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

SWDEV-422207 - Optimize graph end detection

Browse Source 
- Do not use extra barrier to detect graph end. If its a kernel node we
can use a completion signal for the last packet. Saves roughly 6us for
Phantom testcase per graph launch.

Change-Id: I5e0c2479d9964fbeda86ed97533f6718f49a7f91
This commit is contained in:
Saleel Kudchadker
2023-11-09 23:52:40 +00:00
parent f06368fd04
commit c3bd229f4f
3 changed files with 42 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hipamd/src/hip_graph_internal.cpp
+19 -3
View File
@@ -541,11 +541,14 @@ hipError_t GraphExec::Run(hipStream_t stream) {
if (parallelLists_.size() == 1) {
amd::AccumulateCommand* accumulate = nullptr;
bool isLastPacketKernel = false;
if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
accumulate = new amd::AccumulateCommand(*hip_stream);
uint8_t* lastCapturedPacket = (topoOrder_.back()->GetType() == hipGraphNodeTypeKernel) ?
topoOrder_.back()->GetAqlPacket() : nullptr;
accumulate = new amd::AccumulateCommand(*hip_stream, {}, nullptr, lastCapturedPacket);
}
for (int i = 0; i < topoOrder_.size(); i++) {
for (int i = 0; i < topoOrder_.size() - 1; i++) {
if (DEBUG_CLR_GRAPH_PACKET_CAPTURE && topoOrder_[i]->GetType() == hipGraphNodeTypeKernel) {
hip_stream->vdev()->dispatchAqlPacket(topoOrder_[i]->GetAqlPacket(), accumulate);
} else {
@@ -555,7 +558,20 @@ hipError_t GraphExec::Run(hipStream_t stream) {
}
}
if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
// If last captured packet is kernel, optimize to detect completion of last kernel
// This saves on extra packet submitted to determine end of graph
if (DEBUG_CLR_GRAPH_PACKET_CAPTURE && topoOrder_.back()->GetType() == hipGraphNodeTypeKernel) {
accumulate->enqueue();
accumulate->release();
isLastPacketKernel = true;
} else {
topoOrder_.back()->SetStream(hip_stream, this);
status = topoOrder_.back()->CreateCommand(topoOrder_.back()->GetQueue());
topoOrder_.back()->EnqueueCommands(stream);
}
// If last packet is not kernel, submit a marker to detect end of graph
if (DEBUG_CLR_GRAPH_PACKET_CAPTURE && !isLastPacketKernel) {
accumulate->enqueue();
accumulate->release();
}