[ScheduleOptimizer] Add -polly-opt-outer-coincidence option.

Add a command line switch to set the
isl_options_set_schedule_outer_coincidence option. ISL then tries to
build schedules where the outer member of a band satisfies the
coincidence constraints.

In practice this allows loop skewing for more parallelism in inner
loops.

llvm-svn: 268222
This commit is contained in:
Michael Kruse 2016-05-02 11:35:27 +00:00
parent 90f5fed10b
commit 315aa3278e
2 changed files with 89 additions and 0 deletions

View File

@ -102,6 +102,12 @@ static cl::opt<std::string>
cl::desc("Maximize the band depth (yes/no)"), cl::Hidden,
cl::init("yes"), cl::ZeroOrMore, cl::cat(PollyCategory));
static cl::opt<std::string> OuterCoincidence(
"polly-opt-outer-coincidence",
cl::desc("Try to construct schedules where the outer member of each band "
"satisfies the coincidence constraints (yes/no)"),
cl::Hidden, cl::init("no"), cl::ZeroOrMore, cl::cat(PollyCategory));
static cl::opt<int> PrevectorWidth(
"polly-prevect-width",
cl::desc(
@ -543,6 +549,20 @@ bool IslScheduleOptimizer::runOnScop(Scop &S) {
IslMaximizeBands = 1;
}
int IslOuterCoincidence;
if (OuterCoincidence == "yes") {
IslOuterCoincidence = 1;
} else if (OuterCoincidence == "no") {
IslOuterCoincidence = 0;
} else {
errs() << "warning: Option -polly-opt-outer-coincidence should either be "
"'yes' or 'no'. Falling back to default: 'no'\n";
IslOuterCoincidence = 0;
}
isl_options_set_schedule_outer_coincidence(S.getIslCtx(),
IslOuterCoincidence);
isl_options_set_schedule_serialize_sccs(S.getIslCtx(), IslSerializeSCCs);
isl_options_set_schedule_maximize_band_depth(S.getIslCtx(), IslMaximizeBands);
isl_options_set_schedule_max_constant_term(S.getIslCtx(), MaxConstantTerm);

View File

@ -0,0 +1,69 @@
; RUN: opt %loadPolly -polly-opt-isl -polly-ast -polly-tiling=0 -polly-parallel -polly-opt-outer-coincidence=no -analyze < %s | FileCheck %s
; RUN: opt %loadPolly -polly-opt-isl -polly-ast -polly-tiling=0 -polly-parallel -polly-opt-outer-coincidence=yes -analyze < %s | FileCheck %s --check-prefix=OUTER
; By skewing, the diagonal can be made parallel. ISL does this when the Check
; the 'outer_coincidence' option is enabled.
;
; void func(int m, int n, float A[static const restrict m][n]) {
; for (int i = 1; i < m; i+=1)
; for (int j = 1; j < n; j+=1)
; A[i][j] = A[i-1][j] + A[i][j-1];
;}
define void @func(i64 %m, i64 %n, float* noalias nonnull %A) #0 {
entry:
br label %for.cond
for.cond: ; preds = %for.inc11, %entry
%i.0 = phi i64 [ 1, %entry ], [ %add12, %for.inc11 ]
%cmp = icmp slt i64 %i.0, %m
br i1 %cmp, label %for.cond1.preheader, label %for.end13
for.cond1.preheader: ; preds = %for.cond
br label %for.cond1
for.cond1: ; preds = %for.cond1.preheader, %for.body3
%j.0 = phi i64 [ %add10, %for.body3 ], [ 1, %for.cond1.preheader ]
%cmp2 = icmp slt i64 %j.0, %n
br i1 %cmp2, label %for.body3, label %for.inc11
for.body3: ; preds = %for.cond1
%sub = add nsw i64 %i.0, -1
%tmp = mul nsw i64 %sub, %n
%arrayidx = getelementptr inbounds float, float* %A, i64 %tmp
%arrayidx4 = getelementptr inbounds float, float* %arrayidx, i64 %j.0
%tmp13 = load float, float* %arrayidx4, align 4
%sub5 = add nsw i64 %j.0, -1
%tmp14 = mul nsw i64 %i.0, %n
%arrayidx6 = getelementptr inbounds float, float* %A, i64 %tmp14
%arrayidx7 = getelementptr inbounds float, float* %arrayidx6, i64 %sub5
%tmp15 = load float, float* %arrayidx7, align 4
%add = fadd float %tmp13, %tmp15
%tmp16 = mul nsw i64 %i.0, %n
%arrayidx8 = getelementptr inbounds float, float* %A, i64 %tmp16
%arrayidx9 = getelementptr inbounds float, float* %arrayidx8, i64 %j.0
store float %add, float* %arrayidx9, align 4
%add10 = add nuw nsw i64 %j.0, 1
br label %for.cond1
for.inc11: ; preds = %for.cond1
%add12 = add nuw nsw i64 %i.0, 1
br label %for.cond
for.end13: ; preds = %for.cond
ret void
}
; CHECK: #pragma minimal dependence distance: 1
; CHECK-NEXT: for (int c0 = 0; c0 < m - 1; c0 += 1)
; CHECK-NEXT: #pragma minimal dependence distance: 1
; CHECK-NEXT: for (int c1 = 0; c1 < n - 1; c1 += 1)
; CHECK-NEXT: Stmt_for_body3(c0, c1);
; OUTER: #pragma minimal dependence distance: 1
; OUTER-NEXT: for (int c0 = 0; c0 < m + n - 3; c0 += 1)
; OUTER-NEXT: #pragma simd
; OUTER-NEXT: #pragma known-parallel
; OUTER-NEXT: for (int c1 = max(0, -m + c0 + 2); c1 <= min(n - 2, c0); c1 += 1)
; OUTER-NEXT: Stmt_for_body3(c0 - c1, c1);