void test_vst2Qs16 (void)
{
int16_t *arg0_int16_t;
int16x8x2_t arg1_int16x8x2_t;
vst2q_s16 (arg0_int16_t, arg1_int16x8x2_t);
}
void vp8_dequantize_b_neon(BLOCKD *d, short *DQC) {
int16x8x2_t qQ, qDQC, qDQ;
qQ = vld2q_s16(d->qcoeff);
qDQC = vld2q_s16(DQC);
qDQ.val[0] = vmulq_s16(qQ.val[0], qDQC.val[0]);
qDQ.val[1] = vmulq_s16(qQ.val[1], qDQC.val[1]);
vst2q_s16(d->dqcoeff, qDQ);
}
void vp8_dequantize_b_loop_neon(
int16_t *Q,
int16_t *DQC,
int16_t *DQ) {
int16x8x2_t qQ, qDQC, qDQ;
qQ = vld2q_s16(Q);
qDQC = vld2q_s16(DQC);
qDQ.val[0] = vmulq_s16(qQ.val[0], qDQC.val[0]);
qDQ.val[1] = vmulq_s16(qQ.val[1], qDQC.val[1]);
vst2q_s16(DQ, qDQ);
return;
}
rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband_width)
{
int y, n;
INT16 * l_ptr = l;
INT16 * h_ptr = h;
INT16 * dst_ptr = dst;
for (y = 0; y < subband_width; y++)
{
/* Even coefficients */
for (n = 0; n < subband_width; n+=8)
{
// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
int16x8_t l_n = vld1q_s16(l_ptr);
int16x8_t h_n = vld1q_s16(h_ptr);
int16x8_t h_n_m = vld1q_s16(h_ptr - 1);
if (n == 0)
{
int16_t first = vgetq_lane_s16(h_n_m, 1);
h_n_m = vsetq_lane_s16(first, h_n_m, 0);
}
int16x8_t tmp_n = vaddq_s16(h_n, h_n_m);
tmp_n = vaddq_s16(tmp_n, vdupq_n_s16(1));
tmp_n = vshrq_n_s16(tmp_n, 1);
int16x8_t dst_n = vsubq_s16(l_n, tmp_n);
vst1q_s16(l_ptr, dst_n);
l_ptr+=8;
h_ptr+=8;
}
l_ptr -= subband_width;
h_ptr -= subband_width;
/* Odd coefficients */
for (n = 0; n < subband_width; n+=8)
{
// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
int16x8_t h_n = vld1q_s16(h_ptr);
h_n = vshlq_n_s16(h_n, 1);
int16x8x2_t dst_n;
dst_n.val[0] = vld1q_s16(l_ptr);
int16x8_t dst_n_p = vld1q_s16(l_ptr + 1);
if (n == subband_width - 8)
{
int16_t last = vgetq_lane_s16(dst_n_p, 6);
dst_n_p = vsetq_lane_s16(last, dst_n_p, 7);
}
dst_n.val[1] = vaddq_s16(dst_n_p, dst_n.val[0]);
dst_n.val[1] = vshrq_n_s16(dst_n.val[1], 1);
dst_n.val[1] = vaddq_s16(dst_n.val[1], h_n);
vst2q_s16(dst_ptr, dst_n);
l_ptr+=8;
h_ptr+=8;
dst_ptr+=16;
}
}
}
inline void vst2q(s16 * ptr, const int16x8x2_t & v) { return vst2q_s16(ptr, v); }