int main()
{
int i, j, rank, nranks, msgsize, dest;
int xdim, ydim;
long bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total, bw;
int count[2], src_stride, trg_stride, stride_level;
OSP_handle_t osp_handle;
OSP_Initialize(OSP_THREAD_SINGLE);
rank = OSP_Process_id(OSP_GROUP_WORLD);
nranks = OSP_Process_total(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
bufsize = MAX_XDIM * MAX_YDIM * sizeof(double);
buffer = (double **) malloc(sizeof(double *) * nranks);
OSP_Alloc_segment((void **) &(buffer[rank]), bufsize);
OSP_Exchange_segments(OSP_GROUP_WORLD, (void **) buffer);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Allocate_handle(&osp_handle);
OSP_Barrier_group(OSP_GROUP_WORLD);
if (rank == 0)
{
printf("OSP_PutS Bandwidth in MBPS \n");
printf("%30s %22s \n", "Dimensions(array of doubles)", "Latency");
fflush(stdout);
dest = 1;
src_stride = MAX_YDIM * sizeof(double);
trg_stride = MAX_YDIM * sizeof(double);
stride_level = 1;
for (xdim = 1; xdim <= MAX_XDIM; xdim *= 2)
{
count[1] = xdim;
for (ydim = 1; ydim <= MAX_YDIM; ydim *= 2)
{
count[0] = ydim * sizeof(double);
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP)
t_start = OSP_Time_seconds();
OSP_NbPutS(1,
stride_level,
count,
(void *) buffer[dest],
&src_stride,
(void *) buffer[rank],
&trg_stride,
osp_handle);
}
OSP_Wait_handle(osp_handle);
t_stop = OSP_Time_seconds();
OSP_Flush(1);
char temp[10];
sprintf(temp, "%dX%d", xdim, ydim);
t_total = t_stop - t_start;
d_total = (xdim*ydim*sizeof(double)*ITERATIONS)/(1024*1024);
bw = d_total/t_total;
printf("%30s %20.2f \n", temp, bw);
fflush(stdout);
}
}
}
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Release_segments(OSP_GROUP_WORLD, (void *) buffer[rank]);
OSP_Free_segment((void *) buffer[rank]);
OSP_Finalize();
return 0;
}
int main() {
int i, j, rank, nranks, msgsize;
int xdim, ydim;
long bufsize;
double **buffer;
double t_start, t_stop, t_latency;
int count[2], src_stride, trg_stride, stride_level, peer;
double expected, actual;
OSP_Initialize(OSP_THREAD_SINGLE);
rank = OSP_Process_id(OSP_GROUP_WORLD);
nranks = OSP_Process_total(OSP_GROUP_WORLD);
buffer = (double **) malloc (sizeof(double *) * nranks);
OSP_Barrier_group(OSP_GROUP_WORLD);
bufsize = MAX_XDIM * MAX_YDIM * sizeof(double);
OSP_Alloc_segment((void **) &(buffer[rank]), bufsize);
OSP_Exchange_segments(OSP_GROUP_WORLD, (void **) buffer);
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
if(rank == 0) {
printf("OSP_PutS Latency - local and remote completions - in usec \n");
printf("%30s %22s \n", "Dimensions(array of doubles)", "Latency-LocalCompeltion", "Latency-RemoteCompletion");
fflush(stdout);
}
src_stride = MAX_YDIM*sizeof(double);
trg_stride = MAX_YDIM*sizeof(double);
stride_level = 1;
for(xdim=1; xdim<=MAX_XDIM; xdim*=2) {
count[1] = xdim;
for(ydim=1; ydim<=MAX_YDIM; ydim*=2) {
count[0] = ydim*sizeof(double);
if(rank == 0)
{
peer = 1;
for(i=0; i<ITERATIONS+SKIP; i++) {
if(i == SKIP)
t_start = OSP_Time_seconds();
OSP_PutS(peer, stride_level, count, (void *) buffer[rank], &src_stride, (void *) buffer[peer], &trg_stride);
}
t_stop = OSP_Time_seconds();
OSP_Flush(peer);
char temp[10];
sprintf(temp,"%dX%d", xdim, ydim);
printf("%30s %20.2f", temp, ((t_stop-t_start)*1000000)/ITERATIONS);
fflush(stdout);
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
for(i=0; i<ITERATIONS+SKIP; i++) {
if(i == SKIP)
t_start = OSP_Time_seconds();
OSP_PutS(peer, stride_level, count, (void *) buffer[rank], &src_stride, (void *) buffer[peer], &trg_stride);
OSP_Flush(peer);
}
t_stop = OSP_Time_seconds();
printf("%20.2f \n", ((t_stop-t_start)*1000000)/ITERATIONS);
fflush(stdout);
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
}
else
{
peer = 0;
expected = (1.0 + (double) peer);
OSP_Barrier_group(OSP_GROUP_WORLD);
for(i=0; i<xdim; i++)
{
for(j=0; j<ydim; j++)
{
actual = *(buffer[rank] + i*MAX_YDIM + j);
if(actual != expected)
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i, j, expected, actual);
fflush(stdout);
return -1;
}
}
}
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
for(i=0; i<xdim; i++)
{
for(j=0; j<ydim; j++)
{
actual = *(buffer[rank] + i*MAX_YDIM + j);
if(actual != expected)
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i, j, expected, actual);
fflush(stdout);
return -1;
}
}
}
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Barrier_group(OSP_GROUP_WORLD);
}
}
}
OSP_Release_segments(OSP_GROUP_WORLD, (void *) buffer[rank]);
OSP_Free_segment((void *) buffer[rank]);
OSP_Finalize();
return 0;
}
/*********************************************************************
* @fn AlgorithmTask
* This task is responsible for running the sensor algorithms
* on the incoming sensor data (could be raw or filtered) and
* processing output results
*
* @param none
*
* @return none
*
**********************************************************************/
ASF_TASK void AlgorithmTask (ASF_TASK_ARG)
{
MessageBuffer *rcvMsg = NULLP;
OSP_STATUS_t OSP_Status;
int alg_count;
OSP_GetLibraryVersion(&version);
D1_printf("OSP Version: %s\r\n", version->VersionString);
/* Initialize the mutex */
mutex_id = osMutexCreate(osMutex(mutexCritSection));
OSP_Status = OSP_Initialize(&gSystemDesc);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "OSP_Initialize Failed");
OSP_SetCalibrationConfig( 0x1); // disable rotational cal.
D0_printf("--Alg Task %i\r\n", __LINE__);
// Register the input sensors
OSP_Status = OSP_RegisterInputSensor(&_AccSensDesc, &_AccHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "OSP_RegisterInputSensor (accel) Failed");
OSP_Status = OSP_RegisterInputSensor(&_MagSensDesc, &_MagHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "OSP_RegisterInputSensor (mag) Failed");
OSP_Status = OSP_RegisterInputSensor(&_GyroSensDesc, &_GyroHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "OSP_RegisterInputSensor (gyro) Failed");
#if 0
SENSOR_SUBSCRIBE(SENSOR_STEP_COUNTER);
SENSOR_SUBSCRIBE(SENSOR_STEP_DETECTOR);
SENSOR_SUBSCRIBE(SENSOR_SIGNIFICANT_MOTION);
SENSOR_SUBSCRIBE(SENSOR_GYROSCOPE_UNCALIBRATED);
SENSOR_SUBSCRIBE(SENSOR_MAGNETIC_FIELD_UNCALIBRATED);
SENSOR_SUBSCRIBE(SENSOR_GYROSCOPE);
SENSOR_SUBSCRIBE(SENSOR_ACCELEROMETER);
SENSOR_SUBSCRIBE(SENSOR_MAGNETIC_FIELD);
SENSOR_SUBSCRIBE(SENSOR_ORIENTATION);
SENSOR_SUBSCRIBE(SENSOR_GRAVITY);
SENSOR_SUBSCRIBE(SENSOR_LINEAR_ACCELERATION);
SENSOR_SUBSCRIBE(SENSOR_ROTATION_VECTOR);
SENSOR_SUBSCRIBE(SENSOR_GAME_ROTATION_VECTOR);
SENSOR_SUBSCRIBE(SENSOR_GEOMAGNETIC_ROTATION_VECTOR);
// Subscribing private sensor results
PRIVATE_SENSOR_SUBSCRIBE(AP_PSENSOR_ACCELEROMETER_UNCALIBRATED);
#endif
D0_printf("%s: --Alg Task init done\r\n", __func__);
while (1) {
ASFReceiveMessage(ALGORITHM_TASK_ID, &rcvMsg);
if (!(mycount % 64)) {
LED_Toggle(LED_GREEN);
}
switch (rcvMsg->msgId) {
case MSG_MAG_DATA:
// SendBgTrigger();
case MSG_ACC_DATA:
case MSG_GYRO_DATA:
mycount++;
HandleSensorData(rcvMsg);
//keep doing foreground computation until its finished
/* Bump clock speed while processing? HY-DBG */
alg_count = 0;
do {
OSP_Status = OSP_DoForegroundProcessing();
ASF_assert(OSP_Status != OSP_STATUS_UNSPECIFIED_ERROR);
alg_count++;
if (alg_count > 5) {
D0_printf("%s:%i Taking too long\r\n", __func__, __LINE__);
break;
}
} while(OSP_Status != OSP_STATUS_IDLE);
/* DBG:
* Run background here as the backgound taks doesn't seem to run enough */
while(OSP_DoBackgroundProcessing() != OSP_STATUS_IDLE);
break;
case MSG_PRESS_DATA:
PressureDataResultCallback(&rcvMsg->msg.msgPressData);
break;
default:
/* Unhandled messages */
D1_printf("Alg-FG:!!!UNHANDLED MESSAGE:%d!!!\r\n", rcvMsg->msgId);
break;
}
ASFDeleteMessage( ALGORITHM_TASK_ID, &rcvMsg );
#ifdef DEBUG_TEST_SENSOR_SUBSCRIPTION
// Testing subscribe and unsubscribe sensors
DebugTestSensorSubscription();
#endif
}
}
/****************************************************************************************************
* @fn AlgorithmTask
* This task is responsible for running the sensor algorithms on the incoming sensor
* data (could be raw or filtered) and processing output results
*
* @param none
*
* @return none
*
***************************************************************************************************/
ASF_TASK void AlgorithmTask ( ASF_TASK_ARG )
{
MessageBuffer *rcvMsg = NULLP;
osp_status_t OSP_Status;
OSP_GetVersion(&version);
D1_printf("OSP Version: %s\r\n", version->VersionString);
OSP_Status = OSP_Initialize(&gSystemDesc);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_Initialize Failed");
// Register the input sensors
OSP_Status = OSP_RegisterInputSensor(&_AccSensDesc, &_AccHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_RegisterSensor (accel) Failed");
OSP_Status = OSP_RegisterInputSensor(&_MagSensDesc, &_MagHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_RegisterSensor (mag) Failed");
OSP_Status = OSP_RegisterInputSensor(&_GyroSensDesc, &_GyroHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_RegisterSensor (gyro) Failed");
// Register output sensors/results
OSP_Status = OSP_SubscribeOutputSensor(&stepCounterRequest, &_stepCounterHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_SubscribeResult (SENSOR_STEP_COUNTER) Failed");
OSP_Status = OSP_SubscribeOutputSensor(&sigMotionRequest, &_sigMotionHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_SubscribeResult (SENSOR_CONTEXT_DEVICE_MOTION) Failed");
OSP_Status = OSP_SubscribeOutputSensor(&UnCalAccelRequest, &_unCalAccelHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_SubscribeResult (SENSOR_ACCELEROMETER) Failed");
OSP_Status = OSP_SubscribeOutputSensor(&UnCalMagRequest, &_unCalMagHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_SubscribeResult (SENSOR_MAGNETIC_FIELD) Failed");
OSP_Status = OSP_SubscribeOutputSensor(&UnCalGyroRequest, &_unCalGyroHandle);
ASF_assert_msg(OSP_STATUS_OK == OSP_Status, "SensorManager: OSP_SubscribeResult (SENSOR_GYROSCOPE) Failed");
while (1)
{
ASFReceiveMessage( ALGORITHM_TASK_ID, &rcvMsg );
switch (rcvMsg->msgId)
{
case MSG_MAG_DATA:
SendBgTrigger();
case MSG_ACC_DATA:
case MSG_GYRO_DATA:
HandleSensorData(rcvMsg);
do
{
OSP_Status = OSP_DoForegroundProcessing();
ASF_assert(OSP_Status != OSP_STATUS_ERROR);
} while(OSP_Status != OSP_STATUS_IDLE)
; //keep doing foreground computation until its finished
break;
default:
/* Unhandled messages */
D1_printf("Alg-FG:!!!UNHANDLED MESSAGE:%d!!!\r\n", rcvMsg->msgId);
break;
}
}
}
int main()
{
size_t i, rank, nranks, msgsize, peer;
long bufsize;
double **buffer;
double scaling;
double t_start, t_stop, t_latency;
OSP_Initialize(OSP_THREAD_SINGLE);
rank = OSP_Process_id(OSP_GROUP_WORLD);
nranks = OSP_Process_total(OSP_GROUP_WORLD);
bufsize = MAX_MSG_SIZE * (ITERATIONS + SKIP);
buffer = (double **) malloc(sizeof(double *) * nranks);
OSP_Alloc_segment((void **) &(buffer[rank]), bufsize);
OSP_Exchange_segments(OSP_GROUP_WORLD, (void **) buffer);
if (rank == 0)
{
printf("OSP_PutAcc Latency in usec \n");
printf("%20s %22s %22s\n",
"Message Size",
"Local Completion",
"Remote Completion");
fflush(stdout);
}
for (i = 0; i < (((ITERATIONS + SKIP) * MAX_MSG_SIZE) / sizeof(double)); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
scaling = 2.0;
OSP_Barrier_group(OSP_GROUP_WORLD);
for (msgsize = sizeof(double); msgsize < MAX_MSG_SIZE; msgsize *= 2)
{
if (rank == 0)
{
peer = 1;
/** Local Completion **/
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP) t_start = OSP_Time_seconds();
OSP_PutAcc(peer,
(void *) ((size_t) buffer[rank] + (size_t)(i
* msgsize)),
(void *) ((size_t) buffer[peer] + (size_t)(i
* msgsize)),
msgsize,
OSP_DOUBLE,
(void *) &scaling);
}
t_stop = OSP_Time_seconds();
OSP_Flush(1);
printf("%20d %20.2f ", msgsize, ((t_stop - t_start) * 1000000)
/ ITERATIONS);
fflush(stdout);
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP) t_start = OSP_Time_seconds();
OSP_PutAcc(1, (void *) ((size_t) buffer[0] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[1] + (size_t)(i
* msgsize)), msgsize, OSP_DOUBLE, (void *) &scaling);
OSP_Flush(1);
}
t_stop = OSP_Time_seconds();
printf("%20.2f \n", ((t_stop - t_start) * 1000000) / ITERATIONS);
fflush(stdout);
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
}
else
{
peer = 0;
OSP_Barrier_group(OSP_GROUP_WORLD);
/** Data Validation **/
for (i = 0; i < (((ITERATIONS + SKIP) * msgsize) / sizeof(double)); i++)
{
if (*(buffer[rank] + i) != ((1.0 + rank) + scaling * (1.0
+ peer)))
{
printf("Data validation failed At displacement : %d Expected : %f Actual : %f \n",
i,
((1.0 + rank) + scaling * (1.0 + peer)),
*(buffer[rank] + i));
fflush(stdout);
return -1;
}
}
for (i = 0; i < (((ITERATIONS + SKIP) * MAX_MSG_SIZE)
/ sizeof(double)); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
/** Data Validation **/
for (i = 0; i < (((ITERATIONS + SKIP) * msgsize) / sizeof(double)); i++)
{
if (*(buffer[rank] + i) != ((1.0 + rank) + scaling * (1.0
+ peer)))
{
printf("Data validation failed At displacement : %d Expected : %f Actual : %f \n",
i,
((1.0 + rank) + scaling * (1.0 + peer)),
*(buffer[rank] + i));
fflush(stdout);
return -1;
}
}
for (i = 0; i < (((ITERATIONS + SKIP) * MAX_MSG_SIZE)
/ sizeof(double)); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Barrier_group(OSP_GROUP_WORLD);
}
}
OSP_Release_segments(OSP_GROUP_WORLD, buffer[rank]);
OSP_Finalize();
return 0;
}
int main()
{
int i, j, rank, nranks, msgsize, dest;
int dim;
long bufsize;
double **buffer;
int iterations;
double t_start, t_stop, t_total, d_total, bw;
int count[2], src_stride, trg_stride, stride_level;
OSP_handle_t osp_handle;
OSP_Initialize(OSP_THREAD_SINGLE);
rank = OSP_Process_id(OSP_GROUP_WORLD);
nranks = OSP_Process_total(OSP_GROUP_WORLD);
OSP_Barrier_group(OSP_GROUP_WORLD);
bufsize = MAX_DIM * MAX_DIM * sizeof(double);
buffer = (double **) malloc(sizeof(double *) * nranks);
OSP_Alloc_segment((void **) &(buffer[rank]), bufsize);
OSP_Exchange_segments(OSP_GROUP_WORLD, (void **) buffer);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Allocate_handle(&osp_handle);
OSP_Barrier_group(OSP_GROUP_WORLD);
if (rank == 0)
{
printf("OSP_GetS Bandwidth in MBPS \n");
printf("%30s %30s %22s \n", "MsgSize", "Dimensions(array of doubles)", "Latency");
fflush(stdout);
dest = 1;
src_stride = MAX_DIM * sizeof(double);
trg_stride = MAX_DIM * sizeof(double);
stride_level = 1;
for (dim = 1; dim <= MAX_DIM; dim *= 2)
{
count[0] = dim*sizeof(double);
count[1] = dim;
iterations = (MAX_DIM * MAX_DIM)/(dim*dim);
t_start = OSP_Time_seconds();
for (i = 0; i < iterations; i++)
{
OSP_NbGetS(1,
stride_level,
count,
(void *) buffer[dest],
&src_stride,
(void *) buffer[rank],
&trg_stride,
osp_handle);
}
OSP_Wait_handle(osp_handle);
t_stop = OSP_Time_seconds();
char temp[10];
sprintf(temp, "%dX%d", dim, dim);
t_total = t_stop - t_start;
d_total = (dim*dim*sizeof(double)*iterations)/(1024*1024);
bw = d_total/t_total;
printf("%30d %30s %20.2f \n", dim*dim*sizeof(double), temp, bw);
fflush(stdout);
for (i = 0; i < dim; i++)
{
for (j = 0; j < dim; j++)
{
if (*(buffer[rank] + i * MAX_DIM + j) != (1.0 + dest))
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i,
j,
(1.0 + dest),
*(buffer[rank] + i * MAX_DIM + j));
fflush(stdout);
return -1;
}
}
}
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
}
}
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Release_segments(OSP_GROUP_WORLD, (void *) buffer[rank]);
OSP_Free_segment((void *) buffer[rank]);
OSP_Finalize();
return 0;
}
int main()
{
size_t i, rank, nranks, msgsize, dest;
long bufsize;
double **buffer;
double t_start, t_stop, t_latency;
OSP_Initialize(OSP_THREAD_SINGLE);
rank = OSP_Process_id(OSP_GROUP_WORLD);
nranks = OSP_Process_total(OSP_GROUP_WORLD);
bufsize = MAX_MSG_SIZE * (ITERATIONS + SKIP);
buffer = (double **) malloc(sizeof(double *) * nranks);
OSP_Alloc_segment((void **) &(buffer[rank]), bufsize);
OSP_Exchange_segments(OSP_GROUP_WORLD, (void **) buffer);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
OSP_Barrier_group(OSP_GROUP_WORLD);
if (rank == 0)
{
printf("OSP_Get Latency in usec \n");
printf("%20s %22s \n", "Message Size", "Latency");
fflush(stdout);
dest = 1;
for (msgsize = sizeof(double); msgsize <= MAX_MSG_SIZE; msgsize *= 2)
{
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP) t_start = OSP_Time_seconds();
OSP_Get(1, (void *) ((size_t) buffer[dest] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[rank]
+ (size_t)(i * msgsize)), msgsize);
}
t_stop = OSP_Time_seconds();
printf("%20d %20.2f \n", msgsize, ((t_stop - t_start) * 1000000)
/ ITERATIONS);
fflush(stdout);
for (i = 0; i < (((ITERATIONS + SKIP) * msgsize) / sizeof(double)); i++)
{
if (*(buffer[rank] + i) != (1.0 + dest))
{
printf("Data validation failed At displacement : %d Expected : %f Actual : %f \n",
i,
(1.0 + dest),
*(buffer[rank] + i));
fflush(stdout);
return -1;
}
}
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
}
}
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Release_segments(OSP_GROUP_WORLD, buffer[rank]);
OSP_Finalize();
return 0;
}
int main() {
int i, j, k, rank, nranks, msgsize;
int dim;
long bufsize;
double **buffer;
unsigned long long t_start, t_stop, t_latency, t_overlap;
unsigned long long wait_start, wait_stop;
int count[2], src_stride, trg_stride, stride_level, peer;
double A[1024][1024], B[1024][1024], C[1024][1024];
int m1,m2,m3;
double expected, actual;
OSP_handle_t osp_handle;
OSP_Initialize(OSP_THREAD_SINGLE);
rank = OSP_Process_id(OSP_GROUP_WORLD);
nranks = OSP_Process_total(OSP_GROUP_WORLD);
buffer = (double **) malloc (sizeof(double *) * nranks);
OSP_Allocate_handle(&osp_handle);
OSP_Barrier_group(OSP_GROUP_WORLD);
bufsize = MAX_DIM * MAX_DIM * sizeof(double);
OSP_Alloc_segment((void **) &(buffer[rank]), bufsize);
OSP_Exchange_segments(OSP_GROUP_WORLD, (void **) buffer);
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
if(rank == 0) {
printf("OSP_PutS Overlap - NbPutS + DGEMM + Wait. Time in cycles\n");
printf("%30s %30s %22s %22s\n", "Msg Size", "Dimensions(array of doubles)", "Base Latency", "Overlaped Latency");
fflush(stdout);
src_stride = MAX_DIM*sizeof(double);
trg_stride = MAX_DIM*sizeof(double);
stride_level = 1;
for(dim=1; dim<=MAX_DIM; dim*=2) {
count[0] = dim*sizeof(double);
count[1] = 512;
peer = 1;
for(i=0; i<ITERATIONS+SKIP; i++) {
if(i == SKIP)
t_start = OSP_Time_cycles();
for(k=0; k<WINDOW; k++)
{
OSP_NbPutS(peer, stride_level, count, (void *) buffer[rank],
&src_stride, (void *) buffer[peer], &trg_stride, osp_handle);
}
OSP_Wait_handle(osp_handle);
}
t_stop = OSP_Time_cycles();
OSP_Flush(peer);
t_latency = (t_stop-t_start)/ITERATIONS;
char temp[10];
sprintf(temp,"%dX%d", count[1], dim);
printf("%30d %30s %20lld", count[1]*count[0], temp, t_latency);
fflush(stdout);
t_start = OSP_Time_cycles();
for(i=0; i<ITERATIONS; i++) {
for(k=0; k<WINDOW; k++)
{
OSP_NbPutS(peer, stride_level, count, (void *) buffer[rank],
&src_stride, (void *) buffer[peer], &trg_stride, osp_handle);
}
wait_start = OSP_Time_cycles();
for(m1=0; m1<1024; m1++)
{
for(m2=0; m2<1024; m2++)
{
for(m3=0; m3<1024; m3++)
{
C[m1][m2] += A[m1][m3] * B[m3][m2];
wait_stop = OSP_Time_cycles();
if((wait_stop - wait_start) > t_latency)
break;
}
if((wait_stop - wait_start) > t_latency)
break;
}
if((wait_stop - wait_start) > t_latency)
break;
}
OSP_Wait_handle(osp_handle);
}
t_stop = OSP_Time_cycles();
OSP_Flush(peer);
t_overlap = (t_stop - t_start)/ITERATIONS;
printf("%20lld \n", t_overlap);
}
}
OSP_Barrier_group(OSP_GROUP_WORLD);
OSP_Release_handle(osp_handle);
OSP_Release_segments(OSP_GROUP_WORLD, (void *) buffer[rank]);
OSP_Free_segment((void *) buffer[rank]);
OSP_Finalize();
return 0;
}