I am wondering if cuBLAS supports callback functions similar to the cuFFT library as described here.
I would like to use the cuBLAS library for a general matrix-matrix multiplication. But before and after the GEMM I need to execute some operations, as listed below.
- complex matrix-matrix multiplication (GEMM)
- calculate absolute values
- integration over time
It would be desirable to call processing steps (1,3,4) by callback device functions during the execution of the cuBLAS GEMM Kernel to prevent/reduce the loading of in and output data from and to the global memory. Until now I could not find any information about this in the documentation of cuBLAS or in other sources. I would like to know if cuBLAS supports callback function or will support it in future releases. If not, is there possibly another alternative without writing a complete GEMM kernel?
Thanks in advance!