xref: /aosp_15_r20/external/libaom/aom_dsp/arm/sum_squares_neon_dotprod.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2023, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include <arm_neon.h>
13 #include <assert.h>
14 
15 #include "aom_dsp/arm/mem_neon.h"
16 #include "aom_dsp/arm/sum_neon.h"
17 #include "config/aom_dsp_rtcd.h"
18 
aom_var_2d_u8_4xh_neon_dotprod(uint8_t * src,int src_stride,int width,int height)19 static inline uint64_t aom_var_2d_u8_4xh_neon_dotprod(uint8_t *src,
20                                                       int src_stride, int width,
21                                                       int height) {
22   uint64_t sum = 0;
23   uint64_t sse = 0;
24   uint32x2_t sum_u32 = vdup_n_u32(0);
25   uint32x2_t sse_u32 = vdup_n_u32(0);
26 
27   int h = height / 2;
28   do {
29     int w = width;
30     uint8_t *src_ptr = src;
31     do {
32       uint8x8_t s0 = load_unaligned_u8(src_ptr, src_stride);
33 
34       sum_u32 = vdot_u32(sum_u32, s0, vdup_n_u8(1));
35 
36       sse_u32 = vdot_u32(sse_u32, s0, s0);
37 
38       src_ptr += 8;
39       w -= 8;
40     } while (w >= 8);
41 
42     // Process remaining columns in the row using C.
43     while (w > 0) {
44       int idx = width - w;
45       const uint8_t v = src[idx];
46       sum += v;
47       sse += v * v;
48       w--;
49     }
50 
51     src += 2 * src_stride;
52   } while (--h != 0);
53 
54   sum += horizontal_long_add_u32x2(sum_u32);
55   sse += horizontal_long_add_u32x2(sse_u32);
56 
57   return sse - sum * sum / (width * height);
58 }
59 
aom_var_2d_u8_8xh_neon_dotprod(uint8_t * src,int src_stride,int width,int height)60 static inline uint64_t aom_var_2d_u8_8xh_neon_dotprod(uint8_t *src,
61                                                       int src_stride, int width,
62                                                       int height) {
63   uint64_t sum = 0;
64   uint64_t sse = 0;
65   uint32x2_t sum_u32 = vdup_n_u32(0);
66   uint32x2_t sse_u32 = vdup_n_u32(0);
67 
68   int h = height;
69   do {
70     int w = width;
71     uint8_t *src_ptr = src;
72     do {
73       uint8x8_t s0 = vld1_u8(src_ptr);
74 
75       sum_u32 = vdot_u32(sum_u32, s0, vdup_n_u8(1));
76 
77       sse_u32 = vdot_u32(sse_u32, s0, s0);
78 
79       src_ptr += 8;
80       w -= 8;
81     } while (w >= 8);
82 
83     // Process remaining columns in the row using C.
84     while (w > 0) {
85       int idx = width - w;
86       const uint8_t v = src[idx];
87       sum += v;
88       sse += v * v;
89       w--;
90     }
91 
92     src += src_stride;
93   } while (--h != 0);
94 
95   sum += horizontal_long_add_u32x2(sum_u32);
96   sse += horizontal_long_add_u32x2(sse_u32);
97 
98   return sse - sum * sum / (width * height);
99 }
100 
aom_var_2d_u8_16xh_neon_dotprod(uint8_t * src,int src_stride,int width,int height)101 static inline uint64_t aom_var_2d_u8_16xh_neon_dotprod(uint8_t *src,
102                                                        int src_stride,
103                                                        int width, int height) {
104   uint64_t sum = 0;
105   uint64_t sse = 0;
106   uint32x4_t sum_u32 = vdupq_n_u32(0);
107   uint32x4_t sse_u32 = vdupq_n_u32(0);
108 
109   int h = height;
110   do {
111     int w = width;
112     uint8_t *src_ptr = src;
113     do {
114       uint8x16_t s0 = vld1q_u8(src_ptr);
115 
116       sum_u32 = vdotq_u32(sum_u32, s0, vdupq_n_u8(1));
117 
118       sse_u32 = vdotq_u32(sse_u32, s0, s0);
119 
120       src_ptr += 16;
121       w -= 16;
122     } while (w >= 16);
123 
124     // Process remaining columns in the row using C.
125     while (w > 0) {
126       int idx = width - w;
127       const uint8_t v = src[idx];
128       sum += v;
129       sse += v * v;
130       w--;
131     }
132 
133     src += src_stride;
134   } while (--h != 0);
135 
136   sum += horizontal_long_add_u32x4(sum_u32);
137   sse += horizontal_long_add_u32x4(sse_u32);
138 
139   return sse - sum * sum / (width * height);
140 }
141 
aom_var_2d_u8_neon_dotprod(uint8_t * src,int src_stride,int width,int height)142 uint64_t aom_var_2d_u8_neon_dotprod(uint8_t *src, int src_stride, int width,
143                                     int height) {
144   if (width >= 16) {
145     return aom_var_2d_u8_16xh_neon_dotprod(src, src_stride, width, height);
146   }
147   if (width >= 8) {
148     return aom_var_2d_u8_8xh_neon_dotprod(src, src_stride, width, height);
149   }
150   if (width >= 4 && height % 2 == 0) {
151     return aom_var_2d_u8_4xh_neon_dotprod(src, src_stride, width, height);
152   }
153   return aom_var_2d_u8_c(src, src_stride, width, height);
154 }
155