// 'do_above' and 'do_left' facilitate branch removal when inlined.
static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left,
int do_above, int do_left) {
uint16x8_t sum_top;
uint16x8_t sum_left;
uint8x8_t dc0;
if (do_above) {
const uint8x16_t A0 = vld1q_u8(above); // top row
const uint8x16_t A1 = vld1q_u8(above + 16);
const uint16x8_t p0 = vpaddlq_u8(A0); // cascading summation of the top
const uint16x8_t p1 = vpaddlq_u8(A1);
const uint16x8_t p2 = vaddq_u16(p0, p1);
const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
const uint16x4_t p4 = vpadd_u16(p3, p3);
const uint16x4_t p5 = vpadd_u16(p4, p4);
sum_top = vcombine_u16(p5, p5);
}
if (do_left) {
const uint8x16_t L0 = vld1q_u8(left); // left row
const uint8x16_t L1 = vld1q_u8(left + 16);
const uint16x8_t p0 = vpaddlq_u8(L0); // cascading summation of the left
const uint16x8_t p1 = vpaddlq_u8(L1);
const uint16x8_t p2 = vaddq_u16(p0, p1);
const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
const uint16x4_t p4 = vpadd_u16(p3, p3);
const uint16x4_t p5 = vpadd_u16(p4, p4);
sum_left = vcombine_u16(p5, p5);
}
if (do_above && do_left) {
const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
dc0 = vrshrn_n_u16(sum, 6);
} else if (do_above) {
dc0 = vrshrn_n_u16(sum_top, 5);
} else if (do_left) {
dc0 = vrshrn_n_u16(sum_left, 5);
} else {
dc0 = vdup_n_u8(0x80);
}
{
const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
int i;
for (i = 0; i < 32; ++i) {
vst1q_u8(dst + i * stride, dc);
vst1q_u8(dst + i * stride + 16, dc);
}
}
}
void vp9_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint8x8_t d2u8 = vdup_n_u8(0);
uint8x16_t q0u8 = vdupq_n_u8(0);
uint8x16_t q1u8 = vdupq_n_u8(0);
(void)above;
for (k = 0; k < 2; k++, left += 16) {
q1u8 = vld1q_u8(left);
d2u8 = vget_low_u8(q1u8);
for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
q0u8 = vdupq_lane_u8(d2u8, 0);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 1);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 2);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 3);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 4);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 5);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 6);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
q0u8 = vdupq_lane_u8(d2u8, 7);
vst1q_u8(dst, q0u8);
vst1q_u8(dst + 16, q0u8);
dst += stride;
}
}
}
// 'do_above' and 'do_left' facilitate branch removal when inlined.
static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left,
int do_above, int do_left) {
uint16x8_t sum_top;
uint16x8_t sum_left;
uint8x8_t dc0;
if (do_above) {
const uint8x16_t A = vld1q_u8(above); // top row
const uint16x8_t p0 = vpaddlq_u8(A); // cascading summation of the top
const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
const uint16x4_t p2 = vpadd_u16(p1, p1);
const uint16x4_t p3 = vpadd_u16(p2, p2);
sum_top = vcombine_u16(p3, p3);
}
if (do_left) {
const uint8x16_t L = vld1q_u8(left); // left row
const uint16x8_t p0 = vpaddlq_u8(L); // cascading summation of the left
const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
const uint16x4_t p2 = vpadd_u16(p1, p1);
const uint16x4_t p3 = vpadd_u16(p2, p2);
sum_left = vcombine_u16(p3, p3);
}
if (do_above && do_left) {
const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
dc0 = vrshrn_n_u16(sum, 5);
} else if (do_above) {
dc0 = vrshrn_n_u16(sum_top, 4);
} else if (do_left) {
dc0 = vrshrn_n_u16(sum_left, 4);
} else {
dc0 = vdup_n_u8(0x80);
}
{
const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
int i;
for (i = 0; i < 16; ++i) {
vst1q_u8(dst + i * stride, dc);
}
}
}
uint8x16_t test_vdupq_lane_u8(uint8x8_t v1) {
// CHECK: test_vdupq_lane_u8
return vdupq_lane_u8(v1, 5);
// CHECK: dup {{v[0-9]+}}.16b, {{v[0-9]+}}.b[5]
}
