camm_strat1.h

基于Blas CLapck的.用过的人知道是干啥的

#undef p2_axpby_1
#define p2_axpby_1(a_) \
      plq(SS(a_,RS4),ax,3) \
      pl(SS(a_,RS4),cx,4) \
      pm(5,1) \
      pm(6,2) \
      pa(2,1) \
      puq(1,a_,ax) \
      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,2) \
      pm(5,3) \
      pm(6,4) \
      pa(4,3) \
      puq(3,SS(a_,RS4),ax)
#undef lpaxpby_1
#define lpaxpby_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,2) 
#undef dpaxpby_1
#define dpaxpby_1(a_) \
      plq(SS(a_,RS4),ax,3) \
      pl(SS(a_,RS4),cx,4) \
      pm(5,1) \
      pm(6,2) \
      pa(2,1) \
      puq(1,a_,ax) \
      pm(5,3) \
      pm(6,4) \
      pa(4,3) \
      puq(3,SS(a_,RS4),ax)
#undef plaxpby_1
#define plaxpby_1 8

#undef p1_4_axpy_0
#define p1_4_axpy_0(a_) \
      pls(a_,cx,2) \
      pls(a_,ax,1) \
      pmsr(6,2) \
      pasr(2,1) \
      pus(1,a_,ax)
#undef p1_2_axpy_0
#define p1_2_axpy_0(a_) \
      pld(a_,cx,2) \
      pld(a_,ax,1) \
      pm(6,2) \
      pa(2,1) \
      pud(1,a_,ax)
#undef p1_axpy_0
#define p1_axpy_0(a_) \
      pl(a_,cx,2) \
      plq(a_,ax,1) \
      pm(6,2) \
      pa(2,1) \
      puq(1,a_,ax)
#undef p2_axpy_0
#define p2_axpy_0(a_) \
      pl(SS(a_,RS4),cx,4) \
      pm(6,2) \
      pa(2,1) \
      plq(SS(a_,RS4),ax,3) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      puq(1,a_,ax) \
      pm(6,4) \
      pa(4,3) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      puq(3,SS(a_,RS4),ax) 
#undef lpaxpy_0
#define lpaxpy_0(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,2) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,1) 
#undef dpaxpy_0
#define dpaxpy_0(a_) \
      pl(SS(a_,RS4),cx,4) \
      pm(6,2) \
      pa(2,1) \
      plq(SS(a_,RS4),ax,3) \
      puq(1,a_,ax) \
      pm(6,4) \
      pa(4,3) \
      puq(3,SS(a_,RS4),ax)
#undef plaxpy_0
#define plaxpy_0 8

#undef p1_4_axpy_1
#define p1_4_axpy_1(a_) \
      pls(a_,cx,2) \
      pls(a_,ax,1) \
      pmsr(6,2) \
      pasr(2,1) \
      pus(1,a_,ax)
#undef p1_2_axpy_1
#define p1_2_axpy_1(a_) \
      pld(a_,cx,2) \
      pld(a_,ax,1) \
      pm(6,2) \
      pa(2,1) \
      pud(1,a_,ax)
#undef p1_axpy_1
#define p1_axpy_1(a_) \
      pl(a_,cx,2) \
      pm(6,2) \
      pam(a_,ax,2) \
      puq(2,a_,ax)
#undef p2_axpy_1
#define p2_axpy_1(a_) \
      pl(a_,cx,2) \
      pm(6,2) \
      pl(SS(a_,RS4),cx,4) \
      pam(a_,ax,2) \
      pm(6,4) \
      puq(2,a_,ax) \
      pam(SS(a_,RS4),ax,4) \
      puq(4,SS(a_,RS4),ax) 
#undef p4_axpy_1
#define p4_axpy_1(a_) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(3,RS4)),cx,3) \
      pm(6,2) \
      pam(SS(a_,MM(2,RS4)),ax,2) \
      puq(0,a_,ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      pl(SS(a_,MM(4,RS4)),cx,0) \
      pm(6,3) \
      pam(SS(a_,MM(3,RS4)),ax,3) \
      puq(1,SS(a_,RS4),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),cx) \
      pl(SS(a_,MM(5,RS4)),cx,1) \
      pm(6,0) \
      pam(SS(a_,MM(4,RS4)),ax,0) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) \
      pl(SS(a_,MM(6,RS4)),cx,2) \
      pm(6,1) \
      pam(SS(a_,MM(5,RS4)),ax,1) \
      puq(3,SS(a_,MM(3,RS4)),ax)
#undef lpaxpy_1
#define lpaxpy_1(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(a_,cx,0) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      pl(SS(a_,RS4),cx,1) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pm(6,0) \
      pam(a_,ax,0) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      pm(6,1) \
      pam(SS(a_,RS4),ax,1)
#undef dpaxpy_1
#define dpaxpy_1(a_) \
      pl(SS(a_,MM(3,RS4)),cx,3) \
      pm(6,2) \
      pam(SS(a_,MM(2,RS4)),ax,2) \
      puq(0,a_,ax) \
      pm(6,3) \
      pam(SS(a_,MM(3,RS4)),ax,3) \
      puq(1,SS(a_,RS4),ax) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      puq(3,SS(a_,MM(3,RS4)),ax)
#undef plaxpy_1
#define plaxpy_1 16

#undef p1_4_axpy_2
#define p1_4_axpy_2(a_) \
      pls(a_,cx,5) \
      pls(a_,ax,0) \
      pmsr(6,5) \
      pasr(5,0) \
      pus(0,a_,ax)
#undef p1_2_axpy_2
#define p1_2_axpy_2(a_) \
      pld(a_,cx,5) \
      pld(a_,ax,0) \
      pm(6,5) \
      pa(5,0) \
      pud(0,a_,ax)
#undef p1_axpy_2
#define p1_axpy_2(a_) \
      pl(a_,cx,5) \
      plq(a_,ax,0) \
      pm(6,5) \
      pa(5,0) \
      puq(0,a_,ax)
#undef p2_axpy_2
#define p2_axpy_2(a_) \
      pl(a_,cx,5) \
      plq(a_,ax,0) \
      pl(SS(a_,RS4),cx,4) \
      pm(6,5) \
      pa(5,0) \
      plq(SS(a_,RS4),ax,1) \
      puq(0,a_,ax) \
      pm(6,4) \
      pa(4,1) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_axpy_2
#define p4_axpy_2(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pl(SS(a_,MM(3,RS4)),cx,5) \
      pm(6,4) \
      pa(4,2) \
      puq(0,a_,ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(4,RS4)),cx,4) \
      plq(SS(a_,MM(4,RS4)),ax,0) \
      pm(6,5) \
      pa(5,3) \
      puq(1,SS(a_,RS4),ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(5,RS4)),ax,1) \
      pl(SS(a_,MM(5,RS4)),cx,5) \
      pm(6,4) \
      pa(4,0) \
      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,4) \
      plq(SS(a_,MM(6,RS4)),ax,2) \
      pm(6,5) \
      pa(5,1) \
      puq(3,SS(a_,MM(3,RS4)),ax) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) 
#undef lpaxpy_2
#define lpaxpy_2(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,4) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      pl(SS(a_,MM(1,RS4)),cx,5) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      pm(6,4) \
      pa(4,0) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      pl(SS(a_,MM(2,RS4)),cx,4) \
      pm(6,5) \
      pa(5,1)
#undef dpaxpy_2
#define dpaxpy_2(a_) \
      pl(SS(a_,MM(3,RS4)),cx,5) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      pm(6,4) \
      pa(4,2) \
      puq(0,a_,ax) \
      pm(6,5) \
      pa(5,3) \
      puq(1,SS(a_,RS4),ax) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      puq(3,SS(a_,MM(3,RS4)),ax)
#undef plaxpy_2
#define plaxpy_2 16

#undef p1_4_axpy_2c
#define p1_4_axpy_2c(a_) 
#undef p1_2_axpy_2c
#define p1_2_axpy_2c(a_) \
      pld(a_,cx,4) \
      pld(a_,ax,0) \
      pc(4,2) \
      pm(6,4) \
      pa(4,0) \
      ps(CSHUF,2,2) \
      pm(7,2) \
      pa(2,0) \
      pud(0,a_,ax)
#undef p1_axpy_2c
#define p1_axpy_2c(a_) \
      pl(a_,cx,4) \
      plq(a_,ax,0) \
      pc(4,2) \
      pm(6,4) \
      pa(4,0) \
      ps(CSHUF,2,2) \
      pm(7,2) \
      pa(2,0) \
      puq(0,a_,ax)
#undef p2_axpy_2c
#define p2_axpy_2c(a_) \
      pl(a_,cx,4) \
      plq(a_,ax,0) \
      pl(SS(a_,RS4),cx,5) \
      pc(4,2) \
      pm(6,4) \
      pa(4,0) \
      ps(CSHUF,2,2) \
      pm(7,2) \
      pa(2,0) \
      plq(SS(a_,RS4),ax,1) \
      puq(0,a_,ax) \
      pc(5,3) \
      pm(6,5) \
      pa(5,1) \
      ps(CSHUF,3,3) \
      pm(7,3) \
      pa(3,1) \
      puq(1,SS(a_,RS4),ax) 
#undef p4_axpy_2c
#define p4_axpy_2c(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      puq(0,a_,ax) \
      pc(4,0) \
      pm(6,4) \
      pa(4,2) \
      ps(CSHUF,0,0) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(4,RS4)),cx,4) \
      pm(7,0) \
      pa(0,2) \
      plq(SS(a_,MM(4,RS4)),ax,0) \
      puq(1,SS(a_,RS4),ax) \
      pc(5,1) \
      pm(6,5) \
      pa(5,3) \
      ps(CSHUF,1,1) \
      pl(SS(a_,MM(5,RS4)),cx,5) \
      pm(7,1) \
      pa(1,3) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(5,RS4)),ax,1) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      pc(4,2) \
      pm(6,4) \
      pa(4,0) \
      ps(CSHUF,2,2) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),cx) \
      pl(SS(a_,MM(6,RS4)),cx,4) \
      pm(7,2) \
      pa(2,0) \
      plq(SS(a_,MM(6,RS4)),ax,2) \
      puq(3,SS(a_,MM(3,RS4)),ax) \
      pc(5,3) \
      pm(6,5) \
      pa(5,1) \
      ps(CSHUF,3,3) \
      pl(SS(a_,MM(7,RS4)),cx,5) \
      pm(7,3) \
      pa(3,1) \
      f(nta,SS(a_,MM((SS(4,CL)),RS4)),ax) 
#undef lpaxpy_2c
#define lpaxpy_2c(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,4) \
      plq(SS(a_,MM(0,RS4)),ax,0) \
      pl(SS(a_,MM(1,RS4)),cx,5) \
      plq(SS(a_,MM(1,RS4)),ax,1) \
      pc(4,2) \
      pm(6,4) \
      pa(4,0) \
      ps(CSHUF,2,2) \
      pl(SS(a_,MM(2,RS4)),cx,4) \
      pm(7,2) \
      pa(2,0) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,2) \
      pc(5,3) \
      pm(6,5) \
      pa(5,1) \
      ps(CSHUF,3,3) \
      pl(SS(a_,MM(3,RS4)),cx,5) \
      pm(7,3) \
      pa(3,1)
#undef dpaxpy_2c
#define dpaxpy_2c(a_) \
      plq(SS(a_,MM(3,RS4)),ax,3) \
      puq(0,a_,ax) \
      pc(4,0) \
      pm(6,4) \
      pa(4,2) \
      ps(CSHUF,0,0) \
      puq(1,SS(a_,RS4),ax) \
      pm(7,0) \
      pa(0,2) \
      pc(5,1) \
      pm(6,5) \
      pa(5,3) \
      ps(CSHUF,1,1) \
      puq(2,SS(a_,MM(2,RS4)),ax) \
      pm(7,1) \
      pa(1,3) \
      puq(3,SS(a_,MM(3,RS4)),ax)
#undef plaxpy_2c
#define plaxpy_2c 16

#undef p1_4_axpy_1c
#define p1_4_axpy_1c(a_)
#undef p1_2_axpy_1c
#define p1_2_axpy_1c(a_) \
      pld(a_,cx,2) \
      pc(2,0) \
      pld(a_,ax,1) \
      ps(CSHUF,0,0) \
      pm(6,2) \
      pa(2,1) \
      pm(7,0) \
      pa(0,1) \
      pud(1,a_,ax)
#undef p1_axpy_1c
#define p1_axpy_1c(a_) \
      pl(a_,cx,2) \
      pc(2,0) \
      plq(a_,ax,1) \
      ps(CSHUF,0,0) \
      pm(6,2) \
      pa(2,1) \
      pm(7,0) \
      pa(0,1) \
      puq(1,a_,ax)
#undef p2_axpy_1c
#define p2_axpy_1c(a_) \
      plq(SS(a_,RS4),ax,3) \
      ps(CSHUF,0,0) \
      pl(SS(a_,RS4),cx,4) \
      pm(6,2) \
      pa(2,1) \
      pm(7,0) \
      pa(0,1) \
      pc(4,0) \
      puq(1,a_,ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      plq(SS(a_,MM(2,RS4)),ax,1) \
      ps(CSHUF,0,0) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      pm(6,4) \
      pa(4,3) \
      pm(7,0) \
      pa(0,3) \
      pc(2,0) \
      puq(3,SS(a_,RS4),ax)
#undef lpaxpy_1c
#define lpaxpy_1c(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,2) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      plq(SS(a_,MM(0,RS4)),ax,1) \
      pc(2,0) 
#undef dpaxpy_1c
#define dpaxpy_1c(a_) \
      plq(SS(a_,RS4),ax,3) \
      ps(CSHUF,0,0) \
      pl(SS(a_,RS4),cx,4) \
      pm(6,2) \
      pa(2,1) \
      pm(7,0) \
      pa(0,1) \
      pc(4,0) \
      puq(1,a_,ax) \
      ps(CSHUF,0,0) \
      pm(6,4) \
      pa(4,3) \
      pm(7,0) \
      pa(0,3) \
      puq(3,SS(a_,RS4),ax)
#undef plaxpy_1c
#define plaxpy_1c 8

#undef p1_4_copy_1
#define p1_4_copy_1(a_) \
      pls(a_,cx,2) \
      pus(2,a_,ax)
#undef p1_2_copy_1
#define p1_2_copy_1(a_) \
      pld(a_,cx,2) \
      pud(2,a_,ax)
#undef p1_copy_1
#define p1_copy_1(a_) \
      pl(a_,cx,2) \
      puq(2,a_,ax)
#undef p2_copy_1
#define p2_copy_1(a_) \
      pl(SS(a_,RS4),cx,4) \
      puq(2,a_,ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),ax) \
      f(nta,SS(a_,MM((SS(2,CL)),RS4)),cx) \
      pl(SS(a_,MM(2,RS4)),cx,2) \
      puq(4,SS(a_,RS4),ax)
#undef lpcopy_1
#define lpcopy_1(a_) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),ax) \
      f(nta,SS(a_,MM((SS(0,CL)),RS4)),cx) \
      pl(SS(a_,MM(0,RS4)),cx,2) 
#undef dpcopy_1
#define dpcopy_1(a_) \
      pl(SS(a_,RS4),cx,4) \
      puq(2,a_,ax) \
      puq(4,SS(a_,RS4),ax)
#undef plcopy_1
#define plcopy_1 8

#undef p1_4_copy_2
#define p1_4_copy_2(a_) \
      pls(a_,ax,2) \
      pus(2,a_,cx)
#undef p1_2_copy_2
#define p1_2_copy_2(a_) \
      pld(a_,ax,2) \
      pud(2,a_,cx)
#undef p1_copy_2
#define p1_copy_2(a_) \
      plq(a_,ax,2) \
      pu(2,a_,cx)
#undef p2_copy_2
#define p2_copy_2(a_) \
      plq(SS(a_,RS4),ax,4) \
      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) \
      pu(4,SS(a_,RS4),cx)
#undef lpcopy_2
#define lpcopy_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,2) 
#undef dpcopy_2
#define dpcopy_2(a_) \
      plq(SS(a_,RS4),ax,4) \
      pu(2,a_,cx) \
      pu(4,SS(a_,RS4),cx)
#undef plcopy_2
#define plcopy_2 8

#undef p1_4_copy_3
#define p1_4_copy_3(a_) \
      pls(a_,cx,2) \
      pus(2,a_,ax)
#undef p1_2_copy_3
#define p1_2_copy_3(a_) \
      pld(a_,cx,2) \
      pud(2,a_,ax)
#undef p1_copy_3
#define p1_copy_3(a_) \
      pl(a_,cx,2) \
      punt(2,a_,ax)
#undef p2_copy_3
#define p2_copy_3(a_) \
      pl(SS(a_,MM(0,RS4)),cx,0) \
      pl(SS(a_,MM(1,RS4)),cx,1) \
      punt(0,SS(a_,MM(0,RS4)),ax) \
      punt(1,SS(a_,MM(1,RS4)),ax) 
#undef p4_copy_3
#define p4_copy_3(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) \
      punt(0,SS(a_,MM(0,RS4)),ax) \
      punt(1,SS(a_,MM(1,RS4)),ax) \
      punt(2,SS(a_,MM(2,RS4)),ax) \
      punt(3,SS(a_,MM(3,RS4)),ax) 
#undef p8_copy_3
#define p8_copy_3(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) \
      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,6) \
      pl(SS(a_,MM(7,RS4)),cx,7) \
      punt(0,SS(a_,MM(0,RS4)),ax) \
      punt(1,SS(a_,MM(1,RS4)),ax) \
      punt(2,SS(a_,MM(2,RS4)),ax) \
      punt(3,SS(a_,MM(3,RS4)),ax) \
      punt(4,SS(a_,MM(4,RS4)),ax) \
      punt(5,SS(a_,MM(5,RS4)),ax) \