1; RUN: opt %loadNPMPolly -passes=polly-opt-isl -polly-pattern-matching-based-opts=true \ 2; RUN: -debug -polly-tc-opt=true -disable-output < %s 2>&1 | FileCheck %s 3; RUN: opt %loadNPMPolly '-passes=polly-opt-isl,print<polly-ast>' -polly-pattern-matching-based-opts=true \ 4; RUN: -polly-target-throughput-vector-fma=1 \ 5; RUN: -polly-target-latency-vector-fma=8 \ 6; RUN: -polly-target-1st-cache-level-size=32768 \ 7; RUN: -polly-target-vector-register-bitwidth=256 \ 8; RUN: -polly-target-2nd-cache-level-size=262144 \ 9; RUN: -polly-tc-opt=true -disable-output < %s | \ 10; RUN: FileCheck %s --check-prefix=PATTERN-MATCHING-OPTS 11; REQUIRES: asserts 12; 13; C := A * B + C 14; Check that the pattern matching optimizations can detect different 15; permutations of GEMM loop and produce the correct ISL AST. In this case, 16; dimensions of band nodes can be implicitly permuted by the algorithm 17; applied during the schedule generation. It should be taken into the 18; account during the pattern matching optimizations. 19; for (i = 0; i < _PB_NI; i++) 20; for (k = 0; k < _PB_NK; ++k) 21; for (j = 0; j < _PB_NJ; j++) 22; C[i][j] += A[i][k] * B[k][j]; 23; 24; CHECK: The tensor contraction pattern was detected 25; CHECK: The matrix multiplication pattern was detected 26; 27; PATTERN-MATCHING-OPTS: // 1st level tiling - Tiles 28; PATTERN-MATCHING-OPTS-NEXT: for (int c1 = 0; c1 <= 3; c1 += 1) { 29; PATTERN-MATCHING-OPTS-NEXT: for (int c3 = 256 * c1; c3 <= 256 * c1 + 255; c3 += 1) 30; PATTERN-MATCHING-OPTS-NEXT: for (int c4 = 0; c4 <= 1023; c4 += 1) 31; PATTERN-MATCHING-OPTS-NEXT: CopyStmt_0(0, c3, c4); 32; PATTERN-MATCHING-OPTS-NEXT: for (int c2 = 0; c2 <= 10; c2 += 1) { 33; PATTERN-MATCHING-OPTS-NEXT: for (int c6 = 96 * c2; c6 <= min(1023, 96 * c2 + 95); c6 += 1) 34; PATTERN-MATCHING-OPTS-NEXT: for (int c7 = 256 * c1; c7 <= 256 * c1 + 255; c7 += 1) 35; PATTERN-MATCHING-OPTS-NEXT: CopyStmt_1(0, c1, c2, c6, c7); 36; PATTERN-MATCHING-OPTS-NEXT: // 1st level tiling - Points 37; PATTERN-MATCHING-OPTS-NEXT: // Register tiling - Tiles 38; PATTERN-MATCHING-OPTS-NEXT: for (int c3 = 0; c3 <= 127; c3 += 1) 39; PATTERN-MATCHING-OPTS-NEXT: for (int c4 = 0; c4 <= min(23, -24 * c2 + 255); c4 += 1) 40; PATTERN-MATCHING-OPTS-NEXT: for (int c5 = 0; c5 <= 255; c5 += 1) { 41; PATTERN-MATCHING-OPTS-NEXT: // Loop Vectorizer Disabled 42; PATTERN-MATCHING-OPTS-NEXT: // Register tiling - Points 43; PATTERN-MATCHING-OPTS-NEXT: { 44; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3); 45; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 1); 46; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 2); 47; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 3); 48; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 4); 49; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 5); 50; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 6); 51; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4, 256 * c1 + c5, 8 * c3 + 7); 52; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3); 53; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 1); 54; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 2); 55; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 3); 56; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 4); 57; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 5); 58; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 6); 59; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 1, 256 * c1 + c5, 8 * c3 + 7); 60; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3); 61; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 1); 62; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 2); 63; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 3); 64; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 4); 65; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 5); 66; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 6); 67; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 2, 256 * c1 + c5, 8 * c3 + 7); 68; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3); 69; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 1); 70; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 2); 71; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 3); 72; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 4); 73; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 5); 74; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 6); 75; PATTERN-MATCHING-OPTS-NEXT: Stmt_for_body6(96 * c2 + 4 * c4 + 3, 256 * c1 + c5, 8 * c3 + 7); 76; PATTERN-MATCHING-OPTS-NEXT: } 77; PATTERN-MATCHING-OPTS-NEXT: } 78; PATTERN-MATCHING-OPTS-NEXT: } 79; PATTERN-MATCHING-OPTS-NEXT: } 80; 81target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 82target triple = "x86_64-unknown-unknown" 83 84define internal void @kernel_gemm(i32 %ni, i32 %nj, i32 %nk, double %alpha, double %beta, ptr %C, ptr %A, ptr %B) { 85entry: 86 br label %entry.split 87 88entry.split: ; preds = %entry 89 br label %for.cond1.preheader 90 91for.cond1.preheader: ; preds = %for.inc20, %entry.split 92 %indvars.iv41 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next42, %for.inc20 ] 93 br label %for.cond4.preheader 94 95for.cond4.preheader: ; preds = %for.inc17, %for.cond1.preheader 96 %indvars.iv38 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next39, %for.inc17 ] 97 br label %for.body6 98 99for.body6: ; preds = %for.body6, %for.cond4.preheader 100 %indvars.iv = phi i64 [ 0, %for.cond4.preheader ], [ %indvars.iv.next, %for.body6 ] 101 %arrayidx8 = getelementptr inbounds [1024 x double], ptr %A, i64 %indvars.iv41, i64 %indvars.iv38 102 %tmp = load double, ptr %arrayidx8, align 8 103 %arrayidx12 = getelementptr inbounds [1024 x double], ptr %B, i64 %indvars.iv38, i64 %indvars.iv 104 %tmp1 = load double, ptr %arrayidx12, align 8 105 %mul = fmul double %tmp, %tmp1 106 %arrayidx16 = getelementptr inbounds [1024 x double], ptr %C, i64 %indvars.iv41, i64 %indvars.iv 107 %tmp2 = load double, ptr %arrayidx16, align 8 108 %add = fadd double %tmp2, %mul 109 store double %add, ptr %arrayidx16, align 8 110 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 111 %exitcond = icmp ne i64 %indvars.iv.next, 1024 112 br i1 %exitcond, label %for.body6, label %for.inc17 113 114for.inc17: ; preds = %for.body6 115 %indvars.iv.next39 = add nuw nsw i64 %indvars.iv38, 1 116 %exitcond40 = icmp ne i64 %indvars.iv.next39, 1024 117 br i1 %exitcond40, label %for.cond4.preheader, label %for.inc20 118 119for.inc20: ; preds = %for.inc17 120 %indvars.iv.next42 = add nuw nsw i64 %indvars.iv41, 1 121 %exitcond43 = icmp ne i64 %indvars.iv.next42, 1024 122 br i1 %exitcond43, label %for.cond1.preheader, label %for.end22 123 124for.end22: ; preds = %for.inc20 125 ret void 126} 127