The Android Open Source Project | 7f81d9b | 2009-03-03 19:30:08 -0800 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (C) 2008 The Android Open Source Project |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | /* ---- includes ----------------------------------------------------------- */ |
| 18 | |
| 19 | #include "b_TensorEm/Flt16Vec.h" |
| 20 | #include "b_BasicEm/Memory.h" |
| 21 | #include "b_BasicEm/Math.h" |
| 22 | #include "b_BasicEm/Functions.h" |
| 23 | |
| 24 | /* ------------------------------------------------------------------------- */ |
| 25 | |
| 26 | /* ========================================================================= */ |
| 27 | /* */ |
| 28 | /* ---- \ghd{ auxiliary functions } ---------------------------------------- */ |
| 29 | /* */ |
| 30 | /* ========================================================================= */ |
| 31 | |
| 32 | /* ------------------------------------------------------------------------- */ |
| 33 | |
| 34 | /* ========================================================================= */ |
| 35 | /* */ |
| 36 | /* ---- \ghd{ constructor / destructor } ----------------------------------- */ |
| 37 | /* */ |
| 38 | /* ========================================================================= */ |
| 39 | |
| 40 | /* ------------------------------------------------------------------------- */ |
| 41 | |
| 42 | void bts_Flt16Vec_init( struct bbs_Context* cpA, |
| 43 | struct bts_Flt16Vec* ptrA ) |
| 44 | { |
| 45 | bbs_Int16Arr_init( cpA, &ptrA->arrE ); |
| 46 | ptrA->expE = 0; |
| 47 | } |
| 48 | |
| 49 | /* ------------------------------------------------------------------------- */ |
| 50 | |
| 51 | void bts_Flt16Vec_exit( struct bbs_Context* cpA, |
| 52 | struct bts_Flt16Vec* ptrA ) |
| 53 | { |
| 54 | bbs_Int16Arr_exit( cpA, &ptrA->arrE ); |
| 55 | ptrA->expE = 0; |
| 56 | } |
| 57 | |
| 58 | /* ------------------------------------------------------------------------- */ |
| 59 | |
| 60 | /* ========================================================================= */ |
| 61 | /* */ |
| 62 | /* ---- \ghd{ operators } -------------------------------------------------- */ |
| 63 | /* */ |
| 64 | /* ========================================================================= */ |
| 65 | |
| 66 | /* ------------------------------------------------------------------------- */ |
| 67 | |
| 68 | void bts_Flt16Vec_copy( struct bbs_Context* cpA, |
| 69 | struct bts_Flt16Vec* ptrA, |
| 70 | const struct bts_Flt16Vec* srcPtrA ) |
| 71 | { |
| 72 | bbs_Int16Arr_copy( cpA, &ptrA->arrE, &srcPtrA->arrE ); |
| 73 | ptrA->expE = srcPtrA->expE; |
| 74 | } |
| 75 | |
| 76 | /* ------------------------------------------------------------------------- */ |
| 77 | |
| 78 | flag bts_Flt16Vec_equal( struct bbs_Context* cpA, |
| 79 | const struct bts_Flt16Vec* ptrA, |
| 80 | const struct bts_Flt16Vec* srcPtrA ) |
| 81 | { |
| 82 | if( !bbs_Int16Arr_equal( cpA, &ptrA->arrE, &srcPtrA->arrE ) ) return FALSE; |
| 83 | if( ptrA->expE != srcPtrA->expE ) return FALSE; |
| 84 | return TRUE; |
| 85 | } |
| 86 | |
| 87 | /* ------------------------------------------------------------------------- */ |
| 88 | |
| 89 | /* ========================================================================= */ |
| 90 | /* */ |
| 91 | /* ---- \ghd{ query functions } -------------------------------------------- */ |
| 92 | /* */ |
| 93 | /* ========================================================================= */ |
| 94 | |
| 95 | /* ------------------------------------------------------------------------- */ |
| 96 | |
| 97 | int16 bts_Flt16Vec_avg( struct bbs_Context* cpA, const struct bts_Flt16Vec* ptrA ) |
| 98 | { |
| 99 | uint16 iL; |
| 100 | uint16 sizeL = ptrA->arrE.sizeE; |
| 101 | int32 sumL = 0; |
| 102 | const int16* srcL = ptrA->arrE.arrPtrE; |
| 103 | for( iL = 0; iL < sizeL; iL++ ) |
| 104 | { |
| 105 | sumL += srcL[ iL ]; |
| 106 | } |
| 107 | return sumL / ( int32 )sizeL; |
| 108 | } |
| 109 | |
| 110 | /* ------------------------------------------------------------------------- */ |
| 111 | |
| 112 | uint32 bts_Flt16Vec_norm( struct bbs_Context* cpA, const struct bts_Flt16Vec* ptrA ) |
| 113 | { |
| 114 | return bbs_vecNorm16( ptrA->arrE.arrPtrE, ptrA->arrE.sizeE ); |
| 115 | } |
| 116 | |
| 117 | /* ------------------------------------------------------------------------- */ |
| 118 | |
| 119 | uint16 bts_Flt16Vec_maxAbs( struct bbs_Context* cpA, const struct bts_Flt16Vec* ptrA ) |
| 120 | { |
| 121 | uint16 iL; |
| 122 | uint16 sizeL = ptrA->arrE.sizeE; |
| 123 | uint16 maxL = 0; |
| 124 | const int16* srcL = ptrA->arrE.arrPtrE; |
| 125 | for( iL = 0; iL < sizeL; iL++ ) |
| 126 | { |
| 127 | uint16 vL = srcL[ iL ] > 0 ? srcL[ iL ] : -srcL[ iL ]; |
| 128 | maxL = vL > maxL ? vL : maxL; |
| 129 | } |
| 130 | return maxL; |
| 131 | } |
| 132 | |
| 133 | /* ------------------------------------------------------------------------- */ |
| 134 | |
| 135 | /* ========================================================================= */ |
| 136 | /* */ |
| 137 | /* ---- \ghd{ modify functions } ------------------------------------------- */ |
| 138 | /* */ |
| 139 | /* ========================================================================= */ |
| 140 | |
| 141 | /* ------------------------------------------------------------------------- */ |
| 142 | |
| 143 | void bts_Flt16Vec_create( struct bbs_Context* cpA, |
| 144 | struct bts_Flt16Vec* ptrA, |
| 145 | uint32 sizeA, |
| 146 | struct bbs_MemSeg* mspA ) |
| 147 | { |
| 148 | bbs_Int16Arr_create( cpA, &ptrA->arrE, sizeA, mspA ); |
| 149 | } |
| 150 | |
| 151 | /* ------------------------------------------------------------------------- */ |
| 152 | |
| 153 | void bts_Flt16Vec_size( struct bbs_Context* cpA, |
| 154 | struct bts_Flt16Vec* ptrA, |
| 155 | uint32 sizeA ) |
| 156 | { |
| 157 | bbs_Int16Arr_size( cpA, &ptrA->arrE, sizeA ); |
| 158 | } |
| 159 | |
| 160 | /* ------------------------------------------------------------------------- */ |
| 161 | |
| 162 | /* ========================================================================= */ |
| 163 | /* */ |
| 164 | /* ---- \ghd{ I/O } -------------------------------------------------------- */ |
| 165 | /* */ |
| 166 | /* ========================================================================= */ |
| 167 | |
| 168 | /* ------------------------------------------------------------------------- */ |
| 169 | |
| 170 | uint32 bts_Flt16Vec_memSize( struct bbs_Context* cpA, |
| 171 | const struct bts_Flt16Vec *ptrA ) |
| 172 | { |
| 173 | return bbs_SIZEOF16( uint32 ) /* mem size */ |
| 174 | + bbs_Int16Arr_memSize( cpA, &ptrA->arrE ) |
| 175 | + bbs_SIZEOF16( ptrA->expE ); |
| 176 | } |
| 177 | |
| 178 | /* ------------------------------------------------------------------------- */ |
| 179 | |
| 180 | uint32 bts_Flt16Vec_memWrite( struct bbs_Context* cpA, |
| 181 | const struct bts_Flt16Vec* ptrA, |
| 182 | uint16* memPtrA ) |
| 183 | { |
| 184 | uint32 memSizeL = bts_Flt16Vec_memSize( cpA, ptrA ); |
| 185 | memPtrA += bbs_memWrite32( &memSizeL, memPtrA ); |
| 186 | memPtrA += bbs_Int16Arr_memWrite( cpA, &ptrA->arrE, memPtrA ); |
| 187 | memPtrA += bbs_memWrite16( &ptrA->expE, memPtrA ); |
| 188 | return memSizeL; |
| 189 | } |
| 190 | |
| 191 | /* ------------------------------------------------------------------------- */ |
| 192 | |
| 193 | uint32 bts_Flt16Vec_memRead( struct bbs_Context* cpA, |
| 194 | struct bts_Flt16Vec* ptrA, |
| 195 | const uint16* memPtrA, |
| 196 | struct bbs_MemSeg* mspA ) |
| 197 | { |
| 198 | uint32 memSizeL; |
| 199 | if( bbs_Context_error( cpA ) ) return 0; |
| 200 | memPtrA += bbs_memRead32( &memSizeL, memPtrA ); |
| 201 | memPtrA += bbs_Int16Arr_memRead( cpA, &ptrA->arrE, memPtrA, mspA ); |
| 202 | memPtrA += bbs_memRead16( &ptrA->expE, memPtrA ); |
| 203 | |
| 204 | if( memSizeL != bts_Flt16Vec_memSize( cpA, ptrA ) ) |
| 205 | { |
| 206 | bbs_ERR0( bbs_ERR_CORRUPT_DATA, "uint32 bts_Flt16Vec_memRead( const struct bts_Flt16Vec* ptrA, const void* memPtrA ):\n" |
| 207 | "size mismatch" ); |
| 208 | return 0; |
| 209 | } |
| 210 | return memSizeL; |
| 211 | } |
| 212 | |
| 213 | /* ------------------------------------------------------------------------- */ |
| 214 | |
| 215 | /* ========================================================================= */ |
| 216 | /* */ |
| 217 | /* ---- \ghd{ exec functions } --------------------------------------------- */ |
| 218 | /* */ |
| 219 | /* ========================================================================= */ |
| 220 | |
| 221 | /* ------------------------------------------------------------------------- */ |
| 222 | |
| 223 | void bts_Flt16Vec_maximizeMantisse( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA ) |
| 224 | { |
| 225 | uint32 maxAbsL = bts_Flt16Vec_maxAbs( cpA, ptrA ); |
| 226 | int16 shlL = 0; |
| 227 | |
| 228 | if( maxAbsL == 0 ) return; /* cannot maximize 0 */ |
| 229 | |
| 230 | while( maxAbsL < 0x4000 ) |
| 231 | { |
| 232 | shlL++; |
| 233 | maxAbsL <<= 1; |
| 234 | } |
| 235 | |
| 236 | if( shlL > 0 ) |
| 237 | { |
| 238 | uint32 iL; |
| 239 | uint32 sizeL = ptrA->arrE.sizeE; |
| 240 | int16* dstL = ptrA->arrE.arrPtrE; |
| 241 | for( iL = 0; iL < sizeL; iL++ ) dstL[ iL ] <<= shlL; |
| 242 | ptrA->expE -= shlL; |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | /* ------------------------------------------------------------------------- */ |
| 247 | |
| 248 | uint32 bts_Flt16Vec_maximizeAbsValue( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA ) |
| 249 | { |
| 250 | int32 maxAbsL = bts_Flt16Vec_maxAbs( cpA, ptrA ); |
| 251 | int32 fL; |
| 252 | if( maxAbsL == 0 ) return 0; /* vector is zero */ |
| 253 | |
| 254 | fL = ( int32 )0x7FFF0000 / maxAbsL; |
| 255 | |
| 256 | { |
| 257 | uint32 iL; |
| 258 | uint32 sizeL = ptrA->arrE.sizeE; |
| 259 | int16* dstL = ptrA->arrE.arrPtrE; |
| 260 | for( iL = 0; iL < sizeL; iL++ ) dstL[ iL ] = ( ( int32 )dstL[ iL ] * fL + 32768 ) >> 16; |
| 261 | } |
| 262 | |
| 263 | return fL; |
| 264 | } |
| 265 | |
| 266 | /* ------------------------------------------------------------------------- */ |
| 267 | |
| 268 | void bts_Flt16Vec_zeroAverage( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA ) |
| 269 | { |
| 270 | uint16 iL; |
| 271 | uint16 sizeL = ptrA->arrE.sizeE; |
| 272 | int16* dstL = ptrA->arrE.arrPtrE; |
| 273 | int16 avgL = bts_Flt16Vec_avg( cpA, ptrA ); |
| 274 | for( iL = 0; iL < sizeL; iL++ ) dstL[ iL ] -= avgL; |
| 275 | } |
| 276 | |
| 277 | /* ------------------------------------------------------------------------- */ |
| 278 | |
| 279 | void bts_Flt16Vec_normalize( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA ) |
| 280 | { |
| 281 | uint32 normL = bts_Flt16Vec_norm( cpA, ptrA ); |
| 282 | |
| 283 | if( normL == 0 ) |
| 284 | { |
| 285 | /* vector is zero - do nothing */ |
| 286 | return; |
| 287 | } |
| 288 | else |
| 289 | { |
| 290 | int16* dstL = ptrA->arrE.arrPtrE; |
| 291 | uint16 iL; |
| 292 | uint16 sizeL = ptrA->arrE.sizeE; |
| 293 | int16 expL = 0; |
| 294 | int32 fL; |
| 295 | |
| 296 | /* let norm occupy 17 bits */ |
| 297 | if( ( normL & 0xFFFE0000 ) != 0 ) |
| 298 | { |
| 299 | while( ( ( normL >> -expL ) & 0xFFFE0000 ) != 0 ) expL--; |
| 300 | normL >>= -expL; |
| 301 | } |
| 302 | else |
| 303 | { |
| 304 | while( ( ( normL << expL ) & 0xFFFF0000 ) == 0 ) expL++; |
| 305 | normL <<= expL; |
| 306 | } |
| 307 | |
| 308 | /* fL is positive and occupies only 16 bits - a product with int16 fits in int32 */ |
| 309 | fL = ( uint32 )0xFFFFFFFF / normL; |
| 310 | |
| 311 | /* multiply with factor */ |
| 312 | for( iL = 0; iL < sizeL; iL++ ) dstL[ iL ] = ( ( ( ( int32 )dstL[ iL ] * fL ) >> 15 ) + 1 ) >> 1; |
| 313 | |
| 314 | /* set exponent */ |
| 315 | ptrA->expE = expL - 16; |
| 316 | } |
| 317 | /* |
| 318 | { |
| 319 | uint32 testNormL = bts_Flt16Vec_norm( cpA, ptrA ); |
| 320 | printf( "test norm %f\n", ( float )testNormL / ( 1 << -ptrA->expE ) ); |
| 321 | } |
| 322 | */ |
| 323 | } |
| 324 | |
| 325 | /* ------------------------------------------------------------------------- */ |
| 326 | |
| 327 | void bts_Flt16Vec_setZero( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA ) |
| 328 | { |
| 329 | bbs_Int16Arr_fill( cpA, &ptrA->arrE, 0 ); |
| 330 | ptrA->expE = 0; |
| 331 | } |
| 332 | |
| 333 | /* ------------------------------------------------------------------------- */ |
| 334 | |
| 335 | void bts_Flt16Vec_mul( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA, int32 valA, int16 expA ) |
| 336 | { |
| 337 | int32 valL = valA; |
| 338 | int16 expL = expA; |
| 339 | |
| 340 | if( valL == 0 ) |
| 341 | { |
| 342 | bts_Flt16Vec_setZero( cpA, ptrA ); |
| 343 | return; |
| 344 | } |
| 345 | else |
| 346 | { |
| 347 | uint32 iL; |
| 348 | uint32 sizeL = ptrA->arrE.sizeE; |
| 349 | int16* dstL = ptrA->arrE.arrPtrE; |
| 350 | |
| 351 | /* adjust valL to maximum 16 bit accuracy */ |
| 352 | uint32 absValL = valL > 0 ? valL : -valL; |
| 353 | if( ( absValL & 0xFFFF8000 ) != 0 ) |
| 354 | { |
| 355 | int32 shrL = 0; |
| 356 | while( ( absValL & 0xFFFF8000 ) != 0 ) |
| 357 | { |
| 358 | absValL >>= 1; |
| 359 | shrL++; |
| 360 | } |
| 361 | |
| 362 | if( shrL > 0 ) |
| 363 | { |
| 364 | valL = ( ( valL >> ( shrL - 1 ) ) + 1 ) >> 1; |
| 365 | expL += shrL; |
| 366 | if( valL >= 0x08000 ) valL = 0x07FFF; /* saturate */ |
| 367 | } |
| 368 | } |
| 369 | else |
| 370 | { |
| 371 | int32 shlL = 0; |
| 372 | while( ( absValL & 0xFFFFC000 ) == 0 ) |
| 373 | { |
| 374 | absValL <<= 1; |
| 375 | shlL++; |
| 376 | } |
| 377 | |
| 378 | valL <<= shlL; |
| 379 | expL -= shlL; |
| 380 | } |
| 381 | |
| 382 | for( iL = 0; iL < sizeL; iL++ ) |
| 383 | { |
| 384 | dstL[ iL ] = ( ( ( ( int32 )dstL[ iL ] * valL ) >> 15 ) + 1 ) >> 1; |
| 385 | } |
| 386 | ptrA->expE += expL + 16; |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | /* ------------------------------------------------------------------------- */ |
| 391 | |
| 392 | void bts_Flt16Vec_dotPtrd( struct bbs_Context* cpA, struct bts_Flt16Vec* vp1A, struct bts_Flt16Vec* vp2A, int32* manPtrA, int32* expPtrA ) |
| 393 | { |
| 394 | bbs_DEF_fNameL( "void bts_Flt16Vec_dotPtrd( struct bbs_Context* cpA, struct bts_Flt16Vec* vp1A, struct bts_Flt16Vec* vp2A, int32* matPtrA, int32* expPtrA )" ) |
| 395 | uint16 iL; |
| 396 | uint16 sizeL = vp1A->arrE.sizeE; |
| 397 | const int16* arr1L = vp1A->arrE.arrPtrE; |
| 398 | const int16* arr2L = vp2A->arrE.arrPtrE; |
| 399 | int16 shrm1L = -1; /* shift minus 1 */ |
| 400 | int32 sumL; |
| 401 | |
| 402 | if( vp1A->arrE.sizeE != vp2A->arrE.sizeE ) |
| 403 | { |
| 404 | bbs_ERROR1( "%s:\nVectors have different size", fNameL ); |
| 405 | return; |
| 406 | } |
| 407 | |
| 408 | sumL = 0; |
| 409 | /* shrm1L == -1 */ |
| 410 | for( iL = 0; iL < sizeL; iL++ ) |
| 411 | { |
| 412 | sumL += ( int32 )arr1L[ iL ] * ( int32 )arr2L[ iL ]; |
| 413 | if( ( ( ( sumL > 0 ) ? sumL : -sumL ) & 0xC0000000 ) != 0 ) break; |
| 414 | } |
| 415 | |
| 416 | if( iL < sizeL ) |
| 417 | { |
| 418 | /* danger of overflow: increase shift; adjust sum */ |
| 419 | shrm1L++; |
| 420 | sumL = ( ( sumL >> 1 ) + 1 ) >> 1; |
| 421 | |
| 422 | /* shrm1L == 0 */ |
| 423 | for( iL = 0; iL < sizeL; iL++ ) |
| 424 | { |
| 425 | sumL += ( int32 )( ( arr1L[ iL ] + 1 ) >> 1 ) * ( int32 )( ( arr2L[ iL ] + 1 ) >> 1 ); |
| 426 | if( ( ( ( sumL > 0 ) ? sumL : -sumL ) & 0xC0000000 ) != 0 ) break; |
| 427 | } |
| 428 | |
| 429 | for( iL = 0; iL < sizeL; iL++ ) |
| 430 | { |
| 431 | if( ( ( ( sumL > 0 ) ? sumL : -sumL ) & 0xC0000000 ) != 0 ) |
| 432 | { |
| 433 | /* danger of overflow: increase shift; adjust sum */ |
| 434 | shrm1L++; |
| 435 | sumL = ( ( sumL >> 1 ) + 1 ) >> 1; |
| 436 | } |
| 437 | |
| 438 | sumL += ( int32 )( ( ( arr1L[ iL ] >> shrm1L ) + 1 ) >> 1 ) * ( int32 )( ( ( arr2L[ iL ] >> shrm1L ) + 1 ) >> 1 ); |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | if( manPtrA != NULL ) *manPtrA = sumL; |
| 443 | if( expPtrA != NULL ) *expPtrA = vp1A->expE + vp2A->expE + ( ( shrm1L + 1 ) << 1 ); |
| 444 | } |
| 445 | |
| 446 | /* ------------------------------------------------------------------------- */ |
| 447 | |
| 448 | void bts_Flt16Vec_append( struct bbs_Context* cpA, struct bts_Flt16Vec* ptrA, struct bts_Flt16Vec* srcPtrA ) |
| 449 | { |
| 450 | if( ptrA->arrE.sizeE == 0 ) |
| 451 | { |
| 452 | bts_Flt16Vec_copy( cpA, ptrA, srcPtrA ); |
| 453 | } |
| 454 | else |
| 455 | { |
| 456 | uint32 idxL = ptrA->arrE.sizeE; |
| 457 | bts_Flt16Vec_size( cpA, ptrA, idxL + srcPtrA->arrE.sizeE ); |
| 458 | |
| 459 | /* copy data */ |
| 460 | bbs_memcpy16( ptrA->arrE.arrPtrE + idxL, srcPtrA->arrE.arrPtrE, srcPtrA->arrE.sizeE ); |
| 461 | |
| 462 | /* equalize exponent */ |
| 463 | if( ptrA->expE > srcPtrA->expE ) |
| 464 | { |
| 465 | uint32 iL; |
| 466 | uint32 sizeL = srcPtrA->arrE.sizeE; |
| 467 | uint32 shrL = ptrA->expE - srcPtrA->expE; |
| 468 | int16* dstL = ptrA->arrE.arrPtrE + idxL; |
| 469 | for( iL = 0; iL < sizeL; iL++ ) dstL[ iL ] = ( ( dstL[ iL ] >> ( shrL - 1 ) ) + 1 ) >> 1; |
| 470 | } |
| 471 | else if( ptrA->expE < srcPtrA->expE ) |
| 472 | { |
| 473 | uint32 iL; |
| 474 | uint32 sizeL = idxL; |
| 475 | uint32 shrL = srcPtrA->expE - ptrA->expE; |
| 476 | int16* dstL = ptrA->arrE.arrPtrE; |
| 477 | for( iL = 0; iL < sizeL; iL++ ) dstL[ iL ] = ( ( dstL[ iL ] >> ( shrL - 1 ) ) + 1 ) >> 1; |
| 478 | ptrA->expE = srcPtrA->expE; |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | /* ------------------------------------------------------------------------- */ |
| 484 | |
| 485 | /* ========================================================================= */ |
| 486 | |