This is an archive of the discontinued LLVM Phabricator instance.

Differential D142262

[AMDGPU] Use more consistemt way to avoid overflow in the scheduler
ClosedPublic

Authored by rampitec on Jan 20 2023, 3:22 PM.

Details

Reviewers
kerbowa
Commits
rG7d0145cc4748: [AMDGPU] Use more consistemt way to avoid overflow in the scheduler
Summary

Use more consistent way to avoid overflow when calculating SGPR
and VGPR pressure limits.

Diff Detail

Event Timeline

rampitec created this revision.Jan 20 2023, 3:22 PM
Herald added a project: Restricted Project. · View Herald TranscriptJan 20 2023, 3:22 PM
Herald added subscribers: kosarev, foad, hiraditya and 6 others. · View Herald Transcript
rampitec requested review of this revision.Jan 20 2023, 3:22 PM
Herald added a project: Restricted Project. · View Herald TranscriptJan 20 2023, 3:22 PM
Herald added a subscriber: wdng. · View Herald Transcript
rampitec added a parent revision: D141876: [AMDGPU] Tune scheduler on GFX10 and GFX11 for regions with spilling.Jan 20 2023, 3:23 PM
Harbormaster completed remote builds in B209083: Diff 490988.Jan 20 2023, 6:16 PM
arsenm added inline comments.Jan 20 2023, 6:29 PM
llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
91–92

clamp with min of 0 seems weird. Also could use -=

rampitec updated this revision to Diff 491031.Jan 20 2023, 9:28 PM
rampitec marked an inline comment as done.
rampitec retitled this revision from [AMDGPU] Use std::clamp to avoid overflow in the scheduler to [AMDGPU] Use more consistemt way to avoid overflow in the scheduler.
rampitec edited the summary of this revision. (Show Details)
Harbormaster completed remote builds in B209114: Diff 491031.Jan 20 2023, 11:51 PM
kerbowa accepted this revision.Jan 22 2023, 10:27 PM
This revision is now accepted and ready to land.Jan 22 2023, 10:27 PM
Closed by commit rG7d0145cc4748: [AMDGPU] Use more consistemt way to avoid overflow in the scheduler (authored by rampitec). · Explain WhyJan 23 2023, 11:01 AM
This revision was automatically updated to reflect the committed changes.
rampitec added a commit: rG7d0145cc4748: [AMDGPU] Use more consistemt way to avoid overflow in the scheduler.

Revision Contents

PathSize
llvm/
lib/
Target/
AMDGPU/
16 lines

Diff 491457

llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp

Show First 20 Lines • Show All 82 Lines • ▼ Show 20 Lines if (!KnownExcessRP) {
unsigned Granule = AMDGPU::IsaInfo::getVGPRAllocGranule(&ST); unsigned Granule = AMDGPU::IsaInfo::getVGPRAllocGranule(&ST);
unsigned Addressable = AMDGPU::IsaInfo::getAddressableNumVGPRs(&ST); unsigned Addressable = AMDGPU::IsaInfo::getAddressableNumVGPRs(&ST);
unsigned VGPRBudget = alignDown(Addressable / TargetOccupancy, Granule); unsigned VGPRBudget = alignDown(Addressable / TargetOccupancy, Granule);
VGPRBudget = std::max(VGPRBudget, Granule); VGPRBudget = std::max(VGPRBudget, Granule);
VGPRCriticalLimit = std::min(VGPRBudget, VGPRExcessLimit); VGPRCriticalLimit = std::min(VGPRBudget, VGPRExcessLimit);
} }
// Subtract error margin and bias from register limits and avoid overflow. // Subtract error margin and bias from register limits and avoid overflow.
SGPRCriticalLimit = SGPRCriticalLimit -= std::min(SGPRLimitBias + ErrorMargin, SGPRCriticalLimit);
arsenmUnsubmitted
Done
ReplyInline Actions

clamp with min of 0 seems weird. Also could use -=

arsenm: clamp with min of 0 seems weird. Also could use -=
std::min(SGPRCriticalLimit - SGPRLimitBias, SGPRCriticalLimit); VGPRCriticalLimit -= std::min(VGPRLimitBias + ErrorMargin, VGPRCriticalLimit);
SGPRCriticalLimit = SGPRExcessLimit -= std::min(SGPRLimitBias + ErrorMargin, SGPRExcessLimit);
std::min(SGPRCriticalLimit - ErrorMargin, SGPRCriticalLimit); VGPRExcessLimit -= std::min(VGPRLimitBias + ErrorMargin, VGPRExcessLimit);
VGPRCriticalLimit =
std::min(VGPRCriticalLimit - VGPRLimitBias, VGPRCriticalLimit);
VGPRCriticalLimit =
std::min(VGPRCriticalLimit - ErrorMargin, VGPRCriticalLimit);
SGPRExcessLimit = std::min(SGPRExcessLimit - SGPRLimitBias, SGPRExcessLimit);
SGPRExcessLimit = std::min(SGPRExcessLimit - ErrorMargin, SGPRExcessLimit);
VGPRExcessLimit = std::min(VGPRExcessLimit - VGPRLimitBias, VGPRExcessLimit);
VGPRExcessLimit = std::min(VGPRExcessLimit - ErrorMargin, VGPRExcessLimit);
LLVM_DEBUG(dbgs() << "VGPRCriticalLimit = " << VGPRCriticalLimit LLVM_DEBUG(dbgs() << "VGPRCriticalLimit = " << VGPRCriticalLimit
<< ", VGPRExcessLimit = " << VGPRExcessLimit << ", VGPRExcessLimit = " << VGPRExcessLimit
<< ", SGPRCriticalLimit = " << SGPRCriticalLimit << ", SGPRCriticalLimit = " << SGPRCriticalLimit
<< ", SGPRExcessLimit = " << SGPRExcessLimit << "\n\n"); << ", SGPRExcessLimit = " << SGPRExcessLimit << "\n\n");
} }
void GCNSchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU, void GCNSchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU,
▲ Show 20 Lines • Show All 1,439 Lines • Show Last 20 Lines