From 83b15bfcddfe4a1fdc656300c3ede84c534766d6 Mon Sep 17 00:00:00 2001 From: paboyle Date: Wed, 30 Mar 2016 08:39:39 +0100 Subject: [PATCH] Better Avx512 assembly sequence for SU3 using fmaddsub to get the imag imag sign --- lib/simd/Intel512wilson.h | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/lib/simd/Intel512wilson.h b/lib/simd/Intel512wilson.h index 9c746a39..2d0e1e35 100644 --- a/lib/simd/Intel512wilson.h +++ b/lib/simd/Intel512wilson.h @@ -290,17 +290,17 @@ Author: paboyle ZEND2(UChi_02,Z4,Chi_02) \ ZEND2(UChi_12,Z5,Chi_12) ); -#define MULT_2SPINa(ptr) MULT_2SPIN_PF(ptr,ptr,VPREFETCHG); -#define MULT_2SPIN(ptr) MULT_ADDSUB_2SPIN(ptr); +#define MULT_2SPIN_PFXM(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFYM(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFZM(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFTM(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFTP(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFZP(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFYP(ptr,pf) MULT_2SPIN(ptr) +#define MULT_2SPIN_PFXP(ptr,pf) MULT_2SPIN(ptr) -#define MULT_2SPIN_PFXM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) -#define MULT_2SPIN_PFYM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) -#define MULT_2SPIN_PFZM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) -#define MULT_2SPIN_PFTM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) -#define MULT_2SPIN_PFTP(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) -#define MULT_2SPIN_PFZP(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) -#define MULT_2SPIN_PFYP(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCH) -#define MULT_2SPIN_PFXP(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA) +#define MULT_2SPINa(ptr) MULT_2SPIN_PF(ptr,ptr,VPREFETCHG); +#define MULT_2SPIN(ptr) MULT_ADDSUB_2SPIN(ptr); #define MULT_2SPIN_PF(ptr,pf,VPF) \