Using gcgemm from CuBLAS

There is a cublasSgemm() function that uses sgemm related kernels. In the output of profiler, I see some kernels with names that contain gcgemm, e.g gcgemm_64x32_nt. I searched for the high level function call in order to reach such kernels, but found nothing.

These are also related to matrix multiplication. Maybe by setting the correct arguments in cublasSgemm() it is possible to use gcgemm. Is there any information for that. Thanks.

CUBLAS *gemm API functions map to myriad different kernels. The mapping depends on CUDA version, GPU architecture, and arguments passed (in particular: matrix sizes, matrix shapes, and transpose modes). These are internal implementation details that NVIDIA does not publish. They are subject to frequent change, without notice.

1 Like