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

SWDEV-482086 - Fix hipGraphInstantiate leak

Browse Source 
* In a scenario where kernel is launched with hipExtLaunchKernelGGL and stop event is used, hipGraphInstantiate leaks. Since stop event is used, profiling is enabled and Timestamp (ReferencedCountedObject) is created, but it doesn't get released.
* The idea behind this solution is that profiling should be disabled when command is captured, hence the timestamp should not be created. Because information about capturing isn't available when kernel command is created, packet capturing state is used to determine whether to create a timestamp or not.

Change-Id: Ia23adac4592ded4fb5e236acf99e12e729f63692
This commit is contained in:
Vladana Stojiljkovic
2024-09-09 15:52:15 +02:00
parent d6193a2f23
commit da5f1a6146
3 changed files with 15 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/device/rocm/rocvirtual.cpp
+9 -7
View File
@@ -1545,7 +1545,9 @@ address VirtualGPU::allocKernelArguments(size_t size, size_t alignment) {
* and then calls start() to get the current host timestamp.
*/
void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
if (command.profilingInfo().enabled_) {
// Disable profiling when command is being captured to prevent memory leak from created timestamp_
// which won't get freed, since the command is not being executed until graph launch
if (!command.getPktCapturingState() && command.profilingInfo().enabled_) {
if (timestamp_ != nullptr) {
LogWarning("Trying to create a second timestamp in VirtualGPU. \
This could have unintended consequences.");
@@ -1581,7 +1583,7 @@ void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
}
}
}
if (command.getCapturingState()) {
if (command.getPktCapturingState()) {
currCmd_ = &command;
}
}
@@ -1592,7 +1594,7 @@ void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
* current host timestamp if no signal is available.
*/
void VirtualGPU::profilingEnd(amd::Command& command) {
if (command.profilingInfo().enabled_) {
if (!command.getPktCapturingState() && command.profilingInfo().enabled_) {
if (timestamp_->HwProfiling() == false) {
timestamp_->end();
}
@@ -3111,7 +3113,7 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
amd::Memory* const* memories =
reinterpret_cast<amd::Memory* const*>(parameters + kernelParams.memoryObjOffset());
bool isGraphCapture = currCmd_ != nullptr && currCmd_->getCapturingState();
bool isGraphCapture = currCmd_ != nullptr && currCmd_->getPktCapturingState();
for (int j = 0; j < iteration; j++) {
// Reset global size for dimension dim if split is needed
if (dim != -1) {
@@ -3424,9 +3426,9 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
if (isGraphCapture) {
// Dispatch the packet
if (!dispatchAqlPacket(&dispatchPacket, aqlHeaderWithOrder,
(sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS),
GPU_FLUSH_ON_EXECUTION, currCmd_->getCapturingState(),
currCmd_->getAqlPacket())) {
(sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS),
GPU_FLUSH_ON_EXECUTION, currCmd_->getPktCapturingState(),
currCmd_->getAqlPacket())) {
return false;
}
} else {