[FEA] Use cuBLAS's hermitian transpose if hermitian operator is used
cliffburdick opened this issue · comments
Cliff Burdick commented
As commonly done in signal processing code a user can do:
matmul(c, a, hermitian(b));
But this is not currently possible since you cannot pass an operator into matmul
. Instead, a user must make a temporary copy of the hermitian'd matrix into another tensor, then call matmul. Instead, we can detect if this is a primary hermitian operator and pass this to the underlying GEMM library rather than adding the extra copy.