IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Fix sporadic fails in canRun3SyncRequestsConsistentlyFromThreads test (#8377)

Browse Source
This commit is contained in:
Mikhail Letavin
2021-11-03 14:59:38 +03:00
committed by GitHub
parent 5d9a76d2f8
commit 28706c34a1
4 changed files with 54 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
inference-engine/src/cldnn_engine/cldnn_async_infer_request.cpp
View File

@@ -17,29 +17,34 @@ CLDNNPlugin::CLDNNAsyncInferRequest::CLDNNAsyncInferRequest(const CLDNNInferRequ
_pipeline.push_back({taskExecutor,
[this] {
OV_ITT_SCOPED_TASK(itt::domains::CLDNNPlugin, "CLDNNAsyncInferRequest::PreprocessingAndStartPipeline");
_inferRequest->setup_stream_graph();
_inferRequest->preprocess();
_inferRequest->enqueue();
} });
}
_pipeline.push_back({_waitExecutor,
[this] {
OV_ITT_SCOPED_TASK(itt::domains::CLDNNPlugin, "CLDNNAsyncInferRequest::WaitPipeline");
_inferRequest->wait();
}});
} });
} else {
_pipeline.push_back({ _waitExecutor,
[this] {
OV_ITT_SCOPED_TASK(itt::domains::CLDNNPlugin, "CLDNNAsyncInferRequest::WaitPipeline");
_inferRequest->wait_notify();
} });
}
}
void CLDNNPlugin::CLDNNAsyncInferRequest::Infer_ThreadUnsafe() {
if (_inferRequest->use_external_queue()) {
_inferRequest->preprocess();
_inferRequest->enqueue();
_inferRequest->setup_stream_graph();
_inferRequest->preprocess_notify();
_inferRequest->enqueue_notify();
}
Parent::Infer_ThreadUnsafe();
}
void CLDNNPlugin::CLDNNAsyncInferRequest::StartAsync_ThreadUnsafe() {
if (_inferRequest->use_external_queue()) {
_inferRequest->preprocess();
_inferRequest->enqueue();
_inferRequest->setup_stream_graph();
_inferRequest->preprocess_notify();
_inferRequest->enqueue_notify();
}
Parent::StartAsync_ThreadUnsafe();
}

1
inference-engine/src/cldnn_engine/cldnn_graph.h
View File

@@ -70,6 +70,7 @@ public:
}
m_cv.notify_one();
}
std::mutex& get_mutex() { return m_infer_mutex; }
bool use_external_queue() const;

51
inference-engine/src/cldnn_engine/cldnn_infer_request.cpp
View File

@@ -471,28 +471,30 @@ CLDNNInferRequest::CLDNNInferRequest(const std::vector<std::shared_ptr<const ov:
// ----------------------------------------------------------------------------------------- //
// ---------------------------- internal pipeline stages ----------------------------------- //
// ----------------------------------------------------------------------------------------- //
void CLDNNInferRequest::preprocess() {
int streamID = 0;
auto& streamGraphs = static_cast<CLDNNExecNetwork*>(_exeNetwork.get())->m_graphs;
if (nullptr != streamExecutor) {
streamID = streamExecutor->GetStreamId();
int numGraphs = streamGraphs.size();
streamID = streamID % numGraphs;
}
m_graph = streamGraphs[streamID];
void CLDNNInferRequest::preprocess_notify() {
m_graph->wait(CLDNNGraph::Stage::PREPROC);
if (m_graph->GetMaxDynamicBatchSize() > 1) {
preprocess_dynamic();
return;
} else {
execDataPreprocessing(_inputs, true); // "true" stands for serial preprocessing in case of OpenMP
}
execDataPreprocessing(_inputs, true); // "true" stands for serial preprocessing in case of OpenMP
m_graph->notify(CLDNNGraph::Stage::PREPROC);
}
void CLDNNInferRequest::enqueue() {
void CLDNNInferRequest::preprocess() {
if (m_graph->GetMaxDynamicBatchSize() > 1) {
preprocess_dynamic();
} else {
execDataPreprocessing(_inputs, true); // "true" stands for serial preprocessing in case of OpenMP
}
}
void CLDNNInferRequest::enqueue_notify() {
m_graph->wait(CLDNNGraph::Stage::EXECUTE);
enqueue();
}
void CLDNNInferRequest::enqueue() {
if (m_graph->GetMaxDynamicBatchSize() > 1) {
enqueue_dynamic();
return;
@@ -541,6 +543,11 @@ void CLDNNInferRequest::enqueue() {
internal_outputs = m_graph->GetNetwork()->execute(dependencies);
}
void CLDNNInferRequest::wait_notify() {
wait();
m_graph->notify(CLDNNGraph::Stage::EXECUTE);
}
void CLDNNInferRequest::wait() {
if (m_graph->GetMaxDynamicBatchSize() > 1) {
wait_dynamic();
@@ -568,13 +575,11 @@ void CLDNNInferRequest::wait() {
if (m_useProfiling) {
m_graph->UpdatePerfStatistics();
}
m_graph->notify(CLDNNGraph::Stage::EXECUTE);
}
void CLDNNInferRequest::preprocess_dynamic() {
// execute input pre-processing.
execDataPreprocessing(_inputs, true); // "true" stands for serial preprocessing in case of OpenMP
m_graph->notify(CLDNNGraph::Stage::PREPROC);
}
void CLDNNInferRequest::enqueue_dynamic() {
@@ -619,12 +624,21 @@ void CLDNNInferRequest::wait_dynamic() {
}
}
}
m_graph->notify(CLDNNGraph::Stage::EXECUTE);
}
// ----------------------------------------------------------------------------------------- //
// ---------------------------- internal utils --------- ----------------------------------- //
// ----------------------------------------------------------------------------------------- //
void CLDNNInferRequest::setup_stream_graph() {
int streamID = 0;
auto& streamGraphs = static_cast<CLDNNExecNetwork*>(_exeNetwork.get())->m_graphs;
if (nullptr != streamExecutor) {
streamID = streamExecutor->GetStreamId();
int numGraphs = streamGraphs.size();
streamID = streamID % numGraphs;
}
m_graph = streamGraphs[streamID];
}
Blob::Ptr CLDNNInferRequest::create_host_blob(const TensorDesc& desc, uint8_t* mem_ptr) {
OV_ITT_SCOPED_TASK(itt::domains::CLDNNPlugin, "CLDNNInferRequest::create_host_blob");
@@ -852,7 +866,8 @@ void CLDNNInferRequest::allocate_outputs_dynamic() {
void CLDNNInferRequest::InferImpl() {
OV_ITT_SCOPED_TASK(itt::domains::CLDNNPlugin, "CLDNNInferRequest::InferImpl");
setup_stream_graph();
std::lock_guard<std::mutex> lk(m_graph->get_mutex());
preprocess();
enqueue();
wait();

5
inference-engine/src/cldnn_engine/cldnn_infer_request.h
View File

@@ -49,6 +49,11 @@ public:
void EnableProfiling() { m_useProfiling = true; }
void EnableStreams() { m_useStreams = true; }
void setup_stream_graph();
void preprocess_notify();
void enqueue_notify();
void wait_notify();
void preprocess();
void enqueue();
void wait();