1 /*
2 * Copyright (c) 2023 The WebM project authors. All rights reserved.
3 * Copyright (c) 2023, Alliance for Open Media. All rights reserved.
4 *
5 * This source code is subject to the terms of the BSD 2 Clause License and
6 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
7 * was not distributed with this source code in the LICENSE file, you can
8 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
9 * Media Patent License 1.0 was not distributed with this source code in the
10 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
11 */
12
13 #include <arm_neon.h>
14 #include "config/aom_dsp_rtcd.h"
15 #include "aom/aom_integer.h"
16 #include "aom_dsp/arm/mem_neon.h"
17 #include "aom_dsp/arm/transpose_neon.h"
18 #include "aom_dsp/arm/sum_neon.h"
19 #include "aom_ports/mem.h"
20
hadamard_highbd_col8_first_pass(int16x8_t * a0,int16x8_t * a1,int16x8_t * a2,int16x8_t * a3,int16x8_t * a4,int16x8_t * a5,int16x8_t * a6,int16x8_t * a7)21 static inline void hadamard_highbd_col8_first_pass(int16x8_t *a0, int16x8_t *a1,
22 int16x8_t *a2, int16x8_t *a3,
23 int16x8_t *a4, int16x8_t *a5,
24 int16x8_t *a6,
25 int16x8_t *a7) {
26 int16x8_t b0 = vaddq_s16(*a0, *a1);
27 int16x8_t b1 = vsubq_s16(*a0, *a1);
28 int16x8_t b2 = vaddq_s16(*a2, *a3);
29 int16x8_t b3 = vsubq_s16(*a2, *a3);
30 int16x8_t b4 = vaddq_s16(*a4, *a5);
31 int16x8_t b5 = vsubq_s16(*a4, *a5);
32 int16x8_t b6 = vaddq_s16(*a6, *a7);
33 int16x8_t b7 = vsubq_s16(*a6, *a7);
34
35 int16x8_t c0 = vaddq_s16(b0, b2);
36 int16x8_t c2 = vsubq_s16(b0, b2);
37 int16x8_t c1 = vaddq_s16(b1, b3);
38 int16x8_t c3 = vsubq_s16(b1, b3);
39 int16x8_t c4 = vaddq_s16(b4, b6);
40 int16x8_t c6 = vsubq_s16(b4, b6);
41 int16x8_t c5 = vaddq_s16(b5, b7);
42 int16x8_t c7 = vsubq_s16(b5, b7);
43
44 *a0 = vaddq_s16(c0, c4);
45 *a2 = vsubq_s16(c0, c4);
46 *a7 = vaddq_s16(c1, c5);
47 *a6 = vsubq_s16(c1, c5);
48 *a3 = vaddq_s16(c2, c6);
49 *a1 = vsubq_s16(c2, c6);
50 *a4 = vaddq_s16(c3, c7);
51 *a5 = vsubq_s16(c3, c7);
52 }
53
hadamard_highbd_col4_second_pass(int16x4_t a0,int16x4_t a1,int16x4_t a2,int16x4_t a3,int16x4_t a4,int16x4_t a5,int16x4_t a6,int16x4_t a7,tran_low_t * coeff)54 static inline void hadamard_highbd_col4_second_pass(int16x4_t a0, int16x4_t a1,
55 int16x4_t a2, int16x4_t a3,
56 int16x4_t a4, int16x4_t a5,
57 int16x4_t a6, int16x4_t a7,
58 tran_low_t *coeff) {
59 int32x4_t b0 = vaddl_s16(a0, a1);
60 int32x4_t b1 = vsubl_s16(a0, a1);
61 int32x4_t b2 = vaddl_s16(a2, a3);
62 int32x4_t b3 = vsubl_s16(a2, a3);
63 int32x4_t b4 = vaddl_s16(a4, a5);
64 int32x4_t b5 = vsubl_s16(a4, a5);
65 int32x4_t b6 = vaddl_s16(a6, a7);
66 int32x4_t b7 = vsubl_s16(a6, a7);
67
68 int32x4_t c0 = vaddq_s32(b0, b2);
69 int32x4_t c2 = vsubq_s32(b0, b2);
70 int32x4_t c1 = vaddq_s32(b1, b3);
71 int32x4_t c3 = vsubq_s32(b1, b3);
72 int32x4_t c4 = vaddq_s32(b4, b6);
73 int32x4_t c6 = vsubq_s32(b4, b6);
74 int32x4_t c5 = vaddq_s32(b5, b7);
75 int32x4_t c7 = vsubq_s32(b5, b7);
76
77 int32x4_t d0 = vaddq_s32(c0, c4);
78 int32x4_t d2 = vsubq_s32(c0, c4);
79 int32x4_t d7 = vaddq_s32(c1, c5);
80 int32x4_t d6 = vsubq_s32(c1, c5);
81 int32x4_t d3 = vaddq_s32(c2, c6);
82 int32x4_t d1 = vsubq_s32(c2, c6);
83 int32x4_t d4 = vaddq_s32(c3, c7);
84 int32x4_t d5 = vsubq_s32(c3, c7);
85
86 vst1q_s32(coeff + 0, d0);
87 vst1q_s32(coeff + 4, d1);
88 vst1q_s32(coeff + 8, d2);
89 vst1q_s32(coeff + 12, d3);
90 vst1q_s32(coeff + 16, d4);
91 vst1q_s32(coeff + 20, d5);
92 vst1q_s32(coeff + 24, d6);
93 vst1q_s32(coeff + 28, d7);
94 }
95
aom_highbd_hadamard_8x8_neon(const int16_t * src_diff,ptrdiff_t src_stride,tran_low_t * coeff)96 void aom_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride,
97 tran_low_t *coeff) {
98 int16x4_t b0, b1, b2, b3, b4, b5, b6, b7;
99
100 int16x8_t s0 = vld1q_s16(src_diff + 0 * src_stride);
101 int16x8_t s1 = vld1q_s16(src_diff + 1 * src_stride);
102 int16x8_t s2 = vld1q_s16(src_diff + 2 * src_stride);
103 int16x8_t s3 = vld1q_s16(src_diff + 3 * src_stride);
104 int16x8_t s4 = vld1q_s16(src_diff + 4 * src_stride);
105 int16x8_t s5 = vld1q_s16(src_diff + 5 * src_stride);
106 int16x8_t s6 = vld1q_s16(src_diff + 6 * src_stride);
107 int16x8_t s7 = vld1q_s16(src_diff + 7 * src_stride);
108
109 // For the first pass we can stay in 16-bit elements (4095*8 = 32760).
110 hadamard_highbd_col8_first_pass(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
111
112 transpose_elems_inplace_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
113
114 // For the second pass we need to widen to 32-bit elements, so we're
115 // processing 4 columns at a time.
116 // Skip the second transpose because it is not required.
117
118 b0 = vget_low_s16(s0);
119 b1 = vget_low_s16(s1);
120 b2 = vget_low_s16(s2);
121 b3 = vget_low_s16(s3);
122 b4 = vget_low_s16(s4);
123 b5 = vget_low_s16(s5);
124 b6 = vget_low_s16(s6);
125 b7 = vget_low_s16(s7);
126
127 hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff);
128
129 b0 = vget_high_s16(s0);
130 b1 = vget_high_s16(s1);
131 b2 = vget_high_s16(s2);
132 b3 = vget_high_s16(s3);
133 b4 = vget_high_s16(s4);
134 b5 = vget_high_s16(s5);
135 b6 = vget_high_s16(s6);
136 b7 = vget_high_s16(s7);
137
138 hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff + 32);
139 }
140
aom_highbd_hadamard_16x16_neon(const int16_t * src_diff,ptrdiff_t src_stride,tran_low_t * coeff)141 void aom_highbd_hadamard_16x16_neon(const int16_t *src_diff,
142 ptrdiff_t src_stride, tran_low_t *coeff) {
143 // Rearrange 16x16 to 8x32 and remove stride.
144 // Top left first.
145 aom_highbd_hadamard_8x8_neon(src_diff, src_stride, coeff);
146 // Top right.
147 aom_highbd_hadamard_8x8_neon(src_diff + 8, src_stride, coeff + 64);
148 // Bottom left.
149 aom_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride, src_stride,
150 coeff + 128);
151 // Bottom right.
152 aom_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride + 8, src_stride,
153 coeff + 192);
154
155 for (int i = 0; i < 16; i++) {
156 int32x4_t a0 = vld1q_s32(coeff + 4 * i);
157 int32x4_t a1 = vld1q_s32(coeff + 4 * i + 64);
158 int32x4_t a2 = vld1q_s32(coeff + 4 * i + 128);
159 int32x4_t a3 = vld1q_s32(coeff + 4 * i + 192);
160
161 int32x4_t b0 = vhaddq_s32(a0, a1);
162 int32x4_t b1 = vhsubq_s32(a0, a1);
163 int32x4_t b2 = vhaddq_s32(a2, a3);
164 int32x4_t b3 = vhsubq_s32(a2, a3);
165
166 int32x4_t c0 = vaddq_s32(b0, b2);
167 int32x4_t c1 = vaddq_s32(b1, b3);
168 int32x4_t c2 = vsubq_s32(b0, b2);
169 int32x4_t c3 = vsubq_s32(b1, b3);
170
171 vst1q_s32(coeff + 4 * i, c0);
172 vst1q_s32(coeff + 4 * i + 64, c1);
173 vst1q_s32(coeff + 4 * i + 128, c2);
174 vst1q_s32(coeff + 4 * i + 192, c3);
175 }
176 }
177
aom_highbd_hadamard_32x32_neon(const int16_t * src_diff,ptrdiff_t src_stride,tran_low_t * coeff)178 void aom_highbd_hadamard_32x32_neon(const int16_t *src_diff,
179 ptrdiff_t src_stride, tran_low_t *coeff) {
180 // Rearrange 32x32 to 16x64 and remove stride.
181 // Top left first.
182 aom_highbd_hadamard_16x16_neon(src_diff, src_stride, coeff);
183 // Top right.
184 aom_highbd_hadamard_16x16_neon(src_diff + 16, src_stride, coeff + 256);
185 // Bottom left.
186 aom_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride, src_stride,
187 coeff + 512);
188 // Bottom right.
189 aom_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride + 16, src_stride,
190 coeff + 768);
191
192 for (int i = 0; i < 64; i++) {
193 int32x4_t a0 = vld1q_s32(coeff + 4 * i);
194 int32x4_t a1 = vld1q_s32(coeff + 4 * i + 256);
195 int32x4_t a2 = vld1q_s32(coeff + 4 * i + 512);
196 int32x4_t a3 = vld1q_s32(coeff + 4 * i + 768);
197
198 int32x4_t b0 = vshrq_n_s32(vaddq_s32(a0, a1), 2);
199 int32x4_t b1 = vshrq_n_s32(vsubq_s32(a0, a1), 2);
200 int32x4_t b2 = vshrq_n_s32(vaddq_s32(a2, a3), 2);
201 int32x4_t b3 = vshrq_n_s32(vsubq_s32(a2, a3), 2);
202
203 int32x4_t c0 = vaddq_s32(b0, b2);
204 int32x4_t c1 = vaddq_s32(b1, b3);
205 int32x4_t c2 = vsubq_s32(b0, b2);
206 int32x4_t c3 = vsubq_s32(b1, b3);
207
208 vst1q_s32(coeff + 4 * i, c0);
209 vst1q_s32(coeff + 4 * i + 256, c1);
210 vst1q_s32(coeff + 4 * i + 512, c2);
211 vst1q_s32(coeff + 4 * i + 768, c3);
212 }
213 }
214