PERF_TEST_P( EstimateAffine, EstimateAffine2D, ESTIMATE_PARAMS )
{
AffineParams params = GetParam();
const int n = get<0>(params);
const double confidence = get<1>(params);
const int method = get<2>(params);
const size_t refining = get<3>(params);
Mat aff(2, 3, CV_64F);
cv::randu(aff, -2., 2.);
// LMEDS can't handle more than 50% outliers (by design)
int m;
if (method == LMEDS)
m = 3*n/5;
else
m = 2*n/5;
const float shift_outl = 15.f;
const float noise_level = 20.f;
Mat fpts(1, n, CV_32FC2);
Mat tpts(1, n, CV_32FC2);
randu(fpts, 0., 100.);
transform(fpts, tpts, aff);
/* adding noise to some points */
Mat outliers = tpts.colRange(m, n);
outliers.reshape(1) += shift_outl;
Mat noise (outliers.size(), outliers.type());
randu(noise, 0., noise_level);
outliers += noise;
Mat aff_est;
vector<uchar> inliers (n);
warmup(inliers, WARMUP_WRITE);
warmup(fpts, WARMUP_READ);
warmup(tpts, WARMUP_READ);
TEST_CYCLE()
{
aff_est = estimateAffine2D(fpts, tpts, inliers, method, 3, 2000, confidence, refining);
}
// we already have accuracy tests
SANITY_CHECK_NOTHING();
}
double CPUBenchmark::getSystemCallOverhead(bool isCached) {
uint64_t sum = 0;
uint64_t start, end;
warmup();
for(int i = 0; i < TIMES; i++) {
// in order to reduce overhead, leave the time count inside the if block
if(isCached) {
start = rdtscStart();
// this one has cache
getpid();
end = rdtscEnd();
}
else {
start = rdtscStart();
// this one don't have cache
getppid();
end = rdtscEnd();
}
sum += (end - start);
}
return (double) sum / (double) TIMES;
}
int main(int argc, char** argv) {
int budget = atoi(argv[0]);
warmup();
int iterations = 0;
double start = secondtime();
double elapsed;
while (1) {
if (!sample()) {
abort();
}
iterations++;
elapsed = secondtime() - start;
if (elapsed > budget)
break;
}
double score = iterations / elapsed * 1000.0;
printf("%f\n", score);
}
uint64_t CPUBenchmark::getLoopOverhead(int times) {
uint64_t start, end;
warmup();
start = rdtscStart();
for(int i = 0; i < times; i++) {
// end loop to avoid new overhead
}
end = rdtscEnd();
return end - start;
}
double CPUBenchmark::getReadOverhead() {
uint64_t sum = 0;
uint64_t start, end;
warmup();
for(int i = 0; i < TIMES; i++) {
start = rdtscStart();
end = rdtscEnd();
sum += (end - start);
}
return (double)sum / (double)TIMES;
}
void test_BS (cl_uint sizeMin, cl_uint sizeMax, int warmup_num_iters, int run_num_iters,
size_t blockMin, size_t blockMax,
size_t gridMin, size_t gridMax,
char *filename, cl_float(*gen)(cl_float low, cl_float high))
{
BS_test_t t;
t.num = sizeMin;
t.blockSize = blockMin;
t.gridSize = gridMin;
t.Riskfree = R;
t.Volatility = V;
bool result = false;
FILE * csv = fopen(filename, "a");
if (csv == NULL)
fprintf(stderr, "can't open output file\n");
if (csv != NULL)
fprintf(csv, "size, block size, grid size, total time\n");
setup_cl(&t);
setup_buffers(&t, gen);
if (t.maxBlockSize < blockMin)
blockMin = t.maxBlockSize;
if (t.maxBlockSize < blockMax)
blockMax = t.maxBlockSize;
printf("blockMIN: %d blockMAX: %d t.maxBlockSize: %d\n", blockMin, blockMax, t.maxBlockSize);
warmup(&t, warmup_num_iters);
for (t.num = sizeMin; t.num <= sizeMax; t.num <<= 1) {
for (t.gridSize = gridMin; t.gridSize <= gridMax; t.gridSize += 16*1024) {
for (t.blockSize = blockMin; t.blockSize <= blockMax; t.blockSize += 64) {
if (t.gridSize % t.blockSize != 0)
continue;
setup_buffers(&t, gen);
result = run(&t, run_num_iters);
post(&t, run_num_iters, result);
if (result && csv != NULL)
output(&t, csv, run_num_iters);
cleanup(&t);
}
}
}
if (csv != NULL)
fclose(csv);
}
uint64_t CPUBenchmark::calculateThreadSwitchTime(){
uint64_t start;
uint64_t end;
pthread_t thread;
warmup();
// pipe(fd);
pthread_create(&thread, NULL, foo, &end);
start = rdtscStart();
pthread_join(thread, NULL);
// read(fd[0], (void*)&end, sizeof(uint64_t));
return end - start;
}
double CPUBenchmark::getKernelThreadCreationTime() {
double sum = 0;
uint64_t start, end;
pthread_t thread;
warmup();
for(int i = 0; i < TASK_OP_TIMES; i++) {
start = rdtscStart();
pthread_create(&thread, NULL, &threadStartRountine, NULL);
// make the main process to wait new thread
pthread_join(thread, NULL);
end = rdtscEnd();
sum += end - start;
}
return (double) sum / (double) TASK_OP_TIMES;
}
uint64_t CPUBenchmark::calculateProcessSwitchTime(int *fd){
uint64_t start;
uint64_t end;
pid_t cpid;
uint64_t result = 0;
warmup();
if ((cpid = fork()) != 0) {
start = rdtscStart();
wait(NULL);
read(fd[0], (void*)&end, sizeof(uint64_t));
}
else {
end = rdtscEnd();
write(fd[1], (void*)&end, sizeof(uint64_t));
exit(1);
}
if(end > start){
result = end - start;
}
return (result);
}
int begin(void)
{
const char *scheme = use_secure?"https":"http";
char *space;
static const char blanks[] = " ";
static const char stars[] = "**********************************";
i_session = ne_session_create(scheme, i_hostname, i_port);
CALL(init_session(i_session));
ne_hook_pre_send(i_session, i_pre_send, "X-Prestan");
space = ne_concat(i_path, "davtest/", NULL);
ne_delete(i_session, space);
if (ne_mkcol(i_session, space)) {
t_context("Could not create new collection `%s' for tests: %s\n"
"Server must allow `MKCOL %s' for tests to proceed",
space, ne_get_error(i_session), space);
return FAILHARD;
}
free(i_path);
i_path = space;
warmup();
printf("\nStart Testing %s:\n\n",pget_option.URL) ;
printf("\n%s%s\n", blanks, stars);
printf("\n%s* Number of Requests\t\t%d\n", blanks, pget_option.requests);
printf("\n%s* Number of Dead Properties\t%d\n",
blanks, pget_option.numprops);
printf("\n%s* Depth of Collection\t\t%d\n", blanks, pget_option.depth);
printf("\n%s* Width of Collection\t\t%d\n", blanks, pget_option.width);
printf("\n%s* Type of Methods\t\t%s\n", blanks, pget_option.methods);
printf("\n%s%s\n", blanks, stars);
printf("\n\n");
return OK;
}
int main(int argc, char *argv[]){
if(argc<=2){
usage(argv[0]);
exit(-1);
}
int pagesize = getpagesize();
CmdParam cmd;
CacheStat stat;
memset(&stat, 0, sizeof(stat));
memset(&cmd, 0, sizeof(cmd));
stat.isPrint = true; //default true
parseArgs(argc, argv, &cmd);
if(cmd.isDaemon)
daemonMe();
if(cmd.cmd_type == CLEAR){
clear(cmd.path, cmd.file_type);
}else if(cmd.cmd_type == STAT){
normalStat(cmd.path, cmd.file_type, &stat);
if(cmd.file_type == DIRECTORY){
fprintf(stdout, "\nTotal Cache of Directory:%s size:%s cached:%s\n",
cmd.path,
sizeFit(stat.pageCount*pagesize, buf1),
sizeFit(stat.inCache*pagesize, buf2));
}
}else if(cmd.cmd_type == RSTAT){
realStat(cmd.path, cmd.interval, cmd.file_type, cmd.isSuppress);
}else if(cmd.cmd_type == LOCK){
lock(cmd.path, cmd.file_type);
select(0, NULL, NULL, NULL, NULL);
}else if(cmd.cmd_type == WARM){
warmup(cmd.path, cmd.file_type);
}
return 0;
}
double CPUBenchmark::getProcessCreationTime() {
uint64_t sum = 0;
uint64_t start, end;
pid_t pid;
warmup();
for(int i = 0; i < TASK_OP_TIMES; i++) {
start = rdtscStart();
pid = fork();
if(pid == 0) {
// child process, just exit
exit(1);
}
else {
// parent process, wait child exit
wait(NULL);
end = rdtscEnd();
sum += (end - start);
}
}
return (double) sum / (double) TASK_OP_TIMES;
}
int main(int argc, char **argv)
{
#define test_A(i,j) test_A[(size_t)(j)*N+(i)]
#define test_A2(i,j) test_A2[(size_t)(j)*N+(i)]
int N,NB,w,LDA,BB;
size_t memsize; //bytes
int iam, nprocs, mydevice;
int ICTXT, nprow, npcol, myprow, mypcol;
int i_one = 1, i_zero = 0, i_negone = -1;
double d_one = 1.0, d_zero = 0.0, d_negone = -1.0;
int IASEED = 100;
/* printf("N=?\n");
scanf("%ld",&N);
printf("NB=?\n");
scanf("%d", &NB);
printf("width of Y panel=?\n");
scanf("%ld",&w);
*/
if(argc < 4){
printf("invalid arguments N NB memsize(M)\n");
exit(1);
}
N = atoi(argv[1]);
NB = atoi(argv[2]);
memsize = (size_t)atoi(argv[3])*1024*1024;
BB = (N + NB - 1) / NB;
w = memsize/sizeof(double)/BB/NB/NB - 1;
assert(w > 0);
LDA = N + 0; //padding
int do_io = (N <= NSIZE);
double llttime;
double gflops;
nprow = npcol = 1;
blacs_pinfo_(&iam, &nprocs);
blacs_get_(&i_negone, &i_zero, &ICTXT);
blacs_gridinit_(&ICTXT, "R", &nprow, &npcol);
blacs_gridinfo_(&ICTXT, &nprow, &npcol, &myprow, &mypcol);
#ifdef USE_MIC
#ifdef __INTEL_OFFLOAD
printf("offload compilation enabled\ninitialize each MIC\n");
offload_init(&iam, &mydevice);
#pragma offload target(mic:0)
{
mkl_peak_mem_usage(MKL_PEAK_MEM_ENABLE);
}
#else
if(isroot)
printf("offload compilation not enabled\n");
exit(0);
#endif
#else
#ifdef USE_CUBLASV2
{
cublasStatus_t cuStatus;
for(int r = 0; r < OOC_NTHREADS; r++){
cuStatus = cublasCreate(&worker_handle[r]);
assert(cuStatus == CUBLAS_STATUS_SUCCESS);
}
}
#else
cublasInit();
#endif
#endif
double *test_A = (double*)memalign(64,(size_t)LDA*N*sizeof(double)); // for chol
#ifdef VERIFY
double *test_A2 = (double*)memalign(64,(size_t)LDA*N*sizeof(double)); // for verify
#endif
/*Initialize A */
int i,j;
printf("Initialize A ... "); fflush(stdout);
llttime = MPI_Wtime();
pdmatgen(&ICTXT, "Symm", "Diag", &N,
&N, &NB, &NB,
test_A, &LDA, &i_zero, &i_zero,
&IASEED, &i_zero, &N, &i_zero, &N,
&myprow, &mypcol, &nprow, &npcol);
llttime = MPI_Wtime() - llttime;
printf("time %lf\n", llttime);
/*print test_A*/
if(do_io){
printf("Original A=\n\n");
matprint(test_A, N, LDA, 'A');
}
/*Use directed unblocked Cholesky factorization*/
/*
t1 = clock();
Test_dpotrf(test_A2,N);
t2 = clock();
printf ("time for unblocked Cholesky factorization on host %f \n",
((float) (t2 - t1)) / CLOCKS_PER_SEC);
*/
/*print test_A*/
/*
if(do_io){
printf("Unblocked result:\n\n");
matprint(test_A2,N,'L');
}
*/
/*Use tile algorithm*/
Quark *quark = QUARK_New(OOC_NTHREADS);
QUARK_DOT_DAG_Enable(quark, 0);
#ifdef USE_MIC
// mklmem(NB);
printf("QUARK MIC affinity binding\n");
QUARK_bind(quark);
printf("offload warm up\n");
warmup(quark);
#endif
QUARK_DOT_DAG_Enable(quark, quark_getenv_int("QUARK_DOT_DAG_ENABLE", 0));
printf("LLT start %lf\n", MPI_Wtime());
llttime = Cholesky(quark,test_A,N,NB,LDA,memsize);
printf("LLT end %lf\n", MPI_Wtime());
QUARK_Delete(quark);
#ifdef USE_MIC
offload_destroy();
#else
#ifdef USE_CUBLASV2
{
cublasStatus_t cuStatus;
for(int r = 0; r < OOC_NTHREADS; r++){
cuStatus = cublasDestroy(worker_handle[r]);
assert(cuStatus == CUBLAS_STATUS_SUCCESS);
}
}
#else
cublasShutdown();
#endif
#endif
gflops = (double) N;
gflops = gflops/3.0 + 0.5;
gflops = gflops*(double)(N)*(double)(N);
gflops = gflops/llttime/1024.0/1024.0/1024.0;
printf ("N NB memsize(MB) quark_pthreads time Gflops\n%d %d %lf %d %lf %lf\n",
N, NB, (double)memsize/1024/1024, OOC_NTHREADS, llttime, gflops);
#ifdef USE_MIC
#pragma offload target(mic:0)
{
memsize = mkl_peak_mem_usage(MKL_PEAK_MEM_RESET);
}
printf("mkl_peak_mem_usage %lf MB\n", (double)memsize/1024.0/1024.0);
#endif
/*Update and print L*/
if(do_io){
printf("L:\n\n");
matprint(test_A,N,LDA,'L');
}
#ifdef VERIFY
printf("Verify... ");
llttime = MPI_Wtime();
/*
* ------------------------
* check difference betwen
* test_A and test_A2
* ------------------------
*/
/*
{
double maxerr = 0;
double maxerr2 = 0;
for (j = 0; j < N; j++)
{
for (i = j; i < N; i++)
{
double err = (test_A (i, j) - test_A2 (i, j));
err = ABS (err);
maxerr = MAX (err, maxerr);
maxerr2 = maxerr2 + err * err;
};
};
maxerr2 = sqrt (ABS (maxerr2));
printf ("max difference between test_A and test_A2 %lf \n", maxerr);
printf ("L2 difference between test_A and test_A2 %lf \n", maxerr2);
};
*/
/*
* ------------------
* over-write test_A2
* ------------------
*/
pdmatgen(&ICTXT, "Symm", "Diag", &N,
&N, &NB, &NB,
test_A2, &LDA, &i_zero,
&i_zero, &IASEED, &i_zero, &N, &i_zero, &N,
&myprow, &mypcol, &nprow, &npcol);
/*
* ---------------------------------------
* after solve, test_A2 should be identity
* ---------------------------------------
*/
// test_A = chol(B) = L;
// test_A2 = B
// solve L*L'*X = B
// if L is correct, X is identity */
{
int uplo = 'L';
const char *uplo_char = ((uplo == (int) 'U')
|| (uplo == (int) 'u')) ? "U" : "L";
int info = 0;
int nrhs = N;
int LDA = N;
int ldb = N;
dpotrs(uplo_char, &N, &nrhs, test_A, &LDA, test_A2, &ldb, &info);
assert (info == 0);
}
{
double maxerr = 0;
double maxerr2 = 0;
for (j = 0; j < N; j++)
{
for (i = 0; i < N; i++)
{
double eyeij = (i == j) ? 1.0 : 0.0;
double err = (test_A2 (i, j) - eyeij);
err = ABS (err);
maxerr = MAX (maxerr, err);
maxerr2 = maxerr2 + err * err;
};
};
maxerr2 = sqrt (ABS (maxerr2));
printf("time %lf\n", MPI_Wtime() - llttime);
printf ("max error %lf \n", maxerr);
printf ("max L2 error %lf \n", maxerr2);
}
#endif
free(test_A);test_A=NULL;
#ifdef VERIFY
free(test_A2);test_A2=NULL;
#endif
blacs_gridexit_(&ICTXT);
blacs_exit_(&i_zero);
return 0;
#undef test_A
#undef test_A2
}
//#ifdef _MT
void runthreads(long sleep_cnt, int min_threads, int max_threads, int chperthread, int num_rounds)
{
thread_data *de_area = new thread_data[max_threads] ;
thread_data *pdea;
int nperthread ;
int sum_threads ;
unsigned long sum_allocs ;
unsigned long sum_frees ;
double duration ;
#ifdef __WIN32__
_LARGE_INTEGER ticks_per_sec, start_cnt, end_cnt;
#else
long ticks_per_sec ;
long start_cnt, end_cnt ;
#endif
_int64 ticks ;
double rate_1=0, rate_n ;
double reqd_space ;
ULONG used_space ;
int prevthreads ;
int i ;
QueryPerformanceFrequency( &ticks_per_sec ) ;
pdea = &de_area[0] ;
memset(&de_area[0], 0, sizeof(thread_data)) ;
prevthreads = 0 ;
for(num_threads=min_threads; num_threads <= max_threads; num_threads++ )
{
warmup(&blkp[prevthreads*chperthread], (num_threads-prevthreads)*chperthread );
nperthread = chperthread ;
stopflag = FALSE ;
for(i=0; i< num_threads; i++){
de_area[i].threadno = i+1 ;
de_area[i].NumBlocks = num_rounds*nperthread;
de_area[i].array = &blkp[i*nperthread] ;
de_area[i].blksize = &blksize[i*nperthread] ;
de_area[i].asize = nperthread ;
de_area[i].min_size = min_size ;
de_area[i].max_size = max_size ;
de_area[i].seed = lran2(&rgen) ; ;
de_area[i].finished = 0 ;
de_area[i].cAllocs = 0 ;
de_area[i].cFrees = 0 ;
de_area[i].cThreads = 0 ;
de_area[i].finished = FALSE ;
lran2_init(&de_area[i].rgen, de_area[i].seed) ;
#ifdef __WIN32__
_beginthread((void (__cdecl*)(void *)) exercise_heap, 0, &de_area[i]) ;
#else
_beginthread(exercise_heap, 0, &de_area[i]) ;
#endif
}
QueryPerformanceCounter( &start_cnt) ;
// printf ("Sleeping for %ld seconds.\n", sleep_cnt);
Sleep(sleep_cnt * 1000L) ;
stopflag = TRUE ;
for(i=0; i<num_threads; i++){
while( !de_area[i].finished ){
#ifdef __WIN32__
Sleep(1);
#elif defined(__SVR4)
thr_yield();
#else
sched_yield();
#endif
}
}
QueryPerformanceCounter( &end_cnt) ;
sum_frees = sum_allocs =0 ;
sum_threads = 0 ;
for(i=0;i< num_threads; i++){
sum_allocs += de_area[i].cAllocs ;
sum_frees += de_area[i].cFrees ;
sum_threads += de_area[i].cThreads ;
de_area[i].cAllocs = de_area[i].cFrees = 0;
}
#ifdef __WIN32__
ticks = end_cnt.QuadPart - start_cnt.QuadPart ;
duration = (double)ticks/ticks_per_sec.QuadPart ;
#else
ticks = end_cnt - start_cnt ;
duration = (double)ticks/ticks_per_sec ;
#endif
for( i=0; i<num_threads; i++){
if( !de_area[i].finished )
printf("Thread at %d not finished\n", i) ;
}
rate_n = sum_allocs/duration ;
if( rate_1 == 0){
rate_1 = rate_n ;
}
reqd_space = (0.5*(min_size+max_size)*num_threads*chperthread) ;
// used_space = CountReservedSpace() - init_space;
used_space = 0;
printf ("Throughput = %8.0f operations per second.\n", sum_allocs / duration);
#if 0
printf("%2d ", num_threads ) ;
printf("%6.3f", duration ) ;
printf("%6.3f", rate_n/rate_1 ) ;
printf("%8.0f", sum_allocs/duration ) ;
printf(" %6.3f %.3f", (double)used_space/(1024*1024), used_space/reqd_space) ;
printf("\n") ;
#endif
Sleep(5000L) ; // wait 5 sec for old threads to die
prevthreads = num_threads ;
printf ("Done sleeping...\n");
}
delete [] de_area;
}
bool warmup(const char *path, enum FileType file_type){
int fd = 0;
DIR *dir = NULL;
int time_used = 0;
char file_name[256];
struct stat st;
struct timeval begin, end;
struct dirent *dp = NULL;
if(stat(path, &st)<0){
goto ERROR;
}
gettimeofday(&begin, NULL);
if(file_type == REGFILE){
fd = open(path, O_RDONLY);
if(fd<0){
goto ERROR;
}
if(posix_fadvise(fd, 0, st.st_size, POSIX_FADV_WILLNEED) != 0){
goto ERROR;
}
}else if(file_type == DIRECTORY){
if((dir = opendir(path)) == NULL){
goto ERROR;
}
while((dp = readdir(dir)) != NULL){
if(dp->d_name[0] != '.'){
memset(file_name, 0, sizeof(path));
strcat(file_name, path);
strcat(file_name, "/");
strcat(file_name, dp->d_name);
if(dp->d_type == DT_REG){
warmup(file_name, REGFILE);
}else if(dp->d_type == DT_DIR){
warmup(file_name, DIRECTORY);
}else{
fprintf(stdout, "%s:%c type unsupported!\n", dp->d_name, dp->d_type);
}
}
}
}
gettimeofday(&end, NULL);
time_used = getUsedTime(&begin, &end);
if(file_type == REGFILE){
fprintf(stdout, "Warmup File:%s TimeUsed:%d ms\n", path, time_used);
close(fd);
}else if(file_type == DIRECTORY){
fprintf(stdout, "Warmup Dir:%s TimeUsed:%d ms\n", path, time_used);
closedir(dir);
}
return true;
ERROR:
fprintf(stderr, "File:%s %s\n", path, strerror(errno));
if(fd) close(fd);
if(dir) closedir(dir);
return false;
}
void CPUBenchmark::getProcedureOverhead(vector<double> &result){
uint64_t totalTime = 0;
uint64_t start, end;
//0 argument
warmup();
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_0();
end = rdtscEnd();
totalTime += (end - start);
}
result[0] = (double)totalTime / (double)TIMES;
//1 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_1(1);
end = rdtscEnd();
totalTime += (end - start);
}
result[1] = (double)totalTime / (double)TIMES;
//2 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_2(1, 2);
end = rdtscEnd();
totalTime += (end - start);
}
result[2] = (double)totalTime / (double)TIMES;
//3 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_3(1, 2, 3);
end = rdtscEnd();
totalTime += (end - start);
}
result[3] = (double)totalTime / (double)TIMES;
//4 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_4(1, 2, 3, 4);
end = rdtscEnd();
totalTime += (end - start);
}
result[4] = (double)totalTime / (double)TIMES;
//5 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_5(1, 2, 3, 4, 5);
end = rdtscEnd();
totalTime += (end - start);
}
result[5] = (double)totalTime / (double)TIMES;
//6 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_6(1, 2, 3, 4, 5, 6);
end = rdtscEnd();
totalTime += (end - start);
}
result[6] = (double)totalTime / (double)TIMES;
//7 argument
warmup();
totalTime = 0;
for (int i = 0; i < TIMES; i++) {
start = rdtscStart();
fun_7(1, 2, 3, 4, 5, 6, 7);
end = rdtscEnd();
totalTime += (end - start);
}
result[7] = (double)totalTime / (double)TIMES;
}
void runthreads(long sleep_cnt, int min_threads, int max_threads, int chperthread, int num_rounds)
{
thread_data de_area[MAX_THREADS] ;
int nperthread ;
int sum_threads ;
int sum_allocs ;
int sum_frees ;
double duration ;
long ticks_per_sec ;
long start_cnt, end_cnt ;
_int64 ticks ;
double rate_1=0, rate_n ;
double reqd_space ;
ptrdiff_t used_space ;
int prevthreads ;
int i ;
QueryPerformanceFrequency( &ticks_per_sec ) ;
memset(&de_area[0], 0, sizeof(thread_data)) ;
prevthreads = 0 ;
for(num_threads=min_threads; num_threads <= max_threads; num_threads++ )
{
warmup(&blkp[prevthreads*chperthread], (num_threads-prevthreads)*chperthread );
nperthread = chperthread ;
stopflag = FALSE ;
for(i=0; i< num_threads; i++){
de_area[i].threadno = i+1 ;
de_area[i].NumBlocks = num_rounds*nperthread;
de_area[i].array = &blkp[i*nperthread] ;
de_area[i].blksize = &blksize[i*nperthread] ;
de_area[i].asize = nperthread ;
de_area[i].min_size = min_size ;
de_area[i].max_size = max_size ;
de_area[i].seed = lran2(&rgen) ; ;
de_area[i].finished = 0 ;
de_area[i].cAllocs = 0 ;
de_area[i].cFrees = 0 ;
de_area[i].cThreads = 0 ;
de_area[i].finished = FALSE ;
lran2_init(&de_area[i].rgen, de_area[i].seed) ;
_beginthread(exercise_heap, 0, &de_area[i]) ;
}
QueryPerformanceCounter( &start_cnt) ;
//printf ("Sleeping for %ld seconds.\n", sleep_cnt);
Sleep(sleep_cnt * 1000L) ;
stopflag = TRUE ;
for(i=0; i<num_threads; i++){
while( !de_area[i].finished ){
sched_yield();
}
}
QueryPerformanceCounter( &end_cnt) ;
sum_frees = sum_allocs =0 ;
sum_threads = 0 ;
for(i=0;i< num_threads; i++){
sum_allocs += de_area[i].cAllocs ;
sum_frees += de_area[i].cFrees ;
sum_threads += de_area[i].cThreads ;
printf("area %d: %d allocs, %d threads\n",i,de_area[i].cAllocs,de_area[i].cThreads);
de_area[i].cAllocs = de_area[i].cFrees = 0;
}
ticks = end_cnt - start_cnt ;
duration = (double)ticks/ticks_per_sec ;
for( i=0; i<num_threads; i++){
if( !de_area[i].finished )
printf("Thread at %d not finished\n", i) ;
}
rate_n = sum_allocs/duration ;
if( rate_1 == 0){
rate_1 = rate_n ;
}
reqd_space = (0.5*(min_size+max_size)*num_threads*chperthread) ;
used_space = mem_usage();
printf ("Throughput = %8.0f operations per second.\n", sum_allocs / duration);
printf ("Memory used = %ld bytes, required %.0lf, ratio %lf\n",used_space,reqd_space,used_space/reqd_space);
#if 0
printf("%2d ", num_threads ) ;
printf("%6.3f", duration ) ;
printf("%6.3f", rate_n/rate_1 ) ;
printf("%8.0f", sum_allocs/duration ) ;
printf(" %6.3f %.3f", (double)used_space/(1024*1024), used_space/reqd_space) ;
printf("\n") ;
#endif
Sleep(5000L) ; // wait 5 sec for old threads to die
prevthreads = num_threads ;
printf ("Done sleeping...\n");
}
}
