1// RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s 2 3// CHECK-LABEL: func.func @fill( 4// CHECK-SAME: %[[ARG0:.*]]: f32, 5// CHECK-SAME: %[[ARG1:.*]]: memref<32x7xf32> 6// CHECK-NEXT: %[[FLATTENED:.*]] = memref.collapse_shape %[[ARG1]] {{\[}}[0, 1]] 7// CHECK-NEXT: linalg.fill ins(%[[ARG0]] : f32) outs(%[[FLATTENED]] : memref<224xf32>) 8func.func @fill(%cst: f32, %arg: memref<32x7xf32>) { 9 linalg.fill ins(%cst: f32) outs(%arg: memref<32x7xf32>) 10 return 11} 12 13module attributes {transform.with_named_sequence} { 14 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { 15 %0 = transform.structured.match interface{LinalgOp} in %arg1 : (!transform.any_op) -> !transform.any_op 16 %flattened = transform.structured.flatten_elementwise %0 17 : (!transform.any_op) -> !transform.any_op 18 transform.yield 19 } 20} 21 22// ----- 23 24// CHECK-LABEL: func.func @fill_tensor( 25// CHECK-SAME: %[[ARG0:.*]]: f32, 26// CHECK-SAME: %[[ARG1:.*]]: tensor<32x7xf32> 27// CHECK-NEXT: %[[FLATTENED:.*]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0, 1]] 28// CHECK-NEXT: %[[FLATTENED_RESULT:.*]] = linalg.fill ins(%[[ARG0]] : f32) outs(%[[FLATTENED]] : tensor<224xf32>) 29// CHECK-NEXT: %[[RESULT:.*]] = tensor.expand_shape %[[FLATTENED_RESULT]] {{\[}}[0, 1]] output_shape [32, 7] : tensor<224xf32> into tensor<32x7xf32> 30func.func @fill_tensor(%cst: f32, %arg: tensor<32x7xf32>) -> tensor<32x7xf32> { 31 %0 = linalg.fill ins(%cst: f32) outs(%arg: tensor<32x7xf32>) -> tensor<32x7xf32> 32 return %0 : tensor<32x7xf32> 33} 34 35module attributes {transform.with_named_sequence} { 36 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { 37 %0 = transform.structured.match interface{LinalgOp} in %arg1 : (!transform.any_op) -> !transform.any_op 38 %flattened = transform.structured.flatten_elementwise %0 39 : (!transform.any_op) -> !transform.any_op 40 transform.yield 41 } 42} 43 44// ----- 45 46// CHECK-LABEL: func.func @map( 47// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<32x7xf32> 48// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<32x7xf32> 49// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<32x7xf32> 50// CHECK-NEXT: %[[FLATTENED_0:.*]] = memref.collapse_shape %[[ARG0]] {{\[}}[0, 1]] 51// CHECK-NEXT: %[[FLATTENED_1:.*]] = memref.collapse_shape %[[ARG1]] {{\[}}[0, 1]] 52// CHECK-NEXT: %[[FLATTENED_2:.*]] = memref.collapse_shape %[[ARG2]] {{\[}}[0, 1]] 53// CHECK-NEXT: linalg.map { arith.addf } ins(%[[FLATTENED_0]], %[[FLATTENED_1]] : memref<224xf32>, memref<224xf32>) outs(%[[FLATTENED_2]] : memref<224xf32>) 54func.func @map(%arg0: memref<32x7xf32>, %arg1: memref<32x7xf32>, %arg2: memref<32x7xf32>) { 55 linalg.map {arith.addf} ins(%arg0, %arg1: memref<32x7xf32>, memref<32x7xf32>) outs(%arg2: memref<32x7xf32>) 56 return 57} 58 59module attributes {transform.with_named_sequence} { 60 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { 61 %0 = transform.structured.match interface{LinalgOp} in %arg1 : (!transform.any_op) -> !transform.any_op 62 %flattened = transform.structured.flatten_elementwise %0 63 : (!transform.any_op) -> !transform.any_op 64 transform.yield 65 } 66} 67 68// ----- 69 70// CHECK-LABEL: func.func @map_already_flat( 71// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<32xf32> 72// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<32xf32> 73// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<32xf32> 74// CHECK-NEXT: linalg.map { arith.addf } ins(%[[ARG0]], %[[ARG1]] : memref<32xf32>, memref<32xf32>) outs(%[[ARG2]] : memref<32xf32>) 75func.func @map_already_flat(%arg0: memref<32xf32>, %arg1: memref<32xf32>, %arg2: memref<32xf32>) { 76 linalg.map {arith.addf} ins(%arg0, %arg1: memref<32xf32>, memref<32xf32>) outs(%arg2: memref<32xf32>) 77 return 78} 79 80module attributes {transform.with_named_sequence} { 81 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { 82 %0 = transform.structured.match interface{LinalgOp} in %arg1 : (!transform.any_op) -> !transform.any_op 83 %flattened = transform.structured.flatten_elementwise %0 84 : (!transform.any_op) -> !transform.any_op 85 transform.yield 86 } 87} 88 89// ----- 90 91// CHECK: #[[$MAP0:.*]] = affine_map<(d0) -> (d0)> 92// CHECK-LABEL: func.func @generic 93// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<32x7xf32> 94// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<32x7xf32> 95// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<32x7xf32> 96// CHECK-NEXT: %[[FLATTENED_0:.*]] = memref.collapse_shape %[[ARG0]] {{\[}}[0, 1]] 97// CHECK-NEXT: %[[FLATTENED_1:.*]] = memref.collapse_shape %[[ARG1]] {{\[}}[0, 1]] 98// CHECK-NEXT: %[[FLATTENED_2:.*]] = memref.collapse_shape %[[ARG2]] {{\[}}[0, 1]] 99// CHECK-NEXT: linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel"]} ins(%[[FLATTENED_0]], %[[FLATTENED_1]] : memref<224xf32>, memref<224xf32>) outs(%[[FLATTENED_2]] : memref<224xf32>) 100// CHECK-NEXT: ^bb0(%[[A:.*]]: f32, %[[B:.*]]: f32, %[[C:.*]]: f32) 101// CHECK-NEXT: %[[SUM:.*]] = arith.addf %[[A]], %[[B]] 102// CHECK-NEXT: linalg.yield %[[SUM]] 103#map = affine_map<(d0, d1) -> (d0, d1)> 104func.func @generic( %arg0: memref<32x7xf32>, %arg1: memref<32x7xf32>, %arg2: memref<32x7xf32>) { 105 linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = ["parallel", "parallel"]} ins(%arg0, %arg1: memref<32x7xf32>, memref<32x7xf32>) outs(%arg2: memref<32x7xf32>) { 106 ^bb0(%a: f32, %b: f32, %c: f32): 107 %0 = arith.addf %a, %b : f32 108 linalg.yield %0 : f32 109 } 110 return 111} 112 113module attributes {transform.with_named_sequence} { 114 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { 115 %0 = transform.structured.match interface{LinalgOp} in %arg1 : (!transform.any_op) -> !transform.any_op 116 %flattened = transform.structured.flatten_elementwise %0 117 : (!transform.any_op) -> !transform.any_op 118 transform.yield 119 } 120} 121