在DirectX 12中使用compute shader

compute shader是用来把一些适合在gpu上运算的任务从cpu挪到gpu去做。例如一些后处理的效果，对纹理上的每个像素分别进行各自运算，最后得到结果。我们以实现sobel边缘检测算法为例，首先把正常绘制的内容渲染到一张off screen的render target上；然后使用这张render target作为compute shader的输入，在gpu上运行sobel边缘检测算法，输出到一个新的buffer；最后，将buffer作为输入，绘制到屏幕上。

首先，需要为off screen的render target创建对应的buffer和view：

	D3D12_RESOURCE_DESC resDesc;
	resDesc.Alignment = 0;
	resDesc.DepthOrArraySize = 1;
	resDesc.Dimension = D3D12_RESOURCE_DIMENSION_TEXTURE2D;
	resDesc.Flags = D3D12_RESOURCE_FLAG_ALLOW_RENDER_TARGET;
	resDesc.Format = mBackBufferFormat;
	resDesc.Height = windowHeight;
	resDesc.Layout = D3D12_TEXTURE_LAYOUT_UNKNOWN;
	resDesc.MipLevels = 1;
	resDesc.SampleDesc.Count = 1;
	resDesc.SampleDesc.Quality = 0;
	resDesc.Width = windowWidth;
	ThrowIfFailed(mDevice->CreateCommittedResource(&CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_DEFAULT),
		D3D12_HEAP_FLAG_NONE, &resDesc, D3D12_RESOURCE_STATE_GENERIC_READ, nullptr,
		IID_PPV_ARGS(&mRenderTargetBuffer)));

	D3D12_SHADER_RESOURCE_VIEW_DESC srvDesc;
	srvDesc.Format = mBackBufferFormat;
	srvDesc.Shader4ComponentMapping = D3D12_DEFAULT_SHADER_4_COMPONENT_MAPPING;
	srvDesc.Texture2D.MipLevels = 1;
	srvDesc.Texture2D.MostDetailedMip = 0;
	srvDesc.Texture2D.PlaneSlice = 0;
	srvDesc.Texture2D.ResourceMinLODClamp = 0.0f;
	srvDesc.ViewDimension = D3D12_SRV_DIMENSION_TEXTURE2D;

	mDevice->CreateShaderResourceView(mRenderTargetBuffer.Get(), &srvDesc, cpuSrvHandle);

	D3D12_RENDER_TARGET_VIEW_DESC rtvDesc;
	rtvDesc.Format = mBackBufferFormat;
	rtvDesc.Texture2D.MipSlice = 0;
	rtvDesc.Texture2D.PlaneSlice = 0;
	rtvDesc.ViewDimension = D3D12_RTV_DIMENSION_TEXTURE2D;

	mDevice->CreateRenderTargetView(mRenderTargetBuffer.Get(), &rtvDesc, cpuRtvHandle);

我们为buffer创建了一个shader resource view和render target view，前者是用来给后面的sobel算法的输入使用，后者是作为off screen buffer存储正常绘制的内容。

接下来，需要为compute shader去创建对应的buffer和view：

	D3D12_RESOURCE_DESC resDesc;
	resDesc.Alignment = 0;
	resDesc.DepthOrArraySize = 1;
	resDesc.Dimension = D3D12_RESOURCE_DIMENSION_TEXTURE2D;
	resDesc.Flags = D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS;
	resDesc.Format = mBackBufferFormat;
	resDesc.Height = windowHeight;
	resDesc.Layout = D3D12_TEXTURE_LAYOUT_UNKNOWN;
	resDesc.MipLevels = 1;
	resDesc.SampleDesc.Count = 1;
	resDesc.SampleDesc.Quality = 0;
	resDesc.Width = windowWidth;

	ThrowIfFailed(mDevice->CreateCommittedResource(&CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_DEFAULT),
		D3D12_HEAP_FLAG_NONE, &resDesc, D3D12_RESOURCE_STATE_GENERIC_READ, nullptr,
		IID_PPV_ARGS(&mComputeBuffer)));

	D3D12_SHADER_RESOURCE_VIEW_DESC srvDesc;
	srvDesc.Format = mBackBufferFormat;
	srvDesc.Shader4ComponentMapping = D3D12_DEFAULT_SHADER_4_COMPONENT_MAPPING;
	srvDesc.Texture2D.MipLevels = 1;
	srvDesc.Texture2D.MostDetailedMip = 0;
	srvDesc.Texture2D.PlaneSlice = 0;
	srvDesc.Texture2D.ResourceMinLODClamp = 0.0f;
	srvDesc.ViewDimension = D3D12_SRV_DIMENSION_TEXTURE2D;

	mDevice->CreateShaderResourceView(mComputeBuffer.Get(), &srvDesc, cpuSrvHandle);

	D3D12_UNORDERED_ACCESS_VIEW_DESC uavDesc;
	uavDesc.Format = mBackBufferFormat;
	uavDesc.Texture2D.MipSlice = 0;
	uavDesc.Texture2D.PlaneSlice = 0;
	uavDesc.ViewDimension = D3D12_UAV_DIMENSION_TEXTURE2D;

	mDevice->CreateUnorderedAccessView(mComputeBuffer.Get(), nullptr, &uavDesc, cpuUavHandle);

我们为buffer创建了一个shader resource view和unordered access view，前者是compute shader执行完之后，用来提供给最终绘制到屏幕的shader使用的；后者是compute shader如果需要写入到buffer，则buffer需要绑定到unordered access view。

其次，还要为它们创建对应的根签名和pipeline state object，用来在绘制过程中切换：

	CD3DX12_DESCRIPTOR_RANGE cbvSrvTable[3];
	cbvSrvTable[0].Init(D3D12_DESCRIPTOR_RANGE_TYPE_SRV, 1, 0);
	cbvSrvTable[1].Init(D3D12_DESCRIPTOR_RANGE_TYPE_SRV, 1, 1);
	cbvSrvTable[2].Init(D3D12_DESCRIPTOR_RANGE_TYPE_UAV, 2, 0);

	CD3DX12_ROOT_PARAMETER rootParams[3];
	rootParams[0].InitAsDescriptorTable(1, &cbvSrvTable[0]);
	rootParams[1].InitAsDescriptorTable(1, &cbvSrvTable[1]);
	rootParams[2].InitAsDescriptorTable(1, &cbvSrvTable[2]);

	CD3DX12_ROOT_SIGNATURE_DESC sigDesc(_countof(rootParams), rootParams, mStaticSamplers.size(),
		mStaticSamplers.data(), D3D12_ROOT_SIGNATURE_FLAG_ALLOW_INPUT_ASSEMBLER_INPUT_LAYOUT);
	ComPtr<ID3DBlob> signature = nullptr;
	ComPtr<ID3DBlob> error = nullptr;
	HRESULT hr = D3D12SerializeRootSignature(&sigDesc, D3D_ROOT_SIGNATURE_VERSION_1, &signature, &error);
	ThrowIfFailed(hr);
	ThrowIfFailed(mDevice->CreateRootSignature(0, signature->GetBufferPointer(), signature->GetBufferSize(),
		IID_PPV_ARGS(&mSignature)));

	D3D12_COMPUTE_PIPELINE_STATE_DESC computePsoDesc;
	computePsoDesc.CachedPSO.CachedBlobSizeInBytes = 0;
	computePsoDesc.CachedPSO.pCachedBlob = nullptr;
	computePsoDesc.CS = { cs->GetBufferPointer(), cs->GetBufferSize() };
	computePsoDesc.Flags = D3D12_PIPELINE_STATE_FLAG_NONE;
	computePsoDesc.NodeMask = 0;
	computePsoDesc.pRootSignature = mSignature.Get();
	ThrowIfFailed(mDevice->CreateComputePipelineState(&computePsoDesc, 
		IID_PPV_ARGS(&mComputePipelineState)));

	// render target
	D3D12_GRAPHICS_PIPELINE_STATE_DESC graphicPsoDesc;
	ZeroMemory(&graphicPsoDesc, sizeof(D3D12_GRAPHICS_PIPELINE_STATE_DESC));
	graphicPsoDesc.BlendState = CD3DX12_BLEND_DESC(D3D12_DEFAULT);
	CD3DX12_DEPTH_STENCIL_DESC dsDesc = CD3DX12_DEPTH_STENCIL_DESC(D3D12_DEFAULT);
	dsDesc.DepthEnable = false;
	dsDesc.DepthWriteMask = D3D12_DEPTH_WRITE_MASK_ZERO;
	dsDesc.DepthFunc = D3D12_COMPARISON_FUNC_ALWAYS;
	graphicPsoDesc.DepthStencilState = dsDesc;
	graphicPsoDesc.DSVFormat = mDepthStencilBufferFormat;
	graphicPsoDesc.Flags = D3D12_PIPELINE_STATE_FLAG_NONE;
	graphicPsoDesc.IBStripCutValue = D3D12_INDEX_BUFFER_STRIP_CUT_VALUE_DISABLED;
	graphicPsoDesc.NodeMask = 0;
	graphicPsoDesc.NumRenderTargets = 1;
	graphicPsoDesc.pRootSignature = mSignature.Get();
	CD3DX12_RASTERIZER_DESC rastDesc = CD3DX12_RASTERIZER_DESC(D3D12_DEFAULT);
	rastDesc.FillMode = mFillMode;
	rastDesc.CullMode = mCullMode;
	graphicPsoDesc.RasterizerState = rastDesc;
	graphicPsoDesc.RTVFormats[0] = mBackBufferFormat;
	graphicPsoDesc.SampleDesc.Count = mEnableMsaa ? mMsaaCount : 1;
	graphicPsoDesc.SampleDesc.Quality = mEnableMsaa ? mMsaaQuality - 1 : 0;
	graphicPsoDesc.SampleMask = UINT_MAX;
	graphicPsoDesc.PS = { ps->GetBufferPointer(), ps->GetBufferSize() };
	graphicPsoDesc.VS = { vs->GetBufferPointer(), vs->GetBufferSize() };
	graphicPsoDesc.InputLayout = { mInputLayout.data(), mInputLayout.size() };
	graphicPsoDesc.PrimitiveTopologyType = D3D12_PRIMITIVE_TOPOLOGY_TYPE_TRIANGLE;
	ThrowIfFailed(mDevice->CreateGraphicsPipelineState(&graphicPsoDesc,
		IID_PPV_ARGS(&mCoreRenderTarget.mGraphicPipelineState)));

这里为了方便处理，我们只创建了一个根签名给两个shader使用。因为这两个shader都只接收一个buffer的输入，输出一个buffer。

准备过程完成后，我们就可以开始绘制了。先要对已有的逻辑进行修改，即渲染写入的对象从原先的back buffer转移到off screen的render target上：

		mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(
			mRenderTargetBuffer.Get(), D3D12_RESOURCE_STATE_GENERIC_READ, 
			D3D12_RESOURCE_STATE_RENDER_TARGET));
		mCommandList->ClearRenderTargetView(rtv, LightSteelBlue, 0, nullptr);
		mCommandList->OMSetRenderTargets(1, &rtv, true, &dsv);

		// draw objects

		mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(
			mRenderTargetBuffer.Get(), D3D12_RESOURCE_STATE_RENDER_TARGET, 
			D3D12_RESOURCE_STATE_GENERIC_READ));

正常绘制工作完成后，再执行一开始提到的后处理逻辑，先将off screen的render target传给compute shader处理，输出提取边缘后的buffer，然后将buffer直接绘制屏幕上：

	mCommandList->SetComputeRootSignature(mComputeSignature.Get());
	mCommandList->SetComputeRootDescriptorTable(0, mGpuSrv);
	mCommandList->SetComputeRootDescriptorTable(2, mGpuUav);
	mCommandList->SetPipelineState(mComputePipelineState.Get());

	mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(mComputeBuffer.Get(),
		D3D12_RESOURCE_STATE_GENERIC_READ, D3D12_RESOURCE_STATE_UNORDERED_ACCESS));

	UINT numGroupsX = (UINT)ceilf(windowWidth / 16.0f);
	UINT numGroupsY = (UINT)ceilf(windowHeight / 16.0f);
	mCommandList->Dispatch(numGroupsX, numGroupsY, 1);

	mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(mComputeBuffer.Get(),
		D3D12_RESOURCE_STATE_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_GENERIC_READ));


	mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(mBackBuffer[mCurBackBuffer].Get(),
		D3D12_RESOURCE_STATE_PRESENT, D3D12_RESOURCE_STATE_RENDER_TARGET));
	CD3DX12_CPU_DESCRIPTOR_HANDLE rtv = CD3DX12_CPU_DESCRIPTOR_HANDLE(
		mRtvHeap->GetCPUDescriptorHandleForHeapStart(), mCurBackBuffer, mRtvHeapIncSize);
	CD3DX12_CPU_DESCRIPTOR_HANDLE dsv = CD3DX12_CPU_DESCRIPTOR_HANDLE(
		mDsvHeap->GetCPUDescriptorHandleForHeapStart());
	mCommandList->OMSetRenderTargets(1, &rtv, true, &dsv);

	mCommandList->SetGraphicsRootSignature(mGraphicSignature.Get());
	mCommandList->SetGraphicsRootDescriptorTable(0, mGpuSrv);
	mCommandList->SetGraphicsRootDescriptorTable(1, mGpuSrv);
	mCommandList->SetPipelineState(mGraphicPipelineState.Get());

	mCommandList->IASetVertexBuffers(0, 1, nullptr);
	mCommandList->IASetIndexBuffer(nullptr);
	mCommandList->IASetPrimitiveTopology(D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
	mCommandList->DrawInstanced(6, 1, 0, 0);

	mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(mBackBuffer[mCurBackBuffer].Get(),
		D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_PRESENT));

值得一提的是，在最后绘制边缘效果到屏幕上时，我们并没有传入vertex buffer和index buffer，通过设置为nullptr，我们在shader中直接去写死clip space中顶点的位置：

float2 gTexCoords[6] = 
{
	float2(0.0f, 1.0f),
	float2(0.0f, 0.0f),
	float2(1.0f, 0.0f),
	float2(0.0f, 1.0f),
	float2(1.0f, 0.0f),
	float2(1.0f, 1.0f)
};

struct VertexOut
{
	float4 PosH    : SV_POSITION;
	float2 TexC    : TEXCOORD;
};

VertexOut VS(uint vid : SV_VertexID)
{
	VertexOut vout;
	
	vout.TexC = gTexCoords[vid];
	
	// Map [0,1]^2 to NDC space.
	vout.PosH = float4(2.0f*vout.TexC.x - 1.0f, 1.0f - 2.0f*vout.TexC.y, 0.0f, 1.0f);

    return vout;
}

最后绘制的效果如下：

如果你觉得我的文章有帮助，欢迎关注我的微信公众号（大龄社畜的游戏开发之路）^-