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);
}
static WEBP_INLINE uint32_t Average2(const uint32_t* const a,
const uint32_t* const b) {
const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
const uint8x8_t avg = vhadd_u8(a0, b0);
return vget_lane_u32(vreinterpret_u32_u8(avg), 0);
}
static WEBP_INLINE uint32_t Average3(const uint32_t* const a,
const uint32_t* const b,
const uint32_t* const c) {
const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
const uint8x8_t avg1 = vhadd_u8(a0, c0);
const uint8x8_t avg2 = vhadd_u8(avg1, b0);
return vget_lane_u32(vreinterpret_u32_u8(avg2), 0);
}
static WEBP_INLINE uint32_t Average2(const uint32_t* const a,
const uint32_t* const b) {
const uint64x1_t a0 = { *a }, b0 = { *b };
const uint8x8_t a1 = vreinterpret_u8_u64(a0);
const uint8x8_t b1 = vreinterpret_u8_u64(b0);
const uint8x8_t avg = vhadd_u8(a1, b1);
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(avg), 0);
return ret;
}
static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint16x8_t sum0 = vaddl_u8(p0, p1); // add and widen
const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2)); // widen and subtract
const uint8x8_t out = vqmovn_u16(sum1); // narrow and clamp
return vget_lane_u32(vreinterpret_u32_u8(out), 0);
}
static WEBP_INLINE uint32_t ClampedAddSubtractHalf(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint8x8_t avg = vhadd_u8(p0, p1); // Average(c0,c1)
const uint8x8_t ab = vshr_n_u8(vqsub_u8(avg, p2), 1); // (a-b)>>1 saturated
const uint8x8_t ba = vshr_n_u8(vqsub_u8(p2, avg), 1); // (b-a)>>1 saturated
const uint8x8_t out = vqsub_u8(vqadd_u8(avg, ab), ba);
return vget_lane_u32(vreinterpret_u32_u8(out), 0);
}
static WEBP_INLINE uint32_t Average4(const uint32_t* const a,
const uint32_t* const b,
const uint32_t* const c,
const uint32_t* const d) {
const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
const uint8x8_t d0 = vreinterpret_u8_u64(vcreate_u64(*d));
const uint8x8_t avg1 = vhadd_u8(a0, b0);
const uint8x8_t avg2 = vhadd_u8(c0, d0);
const uint8x8_t avg3 = vhadd_u8(avg1, avg2);
return vget_lane_u32(vreinterpret_u32_u8(avg3), 0);
}
static WEBP_INLINE uint32_t Average3(const uint32_t* const a,
const uint32_t* const b,
const uint32_t* const c) {
const uint64x1_t a0 = { *a }, b0 = { *b }, c0 = { *c };
const uint8x8_t a1 = vreinterpret_u8_u64(a0);
const uint8x8_t b1 = vreinterpret_u8_u64(b0);
const uint8x8_t c1 = vreinterpret_u8_u64(c0);
const uint8x8_t avg1 = vhadd_u8(a1, c1);
const uint8x8_t avg2 = vhadd_u8(avg1, b1);
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(avg2), 0);
return ret;
}
static WEBP_INLINE uint32_t Select(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint8x8_t bc = vabd_u8(p1, p2); // |b-c|
const uint8x8_t ac = vabd_u8(p0, p2); // |a-c|
const int16x4_t sum_bc = vreinterpret_s16_u16(vpaddl_u8(bc));
const int16x4_t sum_ac = vreinterpret_s16_u16(vpaddl_u8(ac));
const int32x2_t diff = vpaddl_s16(vsub_s16(sum_bc, sum_ac));
const int32_t pa_minus_pb = vget_lane_s32(diff, 0);
return (pa_minus_pb <= 0) ? *c0 : *c1;
}
static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint64x1_t C0 = { *c0, 0 }, C1 = { *c1, 0 }, C2 = { *c2, 0 };
const uint8x8_t p0 = vreinterpret_u8_u64(C0);
const uint8x8_t p1 = vreinterpret_u8_u64(C1);
const uint8x8_t p2 = vreinterpret_u8_u64(C2);
const uint16x8_t sum0 = vaddl_u8(p0, p1); // add and widen
const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2)); // widen and subtract
const uint8x8_t out = vqmovn_u16(sum1); // narrow and clamp
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(out), 0);
return ret;
}
void vp9_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
uint8x8_t d0u8 = vdup_n_u8(0);
uint64x1_t d1u64 = vdup_n_u64(0);
(void)above;
d1u64 = vld1_u64((const uint64_t *)left);
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
vst1_u8(dst, d0u8);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
vst1_u8(dst, d0u8);
}
void vp9_idct8x8_1_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8x8_t d2u8, d3u8, d30u8, d31u8;
uint64x1_t d2u64, d3u64, d4u64, d5u64;
uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
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, 5);
q0s16 = vdupq_n_s16(a1);
q0u16 = vreinterpretq_u16_s16(q0s16);
d1 = d2 = dest;
for (i = 0; i < 2; i++) {
d2u64 = vld1_u64((const uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((const uint64_t *)d1);
d1 += dest_stride;
d4u64 = vld1_u64((const uint64_t *)d1);
d1 += dest_stride;
d5u64 = vld1_u64((const uint64_t *)d1);
d1 += dest_stride;
q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d31u8));
d2 += dest_stride;
}
return;
}
static WEBP_INLINE uint32_t ClampedAddSubtractHalf(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint64x1_t C0 = { *c0, 0 }, C1 = { *c1, 0 }, C2 = { *c2, 0 };
const uint8x8_t p0 = vreinterpret_u8_u64(C0);
const uint8x8_t p1 = vreinterpret_u8_u64(C1);
const uint8x8_t p2 = vreinterpret_u8_u64(C2);
const uint8x8_t avg = vhadd_u8(p0, p1); // Average(c0,c1)
const uint8x8_t ab = vshr_n_u8(vqsub_u8(avg, p2), 1); // (a-b)>>1 saturated
const uint8x8_t ba = vshr_n_u8(vqsub_u8(p2, avg), 1); // (b-a)>>1 saturated
const uint8x8_t out = vqsub_u8(vqadd_u8(avg, ab), ba);
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(out), 0);
return ret;
}
static WEBP_INLINE uint32_t Select(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint64x1_t C0 = { *c0, 0 }, C1 = { *c1, 0 }, C2 = { *c2, 0 };
const uint8x8_t p0 = vreinterpret_u8_u64(C0);
const uint8x8_t p1 = vreinterpret_u8_u64(C1);
const uint8x8_t p2 = vreinterpret_u8_u64(C2);
const uint8x8_t bc = vabd_u8(p1, p2); // |b-c|
const uint8x8_t ac = vabd_u8(p0, p2); // |a-c|
const int16x4_t sum_bc = vreinterpret_s16_u16(vpaddl_u8(bc));
const int16x4_t sum_ac = vreinterpret_s16_u16(vpaddl_u8(ac));
const int32x2_t diff = vpaddl_s16(vsub_s16(sum_bc, sum_ac));
int32_t pa_minus_pb;
vst1_lane_s32(&pa_minus_pb, diff, 0);
return (pa_minus_pb <= 0) ? *c0 : *c1;
}
void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(above)); // top row
const uint64x1_t A1 = vshr_n_u64(A0, 8);
const uint64x1_t A2 = vshr_n_u64(A0, 16);
const uint8x8_t ABCDEFGH = vreinterpret_u8_u64(A0);
const uint8x8_t BCDEFGH0 = vreinterpret_u8_u64(A1);
const uint8x8_t CDEFGH00 = vreinterpret_u8_u64(A2);
const uint8x8_t avg1 = vhadd_u8(ABCDEFGH, CDEFGH00);
const uint8x8_t avg2 = vrhadd_u8(avg1, BCDEFGH0);
const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
const uint32x2_t r0 = vreinterpret_u32_u8(avg2);
const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
const uint32x2_t r3 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
(void)left;
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);
dst[3 * stride + 3] = above[7];
}
void av1_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride, const TxfmParam *txfm_param) {
int i;
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
uint64x1_t d0u64, d1u64, d2u64, d3u64;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
uint16x8_t q8u16, q9u16, q10u16, q11u16;
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
q10s16 = vld1q_s16(input + 8 * 2);
q11s16 = vld1q_s16(input + 8 * 3);
q12s16 = vld1q_s16(input + 8 * 4);
q13s16 = vld1q_s16(input + 8 * 5);
q14s16 = vld1q_s16(input + 8 * 6);
q15s16 = vld1q_s16(input + 8 * 7);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
const TX_TYPE tx_type = txfm_param->tx_type;
switch (tx_type) {
case DCT_DCT: // idct_idct is not supported. Fall back to C
av1_iht8x8_64_add_c(input, dest, dest_stride, txfm_param);
return;
break;
case ADST_DCT: // iadst_idct
// generate IDCT constants
// GENERATE_IDCT_CONSTANTS
// first transform rows
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// transpose the matrix
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// generate IADST constants
// GENERATE_IADST_CONSTANTS
// then transform columns
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
break;
case DCT_ADST: // idct_iadst
// generate IADST constants
// GENERATE_IADST_CONSTANTS
// first transform rows
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// transpose the matrix
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// generate IDCT constants
// GENERATE_IDCT_CONSTANTS
// then transform columns
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
break;
case ADST_ADST: // iadst_iadst
// generate IADST constants
// GENERATE_IADST_CONSTANTS
// first transform rows
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// transpose the matrix
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// then transform columns
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
break;
default: // iadst_idct
assert(0);
break;
}
q8s16 = vrshrq_n_s16(q8s16, 5);
q9s16 = vrshrq_n_s16(q9s16, 5);
q10s16 = vrshrq_n_s16(q10s16, 5);
q11s16 = vrshrq_n_s16(q11s16, 5);
q12s16 = vrshrq_n_s16(q12s16, 5);
q13s16 = vrshrq_n_s16(q13s16, 5);
q14s16 = vrshrq_n_s16(q14s16, 5);
q15s16 = vrshrq_n_s16(q15s16, 5);
for (d1 = d2 = dest, i = 0; i < 2; i++) {
if (i != 0) {
q8s16 = q12s16;
q9s16 = q13s16;
q10s16 = q14s16;
q11s16 = q15s16;
}
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
q10u16 =
vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
q11u16 =
vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
}
return;
}
void vpx_idct8x8_64_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
uint64x1_t d0u64, d1u64, d2u64, d3u64;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
uint16x8_t q8u16, q9u16, q10u16, q11u16;
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
q10s16 = vld1q_s16(input + 16);
q11s16 = vld1q_s16(input + 24);
q12s16 = vld1q_s16(input + 32);
q13s16 = vld1q_s16(input + 40);
q14s16 = vld1q_s16(input + 48);
q15s16 = vld1q_s16(input + 56);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
q8s16 = vrshrq_n_s16(q8s16, 5);
q9s16 = vrshrq_n_s16(q9s16, 5);
q10s16 = vrshrq_n_s16(q10s16, 5);
q11s16 = vrshrq_n_s16(q11s16, 5);
q12s16 = vrshrq_n_s16(q12s16, 5);
q13s16 = vrshrq_n_s16(q13s16, 5);
q14s16 = vrshrq_n_s16(q14s16, 5);
q15s16 = vrshrq_n_s16(q15s16, 5);
d1 = d2 = dest;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
q8s16 = q12s16;
q9s16 = q13s16;
q10s16 = q14s16;
q11s16 = q15s16;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
return;
}
void vpx_idct8x8_12_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
int16x4_t d26s16, d27s16, d28s16, d29s16;
uint64x1_t d0u64, d1u64, d2u64, d3u64;
int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
uint16x8_t q8u16, q9u16, q10u16, q11u16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
q10s16 = vld1q_s16(input + 16);
q11s16 = vld1q_s16(input + 24);
q12s16 = vld1q_s16(input + 32);
q13s16 = vld1q_s16(input + 40);
q14s16 = vld1q_s16(input + 48);
q15s16 = vld1q_s16(input + 56);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
// First transform rows
// stage 1
q0s16 = vdupq_n_s16(cospi_28_64 * 2);
q1s16 = vdupq_n_s16(cospi_4_64 * 2);
q4s16 = vqrdmulhq_s16(q9s16, q0s16);
q0s16 = vdupq_n_s16(-cospi_20_64 * 2);
q7s16 = vqrdmulhq_s16(q9s16, q1s16);
q1s16 = vdupq_n_s16(cospi_12_64 * 2);
q5s16 = vqrdmulhq_s16(q11s16, q0s16);
q0s16 = vdupq_n_s16(cospi_16_64 * 2);
q6s16 = vqrdmulhq_s16(q11s16, q1s16);
// stage 2 & stage 3 - even half
q1s16 = vdupq_n_s16(cospi_24_64 * 2);
q9s16 = vqrdmulhq_s16(q8s16, q0s16);
q0s16 = vdupq_n_s16(cospi_8_64 * 2);
q13s16 = vqrdmulhq_s16(q10s16, q1s16);
q15s16 = vqrdmulhq_s16(q10s16, q0s16);
// stage 3 -odd half
q0s16 = vaddq_s16(q9s16, q15s16);
q1s16 = vaddq_s16(q9s16, q13s16);
q2s16 = vsubq_s16(q9s16, q13s16);
q3s16 = vsubq_s16(q9s16, q15s16);
// stage 2 - odd half
q13s16 = vsubq_s16(q4s16, q5s16);
q4s16 = vaddq_s16(q4s16, q5s16);
q14s16 = vsubq_s16(q7s16, q6s16);
q7s16 = vaddq_s16(q7s16, q6s16);
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
d16s16 = vdup_n_s16(cospi_16_64);
q9s32 = vmull_s16(d28s16, d16s16);
q10s32 = vmull_s16(d29s16, d16s16);
q11s32 = vmull_s16(d28s16, d16s16);
q12s32 = vmull_s16(d29s16, d16s16);
q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
d10s16 = vqrshrn_n_s32(q9s32, 14);
d11s16 = vqrshrn_n_s32(q10s32, 14);
d12s16 = vqrshrn_n_s32(q11s32, 14);
d13s16 = vqrshrn_n_s32(q12s32, 14);
q5s16 = vcombine_s16(d10s16, d11s16);
q6s16 = vcombine_s16(d12s16, d13s16);
// stage 4
q8s16 = vaddq_s16(q0s16, q7s16);
q9s16 = vaddq_s16(q1s16, q6s16);
q10s16 = vaddq_s16(q2s16, q5s16);
q11s16 = vaddq_s16(q3s16, q4s16);
q12s16 = vsubq_s16(q3s16, q4s16);
q13s16 = vsubq_s16(q2s16, q5s16);
q14s16 = vsubq_s16(q1s16, q6s16);
q15s16 = vsubq_s16(q0s16, q7s16);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
&q12s16, &q13s16, &q14s16, &q15s16);
q8s16 = vrshrq_n_s16(q8s16, 5);
q9s16 = vrshrq_n_s16(q9s16, 5);
q10s16 = vrshrq_n_s16(q10s16, 5);
q11s16 = vrshrq_n_s16(q11s16, 5);
q12s16 = vrshrq_n_s16(q12s16, 5);
q13s16 = vrshrq_n_s16(q13s16, 5);
q14s16 = vrshrq_n_s16(q14s16, 5);
q15s16 = vrshrq_n_s16(q15s16, 5);
d1 = d2 = dest;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
q8s16 = q12s16;
q9s16 = q13s16;
q10s16 = q14s16;
q11s16 = q15s16;
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
vreinterpret_u8_u64(d1u64));
q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
vreinterpret_u8_u64(d2u64));
q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
return;
}