d3d12: Do not create/submit an extra command list for texture upload/rtt state change

rpcs3/Emu/RSX/D3D12/D3D12GSRender.cpp

@@ -431,12 +431,12 @@ void D3D12GSRender::Clear(u32 cmd)
 {
 	assert(cmd == NV4097_CLEAR_SURFACE);
 
-	PrepareRenderTargets();
-
 	ID3D12GraphicsCommandList *commandList;
 	check(m_device->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_DIRECT, getCurrentResourceStorage().m_commandAllocator, nullptr, IID_PPV_ARGS(&commandList)));
 	getCurrentResourceStorage().m_inflightCommandList.push_back(commandList);
 
+	PrepareRenderTargets(commandList);
+
 /*	if (m_set_color_mask)
 	{
 		glColorMask(m_color_mask_r, m_color_mask_g, m_color_mask_b, m_color_mask_a);
@@ -509,7 +509,11 @@ void D3D12GSRender::Clear(u32 cmd)
 
 void D3D12GSRender::Draw()
 {
-	PrepareRenderTargets();
+	ID3D12GraphicsCommandList *commandList;
+	m_device->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_DIRECT, getCurrentResourceStorage().m_commandAllocator, nullptr, IID_PPV_ARGS(&commandList));
+	getCurrentResourceStorage().m_inflightCommandList.push_back(commandList);
+
+	PrepareRenderTargets(commandList);
 
 	// Init vertex count
 	// TODO: Very hackish, clean this
@@ -536,11 +540,6 @@ void D3D12GSRender::Draw()
 		}
 	}
 
-	ID3D12GraphicsCommandList *commandList;
-	m_device->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_DIRECT, getCurrentResourceStorage().m_commandAllocator, nullptr, IID_PPV_ARGS(&commandList));
-	getCurrentResourceStorage().m_inflightCommandList.push_back(commandList);
-
-
 	std::chrono::time_point<std::chrono::system_clock> startVertexTime = std::chrono::system_clock::now();
 	if (m_indexed_array.m_count || m_draw_array_count)
 	{
@@ -588,7 +587,7 @@ void D3D12GSRender::Draw()
 	if (m_PSO->second > 0)
 	{
 		std::chrono::time_point<std::chrono::system_clock> startTextureTime = std::chrono::system_clock::now();
-		size_t usedTexture = UploadTextures();
+		size_t usedTexture = UploadTextures(commandList);
 
 		// Fill empty slots
 		for (; usedTexture < m_PSO->second; usedTexture++)

rpcs3/Emu/RSX/D3D12/D3D12GSRender.h

@@ -402,12 +402,19 @@ private:
 	void FillVertexShaderConstantsBuffer();
 	void FillPixelShaderConstantsBuffer();
 	/**
-	 * Upload textures to Data heap if necessary and create necessary descriptor in the per frame storage struct.
+	 * Fetch all textures recorded in the state in the render target cache and in the texture cache.
-	 * returns the number of texture uploaded
+	 * If a texture is not cached, populate cmdlist with uploads command.
+	 * Create necessary resource view/sampler descriptors in the per frame storage struct.
+	 * returns the number of texture uploaded.
 	 */
-	size_t UploadTextures();
+	size_t UploadTextures(ID3D12GraphicsCommandList *cmdlist);
 
-	void PrepareRenderTargets();
+	/**
+	 * Creates render target if necessary.
+	 * Populate cmdlist with render target state change (from RTT to generic read for previous rtt,
+	 * from generic to rtt for rtt in cache).
+	 */
+	void PrepareRenderTargets(ID3D12GraphicsCommandList *cmdlist);
 protected:
 	virtual void OnInit() override;
 	virtual void OnInitThread() override;

rpcs3/Emu/RSX/D3D12/D3D12RenderTargetSets.cpp

@@ -12,7 +12,7 @@
 #include "D3D12.h"
 #include "D3D12GSRender.h"
 
-void D3D12GSRender::PrepareRenderTargets()
+void D3D12GSRender::PrepareRenderTargets(ID3D12GraphicsCommandList *copycmdlist)
 {
 	// FBO location has changed, previous data might be copied
 	u32 address_a = m_set_context_dma_color_a ? GetAddress(m_surface_offset_a, m_context_dma_color_a - 0xfeed0000) : 0;
@@ -21,10 +21,6 @@ void D3D12GSRender::PrepareRenderTargets()
 	u32 address_d = m_set_context_dma_color_d ? GetAddress(m_surface_offset_d, m_context_dma_color_d - 0xfeed0000) : 0;
 	u32 address_z = m_set_context_dma_z ? GetAddress(m_surface_offset_z, m_context_dma_z - 0xfeed0000) : 0;
 
-	ID3D12GraphicsCommandList *copycmdlist;
-	check(m_device->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_DIRECT, getCurrentResourceStorage().m_commandAllocator, nullptr, IID_PPV_ARGS(&copycmdlist)));
-	getCurrentResourceStorage().m_inflightCommandList.push_back(copycmdlist);
-
 	// Make previous RTTs sampleable
 	for (unsigned i = 0; i < 4; i++)
 	{
@@ -141,9 +137,6 @@ void D3D12GSRender::PrepareRenderTargets()
 	}
 	depthStencilViewDesc.ViewDimension = D3D12_DSV_DIMENSION_TEXTURE2D;
 	m_device->CreateDepthStencilView(ds, &depthStencilViewDesc, m_rtts.m_depthStencilDescriptorHeap->GetCPUDescriptorHandleForHeapStart());
-
-	check(copycmdlist->Close());
-	m_commandQueueGraphic->ExecuteCommandLists(1, (ID3D12CommandList**)&copycmdlist);
 }
 
 ID3D12Resource *RenderTargets::bindAddressAsRenderTargets(ID3D12Device *device, ID3D12GraphicsCommandList *cmdList, size_t slot, u32 address,

rpcs3/Emu/RSX/D3D12/D3D12Texture.cpp

@@ -726,7 +726,7 @@ size_t getTextureSize(const RSXTexture &texture)
 	}
 }
 
-size_t D3D12GSRender::UploadTextures()
+size_t D3D12GSRender::UploadTextures(ID3D12GraphicsCommandList *cmdlist)
 {
 	std::lock_guard<std::mutex> lock(mut);
 	size_t usedTexture = 0;
@@ -758,15 +758,7 @@ size_t D3D12GSRender::UploadTextures()
 		}
 		else
 		{
-			// Upload at each iteration to take advantage of overlapping transfer
+			vramTexture = uploadSingleTexture(m_textures[i], m_device, cmdlist, m_textureUploadData);
-			ID3D12GraphicsCommandList *commandList;
-			check(m_device->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_DIRECT, getCurrentResourceStorage().m_textureUploadCommandAllocator, nullptr, IID_PPV_ARGS(&commandList)));
-
-			vramTexture = uploadSingleTexture(m_textures[i], m_device, commandList, m_textureUploadData);
-
-			check(commandList->Close());
-			m_commandQueueGraphic->ExecuteCommandLists(1, (ID3D12CommandList**)&commandList);
-			getCurrentResourceStorage().m_inflightCommandList.push_back(commandList);
 			m_texturesCache[texaddr] = vramTexture;
 
 			u32 s = (u32)align(getTextureSize(m_textures[i]), 4096);