Re:Flavors of SIMD
by gnufx on Tuesday February 19, 2019 @12:58PM
(#19201124)
I can't comment on the GPU comments, but you may be better off leaving the vectorization to gcc than using the simd pragma. On haswell, it uses avx2, but not fma, so you loose a factor of two on GEMM, for instance. The GCC manual also gives an example for the ivdep pragma.
[ Parent
| on HN ]
Would not that be just a question of missing improvements?
If I recall correctly, OpenJDK can use FMA thanks to Intel contributions.
I don't see how openjdk is related to the openmp pragma. GCC has no problem using FMA if you just let it, avoiding the pragma which simply says "simd".