Load from multisampled texture fails on some Intel GPUs

Question

I’ve implemented mouse picking in my app by using a stencil buffer.

Here’s the pixel shader that reads value under the mouse:

Texture2DMS<uint2> depthStencilTexture : register( t0 );

cbuffer ReadDepthInput : register( b2 )
{
    int2 readPosition;
}

// Produces vector3( x, y, stencilValue )
float4 main() : SV_Target0
{
    uint2 res = depthStencilTexture.Load( readPosition, 1 );
    float stencil = res.y;
    // Scale the result from integers to 0..1
    stencil /= 255.0f;
    return float4( float( readPosition.x ) / 65535.0f, float( readPosition.y ) / 65535.0f, stencil, 0 );
}

The code works on nVidia and AMD GPUs, works on Intel Iris 550.

The code fails on Intel Haswell GPUs (Intel HD 5000, Intel HD 4600). Texture2DMS.Load just returns 0. Any ideas what might be wrong?

Update: Only happens with 8x MSAA. Decreasing it to 4x works OK even on those affected Intel GPUs.

Did you confirm that the driver reports support for ``D3D11_FORMAT_SUPPORT_MULTISAMPLE_LOAD`` using ``CheckFormatSupport`` and 8x MSAA using ``CheckMultisampleQualityLevels`` for that format? — Chuck Walbourn, Apr 16 '17 at 06:13
@ChuckWalbourn yes, my app calls them both at startup, and disables MSAA / reduce the level. — Soonts, Apr 16 '17 at 12:16

Soonts · Accepted Answer · 2017-07-31T21:45:31.750

Looks like there’s a bug in Intel GPU driver or hardware.

Here’s a workaround:

HRESULT isIntelHaswellGPU( ID3D11Device* device )
{
    CComQIPtr<IDXGIDevice> dxgiDev = device;
    if( !dxgiDev )
        return E_NOINTERFACE;

    CComPtr<IDXGIAdapter> adapter;
    CHECK( dxgiDev->GetAdapter( &adapter ) );

    DXGI_ADAPTER_DESC desc;
    CHECK( adapter->GetDesc( &desc ) );

    if( desc.VendorId != 0x8086 )
        return S_FALSE; // nVidia or AMD GPU

    // https://en.wikipedia.org/wiki/List_of_Intel_graphics_processing_units#Seventh_generation
    // https://github.com/GameTechDev/gpudetect/blob/master/DeviceId.cpp
    const UINT maskedDeviceId = ( desc.DeviceId & 0xFF00 );
    if( maskedDeviceId == 0x0400 || maskedDeviceId == 0x0A00 || maskedDeviceId == 0x0D00 )
        return S_OK;    // Detected Intel Haswell GPU
    return S_FALSE;
}

If this function returns S_OK, I limit max. MSAA level to 4.

Load from multisampled texture fails on some Intel GPUs

1 Answers1

Linked