1/*
2 *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include <arm_neon.h>
12
13#include "./vpx_dsp_rtcd.h"
14#include "vpx_dsp/arm/idct_neon.h"
15#include "vpx_dsp/inv_txfm.h"
16
17static INLINE void highbd_idct32x32_1_add_pos_kernel(uint16_t **dest,
18                                                     const int stride,
19                                                     const int16x8_t res,
20                                                     const int16x8_t max) {
21  const uint16x8_t a0 = vld1q_u16(*dest);
22  const uint16x8_t a1 = vld1q_u16(*dest + 8);
23  const uint16x8_t a2 = vld1q_u16(*dest + 16);
24  const uint16x8_t a3 = vld1q_u16(*dest + 24);
25  const int16x8_t b0 = vaddq_s16(res, vreinterpretq_s16_u16(a0));
26  const int16x8_t b1 = vaddq_s16(res, vreinterpretq_s16_u16(a1));
27  const int16x8_t b2 = vaddq_s16(res, vreinterpretq_s16_u16(a2));
28  const int16x8_t b3 = vaddq_s16(res, vreinterpretq_s16_u16(a3));
29  const int16x8_t c0 = vminq_s16(b0, max);
30  const int16x8_t c1 = vminq_s16(b1, max);
31  const int16x8_t c2 = vminq_s16(b2, max);
32  const int16x8_t c3 = vminq_s16(b3, max);
33  vst1q_u16(*dest, vreinterpretq_u16_s16(c0));
34  vst1q_u16(*dest + 8, vreinterpretq_u16_s16(c1));
35  vst1q_u16(*dest + 16, vreinterpretq_u16_s16(c2));
36  vst1q_u16(*dest + 24, vreinterpretq_u16_s16(c3));
37  *dest += stride;
38}
39
40static INLINE void highbd_idct32x32_1_add_neg_kernel(uint16_t **dest,
41                                                     const int stride,
42                                                     const int16x8_t res) {
43  const uint16x8_t a0 = vld1q_u16(*dest);
44  const uint16x8_t a1 = vld1q_u16(*dest + 8);
45  const uint16x8_t a2 = vld1q_u16(*dest + 16);
46  const uint16x8_t a3 = vld1q_u16(*dest + 24);
47  const int16x8_t b0 = vaddq_s16(res, vreinterpretq_s16_u16(a0));
48  const int16x8_t b1 = vaddq_s16(res, vreinterpretq_s16_u16(a1));
49  const int16x8_t b2 = vaddq_s16(res, vreinterpretq_s16_u16(a2));
50  const int16x8_t b3 = vaddq_s16(res, vreinterpretq_s16_u16(a3));
51  const uint16x8_t c0 = vqshluq_n_s16(b0, 0);
52  const uint16x8_t c1 = vqshluq_n_s16(b1, 0);
53  const uint16x8_t c2 = vqshluq_n_s16(b2, 0);
54  const uint16x8_t c3 = vqshluq_n_s16(b3, 0);
55  vst1q_u16(*dest, c0);
56  vst1q_u16(*dest + 8, c1);
57  vst1q_u16(*dest + 16, c2);
58  vst1q_u16(*dest + 24, c3);
59  *dest += stride;
60}
61
62void vpx_highbd_idct32x32_1_add_neon(const tran_low_t *input, uint16_t *dest,
63                                     int stride, int bd) {
64  const tran_low_t out0 =
65      HIGHBD_WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), bd);
66  const tran_low_t out1 =
67      HIGHBD_WRAPLOW(dct_const_round_shift(out0 * cospi_16_64), bd);
68  const int16_t a1 = ROUND_POWER_OF_TWO(out1, 6);
69  const int16x8_t dc = vdupq_n_s16(a1);
70  int i;
71
72  if (a1 >= 0) {
73    const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
74    for (i = 0; i < 8; ++i) {
75      highbd_idct32x32_1_add_pos_kernel(&dest, stride, dc, max);
76      highbd_idct32x32_1_add_pos_kernel(&dest, stride, dc, max);
77      highbd_idct32x32_1_add_pos_kernel(&dest, stride, dc, max);
78      highbd_idct32x32_1_add_pos_kernel(&dest, stride, dc, max);
79    }
80  } else {
81    for (i = 0; i < 8; ++i) {
82      highbd_idct32x32_1_add_neg_kernel(&dest, stride, dc);
83      highbd_idct32x32_1_add_neg_kernel(&dest, stride, dc);
84      highbd_idct32x32_1_add_neg_kernel(&dest, stride, dc);
85      highbd_idct32x32_1_add_neg_kernel(&dest, stride, dc);
86    }
87  }
88}
89