HLSL写的Compute Shader处理简单贴图时产生的问题

tigerccx 2018-10-26 11:44:33

在尝试编写《Introduction to 3D Game Programming with DirectX 11》中12章习题5时产生的问题。
再第一次初始化两张格式为DXGI_FORMAT_R32_FLOAT的贴图作为现在时刻和过去时刻的波浪高度贴图时，使用全0初始化：



    D3D11_TEXTURE2D_DESC tex2DDesc;

	tex2DDesc.Height = hCount;

	tex2DDesc.Width = wCount;

	tex2DDesc.MipLevels = 1;

	tex2DDesc.ArraySize = 1;

	tex2DDesc.Format = DXGI_FORMAT_R32_FLOAT;

	tex2DDesc.SampleDesc.Count = 1;

	tex2DDesc.SampleDesc.Quality = 0;

	tex2DDesc.Usage = D3D11_USAGE_DEFAULT;

	tex2DDesc.BindFlags = D3D11_BIND_SHADER_RESOURCE | D3D11_BIND_UNORDERED_ACCESS;

	tex2DDesc.CPUAccessFlags = 0;

	tex2DDesc.MiscFlags = 0;

	// Set init data for both height maps to 0.0f

	std::vector<float> vecInitHeight;

	for (int i = 0; i < hCount*wCount; ++i)

	{

		vecInitHeight.push_back(0.0f);

	}

	D3D11_SUBRESOURCE_DATA subHeightData;

	subHeightData.pSysMem = &vecInitHeight[0];

	subHeightData.SysMemPitch = wCount*sizeof(float);

	subHeightData.SysMemSlicePitch = hCount*wCount*sizeof(float);

	// Sampled in VS to do transform

	HR(device->CreateTexture2D(&tex2DDesc, &subHeightData, &mCurHeightMap));

	HR(device->CreateTexture2D(&tex2DDesc, &subHeightData, &mLasHeightMap));

然后创建了相应的SRV和UAV，之后使用Compute Shader进行更新：

#define THREAD_NUM 32

#define MEM_SIZE (THREAD_NUM+2)



// Only need 3 params for CSUpdateHeight

cbuffer cbUpdateHeightPerObject

{

	float gK1;

	float gK2;

	float gK3;

};



cbuffer cbUpdateNormAndTangPerObject

{

	float gSpatialStep;

};



Texture2D<float> gCurTex;

RWTexture2D<float> gLasTex;

RWTexture2D<float2> gNormTex;

RWTexture2D<float2> gTangTex;



groupshared float gCache1[MEM_SIZE][MEM_SIZE]; //curtex

groupshared float gCache2[MEM_SIZE][MEM_SIZE]; //prevtex



int2 MaxInt2(int2 src, int2 border)

{

	return int2(max(src.x, border.x), max(src.y, border.y));

}



int2 MinInt2(int2 src, int2 border)

{

	return int2(min(src.x, border.x), min(src.y, border.y));

}



[numthreads(THREAD_NUM, THREAD_NUM, 1)]

void CSUpdateHeight(int3 groupThreadID:SV_GroupThreadID,

	int3 dispatchThreadID : SV_DispatchThreadID)

{

	// Update gCurTex data to gChache1

	// Update gLasTex data to gCache2

	// Clamp left up border

    if (groupThreadID.x == 0 || groupThreadID.y == 0)

    {

		int2 xy0 = int2(dispatchThreadID.x - 1, dispatchThreadID.y - 1);//-1 needed?

		int2 xy1 = int2(0, 0);

        int2 xy = MaxInt2(xy0,int2(0,0));

        gCache1[groupThreadID.x][groupThreadID.y] = gCurTex[xy];

		gCache2[groupThreadID.x][groupThreadID.y] = gLasTex[xy];

    }

	// Clamp right down border

    if (groupThreadID.x == (THREAD_NUM - 1) || groupThreadID.y == (THREAD_NUM - 1))//-1 needed?

    {

		int2 xy0 = int2(dispatchThreadID.x + 1, dispatchThreadID.y + 1);

		int2 xy1 = int2(gCurTex.Length.x - 1, gCurTex.Length.y - 1);

        int2 xy = MinInt2(xy0, xy1);

        gCache1[groupThreadID.x + 2][groupThreadID.y + 2] = gCurTex[xy];

		gCache2[groupThreadID.x + 2][groupThreadID.y + 2] = gLasTex[xy];

    }

	// Other

	int2 xy0 = int2(gCurTex.Length.x - 1, gCurTex.Length.y - 1);

	int2 xy = MinInt2(dispatchThreadID.xy, xy0);

    gCache1[groupThreadID.x + 1][groupThreadID.y + 1] = gCurTex[xy];

	gCache2[groupThreadID.x + 1][groupThreadID.y + 1] = gLasTex[xy];



	// Wait for cache synchronize

    GroupMemoryBarrierWithGroupSync();



	int cx = groupThreadID.x + 1;

	int cy = groupThreadID.y + 1;

	// Restrict border from changing

	if(dispatchThreadID.x!=0 && 

		dispatchThreadID.y!=0 && 

		dispatchThreadID.x<gCurTex.Length.x - 1 && 

		dispatchThreadID.y<gCurTex.Length.y- 1)

	gLasTex[dispatchThreadID.xy] =

		gK1*gCache2[cx][cy] +

		gK2*gCache1[cx][cy] +

		gK3*(gCache1[cx + 1][cy] +

			gCache1[cx - 1][cy] +

			gCache1[cx][cy + 1] +

			gCache1[cx][cy - 1]);

}



technique11 UpdateHeight

{

	pass P0

	{

		SetVertexShader(NULL);

		SetPixelShader(NULL);

		SetComputeShader(CompileShader(cs_5_0, CSUpdateHeight()));

	}

}

并在代码中调用



    t += dt;

	if (t >= mTimeStep)

	{

		++testcount;

		// Update current and last height map

		D3DX11_TECHNIQUE_DESC techDesc;

		Effects::WavesUpdateFX->Test->GetDesc(&techDesc);

		for (int p = 0; p < techDesc.Passes; ++p)

		{

			Effects::WavesUpdateFX->SetTestMap(mLasHeightSRV);

			Effects::WavesUpdateFX->SetTestCount(testcount);



			Effects::WavesUpdateFX->Test->GetPassByIndex(p)->Apply(0, dc);

			int cxGroups = ceilf(1.0f * mNumRows / THREAD_NUM);

			int cyGroups = ceilf(1.0f * mNumCols / THREAD_NUM);

			HR(dc->Dispatch(cxGroups, cyGroups, 1));



			ID3D11ShaderResourceView* clearSRVs[] = { NULL };

			dc->CSSetShaderResources(0, 1, clearSRVs);

		}



		Effects::WavesUpdateFX->UpdateHeightTech->GetDesc(&techDesc);

		for (int p = 0; p < techDesc.Passes; ++p)

		{

			Effects::WavesUpdateFX->SetK1(mK1);

			Effects::WavesUpdateFX->SetK2(mK2);

			Effects::WavesUpdateFX->SetK3(mK3);

			Effects::WavesUpdateFX->SetCurTex(mCurHeightSRV);

			Effects::WavesUpdateFX->SetLasTex(mLasHeightUAV);



			Effects::WavesUpdateFX->UpdateHeightTech->GetPassByIndex(p)->Apply(0, dc);

			int cxGroups = ceilf(1.0f * mNumRows / THREAD_NUM);

			int cyGroups = ceilf(1.0f * mNumCols / THREAD_NUM);

			HR(dc->Dispatch(cxGroups, cyGroups, 1));

		}



		// Good house keeping

		ID3D11ShaderResourceView* clearSRVs[] = { NULL };

		dc->CSSetShaderResources(0, 1, clearSRVs);



		ID3D11UnorderedAccessView* clearUAVs[] = { NULL };

		dc->CSSetUnorderedAccessViews(0, 1, clearUAVs, 0);

		

		t = 0.0f;

	}

	dc->CSSetShader(0, 0, 0);