diff --git a/configure.ac b/configure.ac index 072c610..16b31d5 100755 --- a/configure.ac +++ b/configure.ac @@ -13,7 +13,7 @@ dnl Init Autoconf/Automake/Libtool -AC_INIT([Sandia OpenSHMEM Test Suite], [1.4.0], [https://github.com/Sandia-OpenSHMEM/SOS]) +AC_INIT([Sandia OpenSHMEM Test Suite], [1.4.1], [https://github.com/Sandia-OpenSHMEM/SOS]) AC_PREREQ([2.60]) AC_CONFIG_AUX_DIR([config]) AC_CONFIG_MACRO_DIR([config]) @@ -198,6 +198,7 @@ AC_CONFIG_FILES([Makefile test/Makefile test/unit/Makefile test/shmemx/Makefile + test/include/Makefile test/performance/Makefile test/performance/shmem_perf_suite/Makefile test/performance/tests/Makefile diff --git a/test/Makefile.am b/test/Makefile.am index 454d797..419484e 100644 --- a/test/Makefile.am +++ b/test/Makefile.am @@ -11,4 +11,4 @@ # information, see the LICENSE file in the top level directory of the # distribution. -SUBDIRS = unit shmemx performance apps +SUBDIRS = unit shmemx performance apps include diff --git a/test/apps/gups.c b/test/apps/gups.c index 46fb90e..f7fc03e 100644 --- a/test/apps/gups.c +++ b/test/apps/gups.c @@ -309,6 +309,7 @@ int SHMEMRandomAccess(void) { int64_t i; + uint64_t i_u; static int64_t NumErrors, GlbNumErrors; int NumProcs, MyProc; @@ -435,8 +436,8 @@ SHMEMRandomAccess(void) } /* Initialize main table */ - for (i=0; i /* memcmp,strlen */ +#include /* ptrdiff_t */ +#include /* exit() */ + +/* These macros use decltype or the earlier __typeof GNU extension. + As decltype is only available in newer compilers (VS2010 or gcc 4.3+ + when compiling c++ source) this code uses whatever method is needed + or, for VS2008 where neither is available, uses casting workarounds. */ +#if defined(_MSC_VER) /* MS compiler */ +#if _MSC_VER >= 1600 && defined(__cplusplus) /* VS2010 or newer in C++ mode */ +#define DECLTYPE(x) (decltype(x)) +#else /* VS2008 or older (or VS2010 in C mode) */ +#define NO_DECLTYPE +#define DECLTYPE(x) +#endif +#elif defined(__BORLANDC__) || defined(__LCC__) || defined(__WATCOMC__) +#define NO_DECLTYPE +#define DECLTYPE(x) +#else /* GNU, Sun and other compilers */ +#define DECLTYPE(x) (__typeof(x)) +#endif + +#ifdef NO_DECLTYPE +#define DECLTYPE_ASSIGN(dst,src) \ +do { \ + char **_da_dst = (char**)(&(dst)); \ + *_da_dst = (char*)(src); \ +} while (0) +#else +#define DECLTYPE_ASSIGN(dst,src) \ +do { \ + (dst) = DECLTYPE(dst)(src); \ +} while (0) +#endif + +/* a number of the hash function use uint32_t which isn't defined on Pre VS2010 */ +#if defined(_WIN32) +#if defined(_MSC_VER) && _MSC_VER >= 1600 +#include +#elif defined(__WATCOMC__) || defined(__MINGW32__) || defined(__CYGWIN__) +#include +#else +typedef unsigned int uint32_t; +typedef unsigned char uint8_t; +#endif +#elif defined(__GNUC__) && !defined(__VXWORKS__) +#include +#else +typedef unsigned int uint32_t; +typedef unsigned char uint8_t; +#endif + +#ifndef uthash_fatal +#define uthash_fatal(msg) exit(-1) /* fatal error (out of memory,etc) */ +#endif +#ifndef uthash_malloc +#define uthash_malloc(sz) malloc(sz) /* malloc fcn */ +#endif +#ifndef uthash_free +#define uthash_free(ptr,sz) free(ptr) /* free fcn */ +#endif +#ifndef uthash_strlen +#define uthash_strlen(s) strlen(s) +#endif +#ifndef uthash_memcmp +#define uthash_memcmp(a,b,n) memcmp(a,b,n) +#endif + +#ifndef uthash_noexpand_fyi +#define uthash_noexpand_fyi(tbl) /* can be defined to log noexpand */ +#endif +#ifndef uthash_expand_fyi +#define uthash_expand_fyi(tbl) /* can be defined to log expands */ +#endif + +/* initial number of buckets */ +#define HASH_INITIAL_NUM_BUCKETS 32U /* initial number of buckets */ +#define HASH_INITIAL_NUM_BUCKETS_LOG2 5U /* lg2 of initial number of buckets */ +#define HASH_BKT_CAPACITY_THRESH 10U /* expand when bucket count reaches */ + +/* calculate the element whose hash handle address is hhp */ +#define ELMT_FROM_HH(tbl,hhp) ((void*)(((char*)(hhp)) - ((tbl)->hho))) +/* calculate the hash handle from element address elp */ +#define HH_FROM_ELMT(tbl,elp) ((UT_hash_handle *)(((char*)(elp)) + ((tbl)->hho))) + +#define HASH_VALUE(keyptr,keylen,hashv) \ +do { \ + HASH_FCN(keyptr, keylen, hashv); \ +} while (0) + +#define HASH_FIND_BYHASHVALUE(hh,head,keyptr,keylen,hashval,out) \ +do { \ + (out) = NULL; \ + if (head) { \ + unsigned _hf_bkt; \ + HASH_TO_BKT(hashval, (head)->hh.tbl->num_buckets, _hf_bkt); \ + if (HASH_BLOOM_TEST((head)->hh.tbl, hashval) != 0) { \ + HASH_FIND_IN_BKT((head)->hh.tbl, hh, (head)->hh.tbl->buckets[ _hf_bkt ], keyptr, keylen, hashval, out); \ + } \ + } \ +} while (0) + +#define HASH_FIND(hh,head,keyptr,keylen,out) \ +do { \ + unsigned _hf_hashv; \ + HASH_VALUE(keyptr, keylen, _hf_hashv); \ + HASH_FIND_BYHASHVALUE(hh, head, keyptr, keylen, _hf_hashv, out); \ +} while (0) + +#ifdef HASH_BLOOM +#define HASH_BLOOM_BITLEN (1UL << HASH_BLOOM) +#define HASH_BLOOM_BYTELEN (HASH_BLOOM_BITLEN/8UL) + (((HASH_BLOOM_BITLEN%8UL)!=0UL) ? 1UL : 0UL) +#define HASH_BLOOM_MAKE(tbl) \ +do { \ + (tbl)->bloom_nbits = HASH_BLOOM; \ + (tbl)->bloom_bv = (uint8_t*)uthash_malloc(HASH_BLOOM_BYTELEN); \ + if (!((tbl)->bloom_bv)) { uthash_fatal( "out of memory"); } \ + memset((tbl)->bloom_bv, 0, HASH_BLOOM_BYTELEN); \ + (tbl)->bloom_sig = HASH_BLOOM_SIGNATURE; \ +} while (0) + +#define HASH_BLOOM_FREE(tbl) \ +do { \ + uthash_free((tbl)->bloom_bv, HASH_BLOOM_BYTELEN); \ +} while (0) + +#define HASH_BLOOM_BITSET(bv,idx) (bv[(idx)/8U] |= (1U << ((idx)%8U))) +#define HASH_BLOOM_BITTEST(bv,idx) (bv[(idx)/8U] & (1U << ((idx)%8U))) + +#define HASH_BLOOM_ADD(tbl,hashv) \ + HASH_BLOOM_BITSET((tbl)->bloom_bv, (hashv & (uint32_t)((1ULL << (tbl)->bloom_nbits) - 1U))) + +#define HASH_BLOOM_TEST(tbl,hashv) \ + HASH_BLOOM_BITTEST((tbl)->bloom_bv, (hashv & (uint32_t)((1ULL << (tbl)->bloom_nbits) - 1U))) + +#else +#define HASH_BLOOM_MAKE(tbl) +#define HASH_BLOOM_FREE(tbl) +#define HASH_BLOOM_ADD(tbl,hashv) +#define HASH_BLOOM_TEST(tbl,hashv) (1) +#define HASH_BLOOM_BYTELEN 0U +#endif + +#define HASH_MAKE_TABLE(hh,head) \ +do { \ + (head)->hh.tbl = (UT_hash_table*)uthash_malloc( \ + sizeof(UT_hash_table)); \ + if (!((head)->hh.tbl)) { uthash_fatal( "out of memory"); } \ + memset((head)->hh.tbl, 0, sizeof(UT_hash_table)); \ + (head)->hh.tbl->tail = &((head)->hh); \ + (head)->hh.tbl->num_buckets = HASH_INITIAL_NUM_BUCKETS; \ + (head)->hh.tbl->log2_num_buckets = HASH_INITIAL_NUM_BUCKETS_LOG2; \ + (head)->hh.tbl->hho = (char*)(&(head)->hh) - (char*)(head); \ + (head)->hh.tbl->buckets = (UT_hash_bucket*)uthash_malloc( \ + HASH_INITIAL_NUM_BUCKETS*sizeof(struct UT_hash_bucket)); \ + if (! (head)->hh.tbl->buckets) { uthash_fatal( "out of memory"); } \ + memset((head)->hh.tbl->buckets, 0, \ + HASH_INITIAL_NUM_BUCKETS*sizeof(struct UT_hash_bucket)); \ + HASH_BLOOM_MAKE((head)->hh.tbl); \ + (head)->hh.tbl->signature = HASH_SIGNATURE; \ +} while (0) + +#define HASH_REPLACE_BYHASHVALUE_INORDER(hh,head,fieldname,keylen_in,hashval,add,replaced,cmpfcn) \ +do { \ + (replaced) = NULL; \ + HASH_FIND_BYHASHVALUE(hh, head, &((add)->fieldname), keylen_in, hashval, replaced); \ + if (replaced) { \ + HASH_DELETE(hh, head, replaced); \ + } \ + HASH_ADD_KEYPTR_BYHASHVALUE_INORDER(hh, head, &((add)->fieldname), keylen_in, hashval, add, cmpfcn); \ +} while (0) + +#define HASH_REPLACE_BYHASHVALUE(hh,head,fieldname,keylen_in,hashval,add,replaced) \ +do { \ + (replaced) = NULL; \ + HASH_FIND_BYHASHVALUE(hh, head, &((add)->fieldname), keylen_in, hashval, replaced); \ + if (replaced) { \ + HASH_DELETE(hh, head, replaced); \ + } \ + HASH_ADD_KEYPTR_BYHASHVALUE(hh, head, &((add)->fieldname), keylen_in, hashval, add); \ +} while (0) + +#define HASH_REPLACE(hh,head,fieldname,keylen_in,add,replaced) \ +do { \ + unsigned _hr_hashv; \ + HASH_VALUE(&((add)->fieldname), keylen_in, _hr_hashv); \ + HASH_REPLACE_BYHASHVALUE(hh, head, fieldname, keylen_in, _hr_hashv, add, replaced); \ +} while (0) + +#define HASH_REPLACE_INORDER(hh,head,fieldname,keylen_in,add,replaced,cmpfcn) \ +do { \ + unsigned _hr_hashv; \ + HASH_VALUE(&((add)->fieldname), keylen_in, _hr_hashv); \ + HASH_REPLACE_BYHASHVALUE_INORDER(hh, head, fieldname, keylen_in, _hr_hashv, add, replaced, cmpfcn); \ +} while (0) + +#define HASH_APPEND_LIST(hh, head, add) \ +do { \ + (add)->hh.next = NULL; \ + (add)->hh.prev = ELMT_FROM_HH((head)->hh.tbl, (head)->hh.tbl->tail); \ + (head)->hh.tbl->tail->next = (add); \ + (head)->hh.tbl->tail = &((add)->hh); \ +} while (0) + +#define HASH_ADD_KEYPTR_BYHASHVALUE_INORDER(hh,head,keyptr,keylen_in,hashval,add,cmpfcn) \ +do { \ + unsigned _ha_bkt; \ + (add)->hh.hashv = (hashval); \ + (add)->hh.key = (char*) (keyptr); \ + (add)->hh.keylen = (unsigned) (keylen_in); \ + if (!(head)) { \ + (add)->hh.next = NULL; \ + (add)->hh.prev = NULL; \ + (head) = (add); \ + HASH_MAKE_TABLE(hh, head); \ + } else { \ + struct UT_hash_handle *_hs_iter = &(head)->hh; \ + (add)->hh.tbl = (head)->hh.tbl; \ + do { \ + if (cmpfcn(DECLTYPE(head) ELMT_FROM_HH((head)->hh.tbl, _hs_iter), add) > 0) \ + break; \ + } while ((_hs_iter = _hs_iter->next)); \ + if (_hs_iter) { \ + (add)->hh.next = _hs_iter; \ + if (((add)->hh.prev = _hs_iter->prev)) { \ + HH_FROM_ELMT((head)->hh.tbl, _hs_iter->prev)->next = (add); \ + } else { \ + (head) = (add); \ + } \ + _hs_iter->prev = (add); \ + } else { \ + HASH_APPEND_LIST(hh, head, add); \ + } \ + } \ + (head)->hh.tbl->num_items++; \ + HASH_TO_BKT(hashval, (head)->hh.tbl->num_buckets, _ha_bkt); \ + HASH_ADD_TO_BKT((head)->hh.tbl->buckets[_ha_bkt], &(add)->hh); \ + HASH_BLOOM_ADD((head)->hh.tbl, hashval); \ + HASH_EMIT_KEY(hh, head, keyptr, keylen_in); \ + HASH_FSCK(hh, head); \ +} while (0) + +#define HASH_ADD_KEYPTR_INORDER(hh,head,keyptr,keylen_in,add,cmpfcn) \ +do { \ + unsigned _hs_hashv; \ + HASH_VALUE(keyptr, keylen_in, _hs_hashv); \ + HASH_ADD_KEYPTR_BYHASHVALUE_INORDER(hh, head, keyptr, keylen_in, _hs_hashv, add, cmpfcn); \ +} while (0) + +#define HASH_ADD_BYHASHVALUE_INORDER(hh,head,fieldname,keylen_in,hashval,add,cmpfcn) \ + HASH_ADD_KEYPTR_BYHASHVALUE_INORDER(hh, head, &((add)->fieldname), keylen_in, hashval, add, cmpfcn) + +#define HASH_ADD_INORDER(hh,head,fieldname,keylen_in,add,cmpfcn) \ + HASH_ADD_KEYPTR_INORDER(hh, head, &((add)->fieldname), keylen_in, add, cmpfcn) + +#define HASH_ADD_KEYPTR_BYHASHVALUE(hh,head,keyptr,keylen_in,hashval,add) \ +do { \ + unsigned _ha_bkt; \ + (add)->hh.hashv = (hashval); \ + (add)->hh.key = (char*) (keyptr); \ + (add)->hh.keylen = (unsigned) (keylen_in); \ + if (!(head)) { \ + (add)->hh.next = NULL; \ + (add)->hh.prev = NULL; \ + (head) = (add); \ + HASH_MAKE_TABLE(hh, head); \ + } else { \ + (add)->hh.tbl = (head)->hh.tbl; \ + HASH_APPEND_LIST(hh, head, add); \ + } \ + (head)->hh.tbl->num_items++; \ + HASH_TO_BKT(hashval, (head)->hh.tbl->num_buckets, _ha_bkt); \ + HASH_ADD_TO_BKT((head)->hh.tbl->buckets[_ha_bkt], &(add)->hh); \ + HASH_BLOOM_ADD((head)->hh.tbl, hashval); \ + HASH_EMIT_KEY(hh, head, keyptr, keylen_in); \ + HASH_FSCK(hh, head); \ +} while (0) + +#define HASH_ADD_KEYPTR(hh,head,keyptr,keylen_in,add) \ +do { \ + unsigned _ha_hashv; \ + HASH_VALUE(keyptr, keylen_in, _ha_hashv); \ + HASH_ADD_KEYPTR_BYHASHVALUE(hh, head, keyptr, keylen_in, _ha_hashv, add); \ +} while (0) + +#define HASH_ADD_BYHASHVALUE(hh,head,fieldname,keylen_in,hashval,add) \ + HASH_ADD_KEYPTR_BYHASHVALUE(hh, head, &((add)->fieldname), keylen_in, hashval, add) + +#define HASH_ADD(hh,head,fieldname,keylen_in,add) \ + HASH_ADD_KEYPTR(hh, head, &((add)->fieldname), keylen_in, add) + +#define HASH_TO_BKT(hashv,num_bkts,bkt) \ +do { \ + bkt = ((hashv) & ((num_bkts) - 1U)); \ +} while (0) + +/* delete "delptr" from the hash table. + * "the usual" patch-up process for the app-order doubly-linked-list. + * The use of _hd_hh_del below deserves special explanation. + * These used to be expressed using (delptr) but that led to a bug + * if someone used the same symbol for the head and deletee, like + * HASH_DELETE(hh,users,users); + * We want that to work, but by changing the head (users) below + * we were forfeiting our ability to further refer to the deletee (users) + * in the patch-up process. Solution: use scratch space to + * copy the deletee pointer, then the latter references are via that + * scratch pointer rather than through the repointed (users) symbol. + */ +#define HASH_DELETE(hh,head,delptr) \ +do { \ + struct UT_hash_handle *_hd_hh_del; \ + if ( ((delptr)->hh.prev == NULL) && ((delptr)->hh.next == NULL) ) { \ + uthash_free((head)->hh.tbl->buckets, \ + (head)->hh.tbl->num_buckets*sizeof(struct UT_hash_bucket) ); \ + HASH_BLOOM_FREE((head)->hh.tbl); \ + uthash_free((head)->hh.tbl, sizeof(UT_hash_table)); \ + head = NULL; \ + } else { \ + unsigned _hd_bkt; \ + _hd_hh_del = &((delptr)->hh); \ + if ((delptr) == ELMT_FROM_HH((head)->hh.tbl,(head)->hh.tbl->tail)) { \ + (head)->hh.tbl->tail = \ + (UT_hash_handle*)((ptrdiff_t)((delptr)->hh.prev) + \ + (head)->hh.tbl->hho); \ + } \ + if ((delptr)->hh.prev != NULL) { \ + ((UT_hash_handle*)((ptrdiff_t)((delptr)->hh.prev) + \ + (head)->hh.tbl->hho))->next = (delptr)->hh.next; \ + } else { \ + DECLTYPE_ASSIGN(head,(delptr)->hh.next); \ + } \ + if (_hd_hh_del->next != NULL) { \ + ((UT_hash_handle*)((ptrdiff_t)_hd_hh_del->next + \ + (head)->hh.tbl->hho))->prev = \ + _hd_hh_del->prev; \ + } \ + HASH_TO_BKT( _hd_hh_del->hashv, (head)->hh.tbl->num_buckets, _hd_bkt); \ + HASH_DEL_IN_BKT(hh,(head)->hh.tbl->buckets[_hd_bkt], _hd_hh_del); \ + (head)->hh.tbl->num_items--; \ + } \ + HASH_FSCK(hh,head); \ +} while (0) + + +/* convenience forms of HASH_FIND/HASH_ADD/HASH_DEL */ +#define HASH_FIND_STR(head,findstr,out) \ + HASH_FIND(hh,head,findstr,(unsigned)uthash_strlen(findstr),out) +#define HASH_ADD_STR(head,strfield,add) \ + HASH_ADD(hh,head,strfield[0],(unsigned)uthash_strlen(add->strfield),add) +#define HASH_REPLACE_STR(head,strfield,add,replaced) \ + HASH_REPLACE(hh,head,strfield[0],(unsigned)uthash_strlen(add->strfield),add,replaced) +#define HASH_FIND_INT(head,findint,out) \ + HASH_FIND(hh,head,findint,sizeof(int),out) +#define HASH_ADD_INT(head,intfield,add) \ + HASH_ADD(hh,head,intfield,sizeof(int),add) +#define HASH_REPLACE_INT(head,intfield,add,replaced) \ + HASH_REPLACE(hh,head,intfield,sizeof(int),add,replaced) +#define HASH_FIND_PTR(head,findptr,out) \ + HASH_FIND(hh,head,findptr,sizeof(void *),out) +#define HASH_ADD_PTR(head,ptrfield,add) \ + HASH_ADD(hh,head,ptrfield,sizeof(void *),add) +#define HASH_REPLACE_PTR(head,ptrfield,add,replaced) \ + HASH_REPLACE(hh,head,ptrfield,sizeof(void *),add,replaced) +#define HASH_DEL(head,delptr) \ + HASH_DELETE(hh,head,delptr) + +/* HASH_FSCK checks hash integrity on every add/delete when HASH_DEBUG is defined. + * This is for uthash developer only; it compiles away if HASH_DEBUG isn't defined. + */ +#ifdef HASH_DEBUG +#define HASH_OOPS(...) do { fprintf(stderr,__VA_ARGS__); exit(-1); } while (0) +#define HASH_FSCK(hh,head) \ +do { \ + struct UT_hash_handle *_thh; \ + if (head) { \ + unsigned _bkt_i; \ + unsigned _count; \ + char *_prev; \ + _count = 0; \ + for( _bkt_i = 0; _bkt_i < (head)->hh.tbl->num_buckets; _bkt_i++) { \ + unsigned _bkt_count = 0; \ + _thh = (head)->hh.tbl->buckets[_bkt_i].hh_head; \ + _prev = NULL; \ + while (_thh) { \ + if (_prev != (char*)(_thh->hh_prev)) { \ + HASH_OOPS("invalid hh_prev %p, actual %p\n", \ + _thh->hh_prev, _prev ); \ + } \ + _bkt_count++; \ + _prev = (char*)(_thh); \ + _thh = _thh->hh_next; \ + } \ + _count += _bkt_count; \ + if ((head)->hh.tbl->buckets[_bkt_i].count != _bkt_count) { \ + HASH_OOPS("invalid bucket count %u, actual %u\n", \ + (head)->hh.tbl->buckets[_bkt_i].count, _bkt_count); \ + } \ + } \ + if (_count != (head)->hh.tbl->num_items) { \ + HASH_OOPS("invalid hh item count %u, actual %u\n", \ + (head)->hh.tbl->num_items, _count ); \ + } \ + /* traverse hh in app order; check next/prev integrity, count */ \ + _count = 0; \ + _prev = NULL; \ + _thh = &(head)->hh; \ + while (_thh) { \ + _count++; \ + if (_prev !=(char*)(_thh->prev)) { \ + HASH_OOPS("invalid prev %p, actual %p\n", \ + _thh->prev, _prev ); \ + } \ + _prev = (char*)ELMT_FROM_HH((head)->hh.tbl, _thh); \ + _thh = ( _thh->next ? (UT_hash_handle*)((char*)(_thh->next) + \ + (head)->hh.tbl->hho) : NULL ); \ + } \ + if (_count != (head)->hh.tbl->num_items) { \ + HASH_OOPS("invalid app item count %u, actual %u\n", \ + (head)->hh.tbl->num_items, _count ); \ + } \ + } \ +} while (0) +#else +#define HASH_FSCK(hh,head) +#endif + +/* When compiled with -DHASH_EMIT_KEYS, length-prefixed keys are emitted to + * the descriptor to which this macro is defined for tuning the hash function. + * The app can #include to get the prototype for write(2). */ +#ifdef HASH_EMIT_KEYS +#define HASH_EMIT_KEY(hh,head,keyptr,fieldlen) \ +do { \ + unsigned _klen = fieldlen; \ + write(HASH_EMIT_KEYS, &_klen, sizeof(_klen)); \ + write(HASH_EMIT_KEYS, keyptr, (unsigned long)fieldlen); \ +} while (0) +#else +#define HASH_EMIT_KEY(hh,head,keyptr,fieldlen) +#endif + +/* default to Jenkin's hash unless overridden e.g. DHASH_FUNCTION=HASH_SAX */ +#ifdef HASH_FUNCTION +#define HASH_FCN HASH_FUNCTION +#else +#define HASH_FCN HASH_JEN +#endif + +/* The Bernstein hash function, used in Perl prior to v5.6. Note (x<<5+x)=x*33. */ +#define HASH_BER(key,keylen,hashv) \ +do { \ + unsigned _hb_keylen=(unsigned)keylen; \ + const unsigned char *_hb_key=(const unsigned char*)(key); \ + (hashv) = 0; \ + while (_hb_keylen-- != 0U) { \ + (hashv) = (((hashv) << 5) + (hashv)) + *_hb_key++; \ + } \ +} while (0) + + +/* SAX/FNV/OAT/JEN hash functions are macro variants of those listed at + * http://eternallyconfuzzled.com/tuts/algorithms/jsw_tut_hashing.aspx */ +#define HASH_SAX(key,keylen,hashv) \ +do { \ + unsigned _sx_i; \ + const unsigned char *_hs_key=(const unsigned char*)(key); \ + hashv = 0; \ + for(_sx_i=0; _sx_i < keylen; _sx_i++) { \ + hashv ^= (hashv << 5) + (hashv >> 2) + _hs_key[_sx_i]; \ + } \ +} while (0) +/* FNV-1a variation */ +#define HASH_FNV(key,keylen,hashv) \ +do { \ + unsigned _fn_i; \ + const unsigned char *_hf_key=(const unsigned char*)(key); \ + hashv = 2166136261U; \ + for(_fn_i=0; _fn_i < keylen; _fn_i++) { \ + hashv = hashv ^ _hf_key[_fn_i]; \ + hashv = hashv * 16777619U; \ + } \ +} while (0) + +#define HASH_OAT(key,keylen,hashv) \ +do { \ + unsigned _ho_i; \ + const unsigned char *_ho_key=(const unsigned char*)(key); \ + hashv = 0; \ + for(_ho_i=0; _ho_i < keylen; _ho_i++) { \ + hashv += _ho_key[_ho_i]; \ + hashv += (hashv << 10); \ + hashv ^= (hashv >> 6); \ + } \ + hashv += (hashv << 3); \ + hashv ^= (hashv >> 11); \ + hashv += (hashv << 15); \ +} while (0) + +#define HASH_JEN_MIX(a,b,c) \ +do { \ + a -= b; a -= c; a ^= ( c >> 13 ); \ + b -= c; b -= a; b ^= ( a << 8 ); \ + c -= a; c -= b; c ^= ( b >> 13 ); \ + a -= b; a -= c; a ^= ( c >> 12 ); \ + b -= c; b -= a; b ^= ( a << 16 ); \ + c -= a; c -= b; c ^= ( b >> 5 ); \ + a -= b; a -= c; a ^= ( c >> 3 ); \ + b -= c; b -= a; b ^= ( a << 10 ); \ + c -= a; c -= b; c ^= ( b >> 15 ); \ +} while (0) + +#define HASH_JEN(key,keylen,hashv) \ +do { \ + unsigned _hj_i,_hj_j,_hj_k; \ + unsigned const char *_hj_key=(unsigned const char*)(key); \ + hashv = 0xfeedbeefu; \ + _hj_i = _hj_j = 0x9e3779b9u; \ + _hj_k = (unsigned)(keylen); \ + while (_hj_k >= 12U) { \ + _hj_i += (_hj_key[0] + ( (unsigned)_hj_key[1] << 8 ) \ + + ( (unsigned)_hj_key[2] << 16 ) \ + + ( (unsigned)_hj_key[3] << 24 ) ); \ + _hj_j += (_hj_key[4] + ( (unsigned)_hj_key[5] << 8 ) \ + + ( (unsigned)_hj_key[6] << 16 ) \ + + ( (unsigned)_hj_key[7] << 24 ) ); \ + hashv += (_hj_key[8] + ( (unsigned)_hj_key[9] << 8 ) \ + + ( (unsigned)_hj_key[10] << 16 ) \ + + ( (unsigned)_hj_key[11] << 24 ) ); \ + \ + HASH_JEN_MIX(_hj_i, _hj_j, hashv); \ + \ + _hj_key += 12; \ + _hj_k -= 12U; \ + } \ + hashv += (unsigned)(keylen); \ + switch ( _hj_k ) { \ + case 11: hashv += ( (unsigned)_hj_key[10] << 24 ); /* FALLTHROUGH */ \ + case 10: hashv += ( (unsigned)_hj_key[9] << 16 ); /* FALLTHROUGH */ \ + case 9: hashv += ( (unsigned)_hj_key[8] << 8 ); /* FALLTHROUGH */ \ + case 8: _hj_j += ( (unsigned)_hj_key[7] << 24 ); /* FALLTHROUGH */ \ + case 7: _hj_j += ( (unsigned)_hj_key[6] << 16 ); /* FALLTHROUGH */ \ + case 6: _hj_j += ( (unsigned)_hj_key[5] << 8 ); /* FALLTHROUGH */ \ + case 5: _hj_j += _hj_key[4]; /* FALLTHROUGH */ \ + case 4: _hj_i += ( (unsigned)_hj_key[3] << 24 ); /* FALLTHROUGH */ \ + case 3: _hj_i += ( (unsigned)_hj_key[2] << 16 ); /* FALLTHROUGH */ \ + case 2: _hj_i += ( (unsigned)_hj_key[1] << 8 ); /* FALLTHROUGH */ \ + case 1: _hj_i += _hj_key[0]; \ + } \ + HASH_JEN_MIX(_hj_i, _hj_j, hashv); \ +} while (0) + +/* The Paul Hsieh hash function */ +#undef get16bits +#if (defined(__GNUC__) && defined(__i386__)) || defined(__WATCOMC__) \ + || defined(_MSC_VER) || defined (__BORLANDC__) || defined (__TURBOC__) +#define get16bits(d) (*((const uint16_t *) (d))) +#endif + +#if !defined (get16bits) +#define get16bits(d) ((((uint32_t)(((const uint8_t *)(d))[1])) << 8) \ + +(uint32_t)(((const uint8_t *)(d))[0]) ) +#endif +#define HASH_SFH(key,keylen,hashv) \ +do { \ + unsigned const char *_sfh_key=(unsigned const char*)(key); \ + uint32_t _sfh_tmp, _sfh_len = (uint32_t)keylen; \ + \ + unsigned _sfh_rem = _sfh_len & 3U; \ + _sfh_len >>= 2; \ + hashv = 0xcafebabeu; \ + \ + /* Main loop */ \ + for (;_sfh_len > 0U; _sfh_len--) { \ + hashv += get16bits (_sfh_key); \ + _sfh_tmp = ((uint32_t)(get16bits (_sfh_key+2)) << 11) ^ hashv; \ + hashv = (hashv << 16) ^ _sfh_tmp; \ + _sfh_key += 2U*sizeof (uint16_t); \ + hashv += hashv >> 11; \ + } \ + \ + /* Handle end cases */ \ + switch (_sfh_rem) { \ + case 3: hashv += get16bits (_sfh_key); \ + hashv ^= hashv << 16; \ + hashv ^= (uint32_t)(_sfh_key[sizeof (uint16_t)]) << 18; \ + hashv += hashv >> 11; \ + break; \ + case 2: hashv += get16bits (_sfh_key); \ + hashv ^= hashv << 11; \ + hashv += hashv >> 17; \ + break; \ + case 1: hashv += *_sfh_key; \ + hashv ^= hashv << 10; \ + hashv += hashv >> 1; \ + } \ + \ + /* Force "avalanching" of final 127 bits */ \ + hashv ^= hashv << 3; \ + hashv += hashv >> 5; \ + hashv ^= hashv << 4; \ + hashv += hashv >> 17; \ + hashv ^= hashv << 25; \ + hashv += hashv >> 6; \ +} while (0) + +#ifdef HASH_USING_NO_STRICT_ALIASING +/* The MurmurHash exploits some CPU's (x86,x86_64) tolerance for unaligned reads. + * For other types of CPU's (e.g. Sparc) an unaligned read causes a bus error. + * MurmurHash uses the faster approach only on CPU's where we know it's safe. + * + * Note the preprocessor built-in defines can be emitted using: + * + * gcc -m64 -dM -E - < /dev/null (on gcc) + * cc -## a.c (where a.c is a simple test file) (Sun Studio) + */ +#if (defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)) +#define MUR_GETBLOCK(p,i) p[i] +#else /* non intel */ +#define MUR_PLUS0_ALIGNED(p) (((unsigned long)p & 3UL) == 0UL) +#define MUR_PLUS1_ALIGNED(p) (((unsigned long)p & 3UL) == 1UL) +#define MUR_PLUS2_ALIGNED(p) (((unsigned long)p & 3UL) == 2UL) +#define MUR_PLUS3_ALIGNED(p) (((unsigned long)p & 3UL) == 3UL) +#define WP(p) ((uint32_t*)((unsigned long)(p) & ~3UL)) +#if (defined(__BIG_ENDIAN__) || defined(SPARC) || defined(__ppc__) || defined(__ppc64__)) +#define MUR_THREE_ONE(p) ((((*WP(p))&0x00ffffff) << 8) | (((*(WP(p)+1))&0xff000000) >> 24)) +#define MUR_TWO_TWO(p) ((((*WP(p))&0x0000ffff) <<16) | (((*(WP(p)+1))&0xffff0000) >> 16)) +#define MUR_ONE_THREE(p) ((((*WP(p))&0x000000ff) <<24) | (((*(WP(p)+1))&0xffffff00) >> 8)) +#else /* assume little endian non-intel */ +#define MUR_THREE_ONE(p) ((((*WP(p))&0xffffff00) >> 8) | (((*(WP(p)+1))&0x000000ff) << 24)) +#define MUR_TWO_TWO(p) ((((*WP(p))&0xffff0000) >>16) | (((*(WP(p)+1))&0x0000ffff) << 16)) +#define MUR_ONE_THREE(p) ((((*WP(p))&0xff000000) >>24) | (((*(WP(p)+1))&0x00ffffff) << 8)) +#endif +#define MUR_GETBLOCK(p,i) (MUR_PLUS0_ALIGNED(p) ? ((p)[i]) : \ + (MUR_PLUS1_ALIGNED(p) ? MUR_THREE_ONE(p) : \ + (MUR_PLUS2_ALIGNED(p) ? MUR_TWO_TWO(p) : \ + MUR_ONE_THREE(p)))) +#endif +#define MUR_ROTL32(x,r) (((x) << (r)) | ((x) >> (32 - (r)))) +#define MUR_FMIX(_h) \ +do { \ + _h ^= _h >> 16; \ + _h *= 0x85ebca6bu; \ + _h ^= _h >> 13; \ + _h *= 0xc2b2ae35u; \ + _h ^= _h >> 16; \ +} while (0) + +#define HASH_MUR(key,keylen,hashv) \ +do { \ + const uint8_t *_mur_data = (const uint8_t*)(key); \ + const int _mur_nblocks = (int)(keylen) / 4; \ + uint32_t _mur_h1 = 0xf88D5353u; \ + uint32_t _mur_c1 = 0xcc9e2d51u; \ + uint32_t _mur_c2 = 0x1b873593u; \ + uint32_t _mur_k1 = 0; \ + const uint8_t *_mur_tail; \ + const uint32_t *_mur_blocks = (const uint32_t*)(_mur_data+(_mur_nblocks*4)); \ + int _mur_i; \ + for(_mur_i = -_mur_nblocks; _mur_i!=0; _mur_i++) { \ + _mur_k1 = MUR_GETBLOCK(_mur_blocks,_mur_i); \ + _mur_k1 *= _mur_c1; \ + _mur_k1 = MUR_ROTL32(_mur_k1,15); \ + _mur_k1 *= _mur_c2; \ + \ + _mur_h1 ^= _mur_k1; \ + _mur_h1 = MUR_ROTL32(_mur_h1,13); \ + _mur_h1 = (_mur_h1*5U) + 0xe6546b64u; \ + } \ + _mur_tail = (const uint8_t*)(_mur_data + (_mur_nblocks*4)); \ + _mur_k1=0; \ + switch((keylen) & 3U) { \ + case 3: _mur_k1 ^= (uint32_t)_mur_tail[2] << 16; /* FALLTHROUGH */ \ + case 2: _mur_k1 ^= (uint32_t)_mur_tail[1] << 8; /* FALLTHROUGH */ \ + case 1: _mur_k1 ^= (uint32_t)_mur_tail[0]; \ + _mur_k1 *= _mur_c1; \ + _mur_k1 = MUR_ROTL32(_mur_k1,15); \ + _mur_k1 *= _mur_c2; \ + _mur_h1 ^= _mur_k1; \ + } \ + _mur_h1 ^= (uint32_t)(keylen); \ + MUR_FMIX(_mur_h1); \ + hashv = _mur_h1; \ +} while (0) +#endif /* HASH_USING_NO_STRICT_ALIASING */ + +/* iterate over items in a known bucket to find desired item */ +#define HASH_FIND_IN_BKT(tbl,hh,head,keyptr,keylen_in,hashval,out) \ +do { \ + if ((head).hh_head != NULL) { \ + DECLTYPE_ASSIGN(out, ELMT_FROM_HH(tbl, (head).hh_head)); \ + } else { \ + (out) = NULL; \ + } \ + while ((out) != NULL) { \ + if ((out)->hh.hashv == (hashval) && (out)->hh.keylen == (keylen_in)) { \ + if (uthash_memcmp((out)->hh.key, keyptr, keylen_in) == 0) { \ + break; \ + } \ + } \ + if ((out)->hh.hh_next != NULL) { \ + DECLTYPE_ASSIGN(out, ELMT_FROM_HH(tbl, (out)->hh.hh_next)); \ + } else { \ + (out) = NULL; \ + } \ + } \ +} while (0) + +/* add an item to a bucket */ +#define HASH_ADD_TO_BKT(head,addhh) \ +do { \ + head.count++; \ + (addhh)->hh_next = head.hh_head; \ + (addhh)->hh_prev = NULL; \ + if (head.hh_head != NULL) { (head).hh_head->hh_prev = (addhh); } \ + (head).hh_head=addhh; \ + if ((head.count >= ((head.expand_mult+1U) * HASH_BKT_CAPACITY_THRESH)) \ + && ((addhh)->tbl->noexpand != 1U)) { \ + HASH_EXPAND_BUCKETS((addhh)->tbl); \ + } \ +} while (0) + +/* remove an item from a given bucket */ +#define HASH_DEL_IN_BKT(hh,head,hh_del) \ + (head).count--; \ + if ((head).hh_head == hh_del) { \ + (head).hh_head = hh_del->hh_next; \ + } \ + if (hh_del->hh_prev) { \ + hh_del->hh_prev->hh_next = hh_del->hh_next; \ + } \ + if (hh_del->hh_next) { \ + hh_del->hh_next->hh_prev = hh_del->hh_prev; \ + } + +/* Bucket expansion has the effect of doubling the number of buckets + * and redistributing the items into the new buckets. Ideally the + * items will distribute more or less evenly into the new buckets + * (the extent to which this is true is a measure of the quality of + * the hash function as it applies to the key domain). + * + * With the items distributed into more buckets, the chain length + * (item count) in each bucket is reduced. Thus by expanding buckets + * the hash keeps a bound on the chain length. This bounded chain + * length is the essence of how a hash provides constant time lookup. + * + * The calculation of tbl->ideal_chain_maxlen below deserves some + * explanation. First, keep in mind that we're calculating the ideal + * maximum chain length based on the *new* (doubled) bucket count. + * In fractions this is just n/b (n=number of items,b=new num buckets). + * Since the ideal chain length is an integer, we want to calculate + * ceil(n/b). We don't depend on floating point arithmetic in this + * hash, so to calculate ceil(n/b) with integers we could write + * + * ceil(n/b) = (n/b) + ((n%b)?1:0) + * + * and in fact a previous version of this hash did just that. + * But now we have improved things a bit by recognizing that b is + * always a power of two. We keep its base 2 log handy (call it lb), + * so now we can write this with a bit shift and logical AND: + * + * ceil(n/b) = (n>>lb) + ( (n & (b-1)) ? 1:0) + * + */ +#define HASH_EXPAND_BUCKETS(tbl) \ +do { \ + unsigned _he_bkt; \ + unsigned _he_bkt_i; \ + struct UT_hash_handle *_he_thh, *_he_hh_nxt; \ + UT_hash_bucket *_he_new_buckets, *_he_newbkt; \ + _he_new_buckets = (UT_hash_bucket*)uthash_malloc( \ + 2UL * tbl->num_buckets * sizeof(struct UT_hash_bucket)); \ + if (!_he_new_buckets) { uthash_fatal( "out of memory"); } \ + memset(_he_new_buckets, 0, \ + 2UL * tbl->num_buckets * sizeof(struct UT_hash_bucket)); \ + tbl->ideal_chain_maxlen = \ + (tbl->num_items >> (tbl->log2_num_buckets+1U)) + \ + (((tbl->num_items & ((tbl->num_buckets*2U)-1U)) != 0U) ? 1U : 0U); \ + tbl->nonideal_items = 0; \ + for(_he_bkt_i = 0; _he_bkt_i < tbl->num_buckets; _he_bkt_i++) \ + { \ + _he_thh = tbl->buckets[ _he_bkt_i ].hh_head; \ + while (_he_thh != NULL) { \ + _he_hh_nxt = _he_thh->hh_next; \ + HASH_TO_BKT( _he_thh->hashv, tbl->num_buckets*2U, _he_bkt); \ + _he_newbkt = &(_he_new_buckets[ _he_bkt ]); \ + if (++(_he_newbkt->count) > tbl->ideal_chain_maxlen) { \ + tbl->nonideal_items++; \ + _he_newbkt->expand_mult = _he_newbkt->count / \ + tbl->ideal_chain_maxlen; \ + } \ + _he_thh->hh_prev = NULL; \ + _he_thh->hh_next = _he_newbkt->hh_head; \ + if (_he_newbkt->hh_head != NULL) { _he_newbkt->hh_head->hh_prev = \ + _he_thh; } \ + _he_newbkt->hh_head = _he_thh; \ + _he_thh = _he_hh_nxt; \ + } \ + } \ + uthash_free( tbl->buckets, tbl->num_buckets*sizeof(struct UT_hash_bucket) ); \ + tbl->num_buckets *= 2U; \ + tbl->log2_num_buckets++; \ + tbl->buckets = _he_new_buckets; \ + tbl->ineff_expands = (tbl->nonideal_items > (tbl->num_items >> 1)) ? \ + (tbl->ineff_expands+1U) : 0U; \ + if (tbl->ineff_expands > 1U) { \ + tbl->noexpand=1; \ + uthash_noexpand_fyi(tbl); \ + } \ + uthash_expand_fyi(tbl); \ +} while (0) + + +/* This is an adaptation of Simon Tatham's O(n log(n)) mergesort */ +/* Note that HASH_SORT assumes the hash handle name to be hh. + * HASH_SRT was added to allow the hash handle name to be passed in. */ +#define HASH_SORT(head,cmpfcn) HASH_SRT(hh,head,cmpfcn) +#define HASH_SRT(hh,head,cmpfcn) \ +do { \ + unsigned _hs_i; \ + unsigned _hs_looping,_hs_nmerges,_hs_insize,_hs_psize,_hs_qsize; \ + struct UT_hash_handle *_hs_p, *_hs_q, *_hs_e, *_hs_list, *_hs_tail; \ + if (head != NULL) { \ + _hs_insize = 1; \ + _hs_looping = 1; \ + _hs_list = &((head)->hh); \ + while (_hs_looping != 0U) { \ + _hs_p = _hs_list; \ + _hs_list = NULL; \ + _hs_tail = NULL; \ + _hs_nmerges = 0; \ + while (_hs_p != NULL) { \ + _hs_nmerges++; \ + _hs_q = _hs_p; \ + _hs_psize = 0; \ + for ( _hs_i = 0; _hs_i < _hs_insize; _hs_i++ ) { \ + _hs_psize++; \ + _hs_q = (UT_hash_handle*)((_hs_q->next != NULL) ? \ + ((void*)((char*)(_hs_q->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + if (! (_hs_q) ) { break; } \ + } \ + _hs_qsize = _hs_insize; \ + while ((_hs_psize > 0U) || ((_hs_qsize > 0U) && (_hs_q != NULL))) {\ + if (_hs_psize == 0U) { \ + _hs_e = _hs_q; \ + _hs_q = (UT_hash_handle*)((_hs_q->next != NULL) ? \ + ((void*)((char*)(_hs_q->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + _hs_qsize--; \ + } else if ( (_hs_qsize == 0U) || (_hs_q == NULL) ) { \ + _hs_e = _hs_p; \ + if (_hs_p != NULL){ \ + _hs_p = (UT_hash_handle*)((_hs_p->next != NULL) ? \ + ((void*)((char*)(_hs_p->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + } \ + _hs_psize--; \ + } else if (( \ + cmpfcn(DECLTYPE(head)(ELMT_FROM_HH((head)->hh.tbl,_hs_p)), \ + DECLTYPE(head)(ELMT_FROM_HH((head)->hh.tbl,_hs_q))) \ + ) <= 0) { \ + _hs_e = _hs_p; \ + if (_hs_p != NULL){ \ + _hs_p = (UT_hash_handle*)((_hs_p->next != NULL) ? \ + ((void*)((char*)(_hs_p->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + } \ + _hs_psize--; \ + } else { \ + _hs_e = _hs_q; \ + _hs_q = (UT_hash_handle*)((_hs_q->next != NULL) ? \ + ((void*)((char*)(_hs_q->next) + \ + (head)->hh.tbl->hho)) : NULL); \ + _hs_qsize--; \ + } \ + if ( _hs_tail != NULL ) { \ + _hs_tail->next = ((_hs_e != NULL) ? \ + ELMT_FROM_HH((head)->hh.tbl,_hs_e) : NULL); \ + } else { \ + _hs_list = _hs_e; \ + } \ + if (_hs_e != NULL) { \ + _hs_e->prev = ((_hs_tail != NULL) ? \ + ELMT_FROM_HH((head)->hh.tbl,_hs_tail) : NULL); \ + } \ + _hs_tail = _hs_e; \ + } \ + _hs_p = _hs_q; \ + } \ + if (_hs_tail != NULL){ \ + _hs_tail->next = NULL; \ + } \ + if ( _hs_nmerges <= 1U ) { \ + _hs_looping=0; \ + (head)->hh.tbl->tail = _hs_tail; \ + DECLTYPE_ASSIGN(head,ELMT_FROM_HH((head)->hh.tbl, _hs_list)); \ + } \ + _hs_insize *= 2U; \ + } \ + HASH_FSCK(hh,head); \ + } \ +} while (0) + +/* This function selects items from one hash into another hash. + * The end result is that the selected items have dual presence + * in both hashes. There is no copy of the items made; rather + * they are added into the new hash through a secondary hash + * hash handle that must be present in the structure. */ +#define HASH_SELECT(hh_dst, dst, hh_src, src, cond) \ +do { \ + unsigned _src_bkt, _dst_bkt; \ + void *_last_elt=NULL, *_elt; \ + UT_hash_handle *_src_hh, *_dst_hh, *_last_elt_hh=NULL; \ + ptrdiff_t _dst_hho = ((char*)(&(dst)->hh_dst) - (char*)(dst)); \ + if (src != NULL) { \ + for(_src_bkt=0; _src_bkt < (src)->hh_src.tbl->num_buckets; _src_bkt++) { \ + for(_src_hh = (src)->hh_src.tbl->buckets[_src_bkt].hh_head; \ + _src_hh != NULL; \ + _src_hh = _src_hh->hh_next) { \ + _elt = ELMT_FROM_HH((src)->hh_src.tbl, _src_hh); \ + if (cond(_elt)) { \ + _dst_hh = (UT_hash_handle*)(((char*)_elt) + _dst_hho); \ + _dst_hh->key = _src_hh->key; \ + _dst_hh->keylen = _src_hh->keylen; \ + _dst_hh->hashv = _src_hh->hashv; \ + _dst_hh->prev = _last_elt; \ + _dst_hh->next = NULL; \ + if (_last_elt_hh != NULL) { _last_elt_hh->next = _elt; } \ + if (dst == NULL) { \ + DECLTYPE_ASSIGN(dst,_elt); \ + HASH_MAKE_TABLE(hh_dst,dst); \ + } else { \ + _dst_hh->tbl = (dst)->hh_dst.tbl; \ + } \ + HASH_TO_BKT(_dst_hh->hashv, _dst_hh->tbl->num_buckets, _dst_bkt); \ + HASH_ADD_TO_BKT(_dst_hh->tbl->buckets[_dst_bkt],_dst_hh); \ + (dst)->hh_dst.tbl->num_items++; \ + _last_elt = _elt; \ + _last_elt_hh = _dst_hh; \ + } \ + } \ + } \ + } \ + HASH_FSCK(hh_dst,dst); \ +} while (0) + +#define HASH_CLEAR(hh,head) \ +do { \ + if (head != NULL) { \ + uthash_free((head)->hh.tbl->buckets, \ + (head)->hh.tbl->num_buckets*sizeof(struct UT_hash_bucket)); \ + HASH_BLOOM_FREE((head)->hh.tbl); \ + uthash_free((head)->hh.tbl, sizeof(UT_hash_table)); \ + (head)=NULL; \ + } \ +} while (0) + +#define HASH_OVERHEAD(hh,head) \ + ((head != NULL) ? ( \ + (size_t)(((head)->hh.tbl->num_items * sizeof(UT_hash_handle)) + \ + ((head)->hh.tbl->num_buckets * sizeof(UT_hash_bucket)) + \ + sizeof(UT_hash_table) + \ + (HASH_BLOOM_BYTELEN))) : 0U) + +#ifdef NO_DECLTYPE +#define HASH_ITER(hh,head,el,tmp) \ +for(((el)=(head)), ((*(char**)(&(tmp)))=(char*)((head!=NULL)?(head)->hh.next:NULL)); \ + (el) != NULL; ((el)=(tmp)), ((*(char**)(&(tmp)))=(char*)((tmp!=NULL)?(tmp)->hh.next:NULL))) +#else +#define HASH_ITER(hh,head,el,tmp) \ +for(((el)=(head)), ((tmp)=DECLTYPE(el)((head!=NULL)?(head)->hh.next:NULL)); \ + (el) != NULL; ((el)=(tmp)), ((tmp)=DECLTYPE(el)((tmp!=NULL)?(tmp)->hh.next:NULL))) +#endif + +/* obtain a count of items in the hash */ +#define HASH_COUNT(head) HASH_CNT(hh,head) +#define HASH_CNT(hh,head) ((head != NULL)?((head)->hh.tbl->num_items):0U) + +typedef struct UT_hash_bucket { + struct UT_hash_handle *hh_head; + unsigned count; + + /* expand_mult is normally set to 0. In this situation, the max chain length + * threshold is enforced at its default value, HASH_BKT_CAPACITY_THRESH. (If + * the bucket's chain exceeds this length, bucket expansion is triggered). + * However, setting expand_mult to a non-zero value delays bucket expansion + * (that would be triggered by additions to this particular bucket) + * until its chain length reaches a *multiple* of HASH_BKT_CAPACITY_THRESH. + * (The multiplier is simply expand_mult+1). The whole idea of this + * multiplier is to reduce bucket expansions, since they are expensive, in + * situations where we know that a particular bucket tends to be overused. + * It is better to let its chain length grow to a longer yet-still-bounded + * value, than to do an O(n) bucket expansion too often. + */ + unsigned expand_mult; + +} UT_hash_bucket; + +/* random signature used only to find hash tables in external analysis */ +#define HASH_SIGNATURE 0xa0111fe1u +#define HASH_BLOOM_SIGNATURE 0xb12220f2u + +typedef struct UT_hash_table { + UT_hash_bucket *buckets; + unsigned num_buckets, log2_num_buckets; + unsigned num_items; + struct UT_hash_handle *tail; /* tail hh in app order, for fast append */ + ptrdiff_t hho; /* hash handle offset (byte pos of hash handle in element */ + + /* in an ideal situation (all buckets used equally), no bucket would have + * more than ceil(#items/#buckets) items. that's the ideal chain length. */ + unsigned ideal_chain_maxlen; + + /* nonideal_items is the number of items in the hash whose chain position + * exceeds the ideal chain maxlen. these items pay the penalty for an uneven + * hash distribution; reaching them in a chain traversal takes >ideal steps */ + unsigned nonideal_items; + + /* ineffective expands occur when a bucket doubling was performed, but + * afterward, more than half the items in the hash had nonideal chain + * positions. If this happens on two consecutive expansions we inhibit any + * further expansion, as it's not helping; this happens when the hash + * function isn't a good fit for the key domain. When expansion is inhibited + * the hash will still work, albeit no longer in constant time. */ + unsigned ineff_expands, noexpand; + + uint32_t signature; /* used only to find hash tables in external analysis */ +#ifdef HASH_BLOOM + uint32_t bloom_sig; /* used only to test bloom exists in external analysis */ + uint8_t *bloom_bv; + uint8_t bloom_nbits; +#endif + +} UT_hash_table; + +typedef struct UT_hash_handle { + struct UT_hash_table *tbl; + void *prev; /* prev element in app order */ + void *next; /* next element in app order */ + struct UT_hash_handle *hh_prev; /* previous hh in bucket order */ + struct UT_hash_handle *hh_next; /* next hh in bucket order */ + void *key; /* ptr to enclosing struct's key */ + unsigned keylen; /* enclosing struct's key len */ + unsigned hashv; /* result of hash-fcn(key) */ +} UT_hash_handle; + +#endif /* UTHASH_H */ diff --git a/test/performance/shmem_perf_suite/Makefile.am b/test/performance/shmem_perf_suite/Makefile.am index 38e0866..104a67e 100644 --- a/test/performance/shmem_perf_suite/Makefile.am +++ b/test/performance/shmem_perf_suite/Makefile.am @@ -38,7 +38,9 @@ noinst_HEADERS = \ int_element_latency.h \ bw_common.h \ uni_dir.h \ + uni_dir_ctx.h \ bi_dir.h \ + bi_dir_ctx.h \ target_put.h if ENABLE_LENGTHY_TESTS diff --git a/test/performance/shmem_perf_suite/bi_dir.h b/test/performance/shmem_perf_suite/bi_dir.h index fda8119..d46437e 100644 --- a/test/performance/shmem_perf_suite/bi_dir.h +++ b/test/performance/shmem_perf_suite/bi_dir.h @@ -29,7 +29,21 @@ void static inline bi_bw_put(int len, perf_metrics_t *metric_info) { double start = 0.0, end = 0.0; int dest = partner_node(*metric_info); - int i = 0, j = 0; + unsigned long int i = 0, j = 0; + static int check_once = 0; + static int fin = -1; + + if (!check_once) { + /* check to see whether sender and receiver are the same process */ + if (dest == metric_info->my_node) { + fprintf(stderr, "Warning: Sender and receiver are the same process (%d)\n", + dest); + } + /* hostname validation for all sender and receiver processes */ + int status = check_hostname_validation(*metric_info); + if (status != 0) return; + check_once++; + } shmem_barrier_all(); @@ -60,18 +74,37 @@ void static inline bi_bw_put(int len, perf_metrics_t *metric_info) shmem_quiet(); } - shmem_barrier_all(); if (streaming_node(*metric_info)) { + shmem_int_p(&fin, 1, dest); + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 0); end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, *metric_info); + calc_and_print_results(end, start, len, *metric_info); + } else { + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 1); + shmem_int_p(&fin, 0, dest); } + } void static inline bi_bw_get(int len, perf_metrics_t *metric_info) { double start = 0.0, end = 0.0; int dest = partner_node(*metric_info); - int i = 0, j = 0; + unsigned long int i = 0, j = 0; + static int check_once = 0; + static int fin = -1; + + if (!check_once) { + /* check to see whether sender and receiver are the same process */ + if (dest == metric_info->my_node) { + fprintf(stderr, "Warning: Sender and receiver are the same process (%d)\n", + dest); + } + /* hostname validation for all sender and receiver processes */ + int status = check_hostname_validation(*metric_info); + if (status != 0) return; + check_once++; + } shmem_barrier_all(); @@ -112,10 +145,14 @@ void static inline bi_bw_get(int len, perf_metrics_t *metric_info) #endif } - shmem_barrier_all(); if (streaming_node(*metric_info)) { + shmem_int_p(&fin, 1, dest); + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 0); end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, *metric_info); + calc_and_print_results(end, start, len, *metric_info); + } else { + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 1); + shmem_int_p(&fin, 0, dest); } } diff --git a/test/performance/shmem_perf_suite/bi_dir_ctx.h b/test/performance/shmem_perf_suite/bi_dir_ctx.h index 7043b38..8b9fe10 100644 --- a/test/performance/shmem_perf_suite/bi_dir_ctx.h +++ b/test/performance/shmem_perf_suite/bi_dir_ctx.h @@ -34,6 +34,19 @@ void static inline bi_bw_ctx (int len, perf_metrics_t *metric_info) char *src = aligned_buffer_alloc(metric_info->nthreads * len); char *dst = aligned_buffer_alloc(metric_info->nthreads * len); assert(src && dst); + static int check_once = 0; + + if (!check_once) { + /* check to see whether sender and receiver are the same process */ + if (dest == metric_info->my_node) { + fprintf(stderr, "Warning: Sender and receiver are the same process (%d)\n", + dest); + } + /* hostname validation for all sender and receiver processes */ + int status = check_hostname_validation(*metric_info); + if (status != 0) return; + check_once++; + } shmem_barrier_all(); @@ -111,7 +124,7 @@ void static inline bi_bw_ctx (int len, perf_metrics_t *metric_info) shmem_barrier_all(); if (streaming_node(*metric_info)) { end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, *metric_info); + calc_and_print_results(end, start, len, *metric_info); } shmem_barrier_all(); diff --git a/test/performance/shmem_perf_suite/bw_common.h b/test/performance/shmem_perf_suite/bw_common.h index 12beef8..847c6ec 100644 --- a/test/performance/shmem_perf_suite/bw_common.h +++ b/test/performance/shmem_perf_suite/bw_common.h @@ -98,9 +98,11 @@ typedef struct perf_metrics { bw_style bwstyle; int thread_safety; int nthreads; + int individual_report; } perf_metrics_t; long red_psync[SHMEM_REDUCE_SYNC_SIZE]; +long bar_psync[SHMEM_BARRIER_SYNC_SIZE]; /*default settings if no input is provided */ void static data_set_defaults(perf_metrics_t * data) { @@ -124,6 +126,7 @@ void static data_set_defaults(perf_metrics_t * data) { data->bwstyle = STYLE_RMA; data->thread_safety = SHMEM_THREAD_SINGLE; data->nthreads = 1; + data->individual_report = -1; } static int error_checking_init_target_usage(perf_metrics_t *metric_info) { @@ -168,12 +171,12 @@ static int data_runtime_update(perf_metrics_t *data) { static const char * dt_names [] = { "int", "long", "longlong" }; void static bi_dir_data_init(perf_metrics_t * data) { - data->bw_type = "Bi-directional Bandwidth"; + data->bw_type = "Bi-dir"; data->type = BI_DIR; } void static uni_dir_data_init(perf_metrics_t * data) { - data->bw_type = "Uni-directional Bandwidth"; + data->bw_type = "Uni-dir"; data->type = UNI_DIR; } @@ -204,6 +207,16 @@ int static inline streaming_node(perf_metrics_t my_info) } } +static int inline is_streaming_node(perf_metrics_t my_info, int node) +{ + if(my_info.cstyle == COMM_PAIRWISE) { + return (node < my_info.szinitiator); + } else { + assert(my_info.cstyle == COMM_INCAST); + return true; + } +} + int static inline target_node(perf_metrics_t my_info) { return (my_info.my_node >= my_info.midpt && @@ -242,7 +255,7 @@ static int command_line_arg_check(int argc, char *argv[], extern char *optarg; /* check command line args */ - while ((ch = getopt(argc, argv, "e:s:n:w:p:r:l:kbvtC:T:")) != EOF) { + while ((ch = getopt(argc, argv, "e:s:n:w:p:r:l:kbivtC:T:")) != EOF) { switch (ch) { case 's': metric_info->start_len = strtoul(optarg, (char **)NULL, 0); @@ -251,6 +264,10 @@ static int command_line_arg_check(int argc, char *argv[], fprintf(stderr, "Error: start_length must be a power of two\n"); error = true; } + if (metric_info->start_len > INT_MAX) { + fprintf(stderr, "Error: start_length is out of integer range\n"); + error = true; + } break; case 'e': metric_info->max_len = strtoul(optarg, (char **)NULL, 0); @@ -264,6 +281,10 @@ static int command_line_arg_check(int argc, char *argv[], metric_info->start_len); error = true; } + if (metric_info->max_len > INT_MAX) { + fprintf(stderr, "Error: end_length is out of integer range\n"); + error = true; + } break; case 'n': metric_info->trials = strtoul(optarg, (char **)NULL, 0); @@ -323,6 +344,9 @@ static int command_line_arg_check(int argc, char *argv[], case 'T': metric_info->nthreads = atoi(optarg); break; + case 'i': + metric_info->individual_report = 1; + break; default: error = true; break; @@ -349,6 +373,7 @@ static int command_line_arg_check(int argc, char *argv[], "[-w window size - iterations between completion, cannot use with -t] \n" "[-k (kilobytes/second)] [-b (bytes/second)] \n" "[-v (validate data stream)] \n" + "[-i (turn on individual bandwidth reporting)] \n" "[-t output data for target side (default is initiator," " only use with put_bw),\n cannot be used in conjunction " "with validate, special sizes used, \ntrials" @@ -411,54 +436,60 @@ static void inline thread_safety_validation_check(perf_metrics_t *metric_info) { } void static print_atomic_results_header(perf_metrics_t metric_info) { - printf("\nResults for %d PEs %lu trials with window size %lu ", - metric_info.num_pes, metric_info.trials, metric_info.window_size); + printf("\nSandia OpenSHMEM Performance Suite\n"); + printf("==================================\n"); + printf("Total Number of PEs: %10d\n", metric_info.num_pes); + printf("Iteration count: %10lu\n", metric_info.trials); + printf("Window size: %10lu\n", metric_info.window_size); + printf("Bandwidth test type: %10s\n", metric_info.bw_type); if (metric_info.cstyle == COMM_INCAST) { - printf("using incast communication style\n"); + printf("Communication style: INCAST\n"); } else { assert(metric_info.cstyle == COMM_PAIRWISE); - printf("using pairwise communication style\n"); + printf("Communication style: PAIRWISE\n"); } - printf("\nOperation %s " - "Message Rate%17sLatency\n", metric_info.bw_type, " "); + printf("\nOperation%15sBandwidth%15sMessage Rate%15sLatency\n", + " ", " ", " "); if (metric_info.unit == MB) { - printf("%19s in megabytes per second"," "); + printf("%19s in mbytes/sec"," "); } else if (metric_info.unit == KB) { - printf("%19s in kilobytes per second", " "); + printf("%19s in kbytes/sec", " "); } else { - printf("%19s in bytes per second", " "); + printf("%20s in bytes/sec", " "); } - printf(" in Million ops/second%8sin microseconds\n", " "); - - /* hack */ - printf("shmem_add\n"); + printf("%15s in Mops/sec%15s in us\n", " ", " "); } void static print_results_header(perf_metrics_t metric_info) { - printf("\nResults for %d PEs %lu trials with window size %lu " - "max message size %lu with multiple of %lu increments, " - "\ntargeting %d remote PEs initiated from %d PEs", metric_info.num_pes, - metric_info.trials, metric_info.window_size, metric_info.max_len, - metric_info.size_inc, metric_info.sztarget, metric_info.szinitiator); - printf(", thread safety %s (%d threads)\n", - thread_safety_str(&metric_info), metric_info.nthreads); - printf("\nLength %s " - "Message Rate\n", metric_info.bw_type); - - printf("in bytes "); + printf("\nSandia OpenSHMEM Performance Suite\n"); + printf("==================================\n"); + printf("Total Number of PEs: %10d\n", metric_info.num_pes); + printf("Number of source PEs: %10d\n", metric_info.szinitiator); + printf("Number of target PEs: %10d\n", metric_info.sztarget); + printf("Iteration count: %10lu\n", metric_info.trials); + printf("Window size: %10lu\n", metric_info.window_size); + printf("Maximum message size: %10lu\n", metric_info.max_len); + printf("Number of threads: %10d\n", metric_info.nthreads); + printf("Thread safety: %10s\n", thread_safety_str(&metric_info)); + printf("Bandwidth test type: %10s\n", metric_info.bw_type); + + printf("\nMessage Size%15sBandwidth%15sMessage Rate\n", + " ", " "); + + printf("%4sin bytes", " "); if (metric_info.unit == MB) { - printf("in megabytes per second"); + printf("%11sin mbytes/sec", " "); } else if (metric_info.unit == KB) { - printf("in kilobytes per second"); + printf("%11sin kbytes/sec", " "); } else { - printf("in bytes per second"); + printf("%12sin bytes/sec", " "); } - printf(" in messages/seconds\n"); + printf("%16sin msgs/sec\n", " "); } void static print_data_results(double bw, double mr, perf_metrics_t data, @@ -474,10 +505,10 @@ void static print_data_results(double bw, double mr, perf_metrics_t data, } if (data.bwstyle == STYLE_ATOMIC) { - printf("%-10s ", dt_names[atomic_type_index]); + printf("%-10s", dt_names[atomic_type_index]); atomic_type_index = (atomic_type_index + 1) % ATOMICS_N_DTs; } else - printf("%9d ", len); + printf("%2s%10d", " ", len); if(data.unit == KB) { bw = bw * 1.0e3; @@ -486,10 +517,10 @@ void static print_data_results(double bw, double mr, perf_metrics_t data, } if (data.bwstyle == STYLE_ATOMIC) { - printf("%5s%10.2f %10.2f%14s%10.2f\n", " ", bw, - mr/1.0e6, " ", total_t/(data.trials * data.window_size)); + printf("%13s%10.2f%15s%12.2f%12s%10.2f\n", " ", bw, " ", + mr/1.0e6, " ", total_t/(data.trials * data.window_size)); } else - printf("%10.2f %10.2f\n", bw, mr); + printf("%14s%10.2f%15s%12.2f\n", " ", bw, " ", mr); } @@ -512,7 +543,7 @@ void static inline PE_set_used_adjustments(int *nPEs, int *stride, int *start_pe } -void static inline calc_and_print_results(double total_t, int len, +void static inline calc_and_print_results(double end_t, double start_t, int len, perf_metrics_t metric_info) { int stride = 0, start_pe = 0, nPEs = 0; @@ -520,31 +551,74 @@ void static inline calc_and_print_results(double total_t, int len, double pe_bw_avg = 0.0, pe_mr_avg = 0.0; int nred_elements = 1; static double pwrk[SHMEM_REDUCE_MIN_WRKDATA_SIZE]; + static double pe_time_start, pe_time_end, end_time_max = 0.0, start_time_min = 0.0; + double total_t = 0.0, total_t_max = 0.0; + int multiplier = 1; PE_set_used_adjustments(&nPEs, &stride, &start_pe, metric_info); - if (total_t > 0 ) { + /* 2x as many messages at once for bi-directional */ + if(metric_info.type == BI_DIR) + multiplier = 2; + if (end_t > 0 && start_t > 0 && (end_t - start_t) > 0) { + total_t = end_t - start_t; #ifdef ENABLE_OPENMP - bw = (len / 1.0e6 * metric_info.window_size * metric_info.trials * - (double)metric_info.nthreads) / (total_t / 1.0e6); + bw = ((double) len * (double) multiplier / 1.0e6 * metric_info.window_size * metric_info.trials * + (double) metric_info.nthreads) / (total_t / 1.0e6); #else - bw = (len / 1.0e6 * metric_info.window_size * metric_info.trials) / + bw = ((double) len * (double) multiplier / 1.0e6 * metric_info.window_size * metric_info.trials) / (total_t / 1.0e6); #endif + } else { + fprintf(stderr, "Incorrect time measured from bandwidth test: " + "start = %lf, end = %lf\n", start_t, end_t); } - /* 2x as many messages/bytes at once for bi-directional */ - if(metric_info.type == BI_DIR) - bw *= 2.0; - /* base case: will be overwritten by collective if num_pes > 2 */ pe_bw_sum = bw; - if(nPEs >= 2) - shmem_double_sum_to_all(&pe_bw_sum, &bw, nred_elements, start_pe, - stride, nPEs, pwrk, + if (metric_info.individual_report == 1) { + printf("Individual bandwith for PE %6d is %10.2f\n", + metric_info.my_node, pe_bw_sum); + } + + pe_time_start = start_t; + pe_time_end = end_t; + shmem_barrier(start_pe, stride, nPEs, bar_psync); + if (nPEs >= 2) { + shmem_double_min_to_all(&start_time_min, &pe_time_start, nred_elements, + start_pe, stride, nPEs, pwrk, + red_psync); + shmem_barrier(start_pe, stride, nPEs, bar_psync); + shmem_double_max_to_all(&end_time_max, &pe_time_end, nred_elements, + start_pe, stride, nPEs, pwrk, red_psync); + } else if (nPEs == 1) { + start_time_min = pe_time_start; + end_time_max = pe_time_end; + } + + /* calculating bandwidth based on the highest time duration across all PEs */ + if (end_time_max > 0 && start_time_min > 0 && + (end_time_max - start_time_min) > 0) { + + total_t_max = (end_time_max - start_time_min); +#ifdef ENABLE_OPENMP + bw = ((double) len * (double) multiplier * (double) metric_info.midpt / 1.0e6 * metric_info.window_size * + metric_info.trials * (double) metric_info.nthreads) / + (total_t_max / 1.0e6); +#else + bw = ((double) len * (double) multiplier * (double) metric_info.midpt / 1.0e6 * metric_info.window_size * + metric_info.trials) / (total_t_max / 1.0e6); +#endif + } else { + fprintf(stderr, "Incorrect time measured from bandwidth test: " + "start_min = %lf, end_max = %lf\n", + start_time_min, end_time_max); + } + + pe_bw_sum = bw; /* aggregate bw since bw op pairs are communicating simultaneously */ if(metric_info.my_node == start_pe) { @@ -566,7 +640,7 @@ static void validate_atomics(perf_metrics_t m_info) { int snode = streaming_node(m_info); int * my_buf = (int *)m_info.dest; bw_type tbw = m_info.type; - unsigned int expected_val = 0; + int expected_val = 0; unsigned int ppe_exp_val = ((m_info.trials + m_info.warmup) * m_info.window_size * ATOMICS_N_DTs * ATOMICS_N_OPs) + m_info.my_node; @@ -600,7 +674,8 @@ static void validate_atomics(perf_metrics_t m_info) { extern void bi_dir_bw(int len, perf_metrics_t *metric_info); void static inline bi_dir_bw_test_and_output(perf_metrics_t metric_info) { - int len = 0, partner_pe = partner_node(metric_info); + int partner_pe = partner_node(metric_info); + unsigned long int len; if(metric_info.my_node == 0) { if (metric_info.bwstyle == STYLE_ATOMIC) @@ -638,7 +713,8 @@ void static inline bi_dir_bw_test_and_output(perf_metrics_t metric_info) { extern void uni_dir_bw(int len, perf_metrics_t *metric_info); void static inline uni_dir_bw_test_and_output(perf_metrics_t metric_info) { - int len = 0, partner_pe = partner_node(metric_info); + int partner_pe = partner_node(metric_info); + unsigned long int len = 0; if(metric_info.my_node == 0) { if (metric_info.bwstyle == STYLE_ATOMIC) @@ -700,9 +776,12 @@ static inline int bw_init_data_stream(perf_metrics_t *metric_info, metric_info->sztarget = metric_info->midpt; metric_info->szinitiator = metric_info->midpt; - for(i = 0; i < SHMEM_REDUCE_MIN_WRKDATA_SIZE; i++) + for(i = 0; i < SHMEM_REDUCE_SYNC_SIZE; i++) red_psync[i] = SHMEM_SYNC_VALUE; + for(i = 0; i < SHMEM_BARRIER_SYNC_SIZE; i++) + bar_psync[i] = SHMEM_SYNC_VALUE; + if (only_even_PEs_check(metric_info->my_node, metric_info->num_pes) != 0) { return -2; } @@ -781,3 +860,74 @@ void static inline uni_dir_bw_main(int argc, char *argv[], bw_style bwstyl) { if (ret != -1) bw_finalize(); } /*main() */ + +static inline int check_hostname_validation(perf_metrics_t my_info) { + + int hostname_status = -1; + + /* hostname_size should be a length divisible by 4 */ + int hostname_size = (MAX_HOSTNAME_LEN % 4 == 0) ? MAX_HOSTNAME_LEN : + MAX_HOSTNAME_LEN + (4 - MAX_HOSTNAME_LEN % 4); + int i, errors = 0; + + /* pSync for fcollect of hostnames */ + static long pSync_collect[SHMEM_COLLECT_SYNC_SIZE]; + for (i = 0; i < SHMEM_COLLECT_SYNC_SIZE; i++) + pSync_collect[i] = SHMEM_SYNC_VALUE; + + char *hostname = (char *) shmem_malloc (hostname_size * sizeof(char)); + char *dest = (char *) shmem_malloc (my_info.num_pes * hostname_size * sizeof(char)); + + hostname_status = gethostname(hostname, hostname_size); + if (hostname_status != 0) { + fprintf(stderr, "gethostname failed (%d)\n", hostname_status); + return -1; + } + shmem_barrier_all(); + + /* nelems needs to be updated based on 32-bit API */ + shmem_fcollect32(dest, hostname, hostname_size/4, 0, 0, my_info.num_pes, pSync_collect); + + char *snode_name = NULL; + char *tnode_name = NULL; + for (i = 0; i < my_info.num_pes; i++) { + char *curr_name = &dest[i * hostname_size]; + + if (is_streaming_node(my_info, i)) { + if (snode_name == NULL) { + snode_name = curr_name; + } + + if (strncmp(snode_name, curr_name, hostname_size) != 0) { + fprintf(stderr, "PE %d on %s is a streaming node " + "but not placed on %s\n", i, curr_name, snode_name); + errors++; + } + } else { + if (tnode_name == NULL) { + tnode_name = curr_name; + } + + if (strncmp(tnode_name, curr_name, hostname_size) != 0) { + fprintf(stderr, "PE %d on %s is a target node " + "but not placed on %s\n", i, curr_name, tnode_name); + errors++; + } + } + } + + if (snode_name == NULL || tnode_name == NULL) { + fprintf(stderr, "Error: no streaming or target node\n"); + return -1; + } + + if (strncmp(snode_name, tnode_name, hostname_size) == 0) { + fprintf(stderr, "Warning: senders and receivers are running on the " + "same node %s\n", snode_name); + } + + shmem_free(dest); + shmem_free(hostname); + + return errors; +} diff --git a/test/performance/shmem_perf_suite/common.h b/test/performance/shmem_perf_suite/common.h index 4c444ba..c48f66e 100644 --- a/test/performance/shmem_perf_suite/common.h +++ b/test/performance/shmem_perf_suite/common.h @@ -36,6 +36,15 @@ #include #include +#include +#include + +#ifdef MAXHOSTNAMELEN +#define MAX_HOSTNAME_LEN MAXHOSTNAMELEN +#else +#define MAX_HOSTNAME_LEN HOST_NAME_MAX +#endif + #define ONE 1 /* return microseconds */ diff --git a/test/performance/shmem_perf_suite/int_element_latency.h b/test/performance/shmem_perf_suite/int_element_latency.h index df9e8d6..45e857e 100644 --- a/test/performance/shmem_perf_suite/int_element_latency.h +++ b/test/performance/shmem_perf_suite/int_element_latency.h @@ -30,7 +30,7 @@ int_p_latency(perf_metrics_t data) { double start = 0.0; double end = 0.0; - int i = 0; + unsigned int i = 0; if (data.my_node == PUT_IO_NODE) { printf("\nStream shmem_int_p results:\n"); @@ -65,7 +65,7 @@ int_g_latency(perf_metrics_t data) { double start = 0.0; double end = 0.0; - int i = 0; + unsigned int i = 0; int rtnd = -1; if (data.my_node == GET_IO_NODE) { diff --git a/test/performance/shmem_perf_suite/latency_common.h b/test/performance/shmem_perf_suite/latency_common.h index 6f2ae80..482b47c 100644 --- a/test/performance/shmem_perf_suite/latency_common.h +++ b/test/performance/shmem_perf_suite/latency_common.h @@ -153,7 +153,7 @@ extern void int_element_latency(perf_metrics_t data); extern void streaming_latency(int len, perf_metrics_t *data); void static inline multi_size_latency(perf_metrics_t data, char *argv[]) { - int len; + unsigned int len; int partner_pe = partner_node(data.my_node); for (len = data.start_len; len <= data.max_len; len *= data.inc) { diff --git a/test/performance/shmem_perf_suite/round_t_latency.h b/test/performance/shmem_perf_suite/round_t_latency.h index 7552223..6b3d86e 100644 --- a/test/performance/shmem_perf_suite/round_t_latency.h +++ b/test/performance/shmem_perf_suite/round_t_latency.h @@ -30,7 +30,7 @@ long_element_round_trip_latency_get(perf_metrics_t data) { double start = 0.0; double end = 0.0; - int dest = 1, i = 0; + int dest = 1; int partner_pe = partner_node(data.my_node); *data.target = data.my_node; @@ -42,6 +42,7 @@ long_element_round_trip_latency_get(perf_metrics_t data) shmem_barrier_all(); if (data.my_node == GET_IO_NODE) { + unsigned int i; for (i = 0; i < data.trials + data.warmup; i++) { if(i == data.warmup) start = perf_shmemx_wtime(); @@ -66,7 +67,8 @@ long_element_round_trip_latency_put(perf_metrics_t data) double start = 0.0; double end = 0.0; long tmp; - int dest = (data.my_node + 1) % data.npes, i = 0; + int dest = (data.my_node + 1) % data.npes; + unsigned int i; tmp = *data.target = INIT_VALUE; if (data.my_node == PUT_IO_NODE) { diff --git a/test/performance/shmem_perf_suite/shmem_bibw_atomics_perf.c b/test/performance/shmem_perf_suite/shmem_bibw_atomics_perf.c index b0ad0bd..f287646 100644 --- a/test/performance/shmem_perf_suite/shmem_bibw_atomics_perf.c +++ b/test/performance/shmem_perf_suite/shmem_bibw_atomics_perf.c @@ -41,7 +41,7 @@ #define bi_bw(len, metric_info, snode, NAME, TYPE, op) \ do { \ double start = 0.0, end = 0.0; \ - int i = 0, j = 0, num_itr = metric_info->trials + metric_info->warmup; \ + unsigned long int i = 0, j = 0, num_itr = metric_info->trials + metric_info->warmup; \ int dest = partner_node(*metric_info); \ shmem_barrier_all(); \ \ @@ -128,7 +128,7 @@ break; \ } \ if(snode) \ - calc_and_print_results((end - start), len, *metric_info); \ + calc_and_print_results(end, start, len, *metric_info); \ } while(0) diff --git a/test/performance/shmem_perf_suite/shmem_bw_atomics_perf.c b/test/performance/shmem_perf_suite/shmem_bw_atomics_perf.c index ce3c618..79cd499 100644 --- a/test/performance/shmem_perf_suite/shmem_bw_atomics_perf.c +++ b/test/performance/shmem_perf_suite/shmem_bw_atomics_perf.c @@ -42,7 +42,7 @@ #define uni_bw(len, metric_info, snode, NAME, TYPE, op) \ do { \ double start = 0.0, end = 0.0; \ - int i = 0, j = 0, num_itr = metric_info->trials + metric_info->warmup; \ + unsigned long int i = 0, j = 0, num_itr = metric_info->trials + metric_info->warmup; \ int dest = partner_node(*metric_info); \ shmem_barrier_all(); \ \ @@ -125,7 +125,7 @@ for atomics\n", op); \ break; \ } \ - calc_and_print_results((end - start), len, *metric_info); \ + calc_and_print_results(end, start, len, *metric_info); \ } \ } while(0) @@ -154,7 +154,7 @@ static inline void bw_set_metric_info_len(perf_metrics_t *metric_info) atomic_op_type op_type = OP_ADD; for(op_type = OP_ADD; op_type < SIZE_OF_OP; op_type++) { - if(metric_info->my_node == 0 && op_type != OP_ADD) + if(metric_info->my_node == 0) printf("\nshmem_%s\n", op_names[op_type]); metric_info->start_len = atomic_sizes[0]; diff --git a/test/performance/shmem_perf_suite/shmem_latency_get_perf.c b/test/performance/shmem_perf_suite/shmem_latency_get_perf.c index 1d1a595..a7ce4a8 100644 --- a/test/performance/shmem_perf_suite/shmem_latency_get_perf.c +++ b/test/performance/shmem_perf_suite/shmem_latency_get_perf.c @@ -65,7 +65,7 @@ streaming_latency(int len, perf_metrics_t *data) { double start = 0.0; double end = 0.0; - int i = 0; + unsigned long int i = 0; static int print_once = 0; if(!print_once && data->my_node == GET_IO_NODE) { printf("\nStreaming results for %d trials each of length %d through %d in"\ diff --git a/test/performance/shmem_perf_suite/shmem_latency_nb_get_perf.c b/test/performance/shmem_perf_suite/shmem_latency_nb_get_perf.c index f636e55..7346580 100644 --- a/test/performance/shmem_perf_suite/shmem_latency_nb_get_perf.c +++ b/test/performance/shmem_perf_suite/shmem_latency_nb_get_perf.c @@ -56,7 +56,7 @@ streaming_latency(int len, perf_metrics_t *data) { double start = 0.0; double end = 0.0; - int i = 0; + unsigned long int i = 0; static int print_once = 0; if(!print_once && data->my_node == GET_IO_NODE) { printf("\nStreaming results for %d trials each of length %d through %d in"\ diff --git a/test/performance/shmem_perf_suite/shmem_latency_nb_put_perf.c b/test/performance/shmem_perf_suite/shmem_latency_nb_put_perf.c index 7138b85..c9f4c3d 100644 --- a/test/performance/shmem_perf_suite/shmem_latency_nb_put_perf.c +++ b/test/performance/shmem_perf_suite/shmem_latency_nb_put_perf.c @@ -58,7 +58,7 @@ streaming_latency(int len, perf_metrics_t *data) { double start = 0.0; double end = 0.0; - int i = 0; + unsigned long int i = 0; static int print_once = 0; if(!print_once && data->my_node == PUT_IO_NODE) { printf("\nStreaming results for %d trials each of length %d through %d in"\ diff --git a/test/performance/shmem_perf_suite/shmem_latency_put_perf.c b/test/performance/shmem_perf_suite/shmem_latency_put_perf.c index 2bf9f23..97b2bd2 100644 --- a/test/performance/shmem_perf_suite/shmem_latency_put_perf.c +++ b/test/performance/shmem_perf_suite/shmem_latency_put_perf.c @@ -65,7 +65,7 @@ streaming_latency(int len, perf_metrics_t *data) { double start = 0.0; double end = 0.0; - int i = 0; + unsigned long int i = 0; static int print_once = 0; if(!print_once && data->my_node == PUT_IO_NODE) { printf("\nStreaming results for %d trials each of length %d through %d in"\ diff --git a/test/performance/shmem_perf_suite/target_put.h b/test/performance/shmem_perf_suite/target_put.h index 984b687..223a457 100644 --- a/test/performance/shmem_perf_suite/target_put.h +++ b/test/performance/shmem_perf_suite/target_put.h @@ -73,7 +73,8 @@ int static inline *get_initiators_partners(perf_metrics_t my_info, int num_partn void static inline target_data_uni_bw(int len, perf_metrics_t metric_info) { double start = 0.0, end = 0.0; - int i = 0, j = 0; + int i = 0; + unsigned long int j = 0; int snode = (metric_info.num_pes != 1)? streaming_node(metric_info) : true; int num_partners = get_num_partners(metric_info); static int completion_signal = 0; @@ -120,7 +121,7 @@ void static inline target_data_uni_bw(int len, perf_metrics_t metric_info) shmem_barrier_all(); if (snode || target_node(metric_info)) { end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, metric_info); + calc_and_print_results(end, start, len, metric_info); } free(my_PE_partners); } diff --git a/test/performance/shmem_perf_suite/uni_dir.h b/test/performance/shmem_perf_suite/uni_dir.h index 627b2b7..c47aef5 100644 --- a/test/performance/shmem_perf_suite/uni_dir.h +++ b/test/performance/shmem_perf_suite/uni_dir.h @@ -29,9 +29,23 @@ void static inline uni_bw_put(int len, perf_metrics_t *metric_info) { double start = 0.0, end = 0.0; - int i = 0, j = 0; + unsigned long int i = 0, j = 0; int dest = partner_node(*metric_info); int snode = (metric_info->num_pes != 1)? streaming_node(*metric_info) : true; + static int check_once = 0; + static int fin = -1; + + if (!check_once) { + /* check to see whether sender and receiver are the same process */ + if (dest == metric_info->my_node) { + fprintf(stderr, "Warning: Sender and receiver are the same process (%d)\n", + dest); + } + /* hostname validation for all sender and receiver processes */ + int status = check_hostname_validation(*metric_info); + if (status != 0) return; + check_once++; + } if(metric_info->target_data) { target_bw_itr(len, metric_info); @@ -66,21 +80,36 @@ void static inline uni_bw_put(int len, perf_metrics_t *metric_info) } shmem_quiet(); } - } - - shmem_barrier_all(); - if (snode) { + shmem_int_p(&fin, 1, dest); + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 0); end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, *metric_info); + calc_and_print_results(end, start, len, *metric_info); + } else { + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 1); + shmem_int_p(&fin, 0, dest); } } void static inline uni_bw_get(int len, perf_metrics_t *metric_info) { double start = 0.0, end = 0.0; - int i = 0, j = 0; + unsigned long int i = 0, j = 0; int dest = partner_node(*metric_info); int snode = (metric_info->num_pes != 1)? streaming_node(*metric_info) : true; + static int check_once = 0; + static int fin = -1; + + if (!check_once) { + /* check to see whether sender and receiver are the same process */ + if (dest == metric_info->my_node) { + fprintf(stderr, "Warning: Sender and receiver are the same process (%d)\n", + dest); + } + /* hostname validation for all sender and receiver processes */ + int status = check_hostname_validation(*metric_info); + if (status != 0) return; + check_once++; + } if(metric_info->target_data) { target_bw_itr(len, metric_info); @@ -122,12 +151,13 @@ void static inline uni_bw_get(int len, perf_metrics_t *metric_info) shmem_quiet(); #endif } - } - - shmem_barrier_all(); - if (snode) { + shmem_int_p(&fin, 1, dest); + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 0); end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, *metric_info); + calc_and_print_results(end, start, len, *metric_info); + } else { + shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 1); + shmem_int_p(&fin, 0, dest); } } diff --git a/test/performance/shmem_perf_suite/uni_dir_ctx.h b/test/performance/shmem_perf_suite/uni_dir_ctx.h index 199815c..fd46960 100644 --- a/test/performance/shmem_perf_suite/uni_dir_ctx.h +++ b/test/performance/shmem_perf_suite/uni_dir_ctx.h @@ -35,6 +35,19 @@ void static inline uni_bw_ctx(int len, perf_metrics_t *metric_info, char *src = aligned_buffer_alloc(metric_info->nthreads * len); char *dst = aligned_buffer_alloc(metric_info->nthreads * len); assert(src && dst); + static int check_once = 0; + + if (!check_once) { + /* check to see whether sender and receiver are the same process */ + if (dest == metric_info->my_node) { + fprintf(stderr, "Warning: Sender and receiver are the same process (%d)\n", + dest); + } + /* hostname validation for all sender and receiver processes */ + int status = check_hostname_validation(*metric_info); + if (status != 0) return; + check_once++; + } shmem_barrier_all(); @@ -94,7 +107,7 @@ void static inline uni_bw_ctx(int len, perf_metrics_t *metric_info, shmem_barrier_all(); if (streaming_node) { end = perf_shmemx_wtime(); - calc_and_print_results((end - start), len, *metric_info); + calc_and_print_results(end, start, len, *metric_info); } shmem_barrier_all(); diff --git a/test/performance/tests/msgrate.c b/test/performance/tests/msgrate.c index 5400237..941475f 100644 --- a/test/performance/tests/msgrate.c +++ b/test/performance/tests/msgrate.c @@ -299,9 +299,9 @@ main(int argc, char *argv[]) shmem_barrier_all(); /* broadcast results */ - printf("%d: psync: 0x%lu\n", rank, (unsigned long) bcast_pSync); shmem_broadcast32(&start_err, &start_err, 1, 0, 0, 0, world_size, bcast_pSync); if (0 != start_err) { + shmem_finalize(); exit(start_err); } shmem_barrier_all(); diff --git a/test/unit/Makefile.am b/test/unit/Makefile.am index 6825da9..d7a7580 100644 --- a/test/unit/Makefile.am +++ b/test/unit/Makefile.am @@ -126,13 +126,13 @@ endif if HAVE_FORTRAN check_PROGRAMS += \ - hello_f \ - shmem_info_f + hello_f if !HAVE_LONG_FORTRAN_HEADER check_PROGRAMS += \ complex_reductions_f \ - set_fetch_f + set_fetch_f \ + shmem_info_f endif endif @@ -141,6 +141,7 @@ check_PROGRAMS += \ mt_a2a \ mt_contention \ mt_membar \ + mt_lock_test \ threading \ web \ thread_wait \ @@ -161,11 +162,11 @@ AM_LDFLAGS = $(LIBTOOL_WRAPPER_LDFLAGS) if EXTERNAL_TESTS bin_PROGRAMS = $(check_PROGRAMS) -AM_CPPFLAGS = +AM_CPPFLAGS = -I$(top_srcdir)/test/include AM_FCFLAGS = LDADD = else -AM_CPPFLAGS = -I$(top_builddir)/mpp +AM_CPPFLAGS = -I$(top_builddir)/mpp -I$(top_srcdir)/test/include AM_FCFLAGS = -I$(top_builddir)/mpp LDADD = $(top_builddir)/src/libsma.la endif @@ -190,6 +191,11 @@ mt_contention_LDFLAGS = $(AM_LDFLAGS) $(PTHREAD_LIBS) mt_contention_CFLAGS = $(PTHREAD_CFLAGS) mt_contention_LDADD = $(LDADD) $(PTHREAD_CFLAGS) +mt_lock_test_SOURCES = mt_lock_test.c mt_lock.c mt_lock.h +mt_lock_test_LDFLAGS = $(AM_LDFLAGS) $(PTHREAD_LIBS) +mt_lock_test_CFLAGS = $(PTHREAD_CFLAGS) +mt_lock_test_LDADD = $(LDADD) $(PTHREAD_CFLAGS) + mt_membar_LDFLAGS = $(AM_LDFLAGS) $(PTHREAD_LIBS) mt_membar_CFLAGS = -I$(top_srcdir)/test/unit $(PTHREAD_CFLAGS) mt_membar_LDADD = $(LDADD) $(PTHREAD_CFLAGS) diff --git a/test/unit/c11_test_shmem_atomic_and.c b/test/unit/c11_test_shmem_atomic_and.c index b8e9d3f..cee751d 100644 --- a/test/unit/c11_test_shmem_atomic_and.c +++ b/test/unit/c11_test_shmem_atomic_and.c @@ -30,6 +30,7 @@ * SOFTWARE. */ +#include #include #include #include @@ -51,7 +52,7 @@ enum op { AND = 0, CTX_AND, FETCH_AND, CTX_FETCH_AND }; const int mype = shmem_my_pe(); \ const int npes = shmem_n_pes(); \ TYPE old = (TYPE)0; \ - if (npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ + if ((size_t) npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ for (int i = 0; i < npes; i++) \ switch (OP) { \ case AND: \ diff --git a/test/unit/c11_test_shmem_atomic_cswap.c b/test/unit/c11_test_shmem_atomic_cswap.c index ff8f811..e06e260 100644 --- a/test/unit/c11_test_shmem_atomic_cswap.c +++ b/test/unit/c11_test_shmem_atomic_cswap.c @@ -78,7 +78,7 @@ enum op { CSWAP = 0, ATOMIC_COMPARE_SWAP, CTX_ATOMIC_COMPARE_SWAP }; mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ } \ - if (old != npes) { \ + if (old != (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ diff --git a/test/unit/c11_test_shmem_atomic_inc.c b/test/unit/c11_test_shmem_atomic_inc.c index d3bf945..9d9d823 100644 --- a/test/unit/c11_test_shmem_atomic_inc.c +++ b/test/unit/c11_test_shmem_atomic_inc.c @@ -70,7 +70,7 @@ enum op { INC = 0, ATOMIC_INC, CTX_ATOMIC_INC, FINC, ATOMIC_FETCH_INC, break; \ case FINC: \ old = DEPRECATED_FINC(&remote, i); \ - if (old > npes) { \ + if (old > (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ @@ -78,7 +78,7 @@ enum op { INC = 0, ATOMIC_INC, CTX_ATOMIC_INC, FINC, ATOMIC_FETCH_INC, break; \ case ATOMIC_FETCH_INC: \ old = shmem_atomic_fetch_inc(&remote, i); \ - if (old > npes) { \ + if (old > (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ @@ -86,7 +86,7 @@ enum op { INC = 0, ATOMIC_INC, CTX_ATOMIC_INC, FINC, ATOMIC_FETCH_INC, break; \ case CTX_ATOMIC_FETCH_INC: \ old = shmem_atomic_fetch_inc(SHMEM_CTX_DEFAULT, &remote, i); \ - if (old > npes) { \ + if (old > (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ diff --git a/test/unit/c11_test_shmem_atomic_or.c b/test/unit/c11_test_shmem_atomic_or.c index 2215289..94c535e 100644 --- a/test/unit/c11_test_shmem_atomic_or.c +++ b/test/unit/c11_test_shmem_atomic_or.c @@ -30,6 +30,7 @@ * SOFTWARE. */ +#include #include #include #include @@ -51,7 +52,7 @@ enum op { OR = 0, CTX_OR, FETCH_OR, CTX_FETCH_OR }; const int mype = shmem_my_pe(); \ const int npes = shmem_n_pes(); \ TYPE old = (TYPE)0; \ - if (npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ + if ((size_t) npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ for (int i = 0; i < npes; i++) \ switch (OP) { \ case OR: \ diff --git a/test/unit/c11_test_shmem_atomic_swap.c b/test/unit/c11_test_shmem_atomic_swap.c index 646441b..e844ae4 100644 --- a/test/unit/c11_test_shmem_atomic_swap.c +++ b/test/unit/c11_test_shmem_atomic_swap.c @@ -75,7 +75,7 @@ enum op { SWAP = 0, ATOMIC_SWAP, CTX_ATOMIC_SWAP }; "TEST_SHMEM_SWAP(%s, %s)\n", mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ } \ - if (old != npes) { \ + if (old != (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ diff --git a/test/unit/c11_test_shmem_atomic_xor.c b/test/unit/c11_test_shmem_atomic_xor.c index 67e1562..5ee2f8b 100644 --- a/test/unit/c11_test_shmem_atomic_xor.c +++ b/test/unit/c11_test_shmem_atomic_xor.c @@ -30,6 +30,7 @@ * SOFTWARE. */ +#include #include #include #include @@ -51,7 +52,7 @@ enum op { XOR = 0, CTX_XOR, FETCH_XOR, CTX_FETCH_XOR }; const int mype = shmem_my_pe(); \ const int npes = shmem_n_pes(); \ TYPE old; \ - if (npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ + if ((size_t) npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ for (int i = 0; i < npes; i++) \ switch (OP) { \ case XOR: \ diff --git a/test/unit/cxx_test_shmem_atomic_and.cpp b/test/unit/cxx_test_shmem_atomic_and.cpp index a399c29..b5d5af6 100644 --- a/test/unit/cxx_test_shmem_atomic_and.cpp +++ b/test/unit/cxx_test_shmem_atomic_and.cpp @@ -30,6 +30,7 @@ * SOFTWARE. */ +#include #include #include #include @@ -47,7 +48,7 @@ enum op { AND = 0, CTX_AND, FETCH_AND, CTX_FETCH_AND }; do { \ static TYPE remote = ~(TYPE)0; \ TYPE old = (TYPE)0; \ - if (npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ + if ((size_t) npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ for (int i = 0; i < npes; i++) \ switch (OP) { \ case AND: \ diff --git a/test/unit/cxx_test_shmem_atomic_cswap.cpp b/test/unit/cxx_test_shmem_atomic_cswap.cpp index 1a4ef0d..8901c95 100644 --- a/test/unit/cxx_test_shmem_atomic_cswap.cpp +++ b/test/unit/cxx_test_shmem_atomic_cswap.cpp @@ -76,7 +76,7 @@ enum op { CSWAP = 0, ATOMIC_COMPARE_SWAP, CTX_ATOMIC_COMPARE_SWAP }; mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ } \ - if (old != npes) { \ + if (old != (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ diff --git a/test/unit/cxx_test_shmem_atomic_inc.cpp b/test/unit/cxx_test_shmem_atomic_inc.cpp index 3090454..b3288ec 100644 --- a/test/unit/cxx_test_shmem_atomic_inc.cpp +++ b/test/unit/cxx_test_shmem_atomic_inc.cpp @@ -68,7 +68,7 @@ enum op { INC = 0, ATOMIC_INC, CTX_ATOMIC_INC, FINC, ATOMIC_FETCH_INC, break; \ case FINC: \ old = DEPRECATED_FINC(&remote, i); \ - if (old > npes) { \ + if (old > (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ @@ -76,7 +76,7 @@ enum op { INC = 0, ATOMIC_INC, CTX_ATOMIC_INC, FINC, ATOMIC_FETCH_INC, break; \ case ATOMIC_FETCH_INC: \ old = shmem_atomic_fetch_inc(&remote, i); \ - if (old > npes) { \ + if (old > (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ @@ -84,7 +84,7 @@ enum op { INC = 0, ATOMIC_INC, CTX_ATOMIC_INC, FINC, ATOMIC_FETCH_INC, break; \ case CTX_ATOMIC_FETCH_INC: \ old = shmem_atomic_fetch_inc(SHMEM_CTX_DEFAULT, &remote, i); \ - if (old > npes) { \ + if (old > (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ diff --git a/test/unit/cxx_test_shmem_atomic_or.cpp b/test/unit/cxx_test_shmem_atomic_or.cpp index 5623a9a..9b35722 100644 --- a/test/unit/cxx_test_shmem_atomic_or.cpp +++ b/test/unit/cxx_test_shmem_atomic_or.cpp @@ -30,6 +30,7 @@ * SOFTWARE. */ +#include #include #include #include @@ -47,7 +48,7 @@ enum op { OR = 0, CTX_OR, FETCH_OR, CTX_FETCH_OR }; do { \ static TYPE remote = (TYPE)0; \ TYPE old = (TYPE)0; \ - if (npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ + if ((size_t) npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ for (int i = 0; i < npes; i++) \ switch (OP) { \ case OR: \ diff --git a/test/unit/cxx_test_shmem_atomic_swap.cpp b/test/unit/cxx_test_shmem_atomic_swap.cpp index 8efddf2..a7f9c2b 100644 --- a/test/unit/cxx_test_shmem_atomic_swap.cpp +++ b/test/unit/cxx_test_shmem_atomic_swap.cpp @@ -73,7 +73,7 @@ enum op { SWAP = 0, ATOMIC_SWAP, CTX_ATOMIC_SWAP }; "TEST_SHMEM_SWAP(%s, %s)\n", mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ } \ - if (old != npes) { \ + if (old != (TYPE) npes) { \ printf("PE %i error inconsistent value of old (%s, %s)\n", \ mype, #OP, #TYPE); \ rc = EXIT_FAILURE; \ diff --git a/test/unit/cxx_test_shmem_atomic_xor.cpp b/test/unit/cxx_test_shmem_atomic_xor.cpp index c52e3c3..d0dde4a 100644 --- a/test/unit/cxx_test_shmem_atomic_xor.cpp +++ b/test/unit/cxx_test_shmem_atomic_xor.cpp @@ -30,6 +30,7 @@ * SOFTWARE. */ +#include #include #include #include @@ -47,7 +48,7 @@ enum op { XOR = 0, CTX_XOR, FETCH_XOR, CTX_FETCH_XOR }; do { \ static TYPE remote = ~(TYPE)0; \ TYPE old; \ - if (npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ + if ((size_t) npes-1 > sizeof(TYPE)) break; /* Avoid overflow */ \ for (int i = 0; i < npes; i++) \ switch (OP) { \ case XOR: \ diff --git a/test/unit/mt_lock.c b/test/unit/mt_lock.c new file mode 100644 index 0000000..44ab935 --- /dev/null +++ b/test/unit/mt_lock.c @@ -0,0 +1,235 @@ +/* -*- C -*- + * + * Copyright 2011 Sandia Corporation. Under the terms of Contract + * DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government + * retains certain rights in this software. + * + * Copyright (c) 2018 Intel Corporation. All rights reserved. + * This software is available to you under the BSD license. + * + * This file is part of the Sandia OpenSHMEM software package. For license + * information, see the LICENSE file in the top level directory of the + * distribution. + * + */ + +#include +#include +#include +#include +#include "mt_lock.h" +#include "uthash.h" + +#define MTL_QUEUE_SIZE 4 +#define DEBUG 0 + +typedef struct { + UT_hash_handle hh; + long *key; + uint64_t next_ticket; + uint64_t cur_ticket; + pthread_mutex_t mutex; + pthread_cond_t cond[]; +} mtl_guard_t; + +static mtl_guard_t *guards = NULL; +pthread_mutex_t guards_mutex = PTHREAD_MUTEX_INITIALIZER; + +#define RAISE_ERROR(err, msg) \ + do { \ + char errmsg[256]; \ + strerror_r(err, errmsg, 256); \ + fprintf(stderr, "%d: %s - %s\n", \ + shmem_my_pe(), msg, errmsg); \ + shmem_global_exit(1); \ + } while(0) + + +/* Simple queueing lock using Lamport's bakery algorithm. Uses MTL_QUEUE_SIZE + * condition variables per lock to reduce the number of threads that are woken + * up when the lock is released. */ + +static inline void mtl_qlock_lock(mtl_guard_t *g) { + int ret; + uint64_t my_ticket; + + ret = pthread_mutex_lock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_lock failed"); + + my_ticket = g->next_ticket++; + while (my_ticket != g->cur_ticket) { + ret = pthread_cond_wait(&g->cond[my_ticket % MTL_QUEUE_SIZE], + &g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_cond_wait failed"); + } + ret = pthread_mutex_unlock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_unlock failed"); +} + + +static inline void mtl_qlock_unlock(mtl_guard_t *g) { + int ret; + + ret = pthread_mutex_lock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_lock failed"); + + g->cur_ticket++; + ret = pthread_cond_broadcast(&g->cond[g->cur_ticket % MTL_QUEUE_SIZE]); + if (ret) RAISE_ERROR(ret, "pthread_cond_broadcast failed"); + ret = pthread_mutex_unlock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_unlock failed"); +} + + +static inline int mtl_qlock_trylock(mtl_guard_t *g) { + int ret; + + ret = pthread_mutex_lock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_lock failed"); + + if (g->next_ticket == g->cur_ticket) { + g->next_ticket++; + ret = pthread_mutex_unlock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_unlock failed"); + return 0; + } + else { + ret = pthread_mutex_unlock(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_unlock failed"); + return 1; + } +} + +/** Guard Implementation **/ + +static mtl_guard_t* mtl_guard_locate(long *lockp) { + int ret, i; + mtl_guard_t *g; + + pthread_mutex_lock(&guards_mutex); + HASH_FIND_PTR(guards, &lockp, g); + + if (g == NULL) { + if (DEBUG) + printf("%d: Registering guard for lock %p\n", shmem_my_pe(), + (void*) lockp); + + g = (mtl_guard_t*) + malloc(sizeof(mtl_guard_t) + sizeof(pthread_cond_t) + * MTL_QUEUE_SIZE); + if (g == NULL) { + fprintf(stderr, "%d: Out of memory allocating lock guard\n", + shmem_my_pe()); + shmem_global_exit(2); + } + + g->key = lockp; + g->next_ticket = 0; + g->cur_ticket = 0; + ret = pthread_mutex_init(&g->mutex, NULL); + if (ret) RAISE_ERROR(ret, "pthread_mutex_init failed"); + for (i = 0; i < MTL_QUEUE_SIZE ; i++) { + ret = pthread_cond_init(&g->cond[i], NULL); + if (ret) RAISE_ERROR(ret, "pthread_cond_init failed"); + } + + HASH_ADD_PTR(guards, key, g); + } + pthread_mutex_unlock(&guards_mutex); + return g; +} + + +static void mtl_guard_enter(long *lockp) { + mtl_guard_t *g = mtl_guard_locate(lockp); + mtl_qlock_lock(g); +} + + +static int mtl_guard_test_enter(long *lockp) { + mtl_guard_t *g = mtl_guard_locate(lockp); + return mtl_qlock_trylock(g); +} + + +static void mtl_guard_exit(long *lockp) { + mtl_guard_t *g; + + pthread_mutex_lock(&guards_mutex); + HASH_FIND_PTR(guards, &lockp, g); + pthread_mutex_unlock(&guards_mutex); + + if (g == NULL) { + fprintf(stderr, "%d: Attempted to clear a lock that is not held (%p)\n", + shmem_my_pe(), (void*) lockp); + shmem_global_exit(3); + } + + mtl_qlock_unlock(g); +} + +/** User-facing API **/ + +void mtl_cleanup(void) { + int ret, i; + mtl_guard_t *g; + + for (g = guards; g != NULL ; ) { + mtl_guard_t *nextg = g->hh.next; + ret = pthread_mutex_destroy(&g->mutex); + if (ret) RAISE_ERROR(ret, "pthread_mutex_destroy failed"); + for (i = 0; i < MTL_QUEUE_SIZE; i++) { + ret = pthread_cond_destroy(&g->cond[i]); + if (ret) RAISE_ERROR(ret, "pthread_cond_destroy failed"); + } + free(g); + g = nextg; + } +} + + +void mtl_clear_lock(long *lockp) { + int prov; + + shmem_clear_lock(lockp); + shmem_query_thread(&prov); + + if (prov == SHMEM_THREAD_SERIALIZED || prov == SHMEM_THREAD_MULTIPLE) + mtl_guard_exit(lockp); +} + + +void mtl_set_lock(long *lockp) { + int prov; + + /* Guards use a local mutex to prevent concurrent requests on the same + * lock. The guard is "entered" when the SHMEM lock is acquired and + * "exited" when the SHMEM lock is released. */ + shmem_query_thread(&prov); + + if (prov == SHMEM_THREAD_SERIALIZED || prov == SHMEM_THREAD_MULTIPLE) + mtl_guard_enter(lockp); + + shmem_set_lock(lockp); +} + + +int mtl_test_lock(long *lockp) { + int status, prov; + + shmem_query_thread(&prov); + + if (prov == SHMEM_THREAD_SERIALIZED || prov == SHMEM_THREAD_MULTIPLE) { + if (mtl_guard_test_enter(lockp)) + return 1; + } + + status = shmem_test_lock(lockp); + if (status == 0) + return status; + + if (prov == SHMEM_THREAD_SERIALIZED || prov == SHMEM_THREAD_MULTIPLE) + mtl_guard_exit(lockp); + + return 1; +} diff --git a/test/unit/mt_lock.h b/test/unit/mt_lock.h new file mode 100644 index 0000000..8e183be --- /dev/null +++ b/test/unit/mt_lock.h @@ -0,0 +1,34 @@ +/* -*- C -*- + * + * Copyright 2011 Sandia Corporation. Under the terms of Contract + * DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government + * retains certain rights in this software. + * + * Copyright (c) 2018 Intel Corporation. All rights reserved. + * This software is available to you under the BSD license. + * + * This file is part of the Sandia OpenSHMEM software package. For license + * information, see the LICENSE file in the top level directory of the + * distribution. + * + */ + +/* MTL -- MultiThreaded Lock library for OpenSHMEM + * + * Wraps the OpenSHMEM lock API, allowing multiple threads at each PE to + * request the same SHMEM lock concurrently. */ + +#ifndef MT_LOCK_H +#define MT_LOCK_H + +#include "shmem.h" + +void mtl_clear_lock(long *lockp); +void mtl_set_lock(long *lockp); +int mtl_test_lock(long *lockp); + +/* When finished using MTL, the following routine can be used to free memory + * used by the library. */ +void mtl_cleanup(void); + +#endif /* #ifndef MT_LOCK_H */ diff --git a/test/unit/mt_lock_test.c b/test/unit/mt_lock_test.c new file mode 100644 index 0000000..db7342c --- /dev/null +++ b/test/unit/mt_lock_test.c @@ -0,0 +1,124 @@ +/* + * Copyright (c) 2018 Intel Corporation. All rights reserved. + * This software is available to you under the BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +/* Multithreaded locking test -- Ensure the SHMEM lock API behaves correctly + * when used in a multithreaded setting. */ + +#include +#include +#include +#include +#include +#include "mt_lock.h" + +#define T 4 +#define N 4 + +long dest = 0; +long *locks; + +int me, npes; +int errors = 0; + +static void * thread_main(void *arg) { + int i, j; + + for (j = 0; j < N; j++) { + for (i = 0 ; i < npes; i++) { + /* Alternate even/odd PEs using set_lock versus a test_lock loop to + * acquire the lock */ + if ((me+j) % 2 == 0) { + mtl_set_lock(&locks[i]); + long d = shmem_long_g(&dest, i); + shmem_long_p(&dest, d+1, i); + mtl_clear_lock(&locks[i]); + } else { + while (mtl_test_lock(&locks[i])) ; + long d = shmem_long_g(&dest, i); + shmem_long_p(&dest, d+1, i); + mtl_clear_lock(&locks[i]); + } + } + } + + return NULL; +} + + +int main(int argc, char **argv) { + int tl, i, ret; + pthread_t threads[T]; + int t_arg[T]; + + ret = shmem_init_thread(SHMEM_THREAD_MULTIPLE, &tl); + + if (tl != SHMEM_THREAD_MULTIPLE || ret != 0) { + printf("Init failed (requested thread level %d, got %d, ret %d)\n", + SHMEM_THREAD_MULTIPLE, tl, ret); + + if (ret == 0) { + shmem_global_exit(1); + } else { + return ret; + } + } + + me = shmem_my_pe(); + npes = shmem_n_pes(); + locks = shmem_calloc(npes, sizeof(long)); + + if (me == 0) printf("Starting MT locking test on %d PEs, %d threads/PE\n", npes, T); + + for (i = 0; i < T; i++) { + int err; + t_arg[i] = i; + err = pthread_create(&threads[i], NULL, thread_main, (void*) &t_arg[i]); + assert(0 == err); + } + + for (i = 0; i < T; i++) { + int err; + err = pthread_join(threads[i], NULL); + assert(0 == err); + } + + mtl_cleanup(); + shmem_barrier_all(); + + if (dest != npes*T*N) { + printf("%d: Error expected %d, got %ld\n", me, npes*T*N, dest); + errors++; + } + + if (me == 0) { + if (errors) printf("Encountered %d errors\n", errors); + else printf("Success\n"); + } + + shmem_finalize(); + return (errors == 0) ? 0 : 1; +} diff --git a/test/unit/shmem_info_f.f90 b/test/unit/shmem_info_f.f90 index 8c3e4d9..c1e2ba2 100644 --- a/test/unit/shmem_info_f.f90 +++ b/test/unit/shmem_info_f.f90 @@ -29,6 +29,7 @@ program shmem_info include 'shmem.fh' integer me, major_ver, minor_ver + integer shmem_my_pe character (len=SHMEM_MAX_NAME_LEN) lib_name call shmem_init()