Debian Codegraph @

mlir/test/Dialect/Linalg/vectorization/insert-slice-with-patterns.mlir

raw
1 // RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s
2
3 ///----------------------------------------------------------------------------------------
4 /// Tests for tensor.insert_slice
5 ///----------------------------------------------------------------------------------------
6
7 // The pad value for xfer-read is neither needed nor available - use the default (0.0).
8
9 // CHECK-LABEL: func @insert_static_slice_default_pad
10 // CHECK-SAME: %[[ARG_0:.*]]: tensor<1x2x3xf32>,
11 // CHECK-SAME: %[[ARG_1:.*]]: tensor<9x8x7x1x2x3xf32>) -> tensor<9x8x7x1x2x3xf32> {
12 // CHECK: %[[PAD:.*]] = arith.constant 0.000000e+00 : f32
13 // CHECK: %[[C0:.*]] = arith.constant 0 : index
14 // CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG_0]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : tensor<1x2x3xf32>, vector<1x2x3xf32>
15 // CHECK: %[[WRITE:.*]] = vector.transfer_write %[[READ]], %[[ARG_1]]{{\[}}%[[C0]], %[[C0]], %[[C0]], %[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<1x2x3xf32>, tensor<9x8x7x1x2x3xf32>
16 // CHECK: return %[[WRITE]] : tensor<9x8x7x1x2x3xf32>
17 func.func @insert_static_slice_default_pad(%arg0: tensor<1x2x3xf32>, %arg1: tensor<9x8x7x1x2x3xf32>) -> tensor<9x8x7x1x2x3xf32> {
18 %res = tensor.insert_slice %arg0 into %arg1[0, 0, 0, 0, 0, 0] [1, 1, 1, 1, 2, 3][1, 1, 1, 1, 1, 1] : tensor<1x2x3xf32> into tensor<9x8x7x1x2x3xf32>
19 return %res : tensor<9x8x7x1x2x3xf32>
20 }
21
22 module attributes {transform.with_named_sequence} {
23 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
24 %0 = transform.structured.match ops{["tensor.insert_slice"]} in %arg1 : (!transform.any_op) -> !transform.any_op
25 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
26 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { vectorize_padding } : (!transform.any_op) -> !transform.any_op
27 transform.yield
28 }
29 }
30
31 // -----
32
33 // Same as above, but there's a pad value available that should be used instead of the default value.
34
35 // CHECK-LABEL: func.func @insert_static_slice_non_zero_pad
36 // CHECK-SAME: %[[ARG_0:.*]]: tensor<1x2x3xf32>,
37 // CHECK-SAME: %[[PAD:.*]]: f32) -> tensor<9x8x7x1x2x3xf32> {
38 // CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<9x8x7x1x2x3xf32>
39 // CHECK: %[[BC:.*]] = vector.broadcast %[[PAD]] : f32 to vector<9x8x7x1x2x3xf32>
40 // CHECK: %[[WRITE:.*]] = vector.transfer_write %[[BC]], %[[EMPTY]]{{.*}} {in_bounds = [true, true, true, true, true, true]} : vector<9x8x7x1x2x3xf32>, tensor<9x8x7x1x2x3xf32>
41 // CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG_0]]{{.*}}, %[[PAD]] {in_bounds = [true, true, true]} : tensor<1x2x3xf32>, vector<1x2x3xf32>
42 // CHECK: %[[RES:.*]] = vector.transfer_write %[[READ]], %[[WRITE]]{{.*}} {in_bounds = [true, true, true]} : vector<1x2x3xf32>, tensor<9x8x7x1x2x3xf32>
43 // CHECK: return %[[RES]] : tensor<9x8x7x1x2x3xf32>
44 func.func @insert_static_slice_non_zero_pad(%arg0: tensor<1x2x3xf32>, %pad : f32) -> tensor<9x8x7x1x2x3xf32> {
45 %init = tensor.empty() : tensor<9x8x7x1x2x3xf32>
46 %fill = linalg.fill ins(%pad : f32) outs(%init : tensor<9x8x7x1x2x3xf32>) -> tensor<9x8x7x1x2x3xf32>
47 %res = tensor.insert_slice %arg0 into %fill[0, 0, 0, 0, 0, 0] [1, 1, 1, 1, 2, 3][1, 1, 1, 1, 1, 1] : tensor<1x2x3xf32> into tensor<9x8x7x1x2x3xf32>
48 return %res : tensor<9x8x7x1x2x3xf32>
49 }
50
51 module attributes {transform.with_named_sequence} {
52 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
53 %0 = transform.structured.match ops{["tensor.insert_slice"]} in %arg1 : (!transform.any_op) -> !transform.any_op
54 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
55 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
56 transform.yield
57 }
58 }
59
60 // -----
61
62 // Same as above, but the source type has is dynamically shaped. This means
63 // that the pad value is now required and the vector dim corresponding to the
64 // dynamic shape has to be inferred from the shape of the destination tensor.
65
66 // CHECK-LABEL: func.func @insert_dynamic_slice_non_zero_pad(
67 // CHECK-SAME: %[[ARG_0:.*]]: tensor<1x?x3xf32>,
68 // CHECK-SAME: %[[PAD:.*]]: f32,
69 // CHECK-SAME: %[[SIZE:.*]]: index) -> tensor<9x8x7x1x2x3xf32> {
70 // CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<9x8x7x1x2x3xf32>
71 // CHECK: %[[BC:.*]] = vector.broadcast %[[PAD]] : f32 to vector<9x8x7x1x2x3xf32>
72 // CHECK: %[[WRITE:.*]] = vector.transfer_write %[[BC]], %[[EMPTY]]{{.*}} {in_bounds = [true, true, true, true, true, true]} : vector<9x8x7x1x2x3xf32>, tensor<9x8x7x1x2x3xf32>
73 // CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG_0]]{{.*}}, %[[PAD]] {in_bounds = [true, false, true]} : tensor<1x?x3xf32>, vector<1x2x3xf32>
74 // CHECK: %[[RES:.*]] = vector.transfer_write %[[READ]], %[[WRITE]]{{.*}} {in_bounds = [true, true, true]} : vector<1x2x3xf32>, tensor<9x8x7x1x2x3xf32>
75 // CHECK: return %[[RES]] : tensor<9x8x7x1x2x3xf32>
76 func.func @insert_dynamic_slice_non_zero_pad(%arg0: tensor<1x?x3xf32>, %pad : f32, %size: index) -> tensor<9x8x7x1x2x3xf32> {
77 %init = tensor.empty() : tensor<9x8x7x1x2x3xf32>
78 %fill = linalg.fill ins(%pad : f32) outs(%init : tensor<9x8x7x1x2x3xf32>) -> tensor<9x8x7x1x2x3xf32>
79 %res = tensor.insert_slice %arg0 into %fill[0, 0, 0, 0, 0, 0] [1, 1, 1, 1, %size, 3][1, 1, 1, 1, 1, 1] : tensor<1x?x3xf32> into tensor<9x8x7x1x2x3xf32>
80 return %res : tensor<9x8x7x1x2x3xf32>
81 }
82
83 module attributes {transform.with_named_sequence} {
84 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
85 %0 = transform.structured.match ops{["tensor.insert_slice"]} in %arg1 : (!transform.any_op) -> !transform.any_op
86 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
87 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
88 transform.yield
89 }
90 }