Right now, the GEMM function for this library relies on kernels ported from CLBlast.
While these have reasonable-enough performance, they do not take advantage of dedicated matmul acceleration instruction sets present on some modern GPUs, provided by VK_KHR_cooperative_matrix.
Implementing a GEMM kernel that uses this capability would result in significant performance improvements on applicable hardware.
Right now, the GEMM function for this library relies on kernels ported from CLBlast.
While these have reasonable-enough performance, they do not take advantage of dedicated matmul acceleration instruction sets present on some modern GPUs, provided by
VK_KHR_cooperative_matrix.Implementing a GEMM kernel that uses this capability would result in significant performance improvements on applicable hardware.