Performance Overhead for Irregular Matrices

[EJainDev]

The CLBlast API converts hermitian, symmetric, and triangular matrices to regular matrices, and then calls regular GEMM. Doesn’t this incur 2 types of overhead: extra computation + bandwidth and memory allocation + copy, that could be avoided by custom kernels?

[CNugteren]

Indeed, this conversion incurs overhead. However, a custom kernel might be much slower than paying the overhead, so not sure if there is anything to gain. Furthermore, such computations are perhaps not the most common either, so the question arise whether it is worth spending effort into these routines.

But we can leave this open in case someone wants to investigate.