1 /*
2  *  Copyright (c) 2014 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 #include "./vp9_rtcd.h"
13 #include "./vpx_config.h"
14 
15 #include "vpx/vpx_integer.h"
16 
horizontal_long_add_16x8(const uint16x8_t vec_lo,const uint16x8_t vec_hi)17 static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
18                                                     const uint16x8_t vec_hi) {
19   const uint32x4_t vec_l_lo = vaddl_u16(vget_low_u16(vec_lo),
20                                         vget_high_u16(vec_lo));
21   const uint32x4_t vec_l_hi = vaddl_u16(vget_low_u16(vec_hi),
22                                         vget_high_u16(vec_hi));
23   const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
24   const uint64x2_t b = vpaddlq_u32(a);
25   const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
26                                 vreinterpret_u32_u64(vget_high_u64(b)));
27   return vget_lane_u32(c, 0);
28 }
horizontal_add_16x8(const uint16x8_t vec_16x8)29 static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) {
30   const uint32x4_t a = vpaddlq_u16(vec_16x8);
31   const uint64x2_t b = vpaddlq_u32(a);
32   const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
33                                 vreinterpret_u32_u64(vget_high_u64(b)));
34   return vget_lane_u32(c, 0);
35 }
36 
vp9_sad64x64_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride)37 unsigned int vp9_sad64x64_neon(const uint8_t *src, int src_stride,
38                                const uint8_t *ref, int ref_stride) {
39   int i;
40   uint16x8_t vec_accum_lo = vdupq_n_u16(0);
41   uint16x8_t vec_accum_hi = vdupq_n_u16(0);
42   for (i = 0; i < 64; ++i) {
43     const uint8x16_t vec_src_00 = vld1q_u8(src);
44     const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
45     const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
46     const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
47     const uint8x16_t vec_ref_00 = vld1q_u8(ref);
48     const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
49     const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
50     const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
51     src += src_stride;
52     ref += ref_stride;
53     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
54                             vget_low_u8(vec_ref_00));
55     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
56                             vget_high_u8(vec_ref_00));
57     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
58                             vget_low_u8(vec_ref_16));
59     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
60                             vget_high_u8(vec_ref_16));
61     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32),
62                             vget_low_u8(vec_ref_32));
63     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32),
64                             vget_high_u8(vec_ref_32));
65     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48),
66                             vget_low_u8(vec_ref_48));
67     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48),
68                             vget_high_u8(vec_ref_48));
69   }
70   return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi);
71 }
72 
vp9_sad32x32_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride)73 unsigned int vp9_sad32x32_neon(const uint8_t *src, int src_stride,
74                                const uint8_t *ref, int ref_stride) {
75   int i;
76   uint16x8_t vec_accum_lo = vdupq_n_u16(0);
77   uint16x8_t vec_accum_hi = vdupq_n_u16(0);
78 
79   for (i = 0; i < 32; ++i) {
80     const uint8x16_t vec_src_00 = vld1q_u8(src);
81     const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
82     const uint8x16_t vec_ref_00 = vld1q_u8(ref);
83     const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
84     src += src_stride;
85     ref += ref_stride;
86     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
87                             vget_low_u8(vec_ref_00));
88     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
89                             vget_high_u8(vec_ref_00));
90     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
91                             vget_low_u8(vec_ref_16));
92     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
93                             vget_high_u8(vec_ref_16));
94   }
95   return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
96 }
97 
vp9_sad16x16_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride)98 unsigned int vp9_sad16x16_neon(const uint8_t *src, int src_stride,
99                                const uint8_t *ref, int ref_stride) {
100   int i;
101   uint16x8_t vec_accum_lo = vdupq_n_u16(0);
102   uint16x8_t vec_accum_hi = vdupq_n_u16(0);
103 
104   for (i = 0; i < 16; ++i) {
105     const uint8x16_t vec_src = vld1q_u8(src);
106     const uint8x16_t vec_ref = vld1q_u8(ref);
107     src += src_stride;
108     ref += ref_stride;
109     vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src),
110                             vget_low_u8(vec_ref));
111     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src),
112                             vget_high_u8(vec_ref));
113   }
114   return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
115 }
116 
vp9_sad8x8_neon(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride)117 unsigned int vp9_sad8x8_neon(const uint8_t *src, int src_stride,
118                              const uint8_t *ref, int ref_stride) {
119   int i;
120   uint16x8_t vec_accum = vdupq_n_u16(0);
121 
122   for (i = 0; i < 8; ++i) {
123     const uint8x8_t vec_src = vld1_u8(src);
124     const uint8x8_t vec_ref = vld1_u8(ref);
125     src += src_stride;
126     ref += ref_stride;
127     vec_accum = vabal_u8(vec_accum, vec_src, vec_ref);
128   }
129   return horizontal_add_16x8(vec_accum);
130 }
131