DirectX12的初始化

DirectX12的初始化主要分为以下若干步骤：

1. 创建device和gifactory
2. 创建与GPU通信同步相关的objects，command和fence
3. 创建swap chain
4. 为render target view和depth stencil view创建descriptor heap
5. 创建back buffer和与之绑定的render target view
6. 创建depth stencil buffer和与之绑定的depth stencil view
7. 填充viewport和屏幕剪裁区域结构
8. 绘制前：设置buffer，设置view，设置viewport和屏幕剪裁区域
9. 绘制后：提交绘制，设置buffer

创建device和gifactory

主要调用如下两个接口：

HRESULT D3D12CreateDevice(
  IUnknown          *pAdapter,
  D3D_FEATURE_LEVEL MinimumFeatureLevel,
  REFIID            riid,
  void              **ppDevice
);

HRESULT CreateDXGIFactory1(
  REFIID riid,
  void   **ppFactory
);

这里，我们可以借用微软提供的ComPtr来管理要创建的相关对象：

ComPtr提供了一些常用的接口，重载了一些常用的操作符：

• Get()：获取原始的指针
• GetAddressOf()：获取原始指针的地址
• operator&：释放原始指针指向的对象，然而再返回原始指针的地址

			ComPtr<ID3D12Device> mDevice = nullptr;
			ComPtr<IDXGISwapChain> mSwapChain = nullptr;

同时，为了方便填充这些创建接口，微软提供了一个宏定义IID_PPV_ARGS：

	ThrowIfFailed(D3D12CreateDevice(nullptr, D3D_FEATURE_LEVEL_11_0, IID_PPV_ARGS(&mDevice)));
	ThrowIfFailed(CreateDXGIFactory1(IID_PPV_ARGS(&mDxGiFactory)));

创建与GPU通信同步相关的objects，command和fence

command是用来从CPU提交相关指令给GPU执行的。这里需要三个objects：queue，allocator，list。queue是用来最终提交给CPU的数据结构，allocator负责为list提供空间管理，list提供了若干CPU提交相关指令的接口。

ThrowIfFailed(mDevice->CreateCommandQueue(&queueDesc, IID_PPV_ARGS(&mCommandQueue)));
ThrowIfFailed(mDevice->CreateCommandAllocator(D3D12_COMMAND_LIST_TYPE_DIRECT, IID_PPV_ARGS(&mCommandAlloc)));
ThrowIfFailed(mDevice->CreateCommandList(0, D3D12_COMMAND_LIST_TYPE_DIRECT, mCommandAlloc.Get(), nullptr, IID_PPV_ARGS(&mCommandList)));
// must close otherwise call reset will raise a error
mCommandList->Close();

创建完之后记得调用Close()方法，否则后面调用Reset()时会因为状态未关闭而出错。

fence是用来管理CPU和GPU同步的对象，防止对同一资源进行冲突访问，创建接口如下：

HRESULT CreateFence(
  UINT64            InitialValue,
  D3D12_FENCE_FLAGS Flags,
  REFIID            riid,
  void              **ppFence
);

创建swap chain

HRESULT CreateSwapChain(
  IUnknown             *pDevice,
  DXGI_SWAP_CHAIN_DESC *pDesc,
  IDXGISwapChain       **ppSwapChain
);

值得一提的是，DirectX12不支持用该API创建MSAA swap chain：

Direct3D 12 don’t support creating MSAA swap chains–attempts to create a swap chain with SampleDesc.Count > 1 will fail. Instead, you create your own MSAA render target and explicitly resolve to the DXGI back-buffer for presentation as shown here.

为render target view和depth stencil view创建descriptor heap

heap就是用来存放管理view的结构，相关接口如下：

UINT GetDescriptorHandleIncrementSize(
  D3D12_DESCRIPTOR_HEAP_TYPE DescriptorHeapType
);

HRESULT CreateDescriptorHeap(
  const D3D12_DESCRIPTOR_HEAP_DESC *pDescriptorHeapDesc,
  REFIID                           riid,
  void                             **ppvHeap
);

通过GetDescriptorHandleIncrementSize可以知道heap上每个handle的大小，方便在取一个heap上多个view进行偏移。

创建back buffer和与之绑定的render target view

	ThrowIfFailed(mSwapChain->ResizeBuffers(mBackBufferCount, windowWidth, windowHeight, mBackBufferFormat, 
		DXGI_SWAP_CHAIN_FLAG_ALLOW_MODE_SWITCH));
	CD3DX12_CPU_DESCRIPTOR_HANDLE handle = CD3DX12_CPU_DESCRIPTOR_HANDLE(
		mRtvHeap->GetCPUDescriptorHandleForHeapStart());

	for (int i = 0; i < mBackBufferCount; i++)
	{
		ThrowIfFailed(mSwapChain->GetBuffer(i, IID_PPV_ARGS(&mBackBuffer[i])));
		mDevice->CreateRenderTargetView(mBackBuffer[i].Get(), nullptr, handle);
		handle.Offset(1, mRtvHeapIncSize);
	}

CD3DX12_CPU_DESCRIPTOR_HANDLE也是微软提供的便利数据结构，包含一些简单的接口。这里用了Offset方法来取得正确的heap位置。

创建depth stencil buffer和与之绑定的depth stencil view

	ThrowIfFailed(mDevice->CreateCommittedResource(&CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_DEFAULT), 
		D3D12_HEAP_FLAG_NONE, &desc, D3D12_RESOURCE_STATE_COMMON, &clearValue, IID_PPV_ARGS(&mDepthStencilBuffer)));

	handle = CD3DX12_CPU_DESCRIPTOR_HANDLE(mDsvHeap->GetCPUDescriptorHandleForHeapStart());
	mDevice->CreateDepthStencilView(mDepthStencilBuffer.Get(), &viewDesc, handle);

与back buffer的过程类似，只不过depth stencil buffer只有一个，而back buffer可能有多个。CD3DX12_HEAP_PROPERTIES同样也是便利数据结构。

注意到buffer创建完后处于D3D12_RESOURCE_STATE_COMMON状态，我们需要将其转为D3D12_RESOURCE_STATE_DEPTH_WRITE状态，使用CD3DX12_RESOURCE_BARRIER::Transition进行便捷操作：

	mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(mDepthStencilBuffer.Get(), 
		D3D12_RESOURCE_STATE_COMMON, D3D12_RESOURCE_STATE_DEPTH_WRITE));
	ThrowIfFailed(mCommandList->Close());

	ID3D12CommandList *cmdList[] = { mCommandList.Get() };
	mCommandQueue->ExecuteCommandLists(_countof(cmdList), cmdList);

	FlushCommandQueue();

void DirectX12::FlushCommandQueue()
{
	mCurrentFence++;
	ThrowIfFailed(mCommandQueue->Signal(mFence.Get(), mCurrentFence));
	if (mFence->GetCompletedValue() < mCurrentFence)
	{
		HANDLE handle = CreateEventEx(nullptr, false, false, EVENT_ALL_ACCESS);
		ThrowIfFailed(mFence->SetEventOnCompletion(mCurrentFence, handle));

		WaitForSingleObject(handle, INFINITE);
		CloseHandle(handle);
	}
}

资源状态转换需要向GPU提交指令，因此涉及到同步问题，即CPU需要等待GPU完成指令之后才能继续执行其他操作。这里用了FlushCommandQueue来完成同步。这个方法的实现一目了然：

首先将当前fence计数值增加，signal提交到GPU，然后一直等待，直到获取到与计数值相同的值为止。

填充viewport和屏幕剪裁区域结构

这一步骤只需填充D3D12_VIEWPORT和D3D12_RECT结构即可。

	mViewport.TopLeftX = 0;
	mViewport.TopLeftY = 0;
	mViewport.Width = windowWidth;
	mViewport.Height = windowHeight;
	mViewport.MinDepth = 0.0f;
	mViewport.MaxDepth = 1.0f;

	mScissorRect.left = 0;
	mScissorRect.right = windowWidth;
	mScissorRect.top = 0;
	mScissorRect.bottom = windowHeight;

绘制前：设置buffer，设置view，设置viewport和屏幕剪裁区域

	mCommandList->ResourceBarrier(1, &CD3DX12_RESOURCE_BARRIER::Transition(mBackBuffer[mCurBackBuffer].Get(), 
		D3D12_RESOURCE_STATE_PRESENT, D3D12_RESOURCE_STATE_RENDER_TARGET));
	mCommandList->RSSetScissorRects(1, &mScissorRect);
	mCommandList->RSSetViewports(1, &mViewport);
	CD3DX12_CPU_DESCRIPTOR_HANDLE dsv = CD3DX12_CPU_DESCRIPTOR_HANDLE(
		mDsvHeap->GetCPUDescriptorHandleForHeapStart());
	mCommandList->ClearDepthStencilView(dsv, D3D12_CLEAR_FLAG_DEPTH | D3D12_CLEAR_FLAG_STENCIL, 1.0f, 0, 0, 
		nullptr);
	CD3DX12_CPU_DESCRIPTOR_HANDLE rtv = CD3DX12_CPU_DESCRIPTOR_HANDLE(
		mRtvHeap->GetCPUDescriptorHandleForHeapStart(), mCurBackBuffer, mRtvHeapIncSize);
	mCommandList->ClearRenderTargetView(rtv, LightSteelBlue, 0, nullptr);
	mCommandList->OMSetRenderTargets(1, &rtv, true, &dsv);

这里将backBuffer的状态从D3D12_RESOURCE_STATE_PRESENT转移到了D3D12_RESOURCE_STATE_RENDER_TARGET，这样后面的设置和绘制工作才会有效，否则会出错。实际上，状态定义里，D3D12_RESOURCE_STATE_PRESENT和D3D12_RESOURCE_STATE_COMMON是等价的，都为0，这里只是换了一种说法。

绘制后：提交绘制，设置buffer

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

	ID3D12CommandList *cmdList[] = { mCommandList.Get() };
	mCommandQueue->ExecuteCommandLists(_countof(cmdList), cmdList);

	ThrowIfFailed(mSwapChain->Present(0, 0));
	mCurBackBuffer = (mCurBackBuffer + 1) % mBackBufferCount;

在调用swapChain的Present方法之前，必须将backbuffer的状态设置为D3D12_RESOURCE_STATE_PRESENT。如果使用了多个backbuffer，别忘记切换到下一个backbuffer。

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）^-