static int lua_likwid_pinProcess(lua_State* L)
{
int cpuID = luaL_checknumber(L,-2);
int silent = luaL_checknumber(L,-1);
luaL_argcheck(L, cpuID >= 0, 1, "CPU ID must be greater or equal 0");
if (affinity_isInitialized == 0)
{
affinity_init();
affinity_isInitialized = 1;
affinity = get_affinityDomains();
}
affinity_pinProcess(cpuID);
if (!silent)
{
#ifdef COLOR
color_on(BRIGHT, COLOR);
#endif
printf("[likwid-pin] Main PID -> core %d - OK", cpuID);
#ifdef COLOR
color_reset();
#endif
printf("\n");
}
return 0;
}
static int cpustr_to_cpulist_scatter(bstring bcpustr, int* cpulist, int length)
{
topology_init();
CpuTopology_t cpuid_topology = get_cpuTopology();
affinity_init();
AffinityDomains_t affinity = get_affinityDomains();
char* cpustring = bstr2cstr(bcpustr, '\0');
if (bstrchrp(bcpustr, ':', 0) != BSTR_ERR)
{
int insert = 0;
int suitidx = 0;
int* suitable = (int*)malloc(affinity->numberOfAffinityDomains*sizeof(int));
if (!suitable)
{
bcstrfree(cpustring);
return -ENOMEM;
}
for (int i=0; i<affinity->numberOfAffinityDomains; i++)
{
if (bstrchrp(affinity->domains[i].tag, cpustring[0], 0) != BSTR_ERR)
{
suitable[suitidx] = i;
suitidx++;
}
}
int* sortedList = (int*) malloc(affinity->domains[suitable[0]].numberOfProcessors * sizeof(int));
if (!sortedList)
{
free(suitable);
bcstrfree(cpustring);
return -ENOMEM;
}
for (int off=0;off<affinity->domains[suitable[0]].numberOfProcessors;off++)
{
for(int i=0;i < suitidx; i++)
{
cpulist_sort(affinity->domains[suitable[i]].processorList, sortedList, affinity->domains[suitable[i]].numberOfProcessors);
cpulist[insert] = sortedList[off];
insert++;
if (insert == length)
goto scatter_done;
}
}
scatter_done:
bcstrfree(cpustring);
free(sortedList);
free(suitable);
return insert;
}
bcstrfree(cpustring);
return 0;
}
static int cpu_in_domain(int domainidx, int cpu)
{
affinity_init();
AffinityDomains_t affinity = get_affinityDomains();
for (int i=0;i<affinity->domains[domainidx].numberOfProcessors; i++)
{
if (cpu == affinity->domains[domainidx].processorList[i])
{
return 1;
}
}
return 0;
}
int test_affinityinit()
{
int i = 0;
topology_init();
CpuTopology_t cputopo = get_cpuTopology();
numa_init();
affinity_init();
AffinityDomains_t doms = get_affinityDomains();
if (doms == NULL)
goto fail;
if (doms->numberOfSocketDomains != cputopo->numSockets)
goto fail;
if (doms->numberOfNumaDomains == 0)
goto fail;
if (doms->numberOfProcessorsPerSocket == 0)
goto fail;
if (doms->numberOfAffinityDomains == 0)
goto fail;
if (doms->numberOfCacheDomains == 0)
goto fail;
if (doms->numberOfCoresPerCache == 0)
goto fail;
if (doms->numberOfProcessorsPerCache == 0)
goto fail;
if (doms->numberOfProcessorsPerCache < doms->numberOfCoresPerCache)
goto fail;
if (doms->domains == NULL)
goto fail;
for (i = 0; i < doms->numberOfAffinityDomains; i++)
{
if (doms->domains[i].numberOfProcessors == 0)
goto fail;
if (doms->domains[i].numberOfCores == 0)
goto fail;
if (doms->domains[i].numberOfProcessors < doms->domains[i].numberOfCores)
goto fail;
if (doms->domains[i].processorList == NULL)
goto fail;
}
affinity_finalize();
topology_finalize();
return 1;
fail:
affinity_finalize();
topology_finalize();
return 0;
}
static int cpuexpr_to_list(bstring bcpustr, bstring prefix, int* list, int length)
{
topology_init();
CpuTopology_t cpuid_topology = get_cpuTopology();
affinity_init();
AffinityDomains_t affinity = get_affinityDomains();
struct bstrList* strlist = bstrListCreate();
strlist = bsplit(bcpustr, ',');
int oldinsert = 0;
int insert = 0;
for (int i=0;i < strlist->qty; i++)
{
bstring newstr = bstrcpy(prefix);
bconcat(newstr, strlist->entry[i]);
oldinsert = insert;
for (int j = 0; j < affinity->numberOfAffinityDomains; j++)
{
if (bstrcmp(affinity->domains[j].tag, newstr) == 0)
{
list[insert] = atoi(bdata(strlist->entry[i]));
insert++;
if (insert == length)
goto list_done;
break;
}
}
if (insert == oldinsert)
{
fprintf(stderr,"Domain %s cannot be found\n", bdata(newstr));
}
bdestroy(newstr);
}
list_done:
bstrListDestroy(strlist);
return insert;
}
static int lua_likwid_getAffinityInfo(lua_State* L)
{
int i,j;
if (topology_isInitialized == 0)
{
topology_init();
topology_isInitialized = 1;
cpuinfo = get_cpuInfo();
cputopo = get_cpuTopology();
}
if ((topology_isInitialized) && (cpuinfo == NULL))
{
cpuinfo = get_cpuInfo();
}
if ((topology_isInitialized) && (cputopo == NULL))
{
cputopo = get_cpuTopology();
}
if (numa_isInitialized == 0)
{
if (numa_init() == 0)
{
numa_isInitialized = 1;
numainfo = get_numaTopology();
}
}
if ((numa_isInitialized) && (numainfo == NULL))
{
numainfo = get_numaTopology();
}
if (affinity_isInitialized == 0)
{
affinity_init();
affinity_isInitialized = 1;
affinity = get_affinityDomains();
}
if ((affinity_isInitialized) && (affinity == NULL))
{
affinity = get_affinityDomains();
}
if (!affinity)
{
lua_pushstring(L,"Cannot initialize affinity groups");
lua_error(L);
}
lua_newtable(L);
lua_pushstring(L,"numberOfAffinityDomains");
lua_pushunsigned(L,affinity->numberOfAffinityDomains);
lua_settable(L,-3);
lua_pushstring(L,"numberOfSocketDomains");
lua_pushunsigned(L,affinity->numberOfSocketDomains);
lua_settable(L,-3);
lua_pushstring(L,"numberOfNumaDomains");
lua_pushunsigned(L,affinity->numberOfNumaDomains);
lua_settable(L,-3);
lua_pushstring(L,"numberOfProcessorsPerSocket");
lua_pushunsigned(L,affinity->numberOfProcessorsPerSocket);
lua_settable(L,-3);
lua_pushstring(L,"numberOfCacheDomains");
lua_pushunsigned(L,affinity->numberOfCacheDomains);
lua_settable(L,-3);
lua_pushstring(L,"numberOfCoresPerCache");
lua_pushunsigned(L,affinity->numberOfCoresPerCache);
lua_settable(L,-3);
lua_pushstring(L,"numberOfProcessorsPerCache");
lua_pushunsigned(L,affinity->numberOfProcessorsPerCache);
lua_settable(L,-3);
lua_pushstring(L,"domains");
lua_newtable(L);
for(i=0;i<affinity->numberOfAffinityDomains;i++)
{
lua_pushunsigned(L, i+1);
lua_newtable(L);
lua_pushstring(L,"tag");
lua_pushstring(L,bdata(affinity->domains[i].tag));
lua_settable(L,-3);
lua_pushstring(L,"numberOfProcessors");
lua_pushunsigned(L,affinity->domains[i].numberOfProcessors);
lua_settable(L,-3);
lua_pushstring(L,"numberOfCores");
lua_pushunsigned(L,affinity->domains[i].numberOfCores);
lua_settable(L,-3);
lua_pushstring(L,"processorList");
lua_newtable(L);
for(j=0;j<affinity->domains[i].numberOfProcessors;j++)
{
lua_pushunsigned(L,j+1);
lua_pushunsigned(L,affinity->domains[i].processorList[j]);
lua_settable(L,-3);
}
lua_settable(L,-3);
lua_settable(L,-3);
}
lua_settable(L,-3);
return 1;
}
static int lua_likwid_getNumaInfo(lua_State* L)
{
uint32_t i,j;
if (topology_isInitialized == 0)
{
topology_init();
topology_isInitialized = 1;
cpuinfo = get_cpuInfo();
cputopo = get_cpuTopology();
}
if ((topology_isInitialized) && (cpuinfo == NULL))
{
cpuinfo = get_cpuInfo();
}
if ((topology_isInitialized) && (cputopo == NULL))
{
cputopo = get_cpuTopology();
}
if (numa_isInitialized == 0)
{
if (numa_init() == 0)
{
numa_isInitialized = 1;
numainfo = get_numaTopology();
}
else
{
lua_newtable(L);
lua_pushstring(L,"numberOfNodes");
lua_pushunsigned(L,0);
lua_settable(L,-3);
lua_pushstring(L,"nodes");
lua_newtable(L);
lua_settable(L,-3);
return 1;
}
}
if ((numa_isInitialized) && (numainfo == NULL))
{
numainfo = get_numaTopology();
}
if (affinity_isInitialized == 0)
{
affinity_init();
affinity_isInitialized = 1;
affinity = get_affinityDomains();
}
if ((affinity_isInitialized) && (affinity == NULL))
{
affinity = get_affinityDomains();
}
lua_newtable(L);
lua_pushstring(L,"numberOfNodes");
lua_pushunsigned(L,numainfo->numberOfNodes);
lua_settable(L,-3);
lua_pushstring(L,"nodes");
lua_newtable(L);
for(i=0;i<numainfo->numberOfNodes;i++)
{
lua_pushinteger(L, i+1);
lua_newtable(L);
lua_pushstring(L,"id");
lua_pushunsigned(L,numainfo->nodes[i].id);
lua_settable(L,-3);
lua_pushstring(L,"totalMemory");
lua_pushunsigned(L,numainfo->nodes[i].totalMemory);
lua_settable(L,-3);
lua_pushstring(L,"freeMemory");
lua_pushunsigned(L,numainfo->nodes[i].freeMemory);
lua_settable(L,-3);
lua_pushstring(L,"numberOfProcessors");
lua_pushunsigned(L,numainfo->nodes[i].numberOfProcessors);
lua_settable(L,-3);
lua_pushstring(L,"numberOfDistances");
lua_pushunsigned(L,numainfo->nodes[i].numberOfDistances);
lua_settable(L,-3);
lua_pushstring(L,"processors");
lua_newtable(L);
for(j=0;j<numainfo->nodes[i].numberOfProcessors;j++)
{
lua_pushunsigned(L,j+1);
lua_pushunsigned(L,numainfo->nodes[i].processors[j]);
lua_settable(L,-3);
}
lua_settable(L,-3);
/*lua_pushstring(L,"processorsCompact");
lua_newtable(L);
for(j=0;j<numa->nodes[i].numberOfProcessors;j++)
{
lua_pushunsigned(L,j);
lua_pushunsigned(L,numa->nodes[i].processorsCompact[j]);
lua_settable(L,-3);
}
lua_settable(L,-3);*/
lua_pushstring(L,"distances");
lua_newtable(L);
for(j=0;j<numainfo->nodes[i].numberOfDistances;j++)
{
lua_pushinteger(L,j+1);
lua_newtable(L);
lua_pushinteger(L,j);
lua_pushunsigned(L,numainfo->nodes[i].distances[j]);
lua_settable(L,-3);
lua_settable(L,-3);
}
lua_settable(L,-3);
lua_settable(L,-3);
}
lua_settable(L,-3);
return 1;
}
void
allocator_init(int numVectors)
{
allocList = (allocation*) malloc(numVectors * sizeof(allocation));
domains = get_affinityDomains();
}
static int cpustr_to_cpulist_physical(bstring bcpustr, int* cpulist, int length)
{
topology_init();
CpuTopology_t cpuid_topology = get_cpuTopology();
affinity_init();
AffinityDomains_t affinity = get_affinityDomains();
bstring bdomain;
bstring blist;
int domainidx = -1;
if (bstrchrp(bcpustr, ':', 0) != BSTR_ERR)
{
struct bstrList* strlist = bstrListCreate();
strlist = bsplit(bcpustr, ':');
bdomain = bstrcpy(strlist->entry[0]);
blist = bstrcpy(strlist->entry[1]);
bstrListDestroy(strlist);
}
else
{
bdomain = bformat("N");
blist = bstrcpy(bcpustr);
}
for (int i=0; i<affinity->numberOfAffinityDomains; i++)
{
if (bstrcmp(bdomain, affinity->domains[i].tag) == 0)
{
domainidx = i;
break;
}
}
if (domainidx < 0)
{
fprintf(stderr, "Cannot find domain %s\n", bdata(bdomain));
bdestroy(bdomain);
bdestroy(blist);
return 0;
}
struct bstrList* strlist = bstrListCreate();
strlist = bsplit(blist, ',');
int insert = 0;
for (int i=0;i< strlist->qty; i++)
{
if (bstrchrp(strlist->entry[i], '-', 0) != BSTR_ERR)
{
struct bstrList* indexlist = bstrListCreate();
indexlist = bsplit(strlist->entry[i], '-');
if (atoi(bdata(indexlist->entry[0])) <= atoi(bdata(indexlist->entry[1])))
{
for (int j=atoi(bdata(indexlist->entry[0])); j<=atoi(bdata(indexlist->entry[1]));j++)
{
if (cpu_in_domain(domainidx, j))
{
cpulist[insert] = j;
insert++;
if (insert == length)
{
bstrListDestroy(indexlist);
goto physical_done;
}
}
else
{
fprintf(stderr, "CPU %d not in domain %s\n", j, bdata(affinity->domains[domainidx].tag));
}
}
}
else
{
for (int j=atoi(bdata(indexlist->entry[0])); j>=atoi(bdata(indexlist->entry[1]));j--)
{
if (cpu_in_domain(domainidx, j))
{
cpulist[insert] = j;
insert++;
if (insert == length)
{
bstrListDestroy(indexlist);
goto physical_done;
}
}
else
{
fprintf(stderr, "CPU %d not in domain %s\n", j, bdata(affinity->domains[domainidx].tag));
}
}
}
bstrListDestroy(indexlist);
}
else
{
int cpu = atoi(bdata(strlist->entry[i]));
if (cpu_in_domain(domainidx, cpu))
{
cpulist[insert] = cpu;
insert++;
if (insert == length)
{
goto physical_done;
}
}
else
{
fprintf(stderr, "CPU %d not in domain %s\n", cpu, bdata(affinity->domains[domainidx].tag));
}
}
}
physical_done:
bstrListDestroy(strlist);
bdestroy(bdomain);
bdestroy(blist);
return insert;
}
static int cpustr_to_cpulist_logical(bstring bcpustr, int* cpulist, int length)
{
topology_init();
CpuTopology_t cpuid_topology = get_cpuTopology();
affinity_init();
AffinityDomains_t affinity = get_affinityDomains();
int domainidx = -1;
bstring bdomain;
bstring blist;
if (bstrchrp(bcpustr, 'L', 0) != 0)
{
fprintf(stderr, "Not a valid CPU expression\n");
return 0;
}
struct bstrList* strlist = bstrListCreate();
strlist = bsplit(bcpustr, ':');
if (strlist->qty != 3)
{
fprintf(stderr, "ERROR: Invalid expression, should look like L:<domain>:<indexlist> or be in a cpuset\n");
bstrListDestroy(strlist);
return 0;
}
bdomain = bstrcpy(strlist->entry[1]);
blist = bstrcpy(strlist->entry[2]);
bstrListDestroy(strlist);
for (int i=0; i<affinity->numberOfAffinityDomains; i++)
{
if (bstrcmp(bdomain, affinity->domains[i].tag) == 0)
{
domainidx = i;
break;
}
}
if (domainidx < 0)
{
fprintf(stderr, "Cannot find domain %s\n", bdata(bdomain));
return 0;
}
int *inlist = malloc(affinity->domains[domainidx].numberOfProcessors * sizeof(int));
if (inlist == NULL)
{
return -ENOMEM;
}
int ret = cpulist_sort(affinity->domains[domainidx].processorList, inlist, affinity->domains[domainidx].numberOfProcessors);
strlist = bstrListCreate();
strlist = bsplit(blist, ',');
int insert = 0;
for (int i=0; i< strlist->qty; i++)
{
if (bstrchrp(strlist->entry[i], '-', 0) != BSTR_ERR)
{
struct bstrList* indexlist = bstrListCreate();
indexlist = bsplit(strlist->entry[i], '-');
if (atoi(bdata(indexlist->entry[0])) <= atoi(bdata(indexlist->entry[1])))
{
for (int j=atoi(bdata(indexlist->entry[0])); j<=atoi(bdata(indexlist->entry[1]));j++)
{
cpulist[insert] = inlist[j];
insert++;
if (insert == length)
{
bstrListDestroy(indexlist);
goto logical_done;
}
}
}
else
{
for (int j=atoi(bdata(indexlist->entry[0])); j>=atoi(bdata(indexlist->entry[1]));j--)
{
cpulist[insert] = inlist[j];
insert++;
if (insert == length)
{
bstrListDestroy(indexlist);
goto logical_done;
}
}
}
bstrListDestroy(indexlist);
}
else
{
cpulist[insert] = inlist[atoi(bdata(strlist->entry[i]))];
insert++;
if (insert == length)
{
goto logical_done;
}
}
}
logical_done:
free(inlist);
bstrListDestroy(strlist);
return insert;
}
static int cpustr_to_cpulist_expression(bstring bcpustr, int* cpulist, int length)
{
topology_init();
CpuTopology_t cpuid_topology = get_cpuTopology();
affinity_init();
AffinityDomains_t affinity = get_affinityDomains();
bstring bdomain;
int domainidx = -1;
int count = 0;
int stride = 0;
int chunk = 0;
if (bstrchrp(bcpustr, 'E', 0) != 0)
{
fprintf(stderr, "Not a valid CPU expression\n");
return 0;
}
struct bstrList* strlist = bstrListCreate();
strlist = bsplit(bcpustr, ':');
if (strlist->qty == 3)
{
bdomain = bstrcpy(strlist->entry[1]);
count = atoi(bdata(strlist->entry[2]));
stride = 1;
chunk = 1;
}
else if (strlist->qty == 5)
{
bdomain = bstrcpy(strlist->entry[1]);
count = atoi(bdata(strlist->entry[2]));
chunk = atoi(bdata(strlist->entry[3]));
stride = atoi(bdata(strlist->entry[4]));
}
for (int i=0; i<affinity->numberOfAffinityDomains; i++)
{
if (bstrcmp(bdomain, affinity->domains[i].tag) == 0)
{
domainidx = i;
break;
}
}
if (domainidx < 0)
{
fprintf(stderr, "Cannot find domain %s\n", bdata(bdomain));
bstrListDestroy(strlist);
return 0;
}
int offset = 0;
int insert = 0;
for (int i=0;i<count;i++)
{
for (int j=0;j<chunk && offset+j<affinity->domains[domainidx].numberOfProcessors;j++)
{
cpulist[insert] = affinity->domains[domainidx].processorList[offset + j];
insert++;
if (insert == length)
goto expression_done;
}
offset += stride;
if (offset >= affinity->domains[domainidx].numberOfProcessors)
{
offset = 0;
}
if (insert >= count)
goto expression_done;
}
bstrListDestroy(strlist);
return 0;
expression_done:
bstrListDestroy(strlist);
return insert;
}
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;
}
void bstr_to_workgroup(Workgroup* group, const_bstring str, DataType type, int numberOfStreams)
{
uint32_t i;
int parseStreams = 0;
bstring threadInfo;
bstring streams= bformat("0");
struct bstrList* tokens;
struct bstrList* subtokens;
AffinityDomains_t domains;
AffinityDomain* domain = NULL;
/* split the workgroup into the thread and the streams part */
tokens = bsplit(str,'-');
if (tokens->qty == 2)
{
threadInfo = bstrcpy(tokens->entry[0]);
streams = bstrcpy(tokens->entry[1]);
parseStreams = 1;
}
else if (tokens->qty == 1)
{
threadInfo = bstrcpy(tokens->entry[0]);
}
else
{
fprintf(stderr, "Error in parsing workgroup string\n");
}
bstrListDestroy (tokens);
tokens = bsplit(threadInfo,':');
if (tokens->qty == 5)
{
uint32_t maxNumThreads;
int chunksize;
int stride;
int counter;
int currentId = 0;
int startId = 0;
domains = get_affinityDomains();
for (i = 0; i < domains->numberOfAffinityDomains; i++)
{
if (bstrcmp(domains->domains[i].tag, tokens->entry[0]) == BSTR_OK)
{
domain = &(domains->domains[i]);
break;
}
}
if (domain == NULL)
{
fprintf(stderr, "Error: Domain %s not available on current machine.\nTry likwid-bench -p for supported domains.",
bdata(tokens->entry[0]));
exit(EXIT_FAILURE);
}
group->size = bstr_to_doubleSize(tokens->entry[1], type);
group->numberOfThreads = str2int(bdata(tokens->entry[2]));
chunksize = str2int(bdata(tokens->entry[3]));
stride = str2int(bdata(tokens->entry[4]));
maxNumThreads = ceil((double)domain->numberOfProcessors / stride) * chunksize;
if (group->numberOfThreads > maxNumThreads)
{
fprintf(stderr, "Warning: More threads (%d) requested than CPUs in domain %s fulfilling expression (%d).\n",
group->numberOfThreads, bdata(tokens->entry[0]), maxNumThreads);
}
group->processorIds = (int*) malloc(group->numberOfThreads * sizeof(int));
counter = chunksize;
counter = 0;
for (int j=0; j<group->numberOfThreads; j+=chunksize)
{
for(i=0;i<chunksize && j+i<group->numberOfThreads ;i++)
{
group->processorIds[startId++] = domain->processorList[counter+i];
}
counter += stride;
if (counter >= domain->numberOfProcessors)
{
counter = counter-domain->numberOfProcessors;
}
}
}
else if (tokens->qty == 3)
{
domains = get_affinityDomains();
for (i = 0; i < domains->numberOfAffinityDomains; i++)
{
if (bstrcmp(domains->domains[i].tag, tokens->entry[0]) == BSTR_OK)
{
domain = &(domains->domains[i]);
break;
}
}
if (domain == NULL)
{
fprintf(stderr, "Error: Domain %s not available on current machine.\nTry likwid-bench -p for supported domains.",
bdata(tokens->entry[0]));
exit(EXIT_FAILURE);
}
group->size = bstr_to_doubleSize(tokens->entry[1], type);
group->numberOfThreads = str2int(bdata(tokens->entry[2]));
if (group->numberOfThreads > domain->numberOfProcessors)
{
fprintf(stderr, "Warning: More threads (%d) requested than CPUs in domain %s (%d).\n",
group->numberOfThreads, bdata(tokens->entry[0]), domain->numberOfProcessors);
}
group->processorIds = (int*) malloc(group->numberOfThreads * sizeof(int));
for (i=0; i<group->numberOfThreads; i++)
{
group->processorIds[i] = domain->processorList[i % domain->numberOfProcessors];
}
}
else if (tokens->qty == 2)
{
domains = get_affinityDomains();
for (i = 0; i < domains->numberOfAffinityDomains; i++)
{
if (bstrcmp(domains->domains[i].tag, tokens->entry[0]) == BSTR_OK)
{
domain = &(domains->domains[i]);
break;
}
}
if (domain == NULL)
{
fprintf(stderr, "Error: Domain %s not available on current machine.\nTry likwid-bench -p for supported domains.",
bdata(tokens->entry[0]));
exit(EXIT_FAILURE);
}
group->size = bstr_to_doubleSize(tokens->entry[1], type);
group->numberOfThreads = domain->numberOfProcessors;
group->processorIds = (int*) malloc(group->numberOfThreads * sizeof(int));
for (i=0; i<group->numberOfThreads; i++)
{
group->processorIds[i] = domain->processorList[i];
}
}
else
{
fprintf(stderr, "Error in parsing workgroup string\n");
}
bstrListDestroy(tokens);
/* parse stream list */
if (parseStreams)
{
tokens = bsplit(streams,',');
if (tokens->qty < numberOfStreams)
{
fprintf(stderr, "Error: Testcase requires at least %d streams\n", numberOfStreams);
}
group->streams = (Stream*) malloc(numberOfStreams * sizeof(Stream));
for (i=0;i<(uint32_t) tokens->qty;i++)
{
subtokens = bsplit(tokens->entry[i],':');
if ( subtokens->qty == 3 )
{
int index = str2int(bdata(subtokens->entry[0]));
if (index >= numberOfStreams)
{
fprintf(stderr, "Error: Stream index %d out of range\n",index);
}
group->streams[index].domain = bstrcpy(subtokens->entry[1]);
group->streams[index].offset = str2int(bdata(subtokens->entry[2]));
}
else if ( subtokens->qty == 2 )
{
int index = str2int(bdata(subtokens->entry[0]));
if (index >= numberOfStreams)
{
fprintf(stderr, "Error: Stream index %d out of range\n",index);
}
group->streams[index].domain = bstrcpy(subtokens->entry[1]);
group->streams[index].offset = 0;
}
else
{
fprintf(stderr, "Error: Cannot parse stream placement defintition in %s\n", bdata(str));
}
bstrListDestroy(subtokens);
}
bstrListDestroy(tokens);
}
else
{
group->streams = (Stream*) malloc(numberOfStreams * sizeof(Stream));
for (i=0; i< (uint32_t)numberOfStreams; i++)
{
group->streams[i].domain = domain->tag;
group->streams[i].offset = 0;
}
}
group->size /= numberOfStreams;
return;
}