ICX AVX512 support
Tests with the Intel icx compiler show that performance on COSMA7 is somewhat below that of the old icc, this is not seen on COSMA8. This suggests that AVX512 vectorization may not be working as well.
We currently use the CFLAGS -march=skylake-avx512
and -mavx512dq
, but no GRAVITY_CFLAGS
as we do
for icc, which uses -qopt-zmm-usage=high
and GCC for which we use -mprefer-vector-width=512
. These
are supposed to enable greater use of the zmm registers.
Being closer to GCC (as it is clang) icx supports the -mprefer-vector-width=512
option, however,
tests with an EAGLE 50 lowz gravity/hydro volume show that the gravity tasks are slower, this is
also seen in the total runtime.
So for now the gap remains unexplained, and we should continue without any GRAVITY_CFLAGS
.