101 lines · plain
1; RUN: opt %loadNPMPolly -polly-pattern-matching-based-opts=true '-passes=polly-custom<optree;delicm;simplify;opt-isl>' -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s2; REQUIRES: asserts3 4; Check that the pattern matching detects the matrix multiplication pattern5; after a full run of -polly-optree and -polly-delicm, where the write access6; is not through the original memory access, but through a PHI node that was7; delicmed. This test covers the polybench 2mm and 3mm cases.8;9; This test case generates the following schedule, which contains filters:10;11; domain: "{ Stmt_for_body8[i0, i1, i2] : 0 <= i0 <= 1599 and12; 0 <= i1 <= 1799 and13; 0 <= i2 <= 2199;14; Stmt_for_body3[i0, i1] : 0 <= i0 <= 1599 and15; 0 <= i1 <= 1799;16; Stmt_for_body3_last[i0, i1] : 0 <= i0 <= 1599 and17; 0 <= i1 <= 1799 }"18; child:19; sequence:20; - filter: "{ Stmt_for_body3[i0, i1] }"21; child:22; schedule: "[{ Stmt_for_body3[i0, i1] -> [(i0)] }, { Stmt_for_body3[i0, i1] -> [(i1)] }]"23; permutable: 124; coincident: [ 1, 1 ]25; - filter: "{ Stmt_for_body3_last[i0, i1] }"26; child:27; schedule: "[{ Stmt_for_body3_last[i0, i1] -> [(i0)] }, { Stmt_for_body3_last[i0, i1] -> [(i1)] }]"28; permutable: 129; coincident: [ 1, 1 ]30; - filter: "{ Stmt_for_body8[i0, i1, i2] }"31; child:32; schedule: "[{ Stmt_for_body8[i0, i1, i2] -> [(i0)] },33; { Stmt_for_body8[i0, i1, i2] -> [(i1)] },34; { Stmt_for_body8[i0, i1, i2] -> [(i2)] }]"35; permutable: 136; coincident: [ 1, 1, 0 ]37;38; CHECK: The tensor contraction pattern was detected39; CHECK: The matrix multiplication pattern was detected40;41 42target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"43target triple = "x86_64-unknown-linux-gnu"44 45; Function Attrs: norecurse nounwind uwtable46define void @kernel_2mm(i32 %ni, i32 %nj, i32 %nk, i32 %nl, double %alpha, double %beta, ptr nocapture %tmp, ptr nocapture readonly %A, ptr nocapture readonly %B, ptr nocapture readnone %C, ptr nocapture readnone %D) local_unnamed_addr {47entry:48 br label %entry.split49 50entry.split: ; preds = %entry51 br label %for.body52 53for.body: ; preds = %for.inc25, %entry.split54 %indvars.iv50 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next51, %for.inc25 ]55 br label %for.body356 57for.body3: ; preds = %for.inc22, %for.body58 %indvars.iv46 = phi i64 [ 0, %for.body ], [ %indvars.iv.next47, %for.inc22 ]59 %arrayidx5 = getelementptr inbounds [1800 x double], ptr %tmp, i64 %indvars.iv50, i64 %indvars.iv4660 store double 0.000000e+00, ptr %arrayidx5, align 8, !tbaa !261 br label %for.body862 63for.body8: ; preds = %for.body8, %for.body364 %0 = phi double [ 0.000000e+00, %for.body3 ], [ %add, %for.body8 ]65 %indvars.iv = phi i64 [ 0, %for.body3 ], [ %indvars.iv.next, %for.body8 ]66 %arrayidx12 = getelementptr inbounds [2200 x double], ptr %A, i64 %indvars.iv50, i64 %indvars.iv67 %1 = load double, ptr %arrayidx12, align 8, !tbaa !268 %mul = fmul double %1, %alpha69 %arrayidx16 = getelementptr inbounds [1800 x double], ptr %B, i64 %indvars.iv, i64 %indvars.iv4670 %2 = load double, ptr %arrayidx16, align 8, !tbaa !271 %mul17 = fmul double %mul, %272 %add = fadd double %0, %mul1773 store double %add, ptr %arrayidx5, align 8, !tbaa !274 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 175 %exitcond = icmp eq i64 %indvars.iv.next, 220076 br i1 %exitcond, label %for.inc22, label %for.body877 78for.inc22: ; preds = %for.body879 %indvars.iv.next47 = add nuw nsw i64 %indvars.iv46, 180 %exitcond48 = icmp eq i64 %indvars.iv.next47, 180081 br i1 %exitcond48, label %for.inc25, label %for.body382 83for.inc25: ; preds = %for.inc2284 %indvars.iv.next51 = add nuw nsw i64 %indvars.iv50, 185 %exitcond52 = icmp eq i64 %indvars.iv.next51, 160086 br i1 %exitcond52, label %for.end27, label %for.body87 88for.end27: ; preds = %for.inc2589 ret void90}91 92!llvm.module.flags = !{!0}93!llvm.ident = !{!1}94 95!0 = !{i32 1, !"wchar_size", i32 4}96!1 = !{!"clang version 6.0.0 (trunk 309912) (llvm/trunk 309933)"}97!2 = !{!3, !3, i64 0}98!3 = !{!"double", !4, i64 0}99!4 = !{!"omnipotent char", !5, i64 0}100!5 = !{!"Simple C/C++ TBAA"}101