static inline int16x8_t qvsource_over_s16(int16x8_t src16, int16x8_t dst16, int16x8_t half, int16x8_t full)
{
const int16x4_t alpha16_high = vdup_lane_s16(vget_high_s16(src16), 3);
const int16x4_t alpha16_low = vdup_lane_s16(vget_low_s16(src16), 3);
const int16x8_t alpha16 = vsubq_s16(full, vcombine_s16(alpha16_low, alpha16_high));
return vaddq_s16(src16, qvbyte_mul_s16(dst16, alpha16, half));
}
static INLINE void scaledconvolve_vert_w4(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
const int y0_q4, const int y_step_q4, const int w, const int h) {
int y;
int y_q4 = y0_q4;
src -= src_stride * (SUBPEL_TAPS / 2 - 1);
y = h;
do {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
if (y_q4 & SUBPEL_MASK) {
const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
uint8x8_t s[8], d;
int16x4_t t[8], tt;
load_u8_8x8(src_y, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
&s[6], &s[7]);
t[0] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[0])));
t[1] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[1])));
t[2] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[2])));
t[3] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[3])));
t[4] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[4])));
t[5] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[5])));
t[6] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[6])));
t[7] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[7])));
tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7], filters,
filter3, filter4);
d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
} else {
memcpy(dst, &src_y[3 * src_stride], w);
}
dst += dst_stride;
y_q4 += y_step_q4;
} while (--y);
}
int16x4_t test_vdup_lane_s16(int16x4_t v1) {
// CHECK: test_vdup_lane_s16
return vdup_lane_s16(v1, 2);
// CHECK: dup {{v[0-9]+}}.4h, {{v[0-9]+}}.h[2]
}
static INLINE void scaledconvolve_horiz_w4(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const x_filters,
const int x0_q4, const int x_step_q4, const int w, const int h) {
DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
int x, y, z;
src -= SUBPEL_TAPS / 2 - 1;
y = h;
do {
int x_q4 = x0_q4;
x = 0;
do {
// process 4 src_x steps
for (z = 0; z < 4; ++z) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
if (x_q4 & SUBPEL_MASK) {
const int16x8_t filters = vld1q_s16(x_filters[x_q4 & SUBPEL_MASK]);
const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
uint8x8_t s[8], d;
int16x8_t ss[4];
int16x4_t t[8], tt;
load_u8_8x4(src_x, src_stride, &s[0], &s[1], &s[2], &s[3]);
transpose_u8_8x4(&s[0], &s[1], &s[2], &s[3]);
ss[0] = vreinterpretq_s16_u16(vmovl_u8(s[0]));
ss[1] = vreinterpretq_s16_u16(vmovl_u8(s[1]));
ss[2] = vreinterpretq_s16_u16(vmovl_u8(s[2]));
ss[3] = vreinterpretq_s16_u16(vmovl_u8(s[3]));
t[0] = vget_low_s16(ss[0]);
t[1] = vget_low_s16(ss[1]);
t[2] = vget_low_s16(ss[2]);
t[3] = vget_low_s16(ss[3]);
t[4] = vget_high_s16(ss[0]);
t[5] = vget_high_s16(ss[1]);
t[6] = vget_high_s16(ss[2]);
t[7] = vget_high_s16(ss[3]);
tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7],
filters, filter3, filter4);
d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
vst1_lane_u32((uint32_t *)&temp[4 * z], vreinterpret_u32_u8(d), 0);
} else {
int i;
for (i = 0; i < 4; ++i) {
temp[z * 4 + i] = src_x[i * src_stride + 3];
}
}
x_q4 += x_step_q4;
}
// transpose the 4x4 filters values back to dst
{
const uint8x8x4_t d4 = vld4_u8(temp);
vst1_lane_u32((uint32_t *)&dst[x + 0 * dst_stride],
vreinterpret_u32_u8(d4.val[0]), 0);
vst1_lane_u32((uint32_t *)&dst[x + 1 * dst_stride],
vreinterpret_u32_u8(d4.val[1]), 0);
vst1_lane_u32((uint32_t *)&dst[x + 2 * dst_stride],
vreinterpret_u32_u8(d4.val[2]), 0);
vst1_lane_u32((uint32_t *)&dst[x + 3 * dst_stride],
vreinterpret_u32_u8(d4.val[3]), 0);
}
x += 4;
} while (x < w);
src += src_stride * 4;
dst += dst_stride * 4;
y -= 4;
} while (y > 0);
}
wrap_vdup_lane_s16_1 (int16x4_t a)
{
return vdup_lane_s16 (a, 1);
}
wrap_vdup_lane_s16_0 (int16x4_t a)
{
return vdup_lane_s16 (a, 0);
}
