camm_dpa.h

基于Blas CLapck的.用过的人知道是干啥的

#undef dp
#define dp(a_,b_,c_)        fl(a_ ## 0,b_) fd(0) fm(srr(c_),0) fap(0,trr(c_)) \
                                                 fm(sri(c_),0) fap(0,tri(c_))\
                            fl(a_ ## 8,b_)       fm(0,sir(c_)) fmp(0,M(sir(c_),1)) \
                                                 fspi(0,M(tir(c_),2)) fapi(0,M(tii(c_),2))

#undef dpp
#define dpp(a_,b_,c_,d_,e_) fl(a_ ## 0,b_) fd(0) fm(srr(c_),0) fap(0,trr(c_)) \
                                                 pf(d_,e_) fm(sri(c_),0) fap(0,tri(c_))\
                            fl(a_ ## 8,b_)       fm(0,sir(c_)) fmp(0,M(sir(c_),1)) \
                                                 fspi(0,M(tir(c_),2)) fapi(0,M(tii(c_),2))


#endif

#else

#undef plaa
#define plaa(a_)            fl(a_ ## 0,si) fl(a_ ## 8,si) plax
#undef wa
#define wa(a_)

#undef ddprr
#define ddprr(a_,b_,c_)     fl(a_ ## 0,b_) \
                                              fd(tri(c_))           fm(P(sri(c_),1),0)      fap(0,1) \
                                              fd(M(trr(c_),1))      fm(srr(c_),0)           fspi(0,1) \
                            fp(a_ ## 0,b_) 
#undef ddpri
#define ddpri(a_,b_,c_)     fl(a_ ## 8,b_) \
                                              fd(tii(c_))           fm(P(sii(c_),1),0)      fap(0,1) \
                                              fd(M(tir(c_),1))      fm(sir(c_),0)           fapi(0,1) \
                            fp(a_ ## 8,b_) 
#undef dpri
#define dpri(a_,b_,c_)      fl(a_ ## 8,b_) \
                                              fx(2)                 fm(sir(c_),0)           fap(0,2) \
                                                                    fm(M(sii(c_),2),0)      fapi(0,1) \
                            fp(a_ ## 8,b_)


#undef ddpp
#define ddpp(a_,b_,c_,d_,e_) ddprr(a_,b_,c_) pf(d_,e_) ddpri(a_,b_,c_)
#undef ddp
#define ddp(a_,b_,c_)        ddprr(a_,b_,c_)           ddpri(a_,b_,c_)
#undef dpp
#define dpp(a_,b_,c_,d_,e_)  ddprr(a_,b_,c_) pf(d_,e_) dpri(a_,b_,c_)
#undef dp
#define dp(a_,b_,c_)         ddprr(a_,b_,c_)           dpri(a_,b_,c_)

#endif


#undef R1
#define R1 4
#undef R2
#define R2 6
#undef R3
#define R3 6
#undef R4
#define R4 6

#endif

#endif


/******************************************************************************
 *  General Macros
 ******************************************************************************/  




#undef bla1
#define bla1(a_,b_)          plaa(a_) dpp(a_,ax,R1,b_,si) wa(a_) 
#undef blb1
#define blb1(a_,b_)          plaa(a_) dpp(a_,ax,R1,b_,ax) wa(a_)
			     
#undef bla2
#undef bla2
#define bla2(a_,b_)          pf(b_,si) plaa(a_) ddp(a_,ax,R1)        pf(b_,ax) dp(a_,bx,R2) wa(a_)
#undef blb2
#undef blb2
#define blb2(a_,b_)                    plaa(a_) ddpp(a_,ax,R1,b_,bx)           dp(a_,bx,R2) wa(a_) 
			     
#undef bla3
#define bla3(a_,b_)          plaa(a_) ddpp(a_,ax,R1,b_,si) ddp(a_,bx,R2) \
                             dpp(a_,cx,R3,b_,ax) wa(a_)
#undef blb3
#define blb3(a_,b_)          plaa(a_) ddpp(a_,ax,R1,b_,bx) ddp(a_,bx,R2) \
                             dpp(a_,cx,R3,b_,cx) wa(a_)
			     
#undef bla4
#define bla4(a_,b_)          plaa(a_) ddpp(a_,ax,R1,b_,si) ddpp(a_,bx,R2,b_,ax) \
                             ddp(a_,cx,R3) dpp(a_,dx,R4,b_,bx) wa(a_)
#undef blb4
#define blb4(a_,b_)          plaa(a_) ddp(a_,ax,R1)        ddpp(a_,bx,R2,b_,cx) \
                             ddp(a_,cx,R3) dpp(a_,dx,R4,b_,dx) wa(a_)

#undef bla
#define bla(a_,b_)      Mjoin(bla,NDP)(a_,b_)
#undef blb
#define blb(a_,b_)      Mjoin(blb,NDP)(a_,b_)



#undef bla11_2
#define bla11_2(a_)    plaa1_2(a_) dp1_2(a_,ax,R1) wa1_2(a_) 
#undef bla21_2
#define bla21_2(a_)    plaa1_2(a_) ddp1_2(a_,ax,R1) dp1_2(a_,bx,R2) wa1_2(a_)
#undef bla31_2
#define bla31_2(a_)    plaa1_2(a_) ddp1_2(a_,ax,R1) ddp1_2(a_,bx,R2) \
                          dp1_2(a_,cx,R3) wa1_2(a_)
#undef bla41_2
#define bla41_2(a_)    plaa1_2(a_) ddp1_2(a_,ax,R1) ddp1_2(a_,bx,R2) \
                          ddp1_2(a_,cx,R3) dp1_2(a_,dx,R4) wa1_2(a_)

#undef bla1_2
#define bla1_2(a_)     Mjoin(Mjoin(bla,NDP),1_2)(a_)



#undef bla11_4
#define bla11_4(a_)    plaa1_4(a_) dp1_4(a_,ax,R1) wa1_4(a_) 
#undef bla21_4
#define bla21_4(a_)    plaa1_4(a_) ddp1_4(a_,ax,R1) dp1_4(a_,bx,R2) wa1_4(a_)
#undef bla31_4
#define bla31_4(a_)    plaa1_4(a_) ddp1_4(a_,ax,R1) ddp1_4(a_,bx,R2) \
                          dp1_4(a_,cx,R3) wa1_4(a_)
#undef bla41_4
#define bla41_4(a_)    plaa1_4(a_) ddp1_4(a_,ax,R1) ddp1_4(a_,bx,R2) \
                          ddp1_4(a_,cx,R3) dp1_4(a_,dx,R4) wa1_4(a_)

#undef bla1_4
#define bla1_4(a_)     Mjoin(Mjoin(bla,NDP),1_4)(a_)



#undef inc1
#define inc1(a_)        a(a_,si) a(a_,ax)
#undef inc2
#define inc2(a_)        inc1(a_) a(a_,bx)
#undef inc3
#define inc3(a_)        inc2(a_) a(a_,cx)
#undef inc4
#define inc4(a_)        inc3(a_) a(a_,dx)

#undef inc
#define inc(a_)         Mjoin(inc,NDP)(a_)


#ifdef PREFETCH
/* #include "camm_arith.h" */
#undef S
#define S(a_,b_) (a_) + (b_)
#undef PF1
#define PF1 PREFETCH
#undef PF2
#define PF2 S(PF1,32)
#undef PF3
#define PF3 S(PF1,64)
#undef PF4
#define PF4 S(PF1,96)
#undef PF5
#define PF5 S(PF1,128)
#undef PF6
#define PF6 S(PF1,160)
#undef PF7
#define PF7 S(PF1,192)
#undef PF8
#define PF8 S(PF1,224)
#else
#undef PF1
#define PF1 64
#undef PF2
#define PF2 96
#undef PF3
#define PF3 128
#undef PF4
#define PF4 160
#undef PF5
#define PF5 192
#undef PF6
#define PF6 224
#undef PF7
#define PF7 256
#undef PF8
#define PF8 288
#endif


#if defined(NO_TRANSPOSE) && !defined(SREAL) && !defined(GER)
#undef pf
#define pf(a_,b_)  f(t0,a_,b_)
#else
#undef pf
#define pf(a_,b_)  f(nta,a_,b_)
#endif

#undef bl1
#define bl1            bla1_4(0x0) inc(4)
#undef bl2
#define bl2            bla1_2(0x0) inc(8)
#undef bl4
#define bl4            bla(0x0,PF1) inc(16)
#undef bl8
#define bl8            bla(0x0,PF1) blb(0x1,PF1) inc(32) 
#undef bl16
#define bl16           bla(0x0,PF1) blb(0x1,PF1) bla(0x2,PF2) blb(0x3,PF2) inc(64)
#undef bl32
#define bl32           bla(0x0,PF1) blb(0x1,PF1) bla(0x2,PF2) blb(0x3,PF2) \
                       bla(0x4,PF3) blb(0x5,PF3) bla(0x6,PF4) blb(0x7,PF4) inc(128)
#undef bl64
#define bl64           bla(0x0,PF1) blb(0x1,PF1) bla(0x2,PF2) blb(0x3,PF2) \
                       bla(0x4,PF3) blb(0x5,PF3) bla(0x6,PF4) blb(0x7,PF4) \
                       bla(0x8,PF5) blb(0x9,PF5) bla(0xa,PF6) blb(0xb,PF6) \
                       bla(0xc,PF7) blb(0xd,PF7) bla(0xe,PF8) blb(0xf,PF8) inc(256)

/* #define in2           inc(8) */
/* #define in4           inc(16) */
/* #define in8           inc(32) */
/* #define in16          inc(64) */

#undef in2
#define in2  
#undef in4
#define in4  
#undef in8
#define in8  
#undef in16
#define in16 

#ifdef NO_TRANSPOSE
#undef incf
#define incf           ra(di,si)
#else
#undef incf
#define incf
#endif

#undef lf1
#define lf1            mpx(R1)
#undef lf2
#define lf2            lf1 incf mpx(R2)
#undef lf3
#define lf3            lf2 incf mpx(R3)
#undef lf4
#define lf4            lf3 incf mpx(R4)

#undef lf
#define lf             Mjoin(lf,NDP)


#undef ulf1
#define ulf1           ulfa(R1)
#undef ulf2
#define ulf2           ulf1 ra(di,si) ulfa(R2) 
#undef ulf3
#define ulf3           ulf2 ra(di,si) ulfa(R3) 
#undef ulf4
#define ulf4           ulf3 ra(di,si) ulfa(R4) 

#undef ulf
#define ulf            Mjoin(ulf,NDP)

#undef lpba
#define lpba(a_)      "movl %%esi,%%e" #a_ "\n\t"

#undef lpb1
#define lpb1          lpba(ax)
#undef lpb2
#define lpb2          lpb1 ra(di,si) lpba(bx)
#undef lpb3
#define lpb3          lpb2 ra(di,si) lpba(cx)
#undef lpb4
#define lpb4          lpb3 ra(di,si) lpba(dx)

#undef lpb
#define lpb           Mjoin(lpb,NDP)

#undef ipf1
#define ipf1(a_)   pf(a_,si) pf(a_,ax)
#undef ipf2
#define ipf2(a_)   ipf1(a_)  pf(a_,bx) 
#undef ipf3
#define ipf3(a_)   ipf2(a_)  pf(a_,cx) 
#undef ipf4
#define ipf4(a_)   ipf3(a_)  pf(a_,dx) 

#undef ipf
#define ipf(a_)     Mjoin(ipf,NDP)(a_)

#ifdef LUNROLL
#undef UNROLL
#ifdef SREAL
#undef UNROLL
#define UNROLL LUNROLL
#elif defined(DREAL) || defined(SCPLX)
#undef UNROLL
#define UNROLL LUNROLL*2
#elif defined(DCPLX)
#undef UNROLL
#define UNROLL LUNROLL*4
#endif
#else
#undef UNROLL
#define UNROLL 16
#endif

#undef UNROLL1_2
#if UNROLL == 64
#undef blUNROLL
#define blUNROLL bl64
#undef UNROLL1_2
#define UNROLL1_2 32
#elif UNROLL == 32
#undef blUNROLL
#define blUNROLL bl32
#undef UNROLL1_2
#define UNROLL1_2 16
#elif UNROLL == 16
#undef blUNROLL
#define blUNROLL bl16
#undef UNROLL1_2
#define UNROLL1_2 8
#elif UNROLL == 8
#undef blUNROLL
#define blUNROLL bl8
#undef UNROLL1_2
#define UNROLL1_2 4
#elif UNROLL == 4
#undef blUNROLL
#define blUNROLL bl4
#undef UNROLL1_2
#define UNROLL1_2 2
#elif UNROLL == 2
#undef blUNROLL
#define blUNROLL bl2
#undef UNROLL1_2
#define UNROLL1_2 1
#elif UNROLL == 1
#undef blUNROLL
#define blUNROLL bl1
#undef UNROLL1_2
#define UNROLL1_2 stop
#endif
#ifndef UNROLL1_2
#error UNROLL must be set to power of 2 < 128
#endif


#ifdef GER
#undef aconst
#define aconst
#undef cconst
#define cconst const
#else
#undef aconst
#define aconst const
#undef cconst
#define cconst
#endif

#undef MY_FUNCTION
#define MY_FUNCTION Mjoin(dp,EXT)

static void
MY_FUNCTION(aconst TYPE *a,int lda,
	      const TYPE *b,
	      cconst TYPE *c,int stride,int len) {

#ifdef SCPLX
#if defined(GER) && defined(Conj_)
    const TYPE w1[2]={{-1.0,1.0},{-1.0,1.0}},*w=w1;
#else
    const TYPE w1[2]={{1.0,-1.0},{1.0,-1.0}},*w=w1;
#endif
#endif

#if defined(DCPLX) && defined(ATL_SSE2)
#if defined(GER) && defined(Conj_)
    const TYPE w1[1]={{-1.0,1.0}},*w=w1;
#else
    const TYPE w1[1]={{1.0,-1.0}},*w=w1;
#endif
#endif

#ifdef NO_TRANSPOSE
#undef movm
#define movm c
#undef fixm
#define fixm b
#else
#undef movm
#define movm b
#undef fixm
#define fixm c
#endif    
    NO_INLINE
    unsigned u1=stride*sizeof(*fixm),u2=lda*sizeof(*a),u3=len*sizeof(*movm)/sizeof(float);

    ASM (

	 "pushl %%ebx\n\t"
	 a(4,sp)

#if defined(SCPLX) || (defined(DCPLX) && defined(ATL_SSE2))
	 "movl %6,%%esi\n\t"
	 pl(0,si,SREG)
#endif
	 
#ifdef NO_TRANSPOSE
	 "movl %1,%%esi\n\t"  /* fixm */
	 "movl %2,%%edi\n\t"  /* fixm2fixm */
#endif

	 lf

	 "movl %3,%%esi\n\t"  /* a */
	 "movl %4,%%edi\n\t"  /* a2a */

	 lpb

	 ipf(0)

	 "movl %0,%%esi\n\t"  /* movm */
	 "movl %5,%%edi\n\t"  /* len */

#if defined(ALIGN)

#if defined(SREAL)

	 test(4,ax)
	 je(Mjoin(a1,EXT))
	 test(-1,di)
	 je(Mjoin(a1,EXT))
	 sub(1,di)
	 bl1

	 lab(Mjoin(a1,EXT))

#endif

#if defined(DREAL) || defined(SREAL)

	 test(8,ax)
	 je(Mjoin(as,EXT))
	 test(-2,di)
	 je(Mjoin(as,EXT))
	 sub(2,di)
	 bl2

	 lab(Mjoin(as,EXT))

#endif

#endif
	      

	 ipf(32)

	 lab(Mjoin(loop,EXT))

	 test(-UNROLL,di)
	 je(Mjoin(UNROLL1_2,EXT))
	 sub(UNROLL,di)

	 blUNROLL
	 
	 jmp(Mjoin(loop,EXT))

#if UNROLL > 32
	 lab(Mjoin(32,EXT))
	 test(32,di)
	 je(Mjoin(16,EXT))
	 bl32
#endif	 

#if UNROLL > 16
	 lab(Mjoin(16,EXT))
	 test(16,di)
	 je(Mjoin(8,EXT))
	 bl16
#endif	 

#if UNROLL > 8
	 lab(Mjoin(8,EXT))
	 test(8,di)
	 je(Mjoin(4,EXT))
	 bl8
#endif	 

#if UNROLL > 4
	 lab(Mjoin(4,EXT))
	 test(4,di)
	 je(Mjoin(2,EXT))
	 bl4
#endif

#if UNROLL > 2	 
	 lab(Mjoin(2,EXT))
#ifndef DCPLX
	 test(2,di)
	 je(Mjoin(1,EXT))
	 bl2
#endif
#endif

#if UNROLL > 1
	 lab(Mjoin(1,EXT))
#ifdef SREAL
	 test(1,di)
	 je(Mjoin(stop,EXT))
	 bl1
#endif
#endif

	 lab(Mjoin(stop,EXT))

#ifndef NO_TRANSPOSE
	 "movl %1,%%esi\n\t"  /* fixm */
	 "movl %2,%%edi\n\t"  /* fixm2fixm */
#endif

	 ulf

	 a(-4,sp)
	 "popl %%ebx\n\t"


	 ::"m" (movm),"m" (fixm),"m" (u1),"m" (a),"m" (u2),"m" (u3)

#if defined(SCPLX) || (defined(DCPLX) && defined(ATL_SSE2))
	 ,"m" (w)
#endif
	 :"ax","bx","cx","dx","si","di");


}