int test_timercpuclock_noinit()
{
uint64_t cyc = timer_getCpuClock();
if (cyc != 0)
return 0;
return 1;
}
static int lua_likwid_getCpuClock(lua_State* L)
{
if (timer_isInitialized == 0)
{
timer_init();
timer_isInitialized = 1;
}
lua_pushnumber(L,timer_getCpuClock());
return 1;
}
int test_timercpuclock()
{
timer_init();
uint64_t cyc = timer_getCpuClock();
if (cyc == 0)
return 0;
timer_finalize();
return 1;
fail:
timer_finalize();
return 0;
}
int test_timerinit()
{
timer_init();
uint64_t clock = timer_getCpuClock();
if (clock == 0)
goto fail;
timer_finalize();
return 1;
fail:
timer_finalize();
return 0;
}
int test_timerprint_stop()
{
TimerData timer;
timer_init();
timer_reset(&timer);
timer_start(&timer);
timer_stop(&timer);
double time = timer_print(&timer);
if (time > 1)
goto fail;
if (time == 0)
goto fail;
uint64_t cycles = timer_printCycles(&timer);
if (cycles == 0)
goto fail;
if (cycles > timer_getCpuClock())
goto fail;
timer_finalize();
return 1;
fail:
timer_finalize();
return 0;
}
int main(int argc, char** argv)
{
int iter = 100;
uint32_t i;
uint32_t j;
int globalNumberOfThreads = 0;
int optPrintDomains = 0;
int c;
ThreadUserData myData;
bstring testcase = bfromcstr("none");
uint32_t numberOfWorkgroups = 0;
int tmp = 0;
double time;
const TestCase* test = NULL;
Workgroup* currentWorkgroup = NULL;
Workgroup* groups = NULL;
cpuid_init();
numa_init();
affinity_init();
/* Handling of command line options */
if (argc == 1) { HELP_MSG; }
while ((c = getopt (argc, argv, "g:w:t:i:l:aphv")) != -1) {
switch (c)
{
case 'h':
HELP_MSG;
exit (EXIT_SUCCESS);
case 'v':
VERSION_MSG;
exit (EXIT_SUCCESS);
case 'a':
printf(TESTS"\n");
exit (EXIT_SUCCESS);
case 'w':
tmp--;
if (tmp == -1)
{
fprintf (stderr, "More workgroups configured than allocated!\n");
return EXIT_FAILURE;
}
if (!test)
{
fprintf (stderr, "You need to specify a test case first!\n");
return EXIT_FAILURE;
}
testcase = bfromcstr(optarg);
currentWorkgroup = groups+tmp; /*FIXME*/
bstr_to_workgroup(currentWorkgroup, testcase, test->type, test->streams);
bdestroy(testcase);
for (i=0; i< test->streams; i++)
{
if (currentWorkgroup->streams[i].offset%test->stride)
{
fprintf (stderr, "Stream %d: offset is not a multiple of stride!\n",i);
return EXIT_FAILURE;
}
allocator_allocateVector(&(currentWorkgroup->streams[i].ptr),
PAGE_ALIGNMENT,
currentWorkgroup->size,
currentWorkgroup->streams[i].offset,
test->type,
currentWorkgroup->streams[i].domain);
}
break;
case 'i':
iter = atoi(optarg);
break;
case 'l':
testcase = bfromcstr(optarg);
for (i=0; i<NUMKERNELS; i++)
{
if (biseqcstr(testcase, kernels[i].name))
{
test = kernels+i;
break;
}
}
if (biseqcstr(testcase,"none"))
{
fprintf (stderr, "Unknown test case %s\n",optarg);
return EXIT_FAILURE;
}
else
{
printf("Name: %s\n",test->name);
printf("Number of streams: %d\n",test->streams);
printf("Loop stride: %d\n",test->stride);
printf("Flops: %d\n",test->flops);
printf("Bytes: %d\n",test->bytes);
switch (test->type)
{
case SINGLE:
printf("Data Type: Single precision float\n");
break;
case DOUBLE:
printf("Data Type: Double precision float\n");
break;
}
}
bdestroy(testcase);
exit (EXIT_SUCCESS);
break;
case 'p':
optPrintDomains = 1;
break;
case 'g':
numberOfWorkgroups = atoi(optarg);
allocator_init(numberOfWorkgroups * MAX_STREAMS);
tmp = numberOfWorkgroups;
groups = (Workgroup*) malloc(numberOfWorkgroups*sizeof(Workgroup));
break;
case 't':
testcase = bfromcstr(optarg);
for (i=0; i<NUMKERNELS; i++)
{
if (biseqcstr(testcase, kernels[i].name))
{
test = kernels+i;
break;
}
}
if (biseqcstr(testcase,"none"))
{
fprintf (stderr, "Unknown test case %s\n",optarg);
return EXIT_FAILURE;
}
bdestroy(testcase);
break;
case '?':
if (isprint (optopt))
fprintf (stderr, "Unknown option `-%c'.\n", optopt);
else
fprintf (stderr,
"Unknown option character `\\x%x'.\n",
optopt);
return EXIT_FAILURE;
default:
HELP_MSG;
}
}
if (optPrintDomains)
{
affinity_printDomains();
exit (EXIT_SUCCESS);
}
timer_init();
/* :WARNING:05/04/2010 08:58:05 AM:jt: At the moment the thread
* module only allows equally sized thread groups*/
for (i=0; i<numberOfWorkgroups; i++)
{
globalNumberOfThreads += groups[i].numberOfThreads;
}
threads_init(globalNumberOfThreads);
threads_createGroups(numberOfWorkgroups);
/* we configure global barriers only */
barrier_init(1);
barrier_registerGroup(globalNumberOfThreads);
#ifdef PERFMON
printf("Using likwid\n");
likwid_markerInit();
#endif
/* initialize data structures for threads */
for (i=0; i<numberOfWorkgroups; i++)
{
myData.iter = iter;
myData.size = groups[i].size;
myData.test = test;
myData.numberOfThreads = groups[i].numberOfThreads;
myData.processors = (int*) malloc(myData.numberOfThreads * sizeof(int));
myData.streams = (void**) malloc(test->streams * sizeof(void*));
for (j=0; j<groups[i].numberOfThreads; j++)
{
myData.processors[j] = groups[i].processorIds[j];
}
for (j=0; j< test->streams; j++)
{
myData.streams[j] = groups[i].streams[j].ptr;
}
threads_registerDataGroup(i, &myData, copyThreadData);
free(myData.processors);
free(myData.streams);
}
printf(HLINE);
printf("LIKWID MICRO BENCHMARK\n");
printf("Test: %s\n",test->name);
printf(HLINE);
printf("Using %d work groups\n",numberOfWorkgroups);
printf("Using %d threads\n",globalNumberOfThreads);
printf(HLINE);
threads_create(runTest);
threads_destroy();
allocator_finalize();
time = (double) threads_data[0].cycles / (double) timer_getCpuClock();
printf("Cycles: %llu \n", LLU_CAST threads_data[0].cycles);
printf("Iterations: %llu \n", LLU_CAST iter);
printf("Size: %d \n", currentWorkgroup->size );
printf("Vectorlength: %d \n", threads_data[0].data.size);
printf("Time: %e sec\n", time);
printf("MFlops/s:\t%.2f\n",
1.0E-06 * ((double) numberOfWorkgroups * iter * currentWorkgroup->size * test->flops/ time));
printf("MByte/s:\t%.2f\n",
1.0E-06 * ( (double) numberOfWorkgroups * iter * currentWorkgroup->size * test->bytes/ time));
printf("Cycles per update:\t%f\n",
((double) threads_data[0].cycles / (double) (iter * threads_data[0].data.size)));
switch ( test->type )
{
case SINGLE:
printf("Cycles per cacheline:\t%f\n",
(16.0 * (double) threads_data[0].cycles / (double) (iter * threads_data[0].data.size)));
break;
case DOUBLE:
printf("Cycles per cacheline:\t%f\n",
(8.0 * (double) threads_data[0].cycles / (double) (iter * threads_data[0].data.size)));
break;
}
printf(HLINE);
#ifdef PERFMON
likwid_markerClose();
#endif
return EXIT_SUCCESS;
}
int main (int argc, char** argv)
{
int socket_fd = -1;
int optInfo = 0;
int optClock = 0;
int optStethoscope = 0;
int optSockets = 0;
double runtime;
int hasDRAM = 0;
int c;
bstring argString;
bstring eventString = bfromcstr("CLOCK");
int numSockets=1;
int numThreads=0;
int threadsSockets[MAX_NUM_NODES*2];
int threads[MAX_NUM_THREADS];
threadsSockets[0] = 0;
if (argc == 1)
{
HELP_MSG;
exit (EXIT_SUCCESS);
}
while ((c = getopt (argc, argv, "+c:hiM:ps:v")) != -1)
{
switch (c)
{
case 'c':
CHECK_OPTION_STRING;
numSockets = bstr_to_cpuset_physical((uint32_t*) threadsSockets, argString);
bdestroy(argString);
optSockets = 1;
break;
case 'h':
HELP_MSG;
exit (EXIT_SUCCESS);
case 'i':
optInfo = 1;
break;
case 'M': /* Set MSR Access mode */
CHECK_OPTION_STRING;
accessClient_setaccessmode(str2int((char*) argString->data));
bdestroy(argString);
break;
case 'p':
optClock = 1;
break;
case 's':
CHECK_OPTION_STRING;
optStethoscope = str2int((char*) argString->data);
bdestroy(argString);
break;
case 'v':
VERSION_MSG;
exit (EXIT_SUCCESS);
case '?':
if (optopt == 's' || optopt == 'M' || optopt == 'c')
{
HELP_MSG;
}
else if (isprint (optopt))
{
fprintf (stderr, "Unknown option `-%c'.\n", optopt);
}
else
{
fprintf (stderr,
"Unknown option character `\\x%x'.\n",
optopt);
}
exit( EXIT_FAILURE);
default:
HELP_MSG;
exit (EXIT_SUCCESS);
}
}
if (!lock_check())
{
fprintf(stderr,"Access to performance counters is locked.\n");
exit(EXIT_FAILURE);
}
if (optClock && optind == argc)
{
fprintf(stderr,"Commandline option -p requires an executable.\n");
exit(EXIT_FAILURE);
}
if (optSockets && !optStethoscope && optind == argc)
{
fprintf(stderr,"Commandline option -c requires an executable if not used in combination with -s.\n");
exit(EXIT_FAILURE);
}
if (cpuid_init() == EXIT_FAILURE)
{
fprintf(stderr, "CPU not supported\n");
exit(EXIT_FAILURE);
}
if (numSockets > cpuid_topology.numSockets)
{
fprintf(stderr, "System has only %d sockets but %d are given on commandline\n",
cpuid_topology.numSockets, numSockets);
exit(EXIT_FAILURE);
}
numa_init(); /* consider NUMA node as power unit for the moment */
accessClient_init(&socket_fd);
msr_init(socket_fd);
timer_init();
/* check for supported processors */
if ((cpuid_info.model == SANDYBRIDGE_EP) ||
(cpuid_info.model == SANDYBRIDGE) ||
(cpuid_info.model == IVYBRIDGE) ||
(cpuid_info.model == IVYBRIDGE_EP) ||
(cpuid_info.model == HASWELL) ||
(cpuid_info.model == NEHALEM_BLOOMFIELD) ||
(cpuid_info.model == NEHALEM_LYNNFIELD) ||
(cpuid_info.model == NEHALEM_WESTMERE))
{
power_init(numa_info.nodes[0].processors[0]);
}
else
{
fprintf (stderr, "Query Turbo Mode only supported on Intel Nehalem/Westmere/SandyBridge/IvyBridge/Haswell processors!\n");
exit(EXIT_FAILURE);
}
double clock = (double) timer_getCpuClock();
printf(HLINE);
printf("CPU name:\t%s \n",cpuid_info.name);
printf("CPU clock:\t%3.2f GHz \n", (float) clock * 1.E-09);
printf(HLINE);
if (optInfo)
{
if (power_info.turbo.numSteps != 0)
{
printf("Base clock:\t%.2f MHz \n", power_info.baseFrequency );
printf("Minimal clock:\t%.2f MHz \n", power_info.minFrequency );
printf("Turbo Boost Steps:\n");
for (int i=0; i < power_info.turbo.numSteps; i++ )
{
printf("C%d %.2f MHz \n",i+1, power_info.turbo.steps[i] );
}
}
printf(HLINE);
}
if (cpuid_info.model == SANDYBRIDGE_EP)
{
hasDRAM = 1;
}
else if ((cpuid_info.model != SANDYBRIDGE) &&
(cpuid_info.model != SANDYBRIDGE_EP) &&
(cpuid_info.model != IVYBRIDGE) &&
(cpuid_info.model != IVYBRIDGE_EP) &&
(cpuid_info.model != HASWELL))
{
fprintf (stderr, "RAPL not supported on this processor!\n");
exit(EXIT_FAILURE);
}
if (optInfo)
{
printf("Thermal Spec Power: %g Watts \n", power_info.tdp );
printf("Minimum Power: %g Watts \n", power_info.minPower);
printf("Maximum Power: %g Watts \n", power_info.maxPower);
printf("Maximum Time Window: %g micro sec \n", power_info.maxTimeWindow);
printf(HLINE);
exit(EXIT_SUCCESS);
}
if (optClock)
{
affinity_init();
argString = bformat("S%u:0-%u", threadsSockets[0], cpuid_topology.numCoresPerSocket-1);
for (int i=1; i<numSockets; i++)
{
bstring tExpr = bformat("@S%u:0-%u", threadsSockets[i], cpuid_topology.numCoresPerSocket-1);
bconcat(argString, tExpr);
}
numThreads = bstr_to_cpuset(threads, argString);
bdestroy(argString);
perfmon_init(numThreads, threads, stdout);
perfmon_setupEventSet(eventString, NULL);
}
{
PowerData pDataPkg[MAX_NUM_NODES*2];
PowerData pDataDram[MAX_NUM_NODES*2];
printf("Measure on sockets: %d", threadsSockets[0]);
for (int i=1; i<numSockets; i++)
{
printf(", %d", threadsSockets[i]);
}
printf("\n");
if (optStethoscope)
{
if (optClock)
{
perfmon_startCounters();
}
else
{
for (int i=0; i<numSockets; i++)
{
int cpuId = numa_info.nodes[threadsSockets[i]].processors[0];
if (hasDRAM) power_start(pDataDram+i, cpuId, DRAM);
power_start(pDataPkg+i, cpuId, PKG);
}
}
sleep(optStethoscope);
if (optClock)
{
perfmon_stopCounters();
perfmon_printCounterResults();
perfmon_finalize();
}
else
{
for (int i=0; i<numSockets; i++)
{
int cpuId = numa_info.nodes[threadsSockets[i]].processors[0];
power_stop(pDataPkg+i, cpuId, PKG);
if (hasDRAM) power_stop(pDataDram+i, cpuId, DRAM);
}
}
runtime = (double) optStethoscope;
}
else
{
TimerData time;
argv += optind;
bstring exeString = bfromcstr(argv[0]);
for (int i=1; i<(argc-optind); i++)
{
bconchar(exeString, ' ');
bcatcstr(exeString, argv[i]);
}
printf("%s\n",bdata(exeString));
if (optClock)
{
perfmon_startCounters();
}
else
{
for (int i=0; i<numSockets; i++)
{
int cpuId = numa_info.nodes[threadsSockets[i]].processors[0];
if (hasDRAM) power_start(pDataDram+i, cpuId, DRAM);
power_start(pDataPkg+i, cpuId, PKG);
}
timer_start(&time);
}
if (system(bdata(exeString)) == EOF)
{
fprintf(stderr, "Failed to execute %s!\n", bdata(exeString));
exit(EXIT_FAILURE);
}
if (optClock)
{
perfmon_stopCounters();
perfmon_printCounterResults();
perfmon_finalize();
}
else
{
timer_stop(&time);
for (int i=0; i<numSockets; i++)
{
int cpuId = numa_info.nodes[threadsSockets[i]].processors[0];
power_stop(pDataPkg+i, cpuId, PKG);
if (hasDRAM) power_stop(pDataDram+i, cpuId, DRAM);
}
runtime = timer_print(&time);
}
}
if (!optClock)
{
printf("Runtime: %g second \n",runtime);
printf(HLINE);
for (int i=0; i<numSockets; i++)
{
printf("Socket %d\n",threadsSockets[i]);
printf("Domain: PKG \n");
printf("Energy consumed: %g Joules \n", power_printEnergy(pDataPkg+i));
printf("Power consumed: %g Watts \n", power_printEnergy(pDataPkg+i) / runtime );
if (hasDRAM)
{
printf("Domain: DRAM \n");
printf("Energy consumed: %g Joules \n", power_printEnergy(pDataDram+i));
printf("Power consumed: %g Watts \n", power_printEnergy(pDataDram+i) / runtime );
}
printf("\n");
}
}
}
#if 0
if ( cpuid_hasFeature(TM2) )
{
thermal_init(0);
printf("Current core temperatures:\n");
for (uint32_t i = 0; i < cpuid_topology.numCoresPerSocket; i++ )
{
printf("Core %d: %u C\n",
numa_info.nodes[socketId].processors[i],
thermal_read(numa_info.nodes[socketId].processors[i]));
}
}
#endif
msr_finalize();
return EXIT_SUCCESS;
}
int main(int argc, char** argv)
{
uint64_t iter = 100;
uint32_t i;
uint32_t j;
int globalNumberOfThreads = 0;
int optPrintDomains = 0;
int c;
ThreadUserData myData;
bstring testcase = bfromcstr("none");
uint64_t numberOfWorkgroups = 0;
int tmp = 0;
double time;
double cycPerUp = 0.0;
const TestCase* test = NULL;
uint64_t realSize = 0;
uint64_t realIter = 0;
uint64_t maxCycles = 0;
uint64_t minCycles = UINT64_MAX;
uint64_t cyclesClock = 0;
uint64_t demandIter = 0;
TimerData itertime;
Workgroup* currentWorkgroup = NULL;
Workgroup* groups = NULL;
uint32_t min_runtime = 1; /* 1s */
bstring HLINE = bfromcstr("");
binsertch(HLINE, 0, 80, '-');
binsertch(HLINE, 80, 1, '\n');
int (*ownprintf)(const char *format, ...);
ownprintf = &printf;
/* Handling of command line options */
if (argc == 1)
{
HELP_MSG;
exit(EXIT_SUCCESS);
}
while ((c = getopt (argc, argv, "w:t:s:l:aphvi:")) != -1) {
switch (c)
{
case 'h':
HELP_MSG;
exit (EXIT_SUCCESS);
case 'v':
VERSION_MSG;
exit (EXIT_SUCCESS);
case 'a':
ownprintf(TESTS"\n");
exit (EXIT_SUCCESS);
case 'w':
numberOfWorkgroups++;
break;
case 's':
min_runtime = atoi(optarg);
break;
case 'i':
demandIter = strtoul(optarg, NULL, 10);
if (demandIter <= 0)
{
fprintf (stderr, "Error: Iterations must be greater than 0\n");
return EXIT_FAILURE;
}
break;
case 'l':
bdestroy(testcase);
testcase = bfromcstr(optarg);
for (i=0; i<NUMKERNELS; i++)
{
if (biseqcstr(testcase, kernels[i].name))
{
test = kernels+i;
break;
}
}
if (test == NULL)
{
fprintf (stderr, "Error: Unknown test case %s\n",optarg);
return EXIT_FAILURE;
}
else
{
ownprintf("Name: %s\n",test->name);
ownprintf("Number of streams: %d\n",test->streams);
ownprintf("Loop stride: %d\n",test->stride);
ownprintf("Flops: %d\n",test->flops);
ownprintf("Bytes: %d\n",test->bytes);
switch (test->type)
{
case INT:
ownprintf("Data Type: Integer\n");
break;
case SINGLE:
ownprintf("Data Type: Single precision float\n");
break;
case DOUBLE:
ownprintf("Data Type: Double precision float\n");
break;
}
if (test->loads >= 0)
{
ownprintf("Load Ops: %d\n",test->loads);
}
if (test->stores >= 0)
{
ownprintf("Store Ops: %d\n",test->stores);
}
if (test->branches >= 0)
{
ownprintf("Branches: %d\n",test->branches);
}
if (test->instr_const >= 0)
{
ownprintf("Constant instructions: %d\n",test->instr_const);
}
if (test->instr_loop >= 0)
{
ownprintf("Loop instructions: %d\n",test->instr_loop);
}
}
bdestroy(testcase);
exit (EXIT_SUCCESS);
break;
case 'p':
optPrintDomains = 1;
break;
case 'g':
numberOfWorkgroups = LLU_CAST atol(optarg);
tmp = numberOfWorkgroups;
break;
case 't':
bdestroy(testcase);
testcase = bfromcstr(optarg);
for (i=0; i<NUMKERNELS; i++)
{
if (biseqcstr(testcase, kernels[i].name))
{
test = kernels+i;
break;
}
}
if (test == NULL)
{
fprintf (stderr, "Error: Unknown test case %s\n",optarg);
return EXIT_FAILURE;
}
bdestroy(testcase);
break;
case '?':
if (isprint (optopt))
fprintf (stderr, "Unknown option `-%c'.\n", optopt);
else
fprintf (stderr,
"Unknown option character `\\x%x'.\n",
optopt);
return EXIT_FAILURE;
default:
HELP_MSG;
}
}
if ((numberOfWorkgroups == 0) && (!optPrintDomains))
{
fprintf(stderr, "Error: At least one workgroup (-w) must be set on commandline\n");
exit (EXIT_FAILURE);
}
if (topology_init() != EXIT_SUCCESS)
{
fprintf(stderr, "Error: Unsupported processor!\n");
exit(EXIT_FAILURE);
}
if ((test == NULL) && (!optPrintDomains))
{
fprintf(stderr, "Unknown test case. Please check likwid-bench -a for available tests\n");
fprintf(stderr, "and select one using the -t commandline option\n");
exit(EXIT_FAILURE);
}
numa_init();
affinity_init();
timer_init();
if (optPrintDomains)
{
bdestroy(testcase);
AffinityDomains_t affinity = get_affinityDomains();
ownprintf("Number of Domains %d\n",affinity->numberOfAffinityDomains);
for (i=0; i < affinity->numberOfAffinityDomains; i++ )
{
ownprintf("Domain %d:\n",i);
ownprintf("\tTag %s:",bdata(affinity->domains[i].tag));
for ( uint32_t j=0; j < affinity->domains[i].numberOfProcessors; j++ )
{
ownprintf(" %d",affinity->domains[i].processorList[j]);
}
ownprintf("\n");
}
exit (EXIT_SUCCESS);
}
allocator_init(numberOfWorkgroups * MAX_STREAMS);
groups = (Workgroup*) malloc(numberOfWorkgroups*sizeof(Workgroup));
tmp = 0;
optind = 0;
while ((c = getopt (argc, argv, "w:t:s:l:i:aphv")) != -1)
{
switch (c)
{
case 'w':
currentWorkgroup = groups+tmp;
bstring groupstr = bfromcstr(optarg);
i = bstr_to_workgroup(currentWorkgroup, groupstr, test->type, test->streams);
bdestroy(groupstr);
if (i == 0)
{
for (i=0; i< test->streams; i++)
{
if (currentWorkgroup->streams[i].offset%test->stride)
{
fprintf (stderr, "Error: Stream %d: offset is not a multiple of stride!\n",i);
return EXIT_FAILURE;
}
allocator_allocateVector(&(currentWorkgroup->streams[i].ptr),
PAGE_ALIGNMENT,
currentWorkgroup->size,
currentWorkgroup->streams[i].offset,
test->type,
currentWorkgroup->streams[i].domain);
}
tmp++;
}
else
{
exit(EXIT_FAILURE);
}
break;
default:
continue;
break;
}
}
/* :WARNING:05/04/2010 08:58:05 AM:jt: At the moment the thread
* module only allows equally sized thread groups*/
for (i=0; i<numberOfWorkgroups; i++)
{
globalNumberOfThreads += groups[i].numberOfThreads;
}
ownprintf(bdata(HLINE));
ownprintf("LIKWID MICRO BENCHMARK\n");
ownprintf("Test: %s\n",test->name);
ownprintf(bdata(HLINE));
ownprintf("Using %" PRIu64 " work groups\n",numberOfWorkgroups);
ownprintf("Using %d threads\n",globalNumberOfThreads);
ownprintf(bdata(HLINE));
threads_init(globalNumberOfThreads);
threads_createGroups(numberOfWorkgroups);
/* we configure global barriers only */
barrier_init(1);
barrier_registerGroup(globalNumberOfThreads);
cyclesClock = timer_getCycleClock();
#ifdef LIKWID_PERFMON
if (getenv("LIKWID_FILEPATH") != NULL)
{
ownprintf("Using Likwid Marker API\n");
}
LIKWID_MARKER_INIT;
ownprintf(bdata(HLINE));
#endif
/* initialize data structures for threads */
for (i=0; i<numberOfWorkgroups; i++)
{
myData.iter = iter;
if (demandIter > 0)
{
myData.iter = demandIter;
}
myData.min_runtime = min_runtime;
myData.size = groups[i].size;
myData.test = test;
myData.cycles = 0;
myData.numberOfThreads = groups[i].numberOfThreads;
myData.processors = (int*) malloc(myData.numberOfThreads * sizeof(int));
myData.streams = (void**) malloc(test->streams * sizeof(void*));
for (j=0; j<groups[i].numberOfThreads; j++)
{
myData.processors[j] = groups[i].processorIds[j];
}
for (j=0; j< test->streams; j++)
{
myData.streams[j] = groups[i].streams[j].ptr;
}
threads_registerDataGroup(i, &myData, copyThreadData);
free(myData.processors);
free(myData.streams);
}
if (demandIter == 0)
{
getIterSingle((void*) &threads_data[0]);
for (i=0; i<numberOfWorkgroups; i++)
{
iter = threads_updateIterations(i, demandIter);
}
}
#ifdef DEBUG_LIKWID
else
{
ownprintf("Using manually selected iterations per thread\n");
}
#endif
threads_create(runTest);
threads_join();
for (int i=0; i<globalNumberOfThreads; i++)
{
realSize += threads_data[i].data.size;
realIter += threads_data[i].data.iter;
if (threads_data[i].cycles > maxCycles)
{
maxCycles = threads_data[i].cycles;
}
if (threads_data[i].cycles < minCycles)
{
minCycles = threads_data[i].cycles;
}
}
time = (double) maxCycles / (double) cyclesClock;
ownprintf(bdata(HLINE));
ownprintf("Cycles:\t\t\t%" PRIu64 "\n", maxCycles);
ownprintf("CPU Clock:\t\t%" PRIu64 "\n", timer_getCpuClock());
ownprintf("Cycle Clock:\t\t%" PRIu64 "\n", cyclesClock);
ownprintf("Time:\t\t\t%e sec\n", time);
ownprintf("Iterations:\t\t%" PRIu64 "\n", realIter);
ownprintf("Iterations per thread:\t%" PRIu64 "\n",threads_data[0].data.iter);
ownprintf("Inner loop executions:\t%.0f\n", ((double)realSize)/((double)test->stride));
ownprintf("Size:\t\t\t%" PRIu64 "\n", realSize*test->bytes );
ownprintf("Size per thread:\t%" PRIu64 "\n", threads_data[0].data.size*test->bytes);
ownprintf("Number of Flops:\t%" PRIu64 "\n", (threads_data[0].data.iter * realSize * test->flops));
ownprintf("MFlops/s:\t\t%.2f\n",
1.0E-06 * ((double) threads_data[0].data.iter * realSize * test->flops/ time));
ownprintf("Data volume (Byte):\t%llu\n", LLU_CAST (threads_data[0].data.iter * realSize * test->bytes));
ownprintf("MByte/s:\t\t%.2f\n",
1.0E-06 * ( (double) threads_data[0].data.iter * realSize * test->bytes/ time));
cycPerUp = ((double) maxCycles / (double) (threads_data[0].data.iter * realSize));
ownprintf("Cycles per update:\t%f\n", cycPerUp);
switch ( test->type )
{
case INT:
case SINGLE:
ownprintf("Cycles per cacheline:\t%f\n", (16.0 * cycPerUp));
break;
case DOUBLE:
ownprintf("Cycles per cacheline:\t%f\n", (8.0 * cycPerUp));
break;
}
ownprintf("Loads per update:\t%ld\n", test->loads );
ownprintf("Stores per update:\t%ld\n", test->stores );
if ((test->loads > 0) && (test->stores > 0))
{
ownprintf("Load/store ratio:\t%.2f\n", ((double)test->loads)/((double)test->stores) );
}
if ((test->instr_loop > 0) && (test->instr_const > 0))
{
ownprintf("Instructions:\t\t%" PRIu64 "\n", LLU_CAST ((double)realSize/test->stride)*test->instr_loop*threads_data[0].data.iter + test->instr_const );
}
if (test->uops > 0)
{
ownprintf("UOPs:\t\t\t%" PRIu64 "\n", LLU_CAST ((double)realSize/test->stride)*test->uops*threads_data[0].data.iter);
}
ownprintf(bdata(HLINE));
threads_destroy(numberOfWorkgroups, test->streams);
allocator_finalize();
workgroups_destroy(&groups, numberOfWorkgroups, test->streams);
#ifdef LIKWID_PERFMON
if (getenv("LIKWID_FILEPATH") != NULL)
{
ownprintf("Writing Likwid Marker API results to file %s\n", getenv("LIKWID_FILEPATH"));
}
LIKWID_MARKER_CLOSE;
#endif
bdestroy(HLINE);
return EXIT_SUCCESS;
}
