void test_vreinterpretQu16_s32 (void)
{
uint16x8_t out_uint16x8_t;
int32x4_t arg0_int32x4_t;
out_uint16x8_t = vreinterpretq_u16_s32 (arg0_int32x4_t);
}
void decal_nofilter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count)
{
int i;
if (count >= 8) {
/* SkFixed is 16.16 fixed point */
SkFixed dx2 = dx+dx;
SkFixed dx4 = dx2+dx2;
SkFixed dx8 = dx4+dx4;
/* now build fx/fx+dx/fx+2dx/fx+3dx */
SkFixed fx1, fx2, fx3;
int32x4_t lbase, hbase;
uint16_t *dst16 = (uint16_t *)dst;
fx1 = fx+dx;
fx2 = fx1+dx;
fx3 = fx2+dx;
/* avoid an 'lbase unitialized' warning */
lbase = vdupq_n_s32(fx);
lbase = vsetq_lane_s32(fx1, lbase, 1);
lbase = vsetq_lane_s32(fx2, lbase, 2);
lbase = vsetq_lane_s32(fx3, lbase, 3);
hbase = vaddq_s32(lbase, vdupq_n_s32(dx4));
/* take upper 16 of each, store, and bump everything */
do {
int32x4_t lout, hout;
uint16x8_t hi16;
lout = lbase;
hout = hbase;
/* gets hi's of all louts then hi's of all houts */
asm ("vuzpq.16 %q0, %q1" : "+w" (lout), "+w" (hout));
hi16 = vreinterpretq_u16_s32(hout);
vst1q_u16(dst16, hi16);
/* on to the next */
lbase = vaddq_s32 (lbase, vdupq_n_s32(dx8));
hbase = vaddq_s32 (hbase, vdupq_n_s32(dx8));
dst16 += 8;
count -= 8;
fx += dx8;
} while (count >= 8);
dst = (uint32_t *) dst16;
}
