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

SWDEV-301667 - Optimize performance when graph has single branch

Browse Source 
Three for loops iterate over all graph nodes for UpdateStream, FillCommands and
EnqueueCommands has performance drop for large graphs.

Change-Id: I077accf3a4680d5d944b73200fd6498a7a48f25c


[ROCm/clr commit: 530dc6de2a]
This commit is contained in:
Anusha GodavarthySurya
2023-08-28 12:13:20 +00:00
committed by Anusha Godavarthy Surya
parent a0ee7e2784
commit f464a6ade6
1 changed files with 27 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/src/hip_graph_internal.cpp
+27 -22
View File
@@ -602,23 +602,12 @@ hipError_t GraphExec::Run(hipStream_t stream) {
} else {
repeatLaunch_ = true;
}
UpdateStream(parallelLists_, hip_stream, this);
amd::Command* rootCommand = nullptr;
amd::Command* endCommand = nullptr;
status = FillCommands(parallelLists_, nodeWaitLists_, topoOrder_, clonedGraph_, rootCommand,
endCommand, hip_stream);
if (status != hipSuccess) {
return status;
}
if (rootCommand != nullptr) {
rootCommand->enqueue();
rootCommand->release();
}
for (int i = 0; i < topoOrder_.size(); i++) {
if (DEBUG_CLR_GRAPH_ENABLE_BUFFERING) {
// Enable buffering for graph with single branch
if (parallelLists_.size() == 1) {
if (parallelLists_.size() == 1) {
for (int i = 0; i < topoOrder_.size(); i++) {
topoOrder_[i]->SetStream(hip_stream, this);
status = topoOrder_[i]->CreateCommand(topoOrder_[i]->GetQueue());
if (DEBUG_CLR_GRAPH_ENABLE_BUFFERING) {
// Enable buffering for graph with single branch
// Peep through the next node. If current and next node are kernel then enable AQL
// buffering
if (((i + 1) != topoOrder_.size()) && topoOrder_[i]->GetType() == hipGraphNodeTypeKernel &&
@@ -626,12 +615,28 @@ hipError_t GraphExec::Run(hipStream_t stream) {
topoOrder_[i]->EnableBuffering();
}
}
topoOrder_[i]->EnqueueCommands(stream);
}
} else {
UpdateStream(parallelLists_, hip_stream, this);
amd::Command* rootCommand = nullptr;
amd::Command* endCommand = nullptr;
status = FillCommands(parallelLists_, nodeWaitLists_, topoOrder_, clonedGraph_, rootCommand,
endCommand, hip_stream);
if (status != hipSuccess) {
return status;
}
if (rootCommand != nullptr) {
rootCommand->enqueue();
rootCommand->release();
}
for (int i = 0; i < topoOrder_.size(); i++) {
topoOrder_[i]->EnqueueCommands(stream);
}
if (endCommand != nullptr) {
endCommand->enqueue();
endCommand->release();
}
topoOrder_[i]->EnqueueCommands(stream);
}
if (endCommand != nullptr) {
endCommand->enqueue();
endCommand->release();
}
ResetQueueIndex();
return status;