void byte2float48_neon(const uint8_t *t, const int pitch, float *p) {
uint16x8_t m0, m1, m2, m3, m4, m5;
uint32x2_t temp1, temp4;
m0 = vmovl_u8(vld1_u8(t));
temp1 = vld1_lane_u32((const uint32_t *)(t + 8), temp1, 0);
temp1 = vld1_lane_u32((const uint32_t *)(t + pitch * 2), temp1, 1);
m1 = vmovl_u8(vreinterpret_u8_u32(temp1));
m2 = vmovl_u8(vld1_u8(t + pitch * 2 + 4));
t += pitch * 4;
m3 = vmovl_u8(vld1_u8(t));
temp4 = vld1_lane_u32((const uint32_t *)(t + 8), temp4, 0);
temp4 = vld1_lane_u32((const uint32_t *)(t + pitch * 2), temp4, 1);
m4 = vmovl_u8(vreinterpret_u8_u32(temp4));
m5 = vmovl_u8(vld1_u8(t + pitch * 2 + 4));
vst1q_f32(p, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m0))));
vst1q_f32(p + 4, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m0))));
vst1q_f32(p + 8, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m1))));
vst1q_f32(p + 12, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m1))));
vst1q_f32(p + 16, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m2))));
vst1q_f32(p + 20, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m2))));
vst1q_f32(p + 24, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m3))));
vst1q_f32(p + 28, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m3))));
vst1q_f32(p + 32, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m4))));
vst1q_f32(p + 36, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m4))));
vst1q_f32(p + 40, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m5))));
vst1q_f32(p + 44, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m5))));
}
void vpx_idct4x4_1_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8x8_t d6u8;
uint32x2_t d2u32 = vdup_n_u32(0);
uint16x8_t q8u16;
int16x8_t q0s16;
uint8_t *d1, *d2;
int16_t i, a1, cospi_16_64 = 11585;
int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
out = dct_const_round_shift(out * cospi_16_64);
a1 = ROUND_POWER_OF_TWO(out, 4);
q0s16 = vdupq_n_s16(a1);
// dc_only_idct_add
d1 = d2 = dest;
for (i = 0; i < 2; i++) {
d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
d1 += dest_stride;
d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16),
vreinterpret_u8_u32(d2u32));
d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
d2 += dest_stride;
vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
d2 += dest_stride;
}
return;
}
unsigned int aom_avg_4x4_neon(const uint8_t *s, int p) {
uint16x8_t v_sum;
uint32x2_t v_s0 = vdup_n_u32(0);
uint32x2_t v_s1 = vdup_n_u32(0);
v_s0 = vld1_lane_u32((const uint32_t *)s, v_s0, 0);
v_s0 = vld1_lane_u32((const uint32_t *)(s + p), v_s0, 1);
v_s1 = vld1_lane_u32((const uint32_t *)(s + 2 * p), v_s1, 0);
v_s1 = vld1_lane_u32((const uint32_t *)(s + 3 * p), v_s1, 1);
v_sum = vaddl_u8(vreinterpret_u8_u32(v_s0), vreinterpret_u8_u32(v_s1));
return (horizontal_add_u16x8(v_sum) + 8) >> 4;
}
static INLINE void idct4x4_1_add_kernel(uint8_t **dest, const int stride,
const int16x8_t res,
uint32x2_t *const d) {
uint16x8_t a;
uint8x8_t b;
*d = vld1_lane_u32((const uint32_t *)*dest, *d, 0);
*d = vld1_lane_u32((const uint32_t *)(*dest + stride), *d, 1);
a = vaddw_u8(vreinterpretq_u16_s16(res), vreinterpret_u8_u32(*d));
b = vqmovun_s16(vreinterpretq_s16_u16(a));
vst1_lane_u32((uint32_t *)*dest, vreinterpret_u32_u8(b), 0);
*dest += stride;
vst1_lane_u32((uint32_t *)*dest, vreinterpret_u32_u8(b), 1);
*dest += stride;
}
void test_vld1_laneu32 (void)
{
uint32x2_t out_uint32x2_t;
uint32x2_t arg1_uint32x2_t;
out_uint32x2_t = vld1_lane_u32 (0, arg1_uint32x2_t, 1);
}
void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
const uint8x8_t XABCD_u8 = vld1_u8(above - 1);
const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8);
const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32);
const uint32x2_t zero = vdup_n_u32(0);
const uint32x2_t IJKL = vld1_lane_u32((const uint32_t *)left, zero, 0);
const uint8x8_t IJKL_u8 = vreinterpret_u8_u32(IJKL);
const uint64x1_t LKJI____ = vreinterpret_u64_u8(vrev32_u8(IJKL_u8));
const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC);
const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8));
const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16));
const uint8_t D = vget_lane_u8(XABCD_u8, 4);
const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6);
const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC);
const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8);
const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_);
const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
}
void vp9_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int i;
uint32x2_t d0u32 = vdup_n_u32(0);
(void)left;
d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0);
for (i = 0; i < 4; i++, dst += stride)
vst1_lane_u32((uint32_t *)dst, d0u32, 0);
}
void byte2word48_neon(const uint8_t *t, const int pitch, float *pf) {
uint16_t *p = (uint16_t *)pf;
uint8x8_t m0, m1, m2, m3, m4, m5;
m0 = vld1_u8(t);
m1 = vreinterpret_u8_u32(vld1_lane_u32((const uint32_t *)(t + 8), vreinterpret_u32_u8(m1), 0));
m1 = vreinterpret_u8_u32(vld1_lane_u32((const uint32_t *)(t + pitch * 2), vreinterpret_u32_u8(m1), 1));
m2 = vld1_u8(t + pitch * 2 + 4);
t += pitch * 4;
m3 = vld1_u8(t);
m4 = vreinterpret_u8_u32(vld1_lane_u32((const uint32_t *)(t + 8), vreinterpret_u32_u8(m4), 0));
m4 = vreinterpret_u8_u32(vld1_lane_u32((const uint32_t *)(t + pitch * 2), vreinterpret_u32_u8(m4), 1));
m5 = vld1_u8(t + pitch * 2 + 4);
vst1q_u16(p, vmovl_u8(m0));
vst1q_u16(p + 8, vmovl_u8(m1));
vst1q_u16(p + 16, vmovl_u8(m2));
vst1q_u16(p + 24, vmovl_u8(m3));
vst1q_u16(p + 32, vmovl_u8(m4));
vst1q_u16(p + 40, vmovl_u8(m5));
}
void vp9_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int i;
uint16x8_t q1u16, q3u16;
int16x8_t q1s16;
uint8x8_t d0u8 = vdup_n_u8(0);
uint32x2_t d2u32 = vdup_n_u32(0);
d0u8 = vld1_dup_u8(above - 1);
d2u32 = vld1_lane_u32((const uint32_t *)above, d2u32, 0);
q3u16 = vsubl_u8(vreinterpret_u8_u32(d2u32), d0u8);
for (i = 0; i < 4; i++, dst += stride) {
q1u16 = vdupq_n_u16((uint16_t)left[i]);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q1u16),
vreinterpretq_s16_u16(q3u16));
d0u8 = vqmovun_s16(q1s16);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
}
}
void vp9_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
uint8x8_t d0u8 = vdup_n_u8(0);
uint32x2_t d1u32 = vdup_n_u32(0);
(void)above;
d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
}
void vp8_short_idct4x4llm_neon(int16_t *input, unsigned char *pred_ptr,
int pred_stride, unsigned char *dst_ptr,
int dst_stride) {
int i;
uint32x2_t d6u32 = vdup_n_u32(0);
uint8x8_t d1u8;
int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
uint16x8_t q1u16;
int16x8_t q1s16, q2s16, q3s16, q4s16;
int32x2x2_t v2tmp0, v2tmp1;
int16x4x2_t v2tmp2, v2tmp3;
d2 = vld1_s16(input);
d3 = vld1_s16(input + 4);
d4 = vld1_s16(input + 8);
d5 = vld1_s16(input + 12);
// 1st for loop
q1s16 = vcombine_s16(d2, d4); // Swap d3 d4 here
q2s16 = vcombine_s16(d3, d5);
q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // a1
d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // b1
q3s16 = vshrq_n_s16(q3s16, 1);
q4s16 = vshrq_n_s16(q4s16, 1);
q3s16 = vqaddq_s16(q3s16, q2s16);
q4s16 = vqaddq_s16(q4s16, q2s16);
d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16)); // c1
d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16)); // d1
d2 = vqadd_s16(d12, d11);
d3 = vqadd_s16(d13, d10);
d4 = vqsub_s16(d13, d10);
d5 = vqsub_s16(d12, d11);
v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
vreinterpret_s16_s32(v2tmp1.val[0]));
v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
vreinterpret_s16_s32(v2tmp1.val[1]));
// 2nd for loop
q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp3.val[0]);
q2s16 = vcombine_s16(v2tmp2.val[1], v2tmp3.val[1]);
q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // a1
d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // b1
q3s16 = vshrq_n_s16(q3s16, 1);
q4s16 = vshrq_n_s16(q4s16, 1);
q3s16 = vqaddq_s16(q3s16, q2s16);
q4s16 = vqaddq_s16(q4s16, q2s16);
d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16)); // c1
d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16)); // d1
d2 = vqadd_s16(d12, d11);
d3 = vqadd_s16(d13, d10);
d4 = vqsub_s16(d13, d10);
d5 = vqsub_s16(d12, d11);
d2 = vrshr_n_s16(d2, 3);
d3 = vrshr_n_s16(d3, 3);
d4 = vrshr_n_s16(d4, 3);
d5 = vrshr_n_s16(d5, 3);
v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
vreinterpret_s16_s32(v2tmp1.val[0]));
v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
vreinterpret_s16_s32(v2tmp1.val[1]));
q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp2.val[1]);
q2s16 = vcombine_s16(v2tmp3.val[0], v2tmp3.val[1]);
// dc_only_idct_add
for (i = 0; i < 2; i++, q1s16 = q2s16) {
d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 0);
pred_ptr += pred_stride;
d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 1);
pred_ptr += pred_stride;
q1u16 = vaddw_u8(vreinterpretq_u16_s16(q1s16), vreinterpret_u8_u32(d6u32));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 0);
dst_ptr += dst_stride;
vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 1);
dst_ptr += dst_stride;
}
return;
}
void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
uint8x8_t d26u8, d27u8;
uint32x2_t d26u32, d27u32;
uint16x8_t q8u16, q9u16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
int16x8_t q8s16, q9s16, q13s16, q14s16;
int32x4_t q1s32, q13s32, q14s32, q15s32;
int16x4x2_t d0x2s16, d1x2s16;
int32x4x2_t q0x2s32;
uint8_t *d;
d26u32 = d27u32 = vdup_n_u32(0);
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_low_s16(q9s16);
d19s16 = vget_high_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
d20s16 = vdup_n_s16((int16_t)cospi_8_64);
d21s16 = vdup_n_s16((int16_t)cospi_16_64);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d22s16 = vdup_n_s16((int16_t)cospi_24_64);
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_high_s16(q9s16); // vswp d18 d19
d19s16 = vget_low_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
// do the transform on columns
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
q8s16 = vrshrq_n_s16(q8s16, 4);
q9s16 = vrshrq_n_s16(q9s16, 4);
d = dest;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
d += dest_stride;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d = dest;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
return;
}
void vp10_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride, int tx_type) {
uint8x8_t d26u8, d27u8;
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
uint32x2_t d26u32, d27u32;
int16x8_t q3s16, q8s16, q9s16;
uint16x8_t q8u16, q9u16;
d26u32 = d27u32 = vdup_n_u32(0);
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
TRANSPOSE4X4(&q8s16, &q9s16);
switch (tx_type) {
case 0: // idct_idct is not supported. Fall back to C
vp10_iht4x4_16_add_c(input, dest, dest_stride, tx_type);
return;
break;
case 1: // iadst_idct
// generate constants
GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
// first transform rows
IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
// transpose the matrix
TRANSPOSE4X4(&q8s16, &q9s16);
// then transform columns
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
break;
case 2: // idct_iadst
// generate constantsyy
GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
// first transform rows
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
// transpose the matrix
TRANSPOSE4X4(&q8s16, &q9s16);
// then transform columns
IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
break;
case 3: // iadst_iadst
// generate constants
GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
// first transform rows
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
// transpose the matrix
TRANSPOSE4X4(&q8s16, &q9s16);
// then transform columns
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
break;
default: // iadst_idct
assert(0);
break;
}
q8s16 = vrshrq_n_s16(q8s16, 4);
q9s16 = vrshrq_n_s16(q9s16, 4);
d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 0);
dest += dest_stride;
d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 1);
dest += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 0);
dest += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 1);
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 1);
dest -= dest_stride;
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 0);
dest -= dest_stride;
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 1);
dest -= dest_stride;
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 0);
return;
}
void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, // unused
int x_step_q4, // unused
const int16_t *filter_y, int y_step_q4, int w,
int h) {
int height;
const uint8_t *s;
uint8_t *d;
uint32x2_t d2u32, d3u32;
uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
int16x4_t d24s16, d25s16, d26s16, d27s16;
uint16x4_t d2u16, d3u16, d4u16, d5u16;
int16x8_t q0s16;
uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
int32x4_t q1s32, q2s32, q14s32, q15s32;
assert(y_step_q4 == 16);
(void)x_step_q4;
(void)y_step_q4;
(void)filter_x;
src -= src_stride * 3;
q0s16 = vld1q_s16(filter_y);
for (; w > 0; w -= 4, src += 4, dst += 4) { // loop_vert_h
s = src;
d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
s += src_stride;
d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
s += src_stride;
d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
s += src_stride;
d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
s += src_stride;
d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
s += src_stride;
d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
s += src_stride;
d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
s += src_stride;
q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d = dst;
for (height = h; height > 0; height -= 4) { // loop_vert
d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
s += src_stride;
d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
s += src_stride;
d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
s += src_stride;
d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
s += src_stride;
q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
__builtin_prefetch(d);
__builtin_prefetch(d + dst_stride);
q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
d22s16, d24s16, q0s16);
__builtin_prefetch(d + dst_stride * 2);
__builtin_prefetch(d + dst_stride * 3);
q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
d24s16, d26s16, q0s16);
__builtin_prefetch(s);
__builtin_prefetch(s + src_stride);
q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
d26s16, d27s16, q0s16);
__builtin_prefetch(s + src_stride * 2);
__builtin_prefetch(s + src_stride * 3);
q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
d27s16, d25s16, q0s16);
d2u16 = vqrshrun_n_s32(q1s32, 7);
d3u16 = vqrshrun_n_s32(q2s32, 7);
d4u16 = vqrshrun_n_s32(q14s32, 7);
d5u16 = vqrshrun_n_s32(q15s32, 7);
q1u16 = vcombine_u16(d2u16, d3u16);
q2u16 = vcombine_u16(d4u16, d5u16);
d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
vst1_lane_u32((uint32_t *)d, d2u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d2u32, 1);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 1);
d += dst_stride;
q8u16 = q10u16;
d18s16 = d22s16;
d19s16 = d24s16;
q10u16 = q13u16;
d22s16 = d25s16;
}
}
return;
}
