static void CUPTIAPI
bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, size_t size, size_t validSize)
{
CUptiResult status;
CUpti_Activity *record = NULL;
do {
status = cuptiActivityGetNextRecord(buffer, validSize, &record);
if(status == CUPTI_SUCCESS) {
printActivity(record);
}
else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) {
break;
}
else {
CUPTI_CALL(status);
}
} while (1);
size_t dropped;
CUPTI_CALL(cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped));
if (dropped != 0) {
printf("Dropped %u activity records\n", (unsigned int)dropped);
}
printf("\n\n\n\n\n\n");
printf("************* STALL SUMMARY ********************\n");
int i;
for(i=0;i<12;++i)
if(stall_name[i] != NULL)
printf("%s = %d \n",stall_name[i],val[i]);
printf("*************************************************\n\n");
}
static void
handleSync(CUpti_CallbackId cbid, const CUpti_SynchronizeData *syncData)
{
// check the top buffer of the global queue and dequeue if full. If
// we dump a buffer add it back to the queue
uint8_t *buffer = dumpIfFull(NULL, 0);
if (buffer != NULL) {
CUPTI_CALL(cuptiActivityEnqueueBuffer(NULL, 0, buffer, BUF_SIZE));
}
// dump context buffer on context sync
if (cbid == CUPTI_CBID_SYNCHRONIZE_CONTEXT_SYNCHRONIZED) {
buffer = dumpIfFull(syncData->context, 0);
if (buffer != NULL) {
CUPTI_CALL(cuptiActivityEnqueueBuffer(syncData->context, 0, buffer, BUF_SIZE));
}
}
// dump stream buffer on stream sync
else if (cbid == CUPTI_CBID_SYNCHRONIZE_STREAM_SYNCHRONIZED) {
uint32_t streamId;
CUPTI_CALL(cuptiGetStreamId(syncData->context, syncData->stream, &streamId));
buffer = dumpIfFull(syncData->context, streamId);
if (buffer != NULL) {
CUPTI_CALL(cuptiActivityEnqueueBuffer(syncData->context, streamId, buffer, BUF_SIZE));
}
}
}
static void
handleResource(CUpti_CallbackId cbid, const CUpti_ResourceData *resourceData)
{
// enqueue buffers on a context's queue when the context is created
if (cbid == CUPTI_CBID_RESOURCE_CONTEXT_CREATED) {
queueNewBuffer(resourceData->context, 0);
queueNewBuffer(resourceData->context, 0);
}
// dump all buffers on a context destroy
else if (cbid == CUPTI_CBID_RESOURCE_CONTEXT_DESTROY_STARTING) {
while (dump(resourceData->context, 0) != NULL) ;
}
// enqueue buffers on a stream's queue when a non-default stream is created
if (cbid == CUPTI_CBID_RESOURCE_STREAM_CREATED) {
uint32_t streamId;
CUPTI_CALL(cuptiGetStreamId(resourceData->context, resourceData->resourceHandle.stream, &streamId));
queueNewBuffer(resourceData->context, streamId);
queueNewBuffer(resourceData->context, streamId);
}
// dump all buffers on a stream destroy
else if (cbid == CUPTI_CBID_RESOURCE_STREAM_DESTROY_STARTING) {
uint32_t streamId;
CUPTI_CALL(cuptiGetStreamId(resourceData->context, resourceData->resourceHandle.stream, &streamId));
while (dump(resourceData->context, streamId) != NULL) ;
}
}
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));
}
/**
* Dump the contents of the top buffer in the queue specified by
* 'context' and 'streamId', and return the top buffer. If the queue
* is empty return NULL.
*/
static uint8_t *
dump(CUcontext context, uint32_t streamId)
{
uint8_t *buffer = NULL;
size_t validBufferSizeBytes;
CUptiResult status;
status = cuptiActivityDequeueBuffer(context, streamId, &buffer, &validBufferSizeBytes);
if (status == CUPTI_ERROR_QUEUE_EMPTY) {
return NULL;
}
CUPTI_CALL(status);
if (context == NULL) {
printf("[CUPTI] Starting dump for global\n");
} else if (streamId == 0) {
printf("[CUPTI] Starting dump for context %p\n", context);
} else {
printf("[CUPTI] Starting dump for context %p, stream %u\n", context, streamId);
}
CUpti_Activity *record = NULL;
do {
status = cuptiActivityGetNextRecord(buffer, validBufferSizeBytes, &record);
if(status == CUPTI_SUCCESS) {
printActivity(record);
}
else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) {
printf ("[CUPTI] CUPTI_ERROR_MAX_LIMIT_REACHED\n");
break;
}
else {
CUPTI_CALL(status);
}
} while (1);
// report any records dropped from the queue
size_t dropped;
CUPTI_CALL(cuptiActivityGetNumDroppedRecords(context, streamId, &dropped));
if (dropped != 0) {
printf("Dropped %u activity records\n", (unsigned int)dropped);
}
if (context == NULL) {
printf("[CUPTI] Finished dump for global\n");
} else if (streamId == 0) {
printf("[CUPTI] Finished dump for context %p \n", context);
} else {
printf("[CUPTI] Finished dump for context %p, stream %u\n", context, streamId);
}
return buffer;
}
/**
* Allocate a new BUF_SIZE buffer and add it to the queue specified by
* 'context' and 'streamId'.
*/
static void
queueNewBuffer(CUcontext context, uint32_t streamId)
{
size_t size = BUF_SIZE;
uint8_t *buffer = (uint8_t *)malloc(size+ALIGN_SIZE);
CUPTI_CALL(cuptiActivityEnqueueBuffer(context, streamId, ALIGN_BUFFER(buffer, ALIGN_SIZE), size));
}
/**
* If the top buffer in the queue specified by 'context' and
* 'streamId' is full, then dump its contents and return the
* buffer. If the top buffer is not full, return NULL.
*/
static uint8_t *
dumpIfFull(CUcontext context, uint32_t streamId)
{
size_t validBufferSizeBytes;
CUptiResult status;
status = cuptiActivityQueryBuffer(context, streamId, &validBufferSizeBytes);
if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) {
return dump(context, streamId);
} else if ((status != CUPTI_SUCCESS) && (status != CUPTI_ERROR_QUEUE_EMPTY)) {
CUPTI_CALL(status);
}
return NULL;
}
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
}
}
__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));
}
__attribute__((destructor)) void
finiTrace()
{
// printf("FLushing CUPTI \n");
CUPTI_CALL(cuptiActivityFlushAll(0));
}
