On Thu, Mar 2, 2017 at 9:50 AM, Thomas Koenig <tkoe...@netcologne.de> wrote: > Am 02.03.2017 um 08:32 schrieb Janne Blomqvist: >> >> On Wed, Mar 1, 2017 at 11:00 PM, Thomas Koenig <tkoe...@netcologne.de> >> wrote: >>> >>> Hello world, >>> >>> the attached patch enables FMA for the AVX2 and AVX512F variants of >>> matmul. This should bring a very nice speedup (although I have >>> been unable to run benchmarks due to lack of a suitable machine). >> >> >> In lieu of benchmarks, have you looked at the generated asm to verify >> that fma is actually used? > > > Yes, I did. > > Here's something from the new matmul_r8_avx2: > > 156c: c4 62 e5 b8 fd vfmadd231pd %ymm5,%ymm3,%ymm15 > 1571: c4 c1 79 10 04 06 vmovupd (%r14,%rax,1),%xmm0 > 1577: c4 62 dd b8 db vfmadd231pd %ymm3,%ymm4,%ymm11 > 157c: c4 c3 7d 18 44 06 10 vinsertf128 > $0x1,0x10(%r14,%rax,1),%ymm0,%ymm0 > 1583: 01 > 1584: c4 62 ed b8 ed vfmadd231pd %ymm5,%ymm2,%ymm13 > 1589: c4 e2 ed b8 fc vfmadd231pd %ymm4,%ymm2,%ymm7 > 158e: c4 e2 fd a8 ad 30 ff vfmadd213pd > -0x800d0(%rbp),%ymm0,%ymm5
Great, looks good! > ... and here from matmul_r8_avx512f: > > 1da8: c4 a1 7b 10 14 d6 vmovsd (%rsi,%r10,8),%xmm2 > 1dae: c4 c2 b1 b9 f0 vfmadd231sd %xmm8,%xmm9,%xmm6 > 1db3: 62 62 ed 08 b9 e5 vfmadd231sd %xmm5,%xmm2,%xmm28 > 1db9: 62 62 ed 08 b9 ec vfmadd231sd %xmm4,%xmm2,%xmm29 > 1dbf: 62 62 ed 08 b9 f3 vfmadd231sd %xmm3,%xmm2,%xmm30 > 1dc5: c4 e2 91 99 e8 vfmadd132sd %xmm0,%xmm13,%xmm5 > 1dca: c4 e2 99 99 e0 vfmadd132sd %xmm0,%xmm12,%xmm4 > 1dcf: c4 e2 a1 99 d8 vfmadd132sd %xmm0,%xmm11,%xmm3 > 1dd4: c4 c2 a9 99 d1 vfmadd132sd %xmm9,%xmm10,%xmm2 > 1dd9: c4 c2 89 99 c1 vfmadd132sd %xmm9,%xmm14,%xmm0 > 1dde: 0f 8e d3 fe ff ff jle 1cb7 > <matmul_r8_avx512f+0x1cb7> Good, it's using fma, but why is this using xmm registers? That would mean it's operating only on 128 bit blocks at a time so no better than plain AVX. AFAIU avx512 should use zmm registers to operate on 512 bit chunks. I guess this is not due to your patch, but some other issue. -- Janne Blomqvist
