Matrix layout modifiers ignored

[zopsicle]

Consider the following shaders:

[shader("vertex")]
void Main(
    uniform row_major float4x4 worldToClip,
    out float4 SV_Position : SV_Position,
    in row_major float4x4 instanceMeshToWorld : instanceMeshToWorld,
    in float3 vertexPositionInMesh : vertexPositionInMesh
)
{
    SV_Position = mul(
        mul(worldToClip, instanceMeshToWorld),
        float4(vertexPositionInMesh, 1.f),
    );
}

[shader("vertex")]
void Main(
    uniform column_major float4x4 worldToClip,
    out float4 SV_Position : SV_Position,
    in column_major float4x4 instanceMeshToWorld : instanceMeshToWorld,
    in float3 vertexPositionInMesh : vertexPositionInMesh
)
{
    SV_Position = mul(
        mul(worldToClip, instanceMeshToWorld),
        float4(vertexPositionInMesh, 1.f),
    );
}

and their diff:

 void Main(
-    uniform row_major float4x4 worldToClip,
+    uniform column_major float4x4 worldToClip,
     out float4 SV_Position : SV_Position,
-    in row_major float4x4 instanceMeshToWorld : instanceMeshToWorld,
+    in column_major float4x4 instanceMeshToWorld : instanceMeshToWorld,
     in float3 vertexPositionInMesh : vertexPositionInMesh
 )
 {
     SV_Position = mul(
         mul(worldToClip, instanceMeshToWorld),
         float4(vertexPositionInMesh, 1.f),
     );
 }

Now let's compile them to HLSL:

$ slangc -entry Main -profile sm_5_0 -o row_major.hlsl row_major.slang
$ slangc -entry Main -profile sm_5_0 -o column_major.hlsl column_major.slang

and compare the results:

$ diff row_major.hlsl column_major.hlsl
$ echo $?
0

The same code is generated for both; the matrix layout modifiers are ignored. As a result, the shader generates the wrong output.

Using Slang v2024.9.2.

[zopsicle]

(I am under the impression that the purpose of these modifiers is to specify how the matrices should be loaded from the constant and vertex buffers. Maybe I am mistaken and their purpose is something else.)

[jkwak-work]

I don't think there is a reason to support column major memory layout for HLSL, because DirectX API doesn't support a way to upload a matrix in a column major memory layout from CPU to GPU.

We may need to print a warning for this.

[zopsicle]

Uploading matrices in column major layout is definitely supported. After all, the uploading API just takes an array of bytes. For instance:

class ColumnMajorMatrix
{
public:
    float col1[4];
    float col2[4];
    float col3[4];
    float col4[4];
};

ColumnMajorMatrix myMatrix{...};

D3D11_BUFFER_DESC{
    .ByteWidth{64},
    .Usage{D3D11_USAGE_IMMUTABLE},
    .BindFlags{D3D11_BIND_CONSTANT_BUFFER},
    .CPUAccessFlags{0},
    .MiscFlags{0},
    .StructureByteStride{},
};

D3D11_SUBRESOURCE_DATA initialData{
    .pSysMem{&myMatrix},
    .SysMemPitch{},
    .SysMemSlicePitch{},
};

ComPtr<ID3D11Buffer> buffer;
auto result = device->CreateBuffer(&desc, &initialData, &buffer);
_com_util::CheckError(result);

Now, reading this matrix from the buffer in a shader works as expected if the shader is compiled with a default matrix layout of column-major (set by sessionDesc.defaultMatrixLayoutMode = SLANG_MATRIX_LAYOUT_COLUMN_MAJOR; or the equivalent CLI option).

[jkwak-work]

I see. I think we should fix it with the HLSL keywords, row_major and column_major described in the following link. https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-variable-syntax