4 ms·
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 f
by gnufx 8y ago
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.
- pjmlp 8y agoWould not that be just a question of missing improvements? If I recall correctly, OpenJDK can use FMA thanks to Intel contributions.
- gnufx 8y agoI 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".
- pjmlp 8y agoI understood that GCC auto-vectorization wouldn't do it currently, and hence gave an example where auto-vectorization does make use of it, assuming I remember Intel's session at CodeONE correctly.