Matrix in constant buffer on the GPU contains NaNs but not on the CPU

Question

I've got a shader with the following constant buffer:

cbuffer ShaderParameters
{
    matrix inverseViewProjectionMatrix;
    float4 cameraPosition;
};

When using Visual Studio 2012's graphic debugger I see that the last column of this matrix contains only NaNs, the other values are correct. So somehow my program is passing wrong values to the GPU.

However when I debugged my C++ code I see no NaNs in the matrix that I upload to the GPU. So what is going on? Below are the relevant parts of code

Struct for the constant buffer on the CPU:

struct ShaderParameters
{
    XMMATRIX inverseViewProjectionMatrix;
    XMFLOAT4 cameraPosition;
};

Buffer creation:

D3D11_BUFFER_DESC shaderParametersDesc;
shaderParametersDesc.BindFlags = D3D11_BIND_CONSTANT_BUFFER;
shaderParametersDesc.ByteWidth = sizeof(ShaderParameters);
shaderParametersDesc.CPUAccessFlags = D3D11_CPU_ACCESS_WRITE;
shaderParametersDesc.MiscFlags = 0;
shaderParametersDesc.StructureByteStride = 0;
device->CreateBuffer(&shaderParametersDesc, NULL, &m_shaderParameters);

Buffer setting (every frame)

D3D11_MAPPED_SUBRESOURCE mappedResource;
ShaderParameters* shaderParameters;
deviceContext->Map(m_shaderParameters, 0, D3D11_MAP_WRITE_DISCARD, 0, &mappedResource);
shaderParameters = (ShaderParameters*)mappedResource.pData;

XMMATRIX view = camera->GetViewMatrix();
XMMATRIX proj = camera->GetProjectionMatrix();
XMMATRIX mult = XMMatrixMultiply(view, proj);
XMMATRIX invp = XMMatrixInverse(nullptr, mult);

if(XMMatrixIsNaN(invp))
    throw std::exception("invp is NAN");

shaderParameters->inverseViewProjectionMatrix = XMMatrixTranspose(invp);
shaderParameters->cameraPosition = camera->GetPosition();

deviceContext->Unmap(m_shaderParameters, 0);
deviceContext->VSSetConstantBuffers(0, 1, &m_shaderParameters);

Note that the exception is never thrown and if I inspect the matrix manually there are sane values in there (see image)

EDIT: some asked (via twitter) to include some more shader source code as unused values might be seen by the debugger as NaN. So here is the complete vertex shader:

cbuffer ShaderParameters
{
    matrix inverseViewProjectionMatrix;
    float4 cameraPosition;
};

struct VertexInputType
{
    float4 position : POSITION;
    float4 uv       : TEXCOORD0;
};

struct PixelInputType
{
    float4 position         : SV_POSITION;      
    float4 direction        : TEXCOORD0;
    float4 cameraPosition   : TEXCOORD1;
};

PixelInputType main(VertexInputType input)
{
        PixelInputType output;

        input.position.w = 1.0f;        
        output.position = input.position;

        output.direction = mul(input.position, inverseViewProjectionMatrix) - cameraPosition;
        output.direction.xyz = normalize(output.direction.xyz);

        output.direction.w = 1.0f;  

        output.cameraPosition = cameraPosition;

        return output;
}

Answer 1

The person on Twitter was right.

Because you overwrite output.direction.w with 1.0f the shader optimizer will see that it doesn't need to use that row of the matrix to calculate it.

Try commenting out that line to see the effect - the disassembly will include an extra dp4 instruction when it's gone.