xref: /aosp_15_r20/external/libaom/av1/common/arm/highbd_convolve_neon.h (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 #ifndef AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_NEON_H_
13 #define AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_NEON_H_
14 
15 #include <arm_neon.h>
16 
17 #include "aom_dsp/arm/mem_neon.h"
18 #include "aom_dsp/arm/transpose_neon.h"
19 #include "av1/common/convolve.h"
20 
highbd_convolve8_4_s32(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x8_t y_filter,const int32x4_t offset)21 static inline int32x4_t highbd_convolve8_4_s32(
22     const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
23     const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
24     const int16x4_t s6, const int16x4_t s7, const int16x8_t y_filter,
25     const int32x4_t offset) {
26   const int16x4_t y_filter_lo = vget_low_s16(y_filter);
27   const int16x4_t y_filter_hi = vget_high_s16(y_filter);
28 
29   int32x4_t sum = vmlal_lane_s16(offset, s0, y_filter_lo, 0);
30   sum = vmlal_lane_s16(sum, s1, y_filter_lo, 1);
31   sum = vmlal_lane_s16(sum, s2, y_filter_lo, 2);
32   sum = vmlal_lane_s16(sum, s3, y_filter_lo, 3);
33   sum = vmlal_lane_s16(sum, s4, y_filter_hi, 0);
34   sum = vmlal_lane_s16(sum, s5, y_filter_hi, 1);
35   sum = vmlal_lane_s16(sum, s6, y_filter_hi, 2);
36   sum = vmlal_lane_s16(sum, s7, y_filter_hi, 3);
37 
38   return sum;
39 }
40 
highbd_convolve8_4_sr_s32_s16(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x8_t y_filter,const int32x4_t shift_s32,const int32x4_t offset)41 static inline uint16x4_t highbd_convolve8_4_sr_s32_s16(
42     const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
43     const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
44     const int16x4_t s6, const int16x4_t s7, const int16x8_t y_filter,
45     const int32x4_t shift_s32, const int32x4_t offset) {
46   int32x4_t sum =
47       highbd_convolve8_4_s32(s0, s1, s2, s3, s4, s5, s6, s7, y_filter, offset);
48 
49   sum = vqrshlq_s32(sum, shift_s32);
50   return vqmovun_s32(sum);
51 }
52 
53 // Like above but also perform round shifting and subtract correction term
highbd_convolve8_4_srsub_s32_s16(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x8_t y_filter,const int32x4_t round_shift,const int32x4_t offset,const int32x4_t correction)54 static inline uint16x4_t highbd_convolve8_4_srsub_s32_s16(
55     const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
56     const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
57     const int16x4_t s6, const int16x4_t s7, const int16x8_t y_filter,
58     const int32x4_t round_shift, const int32x4_t offset,
59     const int32x4_t correction) {
60   int32x4_t sum =
61       highbd_convolve8_4_s32(s0, s1, s2, s3, s4, s5, s6, s7, y_filter, offset);
62 
63   sum = vsubq_s32(vqrshlq_s32(sum, round_shift), correction);
64   return vqmovun_s32(sum);
65 }
66 
highbd_convolve8_8_s32(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t s6,const int16x8_t s7,const int16x8_t y_filter,const int32x4_t offset,int32x4_t * sum0,int32x4_t * sum1)67 static inline void highbd_convolve8_8_s32(
68     const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
69     const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
70     const int16x8_t s6, const int16x8_t s7, const int16x8_t y_filter,
71     const int32x4_t offset, int32x4_t *sum0, int32x4_t *sum1) {
72   const int16x4_t y_filter_lo = vget_low_s16(y_filter);
73   const int16x4_t y_filter_hi = vget_high_s16(y_filter);
74 
75   *sum0 = vmlal_lane_s16(offset, vget_low_s16(s0), y_filter_lo, 0);
76   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s1), y_filter_lo, 1);
77   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s2), y_filter_lo, 2);
78   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s3), y_filter_lo, 3);
79   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s4), y_filter_hi, 0);
80   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s5), y_filter_hi, 1);
81   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s6), y_filter_hi, 2);
82   *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s7), y_filter_hi, 3);
83 
84   *sum1 = vmlal_lane_s16(offset, vget_high_s16(s0), y_filter_lo, 0);
85   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s1), y_filter_lo, 1);
86   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s2), y_filter_lo, 2);
87   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s3), y_filter_lo, 3);
88   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s4), y_filter_hi, 0);
89   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s5), y_filter_hi, 1);
90   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s6), y_filter_hi, 2);
91   *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s7), y_filter_hi, 3);
92 }
93 
94 // Like above but also perform round shifting and subtract correction term
highbd_convolve8_8_srsub_s32_s16(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t s6,const int16x8_t s7,const int16x8_t y_filter,const int32x4_t round_shift,const int32x4_t offset,const int32x4_t correction)95 static inline uint16x8_t highbd_convolve8_8_srsub_s32_s16(
96     const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
97     const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
98     const int16x8_t s6, const int16x8_t s7, const int16x8_t y_filter,
99     const int32x4_t round_shift, const int32x4_t offset,
100     const int32x4_t correction) {
101   int32x4_t sum0;
102   int32x4_t sum1;
103   highbd_convolve8_8_s32(s0, s1, s2, s3, s4, s5, s6, s7, y_filter, offset,
104                          &sum0, &sum1);
105 
106   sum0 = vsubq_s32(vqrshlq_s32(sum0, round_shift), correction);
107   sum1 = vsubq_s32(vqrshlq_s32(sum1, round_shift), correction);
108 
109   return vcombine_u16(vqmovun_s32(sum0), vqmovun_s32(sum1));
110 }
111 
highbd_convolve8_2d_scale_horiz4x8_s32(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x4_t * filters_lo,const int16x4_t * filters_hi,const int32x4_t offset)112 static inline int32x4_t highbd_convolve8_2d_scale_horiz4x8_s32(
113     const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
114     const int16x8_t s3, const int16x4_t *filters_lo,
115     const int16x4_t *filters_hi, const int32x4_t offset) {
116   int16x4_t s_lo[] = { vget_low_s16(s0), vget_low_s16(s1), vget_low_s16(s2),
117                        vget_low_s16(s3) };
118   int16x4_t s_hi[] = { vget_high_s16(s0), vget_high_s16(s1), vget_high_s16(s2),
119                        vget_high_s16(s3) };
120 
121   transpose_array_inplace_u16_4x4((uint16x4_t *)s_lo);
122   transpose_array_inplace_u16_4x4((uint16x4_t *)s_hi);
123 
124   int32x4_t sum = vmlal_s16(offset, s_lo[0], filters_lo[0]);
125   sum = vmlal_s16(sum, s_lo[1], filters_lo[1]);
126   sum = vmlal_s16(sum, s_lo[2], filters_lo[2]);
127   sum = vmlal_s16(sum, s_lo[3], filters_lo[3]);
128   sum = vmlal_s16(sum, s_hi[0], filters_hi[0]);
129   sum = vmlal_s16(sum, s_hi[1], filters_hi[1]);
130   sum = vmlal_s16(sum, s_hi[2], filters_hi[2]);
131   sum = vmlal_s16(sum, s_hi[3], filters_hi[3]);
132 
133   return sum;
134 }
135 
highbd_convolve8_2d_scale_horiz4x8_s32_s16(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x4_t * filters_lo,const int16x4_t * filters_hi,const int32x4_t shift_s32,const int32x4_t offset)136 static inline uint16x4_t highbd_convolve8_2d_scale_horiz4x8_s32_s16(
137     const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
138     const int16x8_t s3, const int16x4_t *filters_lo,
139     const int16x4_t *filters_hi, const int32x4_t shift_s32,
140     const int32x4_t offset) {
141   int32x4_t sum = highbd_convolve8_2d_scale_horiz4x8_s32(
142       s0, s1, s2, s3, filters_lo, filters_hi, offset);
143 
144   sum = vqrshlq_s32(sum, shift_s32);
145   return vqmovun_s32(sum);
146 }
147 
148 #endif  // AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_NEON_H_
149