Wgpu: [DX12] Readonly StorageTexture binding does not create UAV binding

[Wumpf]

Minimal repro here:
https://github.com/Wumpf/wgpu-rs/commit/a00e08ceced718cf90553fc544052f023fb78516
Note that this builds on top of the fix from #572/#573 - haven't tried again, but I'm quite sure the particular case would work before the fix.

Steps:

• Use a uniform readonly image2D image. Note that you actually need to use the image, otherwise DX validation layer won't complain (does not influence hlsl conversion though).
  This is translated in hlsl to a RWTexture2D image, so a UAV, not a SRV as one might expect!
• Use then then the correct binding:

wgpu::BindingType::StorageTexture {
                dimension: wgpu::TextureViewDimension::D3,
                component_type: wgpu::TextureComponentType::Uint,
                format: wgpu::TextureFormat::R32Uint,
                readonly: true,
             }

• The resulting root signature apparently doesn't have a UAV: DX12 validation layer gets rather unhappy:

D3D12 ERROR: ID3D12Device::CreateGraphicsPipelineState: Root Signature doesn't match Vertex Shader: Shader UAV descriptor range (BaseShaderRegister=0, NumDescriptors=1, RegisterSpace=0) is not fully bound in root signature [ STATE_CREATION ERROR #688: CREATEGRAPHICSPIPELINESTATE_VS_ROOT_SIGNATURE_MISMATCH]

Without diving into wgpu/gfx-hal it looks like either a UAV binding is simply missing or the assumption was that the HLSL shader wouldn't use a UAV to begin with? Arguably it would have made more sense if it would convert everything to a SRV and changes the respective loads to texture.Load

Full converted hlsl shader code from the example repro for reference:

   RWTexture2D<float4> testimage : register(u0, space0);

    static float4 gl_Position;
    static int gl_VertexIndex;
    struct SPIRV_Cross_Input
    {
        uint gl_VertexIndex : SV_VertexID;
    };

    struct SPIRV_Cross_Output
    {
        float4 gl_Position : SV_Position;
    };

    void vert_main()
    {
        float2 position = float2(float(gl_VertexIndex), float((gl_VertexIndex & 1) * 2)) - 1.0f.xx;
        gl_Position = float4(position, 0.0f, 1.0f) + testimage[int2(0, 0)];
    }

    SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
    {
        gl_VertexIndex = int(stage_input.gl_VertexIndex);
        vert_main();
        SPIRV_Cross_Output stage_output;
        stage_output.gl_Position = gl_Position;
        return stage_output;
    }

[grovesNL]

Probably related to https://github.com/KhronosGroup/SPIRV-Cross/issues/1306, we might just need to update spirv_cross to use this flag

[kvark]

Oh snap, I thought this is already the way it works. I recall working on it in https://github.com/gfx-rs/gfx/pull/3176 and testing on gfx-ocean, which uses a bunch of read-only descriptors now :thinking:

[cwfitzgerald]

Is this still an issue on master?

[cwfitzgerald]

Looking at your repro closely it looks like you are creating a bind group that doesn't match the bind group layout (layout has one binding, bind group has none) so I'm going to call this closed for now, if you have an updated example that still has this problem, feel free to reopen!

[Wumpf]

apologies for not following up on this one. Thanks for digging in and pointing out that mistake!