|
| 1 | +/******************************************************************************* |
| 2 | +Copyright (c) 2024, The OpenBLAS Project |
| 3 | +All rights reserved. |
| 4 | +Redistribution and use in source and binary forms, with or without |
| 5 | +modification, are permitted provided that the following conditions are |
| 6 | +met: |
| 7 | +1. Redistributions of source code must retain the above copyright |
| 8 | +notice, this list of conditions and the following disclaimer. |
| 9 | +2. 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 | +3. Neither the name of the OpenBLAS project nor the names of |
| 14 | +its contributors may be used to endorse or promote products |
| 15 | +derived from this software without specific prior written permission. |
| 16 | +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 17 | +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 18 | +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 19 | +ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE |
| 20 | +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 21 | +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR |
| 22 | +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER |
| 23 | +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 24 | +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE |
| 25 | +USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 26 | +*******************************************************************************/ |
| 27 | + |
| 28 | +#include <stdio.h> |
| 29 | +#include "common.h" |
| 30 | + |
| 31 | +int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){ |
| 32 | + |
| 33 | + BLASLONG i, js, offset; |
| 34 | + |
| 35 | + FLOAT data01, data02, data03, data04, data05, data06, data07, data08; |
| 36 | + FLOAT data09, data10, data11, data12, data13, data14, data15, data16; |
| 37 | + FLOAT data17, data18, data19, data20, data21, data22, data23, data24; |
| 38 | + FLOAT data25, data26, data27, data28, data29, data30, data31, data32; |
| 39 | + |
| 40 | + FLOAT *ao1, *ao2, *ao3, *ao4, *ao5, *ao6, *ao7, *ao8; |
| 41 | + FLOAT *ao9, *ao10, *ao11, *ao12, *ao13, *ao14, *ao15, *ao16; |
| 42 | + |
| 43 | + lda *= 2; |
| 44 | + |
| 45 | + js = (n >> 4); |
| 46 | + while (js > 0){ |
| 47 | + |
| 48 | + offset = posX - posY; |
| 49 | + |
| 50 | + if (offset > 0) ao1 = a + (posX + 0) * 2 + posY * lda; else ao1 = a + posY * 2 + (posX + 0) * lda; |
| 51 | + if (offset > -1) ao2 = a + (posX + 1) * 2 + posY * lda; else ao2 = a + posY * 2 + (posX + 1) * lda; |
| 52 | + if (offset > -2) ao3 = a + (posX + 2) * 2 + posY * lda; else ao3 = a + posY * 2 + (posX + 2) * lda; |
| 53 | + if (offset > -3) ao4 = a + (posX + 3) * 2 + posY * lda; else ao4 = a + posY * 2 + (posX + 3) * lda; |
| 54 | + if (offset > -4) ao5 = a + (posX + 4) * 2 + posY * lda; else ao5 = a + posY * 2 + (posX + 4) * lda; |
| 55 | + if (offset > -5) ao6 = a + (posX + 5) * 2 + posY * lda; else ao6 = a + posY * 2 + (posX + 5) * lda; |
| 56 | + if (offset > -6) ao7 = a + (posX + 6) * 2 + posY * lda; else ao7 = a + posY * 2 + (posX + 6) * lda; |
| 57 | + if (offset > -7) ao8 = a + (posX + 7) * 2 + posY * lda; else ao8 = a + posY * 2 + (posX + 7) * lda; |
| 58 | + if (offset > -8) ao9 = a + (posX + 8) * 2 + posY * lda; else ao9 = a + posY * 2 + (posX + 8) * lda; |
| 59 | + if (offset > -9) ao10 = a + (posX + 9) * 2 + posY * lda; else ao10 = a + posY * 2 + (posX + 9) * lda; |
| 60 | + if (offset > -10) ao11 = a + (posX + 10) * 2 + posY * lda; else ao11 = a + posY * 2 + (posX + 10) * lda; |
| 61 | + if (offset > -11) ao12 = a + (posX + 11) * 2 + posY * lda; else ao12 = a + posY * 2 + (posX + 11) * lda; |
| 62 | + if (offset > -12) ao13 = a + (posX + 12) * 2 + posY * lda; else ao13 = a + posY * 2 + (posX + 12) * lda; |
| 63 | + if (offset > -13) ao14 = a + (posX + 13) * 2 + posY * lda; else ao14 = a + posY * 2 + (posX + 13) * lda; |
| 64 | + if (offset > -14) ao15 = a + (posX + 14) * 2 + posY * lda; else ao15 = a + posY * 2 + (posX + 14) * lda; |
| 65 | + if (offset > -15) ao16 = a + (posX + 15) * 2 + posY * lda; else ao16 = a + posY * 2 + (posX + 15) * lda; |
| 66 | + |
| 67 | + |
| 68 | + i = m; |
| 69 | + |
| 70 | + while (i > 0) { |
| 71 | + data01 = *(ao1 + 0); |
| 72 | + data02 = *(ao1 + 1); |
| 73 | + data03 = *(ao2 + 0); |
| 74 | + data04 = *(ao2 + 1); |
| 75 | + data05 = *(ao3 + 0); |
| 76 | + data06 = *(ao3 + 1); |
| 77 | + data07 = *(ao4 + 0); |
| 78 | + data08 = *(ao4 + 1); |
| 79 | + data09 = *(ao5 + 0); |
| 80 | + data10 = *(ao5 + 1); |
| 81 | + data11 = *(ao6 + 0); |
| 82 | + data12 = *(ao6 + 1); |
| 83 | + data13 = *(ao7 + 0); |
| 84 | + data14 = *(ao7 + 1); |
| 85 | + data15 = *(ao8 + 0); |
| 86 | + data16 = *(ao8 + 1); |
| 87 | + data17 = *(ao9 + 0); |
| 88 | + data18 = *(ao9 + 1); |
| 89 | + data19 = *(ao10 + 0); |
| 90 | + data20 = *(ao10 + 1); |
| 91 | + data21 = *(ao11 + 0); |
| 92 | + data22 = *(ao11 + 1); |
| 93 | + data23 = *(ao12 + 0); |
| 94 | + data24 = *(ao12 + 1); |
| 95 | + data25 = *(ao13 + 0); |
| 96 | + data26 = *(ao13 + 1); |
| 97 | + data27 = *(ao14 + 0); |
| 98 | + data28 = *(ao14 + 1); |
| 99 | + data29 = *(ao15 + 0); |
| 100 | + data30 = *(ao15 + 1); |
| 101 | + data31 = *(ao16 + 0); |
| 102 | + data32 = *(ao16 + 1); |
| 103 | + |
| 104 | + if (offset > 0) ao1 += lda; else ao1 += 2; |
| 105 | + if (offset > -1) ao2 += lda; else ao2 += 2; |
| 106 | + if (offset > -2) ao3 += lda; else ao3 += 2; |
| 107 | + if (offset > -3) ao4 += lda; else ao4 += 2; |
| 108 | + if (offset > -4) ao5 += lda; else ao5 += 2; |
| 109 | + if (offset > -5) ao6 += lda; else ao6 += 2; |
| 110 | + if (offset > -6) ao7 += lda; else ao7 += 2; |
| 111 | + if (offset > -7) ao8 += lda; else ao8 += 2; |
| 112 | + if (offset > -8) ao9 += lda; else ao9 += 2; |
| 113 | + if (offset > -9) ao10 += lda; else ao10 += 2; |
| 114 | + if (offset > -10) ao11 += lda; else ao11 += 2; |
| 115 | + if (offset > -11) ao12 += lda; else ao12 += 2; |
| 116 | + if (offset > -12) ao13 += lda; else ao13 += 2; |
| 117 | + if (offset > -13) ao14 += lda; else ao14 += 2; |
| 118 | + if (offset > -14) ao15 += lda; else ao15 += 2; |
| 119 | + if (offset > -15) ao16 += lda; else ao16 += 2; |
| 120 | + |
| 121 | + b[ 0] = data01; |
| 122 | + b[ 1] = data02; |
| 123 | + b[ 2] = data03; |
| 124 | + b[ 3] = data04; |
| 125 | + b[ 4] = data05; |
| 126 | + b[ 5] = data06; |
| 127 | + b[ 6] = data07; |
| 128 | + b[ 7] = data08; |
| 129 | + b[ 8] = data09; |
| 130 | + b[ 9] = data10; |
| 131 | + b[10] = data11; |
| 132 | + b[11] = data12; |
| 133 | + b[12] = data13; |
| 134 | + b[13] = data14; |
| 135 | + b[14] = data15; |
| 136 | + b[15] = data16; |
| 137 | + b[16] = data17; |
| 138 | + b[17] = data18; |
| 139 | + b[18] = data19; |
| 140 | + b[19] = data20; |
| 141 | + b[20] = data21; |
| 142 | + b[21] = data22; |
| 143 | + b[22] = data23; |
| 144 | + b[23] = data24; |
| 145 | + b[24] = data25; |
| 146 | + b[25] = data26; |
| 147 | + b[26] = data27; |
| 148 | + b[27] = data28; |
| 149 | + b[28] = data29; |
| 150 | + b[29] = data30; |
| 151 | + b[30] = data31; |
| 152 | + b[31] = data32; |
| 153 | + |
| 154 | + b += 32; |
| 155 | + |
| 156 | + offset --; |
| 157 | + i --; |
| 158 | + } |
| 159 | + |
| 160 | + posX += 16; |
| 161 | + js --; |
| 162 | + } |
| 163 | + |
| 164 | + if (n & 8) { |
| 165 | + offset = posX - posY; |
| 166 | + |
| 167 | + if (offset > 0) ao1 = a + (posX + 0) * 2 + posY * lda; else ao1 = a + posY * 2 + (posX + 0) * lda; |
| 168 | + if (offset > -1) ao2 = a + (posX + 1) * 2 + posY * lda; else ao2 = a + posY * 2 + (posX + 1) * lda; |
| 169 | + if (offset > -2) ao3 = a + (posX + 2) * 2 + posY * lda; else ao3 = a + posY * 2 + (posX + 2) * lda; |
| 170 | + if (offset > -3) ao4 = a + (posX + 3) * 2 + posY * lda; else ao4 = a + posY * 2 + (posX + 3) * lda; |
| 171 | + if (offset > -4) ao5 = a + (posX + 4) * 2 + posY * lda; else ao5 = a + posY * 2 + (posX + 4) * lda; |
| 172 | + if (offset > -5) ao6 = a + (posX + 5) * 2 + posY * lda; else ao6 = a + posY * 2 + (posX + 5) * lda; |
| 173 | + if (offset > -6) ao7 = a + (posX + 6) * 2 + posY * lda; else ao7 = a + posY * 2 + (posX + 6) * lda; |
| 174 | + if (offset > -7) ao8 = a + (posX + 7) * 2 + posY * lda; else ao8 = a + posY * 2 + (posX + 7) * lda; |
| 175 | + |
| 176 | + i = m; |
| 177 | + |
| 178 | + while (i > 0) { |
| 179 | + data01 = *(ao1 + 0); |
| 180 | + data02 = *(ao1 + 1); |
| 181 | + data03 = *(ao2 + 0); |
| 182 | + data04 = *(ao2 + 1); |
| 183 | + data05 = *(ao3 + 0); |
| 184 | + data06 = *(ao3 + 1); |
| 185 | + data07 = *(ao4 + 0); |
| 186 | + data08 = *(ao4 + 1); |
| 187 | + data09 = *(ao5 + 0); |
| 188 | + data10 = *(ao5 + 1); |
| 189 | + data11 = *(ao6 + 0); |
| 190 | + data12 = *(ao6 + 1); |
| 191 | + data13 = *(ao7 + 0); |
| 192 | + data14 = *(ao7 + 1); |
| 193 | + data15 = *(ao8 + 0); |
| 194 | + data16 = *(ao8 + 1); |
| 195 | + |
| 196 | + if (offset > 0) ao1 += lda; else ao1 += 2; |
| 197 | + if (offset > -1) ao2 += lda; else ao2 += 2; |
| 198 | + if (offset > -2) ao3 += lda; else ao3 += 2; |
| 199 | + if (offset > -3) ao4 += lda; else ao4 += 2; |
| 200 | + if (offset > -4) ao5 += lda; else ao5 += 2; |
| 201 | + if (offset > -5) ao6 += lda; else ao6 += 2; |
| 202 | + if (offset > -6) ao7 += lda; else ao7 += 2; |
| 203 | + if (offset > -7) ao8 += lda; else ao8 += 2; |
| 204 | + |
| 205 | + b[ 0] = data01; |
| 206 | + b[ 1] = data02; |
| 207 | + b[ 2] = data03; |
| 208 | + b[ 3] = data04; |
| 209 | + b[ 4] = data05; |
| 210 | + b[ 5] = data06; |
| 211 | + b[ 6] = data07; |
| 212 | + b[ 7] = data08; |
| 213 | + b[ 8] = data09; |
| 214 | + b[ 9] = data10; |
| 215 | + b[10] = data11; |
| 216 | + b[11] = data12; |
| 217 | + b[12] = data13; |
| 218 | + b[13] = data14; |
| 219 | + b[14] = data15; |
| 220 | + b[15] = data16; |
| 221 | + |
| 222 | + b += 16; |
| 223 | + |
| 224 | + offset --; |
| 225 | + i --; |
| 226 | + } |
| 227 | + |
| 228 | + posX += 8; |
| 229 | + } |
| 230 | + |
| 231 | + if (n & 4) { |
| 232 | + offset = posX - posY; |
| 233 | + |
| 234 | + if (offset > 0) ao1 = a + (posX + 0) * 2 + posY * lda; else ao1 = a + posY * 2 + (posX + 0) * lda; |
| 235 | + if (offset > -1) ao2 = a + (posX + 1) * 2 + posY * lda; else ao2 = a + posY * 2 + (posX + 1) * lda; |
| 236 | + if (offset > -2) ao3 = a + (posX + 2) * 2 + posY * lda; else ao3 = a + posY * 2 + (posX + 2) * lda; |
| 237 | + if (offset > -3) ao4 = a + (posX + 3) * 2 + posY * lda; else ao4 = a + posY * 2 + (posX + 3) * lda; |
| 238 | + |
| 239 | + i = m; |
| 240 | + |
| 241 | + while (i > 0) { |
| 242 | + data01 = *(ao1 + 0); |
| 243 | + data02 = *(ao1 + 1); |
| 244 | + data03 = *(ao2 + 0); |
| 245 | + data04 = *(ao2 + 1); |
| 246 | + data05 = *(ao3 + 0); |
| 247 | + data06 = *(ao3 + 1); |
| 248 | + data07 = *(ao4 + 0); |
| 249 | + data08 = *(ao4 + 1); |
| 250 | + |
| 251 | + if (offset > 0) ao1 += lda; else ao1 += 2; |
| 252 | + if (offset > -1) ao2 += lda; else ao2 += 2; |
| 253 | + if (offset > -2) ao3 += lda; else ao3 += 2; |
| 254 | + if (offset > -3) ao4 += lda; else ao4 += 2; |
| 255 | + |
| 256 | + b[ 0] = data01; |
| 257 | + b[ 1] = data02; |
| 258 | + b[ 2] = data03; |
| 259 | + b[ 3] = data04; |
| 260 | + b[ 4] = data05; |
| 261 | + b[ 5] = data06; |
| 262 | + b[ 6] = data07; |
| 263 | + b[ 7] = data08; |
| 264 | + |
| 265 | + b += 8; |
| 266 | + |
| 267 | + offset --; |
| 268 | + i --; |
| 269 | + } |
| 270 | + |
| 271 | + posX += 4; |
| 272 | + } |
| 273 | + |
| 274 | + if (n & 2) { |
| 275 | + |
| 276 | + offset = posX - posY; |
| 277 | + |
| 278 | + if (offset > 0) ao1 = a + (posX + 0) * 2 + posY * lda; else ao1 = a + posY * 2 + (posX + 0) * lda; |
| 279 | + if (offset > -1) ao2 = a + (posX + 1) * 2 + posY * lda; else ao2 = a + posY * 2 + (posX + 1) * lda; |
| 280 | + |
| 281 | + i = m; |
| 282 | + |
| 283 | + while (i > 0) { |
| 284 | + data01 = *(ao1 + 0); |
| 285 | + data02 = *(ao1 + 1); |
| 286 | + data03 = *(ao2 + 0); |
| 287 | + data04 = *(ao2 + 1); |
| 288 | + |
| 289 | + if (offset > 0) ao1 += lda; else ao1 += 2; |
| 290 | + if (offset > -1) ao2 += lda; else ao2 += 2; |
| 291 | + |
| 292 | + b[ 0] = data01; |
| 293 | + b[ 1] = data02; |
| 294 | + b[ 2] = data03; |
| 295 | + b[ 3] = data04; |
| 296 | + |
| 297 | + b += 4; |
| 298 | + |
| 299 | + offset --; |
| 300 | + i --; |
| 301 | + } |
| 302 | + |
| 303 | + posX += 2; |
| 304 | + |
| 305 | + } |
| 306 | + |
| 307 | + if (n & 1) { |
| 308 | + |
| 309 | + offset = posX - posY; |
| 310 | + |
| 311 | + if (offset > 0) ao1 = a + (posX + 0) * 2 + posY * lda; else ao1 = a + posY * 2 + (posX + 0) * lda; |
| 312 | + |
| 313 | + i = m; |
| 314 | + |
| 315 | + while (i > 0) { |
| 316 | + data01 = *(ao1 + 0); |
| 317 | + data02 = *(ao1 + 1); |
| 318 | + |
| 319 | + if (offset > 0) ao1 += lda; else ao1 += 2; |
| 320 | + |
| 321 | + b[ 0] = data01; |
| 322 | + b[ 1] = data02; |
| 323 | + |
| 324 | + b += 2; |
| 325 | + |
| 326 | + offset --; |
| 327 | + i --; |
| 328 | + } |
| 329 | + |
| 330 | + } |
| 331 | + |
| 332 | + return 0; |
| 333 | +} |
0 commit comments