150
|
1 ; RUN: opt %loadPolly -polly-opt-isl -polly-pattern-matching-based-opts=true \
|
|
2 ; RUN: -polly-target-throughput-vector-fma=1 \
|
|
3 ; RUN: -polly-target-latency-vector-fma=8 \
|
|
4 ; RUN: -polly-target-1st-cache-level-associativity=8 \
|
|
5 ; RUN: -polly-target-2nd-cache-level-associativity=8 \
|
|
6 ; RUN: -polly-target-1st-cache-level-size=32768 \
|
|
7 ; RUN: -polly-target-2nd-cache-level-size=262144 \
|
|
8 ; RUN: -polly-optimized-scops \
|
|
9 ; RUN: -polly-target-vector-register-bitwidth=256 \
|
223
|
10 ; RUN: -disable-output < %s
|
150
|
11 ;
|
|
12 ; /* C := alpha*A*B + beta*C */
|
|
13 ; for (i = 0; i < _PB_NI; i++)
|
|
14 ; for (j = 0; j < _PB_NJ; j++)
|
|
15 ; {
|
|
16 ; C[i][j] *= beta;
|
|
17 ; for (k = 0; k < _PB_NK; ++k)
|
|
18 ; C[i][j] += alpha * A[i][k] * B[k][j];
|
|
19 ; }
|
|
20 ;
|
|
21 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
|
|
22 target triple = "x86_64-unknown-unknown"
|
|
23
|
252
|
24 define internal void @kernel_gemm(i32 %arg, i32 %arg1, i32 %arg2, double %arg3, double %arg4, ptr %arg5, ptr %arg6, ptr %arg7) #0 {
|
150
|
25 bb:
|
|
26 br label %bb8
|
|
27
|
|
28 bb8: ; preds = %bb29, %bb
|
|
29 %tmp = phi i64 [ 0, %bb ], [ %tmp30, %bb29 ]
|
|
30 br label %bb9
|
|
31
|
|
32 bb9: ; preds = %bb26, %bb8
|
|
33 %tmp10 = phi i64 [ 0, %bb8 ], [ %tmp27, %bb26 ]
|
252
|
34 %tmp11 = getelementptr inbounds [1056 x double], ptr %arg5, i64 %tmp, i64 %tmp10
|
|
35 %tmp12 = load double, ptr %tmp11, align 8
|
150
|
36 %tmp13 = fmul double %tmp12, %arg4
|
252
|
37 store double %tmp13, ptr %tmp11, align 8
|
150
|
38 br label %Copy_0
|
|
39
|
|
40 Copy_0: ; preds = %Copy_0, %bb9
|
|
41 %tmp15 = phi i64 [ 0, %bb9 ], [ %tmp24, %Copy_0 ]
|
252
|
42 %tmp16 = getelementptr inbounds [1024 x double], ptr %arg6, i64 %tmp, i64 %tmp15
|
|
43 %tmp17 = load double, ptr %tmp16, align 8
|
150
|
44 %tmp18 = fmul double %tmp17, %arg3
|
252
|
45 %tmp19 = getelementptr inbounds [1056 x double], ptr %arg7, i64 %tmp15, i64 %tmp10
|
|
46 %tmp20 = load double, ptr %tmp19, align 8
|
150
|
47 %tmp21 = fmul double %tmp18, %tmp20
|
252
|
48 %tmp22 = load double, ptr %tmp11, align 8
|
150
|
49 %tmp23 = fadd double %tmp22, %tmp21
|
252
|
50 store double %tmp23, ptr %tmp11, align 8
|
150
|
51 %tmp24 = add nuw nsw i64 %tmp15, 1
|
|
52 %tmp25 = icmp ne i64 %tmp24, 1024
|
|
53 br i1 %tmp25, label %Copy_0, label %bb26
|
|
54
|
|
55 bb26: ; preds = %Copy_0
|
|
56 %tmp27 = add nuw nsw i64 %tmp10, 1
|
|
57 %tmp28 = icmp ne i64 %tmp27, 1056
|
|
58 br i1 %tmp28, label %bb9, label %bb29
|
|
59
|
|
60 bb29: ; preds = %bb26
|
|
61 %tmp30 = add nuw nsw i64 %tmp, 1
|
|
62 %tmp31 = icmp ne i64 %tmp30, 1056
|
|
63 br i1 %tmp31, label %bb8, label %bb32
|
|
64
|
|
65 bb32: ; preds = %bb29
|
|
66 ret void
|
|
67 }
|
|
68
|
|
69 attributes #0 = { nounwind uwtable "target-cpu"="x86-64" "target-features"="+aes,+avx,+cmov,+cx16,+fxsr,+mmx,+pclmul,+popcnt,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave,+xsaveopt" }
|