0

I try to compile the netlib hpl linpack benchmark, but I want to disable vectorization.

I removed all optimization flags in the make file and only have the flags -qno-openmp-simd -no-vec -no-simd set.

However, when I execute the binary xhpl, the walltime and FLOPs are the same as whenn all optimization flags are set and the code is vectorized. Further, likwid shows a vectorization ratio of ~98%.

I read that these flags do not inhibit libraries like MKL from being vectorized. The netlib HPL uses the Linear Algebra library (BLAS or VSIPL) from MKLROOT.

Can I force the compiler to disable vectorization when compiling this specific benchmark and if so, how?

Luke
  • 31
  • 9

1 Answers1

1

this is expected behavior as the most part of computations happen from withing BLAS or BSIPL which are already built.

Gennady.F
  • 571
  • 2
  • 7
  • Thanks. That is what I assumed. Simply put, if our system had problems with vectorization and the avx units, we would notice them with significantly less performance? What I tried to do is make one hpl run with and one with no vectorization at all. Is there any way to pull this off without root access on the machine? – Luke Jun 02 '20 at 08:06