blob: 2f3567aea30ef6bae0aebbe1327552271f21e36a [file] [log] [blame]
/*
* Copyright (c) 2018, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <arm_neon.h>
#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "config/av1_rtcd.h"
#include "av1/common/av1_inv_txfm1d.h"
#include "av1/common/av1_inv_txfm1d_cfg.h"
#include "av1/common/av1_txfm.h"
#include "av1/common/enums.h"
#include "av1/common/idct.h"
#include "av1/common/arm/av1_inv_txfm_neon.h"
#include "av1/common/arm/transpose_neon.h"
// 1D itx types
typedef enum ATTRIBUTE_PACKED {
IDCT_1D,
IADST_1D,
IFLIPADST_1D = IADST_1D,
IIDENTITY_1D,
ITX_TYPES_1D,
} ITX_TYPE_1D;
static const ITX_TYPE_1D vitx_1d_tab[TX_TYPES] = {
IDCT_1D, IADST_1D, IDCT_1D, IADST_1D,
IFLIPADST_1D, IDCT_1D, IFLIPADST_1D, IADST_1D,
IFLIPADST_1D, IIDENTITY_1D, IDCT_1D, IIDENTITY_1D,
IADST_1D, IIDENTITY_1D, IFLIPADST_1D, IIDENTITY_1D,
};
static const ITX_TYPE_1D hitx_1d_tab[TX_TYPES] = {
IDCT_1D, IDCT_1D, IADST_1D, IADST_1D,
IDCT_1D, IFLIPADST_1D, IFLIPADST_1D, IFLIPADST_1D,
IADST_1D, IIDENTITY_1D, IIDENTITY_1D, IDCT_1D,
IIDENTITY_1D, IADST_1D, IIDENTITY_1D, IFLIPADST_1D,
};
// 1D functions
static const transform_1d_neon lowbd_txfm_all_1d_arr[TX_SIZES][ITX_TYPES_1D] = {
{ av1_idct4, av1_iadst4, av1_iidentity4_c },
{ av1_idct8, av1_iadst8, av1_iidentity8_c },
{ av1_idct16, av1_iadst16, av1_iidentity16_c },
{ av1_idct32, NULL, NULL },
{ av1_idct64, NULL, NULL },
};
static INLINE void lowbd_add_flip_buffer_8xn_neon(int16x8_t *in,
uint8_t *output, int stride,
int flipud,
const int height) {
int j = flipud ? (height - 1) : 0;
const int step = flipud ? -1 : 1;
int16x8_t temp_output;
for (int i = 0; i < height; ++i, j += step) {
temp_output = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(output)));
temp_output = vaddq_s16(temp_output, in[j]);
vst1_u8(output, vqmovun_s16(temp_output));
output += stride;
}
}
static INLINE uint8x16_t lowbd_get_recon_16x16_neon(const uint8x16_t pred,
int16x8_t res0,
int16x8_t res1) {
int16x8_t temp_output[2];
uint8x16_t temp_output_8q;
temp_output[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pred)));
temp_output[0] = vaddq_s16(temp_output[0], res0);
temp_output[1] = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pred)));
temp_output[1] = vaddq_s16(temp_output[1], res1);
temp_output_8q =
vcombine_u8(vqmovun_s16(temp_output[0]), vqmovun_s16(temp_output[1]));
return temp_output_8q;
}
static INLINE void lowbd_add_flip_buffer_16xn_neon(int16x8_t *in,
uint8_t *output, int stride,
int flipud, int height) {
uint8x16_t temp_output_8q;
int j = flipud ? (height - 1) : 0;
const int step = flipud ? -1 : 1;
for (int i = 0; i < height; ++i, j += step) {
temp_output_8q = vld1q_u8(output + i * stride);
temp_output_8q =
lowbd_get_recon_16x16_neon(temp_output_8q, in[j], in[j + height]);
vst1q_u8((output + i * stride), temp_output_8q);
}
}
static INLINE void lowbd_inv_txfm2d_memset_neon(int16x8_t *a, int size,
int value) {
for (int i = 0; i < size; i++) {
a[i] = vdupq_n_s16((int16_t)value);
}
}
static INLINE void btf_16_lane_0_1_neon(const int16x8_t in0,
const int16x8_t in1, const int16x4_t c,
int16x8_t *t0, int16x8_t *t1) {
int32x4_t s0[2], s1[2];
int16x4_t v0[2], v1[2];
s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 0);
s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 0);
s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 1);
s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 1);
s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 1);
s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 1);
s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 0);
s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 0);
v0[0] = vrshrn_n_s32(s0[0], INV_COS_BIT);
v0[1] = vrshrn_n_s32(s0[1], INV_COS_BIT);
v1[0] = vrshrn_n_s32(s1[0], INV_COS_BIT);
v1[1] = vrshrn_n_s32(s1[1], INV_COS_BIT);
*t0 = vcombine_s16(v0[0], v0[1]);
*t1 = vcombine_s16(v1[0], v1[1]);
}
static INLINE void btf_16_lane_1_0_neon(const int16x8_t in0,
const int16x8_t in1, const int16x4_t c,
int16x8_t *t0, int16x8_t *t1) {
int32x4_t s0[2], s1[2];
int16x4_t v0[2], v1[2];
s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 1);
s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 1);
s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 0);
s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 0);
s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 0);
s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 0);
s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 1);
s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 1);
v0[0] = vrshrn_n_s32(s0[0], INV_COS_BIT);
v0[1] = vrshrn_n_s32(s0[1], INV_COS_BIT);
v1[0] = vrshrn_n_s32(s1[0], INV_COS_BIT);
v1[1] = vrshrn_n_s32(s1[1], INV_COS_BIT);
*t0 = vcombine_s16(v0[0], v0[1]);
*t1 = vcombine_s16(v1[0], v1[1]);
}
static INLINE void btf_16_lane_2_3_neon(const int16x8_t in0,
const int16x8_t in1, const int16x4_t c,
int16x8_t *t0, int16x8_t *t1) {
int32x4_t s0[2], s1[2];
int16x4_t v0[2], v1[2];
s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 2);
s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 2);
s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 3);
s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 3);
s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 3);
s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 3);
s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 2);
s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 2);
v0[0] = vrshrn_n_s32(s0[0], INV_COS_BIT);
v0[1] = vrshrn_n_s32(s0[1], INV_COS_BIT);
v1[0] = vrshrn_n_s32(s1[0], INV_COS_BIT);
v1[1] = vrshrn_n_s32(s1[1], INV_COS_BIT);
*t0 = vcombine_s16(v0[0], v0[1]);
*t1 = vcombine_s16(v1[0], v1[1]);
}
static INLINE void btf_16_neon(const int16x8_t in0, int16_t coef1,
int16_t coef2, int16x8_t *t0, int16x8_t *t1) {
int32x4_t s0_l, s0_h, s1_l, s1_h;
int16x4_t v0[2], v1[2];
s0_l = vmull_n_s16(vget_low_s16(in0), coef1);
s0_h = vmull_n_s16(vget_high_s16(in0), coef1);
s1_l = vmull_n_s16(vget_low_s16(in0), coef2);
s1_h = vmull_n_s16(vget_high_s16(in0), coef2);
v0[0] = vrshrn_n_s32(s0_l, INV_COS_BIT);
v0[1] = vrshrn_n_s32(s0_h, INV_COS_BIT);
v1[0] = vrshrn_n_s32(s1_l, INV_COS_BIT);
v1[1] = vrshrn_n_s32(s1_h, INV_COS_BIT);
*t0 = vcombine_s16(v0[0], v0[1]);
*t1 = vcombine_s16(v1[0], v1[1]);
}
static INLINE void btf_16_lane_3_2_neon(const int16x8_t in0,
const int16x8_t in1, const int16x4_t c,
int16x8_t *t0, int16x8_t *t1) {
int32x4_t s0[2], s1[2];
int16x4_t v0[2], v1[2];
s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 3);
s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 3);
s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 2);
s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 2);
s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 2);
s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 2);
s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 3);
s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 3);
v0[0] = vrshrn_n_s32(s0[0], INV_COS_BIT);
v0[1] = vrshrn_n_s32(s0[1], INV_COS_BIT);
v1[0] = vrshrn_n_s32(s1[0], INV_COS_BIT);
v1[1] = vrshrn_n_s32(s1[1], INV_COS_BIT);
*t0 = vcombine_s16(v0[0], v0[1]);
*t1 = vcombine_s16(v1[0], v1[1]);
}
static INLINE void btf_16_half_neon(int16x8_t *const x, const int16x4_t c) {
int32x4_t t0[2], t1[2];
int16x4_t v0[2], v1[2];
// Don't add/sub before multiply, which will overflow in iadst8.
const int32x4_t x0_lo = vmull_lane_s16(vget_low_s16(x[0]), c, 0);
const int32x4_t x0_hi = vmull_lane_s16(vget_high_s16(x[0]), c, 0);
const int32x4_t x1_lo = vmull_lane_s16(vget_low_s16(x[1]), c, 0);
const int32x4_t x1_hi = vmull_lane_s16(vget_high_s16(x[1]), c, 0);
t0[0] = vaddq_s32(x0_lo, x1_lo);
t0[1] = vaddq_s32(x0_hi, x1_hi);
t1[0] = vsubq_s32(x0_lo, x1_lo);
t1[1] = vsubq_s32(x0_hi, x1_hi);
v0[0] = vrshrn_n_s32(t0[0], INV_COS_BIT);
v0[1] = vrshrn_n_s32(t0[1], INV_COS_BIT);
v1[0] = vrshrn_n_s32(t1[0], INV_COS_BIT);
v1[1] = vrshrn_n_s32(t1[1], INV_COS_BIT);
x[0] = vcombine_s16(v0[0], v0[1]);
x[1] = vcombine_s16(v1[0], v1[1]);
}
static INLINE int16x4_t set_s16x4_neon(const int16_t c0, const int16_t c1,
const int16_t c2, const int16_t c3) {
int16x4_t val = vdup_n_s16((int16_t)0);
val = vset_lane_s16(c0, val, 0);
val = vset_lane_s16(c1, val, 1);
val = vset_lane_s16(c2, val, 2);
val = vset_lane_s16(c3, val, 3);
return val;
}
static INLINE void iadst8_neon(int16x8_t *const in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[4], (int16_t)cospi[60],
(int16_t)cospi[20], (int16_t)cospi[44]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[36], (int16_t)cospi[28],
(int16_t)cospi[52], (int16_t)cospi[12]);
const int16x4_t c2 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
int16x8_t x[8];
int16x8_t s0, s1, s2, s3, s4, s5, s6, s7;
// Stage 1
x[0] = in[7];
x[1] = in[0];
x[2] = in[5];
x[3] = in[2];
x[4] = in[3];
x[5] = in[4];
x[6] = in[1];
x[7] = in[6];
// Stage 2
btf_16_lane_0_1_neon(x[0], x[1], c0, &s0, &s1);
btf_16_lane_2_3_neon(x[2], x[3], c0, &s2, &s3);
btf_16_lane_0_1_neon(x[4], x[5], c1, &s4, &s5);
btf_16_lane_2_3_neon(x[6], x[7], c1, &s6, &s7);
// Stage 3
x[0] = vqaddq_s16(s0, s4);
x[1] = vqaddq_s16(s1, s5);
x[2] = vqaddq_s16(s2, s6);
x[3] = vqaddq_s16(s3, s7);
x[4] = vqsubq_s16(s0, s4);
x[5] = vqsubq_s16(s1, s5);
x[6] = vqsubq_s16(s2, s6);
x[7] = vqsubq_s16(s3, s7);
// Stage 4
s0 = x[0];
s1 = x[1];
s2 = x[2];
s3 = x[3];
btf_16_lane_2_3_neon(x[4], x[5], c2, &s4, &s5);
btf_16_lane_3_2_neon(x[7], x[6], c2, &s7, &s6);
// Stage 5
x[0] = vqaddq_s16(s0, s2);
x[1] = vqaddq_s16(s1, s3);
x[2] = vqsubq_s16(s0, s2);
x[3] = vqsubq_s16(s1, s3);
x[4] = vqaddq_s16(s4, s6);
x[5] = vqaddq_s16(s5, s7);
x[6] = vqsubq_s16(s4, s6);
x[7] = vqsubq_s16(s5, s7);
// stage 6
btf_16_half_neon(x + 2, c2);
btf_16_half_neon(x + 6, c2);
// Stage 7
out[0] = x[0];
out[1] = vqnegq_s16(x[4]);
out[2] = x[6];
out[3] = vqnegq_s16(x[2]);
out[4] = x[3];
out[5] = vqnegq_s16(x[7]);
out[6] = x[5];
out[7] = vqnegq_s16(x[1]);
}
static INLINE void iadst8_low1_neon(int16x8_t *const in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
const int16x4_t c2 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
int16x8_t x[8];
int16x8_t s0, s1, s4, s5;
// Stage 1
x[1] = in[0];
// Stage 2
btf_16_neon(x[1], cospi[60], -cospi[4], &s0, &s1);
// Stage 3
x[0] = s0;
x[1] = s1;
x[4] = s0;
x[5] = s1;
// Stage 4
s0 = x[0];
s1 = x[1];
btf_16_lane_2_3_neon(x[4], x[5], c2, &s4, &s5);
// Stage 5
x[0] = s0;
x[1] = s1;
x[2] = s0;
x[3] = s1;
x[4] = s4;
x[5] = s5;
x[6] = s4;
x[7] = s5;
// stage 6
btf_16_half_neon(x + 2, c2);
btf_16_half_neon(x + 6, c2);
// Stage 7
out[0] = x[0];
out[1] = vqnegq_s16(x[4]);
out[2] = x[6];
out[3] = vqnegq_s16(x[2]);
out[4] = x[3];
out[5] = vqnegq_s16(x[7]);
out[6] = x[5];
out[7] = vqnegq_s16(x[1]);
}
static INLINE void idct8_neon(int16x8_t *in, int16x8_t *out, int8_t cos_bit,
int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1[8], step2[8];
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
// stage 2
btf_16_lane_0_1_neon(in[1], in[7], c0, &step1[7], &step1[4]);
btf_16_lane_2_3_neon(in[5], in[3], c0, &step1[6], &step1[5]);
// stage 3
btf_16_lane_0_1_neon(in[0], in[4], c1, &step2[0], &step2[1]);
btf_16_lane_2_3_neon(in[2], in[6], c1, &step2[3], &step2[2]);
step2[4] = vqaddq_s16(step1[4], step1[5]);
step2[5] = vqsubq_s16(step1[4], step1[5]);
step2[6] = vqsubq_s16(step1[7], step1[6]);
step2[7] = vqaddq_s16(step1[7], step1[6]);
// stage 4
step1[0] = vqaddq_s16(step2[0], step2[3]);
step1[1] = vqaddq_s16(step2[1], step2[2]);
step1[2] = vqsubq_s16(step2[1], step2[2]);
step1[3] = vqsubq_s16(step2[0], step2[3]);
btf_16_lane_0_1_neon(step2[6], step2[5], c1, &step1[6], &step1[5]);
// stage 5
out[0] = vqaddq_s16(step1[0], step2[7]);
out[1] = vqaddq_s16(step1[1], step1[6]);
out[2] = vqaddq_s16(step1[2], step1[5]);
out[3] = vqaddq_s16(step1[3], step2[4]);
out[4] = vqsubq_s16(step1[3], step2[4]);
out[5] = vqsubq_s16(step1[2], step1[5]);
out[6] = vqsubq_s16(step1[1], step1[6]);
out[7] = vqsubq_s16(step1[0], step2[7]);
}
static INLINE void idct8_low1_neon(int16x8_t *in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1;
int32x4_t t32[2];
// stage 1
// stage 2
// stage 3
t32[0] = vmull_n_s16(vget_low_s16(in[0]), (int16_t)cospi[32]);
t32[1] = vmull_n_s16(vget_high_s16(in[0]), (int16_t)cospi[32]);
step1 = vcombine_s16(vrshrn_n_s32(t32[0], INV_COS_BIT),
vrshrn_n_s32(t32[1], INV_COS_BIT));
// stage 4
// stage 5
out[0] = step1;
out[1] = step1;
out[2] = step1;
out[3] = step1;
out[4] = step1;
out[5] = step1;
out[6] = step1;
out[7] = step1;
}
void av1_round_shift_array_16_neon(int16x8_t *arr, int size, int bit) {
assert(!(size % 4));
if (!bit) return;
const int16x8_t dup_bits_n_16x8 = vdupq_n_s16((int16_t)(-bit));
for (int i = 0; i < size; i++) {
arr[i] = vrshlq_s16(arr[i], dup_bits_n_16x8);
}
}
static INLINE void flip_buf_ud_neon(int16x8_t *input, int size) {
int16x8_t temp[8];
for (int i = 0; i < size; ++i) {
temp[i] = input[size - 1 - i];
}
for (int i = 0; i < size; ++i) {
input[i] = temp[i];
}
}
static INLINE void load_buffer_32bit_to_16bit_neon(const int32_t *input,
int16x8_t *const a,
int out_size) {
for (int i = 0; i < 8; ++i) {
a[i] = vcombine_s16(vmovn_s32(vld1q_s32(input)),
vmovn_s32(vld1q_s32(input + 4)));
input += out_size;
}
}
static int16_t sqrt_2_list[TX_SIZES] = { 5793, 2 * 4096, 2 * 5793, 4 * 4096,
4 * 5793 };
static INLINE void identity_txfm_round_neon(int16x8_t *input, int16x8_t *output,
int txw_idx, int8_t size, int bit) {
const int32x4_t dup_bits_n_32x4 = vdupq_n_s32((int32_t)(-bit));
int16x4_t scale = vdup_n_s16(sqrt_2_list[txw_idx]);
int16x4_t low_i16, high_i16;
int32x4_t low_i32, high_i32;
for (int i = 0; i < size; i++) {
int32x4_t temp_out_low = vmull_s16(vget_low_s16(input[i]), scale);
int32x4_t temp_out_high = vmull_s16(vget_high_s16(input[i]), scale);
low_i32 = vrshlq_s32(vrshrq_n_s32(temp_out_low, 12), dup_bits_n_32x4);
high_i32 = vrshlq_s32(vrshrq_n_s32(temp_out_high, 12), dup_bits_n_32x4);
low_i16 = vqmovn_s32(low_i32);
high_i16 = vqmovn_s32(high_i32);
output[i] = vcombine_s16(low_i16, high_i16);
}
}
static INLINE void round_shift_for_rect(int16x8_t *input, int16x8_t *output,
int size) {
int32x4_t out_low, out_high;
int16x4_t low, high;
for (int z = 0; z < size; ++z) {
out_low = vmull_n_s16(vget_low_s16(input[z]), (int16_t)NewInvSqrt2);
out_high = vmull_n_s16(vget_high_s16(input[z]), (int16_t)NewInvSqrt2);
low = vqrshrn_n_s32(out_low, (int32_t)NewSqrt2Bits);
high = vqrshrn_n_s32(out_high, (int32_t)NewSqrt2Bits);
output[z] = vcombine_s16(low, high);
}
}
static INLINE void idct16_low1_neon(int16x8_t *in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1;
int32x4_t t32[2];
// stage 4
t32[0] = vmull_n_s16(vget_low_s16(in[0]), cospi[32]);
t32[1] = vmull_n_s16(vget_high_s16(in[0]), cospi[32]);
step1 = vcombine_s16(vrshrn_n_s32(t32[0], INV_COS_BIT),
vrshrn_n_s32(t32[1], INV_COS_BIT));
// stage 6
// stage 7
out[0] = step1;
out[1] = step1;
out[2] = step1;
out[3] = step1;
out[4] = step1;
out[5] = step1;
out[6] = step1;
out[7] = step1;
out[8] = step1;
out[9] = step1;
out[10] = step1;
out[11] = step1;
out[12] = step1;
out[13] = step1;
out[14] = step1;
out[15] = step1;
}
static INLINE void idct16_neon(int16x8_t *in, int16x8_t *out, int8_t cos_bit,
int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1[16], step2[16];
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[4], (int16_t)cospi[60],
(int16_t)cospi[36], (int16_t)cospi[28]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[20], (int16_t)cospi[44],
(int16_t)cospi[52], (int16_t)cospi[12]);
const int16x4_t c2 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c3 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
const int16x4_t c4 =
set_s16x4_neon((int16_t)(-cospi[32]), (int16_t)(-cospi[32]),
(int16_t)(-cospi[16]), (int16_t)(-cospi[48]));
// stage 2
btf_16_lane_0_1_neon(in[1], in[15], c0, &step2[15], &step2[8]);
btf_16_lane_2_3_neon(in[9], in[7], c0, &step2[14], &step2[9]);
btf_16_lane_0_1_neon(in[5], in[11], c1, &step2[13], &step2[10]);
btf_16_lane_2_3_neon(in[13], in[3], c1, &step2[12], &step2[11]);
step2[0] = in[0];
step2[1] = in[8];
step2[2] = in[4];
step2[3] = in[12];
step2[4] = in[2];
step2[5] = in[10];
step2[6] = in[6];
step2[7] = in[14];
// stage 3
btf_16_lane_0_1_neon(step2[4], step2[7], c2, &step1[7], &step1[4]);
btf_16_lane_2_3_neon(step2[5], step2[6], c2, &step1[6], &step1[5]);
step1[0] = step2[0];
step1[1] = step2[1];
step1[2] = step2[2];
step1[3] = step2[3];
step1[8] = vqaddq_s16(step2[8], step2[9]);
step1[9] = vqsubq_s16(step2[8], step2[9]);
step1[10] = vqsubq_s16(step2[11], step2[10]);
step1[11] = vqaddq_s16(step2[11], step2[10]);
step1[12] = vqaddq_s16(step2[12], step2[13]);
step1[13] = vqsubq_s16(step2[12], step2[13]);
step1[14] = vqsubq_s16(step2[15], step2[14]);
step1[15] = vqaddq_s16(step2[15], step2[14]);
// stage 4
btf_16_lane_0_1_neon(step1[0], step1[1], c3, &step2[0], &step2[1]);
btf_16_lane_2_3_neon(step1[2], step1[3], c3, &step2[3], &step2[2]);
btf_16_lane_2_3_neon(step1[14], step1[9], c3, &step2[14], &step2[9]);
btf_16_lane_3_2_neon(step1[10], step1[13], c4, &step2[10], &step2[13]);
step2[4] = vqaddq_s16(step1[4], step1[5]);
step2[5] = vqsubq_s16(step1[4], step1[5]);
step2[6] = vqsubq_s16(step1[7], step1[6]);
step2[7] = vqaddq_s16(step1[7], step1[6]);
step2[8] = step1[8];
step2[11] = step1[11];
step2[12] = step1[12];
step2[15] = step1[15];
// stage 5
btf_16_lane_0_1_neon(step2[6], step2[5], c3, &step1[6], &step1[5]);
step1[0] = vqaddq_s16(step2[0], step2[3]);
step1[1] = vqaddq_s16(step2[1], step2[2]);
step1[2] = vqsubq_s16(step2[1], step2[2]);
step1[3] = vqsubq_s16(step2[0], step2[3]);
step1[4] = step2[4];
step1[7] = step2[7];
step1[8] = vqaddq_s16(step2[8], step2[11]);
step1[9] = vqaddq_s16(step2[9], step2[10]);
step1[10] = vqsubq_s16(step2[9], step2[10]);
step1[11] = vqsubq_s16(step2[8], step2[11]);
step1[12] = vqsubq_s16(step2[15], step2[12]);
step1[13] = vqsubq_s16(step2[14], step2[13]);
step1[14] = vqaddq_s16(step2[14], step2[13]);
step1[15] = vqaddq_s16(step2[15], step2[12]);
// stage 6
btf_16_lane_0_1_neon(step1[13], step1[10], c3, &step2[13], &step2[10]);
btf_16_lane_0_1_neon(step1[12], step1[11], c3, &step2[12], &step2[11]);
step2[0] = vqaddq_s16(step1[0], step1[7]);
step2[1] = vqaddq_s16(step1[1], step1[6]);
step2[2] = vqaddq_s16(step1[2], step1[5]);
step2[3] = vqaddq_s16(step1[3], step1[4]);
step2[4] = vqsubq_s16(step1[3], step1[4]);
step2[5] = vqsubq_s16(step1[2], step1[5]);
step2[6] = vqsubq_s16(step1[1], step1[6]);
step2[7] = vqsubq_s16(step1[0], step1[7]);
step2[8] = step1[8];
step2[9] = step1[9];
step2[14] = step1[14];
step2[15] = step1[15];
// stage 7
out[0] = vqaddq_s16(step2[0], step2[15]);
out[1] = vqaddq_s16(step2[1], step2[14]);
out[2] = vqaddq_s16(step2[2], step2[13]);
out[3] = vqaddq_s16(step2[3], step2[12]);
out[4] = vqaddq_s16(step2[4], step2[11]);
out[5] = vqaddq_s16(step2[5], step2[10]);
out[6] = vqaddq_s16(step2[6], step2[9]);
out[7] = vqaddq_s16(step2[7], step2[8]);
out[8] = vqsubq_s16(step2[7], step2[8]);
out[9] = vqsubq_s16(step2[6], step2[9]);
out[10] = vqsubq_s16(step2[5], step2[10]);
out[11] = vqsubq_s16(step2[4], step2[11]);
out[12] = vqsubq_s16(step2[3], step2[12]);
out[13] = vqsubq_s16(step2[2], step2[13]);
out[14] = vqsubq_s16(step2[1], step2[14]);
out[15] = vqsubq_s16(step2[0], step2[15]);
}
static INLINE void idct16_low8_neon(int16x8_t *in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1[16], step2[16];
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
const int16x4_t c1 =
set_s16x4_neon((int16_t)(-cospi[32]), (int16_t)(-cospi[32]),
(int16_t)(-cospi[16]), (int16_t)(-cospi[48]));
// stage 1
// stage 2
step2[0] = in[0];
step2[2] = in[4];
step2[4] = in[2];
step2[6] = in[6];
btf_16_neon(in[1], cospi[60], cospi[4], &step2[8], &step2[15]);
btf_16_neon(in[7], -cospi[36], cospi[28], &step2[9], &step2[14]);
btf_16_neon(in[5], cospi[44], cospi[20], &step2[10], &step2[13]);
btf_16_neon(in[3], -cospi[52], cospi[12], &step2[11], &step2[12]);
// stage 3
btf_16_neon(step2[4], cospi[56], cospi[8], &step1[4], &step1[7]);
btf_16_neon(step2[6], -cospi[40], cospi[24], &step1[5], &step1[6]);
step1[0] = step2[0];
step1[2] = step2[2];
step1[8] = vqaddq_s16(step2[8], step2[9]);
step1[9] = vqsubq_s16(step2[8], step2[9]);
step1[10] = vqsubq_s16(step2[11], step2[10]);
step1[11] = vqaddq_s16(step2[11], step2[10]);
step1[12] = vqaddq_s16(step2[12], step2[13]);
step1[13] = vqsubq_s16(step2[12], step2[13]);
step1[14] = vqsubq_s16(step2[15], step2[14]);
step1[15] = vqaddq_s16(step2[15], step2[14]);
// stage 4
btf_16_neon(step1[0], cospi[32], cospi[32], &step2[0], &step2[1]);
btf_16_neon(step1[2], cospi[48], cospi[16], &step2[2], &step2[3]);
btf_16_lane_2_3_neon(step1[14], step1[9], c0, &step2[14], &step2[9]);
btf_16_lane_3_2_neon(step1[10], step1[13], c1, &step2[10], &step2[13]);
step2[4] = vqaddq_s16(step1[4], step1[5]);
step2[5] = vqsubq_s16(step1[4], step1[5]);
step2[6] = vqsubq_s16(step1[7], step1[6]);
step2[7] = vqaddq_s16(step1[7], step1[6]);
step2[8] = step1[8];
step2[11] = step1[11];
step2[12] = step1[12];
step2[15] = step1[15];
// stage 5
btf_16_lane_0_1_neon(step2[6], step2[5], c0, &step1[6], &step1[5]);
step1[0] = vqaddq_s16(step2[0], step2[3]);
step1[1] = vqaddq_s16(step2[1], step2[2]);
step1[2] = vqsubq_s16(step2[1], step2[2]);
step1[3] = vqsubq_s16(step2[0], step2[3]);
step1[4] = step2[4];
step1[7] = step2[7];
step1[8] = vqaddq_s16(step2[8], step2[11]);
step1[9] = vqaddq_s16(step2[9], step2[10]);
step1[10] = vqsubq_s16(step2[9], step2[10]);
step1[11] = vqsubq_s16(step2[8], step2[11]);
step1[12] = vqsubq_s16(step2[15], step2[12]);
step1[13] = vqsubq_s16(step2[14], step2[13]);
step1[14] = vqaddq_s16(step2[14], step2[13]);
step1[15] = vqaddq_s16(step2[15], step2[12]);
// stage 6
btf_16_lane_0_1_neon(step1[13], step1[10], c0, &step2[13], &step2[10]);
btf_16_lane_0_1_neon(step1[12], step1[11], c0, &step2[12], &step2[11]);
step2[0] = vqaddq_s16(step1[0], step1[7]);
step2[1] = vqaddq_s16(step1[1], step1[6]);
step2[2] = vqaddq_s16(step1[2], step1[5]);
step2[3] = vqaddq_s16(step1[3], step1[4]);
step2[4] = vqsubq_s16(step1[3], step1[4]);
step2[5] = vqsubq_s16(step1[2], step1[5]);
step2[6] = vqsubq_s16(step1[1], step1[6]);
step2[7] = vqsubq_s16(step1[0], step1[7]);
step2[8] = step1[8];
step2[9] = step1[9];
step2[14] = step1[14];
step2[15] = step1[15];
// stage 7
out[0] = vqaddq_s16(step2[0], step2[15]);
out[1] = vqaddq_s16(step2[1], step2[14]);
out[2] = vqaddq_s16(step2[2], step2[13]);
out[3] = vqaddq_s16(step2[3], step2[12]);
out[4] = vqaddq_s16(step2[4], step2[11]);
out[5] = vqaddq_s16(step2[5], step2[10]);
out[6] = vqaddq_s16(step2[6], step2[9]);
out[7] = vqaddq_s16(step2[7], step2[8]);
out[8] = vqsubq_s16(step2[7], step2[8]);
out[9] = vqsubq_s16(step2[6], step2[9]);
out[10] = vqsubq_s16(step2[5], step2[10]);
out[11] = vqsubq_s16(step2[4], step2[11]);
out[12] = vqsubq_s16(step2[3], step2[12]);
out[13] = vqsubq_s16(step2[2], step2[13]);
out[14] = vqsubq_s16(step2[1], step2[14]);
out[15] = vqsubq_s16(step2[0], step2[15]);
}
static INLINE void iadst16_neon(int16x8_t *const in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[2], (int16_t)cospi[62],
(int16_t)cospi[10], (int16_t)cospi[54]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[18], (int16_t)cospi[46],
(int16_t)cospi[26], (int16_t)cospi[38]);
const int16x4_t c2 = set_s16x4_neon((int16_t)cospi[34], (int16_t)cospi[30],
(int16_t)cospi[42], (int16_t)cospi[22]);
const int16x4_t c3 = set_s16x4_neon((int16_t)cospi[50], (int16_t)cospi[14],
(int16_t)cospi[58], (int16_t)cospi[6]);
const int16x4_t c4 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c5 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
int16x8_t x[16];
int16x8_t t[14];
int16x8_t s0, s1, s2, s3, s4, s5, s6, s7;
int16x8_t s8, s9, s10, s11, s12, s13, s14, s15;
// Stage 1
x[0] = in[15];
x[1] = in[0];
x[2] = in[13];
x[3] = in[2];
x[4] = in[11];
x[5] = in[4];
x[6] = in[9];
x[7] = in[6];
x[8] = in[7];
x[9] = in[8];
x[10] = in[5];
x[11] = in[10];
x[12] = in[3];
x[13] = in[12];
x[14] = in[1];
x[15] = in[14];
// Stage 2
btf_16_lane_0_1_neon(x[0], x[1], c0, &s0, &s1);
btf_16_lane_2_3_neon(x[2], x[3], c0, &s2, &s3);
btf_16_lane_0_1_neon(x[4], x[5], c1, &s4, &s5);
btf_16_lane_2_3_neon(x[6], x[7], c1, &s6, &s7);
btf_16_lane_0_1_neon(x[8], x[9], c2, &s8, &s9);
btf_16_lane_2_3_neon(x[10], x[11], c2, &s10, &s11);
btf_16_lane_0_1_neon(x[12], x[13], c3, &s12, &s13);
btf_16_lane_2_3_neon(x[14], x[15], c3, &s14, &s15);
// Stage 3
x[0] = vqaddq_s16(s0, s8);
x[1] = vqaddq_s16(s1, s9);
x[2] = vqaddq_s16(s2, s10);
x[3] = vqaddq_s16(s3, s11);
x[4] = vqaddq_s16(s4, s12);
x[5] = vqaddq_s16(s5, s13);
x[6] = vqaddq_s16(s6, s14);
x[7] = vqaddq_s16(s7, s15);
x[8] = vqsubq_s16(s0, s8);
x[9] = vqsubq_s16(s1, s9);
x[10] = vqsubq_s16(s2, s10);
x[11] = vqsubq_s16(s3, s11);
x[12] = vqsubq_s16(s4, s12);
x[13] = vqsubq_s16(s5, s13);
x[14] = vqsubq_s16(s6, s14);
x[15] = vqsubq_s16(s7, s15);
// Stage 4
t[0] = x[0];
t[1] = x[1];
t[2] = x[2];
t[3] = x[3];
t[4] = x[4];
t[5] = x[5];
t[6] = x[6];
t[7] = x[7];
btf_16_lane_0_1_neon(x[8], x[9], c4, &s8, &s9);
btf_16_lane_2_3_neon(x[10], x[11], c4, &s10, &s11);
btf_16_lane_1_0_neon(x[13], x[12], c4, &s13, &s12);
btf_16_lane_3_2_neon(x[15], x[14], c4, &s15, &s14);
// Stage 5
x[0] = vqaddq_s16(t[0], t[4]);
x[1] = vqaddq_s16(t[1], t[5]);
x[2] = vqaddq_s16(t[2], t[6]);
x[3] = vqaddq_s16(t[3], t[7]);
x[4] = vqsubq_s16(t[0], t[4]);
x[5] = vqsubq_s16(t[1], t[5]);
x[6] = vqsubq_s16(t[2], t[6]);
x[7] = vqsubq_s16(t[3], t[7]);
x[8] = vqaddq_s16(s8, s12);
x[9] = vqaddq_s16(s9, s13);
x[10] = vqaddq_s16(s10, s14);
x[11] = vqaddq_s16(s11, s15);
x[12] = vqsubq_s16(s8, s12);
x[13] = vqsubq_s16(s9, s13);
x[14] = vqsubq_s16(s10, s14);
x[15] = vqsubq_s16(s11, s15);
// stage 6
t[0] = x[0];
t[1] = x[1];
t[2] = x[2];
t[3] = x[3];
btf_16_lane_2_3_neon(x[4], x[5], c5, &s4, &s5);
btf_16_lane_3_2_neon(x[7], x[6], c5, &s7, &s6);
t[8] = x[8];
t[9] = x[9];
t[10] = x[10];
t[11] = x[11];
btf_16_lane_2_3_neon(x[12], x[13], c5, &s12, &s13);
btf_16_lane_3_2_neon(x[15], x[14], c5, &s15, &s14);
// Stage 7
x[0] = vqaddq_s16(t[0], t[2]);
x[1] = vqaddq_s16(t[1], t[3]);
x[2] = vqsubq_s16(t[0], t[2]);
x[3] = vqsubq_s16(t[1], t[3]);
x[4] = vqaddq_s16(s4, s6);
x[5] = vqaddq_s16(s5, s7);
x[6] = vqsubq_s16(s4, s6);
x[7] = vqsubq_s16(s5, s7);
x[8] = vqaddq_s16(t[8], t[10]);
x[9] = vqaddq_s16(t[9], t[11]);
x[10] = vqsubq_s16(t[8], t[10]);
x[11] = vqsubq_s16(t[9], t[11]);
x[12] = vqaddq_s16(s12, s14);
x[13] = vqaddq_s16(s13, s15);
x[14] = vqsubq_s16(s12, s14);
x[15] = vqsubq_s16(s13, s15);
// Stage 8
btf_16_half_neon(x + 2, c5);
btf_16_half_neon(x + 6, c5);
btf_16_half_neon(x + 10, c5);
btf_16_half_neon(x + 14, c5);
// Stage 9
out[0] = x[0];
out[1] = vqnegq_s16(x[8]);
out[2] = x[12];
out[3] = vqnegq_s16(x[4]);
out[4] = x[6];
out[5] = vqnegq_s16(x[14]);
out[6] = x[10];
out[7] = vqnegq_s16(x[2]);
out[8] = x[3];
out[9] = vqnegq_s16(x[11]);
out[10] = x[15];
out[11] = vqnegq_s16(x[7]);
out[12] = x[5];
out[13] = vqnegq_s16(x[13]);
out[14] = x[9];
out[15] = vqnegq_s16(x[1]);
}
static INLINE void iadst16_low1_neon(int16x8_t *const in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
int16x8_t x[16];
int16x8_t t[10];
int16x8_t s0, s1, s4, s5;
int16x8_t s8, s9, s12, s13;
// Stage 1
x[1] = in[0];
// Stage 2
btf_16_neon(x[1], cospi[62], -cospi[2], &s0, &s1);
// Stage 3
x[0] = s0;
x[1] = s1;
x[8] = s0;
x[9] = s1;
// Stage 4
t[0] = x[0];
t[1] = x[1];
btf_16_lane_0_1_neon(x[8], x[9], c0, &s8, &s9);
// Stage 5
x[0] = t[0];
x[1] = t[1];
x[4] = t[0];
x[5] = t[1];
x[8] = s8;
x[9] = s9;
x[12] = s8;
x[13] = s9;
// stage 6
t[0] = x[0];
t[1] = x[1];
btf_16_lane_2_3_neon(x[4], x[5], c1, &s4, &s5);
t[8] = x[8];
t[9] = x[9];
btf_16_lane_2_3_neon(x[12], x[13], c1, &s12, &s13);
// Stage 7
x[0] = t[0];
x[1] = t[1];
x[2] = t[0];
x[3] = t[1];
x[4] = s4;
x[5] = s5;
x[6] = s4;
x[7] = s5;
x[8] = t[8];
x[9] = t[9];
x[10] = t[8];
x[11] = t[9];
x[12] = s12;
x[13] = s13;
x[14] = s12;
x[15] = s13;
// Stage 8
btf_16_half_neon(x + 2, c1);
btf_16_half_neon(x + 6, c1);
btf_16_half_neon(x + 10, c1);
btf_16_half_neon(x + 14, c1);
// Stage 9
out[0] = x[0];
out[1] = vqnegq_s16(x[8]);
out[2] = x[12];
out[3] = vqnegq_s16(x[4]);
out[4] = x[6];
out[5] = vqnegq_s16(x[14]);
out[6] = x[10];
out[7] = vqnegq_s16(x[2]);
out[8] = x[3];
out[9] = vqnegq_s16(x[11]);
out[10] = x[15];
out[11] = vqnegq_s16(x[7]);
out[12] = x[5];
out[13] = vqnegq_s16(x[13]);
out[14] = x[9];
out[15] = vqnegq_s16(x[1]);
}
static INLINE void iadst16_low8_neon(int16x8_t *const in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
int16x8_t x[16];
int16x8_t t[14];
int16x8_t s0, s1, s2, s3, s4, s5, s6, s7;
int16x8_t s8, s9, s10, s11, s12, s13, s14, s15;
// Stage 1
x[1] = in[0];
x[3] = in[2];
x[5] = in[4];
x[7] = in[6];
x[8] = in[7];
x[10] = in[5];
x[12] = in[3];
x[14] = in[1];
// Stage 2
btf_16_neon(x[1], cospi[62], -cospi[2], &s0, &s1);
btf_16_neon(x[3], cospi[54], -cospi[10], &s2, &s3);
btf_16_neon(x[5], cospi[46], -cospi[18], &s4, &s5);
btf_16_neon(x[7], cospi[38], -cospi[26], &s6, &s7);
btf_16_neon(x[8], cospi[34], cospi[30], &s8, &s9);
btf_16_neon(x[10], cospi[42], cospi[22], &s10, &s11);
btf_16_neon(x[12], cospi[50], cospi[14], &s12, &s13);
btf_16_neon(x[14], cospi[58], cospi[6], &s14, &s15);
// Stage 3
x[0] = vqaddq_s16(s0, s8);
x[1] = vqaddq_s16(s1, s9);
x[2] = vqaddq_s16(s2, s10);
x[3] = vqaddq_s16(s3, s11);
x[4] = vqaddq_s16(s4, s12);
x[5] = vqaddq_s16(s5, s13);
x[6] = vqaddq_s16(s6, s14);
x[7] = vqaddq_s16(s7, s15);
x[8] = vqsubq_s16(s0, s8);
x[9] = vqsubq_s16(s1, s9);
x[10] = vqsubq_s16(s2, s10);
x[11] = vqsubq_s16(s3, s11);
x[12] = vqsubq_s16(s4, s12);
x[13] = vqsubq_s16(s5, s13);
x[14] = vqsubq_s16(s6, s14);
x[15] = vqsubq_s16(s7, s15);
// Stage 4
t[0] = x[0];
t[1] = x[1];
t[2] = x[2];
t[3] = x[3];
t[4] = x[4];
t[5] = x[5];
t[6] = x[6];
t[7] = x[7];
btf_16_lane_0_1_neon(x[8], x[9], c0, &s8, &s9);
btf_16_lane_2_3_neon(x[10], x[11], c0, &s10, &s11);
btf_16_lane_1_0_neon(x[13], x[12], c0, &s13, &s12);
btf_16_lane_3_2_neon(x[15], x[14], c0, &s15, &s14);
// Stage 5
x[0] = vqaddq_s16(t[0], t[4]);
x[1] = vqaddq_s16(t[1], t[5]);
x[2] = vqaddq_s16(t[2], t[6]);
x[3] = vqaddq_s16(t[3], t[7]);
x[4] = vqsubq_s16(t[0], t[4]);
x[5] = vqsubq_s16(t[1], t[5]);
x[6] = vqsubq_s16(t[2], t[6]);
x[7] = vqsubq_s16(t[3], t[7]);
x[8] = vqaddq_s16(s8, s12);
x[9] = vqaddq_s16(s9, s13);
x[10] = vqaddq_s16(s10, s14);
x[11] = vqaddq_s16(s11, s15);
x[12] = vqsubq_s16(s8, s12);
x[13] = vqsubq_s16(s9, s13);
x[14] = vqsubq_s16(s10, s14);
x[15] = vqsubq_s16(s11, s15);
// stage 6
t[0] = x[0];
t[1] = x[1];
t[2] = x[2];
t[3] = x[3];
btf_16_lane_2_3_neon(x[4], x[5], c1, &s4, &s5);
btf_16_lane_3_2_neon(x[7], x[6], c1, &s7, &s6);
t[8] = x[8];
t[9] = x[9];
t[10] = x[10];
t[11] = x[11];
btf_16_lane_2_3_neon(x[12], x[13], c1, &s12, &s13);
btf_16_lane_3_2_neon(x[15], x[14], c1, &s15, &s14);
// Stage 7
x[0] = vqaddq_s16(t[0], t[2]);
x[1] = vqaddq_s16(t[1], t[3]);
x[2] = vqsubq_s16(t[0], t[2]);
x[3] = vqsubq_s16(t[1], t[3]);
x[4] = vqaddq_s16(s4, s6);
x[5] = vqaddq_s16(s5, s7);
x[6] = vqsubq_s16(s4, s6);
x[7] = vqsubq_s16(s5, s7);
x[8] = vqaddq_s16(t[8], t[10]);
x[9] = vqaddq_s16(t[9], t[11]);
x[10] = vqsubq_s16(t[8], t[10]);
x[11] = vqsubq_s16(t[9], t[11]);
x[12] = vqaddq_s16(s12, s14);
x[13] = vqaddq_s16(s13, s15);
x[14] = vqsubq_s16(s12, s14);
x[15] = vqsubq_s16(s13, s15);
// Stage 8
btf_16_half_neon(x + 2, c1);
btf_16_half_neon(x + 6, c1);
btf_16_half_neon(x + 10, c1);
btf_16_half_neon(x + 14, c1);
// Stage 9
out[0] = x[0];
out[1] = vqnegq_s16(x[8]);
out[2] = x[12];
out[3] = vqnegq_s16(x[4]);
out[4] = x[6];
out[5] = vqnegq_s16(x[14]);
out[6] = x[10];
out[7] = vqnegq_s16(x[2]);
out[8] = x[3];
out[9] = vqnegq_s16(x[11]);
out[10] = x[15];
out[11] = vqnegq_s16(x[7]);
out[12] = x[5];
out[13] = vqnegq_s16(x[13]);
out[14] = x[9];
out[15] = vqnegq_s16(x[1]);
}
static INLINE void idct32_neon(int16x8_t *in, int16x8_t *out, int8_t cos_bit,
int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1[32], step2[32];
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[2], (int16_t)cospi[62],
(int16_t)cospi[34], (int16_t)cospi[30]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[18], (int16_t)cospi[46],
(int16_t)cospi[50], (int16_t)cospi[14]);
const int16x4_t c2 = set_s16x4_neon((int16_t)cospi[10], (int16_t)cospi[54],
(int16_t)cospi[42], (int16_t)cospi[22]);
const int16x4_t c3 = set_s16x4_neon((int16_t)cospi[26], (int16_t)cospi[38],
(int16_t)cospi[58], (int16_t)cospi[6]);
const int16x4_t c4 = set_s16x4_neon((int16_t)cospi[4], (int16_t)cospi[60],
(int16_t)cospi[36], (int16_t)cospi[28]);
const int16x4_t c5 = set_s16x4_neon((int16_t)cospi[20], (int16_t)cospi[44],
(int16_t)cospi[52], (int16_t)cospi[12]);
const int16x4_t c6 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c7 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
const int16x4_t c8 =
set_s16x4_neon((int16_t)(-cospi[8]), (int16_t)(-cospi[56]),
(int16_t)(-cospi[40]), (int16_t)(-cospi[24]));
const int16x4_t c9 =
set_s16x4_neon((int16_t)(-cospi[32]), (int16_t)(-cospi[32]),
(int16_t)(-cospi[16]), (int16_t)(-cospi[48]));
// stage 2
btf_16_lane_0_1_neon(in[1], in[31], c0, &step2[31], &step2[16]);
btf_16_lane_2_3_neon(in[17], in[15], c0, &step2[30], &step2[17]);
btf_16_lane_0_1_neon(in[9], in[23], c1, &step2[29], &step2[18]);
btf_16_lane_2_3_neon(in[25], in[7], c1, &step2[28], &step2[19]);
btf_16_lane_0_1_neon(in[5], in[27], c2, &step2[27], &step2[20]);
btf_16_lane_2_3_neon(in[21], in[11], c2, &step2[26], &step2[21]);
btf_16_lane_0_1_neon(in[13], in[19], c3, &step2[25], &step2[22]);
btf_16_lane_2_3_neon(in[29], in[3], c3, &step2[24], &step2[23]);
step2[0] = in[0];
step2[1] = in[16];
step2[2] = in[8];
step2[3] = in[24];
step2[4] = in[4];
step2[5] = in[20];
step2[6] = in[12];
step2[7] = in[28];
step2[8] = in[2];
step2[9] = in[18];
step2[10] = in[10];
step2[11] = in[26];
step2[12] = in[6];
step2[13] = in[22];
step2[14] = in[14];
step2[15] = in[30];
// stage 3
btf_16_lane_0_1_neon(step2[8], step2[15], c4, &step1[15], &step1[8]);
btf_16_lane_2_3_neon(step2[9], step2[14], c4, &step1[14], &step1[9]);
btf_16_lane_0_1_neon(step2[10], step2[13], c5, &step1[13], &step1[10]);
btf_16_lane_2_3_neon(step2[11], step2[12], c5, &step1[12], &step1[11]);
step1[0] = step2[0];
step1[1] = step2[1];
step1[2] = step2[2];
step1[3] = step2[3];
step1[4] = step2[4];
step1[5] = step2[5];
step1[6] = step2[6];
step1[7] = step2[7];
step1[16] = vqaddq_s16(step2[16], step2[17]);
step1[17] = vqsubq_s16(step2[16], step2[17]);
step1[18] = vqsubq_s16(step2[19], step2[18]);
step1[19] = vqaddq_s16(step2[19], step2[18]);
step1[20] = vqaddq_s16(step2[20], step2[21]);
step1[21] = vqsubq_s16(step2[20], step2[21]);
step1[22] = vqsubq_s16(step2[23], step2[22]);
step1[23] = vqaddq_s16(step2[23], step2[22]);
step1[24] = vqaddq_s16(step2[24], step2[25]);
step1[25] = vqsubq_s16(step2[24], step2[25]);
step1[26] = vqsubq_s16(step2[27], step2[26]);
step1[27] = vqaddq_s16(step2[27], step2[26]);
step1[28] = vqaddq_s16(step2[28], step2[29]);
step1[29] = vqsubq_s16(step2[28], step2[29]);
step1[30] = vqsubq_s16(step2[31], step2[30]);
step1[31] = vqaddq_s16(step2[31], step2[30]);
// stage 4
btf_16_lane_0_1_neon(step1[4], step1[7], c6, &step2[7], &step2[4]);
btf_16_lane_2_3_neon(step1[5], step1[6], c6, &step2[6], &step2[5]);
btf_16_lane_0_1_neon(step1[30], step1[17], c6, &step2[30], &step2[17]);
btf_16_lane_1_0_neon(step1[18], step1[29], c8, &step2[18], &step2[29]);
btf_16_lane_2_3_neon(step1[26], step1[21], c6, &step2[26], &step2[21]);
btf_16_lane_3_2_neon(step1[22], step1[25], c8, &step2[22], &step2[25]);
step2[0] = step1[0];
step2[1] = step1[1];
step2[2] = step1[2];
step2[3] = step1[3];
step2[8] = vqaddq_s16(step1[8], step1[9]);
step2[9] = vqsubq_s16(step1[8], step1[9]);
step2[10] = vqsubq_s16(step1[11], step1[10]);
step2[11] = vqaddq_s16(step1[11], step1[10]);
step2[12] = vqaddq_s16(step1[12], step1[13]);
step2[13] = vqsubq_s16(step1[12], step1[13]);
step2[14] = vqsubq_s16(step1[15], step1[14]);
step2[15] = vqaddq_s16(step1[15], step1[14]);
step2[16] = step1[16];
step2[19] = step1[19];
step2[20] = step1[20];
step2[23] = step1[23];
step2[24] = step1[24];
step2[27] = step1[27];
step2[28] = step1[28];
step2[31] = step1[31];
// stage 5
btf_16_lane_0_1_neon(step2[0], step2[1], c7, &step1[0], &step1[1]);
btf_16_lane_2_3_neon(step2[2], step2[3], c7, &step1[3], &step1[2]);
btf_16_lane_2_3_neon(step2[14], step2[9], c7, &step1[14], &step1[9]);
btf_16_lane_3_2_neon(step2[10], step2[13], c9, &step1[10], &step1[13]);
step1[4] = vqaddq_s16(step2[4], step2[5]);
step1[5] = vqsubq_s16(step2[4], step2[5]);
step1[6] = vqsubq_s16(step2[7], step2[6]);
step1[7] = vqaddq_s16(step2[7], step2[6]);
step1[8] = step2[8];
step1[11] = step2[11];
step1[12] = step2[12];
step1[15] = step2[15];
step1[16] = vqaddq_s16(step2[16], step2[19]);
step1[17] = vqaddq_s16(step2[17], step2[18]);
step1[18] = vqsubq_s16(step2[17], step2[18]);
step1[19] = vqsubq_s16(step2[16], step2[19]);
step1[20] = vqsubq_s16(step2[23], step2[20]);
step1[21] = vqsubq_s16(step2[22], step2[21]);
step1[22] = vqaddq_s16(step2[22], step2[21]);
step1[23] = vqaddq_s16(step2[23], step2[20]);
step1[24] = vqaddq_s16(step2[24], step2[27]);
step1[25] = vqaddq_s16(step2[25], step2[26]);
step1[26] = vqsubq_s16(step2[25], step2[26]);
step1[27] = vqsubq_s16(step2[24], step2[27]);
step1[28] = vqsubq_s16(step2[31], step2[28]);
step1[29] = vqsubq_s16(step2[30], step2[29]);
step1[30] = vqaddq_s16(step2[30], step2[29]);
step1[31] = vqaddq_s16(step2[31], step2[28]);
// stage 6
btf_16_lane_0_1_neon(step1[6], step1[5], c7, &step2[6], &step2[5]);
btf_16_lane_2_3_neon(step1[29], step1[18], c7, &step2[29], &step2[18]);
btf_16_lane_2_3_neon(step1[28], step1[19], c7, &step2[28], &step2[19]);
btf_16_lane_3_2_neon(step1[20], step1[27], c9, &step2[20], &step2[27]);
btf_16_lane_3_2_neon(step1[21], step1[26], c9, &step2[21], &step2[26]);
step2[0] = vqaddq_s16(step1[0], step1[3]);
step2[1] = vqaddq_s16(step1[1], step1[2]);
step2[2] = vqsubq_s16(step1[1], step1[2]);
step2[3] = vqsubq_s16(step1[0], step1[3]);
step2[4] = step1[4];
step2[7] = step1[7];
step2[8] = vqaddq_s16(step1[8], step1[11]);
step2[9] = vqaddq_s16(step1[9], step1[10]);
step2[10] = vqsubq_s16(step1[9], step1[10]);
step2[11] = vqsubq_s16(step1[8], step1[11]);
step2[12] = vqsubq_s16(step1[15], step1[12]);
step2[13] = vqsubq_s16(step1[14], step1[13]);
step2[14] = vqaddq_s16(step1[14], step1[13]);
step2[15] = vqaddq_s16(step1[15], step1[12]);
step2[16] = step1[16];
step2[17] = step1[17];
step2[22] = step1[22];
step2[23] = step1[23];
step2[24] = step1[24];
step2[25] = step1[25];
step2[30] = step1[30];
step2[31] = step1[31];
// stage 7
btf_16_lane_0_1_neon(step2[13], step2[10], c7, &step1[13], &step1[10]);
btf_16_lane_0_1_neon(step2[12], step2[11], c7, &step1[12], &step1[11]);
step1[0] = vqaddq_s16(step2[0], step2[7]);
step1[1] = vqaddq_s16(step2[1], step2[6]);
step1[2] = vqaddq_s16(step2[2], step2[5]);
step1[3] = vqaddq_s16(step2[3], step2[4]);
step1[4] = vqsubq_s16(step2[3], step2[4]);
step1[5] = vqsubq_s16(step2[2], step2[5]);
step1[6] = vqsubq_s16(step2[1], step2[6]);
step1[7] = vqsubq_s16(step2[0], step2[7]);
step1[8] = step2[8];
step1[9] = step2[9];
step1[14] = step2[14];
step1[15] = step2[15];
step1[16] = vqaddq_s16(step2[16], step2[23]);
step1[17] = vqaddq_s16(step2[17], step2[22]);
step1[18] = vqaddq_s16(step2[18], step2[21]);
step1[19] = vqaddq_s16(step2[19], step2[20]);
step1[20] = vqsubq_s16(step2[19], step2[20]);
step1[21] = vqsubq_s16(step2[18], step2[21]);
step1[22] = vqsubq_s16(step2[17], step2[22]);
step1[23] = vqsubq_s16(step2[16], step2[23]);
step1[24] = vqsubq_s16(step2[31], step2[24]);
step1[25] = vqsubq_s16(step2[30], step2[25]);
step1[26] = vqsubq_s16(step2[29], step2[26]);
step1[27] = vqsubq_s16(step2[28], step2[27]);
step1[28] = vqaddq_s16(step2[27], step2[28]);
step1[29] = vqaddq_s16(step2[26], step2[29]);
step1[30] = vqaddq_s16(step2[25], step2[30]);
step1[31] = vqaddq_s16(step2[24], step2[31]);
// stage 8
btf_16_lane_0_1_neon(step1[27], step1[20], c7, &step2[27], &step2[20]);
btf_16_lane_0_1_neon(step1[26], step1[21], c7, &step2[26], &step2[21]);
btf_16_lane_0_1_neon(step1[25], step1[22], c7, &step2[25], &step2[22]);
btf_16_lane_0_1_neon(step1[24], step1[23], c7, &step2[24], &step2[23]);
step2[0] = vqaddq_s16(step1[0], step1[15]);
step2[1] = vqaddq_s16(step1[1], step1[14]);
step2[2] = vqaddq_s16(step1[2], step1[13]);
step2[3] = vqaddq_s16(step1[3], step1[12]);
step2[4] = vqaddq_s16(step1[4], step1[11]);
step2[5] = vqaddq_s16(step1[5], step1[10]);
step2[6] = vqaddq_s16(step1[6], step1[9]);
step2[7] = vqaddq_s16(step1[7], step1[8]);
step2[8] = vqsubq_s16(step1[7], step1[8]);
step2[9] = vqsubq_s16(step1[6], step1[9]);
step2[10] = vqsubq_s16(step1[5], step1[10]);
step2[11] = vqsubq_s16(step1[4], step1[11]);
step2[12] = vqsubq_s16(step1[3], step1[12]);
step2[13] = vqsubq_s16(step1[2], step1[13]);
step2[14] = vqsubq_s16(step1[1], step1[14]);
step2[15] = vqsubq_s16(step1[0], step1[15]);
step2[16] = step1[16];
step2[17] = step1[17];
step2[18] = step1[18];
step2[19] = step1[19];
step2[28] = step1[28];
step2[29] = step1[29];
step2[30] = step1[30];
step2[31] = step1[31];
// stage 9
out[0] = vqaddq_s16(step2[0], step2[31]);
out[1] = vqaddq_s16(step2[1], step2[30]);
out[2] = vqaddq_s16(step2[2], step2[29]);
out[3] = vqaddq_s16(step2[3], step2[28]);
out[4] = vqaddq_s16(step2[4], step2[27]);
out[5] = vqaddq_s16(step2[5], step2[26]);
out[6] = vqaddq_s16(step2[6], step2[25]);
out[7] = vqaddq_s16(step2[7], step2[24]);
out[8] = vqaddq_s16(step2[8], step2[23]);
out[9] = vqaddq_s16(step2[9], step2[22]);
out[10] = vqaddq_s16(step2[10], step2[21]);
out[11] = vqaddq_s16(step2[11], step2[20]);
out[12] = vqaddq_s16(step2[12], step2[19]);
out[13] = vqaddq_s16(step2[13], step2[18]);
out[14] = vqaddq_s16(step2[14], step2[17]);
out[15] = vqaddq_s16(step2[15], step2[16]);
out[16] = vqsubq_s16(step2[15], step2[16]);
out[17] = vqsubq_s16(step2[14], step2[17]);
out[18] = vqsubq_s16(step2[13], step2[18]);
out[19] = vqsubq_s16(step2[12], step2[19]);
out[20] = vqsubq_s16(step2[11], step2[20]);
out[21] = vqsubq_s16(step2[10], step2[21]);
out[22] = vqsubq_s16(step2[9], step2[22]);
out[23] = vqsubq_s16(step2[8], step2[23]);
out[24] = vqsubq_s16(step2[7], step2[24]);
out[25] = vqsubq_s16(step2[6], step2[25]);
out[26] = vqsubq_s16(step2[5], step2[26]);
out[27] = vqsubq_s16(step2[4], step2[27]);
out[28] = vqsubq_s16(step2[3], step2[28]);
out[29] = vqsubq_s16(step2[2], step2[29]);
out[30] = vqsubq_s16(step2[1], step2[30]);
out[31] = vqsubq_s16(step2[0], step2[31]);
}
static INLINE void idct32_low1_neon(int16x8_t *in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1;
int32x4_t t32[2];
// stage 1
// stage 2
// stage 3
// stage 4
// stage 5
t32[0] = vmull_n_s16(vget_low_s16(in[0]), cospi[32]);
t32[1] = vmull_n_s16(vget_high_s16(in[0]), cospi[32]);
step1 = vcombine_s16(vrshrn_n_s32(t32[0], INV_COS_BIT),
vrshrn_n_s32(t32[1], INV_COS_BIT));
// stage 6
// stage 7
// stage 8
// stage 9
out[0] = step1;
out[1] = step1;
out[2] = step1;
out[3] = step1;
out[4] = step1;
out[5] = step1;
out[6] = step1;
out[7] = step1;
out[8] = step1;
out[9] = step1;
out[10] = step1;
out[11] = step1;
out[12] = step1;
out[13] = step1;
out[14] = step1;
out[15] = step1;
out[16] = step1;
out[17] = step1;
out[18] = step1;
out[19] = step1;
out[20] = step1;
out[21] = step1;
out[22] = step1;
out[23] = step1;
out[24] = step1;
out[25] = step1;
out[26] = step1;
out[27] = step1;
out[28] = step1;
out[29] = step1;
out[30] = step1;
out[31] = step1;
}
static INLINE void idct32_low8_neon(int16x8_t *in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1[32], step2[32];
int32x4_t t32[16];
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], cospi[48]);
const int16x4_t c2 =
set_s16x4_neon((int16_t)(-cospi[8]), (int16_t)(-cospi[56]),
(int16_t)(-cospi[40]), (int16_t)(-cospi[24]));
const int16x4_t c3 =
set_s16x4_neon((int16_t)(-cospi[32]), (int16_t)(-cospi[32]),
(int16_t)(-cospi[16]), (int16_t)(-cospi[48]));
// stage 1
// stage 2
step2[0] = in[0];
step2[4] = in[4];
step2[8] = in[2];
step2[12] = in[6];
btf_16_neon(in[1], cospi[62], cospi[2], &step2[16], &step2[31]);
btf_16_neon(in[7], -cospi[50], cospi[14], &step2[19], &step2[28]);
btf_16_neon(in[5], cospi[54], cospi[10], &step2[20], &step2[27]);
btf_16_neon(in[3], -cospi[58], cospi[6], &step2[23], &step2[24]);
// stage 3
step1[0] = step2[0];
step1[4] = step2[4];
btf_16_neon(step2[8], cospi[60], cospi[4], &step1[8], &step1[15]);
btf_16_neon(step2[12], -cospi[52], cospi[12], &step1[11], &step1[12]);
step1[16] = step2[16];
step1[17] = step2[16];
step1[18] = step2[19];
step1[19] = step2[19];
step1[20] = step2[20];
step1[21] = step2[20];
step1[22] = step2[23];
step1[23] = step2[23];
step1[24] = step2[24];
step1[25] = step2[24];
step1[26] = step2[27];
step1[27] = step2[27];
step1[28] = step2[28];
step1[29] = step2[28];
step1[30] = step2[31];
step1[31] = step2[31];
// stage 4
btf_16_neon(step1[4], cospi[56], cospi[8], &step2[4], &step2[7]);
btf_16_lane_0_1_neon(step1[30], step1[17], c0, &step2[30], &step2[17]);
btf_16_lane_1_0_neon(step1[18], step1[29], c2, &step2[18], &step2[29]);
btf_16_lane_2_3_neon(step1[26], step1[21], c0, &step2[26], &step2[21]);
btf_16_lane_3_2_neon(step1[22], step1[25], c2, &step2[22], &step2[25]);
step2[0] = step1[0];
step2[8] = step1[8];
step2[9] = step1[8];
step2[10] = step1[11];
step2[11] = step1[11];
step2[12] = step1[12];
step2[13] = step1[12];
step2[14] = step1[15];
step2[15] = step1[15];
step2[16] = step1[16];
step2[19] = step1[19];
step2[20] = step1[20];
step2[23] = step1[23];
step2[24] = step1[24];
step2[27] = step1[27];
step2[28] = step1[28];
step2[31] = step1[31];
// stage 5
t32[0] = vmull_n_s16(vget_low_s16(step2[0]), cospi[32]);
t32[1] = vmull_n_s16(vget_high_s16(step2[0]), cospi[32]);
step1[0] = vcombine_s16(vrshrn_n_s32(t32[0], INV_COS_BIT),
vrshrn_n_s32(t32[1], INV_COS_BIT));
btf_16_lane_2_3_neon(step2[14], step2[9], c1, &step1[14], &step1[9]);
btf_16_lane_3_2_neon(step2[10], step2[13], c3, &step1[10], &step1[13]);
step1[4] = step2[4];
step1[5] = step2[4];
step1[6] = step2[7];
step1[7] = step2[7];
step1[8] = step2[8];
step1[11] = step2[11];
step1[12] = step2[12];
step1[15] = step2[15];
step1[16] = vqaddq_s16(step2[16], step2[19]);
step1[17] = vqaddq_s16(step2[17], step2[18]);
step1[18] = vqsubq_s16(step2[17], step2[18]);
step1[19] = vqsubq_s16(step2[16], step2[19]);
step1[20] = vqsubq_s16(step2[23], step2[20]);
step1[21] = vqsubq_s16(step2[22], step2[21]);
step1[22] = vqaddq_s16(step2[22], step2[21]);
step1[23] = vqaddq_s16(step2[23], step2[20]);
step1[24] = vqaddq_s16(step2[24], step2[27]);
step1[25] = vqaddq_s16(step2[25], step2[26]);
step1[26] = vqsubq_s16(step2[25], step2[26]);
step1[27] = vqsubq_s16(step2[24], step2[27]);
step1[28] = vqsubq_s16(step2[31], step2[28]);
step1[29] = vqsubq_s16(step2[30], step2[29]);
step1[30] = vqaddq_s16(step2[30], step2[29]);
step1[31] = vqaddq_s16(step2[31], step2[28]);
// stage 6
btf_16_lane_0_1_neon(step1[6], step1[5], c1, &step2[6], &step2[5]);
btf_16_lane_2_3_neon(step1[29], step1[18], c1, &step2[29], &step2[18]);
btf_16_lane_2_3_neon(step1[28], step1[19], c1, &step2[28], &step2[19]);
btf_16_lane_3_2_neon(step1[20], step1[27], c3, &step2[20], &step2[27]);
btf_16_lane_3_2_neon(step1[21], step1[26], c3, &step2[21], &step2[26]);
step2[0] = step1[0];
step2[1] = step1[0];
step2[2] = step1[0];
step2[3] = step1[0];
step2[4] = step1[4];
step2[7] = step1[7];
step2[8] = vqaddq_s16(step1[8], step1[11]);
step2[9] = vqaddq_s16(step1[9], step1[10]);
step2[10] = vqsubq_s16(step1[9], step1[10]);
step2[11] = vqsubq_s16(step1[8], step1[11]);
step2[12] = vqsubq_s16(step1[15], step1[12]);
step2[13] = vqsubq_s16(step1[14], step1[13]);
step2[14] = vqaddq_s16(step1[14], step1[13]);
step2[15] = vqaddq_s16(step1[15], step1[12]);
step2[16] = step1[16];
step2[17] = step1[17];
step2[22] = step1[22];
step2[23] = step1[23];
step2[24] = step1[24];
step2[25] = step1[25];
step2[30] = step1[30];
step2[31] = step1[31];
// stage 7
btf_16_lane_0_1_neon(step2[13], step2[10], c1, &step1[13], &step1[10]);
btf_16_lane_0_1_neon(step2[12], step2[11], c1, &step1[12], &step1[11]);
step1[0] = vqaddq_s16(step2[0], step2[7]);
step1[1] = vqaddq_s16(step2[1], step2[6]);
step1[2] = vqaddq_s16(step2[2], step2[5]);
step1[3] = vqaddq_s16(step2[3], step2[4]);
step1[4] = vqsubq_s16(step2[3], step2[4]);
step1[5] = vqsubq_s16(step2[2], step2[5]);
step1[6] = vqsubq_s16(step2[1], step2[6]);
step1[7] = vqsubq_s16(step2[0], step2[7]);
step1[8] = step2[8];
step1[9] = step2[9];
step1[14] = step2[14];
step1[15] = step2[15];
step1[16] = vqaddq_s16(step2[16], step2[23]);
step1[17] = vqaddq_s16(step2[17], step2[22]);
step1[18] = vqaddq_s16(step2[18], step2[21]);
step1[19] = vqaddq_s16(step2[19], step2[20]);
step1[20] = vqsubq_s16(step2[19], step2[20]);
step1[21] = vqsubq_s16(step2[18], step2[21]);
step1[22] = vqsubq_s16(step2[17], step2[22]);
step1[23] = vqsubq_s16(step2[16], step2[23]);
step1[24] = vqsubq_s16(step2[31], step2[24]);
step1[25] = vqsubq_s16(step2[30], step2[25]);
step1[26] = vqsubq_s16(step2[29], step2[26]);
step1[27] = vqsubq_s16(step2[28], step2[27]);
step1[28] = vqaddq_s16(step2[27], step2[28]);
step1[29] = vqaddq_s16(step2[26], step2[29]);
step1[30] = vqaddq_s16(step2[25], step2[30]);
step1[31] = vqaddq_s16(step2[24], step2[31]);
// stage 8
btf_16_lane_0_1_neon(step1[27], step1[20], c1, &step2[27], &step2[20]);
btf_16_lane_0_1_neon(step1[26], step1[21], c1, &step2[26], &step2[21]);
btf_16_lane_0_1_neon(step1[25], step1[22], c1, &step2[25], &step2[22]);
btf_16_lane_0_1_neon(step1[24], step1[23], c1, &step2[24], &step2[23]);
step2[0] = vqaddq_s16(step1[0], step1[15]);
step2[1] = vqaddq_s16(step1[1], step1[14]);
step2[2] = vqaddq_s16(step1[2], step1[13]);
step2[3] = vqaddq_s16(step1[3], step1[12]);
step2[4] = vqaddq_s16(step1[4], step1[11]);
step2[5] = vqaddq_s16(step1[5], step1[10]);
step2[6] = vqaddq_s16(step1[6], step1[9]);
step2[7] = vqaddq_s16(step1[7], step1[8]);
step2[8] = vqsubq_s16(step1[7], step1[8]);
step2[9] = vqsubq_s16(step1[6], step1[9]);
step2[10] = vqsubq_s16(step1[5], step1[10]);
step2[11] = vqsubq_s16(step1[4], step1[11]);
step2[12] = vqsubq_s16(step1[3], step1[12]);
step2[13] = vqsubq_s16(step1[2], step1[13]);
step2[14] = vqsubq_s16(step1[1], step1[14]);
step2[15] = vqsubq_s16(step1[0], step1[15]);
step2[16] = step1[16];
step2[17] = step1[17];
step2[18] = step1[18];
step2[19] = step1[19];
step2[28] = step1[28];
step2[29] = step1[29];
step2[30] = step1[30];
step2[31] = step1[31];
// stage 9
out[0] = vqaddq_s16(step2[0], step2[31]);
out[1] = vqaddq_s16(step2[1], step2[30]);
out[2] = vqaddq_s16(step2[2], step2[29]);
out[3] = vqaddq_s16(step2[3], step2[28]);
out[4] = vqaddq_s16(step2[4], step2[27]);
out[5] = vqaddq_s16(step2[5], step2[26]);
out[6] = vqaddq_s16(step2[6], step2[25]);
out[7] = vqaddq_s16(step2[7], step2[24]);
out[8] = vqaddq_s16(step2[8], step2[23]);
out[9] = vqaddq_s16(step2[9], step2[22]);
out[10] = vqaddq_s16(step2[10], step2[21]);
out[11] = vqaddq_s16(step2[11], step2[20]);
out[12] = vqaddq_s16(step2[12], step2[19]);
out[13] = vqaddq_s16(step2[13], step2[18]);
out[14] = vqaddq_s16(step2[14], step2[17]);
out[15] = vqaddq_s16(step2[15], step2[16]);
out[16] = vqsubq_s16(step2[15], step2[16]);
out[17] = vqsubq_s16(step2[14], step2[17]);
out[18] = vqsubq_s16(step2[13], step2[18]);
out[19] = vqsubq_s16(step2[12], step2[19]);
out[20] = vqsubq_s16(step2[11], step2[20]);
out[21] = vqsubq_s16(step2[10], step2[21]);
out[22] = vqsubq_s16(step2[9], step2[22]);
out[23] = vqsubq_s16(step2[8], step2[23]);
out[24] = vqsubq_s16(step2[7], step2[24]);
out[25] = vqsubq_s16(step2[6], step2[25]);
out[26] = vqsubq_s16(step2[5], step2[26]);
out[27] = vqsubq_s16(step2[4], step2[27]);
out[28] = vqsubq_s16(step2[3], step2[28]);
out[29] = vqsubq_s16(step2[2], step2[29]);
out[30] = vqsubq_s16(step2[1], step2[30]);
out[31] = vqsubq_s16(step2[0], step2[31]);
}
static INLINE void idct32_low16_neon(int16x8_t *in, int16x8_t *out,
int8_t cos_bit, int bit) {
(void)bit;
const int32_t *cospi = cospi_arr(cos_bit);
int16x8_t step1[32], step2[32];
int32x4_t t32[16];
const int16x4_t c0 = set_s16x4_neon((int16_t)cospi[8], (int16_t)cospi[56],
(int16_t)cospi[40], (int16_t)cospi[24]);
const int16x4_t c1 = set_s16x4_neon((int16_t)cospi[32], (int16_t)cospi[32],
(int16_t)cospi[16], (int16_t)cospi[48]);
const int16x4_t c2 =
set_s16x4_neon((int16_t)(-cospi[8]), (int16_t)(-cospi[56]),
(int16_t)(-cospi[40]), (int16_t)(-cospi[24]));
const int16x4_t c3 =
set_s16x4_neon((int16_t)(-cospi[32]), (int16_t)(-cospi[32]),
(int16_t)(-cospi[16]), (int16_t)(-cospi[48]));
// stage 1
// stage 2
btf_16_neon(in[1], cospi[62], cospi[2], &step2[16], &step2[31]);
btf_16_neon(in[15], -cospi[34], cospi[30], &step2[17], &step2[30]);
btf_16_neon(in[9], cospi[46], cospi[18], &step2[18], &step2[29]);
btf_16_neon(in[7], -cospi[50], cospi[14], &step2[19], &step2[28]);
btf_16_neon(in[5], cospi[54], cospi[10], &step2[20], &step2[27]);
btf_16_neon(in[11], -cospi[42], cospi[22], &step2[21], &step2[26]);
btf_16_neon(in[13], cospi[38], cospi[26], &step2[22], &step2[25]);
btf_16_neon(in[3], -cospi[58], cospi[6], &step2[23], &step2[24]);
step2[0] = in[0];
step2[2] = in[8];
step2[4] = in[4];
step2[6] = in[12];
step2[8] = in[2];
step2[10] = in[10];
step2[12] = in[6];
step2[14] = in[14];
// stage 3
btf_16_neon(step2[8], cospi[60], cospi[4], &step1[8], &step1[15]);
btf_16_neon(step2[14], -cospi[36], cospi[28], &step1[9], &step1[14]);
btf_16_neon(step2[10], cospi[44], cospi[20], &step1[10], &step1[13]);
btf_16_neon(step2[12], -cospi[52], cospi[12], &step1[11], &step1[12]);
step1[0] = step2[0];
step1[2] = step2[2];
step1[4] = step2[4];
step1[6] = step2[6];
step1[16] = vqaddq_s16(step2[16], step2[17]);
step1[17] = vqsubq_s16(step2[16], step2[17]);
step1[18] = vqsubq_s16(step2[19], step2[18]);
step1[19] = vqaddq_s16(step2[19], step2[18]);
step1[20] = vqaddq_s16(step2[20], step2[21]);
step1[21] = vqsubq_s16(step2[20], step2[21]);
step1[22] = vqsubq_s16(step2[23], step2[22]);
step1[23] = vqaddq_s16(step2[23], step2[22]);
step1[24] = vqaddq_s16(step2[24], step2[25]);
step1[25] = vqsubq_s16(step2[24], step2[25]);
step1[26] = vqsubq_s16(step2[27], step2[26]);
step1[27] = vqaddq_s16(step2[27], step2[26]);
step1[28] = vqaddq_s16(step2[28], step2[29]);
step1[29] = vqsubq_s16(step2[28], step2[29]);
step1[30] = vqsubq_s16(step2[31], step2[30]);
step1[31] = vqaddq_s16(step2[31], step2[30]);
// stage 4
btf_16_neon(step1[4], cospi[56], cospi[8], &step2[4], &step2[7]);
btf_16_neon(step1[6], -cospi[40], cospi[24], &step2[5], &step2[6]);
btf_16_lane_0_1_neon(step1[30], step1[17], c0, &step2[30], &step2[17]);
btf_16_lane_1_0_neon(step1[18], step1[29], c2, &step2[18], &step2[29]);
btf_16_lane_2_3_neon(step1[26], step1[21], c0, &step2[26], &step2[21]);
btf_16_lane_3_2_neon(step1[22], step1[25], c2, &step2[22], &step2[25]);
step2[0] = step1[0];
step2[2] = step1[2];
step2[8] = vqaddq_s16(step1[8], step1[9]);
step2[9] = vqsubq_s16(step1[8], step1[9]);
step2[10] = vqsubq_s16(step1[11], step1[10]);
step2[11] = vqaddq_s16(step1[11], step1[10]);
step2[12] = vqaddq_s16(step1[12], step1[13]);
step2[13] = vqsubq_s16(step1[12], step1[13]);
step2[14] = vqsubq_s16(step1[15], step1[14]);
step2[15] = vqaddq_s16(step1[15], step1[14]);
step2[16] = step1[16];
step2[19] = step1[19];
step2[20] = step1[20];
step2[23] = step1[23];
step2[24] = step1[24];
step2[27] = step1[27];
step2[28] = step1[28];
step2[31] = step1[31];
// stage 5
t32[0] = vmull_n_s16(vget_low_s16(step2[0]), cospi[32]);
t32[1] = vmull_n_s16(vget_high_s16(step2[0]), cospi[32]);
step1[0] = vcombine_s16(vrshrn_n_s32(t32[0], INV_COS_BIT),
vrshrn_n_s32(t32[1], INV_COS_BIT));
btf_16_neon(step2[2], cospi[48], cospi[16], &step1[2], &step1[3]);
btf_16_lane_2_3_neon(step2[14], step2[9], c1, &step1[14], &step1[9]);
btf_16_lane_3_2_neon(step2[10], step2[13], c3, &step1[10], &step1[13]);
step1[4] = vqaddq_s16(step2[4], step2[5]);
step1[5] = vqsubq_s16(step2[4], step2[5]);
step1[6] = vqsubq_s16(step2[7], step2[6]);
step1[7] = vqaddq_s16(step2[7], step2[6]);
step1[8] = step2[8];
step1[11] = step2[11];
step1[12] = step2[12];
step1[15] = step2[15];
step1[16] = vqaddq_s16(step2[16], step2[19]);
step1[17] = vqaddq_s16(step2[17], step2[18]);
step1[18] = vqsubq_s16(step2[17], step2[18]);
step1[19] = vqsubq_s16(step2[16], step2[19]);
step1[20] = vqsubq_s16(step2[23], step2[20]);
step1[21] = vqsubq_s16(step2[22], step2[21]);
step1[22] = vqaddq_s16(step2[22], step2[21]);
step1[23] = vqaddq_s16(step2[23], step2[20]);
step1[24] = vqaddq_s16(step2[24], step2[27]);
step1[25] = vqaddq_s16(step2[25], step2[26]);
step1[26] = vqsubq_s16(step2[25], step2[26]);
step1[27] = vqsubq_s16(step2[24], step2[27]);
step1[28] = vqsubq_s16(step2[31], step2[28]);
step1[29] = vqsubq_s16(step2[30], step2[29]);
step1[30] = vqaddq_s16(step2[30], step2[29]);
step1[31] = vqaddq_s16(step2[31], step2[28]);
// stage 6
btf_16_lane_0_1_neon(step1[6], step1[5], c1, &step2[6], &step2[5]);
btf_16_lane_2_3_neon(step1[29], step1[18], c1, &step2[29], &step2[18]);
btf_16_lane_2_3_neon(step1[28], step1[19], c1, &step2[28], &step2[19]);
btf_16_lane_3_2_neon(step1[20], step1[27], c3, &step2[20], &step2[27]);
btf_16_lane_3_2_neon(step1[21], step1[26], c3, &step2[21], &step2[26]);
step2[0] = vqaddq_s16(step1[0], step1[3]);
step2[1] = vqaddq_s16(step1[0], step1[2]);
step2[2] = vqsubq_s16(step1[0], step1[2]);
step2[3] = vqsubq_s16(step1[0], step1[3]);
step2[4] = step1[4];
step2[7] = step1[7];
step2[8] = vqaddq_s16(step1[8], step1[11]);
step2[9] = vqaddq_s16(step1[9], step1[10]);
step2[10] = vqsubq_s16(step1[9], step1[10]);
step2[11] = vqsubq_s16(step1[8], step1[11]);
step2[12] = vqsubq_s16(step1[15], step1[12]);
step2[13] = vqsubq_s16(step1[14], step1[13]);
step2[14] = vqaddq_s16(step1[14], step1[13]);
step2[15] = vqaddq_s16(step1[15], step1[12]);
step2[16] = step1[16];
step2[17] = step1[17];
step2[22] = step1[22];
step2[23] = step1[23];
step2[24] = step1[24];
step2[25] = step1