1 /*
2 * Copyright (c) 2023 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_config.h"
15
16 #include "vpx_dsp/arm/mem_neon.h"
17 #include "vpx_dsp/arm/transpose_neon.h"
18
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)19 static INLINE void hadamard_highbd_col8_first_pass(int16x8_t *a0, int16x8_t *a1,
20 int16x8_t *a2, int16x8_t *a3,
21 int16x8_t *a4, int16x8_t *a5,
22 int16x8_t *a6,
23 int16x8_t *a7) {
24 int16x8_t b0 = vaddq_s16(*a0, *a1);
25 int16x8_t b1 = vsubq_s16(*a0, *a1);
26 int16x8_t b2 = vaddq_s16(*a2, *a3);
27 int16x8_t b3 = vsubq_s16(*a2, *a3);
28 int16x8_t b4 = vaddq_s16(*a4, *a5);
29 int16x8_t b5 = vsubq_s16(*a4, *a5);
30 int16x8_t b6 = vaddq_s16(*a6, *a7);
31 int16x8_t b7 = vsubq_s16(*a6, *a7);
32
33 int16x8_t c0 = vaddq_s16(b0, b2);
34 int16x8_t c2 = vsubq_s16(b0, b2);
35 int16x8_t c1 = vaddq_s16(b1, b3);
36 int16x8_t c3 = vsubq_s16(b1, b3);
37 int16x8_t c4 = vaddq_s16(b4, b6);
38 int16x8_t c6 = vsubq_s16(b4, b6);
39 int16x8_t c5 = vaddq_s16(b5, b7);
40 int16x8_t c7 = vsubq_s16(b5, b7);
41
42 *a0 = vaddq_s16(c0, c4);
43 *a2 = vsubq_s16(c0, c4);
44 *a7 = vaddq_s16(c1, c5);
45 *a6 = vsubq_s16(c1, c5);
46 *a3 = vaddq_s16(c2, c6);
47 *a1 = vsubq_s16(c2, c6);
48 *a4 = vaddq_s16(c3, c7);
49 *a5 = vsubq_s16(c3, c7);
50 }
51
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)52 static INLINE void hadamard_highbd_col4_second_pass(int16x4_t a0, int16x4_t a1,
53 int16x4_t a2, int16x4_t a3,
54 int16x4_t a4, int16x4_t a5,
55 int16x4_t a6, int16x4_t a7,
56 tran_low_t *coeff) {
57 int32x4_t b0 = vaddl_s16(a0, a1);
58 int32x4_t b1 = vsubl_s16(a0, a1);
59 int32x4_t b2 = vaddl_s16(a2, a3);
60 int32x4_t b3 = vsubl_s16(a2, a3);
61 int32x4_t b4 = vaddl_s16(a4, a5);
62 int32x4_t b5 = vsubl_s16(a4, a5);
63 int32x4_t b6 = vaddl_s16(a6, a7);
64 int32x4_t b7 = vsubl_s16(a6, a7);
65
66 int32x4_t c0 = vaddq_s32(b0, b2);
67 int32x4_t c2 = vsubq_s32(b0, b2);
68 int32x4_t c1 = vaddq_s32(b1, b3);
69 int32x4_t c3 = vsubq_s32(b1, b3);
70 int32x4_t c4 = vaddq_s32(b4, b6);
71 int32x4_t c6 = vsubq_s32(b4, b6);
72 int32x4_t c5 = vaddq_s32(b5, b7);
73 int32x4_t c7 = vsubq_s32(b5, b7);
74
75 int32x4_t d0 = vaddq_s32(c0, c4);
76 int32x4_t d2 = vsubq_s32(c0, c4);
77 int32x4_t d7 = vaddq_s32(c1, c5);
78 int32x4_t d6 = vsubq_s32(c1, c5);
79 int32x4_t d3 = vaddq_s32(c2, c6);
80 int32x4_t d1 = vsubq_s32(c2, c6);
81 int32x4_t d4 = vaddq_s32(c3, c7);
82 int32x4_t d5 = vsubq_s32(c3, c7);
83
84 store_s32q_to_tran_low(coeff + 0, d0);
85 store_s32q_to_tran_low(coeff + 4, d1);
86 store_s32q_to_tran_low(coeff + 8, d2);
87 store_s32q_to_tran_low(coeff + 12, d3);
88 store_s32q_to_tran_low(coeff + 16, d4);
89 store_s32q_to_tran_low(coeff + 20, d5);
90 store_s32q_to_tran_low(coeff + 24, d6);
91 store_s32q_to_tran_low(coeff + 28, d7);
92 }
93
vpx_highbd_hadamard_8x8_neon(const int16_t * src_diff,ptrdiff_t src_stride,tran_low_t * coeff)94 void vpx_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride,
95 tran_low_t *coeff) {
96 int16x4_t b0, b1, b2, b3, b4, b5, b6, b7;
97
98 int16x8_t s0 = vld1q_s16(src_diff + 0 * src_stride);
99 int16x8_t s1 = vld1q_s16(src_diff + 1 * src_stride);
100 int16x8_t s2 = vld1q_s16(src_diff + 2 * src_stride);
101 int16x8_t s3 = vld1q_s16(src_diff + 3 * src_stride);
102 int16x8_t s4 = vld1q_s16(src_diff + 4 * src_stride);
103 int16x8_t s5 = vld1q_s16(src_diff + 5 * src_stride);
104 int16x8_t s6 = vld1q_s16(src_diff + 6 * src_stride);
105 int16x8_t s7 = vld1q_s16(src_diff + 7 * src_stride);
106
107 // For the first pass we can stay in 16-bit elements (4095*8 = 32760).
108 hadamard_highbd_col8_first_pass(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
109
110 transpose_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
111
112 // For the second pass we need to widen to 32-bit elements, so we're
113 // processing 4 columns at a time.
114 // Skip the second transpose because it is not required.
115
116 b0 = vget_low_s16(s0);
117 b1 = vget_low_s16(s1);
118 b2 = vget_low_s16(s2);
119 b3 = vget_low_s16(s3);
120 b4 = vget_low_s16(s4);
121 b5 = vget_low_s16(s5);
122 b6 = vget_low_s16(s6);
123 b7 = vget_low_s16(s7);
124
125 hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff);
126
127 b0 = vget_high_s16(s0);
128 b1 = vget_high_s16(s1);
129 b2 = vget_high_s16(s2);
130 b3 = vget_high_s16(s3);
131 b4 = vget_high_s16(s4);
132 b5 = vget_high_s16(s5);
133 b6 = vget_high_s16(s6);
134 b7 = vget_high_s16(s7);
135
136 hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff + 32);
137 }
138
vpx_highbd_hadamard_16x16_neon(const int16_t * src_diff,ptrdiff_t src_stride,tran_low_t * coeff)139 void vpx_highbd_hadamard_16x16_neon(const int16_t *src_diff,
140 ptrdiff_t src_stride, tran_low_t *coeff) {
141 int i = 0;
142
143 // Rearrange 16x16 to 8x32 and remove stride.
144 // Top left first.
145 vpx_highbd_hadamard_8x8_neon(src_diff, src_stride, coeff);
146 // Top right.
147 vpx_highbd_hadamard_8x8_neon(src_diff + 8, src_stride, coeff + 64);
148 // Bottom left.
149 vpx_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride, src_stride,
150 coeff + 128);
151 // Bottom right.
152 vpx_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride + 8, src_stride,
153 coeff + 192);
154
155 do {
156 int32x4_t a0 = load_tran_low_to_s32q(coeff + 4 * i);
157 int32x4_t a1 = load_tran_low_to_s32q(coeff + 4 * i + 64);
158 int32x4_t a2 = load_tran_low_to_s32q(coeff + 4 * i + 128);
159 int32x4_t a3 = load_tran_low_to_s32q(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 store_s32q_to_tran_low(coeff + 4 * i, c0);
172 store_s32q_to_tran_low(coeff + 4 * i + 64, c1);
173 store_s32q_to_tran_low(coeff + 4 * i + 128, c2);
174 store_s32q_to_tran_low(coeff + 4 * i + 192, c3);
175 } while (++i < 16);
176 }
177
vpx_highbd_hadamard_32x32_neon(const int16_t * src_diff,ptrdiff_t src_stride,tran_low_t * coeff)178 void vpx_highbd_hadamard_32x32_neon(const int16_t *src_diff,
179 ptrdiff_t src_stride, tran_low_t *coeff) {
180 int i = 0;
181
182 // Rearrange 32x32 to 16x64 and remove stride.
183 // Top left first.
184 vpx_highbd_hadamard_16x16_neon(src_diff, src_stride, coeff);
185 // Top right.
186 vpx_highbd_hadamard_16x16_neon(src_diff + 16, src_stride, coeff + 256);
187 // Bottom left.
188 vpx_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride, src_stride,
189 coeff + 512);
190 // Bottom right.
191 vpx_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride + 16, src_stride,
192 coeff + 768);
193
194 do {
195 int32x4_t a0 = load_tran_low_to_s32q(coeff + 4 * i);
196 int32x4_t a1 = load_tran_low_to_s32q(coeff + 4 * i + 256);
197 int32x4_t a2 = load_tran_low_to_s32q(coeff + 4 * i + 512);
198 int32x4_t a3 = load_tran_low_to_s32q(coeff + 4 * i + 768);
199
200 int32x4_t b0 = vshrq_n_s32(vaddq_s32(a0, a1), 2);
201 int32x4_t b1 = vshrq_n_s32(vsubq_s32(a0, a1), 2);
202 int32x4_t b2 = vshrq_n_s32(vaddq_s32(a2, a3), 2);
203 int32x4_t b3 = vshrq_n_s32(vsubq_s32(a2, a3), 2);
204
205 int32x4_t c0 = vaddq_s32(b0, b2);
206 int32x4_t c1 = vaddq_s32(b1, b3);
207 int32x4_t c2 = vsubq_s32(b0, b2);
208 int32x4_t c3 = vsubq_s32(b1, b3);
209
210 store_s32q_to_tran_low(coeff + 4 * i, c0);
211 store_s32q_to_tran_low(coeff + 4 * i + 256, c1);
212 store_s32q_to_tran_low(coeff + 4 * i + 512, c2);
213 store_s32q_to_tran_low(coeff + 4 * i + 768, c3);
214 } while (++i < 64);
215 }
216