AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

pmc: add support of concurrent kernel profiling

Browse Source 
The profiling was only enabled in serial mode, i.e., kernels
are serialized in execution, and counters are reset at each
kernel start and read at kernel completion. This patch adds
the concurrent mode, by issuing the process-level start
packet to reset counters, and then reading twice at kernel
start and end time to obtain the counter value difference.
The new concurrent profiling usage needs the integration
with the corresponding augment at aqlprofile side.

Change-Id: I94b4442eadc8c64b8fba51b1e4916fc8b895ad21
This commit is contained in:
Xianwei Zhang
2020-07-07 15:39:08 -04:00
committed by Evgeny
parent 2a7f77b290
commit 61c9df4631
6 changed files with 154 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/core/rocprofiler.cpp
+20 -1
View File
@@ -150,6 +150,20 @@ void RestoreHsaApi() {
table->amd_ext_->hsa_amd_queue_intercept_register_fn = hsa_amd_queue_intercept_register_fn;
}
void PmcStarter(Context* context) {
hsa_agent_t agent = context->GetAgent();
// Create queue
hsa_queue_t* queue;
hsa_status_t status = rocprofiler::CreateQueuePro(agent, 1,
HSA_QUEUE_TYPE_MULTI, NULL, NULL, UINT32_MAX, UINT32_MAX, &queue);
if (status != HSA_STATUS_SUCCESS) EXC_RAISING(status, "CreateQueuePro");
HsaQueue hsa_queue(NULL, queue);
context->Start(0, &hsa_queue);
context->Read(0, &hsa_queue);
context->GetData(0);
hsa_queue_destroy(queue);
}
void StandaloneIntercept() {
::HsaApiTable* table = kHsaApiTable;
table->core_->hsa_queue_create_fn = rocprofiler::CreateQueuePro;
@@ -216,7 +230,10 @@ uint32_t LoadTool() {
if (settings.code_obj_tracking) intercept_mode |= CODE_OBJ_TRACKING_MODE;
if (settings.memcopy_tracking) intercept_mode |= MEMCOPY_INTERCEPT_MODE;
if (settings.hsa_intercepting) intercept_mode |= HSA_INTERCEPT_MODE;
if (settings.k_concurrent) InterceptQueue::k_concurrent_ = true;
if (settings.k_concurrent) {
Context::k_concurrent_ = settings.k_concurrent;
InterceptQueue::k_concurrent_ = settings.k_concurrent;
}
if (settings.opt_mode) InterceptQueue::opt_mode_ = true;
}
@@ -429,6 +446,8 @@ util::Logger::mutex_t util::Logger::mutex_;
std::atomic<util::Logger*> util::Logger::instance_{};
}
CONTEXT_INSTANTIATE();
///////////////////////////////////////////////////////////////////////////////////////////////////
// Public library methods
//