GS/VK/GL/DX12/DX11: Use default buffer instead of upload buffer for accurate prims data.

Should hopefully give better performance.

Also refactor some upload/staging buffer handling in VK/DX12.

Author: TJnotJT

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -396,8 +396,12 @@ bool GSDevice11::Create(GSVSyncMode vsync_mode, bool allow_present_throttle)
 		}
 	}
 
+	bd = {};
+
 	if (m_features.accurate_prims)
 	{
+		bd.Usage = D3D11_USAGE_DEFAULT;
+		bd.CPUAccessFlags = 0;
 		bd.ByteWidth = ACCURATE_PRIMS_BUFFER_SIZE;
 		bd.BindFlags = D3D11_BIND_SHADER_RESOURCE;
 		bd.StructureByteStride = sizeof(AccuratePrimsEdgeData);
@@ -410,16 +414,19 @@ bool GSDevice11::Create(GSVSyncMode vsync_mode, bool allow_present_throttle)
 		}
 
 		const CD3D11_SHADER_RESOURCE_VIEW_DESC accurate_prims_b_srv_desc(
-			D3D11_SRV_DIMENSION_BUFFER, DXGI_FORMAT_UNKNOWN, 0, ACCURATE_PRIMS_BUFFER_SIZE / sizeof(AccuratePrimsEdgeData));
-		if (FAILED(m_dev->CreateShaderResourceView(m_accurate_prims_b.get(), &accurate_prims_b_srv_desc, m_accurate_prims_b_srv.put())))
+			D3D11_SRV_DIMENSION_BUFFER, DXGI_FORMAT_UNKNOWN, 0,
+			ACCURATE_PRIMS_BUFFER_SIZE / sizeof(AccuratePrimsEdgeData));
+		
+		if (FAILED(m_dev->CreateShaderResourceView(m_accurate_prims_b.get(), &accurate_prims_b_srv_desc,
+		m_accurate_prims_b_srv.put())))
 		{
 			Console.Error("D3D11: Failed to create accurate prims buffer SRV.");
 			return false;
 		}
 
 		// If MAX_TEXTURES changes, please change the register for this buffer in the shader.
 		static_assert(MAX_TEXTURES == 5);
-		m_ctx->PSSetShaderResources(MAX_TEXTURES, 1, m_accurate_prims_b_srv.addressof());
+		m_ctx->PSSetShaderResources(5, 1, m_accurate_prims_b_srv.addressof());
 	}
 
 	// rasterizer
@@ -2326,29 +2333,18 @@ bool GSDevice11::SetupAccuratePrims(GSHWDrawConfig& config)
 		if (size > ACCURATE_PRIMS_BUFFER_SIZE)
 			return false;
 
-		D3D11_MAP type = D3D11_MAP_WRITE_NO_OVERWRITE;
-
-		pxAssert(m_accurate_prims_b_pos % sizeof(AccuratePrimsEdgeData) == 0);
-
-		if (m_accurate_prims_b_pos + size > ACCURATE_PRIMS_BUFFER_SIZE)
-		{
-			m_accurate_prims_b_pos = 0;
-			type = D3D11_MAP_WRITE_DISCARD;
-		}
-
-		D3D11_MAPPED_SUBRESOURCE m;
-		if (FAILED(m_ctx->Map(m_accurate_prims_b.get(), 0, type, 0, &m)))
-			return false;
-
-		void* map = static_cast<u8*>(m.pData) + m_accurate_prims_b_pos;
-
-		GSVector4i::storent(map, config.accurate_prims_edge_data->data(), size);
-
-		m_ctx->Unmap(m_accurate_prims_b.get(), 0);
-
-		config.cb_ps.accurate_prims_base_index.x = m_accurate_prims_b_pos / sizeof(AccuratePrimsEdgeData);
-		
-		m_accurate_prims_b_pos += size;
+		// Performance note: UpdateSubresource() copies data to a temp staging buffer to avoid stalling the GPU,
+		// so a manual ring buffer is not needed here like VK/DX12.
+		D3D11_BOX dst_region{};
+		dst_region.left = 0;
+		dst_region.right = size;
+		dst_region.top = 0;
+		dst_region.bottom = 1;
+		dst_region.front = 0;
+		dst_region.back = 1;
+		m_ctx->UpdateSubresource(m_accurate_prims_b.get(), 0, &dst_region, config.accurate_prims_edge_data->data(), size, 0);
+
+		config.cb_ps.accurate_prims_base_index.x = 0; // No offsetting needed like DX12/VK since we don't use a ring buffer.
 	}
 	return true;
 }

pcsx2/GS/Renderers/DX11/GSDevice11.h

@@ -137,7 +137,6 @@ class GSDevice11 final : public GSDevice
 	u32 m_vb_pos = 0; // bytes
 	u32 m_ib_pos = 0; // indices/sizeof(u32)
 	u32 m_structured_vb_pos = 0; // bytes
-	u32 m_accurate_prims_b_pos = 0; // bytes/sizeof(AccuratePrimsEdgeData)
 
 	bool m_allow_tearing_supported = false;
 	bool m_using_flip_model_swap_chain = true;

pcsx2/GS/Renderers/DX12/D3D12StreamBuffer.cpp

@@ -20,29 +20,33 @@ D3D12StreamBuffer::~D3D12StreamBuffer()
 	Destroy();
 }
 
-bool D3D12StreamBuffer::Create(u32 size)
+bool D3D12StreamBuffer::Create(u32 size, bool default_heap)
 {
 	const D3D12_RESOURCE_DESC resource_desc = {D3D12_RESOURCE_DIMENSION_BUFFER, 0, size, 1, 1, 1, DXGI_FORMAT_UNKNOWN,
 		{1, 0}, D3D12_TEXTURE_LAYOUT_ROW_MAJOR, D3D12_RESOURCE_FLAG_NONE};
 
 	D3D12MA::ALLOCATION_DESC allocationDesc = {};
 	allocationDesc.Flags = D3D12MA::ALLOCATION_FLAG_COMMITTED;
-	allocationDesc.HeapType = D3D12_HEAP_TYPE_UPLOAD;
+	allocationDesc.HeapType = default_heap ? D3D12_HEAP_TYPE_DEFAULT : D3D12_HEAP_TYPE_UPLOAD;
 
 	wil::com_ptr_nothrow<ID3D12Resource> buffer;
 	wil::com_ptr_nothrow<D3D12MA::Allocation> allocation;
 	HRESULT hr = GSDevice12::GetInstance()->GetAllocator()->CreateResource(&allocationDesc, &resource_desc,
-		D3D12_RESOURCE_STATE_GENERIC_READ, nullptr, allocation.put(), IID_PPV_ARGS(buffer.put()));
+		default_heap ? D3D12_RESOURCE_STATE_COMMON : D3D12_RESOURCE_STATE_GENERIC_READ,
+		nullptr, allocation.put(), IID_PPV_ARGS(buffer.put()));
 	pxAssertMsg(SUCCEEDED(hr), "Allocate buffer");
 	if (FAILED(hr))
 		return false;
 
 	static const D3D12_RANGE read_range = {};
-	u8* host_pointer;
-	hr = buffer->Map(0, &read_range, reinterpret_cast<void**>(&host_pointer));
-	pxAssertMsg(SUCCEEDED(hr), "Map buffer");
-	if (FAILED(hr))
-		return false;
+	u8* host_pointer = nullptr;
+	if (!default_heap)
+	{
+		hr = buffer->Map(0, &read_range, reinterpret_cast<void**>(&host_pointer));
+		pxAssertMsg(SUCCEEDED(hr), "Map buffer");
+		if (FAILED(hr))
+			return false;
+	}
 
 	Destroy(true);
 
@@ -51,6 +55,7 @@ bool D3D12StreamBuffer::Create(u32 size)
 	m_host_pointer = host_pointer;
 	m_size = size;
 	m_gpu_pointer = m_buffer->GetGPUVirtualAddress();
+	m_default_heap = default_heap;
 	return true;
 }
 
@@ -148,6 +153,7 @@ void D3D12StreamBuffer::Destroy(bool defer)
 	m_current_offset = 0;
 	m_current_space = 0;
 	m_current_gpu_position = 0;
+	m_default_heap = false;
 	m_tracked_fences.clear();
 }
 

pcsx2/GS/Renderers/DX12/D3D12StreamBuffer.h

@@ -22,7 +22,7 @@ class D3D12StreamBuffer
 	D3D12StreamBuffer();
 	~D3D12StreamBuffer();
 
-	bool Create(u32 size);
+	bool Create(u32 size, bool default_heap = false);
 
 	__fi bool IsValid() const { return static_cast<bool>(m_buffer); }
 	__fi ID3D12Resource* GetBuffer() const { return m_buffer.get(); }
@@ -54,7 +54,8 @@ class D3D12StreamBuffer
 	wil::com_ptr_nothrow<ID3D12Resource> m_buffer;
 	wil::com_ptr_nothrow<D3D12MA::Allocation> m_allocation;
 	D3D12_GPU_VIRTUAL_ADDRESS m_gpu_pointer = {};
-	u8* m_host_pointer = nullptr;
+	u8* m_host_pointer = nullptr; // Only used for upload heaps.
+	bool m_default_heap = false; // False for upload heap; true for default heap.
 
 	// List of fences and the corresponding positions in the buffer
 	std::deque<std::pair<u64, u32>> m_tracked_fences;