1 /* 2 * Copyright (C) 2010 The Android Open Source Project 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * * Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * * Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in 12 * the documentation and/or other materials provided with the 13 * distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 16 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 17 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 18 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 19 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 21 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 22 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 23 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 24 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 25 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 /* ChangeLog for this library: 30 * 31 * NDK r8d: Add android_setCpu(). 32 * 33 * NDK r8c: Add new ARM CPU features: VFPv2, VFP_D32, VFP_FP16, 34 * VFP_FMA, NEON_FMA, IDIV_ARM, IDIV_THUMB2 and iWMMXt. 35 * 36 * Rewrite the code to parse /proc/self/auxv instead of 37 * the "Features" field in /proc/cpuinfo. 38 * 39 * Dynamically allocate the buffer that hold the content 40 * of /proc/cpuinfo to deal with newer hardware. 41 * 42 * NDK r7c: Fix CPU count computation. The old method only reported the 43 * number of _active_ CPUs when the library was initialized, 44 * which could be less than the real total. 45 * 46 * NDK r5: Handle buggy kernels which report a CPU Architecture number of 7 47 * for an ARMv6 CPU (see below). 48 * 49 * Handle kernels that only report 'neon', and not 'vfpv3' 50 * (VFPv3 is mandated by the ARM architecture is Neon is implemented) 51 * 52 * Handle kernels that only report 'vfpv3d16', and not 'vfpv3' 53 * 54 * Fix x86 compilation. Report ANDROID_CPU_FAMILY_X86 in 55 * android_getCpuFamily(). 56 * 57 * NDK r4: Initial release 58 */ 59 #include <sys/system_properties.h> 60 #ifdef __arm__ 61 #include <machine/cpu-features.h> 62 #endif 63 #include <pthread.h> 64 #include "cpu-features.h" 65 #include <stdio.h> 66 #include <stdlib.h> 67 #include <fcntl.h> 68 #include <errno.h> 69 70 static pthread_once_t g_once; 71 static int g_inited; 72 static AndroidCpuFamily g_cpuFamily; 73 static uint64_t g_cpuFeatures; 74 static int g_cpuCount; 75 76 static const int android_cpufeatures_debug = 0; 77 78 #ifdef __arm__ 79 # define DEFAULT_CPU_FAMILY ANDROID_CPU_FAMILY_ARM 80 #elif defined __i386__ 81 # define DEFAULT_CPU_FAMILY ANDROID_CPU_FAMILY_X86 82 #else 83 # define DEFAULT_CPU_FAMILY ANDROID_CPU_FAMILY_UNKNOWN 84 #endif 85 86 #define D(...) \ 87 do { \ 88 if (android_cpufeatures_debug) { \ 89 printf(__VA_ARGS__); fflush(stdout); \ 90 } \ 91 } while (0) 92 93 #ifdef __i386__ 94 static __inline__ void x86_cpuid(int func, int values[4]) 95 { 96 int a, b, c, d; 97 /* We need to preserve ebx since we're compiling PIC code */ 98 /* this means we can't use "=b" for the second output register */ 99 __asm__ __volatile__ ( \ 100 "push %%ebx\n" 101 "cpuid\n" \ 102 "mov %%ebx, %1\n" 103 "pop %%ebx\n" 104 : "=a" (a), "=r" (b), "=c" (c), "=d" (d) \ 105 : "a" (func) \ 106 ); 107 values[0] = a; 108 values[1] = b; 109 values[2] = c; 110 values[3] = d; 111 } 112 #endif 113 114 /* Get the size of a file by reading it until the end. This is needed 115 * because files under /proc do not always return a valid size when 116 * using fseek(0, SEEK_END) + ftell(). Nor can they be mmap()-ed. 117 */ 118 static int 119 get_file_size(const char* pathname) 120 { 121 int fd, ret, result = 0; 122 char buffer[256]; 123 124 fd = open(pathname, O_RDONLY); 125 if (fd < 0) { 126 D("Can't open %s: %s\n", pathname, strerror(errno)); 127 return -1; 128 } 129 130 for (;;) { 131 int ret = read(fd, buffer, sizeof buffer); 132 if (ret < 0) { 133 if (errno == EINTR) 134 continue; 135 D("Error while reading %s: %s\n", pathname, strerror(errno)); 136 break; 137 } 138 if (ret == 0) 139 break; 140 141 result += ret; 142 } 143 close(fd); 144 return result; 145 } 146 147 /* Read the content of /proc/cpuinfo into a user-provided buffer. 148 * Return the length of the data, or -1 on error. Does *not* 149 * zero-terminate the content. Will not read more 150 * than 'buffsize' bytes. 151 */ 152 static int 153 read_file(const char* pathname, char* buffer, size_t buffsize) 154 { 155 int fd, count; 156 157 fd = open(pathname, O_RDONLY); 158 if (fd < 0) { 159 D("Could not open %s: %s\n", pathname, strerror(errno)); 160 return -1; 161 } 162 count = 0; 163 while (count < (int)buffsize) { 164 int ret = read(fd, buffer + count, buffsize - count); 165 if (ret < 0) { 166 if (errno == EINTR) 167 continue; 168 D("Error while reading from %s: %s\n", pathname, strerror(errno)); 169 if (count == 0) 170 count = -1; 171 break; 172 } 173 if (ret == 0) 174 break; 175 count += ret; 176 } 177 close(fd); 178 return count; 179 } 180 181 /* Extract the content of a the first occurence of a given field in 182 * the content of /proc/cpuinfo and return it as a heap-allocated 183 * string that must be freed by the caller. 184 * 185 * Return NULL if not found 186 */ 187 static char* 188 extract_cpuinfo_field(char* buffer, int buflen, const char* field) 189 { 190 int fieldlen = strlen(field); 191 char* bufend = buffer + buflen; 192 char* result = NULL; 193 int len, ignore; 194 const char *p, *q; 195 196 /* Look for first field occurence, and ensures it starts the line. */ 197 p = buffer; 198 bufend = buffer + buflen; 199 for (;;) { 200 p = memmem(p, bufend-p, field, fieldlen); 201 if (p == NULL) 202 goto EXIT; 203 204 if (p == buffer || p[-1] == '\n') 205 break; 206 207 p += fieldlen; 208 } 209 210 /* Skip to the first column followed by a space */ 211 p += fieldlen; 212 p = memchr(p, ':', bufend-p); 213 if (p == NULL || p[1] != ' ') 214 goto EXIT; 215 216 /* Find the end of the line */ 217 p += 2; 218 q = memchr(p, '\n', bufend-p); 219 if (q == NULL) 220 q = bufend; 221 222 /* Copy the line into a heap-allocated buffer */ 223 len = q-p; 224 result = malloc(len+1); 225 if (result == NULL) 226 goto EXIT; 227 228 memcpy(result, p, len); 229 result[len] = '\0'; 230 231 EXIT: 232 return result; 233 } 234 235 /* Like strlen(), but for constant string literals */ 236 #define STRLEN_CONST(x) ((sizeof(x)-1) 237 238 239 /* Checks that a space-separated list of items contains one given 'item'. 240 * Returns 1 if found, 0 otherwise. 241 */ 242 static int 243 has_list_item(const char* list, const char* item) 244 { 245 const char* p = list; 246 int itemlen = strlen(item); 247 248 if (list == NULL) 249 return 0; 250 251 while (*p) { 252 const char* q; 253 254 /* skip spaces */ 255 while (*p == ' ' || *p == '\t') 256 p++; 257 258 /* find end of current list item */ 259 q = p; 260 while (*q && *q != ' ' && *q != '\t') 261 q++; 262 263 if (itemlen == q-p && !memcmp(p, item, itemlen)) 264 return 1; 265 266 /* skip to next item */ 267 p = q; 268 } 269 return 0; 270 } 271 272 /* Parse an decimal integer starting from 'input', but not going further 273 * than 'limit'. Return the value into '*result'. 274 * 275 * NOTE: Does not skip over leading spaces, or deal with sign characters. 276 * NOTE: Ignores overflows. 277 * 278 * The function returns NULL in case of error (bad format), or the new 279 * position after the decimal number in case of success (which will always 280 * be <= 'limit'). 281 */ 282 static const char* 283 parse_decimal(const char* input, const char* limit, int* result) 284 { 285 const char* p = input; 286 int val = 0; 287 while (p < limit) { 288 int d = (*p - '0'); 289 if ((unsigned)d >= 10U) 290 break; 291 val = val*10 + d; 292 p++; 293 } 294 if (p == input) 295 return NULL; 296 297 *result = val; 298 return p; 299 } 300 301 /* This small data type is used to represent a CPU list / mask, as read 302 * from sysfs on Linux. See http://www.kernel.org/doc/Documentation/cputopology.txt 303 * 304 * For now, we don't expect more than 32 cores on mobile devices, so keep 305 * everything simple. 306 */ 307 typedef struct { 308 uint32_t mask; 309 } CpuList; 310 311 static __inline__ void 312 cpulist_init(CpuList* list) { 313 list->mask = 0; 314 } 315 316 static __inline__ void 317 cpulist_and(CpuList* list1, CpuList* list2) { 318 list1->mask &= list2->mask; 319 } 320 321 static __inline__ void 322 cpulist_set(CpuList* list, int index) { 323 if ((unsigned)index < 32) { 324 list->mask |= (uint32_t)(1U << index); 325 } 326 } 327 328 static __inline__ int 329 cpulist_count(CpuList* list) { 330 return __builtin_popcount(list->mask); 331 } 332 333 /* Parse a textual list of cpus and store the result inside a CpuList object. 334 * Input format is the following: 335 * - comma-separated list of items (no spaces) 336 * - each item is either a single decimal number (cpu index), or a range made 337 * of two numbers separated by a single dash (-). Ranges are inclusive. 338 * 339 * Examples: 0 340 * 2,4-127,128-143 341 * 0-1 342 */ 343 static void 344 cpulist_parse(CpuList* list, const char* line, int line_len) 345 { 346 const char* p = line; 347 const char* end = p + line_len; 348 const char* q; 349 350 /* NOTE: the input line coming from sysfs typically contains a 351 * trailing newline, so take care of it in the code below 352 */ 353 while (p < end && *p != '\n') 354 { 355 int val, start_value, end_value; 356 357 /* Find the end of current item, and put it into 'q' */ 358 q = memchr(p, ',', end-p); 359 if (q == NULL) { 360 q = end; 361 } 362 363 /* Get first value */ 364 p = parse_decimal(p, q, &start_value); 365 if (p == NULL) 366 goto BAD_FORMAT; 367 368 end_value = start_value; 369 370 /* If we're not at the end of the item, expect a dash and 371 * and integer; extract end value. 372 */ 373 if (p < q && *p == '-') { 374 p = parse_decimal(p+1, q, &end_value); 375 if (p == NULL) 376 goto BAD_FORMAT; 377 } 378 379 /* Set bits CPU list bits */ 380 for (val = start_value; val <= end_value; val++) { 381 cpulist_set(list, val); 382 } 383 384 /* Jump to next item */ 385 p = q; 386 if (p < end) 387 p++; 388 } 389 390 BAD_FORMAT: 391 ; 392 } 393 394 /* Read a CPU list from one sysfs file */ 395 static void 396 cpulist_read_from(CpuList* list, const char* filename) 397 { 398 char file[64]; 399 int filelen; 400 401 cpulist_init(list); 402 403 filelen = read_file(filename, file, sizeof file); 404 if (filelen < 0) { 405 D("Could not read %s: %s\n", filename, strerror(errno)); 406 return; 407 } 408 409 cpulist_parse(list, file, filelen); 410 } 411 412 // See <asm/hwcap.h> kernel header. 413 #define HWCAP_VFP (1 << 6) 414 #define HWCAP_IWMMXT (1 << 9) 415 #define HWCAP_NEON (1 << 12) 416 #define HWCAP_VFPv3 (1 << 13) 417 #define HWCAP_VFPv3D16 (1 << 14) 418 #define HWCAP_VFPv4 (1 << 16) 419 #define HWCAP_IDIVA (1 << 17) 420 #define HWCAP_IDIVT (1 << 18) 421 422 #define AT_HWCAP 16 423 424 /* Read the ELF HWCAP flags by parsing /proc/self/auxv 425 */ 426 static uint32_t 427 get_elf_hwcap(void) 428 { 429 uint32_t result = 0; 430 const char filepath[] = "/proc/self/auxv"; 431 int fd = open(filepath, O_RDONLY); 432 if (fd < 0) { 433 D("Could not open %s: %s\n", filepath, strerror(errno)); 434 return 0; 435 } 436 437 struct { uint32_t tag; uint32_t value; } entry; 438 439 for (;;) { 440 int ret = read(fd, (char*)&entry, sizeof entry); 441 if (ret < 0) { 442 if (errno == EINTR) 443 continue; 444 D("Error while reading %s: %s\n", filepath, strerror(errno)); 445 break; 446 } 447 // Detect end of list. 448 if (ret == 0 || (entry.tag == 0 && entry.value == 0)) 449 break; 450 if (entry.tag == AT_HWCAP) { 451 result = entry.value; 452 break; 453 } 454 } 455 close(fd); 456 return result; 457 } 458 459 /* Return the number of cpus present on a given device. 460 * 461 * To handle all weird kernel configurations, we need to compute the 462 * intersection of the 'present' and 'possible' CPU lists and count 463 * the result. 464 */ 465 static int 466 get_cpu_count(void) 467 { 468 CpuList cpus_present[1]; 469 CpuList cpus_possible[1]; 470 471 cpulist_read_from(cpus_present, "/sys/devices/system/cpu/present"); 472 cpulist_read_from(cpus_possible, "/sys/devices/system/cpu/possible"); 473 474 /* Compute the intersection of both sets to get the actual number of 475 * CPU cores that can be used on this device by the kernel. 476 */ 477 cpulist_and(cpus_present, cpus_possible); 478 479 return cpulist_count(cpus_present); 480 } 481 482 static void 483 android_cpuInitFamily(void) 484 { 485 #if defined(__ARM_ARCH__) 486 g_cpuFamily = ANDROID_CPU_FAMILY_ARM; 487 #elif defined(__i386__) 488 g_cpuFamily = ANDROID_CPU_FAMILY_X86; 489 #elif defined(_MIPS_ARCH) 490 g_cpuFamily = ANDROID_CPU_FAMILY_MIPS; 491 #else 492 g_cpuFamily = ANDROID_CPU_FAMILY_UNKNOWN; 493 #endif 494 } 495 496 static void 497 android_cpuInit(void) 498 { 499 char* cpuinfo = NULL; 500 int cpuinfo_len; 501 502 android_cpuInitFamily(); 503 504 g_cpuFeatures = 0; 505 g_cpuCount = 1; 506 g_inited = 1; 507 508 cpuinfo_len = get_file_size("/proc/cpuinfo"); 509 if (cpuinfo_len < 0) { 510 D("cpuinfo_len cannot be computed!"); 511 return; 512 } 513 cpuinfo = malloc(cpuinfo_len); 514 if (cpuinfo == NULL) { 515 D("cpuinfo buffer could not be allocated"); 516 return; 517 } 518 cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, cpuinfo_len); 519 D("cpuinfo_len is (%d):\n%.*s\n", cpuinfo_len, 520 cpuinfo_len >= 0 ? cpuinfo_len : 0, cpuinfo); 521 522 if (cpuinfo_len < 0) /* should not happen */ { 523 free(cpuinfo); 524 return; 525 } 526 527 /* Count the CPU cores, the value may be 0 for single-core CPUs */ 528 g_cpuCount = get_cpu_count(); 529 if (g_cpuCount == 0) { 530 g_cpuCount = 1; 531 } 532 533 D("found cpuCount = %d\n", g_cpuCount); 534 535 #ifdef __ARM_ARCH__ 536 { 537 char* features = NULL; 538 char* architecture = NULL; 539 540 /* Extract architecture from the "CPU Architecture" field. 541 * The list is well-known, unlike the the output of 542 * the 'Processor' field which can vary greatly. 543 * 544 * See the definition of the 'proc_arch' array in 545 * $KERNEL/arch/arm/kernel/setup.c and the 'c_show' function in 546 * same file. 547 */ 548 char* cpuArch = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "CPU architecture"); 549 550 if (cpuArch != NULL) { 551 char* end; 552 long archNumber; 553 int hasARMv7 = 0; 554 555 D("found cpuArch = '%s'\n", cpuArch); 556 557 /* read the initial decimal number, ignore the rest */ 558 archNumber = strtol(cpuArch, &end, 10); 559 560 /* Here we assume that ARMv8 will be upwards compatible with v7 561 * in the future. Unfortunately, there is no 'Features' field to 562 * indicate that Thumb-2 is supported. 563 */ 564 if (end > cpuArch && archNumber >= 7) { 565 hasARMv7 = 1; 566 } 567 568 /* Unfortunately, it seems that certain ARMv6-based CPUs 569 * report an incorrect architecture number of 7! 570 * 571 * See http://code.google.com/p/android/issues/detail?id=10812 572 * 573 * We try to correct this by looking at the 'elf_format' 574 * field reported by the 'Processor' field, which is of the 575 * form of "(v7l)" for an ARMv7-based CPU, and "(v6l)" for 576 * an ARMv6-one. 577 */ 578 if (hasARMv7) { 579 char* cpuProc = extract_cpuinfo_field(cpuinfo, cpuinfo_len, 580 "Processor"); 581 if (cpuProc != NULL) { 582 D("found cpuProc = '%s'\n", cpuProc); 583 if (has_list_item(cpuProc, "(v6l)")) { 584 D("CPU processor and architecture mismatch!!\n"); 585 hasARMv7 = 0; 586 } 587 free(cpuProc); 588 } 589 } 590 591 if (hasARMv7) { 592 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_ARMv7; 593 } 594 595 /* The LDREX / STREX instructions are available from ARMv6 */ 596 if (archNumber >= 6) { 597 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_LDREX_STREX; 598 } 599 600 free(cpuArch); 601 } 602 603 /* Extract the list of CPU features from ELF hwcaps */ 604 uint32_t hwcaps = get_elf_hwcap(); 605 606 if (hwcaps != 0) { 607 int has_vfp = (hwcaps & HWCAP_VFP); 608 int has_vfpv3 = (hwcaps & HWCAP_VFPv3); 609 int has_vfpv3d16 = (hwcaps & HWCAP_VFPv3D16); 610 int has_vfpv4 = (hwcaps & HWCAP_VFPv4); 611 int has_neon = (hwcaps & HWCAP_NEON); 612 int has_idiva = (hwcaps & HWCAP_IDIVA); 613 int has_idivt = (hwcaps & HWCAP_IDIVT); 614 int has_iwmmxt = (hwcaps & HWCAP_IWMMXT); 615 616 // The kernel does a poor job at ensuring consistency when 617 // describing CPU features. So lots of guessing is needed. 618 619 // 'vfpv4' implies VFPv3|VFP_FMA|FP16 620 if (has_vfpv4) 621 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3 | 622 ANDROID_CPU_ARM_FEATURE_VFP_FP16 | 623 ANDROID_CPU_ARM_FEATURE_VFP_FMA; 624 625 // 'vfpv3' or 'vfpv3d16' imply VFPv3. Note that unlike GCC, 626 // a value of 'vfpv3' doesn't necessarily mean that the D32 627 // feature is present, so be conservative. All CPUs in the 628 // field that support D32 also support NEON, so this should 629 // not be a problem in practice. 630 if (has_vfpv3 || has_vfpv3d16) 631 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3; 632 633 // 'vfp' is super ambiguous. Depending on the kernel, it can 634 // either mean VFPv2 or VFPv3. Make it depend on ARMv7. 635 if (has_vfp) { 636 if (g_cpuFeatures & ANDROID_CPU_ARM_FEATURE_ARMv7) 637 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3; 638 else 639 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv2; 640 } 641 642 // Neon implies VFPv3|D32, and if vfpv4 is detected, NEON_FMA 643 if (has_neon) { 644 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3 | 645 ANDROID_CPU_ARM_FEATURE_NEON | 646 ANDROID_CPU_ARM_FEATURE_VFP_D32; 647 if (has_vfpv4) 648 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_NEON_FMA; 649 } 650 651 // VFPv3 implies VFPv2 and ARMv7 652 if (g_cpuFeatures & ANDROID_CPU_ARM_FEATURE_VFPv3) 653 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv2 | 654 ANDROID_CPU_ARM_FEATURE_ARMv7; 655 656 // Note that some buggy kernels do not report these even when 657 // the CPU actually support the division instructions. However, 658 // assume that if 'vfpv4' is detected, then the CPU supports 659 // sdiv/udiv properly. 660 if (has_idiva || has_vfpv4) 661 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_ARM; 662 if (has_idivt || has_vfpv4) 663 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2; 664 665 if (has_iwmmxt) 666 g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_iWMMXt; 667 } 668 } 669 #endif /* __ARM_ARCH__ */ 670 671 #ifdef __i386__ 672 int regs[4]; 673 674 /* According to http://en.wikipedia.org/wiki/CPUID */ 675 #define VENDOR_INTEL_b 0x756e6547 676 #define VENDOR_INTEL_c 0x6c65746e 677 #define VENDOR_INTEL_d 0x49656e69 678 679 x86_cpuid(0, regs); 680 int vendorIsIntel = (regs[1] == VENDOR_INTEL_b && 681 regs[2] == VENDOR_INTEL_c && 682 regs[3] == VENDOR_INTEL_d); 683 684 x86_cpuid(1, regs); 685 if ((regs[2] & (1 << 9)) != 0) { 686 g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_SSSE3; 687 } 688 if ((regs[2] & (1 << 23)) != 0) { 689 g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_POPCNT; 690 } 691 if (vendorIsIntel && (regs[2] & (1 << 22)) != 0) { 692 g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_MOVBE; 693 } 694 #endif 695 696 free(cpuinfo); 697 } 698 699 700 AndroidCpuFamily 701 android_getCpuFamily(void) 702 { 703 pthread_once(&g_once, android_cpuInit); 704 return g_cpuFamily; 705 } 706 707 708 uint64_t 709 android_getCpuFeatures(void) 710 { 711 pthread_once(&g_once, android_cpuInit); 712 return g_cpuFeatures; 713 } 714 715 716 int 717 android_getCpuCount(void) 718 { 719 pthread_once(&g_once, android_cpuInit); 720 return g_cpuCount; 721 } 722 723 static void 724 android_cpuInitDummy(void) 725 { 726 g_inited = 1; 727 } 728 729 int 730 android_setCpu(int cpu_count, uint64_t cpu_features) 731 { 732 /* Fail if the library was already initialized. */ 733 if (g_inited) 734 return 0; 735 736 android_cpuInitFamily(); 737 g_cpuCount = (cpu_count <= 0 ? 1 : cpu_count); 738 g_cpuFeatures = cpu_features; 739 pthread_once(&g_once, android_cpuInitDummy); 740 741 return 1; 742 } 743 744 /* 745 * Technical note: Making sense of ARM's FPU architecture versions. 746 * 747 * FPA was ARM's first attempt at an FPU architecture. There is no Android 748 * device that actually uses it since this technology was already obsolete 749 * when the project started. If you see references to FPA instructions 750 * somewhere, you can be sure that this doesn't apply to Android at all. 751 * 752 * FPA was followed by "VFP", soon renamed "VFPv1" due to the emergence of 753 * new versions / additions to it. ARM considers this obsolete right now, 754 * and no known Android device implements it either. 755 * 756 * VFPv2 added a few instructions to VFPv1, and is an *optional* extension 757 * supported by some ARMv5TE, ARMv6 and ARMv6T2 CPUs. Note that a device 758 * supporting the 'armeabi' ABI doesn't necessarily support these. 759 * 760 * VFPv3-D16 adds a few instructions on top of VFPv2 and is typically used 761 * on ARMv7-A CPUs which implement a FPU. Note that it is also mandated 762 * by the Android 'armeabi-v7a' ABI. The -D16 suffix in its name means 763 * that it provides 16 double-precision FPU registers (d0-d15) and 32 764 * single-precision ones (s0-s31) which happen to be mapped to the same 765 * register banks. 766 * 767 * VFPv3-D32 is the name of an extension to VFPv3-D16 that provides 16 768 * additional double precision registers (d16-d31). Note that there are 769 * still only 32 single precision registers. 770 * 771 * VFPv3xD is a *subset* of VFPv3-D16 that only provides single-precision 772 * registers. It is only used on ARMv7-M (i.e. on micro-controllers) which 773 * are not supported by Android. Note that it is not compatible with VFPv2. 774 * 775 * NOTE: The term 'VFPv3' usually designate either VFPv3-D16 or VFPv3-D32 776 * depending on context. For example GCC uses it for VFPv3-D32, but 777 * the Linux kernel code uses it for VFPv3-D16 (especially in 778 * /proc/cpuinfo). Always try to use the full designation when 779 * possible. 780 * 781 * NEON, a.k.a. "ARM Advanced SIMD" is an extension that provides 782 * instructions to perform parallel computations on vectors of 8, 16, 783 * 32, 64 and 128 bit quantities. NEON requires VFPv32-D32 since all 784 * NEON registers are also mapped to the same register banks. 785 * 786 * VFPv4-D16, adds a few instructions on top of VFPv3-D16 in order to 787 * perform fused multiply-accumulate on VFP registers, as well as 788 * half-precision (16-bit) conversion operations. 789 * 790 * VFPv4-D32 is VFPv4-D16 with 32, instead of 16, FPU double precision 791 * registers. 792 * 793 * VPFv4-NEON is VFPv4-D32 with NEON instructions. It also adds fused 794 * multiply-accumulate instructions that work on the NEON registers. 795 * 796 * NOTE: Similarly, "VFPv4" might either reference VFPv4-D16 or VFPv4-D32 797 * depending on context. 798 * 799 * The following information was determined by scanning the binutils-2.22 800 * sources: 801 * 802 * Basic VFP instruction subsets: 803 * 804 * #define FPU_VFP_EXT_V1xD 0x08000000 // Base VFP instruction set. 805 * #define FPU_VFP_EXT_V1 0x04000000 // Double-precision insns. 806 * #define FPU_VFP_EXT_V2 0x02000000 // ARM10E VFPr1. 807 * #define FPU_VFP_EXT_V3xD 0x01000000 // VFPv3 single-precision. 808 * #define FPU_VFP_EXT_V3 0x00800000 // VFPv3 double-precision. 809 * #define FPU_NEON_EXT_V1 0x00400000 // Neon (SIMD) insns. 810 * #define FPU_VFP_EXT_D32 0x00200000 // Registers D16-D31. 811 * #define FPU_VFP_EXT_FP16 0x00100000 // Half-precision extensions. 812 * #define FPU_NEON_EXT_FMA 0x00080000 // Neon fused multiply-add 813 * #define FPU_VFP_EXT_FMA 0x00040000 // VFP fused multiply-add 814 * 815 * FPU types (excluding NEON) 816 * 817 * FPU_VFP_V1xD (EXT_V1xD) 818 * | 819 * +--------------------------+ 820 * | | 821 * FPU_VFP_V1 (+EXT_V1) FPU_VFP_V3xD (+EXT_V2+EXT_V3xD) 822 * | | 823 * | | 824 * FPU_VFP_V2 (+EXT_V2) FPU_VFP_V4_SP_D16 (+EXT_FP16+EXT_FMA) 825 * | 826 * FPU_VFP_V3D16 (+EXT_Vx3D+EXT_V3) 827 * | 828 * +--------------------------+ 829 * | | 830 * FPU_VFP_V3 (+EXT_D32) FPU_VFP_V4D16 (+EXT_FP16+EXT_FMA) 831 * | | 832 * | FPU_VFP_V4 (+EXT_D32) 833 * | 834 * FPU_VFP_HARD (+EXT_FMA+NEON_EXT_FMA) 835 * 836 * VFP architectures: 837 * 838 * ARCH_VFP_V1xD (EXT_V1xD) 839 * | 840 * +------------------+ 841 * | | 842 * | ARCH_VFP_V3xD (+EXT_V2+EXT_V3xD) 843 * | | 844 * | ARCH_VFP_V3xD_FP16 (+EXT_FP16) 845 * | | 846 * | ARCH_VFP_V4_SP_D16 (+EXT_FMA) 847 * | 848 * ARCH_VFP_V1 (+EXT_V1) 849 * | 850 * ARCH_VFP_V2 (+EXT_V2) 851 * | 852 * ARCH_VFP_V3D16 (+EXT_V3xD+EXT_V3) 853 * | 854 * +-------------------+ 855 * | | 856 * | ARCH_VFP_V3D16_FP16 (+EXT_FP16) 857 * | 858 * +-------------------+ 859 * | | 860 * | ARCH_VFP_V4_D16 (+EXT_FP16+EXT_FMA) 861 * | | 862 * | ARCH_VFP_V4 (+EXT_D32) 863 * | | 864 * | ARCH_NEON_VFP_V4 (+EXT_NEON+EXT_NEON_FMA) 865 * | 866 * ARCH_VFP_V3 (+EXT_D32) 867 * | 868 * +-------------------+ 869 * | | 870 * | ARCH_VFP_V3_FP16 (+EXT_FP16) 871 * | 872 * ARCH_VFP_V3_PLUS_NEON_V1 (+EXT_NEON) 873 * | 874 * ARCH_NEON_FP16 (+EXT_FP16) 875 * 876 * -fpu=<name> values and their correspondance with FPU architectures above: 877 * 878 * {"vfp", FPU_ARCH_VFP_V2}, 879 * {"vfp9", FPU_ARCH_VFP_V2}, 880 * {"vfp3", FPU_ARCH_VFP_V3}, // For backwards compatbility. 881 * {"vfp10", FPU_ARCH_VFP_V2}, 882 * {"vfp10-r0", FPU_ARCH_VFP_V1}, 883 * {"vfpxd", FPU_ARCH_VFP_V1xD}, 884 * {"vfpv2", FPU_ARCH_VFP_V2}, 885 * {"vfpv3", FPU_ARCH_VFP_V3}, 886 * {"vfpv3-fp16", FPU_ARCH_VFP_V3_FP16}, 887 * {"vfpv3-d16", FPU_ARCH_VFP_V3D16}, 888 * {"vfpv3-d16-fp16", FPU_ARCH_VFP_V3D16_FP16}, 889 * {"vfpv3xd", FPU_ARCH_VFP_V3xD}, 890 * {"vfpv3xd-fp16", FPU_ARCH_VFP_V3xD_FP16}, 891 * {"neon", FPU_ARCH_VFP_V3_PLUS_NEON_V1}, 892 * {"neon-fp16", FPU_ARCH_NEON_FP16}, 893 * {"vfpv4", FPU_ARCH_VFP_V4}, 894 * {"vfpv4-d16", FPU_ARCH_VFP_V4D16}, 895 * {"fpv4-sp-d16", FPU_ARCH_VFP_V4_SP_D16}, 896 * {"neon-vfpv4", FPU_ARCH_NEON_VFP_V4}, 897 * 898 * 899 * Simplified diagram that only includes FPUs supported by Android: 900 * Only ARCH_VFP_V3D16 is actually mandated by the armeabi-v7a ABI, 901 * all others are optional and must be probed at runtime. 902 * 903 * ARCH_VFP_V3D16 (EXT_V1xD+EXT_V1+EXT_V2+EXT_V3xD+EXT_V3) 904 * | 905 * +-------------------+ 906 * | | 907 * | ARCH_VFP_V3D16_FP16 (+EXT_FP16) 908 * | 909 * +-------------------+ 910 * | | 911 * | ARCH_VFP_V4_D16 (+EXT_FP16+EXT_FMA) 912 * | | 913 * | ARCH_VFP_V4 (+EXT_D32) 914 * | | 915 * | ARCH_NEON_VFP_V4 (+EXT_NEON+EXT_NEON_FMA) 916 * | 917 * ARCH_VFP_V3 (+EXT_D32) 918 * | 919 * +-------------------+ 920 * | | 921 * | ARCH_VFP_V3_FP16 (+EXT_FP16) 922 * | 923 * ARCH_VFP_V3_PLUS_NEON_V1 (+EXT_NEON) 924 * | 925 * ARCH_NEON_FP16 (+EXT_FP16) 926 * 927 */ 928
