Mike Klein | f720098 | 2017-01-15 18:14:07 -0500 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright 2017 Google Inc. |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the LICENSE file. |
| 6 | */ |
| 7 | |
| 8 | #ifndef SkSplicer_generated_lowp_DEFINED |
| 9 | #define SkSplicer_generated_lowp_DEFINED |
| 10 | |
| 11 | // This file is generated semi-automatically with this command: |
| 12 | // $ src/splicer/build_stages.py |
| 13 | |
| 14 | #if defined(__aarch64__) |
| 15 | |
| 16 | static const unsigned int kSplice_inc_x_lowp[] = { |
| 17 | 0x91002000, // add x0, x0, #0x8 |
| 18 | }; |
| 19 | static const unsigned int kSplice_clear_lowp[] = { |
| 20 | 0x6f00e400, // movi v0.2d, #0x0 |
| 21 | 0x6f00e401, // movi v1.2d, #0x0 |
| 22 | 0x6f00e402, // movi v2.2d, #0x0 |
| 23 | 0x6f00e403, // movi v3.2d, #0x0 |
| 24 | }; |
| 25 | static const unsigned int kSplice_plus__lowp[] = { |
| 26 | 0x6e640c00, // uqadd v0.8h, v0.8h, v4.8h |
| 27 | 0x6e650c21, // uqadd v1.8h, v1.8h, v5.8h |
| 28 | 0x6e660c42, // uqadd v2.8h, v2.8h, v6.8h |
| 29 | 0x6e670c63, // uqadd v3.8h, v3.8h, v7.8h |
| 30 | }; |
| 31 | static const unsigned int kSplice_srcover_lowp[] = { |
| 32 | 0x91000868, // add x8, x3, #0x2 |
| 33 | 0x4d40c510, // ld1r {v16.8h}, [x8] |
| 34 | 0x6e632e10, // uqsub v16.8h, v16.8h, v3.8h |
| 35 | 0x6e70b491, // sqrdmulh v17.8h, v4.8h, v16.8h |
| 36 | 0x4e241e12, // and v18.16b, v16.16b, v4.16b |
| 37 | 0x6e70b4b3, // sqrdmulh v19.8h, v5.8h, v16.8h |
| 38 | 0x4e60ba31, // abs v17.8h, v17.8h |
| 39 | 0x4e251e14, // and v20.16b, v16.16b, v5.16b |
| 40 | 0x6f111651, // usra v17.8h, v18.8h, #15 |
| 41 | 0x6e70b4d2, // sqrdmulh v18.8h, v6.8h, v16.8h |
| 42 | 0x4e60ba73, // abs v19.8h, v19.8h |
| 43 | 0x6f111693, // usra v19.8h, v20.8h, #15 |
| 44 | 0x4e261e14, // and v20.16b, v16.16b, v6.16b |
| 45 | 0x4e60ba52, // abs v18.8h, v18.8h |
| 46 | 0x6f111692, // usra v18.8h, v20.8h, #15 |
| 47 | 0x6e70b4f4, // sqrdmulh v20.8h, v7.8h, v16.8h |
| 48 | 0x4e271e10, // and v16.16b, v16.16b, v7.16b |
| 49 | 0x4e60ba94, // abs v20.8h, v20.8h |
| 50 | 0x6f111614, // usra v20.8h, v16.8h, #15 |
| 51 | 0x6e600e20, // uqadd v0.8h, v17.8h, v0.8h |
| 52 | 0x6e610e61, // uqadd v1.8h, v19.8h, v1.8h |
| 53 | 0x6e620e42, // uqadd v2.8h, v18.8h, v2.8h |
| 54 | 0x6e630e83, // uqadd v3.8h, v20.8h, v3.8h |
| 55 | }; |
| 56 | static const unsigned int kSplice_dstover_lowp[] = { |
| 57 | 0x91000868, // add x8, x3, #0x2 |
| 58 | 0x4d40c510, // ld1r {v16.8h}, [x8] |
| 59 | 0x6e672e10, // uqsub v16.8h, v16.8h, v7.8h |
| 60 | 0x6e70b411, // sqrdmulh v17.8h, v0.8h, v16.8h |
| 61 | 0x4e201e12, // and v18.16b, v16.16b, v0.16b |
| 62 | 0x6e70b433, // sqrdmulh v19.8h, v1.8h, v16.8h |
| 63 | 0x4e60ba31, // abs v17.8h, v17.8h |
| 64 | 0x4e211e14, // and v20.16b, v16.16b, v1.16b |
| 65 | 0x6f111651, // usra v17.8h, v18.8h, #15 |
| 66 | 0x6e70b452, // sqrdmulh v18.8h, v2.8h, v16.8h |
| 67 | 0x4e60ba73, // abs v19.8h, v19.8h |
| 68 | 0x6f111693, // usra v19.8h, v20.8h, #15 |
| 69 | 0x4e221e14, // and v20.16b, v16.16b, v2.16b |
| 70 | 0x4e60ba52, // abs v18.8h, v18.8h |
| 71 | 0x6f111692, // usra v18.8h, v20.8h, #15 |
| 72 | 0x6e70b474, // sqrdmulh v20.8h, v3.8h, v16.8h |
| 73 | 0x4e231e10, // and v16.16b, v16.16b, v3.16b |
| 74 | 0x4e60ba94, // abs v20.8h, v20.8h |
| 75 | 0x6f111614, // usra v20.8h, v16.8h, #15 |
| 76 | 0x6e640e24, // uqadd v4.8h, v17.8h, v4.8h |
| 77 | 0x6e650e65, // uqadd v5.8h, v19.8h, v5.8h |
| 78 | 0x6e660e46, // uqadd v6.8h, v18.8h, v6.8h |
| 79 | 0x6e670e87, // uqadd v7.8h, v20.8h, v7.8h |
| 80 | }; |
| 81 | static const unsigned int kSplice_clamp_1_lowp[] = { |
| 82 | 0x91000868, // add x8, x3, #0x2 |
| 83 | 0x4d40c510, // ld1r {v16.8h}, [x8] |
| 84 | 0x6e706c00, // umin v0.8h, v0.8h, v16.8h |
| 85 | 0x6e706c21, // umin v1.8h, v1.8h, v16.8h |
| 86 | 0x6e706c42, // umin v2.8h, v2.8h, v16.8h |
| 87 | 0x6e706c63, // umin v3.8h, v3.8h, v16.8h |
| 88 | }; |
| 89 | static const unsigned int kSplice_clamp_a_lowp[] = { |
| 90 | 0x91000868, // add x8, x3, #0x2 |
| 91 | 0x4d40c510, // ld1r {v16.8h}, [x8] |
| 92 | 0x6e706c63, // umin v3.8h, v3.8h, v16.8h |
| 93 | 0x6e636c00, // umin v0.8h, v0.8h, v3.8h |
| 94 | 0x6e636c21, // umin v1.8h, v1.8h, v3.8h |
| 95 | 0x6e636c42, // umin v2.8h, v2.8h, v3.8h |
| 96 | }; |
| 97 | static const unsigned int kSplice_swap_lowp[] = { |
| 98 | 0x4ea31c70, // mov v16.16b, v3.16b |
| 99 | 0x4ea21c51, // mov v17.16b, v2.16b |
| 100 | 0x4ea11c32, // mov v18.16b, v1.16b |
| 101 | 0x4ea01c13, // mov v19.16b, v0.16b |
| 102 | 0x4ea41c80, // mov v0.16b, v4.16b |
| 103 | 0x4ea51ca1, // mov v1.16b, v5.16b |
| 104 | 0x4ea61cc2, // mov v2.16b, v6.16b |
| 105 | 0x4ea71ce3, // mov v3.16b, v7.16b |
| 106 | 0x4eb31e64, // mov v4.16b, v19.16b |
| 107 | 0x4eb21e45, // mov v5.16b, v18.16b |
| 108 | 0x4eb11e26, // mov v6.16b, v17.16b |
| 109 | 0x4eb01e07, // mov v7.16b, v16.16b |
| 110 | }; |
| 111 | static const unsigned int kSplice_move_src_dst_lowp[] = { |
| 112 | 0x4ea01c04, // mov v4.16b, v0.16b |
| 113 | 0x4ea11c25, // mov v5.16b, v1.16b |
| 114 | 0x4ea21c46, // mov v6.16b, v2.16b |
| 115 | 0x4ea31c67, // mov v7.16b, v3.16b |
| 116 | }; |
| 117 | static const unsigned int kSplice_move_dst_src_lowp[] = { |
| 118 | 0x4ea41c80, // mov v0.16b, v4.16b |
| 119 | 0x4ea51ca1, // mov v1.16b, v5.16b |
| 120 | 0x4ea61cc2, // mov v2.16b, v6.16b |
| 121 | 0x4ea71ce3, // mov v3.16b, v7.16b |
| 122 | }; |
| 123 | static const unsigned int kSplice_premul_lowp[] = { |
| 124 | 0x6e63b410, // sqrdmulh v16.8h, v0.8h, v3.8h |
| 125 | 0x4e201c71, // and v17.16b, v3.16b, v0.16b |
| 126 | 0x4e60ba00, // abs v0.8h, v16.8h |
| 127 | 0x6e63b430, // sqrdmulh v16.8h, v1.8h, v3.8h |
| 128 | 0x6f111620, // usra v0.8h, v17.8h, #15 |
| 129 | 0x4e211c71, // and v17.16b, v3.16b, v1.16b |
| 130 | 0x4e60ba01, // abs v1.8h, v16.8h |
| 131 | 0x6e63b450, // sqrdmulh v16.8h, v2.8h, v3.8h |
| 132 | 0x6f111621, // usra v1.8h, v17.8h, #15 |
| 133 | 0x4e221c71, // and v17.16b, v3.16b, v2.16b |
| 134 | 0x4e60ba02, // abs v2.8h, v16.8h |
| 135 | 0x6f111622, // usra v2.8h, v17.8h, #15 |
| 136 | }; |
| 137 | static const unsigned int kSplice_load_8888_lowp[] = { |
| 138 | 0xf9400048, // ldr x8, [x2] |
| 139 | 0x8b000908, // add x8, x8, x0, lsl #2 |
| 140 | 0x0c400110, // ld4 {v16.8b-v19.8b}, [x8] |
| 141 | 0x2f0fa600, // ushll v0.8h, v16.8b, #7 |
| 142 | 0x2f0fa621, // ushll v1.8h, v17.8b, #7 |
| 143 | 0x2f0fa642, // ushll v2.8h, v18.8b, #7 |
| 144 | 0x2f0fa663, // ushll v3.8h, v19.8b, #7 |
| 145 | 0x6f183400, // ursra v0.8h, v0.8h, #8 |
| 146 | 0x6f183421, // ursra v1.8h, v1.8h, #8 |
| 147 | 0x6f183442, // ursra v2.8h, v2.8h, #8 |
| 148 | 0x6f183463, // ursra v3.8h, v3.8h, #8 |
| 149 | }; |
| 150 | static const unsigned int kSplice_store_8888_lowp[] = { |
| 151 | 0xf9400048, // ldr x8, [x2] |
| 152 | 0x2f099410, // uqshrn v16.8b, v0.8h, #7 |
| 153 | 0x2f099431, // uqshrn v17.8b, v1.8h, #7 |
| 154 | 0x2f099452, // uqshrn v18.8b, v2.8h, #7 |
| 155 | 0x8b000908, // add x8, x8, x0, lsl #2 |
| 156 | 0x2f099473, // uqshrn v19.8b, v3.8h, #7 |
| 157 | 0x0c000110, // st4 {v16.8b-v19.8b}, [x8] |
| 158 | }; |
| 159 | |
| 160 | #elif defined(__ARM_NEON__) |
| 161 | |
| 162 | static const unsigned int kSplice_inc_x_lowp[] = { |
| 163 | 0xe2800004, // add r0, r0, #4 |
| 164 | }; |
| 165 | static const unsigned int kSplice_clear_lowp[] = { |
| 166 | 0xf2800010, // vmov.i32 d0, #0 |
| 167 | 0xf2801010, // vmov.i32 d1, #0 |
| 168 | 0xf2802010, // vmov.i32 d2, #0 |
| 169 | 0xf2803010, // vmov.i32 d3, #0 |
| 170 | }; |
| 171 | static const unsigned int kSplice_plus__lowp[] = { |
| 172 | 0xf3100014, // vqadd.u16 d0, d0, d4 |
| 173 | 0xf3111015, // vqadd.u16 d1, d1, d5 |
| 174 | 0xf3122016, // vqadd.u16 d2, d2, d6 |
| 175 | 0xf3133017, // vqadd.u16 d3, d3, d7 |
| 176 | }; |
| 177 | static const unsigned int kSplice_srcover_lowp[] = { |
| 178 | 0xe283c002, // add ip, r3, #2 |
| 179 | 0xf4ec0c5f, // vld1.16 {d16[]}, [ip :16] |
| 180 | 0xf3500293, // vqsub.u16 d16, d16, d3 |
| 181 | 0xf3541b20, // vqrdmulh.s16 d17, d4, d16 |
| 182 | 0xf3552b20, // vqrdmulh.s16 d18, d5, d16 |
| 183 | 0xf3563b20, // vqrdmulh.s16 d19, d6, d16 |
| 184 | 0xf3574b20, // vqrdmulh.s16 d20, d7, d16 |
| 185 | 0xf2405194, // vand d21, d16, d4 |
| 186 | 0xf2406195, // vand d22, d16, d5 |
| 187 | 0xf2407196, // vand d23, d16, d6 |
| 188 | 0xf2400197, // vand d16, d16, d7 |
| 189 | 0xf3f51321, // vabs.s16 d17, d17 |
| 190 | 0xf3f52322, // vabs.s16 d18, d18 |
| 191 | 0xf3f53323, // vabs.s16 d19, d19 |
| 192 | 0xf3f54324, // vabs.s16 d20, d20 |
| 193 | 0xf3d11135, // vsra.u16 d17, d21, #15 |
| 194 | 0xf3d12136, // vsra.u16 d18, d22, #15 |
| 195 | 0xf3d13137, // vsra.u16 d19, d23, #15 |
| 196 | 0xf3d14130, // vsra.u16 d20, d16, #15 |
| 197 | 0xf3110090, // vqadd.u16 d0, d17, d0 |
| 198 | 0xf3121091, // vqadd.u16 d1, d18, d1 |
| 199 | 0xf3132092, // vqadd.u16 d2, d19, d2 |
| 200 | 0xf3143093, // vqadd.u16 d3, d20, d3 |
| 201 | }; |
| 202 | static const unsigned int kSplice_dstover_lowp[] = { |
| 203 | 0xe283c002, // add ip, r3, #2 |
| 204 | 0xf4ec0c5f, // vld1.16 {d16[]}, [ip :16] |
| 205 | 0xf3500297, // vqsub.u16 d16, d16, d7 |
| 206 | 0xf3501b20, // vqrdmulh.s16 d17, d0, d16 |
| 207 | 0xf3512b20, // vqrdmulh.s16 d18, d1, d16 |
| 208 | 0xf3523b20, // vqrdmulh.s16 d19, d2, d16 |
| 209 | 0xf3534b20, // vqrdmulh.s16 d20, d3, d16 |
| 210 | 0xf2405190, // vand d21, d16, d0 |
| 211 | 0xf2406191, // vand d22, d16, d1 |
| 212 | 0xf2407192, // vand d23, d16, d2 |
| 213 | 0xf2400193, // vand d16, d16, d3 |
| 214 | 0xf3f51321, // vabs.s16 d17, d17 |
| 215 | 0xf3f52322, // vabs.s16 d18, d18 |
| 216 | 0xf3f53323, // vabs.s16 d19, d19 |
| 217 | 0xf3f54324, // vabs.s16 d20, d20 |
| 218 | 0xf3d11135, // vsra.u16 d17, d21, #15 |
| 219 | 0xf3d12136, // vsra.u16 d18, d22, #15 |
| 220 | 0xf3d13137, // vsra.u16 d19, d23, #15 |
| 221 | 0xf3d14130, // vsra.u16 d20, d16, #15 |
| 222 | 0xf3114094, // vqadd.u16 d4, d17, d4 |
| 223 | 0xf3125095, // vqadd.u16 d5, d18, d5 |
| 224 | 0xf3136096, // vqadd.u16 d6, d19, d6 |
| 225 | 0xf3147097, // vqadd.u16 d7, d20, d7 |
| 226 | }; |
| 227 | static const unsigned int kSplice_clamp_1_lowp[] = { |
| 228 | 0xe283c002, // add ip, r3, #2 |
| 229 | 0xf4ec0c5f, // vld1.16 {d16[]}, [ip :16] |
| 230 | 0xf3100630, // vmin.u16 d0, d0, d16 |
| 231 | 0xf3111630, // vmin.u16 d1, d1, d16 |
| 232 | 0xf3122630, // vmin.u16 d2, d2, d16 |
| 233 | 0xf3133630, // vmin.u16 d3, d3, d16 |
| 234 | }; |
| 235 | static const unsigned int kSplice_clamp_a_lowp[] = { |
| 236 | 0xe283c002, // add ip, r3, #2 |
| 237 | 0xf4ec0c5f, // vld1.16 {d16[]}, [ip :16] |
| 238 | 0xf3133630, // vmin.u16 d3, d3, d16 |
| 239 | 0xf3100613, // vmin.u16 d0, d0, d3 |
| 240 | 0xf3111613, // vmin.u16 d1, d1, d3 |
| 241 | 0xf3122613, // vmin.u16 d2, d2, d3 |
| 242 | }; |
| 243 | static const unsigned int kSplice_swap_lowp[] = { |
| 244 | 0xeef00b43, // vmov.f64 d16, d3 |
| 245 | 0xeef01b42, // vmov.f64 d17, d2 |
| 246 | 0xeef02b41, // vmov.f64 d18, d1 |
| 247 | 0xeef03b40, // vmov.f64 d19, d0 |
| 248 | 0xeeb00b44, // vmov.f64 d0, d4 |
| 249 | 0xeeb01b45, // vmov.f64 d1, d5 |
| 250 | 0xeeb02b46, // vmov.f64 d2, d6 |
| 251 | 0xeeb03b47, // vmov.f64 d3, d7 |
| 252 | 0xeeb04b63, // vmov.f64 d4, d19 |
| 253 | 0xeeb05b62, // vmov.f64 d5, d18 |
| 254 | 0xeeb06b61, // vmov.f64 d6, d17 |
| 255 | 0xeeb07b60, // vmov.f64 d7, d16 |
| 256 | }; |
| 257 | static const unsigned int kSplice_move_src_dst_lowp[] = { |
| 258 | 0xeeb04b40, // vmov.f64 d4, d0 |
| 259 | 0xeeb05b41, // vmov.f64 d5, d1 |
| 260 | 0xeeb06b42, // vmov.f64 d6, d2 |
| 261 | 0xeeb07b43, // vmov.f64 d7, d3 |
| 262 | }; |
| 263 | static const unsigned int kSplice_move_dst_src_lowp[] = { |
| 264 | 0xeeb00b44, // vmov.f64 d0, d4 |
| 265 | 0xeeb01b45, // vmov.f64 d1, d5 |
| 266 | 0xeeb02b46, // vmov.f64 d2, d6 |
| 267 | 0xeeb03b47, // vmov.f64 d3, d7 |
| 268 | }; |
| 269 | static const unsigned int kSplice_premul_lowp[] = { |
| 270 | 0xf3500b03, // vqrdmulh.s16 d16, d0, d3 |
| 271 | 0xf3511b03, // vqrdmulh.s16 d17, d1, d3 |
| 272 | 0xf3522b03, // vqrdmulh.s16 d18, d2, d3 |
| 273 | 0xf2433110, // vand d19, d3, d0 |
| 274 | 0xf2434111, // vand d20, d3, d1 |
| 275 | 0xf3b50320, // vabs.s16 d0, d16 |
| 276 | 0xf2430112, // vand d16, d3, d2 |
| 277 | 0xf3b51321, // vabs.s16 d1, d17 |
| 278 | 0xf3b52322, // vabs.s16 d2, d18 |
| 279 | 0xf3910133, // vsra.u16 d0, d19, #15 |
| 280 | 0xf3911134, // vsra.u16 d1, d20, #15 |
| 281 | 0xf3912130, // vsra.u16 d2, d16, #15 |
| 282 | }; |
| 283 | static const unsigned int kSplice_load_8888_lowp[] = { |
| 284 | 0xe592c000, // ldr ip, [r2] |
| 285 | 0xe08cc100, // add ip, ip, r0, lsl #2 |
| 286 | 0xf4ec030d, // vld4.8 {d16[0],d17[0],d18[0],d19[0]}, [ip]! |
| 287 | 0xf4ec032d, // vld4.8 {d16[1],d17[1],d18[1],d19[1]}, [ip]! |
| 288 | 0xf4ec034d, // vld4.8 {d16[2],d17[2],d18[2],d19[2]}, [ip]! |
| 289 | 0xf4ec036d, // vld4.8 {d16[3],d17[3],d18[3],d19[3]}, [ip]! |
| 290 | 0xf38f0a30, // vshll.u8 q0, d16, #7 |
| 291 | 0xf38f2a32, // vshll.u8 q1, d18, #7 |
| 292 | 0xf3cf0a31, // vshll.u8 q8, d17, #7 |
| 293 | 0xf3cf2a33, // vshll.u8 q9, d19, #7 |
| 294 | 0xf3980350, // vrsra.u16 q0, q0, #8 |
| 295 | 0xf3d80370, // vrsra.u16 q8, q8, #8 |
| 296 | 0xf3d82372, // vrsra.u16 q9, q9, #8 |
| 297 | 0xf3982352, // vrsra.u16 q1, q1, #8 |
| 298 | 0xf22011b0, // vorr d1, d16, d16 |
| 299 | 0xf22231b2, // vorr d3, d18, d18 |
| 300 | }; |
| 301 | static const unsigned int kSplice_store_8888_lowp[] = { |
| 302 | 0xf2630113, // vorr d16, d3, d3 |
| 303 | 0xe592c000, // ldr ip, [r2] |
| 304 | 0xf2612111, // vorr d18, d1, d1 |
| 305 | 0xf3c94910, // vqshrn.u16 d20, q0, #7 |
| 306 | 0xe08cc100, // add ip, ip, r0, lsl #2 |
| 307 | 0xf3c96912, // vqshrn.u16 d22, q1, #7 |
| 308 | 0xf3c95932, // vqshrn.u16 d21, q9, #7 |
| 309 | 0xf3c97930, // vqshrn.u16 d23, q8, #7 |
| 310 | 0xf4cc430d, // vst4.8 {d20[0],d21[0],d22[0],d23[0]}, [ip]! |
| 311 | 0xf4cc432d, // vst4.8 {d20[1],d21[1],d22[1],d23[1]}, [ip]! |
| 312 | 0xf4cc434d, // vst4.8 {d20[2],d21[2],d22[2],d23[2]}, [ip]! |
| 313 | 0xf4cc436d, // vst4.8 {d20[3],d21[3],d22[3],d23[3]}, [ip]! |
| 314 | }; |
| 315 | |
| 316 | #else |
| 317 | |
| 318 | static const unsigned char kSplice_inc_x_lowp[] = { |
| 319 | 0x48,0x83,0xc7,0x10, // add $0x10,%rdi |
| 320 | }; |
| 321 | static const unsigned char kSplice_clear_lowp[] = { |
| 322 | 0xc5,0xfc,0x57,0xc0, // vxorps %ymm0,%ymm0,%ymm0 |
| 323 | 0xc5,0xf4,0x57,0xc9, // vxorps %ymm1,%ymm1,%ymm1 |
| 324 | 0xc5,0xec,0x57,0xd2, // vxorps %ymm2,%ymm2,%ymm2 |
| 325 | 0xc5,0xe4,0x57,0xdb, // vxorps %ymm3,%ymm3,%ymm3 |
| 326 | }; |
| 327 | static const unsigned char kSplice_plus__lowp[] = { |
| 328 | 0xc5,0xfd,0xdd,0xc4, // vpaddusw %ymm4,%ymm0,%ymm0 |
| 329 | 0xc5,0xf5,0xdd,0xcd, // vpaddusw %ymm5,%ymm1,%ymm1 |
| 330 | 0xc5,0xed,0xdd,0xd6, // vpaddusw %ymm6,%ymm2,%ymm2 |
| 331 | 0xc5,0xe5,0xdd,0xdf, // vpaddusw %ymm7,%ymm3,%ymm3 |
| 332 | }; |
| 333 | static const unsigned char kSplice_srcover_lowp[] = { |
| 334 | 0xc4,0x62,0x7d,0x79,0x41,0x02, // vpbroadcastw 0x2(%rcx),%ymm8 |
| 335 | 0xc5,0x3d,0xd9,0xc3, // vpsubusw %ymm3,%ymm8,%ymm8 |
| 336 | 0xc4,0x42,0x5d,0x0b,0xc8, // vpmulhrsw %ymm8,%ymm4,%ymm9 |
| 337 | 0xc4,0x42,0x7d,0x1d,0xc9, // vpabsw %ymm9,%ymm9 |
| 338 | 0xc5,0xb5,0xdd,0xc0, // vpaddusw %ymm0,%ymm9,%ymm0 |
| 339 | 0xc4,0x42,0x55,0x0b,0xc8, // vpmulhrsw %ymm8,%ymm5,%ymm9 |
| 340 | 0xc4,0x42,0x7d,0x1d,0xc9, // vpabsw %ymm9,%ymm9 |
| 341 | 0xc5,0xb5,0xdd,0xc9, // vpaddusw %ymm1,%ymm9,%ymm1 |
| 342 | 0xc4,0x42,0x4d,0x0b,0xc8, // vpmulhrsw %ymm8,%ymm6,%ymm9 |
| 343 | 0xc4,0x42,0x7d,0x1d,0xc9, // vpabsw %ymm9,%ymm9 |
| 344 | 0xc5,0xb5,0xdd,0xd2, // vpaddusw %ymm2,%ymm9,%ymm2 |
| 345 | 0xc4,0x42,0x45,0x0b,0xc0, // vpmulhrsw %ymm8,%ymm7,%ymm8 |
| 346 | 0xc4,0x42,0x7d,0x1d,0xc0, // vpabsw %ymm8,%ymm8 |
| 347 | 0xc5,0xbd,0xdd,0xdb, // vpaddusw %ymm3,%ymm8,%ymm3 |
| 348 | }; |
| 349 | static const unsigned char kSplice_dstover_lowp[] = { |
| 350 | 0xc4,0x62,0x7d,0x79,0x41,0x02, // vpbroadcastw 0x2(%rcx),%ymm8 |
| 351 | 0xc5,0x3d,0xd9,0xc7, // vpsubusw %ymm7,%ymm8,%ymm8 |
| 352 | 0xc4,0x42,0x7d,0x0b,0xc8, // vpmulhrsw %ymm8,%ymm0,%ymm9 |
| 353 | 0xc4,0x42,0x7d,0x1d,0xc9, // vpabsw %ymm9,%ymm9 |
| 354 | 0xc5,0xb5,0xdd,0xe4, // vpaddusw %ymm4,%ymm9,%ymm4 |
| 355 | 0xc4,0x42,0x75,0x0b,0xc8, // vpmulhrsw %ymm8,%ymm1,%ymm9 |
| 356 | 0xc4,0x42,0x7d,0x1d,0xc9, // vpabsw %ymm9,%ymm9 |
| 357 | 0xc5,0xb5,0xdd,0xed, // vpaddusw %ymm5,%ymm9,%ymm5 |
| 358 | 0xc4,0x42,0x6d,0x0b,0xc8, // vpmulhrsw %ymm8,%ymm2,%ymm9 |
| 359 | 0xc4,0x42,0x7d,0x1d,0xc9, // vpabsw %ymm9,%ymm9 |
| 360 | 0xc5,0xb5,0xdd,0xf6, // vpaddusw %ymm6,%ymm9,%ymm6 |
| 361 | 0xc4,0x42,0x65,0x0b,0xc0, // vpmulhrsw %ymm8,%ymm3,%ymm8 |
| 362 | 0xc4,0x42,0x7d,0x1d,0xc0, // vpabsw %ymm8,%ymm8 |
| 363 | 0xc5,0xbd,0xdd,0xff, // vpaddusw %ymm7,%ymm8,%ymm7 |
| 364 | }; |
| 365 | static const unsigned char kSplice_clamp_1_lowp[] = { |
| 366 | 0xc4,0x62,0x7d,0x79,0x41,0x02, // vpbroadcastw 0x2(%rcx),%ymm8 |
| 367 | 0xc4,0xc2,0x7d,0x3a,0xc0, // vpminuw %ymm8,%ymm0,%ymm0 |
| 368 | 0xc4,0xc2,0x75,0x3a,0xc8, // vpminuw %ymm8,%ymm1,%ymm1 |
| 369 | 0xc4,0xc2,0x6d,0x3a,0xd0, // vpminuw %ymm8,%ymm2,%ymm2 |
| 370 | 0xc4,0xc2,0x65,0x3a,0xd8, // vpminuw %ymm8,%ymm3,%ymm3 |
| 371 | }; |
| 372 | static const unsigned char kSplice_clamp_a_lowp[] = { |
| 373 | 0xc4,0x62,0x7d,0x79,0x41,0x02, // vpbroadcastw 0x2(%rcx),%ymm8 |
| 374 | 0xc4,0xc2,0x65,0x3a,0xd8, // vpminuw %ymm8,%ymm3,%ymm3 |
| 375 | 0xc4,0xe2,0x7d,0x3a,0xc3, // vpminuw %ymm3,%ymm0,%ymm0 |
| 376 | 0xc4,0xe2,0x75,0x3a,0xcb, // vpminuw %ymm3,%ymm1,%ymm1 |
| 377 | 0xc4,0xe2,0x6d,0x3a,0xd3, // vpminuw %ymm3,%ymm2,%ymm2 |
| 378 | }; |
| 379 | static const unsigned char kSplice_swap_lowp[] = { |
| 380 | 0xc5,0x7c,0x28,0xc3, // vmovaps %ymm3,%ymm8 |
| 381 | 0xc5,0x7c,0x28,0xca, // vmovaps %ymm2,%ymm9 |
| 382 | 0xc5,0x7c,0x28,0xd1, // vmovaps %ymm1,%ymm10 |
| 383 | 0xc5,0x7c,0x28,0xd8, // vmovaps %ymm0,%ymm11 |
| 384 | 0xc5,0xfc,0x28,0xc4, // vmovaps %ymm4,%ymm0 |
| 385 | 0xc5,0xfc,0x28,0xcd, // vmovaps %ymm5,%ymm1 |
| 386 | 0xc5,0xfc,0x28,0xd6, // vmovaps %ymm6,%ymm2 |
| 387 | 0xc5,0xfc,0x28,0xdf, // vmovaps %ymm7,%ymm3 |
| 388 | 0xc5,0x7c,0x29,0xdc, // vmovaps %ymm11,%ymm4 |
| 389 | 0xc5,0x7c,0x29,0xd5, // vmovaps %ymm10,%ymm5 |
| 390 | 0xc5,0x7c,0x29,0xce, // vmovaps %ymm9,%ymm6 |
| 391 | 0xc5,0x7c,0x29,0xc7, // vmovaps %ymm8,%ymm7 |
| 392 | }; |
| 393 | static const unsigned char kSplice_move_src_dst_lowp[] = { |
| 394 | 0xc5,0xfc,0x28,0xe0, // vmovaps %ymm0,%ymm4 |
| 395 | 0xc5,0xfc,0x28,0xe9, // vmovaps %ymm1,%ymm5 |
| 396 | 0xc5,0xfc,0x28,0xf2, // vmovaps %ymm2,%ymm6 |
| 397 | 0xc5,0xfc,0x28,0xfb, // vmovaps %ymm3,%ymm7 |
| 398 | }; |
| 399 | static const unsigned char kSplice_move_dst_src_lowp[] = { |
| 400 | 0xc5,0xfc,0x28,0xc4, // vmovaps %ymm4,%ymm0 |
| 401 | 0xc5,0xfc,0x28,0xcd, // vmovaps %ymm5,%ymm1 |
| 402 | 0xc5,0xfc,0x28,0xd6, // vmovaps %ymm6,%ymm2 |
| 403 | 0xc5,0xfc,0x28,0xdf, // vmovaps %ymm7,%ymm3 |
| 404 | }; |
| 405 | static const unsigned char kSplice_premul_lowp[] = { |
| 406 | 0xc4,0xe2,0x7d,0x0b,0xc3, // vpmulhrsw %ymm3,%ymm0,%ymm0 |
| 407 | 0xc4,0xe2,0x7d,0x1d,0xc0, // vpabsw %ymm0,%ymm0 |
| 408 | 0xc4,0xe2,0x75,0x0b,0xcb, // vpmulhrsw %ymm3,%ymm1,%ymm1 |
| 409 | 0xc4,0xe2,0x7d,0x1d,0xc9, // vpabsw %ymm1,%ymm1 |
| 410 | 0xc4,0xe2,0x6d,0x0b,0xd3, // vpmulhrsw %ymm3,%ymm2,%ymm2 |
| 411 | 0xc4,0xe2,0x7d,0x1d,0xd2, // vpabsw %ymm2,%ymm2 |
| 412 | }; |
| 413 | static const unsigned char kSplice_load_8888_lowp[] = { |
| 414 | 0x48,0x8b,0x02, // mov (%rdx),%rax |
| 415 | 0xc5,0xfa,0x6f,0x04,0xb8, // vmovdqu (%rax,%rdi,4),%xmm0 |
| 416 | 0xc5,0xfa,0x6f,0x4c,0xb8,0x10, // vmovdqu 0x10(%rax,%rdi,4),%xmm1 |
| 417 | 0xc5,0xfa,0x6f,0x54,0xb8,0x20, // vmovdqu 0x20(%rax,%rdi,4),%xmm2 |
| 418 | 0xc5,0xfa,0x6f,0x5c,0xb8,0x30, // vmovdqu 0x30(%rax,%rdi,4),%xmm3 |
| 419 | 0xc5,0x79,0x60,0xc1, // vpunpcklbw %xmm1,%xmm0,%xmm8 |
| 420 | 0xc5,0xf9,0x68,0xc1, // vpunpckhbw %xmm1,%xmm0,%xmm0 |
| 421 | 0xc5,0xe9,0x60,0xcb, // vpunpcklbw %xmm3,%xmm2,%xmm1 |
| 422 | 0xc5,0xe9,0x68,0xd3, // vpunpckhbw %xmm3,%xmm2,%xmm2 |
| 423 | 0xc5,0xb9,0x60,0xd8, // vpunpcklbw %xmm0,%xmm8,%xmm3 |
| 424 | 0xc5,0xb9,0x68,0xc0, // vpunpckhbw %xmm0,%xmm8,%xmm0 |
| 425 | 0xc5,0x71,0x60,0xc2, // vpunpcklbw %xmm2,%xmm1,%xmm8 |
| 426 | 0xc5,0xf1,0x68,0xca, // vpunpckhbw %xmm2,%xmm1,%xmm1 |
| 427 | 0xc5,0xe1,0x60,0xd0, // vpunpcklbw %xmm0,%xmm3,%xmm2 |
| 428 | 0xc5,0x61,0x68,0xc8, // vpunpckhbw %xmm0,%xmm3,%xmm9 |
| 429 | 0xc5,0xb9,0x60,0xd9, // vpunpcklbw %xmm1,%xmm8,%xmm3 |
| 430 | 0xc5,0x39,0x68,0xc1, // vpunpckhbw %xmm1,%xmm8,%xmm8 |
| 431 | 0xc5,0xe9,0x6c,0xc3, // vpunpcklqdq %xmm3,%xmm2,%xmm0 |
| 432 | 0xc4,0xe2,0x7d,0x30,0xc0, // vpmovzxbw %xmm0,%ymm0 |
| 433 | 0xc5,0xf5,0x71,0xf0,0x07, // vpsllw $0x7,%ymm0,%ymm1 |
| 434 | 0xc5,0xad,0x71,0xd0,0x01, // vpsrlw $0x1,%ymm0,%ymm10 |
| 435 | 0xc4,0xc1,0x75,0xdd,0xca, // vpaddusw %ymm10,%ymm1,%ymm1 |
| 436 | 0xc4,0x62,0x7d,0x79,0x11, // vpbroadcastw (%rcx),%ymm10 |
| 437 | 0xc4,0xc1,0x7d,0xdd,0xc2, // vpaddusw %ymm10,%ymm0,%ymm0 |
| 438 | 0xc5,0xfd,0x71,0xd0,0x08, // vpsrlw $0x8,%ymm0,%ymm0 |
| 439 | 0xc5,0xf5,0xdd,0xc0, // vpaddusw %ymm0,%ymm1,%ymm0 |
| 440 | 0xc5,0xe9,0x6d,0xcb, // vpunpckhqdq %xmm3,%xmm2,%xmm1 |
| 441 | 0xc4,0xe2,0x7d,0x30,0xc9, // vpmovzxbw %xmm1,%ymm1 |
| 442 | 0xc5,0xed,0x71,0xf1,0x07, // vpsllw $0x7,%ymm1,%ymm2 |
| 443 | 0xc5,0xe5,0x71,0xd1,0x01, // vpsrlw $0x1,%ymm1,%ymm3 |
| 444 | 0xc5,0xed,0xdd,0xd3, // vpaddusw %ymm3,%ymm2,%ymm2 |
| 445 | 0xc4,0xc1,0x75,0xdd,0xca, // vpaddusw %ymm10,%ymm1,%ymm1 |
| 446 | 0xc5,0xf5,0x71,0xd1,0x08, // vpsrlw $0x8,%ymm1,%ymm1 |
| 447 | 0xc5,0xed,0xdd,0xc9, // vpaddusw %ymm1,%ymm2,%ymm1 |
| 448 | 0xc4,0xc1,0x31,0x6c,0xd0, // vpunpcklqdq %xmm8,%xmm9,%xmm2 |
| 449 | 0xc4,0xe2,0x7d,0x30,0xd2, // vpmovzxbw %xmm2,%ymm2 |
| 450 | 0xc5,0xe5,0x71,0xf2,0x07, // vpsllw $0x7,%ymm2,%ymm3 |
| 451 | 0xc5,0xa5,0x71,0xd2,0x01, // vpsrlw $0x1,%ymm2,%ymm11 |
| 452 | 0xc4,0xc1,0x65,0xdd,0xdb, // vpaddusw %ymm11,%ymm3,%ymm3 |
| 453 | 0xc4,0xc1,0x6d,0xdd,0xd2, // vpaddusw %ymm10,%ymm2,%ymm2 |
| 454 | 0xc5,0xed,0x71,0xd2,0x08, // vpsrlw $0x8,%ymm2,%ymm2 |
| 455 | 0xc5,0xe5,0xdd,0xd2, // vpaddusw %ymm2,%ymm3,%ymm2 |
| 456 | 0xc4,0xc1,0x31,0x6d,0xd8, // vpunpckhqdq %xmm8,%xmm9,%xmm3 |
| 457 | 0xc4,0xe2,0x7d,0x30,0xdb, // vpmovzxbw %xmm3,%ymm3 |
| 458 | 0xc5,0xbd,0x71,0xf3,0x07, // vpsllw $0x7,%ymm3,%ymm8 |
| 459 | 0xc5,0xb5,0x71,0xd3,0x01, // vpsrlw $0x1,%ymm3,%ymm9 |
| 460 | 0xc4,0x41,0x3d,0xdd,0xc1, // vpaddusw %ymm9,%ymm8,%ymm8 |
| 461 | 0xc4,0xc1,0x65,0xdd,0xda, // vpaddusw %ymm10,%ymm3,%ymm3 |
| 462 | 0xc5,0xe5,0x71,0xd3,0x08, // vpsrlw $0x8,%ymm3,%ymm3 |
| 463 | 0xc5,0xbd,0xdd,0xdb, // vpaddusw %ymm3,%ymm8,%ymm3 |
| 464 | }; |
| 465 | static const unsigned char kSplice_store_8888_lowp[] = { |
| 466 | 0x48,0x8b,0x02, // mov (%rdx),%rax |
| 467 | 0xc5,0x7d,0xdd,0xc0, // vpaddusw %ymm0,%ymm0,%ymm8 |
| 468 | 0xc4,0xc1,0x3d,0x71,0xd0,0x08, // vpsrlw $0x8,%ymm8,%ymm8 |
| 469 | 0xc4,0x43,0x7d,0x39,0xc1,0x01, // vextracti128 $0x1,%ymm8,%xmm9 |
| 470 | 0xc4,0x41,0x39,0x67,0xc1, // vpackuswb %xmm9,%xmm8,%xmm8 |
| 471 | 0xc5,0x75,0xdd,0xc9, // vpaddusw %ymm1,%ymm1,%ymm9 |
| 472 | 0xc4,0xc1,0x35,0x71,0xd1,0x08, // vpsrlw $0x8,%ymm9,%ymm9 |
| 473 | 0xc4,0x43,0x7d,0x39,0xca,0x01, // vextracti128 $0x1,%ymm9,%xmm10 |
| 474 | 0xc4,0x41,0x31,0x67,0xca, // vpackuswb %xmm10,%xmm9,%xmm9 |
| 475 | 0xc5,0x6d,0xdd,0xd2, // vpaddusw %ymm2,%ymm2,%ymm10 |
| 476 | 0xc4,0xc1,0x2d,0x71,0xd2,0x08, // vpsrlw $0x8,%ymm10,%ymm10 |
| 477 | 0xc4,0x43,0x7d,0x39,0xd3,0x01, // vextracti128 $0x1,%ymm10,%xmm11 |
| 478 | 0xc4,0x41,0x29,0x67,0xd3, // vpackuswb %xmm11,%xmm10,%xmm10 |
| 479 | 0xc5,0x65,0xdd,0xdb, // vpaddusw %ymm3,%ymm3,%ymm11 |
| 480 | 0xc4,0xc1,0x25,0x71,0xd3,0x08, // vpsrlw $0x8,%ymm11,%ymm11 |
| 481 | 0xc4,0x43,0x7d,0x39,0xdc,0x01, // vextracti128 $0x1,%ymm11,%xmm12 |
| 482 | 0xc4,0x41,0x21,0x67,0xdc, // vpackuswb %xmm12,%xmm11,%xmm11 |
| 483 | 0xc4,0x41,0x39,0x60,0xe1, // vpunpcklbw %xmm9,%xmm8,%xmm12 |
| 484 | 0xc4,0x41,0x39,0x68,0xc1, // vpunpckhbw %xmm9,%xmm8,%xmm8 |
| 485 | 0xc4,0x41,0x29,0x60,0xcb, // vpunpcklbw %xmm11,%xmm10,%xmm9 |
| 486 | 0xc4,0x41,0x29,0x68,0xd3, // vpunpckhbw %xmm11,%xmm10,%xmm10 |
| 487 | 0xc4,0x41,0x19,0x61,0xd9, // vpunpcklwd %xmm9,%xmm12,%xmm11 |
| 488 | 0xc5,0x7a,0x7f,0x1c,0xb8, // vmovdqu %xmm11,(%rax,%rdi,4) |
| 489 | 0xc4,0x41,0x19,0x69,0xc9, // vpunpckhwd %xmm9,%xmm12,%xmm9 |
| 490 | 0xc5,0x7a,0x7f,0x4c,0xb8,0x10, // vmovdqu %xmm9,0x10(%rax,%rdi,4) |
| 491 | 0xc4,0x41,0x39,0x61,0xca, // vpunpcklwd %xmm10,%xmm8,%xmm9 |
| 492 | 0xc5,0x7a,0x7f,0x4c,0xb8,0x20, // vmovdqu %xmm9,0x20(%rax,%rdi,4) |
| 493 | 0xc4,0x41,0x39,0x69,0xc2, // vpunpckhwd %xmm10,%xmm8,%xmm8 |
| 494 | 0xc5,0x7a,0x7f,0x44,0xb8,0x30, // vmovdqu %xmm8,0x30(%rax,%rdi,4) |
| 495 | }; |
| 496 | |
| 497 | #endif |
| 498 | |
| 499 | #endif//SkSplicer_generated_lowp_DEFINED |