static void
add_memory_record(void *p,unsigned long size)
{
struct memory_record *mr= (struct memory_record *)malloc(sizeof(struct memory_record));
mr->p= p;
mr->size= size;
grab_stack(mr->ra,MR_CALL_STACK,1);
PR_Lock( mr_tree_lock );
avl_insert( &mr_tree, mr, memory_record_compare, memory_record_duplicate_disallow );
PR_Unlock( mr_tree_lock );
}
void qcdoc_su3_recon( char *name)
{
/**** This section defines all the registers and offsets I need ****/
/*
* This marks the argument registers as defined by ABI as off limits
* to us until they are freed by "getarg()";
*/
int dum = defargcount(4);
/*Handle for the loop entry point*/
int branchsite;
int branchmu;
int retno ;
/*------------------------------------------------------------------
* Floating point registers
*------------------------------------------------------------------
*/
// Reconstruct 8 registers for 4 spinor
// reg_array_2d(PSI,Fregs,4,2);
reg_array_3d(PSI,Fregs,3,4,2);
offset_3d(PSI_IMM,FourSpinType,4,3,2); /*Offsets within 4 spinor*/
// Reconstruct 2 spinor registers
#define NEO 2
reg_array_3d(Atmp,Fregs,1,2,2); /*CHIplus regs */
reg_array_3d(Btmp,Fregs,1,2,2); /*CHIminus regs */
int A[NEO][2][2] = {
Atmp[0][0][0], Atmp[0][0][1],
Atmp[0][1][0], Atmp[0][1][1],
-1,-1,-1,-1
};
int B[NEO][2][2] = {
Btmp[0][0][0], Btmp[0][0][1],
Btmp[0][1][0], Btmp[0][1][1],
-1,-1,-1,-1
};
/*Regs for SU3 two spinor multiply ... overlap with the reconstruct*/
/* registers */
int CHIR[3][2][2] = {
A[0][0][0],A[0][0][1],
A[0][1][0],A[0][1][1],
B[0][0][0],B[0][0][1],
B[0][1][0],B[0][1][1],
PSI[0][0][0],PSI[0][0][1],
PSI[0][1][0],PSI[0][1][1]
};
offset_3d(CHI_IMM,TwoSpinType,3,2,2);
/*Registers for the gauge link (2 rows)*/
int UA[3][2] = {
{PSI[0][2][0],PSI[0][2][1]},
{PSI[2][1][0],PSI[2][1][1]},
{PSI[1][0][0],PSI[1][0][1]}
};
int UB[3][2] = {
{PSI[1][1][0],PSI[1][1][1]},
{PSI[2][0][0],PSI[2][0][1]},
{PSI[1][2][0],PSI[1][2][1]},
};
offset_3d(GIMM , GaugeType, 3, 3 ,2 );
// Other 8 registers used for reduction variables in SU3.
// Could use these in reconstruct??
int E[2] = { PSI[2][2][0],PSI[2][2][1]};
/*
* FCD used for drain of Chi
* Overlap with PSI[*][3][*]
*/
int F[2] = {PSI[0][3][0],PSI[0][3][1]};
int C[2] = {PSI[1][3][0],PSI[1][3][1]};
int D[2] = {PSI[2][3][0],PSI[2][3][1]};
/*
* Integer registers
*/
alreg(psi,Iregs);
alreg(Umu,Iregs);
alreg(Ufetch,Iregs);
alreg(Chiin,Iregs);
alreg(Chiout,Iregs);
alreg(Chifetch,Iregs);
reg_array_1d(Chiplus,Iregs,4);/*Pointers to the 8 2-spinors for recombination*/
reg_array_1d(Chiminus,Iregs,4);
alreg(mu,Iregs);
alreg(Chidrain,Iregs);
alreg(pref,Iregs);
alreg(mem,Iregs);
alreg(length,Iregs);
int Isize = PROC->I_size;
int Fsize = PROC->FP_size;
def_off( ZERO_IMM, Byte,0);
def_off( PSI_ATOM, FourSpinType, 24);
def_off( CHI_ATOM, TwoSpinType, 12);
def_off( PAD_CHI_ATOM, TwoSpinType, 16);
def_off( MAT_IMM, GaugeType, 18);
int Ndim = def_offset(4,Byte,"Ndim");
int Ndimm1 = def_offset(3,Byte,"Ndimm1");
int hbias,bias;
/*Offsets handles to stack*/
int hbitbucket = def_offset(16*Isize,Byte,"hbitbucket");
int Tsize;
if ( TwoSpinType == Double ) Tsize = PROC->FP_size;
else Tsize = PROC->FSP_size;
int hstk0 = def_offset(16*Isize+12*Tsize ,Byte,"hstk0");
int hstk1 = def_offset(16*Isize+2*12*Tsize,Byte,"hstk1");
int hstk2 = def_offset(16*Isize+3*12*Tsize,Byte,"hstk2");
int hstk3 = def_offset(16*Isize+4*12*Tsize,Byte,"hstk3");
int hIsize = def_offset(Isize,Byte,"Isize");
int i,co,j,k,nxt,ri,sp,nxtco,eop,eo_a,eo_b;
/***********************************************************************/
/*
* PROLOGUE
*/
make_inst(DIRECTIVE,Enter_Routine,name);
/*Allocate stack save any callee save registers we need etc...*/
int stack_buf_size;
stack_buf_size = 16*Isize +
12*Fsize * 5 ;
hbias = grab_stack(stack_buf_size);
bias = get_offset(hbias);
save_regs();
queue_iadd_imm(mem,PROC->StackPointer,hbias); /*Pointer to buf on stack*/
/*Define our arguments - all pointers ala fortran*/
getarg(psi);
getarg(Umu);
getarg(Chiin);
getarg(length);
/*{... Process arguments ...*/
queue_iload(length,ZERO_IMM,length); /*Load in sx counter*/
retno = get_target_label(); /*Branch to exit if yzt <1*/
check_iterations(length,retno);
need_cache_line(0);
need_cache_line(1);
need_cache_line(2);
need_cache_line(3);
need_cache_line(4);
pragma(DCBT_SPACE,5);
pragma(DCBT_POST,1);
#define LOAD_U(comin,comax)\
/*Load two link rows*/\
for( i = comin;i<=comax;i++ ){\
for( ri=0;ri<2;ri++){ \
queue_fload(UA[i][ri],GIMM[i][0][ri],Umu,GaugeType);\
queue_fload(UB[i][ri],GIMM[i][1][ri],Umu,GaugeType);\
} \
}
#define PRELOAD_U LOAD_U(0,1)
#define POSTLOAD_U LOAD_U(2,2)
PRELOAD_U
#define LOAD_CHI(comin,comax) \
/*Load Chi column*/\
for( i = comin;i<=comax;i++ ){\
for( ri=0;ri<2;ri++){\
queue_fload(CHIR[i][0][ri],CHI_IMM[i][0][ri],Chiin,TwoSpinType);\
} \
for( ri=0;ri<2;ri++){\
queue_fload(CHIR[i][1][ri],CHI_IMM[i][1][ri],Chiin,TwoSpinType);\
} \
}
#define PRELOAD_CHI LOAD_CHI(0,1)
#define POSTLOAD_CHI LOAD_CHI(2,2)
#define POSTLOAD \
POSTLOAD_CHI \
POSTLOAD_U
do_prefetch(Chiin,0);
do_prefetch(Chiin,1);
if ( SizeofDatum(TwoSpinType) == 8 ) do_prefetch(Chiin,2);
PRELOAD_CHI
/*
* Start site loop
*/
queue_iadd_imm(Chidrain,mem,hbitbucket);
branchsite = start_loop(length);
queue_iadd_imm(Chiout,mem,hstk0);
/*
* Loop over mu in asm
*/
queue_iload_imm(mu,Ndimm1);
#define CHIDRAIN \
queue_fstore(F[0],CHI_IMM[1][1][0],Chidrain,TwoSpinType);\
queue_fstore(F[1],CHI_IMM[1][1][1],Chidrain,TwoSpinType);\
queue_fstore(C[0],CHI_IMM[2][0][0],Chidrain,TwoSpinType);\
queue_fstore(C[1],CHI_IMM[2][0][1],Chidrain,TwoSpinType);\
queue_fstore(D[0],CHI_IMM[2][1][0],Chidrain,TwoSpinType);\
queue_fstore(D[1],CHI_IMM[2][1][1],Chidrain,TwoSpinType);
#define PREFETCH_CHI \
queue_iadd_imm(Chifetch,Chiin,PAD_CHI_ATOM);\
do_prefetch(Chifetch,0);\
do_prefetch(Chifetch,1);\
if ( SizeofDatum(TwoSpinType) == 8 ) do_prefetch(Chifetch,2);
#define PREFETCH_CHIF \
queue_iadd_imm(Chifetch,Chifetch,PAD_CHI_ATOM);\
do_prefetch(Chifetch,0);\
do_prefetch(Chifetch,1);\
if ( SizeofDatum(TwoSpinType) == 8 ) do_prefetch(Chifetch,2);
for ( int unroll=0; unroll<2; unroll++ ) {
if ( unroll==0 ) {
branchmu = start_loop(mu);
pragma(DCBT_SPACE,5);
pragma(STORE_LIM,1);
pragma(LOAD_LIM,2);
} else {
pragma(STORE_LIM,2);
pragma(DCBT_SPACE,5);
pragma(DCBT_POST,1);
pragma(DCBT_PRE,0);
pragma(LOAD_LIM,2);
}
CHIDRAIN
POSTLOAD
if ( unroll == 0 ) {
PREFETCH_CHI
queue_iadd_imm(Ufetch,Umu,MAT_IMM);
do_prefetch(Ufetch,0);
do_prefetch(Ufetch,1);
do_prefetch(Ufetch,2);
if ( GaugeType == Double ) {
do_prefetch(Ufetch,3);
do_prefetch(Ufetch,4);
}
} else {
pragma(DCBT_SPACE,3);
PREFETCH_CHI
PREFETCH_CHIF
PREFETCH_CHIF
PREFETCH_CHIF
}
j=0;
queue_three_cmuls(C[0],C[1],UA[j][0],UA[j][1],CHIR[j][0][0],CHIR[j][0][1],
D[0],D[1],UA[j][0],UA[j][1],CHIR[j][1][0],CHIR[j][1][1],
E[0],E[1],UB[j][0],UB[j][1],CHIR[j][0][0],CHIR[j][0][1]);
j=1;
queue_three_cmadds(C[0],C[1],UA[j][0],UA[j][1],CHIR[j][0][0],CHIR[j][0][1],
D[0],D[1],UA[j][0],UA[j][1],CHIR[j][1][0],CHIR[j][1][1],
E[0],E[1],UB[j][0],UB[j][1],CHIR[j][0][0],CHIR[j][0][1]);
j=2;
queue_three_cmadds(C[0],C[1],UA[j][0],UA[j][1],CHIR[j][0][0],CHIR[j][0][1],
D[0],D[1],UA[j][0],UA[j][1],CHIR[j][1][0],CHIR[j][1][1],
E[0],E[1],UB[j][0],UB[j][1],CHIR[j][0][0],CHIR[j][0][1]);
/*Store the first three results*/
queue_fstore(C[0],CHI_IMM[0][0][0],Chiout,TwoSpinType);
queue_fstore(C[1],CHI_IMM[0][0][1],Chiout,TwoSpinType);
queue_fstore(D[0],CHI_IMM[0][1][0],Chiout,TwoSpinType);
queue_fstore(D[1],CHI_IMM[0][1][1],Chiout,TwoSpinType);
queue_fstore(E[0],CHI_IMM[1][0][0],Chiout,TwoSpinType);
queue_fstore(E[1],CHI_IMM[1][0][1],Chiout,TwoSpinType);
/*Load the third row*/
for(j=0; j<3; j++) {
for(ri=0; ri<2; ri++) {
queue_fload(UA[j][ri],GIMM[j][2][ri],Umu,GaugeType);
}
}
/*Gauge layout is linear, mu faster than site*/
queue_iadd_imm(Umu,Umu,MAT_IMM);
/*Now the second set of three cdots*/
j=0;
queue_three_cmuls(F[0],F[1],UB[j][0],UB[j][1],CHIR[j][1][0],CHIR[j][1][1],
C[0],C[1],UA[j][0],UA[j][1],CHIR[j][0][0],CHIR[j][0][1],
D[0],D[1],UA[j][0],UA[j][1],CHIR[j][1][0],CHIR[j][1][1]);
j=1;
queue_three_cmadds(F[0],F[1],UB[j][0],UB[j][1],CHIR[j][1][0],CHIR[j][1][1],
C[0],C[1],UA[j][0],UA[j][1],CHIR[j][0][0],CHIR[j][0][1],
D[0],D[1],UA[j][0],UA[j][1],CHIR[j][1][0],CHIR[j][1][1]);
j=2;
queue_three_cmadds(F[0],F[1],UB[j][0],UB[j][1],CHIR[j][1][0],CHIR[j][1][1],
C[0],C[1],UA[j][0],UA[j][1],CHIR[j][0][0],CHIR[j][0][1],
D[0],D[1],UA[j][0],UA[j][1],CHIR[j][1][0],CHIR[j][1][1]);
/**************END SU3 CODE *************/
queue_iadd_imm(Chiin,Chiin,PAD_CHI_ATOM);
queue_iadd_imm(Chidrain,Chiout,ZERO_IMM);
queue_iadd_imm(Chiout,Chiout,CHI_ATOM);
if ( unroll == 0 ) {
PRELOAD_U
PRELOAD_CHI
}
/*********************************************************/
/****************** END OF SU3 MULTIPLY ******************/
/*********************************************************/
if ( unroll== 0 ) {
stop_loop(branchmu,mu); /* End loop over mu*/
make_inst(DIRECTIVE,Target,get_target_label() ); /*delineate the sections*/
}
}
/*********************************************************/
/****************** START OF RECONSTRUCT *****************/
/*********************************************************/
//Address calculation...
// Chiminus -> Stack and ChiPlus -> Chiin
pragma(STORE_INORDER,1);
queue_iadd_imm(Chiminus[0],mem,hstk0);
/*For register use reasons loop over colour outermost*/
#define LOAD_CHI_MU0(eo,co) \
for( sp = 0; sp<2;sp++ ){\
for( ri = 0; ri<2;ri++ ){\
queue_fload(A[eo][sp][ri],CHI_IMM[co][sp][ri],Chiminus[0],TwoSpinType);\
if ( co == 0 ) {\
queue_fload(B[eo][sp][ri],CHI_IMM[co][sp][ri],Chiin,TwoSpinType);\
queue_iadd_imm(Chiplus[0],Chiin,ZERO_IMM);\
} else {\
queue_fload(B[eo][sp][ri],CHI_IMM[co][sp][ri],Chiplus [0],TwoSpinType);\
}\
}}
pragma(LOAD_LIM,2);
LOAD_CHI_MU0(0,0)
pragma(DCBT_POST,1);
CHIDRAIN
int neo_a = NEO;
int neo_b = NEO;
eo_a = 0;
eo_b = 0;
for ( co = 0; co <3 ; co ++ ) {
pragma(LOAD_LIM,1);
if ( co == 0 ) {
// Use the third colour for unrolling the loads
A[1][0][0] = PSI[2][0][0];
A[1][0][1] = PSI[2][0][1];
A[1][1][0] = PSI[2][1][0];
A[1][1][1] = PSI[2][1][1];
B[1][0][0] = PSI[2][2][0];
B[1][0][1] = PSI[2][2][1];
B[1][1][0] = PSI[2][3][0];
B[1][1][1] = PSI[2][3][1];
queue_iadd_imm(Chiminus[1],mem,hstk1); // This is invariant of loop
// Take out
queue_iadd_imm(Chiplus[1],Chiin ,PAD_CHI_ATOM);
}
/***************************************************************
* MU = 0 reconstruct *
****************************************************************/
if ( co == 2 ) {
// Flip to not unrolled due to register pressure
neo_b = 1;
neo_a = 2;
A[1][0][0] = PSI[0][0][0];
A[1][0][1] = PSI[0][0][1];
A[1][1][0] = PSI[1][0][0];
A[1][1][1] = PSI[1][0][1];
pragma(DCBT_POST,0);
pragma(DCBT_SPACE,1);
queue_iadd_imm(Ufetch,Umu,ZERO_IMM);
// do_prefetch(Ufetch,0);
do_prefetch(Ufetch,1);
do_prefetch(Ufetch,2);
if ( GaugeType == Double ) {
do_prefetch(Ufetch,3);
do_prefetch(Ufetch,4);
}
}
/* psi_0 = Chiplus[0] + Chiminus[0] */
/* psi_1 = Chiplus[1] + Chiminus[1] */
queue_fadd(PSI[co][0][0],B[eo_b][0][0],A[eo_a][0][0]);
queue_fadd(PSI[co][0][1],B[eo_b][0][1],A[eo_a][0][1]);
queue_fadd(PSI[co][1][0],B[eo_b][1][0],A[eo_a][1][0]);
queue_fadd(PSI[co][1][1],B[eo_b][1][1],A[eo_a][1][1]);
// Dagger = 0:
/* psi_2 =-iChiplus[1] +iChiminus[1] */
/* psi_3 =-iChiplus[0] +iChiminus[0] */
// Dagger = 1:
/* psi_2 = iChiplus[1] -iChiminus[1] */
/* psi_3 = iChiplus[0] -iChiminus[0] */
if ( dagger == 0 ) {
queue_fsub(PSI[co][2][0],B[eo_b][1][1],A[eo_a][1][1]);
queue_fsub(PSI[co][2][1],A[eo_a][1][0],B[eo_b][1][0]);
queue_fsub(PSI[co][3][0],B[eo_b][0][1],A[eo_a][0][1]);
queue_fsub(PSI[co][3][1],A[eo_a][0][0],B[eo_b][0][0]);
} else {
queue_fsub(PSI[co][2][0],A[eo_a][1][1],B[eo_b][1][1]);
queue_fsub(PSI[co][2][1],B[eo_b][1][0],A[eo_a][1][0]);
queue_fsub(PSI[co][3][0],A[eo_a][0][1],B[eo_b][0][1]);
queue_fsub(PSI[co][3][1],B[eo_b][0][0],A[eo_a][0][0]);
}
/***************************************************************
* MU = 1 reconstruct *
****************************************************************/
eo_a = (eo_a+1)%neo_a;
eo_b = (eo_b+1)%neo_b;
for( sp = 0; sp<2; sp++ ) {
for( ri = 0; ri<2; ri++ ) {
queue_fload(A[eo_a][sp][ri],CHI_IMM[co][sp][ri],Chiminus[1],TwoSpinType);
queue_fload(B[eo_b][sp][ri],CHI_IMM[co][sp][ri],Chiplus [1],TwoSpinType);
}
}
if ( co == 0 ) {
queue_iadd_imm(Chiminus[2],mem,hstk2);
queue_iadd_imm(Chiminus[3],mem,hstk3);
queue_iadd_imm(Chiplus[2],Chiplus[1],PAD_CHI_ATOM);
queue_iadd_imm(Chiplus[3],Chiplus[2],PAD_CHI_ATOM);
}
/* psi_0 += Chiplus[0] + Chiminus[0] */
/* psi_1 += Chiplus[1] + Chiminus[1] */
queue_fadd(PSI[co][0][0],PSI[co][0][0],B[eo_b][0][0]);
queue_fadd(PSI[co][0][1],PSI[co][0][1],B[eo_b][0][1]);
queue_fadd(PSI[co][1][0],PSI[co][1][0],B[eo_b][1][0]);
queue_fadd(PSI[co][1][1],PSI[co][1][1],B[eo_b][1][1]);
queue_fadd(PSI[co][0][0],PSI[co][0][0],A[eo_a][0][0]);
queue_fadd(PSI[co][0][1],PSI[co][0][1],A[eo_a][0][1]);
queue_fadd(PSI[co][1][0],PSI[co][1][0],A[eo_a][1][0]);
queue_fadd(PSI[co][1][1],PSI[co][1][1],A[eo_a][1][1]);
//Dagger == 0
/* psi_2 += Chiplus[1] - Chiminus[1] */
/* psi_3 += -Chiplus[0] + Chiminus[0] */
//Dagger == 1
/* psi_2 -= Chiplus[1] - Chiminus[1] */
/* psi_3 -= -Chiplus[0] + Chiminus[0] */
if ( dagger == 0 ) {
queue_fadd(PSI[co][2][0],PSI[co][2][0],B[eo_b][1][0]);
queue_fadd(PSI[co][2][1],PSI[co][2][1],B[eo_b][1][1]);
queue_fsub(PSI[co][2][0],PSI[co][2][0],A[eo_a][1][0]);
queue_fsub(PSI[co][2][1],PSI[co][2][1],A[eo_a][1][1]);
queue_fsub(PSI[co][3][0],PSI[co][3][0],B[eo_b][0][0]);
queue_fsub(PSI[co][3][1],PSI[co][3][1],B[eo_b][0][1]);
queue_fadd(PSI[co][3][0],PSI[co][3][0],A[eo_a][0][0]);
queue_fadd(PSI[co][3][1],PSI[co][3][1],A[eo_a][0][1]);
} else {
queue_fsub(PSI[co][2][0],PSI[co][2][0],B[eo_b][1][0]);
queue_fsub(PSI[co][2][1],PSI[co][2][1],B[eo_b][1][1]);
queue_fadd(PSI[co][2][0],PSI[co][2][0],A[eo_a][1][0]);
queue_fadd(PSI[co][2][1],PSI[co][2][1],A[eo_a][1][1]);
queue_fadd(PSI[co][3][0],PSI[co][3][0],B[eo_b][0][0]);
queue_fadd(PSI[co][3][1],PSI[co][3][1],B[eo_b][0][1]);
queue_fsub(PSI[co][3][0],PSI[co][3][0],A[eo_a][0][0]);
queue_fsub(PSI[co][3][1],PSI[co][3][1],A[eo_a][0][1]);
}
/***************************************************************
* MU = 2 reconstruct *
****************************************************************/
eo_a = (eo_a+1)%neo_a;
eo_b = (eo_b+1)%neo_b;
for( sp = 0; sp<2; sp++ ) {
for( ri = 0; ri<2; ri++ ) {
queue_fload(A[eo_a][sp][ri],CHI_IMM[co][sp][ri],Chiminus[2],TwoSpinType);
queue_fload(B[eo_b][sp][ri],CHI_IMM[co][sp][ri],Chiplus [2],TwoSpinType);
}
}
/* psi_0 += Chiplus[0] + Chiminus[0] */
/* psi_1 += Chiplus[1] + Chiminus[1] */
queue_fadd(PSI[co][0][0],PSI[co][0][0],B[eo_b][0][0]);
queue_fadd(PSI[co][0][1],PSI[co][0][1],B[eo_b][0][1]);
queue_fadd(PSI[co][1][0],PSI[co][1][0],B[eo_b][1][0]);
queue_fadd(PSI[co][1][1],PSI[co][1][1],B[eo_b][1][1]);
queue_fadd(PSI[co][0][0],PSI[co][0][0],A[eo_a][0][0]);
queue_fadd(PSI[co][0][1],PSI[co][0][1],A[eo_a][0][1]);
queue_fadd(PSI[co][1][0],PSI[co][1][0],A[eo_a][1][0]);
queue_fadd(PSI[co][1][1],PSI[co][1][1],A[eo_a][1][1]);
//Dagger == 0
/* psi_2 +=-iChiplus[0] +iChiminus[0] */
/* psi_3 += iChiplus[1] -iChiminus[1] */
//Dagger == 1
/* psi_2 -=-iChiplus[0] +iChiminus[0] */
/* psi_3 -= iChiplus[1] -iChiminus[1] */
if ( dagger == 0 ) {
queue_fadd(PSI[co][2][0],PSI[co][2][0],B[eo_b][0][1]);
queue_fsub(PSI[co][2][1],PSI[co][2][1],B[eo_b][0][0]);
queue_fsub(PSI[co][2][0],PSI[co][2][0],A[eo_a][0][1]);
queue_fadd(PSI[co][2][1],PSI[co][2][1],A[eo_a][0][0]);
queue_fsub(PSI[co][3][0],PSI[co][3][0],B[eo_b][1][1]);
queue_fadd(PSI[co][3][1],PSI[co][3][1],B[eo_b][1][0]);
queue_fadd(PSI[co][3][0],PSI[co][3][0],A[eo_a][1][1]);
queue_fsub(PSI[co][3][1],PSI[co][3][1],A[eo_a][1][0]);
} else {
queue_fsub(PSI[co][2][0],PSI[co][2][0],B[eo_b][0][1]);
queue_fadd(PSI[co][2][1],PSI[co][2][1],B[eo_b][0][0]);
queue_fadd(PSI[co][2][0],PSI[co][2][0],A[eo_a][0][1]);
queue_fsub(PSI[co][2][1],PSI[co][2][1],A[eo_a][0][0]);
queue_fadd(PSI[co][3][0],PSI[co][3][0],B[eo_b][1][1]);
queue_fsub(PSI[co][3][1],PSI[co][3][1],B[eo_b][1][0]);
queue_fsub(PSI[co][3][0],PSI[co][3][0],A[eo_a][1][1]);
queue_fadd(PSI[co][3][1],PSI[co][3][1],A[eo_a][1][0]);
}
/***************************************************************
* MU = 3 reconstruct *
****************************************************************/
pragma(LOAD_LIM,2);
eo_a = (eo_a+1)%neo_a;
eo_b = (eo_b+1)%neo_b;
for( sp = 0; sp<2; sp++ ) {
for( ri = 0; ri<2; ri++ ) {
queue_fload(A[eo_a][sp][ri],CHI_IMM[co][sp][ri],Chiminus[3],TwoSpinType);
queue_fload(B[eo_b][sp][ri],CHI_IMM[co][sp][ri],Chiplus [3],TwoSpinType );
}
}
/* psi_0 += Chiplus[0] + Chiminus[0] */
/* psi_1 += Chiplus[1] + Chiminus[1] */
queue_fadd(PSI[co][0][0],PSI[co][0][0],B[eo_b][0][0]);
queue_fadd(PSI[co][0][1],PSI[co][0][1],B[eo_b][0][1]);
queue_fadd(PSI[co][1][0],PSI[co][1][0],B[eo_b][1][0]);
queue_fadd(PSI[co][1][1],PSI[co][1][1],B[eo_b][1][1]);
//Dagger == 0
/* psi_2 += Chiplus[0] - Chiminus[0] */
/* psi_3 += Chiplus[1] - Chiminus[1] */
//Dagger == 1
/* psi_2 -= Chiplus[0] - Chiminus[0] */
/* psi_3 -= Chiplus[1] - Chiminus[1] */
if ( dagger == 0 ) {
queue_fadd(PSI[co][2][0],PSI[co][2][0],B[eo_b][0][0]);
queue_fadd(PSI[co][2][1],PSI[co][2][1],B[eo_b][0][1]);
queue_fadd(PSI[co][3][0],PSI[co][3][0],B[eo_b][1][0]);
queue_fadd(PSI[co][3][1],PSI[co][3][1],B[eo_b][1][1]);
} else {
queue_fsub(PSI[co][2][0],PSI[co][2][0],B[eo_b][0][0]);
queue_fsub(PSI[co][2][1],PSI[co][2][1],B[eo_b][0][1]);
queue_fsub(PSI[co][3][0],PSI[co][3][0],B[eo_b][1][0]);
queue_fsub(PSI[co][3][1],PSI[co][3][1],B[eo_b][1][1]);
}
queue_fadd(PSI[co][0][0],PSI[co][0][0],A[eo_a][0][0]);
queue_fadd(PSI[co][0][1],PSI[co][0][1],A[eo_a][0][1]);
queue_fadd(PSI[co][1][0],PSI[co][1][0],A[eo_a][1][0]);
queue_fadd(PSI[co][1][1],PSI[co][1][1],A[eo_a][1][1]);
if ( dagger == 0 ) {
queue_fsub(PSI[co][2][0],PSI[co][2][0],A[eo_a][0][0]);
queue_fsub(PSI[co][2][1],PSI[co][2][1],A[eo_a][0][1]);
queue_fsub(PSI[co][3][0],PSI[co][3][0],A[eo_a][1][0]);
queue_fsub(PSI[co][3][1],PSI[co][3][1],A[eo_a][1][1]);
} else {
queue_fadd(PSI[co][2][0],PSI[co][2][0],A[eo_a][0][0]);
queue_fadd(PSI[co][2][1],PSI[co][2][1],A[eo_a][0][1]);
queue_fadd(PSI[co][3][0],PSI[co][3][0],A[eo_a][1][0]);
queue_fadd(PSI[co][3][1],PSI[co][3][1],A[eo_a][1][1]);
}
/*
* Store the spinors. If this is problematic
* in terms of PEC WriteBuf misses, I could
* store to the stack and copy out later.
*/
if ( co != 2 ) {
LOAD_CHI_MU0(0,co+1)
eo_a=0;
eo_b=0;
}
queue_fstore(PSI[co][0][0],PSI_IMM[0][co][0],psi,FourSpinType);
queue_fstore(PSI[co][0][1],PSI_IMM[0][co][1],psi,FourSpinType);
}
/*
* Store out in linear order now
*/
pragma(STORE_LIM,2);
pragma(DCBT_SPACE,8);
for ( co=0; co<3; co ++ ) {
queue_fstore(PSI[co][1][0],PSI_IMM[1][co][0],psi,FourSpinType);
queue_fstore(PSI[co][1][1],PSI_IMM[1][co][1],psi,FourSpinType);
}
for ( co=0; co<3; co ++ ) {
queue_fstore(PSI[co][2][0],PSI_IMM[2][co][0],psi,FourSpinType);
queue_fstore(PSI[co][2][1],PSI_IMM[2][co][1],psi,FourSpinType);
}
if ( TwoSpinType == FourSpinType ) {
queue_iadd_imm(Chidrain,psi,CHI_ATOM);
} else {
queue_iadd_imm(Chidrain,mem,hbitbucket);
for ( co=0; co<3; co ++ ) {
queue_fstore(PSI[co][3][0],PSI_IMM[3][co][0],psi,FourSpinType);
queue_fstore(PSI[co][3][1],PSI_IMM[3][co][1],psi,FourSpinType);
}
}
queue_iadd_imm(psi,psi,PSI_ATOM);
/*
* Put in an artificial dependency here
* to try to stop the preloads getting above the last load of
* reconstruct.
*/
queue_iadd_imm(Chiplus[3],Chiplus[3],ZERO_IMM);
queue_iadd_imm(Chiin ,Chiplus[3],PAD_CHI_ATOM);
pragma(DCBT_SPACE,0);
do_prefetch(Chiin,0);
do_prefetch(Chiin,1);
if ( SizeofDatum(TwoSpinType) == 8 )do_prefetch(Chiin,2);
PRELOAD_U
PRELOAD_CHI
/* TERMINATION point of the loop*/
stop_loop(branchsite,length);
CHIDRAIN
make_inst(DIRECTIVE,Target,retno);
/*
*
* EPILOGUE
*
*/
restore_regs();
free_stack();
make_inst(DIRECTIVE,Exit_Routine,name);
return;
}
void dwf_deriv( char *name)
{
/*
* This marks the argument registers as defined by ABI as off limits
* to us until they are freed by "getarg()";
*/
int dum = defargcount(1);
int retno;
/*
* S=phi^dag (MdagM)^-1 phi
*
* dS = phi^dag (MdagM)^-1 [ dMdag M + Mdag dM ] (MdagM)^-1 phi
*
* Let X = (MdagM)^-1 phi
* Y = M X = M^-dag phi
*
* Want terms: Ydag dM X
* Xdag dMdag Y
*
* Take Xdag 1-gamma Y
*
* Still a bit confused about the 1+g 1-g terms; but this may be simply a factor of two as we add +h.c.
* Will continue to follow Chroma's routine
*/
reg_array_2d(Y,Cregs,4,3); // 4 spinor - 24 regs
reg_array_2d(X,Cregs,4,3); // 4 spinor - 12 regs
reg_array_1d(F,Cregs,3); // Force
alreg(Z,Cregs); // Zero
alreg(creg,Cregs); // Zero
offset_3d(CHIIMM,FourSpinType,2,3,2*nsimd());
offset_3d(PSIIMM,FourSpinType,4,3,2*nsimd());
offset_3d(GIMM ,GaugeType, 3, 3 ,2*nsimd() );
def_off( GAUGE_SITE_IMM, FourSpinType,4*18*nsimd());
def_off( MAT_IMM , GaugeType,18*nsimd());
def_off( PSI_IMM , FourSpinType,24*nsimd());
def_off( CHI_IMM , FourSpinType,12*nsimd());
def_off( CONST_ZERO_OFFSET,Double,2*2*nsimd());
/*
* Integer registers
*/
alreg(F_p,Iregs); /*Pointer to the current cpt of force field */
alreg(F_p_s,Iregs);
alreg(Y_mu,Iregs);
alreg(Y_p,Iregs);
alreg(X_p,Iregs);
alreg(length,Iregs); /*number of sites*/
alreg(tab,Iregs); /*Pointer to current entry in offset table*/
alreg(Complex_i,Iregs);/*Point to (0,1)x Nsimd*/
alreg(Ls,Iregs);
alreg(s,Iregs);
alreg(recbuf_base,Iregs);
alreg(args,Iregs);
alreg(s_offset,Iregs);
/*Useful integer immediate constants, in units of Fsize*/
def_off( ZERO_IMM,Byte,0);
def_off( minusone,Byte,-1);
def_off( one,Byte,1);
// Mask bits for predicating directions
def_off( mask_0,Byte,1);
def_off( mask_1,Byte,2);
def_off( mask_2,Byte,4);
def_off( mask_3,Byte,8);
def_off( mask_4,Byte,16);
def_off( mask_5,Byte,32);
def_off( mask_6,Byte,64);
def_off( mask_7,Byte,128);
int mask_imm[8] = {
mask_0,
mask_1,
mask_2,
mask_3,
mask_4,
mask_5,
mask_6,
mask_7
};
alreg(mask ,Iregs);
offset_1d(TAB_IMM,TableType,17);
// Integer sizes
int Isize = def_offset(PROC->I_size,Byte,"Isize");
int ISsize = def_offset(PROC->IS_size,Byte,"ISsize");
int i,j,co,sp;
/*********************************************************************/
make_inst(DIRECTIVE,Enter_Routine,name);
grab_stack(0);
save_regs();
/*********************************************
* our arguments
*********************************************
*/
getarg(args); /*Pointer to arg list*/
queue_iload(X_p, ZERO_IMM,args); queue_load_addr(args,Isize,args); //0
queue_iload(Y_p, ZERO_IMM,args); queue_load_addr(args,Isize,args); //1
queue_iload(F_p, ZERO_IMM,args); queue_load_addr(args,Isize,args); //2
queue_iload(length,ZERO_IMM,args); queue_load_addr(args,Isize,args); //3
queue_iload(Ls, ZERO_IMM,args); queue_load_addr(args,Isize,args); //4
queue_iload(tab, ZERO_IMM,args); queue_load_addr(args,Isize,args); //5
queue_iload(Complex_i,ZERO_IMM,args);queue_load_addr(args,Isize,args); //6
queue_load_addr(args,Isize,args); //7
queue_iload(recbuf_base,ZERO_IMM,args);queue_load_addr(args,Isize,args); //8
/**************************************************
* Load common constants into Iregs
**************************************************
*/
for (int i =0; i<12; i++ ) {
need_constant(i*2*SizeofDatum(FourSpinType)*nsimd());
}
for (int i =0; i<9; i++ ) {
need_constant(i*2*SizeofDatum(GaugeType)*nsimd());
}
complex_constants_prepare(creg,Complex_i);
complex_load(Z,CONST_ZERO_OFFSET,Complex_i,Double);
// Site loop
retno = get_target_label();
check_iterations(length,retno);
int branchsite = start_loop(length);
// S loop
queue_iload_short(mask,TAB_IMM[10],tab);
queue_iadd_imm (s,Ls,ZERO_IMM);
queue_iload_imm(s_offset,ZERO_IMM);
int branchls = start_loop(s);
queue_iadd_imm(F_p_s,F_p,ZERO_IMM);
// debugI(s);
// Loop over directions
for ( int mu=0;mu<4;mu++ ) {
int dir = mu*2+1; // Always in forward dir
// Complex branch structure for interior/exterior neighbours
int lab_proj_mu = get_target_label();
int lab_continue = get_target_label();
queue_iand_imm (Y_mu,mask,mask_imm[dir]); // non-zero if exterior
check_iterations(Y_mu,lab_proj_mu);
// Exterior points are already projected. Just load.
queue_iload_short(Y_mu,TAB_IMM[dir],tab);
queue_iadd (Y_mu,Y_mu,recbuf_base);
// debugI(Y_mu);
//debugI(recbuf_base);
queue_iadd (Y_mu,Y_mu,s_offset);
for(int sp=0;sp<2;sp++){
for(int co=0;co<3;co++){
complex_load(Y[sp][co],PSIIMM[sp][co][0],Y_mu,FourSpinType);
}
}
jump(lab_continue);
make_inst(DIRECTIVE,Target,lab_proj_mu);
// Interior points are not already projected.
// * Spin project 4 spinor
queue_iload_short(Y_mu,TAB_IMM[dir],tab);
// debugI(tab);
// debugI(Y_mu);
queue_iadd (Y_mu,Y_mu,Y_p);
queue_iadd (Y_mu,Y_mu,s_offset); // offset for this "s"
queue_iadd (Y_mu,Y_mu,s_offset); // offset for this "s"
for(int sp=0;sp<4;sp++){
for(int co=0;co<3;co++){
complex_load(X[sp][co],PSIIMM[sp][co][0],Y_mu,FourSpinType);
// debugC(X[sp][co]);
}
}
int pm = 1; // pm=0 == 1+gamma, pm=1 => 1-gamma
if ( dagger ) pm = 0;
if ( mu == 0 ) {
if ( pm ==0 ) {
for(co=0;co<3;co++) complex_ApiB(Y[0][co],X[0][co],X[3][co]);
for(co=0;co<3;co++) complex_ApiB(Y[1][co],X[1][co],X[2][co]);
} else {
for(co=0;co<3;co++) complex_AmiB(Y[0][co],X[0][co],X[3][co]);
for(co=0;co<3;co++) complex_AmiB(Y[1][co],X[1][co],X[2][co]);
}
} else if ( mu == 1 ) {
if ( pm ==0 ) {
for(co=0;co<3;co++) complex_sub(Y[0][co],X[0][co],X[3][co]);
for(co=0;co<3;co++) complex_add(Y[1][co],X[1][co],X[2][co]);
} else {
for(co=0;co<3;co++) complex_add(Y[0][co],X[0][co],X[3][co]);
for(co=0;co<3;co++) complex_sub(Y[1][co],X[1][co],X[2][co]);
}
} else if ( mu == 2 ) {
if ( pm ==0 ) {
for(co=0;co<3;co++) complex_ApiB(Y[0][co],X[0][co],X[2][co]);
for(co=0;co<3;co++) complex_AmiB(Y[1][co],X[1][co],X[3][co]);
} else {
for(co=0;co<3;co++) complex_AmiB(Y[0][co],X[0][co],X[2][co]);
for(co=0;co<3;co++) complex_ApiB(Y[1][co],X[1][co],X[3][co]);
}
} else if ( mu == 3 ) {
if ( pm ==0 ) {
for(co=0;co<3;co++) complex_add(Y[0][co],X[0][co],X[2][co]);
for(co=0;co<3;co++) complex_add(Y[1][co],X[1][co],X[3][co]);
} else {
for(co=0;co<3;co++) complex_sub(Y[0][co],X[0][co],X[2][co]);
for(co=0;co<3;co++) complex_sub(Y[1][co],X[1][co],X[3][co]);
}
}
make_inst(DIRECTIVE,Target,lab_continue);
///////////////////////////////////////////////////////////////
// Y contains spin projection of forward neighbour in mu direction
// Repromote to Y to 4 spinor
///////////////////////////////////////////////////////////////
for(int co_y=0;co_y<3;co_y++){
if ( (mu==0) && (pm==0) ) complex_AmiB(Y[2][co_y],Z,Y[1][co_y]);
if ( (mu==0) && (pm==1) ) complex_ApiB(Y[2][co_y],Z,Y[1][co_y]);
if ( (mu==1) && (pm==0) ) complex_add (Y[2][co_y],Z,Y[1][co_y]);
if ( (mu==1) && (pm==1) ) complex_sub (Y[2][co_y],Z,Y[1][co_y]);
if ( (mu==2) && (pm==0) ) complex_AmiB(Y[2][co_y],Z,Y[0][co_y]);
if ( (mu==2) && (pm==1) ) complex_ApiB(Y[2][co_y],Z,Y[0][co_y]);
if ( (mu==3) && (pm==0) ) complex_add (Y[2][co_y],Z,Y[0][co_y]);
if ( (mu==3) && (pm==1) ) complex_sub (Y[2][co_y],Z,Y[0][co_y]);
if ( (mu==0) && (pm==0) ) complex_AmiB(Y[3][co_y],Z,Y[0][co_y]);
if ( (mu==0) && (pm==1) ) complex_ApiB(Y[3][co_y],Z,Y[0][co_y]);
if ( (mu==1) && (pm==0) ) complex_sub (Y[3][co_y],Z,Y[0][co_y]);
if ( (mu==1) && (pm==1) ) complex_add (Y[3][co_y],Z,Y[0][co_y]);
if ( (mu==2) && (pm==0) ) complex_ApiB(Y[3][co_y],Z,Y[1][co_y]);
if ( (mu==2) && (pm==1) ) complex_AmiB(Y[3][co_y],Z,Y[1][co_y]);
if ( (mu==3) && (pm==0) ) complex_add (Y[3][co_y],Z,Y[1][co_y]);
if ( (mu==3) && (pm==1) ) complex_sub (Y[3][co_y],Z,Y[1][co_y]);
}
///////////////////////////////////////////////////////////////
// Load X
///////////////////////////////////////////////////////////////
for(int co_x=0;co_x<3;co_x++){
for(int sp=0;sp<4;sp++) {
complex_load(X[sp][co_x],PSIIMM[sp][co_x][0],X_p,FourSpinType);
}
}
///////////////////////////////////////////////////////////////
// Spin trace tensor product
///////////////////////////////////////////////////////////////
for(int co_x=0;co_x<3;co_x++){
// Spin trace outer product
for ( int co_y=0;co_y<3;co_y++) complex_load (F[co_y],GIMM[co_y][co_x][0],F_p_s);
for ( int co_y=0;co_y<3;co_y++) complex_conjmadd(F[co_y],X[0][co_x],Y[0][co_y]);
for ( int co_y=0;co_y<3;co_y++) complex_conjmadd(F[co_y],X[1][co_x],Y[1][co_y]);
for ( int co_y=0;co_y<3;co_y++) complex_conjmadd(F[co_y],X[2][co_x],Y[2][co_y]);
for ( int co_y=0;co_y<3;co_y++) complex_conjmadd(F[co_y],X[3][co_x],Y[3][co_y]);
for ( int co_y=0;co_y<3;co_y++) complex_store(F[co_y],GIMM[co_y][co_x][0],F_p_s);
}
queue_load_addr(F_p_s,MAT_IMM,F_p_s);
}
queue_iadd_imm(X_p,X_p,PSI_IMM);
queue_iadd_imm(s_offset,s_offset,CHI_IMM);
stop_loop(branchls,s);
queue_iadd_imm(F_p,F_p_s,ZERO_IMM);
queue_load_addr(tab,TAB_IMM[16],tab);
stop_loop(branchsite,length);
make_inst(DIRECTIVE,Target,retno);
/*
*
* EPILOGUE
*
*/
restore_regs();
free_stack();
make_inst(DIRECTIVE,Exit_Routine,name);
return;
}
void qcdoc_merge( char *name)
{
int dum = defargcount(5);
/*Integer register usage*/
alreg(outptr,Iregs);
alreg(vec1ptr,Iregs);
alreg(vec2ptr,Iregs);
alreg(counter,Iregs);
/*Floating register usage*/
reg_array_1d(vec1,Cregs,3);
reg_array_1d(vec2,Cregs,3);
reg_array_1d(oreg,Cregs,6);
alreg(permreg,Cregs);
def_off(ZERO,SpinorType,0);
def_off (IN_ATOM,SpinorType,6*nsimd()); // 2spins worth, 3 colors x complex
def_off (OUT_ATOM,SpinorType,12*nsimd());// 2spins worth, 3 colors x complex x simd
def_off(bits16,Byte,0xFFFF);
def_off(thirtytwo,Byte,32);
def_off(sixteen,Byte,16);
offset_2d(CHI_IMM,SpinorType,6,2*nsimd());
int Isize = PROC->I_size;
int word_size = def_offset(Isize,Byte,"word_size");
struct stream *PreOut;
struct stream *PreVec1;
struct stream *PreVec2;
int brchno,retno; /*Branch target handles*/
int co;
make_inst(DIRECTIVE,Enter_Routine,name);
int bias = grab_stack(64);
save_regs();
queue_iadd_imm(PROC->StackPointer,PROC->StackPointer,bias);
getarg(outptr); /*Get args*/
getarg(vec1ptr);
getarg(vec2ptr);
getarg(counter);
alreg(Mask,Iregs);
alreg(Convert1,Iregs);
alreg(Convert2,Iregs);
int memory = PROC->StackPointer;
for (int i =0; i<6; i++ ) {
need_constant(i*2*SizeofDatum(SpinorType)*nsimd());
}
need_constant(64);
complex_simd_init(permreg);
if ( half_precision ) {
queue_iload_imm(Mask,ZERO);
queue_ori(Mask,Mask,bits16);
queue_lshift(Mask,Mask,thirtytwo);
queue_ori(Mask,Mask,bits16);
queue_lshift(Mask,Mask,sixteen);
}
/*
* Insert a label to prevent reordering
*/
make_inst(DIRECTIVE,Target,get_target_label());
PreVec1= create_stream(IN_ATOM,vec1ptr ,counter,STREAM_IN ,LINEAR);
PreVec2= create_stream(IN_ATOM,vec2ptr ,counter,STREAM_IN ,LINEAR);
PreOut = create_stream(OUT_ATOM,outptr ,counter,STREAM_OUT ,LINEAR);
/*Branch to stack restore if length <1*/
retno = get_target_label();
check_iterations(counter,retno);
/*
* Start software pipeline
*/
brchno = start_loop(counter);
int indco[3]={0,1,2};
int permute_mu=3;
for(int ico=0;ico<3;ico++){
co = indco[ico];
// Could do entirely in integer unit for half precision to accelerate this
if ( half_precision ) {
complex_load_half(vec1[co],CHI_IMM[co][0],vec1ptr,memory,Convert1,Convert2,Mask);
complex_load_half(vec2[co],CHI_IMM[co][0],vec2ptr,memory,Convert1,Convert2,Mask);
} else {
complex_load(vec1[co],CHI_IMM[co][0],vec1ptr,SpinorType);
complex_load(vec2[co],CHI_IMM[co][0],vec2ptr,SpinorType);
}
}
{
// Merge the vectors
for(co=0;co<3;co++) complex_simd_merge (0,permute_mu,oreg[co*2] ,vec1[co],vec2[co]);
for(co=0;co<3;co++) complex_simd_merge (1,permute_mu,oreg[co*2+1],vec1[co],vec2[co]);
}
// make_inst(DIRECTIVE,LS_BARRIER);
for(int i=0;i<6;i++){ // 2 SIMD sites, 3 colors, 2 spins 2 complex == 24 floats
if ( half_precision ) {
complex_store_half(oreg[i],CHI_IMM[i][0],outptr,memory,Convert1,Convert2,Mask);
} else {
complex_store(oreg[i],CHI_IMM[i][0],outptr,SpinorType);
}
}
iterate_stream(PreVec1);
iterate_stream(PreVec2);
do_prefetch(vec1ptr,0);
do_prefetch(vec2ptr,0);
do_prefetch(vec1ptr,1);
do_prefetch(vec2ptr,1);
iterate_stream(PreOut);
stop_loop(brchno,counter);
make_inst(DIRECTIVE,Target,retno);
queue_isub_imm(PROC->StackPointer,PROC->StackPointer,bias);
restore_regs();
free_stack();
make_inst(DIRECTIVE,Exit_Routine,name);
return;
}
void clov_apply( char *name)
{
/*
* This marks the argument registers as defined by ABI as off limits
* to us until they are freed by "getarg()";
*/
int dum = defargcount(1);
int retno;
int branchsite;
/*-------------------------------------------------------------------------------
* registers used
*-------------------------------------------------------------------------------
*/
reg_array_1d(Chi ,Cregs,6); // 2 spionr - 6 regs
reg_array_1d(Psi ,Cregs,6); // 2 spinors - 6 regs
reg_array_1d(Clo ,Cregs,15);
offset_2d(PSIIMM,FourSpinType,6,2*nsimd());
offset_2d(CLOIMM,CloverType,21,2*nsimd());
/*
* Integer registers
*/
alreg(Clo_p,Iregs); /*Pointer to the current cpt of gauge field */
alreg(Chi_p,Iregs); /*Pointer to the input four spinor */
alreg(Psi_p,Iregs); /*Pointer to current cpt output PSI field */
alreg(length,Iregs); /*number of sites*/
//alreg(tab,Iregs); /*Pointer to current entry in offset table*/
alreg(args,Iregs);
/*Useful integer immediate constants, in units of Fsize*/
def_off(ZERO_IMM,Byte,0);
def_off(SPINOR, FourSpinType, 12*nsimd());
def_off(CLOVER, CloverType, 42*nsimd());
int Isize = def_offset(PROC->I_size,Byte,"Isize");
/*--------------------------------------------------------------------
* Start of the "pseudo assembler proper.
*--------------------------------------------------------------------
*/
make_inst(DIRECTIVE,Enter_Routine,name);
grab_stack(0);
save_regs();
/* * Define our arguments */
getarg(args); /*Pointer to arg list*/
queue_iload(Chi_p, ZERO_IMM,args);
queue_load_addr(args,Isize,args);
queue_iload(Psi_p, ZERO_IMM,args);
queue_load_addr(args,Isize,args);
queue_iload(Clo_p, ZERO_IMM,args);
queue_load_addr(args,Isize,args);
queue_iload(length,ZERO_IMM,args);
for (int i =0; i<6; i++ )
{
need_constant(i*2*SizeofDatum(FourSpinType)*nsimd());
}
for (int i =0; i<21; i++ )
{
need_constant(i*2*SizeofDatum(CloverType)*nsimd());
}
retno = get_target_label(); /*Branch to exit if length <1*/
check_iterations(length,retno);
/* * Site loop */
branchsite = start_loop(length);
for(int pp=0; pp<6; pp++)
{
complex_load(Psi[pp],PSIIMM[pp][0],Psi_p,FourSpinType);
}
for(int qq=0; qq<6; qq++)
{
complex_load(Clo[qq],CLOIMM[qq][0],Clo_p,CloverType);
}
for(int rr=0; rr<6; rr++)
{
complex_mul (Chi[rr], Clo[rr], Psi[rr]);
}
for(int ss=0; ss<15; ss++)
{
complex_load(Clo[ss],CLOIMM[ss+6][0],Clo_p,CloverType);
}
complex_conjmadd(Chi[0], Clo[0], Psi[1]);
complex_conjmadd(Chi[0], Clo[1], Psi[2]);
complex_conjmadd(Chi[0], Clo[3], Psi[3]);
complex_conjmadd(Chi[0], Clo[6], Psi[4]);
complex_conjmadd(Chi[0], Clo[10], Psi[5]);
complex_madd (Chi[1], Clo[0], Psi[0]);
complex_conjmadd(Chi[1], Clo[2], Psi[2]);
complex_conjmadd(Chi[1], Clo[4], Psi[3]);
complex_conjmadd(Chi[1], Clo[7], Psi[4]);
complex_conjmadd(Chi[1], Clo[11], Psi[5]);
complex_madd (Chi[2], Clo[1], Psi[0]);
complex_madd (Chi[2], Clo[2], Psi[1]);
complex_conjmadd(Chi[2], Clo[5], Psi[3]);
complex_conjmadd(Chi[2], Clo[8], Psi[4]);
complex_conjmadd(Chi[2], Clo[12], Psi[5]);
complex_madd (Chi[3], Clo[3], Psi[0]);
complex_madd (Chi[3], Clo[4], Psi[1]);
complex_madd (Chi[3], Clo[5], Psi[2]);
complex_conjmadd(Chi[3], Clo[9], Psi[4]);
complex_conjmadd(Chi[3], Clo[13], Psi[5]);
complex_madd (Chi[4], Clo[6], Psi[0]);
complex_madd (Chi[4], Clo[7], Psi[1]);
complex_madd (Chi[4], Clo[8], Psi[2]);
complex_madd (Chi[4], Clo[9], Psi[3]);
complex_conjmadd(Chi[4], Clo[14], Psi[5]);
complex_madd (Chi[5], Clo[10], Psi[0]);
complex_madd (Chi[5], Clo[11], Psi[1]);
complex_madd (Chi[5], Clo[12], Psi[2]);
complex_madd (Chi[5], Clo[13], Psi[3]);
complex_madd (Chi[5], Clo[14], Psi[4]);
for(int sp=0; sp<6; sp++)
{
complex_store(Chi[sp],PSIIMM[sp][0],Chi_p,FourSpinType);
}
queue_iadd_imm(Psi_p,Psi_p,SPINOR);
queue_iadd_imm(Chi_p,Chi_p,SPINOR);
queue_iadd_imm(Clo_p,Clo_p,CLOVER);
/* TERMINATION point of the loop*/
stop_loop(branchsite,length);
make_inst(DIRECTIVE,Target,retno);
/* EPILOGUE */
restore_regs();
free_stack();
make_inst(DIRECTIVE,Exit_Routine,name);
return;
}