| 1 |
// RUN: mlir-opt %s -affine-super-vectorize="virtual-vector-size=4,8" | FileCheck %s -check-prefix=VECT |
| 2 |
// RUN: mlir-opt %s -affine-super-vectorize="virtual-vector-size=32,256 test-fastest-varying=1,0" | FileCheck %s |
| 3 |
|
| 4 |
// Permutation maps used in vectorization. |
| 5 |
// CHECK-DAG: #[[$map_id1:map[0-9]*]] = affine_map<(d0) -> (d0)> |
| 6 |
// CHECK-DAG: #[[$map_proj_d0d1_zerod1:map[0-9]*]] = affine_map<(d0, d1) -> (0, d1)> |
| 7 |
// CHECK-DAG: #[[$map_proj_d0d1_d0zero:map[0-9]*]] = affine_map<(d0, d1) -> (d0, 0)> |
| 8 |
// VECT-DAG: #[[$map_id1:map[0-9]*]] = affine_map<(d0) -> (d0)> |
| 9 |
// VECT-DAG: #[[$map_proj_d0d1_zerod1:map[0-9]*]] = affine_map<(d0, d1) -> (0, d1)> |
| 10 |
// VECT-DAG: #[[$map_proj_d0d1_d0zero:map[0-9]*]] = affine_map<(d0, d1) -> (d0, 0)> |
| 11 |
|
| 12 |
func.func @vec2d(%A : memref<?x?x?xf32>) { |
| 13 |
%c0 = arith.constant 0 : index |
| 14 |
%c1 = arith.constant 1 : index |
| 15 |
%c2 = arith.constant 2 : index |
| 16 |
%M = memref.dim %A, %c0 : memref<?x?x?xf32> |
| 17 |
%N = memref.dim %A, %c1 : memref<?x?x?xf32> |
| 18 |
%P = memref.dim %A, %c2 : memref<?x?x?xf32> |
| 19 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 20 |
// CHECK: for {{.*}} = 0 to %{{.*}} step 32 |
| 21 |
// CHECK: for {{.*}} = 0 to %{{.*}} step 256 |
| 22 |
// Example: |
| 23 |
// affine.for %{{.*}} = 0 to %{{.*}} { |
| 24 |
// affine.for %{{.*}} = 0 to %{{.*}} step 32 { |
| 25 |
// affine.for %{{.*}} = 0 to %{{.*}} step 256 { |
| 26 |
// %{{.*}} = "vector.transfer_read"(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) : (memref<?x?x?xf32>, index, index, index) -> vector<32x256xf32> |
| 27 |
affine.for %i0 = 0 to %M { |
| 28 |
affine.for %i1 = 0 to %N { |
| 29 |
affine.for %i2 = 0 to %P { |
| 30 |
%a2 = affine.load %A[%i0, %i1, %i2] : memref<?x?x?xf32> |
| 31 |
} |
| 32 |
} |
| 33 |
} |
| 34 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 35 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 36 |
// CHECK: for {{.*}} = 0 to %{{.*}} { |
| 37 |
// For the case: --test-fastest-varying=1 --test-fastest-varying=0 no |
| 38 |
// vectorization happens because of loop nesting order . |
| 39 |
affine.for %i3 = 0 to %M { |
| 40 |
affine.for %i4 = 0 to %N { |
| 41 |
affine.for %i5 = 0 to %P { |
| 42 |
%a5 = affine.load %A[%i4, %i5, %i3] : memref<?x?x?xf32> |
| 43 |
} |
| 44 |
} |
| 45 |
} |
| 46 |
return |
| 47 |
} |
| 48 |
|
| 49 |
func.func @vector_add_2d(%M : index, %N : index) -> f32 { |
| 50 |
%A = memref.alloc (%M, %N) : memref<?x?xf32, 0> |
| 51 |
%B = memref.alloc (%M, %N) : memref<?x?xf32, 0> |
| 52 |
%C = memref.alloc (%M, %N) : memref<?x?xf32, 0> |
| 53 |
%f1 = arith.constant 1.0 : f32 |
| 54 |
%f2 = arith.constant 2.0 : f32 |
| 55 |
affine.for %i0 = 0 to %M { |
| 56 |
affine.for %i1 = 0 to %N { |
| 57 |
// CHECK: [[C1:%.*]] = arith.constant dense<1.000000e+00> : vector<32x256xf32> |
| 58 |
// CHECK: vector.transfer_write [[C1]], {{.*}} : vector<32x256xf32>, memref<?x?xf32> |
| 59 |
// non-scoped %f1 |
| 60 |
affine.store %f1, %A[%i0, %i1] : memref<?x?xf32, 0> |
| 61 |
} |
| 62 |
} |
| 63 |
affine.for %i2 = 0 to %M { |
| 64 |
affine.for %i3 = 0 to %N { |
| 65 |
// CHECK: [[C3:%.*]] = arith.constant dense<2.000000e+00> : vector<32x256xf32> |
| 66 |
// CHECK: vector.transfer_write [[C3]], {{.*}} : vector<32x256xf32>, memref<?x?xf32> |
| 67 |
// non-scoped %f2 |
| 68 |
affine.store %f2, %B[%i2, %i3] : memref<?x?xf32, 0> |
| 69 |
} |
| 70 |
} |
| 71 |
affine.for %i4 = 0 to %M { |
| 72 |
affine.for %i5 = 0 to %N { |
| 73 |
// CHECK: [[SPLAT2:%.*]] = arith.constant dense<2.000000e+00> : vector<32x256xf32> |
| 74 |
// CHECK: [[SPLAT1:%.*]] = arith.constant dense<1.000000e+00> : vector<32x256xf32> |
| 75 |
// CHECK: [[A5:%.*]] = vector.transfer_read %{{.*}}[{{.*}}], %{{.*}} : memref<?x?xf32>, vector<32x256xf32> |
| 76 |
// CHECK: [[B5:%.*]] = vector.transfer_read %{{.*}}[{{.*}}], %{{.*}} : memref<?x?xf32>, vector<32x256xf32> |
| 77 |
// CHECK: [[S5:%.*]] = arith.addf [[A5]], [[B5]] : vector<32x256xf32> |
| 78 |
// CHECK: [[S6:%.*]] = arith.addf [[S5]], [[SPLAT1]] : vector<32x256xf32> |
| 79 |
// CHECK: [[S7:%.*]] = arith.addf [[S5]], [[SPLAT2]] : vector<32x256xf32> |
| 80 |
// CHECK: [[S8:%.*]] = arith.addf [[S7]], [[S6]] : vector<32x256xf32> |
| 81 |
// CHECK: vector.transfer_write [[S8]], {{.*}} : vector<32x256xf32>, memref<?x?xf32> |
| 82 |
// |
| 83 |
%a5 = affine.load %A[%i4, %i5] : memref<?x?xf32, 0> |
| 84 |
%b5 = affine.load %B[%i4, %i5] : memref<?x?xf32, 0> |
| 85 |
%s5 = arith.addf %a5, %b5 : f32 |
| 86 |
// non-scoped %f1 |
| 87 |
%s6 = arith.addf %s5, %f1 : f32 |
| 88 |
// non-scoped %f2 |
| 89 |
%s7 = arith.addf %s5, %f2 : f32 |
| 90 |
// diamond dependency. |
| 91 |
%s8 = arith.addf %s7, %s6 : f32 |
| 92 |
affine.store %s8, %C[%i4, %i5] : memref<?x?xf32, 0> |
| 93 |
} |
| 94 |
} |
| 95 |
%c7 = arith.constant 7 : index |
| 96 |
%c42 = arith.constant 42 : index |
| 97 |
%res = affine.load %C[%c7, %c42] : memref<?x?xf32, 0> |
| 98 |
return %res : f32 |
| 99 |
} |
| 100 |
|
| 101 |
// VECT-LABEL: func @vectorize_matmul |
| 102 |
func.func @vectorize_matmul(%arg0: memref<?x?xf32>, %arg1: memref<?x?xf32>, %arg2: memref<?x?xf32>) { |
| 103 |
%c0 = arith.constant 0 : index |
| 104 |
%c1 = arith.constant 1 : index |
| 105 |
%M = memref.dim %arg0, %c0 : memref<?x?xf32> |
| 106 |
%K = memref.dim %arg0, %c1 : memref<?x?xf32> |
| 107 |
%N = memref.dim %arg2, %c1 : memref<?x?xf32> |
| 108 |
// VECT: %[[C0:.*]] = arith.constant 0 : index |
| 109 |
// VECT-NEXT: %[[C1:.*]] = arith.constant 1 : index |
| 110 |
// VECT-NEXT: %[[M:.*]] = memref.dim %{{.*}}, %[[C0]] : memref<?x?xf32> |
| 111 |
// VECT-NEXT: %[[K:.*]] = memref.dim %{{.*}}, %[[C1]] : memref<?x?xf32> |
| 112 |
// VECT-NEXT: %[[N:.*]] = memref.dim %{{.*}}, %[[C1]] : memref<?x?xf32> |
| 113 |
// VECT: {{.*}} #[[$map_id1]](%[[M]]) step 4 { |
| 114 |
// VECT-NEXT: {{.*}} #[[$map_id1]](%[[N]]) step 8 { |
| 115 |
// VECT: %[[VC0:.*]] = arith.constant dense<0.000000e+00> : vector<4x8xf32> |
| 116 |
// VECT-NEXT: vector.transfer_write %[[VC0]], %{{.*}}[%{{.*}}, %{{.*}}] : vector<4x8xf32>, memref<?x?xf32> |
| 117 |
affine.for %i0 = affine_map<(d0) -> (d0)>(%c0) to affine_map<(d0) -> (d0)>(%M) { |
| 118 |
affine.for %i1 = affine_map<(d0) -> (d0)>(%c0) to affine_map<(d0) -> (d0)>(%N) { |
| 119 |
%cst = arith.constant 0.000000e+00 : f32 |
| 120 |
affine.store %cst, %arg2[%i0, %i1] : memref<?x?xf32> |
| 121 |
} |
| 122 |
} |
| 123 |
// VECT: affine.for %[[I2:.*]] = #[[$map_id1]](%[[C0]]) to #[[$map_id1]](%[[M]]) step 4 { |
| 124 |
// VECT-NEXT: affine.for %[[I3:.*]] = #[[$map_id1]](%[[C0]]) to #[[$map_id1]](%[[N]]) step 8 { |
| 125 |
// VECT-NEXT: affine.for %[[I4:.*]] = #[[$map_id1]](%[[C0]]) to #[[$map_id1]](%[[K]]) { |
| 126 |
// VECT: %[[A:.*]] = vector.transfer_read %{{.*}}[%[[I4]], %[[I3]]], %{{.*}} {permutation_map = #[[$map_proj_d0d1_zerod1]]} : memref<?x?xf32>, vector<4x8xf32> |
| 127 |
// VECT: %[[B:.*]] = vector.transfer_read %{{.*}}[%[[I2]], %[[I4]]], %{{.*}} {permutation_map = #[[$map_proj_d0d1_d0zero]]} : memref<?x?xf32>, vector<4x8xf32> |
| 128 |
// VECT-NEXT: %[[C:.*]] = arith.mulf %[[B]], %[[A]] : vector<4x8xf32> |
| 129 |
// VECT: %[[D:.*]] = vector.transfer_read %{{.*}}[%[[I2]], %[[I3]]], %{{.*}} : memref<?x?xf32>, vector<4x8xf32> |
| 130 |
// VECT-NEXT: %[[E:.*]] = arith.addf %[[D]], %[[C]] : vector<4x8xf32> |
| 131 |
// VECT: vector.transfer_write %[[E]], %{{.*}}[%[[I2]], %[[I3]]] : vector<4x8xf32>, memref<?x?xf32> |
| 132 |
affine.for %i2 = affine_map<(d0) -> (d0)>(%c0) to affine_map<(d0) -> (d0)>(%M) { |
| 133 |
affine.for %i3 = affine_map<(d0) -> (d0)>(%c0) to affine_map<(d0) -> (d0)>(%N) { |
| 134 |
affine.for %i4 = affine_map<(d0) -> (d0)>(%c0) to affine_map<(d0) -> (d0)>(%K) { |
| 135 |
%6 = affine.load %arg1[%i4, %i3] : memref<?x?xf32> |
| 136 |
%7 = affine.load %arg0[%i2, %i4] : memref<?x?xf32> |
| 137 |
%8 = arith.mulf %7, %6 : f32 |
| 138 |
%9 = affine.load %arg2[%i2, %i3] : memref<?x?xf32> |
| 139 |
%10 = arith.addf %9, %8 : f32 |
| 140 |
affine.store %10, %arg2[%i2, %i3] : memref<?x?xf32> |
| 141 |
} |
| 142 |
} |
| 143 |
} |
| 144 |
return |
| 145 |
} |