double ddot(const int N, const double *a, const int incx, const double *b, const int incy)
{
int i;
vtype q00 = set_vector(0.);
vtype q01 = set_vector(0.);
vtype q0a, q1a;
vtype q0b, q1b;
//
double c;
//
for (i = 0; i < N - N%4; i = i + 4)
{
q0a = LOAD(a + i);
q0b = LOAD(b + i);
q00 = vfmaq_f64(q00, q0a, q0b);
//q0a = vmulq_f64(q0a, q0b);
//q00 = vaddq_f64(q0a, q00);
//
q0a = LOAD(a + i + 2);
q1b = LOAD(b + i + 2);
q01 = vfmaq_f64(q01, q0a, q0b);
//q1a = vmulq_f64(q1a, q1b);
//q01 = vaddq_f64(q1a, q01);
//c += a [i]*b [i];
}
c = q00[0] + q00[1] + q01[0] + q01[1];
return c;
}
float64x2_t test_vfmaq_f64(float64x2_t a1, float64x2_t a2, float64x2_t a3) {
// CHECK: test_vfmaq_f64
return vfmaq_f64(a1, a2, a3);
// CHECK: llvm.fma.v2f64({{.*a2, .*a3, .*a1}})
// CHECK-NEXT: ret
}
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
BLASLONG i=0;
BLASLONG ix=0,iy=0;
FLOAT dot = 0.0 ;
if ( n < 0 ) return(dot);
if ( (inc_x == 1) && (inc_y == 1) )
{
float64x2_t vdot0 = {0.0, 0.0};
float64x2_t vdot1 = {0.0, 0.0};
float64x2_t vdot2 = {0.0, 0.0};
float64x2_t vdot3 = {0.0, 0.0};
float64x2_t *vx = (float64x2_t*)x;
float64x2_t *vy = (float64x2_t*)y;
#if 0
prefetch(x + 128/sizeof(*x));
prefetch(y + 128/sizeof(*y));
#endif
prefetch(x + 2*128/sizeof(*x));
prefetch(y + 2*128/sizeof(*y));
prefetch(x + 3*128/sizeof(*x));
prefetch(y + 3*128/sizeof(*y));
int n1 = n&-8;
while(i < n1)
{
#if 0
vdot0 = vfmaq_f64 (vdot0,
vy[0],
vx[0]);
vdot1 = vfmaq_f64 (vdot1,
vy[1],
vx[1]);
vdot2 = vfmaq_f64 (vdot2,
vy[2],
vx[2]);
vdot3 = vfmaq_f64 (vdot3,
vy[3],
vx[3]);
#else
vdot0 = vy[0] * vx[0] + vdot0;
vdot1 = vy[1] * vx[1] + vdot1;
vdot2 = vy[2] * vx[2] + vdot2;
vdot3 = vy[3] * vx[3] + vdot3;
#endif
vy += 4;
vx += 4;
i += 8;
prefetch(vx + 3*128/sizeof(*x));
prefetch(vy + 3*128/sizeof(*y));
}
dot = vaddvq_f64 (vdot0 + vdot1);
dot += vaddvq_f64 (vdot2 + vdot3);
i = n1;
while(i < n)
{
dot += y[i] * x[i] ;
i++ ;
}
return(dot);
}
while(i < n)
{
dot += y[iy] * x[ix] ;
ix += inc_x ;
iy += inc_y ;
i++ ;
}
return(dot);
}
