Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 1 | /* |
Krishna Rapaka | 7319db5 | 2021-09-28 20:35:29 -0700 | [diff] [blame] | 2 | * Copyright (c) 2021, Alliance for Open Media. All rights reserved |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 3 | * |
Vibhoothi | 41c6dd7 | 2021-10-12 18:48:26 +0000 | [diff] [blame] | 4 | * This source code is subject to the terms of the BSD 3-Clause Clear License |
| 5 | * and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear |
| 6 | * License was not distributed with this source code in the LICENSE file, you |
| 7 | * can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. If the |
| 8 | * Alliance for Open Media Patent License 1.0 was not distributed with this |
| 9 | * source code in the PATENTS file, you can obtain it at |
| 10 | * aomedia.org/license/patent-license/. |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 11 | */ |
| 12 | |
James Zern | e1cbb13 | 2018-08-22 14:10:36 -0700 | [diff] [blame] | 13 | #ifndef AOM_AOM_DSP_SIMD_V256_INTRINSICS_H_ |
| 14 | #define AOM_AOM_DSP_SIMD_V256_INTRINSICS_H_ |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 15 | |
| 16 | #include <stdio.h> |
| 17 | #include <stdlib.h> |
| 18 | #include <string.h> |
Tom Finegan | dd3e2a5 | 2018-05-23 14:33:09 -0700 | [diff] [blame] | 19 | |
| 20 | #include "aom_dsp/simd/v256_intrinsics_c.h" |
| 21 | #include "aom_dsp/simd/v128_intrinsics.h" |
| 22 | #include "aom_dsp/simd/v64_intrinsics.h" |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 23 | |
| 24 | /* Fallback to plain, unoptimised C. */ |
| 25 | |
| 26 | typedef c_v256 v256; |
| 27 | |
| 28 | SIMD_INLINE uint32_t v256_low_u32(v256 a) { return c_v256_low_u32(a); } |
| 29 | SIMD_INLINE v64 v256_low_v64(v256 a) { return c_v256_low_v64(a); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 30 | SIMD_INLINE uint64_t v256_low_u64(v256 a) { return c_v256_low_u64(a); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 31 | SIMD_INLINE v128 v256_low_v128(v256 a) { return c_v256_low_v128(a); } |
| 32 | SIMD_INLINE v128 v256_high_v128(v256 a) { return c_v256_high_v128(a); } |
| 33 | SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) { |
| 34 | return c_v256_from_v128(hi, lo); |
| 35 | } |
| 36 | SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) { |
| 37 | return c_v256_from_64(a, b, c, d); |
| 38 | } |
| 39 | SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) { |
| 40 | return c_v256_from_v64(a, b, c, d); |
| 41 | } |
| 42 | |
| 43 | SIMD_INLINE v256 v256_load_unaligned(const void *p) { |
| 44 | return c_v256_load_unaligned(p); |
| 45 | } |
| 46 | SIMD_INLINE v256 v256_load_aligned(const void *p) { |
| 47 | return c_v256_load_aligned(p); |
| 48 | } |
| 49 | |
| 50 | SIMD_INLINE void v256_store_unaligned(void *p, v256 a) { |
| 51 | c_v256_store_unaligned(p, a); |
| 52 | } |
| 53 | SIMD_INLINE void v256_store_aligned(void *p, v256 a) { |
| 54 | c_v256_store_aligned(p, a); |
| 55 | } |
| 56 | |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 57 | SIMD_INLINE v256 v256_align(v256 a, v256 b, unsigned int c) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 58 | return c_v256_align(a, b, c); |
| 59 | } |
| 60 | |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 61 | SIMD_INLINE v256 v256_zero(void) { return c_v256_zero(); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 62 | SIMD_INLINE v256 v256_dup_8(uint8_t x) { return c_v256_dup_8(x); } |
| 63 | SIMD_INLINE v256 v256_dup_16(uint16_t x) { return c_v256_dup_16(x); } |
| 64 | SIMD_INLINE v256 v256_dup_32(uint32_t x) { return c_v256_dup_32(x); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 65 | SIMD_INLINE v256 v256_dup_64(uint64_t x) { return c_v256_dup_64(x); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 66 | |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 67 | SIMD_INLINE c_sad256_internal v256_sad_u8_init(void) { |
| 68 | return c_v256_sad_u8_init(); |
| 69 | } |
| 70 | SIMD_INLINE c_sad256_internal v256_sad_u8(c_sad256_internal s, v256 a, v256 b) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 71 | return c_v256_sad_u8(s, a, b); |
| 72 | } |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 73 | SIMD_INLINE uint32_t v256_sad_u8_sum(c_sad256_internal s) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 74 | return c_v256_sad_u8_sum(s); |
| 75 | } |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 76 | SIMD_INLINE c_ssd256_internal v256_ssd_u8_init(void) { |
| 77 | return c_v256_ssd_u8_init(); |
| 78 | } |
| 79 | SIMD_INLINE c_ssd256_internal v256_ssd_u8(c_ssd256_internal s, v256 a, v256 b) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 80 | return c_v256_ssd_u8(s, a, b); |
| 81 | } |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 82 | SIMD_INLINE uint32_t v256_ssd_u8_sum(c_ssd256_internal s) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 83 | return c_v256_ssd_u8_sum(s); |
| 84 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 85 | |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 86 | SIMD_INLINE c_ssd256_internal_s16 v256_ssd_s16_init(void) { |
| 87 | return c_v256_ssd_s16_init(); |
| 88 | } |
| 89 | SIMD_INLINE c_ssd256_internal_s16 v256_ssd_s16(c_ssd256_internal_s16 s, v256 a, |
| 90 | v256 b) { |
| 91 | return c_v256_ssd_s16(s, a, b); |
| 92 | } |
Wan-Teh Chang | e63182c | 2020-03-29 08:43:13 -0700 | [diff] [blame] | 93 | SIMD_INLINE uint64_t v256_ssd_s16_sum(c_ssd256_internal_s16 s) { |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 94 | return c_v256_ssd_s16_sum(s); |
| 95 | } |
| 96 | |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 97 | SIMD_INLINE int64_t v256_dotp_su8(v256 a, v256 b) { |
| 98 | return c_v256_dotp_su8(a, b); |
| 99 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 100 | SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) { |
| 101 | return c_v256_dotp_s16(a, b); |
| 102 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 103 | SIMD_INLINE int64_t v256_dotp_s32(v256 a, v256 b) { |
| 104 | return c_v256_dotp_s32(a, b); |
| 105 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 106 | SIMD_INLINE uint64_t v256_hadd_u8(v256 a) { return c_v256_hadd_u8(a); } |
| 107 | |
| 108 | SIMD_INLINE v256 v256_or(v256 a, v256 b) { return c_v256_or(a, b); } |
| 109 | SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return c_v256_xor(a, b); } |
| 110 | SIMD_INLINE v256 v256_and(v256 a, v256 b) { return c_v256_and(a, b); } |
| 111 | SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return c_v256_andn(a, b); } |
| 112 | |
| 113 | SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return c_v256_add_8(a, b); } |
| 114 | SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return c_v256_add_16(a, b); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 115 | SIMD_INLINE v256 v256_sadd_s8(v256 a, v256 b) { return c_v256_sadd_s8(a, b); } |
| 116 | SIMD_INLINE v256 v256_sadd_u8(v256 a, v256 b) { return c_v256_sadd_u8(a, b); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 117 | SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) { return c_v256_sadd_s16(a, b); } |
| 118 | SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return c_v256_add_32(a, b); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 119 | SIMD_INLINE v256 v256_add_64(v256 a, v256 b) { return c_v256_add_64(a, b); } |
| 120 | SIMD_INLINE v256 v256_sub_64(v256 a, v256 b) { return c_v256_sub_64(a, b); } |
| 121 | SIMD_INLINE v256 v256_padd_u8(v256 a) { return c_v256_padd_u8(a); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 122 | SIMD_INLINE v256 v256_padd_s16(v256 a) { return c_v256_padd_s16(a); } |
| 123 | SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return c_v256_sub_8(a, b); } |
| 124 | SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return c_v256_ssub_u8(a, b); } |
| 125 | SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return c_v256_ssub_s8(a, b); } |
| 126 | SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return c_v256_sub_16(a, b); } |
| 127 | SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) { return c_v256_ssub_s16(a, b); } |
Steinar Midtskogen | 9b8444a | 2017-03-31 22:11:06 +0200 | [diff] [blame] | 128 | SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) { return c_v256_ssub_u16(a, b); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 129 | SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return c_v256_sub_32(a, b); } |
| 130 | SIMD_INLINE v256 v256_abs_s16(v256 a) { return c_v256_abs_s16(a); } |
Steinar Midtskogen | 6033fb8 | 2017-04-02 21:32:41 +0200 | [diff] [blame] | 131 | SIMD_INLINE v256 v256_abs_s8(v256 a) { return c_v256_abs_s8(a); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 132 | |
| 133 | SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) { return c_v256_mul_s16(a, b); } |
| 134 | SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) { |
| 135 | return c_v256_mullo_s16(a, b); |
| 136 | } |
| 137 | SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) { |
| 138 | return c_v256_mulhi_s16(a, b); |
| 139 | } |
| 140 | SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) { |
| 141 | return c_v256_mullo_s32(a, b); |
| 142 | } |
| 143 | SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) { return c_v256_madd_s16(a, b); } |
| 144 | SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) { return c_v256_madd_us8(a, b); } |
| 145 | |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 146 | SIMD_INLINE uint32_t v256_movemask_8(v256 a) { return c_v256_movemask_8(a); } |
| 147 | SIMD_INLINE v256 v256_blend_8(v256 a, v256 b, v256 c) { |
| 148 | return c_v256_blend_8(a, b, c); |
| 149 | } |
| 150 | |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 151 | SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return c_v256_avg_u8(a, b); } |
| 152 | SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) { return c_v256_rdavg_u8(a, b); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 153 | SIMD_INLINE v256 v256_rdavg_u16(v256 a, v256 b) { |
| 154 | return c_v256_rdavg_u16(a, b); |
| 155 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 156 | SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return c_v256_avg_u16(a, b); } |
| 157 | SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return c_v256_min_u8(a, b); } |
| 158 | SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return c_v256_max_u8(a, b); } |
| 159 | SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return c_v256_min_s8(a, b); } |
| 160 | SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return c_v256_max_s8(a, b); } |
| 161 | SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return c_v256_min_s16(a, b); } |
| 162 | SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return c_v256_max_s16(a, b); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 163 | SIMD_INLINE v256 v256_min_s32(v256 a, v256 b) { return c_v256_min_s32(a, b); } |
| 164 | SIMD_INLINE v256 v256_max_s32(v256 a, v256 b) { return c_v256_max_s32(a, b); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 165 | |
| 166 | SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) { return c_v256_ziplo_8(a, b); } |
| 167 | SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) { return c_v256_ziphi_8(a, b); } |
| 168 | SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) { return c_v256_ziplo_16(a, b); } |
| 169 | SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) { return c_v256_ziphi_16(a, b); } |
| 170 | SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) { return c_v256_ziplo_32(a, b); } |
| 171 | SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) { return c_v256_ziphi_32(a, b); } |
| 172 | SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) { return c_v256_ziplo_64(a, b); } |
| 173 | SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) { return c_v256_ziphi_64(a, b); } |
| 174 | SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) { |
| 175 | return c_v256_ziplo_128(a, b); |
| 176 | } |
| 177 | SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) { |
| 178 | return c_v256_ziphi_128(a, b); |
| 179 | } |
| 180 | SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) { return c_v256_zip_8(a, b); } |
| 181 | SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) { return c_v256_zip_16(a, b); } |
| 182 | SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) { return c_v256_zip_32(a, b); } |
| 183 | SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) { |
| 184 | return c_v256_unziplo_8(a, b); |
| 185 | } |
| 186 | SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) { |
| 187 | return c_v256_unziphi_8(a, b); |
| 188 | } |
| 189 | SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) { |
| 190 | return c_v256_unziplo_16(a, b); |
| 191 | } |
| 192 | SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) { |
| 193 | return c_v256_unziphi_16(a, b); |
| 194 | } |
| 195 | SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) { |
| 196 | return c_v256_unziplo_32(a, b); |
| 197 | } |
| 198 | SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) { |
| 199 | return c_v256_unziphi_32(a, b); |
| 200 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 201 | SIMD_INLINE v256 v256_unziplo_64(v256 a, v256 b) { |
| 202 | return c_v256_unziplo_64(a, b); |
| 203 | } |
| 204 | SIMD_INLINE v256 v256_unziphi_64(v256 a, v256 b) { |
| 205 | return c_v256_unziphi_64(a, b); |
| 206 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 207 | SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) { return c_v256_unpack_u8_s16(a); } |
| 208 | SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) { |
| 209 | return c_v256_unpacklo_u8_s16(a); |
| 210 | } |
| 211 | SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) { |
| 212 | return c_v256_unpackhi_u8_s16(a); |
| 213 | } |
Steinar Midtskogen | 1b2b739 | 2017-04-11 14:19:20 +0200 | [diff] [blame] | 214 | SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) { return c_v256_unpack_s8_s16(a); } |
| 215 | SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) { |
| 216 | return c_v256_unpacklo_s8_s16(a); |
| 217 | } |
| 218 | SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) { |
| 219 | return c_v256_unpackhi_s8_s16(a); |
| 220 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 221 | SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) { |
| 222 | return c_v256_pack_s32_s16(a, b); |
| 223 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 224 | SIMD_INLINE v256 v256_pack_s32_u16(v256 a, v256 b) { |
| 225 | return c_v256_pack_s32_u16(a, b); |
| 226 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 227 | SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) { |
| 228 | return c_v256_pack_s16_u8(a, b); |
| 229 | } |
| 230 | SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) { |
| 231 | return c_v256_pack_s16_s8(a, b); |
| 232 | } |
| 233 | SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) { |
| 234 | return c_v256_unpack_u16_s32(a); |
| 235 | } |
| 236 | SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) { |
| 237 | return c_v256_unpack_s16_s32(a); |
| 238 | } |
| 239 | SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) { |
| 240 | return c_v256_unpacklo_u16_s32(a); |
| 241 | } |
| 242 | SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) { |
| 243 | return c_v256_unpacklo_s16_s32(a); |
| 244 | } |
| 245 | SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) { |
| 246 | return c_v256_unpackhi_u16_s32(a); |
| 247 | } |
| 248 | SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) { |
| 249 | return c_v256_unpackhi_s16_s32(a); |
| 250 | } |
| 251 | SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) { |
| 252 | return c_v256_shuffle_8(a, pattern); |
| 253 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 254 | SIMD_INLINE v256 v256_wideshuffle_8(v256 a, v256 b, v256 pattern) { |
| 255 | return c_v256_wideshuffle_8(a, b, pattern); |
| 256 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 257 | SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) { |
| 258 | return c_v256_pshuffle_8(a, pattern); |
| 259 | } |
| 260 | |
| 261 | SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) { return c_v256_cmpgt_s8(a, b); } |
| 262 | SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) { return c_v256_cmplt_s8(a, b); } |
| 263 | SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) { return c_v256_cmpeq_8(a, b); } |
| 264 | SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) { |
| 265 | return c_v256_cmpgt_s16(a, b); |
| 266 | } |
| 267 | SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) { |
| 268 | return c_v256_cmplt_s16(a, b); |
| 269 | } |
| 270 | SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) { return c_v256_cmpeq_16(a, b); } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 271 | SIMD_INLINE v256 v256_cmpeq_32(v256 a, v256 b) { return c_v256_cmpeq_32(a, b); } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 272 | |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 273 | SIMD_INLINE v256 v256_cmpgt_s32(v256 a, v256 b) { |
| 274 | return c_v256_cmpgt_s32(a, b); |
| 275 | } |
| 276 | SIMD_INLINE v256 v256_cmplt_s32(v256 a, v256 b) { |
| 277 | return c_v256_cmplt_s32(a, b); |
| 278 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 279 | SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) { |
| 280 | return c_v256_shl_8(a, c); |
| 281 | } |
| 282 | SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) { |
| 283 | return c_v256_shr_u8(a, c); |
| 284 | } |
| 285 | SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) { |
| 286 | return c_v256_shr_s8(a, c); |
| 287 | } |
| 288 | SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) { |
| 289 | return c_v256_shl_16(a, c); |
| 290 | } |
| 291 | SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) { |
| 292 | return c_v256_shr_u16(a, c); |
| 293 | } |
| 294 | SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) { |
| 295 | return c_v256_shr_s16(a, c); |
| 296 | } |
| 297 | SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) { |
| 298 | return c_v256_shl_32(a, c); |
| 299 | } |
| 300 | SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) { |
| 301 | return c_v256_shr_u32(a, c); |
| 302 | } |
| 303 | SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) { |
| 304 | return c_v256_shr_s32(a, c); |
| 305 | } |
Steinar Midtskogen | d8f9521 | 2018-06-27 19:41:11 +0200 | [diff] [blame] | 306 | SIMD_INLINE v256 v256_shl_64(v256 a, unsigned int c) { |
| 307 | return c_v256_shl_64(a, c); |
| 308 | } |
| 309 | SIMD_INLINE v256 v256_shr_u64(v256 a, unsigned int c) { |
| 310 | return c_v256_shr_u64(a, c); |
| 311 | } |
| 312 | SIMD_INLINE v256 v256_shr_s64(v256 a, unsigned int c) { |
| 313 | return c_v256_shr_s64(a, c); |
| 314 | } |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 315 | |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 316 | SIMD_INLINE v256 v256_shr_n_byte(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 317 | return c_v256_shr_n_byte(a, n); |
| 318 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 319 | SIMD_INLINE v256 v256_shl_n_byte(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 320 | return c_v256_shl_n_byte(a, n); |
| 321 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 322 | SIMD_INLINE v256 v256_shl_n_8(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 323 | return c_v256_shl_n_8(a, n); |
| 324 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 325 | SIMD_INLINE v256 v256_shl_n_16(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 326 | return c_v256_shl_n_16(a, n); |
| 327 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 328 | SIMD_INLINE v256 v256_shl_n_32(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 329 | return c_v256_shl_n_32(a, n); |
| 330 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 331 | SIMD_INLINE v256 v256_shl_n_64(v256 a, unsigned int n) { |
| 332 | return c_v256_shl_n_64(a, n); |
| 333 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 334 | SIMD_INLINE v256 v256_shr_n_u8(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 335 | return c_v256_shr_n_u8(a, n); |
| 336 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 337 | SIMD_INLINE v256 v256_shr_n_u16(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 338 | return c_v256_shr_n_u16(a, n); |
| 339 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 340 | SIMD_INLINE v256 v256_shr_n_u32(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 341 | return c_v256_shr_n_u32(a, n); |
| 342 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 343 | SIMD_INLINE v256 v256_shr_n_u64(v256 a, unsigned int n) { |
| 344 | return c_v256_shr_n_u64(a, n); |
| 345 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 346 | SIMD_INLINE v256 v256_shr_n_s8(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 347 | return c_v256_shr_n_s8(a, n); |
| 348 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 349 | SIMD_INLINE v256 v256_shr_n_s16(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 350 | return c_v256_shr_n_s16(a, n); |
| 351 | } |
Yaowu Xu | 032573d | 2017-04-24 15:04:17 -0700 | [diff] [blame] | 352 | SIMD_INLINE v256 v256_shr_n_s32(v256 a, unsigned int n) { |
Steinar Midtskogen | 045d413 | 2016-10-18 12:20:05 +0200 | [diff] [blame] | 353 | return c_v256_shr_n_s32(a, n); |
| 354 | } |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 355 | SIMD_INLINE v256 v256_shr_n_s64(v256 a, unsigned int n) { |
| 356 | return c_v256_shr_n_s64(a, n); |
| 357 | } |
| 358 | |
| 359 | SIMD_INLINE v256 v256_shr_n_word(v256 a, unsigned int n) { |
| 360 | return c_v256_shr_n_word(a, n); |
| 361 | } |
| 362 | SIMD_INLINE v256 v256_shl_n_word(v256 a, unsigned int n) { |
| 363 | return c_v256_shl_n_word(a, n); |
| 364 | } |
| 365 | |
| 366 | typedef uint32_t sad256_internal_u16; |
Steinar Midtskogen | 50b2fc2 | 2020-03-24 14:23:51 +0100 | [diff] [blame] | 367 | SIMD_INLINE sad256_internal_u16 v256_sad_u16_init(void) { |
Steinar Midtskogen | 0578d43 | 2018-05-28 14:47:36 +0200 | [diff] [blame] | 368 | return c_v256_sad_u16_init(); |
| 369 | } |
| 370 | SIMD_INLINE sad256_internal_u16 v256_sad_u16(sad256_internal_u16 s, v256 a, |
| 371 | v256 b) { |
| 372 | return c_v256_sad_u16(s, a, b); |
| 373 | } |
| 374 | SIMD_INLINE uint32_t v256_sad_u16_sum(sad256_internal_u16 s) { |
| 375 | return c_v256_sad_u16_sum(s); |
| 376 | } |
| 377 | |
James Zern | e1cbb13 | 2018-08-22 14:10:36 -0700 | [diff] [blame] | 378 | #endif // AOM_AOM_DSP_SIMD_V256_INTRINSICS_H_ |