portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-16 14:57:05 +01:00
Code Releases Activity

Large change with KNL preparation

Browse Source
This commit is contained in:
paboyle
2016-06-03 03:24:26 -07:00
parent 1c0e922585
commit 139cc5f1ae
26 changed files with 1810 additions and 1705 deletions
Download Patch File Download Diff File Expand all files Collapse all files

592
lib/simd/Intel512wilson.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -27,9 +27,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
#ifndef GRID_ASM_INTEL_512_QCD_H
#define GRID_ASM_INTEL_512_QCD_H
//////////////////////////////////////////////////////////////////////////////////////////
// Register allocations for Wilson Kernel are precision and IMCI/AVX512 indept
// Register allocations for Wilson Kernel are precision indept
//////////////////////////////////////////////////////////////////////////////////////////
#define result_00 %zmm0
#define result_01 %zmm1
@ -64,7 +64,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define UChi_12 %zmm23
#define Uir %zmm24
//#define ONE %zmm24
#define Uri %zmm25
#define T1 %zmm24
#define T2 %zmm25
@ -92,13 +91,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define Chimu_32 UChi_12
#include <simd/Intel512common.h>
#ifdef AVX512
#include <simd/Intel512avx.h>
//#include <simd/Intel512avxAddsub.h> // Alternate implementation
#endif
#ifdef IMCI
#include <simd/Intel512imci.h>
#endif
//////////////////////////////////////////////////////////////////
// Macros used to build wilson kernel -- can rationalise and simplify
@ -193,47 +186,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
VSTORE(11,%r8,result_32) \
);
// auto ptr = &U._odata[sU](A);
// A plan for lifting loads
// can use Z2/3/4/5/U/U for U field in first step.
// can use Chi_00, Chi_10, U U for U field in second step
// can use Chi_00, Chi_10, Chi_01,11, U U for U field in third step
// Enables to lift ALL loads earlier by a few cycles and alleviate OoO pressure if needed.
// KNL is DUAL issue for FP, and lifting these loads is potentially important.
// Need detailed profile data to be sure.
#if 0
#define PREFETCH_U(A) \
LOAD64(%r8,&U._odata[sU](A)) \
__asm__ ( \
VPREFETCHG(0,%r8) \
VPREFETCHG(1,%r8) \
VPREFETCHG(2,%r8) \
VPREFETCHG(3,%r8) \
VPREFETCHG(4,%r8) \
VPREFETCHG(5,%r8) \
VPREFETCHG(6,%r8) \
VPREFETCHG(7,%r8) \
VPREFETCHG(8,%r8) );
#define PREFETCH_R(A) \
LOAD64(%r8,&out._odata[ss]) \
__asm__ ( \
VPREFETCHW(0,%r8) \
VPREFETCHW(1,%r8) \
VPREFETCHW(2,%r8) \
VPREFETCHW(3,%r8) \
VPREFETCHW(4,%r8) \
VPREFETCHW(5,%r8) \
VPREFETCHW(6,%r8) \
VPREFETCHW(7,%r8) \
VPREFETCHW(8,%r8) \
VPREFETCHW(9,%r8) \
VPREFETCHW(10,%r8) \
VPREFETCHW(11,%r8) );
#endif
#define MULT_2SPIN_DIR(A) MULT_2SPIN(&U._odata[sU](A))
#define MULT_2SPIN_DIR_PFXP(A,p) MULT_2SPIN_PFXP(&U._odata[sU](A),p)
#define MULT_2SPIN_DIR_PFYP(A,p) MULT_2SPIN_PFYP(&U._odata[sU](A),p)
#define MULT_2SPIN_DIR_PFZP(A,p) MULT_2SPIN_PFZP(&U._odata[sU](A),p)
@ -244,131 +196,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define MULT_2SPIN_DIR_PFZM(A,p) MULT_2SPIN_PFZM(&U._odata[sU](A),p)
#define MULT_2SPIN_DIR_PFTM(A,p) MULT_2SPIN_PFTM(&U._odata[sU](A),p)
#if 0
#define MULT_2SPIN_UNOPT(ptr) \
LOAD64(%r8,ptr) \
__asm__ ( \
ZLOAD (0,%r8,UChi_01,UChi_11) \
ZLOAD (3,%r8,UChi_02,UChi_12) \
ZLOAD (6,%r8,Uri,Uir) \
ZMUL (UChi_01,UChi_11,Chi_00,UChi_00,Z0) \
ZMUL (UChi_01,UChi_11,Chi_10,UChi_10,Z1) \
ZMUL (UChi_02,UChi_12,Chi_00,UChi_01,Z2) \
ZMUL (UChi_02,UChi_12,Chi_10,UChi_11,Z3) \
ZMUL (Uri,Uir, Chi_00,UChi_02,Z4) \
ZMUL (Uri,Uir, Chi_10,UChi_12,Z5) \
\
ZLOAD (1,%r8,Uri,Uir) \
ZLOAD (4,%r8,Chi_00, Chi_10) \
ZMADD (Uri,Uir, Chi_01,UChi_00,Z0) \
ZMADD (Uri,Uir, Chi_11,UChi_10,Z1) \
ZLOAD (7,%r8,Uri,Uir) \
ZMADD (Chi_00, Chi_10,Chi_01,UChi_01,Z2) \
ZMADD (Chi_00, Chi_10,Chi_11,UChi_11,Z3) \
ZLOAD (2,%r8,Chi_00,Chi_10) \
ZMADD(Uri,Uir, Chi_01,UChi_02,Z4) \
ZMADD(Uri,Uir, Chi_11,UChi_12,Z5) \
\
ZLOAD (5,%r8,Uri,Uir) \
ZMADD (Chi_00,Chi_10, Chi_02,UChi_00,Z0) \
ZMADD (Chi_00,Chi_10, Chi_12,UChi_10,Z1) \
ZLOAD (8,%r8,Chi_00,Chi_10) \
ZMADD (Uri,Uir, Chi_02,UChi_01,Z2) \
ZMADD (Uri,Uir, Chi_12,UChi_11,Z3) \
ZMADD(Chi_00,Chi_10, Chi_02,UChi_02,Z4) \
ZMADD(Chi_00,Chi_10, Chi_12,UChi_12,Z5) \
\
ZEND1(UChi_00,Z0,Chi_01) \
ZEND1(UChi_10,Z1,Chi_11) \
ZEND1(UChi_01,Z2,Chi_00) \
ZEND1(UChi_11,Z3,Chi_10) \
ZEND1(UChi_02,Z4,Chi_02) \
ZEND1(UChi_12,Z5,Chi_12) \
ZEND2(UChi_00,Z0,Chi_01) \
ZEND2(UChi_10,Z1,Chi_11) \
ZEND2(UChi_01,Z2,Chi_00) \
ZEND2(UChi_11,Z3,Chi_10) \
ZEND2(UChi_02,Z4,Chi_02) \
ZEND2(UChi_12,Z5,Chi_12) );
#endif
#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)
// MULT_2SPINa(ptr) MULT_2SPIN_PF(ptr,ptr,VPREFETCHG);
#if 0
#define MULT_2SPIN_PF(ptr,pf,VPF) \
LOAD64(%r8,ptr) \
LOAD64(%r9,pf) \
__asm__ ( \
ZMULMEM2SP(0,%r8,Uri,Chi_00,Chi_10,UChi_00,Z0,UChi_10,Z1) \
VPF(0,%r9) \
ZMULMEM2SP(3,%r8,Uri,Chi_00,Chi_10,UChi_01,Z2,UChi_11,Z3) \
VPF(1,%r9) \
ZMULMEM2SP(6,%r8,Uri,Chi_00,Chi_10,UChi_02,Z4,UChi_12,Z5) \
VPF(2,%r9) \
\
ZMADDMEM2SP(1,%r8,Uri,Chi_01,Chi_11,UChi_00,Z0,UChi_10,Z1) \
VPF(3,%r9) \
ZMADDMEM2SP(4,%r8,Uri,Chi_01,Chi_11,UChi_01,Z2,UChi_11,Z3) \
VPF(4,%r9) \
ZMADDMEM2SP(7,%r8,Uri,Chi_01,Chi_11,UChi_02,Z4,UChi_12,Z5) \
VPF(5,%r9) \
\
ZMADDMEM2SP(2,%r8,Uri,Chi_02,Chi_12,UChi_00,Z0,UChi_10,Z1) \
VPF(6,%r9) \
ZMADDMEM2SP(5,%r8,Uri,Chi_02,Chi_12,UChi_01,Z2,UChi_11,Z3) \
VPF(7,%r9) \
ZMADDMEM2SP(8,%r8,Uri,Chi_02,Chi_12,UChi_02,Z4,UChi_12,Z5) \
VPF(8,%r9) \
\
ZEND1(UChi_00,Z0,Chi_01) \
ZEND1(UChi_10,Z1,Chi_11) \
ZEND1(UChi_01,Z2,Chi_00) \
ZEND1(UChi_11,Z3,Chi_10) \
VPF(9,%r9) \
ZEND1(UChi_02,Z4,Chi_02) \
ZEND1(UChi_12,Z5,Chi_12) \
ZEND2(UChi_00,Z0,Chi_01) \
ZEND2(UChi_10,Z1,Chi_11) \
VPF(10,%r9) \
ZEND2(UChi_01,Z2,Chi_00) \
ZEND2(UChi_11,Z3,Chi_10) \
ZEND2(UChi_02,Z4,Chi_02) \
VPF(11,%r9) \
ZEND2(UChi_12,Z5,Chi_12) );
#endif
#if 0
#define MULT_2SPIN_PFNONE(ptr,pf,VPF) \
LOAD64(%r8,ptr) \
LOAD64(%r9,pf) \
__asm__ ( \
VPF(0,%r9) \
VPF(1,%r9) \
VPF(2,%r9) \
\
VPF(3,%r9) \
VPF(4,%r9) \
VPF(5,%r9) \
\
VPF(6,%r9) \
VPF(7,%r9) \
VPF(8,%r9) \
\
VPF(9,%r9) \
VPF(10,%r9) \
VPF(11,%r9) );
#endif
// Pretty much Perfectly Pipelined
#define MULT_2SPIN_PFXM(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFYM(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFZM(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFTM(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFTP(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFZP(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFYP(ptr,pf) MULT_2SPIN(ptr,pf)
#define MULT_2SPIN_PFXP(ptr,pf) MULT_2SPIN(ptr,pf)
//////////////////////////////////////////////////////////////////
// Dirac algebra
@ -490,7 +325,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
LOAD64(%r8,ptr) \
__asm__ ( \
LOAD_CHIMU01i \
VSUBMEM(6,%r8 ,Chimu_00,Chi_00) \
VSUBMEM(6,%r8,Chimu_00,Chi_00) \
VSUBMEM(7,%r8,Chimu_01,Chi_01) \
VSUBMEM(8,%r8,Chimu_02,Chi_02) \
VSUBMEM(9,%r8,Chimu_10,Chi_10) \
@ -503,18 +338,18 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
// fspin(3)=timesMinusI(hspin(0))
#define XP_RECON __asm__ ( \
VZERO(TMP) \
VMOV(UChi_00,result_00) \
VMOV(UChi_01,result_01) \
VMOV(UChi_02,result_02) \
VMOV(UChi_10,result_10) \
VMOV(UChi_11,result_11) \
VMOV(UChi_12,result_12) \
VTIMESMINUSI0(UChi_10,result_20,TMP) \
VTIMESMINUSI0(UChi_11,result_21,TMP) \
VTIMESMINUSI0(UChi_12,result_22,TMP) \
VTIMESMINUSI0(UChi_00,result_30,TMP) \
VTIMESMINUSI0(UChi_10,result_20,TMP) \
VTIMESMINUSI0(UChi_01,result_31,TMP) \
VTIMESMINUSI0(UChi_11,result_21,TMP) \
VTIMESMINUSI0(UChi_02,result_32,TMP) \
VTIMESMINUSI0(UChi_12,result_22,TMP) \
VMOV(UChi_00,result_00) \
VMOV(UChi_10,result_10) \
VMOV(UChi_01,result_01) \
VMOV(UChi_11,result_11) \
VMOV(UChi_02,result_02) \
VMOV(UChi_12,result_12) \
VTIMESMINUSI1(UChi_10,result_20,TMP) \
VTIMESMINUSI1(UChi_11,result_21,TMP) \
VTIMESMINUSI1(UChi_12,result_22,TMP) \
@ -531,24 +366,24 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
// NB could save 6 ops using addsub => 12 cycles
#define XP_RECON_ACCUM __asm__ ( \
VZERO(TMP)\
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESMINUSI0(UChi_10,result_20,Z0)\
VACCTIMESMINUSI0(UChi_11,result_21,Z1)\
VACCTIMESMINUSI0(UChi_12,result_22,Z2)\
VACCTIMESMINUSI0(UChi_00,result_30,Z3)\
VACCTIMESMINUSI0(UChi_10,result_20,Z0)\
VACCTIMESMINUSI0(UChi_01,result_31,Z4)\
VACCTIMESMINUSI0(UChi_11,result_21,Z1)\
VACCTIMESMINUSI0(UChi_02,result_32,Z5)\
VACCTIMESMINUSI1(UChi_10,result_20,Z0)\
VACCTIMESMINUSI1(UChi_11,result_21,Z1)\
VACCTIMESMINUSI1(UChi_12,result_22,Z2)\
VACCTIMESMINUSI0(UChi_12,result_22,Z2)\
VADD(UChi_00,result_00,result_00)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESMINUSI1(UChi_00,result_30,Z3)\
VACCTIMESMINUSI1(UChi_10,result_20,Z0)\
VACCTIMESMINUSI1(UChi_01,result_31,Z4)\
VACCTIMESMINUSI1(UChi_11,result_21,Z1)\
VACCTIMESMINUSI1(UChi_02,result_32,Z5)\
VACCTIMESMINUSI1(UChi_12,result_22,Z2)\
VACCTIMESMINUSI2(UChi_10,result_20,Z0)\
VACCTIMESMINUSI2(UChi_11,result_21,Z1)\
VACCTIMESMINUSI2(UChi_12,result_22,Z2)\
@ -559,24 +394,24 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define XM_RECON __asm__ ( \
VZERO(TMP)\
VMOV(UChi_00,result_00)\
VMOV(UChi_01,result_01)\
VMOV(UChi_02,result_02)\
VMOV(UChi_10,result_10)\
VMOV(UChi_11,result_11)\
VMOV(UChi_12,result_12)\
VTIMESI0(UChi_10,result_20,TMP)\
VTIMESI0(UChi_11,result_21,TMP)\
VTIMESI0(UChi_12,result_22,TMP)\
VTIMESI0(UChi_00,result_30,TMP)\
VTIMESI0(UChi_10,result_20,TMP)\
VTIMESI0(UChi_01,result_31,TMP)\
VTIMESI0(UChi_11,result_21,TMP)\
VTIMESI0(UChi_02,result_32,TMP)\
VTIMESI1(UChi_10,result_20,TMP)\
VTIMESI1(UChi_11,result_21,TMP)\
VTIMESI1(UChi_12,result_22,TMP)\
VTIMESI0(UChi_12,result_22,TMP)\
VMOV(UChi_00,result_00)\
VMOV(UChi_10,result_10)\
VMOV(UChi_01,result_01)\
VMOV(UChi_11,result_11)\
VMOV(UChi_02,result_02)\
VMOV(UChi_12,result_12)\
VTIMESI1(UChi_00,result_30,TMP)\
VTIMESI1(UChi_10,result_20,TMP)\
VTIMESI1(UChi_01,result_31,TMP)\
VTIMESI1(UChi_11,result_21,TMP)\
VTIMESI1(UChi_02,result_32,TMP)\
VTIMESI1(UChi_12,result_22,TMP)\
VTIMESI2(UChi_10,result_20,TMP)\
VTIMESI2(UChi_11,result_21,TMP)\
VTIMESI2(UChi_12,result_22,TMP)\
@ -586,23 +421,25 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
);
#define XM_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESI0(UChi_10,result_20,Z0)\
VACCTIMESI0(UChi_11,result_21,Z1)\
VACCTIMESI0(UChi_12,result_22,Z2)\
VACCTIMESI0(UChi_00,result_30,Z3)\
VACCTIMESI0(UChi_11,result_21,Z1)\
VACCTIMESI0(UChi_01,result_31,Z4)\
VACCTIMESI0(UChi_12,result_22,Z2)\
VACCTIMESI0(UChi_02,result_32,Z5)\
\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_00,result_00,result_00)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_12,result_12,result_12)\
VADD(UChi_02,result_02,result_02)\
\
VACCTIMESI1(UChi_10,result_20,Z0)\
VACCTIMESI1(UChi_11,result_21,Z1)\
VACCTIMESI1(UChi_12,result_22,Z2)\
VACCTIMESI1(UChi_00,result_30,Z3)\
VACCTIMESI1(UChi_11,result_21,Z1)\
VACCTIMESI1(UChi_01,result_31,Z4)\
VACCTIMESI1(UChi_12,result_22,Z2)\
VACCTIMESI1(UChi_02,result_32,Z5)\
VACCTIMESI2(UChi_10,result_20,Z0)\
VACCTIMESI2(UChi_11,result_21,Z1)\
@ -614,10 +451,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define YP_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VADD(UChi_10,result_20,result_20)\
VADD(UChi_11,result_21,result_21)\
@ -628,10 +465,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define YM_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VSUB(UChi_10,result_20,result_20)\
VSUB(UChi_11,result_21,result_21)\
@ -641,23 +478,23 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
VADD(UChi_02,result_32,result_32) );
#define ZP_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESMINUSI0(UChi_00,result_20,Z0)\
VACCTIMESMINUSI0(UChi_01,result_21,Z1)\
VACCTIMESMINUSI0(UChi_02,result_22,Z2)\
VACCTIMESI0(UChi_10,result_30,Z3)\
VACCTIMESMINUSI0(UChi_01,result_21,Z1)\
VACCTIMESI0(UChi_11,result_31,Z4)\
VACCTIMESMINUSI0(UChi_02,result_22,Z2)\
VACCTIMESI0(UChi_12,result_32,Z5)\
VADD(UChi_00,result_00,result_00)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESMINUSI1(UChi_00,result_20,Z0)\
VACCTIMESMINUSI1(UChi_01,result_21,Z1)\
VACCTIMESMINUSI1(UChi_02,result_22,Z2)\
VACCTIMESI1(UChi_10,result_30,Z3)\
VACCTIMESMINUSI1(UChi_01,result_21,Z1)\
VACCTIMESI1(UChi_11,result_31,Z4)\
VACCTIMESMINUSI1(UChi_02,result_22,Z2)\
VACCTIMESI1(UChi_12,result_32,Z5)\
VACCTIMESMINUSI2(UChi_00,result_20,Z0)\
VACCTIMESMINUSI2(UChi_01,result_21,Z1)\
@ -668,23 +505,23 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
);
#define ZM_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESI0(UChi_00,result_20,Z0)\
VACCTIMESI0(UChi_01,result_21,Z1)\
VACCTIMESI0(UChi_02,result_22,Z2)\
VACCTIMESMINUSI0(UChi_10,result_30,Z3)\
VACCTIMESI0(UChi_01,result_21,Z1)\
VACCTIMESMINUSI0(UChi_11,result_31,Z4)\
VACCTIMESI0(UChi_02,result_22,Z2)\
VACCTIMESMINUSI0(UChi_12,result_32,Z5)\
VADD(UChi_00,result_00,result_00)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VACCTIMESI1(UChi_00,result_20,Z0)\
VACCTIMESI1(UChi_01,result_21,Z1)\
VACCTIMESI1(UChi_02,result_22,Z2)\
VACCTIMESMINUSI1(UChi_10,result_30,Z3)\
VACCTIMESI1(UChi_01,result_21,Z1)\
VACCTIMESMINUSI1(UChi_11,result_31,Z4)\
VACCTIMESI1(UChi_02,result_22,Z2)\
VACCTIMESMINUSI1(UChi_12,result_32,Z5)\
VACCTIMESI2(UChi_00,result_20,Z0)\
VACCTIMESI2(UChi_01,result_21,Z1)\
@ -696,30 +533,30 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define TP_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VADD(UChi_00,result_20,result_20)\
VADD(UChi_01,result_21,result_21)\
VADD(UChi_02,result_22,result_22)\
VADD(UChi_10,result_30,result_30)\
VADD(UChi_01,result_21,result_21)\
VADD(UChi_11,result_31,result_31)\
VADD(UChi_02,result_22,result_22)\
VADD(UChi_12,result_32,result_32) );
#define TM_RECON_ACCUM __asm__ ( \
VADD(UChi_00,result_00,result_00)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_10,result_10,result_10)\
VADD(UChi_01,result_01,result_01)\
VADD(UChi_11,result_11,result_11)\
VADD(UChi_02,result_02,result_02)\
VADD(UChi_12,result_12,result_12)\
VSUB(UChi_00,result_20,result_20)\
VSUB(UChi_01,result_21,result_21)\
VSUB(UChi_02,result_22,result_22)\
VSUB(UChi_10,result_30,result_30)\
VSUB(UChi_01,result_21,result_21)\
VSUB(UChi_11,result_31,result_31)\
VSUB(UChi_02,result_22,result_22)\
VSUB(UChi_12,result_32,result_32) );
//define PREFETCH_CHIMU(A)
@ -758,63 +595,200 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define MULT_ADDSUB_2SPIN1(ptr) \
LOAD64(%r8,ptr)
/*
* __asm__ ( \
);
VMUL(Z0,%zmm2,%zmm3) \
*/
#define MULT_ADDSUB_2SPIN(ptr) \
LOAD64(%r8,ptr) \
__asm__ ( \
VMOVIDUP(0,%r8,Z0 ) \
VMOVIDUP(3,%r8,Z1 )\
VMOVIDUP(6,%r8,Z2 )\
VSHUF(Chi_00,T1) \
VSHUF(Chi_10,T2) \
\
VMUL(Z0,T1,UChi_00) VMOVRDUP(0,%r8,Z3 ) \
VMUL(Z0,T2,UChi_10) VMOVRDUP(3,%r8,Z4 ) \
VMUL(Z1,T1,UChi_01) VMOVRDUP(6,%r8,Z5 ) \
VMUL(Z1,T2,UChi_11) VMOVIDUP(1,%r8,Z0 ) \
VMUL(Z2,T1,UChi_02) VMOVIDUP(4,%r8,Z1 ) \
VMUL(Z2,T2,UChi_12) VMOVIDUP(7,%r8,Z2 ) \
\
VMADDSUB(Z3,Chi_00,UChi_00) VSHUF(Chi_01,T1) \
VMADDSUB(Z3,Chi_10,UChi_10) VSHUF(Chi_11,T2) \
VMADDSUB(Z4,Chi_00,UChi_01) VMOVRDUP(1,%r8,Z3 ) \
VMADDSUB(Z4,Chi_10,UChi_11)\
VMADDSUB(Z5,Chi_00,UChi_02) VMOVRDUP(4,%r8,Z4 ) \
VMADDSUB(Z5,Chi_10,UChi_12)\
\
VMADDSUB(Z0,T1,UChi_00) VMOVRDUP(7,%r8,Z5 ) \
VMADDSUB(Z0,T2,UChi_10)\
VMADDSUB(Z1,T1,UChi_01) VMOVIDUP(2,%r8,Z0 ) \
VMADDSUB(Z1,T2,UChi_11)\
VMADDSUB(Z2,T1,UChi_02) VMOVIDUP(5,%r8,Z1 ) \
VMADDSUB(Z2,T2,UChi_12) VMOVIDUP(8,%r8,Z2 ) \
\
VMADDSUB(Z3,Chi_01,UChi_00) VSHUF(Chi_02,T1) \
VMADDSUB(Z3,Chi_11,UChi_10) VSHUF(Chi_12,T2) \
VMADDSUB(Z4,Chi_01,UChi_01) VMOVRDUP(2,%r8,Z3 ) \
VMADDSUB(Z4,Chi_11,UChi_11)\
VMADDSUB(Z5,Chi_01,UChi_02) VMOVRDUP(5,%r8,Z4 ) \
VMADDSUB(Z5,Chi_11,UChi_12)\
\
VMADDSUB(Z0,T1,UChi_00) VMOVRDUP(8,%r8,Z5 ) \
VMADDSUB(Z0,T2,UChi_10)\
VMADDSUB(Z1,T1,UChi_01)\
VMADDSUB(Z1,T2,UChi_11)\
VMADDSUB(Z2,T1,UChi_02)\
VMADDSUB(Z2,T2,UChi_12)\
\
VMADDSUB(Z3,Chi_02,UChi_00)\
VMADDSUB(Z3,Chi_12,UChi_10)\
VMADDSUB(Z4,Chi_02,UChi_01)\
VMADDSUB(Z4,Chi_12,UChi_11)\
VMADDSUB(Z5,Chi_02,UChi_02)\
VMADDSUB(Z5,Chi_12,UChi_12)\
);
#define MULT_2SPIN(ptr) MULT_ADDSUB_2SPIN(ptr)
#define MULT_ADDSUB_2SPIN(ptr) \
LOAD64(%r8,ptr) \
__asm__ ( \
VSHUF(Chi_00,T1) \
VMOVIDUP(0,%r8,Z0 ) \
VMOVIDUP(3,%r8,Z1 ) \
VMOVIDUP(6,%r8,Z2 ) VSHUF(Chi_10,T2) \
/*6*/ \
VMUL(Z0,T1,UChi_00) VMOVRDUP(0,%r8,Z3 ) \
VMUL(Z0,T2,UChi_10) VMOVRDUP(3,%r8,Z4 ) \
VMUL(Z1,T1,UChi_01) VMOVRDUP(6,%r8,Z5 ) \
VMUL(Z1,T2,UChi_11) VMOVIDUP(1,%r8,Z0 ) \
VMUL(Z2,T1,UChi_02) VMOVIDUP(4,%r8,Z1 ) \
VMUL(Z2,T2,UChi_12) VMOVIDUP(7,%r8,Z2 ) \
/*18*/ \
VMADDSUB(Z3,Chi_00,UChi_00) VSHUF(Chi_01,T1) \
VMADDSUB(Z3,Chi_10,UChi_10) \
VMADDSUB(Z4,Chi_00,UChi_01) VMOVRDUP(1,%r8,Z3 ) \
VMADDSUB(Z4,Chi_10,UChi_11) VSHUF(Chi_11,T2) \
VMADDSUB(Z5,Chi_00,UChi_02) VMOVRDUP(4,%r8,Z4 ) \
VMADDSUB(Z5,Chi_10,UChi_12) \
/*28*/ \
VMADDSUB(Z0,T1,UChi_00) VMOVRDUP(7,%r8,Z5 ) \
VMADDSUB(Z0,T2,UChi_10) \
VMADDSUB(Z1,T1,UChi_01) VMOVIDUP(2,%r8,Z0 ) \
VMADDSUB(Z1,T2,UChi_11) \
VMADDSUB(Z2,T1,UChi_02) VMOVIDUP(5,%r8,Z1 ) \
VMADDSUB(Z2,T2,UChi_12) VMOVIDUP(8,%r8,Z2 ) \
/*38*/ \
VMADDSUB(Z3,Chi_01,UChi_00) VSHUF(Chi_02,T1) \
VMADDSUB(Z3,Chi_11,UChi_10) \
VMADDSUB(Z4,Chi_01,UChi_01) VMOVRDUP(2,%r8,Z3 ) \
VMADDSUB(Z4,Chi_11,UChi_11) VSHUF(Chi_12,T2) \
VMADDSUB(Z5,Chi_01,UChi_02) VMOVRDUP(5,%r8,Z4 ) \
VMADDSUB(Z5,Chi_11,UChi_12) \
/*48*/ \
VMADDSUB(Z0,T1,UChi_00) VMOVRDUP(8,%r8,Z5 ) \
VMADDSUB(Z0,T2,UChi_10) \
VMADDSUB(Z1,T1,UChi_01) \
VMADDSUB(Z1,T2,UChi_11) \
VMADDSUB(Z2,T1,UChi_02) \
VMADDSUB(Z2,T2,UChi_12) \
/*55*/ \
VMADDSUB(Z3,Chi_02,UChi_00) \
VMADDSUB(Z3,Chi_12,UChi_10) \
VMADDSUB(Z4,Chi_02,UChi_01) \
VMADDSUB(Z4,Chi_12,UChi_11) \
VMADDSUB(Z5,Chi_02,UChi_02) \
VMADDSUB(Z5,Chi_12,UChi_12) \
/*61 insns*/ );
#define MULT_ADDSUB_2SPIN_LS(ptr,pf) \
LOAD64(%r8,ptr) \
LOAD64(%r9,pf) \
__asm__ ( \
VSHUF(Chi_00,T1) VSHUF(Chi_10,T2) \
VMULIDUP(0,%r8,T1,UChi_00) VMULIDUP(0,%r8,T2,UChi_10) \
VMULIDUP(3,%r8,T1,UChi_01) VMULIDUP(3,%r8,T2,UChi_11) \
VMULIDUP(6,%r8,T1,UChi_02) VMULIDUP(6,%r8,T2,UChi_12) \
VPREFETCHG(0,%r9) \
VPREFETCHG(1,%r9) \
VPREFETCHG(2,%r9) \
VPREFETCHG(3,%r9) \
/*8*/ \
VSHUF(Chi_01,T1) VSHUF(Chi_11,T2) \
VMADDSUBRDUP(0,%r8,Chi_00,UChi_00) VMADDSUBRDUP(0,%r8,Chi_10,UChi_10) \
VMADDSUBRDUP(3,%r8,Chi_00,UChi_01) VMADDSUBRDUP(3,%r8,Chi_10,UChi_11) \
VMADDSUBRDUP(6,%r8,Chi_00,UChi_02) VMADDSUBRDUP(6,%r8,Chi_10,UChi_12) \
VPREFETCHG(4,%r9) \
VPREFETCHG(5,%r9) \
VPREFETCHG(6,%r9) \
VPREFETCHG(7,%r9) \
/*16*/ \
VMADDSUBIDUP(1,%r8,T1,UChi_00) VMADDSUBIDUP(1,%r8,T2,UChi_10) \
VMADDSUBIDUP(4,%r8,T1,UChi_01) VMADDSUBIDUP(4,%r8,T2,UChi_11) \
VMADDSUBIDUP(7,%r8,T1,UChi_02) VMADDSUBIDUP(7,%r8,T2,UChi_12) \
VPREFETCHG(8,%r9) \
VPREFETCHG(9,%r9) \
VPREFETCHG(10,%r9) \
VPREFETCHG(11,%r9) \
/*22*/ \
VSHUF(Chi_02,T1) VSHUF(Chi_12,T2) \
VMADDSUBRDUP(1,%r8,Chi_01,UChi_00) VMADDSUBRDUP(1,%r8,Chi_11,UChi_10) \
VMADDSUBRDUP(4,%r8,Chi_01,UChi_01) VMADDSUBRDUP(4,%r8,Chi_11,UChi_11) \
VMADDSUBRDUP(7,%r8,Chi_01,UChi_02) VMADDSUBRDUP(7,%r8,Chi_11,UChi_12) \
VPREFETCH2(12,%r9) \
VPREFETCH2(13,%r9) \
VPREFETCH2(14,%r9) \
VPREFETCH2(15,%r9) \
/*30*/ \
VMADDSUBIDUP(2,%r8,T1,UChi_00) VMADDSUBIDUP(2,%r8,T2,UChi_10) \
VMADDSUBIDUP(5,%r8,T1,UChi_01) VMADDSUBIDUP(5,%r8,T2,UChi_11) \
VPREFETCH2(16,%r9) \
VPREFETCH2(17,%r9) \
VPREFETCH2(18,%r9) \
VPREFETCH2(19,%r9) \
VMADDSUBIDUP(8,%r8,T1,UChi_02) VMADDSUBIDUP(8,%r8,T2,UChi_12) \
/*36*/ \
VMADDSUBRDUP(2,%r8,Chi_02,UChi_00) VMADDSUBRDUP(2,%r8,Chi_12,UChi_10) \
VMADDSUBRDUP(5,%r8,Chi_02,UChi_01) VMADDSUBRDUP(5,%r8,Chi_12,UChi_11) \
VMADDSUBRDUP(8,%r8,Chi_02,UChi_02) VMADDSUBRDUP(8,%r8,Chi_12,UChi_12) \
VPREFETCH2(20,%r9) \
VPREFETCH2(21,%r9) \
VPREFETCH2(22,%r9) \
VPREFETCH2(23,%r9) \
VPREFETCHG(2,%r8) \
VPREFETCHG(3,%r8) \
VPREFETCH2(4,%r8) \
VPREFETCH2(5,%r8) \
/*42 insns*/ );
#define MULT_ADDSUB_2SPIN_LSNOPF(ptr,pf) \
LOAD64(%r8,ptr) \
LOAD64(%r9,pf) \
__asm__ ( \
VSHUF(Chi_00,T1) VSHUF(Chi_10,T2) \
VMULIDUP(0,%r8,T1,UChi_00) VMULIDUP(0,%r8,T2,UChi_10) \
VMULIDUP(3,%r8,T1,UChi_01) VMULIDUP(3,%r8,T2,UChi_11) \
VMULIDUP(6,%r8,T1,UChi_02) VMULIDUP(6,%r8,T2,UChi_12) \
/*8*/ \
VSHUF(Chi_01,T1) VSHUF(Chi_11,T2) \
VMADDSUBRDUP(0,%r8,Chi_00,UChi_00) VMADDSUBRDUP(0,%r8,Chi_10,UChi_10) \
VMADDSUBRDUP(3,%r8,Chi_00,UChi_01) VMADDSUBRDUP(3,%r8,Chi_10,UChi_11) \
VMADDSUBRDUP(6,%r8,Chi_00,UChi_02) VMADDSUBRDUP(6,%r8,Chi_10,UChi_12) \
/*16*/ \
VMADDSUBIDUP(1,%r8,T1,UChi_00) VMADDSUBIDUP(1,%r8,T2,UChi_10) \
VMADDSUBIDUP(4,%r8,T1,UChi_01) VMADDSUBIDUP(4,%r8,T2,UChi_11) \
VMADDSUBIDUP(7,%r8,T1,UChi_02) VMADDSUBIDUP(7,%r8,T2,UChi_12) \
/*22*/ \
VSHUF(Chi_02,T1) VSHUF(Chi_12,T2) \
VMADDSUBRDUP(1,%r8,Chi_01,UChi_00) VMADDSUBRDUP(1,%r8,Chi_11,UChi_10) \
VMADDSUBRDUP(4,%r8,Chi_01,UChi_01) VMADDSUBRDUP(4,%r8,Chi_11,UChi_11) \
VMADDSUBRDUP(7,%r8,Chi_01,UChi_02) VMADDSUBRDUP(7,%r8,Chi_11,UChi_12) \
/*30*/ \
VMADDSUBIDUP(2,%r8,T1,UChi_00) VMADDSUBIDUP(2,%r8,T2,UChi_10) \
VMADDSUBIDUP(5,%r8,T1,UChi_01) VMADDSUBIDUP(5,%r8,T2,UChi_11) \
VMADDSUBIDUP(8,%r8,T1,UChi_02) VMADDSUBIDUP(8,%r8,T2,UChi_12) \
/*36*/ \
VMADDSUBRDUP(2,%r8,Chi_02,UChi_00) VMADDSUBRDUP(2,%r8,Chi_12,UChi_10) \
VMADDSUBRDUP(5,%r8,Chi_02,UChi_01) VMADDSUBRDUP(5,%r8,Chi_12,UChi_11) \
VMADDSUBRDUP(8,%r8,Chi_02,UChi_02) VMADDSUBRDUP(8,%r8,Chi_12,UChi_12) \
/* VPREFETCHG(2,%r8)*/ \
/* VPREFETCHG(3,%r8)*/ \
/*42 insns*/ );
#define Z6 Chi_00
#define MULT_ADDSUB_2SPIN_NEW(ptr) \
LOAD64(%r8,ptr) \
__asm__ ( \
VSHUFMEM(0,%r8,Z0) \
VRDUP(Chi_00,T1) VIDUP(Chi_00,Chi_00) \
VRDUP(Chi_10,T2) VIDUP(Chi_10,Chi_10) \
VMUL(Z0,Chi_00,Z1) VMUL(Z0,Chi_10,Z2) \
VSHUFMEM(3,%r8,Z0) \
VMUL(Z0,Chi_00,Z3) VMUL(Z0,Chi_10,Z4) \
VSHUFMEM(6,%r8,Z0) \
VMUL(Z0,Chi_00,Z5) VMUL(Z0,Chi_10,Z6) \
VMULMEM(0,%r8,T1,UChi_00) VMULMEM(0,%r8,T2,UChi_10) \
VMULMEM(3,%r8,T1,UChi_01) VMULMEM(3,%r8,T2,UChi_11) \
VMULMEM(6,%r8,T1,UChi_02) VMULMEM(6,%r8,T2,UChi_12) \
/*11 cycles*/ \
VSHUFMEM(1,%r8,Z0) \
VRDUP(Chi_01,T1) VIDUP(Chi_01,Chi_01) \
VRDUP(Chi_11,T2) VIDUP(Chi_11,Chi_11) \
VMADD(Z0,Chi_01,Z1) VMADD(Z0,Chi_11,Z2) \
VSHUFMEM(4,%r8,Z0) \
VMADD(Z0,Chi_01,Z3) VMADD(Z0,Chi_11,Z4) \
VSHUFMEM(7,%r8,Z0) \
VMADD(Z0,Chi_01,Z5) VMADD(Z0,Chi_11,Z6) \
VMADDMEM(1,%r8,T1,UChi_00) VMADDMEM(1,%r8,T2,UChi_10) \
VMADDMEM(4,%r8,T1,UChi_01) VMADDMEM(4,%r8,T2,UChi_11) \
VMADDMEM(7,%r8,T1,UChi_02) VMADDMEM(7,%r8,T2,UChi_12) \
/*22 cycles*/ \
VSHUFMEM(2,%r8,Z0) \
VRDUP(Chi_02,T1) VIDUP(Chi_02,Chi_02) \
VRDUP(Chi_12,T2) VIDUP(Chi_12,Chi_12) \
VMADD(Z0,Chi_02,Z1) VMADD(Z0,Chi_12,Z2) \
VSHUFMEM(5,%r8,Z0) \
VMADD(Z0,Chi_02,Z3) VMADD(Z0,Chi_12,Z4) \
VSHUFMEM(8,%r8,Z0) \
VMADD(Z0,Chi_02,Z5) VMADD(Z0,Chi_12,Z6) \
/*33 cycles*/ \
VMADDSUBMEM(2,%r8,T1,Z1) VMADDSUBMEM(2,%r8,T2,Z2) \
VMADDSUBMEM(5,%r8,T1,Z3) VMADDSUBMEM(5,%r8,T2,Z4) \
VMADDSUBMEM(8,%r8,T1,Z5) VMADDSUBMEM(8,%r8,T2,Z6) \
/*stall*/ \
/*stall*/ \
/*stall*/ \
VADD(Z1,UChi_00,UChi_00) VADD(Z2,UChi_10,UChi_10) \
VADD(Z3,UChi_01,UChi_01) VADD(Z4,UChi_11,UChi_11) \
VADD(Z5,UChi_02,UChi_02) VADD(Z6,UChi_12,UChi_12) )
#endif