| 1 |
// RUN: mlir-opt %s -affine-super-vectorize="virtual-vector-size=128" -split-input-file | FileCheck %s |
| 2 |
|
| 3 |
// Specific tests to check vectorization of uniform/divergent values. |
| 4 |
|
| 5 |
// CHECK-LABEL: @uniform_arg |
| 6 |
// CHECK-SAME: %[[in:.*]]: memref<512xf32>, |
| 7 |
// CHECK-SAME: %[[uniform:.*]]: f32 |
| 8 |
func.func @uniform_arg(%in : memref<512xf32>, %uniform : f32) { |
| 9 |
affine.for %i = 0 to 512 { |
| 10 |
%ld = affine.load %in[%i] : memref<512xf32> |
| 11 |
%add = arith.addf %ld, %uniform : f32 |
| 12 |
} |
| 13 |
return |
| 14 |
} |
| 15 |
|
| 16 |
// CHECK-NEXT: %[[bcast:.*]] = vector.broadcast %[[uniform]] : f32 to vector<128xf32> |
| 17 |
// CHECK-NEXT: affine.for |
| 18 |
// CHECK: arith.addf %{{.*}}, %[[bcast]] : vector<128xf32> |
| 19 |
|
| 20 |
// ----- |
| 21 |
|
| 22 |
// CHECK-LABEL: @multi_use_uniform_arg |
| 23 |
// CHECK-SAME: %[[in:.*]]: memref<512xf32> |
| 24 |
// CHECK-SAME: %[[uniform:.*]]: f32 |
| 25 |
func.func @multi_use_uniform_arg(%in : memref<512xf32>, %uniform : f32) { |
| 26 |
affine.for %i = 0 to 512 { |
| 27 |
%ld = affine.load %in[%i] : memref<512xf32> |
| 28 |
%user0 = arith.addf %ld, %uniform : f32 |
| 29 |
%user1 = arith.addf %ld, %uniform : f32 |
| 30 |
} |
| 31 |
return |
| 32 |
} |
| 33 |
|
| 34 |
// CHECK-NEXT: %[[bcast:.*]] = vector.broadcast %[[uniform]] : f32 to vector<128xf32> |
| 35 |
// CHECK-NOT: vector.broadcast |
| 36 |
// CHECK-NEXT: affine.for |
| 37 |
// CHECK: arith.addf %{{.*}}, %[[bcast]] : vector<128xf32> |
| 38 |
// CHECK: arith.addf %{{.*}}, %[[bcast]] : vector<128xf32> |
| 39 |
|
| 40 |
// ----- |
| 41 |
|
| 42 |
// CHECK-LABEL: @uniform_load |
| 43 |
func.func @uniform_load(%A : memref<?x?xf32>, %C : memref<?x?xf32>) { |
| 44 |
%c0 = arith.constant 0 : index |
| 45 |
%N = memref.dim %A, %c0 : memref<?x?xf32> |
| 46 |
affine.for %i = 0 to %N { |
| 47 |
%uniform_ld = affine.load %A[%i, %i] : memref<?x?xf32> |
| 48 |
affine.for %j = 0 to %N { |
| 49 |
%b = affine.load %A[%i, %j] : memref<?x?xf32> |
| 50 |
%c = arith.addf %uniform_ld, %b : f32 |
| 51 |
} |
| 52 |
} |
| 53 |
return |
| 54 |
} |
| 55 |
|
| 56 |
// CHECK: affine.for |
| 57 |
// CHECK-NEXT: %[[uniform_ld:.*]] = affine.load %{{.*}}[%{{.*}}, %{{.*}}] : memref<?x?xf32> |
| 58 |
// CHECK-NEXT: %[[bcast:.*]] = vector.broadcast %[[uniform_ld]] : f32 to vector<128xf32> |
| 59 |
// CHECK-NEXT: affine.for |
| 60 |
// CHECK: arith.addf %[[bcast]], %{{.*}} : vector<128xf32> |