From 3a67addd482bdb12fa35cde61d9844fb0b003891 Mon Sep 17 00:00:00 2001 From: Saleel Kudchadker Date: Wed, 8 May 2024 21:11:17 +0000 Subject: [PATCH] SWDEV-459778 - Remove CPU wait for profiler - No cpu wait is needed when profiler is attached, Doing this changes the application profile when roctracer is attached. Change-Id: I2b9cfc48d697cf5ed54bb6a240d8c12bdb079171 [ROCm/clr commit: 51e43687235865fa6e15b79a1712f87b1bc45457] --- projects/clr/rocclr/platform/commandqueue.cpp | 3 --- 1 file changed, 3 deletions(-) diff --git a/projects/clr/rocclr/platform/commandqueue.cpp b/projects/clr/rocclr/platform/commandqueue.cpp index 6dde23d923..960a6d8dbf 100644 --- a/projects/clr/rocclr/platform/commandqueue.cpp +++ b/projects/clr/rocclr/platform/commandqueue.cpp @@ -123,9 +123,6 @@ void HostQueue::finish(bool cpu_wait) { // If command doesn't contain HW event and runtime didn't request CPU wait, // then force marker submit bool force_marker = false; - // Force CPU wait if profiler is enabled. Pytorch tests may use tracer's plugin and rely on - // profiling information to be available right after finish. - cpu_wait |= activity_prof::IsEnabled(OP_ID_DISPATCH); if (AMD_DIRECT_DISPATCH && (command != nullptr) && !cpu_wait) { void* hw_event = (command->NotifyEvent() != nullptr) ? command->NotifyEvent()->HwEvent() : command->HwEvent();