Memset and GEMM kernel


I am checking my trace with nsys-ui. In some cases, I am seeing memset before GEMM kernels and sometimes I am not seeing anything before GEMM kernels. Almost for small size GEMM I am seeing this operation. Why do we have memset and why it is not for all GEMMs kernels?

We really can’t help without more context. What functions are you running? Reproducer code would help. The nsys-ui output would be helpful.