llvm-project/llvm/lib/Target/AMDGPU/GCNSchedStrategy.h

113 lines
3.2 KiB
C
Raw Normal View History

//===-- GCNSchedStrategy.h - GCN Scheduler Strategy -*- C++ -*-------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_LIB_TARGET_AMDGPU_GCNSCHEDSTRATEGY_H
#define LLVM_LIB_TARGET_AMDGPU_GCNSCHEDSTRATEGY_H
#include "GCNRegPressure.h"
#include "llvm/CodeGen/MachineScheduler.h"
namespace llvm {
class SIMachineFunctionInfo;
class SIRegisterInfo;
class SISubtarget;
/// This is a minimal scheduler strategy. The main difference between this
/// and the GenericScheduler is that GCNSchedStrategy uses different
/// heuristics to determine excess/critical pressure sets. Its goal is to
/// maximize kernel occupancy (i.e. maximum number of waves per simd).
class GCNMaxOccupancySchedStrategy : public GenericScheduler {
friend class GCNScheduleDAGMILive;
SUnit *pickNodeBidirectional(bool &IsTopNode);
void pickNodeFromQueue(SchedBoundary &Zone, const CandPolicy &ZonePolicy,
const RegPressureTracker &RPTracker,
SchedCandidate &Cand);
void initCandidate(SchedCandidate &Cand, SUnit *SU,
bool AtTop, const RegPressureTracker &RPTracker,
const SIRegisterInfo *SRI,
unsigned SGPRPressure, unsigned VGPRPressure);
unsigned SGPRExcessLimit;
unsigned VGPRExcessLimit;
unsigned SGPRCriticalLimit;
unsigned VGPRCriticalLimit;
unsigned TargetOccupancy;
MachineFunction *MF;
public:
GCNMaxOccupancySchedStrategy(const MachineSchedContext *C);
SUnit *pickNode(bool &IsTopNode) override;
void initialize(ScheduleDAGMI *DAG) override;
void setTargetOccupancy(unsigned Occ) { TargetOccupancy = Occ; }
};
class GCNScheduleDAGMILive : public ScheduleDAGMILive {
const SISubtarget &ST;
const SIMachineFunctionInfo &MFI;
// Occupancy target at the beginning of function scheduling cycle.
unsigned StartingOccupancy;
// Minimal real occupancy recorder for the function.
unsigned MinOccupancy;
// Scheduling stage number.
unsigned Stage;
[AMDGPU] Cache live-ins and register pressure in scheduler Using LIS can be quite expensive, so caching of calculated region live-ins and pressure is implemented. It does two things: 1. Caches the info for the second stage when we schedule with decreased target occupancy. 2. Tracks the basic block from top to bottom thus eliminating the need to scan whole register file liveness at every region split in the middle of the block. The scheduling is now done in 3 stages instead of two, with the first one being really a no-op and only used to collect scheduling regions as sent by the scheduler driver. There is no functional change to the current behavior, only compilation speed is affected. In general computeBlockPressure() could be simplified if we switch to backward RP tracker, because scheduler sends regions within a block starting from the last upward. We could use a natural order of upward tracker to seamlessly change between regions of the same block, since live reg set of a previous tracked region would become a live-out of the next region. That however requires fixing upward tracker to properly account defs and uses of the same instruction as both are contributing to the current pressure. When we converge on the produced pressure we should be able to switch between them back and forth. In addition, backward tracker is less expensive as it uses LIS in recede less often than forward uses it in advance. At the moment the worst known case compilation time has improved from 26 minutes to 8.5. Differential Revision: https://reviews.llvm.org/D33117 llvm-svn: 303184
2017-05-17 00:11:26 +08:00
// Current region index.
size_t RegionIdx;
// Vecor of regions recorder for later rescheduling
SmallVector<std::pair<MachineBasicBlock::iterator,
MachineBasicBlock::iterator>, 32> Regions;
[AMDGPU] Cache live-ins and register pressure in scheduler Using LIS can be quite expensive, so caching of calculated region live-ins and pressure is implemented. It does two things: 1. Caches the info for the second stage when we schedule with decreased target occupancy. 2. Tracks the basic block from top to bottom thus eliminating the need to scan whole register file liveness at every region split in the middle of the block. The scheduling is now done in 3 stages instead of two, with the first one being really a no-op and only used to collect scheduling regions as sent by the scheduler driver. There is no functional change to the current behavior, only compilation speed is affected. In general computeBlockPressure() could be simplified if we switch to backward RP tracker, because scheduler sends regions within a block starting from the last upward. We could use a natural order of upward tracker to seamlessly change between regions of the same block, since live reg set of a previous tracked region would become a live-out of the next region. That however requires fixing upward tracker to properly account defs and uses of the same instruction as both are contributing to the current pressure. When we converge on the produced pressure we should be able to switch between them back and forth. In addition, backward tracker is less expensive as it uses LIS in recede less often than forward uses it in advance. At the moment the worst known case compilation time has improved from 26 minutes to 8.5. Differential Revision: https://reviews.llvm.org/D33117 llvm-svn: 303184
2017-05-17 00:11:26 +08:00
// Region live-in cache.
SmallVector<GCNRPTracker::LiveRegSet, 32> LiveIns;
// Region pressure cache.
SmallVector<GCNRegPressure, 32> Pressure;
[AMDGPU] Cache live-ins and register pressure in scheduler Using LIS can be quite expensive, so caching of calculated region live-ins and pressure is implemented. It does two things: 1. Caches the info for the second stage when we schedule with decreased target occupancy. 2. Tracks the basic block from top to bottom thus eliminating the need to scan whole register file liveness at every region split in the middle of the block. The scheduling is now done in 3 stages instead of two, with the first one being really a no-op and only used to collect scheduling regions as sent by the scheduler driver. There is no functional change to the current behavior, only compilation speed is affected. In general computeBlockPressure() could be simplified if we switch to backward RP tracker, because scheduler sends regions within a block starting from the last upward. We could use a natural order of upward tracker to seamlessly change between regions of the same block, since live reg set of a previous tracked region would become a live-out of the next region. That however requires fixing upward tracker to properly account defs and uses of the same instruction as both are contributing to the current pressure. When we converge on the produced pressure we should be able to switch between them back and forth. In addition, backward tracker is less expensive as it uses LIS in recede less often than forward uses it in advance. At the moment the worst known case compilation time has improved from 26 minutes to 8.5. Differential Revision: https://reviews.llvm.org/D33117 llvm-svn: 303184
2017-05-17 00:11:26 +08:00
// Temporary basic block live-in cache.
DenseMap<const MachineBasicBlock*, GCNRPTracker::LiveRegSet> MBBLiveIns;
// Return current region pressure.
GCNRegPressure getRealRegPressure() const;
[AMDGPU] Cache live-ins and register pressure in scheduler Using LIS can be quite expensive, so caching of calculated region live-ins and pressure is implemented. It does two things: 1. Caches the info for the second stage when we schedule with decreased target occupancy. 2. Tracks the basic block from top to bottom thus eliminating the need to scan whole register file liveness at every region split in the middle of the block. The scheduling is now done in 3 stages instead of two, with the first one being really a no-op and only used to collect scheduling regions as sent by the scheduler driver. There is no functional change to the current behavior, only compilation speed is affected. In general computeBlockPressure() could be simplified if we switch to backward RP tracker, because scheduler sends regions within a block starting from the last upward. We could use a natural order of upward tracker to seamlessly change between regions of the same block, since live reg set of a previous tracked region would become a live-out of the next region. That however requires fixing upward tracker to properly account defs and uses of the same instruction as both are contributing to the current pressure. When we converge on the produced pressure we should be able to switch between them back and forth. In addition, backward tracker is less expensive as it uses LIS in recede less often than forward uses it in advance. At the moment the worst known case compilation time has improved from 26 minutes to 8.5. Differential Revision: https://reviews.llvm.org/D33117 llvm-svn: 303184
2017-05-17 00:11:26 +08:00
// Compute and cache live-ins and pressure for all regions in block.
void computeBlockPressure(const MachineBasicBlock *MBB);
public:
GCNScheduleDAGMILive(MachineSchedContext *C,
std::unique_ptr<MachineSchedStrategy> S);
void schedule() override;
void finalizeSchedule() override;
};
} // End namespace llvm
#endif // GCNSCHEDSTRATEGY_H