void
initTrace()
{
// Enqueue a couple of buffers in the global queue
queueNewBuffer(NULL, 0);
queueNewBuffer(NULL, 0);
// device activity record is created when CUDA initializes, so we
// want to enable it before cuInit() or any CUDA runtime call
CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE));
// CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY));
cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONTEXT);
// cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER);
// cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME);
cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY);
cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET);
cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL);
CUpti_SubscriberHandle subscriber;
CUPTI_CALL(cuptiSubscribe(&subscriber, (CUpti_CallbackFunc)traceCallback, NULL));
CUPTI_CALL(cuptiEnableDomain(1, subscriber, CUPTI_CB_DOMAIN_RESOURCE));
CUPTI_CALL(cuptiEnableDomain(1, subscriber, CUPTI_CB_DOMAIN_SYNCHRONIZE));
//add by wukai
//CUPIT_CALL(cuptiEnableDomain(1, subscriber, CUPTI_CB_DOMAIN_RUNTIME_API));
CUPTI_CALL(cuptiGetTimestamp(&startTimestamp));
}
__attribute__((constructor)) void Trace_start()
{
cuptierr = cuptiSubscribe(&subscriber, (CUpti_CallbackFunc)getTimestampCallback , &trace);
CHECK_CUPTI_ERROR(cuptierr, "cuptiSubscribe");
cuptierr = cuptiEnableDomain(1, subscriber, CUPTI_CB_DOMAIN_RUNTIME_API);
CHECK_CUPTI_ERROR(cuptierr, "cuptiEnableDomain");
printf("<-----------register Trace_end--------------->\n");
atexit (Trace_end);
}
__attribute__((constructor)) void
initTrace()
{
//get the arguments from the environment variables
int deviceId, sampRate;
CUcontext cuCtx;
deviceId = atoi(getenv("GPU_DEVICE_ID"));
cuInit(0);
cuCtxCreate(&cuCtx,0,deviceId);
CUPTI_CALL(cuptiActivityRegisterCallbacks(bufferRequested, bufferCompleted));
CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_PC_SAMPLING));
//CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_BRANCH));
CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL));
CUPTI_CALL(cuptiSubscribe(&g_subscriber, (CUpti_CallbackFunc)traceCallback, NULL));
CUPTI_CALL(cuptiEnableDomain(1, g_subscriber, CUPTI_CB_DOMAIN_RESOURCE));
CUpti_ActivityPCSamplingConfig config;
sampRate=atoi(getenv("PC_SAMPLING_RATE"));
config.samplingPeriod= sampRate;
CUPTI_CALL(cuptiActivityConfigurePCSampling(cuCtx, &config));
}
int cupti_metrics_enable()
{
CUptiResult res;
if (!initialised || !metricCount) {
return 0;
}
// We cannot initialise anything here - we only hook CUDA context
// creation in order to do the actual work later.
if (subscriber == NULL) {
// Create subscription
CUPTI_CALL(cuptiSubscribe(&subscriber, (CUpti_CallbackFunc)cupti_callback,
NULL));
// Bind callbacks
CUPTI_CALL(cuptiEnableCallback(1, subscriber, CUPTI_CB_DOMAIN_RESOURCE,
CUPTI_CBID_RESOURCE_CONTEXT_CREATED));
CUPTI_CALL(cuptiEnableCallback(1, subscriber, CUPTI_CB_DOMAIN_RESOURCE,
CUPTI_CBID_RESOURCE_CONTEXT_DESTROY_STARTING));
// Hook kernel launch. This feels hacky, but is the approach
// callback_metric.cu takes.
#ifndef LOG_ALL_DRIVER_CALLS
CUPTI_CALL(cuptiEnableCallback(1, subscriber, CUPTI_CB_DOMAIN_DRIVER_API,
CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel));
#else
int i;
for (i = CUPTI_DRIVER_TRACE_CBID_cuInit; i < CUPTI_DRIVER_TRACE_CBID_SIZE; i++) {
CUPTI_CALL(cuptiEnableCallback(1, subscriber, CUPTI_CB_DOMAIN_DRIVER_API, i));
}
#endif
}
}
