[polly] r268222 - [ScheduleOptimizer] Add -polly-opt-outer-coincidence option.
Michael Kruse via llvm-commits
llvm-commits at lists.llvm.org
Mon May 2 04:35:28 PDT 2016
Author: meinersbur
Date: Mon May 2 06:35:27 2016
New Revision: 268222
URL: http://llvm.org/viewvc/llvm-project?rev=268222&view=rev
Log:
[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.
Added:
polly/trunk/test/ScheduleOptimizer/outer_coincidence.ll
Modified:
polly/trunk/lib/Transform/ScheduleOptimizer.cpp
Modified: polly/trunk/lib/Transform/ScheduleOptimizer.cpp
URL: http://llvm.org/viewvc/llvm-project/polly/trunk/lib/Transform/ScheduleOptimizer.cpp?rev=268222&r1=268221&r2=268222&view=diff
==============================================================================
--- polly/trunk/lib/Transform/ScheduleOptimizer.cpp (original)
+++ polly/trunk/lib/Transform/ScheduleOptimizer.cpp Mon May 2 06:35:27 2016
@@ -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(Sco
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);
Added: polly/trunk/test/ScheduleOptimizer/outer_coincidence.ll
URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScheduleOptimizer/outer_coincidence.ll?rev=268222&view=auto
==============================================================================
--- polly/trunk/test/ScheduleOptimizer/outer_coincidence.ll (added)
+++ polly/trunk/test/ScheduleOptimizer/outer_coincidence.ll Mon May 2 06:35:27 2016
@@ -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);
More information about the llvm-commits
mailing list