1/*
2 *  Copyright (c) 2014 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 "./vp8_rtcd.h"
14
15static const int16_t cospi8sqrt2minus1 = 20091;
16// 35468 exceeds INT16_MAX and gets converted to a negative number. Because of
17// the way it is used in vqdmulh, where the result is doubled, it can be divided
18// by 2 beforehand. This saves compensating for the negative value as well as
19// shifting the result.
20static const int16_t sinpi8sqrt2 = 35468 >> 1;
21
22void vp8_dequant_idct_add_neon(int16_t *input, int16_t *dq, unsigned char *dst,
23                               int stride) {
24  unsigned char *dst0;
25  int32x2_t d14, d15;
26  int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
27  int16x8_t q1, q2, q3, q4, q5, q6;
28  int16x8_t qEmpty = vdupq_n_s16(0);
29  int32x2x2_t d2tmp0, d2tmp1;
30  int16x4x2_t d2tmp2, d2tmp3;
31
32  d14 = d15 = vdup_n_s32(0);
33
34  // load input
35  q3 = vld1q_s16(input);
36  vst1q_s16(input, qEmpty);
37  input += 8;
38  q4 = vld1q_s16(input);
39  vst1q_s16(input, qEmpty);
40
41  // load dq
42  q5 = vld1q_s16(dq);
43  dq += 8;
44  q6 = vld1q_s16(dq);
45
46  // load src from dst
47  dst0 = dst;
48  d14 = vld1_lane_s32((const int32_t *)dst0, d14, 0);
49  dst0 += stride;
50  d14 = vld1_lane_s32((const int32_t *)dst0, d14, 1);
51  dst0 += stride;
52  d15 = vld1_lane_s32((const int32_t *)dst0, d15, 0);
53  dst0 += stride;
54  d15 = vld1_lane_s32((const int32_t *)dst0, d15, 1);
55
56  q1 = vreinterpretq_s16_u16(
57      vmulq_u16(vreinterpretq_u16_s16(q3), vreinterpretq_u16_s16(q5)));
58  q2 = vreinterpretq_s16_u16(
59      vmulq_u16(vreinterpretq_u16_s16(q4), vreinterpretq_u16_s16(q6)));
60
61  d12 = vqadd_s16(vget_low_s16(q1), vget_low_s16(q2));
62  d13 = vqsub_s16(vget_low_s16(q1), vget_low_s16(q2));
63
64  q2 = vcombine_s16(vget_high_s16(q1), vget_high_s16(q2));
65
66  q3 = vqdmulhq_n_s16(q2, sinpi8sqrt2);
67  q4 = vqdmulhq_n_s16(q2, cospi8sqrt2minus1);
68
69  q4 = vshrq_n_s16(q4, 1);
70
71  q4 = vqaddq_s16(q4, q2);
72
73  d10 = vqsub_s16(vget_low_s16(q3), vget_high_s16(q4));
74  d11 = vqadd_s16(vget_high_s16(q3), vget_low_s16(q4));
75
76  d2 = vqadd_s16(d12, d11);
77  d3 = vqadd_s16(d13, d10);
78  d4 = vqsub_s16(d13, d10);
79  d5 = vqsub_s16(d12, d11);
80
81  d2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
82  d2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
83  d2tmp2 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[0]),
84                    vreinterpret_s16_s32(d2tmp1.val[0]));
85  d2tmp3 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[1]),
86                    vreinterpret_s16_s32(d2tmp1.val[1]));
87
88  // loop 2
89  q2 = vcombine_s16(d2tmp2.val[1], d2tmp3.val[1]);
90
91  q3 = vqdmulhq_n_s16(q2, sinpi8sqrt2);
92  q4 = vqdmulhq_n_s16(q2, cospi8sqrt2minus1);
93
94  d12 = vqadd_s16(d2tmp2.val[0], d2tmp3.val[0]);
95  d13 = vqsub_s16(d2tmp2.val[0], d2tmp3.val[0]);
96
97  q4 = vshrq_n_s16(q4, 1);
98
99  q4 = vqaddq_s16(q4, q2);
100
101  d10 = vqsub_s16(vget_low_s16(q3), vget_high_s16(q4));
102  d11 = vqadd_s16(vget_high_s16(q3), vget_low_s16(q4));
103
104  d2 = vqadd_s16(d12, d11);
105  d3 = vqadd_s16(d13, d10);
106  d4 = vqsub_s16(d13, d10);
107  d5 = vqsub_s16(d12, d11);
108
109  d2 = vrshr_n_s16(d2, 3);
110  d3 = vrshr_n_s16(d3, 3);
111  d4 = vrshr_n_s16(d4, 3);
112  d5 = vrshr_n_s16(d5, 3);
113
114  d2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
115  d2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
116  d2tmp2 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[0]),
117                    vreinterpret_s16_s32(d2tmp1.val[0]));
118  d2tmp3 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[1]),
119                    vreinterpret_s16_s32(d2tmp1.val[1]));
120
121  q1 = vcombine_s16(d2tmp2.val[0], d2tmp2.val[1]);
122  q2 = vcombine_s16(d2tmp3.val[0], d2tmp3.val[1]);
123
124  q1 = vreinterpretq_s16_u16(
125      vaddw_u8(vreinterpretq_u16_s16(q1), vreinterpret_u8_s32(d14)));
126  q2 = vreinterpretq_s16_u16(
127      vaddw_u8(vreinterpretq_u16_s16(q2), vreinterpret_u8_s32(d15)));
128
129  d14 = vreinterpret_s32_u8(vqmovun_s16(q1));
130  d15 = vreinterpret_s32_u8(vqmovun_s16(q2));
131
132  dst0 = dst;
133  vst1_lane_s32((int32_t *)dst0, d14, 0);
134  dst0 += stride;
135  vst1_lane_s32((int32_t *)dst0, d14, 1);
136  dst0 += stride;
137  vst1_lane_s32((int32_t *)dst0, d15, 0);
138  dst0 += stride;
139  vst1_lane_s32((int32_t *)dst0, d15, 1);
140  return;
141}
142