| 1 |
// RUN: mlir-opt %s -affine-super-vectorize="virtual-vector-size=32,256 test-fastest-varying=0,1" | FileCheck %s |
| 2 |
|
| 3 |
// Permutation maps used in vectorization. |
| 4 |
// CHECK-DAG: #[[map_proj_d0d1d2_d2d1:map[0-9]*]] = affine_map<(d0, d1, d2) -> (d2, d1)> |
| 5 |
|
| 6 |
func.func @vec2d(%A : memref<?x?x?xf32>) { |
| 7 |
%c0 = arith.constant 0 : index |
| 8 |
%c1 = arith.constant 1 : index |
| 9 |
%c2 = arith.constant 2 : index |
| 10 |
%M = memref.dim %A, %c0 : memref<?x?x?xf32> |
| 11 |
%N = memref.dim %A, %c1 : memref<?x?x?xf32> |
| 12 |
%P = memref.dim %A, %c2 : memref<?x?x?xf32> |
| 13 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 14 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 15 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 16 |
// For the case: --test-fastest-varying=0 --test-fastest-varying=1 no |
| 17 |
// vectorization happens because of loop nesting order. |
| 18 |
affine.for %i0 = 0 to %M { |
| 19 |
affine.for %i1 = 0 to %N { |
| 20 |
affine.for %i2 = 0 to %P { |
| 21 |
%a2 = affine.load %A[%i0, %i1, %i2] : memref<?x?x?xf32> |
| 22 |
} |
| 23 |
} |
| 24 |
} |
| 25 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} step 32 |
| 26 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} { |
| 27 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} step 256 |
| 28 |
// CHECK: {{.*}} = vector.transfer_read %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}], %{{.*}} {permutation_map = #[[map_proj_d0d1d2_d2d1]]} : memref<?x?x?xf32>, vector<32x256xf32> |
| 29 |
affine.for %i3 = 0 to %M { |
| 30 |
affine.for %i4 = 0 to %N { |
| 31 |
affine.for %i5 = 0 to %P { |
| 32 |
%a5 = affine.load %A[%i4, %i5, %i3] : memref<?x?x?xf32> |
| 33 |
} |
| 34 |
} |
| 35 |
} |
| 36 |
return |
| 37 |
} |
| 38 |
|
| 39 |
func.func @vec2d_imperfectly_nested(%A : memref<?x?x?xf32>) { |
| 40 |
%c0 = arith.constant 0 : index |
| 41 |
%c1 = arith.constant 1 : index |
| 42 |
%c2 = arith.constant 2 : index |
| 43 |
%0 = memref.dim %A, %c0 : memref<?x?x?xf32> |
| 44 |
%1 = memref.dim %A, %c1 : memref<?x?x?xf32> |
| 45 |
%2 = memref.dim %A, %c2 : memref<?x?x?xf32> |
| 46 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} step 32 { |
| 47 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} step 256 { |
| 48 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} { |
| 49 |
// CHECK: %{{.*}} = vector.transfer_read %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}], %{{.*}} {permutation_map = #[[map_proj_d0d1d2_d2d1]]} : memref<?x?x?xf32>, vector<32x256xf32> |
| 50 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} { |
| 51 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} step 256 { |
| 52 |
// CHECK: %{{.*}} = vector.transfer_read %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}], %{{.*}} {permutation_map = #[[map_proj_d0d1d2_d2d1]]} : memref<?x?x?xf32>, vector<32x256xf32> |
| 53 |
// CHECK: affine.for %{{.*}} = 0 to %{{.*}} step 256 { |
| 54 |
// CHECK: %{{.*}} = vector.transfer_read %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}], %{{.*}} {permutation_map = #[[map_proj_d0d1d2_d2d1]]} : memref<?x?x?xf32>, vector<32x256xf32> |
| 55 |
affine.for %i0 = 0 to %0 { |
| 56 |
affine.for %i1 = 0 to %1 { |
| 57 |
affine.for %i2 = 0 to %2 { |
| 58 |
%a2 = affine.load %A[%i2, %i1, %i0] : memref<?x?x?xf32> |
| 59 |
} |
| 60 |
} |
| 61 |
affine.for %i3 = 0 to %1 { |
| 62 |
affine.for %i4 = 0 to %2 { |
| 63 |
%a4 = affine.load %A[%i3, %i4, %i0] : memref<?x?x?xf32> |
| 64 |
} |
| 65 |
affine.for %i5 = 0 to %2 { |
| 66 |
%a5 = affine.load %A[%i3, %i5, %i0] : memref<?x?x?xf32> |
| 67 |
} |
| 68 |
} |
| 69 |
} |
| 70 |
return |
| 71 |
} |
| 72 |
|