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

[GPU] Fix the new infer request to set output memory normally during enqueue stage (#12938)

Browse Source 
* Fix the new infer request to set output memory normally during enqueue stage
* Fix to use the generated outputsMap instead of graph API to retrieve outputID

Signed-off-by: Andrew Park <andrew.park@intel.com>
This commit is contained in:
Andrew Kwangwoong Park 2022-09-08 16:15:04 +09:00 committed by GitHub
parent 5a42afb45b
commit 90093d5315
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 17 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
src/plugins/intel_gpu/src/plugin/infer_request.cpp
View File

@ -522,15 +522,15 @@ void InferRequest::wait() {
// wait for completion & collect outputs as requested by the model // wait for completion & collect outputs as requested by the model
for (auto& no : _networkOutputs) { for (auto& no : _networkOutputs) {
std::string outputID = m_graph->MapOutputName(no.first); // In dynamic case, graph API must be used to retrieve outputID
// because it does not create outputsMap during SetGraph
std::string outputID = outputsMap.empty() ? m_graph->MapOutputName(no.first) : outputsMap.at(no.first);
auto outputMemory = internal_outputs.at(outputID).get_memory(); auto outputMemory = internal_outputs.at(outputID).get_memory();
auto node = findOutputByNodeName(no.first);
auto out_partial_shape = node->get_output_partial_shape(0);
size_t out_rank = out_partial_shape.rank().get_length();
if (_outputs.find(no.first) == _outputs.end()) { if (_outputs.find(no.first) == _outputs.end()) {
auto node = findOutputByNodeName(no.first);
auto out_partial_shape = node->get_output_partial_shape(0);
size_t out_rank = out_partial_shape.rank().get_length();
auto mem_dims = outputMemory->get_layout().get_shape(); auto mem_dims = outputMemory->get_layout().get_shape();
auto precision = InferenceEngine::Precision::FP32; auto precision = InferenceEngine::Precision::FP32;
auto dims = SizeVector(mem_dims.begin(), mem_dims.end()); auto dims = SizeVector(mem_dims.begin(), mem_dims.end());
@ -891,22 +891,19 @@ void InferRequest::prepare_input(const cldnn::primitive_id& inputName, Blob::Ptr
void InferRequest::prepare_output(const cldnn::primitive_id& outputName, Blob::Ptr& outputBlob) { void InferRequest::prepare_output(const cldnn::primitive_id& outputName, Blob::Ptr& outputBlob) {
OV_ITT_SCOPED_TASK(itt::domains::intel_gpu_plugin, "InferRequest::prepare_output"); OV_ITT_SCOPED_TASK(itt::domains::intel_gpu_plugin, "InferRequest::prepare_output");
Blob::Ptr reqBlob = _deviceOutputs.at(outputName);
cldnn::primitive_id internalName = outputsMap[outputName];
auto _nw_ptr = m_graph->GetNetwork();
auto remote_ptr = outputBlob->as<gpu::ClBlob>(); auto remote_ptr = outputBlob->as<gpu::ClBlob>();
bool is_remote = remote_ptr != nullptr; auto output_blob_ptr = (reqBlob != outputBlob && remote_ptr != nullptr)
if (is_remote) { ? remote_ptr
Blob::Ptr reqBlob = _deviceOutputs.at(outputName); : reqBlob->as<gpu::ClBlob>();
cldnn::primitive_id internalName = outputsMap[outputName]; auto impl = getBlobImpl(output_blob_ptr);
auto _nw_ptr = m_graph->GetNetwork(); if (!impl->is_allocated()) {
auto output_blob_ptr = (reqBlob != outputBlob && remote_ptr != nullptr) IE_THROW(NotAllocated) << str_output_not_allocated;
? remote_ptr
: reqBlob->as<gpu::ClBlob>();
auto impl = getBlobImpl(output_blob_ptr);
if (!impl->is_allocated()) {
IE_THROW(NotAllocated) << str_output_not_allocated;
}
auto outputMem = impl->getMemory();
_nw_ptr->set_output_memory(internalName, outputMem);
} }
auto outputMem = impl->getMemory();
_nw_ptr->set_output_memory(internalName, outputMem);
} }
InferenceEngine::Blob::Ptr InferRequest::create_device_blob(const InferenceEngine::TensorDesc& desc) { InferenceEngine::Blob::Ptr InferRequest::create_device_blob(const InferenceEngine::TensorDesc& desc) {