camm_strat1.h

基于Blas CLapck的.用过的人知道是干啥的

      punt(6,SS(a_,MM(6,RS4)),ax) \
      punt(7,SS(a_,MM(7,RS4)),ax) 
#undef lpcopy_3
#define lpcopy_3(a_) 
#undef dpcopy_3
#define dpcopy_3(a_) p8_copy_3(a_)
#undef plcopy_3
#define plcopy_3 32

#undef p1_4_cpsc_3
#define p1_4_cpsc_3(a_) \
      pls(a_,ax,0) \
      pmsr(6,0) \
      pus(0,a_,cx)
#undef p1_2_cpsc_3
#define p1_2_cpsc_3(a_) \
      pld(a_,ax,0) \
      pm(6,0) \
      pud(0,a_,cx)
#undef p1_cpsc_3
#define p1_cpsc_3(a_) \
      plq(a_,ax,0) \
      pm(6,0) \
      pu(0,a_,cx)
#undef p2_cpsc_3
#define p2_cpsc_3(a_) \
      plq(a_,ax,0) \
      plq(SS(a_,RS4),ax,1) \
      pm(6,0) \
      pm(6,1) \
      pu(0,a_,cx) \
      pu(1,SS(a_,RS4),cx) 
#undef p4_cpsc_3
#define p4_cpsc_3(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(6,2) \
      pu(0,a_,cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,0) \
      pm(6,3) \
      pu(1,SS(a_,RS4),cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      plq(SS(a_,MM(5,RS4)),ax,1) \
      pm(6,0) \
      pu(2,SS(a_,MM(2,RS4)),cx) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      plq(SS(a_,MM(6,RS4)),ax,2) \
      pm(6,1) \
      pu(3,SS(a_,MM(3,RS4)),cx) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),cx) 
#undef lpcpsc_3
#define lpcpsc_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)),cx) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      pm(6,1) 
#undef dpcpsc_3
#define dpcpsc_3(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(6,2) \
      pu(0,a_,cx) \
      pm(6,3) \
      pu(1,SS(a_,RS4),cx) \
      pu(2,SS(a_,MM(2,RS4)),cx) \
      pu(3,SS(a_,MM(3,RS4)),cx)
#undef plcpsc_3
#define plcpsc_3 16

#undef p1_4_cpsc_3c
#define p1_4_cpsc_3c(a_) 
#undef p1_2_cpsc_3c
#define p1_2_cpsc_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_,cx)
#undef p1_cpsc_3c
#define p1_cpsc_3c(a_) \
      plq(a_,ax,0) \
      pc(0,1) \
      pm(6,0) \
      ps(CSHUF,1,1) \
      pm(7,1) \
      pa(1,0) \
      pu(0,a_,cx)
#undef p2_cpsc_3c
#define p2_cpsc_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) \
      pu(0,a_,cx) \
      pc(1,3) \
      pm(6,1) \
      ps(CSHUF,3,3) \
      pm(7,3) \
      pa(3,1) \
      pu(1,SS(a_,RS4),cx) 
#undef p4_cpsc_3c
#define p4_cpsc_3c(a_) \
      pu(0,a_,cx) \
      pc(2,4) \
      pm(6,2) \
      ps(CSHUF,4,4) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,0) \
      pm(7,4) \
      pa(4,2) \
      pu(1,SS(a_,RS4),cx) \
      pc(3,4) \
      pm(6,3) \
      ps(CSHUF,4,4) \
      plq(SS(a_,MM(5,RS4)),ax,1) \
      pm(7,4) \
      pa(4,3) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pu(2,SS(a_,MM(2,RS4)),cx) \
      pc(0,4) \
      pm(6,0) \
      ps(CSHUF,4,4) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      plq(SS(a_,MM(6,RS4)),ax,2) \
      pm(7,4) \
      pa(4,0) \
      pu(3,SS(a_,MM(3,RS4)),cx) \
      pc(1,4) \
      pm(6,1) \
      ps(CSHUF,4,4) \
      plq(SS(a_,MM(7,RS4)),ax,3) \
      pm(7,4) \
      pa(4,1) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),cx) 
#undef lpcpsc_3c
#define lpcpsc_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) \
      pm(7,4) \
      pa(4,0) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pc(1,4) \
      pm(6,1) \
      ps(CSHUF,4,4) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(7,4) \
      pa(4,1)
#undef dpcpsc_3c
#define dpcpsc_3c(a_) \
      pu(0,a_,cx) \
      pc(2,4) \
      pm(6,2) \
      ps(CSHUF,4,4) \
      pu(1,SS(a_,RS4),cx) \
      pm(7,4) \
      pa(4,2) \
      pc(3,4) \
      pm(6,3) \
      ps(CSHUF,4,4) \
      pu(2,SS(a_,MM(2,RS4)),cx) \
      pm(7,4) \
      pa(4,3) \
      pu(3,SS(a_,MM(3,RS4)),cx)
#undef plcpsc_3c
#define plcpsc_3c 16

#undef p1_4_cpsc_4
#define p1_4_cpsc_4(a_) \
      pls(a_,cx,0) \
      pmsr(6,0) \
      pus(0,a_,ax)
#undef p1_2_cpsc_4
#define p1_2_cpsc_4(a_) \
      pld(a_,cx,0) \
      pm(6,0) \
      pud(0,a_,ax)
#undef p1_cpsc_4
#define p1_cpsc_4(a_) \
      pl(a_,cx,0) \
      pm(6,0) \
      puq(0,a_,ax)
#undef p2_cpsc_4
#define p2_cpsc_4(a_) \
      pl(a_,cx,0) \
      pl(SS(a_,RS4),cx,1) \
      pm(6,0) \
      pm(6,1) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_cpsc_4
#define p4_cpsc_4(a_) \
      pl(SS(a_,MM(3,RS4)),cx,3) \
      pm(6,2) \
      puq(0,a_,ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(4,RS4)),cx,0) \
      pm(6,3) \
      puq(1,SS(a_,RS4),ax) \
      pl(SS(a_,MM(5,RS4)),cx,1) \
      pm(6,0) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),cx) \
      pl(SS(a_,MM(6,RS4)),cx,2) \
      pm(6,1) \
      puq(3,SS(a_,MM(3,RS4)),ax) 
#undef lpcpsc_4
#define lpcpsc_4(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,0) \
      pl(SS(a_,MM(1,RS4)),cx,1) \
      pm(6,0) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      pm(6,1) 
#undef dpcpsc_4
#define dpcpsc_4(a_) \
      pl(SS(a_,MM(3,RS4)),cx,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 plcpsc_4
#define plcpsc_4 16

#undef p1_4_cpsc_5
#define p1_4_cpsc_5(a_) \
      pls(a_,cx,0) \
      pmsr(6,0) \
      pus(0,a_,ax)
#undef p1_2_cpsc_5
#define p1_2_cpsc_5(a_) \
      pld(a_,cx,0) \
      pm(6,0) \
      pud(0,a_,ax)
#undef p1_cpsc_5
#define p1_cpsc_5(a_) \
      pl(a_,cx,0) \
      pm(6,0) \
      puq(0,a_,ax)
#undef p2_cpsc_5
#define p2_cpsc_5(a_) \
      pl(a_,cx,0) \
      pl(SS(a_,RS4),cx,1) \
      pm(6,0) \
      pm(6,1) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_cpsc_5
#define p4_cpsc_5(a_) \
      pl(SS(a_,MM(0,RS4)),cx,0) \
      pl(SS(a_,MM(1,RS4)),cx,1) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      pl(SS(a_,MM(3,RS4)),cx,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_cpsc_5
#define p8_cpsc_5(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,0) \
      pl(SS(a_,MM(1,RS4)),cx,1) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      pl(SS(a_,MM(3,RS4)),cx,3) \
      pl(SS(a_,MM(4,RS4)),cx,4) \
      pl(SS(a_,MM(5,RS4)),cx,5) \
      pl(SS(a_,MM(6,RS4)),cx,7) \
      pm(6,0) \
      pm(6,1) \
      pm(6,2) \
      pm(6,3) \
      puq(0,a_,ax) \
      pl(SS(a_,MM(7,RS4)),cx,0) \
      pm(6,4) \
      pm(6,5) \
      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 lpcpsc_5
#define lpcpsc_5(a_) 
#undef dpcpsc_5
#define dpcpsc_5(a_) p8_cpsc_5(a_)
#undef plcpsc_5
#define plcpsc_5 32

#undef cpsc_cdp
#define cpsc_cdp(a_) pc(a_,5) pm(6,a_) ps(CSHUF,5,5) pm(7,5) pa(5,a_)
#undef p1_4_cpsc_5c
#define p1_4_cpsc_5c(a_) 
#undef p1_2_cpsc_5c
#define p1_2_cpsc_5c(a_) \
      pld(a_,cx,0) \
      cpsc_cdp(0) \
      pud(0,a_,ax)
#undef p1_cpsc_5c
#define p1_cpsc_5c(a_) \
      pl(a_,cx,0) \
      cpsc_cdp(0) \
      puq(0,a_,ax)
#undef p2_cpsc_5c
#define p2_cpsc_5c(a_) \
      pl(a_,cx,0) \
      pl(SS(a_,RS4),cx,1) \
      cpsc_cdp(0) \
      cpsc_cdp(1) \
      puq(0,a_,ax) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_cpsc_5c
#define p4_cpsc_5c(a_) \
      pl(SS(a_,MM(0,RS4)),cx,0) \
      pl(SS(a_,MM(1,RS4)),cx,1) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      pl(SS(a_,MM(3,RS4)),cx,3) \
      cpsc_cdp(0) \
      cpsc_cdp(1) \
      cpsc_cdp(2) \
      cpsc_cdp(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_cpsc_5c
#define p8_cpsc_5c(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,0) \
      pl(SS(a_,MM(1,RS4)),cx,1) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      pl(SS(a_,MM(3,RS4)),cx,3) \
      pl(SS(a_,MM(4,RS4)),cx,4) \
      cpsc_cdp(0) \
      cpsc_cdp(1) \
      puq(0,a_,ax) \
      pl(SS(a_,MM(5,RS4)),cx,0) \
      cpsc_cdp(2) \
      cpsc_cdp(3) \
      puq(1,SS(a_,RS4),ax) \
      pl(SS(a_,MM(6,RS4)),cx,1) \
      cpsc_cdp(4) \
      cpsc_cdp(0) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      pl(SS(a_,MM(7,RS4)),cx,2) \
      cpsc_cdp(1) \
      cpsc_cdp(2) \
      puq(3,SS(a_,MM(3,RS4)),ax) \
      puq(4,SS(a_,MM(4,RS4)),ax) \
      puq(0,SS(a_,MM(5,RS4)),ax) \
      puq(1,SS(a_,MM(6,RS4)),ax) \
      puq(2,SS(a_,MM(7,RS4)),ax) 
#undef lpcpsc_5c
#define lpcpsc_5c(a_) 
#undef dpcpsc_5c
#define dpcpsc_5c(a_) p8_cpsc_5c(a_)
#undef plcpsc_5c
#define plcpsc_5c 32

#undef p1_4_cpsc_1
#define p1_4_cpsc_1(a_) \
      pls(a_,ax,2) \
      pmsr(3,2) \
      pus(2,a_,cx)
#undef p1_2_cpsc_1
#define p1_2_cpsc_1(a_) \
      pld(a_,ax,2) \
      pm(3,2) \
      pud(2,a_,cx)
#undef p1_cpsc_1
#define p1_cpsc_1(a_) \
      plq(a_,ax,2) \
      pm(3,2) \
      pu(2,a_,cx)
#undef p2_cpsc_1
#define p2_cpsc_1(a_) \
      plq(SS(a_,RS4),ax,4) \
      pm(3,2) \
      pu(2,a_,cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      pm(3,4) \
      pu(4,SS(a_,RS4),cx)
#undef lpcpsc_1
#define lpcpsc_1(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,2) 
#undef dpcpsc_1
#define dpcpsc_1(a_) \
      plq(SS(a_,RS4),ax,4) \
      pm(3,2) \
      pu(2,a_,cx) \
      pm(3,4) \
      pu(4,SS(a_,RS4),cx)
#undef plcpsc_1
#define plcpsc_1 8

#undef p1_4_cpsc_2
#define p1_4_cpsc_2(a_) \
      pls(a_,ax,2) \
      pmsr(3,2) \
      pus(2,a_,cx)
#undef p1_2_cpsc_2
#define p1_2_cpsc_2(a_) \
      pld(a_,ax,2) \
      pm(3,2) \
      pud(2,a_,cx)
#undef p1_cpsc_2
#define p1_cpsc_2(a_) \
      plq(a_,ax,2) \
      pm(3,2) \
      pu(2,a_,cx)
#undef p2_cpsc_2
#define p2_cpsc_2(a_) \
      plq(a_,ax,2) \
      plq(SS(a_,RS4),ax,4) \
      pm(3,2) \
      pm(3,4) \
      pu(2,a_,cx) \
      pu(4,SS(a_,RS4),cx)
#undef p4_cpsc_2
#define p4_cpsc_2(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,7) \
      pm(3,6) \
      pu(4,a_,cx) \
      plq(SS(a_,MM(3,RS4)),ax,2) \
      pm(3,7) \
      pu(6,SS(a_,RS4),cx) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),cx) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      plq(SS(a_,MM(4,RS4)),ax,4) \
      pm(3,2) \
      pu(7,SS(a_,MM(2,RS4)),cx) \
      plq(SS(a_,MM(5,RS4)),ax,6) \
      pm(3,4) \
      pu(2,SS(a_,MM(3,RS4)),cx) 
#undef lpcpsc_2
#define lpcpsc_2(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,4) \
      plq(SS(a_,MM(1,RS4)),ax,6) \
      pm(3,4) 
#undef dpcpsc_2
#define dpcpsc_2(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,7) \
      pm(3,6) \
      pu(4,a_,cx) \
      plq(SS(a_,MM(3,RS4)),ax,2) \
      pm(3,7) \
      pu(6,SS(a_,RS4),cx) \
      pm(3,2) \
      pu(7,SS(a_,MM(2,RS4)),cx) \
      pu(2,SS(a_,MM(3,RS4)),cx) 
#undef plcpsc_2
#define plcpsc_2 RS4


#undef p1_4_iamax_1
#define p1_4_iamax_1(a_) \
      px(4) \
      pls(a_,ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      paxs(4,3) \
      pan(5,6) \
      pann(0,5) \
      pasr(5,6) \
      pasr(1,0) \
      ps(57,0,0)
#undef p1_2_iamax_1
#define p1_2_iamax_1(a_) \
      px(4) \
      pld(a_,ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      pax(4,3) \
      pan(5,6) \
      pann(0,5) \
      pa(5,6) \
      pasr(1,0) \
      ps(57,0,0)\
      pasr(1,0) \
      ps(57,0,0)
#undef p1_iamax_1
#define p1_iamax_1(a_) \
      plq(a_,ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      pax(4,3) \
      pan(5,6) \
      pann(0,5) \
      pa(5,6) \
      pa(1,0) 
#define p2_iamax_1(a_) \
      plq(SS(a_,RS4),ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      pax(4,3) \
      pan(5,6) \
      pann(0,5) \
      pa(5,6) \
      pa(1,0) \
      f(nta,SS(a_,MM(SS(2,CL),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      pax(4,3) \
      pan(5,6) \
      pann(0,5) \
      pa(5,6) \
      pa(1,0) 
#undef lpiamax_1
#define lpiamax_1(a_) \
      f(nta,SS(a_,MM(CL,RS4)),ax) \
      plq(a_,ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      pax(4,3) \
      pan(5,6) \
      pann(0,5) \
      pa(5,6) \
      pa(1,0) 
#undef dpiamax_1
#define dpiamax_1(a_) \
      plq(SS(a_,RS4),ax,4) \
      pan(2,4) \
      pc(3,5) \
      pcm(6,4,5) \
      pax(4,3) \
      pan(5,6) \
      pann(0,5) \
      pa(5,6) \
      pa(1,0) 
#undef pliamax_1
#define pliamax_1 8

#undef p1_4_iamax_1d
#define p1_4_iamax_1d(