1 /*
2 * Copyright (c) 2015 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_config.h"
14 #include "./vpx_dsp_rtcd.h"
15 #include "vpx/vpx_integer.h"
16 #include "vpx_dsp/arm/mem_neon.h"
17 #include "vpx_dsp/arm/sum_neon.h"
18
vpx_sad4x4x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)19 void vpx_sad4x4x4d_neon(const uint8_t *src, int src_stride,
20 const uint8_t *const ref[4], int ref_stride,
21 uint32_t *res) {
22 int i;
23 const uint8x16_t src_u8 = load_unaligned_u8q(src, src_stride);
24 for (i = 0; i < 4; ++i) {
25 const uint8x16_t ref_u8 = load_unaligned_u8q(ref[i], ref_stride);
26 uint16x8_t abs = vabdl_u8(vget_low_u8(src_u8), vget_low_u8(ref_u8));
27 abs = vabal_u8(abs, vget_high_u8(src_u8), vget_high_u8(ref_u8));
28 res[i] = vget_lane_u32(horizontal_add_uint16x8(abs), 0);
29 }
30 }
31
vpx_sad4x8x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)32 void vpx_sad4x8x4d_neon(const uint8_t *src, int src_stride,
33 const uint8_t *const ref[4], int ref_stride,
34 uint32_t *res) {
35 int i;
36 const uint8x16_t src_0 = load_unaligned_u8q(src, src_stride);
37 const uint8x16_t src_1 = load_unaligned_u8q(src + 4 * src_stride, src_stride);
38 for (i = 0; i < 4; ++i) {
39 const uint8x16_t ref_0 = load_unaligned_u8q(ref[i], ref_stride);
40 const uint8x16_t ref_1 =
41 load_unaligned_u8q(ref[i] + 4 * ref_stride, ref_stride);
42 uint16x8_t abs = vabdl_u8(vget_low_u8(src_0), vget_low_u8(ref_0));
43 abs = vabal_u8(abs, vget_high_u8(src_0), vget_high_u8(ref_0));
44 abs = vabal_u8(abs, vget_low_u8(src_1), vget_low_u8(ref_1));
45 abs = vabal_u8(abs, vget_high_u8(src_1), vget_high_u8(ref_1));
46 res[i] = vget_lane_u32(horizontal_add_uint16x8(abs), 0);
47 }
48 }
49
sad8x_4d(const uint8_t * a,int a_stride,const uint8_t * const b[4],int b_stride,uint32_t * result,const int height)50 static INLINE void sad8x_4d(const uint8_t *a, int a_stride,
51 const uint8_t *const b[4], int b_stride,
52 uint32_t *result, const int height) {
53 int i, j;
54 uint16x8_t sum[4] = { vdupq_n_u16(0), vdupq_n_u16(0), vdupq_n_u16(0),
55 vdupq_n_u16(0) };
56 const uint8_t *b_loop[4] = { b[0], b[1], b[2], b[3] };
57
58 for (i = 0; i < height; ++i) {
59 const uint8x8_t a_u8 = vld1_u8(a);
60 a += a_stride;
61 for (j = 0; j < 4; ++j) {
62 const uint8x8_t b_u8 = vld1_u8(b_loop[j]);
63 b_loop[j] += b_stride;
64 sum[j] = vabal_u8(sum[j], a_u8, b_u8);
65 }
66 }
67
68 for (j = 0; j < 4; ++j) {
69 result[j] = vget_lane_u32(horizontal_add_uint16x8(sum[j]), 0);
70 }
71 }
72
vpx_sad8x4x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)73 void vpx_sad8x4x4d_neon(const uint8_t *src, int src_stride,
74 const uint8_t *const ref[4], int ref_stride,
75 uint32_t *res) {
76 sad8x_4d(src, src_stride, ref, ref_stride, res, 4);
77 }
78
vpx_sad8x8x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)79 void vpx_sad8x8x4d_neon(const uint8_t *src, int src_stride,
80 const uint8_t *const ref[4], int ref_stride,
81 uint32_t *res) {
82 sad8x_4d(src, src_stride, ref, ref_stride, res, 8);
83 }
84
vpx_sad8x16x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)85 void vpx_sad8x16x4d_neon(const uint8_t *src, int src_stride,
86 const uint8_t *const ref[4], int ref_stride,
87 uint32_t *res) {
88 sad8x_4d(src, src_stride, ref, ref_stride, res, 16);
89 }
90
sad16x_4d(const uint8_t * a,int a_stride,const uint8_t * const b[4],int b_stride,uint32_t * result,const int height)91 static INLINE void sad16x_4d(const uint8_t *a, int a_stride,
92 const uint8_t *const b[4], int b_stride,
93 uint32_t *result, const int height) {
94 int i, j;
95 uint16x8_t sum[4] = { vdupq_n_u16(0), vdupq_n_u16(0), vdupq_n_u16(0),
96 vdupq_n_u16(0) };
97 const uint8_t *b_loop[4] = { b[0], b[1], b[2], b[3] };
98
99 for (i = 0; i < height; ++i) {
100 const uint8x16_t a_u8 = vld1q_u8(a);
101 a += a_stride;
102 for (j = 0; j < 4; ++j) {
103 const uint8x16_t b_u8 = vld1q_u8(b_loop[j]);
104 b_loop[j] += b_stride;
105 sum[j] = vabal_u8(sum[j], vget_low_u8(a_u8), vget_low_u8(b_u8));
106 sum[j] = vabal_u8(sum[j], vget_high_u8(a_u8), vget_high_u8(b_u8));
107 }
108 }
109
110 for (j = 0; j < 4; ++j) {
111 result[j] = vget_lane_u32(horizontal_add_uint16x8(sum[j]), 0);
112 }
113 }
114
vpx_sad16x8x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)115 void vpx_sad16x8x4d_neon(const uint8_t *src, int src_stride,
116 const uint8_t *const ref[4], int ref_stride,
117 uint32_t *res) {
118 sad16x_4d(src, src_stride, ref, ref_stride, res, 8);
119 }
120
vpx_sad16x16x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)121 void vpx_sad16x16x4d_neon(const uint8_t *src, int src_stride,
122 const uint8_t *const ref[4], int ref_stride,
123 uint32_t *res) {
124 sad16x_4d(src, src_stride, ref, ref_stride, res, 16);
125 }
126
vpx_sad16x32x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)127 void vpx_sad16x32x4d_neon(const uint8_t *src, int src_stride,
128 const uint8_t *const ref[4], int ref_stride,
129 uint32_t *res) {
130 sad16x_4d(src, src_stride, ref, ref_stride, res, 32);
131 }
132
sad32x_4d(const uint8_t * a,int a_stride,const uint8_t * const b[4],int b_stride,uint32_t * result,const int height)133 static INLINE void sad32x_4d(const uint8_t *a, int a_stride,
134 const uint8_t *const b[4], int b_stride,
135 uint32_t *result, const int height) {
136 int i, j;
137 uint16x8_t sum[4] = { vdupq_n_u16(0), vdupq_n_u16(0), vdupq_n_u16(0),
138 vdupq_n_u16(0) };
139 const uint8_t *b_loop[4] = { b[0], b[1], b[2], b[3] };
140
141 for (i = 0; i < height; ++i) {
142 const uint8x16_t a_0 = vld1q_u8(a);
143 const uint8x16_t a_1 = vld1q_u8(a + 16);
144 a += a_stride;
145 for (j = 0; j < 4; ++j) {
146 const uint8x16_t b_0 = vld1q_u8(b_loop[j]);
147 const uint8x16_t b_1 = vld1q_u8(b_loop[j] + 16);
148 b_loop[j] += b_stride;
149 sum[j] = vabal_u8(sum[j], vget_low_u8(a_0), vget_low_u8(b_0));
150 sum[j] = vabal_u8(sum[j], vget_high_u8(a_0), vget_high_u8(b_0));
151 sum[j] = vabal_u8(sum[j], vget_low_u8(a_1), vget_low_u8(b_1));
152 sum[j] = vabal_u8(sum[j], vget_high_u8(a_1), vget_high_u8(b_1));
153 }
154 }
155
156 for (j = 0; j < 4; ++j) {
157 result[j] = vget_lane_u32(horizontal_add_uint16x8(sum[j]), 0);
158 }
159 }
160
vpx_sad32x16x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)161 void vpx_sad32x16x4d_neon(const uint8_t *src, int src_stride,
162 const uint8_t *const ref[4], int ref_stride,
163 uint32_t *res) {
164 sad32x_4d(src, src_stride, ref, ref_stride, res, 16);
165 }
166
vpx_sad32x32x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)167 void vpx_sad32x32x4d_neon(const uint8_t *src, int src_stride,
168 const uint8_t *const ref[4], int ref_stride,
169 uint32_t *res) {
170 sad32x_4d(src, src_stride, ref, ref_stride, res, 32);
171 }
172
vpx_sad32x64x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)173 void vpx_sad32x64x4d_neon(const uint8_t *src, int src_stride,
174 const uint8_t *const ref[4], int ref_stride,
175 uint32_t *res) {
176 sad32x_4d(src, src_stride, ref, ref_stride, res, 64);
177 }
178
sum64x(const uint8x16_t a_0,const uint8x16_t a_1,const uint8x16_t b_0,const uint8x16_t b_1,uint16x8_t * sum)179 static INLINE void sum64x(const uint8x16_t a_0, const uint8x16_t a_1,
180 const uint8x16_t b_0, const uint8x16_t b_1,
181 uint16x8_t *sum) {
182 *sum = vabal_u8(*sum, vget_low_u8(a_0), vget_low_u8(b_0));
183 *sum = vabal_u8(*sum, vget_high_u8(a_0), vget_high_u8(b_0));
184 *sum = vabal_u8(*sum, vget_low_u8(a_1), vget_low_u8(b_1));
185 *sum = vabal_u8(*sum, vget_high_u8(a_1), vget_high_u8(b_1));
186 }
187
sad64x_4d(const uint8_t * a,int a_stride,const uint8_t * const b[4],int b_stride,uint32_t * result,const int height)188 static INLINE void sad64x_4d(const uint8_t *a, int a_stride,
189 const uint8_t *const b[4], int b_stride,
190 uint32_t *result, const int height) {
191 int i;
192 uint16x8_t sum_0 = vdupq_n_u16(0);
193 uint16x8_t sum_1 = vdupq_n_u16(0);
194 uint16x8_t sum_2 = vdupq_n_u16(0);
195 uint16x8_t sum_3 = vdupq_n_u16(0);
196 uint16x8_t sum_4 = vdupq_n_u16(0);
197 uint16x8_t sum_5 = vdupq_n_u16(0);
198 uint16x8_t sum_6 = vdupq_n_u16(0);
199 uint16x8_t sum_7 = vdupq_n_u16(0);
200 const uint8_t *b_loop[4] = { b[0], b[1], b[2], b[3] };
201
202 for (i = 0; i < height; ++i) {
203 const uint8x16_t a_0 = vld1q_u8(a);
204 const uint8x16_t a_1 = vld1q_u8(a + 16);
205 const uint8x16_t a_2 = vld1q_u8(a + 32);
206 const uint8x16_t a_3 = vld1q_u8(a + 48);
207 a += a_stride;
208 sum64x(a_0, a_1, vld1q_u8(b_loop[0]), vld1q_u8(b_loop[0] + 16), &sum_0);
209 sum64x(a_2, a_3, vld1q_u8(b_loop[0] + 32), vld1q_u8(b_loop[0] + 48),
210 &sum_1);
211 b_loop[0] += b_stride;
212 sum64x(a_0, a_1, vld1q_u8(b_loop[1]), vld1q_u8(b_loop[1] + 16), &sum_2);
213 sum64x(a_2, a_3, vld1q_u8(b_loop[1] + 32), vld1q_u8(b_loop[1] + 48),
214 &sum_3);
215 b_loop[1] += b_stride;
216 sum64x(a_0, a_1, vld1q_u8(b_loop[2]), vld1q_u8(b_loop[2] + 16), &sum_4);
217 sum64x(a_2, a_3, vld1q_u8(b_loop[2] + 32), vld1q_u8(b_loop[2] + 48),
218 &sum_5);
219 b_loop[2] += b_stride;
220 sum64x(a_0, a_1, vld1q_u8(b_loop[3]), vld1q_u8(b_loop[3] + 16), &sum_6);
221 sum64x(a_2, a_3, vld1q_u8(b_loop[3] + 32), vld1q_u8(b_loop[3] + 48),
222 &sum_7);
223 b_loop[3] += b_stride;
224 }
225
226 result[0] = vget_lane_u32(horizontal_add_long_uint16x8(sum_0, sum_1), 0);
227 result[1] = vget_lane_u32(horizontal_add_long_uint16x8(sum_2, sum_3), 0);
228 result[2] = vget_lane_u32(horizontal_add_long_uint16x8(sum_4, sum_5), 0);
229 result[3] = vget_lane_u32(horizontal_add_long_uint16x8(sum_6, sum_7), 0);
230 }
231
vpx_sad64x32x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)232 void vpx_sad64x32x4d_neon(const uint8_t *src, int src_stride,
233 const uint8_t *const ref[4], int ref_stride,
234 uint32_t *res) {
235 sad64x_4d(src, src_stride, ref, ref_stride, res, 32);
236 }
237
vpx_sad64x64x4d_neon(const uint8_t * src,int src_stride,const uint8_t * const ref[4],int ref_stride,uint32_t * res)238 void vpx_sad64x64x4d_neon(const uint8_t *src, int src_stride,
239 const uint8_t *const ref[4], int ref_stride,
240 uint32_t *res) {
241 sad64x_4d(src, src_stride, ref, ref_stride, res, 64);
242 }
243