static INLINE void IDCT4x4_1D(int16x4_t *d0s16, int16x4_t *d1s16,
int16x4_t *d2s16, int16x8_t *q8s16,
int16x8_t *q9s16) {
int16x4_t d16s16, d17s16, d18s16, d19s16, d23s16, d24s16;
int16x4_t d26s16, d27s16, d28s16, d29s16;
int32x4_t q10s32, q13s32, q14s32, q15s32;
int16x8_t q13s16, q14s16;
d16s16 = vget_low_s16(*q8s16);
d17s16 = vget_high_s16(*q8s16);
d18s16 = vget_low_s16(*q9s16);
d19s16 = vget_high_s16(*q9s16);
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, *d2s16);
q10s32 = vmull_s16(d17s16, *d0s16);
q13s32 = vmull_s16(d23s16, *d1s16);
q14s32 = vmull_s16(d24s16, *d1s16);
q15s32 = vmlsl_s16(q15s32, d19s16, *d0s16);
q10s32 = vmlal_s16(q10s32, d19s16, *d2s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q10s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
*q8s16 = vaddq_s16(q13s16, q14s16);
*q9s16 = vsubq_s16(q13s16, q14s16);
*q9s16 = vcombine_s16(vget_high_s16(*q9s16), vget_low_s16(*q9s16)); // vswp
return;
}
void test_vqRshrn_ns32 (void)
{
int16x4_t out_int16x4_t;
int32x4_t arg0_int32x4_t;
out_int16x4_t = vqrshrn_n_s32 (arg0_int32x4_t, 1);
}
static INLINE void IADST4x4_1D(int16x4_t *d3s16, int16x4_t *d4s16,
int16x4_t *d5s16, int16x8_t *q3s16,
int16x8_t *q8s16, int16x8_t *q9s16) {
int16x4_t d6s16, d16s16, d17s16, d18s16, d19s16;
int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
d6s16 = vget_low_s16(*q3s16);
d16s16 = vget_low_s16(*q8s16);
d17s16 = vget_high_s16(*q8s16);
d18s16 = vget_low_s16(*q9s16);
d19s16 = vget_high_s16(*q9s16);
q10s32 = vmull_s16(*d3s16, d16s16);
q11s32 = vmull_s16(*d4s16, d16s16);
q12s32 = vmull_s16(d6s16, d17s16);
q13s32 = vmull_s16(*d5s16, d18s16);
q14s32 = vmull_s16(*d3s16, d18s16);
q15s32 = vmovl_s16(d16s16);
q15s32 = vaddw_s16(q15s32, d19s16);
q8s32 = vmull_s16(*d4s16, d19s16);
q15s32 = vsubw_s16(q15s32, d18s16);
q9s32 = vmull_s16(*d5s16, d19s16);
q10s32 = vaddq_s32(q10s32, q13s32);
q10s32 = vaddq_s32(q10s32, q8s32);
q11s32 = vsubq_s32(q11s32, q14s32);
q8s32 = vdupq_n_s32(sinpi_3_9);
q11s32 = vsubq_s32(q11s32, q9s32);
q15s32 = vmulq_s32(q15s32, q8s32);
q13s32 = vaddq_s32(q10s32, q12s32);
q10s32 = vaddq_s32(q10s32, q11s32);
q14s32 = vaddq_s32(q11s32, q12s32);
q10s32 = vsubq_s32(q10s32, q12s32);
d16s16 = vqrshrn_n_s32(q13s32, 14);
d17s16 = vqrshrn_n_s32(q14s32, 14);
d18s16 = vqrshrn_n_s32(q15s32, 14);
d19s16 = vqrshrn_n_s32(q10s32, 14);
*q8s16 = vcombine_s16(d16s16, d17s16);
*q9s16 = vcombine_s16(d18s16, d19s16);
return;
}
void vpx_highbd_idct4x4_16_add_neon(const tran_low_t *input, uint16_t *dest,
int stride, int bd) {
const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
int32x4_t c0 = vld1q_s32(input);
int32x4_t c1 = vld1q_s32(input + 4);
int32x4_t c2 = vld1q_s32(input + 8);
int32x4_t c3 = vld1q_s32(input + 12);
int16x8_t a0, a1;
if (bd == 8) {
const int16x4_t cospis = vld1_s16(kCospi);
// Rows
a0 = vcombine_s16(vmovn_s32(c0), vmovn_s32(c1));
a1 = vcombine_s16(vmovn_s32(c2), vmovn_s32(c3));
idct4x4_16_kernel_bd8(cospis, &a0, &a1);
// Columns
a1 = vcombine_s16(vget_high_s16(a1), vget_low_s16(a1));
idct4x4_16_kernel_bd8(cospis, &a0, &a1);
a0 = vrshrq_n_s16(a0, 4);
a1 = vrshrq_n_s16(a1, 4);
} else {
const int32x4_t cospis = vld1q_s32(kCospi32);
if (bd == 10) {
idct4x4_16_kernel_bd10(cospis, &c0, &c1, &c2, &c3);
idct4x4_16_kernel_bd10(cospis, &c0, &c1, &c2, &c3);
} else {
idct4x4_16_kernel_bd12(cospis, &c0, &c1, &c2, &c3);
idct4x4_16_kernel_bd12(cospis, &c0, &c1, &c2, &c3);
}
a0 = vcombine_s16(vqrshrn_n_s32(c0, 4), vqrshrn_n_s32(c1, 4));
a1 = vcombine_s16(vqrshrn_n_s32(c3, 4), vqrshrn_n_s32(c2, 4));
}
highbd_idct4x4_1_add_kernel1(&dest, stride, a0, max);
highbd_idct4x4_1_add_kernel2(&dest, stride, a1, max);
}
static INLINE void IDCT8x8_1D(
int16x8_t *q8s16,
int16x8_t *q9s16,
int16x8_t *q10s16,
int16x8_t *q11s16,
int16x8_t *q12s16,
int16x8_t *q13s16,
int16x8_t *q14s16,
int16x8_t *q15s16) {
int16x4_t d0s16, d1s16, d2s16, d3s16;
int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
d0s16 = vdup_n_s16(cospi_28_64);
d1s16 = vdup_n_s16(cospi_4_64);
d2s16 = vdup_n_s16(cospi_12_64);
d3s16 = vdup_n_s16(cospi_20_64);
d16s16 = vget_low_s16(*q8s16);
d17s16 = vget_high_s16(*q8s16);
d18s16 = vget_low_s16(*q9s16);
d19s16 = vget_high_s16(*q9s16);
d20s16 = vget_low_s16(*q10s16);
d21s16 = vget_high_s16(*q10s16);
d22s16 = vget_low_s16(*q11s16);
d23s16 = vget_high_s16(*q11s16);
d24s16 = vget_low_s16(*q12s16);
d25s16 = vget_high_s16(*q12s16);
d26s16 = vget_low_s16(*q13s16);
d27s16 = vget_high_s16(*q13s16);
d28s16 = vget_low_s16(*q14s16);
d29s16 = vget_high_s16(*q14s16);
d30s16 = vget_low_s16(*q15s16);
d31s16 = vget_high_s16(*q15s16);
q2s32 = vmull_s16(d18s16, d0s16);
q3s32 = vmull_s16(d19s16, d0s16);
q5s32 = vmull_s16(d26s16, d2s16);
q6s32 = vmull_s16(d27s16, d2s16);
q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
d8s16 = vqrshrn_n_s32(q2s32, 14);
d9s16 = vqrshrn_n_s32(q3s32, 14);
d10s16 = vqrshrn_n_s32(q5s32, 14);
d11s16 = vqrshrn_n_s32(q6s32, 14);
q4s16 = vcombine_s16(d8s16, d9s16);
q5s16 = vcombine_s16(d10s16, d11s16);
q2s32 = vmull_s16(d18s16, d1s16);
q3s32 = vmull_s16(d19s16, d1s16);
q9s32 = vmull_s16(d26s16, d3s16);
q13s32 = vmull_s16(d27s16, d3s16);
q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
d14s16 = vqrshrn_n_s32(q2s32, 14);
d15s16 = vqrshrn_n_s32(q3s32, 14);
d12s16 = vqrshrn_n_s32(q9s32, 14);
d13s16 = vqrshrn_n_s32(q13s32, 14);
q6s16 = vcombine_s16(d12s16, d13s16);
q7s16 = vcombine_s16(d14s16, d15s16);
d0s16 = vdup_n_s16(cospi_16_64);
q2s32 = vmull_s16(d16s16, d0s16);
q3s32 = vmull_s16(d17s16, d0s16);
q13s32 = vmull_s16(d16s16, d0s16);
q15s32 = vmull_s16(d17s16, d0s16);
q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
d0s16 = vdup_n_s16(cospi_24_64);
d1s16 = vdup_n_s16(cospi_8_64);
d18s16 = vqrshrn_n_s32(q2s32, 14);
d19s16 = vqrshrn_n_s32(q3s32, 14);
d22s16 = vqrshrn_n_s32(q13s32, 14);
d23s16 = vqrshrn_n_s32(q15s32, 14);
*q9s16 = vcombine_s16(d18s16, d19s16);
*q11s16 = vcombine_s16(d22s16, d23s16);
q2s32 = vmull_s16(d20s16, d0s16);
q3s32 = vmull_s16(d21s16, d0s16);
q8s32 = vmull_s16(d20s16, d1s16);
q12s32 = vmull_s16(d21s16, d1s16);
q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
d26s16 = vqrshrn_n_s32(q2s32, 14);
d27s16 = vqrshrn_n_s32(q3s32, 14);
d30s16 = vqrshrn_n_s32(q8s32, 14);
d31s16 = vqrshrn_n_s32(q12s32, 14);
*q13s16 = vcombine_s16(d26s16, d27s16);
*q15s16 = vcombine_s16(d30s16, d31s16);
q0s16 = vaddq_s16(*q9s16, *q15s16);
q1s16 = vaddq_s16(*q11s16, *q13s16);
q2s16 = vsubq_s16(*q11s16, *q13s16);
q3s16 = vsubq_s16(*q9s16, *q15s16);
*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);
*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);
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;
}
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;
}
