1/* Copyright (C) 2011 IBM 2 3 Author: Maynard Johnson <maynardj@us.ibm.com> 4 5 This program is free software; you can redistribute it and/or 6 modify it under the terms of the GNU General Public License as 7 published by the Free Software Foundation; either version 2 of the 8 License, or (at your option) any later version. 9 10 This program is distributed in the hope that it will be useful, but 11 WITHOUT ANY WARRANTY; without even the implied warranty of 12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 13 General Public License for more details. 14 15 You should have received a copy of the GNU General Public License 16 along with this program; if not, write to the Free Software 17 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 18 02111-1307, USA. 19 20 The GNU General Public License is contained in the file COPYING. 21 */ 22 23#ifdef HAS_VSX 24 25#include <stdio.h> 26#include <stdint.h> 27#include <stdlib.h> 28#include <string.h> 29#include <malloc.h> 30#include <altivec.h> 31#include <math.h> 32 33#ifndef __powerpc64__ 34typedef uint32_t HWord_t; 35#else 36typedef uint64_t HWord_t; 37#endif /* __powerpc64__ */ 38 39typedef unsigned char Bool; 40#define True 1 41#define False 0 42register HWord_t r14 __asm__ ("r14"); 43register HWord_t r15 __asm__ ("r15"); 44register HWord_t r16 __asm__ ("r16"); 45register HWord_t r17 __asm__ ("r17"); 46register double f14 __asm__ ("fr14"); 47register double f15 __asm__ ("fr15"); 48register double f16 __asm__ ("fr16"); 49register double f17 __asm__ ("fr17"); 50 51static volatile unsigned int div_flags, div_xer; 52 53#define ALLCR "cr0","cr1","cr2","cr3","cr4","cr5","cr6","cr7" 54 55#define SET_CR(_arg) \ 56 __asm__ __volatile__ ("mtcr %0" : : "b"(_arg) : ALLCR ); 57 58#define SET_XER(_arg) \ 59 __asm__ __volatile__ ("mtxer %0" : : "b"(_arg) : "xer" ); 60 61#define GET_CR(_lval) \ 62 __asm__ __volatile__ ("mfcr %0" : "=b"(_lval) ) 63 64#define GET_XER(_lval) \ 65 __asm__ __volatile__ ("mfxer %0" : "=b"(_lval) ) 66 67#define GET_CR_XER(_lval_cr,_lval_xer) \ 68 do { GET_CR(_lval_cr); GET_XER(_lval_xer); } while (0) 69 70#define SET_CR_ZERO \ 71 SET_CR(0) 72 73#define SET_XER_ZERO \ 74 SET_XER(0) 75 76#define SET_CR_XER_ZERO \ 77 do { SET_CR_ZERO; SET_XER_ZERO; } while (0) 78 79#define SET_FPSCR_ZERO \ 80 do { double _d = 0.0; \ 81 __asm__ __volatile__ ("mtfsf 0xFF, %0" : : "f"(_d) ); \ 82 } while (0) 83 84 85typedef void (*test_func_t)(void); 86typedef struct test_table test_table_t; 87 88 89/* These functions below that construct a table of floating point 90 * values were lifted from none/tests/ppc32/jm-insns.c. 91 */ 92 93#if defined (DEBUG_ARGS_BUILD) 94#define AB_DPRINTF(fmt, args...) do { fprintf(stderr, fmt , ##args); } while (0) 95#else 96#define AB_DPRINTF(fmt, args...) do { } while (0) 97#endif 98 99static inline void register_farg (void *farg, 100 int s, uint16_t _exp, uint64_t mant) 101{ 102 uint64_t tmp; 103 104 tmp = ((uint64_t)s << 63) | ((uint64_t)_exp << 52) | mant; 105 *(uint64_t *)farg = tmp; 106 AB_DPRINTF("%d %03x %013llx => %016llx %0e\n", 107 s, _exp, mant, *(uint64_t *)farg, *(double *)farg); 108} 109 110static inline void register_sp_farg (void *farg, 111 int s, uint16_t _exp, uint32_t mant) 112{ 113 uint32_t tmp; 114 tmp = ((uint32_t)s << 31) | ((uint32_t)_exp << 23) | mant; 115 *(uint32_t *)farg = tmp; 116} 117 118typedef struct fp_test_args { 119 int fra_idx; 120 int frb_idx; 121} fp_test_args_t; 122 123 124fp_test_args_t fp_cmp_tests[] = { 125 {8, 8}, 126 {8, 14}, 127 {8, 6}, 128 {8, 5}, 129 {8, 4}, 130 {8, 7}, 131 {8, 9}, 132 {8, 11}, 133 {14, 8}, 134 {14, 14}, 135 {14, 6}, 136 {14, 5}, 137 {14, 4}, 138 {14, 7}, 139 {14, 9}, 140 {14, 11}, 141 {6, 8}, 142 {6, 14}, 143 {6, 6}, 144 {6, 5}, 145 {6, 4}, 146 {6, 7}, 147 {6, 9}, 148 {6, 11}, 149 {5, 8}, 150 {5, 14}, 151 {5, 6}, 152 {5, 5}, 153 {5, 4}, 154 {5, 7}, 155 {5, 9}, 156 {5, 11}, 157 {4, 8}, 158 {4, 14}, 159 {4, 6}, 160 {4, 5}, 161 {4, 1}, 162 {4, 7}, 163 {4, 9}, 164 {4, 11}, 165 {7, 8}, 166 {7, 14}, 167 {7, 6}, 168 {7, 5}, 169 {7, 4}, 170 {7, 7}, 171 {7, 9}, 172 {7, 11}, 173 {10, 8}, 174 {10, 14}, 175 {10, 6}, 176 {10, 5}, 177 {10, 4}, 178 {10, 7}, 179 {10, 9}, 180 {10, 10}, 181 {12, 8}, 182 {12, 14}, 183 {12, 6}, 184 {12, 5}, 185 {1, 1}, 186 {2, 2}, 187 {3, 3}, 188 {4, 4}, 189}; 190 191 192fp_test_args_t two_arg_fp_tests[] = { 193 {8, 8}, 194 {8, 14}, 195 {15, 16}, 196 {8, 5}, 197 {8, 4}, 198 {8, 7}, 199 {8, 9}, 200 {8, 11}, 201 {14, 8}, 202 {14, 14}, 203 {14, 6}, 204 {14, 5}, 205 {14, 4}, 206 {14, 7}, 207 {14, 9}, 208 {14, 11}, 209 {6, 8}, 210 {6, 14}, 211 {6, 6}, 212 {6, 5}, 213 {6, 4}, 214 {6, 7}, 215 {6, 9}, 216 {6, 11}, 217 {5, 8}, 218 {5, 14}, 219 {5, 6}, 220 {5, 5}, 221 {5, 4}, 222 {5, 7}, 223 {5, 9}, 224 {5, 11}, 225 {4, 8}, 226 {4, 14}, 227 {4, 6}, 228 {4, 5}, 229 {4, 1}, 230 {4, 7}, 231 {4, 9}, 232 {4, 11}, 233 {7, 8}, 234 {7, 14}, 235 {7, 6}, 236 {7, 5}, 237 {7, 4}, 238 {7, 7}, 239 {7, 9}, 240 {7, 11}, 241 {10, 8}, 242 {10, 14}, 243 {12, 6}, 244 {12, 5}, 245 {10, 4}, 246 {10, 7}, 247 {10, 9}, 248 {10, 11}, 249 {12, 8 }, 250 {12, 14}, 251 {12, 6}, 252 {15, 16}, 253 {15, 16}, 254 {9, 11}, 255 {11, 11}, 256 {11, 12} 257}; 258 259 260static int nb_special_fargs; 261static double * spec_fargs; 262static float * spec_sp_fargs; 263 264static void build_special_fargs_table(void) 265{ 266/* 267 Entry Sign Exp fraction Special value 268 0 0 3fd 0x8000000000000ULL Positive finite number 269 1 0 404 0xf000000000000ULL ... 270 2 0 001 0x8000000b77501ULL ... 271 3 0 7fe 0x800000000051bULL ... 272 4 0 012 0x3214569900000ULL ... 273 5 0 000 0x0000000000000ULL +0.0 (+zero) 274 6 1 000 0x0000000000000ULL -0.0 (-zero) 275 7 0 7ff 0x0000000000000ULL +infinity 276 8 1 7ff 0x0000000000000ULL -infinity 277 9 0 7ff 0x7FFFFFFFFFFFFULL +SNaN 278 10 1 7ff 0x7FFFFFFFFFFFFULL -SNaN 279 11 0 7ff 0x8000000000000ULL +QNaN 280 12 1 7ff 0x8000000000000ULL -QNaN 281 13 1 000 0x8340000078000ULL Denormalized val (zero exp and non-zero fraction) 282 14 1 40d 0x0650f5a07b353ULL Negative finite number 283 15 0 412 0x32585a9900000ULL A couple more positive finite numbers 284 16 0 413 0x82511a2000000ULL ... 285*/ 286 287 uint64_t mant; 288 uint32_t mant_sp; 289 uint16_t _exp; 290 int s; 291 int j, i = 0; 292 293 if (spec_fargs) 294 return; 295 296 spec_fargs = malloc( 17 * sizeof(double) ); 297 spec_sp_fargs = malloc( 17 * sizeof(float) ); 298 299 // #0 300 s = 0; 301 _exp = 0x3fd; 302 mant = 0x8000000000000ULL; 303 register_farg(&spec_fargs[i++], s, _exp, mant); 304 305 // #1 306 s = 0; 307 _exp = 0x404; 308 mant = 0xf000000000000ULL; 309 register_farg(&spec_fargs[i++], s, _exp, mant); 310 311 /* None of the ftdiv tests succeed. 312 * FRA = value #0; FRB = value #1 313 * ea_ = -2; e_b = 5 314 * fl_flag || fg_flag || fe_flag = 100 315 */ 316 317 /************************************************* 318 * fe_flag tests 319 * 320 *************************************************/ 321 322 /* fe_flag <- 1 if FRA is a NaN 323 * FRA = value #9; FRB = value #1 324 * e_a = 1024; e_b = 5 325 * fl_flag || fg_flag || fe_flag = 101 326 */ 327 328 /* fe_flag <- 1 if FRB is a NaN 329 * FRA = value #1; FRB = value #12 330 * e_a = 5; e_b = 1024 331 * fl_flag || fg_flag || fe_flag = 101 332 */ 333 334 /* fe_flag <- 1 if e_b <= -1022 335 * FRA = value #0; FRB = value #2 336 * e_a = -2; e_b = -1022 337 * fl_flag || fg_flag || fe_flag = 101 338 * 339 */ 340 // #2 341 s = 0; 342 _exp = 0x001; 343 mant = 0x8000000b77501ULL; 344 register_farg(&spec_fargs[i++], s, _exp, mant); 345 346 /* fe_flag <- 1 if e_b >= 1021 347 * FRA = value #1; FRB = value #3 348 * e_a = 5; e_b = 1023 349 * fl_flag || fg_flag || fe_flag = 101 350 */ 351 // #3 352 s = 0; 353 _exp = 0x7fe; 354 mant = 0x800000000051bULL; 355 register_farg(&spec_fargs[i++], s, _exp, mant); 356 357 /* fe_flag <- 1 if FRA != 0 && e_a - e_b >= 1023 358 * Let FRA = value #3 and FRB be value #0. 359 * e_a = 1023; e_b = -2 360 * fl_flag || fg_flag || fe_flag = 101 361 */ 362 363 /* fe_flag <- 1 if FRA != 0 && e_a - e_b <= -1023 364 * Let FRA = value #0 above and FRB be value #3 above 365 * e_a = -2; e_b = 1023 366 * fl_flag || fg_flag || fe_flag = 101 367 */ 368 369 /* fe_flag <- 1 if FRA != 0 && e_a <= -970 370 * Let FRA = value #4 and FRB be value #0 371 * e_a = -1005; e_b = -2 372 * fl_flag || fg_flag || fe_flag = 101 373 */ 374 // #4 375 s = 0; 376 _exp = 0x012; 377 mant = 0x3214569900000ULL; 378 register_farg(&spec_fargs[i++], s, _exp, mant); 379 380 /************************************************* 381 * fg_flag tests 382 * 383 *************************************************/ 384 /* fg_flag <- 1 if FRA is an Infinity 385 * NOTE: FRA = Inf also sets fe_flag 386 * Do two tests, using values #7 and #8 (+/- Inf) for FRA. 387 * Test 1: 388 * Let FRA be value #7 and FRB be value #1 389 * e_a = 1024; e_b = 5 390 * fl_flag || fg_flag || fe_flag = 111 391 * 392 * Test 2: 393 * Let FRA be value #8 and FRB be value #1 394 * e_a = 1024; e_b = 5 395 * fl_flag || fg_flag || fe_flag = 111 396 * 397 */ 398 399 /* fg_flag <- 1 if FRB is an Infinity 400 * NOTE: FRB = Inf also sets fe_flag 401 * Let FRA be value #1 and FRB be value #7 402 * e_a = 5; e_b = 1024 403 * fl_flag || fg_flag || fe_flag = 111 404 */ 405 406 /* fg_flag <- 1 if FRB is denormalized 407 * NOTE: e_b < -1022 ==> fe_flag <- 1 408 * Let FRA be value #0 and FRB be value #13 409 * e_a = -2; e_b = -1023 410 * fl_flag || fg_flag || fe_flag = 111 411 */ 412 413 /* fg_flag <- 1 if FRB is +zero 414 * NOTE: FRA = Inf also sets fe_flag 415 * Let FRA = val #5; FRB = val #5 416 * ea_ = -1023; e_b = -1023 417 * fl_flag || fg_flag || fe_flag = 111 418 */ 419 420 /* fg_flag <- 1 if FRB is -zero 421 * NOTE: FRA = Inf also sets fe_flag 422 * Let FRA = val #5; FRB = val #6 423 * ea_ = -1023; e_b = -1023 424 * fl_flag || fg_flag || fe_flag = 111 425 */ 426 427 /* Special values */ 428 /* +0.0 : 0 0x000 0x0000000000000 */ 429 // #5 430 s = 0; 431 _exp = 0x000; 432 mant = 0x0000000000000ULL; 433 register_farg(&spec_fargs[i++], s, _exp, mant); 434 435 /* -0.0 : 1 0x000 0x0000000000000 */ 436 // #6 437 s = 1; 438 _exp = 0x000; 439 mant = 0x0000000000000ULL; 440 register_farg(&spec_fargs[i++], s, _exp, mant); 441 442 /* +infinity : 0 0x7FF 0x0000000000000 */ 443 // #7 444 s = 0; 445 _exp = 0x7FF; 446 mant = 0x0000000000000ULL; 447 register_farg(&spec_fargs[i++], s, _exp, mant); 448 449 /* -infinity : 1 0x7FF 0x0000000000000 */ 450 // #8 451 s = 1; 452 _exp = 0x7FF; 453 mant = 0x0000000000000ULL; 454 register_farg(&spec_fargs[i++], s, _exp, mant); 455 456 /* 457 * This comment applies to values #9 and #10 below: 458 * When src is a SNaN, it's converted to a QNaN first before rounding to single-precision, 459 * so we can't just copy the double-precision value to the corresponding slot in the 460 * single-precision array (i.e., in the loop at the end of this function). Instead, we 461 * have to manually set the bits using register_sp_farg(). 462 */ 463 464 /* +SNaN : 0 0x7FF 0x7FFFFFFFFFFFF */ 465 // #9 466 s = 0; 467 _exp = 0x7FF; 468 mant = 0x7FFFFFFFFFFFFULL; 469 register_farg(&spec_fargs[i++], s, _exp, mant); 470 _exp = 0xff; 471 mant_sp = 0x3FFFFF; 472 register_sp_farg(&spec_sp_fargs[i-1], s, _exp, mant_sp); 473 474 /* -SNaN : 1 0x7FF 0x7FFFFFFFFFFFF */ 475 // #10 476 s = 1; 477 _exp = 0x7FF; 478 mant = 0x7FFFFFFFFFFFFULL; 479 register_farg(&spec_fargs[i++], s, _exp, mant); 480 _exp = 0xff; 481 mant_sp = 0x3FFFFF; 482 register_sp_farg(&spec_sp_fargs[i-1], s, _exp, mant_sp); 483 484 /* +QNaN : 0 0x7FF 0x8000000000000 */ 485 // #11 486 s = 0; 487 _exp = 0x7FF; 488 mant = 0x8000000000000ULL; 489 register_farg(&spec_fargs[i++], s, _exp, mant); 490 491 /* -QNaN : 1 0x7FF 0x8000000000000 */ 492 // #12 493 s = 1; 494 _exp = 0x7FF; 495 mant = 0x8000000000000ULL; 496 register_farg(&spec_fargs[i++], s, _exp, mant); 497 498 /* denormalized value */ 499 // #13 500 s = 1; 501 _exp = 0x000; 502 mant = 0x8340000078000ULL; 503 register_farg(&spec_fargs[i++], s, _exp, mant); 504 505 /* Negative finite number */ 506 // #14 507 s = 1; 508 _exp = 0x40d; 509 mant = 0x0650f5a07b353ULL; 510 register_farg(&spec_fargs[i++], s, _exp, mant); 511 512 /* A couple positive finite numbers ... */ 513 // #15 514 s = 0; 515 _exp = 0x412; 516 mant = 0x32585a9900000ULL; 517 register_farg(&spec_fargs[i++], s, _exp, mant); 518 519 // #16 520 s = 0; 521 _exp = 0x413; 522 mant = 0x82511a2000000ULL; 523 register_farg(&spec_fargs[i++], s, _exp, mant); 524 525 nb_special_fargs = i; 526 for (j = 0; j < i; j++) { 527 if (!(j == 9 || j == 10)) 528 spec_sp_fargs[j] = spec_fargs[j]; 529 } 530} 531 532 533struct test_table 534{ 535 test_func_t test_category; 536 char * name; 537}; 538 539typedef enum { 540 SINGLE_TEST, 541 DOUBLE_TEST 542} precision_type_t; 543 544typedef enum { 545 VX_SCALAR_FP_NMSUB = 0, 546 // ALL VECTOR-TYPE OPS SHOULD BE ADDED AFTER THIS LINE 547 VX_VECTOR_FP_MULT_AND_OP2 = 10, 548 // and before this line 549 VX_BASIC_CMP = 30, 550 VX_CONV_WORD, 551 VX_DEFAULT 552} vx_fp_test_type; 553 554typedef struct vx_fp_test 555{ 556 test_func_t test_func; 557 const char * name; 558 fp_test_args_t * targs; 559 int num_tests; 560 precision_type_t precision; 561 vx_fp_test_type type; 562 const char * op; 563} vx_fp_test_t; 564 565static vector unsigned int vec_out, vec_inA, vec_inB, vec_inC; 566 567static Bool do_dot; 568static void test_xvcmpeqdp(void) 569{ 570 if (do_dot) 571 __asm__ __volatile__ ("xvcmpeqdp. %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 572 else 573 __asm__ __volatile__ ("xvcmpeqdp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 574} 575 576static void test_xvcmpgedp(void) 577{ 578 if (do_dot) 579 __asm__ __volatile__ ("xvcmpgedp. %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 580 else 581 __asm__ __volatile__ ("xvcmpgedp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 582} 583 584static void test_xvcmpgtdp(void) 585{ 586 if (do_dot) 587 __asm__ __volatile__ ("xvcmpgtdp. %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 588 else 589 __asm__ __volatile__ ("xvcmpgtdp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 590} 591 592static void test_xvcmpeqsp(void) 593{ 594 if (do_dot) 595 __asm__ __volatile__ ("xvcmpeqsp. %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 596 else 597 __asm__ __volatile__ ("xvcmpeqsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 598} 599 600static void test_xvcmpgesp(void) 601{ 602 if (do_dot) 603 __asm__ __volatile__ ("xvcmpgesp. %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 604 else 605 __asm__ __volatile__ ("xvcmpgesp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 606} 607 608static void test_xvcmpgtsp(void) 609{ 610 if (do_dot) 611 __asm__ __volatile__ ("xvcmpgtsp. %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 612 else 613 __asm__ __volatile__ ("xvcmpgtsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 614} 615 616static Bool do_aXp; 617static Bool do_dp; 618static void test_xsnmsub(void) 619{ 620 if (do_aXp) 621 __asm__ __volatile__ ("xsnmsubadp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 622 else 623 __asm__ __volatile__ ("xsnmsubmdp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 624} 625 626static void test_xvmadd(void) 627{ 628 if (do_aXp) 629 if (do_dp) 630 __asm__ __volatile__ ("xvmaddadp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 631 else 632 __asm__ __volatile__ ("xvmaddasp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 633 else 634 if (do_dp) 635 __asm__ __volatile__ ("xvmaddmdp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 636 else 637 __asm__ __volatile__ ("xvmaddmsp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 638} 639 640static void test_xvnmadd(void) 641{ 642 if (do_aXp) 643 if (do_dp) 644 __asm__ __volatile__ ("xvnmaddadp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 645 else 646 __asm__ __volatile__ ("xvnmaddasp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 647 else 648 if (do_dp) 649 __asm__ __volatile__ ("xvnmaddmdp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 650 else 651 __asm__ __volatile__ ("xvnmaddmsp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 652} 653 654static void test_xvnmsub(void) 655{ 656 if (do_aXp) 657 if (do_dp) 658 __asm__ __volatile__ ("xvnmsubadp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 659 else 660 __asm__ __volatile__ ("xvnmsubasp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 661 else 662 if (do_dp) 663 __asm__ __volatile__ ("xvnmsubmdp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 664 else 665 __asm__ __volatile__ ("xvnmsubmsp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 666} 667 668static void test_xvmsub(void) 669{ 670 if (do_aXp) 671 if (do_dp) 672 __asm__ __volatile__ ("xvmsubadp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 673 else 674 __asm__ __volatile__ ("xvmsubasp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 675 else 676 if (do_dp) 677 __asm__ __volatile__ ("xvmsubmdp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 678 else 679 __asm__ __volatile__ ("xvmsubmsp %x0, %x1, %x2" : "+wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 680} 681 682static void test_xssqrtdp(void) 683{ 684 __asm__ __volatile__ ("xssqrtdp %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB)); 685} 686 687static void test_xsrdpim(void) 688{ 689 __asm__ __volatile__ ("xsrdpim %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB)); 690} 691 692static void test_xsrdpip(void) 693{ 694 __asm__ __volatile__ ("xsrdpip %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB)); 695} 696 697static void test_xstdivdp(void) 698{ 699 __asm__ __volatile__ ("xstdivdp 6, %x0, %x1" : : "wa" (vec_inA), "wa" (vec_inB)); 700} 701 702static void test_xsmaxdp(void) 703{ 704 __asm__ __volatile__ ("xsmaxdp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 705} 706 707static void test_xsmindp(void) 708{ 709 __asm__ __volatile__ ("xsmindp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 710} 711 712static void test_xvadddp(void) 713{ 714 __asm__ __volatile__ ("xvadddp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 715} 716 717static void test_xvaddsp(void) 718{ 719 __asm__ __volatile__ ("xvaddsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 720} 721 722static void test_xvdivdp(void) 723{ 724 __asm__ __volatile__ ("xvdivdp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 725} 726 727static void test_xvdivsp(void) 728{ 729 __asm__ __volatile__ ("xvdivsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 730} 731 732static void test_xvmuldp(void) 733{ 734 __asm__ __volatile__ ("xvmuldp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 735} 736 737static void test_xvmulsp(void) 738{ 739 __asm__ __volatile__ ("xvmulsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 740} 741 742static void test_xvsubdp(void) 743{ 744 __asm__ __volatile__ ("xvsubdp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 745} 746 747static void test_xvmaxdp(void) 748{ 749 __asm__ __volatile__ ("xvmaxdp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 750} 751 752static void test_xvmindp(void) 753{ 754 __asm__ __volatile__ ("xvmindp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 755} 756 757static void test_xvmaxsp(void) 758{ 759 __asm__ __volatile__ ("xvmaxsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 760} 761 762static void test_xvminsp(void) 763{ 764 __asm__ __volatile__ ("xvminsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 765} 766 767static void test_xvsubsp(void) 768{ 769 __asm__ __volatile__ ("xvsubsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 770} 771 772static void test_xvresp(void) 773{ 774 __asm__ __volatile__ ("xvresp %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB)); 775} 776 777static void test_xxsel(void) 778{ 779 unsigned long long * dst; 780 unsigned long long xa[] = { 0xa12bc37de56f9708ULL, 0x3894c1fddeadbeefULL}; 781 unsigned long long xb[] = { 0xfedc432124681235ULL, 0xf1e2d3c4e0057708ULL}; 782 unsigned long long xc[] = { 0xffffffff01020304ULL, 0x128934bd00000000ULL}; 783 784 memcpy(&vec_inA, xa, 16); 785 memcpy(&vec_inB, xb, 16); 786 memcpy(&vec_inC, xc, 16); 787 788 789 __asm__ __volatile__ ("xxsel %x0, %x1, %x2, %x3" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB),"wa" (vec_inC)); 790 dst = (unsigned long long *) &vec_out; 791 printf("xxsel %016llx,%016llx,%016llx => %016llx\n", xa[0], xb[0], xc[0], *dst); 792 dst++; 793 printf("xxsel %016llx,%016llx,%016llx => %016llx\n", xa[1], xb[1], xc[1], *dst); 794 printf("\n"); 795} 796 797static void test_xxspltw(void) 798{ 799 int uim; 800 unsigned long long * dst = NULL; 801 unsigned long long xb[] = { 0xfedc432124681235ULL, 0xf1e2d3c4e0057708ULL}; 802 memcpy(&vec_inB, xb, 16); 803 804 for (uim = 0; uim < 4; uim++) { 805 switch (uim) { 806 case 0: 807 __asm__ __volatile__ ("xxspltw %x0, %x1, 0" : "=wa" (vec_out): "wa" (vec_inB)); 808 break; 809 case 1: 810 __asm__ __volatile__ ("xxspltw %x0, %x1, 1" : "=wa" (vec_out): "wa" (vec_inB)); 811 break; 812 case 2: 813 __asm__ __volatile__ ("xxspltw %x0, %x1, 2" : "=wa" (vec_out): "wa" (vec_inB)); 814 break; 815 case 3: 816 __asm__ __volatile__ ("xxspltw %x0, %x1, 3" : "=wa" (vec_out): "wa" (vec_inB)); 817 break; 818 } 819 dst = (unsigned long long *) &vec_out; 820 printf("xxspltw 0x%016llx%016llx %d=> 0x%016llx", xb[0], xb[1], uim, *dst); 821 dst++; 822 printf("%016llx\n", *dst); 823 } 824 printf("\n"); 825} 826 827static void test_xscvdpsxws(void) 828{ 829 __asm__ __volatile__ ("xscvdpsxws %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB)); 830} 831 832static void test_xscvdpuxds(void) 833{ 834 __asm__ __volatile__ ("xscvdpuxds %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB)); 835} 836 837static void test_xvcpsgndp(void) 838{ 839 __asm__ __volatile__ ("xvcpsgndp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 840} 841 842static void test_xvcpsgnsp(void) 843{ 844 __asm__ __volatile__ ("xvcpsgnsp %x0, %x1, %x2" : "=wa" (vec_out): "wa" (vec_inA),"wa" (vec_inB)); 845} 846 847static void test_xvcvdpsxws(void) 848{ 849 __asm__ __volatile__ ("xvcvdpsxws %x0, %x1 " : "=wa" (vec_out): "wa" (vec_inB)); 850} 851 852static void test_xvcvspsxws(void) 853{ 854 __asm__ __volatile__ ("xvcvspsxws %x0, %x1 " : "=wa" (vec_out): "wa" (vec_inB)); 855} 856 857static vx_fp_test_t 858vx_vector_one_fp_arg_tests[] = { 859 { &test_xvresp, "xvresp", NULL, 16, SINGLE_TEST, VX_BASIC_CMP, "1/x"}, 860 { &test_xvcvdpsxws, "xvcvdpsxws", NULL, 16, DOUBLE_TEST, VX_CONV_WORD, "conv"}, 861 { &test_xvcvspsxws, "xvcvspsxws", NULL, 16, SINGLE_TEST, VX_CONV_WORD, "conv"}, 862 { NULL, NULL, NULL, 0 , 0, 0, NULL} 863}; 864 865static vx_fp_test_t 866vx_vector_fp_tests[] = { 867 { &test_xvcmpeqdp, "xvcmpeqdp", fp_cmp_tests, 64, DOUBLE_TEST, VX_BASIC_CMP, "eq"}, 868 { &test_xvcmpgedp, "xvcmpgedp", fp_cmp_tests, 64, DOUBLE_TEST, VX_BASIC_CMP, "ge"}, 869 { &test_xvcmpgtdp, "xvcmpgtdp", fp_cmp_tests, 64, DOUBLE_TEST, VX_BASIC_CMP, "gt"}, 870 { &test_xvcmpeqsp, "xvcmpeqsp", fp_cmp_tests, 64, SINGLE_TEST, VX_BASIC_CMP, "eq"}, 871 { &test_xvcmpgesp, "xvcmpgesp", fp_cmp_tests, 64, SINGLE_TEST, VX_BASIC_CMP, "ge"}, 872 { &test_xvcmpgtsp, "xvcmpgtsp", fp_cmp_tests, 64, SINGLE_TEST, VX_BASIC_CMP, "gt"}, 873 { &test_xvadddp, "xvadddp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "+" }, 874 { &test_xvaddsp, "xvaddsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "+" }, 875 { &test_xvdivdp, "xvdivdp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "/" }, 876 { &test_xvdivsp, "xvdivsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "/" }, 877 { &test_xvmuldp, "xvmuldp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "*" }, 878 { &test_xvmulsp, "xvmulsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "*" }, 879 { &test_xvsubdp, "xvsubdp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "-" }, 880 { &test_xvsubsp, "xvsubsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "-" }, 881 { &test_xvmaxdp, "xvmaxdp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "@max@" }, 882 { &test_xvmindp, "xvmindp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "@min@" }, 883 { &test_xvmaxsp, "xvmaxsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "@max@" }, 884 { &test_xvminsp, "xvminsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "@min@" }, 885 { &test_xvcpsgndp, "xvcpsgndp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, "+-cp"}, 886 { &test_xvcpsgnsp, "xvcpsgnsp", two_arg_fp_tests, 64, SINGLE_TEST, VX_DEFAULT, "+-cp"}, 887 { NULL, NULL, NULL, 0 , 0, 0, NULL} 888}; 889 890 891static vx_fp_test_t 892vx_aORm_fp_tests[] = { 893 { &test_xsnmsub, "xsnmsub", two_arg_fp_tests, 64, DOUBLE_TEST, VX_SCALAR_FP_NMSUB, "!*-"}, 894 { &test_xvmadd, "xvmadd", two_arg_fp_tests, 64, DOUBLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "*+"}, 895 { &test_xvmadd, "xvmadd", two_arg_fp_tests, 64, SINGLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "*+"}, 896 { &test_xvnmadd, "xvnmadd", two_arg_fp_tests, 64, DOUBLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "!*+"}, 897 { &test_xvnmadd, "xvnmadd", two_arg_fp_tests, 64, SINGLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "!*+"}, 898 { &test_xvmsub, "xvmsub", two_arg_fp_tests, 64, DOUBLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "*-"}, 899 { &test_xvmsub, "xvmsub", two_arg_fp_tests, 64, SINGLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "*-"}, 900 { &test_xvnmsub, "xvnmsub", two_arg_fp_tests, 64, DOUBLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "!*-"}, 901 { &test_xvnmsub, "xvnmsub", two_arg_fp_tests, 64, SINGLE_TEST, VX_VECTOR_FP_MULT_AND_OP2, "!*-"}, 902 { NULL, NULL, NULL, 0, 0, 0, NULL } 903}; 904 905static vx_fp_test_t 906vx_simple_scalar_fp_tests[] = { 907 { &test_xssqrtdp, "xssqrtdp", NULL, 17, DOUBLE_TEST, VX_DEFAULT, NULL}, 908 { &test_xsrdpim, "xsrdpim", NULL, 17, DOUBLE_TEST, VX_DEFAULT, NULL}, 909 { &test_xsrdpip, "xsrdpip", NULL, 17, DOUBLE_TEST, VX_DEFAULT, NULL}, 910 { &test_xstdivdp, "xstdivdp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, NULL}, 911 { &test_xsmaxdp, "xsmaxdp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, NULL}, 912 { &test_xsmindp, "xsmindp", two_arg_fp_tests, 64, DOUBLE_TEST, VX_DEFAULT, NULL}, 913 { &test_xscvdpsxws, "xscvdpsxws", NULL, 17, DOUBLE_TEST, VX_CONV_WORD, NULL}, 914 { &test_xscvdpuxds, "xscvdpuxds", NULL, 17, DOUBLE_TEST, VX_DEFAULT, NULL}, 915 { NULL, NULL, NULL, 0, 0, 0, NULL } 916}; 917 918 919#ifdef __powerpc64__ 920static void test_bpermd(void) 921{ 922 /* NOTE: Bit number is '0 . . . 63' 923 * 924 * Permuted bits are generated bit 0 -7 as follows: 925 * index = (r14)8*i:8*i+7 926 * perm[i] = (r15)index 927 * 928 * So, for i = 0, index is (r14)8*0:8*0+7, or (r14)0:7, which is the MSB 929 * byte of r14, 0x1b(27/base 10). This identifies bit 27 of r15, which is '1'. 930 * For i = 1, index is 0x2c, identifying bit 44 of r15, which is '1'. 931 * So the result of the first two iterations of i are: 932 * perm = 0b01xxxxxx 933 * 934 */ 935 r15 = 0xa12bc37de56f9708ULL; 936 r14 = 0x1b2c31f030000001ULL; 937 __asm__ __volatile__ ("bpermd %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 938 printf("bpermd: 0x%016llx : 0x%016llx => 0x%llx\n", (unsigned long long)r14, 939 (unsigned long long)r15, (unsigned long long)r17); 940 printf("\n"); 941} 942#endif 943 944static Bool do_OE; 945typedef enum { 946 DIV_BASE = 1, 947 DIV_OE = 2, 948 DIV_DOT = 4, 949} div_type_t; 950/* Possible divde type combinations are: 951 * - base 952 * - base+dot 953 * - base+OE 954 * - base+OE+dot 955 */ 956#ifdef __powerpc64__ 957static void test_divde(void) 958{ 959 int divde_type = DIV_BASE; 960 if (do_OE) 961 divde_type |= DIV_OE; 962 if (do_dot) 963 divde_type |= DIV_DOT; 964 965 switch (divde_type) { 966 case 1: 967 SET_CR_XER_ZERO; 968 __asm__ __volatile__ ("divde %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 969 GET_CR_XER(div_flags, div_xer); 970 break; 971 case 3: 972 SET_CR_XER_ZERO; 973 __asm__ __volatile__ ("divdeo %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 974 GET_CR_XER(div_flags, div_xer); 975 break; 976 case 5: 977 SET_CR_XER_ZERO; 978 __asm__ __volatile__ ("divde. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 979 GET_CR_XER(div_flags, div_xer); 980 break; 981 case 7: 982 SET_CR_XER_ZERO; 983 __asm__ __volatile__ ("divdeo. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 984 GET_CR_XER(div_flags, div_xer); 985 break; 986 default: 987 fprintf(stderr, "Invalid divde type. Exiting\n"); 988 exit(1); 989 } 990} 991#endif 992 993static void test_divweu(void) 994{ 995 int divweu_type = DIV_BASE; 996 if (do_OE) 997 divweu_type |= DIV_OE; 998 if (do_dot) 999 divweu_type |= DIV_DOT; 1000 1001 switch (divweu_type) { 1002 case 1: 1003 SET_CR_XER_ZERO; 1004 __asm__ __volatile__ ("divweu %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 1005 GET_CR_XER(div_flags, div_xer); 1006 break; 1007 case 3: 1008 SET_CR_XER_ZERO; 1009 __asm__ __volatile__ ("divweuo %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 1010 GET_CR_XER(div_flags, div_xer); 1011 break; 1012 case 5: 1013 SET_CR_XER_ZERO; 1014 __asm__ __volatile__ ("divweu. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 1015 GET_CR_XER(div_flags, div_xer); 1016 break; 1017 case 7: 1018 SET_CR_XER_ZERO; 1019 __asm__ __volatile__ ("divweuo. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15)); 1020 GET_CR_XER(div_flags, div_xer); 1021 break; 1022 default: 1023 fprintf(stderr, "Invalid divweu type. Exiting\n"); 1024 exit(1); 1025 } 1026} 1027 1028static void test_fctiduz(void) 1029{ 1030 if (do_dot) 1031 __asm__ __volatile__ ("fctiduz. %0, %1" : "=d" (f17) : "d" (f14)); 1032 else 1033 __asm__ __volatile__ ("fctiduz %0, %1" : "=d" (f17) : "d" (f14)); 1034} 1035 1036static void test_fctidu(void) 1037{ 1038 if (do_dot) 1039 __asm__ __volatile__ ("fctidu. %0, %1" : "=d" (f17) : "d" (f14)); 1040 else 1041 __asm__ __volatile__ ("fctidu %0, %1" : "=d" (f17) : "d" (f14)); 1042} 1043 1044static void test_fctiwuz(void) 1045{ 1046 if (do_dot) 1047 __asm__ __volatile__ ("fctiwuz. %0, %1" : "=d" (f17) : "d" (f14)); 1048 else 1049 __asm__ __volatile__ ("fctiwuz %0, %1" : "=d" (f17) : "d" (f14)); 1050} 1051 1052static void test_fctiwu(void) 1053{ 1054 if (do_dot) 1055 __asm__ __volatile__ ("fctiwu. %0, %1" : "=d" (f17) : "d" (f14)); 1056 else 1057 __asm__ __volatile__ ("fctiwu %0, %1" : "=d" (f17) : "d" (f14)); 1058} 1059 1060typedef struct simple_test { 1061 test_func_t test_func; 1062 char * name; 1063 precision_type_t precision; 1064} simple_test_t; 1065 1066static simple_test_t fct_tests[] = { 1067 { &test_fctiduz, "fctiduz", DOUBLE_TEST }, 1068 { &test_fctidu, "fctidu", DOUBLE_TEST }, 1069 { &test_fctiwuz, "fctiwuz", SINGLE_TEST }, 1070 { &test_fctiwu, "fctiwu", SINGLE_TEST }, 1071 { NULL, NULL } 1072}; 1073 1074static void setup_sp_fp_args(fp_test_args_t * targs, Bool swap_inputs) 1075{ 1076 int a_idx, b_idx, i; 1077 void * inA, * inB; 1078 void * vec_src = swap_inputs ? &vec_out : &vec_inB; 1079 1080 for (i = 0; i < 4; i++) { 1081 a_idx = targs->fra_idx; 1082 b_idx = targs->frb_idx; 1083 inA = (void *)&spec_sp_fargs[a_idx]; 1084 inB = (void *)&spec_sp_fargs[b_idx]; 1085 // copy single precision FP into vector element i 1086 memcpy(((void *)&vec_inA) + (i * 4), inA, 4); 1087 memcpy(vec_src + (i * 4), inB, 4); 1088 targs++; 1089 } 1090} 1091 1092static void setup_dp_fp_args(fp_test_args_t * targs, Bool swap_inputs) 1093{ 1094 int a_idx, b_idx, i; 1095 void * inA, * inB; 1096 void * vec_src = swap_inputs ? (void *)&vec_out : (void *)&vec_inB; 1097 1098 for (i = 0; i < 2; i++) { 1099 a_idx = targs->fra_idx; 1100 b_idx = targs->frb_idx; 1101 inA = (void *)&spec_fargs[a_idx]; 1102 inB = (void *)&spec_fargs[b_idx]; 1103 // copy double precision FP into vector element i 1104 memcpy(((void *)&vec_inA) + (i * 8), inA, 8); 1105 memcpy(vec_src + (i * 8), inB, 8); 1106 targs++; 1107 } 1108} 1109 1110#define VX_NOT_CMP_OP 0xffffffff 1111static void print_vector_fp_result(unsigned int cc, vx_fp_test_t * test_group, int i) 1112{ 1113 int a_idx, b_idx, k; 1114 char * name = malloc(20); 1115 int dp = test_group->precision == DOUBLE_TEST ? 1 : 0; 1116 int loops = dp ? 2 : 4; 1117 fp_test_args_t * targs = &test_group->targs[i]; 1118 unsigned long long * frA_dp, * frB_dp, * dst_dp; 1119 unsigned int * frA_sp, *frB_sp, * dst_sp; 1120 strcpy(name, test_group->name); 1121 printf("#%d: %s%s ", dp? i/2 : i/4, name, (do_dot ? "." : "")); 1122 for (k = 0; k < loops; k++) { 1123 a_idx = targs->fra_idx; 1124 b_idx = targs->frb_idx; 1125 if (k) 1126 printf(" AND "); 1127 if (dp) { 1128 frA_dp = (unsigned long long *)&spec_fargs[a_idx]; 1129 frB_dp = (unsigned long long *)&spec_fargs[b_idx]; 1130 printf("%016llx %s %016llx", *frA_dp, test_group->op, *frB_dp); 1131 } else { 1132 frA_sp = (unsigned int *)&spec_sp_fargs[a_idx]; 1133 frB_sp = (unsigned int *)&spec_sp_fargs[b_idx]; 1134 printf("%08x %s %08x", *frA_sp, test_group->op, *frB_sp); 1135 } 1136 targs++; 1137 } 1138 if (cc != VX_NOT_CMP_OP) 1139 printf(" ? cc=%x", cc); 1140 1141 if (dp) { 1142 dst_dp = (unsigned long long *) &vec_out; 1143 printf(" => %016llx %016llx\n", dst_dp[0], dst_dp[1]); 1144 } else { 1145 dst_sp = (unsigned int *) &vec_out; 1146 printf(" => %08x %08x %08x %08x\n", dst_sp[0], dst_sp[1], dst_sp[2], dst_sp[3]); 1147 } 1148 free(name); 1149} 1150 1151 1152static void print_vx_aORm_fp_result(unsigned long long * XT_arg, unsigned long long * XB_arg, 1153 vx_fp_test_t * test_group, int i) 1154{ 1155 int a_idx, k; 1156 char * name = malloc(20); 1157 int dp = test_group->precision == DOUBLE_TEST ? 1 : 0; 1158 int loops = dp ? 2 : 4; 1159 fp_test_args_t * targs = &test_group->targs[i]; 1160 unsigned long long frA_dp, * dst_dp; 1161 unsigned int frA_sp, * dst_sp; 1162 1163 strcpy(name, test_group->name); 1164 if (do_aXp) 1165 if (dp) 1166 strcat(name, "adp"); 1167 else 1168 strcat(name, "asp"); 1169 else 1170 if (dp) 1171 strcat(name, "mdp"); 1172 else 1173 strcat(name, "msp"); 1174 1175 printf("#%d: %s ", dp? i/2 : i/4, name); 1176 for (k = 0; k < loops; k++) { 1177 a_idx = targs->fra_idx; 1178 if (k) 1179 printf(" AND "); 1180 if (dp) { 1181 frA_dp = *((unsigned long long *)&spec_fargs[a_idx]); 1182 printf("%s(%016llx,%016llx,%016llx)", test_group->op, XT_arg[k], frA_dp, XB_arg[k]); 1183 } else { 1184 unsigned int * xt_sp = (unsigned int *)XT_arg; 1185 unsigned int * xb_sp = (unsigned int *)XB_arg; 1186 frA_sp = *((unsigned int *)&spec_sp_fargs[a_idx]); 1187 printf("%s(%08x,%08x,%08x)", test_group->op, xt_sp[k], frA_sp, xb_sp[k]); 1188 } 1189 targs++; 1190 } 1191 1192 if (dp) { 1193 dst_dp = (unsigned long long *) &vec_out; 1194 printf(" => %016llx %016llx\n", dst_dp[0], dst_dp[1]); 1195 } else { 1196 dst_sp = (unsigned int *) &vec_out; 1197 printf(" => %08x %08x %08x %08x\n", dst_sp[0], dst_sp[1], dst_sp[2], dst_sp[3]); 1198 } 1199 free(name); 1200} 1201 1202/* This function currently only supports double precision input arguments. */ 1203static void test_vx_simple_scalar_fp_ops(void) 1204{ 1205 test_func_t func; 1206 int k = 0; 1207 1208 build_special_fargs_table(); 1209 while ((func = vx_simple_scalar_fp_tests[k].test_func)) { 1210 unsigned long long * frap, * frbp, * dst; 1211 unsigned int * pv; 1212 int idx; 1213 vx_fp_test_t test_group = vx_simple_scalar_fp_tests[k]; 1214 Bool convToWord = (test_group.type == VX_CONV_WORD); 1215 if (test_group.precision != DOUBLE_TEST) { 1216 fprintf(stderr, "Unsupported single precision for scalar op in test_vx_aORm_fp_ops\n"); 1217 exit(1); 1218 } 1219 pv = (unsigned int *)&vec_out; 1220 // clear vec_out 1221 for (idx = 0; idx < 4; idx++, pv++) 1222 *pv = 0; 1223 1224 /* If num_tests is exactly equal to nb_special_fargs, this implies the 1225 * instruction being tested only requires one floating point argument 1226 * (e.g. xssqrtdp). 1227 */ 1228 if (test_group.num_tests == nb_special_fargs && !test_group.targs) { 1229 void * inB; 1230 int i; 1231 for (i = 0; i < nb_special_fargs; i++) { 1232 inB = (void *)&spec_fargs[i]; 1233 frbp = (unsigned long long *)&spec_fargs[i]; 1234 memcpy(&vec_inB, inB, 8); 1235 (*func)(); 1236 dst = (unsigned long long *) &vec_out; 1237 printf("#%d: %s %016llx => %016llx\n", i, test_group.name, *frbp, 1238 convToWord ? (*dst & 0x00000000ffffffffULL) : *dst); 1239 } 1240 } else { 1241 void * inA, * inB; 1242 unsigned int condreg, flags; 1243 int isTdiv = (strstr(test_group.name, "xstdivdp") != NULL) ? 1 : 0; 1244 int i; 1245 for (i = 0; i < test_group.num_tests; i++) { 1246 fp_test_args_t aTest = test_group.targs[i]; 1247 inA = (void *)&spec_fargs[aTest.fra_idx]; 1248 inB = (void *)&spec_fargs[aTest.frb_idx]; 1249 frap = (unsigned long long *)&spec_fargs[aTest.fra_idx]; 1250 frbp = (unsigned long long *)&spec_fargs[aTest.frb_idx]; 1251 // Only need to copy one doubleword into each vector's element 0 1252 memcpy(&vec_inA, inA, 8); 1253 memcpy(&vec_inB, inB, 8); 1254 SET_FPSCR_ZERO; 1255 SET_CR_XER_ZERO; 1256 (*func)(); 1257 GET_CR(flags); 1258 if (isTdiv) { 1259 condreg = (flags & 0x000000f0) >> 4; 1260 printf("#%d: %s %016llx,%016llx => cr %x\n", i, test_group.name, *frap, *frbp, condreg); 1261 } else { 1262 dst = (unsigned long long *) &vec_out; 1263 printf("#%d: %s %016llx,%016llx => %016llx\n", i, test_group.name, 1264 *frap, *frbp, *dst); 1265 } 1266 } 1267 } 1268 printf( "\n" ); 1269 k++; 1270 } 1271} 1272 1273static void test_vx_aORm_fp_ops(void) 1274{ 1275 /* These ops need a third src argument, which is stored in element 0 of 1276 * VSX[XT] -- i.e., vec_out. For the xs<ZZZ>m{d|s}p cases, VSX[XT] holds 1277 * src3 and VSX[XB] holds src2; for the xs<ZZZ>a{d|s}p cases, VSX[XT] holds 1278 * src2 and VSX[XB] holds src3. The fp_test_args_t that holds the test 1279 * data (input args, result) contain only two inputs, so I arbitrarily 1280 * choose some spec_fargs elements for the third source argument. 1281 * Note that that by using the same input data for a given pair of 1282 * a{d|s}p/m{d|s}p-type instructions (by swapping the src2 and src3 1283 * arguments), the expected result should be the same. 1284 */ 1285 1286 test_func_t func; 1287 int k; 1288 char * test_name = (char *)malloc(20); 1289 k = 0; 1290 do_dot = False; 1291 1292 build_special_fargs_table(); 1293 while ((func = vx_aORm_fp_tests[k].test_func)) { 1294 int i, stride; 1295 Bool repeat = False; 1296 Bool scalar = False; 1297 unsigned long long * frap, * frbp, * dst; 1298 vx_fp_test_t test_group = vx_aORm_fp_tests[k]; 1299 vx_fp_test_type test_type = test_group.type; 1300 do_dp = test_group.precision == DOUBLE_TEST ? True : False; 1301 frap = frbp = NULL; 1302 1303 if (test_type < VX_VECTOR_FP_MULT_AND_OP2) { 1304 scalar = True; 1305 strcpy(test_name, test_group.name); 1306 if (!repeat) { 1307 repeat = 1; 1308 stride = 1; 1309 // Only support double precision scalar ops in this function 1310 if (do_dp) { 1311 strcat(test_name, "adp"); 1312 } else { 1313 fprintf(stderr, "Unsupported single precision for scalar op in test_vx_aORm_fp_ops\n"); 1314 exit(1); 1315 } 1316 do_aXp = True; 1317 } 1318 } else if (test_type < VX_BASIC_CMP) { 1319 // Then it must be a VX_VECTOR_xxx type 1320 stride = do_dp ? 2 : 4; 1321 if (!repeat) { 1322 // No need to work up the testcase name here, since that will be done in 1323 // the print_vx_aORm_fp_result() function we'll call for vector-type ops. 1324 repeat = 1; 1325 do_aXp = True; 1326 } 1327 } else { 1328 printf("ERROR: Invalid VX FP test type %d\n", test_type); 1329 exit(1); 1330 } 1331 1332again: 1333 for (i = 0; i < test_group.num_tests; i+=stride) { 1334 void * inA, * inB; 1335 int m, fp_idx[4]; 1336 unsigned long long vsr_XT[2]; 1337 unsigned long long vsr_XB[2]; 1338 fp_test_args_t aTest = test_group.targs[i]; 1339 for (m = 0; m < stride; m++) 1340 fp_idx[m] = i % (nb_special_fargs - stride) + m; 1341 1342 /* When repeat == True, we're on the first time through of one of the VX_FP_SMx 1343 * test types, meaning we're testing a xs<ZZZ>adp case, thus we have to swap 1344 * inputs as described above: 1345 * src2 <= VSX[XT] 1346 * src3 <= VSX[XB] 1347 */ 1348 if (scalar) { 1349 // For scalar op, only need to copy one doubleword into each vector's element 0 1350 inA = (void *)&spec_fargs[aTest.fra_idx]; 1351 inB = (void *)&spec_fargs[aTest.frb_idx]; 1352 frap = (unsigned long long *)&spec_fargs[aTest.fra_idx]; 1353 memcpy(&vec_inA, inA, 8); 1354 if (repeat) { 1355 memcpy(&vec_out, inB, 8); // src2 1356 memcpy(&vec_inB, &spec_fargs[fp_idx[0]], 8); //src3 1357 frbp = (unsigned long long *)&spec_fargs[fp_idx[0]]; 1358 } else { 1359 frbp = (unsigned long long *)&spec_fargs[aTest.frb_idx]; 1360 memcpy(&vec_inB, inB, 8); // src2 1361 memcpy(&vec_out, &spec_fargs[fp_idx[0]], 8); //src3 1362 } 1363 memcpy(vsr_XT, &vec_out, 8); 1364 } else { 1365 int j, loops = do_dp ? 2 : 4; 1366 size_t len = do_dp ? 8 : 4; 1367 void * vec_src = repeat ? (void *)&vec_inB : (void *)&vec_out; 1368 for (j = 0; j < loops; j++) { 1369 if (do_dp) 1370 memcpy(vec_src + (j * len), &spec_fargs[fp_idx[j]], len); 1371 else 1372 memcpy(vec_src + (j * len), &spec_sp_fargs[fp_idx[j]], len); 1373 } 1374 if (do_dp) 1375 setup_dp_fp_args(&test_group.targs[i], repeat); 1376 else 1377 setup_sp_fp_args(&test_group.targs[i], repeat); 1378 1379 memcpy(vsr_XT, &vec_out, 16); 1380 memcpy(vsr_XB, &vec_inB, 16); 1381 } 1382 1383 (*func)(); 1384 dst = (unsigned long long *) &vec_out; 1385 if (test_type < VX_VECTOR_FP_MULT_AND_OP2) 1386 printf( "#%d: %s %s(%016llx,%016llx,%016llx) = %016llx\n", i, 1387 test_name, test_group.op, vsr_XT[0], *frap, *frbp, *dst ); 1388 else 1389 print_vx_aORm_fp_result(vsr_XT, vsr_XB, &test_group, i); 1390 } 1391 printf( "\n" ); 1392 1393 if (repeat) { 1394 repeat = 0; 1395 if (test_type < VX_VECTOR_FP_MULT_AND_OP2) { 1396 strcpy(test_name, test_group.name); 1397 strcat(test_name, "mdp"); 1398 } 1399 do_aXp = False; 1400 goto again; 1401 } 1402 k++; 1403 } 1404 printf( "\n" ); 1405 free(test_name); 1406} 1407 1408static void test_vx_vector_one_fp_arg(void) 1409{ 1410 test_func_t func; 1411 int k; 1412 k = 0; 1413 build_special_fargs_table(); 1414 1415 while ((func = vx_vector_one_fp_arg_tests[k].test_func)) { 1416 int idx, i; 1417 vx_fp_test_t test_group = vx_vector_one_fp_arg_tests[k]; 1418 Bool convToWord = (test_group.type == VX_CONV_WORD); 1419 Bool dp = (test_group.precision == DOUBLE_TEST) ? True : False; 1420 Bool xvrespTest = (strstr(test_group.name , "xvresp") != NULL) ? True: False; 1421 int stride = dp ? 2 : 4; 1422 1423 for (i = 0; i < test_group.num_tests; i+=stride) { 1424 unsigned int * pv; 1425 void * inB; 1426 1427 pv = (unsigned int *)&vec_out; 1428 // clear vec_out 1429 for (idx = 0; idx < 4; idx++, pv++) 1430 *pv = 0; 1431 1432 if (dp) { 1433 int j; 1434 unsigned long long * frB_dp, *dst_dp; 1435 for (j = 0; j < 2; j++) { 1436 inB = (void *)&spec_fargs[i + j]; 1437 // copy double precision FP into vector element i 1438 memcpy(((void *)&vec_inB) + (j * 8), inB, 8); 1439 } 1440 // execute test insn 1441 (*func)(); 1442 dst_dp = (unsigned long long *) &vec_out; 1443 printf("#%d: %s ", i/2, test_group.name); 1444 for (j = 0; j < 2; j++) { 1445 if (j) 1446 printf("; "); 1447 frB_dp = (unsigned long long *)&spec_fargs[i + j]; 1448 printf("%s(%016llx)", test_group.op, *frB_dp); 1449 printf(" = %016llx", convToWord ? (dst_dp[j] & 0x00000000ffffffffULL) : dst_dp[j]); 1450 } 1451 printf("\n"); 1452 } else { 1453 int j; 1454 unsigned int * frB_sp, * dst_sp; 1455 1456 for (j = 0; j < 4; j++) { 1457 inB = (void *)&spec_sp_fargs[i + j]; 1458 // copy single precision FP into vector element i 1459 memcpy(((void *)&vec_inB) + (j * 4), inB, 4); 1460 } 1461 // execute test insn 1462 (*func)(); 1463 dst_sp = (unsigned int *) &vec_out; 1464 // print result 1465 printf("#%d: %s ", i/4, test_group.name); 1466 for (j = 0; j < 4; j++) { 1467 if (j) 1468 printf("; "); 1469 frB_sp = (unsigned int *)&spec_sp_fargs[i + j]; 1470 printf("%s(%08x)", test_group.op, *frB_sp); 1471 if (xvrespTest) { 1472 float calc_diff = fabs(spec_sp_fargs[i + j]/256); 1473 float sp_res; 1474 memcpy(&sp_res, &dst_sp[j], 4); 1475 float div_result = 1/spec_sp_fargs[i + j]; 1476 float real_diff = fabs(sp_res - div_result); 1477 printf( " ==> %s", 1478 ( ( sp_res == div_result ) 1479 || ( isnan(sp_res) && isnan(div_result) ) 1480 || ( real_diff <= calc_diff ) ) ? "PASS" 1481 : "FAIL"); 1482 } else { 1483 printf(" = %08x", dst_sp[j]); 1484 } 1485 } 1486 printf("\n"); 1487 } 1488 } 1489 k++; 1490 printf( "\n" ); 1491 } 1492 1493} 1494 1495/* This function assumes the instruction being tested requires two args. */ 1496static void test_vx_vector_fp_ops(void) 1497{ 1498 test_func_t func; 1499 int k; 1500 k = 0; 1501 build_special_fargs_table(); 1502 1503 while ((func = vx_vector_fp_tests[k].test_func)) { 1504 int idx, i, repeat = 1; 1505 vx_fp_test_t test_group = vx_vector_fp_tests[k]; 1506 int stride = test_group.precision == DOUBLE_TEST ? 2 : 4; 1507 do_dot = False; 1508 1509again: 1510 for (i = 0; i < test_group.num_tests; i+=stride) { 1511 unsigned int * pv, condreg; 1512 unsigned int flags; 1513 1514 pv = (unsigned int *)&vec_out; 1515 if (test_group.precision == DOUBLE_TEST) 1516 setup_dp_fp_args(&test_group.targs[i], False); 1517 else 1518 setup_sp_fp_args(&test_group.targs[i], False); 1519 1520 // clear vec_out 1521 for (idx = 0; idx < 4; idx++, pv++) 1522 *pv = 0; 1523 1524 // execute test insn 1525 SET_FPSCR_ZERO; 1526 SET_CR_XER_ZERO; 1527 (*func)(); 1528 GET_CR(flags); 1529 if (test_group.type == VX_BASIC_CMP) { 1530 condreg = (flags & 0x000000f0) >> 4; 1531 } else { 1532 condreg = VX_NOT_CMP_OP; 1533 } 1534 print_vector_fp_result(condreg, &test_group, i); 1535 } 1536 printf("\n"); 1537 if (repeat && test_group.type == VX_BASIC_CMP) { 1538 repeat = 0; 1539 do_dot = True; 1540 goto again; 1541 } 1542 k++; 1543 printf( "\n" ); 1544 } 1545} 1546 1547 1548// The div doubleword test data 1549signed long long div_dw_tdata[13][2] = { 1550 { 4, -4 }, 1551 { 4, -3 }, 1552 { 4, 4 }, 1553 { 4, -5 }, 1554 { 3, 8 }, 1555 { 0x8000000000000000ULL, 0xa }, 1556 { 0x50c, -1 }, 1557 { 0x50c, -4096 }, 1558 { 0x1234fedc, 0x8000a873 }, 1559 { 0xabcd87651234fedcULL, 0xa123b893 }, 1560 { 0x123456789abdcULL, 0 }, 1561 { 0, 2 }, 1562 { 0x77, 0xa3499 } 1563}; 1564#define dw_tdata_len (sizeof(div_dw_tdata)/sizeof(signed long long)/2) 1565 1566// The div word test data 1567unsigned int div_w_tdata[6][2] = { 1568 { 0, 2 }, 1569 { 2, 0 }, 1570 { 0x7abc1234, 0xf0000000 }, 1571 { 0xfabc1234, 5 }, 1572 { 77, 66 }, 1573 { 5, 0xfabc1234 }, 1574}; 1575#define w_tdata_len (sizeof(div_w_tdata)/sizeof(unsigned int)/2) 1576 1577typedef struct div_ext_test 1578{ 1579 test_func_t test_func; 1580 const char *name; 1581 int num_tests; 1582 div_type_t div_type; 1583 precision_type_t precision; 1584} div_ext_test_t; 1585 1586static div_ext_test_t div_tests[] = { 1587#ifdef __powerpc64__ 1588 { &test_divde, "divde", dw_tdata_len, DIV_BASE, DOUBLE_TEST }, 1589 { &test_divde, "divdeo", dw_tdata_len, DIV_OE, DOUBLE_TEST }, 1590#endif 1591 { &test_divweu, "divweu", w_tdata_len, DIV_BASE, SINGLE_TEST }, 1592 { &test_divweu, "divweuo", w_tdata_len, DIV_OE, SINGLE_TEST }, 1593 { NULL, NULL, 0, 0, 0 } 1594}; 1595 1596static void test_div_extensions(void) 1597{ 1598 test_func_t func; 1599 int k; 1600 k = 0; 1601 1602 while ((func = div_tests[k].test_func)) { 1603 int i, repeat = 1; 1604 div_ext_test_t test_group = div_tests[k]; 1605 do_dot = False; 1606 1607again: 1608 for (i = 0; i < test_group.num_tests; i++) { 1609 unsigned int condreg; 1610 1611 if (test_group.div_type == DIV_OE) 1612 do_OE = True; 1613 else 1614 do_OE = False; 1615 1616 if (test_group.precision == DOUBLE_TEST) { 1617 r14 = div_dw_tdata[i][0]; 1618 r15 = div_dw_tdata[i][1]; 1619 } else { 1620 r14 = div_w_tdata[i][0]; 1621 r15 = div_w_tdata[i][1]; 1622 } 1623 // execute test insn 1624 (*func)(); 1625 condreg = (div_flags & 0xf0000000) >> 28; 1626 printf("#%d: %s%s: ", i, test_group.name, do_dot ? "." : ""); 1627 if (test_group.precision == DOUBLE_TEST) { 1628 printf("0x%016llx / 0x%016llx = 0x%016llx;", 1629 div_dw_tdata[i][0], div_dw_tdata[i][1], (signed long long) r17); 1630 } else { 1631 printf("0x%08x / 0x%08x = 0x%08x;", 1632 div_w_tdata[i][0], div_w_tdata[i][1], (unsigned int) r17); 1633 } 1634 printf(" CR=%x; XER=%x\n", condreg, div_xer); 1635 } 1636 printf("\n"); 1637 if (repeat) { 1638 repeat = 0; 1639 do_dot = True; 1640 goto again; 1641 } 1642 k++; 1643 printf( "\n" ); 1644 } 1645 1646} 1647 1648static void test_fct_ops(void) 1649{ 1650 test_func_t func; 1651 int k; 1652 k = 0; 1653 1654 while ((func = fct_tests[k].test_func)) { 1655 int i, repeat = 1; 1656 simple_test_t test_group = fct_tests[k]; 1657 do_dot = False; 1658 1659again: 1660 for (i = 0; i < nb_special_fargs; i++) { 1661 double result; 1662#define SINGLE_MASK 0x00000000FFFFFFFFULL 1663 1664 f14 = spec_fargs[i]; 1665 // execute test insn 1666 SET_FPSCR_ZERO; 1667 (*func)(); 1668 result = f17; 1669 printf("#%d: %s%s: ", i, test_group.name, do_dot ? "." : ""); 1670 printf("0x%016llx (%e) ==> 0x%016llx\n", 1671 *((unsigned long long *)(&spec_fargs[i])), spec_fargs[i], 1672 test_group.precision == SINGLE_TEST ? (SINGLE_MASK & 1673 *((unsigned long long *)(&result))) : 1674 *((unsigned long long *)(&result))); 1675 } 1676 printf("\n"); 1677 if (repeat) { 1678 repeat = 0; 1679 do_dot = True; 1680 goto again; 1681 } 1682 k++; 1683 printf( "\n" ); 1684 } 1685} 1686 1687#ifdef __powerpc64__ 1688void test_stdbrx(void) 1689{ 1690 unsigned long long store, val = 0xdeadbacf12345678ULL; 1691 printf("stdbrx: 0x%llx ==> ", val); 1692 r17 = (HWord_t)val; 1693 r14 = (HWord_t)&store; 1694 __asm__ __volatile__ ("stdbrx %0, 0, %1" : : "r"(r17), "r"(r14)); 1695 printf("0x%llx\n", store); 1696 printf( "\n" ); 1697} 1698#endif 1699 1700static test_table_t 1701 all_tests[] = 1702{ 1703 { &test_vx_vector_one_fp_arg, 1704 "Test VSX vector single arg instructions"}, 1705 { &test_vx_vector_fp_ops, 1706 "Test VSX floating point compare and basic arithmetic instructions" }, 1707#ifdef __powerpc64__ 1708 { &test_bpermd, 1709 "Test bit permute double"}, 1710#endif 1711 { &test_xxsel, 1712 "Test xxsel instruction" }, 1713 { &test_xxspltw, 1714 "Test xxspltw instruction" }, 1715 { &test_div_extensions, 1716 "Test div extensions" }, 1717 { &test_fct_ops, 1718 "Test floating point convert [word | doubleword] unsigned, with round toward zero" }, 1719#ifdef __powerpc64__ 1720 { &test_stdbrx, 1721 "Test stdbrx instruction"}, 1722#endif 1723 { &test_vx_aORm_fp_ops, 1724 "Test floating point arithmetic instructions -- with a{d|s}p or m{d|s}p"}, 1725 { &test_vx_simple_scalar_fp_ops, 1726 "Test scalar floating point arithmetic instructions"}, 1727 { NULL, NULL } 1728}; 1729#endif // HAS_VSX 1730 1731int main(int argc, char *argv[]) 1732{ 1733#ifdef HAS_VSX 1734 1735 test_table_t aTest; 1736 test_func_t func; 1737 int i = 0; 1738 1739 while ((func = all_tests[i].test_category)) { 1740 aTest = all_tests[i]; 1741 printf( "%s\n", aTest.name ); 1742 (*func)(); 1743 i++; 1744 } 1745 if (spec_fargs) 1746 free(spec_fargs); 1747 if (spec_sp_fargs) 1748 free(spec_sp_fargs); 1749 1750#endif // HAS _VSX 1751 1752 return 0; 1753} 1754