1/*
2 *  Copyright (c) 2016 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/vpx_integer.h"
15#include "vpx_dsp/arm/idct_neon.h"
16#include "vpx_dsp/arm/mem_neon.h"
17#include "vpx_dsp/arm/transpose_neon.h"
18
19static void hadamard8x8_one_pass(int16x8_t *a0, int16x8_t *a1, int16x8_t *a2,
20                                 int16x8_t *a3, int16x8_t *a4, int16x8_t *a5,
21                                 int16x8_t *a6, int16x8_t *a7) {
22  const int16x8_t b0 = vaddq_s16(*a0, *a1);
23  const int16x8_t b1 = vsubq_s16(*a0, *a1);
24  const int16x8_t b2 = vaddq_s16(*a2, *a3);
25  const int16x8_t b3 = vsubq_s16(*a2, *a3);
26  const int16x8_t b4 = vaddq_s16(*a4, *a5);
27  const int16x8_t b5 = vsubq_s16(*a4, *a5);
28  const int16x8_t b6 = vaddq_s16(*a6, *a7);
29  const int16x8_t b7 = vsubq_s16(*a6, *a7);
30
31  const int16x8_t c0 = vaddq_s16(b0, b2);
32  const int16x8_t c1 = vaddq_s16(b1, b3);
33  const int16x8_t c2 = vsubq_s16(b0, b2);
34  const int16x8_t c3 = vsubq_s16(b1, b3);
35  const int16x8_t c4 = vaddq_s16(b4, b6);
36  const int16x8_t c5 = vaddq_s16(b5, b7);
37  const int16x8_t c6 = vsubq_s16(b4, b6);
38  const int16x8_t c7 = vsubq_s16(b5, b7);
39
40  *a0 = vaddq_s16(c0, c4);
41  *a1 = vsubq_s16(c2, c6);
42  *a2 = vsubq_s16(c0, c4);
43  *a3 = vaddq_s16(c2, c6);
44  *a4 = vaddq_s16(c3, c7);
45  *a5 = vsubq_s16(c3, c7);
46  *a6 = vsubq_s16(c1, c5);
47  *a7 = vaddq_s16(c1, c5);
48}
49
50void vpx_hadamard_8x8_neon(const int16_t *src_diff, int src_stride,
51                           tran_low_t *coeff) {
52  int16x8_t a0 = vld1q_s16(src_diff);
53  int16x8_t a1 = vld1q_s16(src_diff + src_stride);
54  int16x8_t a2 = vld1q_s16(src_diff + 2 * src_stride);
55  int16x8_t a3 = vld1q_s16(src_diff + 3 * src_stride);
56  int16x8_t a4 = vld1q_s16(src_diff + 4 * src_stride);
57  int16x8_t a5 = vld1q_s16(src_diff + 5 * src_stride);
58  int16x8_t a6 = vld1q_s16(src_diff + 6 * src_stride);
59  int16x8_t a7 = vld1q_s16(src_diff + 7 * src_stride);
60
61  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
62
63  transpose_s16_8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
64
65  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
66
67  // Skip the second transpose because it is not required.
68
69  store_s16q_to_tran_low(coeff + 0, a0);
70  store_s16q_to_tran_low(coeff + 8, a1);
71  store_s16q_to_tran_low(coeff + 16, a2);
72  store_s16q_to_tran_low(coeff + 24, a3);
73  store_s16q_to_tran_low(coeff + 32, a4);
74  store_s16q_to_tran_low(coeff + 40, a5);
75  store_s16q_to_tran_low(coeff + 48, a6);
76  store_s16q_to_tran_low(coeff + 56, a7);
77}
78
79void vpx_hadamard_16x16_neon(const int16_t *src_diff, int src_stride,
80                             tran_low_t *coeff) {
81  int i;
82
83  /* Rearrange 16x16 to 8x32 and remove stride.
84   * Top left first. */
85  vpx_hadamard_8x8_neon(src_diff + 0 + 0 * src_stride, src_stride, coeff + 0);
86  /* Top right. */
87  vpx_hadamard_8x8_neon(src_diff + 8 + 0 * src_stride, src_stride, coeff + 64);
88  /* Bottom left. */
89  vpx_hadamard_8x8_neon(src_diff + 0 + 8 * src_stride, src_stride, coeff + 128);
90  /* Bottom right. */
91  vpx_hadamard_8x8_neon(src_diff + 8 + 8 * src_stride, src_stride, coeff + 192);
92
93  for (i = 0; i < 64; i += 8) {
94    const int16x8_t a0 = load_tran_low_to_s16q(coeff + 0);
95    const int16x8_t a1 = load_tran_low_to_s16q(coeff + 64);
96    const int16x8_t a2 = load_tran_low_to_s16q(coeff + 128);
97    const int16x8_t a3 = load_tran_low_to_s16q(coeff + 192);
98
99    const int16x8_t b0 = vhaddq_s16(a0, a1);
100    const int16x8_t b1 = vhsubq_s16(a0, a1);
101    const int16x8_t b2 = vhaddq_s16(a2, a3);
102    const int16x8_t b3 = vhsubq_s16(a2, a3);
103
104    const int16x8_t c0 = vaddq_s16(b0, b2);
105    const int16x8_t c1 = vaddq_s16(b1, b3);
106    const int16x8_t c2 = vsubq_s16(b0, b2);
107    const int16x8_t c3 = vsubq_s16(b1, b3);
108
109    store_s16q_to_tran_low(coeff + 0, c0);
110    store_s16q_to_tran_low(coeff + 64, c1);
111    store_s16q_to_tran_low(coeff + 128, c2);
112    store_s16q_to_tran_low(coeff + 192, c3);
113
114    coeff += 8;
115  }
116}
117