• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2023, 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/sum_neon.h"
15 
highbd_sse_8x1_init_neon(const uint16_t * src,const uint16_t * ref,uint32x4_t * sse_acc0,uint32x4_t * sse_acc1)16 static INLINE void highbd_sse_8x1_init_neon(const uint16_t *src,
17                                             const uint16_t *ref,
18                                             uint32x4_t *sse_acc0,
19                                             uint32x4_t *sse_acc1) {
20   uint16x8_t s = vld1q_u16(src);
21   uint16x8_t r = vld1q_u16(ref);
22 
23   uint16x8_t abs_diff = vabdq_u16(s, r);
24   uint16x4_t abs_diff_lo = vget_low_u16(abs_diff);
25   uint16x4_t abs_diff_hi = vget_high_u16(abs_diff);
26 
27   *sse_acc0 = vmull_u16(abs_diff_lo, abs_diff_lo);
28   *sse_acc1 = vmull_u16(abs_diff_hi, abs_diff_hi);
29 }
30 
highbd_sse_8x1_neon(const uint16_t * src,const uint16_t * ref,uint32x4_t * sse_acc0,uint32x4_t * sse_acc1)31 static INLINE void highbd_sse_8x1_neon(const uint16_t *src, const uint16_t *ref,
32                                        uint32x4_t *sse_acc0,
33                                        uint32x4_t *sse_acc1) {
34   uint16x8_t s = vld1q_u16(src);
35   uint16x8_t r = vld1q_u16(ref);
36 
37   uint16x8_t abs_diff = vabdq_u16(s, r);
38   uint16x4_t abs_diff_lo = vget_low_u16(abs_diff);
39   uint16x4_t abs_diff_hi = vget_high_u16(abs_diff);
40 
41   *sse_acc0 = vmlal_u16(*sse_acc0, abs_diff_lo, abs_diff_lo);
42   *sse_acc1 = vmlal_u16(*sse_acc1, abs_diff_hi, abs_diff_hi);
43 }
44 
highbd_sse_128xh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int height)45 static INLINE int64_t highbd_sse_128xh_neon(const uint16_t *src, int src_stride,
46                                             const uint16_t *ref, int ref_stride,
47                                             int height) {
48   uint32x4_t sse[16];
49   highbd_sse_8x1_init_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
50   highbd_sse_8x1_init_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
51   highbd_sse_8x1_init_neon(src + 2 * 8, ref + 2 * 8, &sse[4], &sse[5]);
52   highbd_sse_8x1_init_neon(src + 3 * 8, ref + 3 * 8, &sse[6], &sse[7]);
53   highbd_sse_8x1_init_neon(src + 4 * 8, ref + 4 * 8, &sse[8], &sse[9]);
54   highbd_sse_8x1_init_neon(src + 5 * 8, ref + 5 * 8, &sse[10], &sse[11]);
55   highbd_sse_8x1_init_neon(src + 6 * 8, ref + 6 * 8, &sse[12], &sse[13]);
56   highbd_sse_8x1_init_neon(src + 7 * 8, ref + 7 * 8, &sse[14], &sse[15]);
57   highbd_sse_8x1_neon(src + 8 * 8, ref + 8 * 8, &sse[0], &sse[1]);
58   highbd_sse_8x1_neon(src + 9 * 8, ref + 9 * 8, &sse[2], &sse[3]);
59   highbd_sse_8x1_neon(src + 10 * 8, ref + 10 * 8, &sse[4], &sse[5]);
60   highbd_sse_8x1_neon(src + 11 * 8, ref + 11 * 8, &sse[6], &sse[7]);
61   highbd_sse_8x1_neon(src + 12 * 8, ref + 12 * 8, &sse[8], &sse[9]);
62   highbd_sse_8x1_neon(src + 13 * 8, ref + 13 * 8, &sse[10], &sse[11]);
63   highbd_sse_8x1_neon(src + 14 * 8, ref + 14 * 8, &sse[12], &sse[13]);
64   highbd_sse_8x1_neon(src + 15 * 8, ref + 15 * 8, &sse[14], &sse[15]);
65 
66   src += src_stride;
67   ref += ref_stride;
68 
69   while (--height != 0) {
70     highbd_sse_8x1_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
71     highbd_sse_8x1_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
72     highbd_sse_8x1_neon(src + 2 * 8, ref + 2 * 8, &sse[4], &sse[5]);
73     highbd_sse_8x1_neon(src + 3 * 8, ref + 3 * 8, &sse[6], &sse[7]);
74     highbd_sse_8x1_neon(src + 4 * 8, ref + 4 * 8, &sse[8], &sse[9]);
75     highbd_sse_8x1_neon(src + 5 * 8, ref + 5 * 8, &sse[10], &sse[11]);
76     highbd_sse_8x1_neon(src + 6 * 8, ref + 6 * 8, &sse[12], &sse[13]);
77     highbd_sse_8x1_neon(src + 7 * 8, ref + 7 * 8, &sse[14], &sse[15]);
78     highbd_sse_8x1_neon(src + 8 * 8, ref + 8 * 8, &sse[0], &sse[1]);
79     highbd_sse_8x1_neon(src + 9 * 8, ref + 9 * 8, &sse[2], &sse[3]);
80     highbd_sse_8x1_neon(src + 10 * 8, ref + 10 * 8, &sse[4], &sse[5]);
81     highbd_sse_8x1_neon(src + 11 * 8, ref + 11 * 8, &sse[6], &sse[7]);
82     highbd_sse_8x1_neon(src + 12 * 8, ref + 12 * 8, &sse[8], &sse[9]);
83     highbd_sse_8x1_neon(src + 13 * 8, ref + 13 * 8, &sse[10], &sse[11]);
84     highbd_sse_8x1_neon(src + 14 * 8, ref + 14 * 8, &sse[12], &sse[13]);
85     highbd_sse_8x1_neon(src + 15 * 8, ref + 15 * 8, &sse[14], &sse[15]);
86 
87     src += src_stride;
88     ref += ref_stride;
89   }
90 
91   return horizontal_long_add_u32x4_x16(sse);
92 }
93 
highbd_sse_64xh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int height)94 static INLINE int64_t highbd_sse_64xh_neon(const uint16_t *src, int src_stride,
95                                            const uint16_t *ref, int ref_stride,
96                                            int height) {
97   uint32x4_t sse[8];
98   highbd_sse_8x1_init_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
99   highbd_sse_8x1_init_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
100   highbd_sse_8x1_init_neon(src + 2 * 8, ref + 2 * 8, &sse[4], &sse[5]);
101   highbd_sse_8x1_init_neon(src + 3 * 8, ref + 3 * 8, &sse[6], &sse[7]);
102   highbd_sse_8x1_neon(src + 4 * 8, ref + 4 * 8, &sse[0], &sse[1]);
103   highbd_sse_8x1_neon(src + 5 * 8, ref + 5 * 8, &sse[2], &sse[3]);
104   highbd_sse_8x1_neon(src + 6 * 8, ref + 6 * 8, &sse[4], &sse[5]);
105   highbd_sse_8x1_neon(src + 7 * 8, ref + 7 * 8, &sse[6], &sse[7]);
106 
107   src += src_stride;
108   ref += ref_stride;
109 
110   while (--height != 0) {
111     highbd_sse_8x1_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
112     highbd_sse_8x1_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
113     highbd_sse_8x1_neon(src + 2 * 8, ref + 2 * 8, &sse[4], &sse[5]);
114     highbd_sse_8x1_neon(src + 3 * 8, ref + 3 * 8, &sse[6], &sse[7]);
115     highbd_sse_8x1_neon(src + 4 * 8, ref + 4 * 8, &sse[0], &sse[1]);
116     highbd_sse_8x1_neon(src + 5 * 8, ref + 5 * 8, &sse[2], &sse[3]);
117     highbd_sse_8x1_neon(src + 6 * 8, ref + 6 * 8, &sse[4], &sse[5]);
118     highbd_sse_8x1_neon(src + 7 * 8, ref + 7 * 8, &sse[6], &sse[7]);
119 
120     src += src_stride;
121     ref += ref_stride;
122   }
123 
124   return horizontal_long_add_u32x4_x8(sse);
125 }
126 
highbd_sse_32xh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int height)127 static INLINE int64_t highbd_sse_32xh_neon(const uint16_t *src, int src_stride,
128                                            const uint16_t *ref, int ref_stride,
129                                            int height) {
130   uint32x4_t sse[8];
131   highbd_sse_8x1_init_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
132   highbd_sse_8x1_init_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
133   highbd_sse_8x1_init_neon(src + 2 * 8, ref + 2 * 8, &sse[4], &sse[5]);
134   highbd_sse_8x1_init_neon(src + 3 * 8, ref + 3 * 8, &sse[6], &sse[7]);
135 
136   src += src_stride;
137   ref += ref_stride;
138 
139   while (--height != 0) {
140     highbd_sse_8x1_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
141     highbd_sse_8x1_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
142     highbd_sse_8x1_neon(src + 2 * 8, ref + 2 * 8, &sse[4], &sse[5]);
143     highbd_sse_8x1_neon(src + 3 * 8, ref + 3 * 8, &sse[6], &sse[7]);
144 
145     src += src_stride;
146     ref += ref_stride;
147   }
148 
149   return horizontal_long_add_u32x4_x8(sse);
150 }
151 
highbd_sse_16xh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int height)152 static INLINE int64_t highbd_sse_16xh_neon(const uint16_t *src, int src_stride,
153                                            const uint16_t *ref, int ref_stride,
154                                            int height) {
155   uint32x4_t sse[4];
156   highbd_sse_8x1_init_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
157   highbd_sse_8x1_init_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
158 
159   src += src_stride;
160   ref += ref_stride;
161 
162   while (--height != 0) {
163     highbd_sse_8x1_neon(src + 0 * 8, ref + 0 * 8, &sse[0], &sse[1]);
164     highbd_sse_8x1_neon(src + 1 * 8, ref + 1 * 8, &sse[2], &sse[3]);
165 
166     src += src_stride;
167     ref += ref_stride;
168   }
169 
170   return horizontal_long_add_u32x4_x4(sse);
171 }
172 
highbd_sse_8xh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int height)173 static INLINE int64_t highbd_sse_8xh_neon(const uint16_t *src, int src_stride,
174                                           const uint16_t *ref, int ref_stride,
175                                           int height) {
176   uint32x4_t sse[2];
177   highbd_sse_8x1_init_neon(src, ref, &sse[0], &sse[1]);
178 
179   src += src_stride;
180   ref += ref_stride;
181 
182   while (--height != 0) {
183     highbd_sse_8x1_neon(src, ref, &sse[0], &sse[1]);
184 
185     src += src_stride;
186     ref += ref_stride;
187   }
188 
189   return horizontal_long_add_u32x4_x2(sse);
190 }
191 
highbd_sse_4xh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int height)192 static INLINE int64_t highbd_sse_4xh_neon(const uint16_t *src, int src_stride,
193                                           const uint16_t *ref, int ref_stride,
194                                           int height) {
195   // Peel the first loop iteration.
196   uint16x4_t s = vld1_u16(src);
197   uint16x4_t r = vld1_u16(ref);
198 
199   uint16x4_t abs_diff = vabd_u16(s, r);
200   uint32x4_t sse = vmull_u16(abs_diff, abs_diff);
201 
202   src += src_stride;
203   ref += ref_stride;
204 
205   while (--height != 0) {
206     s = vld1_u16(src);
207     r = vld1_u16(ref);
208 
209     abs_diff = vabd_u16(s, r);
210     sse = vmlal_u16(sse, abs_diff, abs_diff);
211 
212     src += src_stride;
213     ref += ref_stride;
214   }
215 
216   return horizontal_long_add_u32x4(sse);
217 }
218 
highbd_sse_wxh_neon(const uint16_t * src,int src_stride,const uint16_t * ref,int ref_stride,int width,int height)219 static INLINE int64_t highbd_sse_wxh_neon(const uint16_t *src, int src_stride,
220                                           const uint16_t *ref, int ref_stride,
221                                           int width, int height) {
222   // { 0, 1, 2, 3, 4, 5, 6, 7 }
223   uint16x8_t k01234567 = vmovl_u8(vcreate_u8(0x0706050403020100));
224   uint16x8_t remainder_mask = vcltq_u16(k01234567, vdupq_n_u16(width & 7));
225   uint64_t sse = 0;
226 
227   do {
228     int w = width;
229     int offset = 0;
230 
231     do {
232       uint16x8_t s = vld1q_u16(src + offset);
233       uint16x8_t r = vld1q_u16(ref + offset);
234 
235       if (w < 8) {
236         // Mask out-of-range elements.
237         s = vandq_u16(s, remainder_mask);
238         r = vandq_u16(r, remainder_mask);
239       }
240 
241       uint16x8_t abs_diff = vabdq_u16(s, r);
242       uint16x4_t abs_diff_lo = vget_low_u16(abs_diff);
243       uint16x4_t abs_diff_hi = vget_high_u16(abs_diff);
244 
245       uint32x4_t sse_u32 = vmull_u16(abs_diff_lo, abs_diff_lo);
246       sse_u32 = vmlal_u16(sse_u32, abs_diff_hi, abs_diff_hi);
247 
248       sse += horizontal_long_add_u32x4(sse_u32);
249 
250       offset += 8;
251       w -= 8;
252     } while (w > 0);
253 
254     src += src_stride;
255     ref += ref_stride;
256   } while (--height != 0);
257 
258   return sse;
259 }
260 
aom_highbd_sse_neon(const uint8_t * src8,int src_stride,const uint8_t * ref8,int ref_stride,int width,int height)261 int64_t aom_highbd_sse_neon(const uint8_t *src8, int src_stride,
262                             const uint8_t *ref8, int ref_stride, int width,
263                             int height) {
264   uint16_t *src = CONVERT_TO_SHORTPTR(src8);
265   uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
266 
267   switch (width) {
268     case 4:
269       return highbd_sse_4xh_neon(src, src_stride, ref, ref_stride, height);
270     case 8:
271       return highbd_sse_8xh_neon(src, src_stride, ref, ref_stride, height);
272     case 16:
273       return highbd_sse_16xh_neon(src, src_stride, ref, ref_stride, height);
274     case 32:
275       return highbd_sse_32xh_neon(src, src_stride, ref, ref_stride, height);
276     case 64:
277       return highbd_sse_64xh_neon(src, src_stride, ref, ref_stride, height);
278     case 128:
279       return highbd_sse_128xh_neon(src, src_stride, ref, ref_stride, height);
280     default:
281       return highbd_sse_wxh_neon(src, src_stride, ref, ref_stride, width,
282                                  height);
283   }
284 }
285