80 lines · plain
1; RUN: opt %loadNPMPolly -polly-pattern-matching-based-opts=true -polly-target-throughput-vector-fma=1 -polly-target-latency-vector-fma=8 -polly-target-1st-cache-level-associativity=8 -polly-target-2nd-cache-level-associativity=8 -polly-target-1st-cache-level-size=32768 -polly-target-vector-register-bitwidth=256 -polly-target-2nd-cache-level-size=262144 '-passes=polly-custom<opt-isl>' -disable-output < %s2;3; RUN: opt %loadNPMPolly '-passes=polly-custom<deps>' -polly-print-deps -disable-output < %s | FileCheck %s --check-prefix=DEPENDENCES4;5; /* C := A * B + C */6; /* Elements of the matrices A, B, C have the char type. */7; /* The type size of elements of the matrix multiplication operands is used8; to determine the parameters of the code produced by the optimization9; of the matrix multiplication (e.g. bounds of the loops of the loop10; nest, the innermost loop body). This test checks the form of11; the generated loop nest. See getMicroKernelParams and12; getMacroKernelParams from lib/Transform/ScheduleOptimizer.cpp13; for details.14;15; This patch also checks that we can detect matrix multiplication16; in case there are reduction dependencies and there are not RAW17; dependencies. */18; for (i = 0; i < _PB_NI; i++)19; for (j = 0; j < _PB_NJ; j++)20; for (k = 0; k < _PB_NK; ++k)21; C[i][j] += A[i][k] * B[k][j];22;23; DEPENDENCES: RAW dependences:24; DEPENDENCES-NEXT: { }25; DEPENDENCES-NEXT: WAR dependences:26; DEPENDENCES-NEXT: { }27; DEPENDENCES-NEXT: WAW dependences:28; DEPENDENCES-NEXT: { }29; DEPENDENCES-NEXT: Reduction dependences:30; DEPENDENCES-NEXT: { Stmt_for_body6[i0, i1, i2] -> Stmt_for_body6[i0, i1, 1 + i2] : 0 <= i0 <= 1023 and 0 <= i1 <= 1023 and 0 <= i2 <= 1022 }31; DEPENDENCES-NEXT: Transitive closure of reduction dependences:32; DEPENDENCES-NEXT: { Stmt_for_body6[i0, i1, i2] -> Stmt_for_body6[i0, i1, o2] : 0 <= i0 <= 1023 and 0 <= i1 <= 1023 and ((i2 >= 0 and i2 < o2 <= 1023) or (i2 <= 1023 and 0 <= o2 < i2)) }33;34target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"35target triple = "x86_64-unknown-unknown"36 37define internal void @kernel_gemm(i32 %ni, i32 %nj, i32 %nk, i8 signext %alpha, i8 signext %beta, ptr %C, ptr %A, ptr %B) {38entry:39 br label %entry.split40 41entry.split: ; preds = %entry42 br label %for.cond1.preheader43 44for.cond1.preheader: ; preds = %for.inc23, %entry.split45 %indvars.iv45 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next46, %for.inc23 ]46 br label %for.cond4.preheader47 48for.cond4.preheader: ; preds = %for.inc20, %for.cond1.preheader49 %indvars.iv42 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next43, %for.inc20 ]50 br label %for.body651 52for.body6: ; preds = %for.body6, %for.cond4.preheader53 %indvars.iv = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next, %for.body6 ]54 %arrayidx8 = getelementptr inbounds [1024 x i8], ptr %A, i64 %indvars.iv45, i64 %indvars.iv55 %tmp = load i8, ptr %arrayidx8, align 156 %arrayidx12 = getelementptr inbounds [1024 x i8], ptr %B, i64 %indvars.iv, i64 %indvars.iv4257 %tmp1 = load i8, ptr %arrayidx12, align 158 %mul = mul i8 %tmp1, %tmp59 %arrayidx17 = getelementptr inbounds [1024 x i8], ptr %C, i64 %indvars.iv45, i64 %indvars.iv4260 %tmp2 = load i8, ptr %arrayidx17, align 161 %add = add i8 %mul, %tmp262 store i8 %add, ptr %arrayidx17, align 163 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 164 %exitcond = icmp ne i64 %indvars.iv.next, 102465 br i1 %exitcond, label %for.body6, label %for.inc2066 67for.inc20: ; preds = %for.body668 %indvars.iv.next43 = add nuw nsw i64 %indvars.iv42, 169 %exitcond44 = icmp ne i64 %indvars.iv.next43, 102470 br i1 %exitcond44, label %for.cond4.preheader, label %for.inc2371 72for.inc23: ; preds = %for.inc2073 %indvars.iv.next46 = add nuw nsw i64 %indvars.iv45, 174 %exitcond47 = icmp ne i64 %indvars.iv.next46, 102475 br i1 %exitcond47, label %for.cond1.preheader, label %for.end2576 77for.end25: ; preds = %for.inc2378 ret void79}80