Why there is always a memset kernel before a cublas matrix multiplication kernel?

Accelerated Computing GPU-Accelerated Libraries
1

system: ubuntu2004
GPU: A100 SXM 40GB
cuda version: 11.3
cuda driver version: 470.42.01
cublas version: 11.4.2.10064

image
image2468×54 9.08 KB

image
image590×520 34.4 KB

image
image1126×908 59.8 KB

2

Maybe BLAS primarily implements multiply-and-add kernels, and in order to just multiply, the result must be nulled out first.