PetscErrorCode MatMultTranspose_SeqBSTRM_5(Mat A,Vec xx,Vec zz)
{
Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
Mat_SeqBSTRM *sbstrm = (Mat_SeqBSTRM*)A->spptr;
PetscScalar zero = 0.0;
PetscScalar *z = 0;
PetscScalar *x,*xb;
const MatScalar *v1, *v2, *v3, *v4, *v5;
PetscScalar x1, x2, x3, x4, x5;
PetscErrorCode ierr;
PetscInt mbs =a->mbs,i,*aj=a->j,*ai=a->i,n,*ib,cval,j;
PetscInt nonzerorow=0;
PetscInt slen;
PetscFunctionBegin;
ierr = VecSet(zz,zero);CHKERRQ(ierr);
ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
ierr = VecGetArray(zz,&z);CHKERRQ(ierr);
slen = 5*(ai[mbs]-ai[0]);
v1 = sbstrm->as;
v2 = v1 + slen;
v3 = v2 + slen;
v4 = v3 + slen;
v5 = v4 + slen;
xb = x;
for (i=0; i<mbs; i++) {
n = ai[i+1] - ai[i];
nonzerorow += (n>0);
x1 = xb[0]; x2 = xb[1]; x3 = xb[2]; x4 = xb[3]; x5 = xb[4]; xb += 5;
ib = aj + ai[i];
PetscPrefetchBlock(ib+n,n,0,PETSC_PREFETCH_HINT_NTA); /* Indices for the next row (assumes same size as this one) */
PetscPrefetchBlock(v1+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v2+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v3+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v4+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v5+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
for (j=0; j<n; j++) {
cval = ib[j]*5;
z[cval] += v1[0]*x1 + v2[0]*x2 + v3[0]*x3 + v4[0]*x4 + v5[0]*x5;
z[cval+1] += v1[1]*x1 + v2[1]*x2 + v3[1]*x3 + v4[1]*x4 + v5[1]*x5;
z[cval+2] += v1[2]*x1 + v2[2]*x2 + v3[2]*x3 + v4[2]*x4 + v5[2]*x5;
z[cval+3] += v1[3]*x1 + v2[3]*x2 + v3[3]*x3 + v4[3]*x4 + v5[3]*x5;
z[cval+4] += v1[4]*x1 + v2[4]*x2 + v3[4]*x3 + v4[4]*x4 + v5[4]*x5;
v1 += 5; v2 += 5; v3 += 5; v4 += 5; v5 += 5;
}
}
ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
ierr = VecRestoreArray(zz,&z);CHKERRQ(ierr);
ierr = PetscLogFlops(50.0*a->nz - 5.0*nonzerorow);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatMult_SeqSBSTRM_5(Mat A,Vec xx,Vec zz)
{
Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data;
Mat_SeqSBSTRM *sbstrm = (Mat_SeqSBSTRM*)A->spptr;
PetscScalar zero = 0.0;
PetscScalar *z = 0;
PetscScalar *x,*xb;
const MatScalar *v1, *v2, *v3, *v4, *v5;
PetscScalar x1, x2, x3, x4, x5;
PetscScalar xr1, xr2, xr3, xr4, xr5;
PetscScalar sum1, sum2, sum3, sum4, sum5;
PetscErrorCode ierr;
PetscInt mbs =a->mbs,i,*aj=a->j,*ai=a->i,n,*ib,cval,j,jmin;
PetscInt nonzerorow=0;
PetscInt slen;
PetscFunctionBegin;
ierr = VecSet(zz,zero);CHKERRQ(ierr);
ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
ierr = VecGetArray(zz,&z);CHKERRQ(ierr);
slen = 5*(ai[mbs]-ai[0]);
v1 = sbstrm->as;
v2 = v1 + slen;
v3 = v2 + slen;
v4 = v3 + slen;
v5 = v4 + slen;
xb = x;
for (i=0; i<mbs; i++) {
n = ai[i+1] - ai[i];
nonzerorow += (n>0);
x1 = xb[0]; x2 = xb[1]; x3 = xb[2]; x4 = xb[3]; x5 = xb[4];
sum1 = z[5*i];
sum2 = z[5*i+1];
sum3 = z[5*i+2];
sum4 = z[5*i+3];
sum5 = z[5*i+4];
jmin = 0;
ib = aj + ai[i];
if (*ib == i) { /* (diag of A)*x */
sum1 += v1[0]*x1 + v1[1]*x2 + v1[2]*x3 + v1[3]*x4 + v1[4]*x5;
sum2 += v1[1]*x1 + v2[1]*x2 + v2[2]*x3 + v2[3]*x4 + v2[4]*x5;
sum3 += v1[2]*x1 + v2[2]*x2 + v3[2]*x3 + v3[3]*x4 + v3[4]*x5;
sum4 += v1[3]*x1 + v2[3]*x2 + v3[3]*x3 + v4[3]*x4 + v4[4]*x5;
sum5 += v1[4]*x1 + v2[4]*x2 + v3[4]*x3 + v4[4]*x4 + v5[4]*x5;
v1 += 5; v2 += 5; v3 += 5; v4 += 5; v5 += 5;
jmin++;
}
PetscPrefetchBlock(ib+jmin+n,n,0,PETSC_PREFETCH_HINT_NTA); /* Indices for the next row (assumes same size as this one) */
PetscPrefetchBlock(v1+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v2+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v3+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v4+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v5+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
for (j=jmin; j<n; j++) {
cval = ib[j]*5;
z[cval] += v1[0]*x1 + v2[0]*x2 + v3[0]*x3 + v4[0]*x4 + v5[0]*x5;
z[cval+1] += v1[1]*x1 + v2[1]*x2 + v3[1]*x3 + v4[1]*x4 + v5[1]*x5;
z[cval+2] += v1[2]*x1 + v2[2]*x2 + v3[2]*x3 + v4[2]*x4 + v5[2]*x5;
z[cval+3] += v1[3]*x1 + v2[3]*x2 + v3[3]*x3 + v4[3]*x4 + v5[3]*x5;
z[cval+4] += v1[4]*x1 + v2[4]*x2 + v3[4]*x3 + v4[4]*x4 + v5[4]*x5;
xr1 = x[cval]; xr2 = x[cval+1]; xr3 = x[cval+2]; xr4 = x[cval+3]; xr5 = x[cval+4];
sum1 += v1[0]*xr1 + v1[1]*xr2 + v1[2]*xr3 + v1[3]*xr4 + v1[4]*xr5;
sum2 += v2[0]*xr1 + v2[1]*xr2 + v2[2]*xr3 + v2[3]*xr4 + v2[4]*xr5;
sum3 += v3[0]*xr1 + v3[1]*xr2 + v3[2]*xr3 + v3[3]*xr4 + v3[4]*xr5;
sum4 += v4[0]*xr1 + v4[1]*xr2 + v4[2]*xr3 + v4[3]*xr4 + v4[4]*xr5;
sum5 += v5[0]*xr1 + v5[1]*xr2 + v5[2]*xr3 + v5[3]*xr4 + v5[4]*xr5;
v1 += 5; v2 += 5; v3 += 5; v4 += 5; v5 += 5;
}
z[5*i] = sum1; z[5*i+1] = sum2; z[5*i+2] = sum3; z[5*i+3] = sum4; z[5*i+4] = sum5;
xb += 5;
}
ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
ierr = VecRestoreArray(zz,&z);CHKERRQ(ierr);
ierr = PetscLogFlops(50.0*(a->nz*2.0 - nonzerorow) - nonzerorow);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatMultAdd_SeqBSTRM_5(Mat A,Vec xx,Vec yy,Vec zz)
{
Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
Mat_SeqBSTRM *bstrm = (Mat_SeqBSTRM*)A->spptr;
PetscScalar *x,*y = 0,*z = 0,*xb,sum1,sum2,sum3,sum4,sum5,x1,x2,x3,x4,x5;
PetscScalar *yarray,*zarray;
MatScalar *v1,*v2,*v3,*v4,*v5;
PetscErrorCode ierr;
PetscInt mbs =a->mbs,i,*idx,*ii,j,n,*ridx=NULL;
PetscBool usecprow=a->compressedrow.use;
PetscInt slen;
PetscFunctionBegin;
ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
ierr = VecGetArray(yy,&yarray);CHKERRQ(ierr);
if (zz != yy) {
ierr = VecGetArray(zz,&zarray);CHKERRQ(ierr);
} else {
zarray = yarray;
}
idx = a->j;
if (usecprow) {
if (zz != yy) {
ierr = PetscMemcpy(zarray,yarray,5*mbs*sizeof(PetscScalar));CHKERRQ(ierr);
}
mbs = a->compressedrow.nrows;
ii = a->compressedrow.i;
ridx = a->compressedrow.rindex;
} else {
ii = a->i;
y = yarray;
z = zarray;
}
slen = 5*(ii[mbs]-ii[0]);
v1 = bstrm->as;
v2 = v1 + slen;
v3 = v2 + slen;
v4 = v3 + slen;
v5 = v4 + slen;
for (i=0; i<mbs; i++) {
n = ii[1] - ii[0]; ii++;
if (usecprow) {
z = zarray + 5*ridx[i];
y = yarray + 5*ridx[i];
}
sum1 = y[0]; sum2 = y[1]; sum3 = y[2]; sum4 = y[3]; sum5 = y[4];
PetscPrefetchBlock(idx+n,n,0,PETSC_PREFETCH_HINT_NTA); /* Indices for the next row (assumes same size as this one) */
PetscPrefetchBlock(v1+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v2+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v3+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v4+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
PetscPrefetchBlock(v5+5*n,5*n,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */
for (j=0; j<n; j++) {
xb = x + 5*(*idx++);
x1 = xb[0]; x2 = xb[1]; x3 = xb[2]; x4 = xb[3]; x5 = xb[4];
sum1 += v1[0]*x1 + v1[1]*x2 + v1[2]*x3 + v1[3]*x4 + v1[4]*x5;
sum2 += v2[0]*x1 + v2[1]*x2 + v2[2]*x3 + v2[3]*x4 + v2[4]*x5;
sum3 += v3[0]*x1 + v3[1]*x2 + v3[2]*x3 + v3[3]*x4 + v3[4]*x5;
sum4 += v4[0]*x1 + v4[1]*x2 + v4[2]*x3 + v4[3]*x4 + v4[4]*x5;
sum5 += v5[0]*x1 + v5[1]*x2 + v5[2]*x3 + v5[3]*x4 + v5[4]*x5;
v1 += 5; v2 += 5; v3 += 5; v4 += 5; v5 += 5;
}
z[0] = sum1; z[1] = sum2; z[2] = sum3; z[3] = sum4; z[4] = sum5;
if (!usecprow) {
z += 5; y += 5;
}
}
ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
ierr = VecRestoreArray(yy,&yarray);CHKERRQ(ierr);
if (zz != yy) {
ierr = VecRestoreArray(zz,&zarray);CHKERRQ(ierr);
}
ierr = PetscLogFlops(50.0*a->nz);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
