replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[ScheduleOptimizer] Make matmul pattern detection work with delicm output 

In certain cases delicm might decide to not leave the original array write in
the loop body, but to remove it and instead leave a transformed phi node as
write access. This commit teached the matmul pattern detection to order the
memory accesses according to when the access actually happens and use this
information to detect the new pattern. This makes pattern based matmul
optimization work for 2mm and 3mm in polybench 4 after
polly-position=before-vectorizer has been enabled.

llvm-svn: 310338
This commit is contained in:
Tobias Grosser 2017-08-08 06:15:15 +00:00
parent 03aa3aee11
commit 327e9ecb0d
4 changed files with 125 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
polly/include/polly/Simplify.h
View File

@ -14,12 +14,33 @@
#ifndef POLLY_TRANSFORM_SIMPLIFY_H
#define POLLY_TRANSFORM_SIMPLIFY_H
#include "llvm/ADT/SmallVector.h"
namespace llvm {
class PassRegistry;
class Pass;
} // namespace llvm
namespace polly {
class MemoryAccess;
class ScopStmt;
/// Return a vector that contains MemoryAccesses in the order in
/// which they are executed.
///
/// The order is:
/// - Implicit reads (BlockGenerator::generateScalarLoads)
/// - Explicit reads and writes (BlockGenerator::generateArrayLoad,
/// BlockGenerator::generateArrayStore)
/// - In block statements, the accesses are in order in which their
/// instructions are executed.
/// - In region statements, that order of execution is not predictable at
/// compile-time.
/// - Implicit writes (BlockGenerator::generateScalarStores)
/// The order in which implicit writes are executed relative to each other is
/// undefined.
llvm::SmallVector<MemoryAccess *, 32> getAccessesInOrder(ScopStmt &Stmt);
llvm::Pass *createSimplifyPass();
} // namespace polly

9
polly/lib/Transform/ScheduleOptimizer.cpp
View File

@ -52,6 +52,7 @@
#include "polly/LinkAllPasses.h"
#include "polly/Options.h"
#include "polly/ScopInfo.h"
#include "polly/Simplify.h"
#include "polly/Support/GICHelper.h"
#include "polly/Support/ISLOStream.h"
#include "llvm/Analysis/TargetTransformInfo.h"
@ -670,7 +671,9 @@ static bool containsOnlyMatrMultAcc(isl::map PartialSchedule,
permuteDimensions(PartialSchedule, isl::dim::out, MMI.j, OutDimNum - 1);
auto MapK =
permuteDimensions(PartialSchedule, isl::dim::out, MMI.k, OutDimNum - 1);
for (auto *MemA = Stmt->begin(); MemA != Stmt->end() - 1; MemA++) {
auto Accesses = getAccessesInOrder(*Stmt);
for (auto *MemA = Accesses.begin(); MemA != Accesses.end() - 1; MemA++) {
auto *MemAccessPtr = *MemA;
if (MemAccessPtr->isLatestArrayKind() && MemAccessPtr != MMI.WriteToC &&
!isMatMulNonScalarReadAccess(MemAccessPtr, MMI) &&
@ -745,7 +748,9 @@ static bool containsMatrMult(isl::map PartialSchedule, const Dependences *D,
auto *Stmt = static_cast<ScopStmt *>(InputDimsId.get_user());
if (Stmt->size() <= 1)
return false;
for (auto *MemA = Stmt->end() - 1; MemA != Stmt->begin(); MemA--) {
auto Accesses = getAccessesInOrder(*Stmt);
for (auto *MemA = Accesses.end() - 1; MemA != Accesses.begin(); MemA--) {
auto *MemAccessPtr = *MemA;
if (!MemAccessPtr->isLatestArrayKind())
continue;

54
polly/lib/Transform/Simplify.cpp
View File

@ -98,39 +98,6 @@ static isl::union_map underapproximatedAddMap(isl::union_map UMap,
return UResult;
}
/// Return a vector that contains MemoryAccesses in the order in
/// which they are executed.
///
/// The order is:
/// - Implicit reads (BlockGenerator::generateScalarLoads)
/// - Explicit reads and writes (BlockGenerator::generateArrayLoad,
/// BlockGenerator::generateArrayStore)
/// - In block statements, the accesses are in order in which their
/// instructions are executed.
/// - In region statements, that order of execution is not predictable at
/// compile-time.
/// - Implicit writes (BlockGenerator::generateScalarStores)
/// The order in which implicit writes are executed relative to each other is
/// undefined.
static SmallVector<MemoryAccess *, 32> getAccessesInOrder(ScopStmt &Stmt) {
SmallVector<MemoryAccess *, 32> Accesses;
for (MemoryAccess *MemAcc : Stmt)
if (isImplicitRead(MemAcc))
Accesses.push_back(MemAcc);
for (MemoryAccess *MemAcc : Stmt)
if (isExplicitAccess(MemAcc))
Accesses.push_back(MemAcc);
for (MemoryAccess *MemAcc : Stmt)
if (isImplicitWrite(MemAcc))
Accesses.push_back(MemAcc);
return Accesses;
}
class Simplify : public ScopPass {
private:
/// The last/current SCoP that is/has been processed.
@ -700,6 +667,27 @@ public:
char Simplify::ID;
} // anonymous namespace
namespace polly {
SmallVector<MemoryAccess *, 32> getAccessesInOrder(ScopStmt &Stmt) {
SmallVector<MemoryAccess *, 32> Accesses;
for (MemoryAccess *MemAcc : Stmt)
if (isImplicitRead(MemAcc))
Accesses.push_back(MemAcc);
for (MemoryAccess *MemAcc : Stmt)
if (isExplicitAccess(MemAcc))
Accesses.push_back(MemAcc);
for (MemoryAccess *MemAcc : Stmt)
if (isImplicitWrite(MemAcc))
Accesses.push_back(MemAcc);
return Accesses;
}
} // namespace polly
Pass *polly::createSimplifyPass() { return new Simplify(); }
INITIALIZE_PASS_BEGIN(Simplify, "polly-simplify", "Polly - Simplify", false,

76
polly/test/ScheduleOptimizer/pattern-matching-based-opts-after-delicm.ll Normal file
View File

@ -0,0 +1,76 @@
; RUN: opt %loadPolly \
; RUN: -polly-pattern-matching-based-opts=true \
; RUN: -polly-optree -polly-delicm -polly-simplify \
; RUN: -polly-opt-isl -debug < %s 2>&1 \
; RUN: | FileCheck %s
; REQUIRES: asserts
; Check that the pattern matching detects the matrix multiplication pattern
; after a full run of -polly-optree and -polly-delicm, where the write access
; is not through the original memory access, but trough a PHI node that was
; delicmed. This test covers the polybench 2mm and 3mm cases.
;
; CHECK: The matrix multiplication pattern was detected
;
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
; Function Attrs: norecurse nounwind uwtable
define void @kernel_2mm(i32 %ni, i32 %nj, i32 %nk, i32 %nl, double %alpha, double %beta, [1800 x double]* nocapture %tmp, [2200 x double]* nocapture readonly %A, [1800 x double]* nocapture readonly %B, [2400 x double]* nocapture readnone %C, [2400 x double]* nocapture readnone %D) local_unnamed_addr #0 {
entry:
br label %entry.split
entry.split: ; preds = %entry
br label %for.body
for.body: ; preds = %for.inc25, %entry.split
%indvars.iv50 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next51, %for.inc25 ]
br label %for.body3
for.body3: ; preds = %for.inc22, %for.body
%indvars.iv46 = phi i64 [ 0, %for.body ], [ %indvars.iv.next47, %for.inc22 ]
%arrayidx5 = getelementptr inbounds [1800 x double], [1800 x double]* %tmp, i64 %indvars.iv50, i64 %indvars.iv46
store double 0.000000e+00, double* %arrayidx5, align 8, !tbaa !2
br label %for.body8
for.body8: ; preds = %for.body8, %for.body3
%0 = phi double [ 0.000000e+00, %for.body3 ], [ %add, %for.body8 ]
%indvars.iv = phi i64 [ 0, %for.body3 ], [ %indvars.iv.next, %for.body8 ]
%arrayidx12 = getelementptr inbounds [2200 x double], [2200 x double]* %A, i64 %indvars.iv50, i64 %indvars.iv
%1 = load double, double* %arrayidx12, align 8, !tbaa !2
%mul = fmul double %1, %alpha
%arrayidx16 = getelementptr inbounds [1800 x double], [1800 x double]* %B, i64 %indvars.iv, i64 %indvars.iv46
%2 = load double, double* %arrayidx16, align 8, !tbaa !2
%mul17 = fmul double %mul, %2
%add = fadd double %0, %mul17
store double %add, double* %arrayidx5, align 8, !tbaa !2
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%exitcond = icmp eq i64 %indvars.iv.next, 2200
br i1 %exitcond, label %for.inc22, label %for.body8
for.inc22: ; preds = %for.body8
%indvars.iv.next47 = add nuw nsw i64 %indvars.iv46, 1
%exitcond48 = icmp eq i64 %indvars.iv.next47, 1800
br i1 %exitcond48, label %for.inc25, label %for.body3
for.inc25: ; preds = %for.inc22
%indvars.iv.next51 = add nuw nsw i64 %indvars.iv50, 1
%exitcond52 = icmp eq i64 %indvars.iv.next51, 1600
br i1 %exitcond52, label %for.end27, label %for.body
for.end27: ; preds = %for.inc25
ret void
}
attributes #0 = { norecurse nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="haswell" "target-features"="+aes,+avx,+avx2,+bmi,+bmi2,+cmov,+cx16,+f16c,+fma,+fsgsbase,+fxsr,+lzcnt,+mmx,+movbe,+pclmul,+popcnt,+rdrnd,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave,+xsaveopt,-adx,-avx512bw,-avx512cd,-avx512dq,-avx512er,-avx512f,-avx512ifma,-avx512pf,-avx512vbmi,-avx512vl,-avx512vpopcntdq,-clflushopt,-clwb,-clzero,-fma4,-lwp,-mwaitx,-pku,-prefetchwt1,-prfchw,-rdseed,-rtm,-sgx,-sha,-sse4a,-tbm,-xop,-xsavec,-xsaves" "unsafe-fp-math"="false" "use-soft-float"="false" }
!llvm.module.flags = !{!0}
!llvm.ident = !{!1}
!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{!"clang version 6.0.0 (trunk 309912) (llvm/trunk 309933)"}
!2 = !{!3, !3, i64 0}
!3 = !{!"double", !4, i64 0}
!4 = !{!"omnipotent char", !5, i64 0}
!5 = !{!"Simple C/C++ TBAA"}