1// RUN: mlir-opt %s -convert-linalg-to-affine-loops | FileCheck %s 2 3// Test that we can lower all the way to LLVM without crashing, don't check results here. 4// RUN: mlir-opt %s -convert-linalg-to-affine-loops -convert-linalg-to-llvm -o=/dev/null 2>&1 5 6// CHECK-DAG: #[[$strided3D:.*]] = affine_map<(d0, d1, d2)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2 + d2)> 7 8// CHECK-DAG: #[[$stride2Dilation1:.*]] = affine_map<(d0, d1) -> (d0 * 2 + d1)> 9 10// CHECK-DAG: #[[$clampMinMap:.*]] = affine_map<(d0) -> (d0, 0)> 11 12func @matmul(%arg0: memref<?xi8>, %M: index, %N: index, %K: index) { 13 %c0 = constant 0 : index 14 %c1 = constant 1 : index 15 %A = view %arg0[%c0][%M, %K] : memref<?xi8> to memref<?x?xf32> 16 %B = view %arg0[%c0][%K, %N] : memref<?xi8> to memref<?x?xf32> 17 %C = view %arg0[%c0][%M, %N] : memref<?xi8> to memref<?x?xf32> 18 linalg.matmul ins(%A, %B: memref<?x?xf32>, memref<?x?xf32>) 19 outs(%C: memref<?x?xf32>) 20 return 21} 22 23// CHECK-LABEL: func @matmul(%{{.*}}: memref<?xi8>, 24// CHECK-SAME: [[M:arg[0-9]+]]: index 25// CHECK-SAME: [[N:arg[0-9]+]]: index 26// CHECK-SAME: [[K:arg[0-9]+]]: index 27// CHECK: %[[A:.*]] = std.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 28// CHECK: %[[B:.*]] = std.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 29// CHECK: %[[C:.*]] = std.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 30// CHECK: affine.for %{{.*}} = 0 to %{{.*}} { 31// CHECK: affine.for %{{.*}} = 0 to %{{.*}} { 32// CHECK: affine.for %{{.*}} = 0 to %{{.*}} { 33// CHECK-DAG: %[[a:.*]] = affine.load %[[A]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 34// CHECK-DAG: %[[b:.*]] = affine.load %[[B]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 35// CHECK-DAG: %[[inc:.*]] = mulf %[[a]], %[[b]] : f32 36// CHECK-DAG: %[[c:.*]] = affine.load %[[C]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 37// CHECK-DAG: %[[res:.*]] = addf %[[c]], %[[inc]] : f32 38// CHECK: affine.store %[[res]], %[[C]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 39 40func @conv_view3(%arg0: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, %arg1: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, %arg2: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>) { 41 linalg.conv(%arg0, %arg1, %arg2) {strides = [2]}: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]> 42 return 43} 44 45// CHECK-LABEL: func @conv_view3( 46// CHECK: %{{.*}}: memref<?x?x?xf32, #[[$strided3D]]>, %{{.*}}: memref<?x?x?xf32, #[[$strided3D]]>, %{{.*}}: memref<?x?x?xf32, #[[$strided3D]]>) { 47// CHECK: %[[Z0:.*]] = dim %arg0, %c0 : memref<?x?x?xf32, #[[$strided3D]]> 48// CHECK: %[[Q:.*]] = dim %arg0, %c1 : memref<?x?x?xf32, #[[$strided3D]]> 49// CHECK: %[[K:.*]] = dim %arg0, %c2 : memref<?x?x?xf32, #[[$strided3D]]> 50// CHECK: %[[B:.*]] = dim %arg1, %c0 : memref<?x?x?xf32, #[[$strided3D]]> 51// CHECK: %[[X0:.*]] = dim %arg2, %c1 : memref<?x?x?xf32, #[[$strided3D]]> 52// CHECK: affine.for %{{.*}} = 0 to %[[B]] { 53// CHECK: affine.for %{{.*}} = 0 to %[[X0]] { 54// CHECK: affine.for %{{.*}} = 0 to %[[K]] { 55// CHECK: affine.for %{{.*}} = 0 to %[[Q]] { 56// CHECK: affine.for %{{.*}} = 0 to %[[Z0]] { 57// CHECK: %[[SUM:.*]] = affine.apply #[[$stride2Dilation1]](%{{.*}}, %{{.*}}) 58// No padding needed here; only affine loads. 59// CHECK-NEXT: affine.load 60// CHECK-NEXT: affine.load 61 62func @conv_padding(%arg0: memref<?x?x?x?xf32>, 63 %arg1: memref<?x?x?x?xf32>, 64 %arg2: memref<?x?x?x?xf32>) { 65 linalg.conv(%arg0, %arg1, %arg2) {dilations = [1, 1], 66 padding = dense<[[0, 1], [1, 1]]> : tensor<2x2xi64>, 67 strides = [1, 1]} : 68 memref<?x?x?x?xf32>, memref<?x?x?x?xf32>, memref<?x?x?x?xf32> 69 return 70} 71// CHECK-LABEL: func @conv_padding 72// CHECK: %{{.*}}: memref<?x?x?x?xf32>, %{{.*}}: memref<?x?x?x?xf32>, %{{.*}}: memref<?x?x?x?xf32>) { 73// CHECK: %[[ZERO:.*]] = constant 0.000000e+00 : f32 74// CHECK: %[[Z0:.*]] = dim %arg0, %c0 : memref<?x?x?x?xf32> 75// CHECK: %[[Z1:.*]] = dim %arg0, %c1 : memref<?x?x?x?xf32> 76// CHECK: %[[Q:.*]] = dim %arg0, %c2 : memref<?x?x?x?xf32> 77// CHECK: %[[K:.*]] = dim %arg0, %c3 : memref<?x?x?x?xf32> 78// CHECK: %[[B:.*]] = dim %arg1, %c0 : memref<?x?x?x?xf32> 79// CHECK: %[[X0:.*]] = dim %arg2, %c1 : memref<?x?x?x?xf32> 80// CHECK: %[[X1:.*]] = dim %arg2, %c2 : memref<?x?x?x?xf32> 81// CHECK: affine.for %{{.*}} = 0 to %[[B]] { 82// CHECK: affine.for %{{.*}} = 0 to %[[X0]] { 83// CHECK: affine.for %{{.*}} = 0 to %[[X1]] { 84// CHECK: affine.for %{{.*}} = 0 to %[[K]] { 85// CHECK: affine.for %{{.*}} = 0 to %[[Q]] { 86// CHECK: affine.for %{{.*}} = 0 to %[[Z0]] { 87// CHECK: affine.for %{{.*}} = 0 to %[[Z1]] { 88// CHECK: %[[SUM0:.*]] = affine.apply #{{.*}}(%{{.*}}, %{{.*}}) 89// CHECK: %[[SUM1:.*]] = affine.apply #{{.*}}(%{{.*}}, %{{.*}}) 90// CHECK: %[[IDX:.*]] = affine.max #[[$clampMinMap]](%[[SUM0]]) 91// CHECK: %[[IDY:.*]] = affine.max #[[$clampMinMap]](%[[SUM1]]) 92// Padded conv involves an affine.max in the memory access and this is not 93// allowed by affine.load. Use std.load in such cases. 94// CHECK: %{{.*}} = load %{{.*}}[%{{.*}}, %[[IDX]], %[[IDY]], %{{.*}}] : memref<?x?x?x?xf32> 95// CHECK: %{{.*}} = select %{{.*}}, %{{.*}}, %{{.*}} : f32 96// CHECK: %{{.*}} = affine.load %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}] : memref<?x?x?x?xf32> 97// CHECK: %{{.*}} = mulf %{{.*}}, %{{.*}} : f32 98// CHECK: %{{.*}} = affine.load %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}] : memref<?x?x?x?xf32> 99// CHECK: %{{.*}} = addf %{{.*}}, %{{.*}} : f32 100// CHECK: affine.store %{{.*}}, %{{.*}}[%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}] : memref<?x?x?x?xf32> 101 102//----------------------------------------------------------------------------// 103// Named ops to loops. 104//----------------------------------------------------------------------------// 105func @named_batch_matmul(%A: memref<?x?x?xf32>, %B: memref<?x?x?xf32>, %C: memref<?x?x?xf32>) { 106 linalg.batch_matmul ins(%A, %B: memref<?x?x?xf32>, memref<?x?x?xf32>) 107 outs(%C : memref<?x?x?xf32>) 108 return 109} 110// CHECK-LABEL: @named_batch_matmul 111// CHECK-SAME: %[[mA:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 112// CHECK-SAME: %[[mB:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 113// CHECK-SAME: %[[mC:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 114// CHECK: %[[B:.*]] = dim %[[mA]], %c0 : memref<?x?x?xf32> 115// CHECK: %[[M:.*]] = dim %[[mA]], %c1 : memref<?x?x?xf32> 116// CHECK: %[[K:.*]] = dim %[[mA]], %c2 : memref<?x?x?xf32> 117// CHECK: %[[N:.*]] = dim %[[mB]], %c2 : memref<?x?x?xf32> 118// CHECK: affine.for %[[b:.*]] = 0 to %[[B]] { 119// CHECK: affine.for %[[m:.*]] = 0 to %[[M]] { 120// CHECK: affine.for %[[n:.*]] = 0 to %[[N]] { 121// CHECK: affine.for %[[k:.*]] = 0 to %[[K]] { 122// CHECK: %[[va:.*]] = affine.load %[[mA]][%[[b]], %[[m]], %[[k]]] : memref<?x?x?xf32> 123// CHECK: %[[vb:.*]] = affine.load %[[mB]][%[[b]], %[[k]], %[[n]]] : memref<?x?x?xf32> 124// CHECK: %[[vc:.*]] = affine.load %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32> 125// CHECK: %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32 126// CHECK: %[[res:.*]] = addf %[[vc]], %[[inc]] : f32 127// CHECK: affine.store %[[res]], %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32> 128 129// CHECK-LABEL: func @pooling_max_min 130func @pooling_max_min(%arg0: memref<?x?xf32>, 131 %arg1: memref<?x?xi32>, 132 %arg2: memref<?x?xf32>) { 133 linalg.pooling_max(%arg0, %arg1, %arg2) { strides = [2, 1] }: 134 memref<?x?xf32>, memref<?x?xi32>, memref<?x?xf32> 135 linalg.pooling_min(%arg0, %arg1, %arg2) { strides = [2, 1] }: 136 memref<?x?xf32>, memref<?x?xi32>, memref<?x?xf32> 137 return 138} 139// This is a basic check to make sure the right load/stores are used. loops.mlir 140// checks for the rest. 141// CHECK: affine.load 142// CHECK-NEXT: affine.load 143// CHECK-NEXT: cmpf 144// CHECK-NEXT: select 145// CHECK-NEXT: affine.store 146// The min pooling body. 147// CHECK: affine.load 148// CHECK-NEXT: affine.load 149// CHECK-NEXT: cmpf 150// CHECK-NEXT: select 151// CHECK-NEXT: affine.store 152