Autotuning for GEMM kernel and combination with other kernels

Accelerated Computing GPU-Accelerated Libraries
1

Hi,

Do we have some package or parameter to do an offline or online optimization on the code? e.g. my GEMM kernel is far from peak performance (25% degradation).

2

If you’re using cuBLAS, internal heuristics will pick the best kernel based on a number of parameters. Peak performance depends on numerous factors.