camm_strat1.h

基于Blas CLapck的.用过的人知道是干啥的

#include "camm_util.h"

#undef p1_4_swap_1
#define p1_4_swap_1(a_) \
      pls(a_,ax,1) \
      pls(a_,cx,0) \
      pus(0,a_,ax) \
      pus(1,a_,cx) 
#undef p1_2_swap_1
#define p1_2_swap_1(a_) \
      px(1) \
      pld(a_,ax,1) \
      px(0) \
      pld(a_,cx,0) \
      pud(0,a_,ax) \
      pud(1,a_,cx) 
#undef p1_swap_1
#define p1_swap_1(a_) \
      plq(a_,ax,1) \
      pl(a_,cx,0) \
      puq(0,a_,ax) \
      pu(1,a_,cx) 
#undef p2_swap_1
#define p2_swap_1(a_) \
      plq(SS(a_,RS4),ax,3) \
      pl(SS(a_,RS4),cx,2) \
      puq(0,a_,ax) \
      pu(1,a_,cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(2,RS4)),cx,0) \
      puq(2,SS(a_,RS4),ax) \
      pu(3,SS(a_,RS4),cx) 
#undef lpswap_1
#define lpswap_1(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,1) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,0) 
#undef dpswap_1
#define dpswap_1(a_) \
      plq(SS(a_,RS4),ax,3) \
      pl(SS(a_,RS4),cx,2) \
      puq(0,a_,ax) \
      pu(1,a_,cx) \
      puq(2,SS(a_,RS4),ax) \
      pu(3,SS(a_,RS4),cx) 
#undef plswap_1
#define plswap_1 8


#undef p1_4_scal_3
#define p1_4_scal_3(a_) \
      pls(a_,ax,0) \
      pmsr(6,0) \
      pus(0,a_,ax)
#undef p1_2_scal_3
#define p1_2_scal_3(a_) \
      pld(a_,ax,0) \
      pm(6,0) \
      pud(0,a_,ax)
#undef p1_scal_3
#define p1_scal_3(a_) \
      plq(a_,ax,0) \
      pm(6,0) \
      puq(0,a_,ax)
#undef p2_scal_3
#define p2_scal_3(a_) \
      plq(a_,ax,0) \
      plq(SS(a_,RS4),ax,1) \
      pm(6,0) \
      pm(6,1) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_scal_3
#define p4_scal_3(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(6,2) \
      puq(0,a_,ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,0) \
      pm(6,3) \
      puq(1,SS(a_,RS4),ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(5,RS4)),ax,1) \
      pm(6,0) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      plq(SS(a_,MM(6,RS4)),ax,2) \
      pm(6,1) \
      puq(3,SS(a_,MM(3,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) 
#undef lpscal_3
#define lpscal_3(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      pm(6,0) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      pm(6,1) 
#undef dpscal_3
#define dpscal_3(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(6,2) \
      puq(0,a_,ax) \
      pm(6,3) \
      puq(1,SS(a_,RS4),ax) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      puq(3,SS(a_,MM(3,RS4)),ax)
#undef plscal_3
#define plscal_3 16

#undef p1_4_scal_3c
#define p1_4_scal_3c(a_) 
#undef p1_2_scal_3c
#define p1_2_scal_3c(a_) \
      pld(a_,ax,0) \
      pc(0,1) \
      pm(6,0) \
      ps(CSHUF,1,1) \
      pm(7,1) \
      pa(1,0) \
      pud(0,a_,ax)
#undef p1_scal_3c
#define p1_scal_3c(a_) \
      plq(a_,ax,0) \
      pc(0,1) \
      pm(6,0) \
      ps(CSHUF,1,1) \
      pm(7,1) \
      pa(1,0) \
      puq(0,a_,ax)
#undef p2_scal_3c
#define p2_scal_3c(a_) \
      plq(a_,ax,0) \
      plq(SS(a_,RS4),ax,1) \
      pc(0,2) \
      pm(6,0) \
      ps(CSHUF,2,2) \
      pm(7,2) \
      pa(2,0) \
      puq(0,a_,ax) \
      pc(1,3) \
      pm(6,1) \
      ps(CSHUF,3,3) \
      pm(7,3) \
      pa(3,1) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_scal_3c
#define p4_scal_3c(a_) \
      pm(7,5) \
      pa(5,1) \
      puq(0,a_,ax) \
      ps(CSHUF,4,4) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,0) \
      pc(3,5) \
      pm(6,3) \
      pm(7,4) \
      pa(4,2) \
      puq(1,SS(a_,RS4),ax) \
      ps(CSHUF,5,5) \
      plq(SS(a_,MM(5,RS4)),ax,1) \
      pc(0,4) \
      pm(6,0) \
      pm(7,5) \
      pa(5,3) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      ps(CSHUF,4,4) \
      plq(SS(a_,MM(6,RS4)),ax,2) \
      pc(1,5) \
      pm(6,1) \
      pm(7,4) \
      pa(4,0) \
      puq(3,SS(a_,MM(3,RS4)),ax) \
      ps(CSHUF,5,5) \
      plq(SS(a_,MM(7,RS4)),ax,3) \
      pc(2,4) \
      pm(6,2) 
#undef lpscal_3c
#define lpscal_3c(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      pc(0,4) \
      pm(6,0) \
      ps(CSHUF,4,4) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      pc(1,5) \
      pm(6,1) \
      pm(7,4) \
      pa(4,0) \
      ps(CSHUF,5,5) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pc(2,4) \
      pm(6,2) 
#undef dpscal_3c
#define dpscal_3c(a_) \
      pm(7,5) \
      pa(5,1) \
      ps(CSHUF,4,4) \
      puq(0,a_,ax) \
      pm(7,4) \
      pa(4,2) \
      pc(3,5) \
      pm(6,3) \
      puq(1,SS(a_,RS4),ax) \
      ps(CSHUF,5,5) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      pm(7,5) \
      pa(5,3) \
      puq(3,SS(a_,MM(3,RS4)),ax)
#undef plscal_3c
#define plscal_3c 16

#undef p1_4_scal_4
#define p1_4_scal_4(a_) \
      pls(SS(a_,MM(0,RS4)),ax,0) \
      pmsr(6,0) \
      pus(0,a_,ax) 
#undef p1_2_scal_4
#define p1_2_scal_4(a_) \
      pld(SS(a_,MM(0,RS4)),ax,0) \
      pm(6,0) \
      pud(0,a_,ax) 
#undef p1_scal_4
#define p1_scal_4(a_) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      pm(6,0) \
      puq(0,a_,ax) 
#undef p2_scal_4
#define p2_scal_4(a_) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      pm(6,0) \
      pm(6,1) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_scal_4
#define p4_scal_4(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(6,0) \
      pm(6,1) \
      pm(6,2) \
      pm(6,3) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      puq(3,SS(a_,MM(3,RS4)),ax) 
#undef p8_scal_4
#define p8_scal_4(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      plq(SS(a_,MM(4,RS4)),ax,4) \
      plq(SS(a_,MM(5,RS4)),ax,5) \
      plq(SS(a_,MM(6,RS4)),ax,7) \
      pm(6,0) \
      pm(6,1) \
      pm(6,2) \
      puq(0,a_,ax) \
      pm(6,3) \
      pm(6,4) \
      pm(6,5) \
      plq(SS(a_,MM(7,RS4)),ax,0) \
      pm(6,7) \
      pm(6,0) \
      puq(1,SS(a_,RS4),ax) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      puq(3,SS(a_,MM(3,RS4)),ax) \
      puq(4,SS(a_,MM(4,RS4)),ax) \
      puq(5,SS(a_,MM(5,RS4)),ax) \
      puq(7,SS(a_,MM(6,RS4)),ax) \
      puq(0,SS(a_,MM(7,RS4)),ax) 
#undef lpscal_4
#define lpscal_4(a_) 
#undef dpscal_4
#define dpscal_4(a_) p4_scal_4(a_)
#undef plscal_4
#define plscal_4 16

#undef p1_4_scal_4c
#define p1_4_scal_4c(a_) 
#undef p1_2_scal_4c
#define p1_2_scal_4c(a_) \
      pld(a_,ax,0) \
      pc(0,1) \
      pm(6,0) \
      ps(CSHUF,1,1) \
      pm(7,1) \
      pa(1,0) \
      pud(0,a_,ax)
#undef p1_scal_4c
#define p1_scal_4c(a_) \
      plq(a_,ax,0) \
      pc(0,1) \
      pm(6,0) \
      ps(CSHUF,1,1) \
      pm(7,1) \
      pa(1,0) \
      puq(0,a_,ax)
#undef p2_scal_4c
#define p2_scal_4c(a_) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      pc(0,4) \
      pc(1,5) \
      pm(6,0) \
      pm(6,1) \
      ps(CSHUF,4,4) \
      ps(CSHUF,5,5) \
      pm(7,4) \
      pa(4,0) \
      pm(7,5) \
      pa(5,1) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_scal_4c
#define p4_scal_4c(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pc(0,4) \
      pc(1,5) \
      pm(6,0) \
      pm(6,1) \
      ps(CSHUF,4,4) \
      ps(CSHUF,5,5) \
      pm(7,4) \
      pa(4,0) \
      pc(2,4) \
      pm(7,5) \
      pa(5,1) \
      pc(3,5) \
      pm(6,2) \
      pm(6,3) \
      ps(CSHUF,4,4) \
      ps(CSHUF,5,5) \
      pm(7,4) \
      pa(4,2) \
      pm(7,5) \
      pa(5,3) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      puq(3,SS(a_,MM(3,RS4)),ax) 
#undef lpscal_4c
#define lpscal_4c(a_) 
#undef dpscal_4c
#define dpscal_4c(a_) p4_scal_4c(a_)
#undef plscal_4c
#define plscal_4c 16

#undef p1_4_scal_1
#define p1_4_scal_1(a_) \
      pls(a_,ax,1) \
      pmsr(0,1) \
      pus(1,a_,ax) 
#undef p1_2_scal_1
#define p1_2_scal_1(a_) \
      px(1) \
      pld(a_,ax,1) \
      pm(0,1) \
      pud(1,a_,ax) 
#undef p1_scal_1
#define p1_scal_1(a_) \
      plq(a_,ax,1) \
      pm(0,1) \
      puq(1,a_,ax) 
#undef p2_scal_1
#define p2_scal_1(a_) \
      plq(a_,ax,1) \
      plq(SS(a_,RS4),ax,2) \
      pm(0,1) \
      pm(0,2) \
      puq(1,a_,ax) \
      puq(2,SS(a_,RS4),ax)
#undef p4_scal_1
#define p4_scal_1(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      pm(0,3) \
      puq(7,a_,ax) \
      plq(SS(a_,MM(3,RS4)),ax,2) \
      pm(0,1) \
      puq(3,SS(a_,MM(1,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,7) \
      pm(0,2) \
      puq(1,SS(a_,MM(2,RS4)),ax) \
      plq(SS(a_,MM(5,RS4)),ax,3) \
      pm(0,7) \
      puq(2,SS(a_,MM(3,RS4)),ax) 
#undef lpscal_1
#define lpscal_1(a_) \
      plq(a_,ax,7) \
      plq(SS(a_,MM(1,RS4)),ax,3) \
      pm(0,7) 
#undef dpscal_1
#define dpscal_1(a_) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      pm(0,3) \
      puq(7,a_,ax) \
      plq(SS(a_,MM(3,RS4)),ax,2) \
      pm(0,1) \
      puq(3,SS(a_,MM(1,RS4)),ax) \
      pm(0,2) \
      puq(1,SS(a_,MM(2,RS4)),ax) \
      puq(2,SS(a_,MM(3,RS4)),ax) 
#undef plscal_1
#define plscal_1 RS4


#undef p1_4_set_1
#define p1_4_set_1(a_) \
      pls(a_,ax,1) \
      pcs(0,1) \
      pus(1,a_,ax) 
#undef p1_2_set_1
#define p1_2_set_1(a_) \
      px(1) \
      pld(a_,ax,1) \
      pc(0,1) \
      pud(1,a_,ax) 
#undef p1_set_1
#define p1_set_1(a_) \
      plq(a_,ax,1) \
      pc(0,1) \
      puq(1,a_,ax) 
#undef p2_set_1
#define p2_set_1(a_) \
      plq(a_,ax,1) \
      plq(SS(a_,RS4),ax,2) \
      pc(0,1) \
      pc(0,2) \
      puq(1,a_,ax) \
      puq(2,SS(a_,RS4),ax)
#undef p4_set_1
#define p4_set_1(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      pc(0,3) \
      puq(7,a_,ax) \
      plq(SS(a_,MM(3,RS4)),ax,2) \
      pc(0,1) \
      puq(3,SS(a_,MM(1,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,7) \
      pc(0,2) \
      puq(1,SS(a_,MM(2,RS4)),ax) \
      plq(SS(a_,MM(5,RS4)),ax,3) \
      pc(0,7) \
      puq(2,SS(a_,MM(3,RS4)),ax) 
#undef lpset_1
#define lpset_1(a_) \
      plq(a_,ax,7) \
      plq(SS(a_,MM(1,RS4)),ax,3) \
      pc(0,7) 
#undef dpset_1
#define dpset_1(a_) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      pc(0,3) \
      puq(7,a_,ax) \
      plq(SS(a_,MM(3,RS4)),ax,2) \
      pc(0,1) \
      puq(3,SS(a_,MM(1,RS4)),ax) \
      pc(0,2) \
      puq(1,SS(a_,MM(2,RS4)),ax) \
      puq(2,SS(a_,MM(3,RS4)),ax) 
#undef plset_1
#define plset_1 RS4


#undef p1_4_set_2
#define p1_4_set_2(a_) \
      pus(0,a_,ax) 
#undef p1_2_set_2
#define p1_2_set_2(a_) \
      pud(0,a_,ax) 
#undef p1_set_2
#define p1_set_2(a_) \
      puq(0,a_,ax) 
#undef p2_set_2
#define p2_set_2(a_) \
      puq(0,a_,ax) \
      puq(0,SS(a_,RS4),ax)
#undef p4_set_2
#define p4_set_2(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      puq(0,a_,ax) \
      puq(0,SS(a_,MM(1,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      puq(0,SS(a_,MM(2,RS4)),ax) \
      puq(0,SS(a_,MM(3,RS4)),ax) 
#undef lpset_2
#define lpset_2(a_) 
#undef dpset_2
#define dpset_2(a_) \
      puq(0,a_,ax) \
      puq(0,SS(a_,MM(1,RS4)),ax) \
      puq(0,SS(a_,MM(2,RS4)),ax) \
      puq(0,SS(a_,MM(3,RS4)),ax) 
#undef plset_2
#define plset_2 RS4


#undef p1_4_set_3
#define p1_4_set_3(a_) \
      pus(0,a_,ax) 
#undef p1_2_set_3
#define p1_2_set_3(a_) \
      pud(0,a_,ax) 
#undef p1_set_3
#define p1_set_3(a_) \
      puq(0,SS(a_,MM(0,RS4)),ax) 
#undef p2_set_3
#define p2_set_3(a_) \
      puq(0,SS(a_,MM(0,RS4)),ax) \
      puq(0,SS(a_,MM(1,RS4)),ax) 
#undef p4_set_3
#define p4_set_3(a_) \
      puq(0,SS(a_,MM(0,RS4)),ax) \
      puq(0,SS(a_,MM(1,RS4)),ax) \
      puq(0,SS(a_,MM(2,RS4)),ax) \
      puq(0,SS(a_,MM(3,RS4)),ax) 
#undef p8_set_3
#define p8_set_3(a_) \
      puq(0,SS(a_,MM(0,RS4)),ax) \
      puq(0,SS(a_,MM(1,RS4)),ax) \
      puq(0,SS(a_,MM(2,RS4)),ax) \
      puq(0,SS(a_,MM(3,RS4)),ax) \
      puq(0,SS(a_,MM(4,RS4)),ax) \
      puq(0,SS(a_,MM(5,RS4)),ax) \
      puq(0,SS(a_,MM(6,RS4)),ax) \
      puq(0,SS(a_,MM(7,RS4)),ax) 
#undef lpset_3
#define lpset_3(a_) 
#undef dpset_3
#define dpset_3(a_) p8_set_3(a_)
#undef plset_3
#define plset_3 32


#undef p1_4_0x1_nrm2_1
#define p1_4_0x1_nrm2_1(a_) \
      pls(a_,ax,1) \
      pmsr(1,1) \
      pasr(1,0) 
#undef p1_2_0x1_nrm2_1
#define p1_2_0x1_nrm2_1(a_) \
      px(1) \
      pld(a_,ax,1) \
      pm(1,1) \
      pa(1,0) 
#undef p1_0x1_nrm2_1
#define p1_0x1_nrm2_1(a_) \
      plq(a_,ax,1) \
      pm(1,1) \
      pa(1,0) 
#undef p2_0x1_nrm2_1
#define p2_0x1_nrm2_1(a_) \
      plq(a_,ax,1) \
      plq(SS(a_,RS4),ax,2) \