1 /*
2 * Copyright (c) 2020, Alliance for Open Media. 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 "config/aom_dsp_rtcd.h"
14 #include "aom_dsp/arm/mem_neon.h"
15 #include "aom_dsp/arm/sum_neon.h"
16
sse_16x1_neon(const uint8_t * src,const uint8_t * ref,uint32x4_t * sse)17 static INLINE void sse_16x1_neon(const uint8_t *src, const uint8_t *ref,
18 uint32x4_t *sse) {
19 uint8x16_t s = vld1q_u8(src);
20 uint8x16_t r = vld1q_u8(ref);
21
22 uint8x16_t abs_diff = vabdq_u8(s, r);
23 uint8x8_t abs_diff_lo = vget_low_u8(abs_diff);
24 uint8x8_t abs_diff_hi = vget_high_u8(abs_diff);
25
26 *sse = vpadalq_u16(*sse, vmull_u8(abs_diff_lo, abs_diff_lo));
27 *sse = vpadalq_u16(*sse, vmull_u8(abs_diff_hi, abs_diff_hi));
28 }
29
sse_8x1_neon(const uint8_t * src,const uint8_t * ref,uint32x4_t * sse)30 static INLINE void sse_8x1_neon(const uint8_t *src, const uint8_t *ref,
31 uint32x4_t *sse) {
32 uint8x8_t s = vld1_u8(src);
33 uint8x8_t r = vld1_u8(ref);
34
35 uint8x8_t abs_diff = vabd_u8(s, r);
36
37 *sse = vpadalq_u16(*sse, vmull_u8(abs_diff, abs_diff));
38 }
39
sse_4x2_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,uint32x4_t * sse)40 static INLINE void sse_4x2_neon(const uint8_t *src, int src_stride,
41 const uint8_t *ref, int ref_stride,
42 uint32x4_t *sse) {
43 uint8x8_t s = load_unaligned_u8(src, src_stride);
44 uint8x8_t r = load_unaligned_u8(ref, ref_stride);
45
46 uint8x8_t abs_diff = vabd_u8(s, r);
47
48 *sse = vpadalq_u16(*sse, vmull_u8(abs_diff, abs_diff));
49 }
50
sse_wxh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int width,int height)51 static INLINE uint32_t sse_wxh_neon(const uint8_t *src, int src_stride,
52 const uint8_t *ref, int ref_stride,
53 int width, int height) {
54 uint32x4_t sse = vdupq_n_u32(0);
55
56 if ((width & 0x07) && ((width & 0x07) < 5)) {
57 int i = height;
58 do {
59 int j = 0;
60 do {
61 sse_8x1_neon(src + j, ref + j, &sse);
62 sse_8x1_neon(src + j + src_stride, ref + j + ref_stride, &sse);
63 j += 8;
64 } while (j + 4 < width);
65
66 sse_4x2_neon(src + j, src_stride, ref + j, ref_stride, &sse);
67 src += 2 * src_stride;
68 ref += 2 * ref_stride;
69 i -= 2;
70 } while (i != 0);
71 } else {
72 int i = height;
73 do {
74 int j = 0;
75 do {
76 sse_8x1_neon(src + j, ref + j, &sse);
77 j += 8;
78 } while (j < width);
79
80 src += src_stride;
81 ref += ref_stride;
82 } while (--i != 0);
83 }
84 return horizontal_add_u32x4(sse);
85 }
86
sse_128xh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int height)87 static INLINE uint32_t sse_128xh_neon(const uint8_t *src, int src_stride,
88 const uint8_t *ref, int ref_stride,
89 int height) {
90 uint32x4_t sse[2] = { vdupq_n_u32(0), vdupq_n_u32(0) };
91
92 int i = height;
93 do {
94 sse_16x1_neon(src, ref, &sse[0]);
95 sse_16x1_neon(src + 16, ref + 16, &sse[1]);
96 sse_16x1_neon(src + 32, ref + 32, &sse[0]);
97 sse_16x1_neon(src + 48, ref + 48, &sse[1]);
98 sse_16x1_neon(src + 64, ref + 64, &sse[0]);
99 sse_16x1_neon(src + 80, ref + 80, &sse[1]);
100 sse_16x1_neon(src + 96, ref + 96, &sse[0]);
101 sse_16x1_neon(src + 112, ref + 112, &sse[1]);
102
103 src += src_stride;
104 ref += ref_stride;
105 } while (--i != 0);
106
107 return horizontal_add_u32x4(vaddq_u32(sse[0], sse[1]));
108 }
109
sse_64xh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int height)110 static INLINE uint32_t sse_64xh_neon(const uint8_t *src, int src_stride,
111 const uint8_t *ref, int ref_stride,
112 int height) {
113 uint32x4_t sse[2] = { vdupq_n_u32(0), vdupq_n_u32(0) };
114
115 int i = height;
116 do {
117 sse_16x1_neon(src, ref, &sse[0]);
118 sse_16x1_neon(src + 16, ref + 16, &sse[1]);
119 sse_16x1_neon(src + 32, ref + 32, &sse[0]);
120 sse_16x1_neon(src + 48, ref + 48, &sse[1]);
121
122 src += src_stride;
123 ref += ref_stride;
124 } while (--i != 0);
125
126 return horizontal_add_u32x4(vaddq_u32(sse[0], sse[1]));
127 }
128
sse_32xh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int height)129 static INLINE uint32_t sse_32xh_neon(const uint8_t *src, int src_stride,
130 const uint8_t *ref, int ref_stride,
131 int height) {
132 uint32x4_t sse[2] = { vdupq_n_u32(0), vdupq_n_u32(0) };
133
134 int i = height;
135 do {
136 sse_16x1_neon(src, ref, &sse[0]);
137 sse_16x1_neon(src + 16, ref + 16, &sse[1]);
138
139 src += src_stride;
140 ref += ref_stride;
141 } while (--i != 0);
142
143 return horizontal_add_u32x4(vaddq_u32(sse[0], sse[1]));
144 }
145
sse_16xh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int height)146 static INLINE uint32_t sse_16xh_neon(const uint8_t *src, int src_stride,
147 const uint8_t *ref, int ref_stride,
148 int height) {
149 uint32x4_t sse[2] = { vdupq_n_u32(0), vdupq_n_u32(0) };
150
151 int i = height;
152 do {
153 sse_16x1_neon(src, ref, &sse[0]);
154 src += src_stride;
155 ref += ref_stride;
156 sse_16x1_neon(src, ref, &sse[1]);
157 src += src_stride;
158 ref += ref_stride;
159 i -= 2;
160 } while (i != 0);
161
162 return horizontal_add_u32x4(vaddq_u32(sse[0], sse[1]));
163 }
164
sse_8xh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int height)165 static INLINE uint32_t sse_8xh_neon(const uint8_t *src, int src_stride,
166 const uint8_t *ref, int ref_stride,
167 int height) {
168 uint32x4_t sse = vdupq_n_u32(0);
169
170 int i = height;
171 do {
172 sse_8x1_neon(src, ref, &sse);
173
174 src += src_stride;
175 ref += ref_stride;
176 } while (--i != 0);
177
178 return horizontal_add_u32x4(sse);
179 }
180
sse_4xh_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int height)181 static INLINE uint32_t sse_4xh_neon(const uint8_t *src, int src_stride,
182 const uint8_t *ref, int ref_stride,
183 int height) {
184 uint32x4_t sse = vdupq_n_u32(0);
185
186 int i = height;
187 do {
188 sse_4x2_neon(src, src_stride, ref, ref_stride, &sse);
189
190 src += 2 * src_stride;
191 ref += 2 * ref_stride;
192 i -= 2;
193 } while (i != 0);
194
195 return horizontal_add_u32x4(sse);
196 }
197
aom_sse_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int width,int height)198 int64_t aom_sse_neon(const uint8_t *src, int src_stride, const uint8_t *ref,
199 int ref_stride, int width, int height) {
200 switch (width) {
201 case 4: return sse_4xh_neon(src, src_stride, ref, ref_stride, height);
202 case 8: return sse_8xh_neon(src, src_stride, ref, ref_stride, height);
203 case 16: return sse_16xh_neon(src, src_stride, ref, ref_stride, height);
204 case 32: return sse_32xh_neon(src, src_stride, ref, ref_stride, height);
205 case 64: return sse_64xh_neon(src, src_stride, ref, ref_stride, height);
206 case 128: return sse_128xh_neon(src, src_stride, ref, ref_stride, height);
207 default:
208 return sse_wxh_neon(src, src_stride, ref, ref_stride, width, height);
209 }
210 }
211