diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalCommon.h b/src/Cafe/HW/Latte/Renderer/Metal/MetalCommon.h
index 44d4d873..a47ecf9a 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/MetalCommon.h
+++ b/src/Cafe/HW/Latte/Renderer/Metal/MetalCommon.h
@@ -60,3 +60,5 @@ inline NS::String* GetLabel(const std::string& label, const void* identifier)
 {
     return ToNSString(label + " (" + std::to_string(reinterpret_cast<uintptr_t>(identifier)) + ")");
 }
+
+constexpr MTL::RenderStages ALL_MTL_RENDER_STAGES = MTL::RenderStageVertex | MTL::RenderStageObject | MTL::RenderStageMesh | MTL::RenderStageFragment;
diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.cpp b/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.cpp
index 07bd2a22..6173532c 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.cpp
@@ -139,7 +139,7 @@ void MetalMemoryManager::UploadToBufferCache(const void* data, size_t offset, si
     // Lock the buffer to make sure it's not deallocated before the copy is done
     m_tempBufferAllocator.LockBuffer(allocation.bufferIndex);
 
-    m_mtlr->CopyBufferToBuffer(buffer, allocation.offset, m_bufferCache, offset, size);
+    m_mtlr->CopyBufferToBuffer(buffer, allocation.offset, m_bufferCache, offset, size, ALL_MTL_RENDER_STAGES, ALL_MTL_RENDER_STAGES);
 
     // Make sure the buffer has the right command buffer
     m_tempBufferAllocator.GetBuffer(allocation.bufferIndex); // TODO: make a helper function for this
@@ -155,5 +155,5 @@ void MetalMemoryManager::CopyBufferCache(size_t srcOffset, size_t dstOffset, siz
 {
     cemu_assert_debug(m_bufferCache);
 
-    m_mtlr->CopyBufferToBuffer(m_bufferCache, srcOffset, m_bufferCache, dstOffset, size);
+    m_mtlr->CopyBufferToBuffer(m_bufferCache, srcOffset, m_bufferCache, dstOffset, size, ALL_MTL_RENDER_STAGES, ALL_MTL_RENDER_STAGES);
 }
diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp b/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp
index 59b18ac5..fbee6f3b 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp
@@ -20,6 +20,7 @@
 #include "HW/Latte/Renderer/Metal/MetalCommon.h"
 #include "HW/Latte/Renderer/Metal/MetalLayerHandle.h"
 #include "HW/Latte/Renderer/Renderer.h"
+#include "Metal/MTLRenderCommandEncoder.hpp"
 #include "imgui.h"
 
 #define IMGUI_IMPL_METAL_CPP
@@ -780,7 +781,7 @@ void MetalRenderer::bufferCache_copy(uint32 srcOffset, uint32 dstOffset, uint32
 
 void MetalRenderer::bufferCache_copyStreamoutToMainBuffer(uint32 srcOffset, uint32 dstOffset, uint32 size)
 {
-    CopyBufferToBuffer(m_xfbRingBuffer, srcOffset, m_memoryManager->GetBufferCache(), dstOffset, size);
+    CopyBufferToBuffer(m_xfbRingBuffer, srcOffset, m_memoryManager->GetBufferCache(), dstOffset, size, MTL::RenderStageVertex, ALL_MTL_RENDER_STAGES);
 }
 
 void MetalRenderer::buffer_bindVertexBuffer(uint32 bufferIndex, uint32 offset, uint32 size)
@@ -1834,7 +1835,7 @@ void MetalRenderer::ClearColorTextureInternal(MTL::Texture* mtlTexture, sint32 s
     EndEncoding();
 }
 
-void MetalRenderer::CopyBufferToBuffer(MTL::Buffer* src, uint32 srcOffset, MTL::Buffer* dst, uint32 dstOffset, uint32 size)
+void MetalRenderer::CopyBufferToBuffer(MTL::Buffer* src, uint32 srcOffset, MTL::Buffer* dst, uint32 dstOffset, uint32 size, MTL::RenderStages after, MTL::RenderStages before)
 {
     // Do the copy in a vertex shader on Apple GPUs
     if (m_isAppleGPU && m_encoderType == MetalEncoderType::Render)
@@ -1842,8 +1843,7 @@ void MetalRenderer::CopyBufferToBuffer(MTL::Buffer* src, uint32 srcOffset, MTL::
         auto renderCommandEncoder = static_cast<MTL::RenderCommandEncoder*>(m_commandEncoder);
 
 		MTL::Resource* barrierBuffers[] = {src};
-		// TODO: let the caller choose the stages
-        renderCommandEncoder->memoryBarrier(barrierBuffers, 1, MTL::RenderStageVertex | MTL::RenderStageFragment | MTL::RenderStageObject | MTL::RenderStageMesh, MTL::RenderStageVertex);
+        renderCommandEncoder->memoryBarrier(barrierBuffers, 1, after, MTL::RenderStageVertex);
 
 		renderCommandEncoder->setRenderPipelineState(m_copyBufferToBufferPipeline->GetRenderPipelineState());
 		m_state.m_encoderState.m_renderPipelineState = m_copyBufferToBufferPipeline->GetRenderPipelineState();
@@ -1854,8 +1854,7 @@ void MetalRenderer::CopyBufferToBuffer(MTL::Buffer* src, uint32 srcOffset, MTL::
 		renderCommandEncoder->drawPrimitives(MTL::PrimitiveTypePoint, NS::UInteger(0), NS::UInteger(size));
 
 		barrierBuffers[0] = dst;
-		// TODO: let the caller choose the stages
-        renderCommandEncoder->memoryBarrier(barrierBuffers, 1, MTL::RenderStageVertex, MTL::RenderStageVertex | MTL::RenderStageFragment | MTL::RenderStageObject | MTL::RenderStageMesh);
+        renderCommandEncoder->memoryBarrier(barrierBuffers, 1, MTL::RenderStageVertex, before);
     }
     else
     {
diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h b/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h
index 3281ccf2..25051a97 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h
+++ b/src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h
@@ -377,7 +377,7 @@ public:
 
     void ClearColorTextureInternal(MTL::Texture* mtlTexture, sint32 sliceIndex, sint32 mipIndex, float r, float g, float b, float a);
 
-    void CopyBufferToBuffer(MTL::Buffer* src, uint32 srcOffset, MTL::Buffer* dst, uint32 dstOffset, uint32 size);
+    void CopyBufferToBuffer(MTL::Buffer* src, uint32 srcOffset, MTL::Buffer* dst, uint32 dstOffset, uint32 size, MTL::RenderStages after, MTL::RenderStages before);
 
     // Getters
     bool IsAppleGPU() const