67 lines · plain
1; RUN: opt %loadNPMPolly '-passes=polly-custom<opt-isl>' -polly-pattern-matching-based-opts=true -polly-tc-opt=true -debug -disable-output < %s 2>&1 | FileCheck %s2; REQUIRES: asserts3;4; /* C := alpha*A*B + beta*C */5; for (i = 0; i < _PB_NI; i++)6; for (j = 0; j < _PB_NJ; j += 2)7; {8; C[i][j] *= beta;9; for (k = 0; k < _PB_NK; ++k)10; C[i][j] += alpha * A[i][k] * B[k][j];11; }12;13; Check that we won’t detect the matrix multiplication pattern,14; if, for example, there are memory accesses that have stride 215; after the interchanging of loops.16;17; CHECK-NOT: The matrix multiplication pattern was detected18; CHECK-NOT: The tensor contraction pattern was detected19;20target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"21target triple = "x86_64-unknown-unknown"22 23define internal void @kernel_gemm(i32 %arg, i32 %arg1, i32 %arg2, double %arg3, double %arg4, ptr %arg5, ptr %arg6, ptr %arg7) #0 {24bb:25 br label %bb826 27bb8: ; preds = %bb29, %bb28 %tmp = phi i64 [ 0, %bb ], [ %tmp30, %bb29 ]29 br label %bb930 31bb9: ; preds = %bb26, %bb832 %tmp10 = phi i64 [ 0, %bb8 ], [ %tmp27, %bb26 ]33 %tmp11 = getelementptr inbounds [1056 x double], ptr %arg5, i64 %tmp, i64 %tmp1034 %tmp12 = load double, ptr %tmp11, align 835 %tmp13 = fmul double %tmp12, %arg436 store double %tmp13, ptr %tmp11, align 837 br label %Copy_038 39Copy_0: ; preds = %Copy_0, %bb940 %tmp15 = phi i64 [ 0, %bb9 ], [ %tmp24, %Copy_0 ]41 %tmp16 = getelementptr inbounds [1024 x double], ptr %arg6, i64 %tmp, i64 %tmp1542 %tmp17 = load double, ptr %tmp16, align 843 %tmp18 = fmul double %tmp17, %arg344 %tmp19 = getelementptr inbounds [1056 x double], ptr %arg7, i64 %tmp15, i64 %tmp1045 %tmp20 = load double, ptr %tmp19, align 846 %tmp21 = fmul double %tmp18, %tmp2047 %tmp22 = load double, ptr %tmp11, align 848 %tmp23 = fadd double %tmp22, %tmp2149 store double %tmp23, ptr %tmp11, align 850 %tmp24 = add nuw nsw i64 %tmp15, 151 %tmp25 = icmp ne i64 %tmp24, 102452 br i1 %tmp25, label %Copy_0, label %bb2653 54bb26: ; preds = %Copy_055 %tmp27 = add nuw nsw i64 %tmp10, 256 %tmp28 = icmp ne i64 %tmp27, 105657 br i1 %tmp28, label %bb9, label %bb2958 59bb29: ; preds = %bb2660 %tmp30 = add nuw nsw i64 %tmp, 161 %tmp31 = icmp ne i64 %tmp30, 105662 br i1 %tmp31, label %bb8, label %bb3263 64bb32: ; preds = %bb2965 ret void66}67