Render)
         {
             auto renderCommandEncoder = static_cast<MTL::RenderCommandEncoder*>(m_mtlr->GetCommandEncoder());
@@ -56,16 +56,12 @@ MetalRestridedBufferRange MetalVertexBufferCache::RestrideBufferIfNeeded(MTL::Bu
 
             renderCommandEncoder->drawPrimitives(MTL::PrimitiveTypeTriangleStrip, NS::UInteger(0), vertexBufferRange.size / stride);
 
-            // TODO: do the barriers in one call?
-            MTL::Resource* barrierBuffers[] = {buffer};
-            renderCommandEncoder->memoryBarrier(barrierBuffers, 1, MTL::RenderStageVertex, MTL::RenderStageVertex);
+            vectorAppendUnique(barrierBuffers, static_cast<MTL::Resource*>(buffer));
         }
         else
         {
-            debug_printf("vertex buffer restride needs an active render command encoder\n");
             cemu_assert_suspicious();
         }
-        */
 
         restrideInfo.memoryInvalidated = false;
         restrideInfo.lastStride = newStride;

diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.h b/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.h
@@ -42,7 +42,7 @@ class MetalVertexBufferCache
         range.offset = INVALID_OFFSET;
     }
 
-    MetalRestridedBufferRange RestrideBufferIfNeeded(MTL::Buffer* bufferCache, uint32 bufferIndex, size_t stride);
+    MetalRestridedBufferRange RestrideBufferIfNeeded(MTL::Buffer* bufferCache, uint32 bufferIndex, size_t stride, std::vector<MTL::Resource*>& barrierBuffers);
 
 private:
     class MetalRenderer* m_mtlr;
@@ -105,9 +105,9 @@ class MetalMemoryManager
         m_vertexBufferCache.UntrackVertexBuffer(bufferIndex);
     }
 
-    MetalRestridedBufferRange RestrideBufferIfNeeded(uint32 bufferIndex, size_t stride)
+    MetalRestridedBufferRange RestrideBufferIfNeeded(uint32 bufferIndex, size_t stride, std::vector<MTL::Resource*>& barrierBuffers)
     {
-        return m_vertexBufferCache.RestrideBufferIfNeeded(m_bufferCache, bufferIndex, stride);
+        return m_vertexBufferCache.RestrideBufferIfNeeded(m_bufferCache, bufferIndex, stride, barrierBuffers);
     }
 
 private:

diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp b/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp
@@ -1141,6 +1141,7 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 	// Resources
 
 	// Vertex buffers
+	std::vector<MTL::Resource*> barrierBuffers;
     for (uint8 i = 0; i < MAX_MTL_BUFFERS; i++)
     {
         auto& vertexBufferRange = m_state.m_vertexBuffers[i];
@@ -1161,7 +1162,7 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
                 uint32 bufferBaseRegisterIndex = mmSQ_VTX_ATTRIBUTE_BLOCK_START + i * 7;
                 uint32 bufferStride = (LatteGPUState.contextNew.GetRawView()[bufferBaseRegisterIndex + 2] >> 11) & 0xFFFF;
 
-                auto restridedBuffer = m_memoryManager->RestrideBufferIfNeeded(i, bufferStride);
+                auto restridedBuffer = m_memoryManager->RestrideBufferIfNeeded(i, bufferStride, barrierBuffers);
 
                 buffer = restridedBuffer.buffer;
                 offset = restridedBuffer.offset;
@@ -1172,6 +1173,11 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
         }
     }
 
+    if (!barrierBuffers.empty())
+    {
+        renderCommandEncoder->memoryBarrier(barrierBuffers.data(), barrierBuffers.size(), MTL::RenderStageVertex, MTL::RenderStageVertex);
+    }
+
 	// Render pipeline state
 	MTL::RenderPipelineState* renderPipelineState;
 	if (usesGeometryShader)