You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Store a matrix to groupshared memory. If the groupshared memory is any type other than i32 or vector of i32, the matrix's component type must match the scalar type or vector component type of the groupshared array. If the groupshared memory is i32 or vector of N by i32 (< N x i32 >) the matrix may be of any type. Validation will verify this requirement.
Validation rules will enforce that:
The matrix scope must be Wave or ThreadGroup
group shared target memory is large enough for the write
Validation should enforce that the groupshared arrays in dx.op.linAlgMatrixStoreToMemory & dx.op.linAlgMatrixLoadFromMemory are either arrays of scalars or vectors with element types that match the matrix element type or that the groupshared array's element type is i32.
It is legal to load and store any type of matrix to i32 groupshared memory, but not to mismatch other types.
Based on microsoft/hlsl-specs#879, new validation should be added for dx.op.linAlgMatrixStoreToMemory, dx.op.linAlgMatrixLoadFromMemory, and dx.op.linAlgMatrixAccumulateToMemory that the stride and offset values (if they are constants) must be properly aligned. The offset must be 128-byte aligned, and the stride must be 16-byte aligned.
Activity
V-FEXrt commentedon Jul 27, 2026
updated per microsoft/hlsl-specs#902
V-FEXrt commentedon Jul 27, 2026
per #8635
Validation should enforce that the groupshared arrays in dx.op.linAlgMatrixStoreToMemory & dx.op.linAlgMatrixLoadFromMemory are either arrays of scalars or vectors with element types that match the matrix element type or that the groupshared array's element type is i32.
It is legal to load and store any type of matrix to i32 groupshared memory, but not to mismatch other types.
V-FEXrt commentedon Jul 27, 2026
per #8636
Based on microsoft/hlsl-specs#879, new validation should be added for dx.op.linAlgMatrixStoreToMemory, dx.op.linAlgMatrixLoadFromMemory, and dx.op.linAlgMatrixAccumulateToMemory that the stride and offset values (if they are constants) must be properly aligned. The offset must be 128-byte aligned, and the stride must be 16-byte aligned.
[SM6.10] LinAlg Validation: MatrixStoreToMemory
[SM6.10] LinAlg Validation: MatrixStoreToMemory (#8834)