1/************************************************************** 2 Copyright (c) 2019 Huawei Technologies Co., Ltd. 3 4 Redistribution and use in source and binary forms, with or without 5 modification, are permitted provided that the following conditions 6 are met: 7 * Redistributions of source code must retain the above copyright 8 notice, this list of conditions and the following disclaimer. 9 * Redistributions in binary form must reproduce the above copyright 10 notice, this list of conditions and the following disclaimer in 11 the documentation and/or other materials provided with the 12 distribution. 13 * Neither the name of Huawei Corporation nor the names of its 14 contributors may be used to endorse or promote products derived 15 from this software without specific prior written permission. 16 17 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 18 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 19 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 20 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 21 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 22 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 23 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 24 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 25 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 26 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 27 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28**********************************************************************/ 29.text 30 31.global gf_2vect_mad_neon 32.type gf_2vect_mad_neon, %function 33 34 35/* arguments */ 36x_len .req x0 37x_vec .req x1 38x_vec_i .req x2 39x_tbl .req x3 40x_src .req x4 41x_dest .req x5 42 43/* returns */ 44w_ret .req w0 45 46/* local variables */ 47x_src_end .req x6 48x_dest1 .req x7 49x_dest2 .req x8 50x_tmp .req x9 51x_tbl1 .req x10 52x_tbl2 .req x11 53x_const .req x12 54 55/* vectors */ 56v_mask0f .req v0 57v_tmp_lo .req v1 58v_tmp_hi .req v2 59v_tmp .req v3 60q_tmp .req q3 61 62v_gft1_lo .req v4 63v_gft1_hi .req v5 64v_gft2_lo .req v6 65v_gft2_hi .req v7 66q_gft1_lo .req q4 67q_gft1_hi .req q5 68q_gft2_lo .req q6 69q_gft2_hi .req q7 70 71v_data_0 .req v8 72v_data_1 .req v9 73v_data_2 .req v10 74v_data_3 .req v11 75v_data_4 .req v12 76v_data_5 .req v13 77v_data_6 .req v14 78v_data_7 .req v15 79q_data_0 .req q8 80q_data_1 .req q9 81q_data_2 .req q10 82q_data_3 .req q11 83q_data_4 .req q12 84q_data_5 .req q13 85q_data_6 .req q14 86q_data_7 .req q15 87 88v_data_0_lo .req v16 89v_data_1_lo .req v17 90v_data_2_lo .req v18 91v_data_3_lo .req v19 92v_data_4_lo .req v20 93v_data_5_lo .req v21 94v_data_6_lo .req v22 95v_data_7_lo .req v23 96v_data_0_hi .req v_data_0 97v_data_1_hi .req v_data_1 98v_data_2_hi .req v_data_2 99v_data_3_hi .req v_data_3 100v_data_4_hi .req v_data_4 101v_data_5_hi .req v_data_5 102v_data_6_hi .req v_data_6 103v_data_7_hi .req v_data_7 104 105v_d0 .req v24 106v_d1 .req v25 107v_d2 .req v26 108v_d3 .req v27 109v_d4 .req v28 110v_d5 .req v29 111v_d6 .req v30 112v_d7 .req v31 113q_d0 .req q24 114q_d1 .req q25 115q_d2 .req q26 116q_d3 .req q27 117q_d4 .req q28 118q_d5 .req q29 119q_d6 .req q30 120q_d7 .req q31 121 122v_data .req v16 123q_data .req q16 124v_data_lo .req v17 125v_data_hi .req v18 126 127 128gf_2vect_mad_neon: 129 /* less than 16 bytes, return_fail */ 130 cmp x_len, #16 131 blt .return_fail 132 133 movi v_mask0f.16b, #0x0f 134 lsl x_vec_i, x_vec_i, #5 135 lsl x_vec, x_vec, #5 136 add x_tbl1, x_tbl, x_vec_i 137 add x_tbl2, x_tbl1, x_vec 138 add x_src_end, x_src, x_len 139 140 ldr x_dest1, [x_dest] 141 ldr x_dest2, [x_dest, #8] 142 ldr q_gft1_lo, [x_tbl1] 143 ldr q_gft1_hi, [x_tbl1, #16] 144 ldr q_gft2_lo, [x_tbl2] 145 ldr q_gft2_hi, [x_tbl2, #16] 146 147.Lloop128_init: 148 /* less than 128 bytes, goto Lloop16_init */ 149 cmp x_len, #128 150 blt .Lloop16_init 151 152 /* save d8 ~ d15 to stack */ 153 sub sp, sp, #64 154 stp d8, d9, [sp] 155 stp d10, d11, [sp, #16] 156 stp d12, d13, [sp, #32] 157 stp d14, d15, [sp, #48] 158 159 sub x_src_end, x_src_end, #128 160 161.Lloop128: 162 ldr q_data_0, [x_src, #16*0] 163 ldr q_data_1, [x_src, #16*1] 164 ldr q_data_2, [x_src, #16*2] 165 ldr q_data_3, [x_src, #16*3] 166 ldr q_data_4, [x_src, #16*4] 167 ldr q_data_5, [x_src, #16*5] 168 ldr q_data_6, [x_src, #16*6] 169 ldr q_data_7, [x_src, #16*7] 170 171 ldr q_d0, [x_dest1, #16*0] 172 ldr q_d1, [x_dest1, #16*1] 173 ldr q_d2, [x_dest1, #16*2] 174 ldr q_d3, [x_dest1, #16*3] 175 ldr q_d4, [x_dest1, #16*4] 176 ldr q_d5, [x_dest1, #16*5] 177 ldr q_d6, [x_dest1, #16*6] 178 ldr q_d7, [x_dest1, #16*7] 179 180 and v_data_0_lo.16b, v_data_0.16b, v_mask0f.16b 181 and v_data_1_lo.16b, v_data_1.16b, v_mask0f.16b 182 and v_data_2_lo.16b, v_data_2.16b, v_mask0f.16b 183 and v_data_3_lo.16b, v_data_3.16b, v_mask0f.16b 184 and v_data_4_lo.16b, v_data_4.16b, v_mask0f.16b 185 and v_data_5_lo.16b, v_data_5.16b, v_mask0f.16b 186 and v_data_6_lo.16b, v_data_6.16b, v_mask0f.16b 187 and v_data_7_lo.16b, v_data_7.16b, v_mask0f.16b 188 189 ushr v_data_0_hi.16b, v_data_0.16b, #4 190 ushr v_data_1_hi.16b, v_data_1.16b, #4 191 ushr v_data_2_hi.16b, v_data_2.16b, #4 192 ushr v_data_3_hi.16b, v_data_3.16b, #4 193 ushr v_data_4_hi.16b, v_data_4.16b, #4 194 ushr v_data_5_hi.16b, v_data_5.16b, #4 195 ushr v_data_6_hi.16b, v_data_6.16b, #4 196 ushr v_data_7_hi.16b, v_data_7.16b, #4 197 198 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_0_lo.16b 199 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_0_hi.16b 200 eor v_d0.16b, v_tmp_lo.16b, v_d0.16b 201 eor v_d0.16b, v_d0.16b, v_tmp_hi.16b 202 203 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_1_lo.16b 204 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_1_hi.16b 205 eor v_d1.16b, v_tmp_lo.16b, v_d1.16b 206 eor v_d1.16b, v_d1.16b, v_tmp_hi.16b 207 208 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_2_lo.16b 209 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_2_hi.16b 210 eor v_d2.16b, v_tmp_lo.16b, v_d2.16b 211 eor v_d2.16b, v_d2.16b, v_tmp_hi.16b 212 213 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_3_lo.16b 214 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_3_hi.16b 215 eor v_d3.16b, v_tmp_lo.16b, v_d3.16b 216 eor v_d3.16b, v_d3.16b, v_tmp_hi.16b 217 218 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_4_lo.16b 219 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_4_hi.16b 220 eor v_d4.16b, v_tmp_lo.16b, v_d4.16b 221 eor v_d4.16b, v_d4.16b, v_tmp_hi.16b 222 223 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_5_lo.16b 224 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_5_hi.16b 225 eor v_d5.16b, v_tmp_lo.16b, v_d5.16b 226 eor v_d5.16b, v_d5.16b, v_tmp_hi.16b 227 228 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_6_lo.16b 229 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_6_hi.16b 230 eor v_d6.16b, v_tmp_lo.16b, v_d6.16b 231 eor v_d6.16b, v_d6.16b, v_tmp_hi.16b 232 233 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_7_lo.16b 234 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_7_hi.16b 235 eor v_d7.16b, v_tmp_lo.16b, v_d7.16b 236 eor v_d7.16b, v_d7.16b, v_tmp_hi.16b 237 238 str q_d0, [x_dest1, #16*0] 239 str q_d1, [x_dest1, #16*1] 240 str q_d2, [x_dest1, #16*2] 241 str q_d3, [x_dest1, #16*3] 242 str q_d4, [x_dest1, #16*4] 243 str q_d5, [x_dest1, #16*5] 244 str q_d6, [x_dest1, #16*6] 245 str q_d7, [x_dest1, #16*7] 246 247 ldr q_d0, [x_dest2, #16*0] 248 ldr q_d1, [x_dest2, #16*1] 249 ldr q_d2, [x_dest2, #16*2] 250 ldr q_d3, [x_dest2, #16*3] 251 ldr q_d4, [x_dest2, #16*4] 252 ldr q_d5, [x_dest2, #16*5] 253 ldr q_d6, [x_dest2, #16*6] 254 ldr q_d7, [x_dest2, #16*7] 255 256 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_0_lo.16b 257 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_0_hi.16b 258 eor v_d0.16b, v_tmp_lo.16b, v_d0.16b 259 eor v_d0.16b, v_d0.16b, v_tmp_hi.16b 260 261 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_1_lo.16b 262 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_1_hi.16b 263 eor v_d1.16b, v_tmp_lo.16b, v_d1.16b 264 eor v_d1.16b, v_d1.16b, v_tmp_hi.16b 265 266 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_2_lo.16b 267 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_2_hi.16b 268 eor v_d2.16b, v_tmp_lo.16b, v_d2.16b 269 eor v_d2.16b, v_d2.16b, v_tmp_hi.16b 270 271 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_3_lo.16b 272 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_3_hi.16b 273 eor v_d3.16b, v_tmp_lo.16b, v_d3.16b 274 eor v_d3.16b, v_d3.16b, v_tmp_hi.16b 275 276 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_4_lo.16b 277 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_4_hi.16b 278 eor v_d4.16b, v_tmp_lo.16b, v_d4.16b 279 eor v_d4.16b, v_d4.16b, v_tmp_hi.16b 280 281 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_5_lo.16b 282 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_5_hi.16b 283 eor v_d5.16b, v_tmp_lo.16b, v_d5.16b 284 eor v_d5.16b, v_d5.16b, v_tmp_hi.16b 285 286 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_6_lo.16b 287 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_6_hi.16b 288 eor v_d6.16b, v_tmp_lo.16b, v_d6.16b 289 eor v_d6.16b, v_d6.16b, v_tmp_hi.16b 290 291 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_7_lo.16b 292 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_7_hi.16b 293 eor v_d7.16b, v_tmp_lo.16b, v_d7.16b 294 eor v_d7.16b, v_d7.16b, v_tmp_hi.16b 295 296 str q_d0, [x_dest2, #16*0] 297 str q_d1, [x_dest2, #16*1] 298 str q_d2, [x_dest2, #16*2] 299 str q_d3, [x_dest2, #16*3] 300 str q_d4, [x_dest2, #16*4] 301 str q_d5, [x_dest2, #16*5] 302 str q_d6, [x_dest2, #16*6] 303 str q_d7, [x_dest2, #16*7] 304 305 add x_src, x_src, #128 306 add x_dest1, x_dest1, #128 307 add x_dest2, x_dest2, #128 308 cmp x_src, x_src_end 309 bls .Lloop128 310 311.Lloop128_end: 312 /* restore d8 ~ d15 */ 313 ldp d8, d9, [sp] 314 ldp d10, d11, [sp, #16] 315 ldp d12, d13, [sp, #32] 316 ldp d14, d15, [sp, #48] 317 add sp, sp, #64 318 add x_src_end, x_src_end, #128 319 320.Lloop16_init: 321 sub x_src_end, x_src_end, #16 322 cmp x_src, x_src_end 323 bhi .lessthan16_init 324 325.Lloop16: 326 ldr q_data, [x_src] 327 328 ldr q_d0, [x_dest1] 329 ldr q_d1, [x_dest2] 330 331 and v_data_lo.16b, v_data.16b, v_mask0f.16b 332 ushr v_data_hi.16b, v_data.16b, #4 333 334 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_lo.16b 335 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_hi.16b 336 eor v_d0.16b, v_tmp_lo.16b, v_d0.16b 337 eor v_d0.16b, v_d0.16b, v_tmp_hi.16b 338 339 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_lo.16b 340 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_hi.16b 341 eor v_d1.16b, v_tmp_lo.16b, v_d1.16b 342 eor v_d1.16b, v_d1.16b, v_tmp_hi.16b 343 344 str q_d0, [x_dest1] 345 str q_d1, [x_dest2] 346 347 add x_dest1, x_dest1, #16 348 add x_dest2, x_dest2, #16 349 add x_src, x_src, #16 350 cmp x_src, x_src_end 351 bls .Lloop16 352 353.lessthan16_init: 354 sub x_tmp, x_src, x_src_end 355 cmp x_tmp, #16 356 beq .return_pass 357 358.lessthan16: 359 mov x_src, x_src_end 360 sub x_dest1, x_dest1, x_tmp 361 sub x_dest2, x_dest2, x_tmp 362 363 ldr x_const, =const_tbl 364 sub x_const, x_const, x_tmp 365 ldr q_tmp, [x_const, #16] 366 367 ldr q_data, [x_src] 368 ldr q_d0, [x_dest1] 369 ldr q_d1, [x_dest2] 370 371 and v_data_lo.16b, v_data.16b, v_mask0f.16b 372 ushr v_data_hi.16b, v_data.16b, #4 373 374 tbl v_tmp_lo.16b, {v_gft1_lo.16b}, v_data_lo.16b 375 tbl v_tmp_hi.16b, {v_gft1_hi.16b}, v_data_hi.16b 376 eor v_tmp_hi.16b, v_tmp_lo.16b, v_tmp_hi.16b 377 and v_tmp_hi.16b, v_tmp_hi.16b, v_tmp.16b 378 eor v_d0.16b, v_d0.16b, v_tmp_hi.16b 379 380 tbl v_tmp_lo.16b, {v_gft2_lo.16b}, v_data_lo.16b 381 tbl v_tmp_hi.16b, {v_gft2_hi.16b}, v_data_hi.16b 382 eor v_tmp_hi.16b, v_tmp_lo.16b, v_tmp_hi.16b 383 and v_tmp_hi.16b, v_tmp_hi.16b, v_tmp.16b 384 eor v_d1.16b, v_d1.16b, v_tmp_hi.16b 385 386 str q_d0, [x_dest1] 387 str q_d1, [x_dest2] 388 389.return_pass: 390 mov w_ret, #0 391 ret 392 393.return_fail: 394 mov w_ret, #1 395 ret 396 397.section .data 398.balign 8 399const_tbl: 400 .dword 0x0000000000000000, 0x0000000000000000 401 .dword 0xffffffffffffffff, 0xffffffffffffffff 402