Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
changeset 168:d7c0c0a8187a perf_counters
Merged default branch
| author | Merten Sach <msach@mailbox.tu-berlin.de> |
|---|---|
| date | Wed, 02 Nov 2011 14:59:29 +0100 |
| parents | 395f58384a5c ad8213a8e916 |
| children | d83f59e6e2db |
| files | CoreLoop.c VMS.c VMS.h VMS__DESIGN_NOTES.txt |
| diffstat | 4 files changed, 631 insertions(+), 638 deletions(-) [+] |
line diff
1.1 --- a/.hgtags Thu Sep 15 17:31:33 2011 +0200 1.2 +++ b/.hgtags Wed Nov 02 14:59:29 2011 +0100 1.3 @@ -1,6 +1,1 @@ 1.4 9c3107044f86c36fea3a8f72f64910b1363555be Dec27_2010_about_to_add_sched_record 1.5 -97e26095c01fef53783d2d89e11575856ce243c5 V0 1.6 -21c95d402fe66570067000e484bfec9a5d92c9d0 malloc_touch 1.7 -42d015d48eebc7ba49b24fa5f24cb1e66244e5e7 malloc_touch 1.8 -42d015d48eebc7ba49b24fa5f24cb1e66244e5e7 malloc_touch 1.9 -0000000000000000000000000000000000000000 malloc_touch
2.1 --- a/VMS.c Thu Sep 15 17:31:33 2011 +0200 2.2 +++ b/VMS.c Wed Nov 02 14:59:29 2011 +0100 2.3 @@ -118,15 +118,15 @@ 2.4 2.5 //============================= MEASUREMENT STUFF ======================== 2.6 #ifdef MEAS__TIME_MALLOC 2.7 - _VMSMasterEnv->mallocTimeHist = makeFixedBinHistExt( 100, 0, 100, 2.8 + _VMSMasterEnv->mallocTimeHist = makeFixedBinHistExt( 100, 0, 30, 2.9 "malloc_time_hist"); 2.10 - _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 80, 0, 100, 2.11 + _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 100, 0, 30, 2.12 "free_time_hist"); 2.13 #endif 2.14 #ifdef MEAS__TIME_PLUGIN 2.15 - _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 1000, 0, 100, 2.16 + _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 100, 0, 200, 2.17 "plugin_low_time_hist"); 2.18 - _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 1000, 0, 100, 2.19 + _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 100, 0, 200, 2.20 "plugin_high_time_hist"); 2.21 #endif 2.22 //========================================================================
3.1 --- a/VMS.h Thu Sep 15 17:31:33 2011 +0200 3.2 +++ b/VMS.h Wed Nov 02 14:59:29 2011 +0100 3.3 @@ -1,627 +1,627 @@ 3.4 -/* 3.5 - * Copyright 2009 OpenSourceStewardshipFoundation.org 3.6 - * Licensed under GNU General Public License version 2 3.7 - * 3.8 - * Author: seanhalle@yahoo.com 3.9 - * 3.10 - */ 3.11 - 3.12 -#ifndef _VMS_H 3.13 -#define _VMS_H 3.14 -#define _GNU_SOURCE 3.15 - 3.16 -#include "VMS_primitive_data_types.h" 3.17 -#include "Queue_impl/PrivateQueue.h" 3.18 -#include "Histogram/Histogram.h" 3.19 -#include "DynArray/DynArray.h" 3.20 -#include "Hash_impl/PrivateHash.h" 3.21 -#include "vmalloc.h" 3.22 -#include "Counters/Counters.h" 3.23 -#include "dependency.h" 3.24 - 3.25 -#include <pthread.h> 3.26 -#include <sys/time.h> 3.27 - 3.28 - 3.29 -//=============================== Debug =================================== 3.30 -// 3.31 -//When SEQUENTIAL is defined, VMS does sequential exe in the main thread 3.32 -// It still does co-routines and all the mechanisms are the same, it just 3.33 -// has only a single thread and animates VPs one at a time 3.34 -//#define SEQUENTIAL 3.35 - 3.36 -//#define USE_WORK_STEALING 3.37 - 3.38 -//turns on the probe-instrumentation in the application -- when not 3.39 -// defined, the calls to the probe functions turn into comments 3.40 -//#define STATS__ENABLE_PROBES 3.41 -//#define TURN_ON_DEBUG_PROBES 3.42 - 3.43 -//These defines turn types of bug messages on and off 3.44 -// be sure debug messages are un-commented (next block of defines) 3.45 -#define dbgAppFlow FALSE /* Top level flow of application code -- general*/ 3.46 -#define dbgProbes FALSE /* for issues inside probes themselves*/ 3.47 -#define dbgB2BMaster FALSE /* in coreloop, back to back master VPs*/ 3.48 -#define dbgRqstHdlr FALSE /* in request handler code*/ 3.49 -#define dbgDependency TRUE /* in request handler code, print dependencies */ 3.50 - 3.51 -//Comment or un- the substitute half to turn on/off types of debug message 3.52 -#define DEBUG( bool, msg) \ 3.53 - if( bool){ printf(msg); fflush(stdin);} 3.54 -#define DEBUG1( bool, msg, param) \ 3.55 - if(bool){printf(msg, param); fflush(stdin);} 3.56 -#define DEBUG2( bool, msg, p1, p2) \ 3.57 - if(bool) {printf(msg, p1, p2); fflush(stdin);} 3.58 - 3.59 -#define ERROR(msg) printf(msg); 3.60 -#define ERROR1(msg, param) printf(msg, param); 3.61 -#define ERROR2(msg, p1, p2) printf(msg, p1, p2); 3.62 - 3.63 -//=========================== STATS ======================= 3.64 - 3.65 - //when MEAS__TIME_STAMP_SUSP is defined, causes code to be inserted and 3.66 - // compiled-in that saves the low part of the time stamp count just before 3.67 - // suspending a processor and just after resuming that processor. It is 3.68 - // saved into a field added to VirtProcr. Have to sanity-check for 3.69 - // rollover of low portion into high portion. 3.70 -//#define MEAS__TIME_STAMP_SUSP 3.71 -//#define MEAS__TIME_MASTER 3.72 -//#define MEAS__TIME_PLUGIN 3.73 -//#define MEAS__TIME_MALLOC 3.74 -//#define MEAS__TIME_MASTER_LOCK 3.75 -//#define MEAS__NUM_TIMES_TO_RUN 100000 3.76 - 3.77 - //For code that calculates normalization-offset between TSC counts of 3.78 - // different cores. 3.79 -//#define NUM_TSC_ROUND_TRIPS 10 3.80 - 3.81 -#define MEAS__PERF_COUNTERS 3.82 -#define DETECT_DEPENDENCIES 3.83 - 3.84 -//========================= Hardware related Constants ===================== 3.85 - //This value is the number of hardware threads in the shared memory 3.86 - // machine 3.87 -#define NUM_CORES 2 3.88 - 3.89 - // tradeoff amortizing master fixed overhead vs imbalance potential 3.90 - // when work-stealing, can make bigger, at risk of losing cache affinity 3.91 -#define NUM_SCHED_SLOTS 5 3.92 - 3.93 -#define MIN_WORK_UNIT_CYCLES 20000 3.94 - 3.95 -#define MASTERLOCK_RETRIES 10000 3.96 - 3.97 - // stack size in virtual processors created 3.98 -#define VIRT_PROCR_STACK_SIZE 0x8000 /* 32K */ 3.99 - 3.100 - // memory for VMS__malloc 3.101 -#define MALLOC_ADDITIONAL_MEM_FROM_OS_SIZE 0x10000000 /* 256M */ 3.102 - 3.103 -#define CACHE_LINE 64 3.104 -#define PAGE_SIZE 4096 3.105 - 3.106 - 3.107 -//============================== 3.108 - 3.109 -#define SUCCESS 0 3.110 - 3.111 -#define writeVMSQ writePrivQ 3.112 -#define readVMSQ readPrivQ 3.113 -#define makeVMSQ makeVMSPrivQ 3.114 -#define numInVMSQ numInPrivQ 3.115 -#define VMSQueueStruc PrivQueueStruc 3.116 - 3.117 - 3.118 - 3.119 -//=========================================================================== 3.120 -typedef unsigned long long TSCount; 3.121 - 3.122 -typedef struct _SchedSlot SchedSlot; 3.123 -typedef struct _VMSReqst VMSReqst; 3.124 -typedef struct _VirtProcr VirtProcr; 3.125 -typedef struct _IntervalProbe IntervalProbe; 3.126 -typedef struct _GateStruc GateStruc; 3.127 - 3.128 - 3.129 -typedef VirtProcr * (*SlaveScheduler) ( void *, int ); //semEnv, coreIdx 3.130 -typedef void (*RequestHandler) ( VirtProcr *, void * ); //prWReqst, semEnv 3.131 -typedef void (*VirtProcrFnPtr) ( void *, VirtProcr * ); //initData, animPr 3.132 -typedef void VirtProcrFn ( void *, VirtProcr * ); //initData, animPr 3.133 -typedef void (*ResumePrFnPtr) ( VirtProcr *, void * ); 3.134 - 3.135 - 3.136 -//============= Requests =========== 3.137 -// 3.138 - 3.139 -enum VMSReqstType //avoid starting enums at 0, for debug reasons 3.140 - { 3.141 - semantic = 1, 3.142 - createReq, 3.143 - dissipate, 3.144 - VMSSemantic //goes with VMSSemReqst below 3.145 - }; 3.146 - 3.147 -struct _VMSReqst 3.148 - { 3.149 - enum VMSReqstType reqType;//used for dissipate and in future for IO requests 3.150 - void *semReqData; 3.151 - 3.152 - VMSReqst *nextReqst; 3.153 - }; 3.154 -//VMSReqst 3.155 - 3.156 -enum VMSSemReqstType //These are equivalent to semantic requests, but for 3.157 - { // VMS's services available directly to app, like OS 3.158 - createProbe = 1, // and probe services -- like a VMS-wide built-in lang 3.159 - openFile, 3.160 - otherIO 3.161 - }; 3.162 - 3.163 -typedef struct 3.164 - { enum VMSSemReqstType reqType; 3.165 - VirtProcr *requestingPr; 3.166 - char *nameStr; //for create probe 3.167 - } 3.168 - VMSSemReq; 3.169 - 3.170 - 3.171 -//==================== Core data structures =================== 3.172 - 3.173 -struct _SchedSlot 3.174 - { 3.175 - int workIsDone; 3.176 - int needsProcrAssigned; 3.177 - VirtProcr *procrAssignedToSlot; 3.178 - }; 3.179 -//SchedSlot 3.180 - 3.181 -/*WARNING: re-arranging this data structure could cause VP switching 3.182 - * assembly code to fail -- hard-codes offsets of fields 3.183 - */ 3.184 -struct _VirtProcr 3.185 - { int procrID; //for debugging -- count up each time create 3.186 - int coreAnimatedBy; 3.187 - void *startOfStack; 3.188 - void *stackPtr; 3.189 - void *framePtr; 3.190 - void *nextInstrPt; 3.191 - 3.192 - void *coreLoopStartPt; //allows proto-runtime to be linked later 3.193 - void *coreLoopFramePtr; //restore before jmp back to core loop 3.194 - void *coreLoopStackPtr; //restore before jmp back to core loop 3.195 - 3.196 - void *initialData; 3.197 - 3.198 - SchedSlot *schedSlot; 3.199 - VMSReqst *requests; 3.200 - 3.201 - void *semanticData; //this livesUSE_GNU here for the life of VP 3.202 - void *dataRetFromReq;//values returned from plugin to VP go here 3.203 - 3.204 - //=========== MEASUREMENT STUFF ========== 3.205 - #ifdef MEAS__TIME_STAMP_SUSP 3.206 - unsigned int preSuspTSCLow; 3.207 - unsigned int postSuspTSCLow; 3.208 - #endif 3.209 - #ifdef MEAS__TIME_MASTER /* in VirtProcr because multiple masterVPs*/ 3.210 - unsigned int startMasterTSCLow;USE_GNU 3.211 - unsigned int endMasterTSCLow; 3.212 - #endif 3.213 - #ifdef MEAS__PERF_COUNTERS // 3.214 - CounterRecord** counter_history; 3.215 - PrivDynArrayInfo* counter_history_array_info; 3.216 - #endif 3.217 - //======================================== 3.218 - 3.219 - float64 createPtInSecs; //have space but don't use on some configs 3.220 - }; 3.221 -//VirtProcr 3.222 - 3.223 - 3.224 -/*WARNING: re-arranging this data structure could cause VP-switching 3.225 - * assembly code to fail -- hard-codes offsets of fields 3.226 - * (because -O3 messes with things otherwise) 3.227 - */ 3.228 -typedef struct 3.229 - { 3.230 - SlaveScheduler slaveScheduler; 3.231 - RequestHandler requestHandler; 3.232 - 3.233 - SchedSlot ***allSchedSlots; 3.234 - VMSQueueStruc **readyToAnimateQs; 3.235 - VirtProcr **masterVPs; 3.236 - 3.237 - void *semanticEnv; 3.238 - void *OSEventStruc; //for future, when add I/O to BLIS 3.239 - MallocProlog *freeListHead; 3.240 - int32 amtOfOutstandingMem; //total currently allocated 3.241 - 3.242 - void *coreLoopReturnPt;//addr to jump to to re-enter coreLoop 3.243 - 3.244 - int32 setupComplete; 3.245 - volatile int32 masterLock; 3.246 - 3.247 - int32 numMasterInARow[NUM_CORES];//detect back-to-back masterVP 3.248 - GateStruc *workStealingGates[ NUM_CORES ]; //concurrent work-steal 3.249 - int32 workStealingLock; 3.250 - 3.251 - int32 numProcrsCreated; //gives ordering to processor creation 3.252 - 3.253 - //=========== MEASUREMENT STUFF ============= 3.254 - IntervalProbe **intervalProbes; 3.255 - PrivDynArrayInfo *dynIntervalProbesInfo; 3.256 - HashTable *probeNameHashTbl; 3.257 - int32 masterCreateProbeID; 3.258 - float64 createPtInSecs; 3.259 - Histogram **measHists; 3.260 - PrivDynArrayInfo *measHistsInfo; 3.261 - #ifdef MEAS__TIME_PLUGIN 3.262 - Histogram *reqHdlrLowTimeHist; 3.263 - Histogram *reqHdlrHighTimeHist; 3.264 - #endif 3.265 - #ifdef MEAS__TIME_MALLOC 3.266 - Histogram *mallocTimeHist; 3.267 - Histogram *freeTimeHist; 3.268 - #endif 3.269 - #ifdef MEAS__TIME_MASTER_LOCK 3.270 - Histogram *masterLockLowTimeHist; 3.271 - Histogram *masterLockHighTimeHist; 3.272 - #endif 3.273 - #ifdef MEAS__PERF_COUNTERS 3.274 - int cycles_counter_fd[NUM_CORES]; 3.275 - int instrs_counter_fd[NUM_CORES]; 3.276 - FILE* counteroutput; 3.277 - #endif 3.278 - #ifdef DETECT_DEPENDENCIES 3.279 - Dependency** dependencies; 3.280 - PrivDynArrayInfo* dependenciesInfo; 3.281 - #endif 3.282 - #ifdef MEAS__PERF_COUNTERS // 3.283 - CounterRecord** counter_history; 3.284 - PrivDynArrayInfo* counter_history_array_info; 3.285 - #endif 3.286 - } 3.287 -MasterEnv; 3.288 - 3.289 -//========================= Extra Stuff Data Strucs ======================= 3.290 -typedef struct 3.291 - { 3.292 - 3.293 - } 3.294 -VMSExcp; 3.295 - 3.296 -struct _GateStruc 3.297 - { 3.298 - int32 gateClosed; 3.299 - int32 preGateProgress; 3.300 - int32 waitProgress; 3.301 - int32 exitProgress; 3.302 - }; 3.303 -//GateStruc 3.304 - 3.305 -//======================= OS Thread related =============================== 3.306 - 3.307 -void * coreLoop( void *paramsIn ); //standard PThreads fn prototype 3.308 -void * coreLoop_Seq( void *paramsIn ); //standard PThreads fn prototype 3.309 -void masterLoop( void *initData, VirtProcr *masterPr ); 3.310 - 3.311 - 3.312 -typedef struct 3.313 - { 3.314 - void *endThdPt; 3.315 - unsigned int coreNum; 3.316 - } 3.317 -ThdParams; 3.318 - 3.319 -pthread_t coreLoopThdHandles[ NUM_CORES ]; //pthread's virt-procr state 3.320 -ThdParams *coreLoopThdParams [ NUM_CORES ]; 3.321 -pthread_mutex_t suspendLock; 3.322 -pthread_cond_t suspend_cond; 3.323 - 3.324 - 3.325 - 3.326 -//===================== Global Vars =================== 3.327 - 3.328 -volatile MasterEnv *_VMSMasterEnv; 3.329 - 3.330 - 3.331 - 3.332 - 3.333 -//=========================== Function Prototypes ========================= 3.334 - 3.335 - 3.336 -//========== Setup and shutdown ========== 3.337 -void 3.338 -VMS__init(); 3.339 - 3.340 -void 3.341 -VMS__init_Seq(); 3.342 - 3.343 -void 3.344 -VMS__start_the_work_then_wait_until_done(); 3.345 - 3.346 -void 3.347 -VMS__start_the_work_then_wait_until_done_Seq(); 3.348 - 3.349 -inline VirtProcr * 3.350 -VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData ); 3.351 - 3.352 -void 3.353 -VMS__dissipate_procr( VirtProcr *procrToDissipate ); 3.354 - 3.355 - //Use this to create processor inside entry point & other places outside 3.356 - // the VMS system boundary (IE, not run in slave nor Master) 3.357 -VirtProcr * 3.358 -VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData ); 3.359 - 3.360 -void 3.361 -VMS_ext__dissipate_procr( VirtProcr *procrToDissipate ); 3.362 - 3.363 -void 3.364 -VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData ); 3.365 - 3.366 -void 3.367 -VMS__shutdown(); 3.368 - 3.369 -void 3.370 -VMS__cleanup_at_end_of_shutdown(); 3.371 - 3.372 -void * 3.373 -VMS__give_sem_env_for( VirtProcr *animPr ); 3.374 - 3.375 - 3.376 -//============== Request Related =============== 3.377 - 3.378 -void 3.379 -VMS__suspend_procr( VirtProcr *callingPr ); 3.380 - 3.381 -inline void 3.382 -VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData, VirtProcr *callingPr ); 3.383 - 3.384 -/*inline*/ __attribute__ ((noinline)) void 3.385 -VMS__send_sem_request( void *semReqData, VirtProcr *callingPr ); 3.386 - 3.387 -void 3.388 -VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr ); 3.389 - 3.390 -void /*inline**/ __attribute__ ((noinline)) 3.391 -VMS__send_dissipate_req( VirtProcr *prToDissipate ); 3.392 - 3.393 -/*inline**/ __attribute__ ((noinline)) void 3.394 -VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr ); 3.395 - 3.396 -VMSReqst * 3.397 -VMS__take_next_request_out_of( VirtProcr *procrWithReq ); 3.398 - 3.399 -inline void * 3.400 -VMS__take_sem_reqst_from( VMSReqst *req ); 3.401 - 3.402 -void inline 3.403 -VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv, 3.404 - ResumePrFnPtr resumePrFnPtr ); 3.405 - 3.406 -//======================== STATS ====================== 3.407 - 3.408 -//===== RDTSC wrapper ===== //Also runs with x86_64 code 3.409 - 3.410 -#define saveTimeStampCountInto(low, high) \ 3.411 - asm volatile("RDTSC; \ 3.412 - movl %%eax, %0; \ 3.413 - movl %%edx, %1;" \ 3.414 - /* outputs */ : "=m" (low), "=m" (high)\ 3.415 - /* inputs */ : \ 3.416 - /* clobber */ : "%eax", "%edx" \ 3.417 - ); 3.418 - 3.419 -#define saveLowTimeStampCountInto(low) \ 3.420 - asm volatile("RDTSC; \ 3.421 - movl %%eax, %0;" \ 3.422 - /* outputs */ : "=m" (low) \ 3.423 - /* inputs */ : \ 3.424 - /* clobber */ : "%eax", "%edx" \ 3.425 - ); 3.426 - 3.427 -//==================== 3.428 -#define makeAMeasHist( idx, name, numBins, startVal, binWidth ) \ 3.429 - makeHighestDynArrayIndexBeAtLeast( _VMSMasterEnv->measHistsInfo, idx ); \ 3.430 - _VMSMasterEnv->measHists[idx] = \ 3.431 - makeFixedBinHist( numBins, startVal, binWidth, name ); 3.432 - 3.433 -#define saveCyclesAndInstrs(core,cycles,instrs) do{ \ 3.434 - int cycles_fd = _VMSMasterEnv->cycles_counter_fd[core]; \ 3.435 - int instrs_fd = _VMSMasterEnv->instrs_counter_fd[core]; \ 3.436 - int nread; \ 3.437 - \ 3.438 - nread = read(cycles_fd,&(cycles),sizeof(cycles)); \ 3.439 - if(nread<0){ \ 3.440 - perror("Error reading cycles counter"); \ 3.441 - cycles = 0; \ 3.442 - } \ 3.443 - \ 3.444 - nread = read(instrs_fd,&(instrs),sizeof(instrs)); \ 3.445 - if(nread<0){ \ 3.446 - perror("Error reading cycles counter"); \ 3.447 - instrs = 0; \ 3.448 - } \ 3.449 -} while (0) 3.450 - 3.451 -#define getReturnAddressBeforeLibraryCall(vp_ptr, res_ptr) do{ \ 3.452 -void* frame_ptr0 = vp_ptr->framePtr; \ 3.453 -void* frame_ptr1 = *((void**)frame_ptr0); \ 3.454 -void* frame_ptr2 = *((void**)frame_ptr1); \ 3.455 -void* frame_ptr3 = *((void**)frame_ptr2); \ 3.456 -void* ret_addr = *((void**)frame_ptr3 + 1); \ 3.457 -*res_ptr = ret_addr; \ 3.458 -} while (0) 3.459 - 3.460 -#define MEAS__SUB_CREATE /*turn on/off subtraction of create from plugin*/ 3.461 - 3.462 -#ifdef VPTHREAD 3.463 - 3.464 -//VPThread 3.465 -#define createHistIdx 1 3.466 -#define mutexLockHistIdx 2 3.467 -#define mutexUnlockHistIdx 3 3.468 -#define condWaitHistIdx 4 3.469 -#define condSignalHistIdx 5 3.470 - 3.471 -#define MakeTheMeasHists() \ 3.472 - _VMSMasterEnv->measHistsInfo = \ 3.473 - makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->measHists), 200); \ 3.474 - makeAMeasHist( createHistIdx, "create", 250, 0, 100 ) \ 3.475 - makeAMeasHist( mutexLockHistIdx, "mutex_lock", 50, 0, 100 ) \ 3.476 - makeAMeasHist( mutexUnlockHistIdx, "mutex_unlock", 50, 0, 100 ) \ 3.477 - makeAMeasHist( condWaitHistIdx, "cond_wait", 50, 0, 100 ) \ 3.478 - makeAMeasHist( condSignalHistIdx, "cond_signal", 50, 0, 100 ) 3.479 - 3.480 -#endif 3.481 - 3.482 - 3.483 -#ifdef VCILK 3.484 - 3.485 -//VCilk 3.486 -#define spawnHistIdx 1 3.487 -#define syncHistIdx 2 3.488 - 3.489 -#define MakeTheMeasHists() \ 3.490 - _VMSMasterEnv->measHistsInfo = \ 3.491 - makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->measHists), 200); \ 3.492 - makeAMeasHist( spawnHistIdx, "Spawn", 50, 0, 200 ) \ 3.493 - makeAMeasHist( syncHistIdx, "Sync", 50, 0, 200 ) 3.494 - 3.495 - 3.496 -#endif 3.497 - 3.498 -#ifdef SSR 3.499 - 3.500 -//SSR 3.501 -#define SendFromToHistIdx 1 3.502 -#define SendOfTypeHistIdx 2 3.503 -#define ReceiveFromToHistIdx 3 3.504 -#define ReceiveOfTypeHistIdx 4 3.505 - 3.506 -#define MakeTheMeasHists() \ 3.507 - _VMSMasterEnv->measHistsInfo = \ 3.508 - makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->measHists), 200); \ 3.509 - makeAMeasHist( SendFromToHistIdx, "SendFromTo", 50, 0, 100 ) \ 3.510 - makeAMeasHist( SendOfTypeHistIdx, "SendOfType", 50, 0, 100 ) \ 3.511 - makeAMeasHist( ReceiveFromToHistIdx,"ReceiveFromTo", 50, 0, 100 ) \ 3.512 - makeAMeasHist( ReceiveOfTypeHistIdx,"ReceiveOfType", 50, 0, 100 ) 3.513 - 3.514 -#endif 3.515 - 3.516 -//=========================================================================== 3.517 -//VPThread 3.518 - 3.519 - 3.520 -#define Meas_startCreate \ 3.521 - int32 startStamp, endStamp; \ 3.522 - saveLowTimeStampCountInto( startStamp ); \ 3.523 - 3.524 -#define Meas_endCreate \ 3.525 - saveLowTimeStampCountInto( endStamp ); \ 3.526 - addIntervalToHist( startStamp, endStamp, \ 3.527 - _VMSMasterEnv->measHists[ createHistIdx ] ); 3.528 - 3.529 -#define Meas_startMutexLock \ 3.530 - int32 startStamp, endStamp; \ 3.531 - saveLowTimeStampCountInto( startStamp ); \ 3.532 - 3.533 -#define Meas_endMutexLock \ 3.534 - saveLowTimeStampCountInto( endStamp ); \ 3.535 - addIntervalToHist( startStamp, endStamp, \ 3.536 - _VMSMasterEnv->measHists[ mutexLockHistIdx ] ); 3.537 - 3.538 -#define Meas_startMutexUnlock \ 3.539 - int32 startStamp, endStamp; \ 3.540 - saveLowTimeStampCountInto( startStamp ); \ 3.541 - 3.542 -#define Meas_endMutexUnlock \ 3.543 - saveLowTimeStampCountInto( endStamp ); \ 3.544 - addIntervalToHist( startStamp, endStamp, \ 3.545 - _VMSMasterEnv->measHists[ mutexUnlockHistIdx ] ); 3.546 - 3.547 -#define Meas_startCondWait \ 3.548 - int32 startStamp, endStamp; \ 3.549 - saveLowTimeStampCountInto( startStamp ); \ 3.550 - 3.551 -#define Meas_endCondWait \ 3.552 - saveLowTimeStampCountInto( endStamp ); \ 3.553 - addIntervalToHist( startStamp, endStamp, \ 3.554 - _VMSMasterEnv->measHists[ condWaitHistIdx ] ); 3.555 - 3.556 -#define Meas_startCondSignal \ 3.557 - int32 startStamp, endStamp; \ 3.558 - saveLowTimeStampCountInto( startStamp ); \ 3.559 - 3.560 -#define Meas_endCondSignal \ 3.561 - saveLowTimeStampCountInto( endStamp ); \ 3.562 - addIntervalToHist( startStamp, endStamp, \ 3.563 - _VMSMasterEnv->measHists[ condSignalHistIdx ] ); 3.564 - 3.565 -//=========================================================================== 3.566 -// VCilk 3.567 -#define Meas_startSpawn \ 3.568 - int32 startStamp, endStamp; \ 3.569 - saveLowTimeStampCountInto( startStamp ); \ 3.570 - 3.571 -#define Meas_endSpawn \ 3.572 - saveLowTimeStampCountInto( endStamp ); \ 3.573 - addIntervalToHist( startStamp, endStamp, \ 3.574 - _VMSMasterEnv->measHists[ spawnHistIdx ] ); 3.575 - 3.576 -#define Meas_startSync \ 3.577 - int32 startStamp, endStamp; \ 3.578 - saveLowTimeStampCountInto( startStamp ); \ 3.579 - 3.580 -#define Meas_endSync \ 3.581 - saveLowTimeStampCountInto( endStamp ); \ 3.582 - addIntervalToHist( startStamp, endStamp, \ 3.583 - _VMSMasterEnv->measHists[ syncHistIdx ] ); 3.584 - 3.585 -//=========================================================================== 3.586 -// SSR 3.587 -#define Meas_startSendFromTo \ 3.588 - int32 startStamp, endStamp; \ 3.589 - saveLowTimeStampCountInto( startStamp ); \ 3.590 - 3.591 -#define Meas_endSendFromTo \ 3.592 - saveLowTimeStampCountInto( endStamp ); \ 3.593 - addIntervalToHist( startStamp, endStamp, \ 3.594 - _VMSMasterEnv->measHists[ SendFromToHistIdx ] ); 3.595 - 3.596 -#define Meas_startSendOfType \ 3.597 - int32 startStamp, endStamp; \ 3.598 - saveLowTimeStampCountInto( startStamp ); \ 3.599 - 3.600 -#define Meas_endSendOfType \ 3.601 - saveLowTimeStampCountInto( endStamp ); \ 3.602 - addIntervalToHist( startStamp, endStamp, \ 3.603 - _VMSMasterEnv->measHists[ SendOfTypeHistIdx ] ); 3.604 - 3.605 -#define Meas_startReceiveFromTo \ 3.606 - int32 startStamp, endStamp; \ 3.607 - saveLowTimeStampCountInto( startStamp ); \ 3.608 - 3.609 -#define Meas_endReceiveFromTo \ 3.610 - saveLowTimeStampCountInto( endStamp ); \ 3.611 - addIntervalToHist( startStamp, endStamp, \ 3.612 - _VMSMasterEnv->measHists[ ReceiveFromToHistIdx ] ); 3.613 - 3.614 -#define Meas_startReceiveOfType \ 3.615 - int32 startStamp, endStamp; \ 3.616 - saveLowTimeStampCountInto( startStamp ); \ 3.617 - 3.618 -#define Meas_endReceiveOfType \ 3.619 - saveLowTimeStampCountInto( endStamp ); \ 3.620 - addIntervalToHist( startStamp, endStamp, \ 3.621 - _VMSMasterEnv->measHists[ReceiveOfTypeHistIdx ] ); 3.622 - 3.623 -//===== 3.624 - 3.625 -#include "ProcrContext.h" 3.626 -#include "probes.h" 3.627 -#include "vutilities.h" 3.628 - 3.629 -#endif /* _VMS_H */ 3.630 - 3.631 +/* 3.632 + * Copyright 2009 OpenSourceStewardshipFoundation.org 3.633 + * Licensed under GNU General Public License version 2 3.634 + * 3.635 + * Author: seanhalle@yahoo.com 3.636 + * 3.637 + */ 3.638 + 3.639 +#ifndef _VMS_H 3.640 +#define _VMS_H 3.641 +#define _GNU_SOURCE 3.642 + 3.643 +#include "VMS_primitive_data_types.h" 3.644 +#include "Queue_impl/PrivateQueue.h" 3.645 +#include "Histogram/Histogram.h" 3.646 +#include "DynArray/DynArray.h" 3.647 +#include "Hash_impl/PrivateHash.h" 3.648 +#include "vmalloc.h" 3.649 +#include "Counters/Counters.h" 3.650 +#include "dependency.h" 3.651 + 3.652 +#include <pthread.h> 3.653 +#include <sys/time.h> 3.654 + 3.655 + 3.656 +//=============================== Debug =================================== 3.657 +// 3.658 +//When SEQUENTIAL is defined, VMS does sequential exe in the main thread 3.659 +// It still does co-routines and all the mechanisms are the same, it just 3.660 +// has only a single thread and animates VPs one at a time 3.661 +//#define SEQUENTIAL 3.662 + 3.663 +//#define USE_WORK_STEALING 3.664 + 3.665 +//turns on the probe-instrumentation in the application -- when not 3.666 +// defined, the calls to the probe functions turn into comments 3.667 +//#define STATS__ENABLE_PROBES 3.668 +//#define TURN_ON_DEBUG_PROBES 3.669 + 3.670 +//These defines turn types of bug messages on and off 3.671 +// be sure debug messages are un-commented (next block of defines) 3.672 +#define dbgAppFlow FALSE /* Top level flow of application code -- general*/ 3.673 +#define dbgProbes FALSE /* for issues inside probes themselves*/ 3.674 +#define dbgB2BMaster FALSE /* in coreloop, back to back master VPs*/ 3.675 +#define dbgRqstHdlr FALSE /* in request handler code*/ 3.676 +#define dbgDependency TRUE /* in request handler code, print dependencies */ 3.677 + 3.678 +//Comment or un- the substitute half to turn on/off types of debug message 3.679 +#define DEBUG( bool, msg) \ 3.680 + if( bool){ printf(msg); fflush(stdin);} 3.681 +#define DEBUG1( bool, msg, param) \ 3.682 + if(bool){printf(msg, param); fflush(stdin);} 3.683 +#define DEBUG2( bool, msg, p1, p2) \ 3.684 + if(bool) {printf(msg, p1, p2); fflush(stdin);} 3.685 + 3.686 +#define ERROR(msg) printf(msg); 3.687 +#define ERROR1(msg, param) printf(msg, param); 3.688 +#define ERROR2(msg, p1, p2) printf(msg, p1, p2); 3.689 + 3.690 +//=========================== STATS ======================= 3.691 + 3.692 + //when MEAS__TIME_STAMP_SUSP is defined, causes code to be inserted and 3.693 + // compiled-in that saves the low part of the time stamp count just before 3.694 + // suspending a processor and just after resuming that processor. It is 3.695 + // saved into a field added to VirtProcr. Have to sanity-check for 3.696 + // rollover of low portion into high portion. 3.697 +//#define MEAS__TIME_STAMP_SUSP 3.698 +//#define MEAS__TIME_MASTER 3.699 +//#define MEAS__TIME_PLUGIN 3.700 +//#define MEAS__TIME_MALLOC 3.701 +//#define MEAS__TIME_MASTER_LOCK 3.702 +//#define MEAS__NUM_TIMES_TO_RUN 100000 3.703 + 3.704 + //For code that calculates normalization-offset between TSC counts of 3.705 + // different cores. 3.706 +//#define NUM_TSC_ROUND_TRIPS 10 3.707 + 3.708 +#define MEAS__PERF_COUNTERS 3.709 +#define DETECT_DEPENDENCIES 3.710 + 3.711 +//========================= Hardware related Constants ===================== 3.712 + //This value is the number of hardware threads in the shared memory 3.713 + // machine 3.714 +//#define NUM_CORES 8 3.715 + 3.716 + // tradeoff amortizing master fixed overhead vs imbalance potential 3.717 + // when work-stealing, can make bigger, at risk of losing cache affinity 3.718 +#define NUM_SCHED_SLOTS 5 3.719 + 3.720 +#define MIN_WORK_UNIT_CYCLES 20000 3.721 + 3.722 +#define MASTERLOCK_RETRIES 10000 3.723 + 3.724 + // stack size in virtual processors created 3.725 +#define VIRT_PROCR_STACK_SIZE 0x8000 /* 32K */ 3.726 + 3.727 + // memory for VMS__malloc 3.728 +#define MALLOC_ADDITIONAL_MEM_FROM_OS_SIZE 0x10000000 /* 256M */ 3.729 + 3.730 +#define CACHE_LINE 64 3.731 +#define PAGE_SIZE 4096 3.732 + 3.733 + 3.734 +//============================== 3.735 + 3.736 +#define SUCCESS 0 3.737 + 3.738 +#define writeVMSQ writePrivQ 3.739 +#define readVMSQ readPrivQ 3.740 +#define makeVMSQ makeVMSPrivQ 3.741 +#define numInVMSQ numInPrivQ 3.742 +#define VMSQueueStruc PrivQueueStruc 3.743 + 3.744 + 3.745 + 3.746 +//=========================================================================== 3.747 +typedef unsigned long long TSCount; 3.748 + 3.749 +typedef struct _SchedSlot SchedSlot; 3.750 +typedef struct _VMSReqst VMSReqst; 3.751 +typedef struct _VirtProcr VirtProcr; 3.752 +typedef struct _IntervalProbe IntervalProbe; 3.753 +typedef struct _GateStruc GateStruc; 3.754 + 3.755 + 3.756 +typedef VirtProcr * (*SlaveScheduler) ( void *, int ); //semEnv, coreIdx 3.757 +typedef void (*RequestHandler) ( VirtProcr *, void * ); //prWReqst, semEnv 3.758 +typedef void (*VirtProcrFnPtr) ( void *, VirtProcr * ); //initData, animPr 3.759 +typedef void VirtProcrFn ( void *, VirtProcr * ); //initData, animPr 3.760 +typedef void (*ResumePrFnPtr) ( VirtProcr *, void * ); 3.761 + 3.762 + 3.763 +//============= Requests =========== 3.764 +// 3.765 + 3.766 +enum VMSReqstType //avoid starting enums at 0, for debug reasons 3.767 + { 3.768 + semantic = 1, 3.769 + createReq, 3.770 + dissipate, 3.771 + VMSSemantic //goes with VMSSemReqst below 3.772 + }; 3.773 + 3.774 +struct _VMSReqst 3.775 + { 3.776 + enum VMSReqstType reqType;//used for dissipate and in future for IO requests 3.777 + void *semReqData; 3.778 + 3.779 + VMSReqst *nextReqst; 3.780 + }; 3.781 +//VMSReqst 3.782 + 3.783 +enum VMSSemReqstType //These are equivalent to semantic requests, but for 3.784 + { // VMS's services available directly to app, like OS 3.785 + createProbe = 1, // and probe services -- like a VMS-wide built-in lang 3.786 + openFile, 3.787 + otherIO 3.788 + }; 3.789 + 3.790 +typedef struct 3.791 + { enum VMSSemReqstType reqType; 3.792 + VirtProcr *requestingPr; 3.793 + char *nameStr; //for create probe 3.794 + } 3.795 + VMSSemReq; 3.796 + 3.797 + 3.798 +//==================== Core data structures =================== 3.799 + 3.800 +struct _SchedSlot 3.801 + { 3.802 + int workIsDone; 3.803 + int needsProcrAssigned; 3.804 + VirtProcr *procrAssignedToSlot; 3.805 + }; 3.806 +//SchedSlot 3.807 + 3.808 +/*WARNING: re-arranging this data structure could cause VP switching 3.809 + * assembly code to fail -- hard-codes offsets of fields 3.810 + */ 3.811 +struct _VirtProcr 3.812 + { int procrID; //for debugging -- count up each time create 3.813 + int coreAnimatedBy; 3.814 + void *startOfStack; 3.815 + void *stackPtr; 3.816 + void *framePtr; 3.817 + void *nextInstrPt; 3.818 + 3.819 + void *coreLoopStartPt; //allows proto-runtime to be linked later 3.820 + void *coreLoopFramePtr; //restore before jmp back to core loop 3.821 + void *coreLoopStackPtr; //restore before jmp back to core loop 3.822 + 3.823 + void *initialData; 3.824 + 3.825 + SchedSlot *schedSlot; 3.826 + VMSReqst *requests; 3.827 + 3.828 + void *semanticData; //this livesUSE_GNU here for the life of VP 3.829 + void *dataRetFromReq;//values returned from plugin to VP go here 3.830 + 3.831 + //=========== MEASUREMENT STUFF ========== 3.832 + #ifdef MEAS__TIME_STAMP_SUSP 3.833 + unsigned int preSuspTSCLow; 3.834 + unsigned int postSuspTSCLow; 3.835 + #endif 3.836 + #ifdef MEAS__TIME_MASTER /* in VirtProcr because multiple masterVPs*/ 3.837 + unsigned int startMasterTSCLow;USE_GNU 3.838 + unsigned int endMasterTSCLow; 3.839 + #endif 3.840 + #ifdef MEAS__PERF_COUNTERS // 3.841 + CounterRecord** counter_history; 3.842 + PrivDynArrayInfo* counter_history_array_info; 3.843 + #endif 3.844 + //======================================== 3.845 + 3.846 + float64 createPtInSecs; //have space but don't use on some configs 3.847 + }; 3.848 +//VirtProcr 3.849 + 3.850 + 3.851 +/*WARNING: re-arranging this data structure could cause VP-switching 3.852 + * assembly code to fail -- hard-codes offsets of fields 3.853 + * (because -O3 messes with things otherwise) 3.854 + */ 3.855 +typedef struct 3.856 + { 3.857 + SlaveScheduler slaveScheduler; 3.858 + RequestHandler requestHandler; 3.859 + 3.860 + SchedSlot ***allSchedSlots; 3.861 + VMSQueueStruc **readyToAnimateQs; 3.862 + VirtProcr **masterVPs; 3.863 + 3.864 + void *semanticEnv; 3.865 + void *OSEventStruc; //for future, when add I/O to BLIS 3.866 + MallocProlog *freeListHead; 3.867 + int32 amtOfOutstandingMem; //total currently allocated 3.868 + 3.869 + void *coreLoopReturnPt;//addr to jump to to re-enter coreLoop 3.870 + 3.871 + int32 setupComplete; 3.872 + volatile int32 masterLock; 3.873 + 3.874 + int32 numMasterInARow[NUM_CORES];//detect back-to-back masterVP 3.875 + GateStruc *workStealingGates[ NUM_CORES ]; //concurrent work-steal 3.876 + int32 workStealingLock; 3.877 + 3.878 + int32 numProcrsCreated; //gives ordering to processor creation 3.879 + 3.880 + //=========== MEASUREMENT STUFF ============= 3.881 + IntervalProbe **intervalProbes; 3.882 + PrivDynArrayInfo *dynIntervalProbesInfo; 3.883 + HashTable *probeNameHashTbl; 3.884 + int32 masterCreateProbeID; 3.885 + float64 createPtInSecs; 3.886 + Histogram **measHists; 3.887 + PrivDynArrayInfo *measHistsInfo; 3.888 + #ifdef MEAS__TIME_PLUGIN 3.889 + Histogram *reqHdlrLowTimeHist; 3.890 + Histogram *reqHdlrHighTimeHist; 3.891 + #endif 3.892 + #ifdef MEAS__TIME_MALLOC 3.893 + Histogram *mallocTimeHist; 3.894 + Histogram *freeTimeHist; 3.895 + #endif 3.896 + #ifdef MEAS__TIME_MASTER_LOCK 3.897 + Histogram *masterLockLowTimeHist; 3.898 + Histogram *masterLockHighTimeHist; 3.899 + #endif 3.900 + #ifdef MEAS__PERF_COUNTERS 3.901 + int cycles_counter_fd[NUM_CORES]; 3.902 + int instrs_counter_fd[NUM_CORES]; 3.903 + FILE* counteroutput; 3.904 + #endif 3.905 + #ifdef DETECT_DEPENDENCIES 3.906 + Dependency** dependencies; 3.907 + PrivDynArrayInfo* dependenciesInfo; 3.908 + #endif 3.909 + #ifdef MEAS__PERF_COUNTERS // 3.910 + CounterRecord** counter_history; 3.911 + PrivDynArrayInfo* counter_history_array_info; 3.912 + #endif 3.913 + } 3.914 +MasterEnv; 3.915 + 3.916 +//========================= Extra Stuff Data Strucs ======================= 3.917 +typedef struct 3.918 + { 3.919 + 3.920 + } 3.921 +VMSExcp; 3.922 + 3.923 +struct _GateStruc 3.924 + { 3.925 + int32 gateClosed; 3.926 + int32 preGateProgress; 3.927 + int32 waitProgress; 3.928 + int32 exitProgress; 3.929 + }; 3.930 +//GateStruc 3.931 + 3.932 +//======================= OS Thread related =============================== 3.933 + 3.934 +void * coreLoop( void *paramsIn ); //standard PThreads fn prototype 3.935 +void * coreLoop_Seq( void *paramsIn ); //standard PThreads fn prototype 3.936 +void masterLoop( void *initData, VirtProcr *masterPr ); 3.937 + 3.938 + 3.939 +typedef struct 3.940 + { 3.941 + void *endThdPt; 3.942 + unsigned int coreNum; 3.943 + } 3.944 +ThdParams; 3.945 + 3.946 +pthread_t coreLoopThdHandles[ NUM_CORES ]; //pthread's virt-procr state 3.947 +ThdParams *coreLoopThdParams [ NUM_CORES ]; 3.948 +pthread_mutex_t suspendLock; 3.949 +pthread_cond_t suspend_cond; 3.950 + 3.951 + 3.952 + 3.953 +//===================== Global Vars =================== 3.954 + 3.955 +volatile MasterEnv *_VMSMasterEnv; 3.956 + 3.957 + 3.958 + 3.959 + 3.960 +//=========================== Function Prototypes ========================= 3.961 + 3.962 + 3.963 +//========== Setup and shutdown ========== 3.964 +void 3.965 +VMS__init(); 3.966 + 3.967 +void 3.968 +VMS__init_Seq(); 3.969 + 3.970 +void 3.971 +VMS__start_the_work_then_wait_until_done(); 3.972 + 3.973 +void 3.974 +VMS__start_the_work_then_wait_until_done_Seq(); 3.975 + 3.976 +inline VirtProcr * 3.977 +VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData ); 3.978 + 3.979 +void 3.980 +VMS__dissipate_procr( VirtProcr *procrToDissipate ); 3.981 + 3.982 + //Use this to create processor inside entry point & other places outside 3.983 + // the VMS system boundary (IE, not run in slave nor Master) 3.984 +VirtProcr * 3.985 +VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData ); 3.986 + 3.987 +void 3.988 +VMS_ext__dissipate_procr( VirtProcr *procrToDissipate ); 3.989 + 3.990 +void 3.991 +VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData ); 3.992 + 3.993 +void 3.994 +VMS__shutdown(); 3.995 + 3.996 +void 3.997 +VMS__cleanup_at_end_of_shutdown(); 3.998 + 3.999 +void * 3.1000 +VMS__give_sem_env_for( VirtProcr *animPr ); 3.1001 + 3.1002 + 3.1003 +//============== Request Related =============== 3.1004 + 3.1005 +void 3.1006 +VMS__suspend_procr( VirtProcr *callingPr ); 3.1007 + 3.1008 +inline void 3.1009 +VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData, VirtProcr *callingPr ); 3.1010 + 3.1011 +/*inline*/ __attribute__ ((noinline)) void 3.1012 +VMS__send_sem_request( void *semReqData, VirtProcr *callingPr ); 3.1013 + 3.1014 +void 3.1015 +VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr ); 3.1016 + 3.1017 +void /*inline**/ __attribute__ ((noinline)) 3.1018 +VMS__send_dissipate_req( VirtProcr *prToDissipate ); 3.1019 + 3.1020 +/*inline**/ __attribute__ ((noinline)) void 3.1021 +VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr ); 3.1022 + 3.1023 +VMSReqst * 3.1024 +VMS__take_next_request_out_of( VirtProcr *procrWithReq ); 3.1025 + 3.1026 +inline void * 3.1027 +VMS__take_sem_reqst_from( VMSReqst *req ); 3.1028 + 3.1029 +void inline 3.1030 +VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv, 3.1031 + ResumePrFnPtr resumePrFnPtr ); 3.1032 + 3.1033 +//======================== STATS ====================== 3.1034 + 3.1035 +//===== RDTSC wrapper ===== //Also runs with x86_64 code 3.1036 + 3.1037 +#define saveTimeStampCountInto(low, high) \ 3.1038 + asm volatile("RDTSC; \ 3.1039 + movl %%eax, %0; \ 3.1040 + movl %%edx, %1;" \ 3.1041 + /* outputs */ : "=m" (low), "=m" (high)\ 3.1042 + /* inputs */ : \ 3.1043 + /* clobber */ : "%eax", "%edx" \ 3.1044 + ); 3.1045 + 3.1046 +#define saveLowTimeStampCountInto(low) \ 3.1047 + asm volatile("RDTSC; \ 3.1048 + movl %%eax, %0;" \ 3.1049 + /* outputs */ : "=m" (low) \ 3.1050 + /* inputs */ : \ 3.1051 + /* clobber */ : "%eax", "%edx" \ 3.1052 + ); 3.1053 + 3.1054 +//==================== 3.1055 +#define makeAMeasHist( idx, name, numBins, startVal, binWidth ) \ 3.1056 + makeHighestDynArrayIndexBeAtLeast( _VMSMasterEnv->measHistsInfo, idx ); \ 3.1057 + _VMSMasterEnv->measHists[idx] = \ 3.1058 + makeFixedBinHist( numBins, startVal, binWidth, name ); 3.1059 + 3.1060 +#define saveCyclesAndInstrs(core,cycles,instrs) do{ \ 3.1061 + int cycles_fd = _VMSMasterEnv->cycles_counter_fd[core]; \ 3.1062 + int instrs_fd = _VMSMasterEnv->instrs_counter_fd[core]; \ 3.1063 + int nread; \ 3.1064 + \ 3.1065 + nread = read(cycles_fd,&(cycles),sizeof(cycles)); \ 3.1066 + if(nread<0){ \ 3.1067 + perror("Error reading cycles counter"); \ 3.1068 + cycles = 0; \ 3.1069 + } \ 3.1070 + \ 3.1071 + nread = read(instrs_fd,&(instrs),sizeof(instrs)); \ 3.1072 + if(nread<0){ \ 3.1073 + perror("Error reading cycles counter"); \ 3.1074 + instrs = 0; \ 3.1075 + } \ 3.1076 +} while (0) 3.1077 + 3.1078 +#define getReturnAddressBeforeLibraryCall(vp_ptr, res_ptr) do{ \ 3.1079 +void* frame_ptr0 = vp_ptr->framePtr; \ 3.1080 +void* frame_ptr1 = *((void**)frame_ptr0); \ 3.1081 +void* frame_ptr2 = *((void**)frame_ptr1); \ 3.1082 +void* frame_ptr3 = *((void**)frame_ptr2); \ 3.1083 +void* ret_addr = *((void**)frame_ptr3 + 1); \ 3.1084 +*res_ptr = ret_addr; \ 3.1085 +} while (0) 3.1086 + 3.1087 +#define MEAS__SUB_CREATE /*turn on/off subtraction of create from plugin*/ 3.1088 + 3.1089 +#ifdef VPTHREAD 3.1090 + 3.1091 +//VPThread 3.1092 +#define createHistIdx 0 3.1093 +#define mutexLockHistIdx 1 3.1094 +#define mutexUnlockHistIdx 2 3.1095 +#define condWaitHistIdx 3 3.1096 +#define condSignalHistIdx 4 3.1097 + 3.1098 +#define MakeTheMeasHists() \ 3.1099 + _VMSMasterEnv->measHistsInfo = \ 3.1100 + makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->measHists), 200); \ 3.1101 + makeAMeasHist( createHistIdx, "create", 250, 0, 100 ) \ 3.1102 + makeAMeasHist( mutexLockHistIdx, "mutex_lock", 50, 0, 100 ) \ 3.1103 + makeAMeasHist( mutexUnlockHistIdx, "mutex_unlock", 50, 0, 100 ) \ 3.1104 + makeAMeasHist( condWaitHistIdx, "cond_wait", 50, 0, 100 ) \ 3.1105 + makeAMeasHist( condSignalHistIdx, "cond_signal", 50, 0, 100 ) 3.1106 + 3.1107 +#endif 3.1108 + 3.1109 + 3.1110 +#ifdef VCILK 3.1111 + 3.1112 +//VCilk 3.1113 +#define spawnHistIdx 0 3.1114 +#define syncHistIdx 1 3.1115 + 3.1116 +#define MakeTheMeasHists() \ 3.1117 + _VMSMasterEnv->measHistsInfo = \ 3.1118 + makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->measHists), 200); \ 3.1119 + makeAMeasHist( spawnHistIdx, "Spawn", 50, 0, 200 ) \ 3.1120 + makeAMeasHist( syncHistIdx, "Sync", 50, 0, 200 ) 3.1121 + 3.1122 + 3.1123 +#endif 3.1124 + 3.1125 +#ifdef SSR 3.1126 + 3.1127 +//SSR 3.1128 +#define SendFromToHistIdx 0 3.1129 +#define SendOfTypeHistIdx 1 3.1130 +#define ReceiveFromToHistIdx 2 3.1131 +#define ReceiveOfTypeHistIdx 3 3.1132 + 3.1133 +#define MakeTheMeasHists() \ 3.1134 + _VMSMasterEnv->measHistsInfo = \ 3.1135 + makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->measHists), 200); \ 3.1136 + makeAMeasHist( SendFromToHistIdx, "SendFromTo", 50, 0, 100 ) \ 3.1137 + makeAMeasHist( SendOfTypeHistIdx, "SendOfType", 50, 0, 100 ) \ 3.1138 + makeAMeasHist( ReceiveFromToHistIdx,"ReceiveFromTo", 50, 0, 100 ) \ 3.1139 + makeAMeasHist( ReceiveOfTypeHistIdx,"ReceiveOfType", 50, 0, 100 ) 3.1140 + 3.1141 +#endif 3.1142 + 3.1143 +//=========================================================================== 3.1144 +//VPThread 3.1145 + 3.1146 + 3.1147 +#define Meas_startCreate \ 3.1148 + int32 startStamp, endStamp; \ 3.1149 + saveLowTimeStampCountInto( startStamp ); \ 3.1150 + 3.1151 +#define Meas_endCreate \ 3.1152 + saveLowTimeStampCountInto( endStamp ); \ 3.1153 + addIntervalToHist( startStamp, endStamp, \ 3.1154 + _VMSMasterEnv->measHists[ createHistIdx ] ); 3.1155 + 3.1156 +#define Meas_startMutexLock \ 3.1157 + int32 startStamp, endStamp; \ 3.1158 + saveLowTimeStampCountInto( startStamp ); \ 3.1159 + 3.1160 +#define Meas_endMutexLock \ 3.1161 + saveLowTimeStampCountInto( endStamp ); \ 3.1162 + addIntervalToHist( startStamp, endStamp, \ 3.1163 + _VMSMasterEnv->measHists[ mutexLockHistIdx ] ); 3.1164 + 3.1165 +#define Meas_startMutexUnlock \ 3.1166 + int32 startStamp, endStamp; \ 3.1167 + saveLowTimeStampCountInto( startStamp ); \ 3.1168 + 3.1169 +#define Meas_endMutexUnlock \ 3.1170 + saveLowTimeStampCountInto( endStamp ); \ 3.1171 + addIntervalToHist( startStamp, endStamp, \ 3.1172 + _VMSMasterEnv->measHists[ mutexUnlockHistIdx ] ); 3.1173 + 3.1174 +#define Meas_startCondWait \ 3.1175 + int32 startStamp, endStamp; \ 3.1176 + saveLowTimeStampCountInto( startStamp ); \ 3.1177 + 3.1178 +#define Meas_endCondWait \ 3.1179 + saveLowTimeStampCountInto( endStamp ); \ 3.1180 + addIntervalToHist( startStamp, endStamp, \ 3.1181 + _VMSMasterEnv->measHists[ condWaitHistIdx ] ); 3.1182 + 3.1183 +#define Meas_startCondSignal \ 3.1184 + int32 startStamp, endStamp; \ 3.1185 + saveLowTimeStampCountInto( startStamp ); \ 3.1186 + 3.1187 +#define Meas_endCondSignal \ 3.1188 + saveLowTimeStampCountInto( endStamp ); \ 3.1189 + addIntervalToHist( startStamp, endStamp, \ 3.1190 + _VMSMasterEnv->measHists[ condSignalHistIdx ] ); 3.1191 + 3.1192 +//=========================================================================== 3.1193 +// VCilk 3.1194 +#define Meas_startSpawn \ 3.1195 + int32 startStamp, endStamp; \ 3.1196 + saveLowTimeStampCountInto( startStamp ); \ 3.1197 + 3.1198 +#define Meas_endSpawn \ 3.1199 + saveLowTimeStampCountInto( endStamp ); \ 3.1200 + addIntervalToHist( startStamp, endStamp, \ 3.1201 + _VMSMasterEnv->measHists[ spawnHistIdx ] ); 3.1202 + 3.1203 +#define Meas_startSync \ 3.1204 + int32 startStamp, endStamp; \ 3.1205 + saveLowTimeStampCountInto( startStamp ); \ 3.1206 + 3.1207 +#define Meas_endSync \ 3.1208 + saveLowTimeStampCountInto( endStamp ); \ 3.1209 + addIntervalToHist( startStamp, endStamp, \ 3.1210 + _VMSMasterEnv->measHists[ syncHistIdx ] ); 3.1211 + 3.1212 +//=========================================================================== 3.1213 +// SSR 3.1214 +#define Meas_startSendFromTo \ 3.1215 + int32 startStamp, endStamp; \ 3.1216 + saveLowTimeStampCountInto( startStamp ); \ 3.1217 + 3.1218 +#define Meas_endSendFromTo \ 3.1219 + saveLowTimeStampCountInto( endStamp ); \ 3.1220 + addIntervalToHist( startStamp, endStamp, \ 3.1221 + _VMSMasterEnv->measHists[ SendFromToHistIdx ] ); 3.1222 + 3.1223 +#define Meas_startSendOfType \ 3.1224 + int32 startStamp, endStamp; \ 3.1225 + saveLowTimeStampCountInto( startStamp ); \ 3.1226 + 3.1227 +#define Meas_endSendOfType \ 3.1228 + saveLowTimeStampCountInto( endStamp ); \ 3.1229 + addIntervalToHist( startStamp, endStamp, \ 3.1230 + _VMSMasterEnv->measHists[ SendOfTypeHistIdx ] ); 3.1231 + 3.1232 +#define Meas_startReceiveFromTo \ 3.1233 + int32 startStamp, endStamp; \ 3.1234 + saveLowTimeStampCountInto( startStamp ); \ 3.1235 + 3.1236 +#define Meas_endReceiveFromTo \ 3.1237 + saveLowTimeStampCountInto( endStamp ); \ 3.1238 + addIntervalToHist( startStamp, endStamp, \ 3.1239 + _VMSMasterEnv->measHists[ ReceiveFromToHistIdx ] ); 3.1240 + 3.1241 +#define Meas_startReceiveOfType \ 3.1242 + int32 startStamp, endStamp; \ 3.1243 + saveLowTimeStampCountInto( startStamp ); \ 3.1244 + 3.1245 +#define Meas_endReceiveOfType \ 3.1246 + saveLowTimeStampCountInto( endStamp ); \ 3.1247 + addIntervalToHist( startStamp, endStamp, \ 3.1248 + _VMSMasterEnv->measHists[ReceiveOfTypeHistIdx ] ); 3.1249 + 3.1250 +//===== 3.1251 + 3.1252 +#include "ProcrContext.h" 3.1253 +#include "probes.h" 3.1254 +#include "vutilities.h" 3.1255 + 3.1256 +#endif /* _VMS_H */ 3.1257 +