brintos

brintos / llvm-project-archived public Read only

0
0
Text · 51.7 KiB · 77c7d7d Raw
936 lines · plain
1// RUN: mlir-opt %s -test-linalg-elementwise-fusion-patterns=fuse-with-reshape-by-collapsing -split-input-file | FileCheck %s2// RUN: mlir-opt %s -test-linalg-elementwise-fusion-patterns=fuse-with-reshape-by-collapsing-control -split-input-file | FileCheck %s --check-prefix=CONTROL3 4// Static problem sizes. Checks all aspects of fusion by collapsing. Rest of the5// tests only check a subset of conditions.6#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>7#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2)>8#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d3, d4, d5, d6)>9#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>10#map4 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d1, d2, d0, d7, d3, d4, d5, d6)>11func.func @fuse_by_collapsing(%arg0 : tensor<2x12x5x336x9xi32>,12    %arg1 : tensor<2x3x4xi32>, %arg2 : tensor<5x6x7x8xi32>) -> (tensor<2x3x4x5x6x7x8x9xi32>, tensor<3x4x2x9x5x6x7x8xi32>) {13  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5, 6], [7]] output_shape [2, 3, 4, 5, 6, 7, 8, 9] : tensor<2x12x5x336x9xi32> into tensor<2x3x4x5x6x7x8x9xi32>14  %init_0 = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>15  %init_1 = tensor.empty() : tensor<3x4x2x9x5x6x7x8xi32>16  %generic:2 = linalg.generic {17    indexing_maps = [#map0, #map1, #map2, #map3, #map4],18    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}19    ins(%expand, %arg1, %arg2 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<2x3x4xi32>, tensor<5x6x7x8xi32>)20    outs(%init_0, %init_1 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<3x4x2x9x5x6x7x8xi32>) {21      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32, %b4 : i32):22        %t0 = arith.addi %b0, %b1 : i3223        %t1 = arith.addi %t0, %b2 : i3224        linalg.yield %t1, %t1 : i32, i3225    } -> (tensor<2x3x4x5x6x7x8x9xi32>, tensor<3x4x2x9x5x6x7x8xi32>)26  return %generic#0, %generic#1 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<3x4x2x9x5x6x7x8xi32>27}28//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>29//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1)>30//  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3)>31//  CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d1, d0, d4, d2, d3)>32//      CHECK: func @fuse_by_collapsing(33// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x12x5x336x9xi32>34// CHECK-SAME:   %[[ARG1:.+]]: tensor<2x3x4xi32>35// CHECK-SAME:   %[[ARG2:.+]]: tensor<5x6x7x8xi32>36//  CHECK-DAG:   %[[INIT0:.+]] = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>37//  CHECK-DAG:   %[[INIT1:.+]] = tensor.empty() : tensor<3x4x2x9x5x6x7x8xi32>38//  CHECK-DAG:   %[[ARG1_RESHAPE:.+]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0], [1, 2]{{\]}}39//  CHECK-DAG:   %[[ARG2_RESHAPE:.+]] = tensor.collapse_shape %[[ARG2]] {{\[}}[0], [1, 2, 3]{{\]}}40//  CHECK-DAG:   %[[INIT0_RESHAPE:.+]] = tensor.collapse_shape %[[INIT0]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]{{\]}}41//  CHECK-DAG:   %[[INIT1_RESHAPE:.+]] = tensor.collapse_shape %[[INIT1]] {{\[}}[0, 1], [2], [3], [4], [5, 6, 7]{{\]}}42//      CHECK:   %[[COLLAPSED_OP:.+]]:2 = linalg.generic43// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]], #[[MAP0]], #[[MAP3]]]44// CHECK-SAME:       iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]45// CHECK-SAME:       ins(%[[ARG0]], %[[ARG1_RESHAPE]], %[[ARG2_RESHAPE]] :46// CHECK-SAME:       outs(%[[INIT0_RESHAPE]], %[[INIT1_RESHAPE]] :47//      CHECK:   %[[RESULT0_RESHAPE:.+]] = tensor.expand_shape %[[COLLAPSED_OP]]#0 {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]{{\]}} output_shape [2, 3, 4, 5, 6, 7, 8, 9]48//      CHECK:   %[[RESULT1_RESHAPE:.+]] = tensor.expand_shape %[[COLLAPSED_OP]]#1 {{\[}}[0, 1], [2], [3], [4], [5, 6, 7]{{\]}} output_shape [3, 4, 2, 9, 5, 6, 7, 8]49//      CHECK:   return %[[RESULT0_RESHAPE]], %[[RESULT1_RESHAPE]]50 51//      CONTROL: func @fuse_by_collapsing(52// CONTROL-SAME:   %[[ARG0:.+]]: tensor<2x12x5x336x9xi32>53// CONTROL-SAME:   %[[ARG1:.+]]: tensor<2x3x4xi32>54// CONTROL-SAME:   %[[ARG2:.+]]: tensor<5x6x7x8xi32>55//      CONTROL:   %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]56//      CONTROL:   %[[GENERIC:.+]]:2 = linalg.generic57// CONTROL-SAME:       ins(%[[EXPAND]],58//      CONTROL:   return %[[GENERIC]]#0, %[[GENERIC]]#159 60// -----61 62#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>63#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2)>64#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d3, d4, d5, d6)>65#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>66func.func @fuse_by_collapsing_indexing_op(%arg0 : tensor<2x12x5x336x9xi32>,67    %arg1 : tensor<2x3x4xi32>, %arg2 : tensor<5x6x7x8xi32>) -> tensor<2x3x4x5x6x7x8x9xi32> {68  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5, 6], [7]] output_shape [2, 3, 4, 5, 6, 7, 8, 9] : tensor<2x12x5x336x9xi32> into tensor<2x3x4x5x6x7x8x9xi32>69  %init = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>70  %generic = linalg.generic {71    indexing_maps = [#map0, #map1, #map2, #map3],72    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}73    ins(%expand, %arg1, %arg2 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<2x3x4xi32>, tensor<5x6x7x8xi32>)74    outs(%init : tensor<2x3x4x5x6x7x8x9xi32>) {75      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32):76        %iv0 = linalg.index 0: index77        %iv1 = linalg.index 1: index78        %t0 = arith.addi %iv0, %iv1 : index79        %iv2 = linalg.index 2 : index80        %t1 = arith.addi %t0, %iv2 : index81        %iv3 = linalg.index 3 : index82        %t2 = arith.addi %t1, %iv3 : index83        %iv4 = linalg.index 4 : index84        %t3 = arith.addi %t2, %iv4 : index85        %iv5 = linalg.index 5 : index86        %t4 = arith.addi %t3, %iv5 : index87        %iv6 = linalg.index 6 : index88        %t5 = arith.addi %t4, %iv6 : index89        %iv7 = linalg.index 7 : index90        %t6 = arith.addi %t5, %iv7 : index91        %yield = arith.index_cast %t6 : index to i3292        linalg.yield %yield : i3293    } -> tensor<2x3x4x5x6x7x8x9xi32>94  return %generic : tensor<2x3x4x5x6x7x8x9xi32>95}96// CHECK-LABEL: func @fuse_by_collapsing_indexing_op(97//   CHECK-DAG:   %[[C4:.+]] = arith.constant 4 : index98//   CHECK-DAG:   %[[C8:.+]] = arith.constant 8 : index99//   CHECK-DAG:   %[[C7:.+]] = arith.constant 7 : index100//       CHECK:     %[[IV0:.+]] = linalg.index 0101//       CHECK:     %[[IV1:.+]] = linalg.index 1102//       CHECK:     %[[REM_IV1:.+]] = arith.remsi %[[IV1]], %[[C4]]103//       CHECK:     %[[DIV_IV1:.+]] = arith.divsi %[[IV1]], %[[C4]]104//       CHECK:     %[[IV2:.+]] = linalg.index 2105//       CHECK:     %[[IV3:.+]] = linalg.index 3106//       CHECK:     %[[REM1_IV3:.+]] = arith.remsi %[[IV3]], %[[C8]]107//       CHECK:     %[[DIV1_IV3:.+]] = arith.divsi %[[IV3]], %[[C8]]108//       CHECK:     %[[REM2_IV3:.+]] = arith.remsi %[[DIV1_IV3]], %[[C7]]109//       CHECK:     %[[DIV2_IV3:.+]] = arith.divsi %[[DIV1_IV3]], %[[C7]]110//       CHECK:     %[[IV4:.+]] = linalg.index 4111//       CHECK:     %[[T0:.+]] = arith.addi %[[IV0]], %[[DIV_IV1]]112//       CHECK:     %[[T1:.+]] = arith.addi %[[T0]], %[[REM_IV1]]113//       CHECK:     %[[T2:.+]] = arith.addi %[[T1]], %[[IV2]]114//       CHECK:     %[[T3:.+]] = arith.addi %[[T2]], %[[DIV2_IV3]]115//       CHECK:     %[[T4:.+]] = arith.addi %[[T3]], %[[REM2_IV3]]116//       CHECK:     %[[T5:.+]] = arith.addi %[[T4]], %[[REM1_IV3]]117//       CHECK:     %[[T6:.+]] = arith.addi %[[T5]], %[[IV4]]118//       CHECK:     %[[YIELD:.+]] = arith.index_cast %[[T6]]119//       CHECK:     linalg.yield %[[YIELD]]120 121// -----122 123#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d7, d5, d6, d0, d1, d2, d3, d4)>124#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d5, d6, d0)>125#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d4, d1, d2, d3)>126#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>127func.func @fuse_by_collapsing_change_reshape_order(%arg0 : tensor<9x56x2x60x6xi32>,128    %arg1 : tensor<7x8x2xi32>, %arg2 : tensor<6x3x4x5xi32>) -> tensor<2x3x4x5x6x7x8x9xi32> {129  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5, 6], [7]] output_shape [9, 7, 8, 2, 3, 4, 5, 6] : tensor<9x56x2x60x6xi32> into tensor<9x7x8x2x3x4x5x6xi32>130  %init = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>131  %generic = linalg.generic {132    indexing_maps = [#map0, #map1, #map2, #map3],133    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}134    ins(%expand, %arg1, %arg2 : tensor<9x7x8x2x3x4x5x6xi32>, tensor<7x8x2xi32>, tensor<6x3x4x5xi32>)135    outs(%init : tensor<2x3x4x5x6x7x8x9xi32>) {136      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32):137        %t0 = arith.addi %b0, %b1 : i32138        %t1 = arith.addi %t0, %b2 : i32139        linalg.yield %t1 : i32140    } -> tensor<2x3x4x5x6x7x8x9xi32>141  return %generic : tensor<2x3x4x5x6x7x8x9xi32>142}143//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d4, d3, d0, d1, d2)>144//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d3, d0)>145//  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d2, d1)>146//  CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>147//      CHECK: func @fuse_by_collapsing_change_reshape_order(148// CHECK-SAME:   %[[ARG0:.+]]: tensor<9x56x2x60x6xi32>149// CHECK-SAME:   %[[ARG1:.+]]: tensor<7x8x2xi32>150// CHECK-SAME:   %[[ARG2:.+]]: tensor<6x3x4x5xi32>151//  CHECK-DAG:   %[[INIT:.+]] = tensor.empty()152//  CHECK-DAG:   %[[ARG1_RESHAPE:.+]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0, 1], [2]{{\]}}153//  CHECK-DAG:   %[[ARG2_RESHAPE:.+]] = tensor.collapse_shape %[[ARG2]] {{\[}}[0], [1, 2, 3]{{\]}}154//  CHECK-DAG:   %[[INIT_RESHAPE:.+]] = tensor.collapse_shape %[[INIT]] {{\[}}[0], [1, 2, 3], [4], [5, 6], [7]{{\]}}155//      CHECK:   %[[COLLAPSED_OP:.+]] = linalg.generic156// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]], #[[MAP3]]]157// CHECK-SAME:       iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]158// CHECK-SAME:       ins(%[[ARG0]], %[[ARG1_RESHAPE]], %[[ARG2_RESHAPE]] :159// CHECK-SAME:       outs(%[[INIT_RESHAPE]] :160//      CHECK:   %[[RESULT_RESHAPE:.+]] = tensor.expand_shape %[[COLLAPSED_OP]] {{\[}}[0], [1, 2, 3], [4], [5, 6], [7]{{\]}} output_shape [2, 3, 4, 5, 6, 7, 8, 9]161//      CHECK:   return %[[RESULT_RESHAPE]]162 163// -----164 165// Dynamic case. Only checks things not covered by `fuse_by_collapsing` test above.166#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d7, d5, d6, d0, d1, d2, d3, d4)>167#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d5, d6, d0)>168#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d4, d1, d2, d3)>169#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>170func.func @fuse_by_collapsing_dynamic(%arg0 : tensor<?x?x?x?x?xi32>,171    %arg1 : tensor<?x?x?xi32>, %arg2 : tensor<?x?x?x?xi32>, %sz0: index, %sz1: index, %sz2: index, %sz3: index, %sz4: index) -> tensor<?x3x?x5x?x7x?x?xi32> {172  %c0 = arith.constant 0 : index173  %c1 = arith.constant 1 : index174  %c2 = arith.constant 2 : index175  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5, 6], [7]] output_shape [%sz0, 7, %sz1, %sz2, 3, %sz3, 5, %sz4]176      : tensor<?x?x?x?x?xi32> into tensor<?x7x?x?x3x?x5x?xi32>177  %d0 = tensor.dim %arg1, %c2 : tensor<?x?x?xi32>178  %d2 = tensor.dim %arg2, %c2 : tensor<?x?x?x?xi32>179  %d4 = tensor.dim %arg2, %c0 : tensor<?x?x?x?xi32>180  %d6 = tensor.dim %arg1, %c1 : tensor<?x?x?xi32>181  %d7 = tensor.dim %arg0, %c0 : tensor<?x?x?x?x?xi32>182  %init = tensor.empty(%d0, %d2, %d4, %d6, %d7) : tensor<?x3x?x5x?x7x?x?xi32>183  %generic = linalg.generic {184    indexing_maps = [#map0, #map1, #map2, #map3],185    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}186    ins(%expand, %arg1, %arg2 : tensor<?x7x?x?x3x?x5x?xi32>, tensor<?x?x?xi32>, tensor<?x?x?x?xi32>)187    outs(%init : tensor<?x3x?x5x?x7x?x?xi32>) {188      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32):189        %iv0 = linalg.index 0: index190        %iv1 = linalg.index 1: index191        %t0 = arith.addi %iv0, %iv1 : index192        %iv2 = linalg.index 2 : index193        %t1 = arith.addi %t0, %iv2 : index194        %iv3 = linalg.index 3 : index195        %t2 = arith.addi %t1, %iv3 : index196        %iv4 = linalg.index 4 : index197        %t3 = arith.addi %t2, %iv4 : index198        %iv5 = linalg.index 5 : index199        %t4 = arith.addi %t3, %iv5 : index200        %iv6 = linalg.index 6 : index201        %t5 = arith.addi %t4, %iv6 : index202        %iv7 = linalg.index 7 : index203        %t6 = arith.addi %t5, %iv7 : index204        %yield = arith.index_cast %t6 : index to i32205        linalg.yield %yield : i32206    } -> tensor<?x3x?x5x?x7x?x?xi32>207  return %generic : tensor<?x3x?x5x?x7x?x?xi32>208}209//      CHECK: func @fuse_by_collapsing_dynamic210// CHECK-SAME:     (%[[ARG0:.+]]: tensor<?x?x?x?x?xi32>, %[[SZ0:.+]]: index, %[[SZ1:.+]]: index, %[[SZ2:.+]]: index, %[[SZ3:.+]]: index, %[[SZ4:.+]]: index)211//  CHECK-DAG:   %[[C2:.+]] = arith.constant 2 : index212//  CHECK-DAG:   %[[C5:.+]] = arith.constant 5 : index213//      CHECK:   %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]214//  CHECK-DAG:   %[[D0:.+]] = tensor.dim %[[EXPAND]], %[[C2]]215//  CHECK-DAG:   %[[D1:.+]] = tensor.dim %[[EXPAND]], %[[C5]]216//      CHECK:   linalg.generic217//      CHECK:     %[[IV0:.+]] = linalg.index 1218//      CHECK:     %[[REM1_IV0:.+]] = arith.remsi %[[IV0]], %[[C5]]219//      CHECK:     %[[DIV1_IV0:.+]] = arith.divsi %[[IV0]], %[[C5]]220//      CHECK:     %[[REM2_IV0:.+]] = arith.remsi %[[DIV1_IV0]], %[[D1]]221//      CHECK:     %[[DIV2_IV0:.+]] = arith.divsi %[[DIV1_IV0]], %[[D1]]222//      CHECK:     %[[IV1:.+]] = linalg.index 3223//      CHECK:     %[[REM1_IV1:.+]] = arith.remsi %[[IV1]], %[[D0]]224//      CHECK:     %[[DIV1_IV1:.+]] = arith.divsi %[[IV1]], %[[D0]]225 226// -----227 228#map0 = affine_map<(d0, d1) -> (d0, d1)>229func.func @fuse_by_collapsing_dynamic_2(%arg0 : tensor<?xf32>, %sz0: index, %sz1: index) -> tensor<?x?xf32> {230  %0 = tensor.expand_shape %arg0 [[0, 1]] output_shape [%sz0, %sz1] : tensor<?xf32> into tensor<?x?xf32>231  %init = tensor.empty(%sz1, %sz0) : tensor<?x?xf32>232  %1 = linalg.generic {233      indexing_maps = [#map0, #map0],234      iterator_types = ["parallel", "parallel"]}235      ins(%0 : tensor<?x?xf32>) 236      outs(%init : tensor<?x?xf32>) {237        ^bb0(%b0 : f32, %b1 : f32):238          %out = arith.negf %b0 : f32239          linalg.yield %out : f32240      } -> tensor<?x?xf32>241  return %1 : tensor<?x?xf32>242}243 244// CHECK-LABEL: func @fuse_by_collapsing_dynamic_2245// CHECK-SAME: %[[ARG0:.+]]: tensor<?xf32>246// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index247// CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index248// CHECK:     %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]]249// CHECK-DAG: %[[DIM0:.+]] = tensor.dim %[[EXPANDED]], %[[C0]]250// CHECK-DAG: %[[DIM1:.+]] = tensor.dim %[[EXPANDED]], %[[C1]]251// CHECK:     %[[OUT:.+]] = linalg.generic252// CHECK-SAME:   ins(%[[ARG0]] : tensor<?xf32>)253// CHECK-SAME:   outs(%{{.*}} : tensor<?xf32>)254// CHECK:     %[[EXPANDED_1:.+]] = tensor.expand_shape %[[OUT]]255// CHECK-SAME:    output_shape [%[[DIM0]], %[[DIM1]]]256// CHECK:      return %[[EXPANDED_1]]257 258// -----259 260#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>261#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d3)>262func.func @fuse_reductions(%arg0 : tensor<2x?x5xf32>, %arg1 : tensor<2x5xf32>, %sz0: index) -> tensor<2x5xf32> {263  %0 = tensor.expand_shape %arg0 [[0], [1, 2], [3]] output_shape [2, 6, %sz0, 5] : tensor<2x?x5xf32> into tensor<2x6x?x5xf32>264  %1 = linalg.generic {265      indexing_maps = [#map0, #map1],266      iterator_types = ["parallel", "reduction", "reduction", "parallel"]}267      ins(%0 : tensor<2x6x?x5xf32>) outs(%arg1 : tensor<2x5xf32>) {268        ^bb0(%b0 : f32, %b1 : f32):269          %2 = arith.addf %b0, %b1 : f32270          linalg.yield %2 : f32271      } -> tensor<2x5xf32>272  return %1 : tensor<2x5xf32>273}274//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>275//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2) -> (d0, d2)>276//      CHECK: func @fuse_reductions(277// CHECK-SAME:     %[[ARG0:.+]]: tensor<2x?x5xf32>278// CHECK-SAME:     %[[ARG1:.+]]: tensor<2x5xf32>279// CHECK-SAME:     %[[SZ0:.+]]: index) -> tensor<2x5xf32>280//      CHECK:   %[[GENERIC:.+]] = linalg.generic281// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]]]282// CHECK-SAME:       iterator_types = ["parallel", "reduction", "parallel"]283// CHECK-SAME:       ins(%[[ARG0]] : tensor<2x?x5xf32>)284// CHECK-SAME:       outs(%[[ARG1]] : tensor<2x5xf32>)285 286// -----287 288// Test no fusion because the folded dimensions are not all preserved.289#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>290#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1)>291func.func @no_fuse_unpreserved_folding(%arg0 : tensor<2x12x5xf32>, %arg1 : tensor<2x3xf32>) -> tensor<2x3x4x5xf32> {292  %0 = tensor.expand_shape %arg0 [[0], [1, 2], [3]] output_shape [2, 3, 4, 5] : tensor<2x12x5xf32> into tensor<2x3x4x5xf32>293  %init = tensor.empty(): tensor<2x3x4x5xf32>294  %1 = linalg.generic {295      indexing_maps = [#map0, #map1, #map0],296      iterator_types = ["parallel", "parallel", "parallel", "parallel"]}297      ins(%0, %arg1 : tensor<2x3x4x5xf32>, tensor<2x3xf32>) outs(%init : tensor<2x3x4x5xf32>) {298        ^bb0(%b0 : f32, %b1 : f32, %b2 : f32):299          %2 = arith.addf %b0, %b1 : f32300          linalg.yield %2 : f32301      } -> tensor<2x3x4x5xf32>302  return %1 : tensor<2x3x4x5xf32>303}304//      CHECK: func @no_fuse_unpreserved_folding305// CHECK-SAME:     %[[ARG0:.+]]: tensor<2x12x5xf32>306// CHECK-SAME:     %[[ARG1:.+]]: tensor<2x3xf32>307//      CHECK:   %[[RESHAPE:.+]] = tensor.expand_shape %[[ARG0]]308//      CHECK:   %[[GENERIC:.+]] = linalg.generic309// CHECK-SAME:       ins(%[[RESHAPE]], %[[ARG1]] :310//      CHECK:   return %[[GENERIC]]311 312// -----313 314// Test no fusion because the folded dimensions are not all preserved.315#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>316#map1 = affine_map<(d0, d1, d2, d3) -> (d0)>317#map2 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>318func.func @no_fuse_unpreserved_folding_transpose(%arg0 : tensor<2x12x5xf32>, %arg1 : tensor<2xf32>) -> tensor<2x4x3x5xf32> {319  %0 = tensor.expand_shape %arg0 [[0], [1, 2], [3]] output_shape [2, 3, 4, 5] : tensor<2x12x5xf32> into tensor<2x3x4x5xf32>320  %init = tensor.empty() : tensor<2x4x3x5xf32>321  %1 = linalg.generic {322      indexing_maps = [#map0, #map1, #map2],323      iterator_types = ["parallel", "parallel", "parallel", "parallel"]}324      ins(%0, %arg1 : tensor<2x3x4x5xf32>, tensor<2xf32>) outs(%init : tensor<2x4x3x5xf32>) {325        ^bb0(%b0 : f32, %b1 : f32, %b2 : f32):326          %2 = arith.addf %b0, %b1 : f32327          linalg.yield %2 : f32328      } -> tensor<2x4x3x5xf32>329  return %1 : tensor<2x4x3x5xf32>330}331//      CHECK: func @no_fuse_unpreserved_folding_transpose332// CHECK-SAME:     %[[ARG0:.+]]: tensor<2x12x5xf32>333// CHECK-SAME:     %[[ARG1:.+]]: tensor<2xf32>334//      CHECK:   %[[RESHAPE:.+]] = tensor.expand_shape %[[ARG0]]335//      CHECK:   %[[GENERIC:.+]] = linalg.generic336// CHECK-SAME:       ins(%[[RESHAPE]], %[[ARG1]] :337//      CHECK:   return %[[GENERIC]]338 339// -----340 341// Test no fusion because the iterator types of folded dims are not preserved.342#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>343#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1)>344#map2 = affine_map<(d0, d1, d2, d3) -> (d0, d3)>345func.func @no_fuse_mismatched_iterator_types(%arg0 : tensor<2x12x5xf32>, %arg1 : tensor<2x3xf32>) -> tensor<2x5xf32> {346  %0 = tensor.expand_shape %arg0 [[0], [1, 2], [3]] output_shape [2, 3, 4, 5] : tensor<2x12x5xf32> into tensor<2x3x4x5xf32>347  %init = tensor.empty() : tensor<2x5xf32>348  %1 = linalg.generic {349      indexing_maps = [#map0, #map1, #map2],350      iterator_types = ["parallel", "reduction", "parallel", "parallel"]}351      ins(%0, %arg1 : tensor<2x3x4x5xf32>, tensor<2x3xf32>) outs(%init : tensor<2x5xf32>) {352        ^bb0(%b0 : f32, %b1 : f32, %b2 : f32):353          %2 = arith.addf %b0, %b1 : f32354          linalg.yield %2 : f32355      } -> tensor<2x5xf32>356  return %1 : tensor<2x5xf32>357}358//      CHECK: func @no_fuse_mismatched_iterator_types359// CHECK-SAME:     %[[ARG0:.+]]: tensor<2x12x5xf32>360// CHECK-SAME:     %[[ARG1:.+]]: tensor<2x3xf32>361//      CHECK:   %[[RESHAPE:.+]] = tensor.expand_shape %[[ARG0]]362//      CHECK:   %[[GENERIC:.+]] = linalg.generic363// CHECK-SAME:       ins(%[[RESHAPE]], %[[ARG1]] :364//      CHECK:   return %[[GENERIC]]365 366// -----367 368// Test control of fusion using control function369// Test no fusion because the folded dimensions are not all preserved.370#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d1)>371#map1 = affine_map<(d0, d1, d2, d3) -> (d2, d3)>372#map2 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>373func.func @control_fusion(%arg0 : tensor<6xf32>, %arg1 : tensor<20xf32>) -> tensor<2x3x4x5xf32> {374  %0 = tensor.expand_shape %arg0 [[0, 1]] output_shape [2, 3] : tensor<6xf32> into tensor<2x3xf32>375  %1 = tensor.expand_shape %arg1 [[0, 1]] output_shape [4, 5] : tensor<20xf32> into tensor<4x5xf32>376    %init = tensor.empty() : tensor<2x3x4x5xf32>377  %2 = linalg.generic {378      indexing_maps = [#map0, #map1, #map2],379      iterator_types = ["parallel", "parallel", "parallel", "parallel"]}380      ins(%0, %1 : tensor<2x3xf32>, tensor<4x5xf32>) outs(%init : tensor<2x3x4x5xf32>) {381        ^bb0(%b0 : f32, %b1 : f32, %b2 : f32):382          %3 = arith.addf %b0, %b1 : f32383          linalg.yield %3 : f32384      } -> tensor<2x3x4x5xf32>385  return %2 : tensor<2x3x4x5xf32>386}387//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1) -> (d0)>388//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1) -> (d1)>389//  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1) -> (d0, d1)>390//      CHECK: func @control_fusion(391// CHECK-SAME:     %[[ARG0:.+]]: tensor<6xf32>392// CHECK-SAME:     %[[ARG1:.+]]: tensor<20xf32>393//      CHECK:   %[[GENERIC:.+]] = linalg.generic394// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]395// CHECK-SAME:       iterator_types = ["parallel", "parallel"]396// CHECK-SAME:       ins(%[[ARG0]], %[[ARG1]] :397// CHECK-SAME:       outs(%{{.+}}: tensor<6x20xf32>)398//      CHECK:   %[[RESHAPE1:.+]] = tensor.expand_shape %[[GENERIC]] {{\[}}[0], [1, 2]{{\]}} output_shape [6, 4, 5]399//      CHECK:   %[[RESHAPE2:.+]] = tensor.expand_shape %[[RESHAPE1]] {{\[}}[0, 1], [2], [3]{{\]}} output_shape [2, 3, 4, 5]400//      CHECK:   return %[[RESHAPE2]]401 402//  CONTROL-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>403//  CONTROL-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2) -> (d2)>404//  CONTROL-DAG: #[[MAP3:.+]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>405//      CONTROL: func @control_fusion(406// CONTROL-SAME:     %[[ARG0:.+]]: tensor<6xf32>407// CONTROL-SAME:     %[[ARG1:.+]]: tensor<20xf32>408//      CONTROL:     %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]409//      CONTROL:     %[[INIT:.+]] = tensor.empty()410//      CONTROL:     %[[INIT_RESHAPE:.+]] = tensor.collapse_shape %[[INIT]] {{\[}}[0], [1], [2, 3]{{\]}}411//      CONTROL:     %[[GENERIC:.+]] = linalg.generic412// CONTROL-SAME:         ins(%[[EXPAND]], %[[ARG1]] :413// CONTROL-SAME:         outs(%[[INIT_RESHAPE]] :414//      CONTROL:     %[[RESULT:.+]] = tensor.expand_shape %[[GENERIC]] {{\[}}[0], [1], [2, 3]{{\]}} output_shape [2, 3, 4, 5]415 416// -----417 418// Corner case that isnt handled currently.419#map = affine_map<(d0) -> (d0)>420func.func @zero_D_test(%arg0: tensor<f32>) -> tensor<1xf32> {421  %0 = tensor.expand_shape %arg0 [] output_shape [1] : tensor<f32> into tensor<1xf32>422  %init = tensor.empty() : tensor<1xf32>423  %1 = linalg.generic {424      indexing_maps = [#map, #map],425      iterator_types = ["parallel"]}426      ins(%0: tensor<1xf32>) outs(%init : tensor<1xf32>) {427        ^bb0(%b0 : f32, %b1 : f32):428          linalg.yield %b0: f32429      } -> tensor<1xf32>430  return %1 : tensor<1xf32>431}432//      CHECK: func @zero_D_test433// CHECK-SAME:     %[[ARG0:.+]]: tensor<f32>434//      CHECK:   %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]435//      CHECK:   %[[GENERIC:.+]] = linalg.generic436// CHECK-SAME:       ins(%[[EXPAND]] :437//      CHECK:   return %[[GENERIC]]438 439// -----440 441#map0 = affine_map<(d0, d1, d2, d3) -> (d1, d0, d2, d3)>442#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>443func.func @fuse_only_one_reassociation(%arg0 : tensor<?x?xf32>, %arg1 : tensor<4x?x?x8xf32>, %sz0: index, %sz1: index) -> tensor<4x?x?x8xf32> {444  %0 = tensor.expand_shape %arg0 [[0, 1], [2, 3]] output_shape [%sz0, 4, %sz1, 8] : tensor<?x?xf32> into tensor<?x4x?x8xf32>445  %1 = linalg.generic {446      indexing_maps = [#map0, #map1, #map1],447      iterator_types = ["parallel", "parallel", "parallel", "parallel"]}448      ins(%0, %arg1 : tensor<?x4x?x8xf32>, tensor<4x?x?x8xf32>)449      outs(%arg1 : tensor<4x?x?x8xf32>) {450    ^bb0(%b0: f32, %b1 : f32, %b2 : f32):451      %2 = arith.addf %b0, %b1 : f32452      linalg.yield %2 : f32453    } -> tensor<4x?x?x8xf32>454  return %1 : tensor<4x?x?x8xf32>455}456//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2) -> (d1, d0, d2)>457//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>458//      CHECK: func @fuse_only_one_reassociation459// CHECK-SAME:     (%[[ARG0:.+]]: tensor<?x?xf32>, %[[ARG1:.+]]: tensor<4x?x?x8xf32>, %[[SZ0:.+]]: index, %[[SZ1:.+]]: index)460//  CHECK-DAG:   %[[C2:.*]] = arith.constant 2 : index461//  CHECK-DAG:   %[[C0:.*]] = arith.constant 0 : index462//  CHECK-DAG:   %[[EXPAND_ARG0:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0, 1], [2, 3]{{\]}} output_shape [%[[SZ0]], 4, %[[SZ1]], 8]463//  CHECK-DAG:   %[[DIM:.+]] = tensor.dim %[[EXPAND_ARG0]], %[[C0]] : tensor<?x4x?x8xf32>464//  CHECK-DAG:   %[[DIM_2:.+]] = tensor.dim %[[EXPAND_ARG0]], %[[C2]] : tensor<?x4x?x8xf32>465//  CHECK-DAG:   %[[COLLAPSE_ARG0:.+]] = tensor.collapse_shape %[[EXPAND_ARG0]] {{\[}}[0], [1], [2, 3]{{\]}}466//  CHECK-DAG:   %[[COLLAPSE_ARG1_0:.+]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0], [1], [2, 3]{{\]}}467//  CHECK-DAG:   %[[COLLAPSE_ARG1_1:.+]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0], [1], [2, 3]{{\]}}468//      CHECK:   %[[GENERIC:.+]] = linalg.generic469// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP1]]]470// CHECK-SAME:       iterator_types = ["parallel", "parallel", "parallel"]471// CHECK-SAME:       ins(%[[COLLAPSE_ARG0]], %[[COLLAPSE_ARG1_0]] :472// CHECK-SAME:       outs(%[[COLLAPSE_ARG1_1]] :473//      CHECK:   %[[EXPANDED_3:.+]] = tensor.expand_shape %[[GENERIC]] {{\[\[}}0], [1], [2, 3]] output_shape [4, %[[DIM]], %[[DIM_2]], 8] : tensor<4x?x?xf32> into tensor<4x?x?x8xf32>474//      CHECK:   return %[[EXPANDED_3]]475 476// -----477 478#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)>479#map1 = affine_map<(d0, d1, d2, d3) -> (d3, d1, d0, d2)>480func.func @fold_non_consecutive_dims(%arg0 : tensor<?x?xi32>, %sz0: index, %sz1: index) -> tensor<?x8x?x4xi32> {481  %c0 = arith.constant 0 : index482  %c2 = arith.constant 2 : index483  %0 = tensor.expand_shape %arg0 [[0, 1], [2, 3]] output_shape [%sz0, 4, %sz1, 8] : tensor<?x?xi32> into tensor<?x4x?x8xi32>484  %d0 = tensor.dim %0, %c0 : tensor<?x4x?x8xi32>485  %d1 = tensor.dim %0, %c2 : tensor<?x4x?x8xi32>486  %init = tensor.empty(%d1, %d0) : tensor<?x8x?x4xi32>487  %1 = linalg.generic {488      indexing_maps = [#map0, #map1],489      iterator_types = ["parallel", "parallel", "parallel", "parallel"]}490      ins(%0 : tensor<?x4x?x8xi32>) outs(%init : tensor<?x8x?x4xi32>) {491    ^bb0(%b0 : i32, %b1 : i32):492      %2 = linalg.index 0 : index493      %3 = linalg.index 1 : index494      %4 = linalg.index 2 : index495      %5 = linalg.index 3 : index496      %6 = arith.addi %2, %3 : index497      %7 = arith.addi %6, %4 : index498      %8 = arith.addi %7, %5 : index499      %9 = arith.index_cast %8 : index to i32500      linalg.yield %9: i32501    } -> tensor<?x8x?x4xi32>502  return %1 : tensor<?x8x?x4xi32>503}504//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1) -> (d0, d1)>505//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1) -> (d1, d0)>506//      CHECK: func @fold_non_consecutive_dims(507// CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xi32>, %[[SZ0:.+]]: index, %[[SZ1:.+]]: index)508//      CHECK-DAG:   %[[C4:.+]] = arith.constant 4 : index509//      CHECK-DAG:   %[[C8:.+]] = arith.constant 8 : index510//      CHECK-DAG:   %[[C0:.+]] = arith.constant 0 : index511//      CHECK-DAG:   %[[C2:.+]] = arith.constant 2 : index512//      CHECK:   %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]] {{\[\[}}0, 1], [2, 3]] output_shape [%[[SZ0]], 4, %[[SZ1]], 8] : tensor<?x?xi32> into tensor<?x4x?x8xi32>513//      CHECK-DAG:   %[[DIM:.+]] = tensor.dim %[[EXPANDED]], %[[C0]]514//      CHECK-DAG:   %[[DIM_0:.+]] = tensor.dim %[[EXPANDED]], %[[C2]]515//      CHECK:   %[[INIT:.+]] = tensor.empty(%[[DIM_0]], %[[DIM]])516//      CHECK-DAG:   %[[DIM_1:.+]] = tensor.dim %[[EXPANDED]], %[[C0]]517//      CHECK-DAG:   %[[DIM_2:.+]] = tensor.dim %[[EXPANDED]], %[[C2]]518//      CHECK:   %[[COLLAPSE_INIT:.+]] = tensor.collapse_shape %[[INIT]] {{\[}}[0, 1], [2, 3]{{\]}}519//      CHECK:   %[[GENERIC:.+]] = linalg.generic520// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]]]521// CHECK-SAME:       iterator_types = ["parallel", "parallel"]522// CHECK-SAME:       ins(%[[ARG0]] :523// CHECK-SAME:       outs(%[[COLLAPSE_INIT]] :524// CHECK-NEXT:   ^bb{{[0-9]}}525//      CHECK:       %[[ID0:.+]] = linalg.index 0526//  CHECK-DAG:       %[[T0:.+]] = arith.remsi %[[ID0]], %[[C4]]527//  CHECK-DAG:       %[[T1:.+]] = arith.divsi %[[ID0]], %[[C4]]528//      CHECK:       %[[ID1:.+]] = linalg.index 1529//  CHECK-DAG:       %[[T2:.+]] = arith.remsi %[[ID1]], %[[C8]]530//  CHECK-DAG:       %[[T3:.+]] = arith.divsi %[[ID1]], %[[C8]]531//  CHECK-DAG:       %[[T4:.+]] = arith.addi %[[T1]], %[[T2]]532//  CHECK-DAG:       %[[T5:.+]] = arith.addi %[[T4]], %[[T0]]533//  CHECK-DAG:       %[[T6:.+]] = arith.addi %[[T5]], %[[T3]]534//  CHECK-DAG:       %[[T7:.+]] = arith.index_cast %[[T6]]535//      CHECK:       linalg.yield %[[T7]]536//      CHECK:   %[[EXPANDED_3:.+]] = tensor.expand_shape %[[GENERIC]] {{\[\[}}0, 1], [2, 3]] output_shape [%[[DIM_2]], 8, %[[DIM_1]], 4] : tensor<?x?xi32> into tensor<?x8x?x4xi32>537//      CHECK:   return %[[EXPANDED_3]]538 539// -----540 541// None of the folded iteration space dims are contiguous reduction dimensions.542// So no change in the code.543#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)>544#map1 = affine_map<(d0, d1, d2, d3) -> ()>545func.func @no_fold_non_consecutive_reduction_dims(%arg0 : tensor<?x?xi32>, %sz0: index, %sz1: index) -> tensor<i32> {546  %c0 = arith.constant 0 : index547  %c2 = arith.constant 2 : index548  %0 = tensor.expand_shape %arg0 [[0, 1], [2, 3]] output_shape [%sz0, 4, %sz1, 8] : tensor<?x?xi32> into tensor<?x4x?x8xi32>549  %init = tensor.empty() : tensor<i32>550  %1 = linalg.generic {551      indexing_maps = [#map0, #map1],552      iterator_types = ["reduction", "reduction", "reduction", "reduction"]}553      ins(%0 : tensor<?x4x?x8xi32>) outs(%init : tensor<i32>) {554    ^bb0(%b0 : i32, %b1 : i32):555      %2 = linalg.index 0 : index556      %3 = linalg.index 1 : index557      %4 = linalg.index 2 : index558      %5 = linalg.index 3 : index559      %6 = arith.addi %2, %3 : index560      %7 = arith.addi %6, %4 : index561      %8 = arith.addi %7, %5 : index562      %9 = arith.index_cast %8 : index to i32563      linalg.yield %9: i32564    } -> tensor<i32>565  return %1 : tensor<i32>566}567//      CHECK: func @no_fold_non_consecutive_reduction_dims(568// CHECK-SAME:   %[[ARG0:.+]]: tensor<?x?xi32>, %[[SZ0:.+]]: index, %[[SZ1:.+]]: index)569//      CHECK:   %[[EXPAND_ARG0:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0, 1], [2, 3]{{\]}} output_shape [%[[SZ0]], 4, %[[SZ1]], 8]570//      CHECK:   %[[GENERIC:.+]] = linalg.generic571// CHECK-SAME:       ins(%[[EXPAND_ARG0]] :572//      CHECK:   return %[[GENERIC]]573 574// -----575 576func.func @fuse_by_collapsing_pad(%arg0 : tensor<2x12x5x336x9xi32>) -> tensor<8x3x4x17x6x7x8x14xi32> {577  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5, 6], [7]] output_shape [2, 3, 4, 5, 6, 7, 8, 9] : tensor<2x12x5x336x9xi32> into tensor<2x3x4x5x6x7x8x9xi32>578  %cst = arith.constant 0 : i32579  %padded_0 = tensor.pad %expand low[1, 0, 0, 8, 0, 0, 0, 3] high[5, 0, 0, 4, 0, 0, 0, 2] {580  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index,581       %arg5: index, %arg6: index, %arg7: index, %arg8: index):582    tensor.yield %cst : i32583  } : tensor<2x3x4x5x6x7x8x9xi32> to tensor<8x3x4x17x6x7x8x14xi32>584  return %padded_0 : tensor<8x3x4x17x6x7x8x14xi32>585}586//      CHECK: func @fuse_by_collapsing_pad(587// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x12x5x336x9xi32>)588//      CHECK:   %[[PAD:.+]] = tensor.pad %[[ARG0]]589// CHECK-SAME:       low[1, 0, 8, 0, 3] high[5, 0, 4, 0, 2]590//      CHECK:       tensor<2x12x5x336x9xi32> to tensor<8x12x17x336x14xi32>591//      CHECK:   %[[EXPAND:.+]] = tensor.expand_shape %[[PAD]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]]592// CHECK-SAME:       output_shape [8, 3, 4, 17, 6, 7, 8, 14] : tensor<8x12x17x336x14xi32> into tensor<8x3x4x17x6x7x8x14xi32>593//      CHECK:   return %[[EXPAND]]594 595// -----596 597func.func @no_fuse_by_collapsing_pad_non_constant_padding(%arg0 : tensor<2x12xi32>) -> tensor<8x3x4xi32> {598  %expand = tensor.expand_shape %arg0 [[0], [1, 2]] output_shape [2, 3, 4] : tensor<2x12xi32> into tensor<2x3x4xi32>599  %cst = arith.constant 0 : i32600  %padded_0 = tensor.pad %expand low[1, 0, 0] high[5, 0, 0] {601  ^bb0(%arg1: index, %arg2: index, %arg3: index):602    %pad_val = arith.index_cast %arg1 : index to i32603    tensor.yield %pad_val : i32604  } : tensor<2x3x4xi32> to tensor<8x3x4xi32>605  return %padded_0 : tensor<8x3x4xi32>606}607//      CHECK: func @no_fuse_by_collapsing_pad_non_constant_padding(608// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x12xi32>)609//      CHECK:   %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]610//      CHECK:   %[[PAD:.+]] = tensor.pad %[[EXPAND]]611//      CHECK:   return %[[PAD]]612 613// -----614 615func.func @no_fuse_by_collapsing_pad(%arg0 : tensor<2x12x5x336x9xi32>) -> tensor<8x5x4x17x6x7x8x14xi32> {616  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5, 6], [7]] output_shape [2, 3, 4, 5, 6, 7, 8, 9] : tensor<2x12x5x336x9xi32> into tensor<2x3x4x5x6x7x8x9xi32>617  %cst = arith.constant 0 : i32618  %padded_0 = tensor.pad %expand low[1, 2, 0, 8, 0, 0, 0, 3] high[5, 0, 0, 4, 0, 0, 0, 2] {619  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index,620       %arg5: index, %arg6: index, %arg7: index, %arg8: index):621    tensor.yield %cst : i32622  } : tensor<2x3x4x5x6x7x8x9xi32> to tensor<8x5x4x17x6x7x8x14xi32>623  return %padded_0 : tensor<8x5x4x17x6x7x8x14xi32>624}625//      CHECK: func @no_fuse_by_collapsing_pad(626// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x12x5x336x9xi32>)627//      CHECK:   %[[EXPAND_ARG0:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]]628// CHECK-SAME:       output_shape [2, 3, 4, 5, 6, 7, 8, 9] : tensor<2x12x5x336x9xi32> into tensor<2x3x4x5x6x7x8x9xi32>629//      CHECK:   %[[PAD:.+]] = tensor.pad %[[EXPAND_ARG0]]630// CHECK-SAME:       low[1, 2, 0, 8, 0, 0, 0, 3] high[5, 0, 0, 4, 0, 0, 0, 2]631//      CHECK:       tensor<2x3x4x5x6x7x8x9xi32> to tensor<8x5x4x17x6x7x8x14xi32>632//      CHECK:   return %[[PAD]]633 634// -----635 636func.func @fuse_by_collapsing_dynamic_pad(%arg0 : tensor<?x?x?x?xf32>,637    %s0 : index, %s1 : index, %s2 : index, %s3 : index, %s4 : index, %s5 : index,638    %l0 : index, %l1 : index, %h0 : index, %h1 : index) -> tensor<?x?x?x?x?x?xf32> {639  %expand = tensor.expand_shape %arg0 [[0], [1, 2], [3], [4, 5]] output_shape [%s0, %s1, %s2, %s3, %s4, %s5] : tensor<?x?x?x?xf32> into tensor<?x?x?x?x?x?xf32>640  %cst = arith.constant 0.0 : f32641  %padded_0 = tensor.pad %expand low[%l0, 0, 0, %l1, 0, 0] high[%h0, 0, 0, %h1, 0, 0] {642  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index, %arg5: index, %arg6: index):643    tensor.yield %cst : f32644  } : tensor<?x?x?x?x?x?xf32> to tensor<?x?x?x?x?x?xf32>645  return %padded_0 : tensor<?x?x?x?x?x?xf32>646}647//  CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0, s1, s2] -> (s0 + s1 + s2)>648//      CHECK: func @fuse_by_collapsing_dynamic_pad(649// CHECK-SAME:   %[[ARG0:.+]]: tensor<?x?x?x?xf32>650// CHECK-SAME:   %[[S0:.+]]: index, %[[S1:.+]]: index, %[[S2:.+]]: index, %[[S3:.+]]: index, %[[S4:.+]]: index, %[[S5:.+]]: index, %[[L0:.+]]: index, %[[L1:.+]]: index, %[[H0:.+]]: index, %[[H1:.+]]: index651//      CHECK:   %[[PAD_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[L0]], %[[H0]], %[[S0]]]652//      CHECK:   %[[PAD_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[L1]], %[[H1]], %[[S3]]]653//      CHECK:   %[[PAD:.+]] = tensor.pad %[[ARG0]]654// CHECK-SAME:       low[%[[L0]], 0, %[[L1]], 0] high[%[[H0]], 0, %[[H1]], 0]655//      CHECK:       tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32>656//      CHECK:   %[[EXPAND:.+]] = tensor.expand_shape %[[PAD]] {{\[}}[0], [1, 2], [3], [4, 5]]657// CHECK-SAME:       output_shape [%[[PAD_SIZE0]], %[[S1]], %[[S2]], %[[PAD_SIZE1]], %[[S4]], %[[S5]]] : tensor<?x?x?x?xf32> into tensor<?x?x?x?x?x?xf32>658//      CHECK:   return %[[EXPAND]]659 660// -----661 662func.func @collapse_shape_with_producer_pad(%arg0: tensor<2x3x4x5x6x7x8x9xi32>) -> tensor<8x12x17x336x14xi32> {663  %cst = arith.constant 0 : i32664  %padded = tensor.pad %arg0 low[1, 0, 0, 8, 0, 0, 0, 3] high[5, 0, 0, 4, 0, 0, 0, 2] {665  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index,666       %arg5: index, %arg6: index, %arg7: index, %arg8: index):667    tensor.yield %cst : i32668  } : tensor<2x3x4x5x6x7x8x9xi32> to tensor<8x3x4x17x6x7x8x14xi32>669  %collapsed = tensor.collapse_shape %padded [[0], [1, 2], [3], [4, 5, 6], [7]]670    : tensor<8x3x4x17x6x7x8x14xi32> into tensor<8x12x17x336x14xi32>671  return %collapsed : tensor<8x12x17x336x14xi32>672}673//      CHECK: func @collapse_shape_with_producer_pad674// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x3x4x5x6x7x8x9xi32>675//      CHECK:   %[[COLLAPSE:.+]] = tensor.collapse_shape %[[ARG0]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]]676//      CHECK:   %[[PAD:.+]] = tensor.pad %[[COLLAPSE]] low[1, 0, 8, 0, 3] high[5, 0, 4, 0, 2]677//      CHECK:   return %[[PAD]]678 679// -----680 681func.func @collapse_shape_with_producer_pad_dynamic(%arg0: tensor<?x?x?x?x?x?xf32>,682    %l0 : index, %l1 : index, %h0 : index, %h1 : index) -> tensor<?x?x?x?xf32> {683  %cst = arith.constant 0.0 : f32684  %padded = tensor.pad %arg0 low[%l0, 0, 0, %l1, 0, 0] high[%h0, 0, 0, %h1, 0, 0] {685  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index, %arg5: index, %arg6: index):686    tensor.yield %cst : f32687  } : tensor<?x?x?x?x?x?xf32> to tensor<?x?x?x?x?x?xf32>688  %collapsed = tensor.collapse_shape %padded [[0], [1, 2], [3], [4, 5]]689    : tensor<?x?x?x?x?x?xf32> into tensor<?x?x?x?xf32>690  return %collapsed : tensor<?x?x?x?xf32>691}692//      CHECK: func @collapse_shape_with_producer_pad_dynamic693// CHECK-SAME:   %[[ARG0:.+]]: tensor<?x?x?x?x?x?xf32>694// CHECK-SAME:   %[[L0:.+]]: index, %[[L1:.+]]: index, %[[H0:.+]]: index, %[[H1:.+]]: index695//      CHECK:   %[[COLLAPSE:.+]] = tensor.collapse_shape %[[ARG0]] {{\[}}[0], [1, 2], [3], [4, 5]]696//      CHECK:   %[[PAD:.+]] = tensor.pad %[[COLLAPSE]] low[%[[L0]], 0, %[[L1]], 0] high[%[[H0]], 0, %[[H1]], 0]697//      CHECK:   return %[[PAD]]698 699// -----700 701func.func @collapse_shape_with_producer_pad_non_constant_padding(%arg0 : tensor<2x3x4xi32>) -> tensor<8x12xi32> {702  %cst = arith.constant 0 : i32703  %padded_0 = tensor.pad %arg0 low[1, 0, 0] high[5, 0, 0] {704  ^bb0(%arg1: index, %arg2: index, %arg3: index):705    %pad_val = arith.index_cast %arg1 : index to i32706    tensor.yield %pad_val : i32707  } : tensor<2x3x4xi32> to tensor<8x3x4xi32>708  %collapsed = tensor.collapse_shape %padded_0 [[0], [1, 2]] : tensor<8x3x4xi32> into tensor<8x12xi32>709  return %collapsed : tensor<8x12xi32>710}711//      CHECK: func @collapse_shape_with_producer_pad_non_constant_padding(712// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x3x4xi32>)713//      CHECK:   %[[PAD:.+]] = tensor.pad %[[ARG0]]714//      CHECK:   %[[COLLAPSED:.+]] = tensor.collapse_shape %[[PAD]]715//      CHECK:   return %[[COLLAPSED]]716 717// -----718// Static problem sizes. Checks all aspects of fusion by collapsing with bubbling up collapse shapes.719#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>720#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2)>721#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d3, d4, d5, d6)>722#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>723#map4 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d1, d2, d0, d7, d3, d4, d5, d6)>724func.func @fuse_by_collapsing_bubblecollapse(%arg0 : tensor<2x3x4x5x6x7x8x9xi32>,725    %arg1 : tensor<2x3x4xi32>, %arg2 : tensor<5x6x7x8xi32>) -> (tensor<2x12x5x336x9xi32>, tensor<12x2x9x5x336xi32>) {726  %init_0 = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>727  %init_1 = tensor.empty() : tensor<3x4x2x9x5x6x7x8xi32>728  %generic:2 = linalg.generic {729    indexing_maps = [#map0, #map1, #map2, #map3, #map4],730    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}731    ins(%arg0, %arg1, %arg2 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<2x3x4xi32>, tensor<5x6x7x8xi32>)732    outs(%init_0, %init_1 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<3x4x2x9x5x6x7x8xi32>) {733      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32, %b4 : i32):734        %t0 = arith.addi %b0, %b1 : i32735        %t1 = arith.addi %t0, %b2 : i32736        linalg.yield %t1, %t1 : i32, i32737    } -> (tensor<2x3x4x5x6x7x8x9xi32>, tensor<3x4x2x9x5x6x7x8xi32>)738  %collapse_1 =  tensor.collapse_shape %generic#0 [[0], [1, 2], [3], [4, 5, 6], [7]] :  tensor<2x3x4x5x6x7x8x9xi32> into tensor<2x12x5x336x9xi32>739  %collapse_2 =  tensor.collapse_shape %generic#1 [[0, 1], [2], [3], [4], [5, 6, 7]] : tensor<3x4x2x9x5x6x7x8xi32> into tensor<12x2x9x5x336xi32>740  return %collapse_1, %collapse_2 : tensor<2x12x5x336x9xi32>, tensor<12x2x9x5x336xi32>741}742//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>743//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1)>744//  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3)>745//  CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d1, d0, d4, d2, d3)>746//      CHECK: func @fuse_by_collapsing_bubblecollapse(747// CHECK-SAME:   %[[ARG0:.+]]: tensor<2x3x4x5x6x7x8x9xi32>748// CHECK-SAME:   %[[ARG1:.+]]: tensor<2x3x4xi32>749// CHECK-SAME:   %[[ARG2:.+]]: tensor<5x6x7x8xi32>750//  CHECK-DAG:   %[[INIT0:.+]] = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>751//  CHECK-DAG:   %[[INIT1:.+]] = tensor.empty() : tensor<3x4x2x9x5x6x7x8xi32>752//  CHECK-DAG:   %[[ARG0_RESHAPE:.+]] = tensor.collapse_shape %[[ARG0]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]{{\]}}753//  CHECK-DAG:   %[[ARG1_RESHAPE:.+]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0], [1, 2]{{\]}}754//  CHECK-DAG:   %[[ARG2_RESHAPE:.+]] = tensor.collapse_shape %[[ARG2]] {{\[}}[0], [1, 2, 3]{{\]}}755//  CHECK-DAG:   %[[INIT0_RESHAPE:.+]] = tensor.collapse_shape %[[INIT0]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]{{\]}}756//  CHECK-DAG:   %[[INIT1_RESHAPE:.+]] = tensor.collapse_shape %[[INIT1]] {{\[}}[0, 1], [2], [3], [4], [5, 6, 7]{{\]}}757//      CHECK:   %[[COLLAPSED_OP:.+]]:2 = linalg.generic758// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]], #[[MAP0]], #[[MAP3]]]759// CHECK-SAME:       iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]760// CHECK-SAME:       ins(%[[ARG0_RESHAPE]], %[[ARG1_RESHAPE]], %[[ARG2_RESHAPE]] :761// CHECK-SAME:       outs(%[[INIT0_RESHAPE]], %[[INIT1_RESHAPE]] :762//      CHECK:   return %[[COLLAPSED_OP]]#0, %[[COLLAPSED_OP]]#1763 764// -----765 766#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>767#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2)>768#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d3, d4, d5, d6)>769#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>770func.func @fuse_by_collapsing_indexing_op_bubblecollapse(%arg0 : tensor<2x3x4x5x6x7x8x9xi32>,771    %arg1 : tensor<2x3x4xi32>, %arg2 : tensor<5x6x7x8xi32>) -> tensor<2x12x5x336x9xi32> {772  %init = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>773  %generic = linalg.generic {774    indexing_maps = [#map0, #map1, #map2, #map3],775    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}776    ins(%arg0, %arg1, %arg2 : tensor<2x3x4x5x6x7x8x9xi32>, tensor<2x3x4xi32>, tensor<5x6x7x8xi32>)777    outs(%init : tensor<2x3x4x5x6x7x8x9xi32>) {778      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32):779        %iv0 = linalg.index 0: index780        %iv1 = linalg.index 1: index781        %t0 = arith.addi %iv0, %iv1 : index782        %iv2 = linalg.index 2 : index783        %t1 = arith.addi %t0, %iv2 : index784        %iv3 = linalg.index 3 : index785        %t2 = arith.addi %t1, %iv3 : index786        %iv4 = linalg.index 4 : index787        %t3 = arith.addi %t2, %iv4 : index788        %iv5 = linalg.index 5 : index789        %t4 = arith.addi %t3, %iv5 : index790        %iv6 = linalg.index 6 : index791        %t5 = arith.addi %t4, %iv6 : index792        %iv7 = linalg.index 7 : index793        %t6 = arith.addi %t5, %iv7 : index794        %yield = arith.index_cast %t6 : index to i32795        linalg.yield %yield : i32796    } -> tensor<2x3x4x5x6x7x8x9xi32>797  %collapse = tensor.collapse_shape %generic [[0], [1, 2], [3], [4, 5, 6], [7]] : tensor<2x3x4x5x6x7x8x9xi32> into tensor<2x12x5x336x9xi32>798  return %collapse : tensor<2x12x5x336x9xi32>799}800// CHECK-LABEL: func @fuse_by_collapsing_indexing_op_bubblecollapse(801//   CHECK-DAG:   %[[C4:.+]] = arith.constant 4 : index802//   CHECK-DAG:   %[[C8:.+]] = arith.constant 8 : index803//   CHECK-DAG:   %[[C7:.+]] = arith.constant 7 : index804//       CHECK:     %[[IV0:.+]] = linalg.index 0805//       CHECK:     %[[IV1:.+]] = linalg.index 1806//       CHECK:     %[[REM_IV1:.+]] = arith.remsi %[[IV1]], %[[C4]]807//       CHECK:     %[[DIV_IV1:.+]] = arith.divsi %[[IV1]], %[[C4]]808//       CHECK:     %[[IV2:.+]] = linalg.index 2809//       CHECK:     %[[IV3:.+]] = linalg.index 3810//       CHECK:     %[[REM1_IV3:.+]] = arith.remsi %[[IV3]], %[[C8]]811//       CHECK:     %[[DIV1_IV3:.+]] = arith.divsi %[[IV3]], %[[C8]]812//       CHECK:     %[[REM2_IV3:.+]] = arith.remsi %[[DIV1_IV3]], %[[C7]]813//       CHECK:     %[[DIV2_IV3:.+]] = arith.divsi %[[DIV1_IV3]], %[[C7]]814//       CHECK:     %[[IV4:.+]] = linalg.index 4815//       CHECK:     %[[T0:.+]] = arith.addi %[[IV0]], %[[DIV_IV1]]816//       CHECK:     %[[T1:.+]] = arith.addi %[[T0]], %[[REM_IV1]]817//       CHECK:     %[[T2:.+]] = arith.addi %[[T1]], %[[IV2]]818//       CHECK:     %[[T3:.+]] = arith.addi %[[T2]], %[[DIV2_IV3]]819//       CHECK:     %[[T4:.+]] = arith.addi %[[T3]], %[[REM2_IV3]]820//       CHECK:     %[[T5:.+]] = arith.addi %[[T4]], %[[REM1_IV3]]821//       CHECK:     %[[T6:.+]] = arith.addi %[[T5]], %[[IV4]]822//       CHECK:     %[[YIELD:.+]] = arith.index_cast %[[T6]]823//       CHECK:     linalg.yield %[[YIELD]]824 825// -----826 827#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d7, d5, d6, d0, d1, d2, d3, d4)>828#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d5, d6, d0)>829#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d4, d1, d2, d3)>830#map3 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>831func.func @fuse_by_collapsing_change_reshape_order_bubblecollapse(%arg0 : tensor<9x7x8x2x3x4x5x6xi32>,832    %arg1 : tensor<7x8x2xi32>, %arg2 : tensor<6x3x4x5xi32>) -> tensor<2x60x6x56x9xi32> {833  %init = tensor.empty() : tensor<2x3x4x5x6x7x8x9xi32>834  %generic = linalg.generic {835    indexing_maps = [#map0, #map1, #map2, #map3],836    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]}837    ins(%arg0, %arg1, %arg2 : tensor<9x7x8x2x3x4x5x6xi32>, tensor<7x8x2xi32>, tensor<6x3x4x5xi32>)838    outs(%init : tensor<2x3x4x5x6x7x8x9xi32>) {839      ^bb0(%b0 : i32, %b1 : i32, %b2 : i32, %b3 : i32):840        %t0 = arith.addi %b0, %b1 : i32841        %t1 = arith.addi %t0, %b2 : i32842        linalg.yield %t1 : i32843    } -> tensor<2x3x4x5x6x7x8x9xi32>844  %collapse = tensor.collapse_shape %generic [[0], [1, 2, 3], [4], [5, 6], [7]] : tensor<2x3x4x5x6x7x8x9xi32> into tensor<2x60x6x56x9xi32>845  return %collapse : tensor<2x60x6x56x9xi32>846}847 848//  CHECK-DAG: #[[MAP0:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d4, d3, d0, d1, d2)>849//  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d3, d0)>850//  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d2, d1)>851//  CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>852//      CHECK: func @fuse_by_collapsing_change_reshape_order_bubblecollapse(853// CHECK-SAME:   %[[ARG0:.+]]: tensor<9x7x8x2x3x4x5x6xi32>854// CHECK-SAME:   %[[ARG1:.+]]: tensor<7x8x2xi32>855// CHECK-SAME:   %[[ARG2:.+]]: tensor<6x3x4x5xi32>856//  CHECK-DAG:   %[[INIT:.+]] = tensor.empty()857//  CHECK-DAG:   %[[ARG0_RESHAPE:.+]] = tensor.collapse_shape %[[ARG0]] {{\[}}[0], [1, 2], [3], [4, 5, 6], [7]{{\]}}858//  CHECK-DAG:   %[[ARG1_RESHAPE:.+]] = tensor.collapse_shape %[[ARG1]] {{\[}}[0, 1], [2]{{\]}}859//  CHECK-DAG:   %[[ARG2_RESHAPE:.+]] = tensor.collapse_shape %[[ARG2]] {{\[}}[0], [1, 2, 3]{{\]}}860//  CHECK-DAG:   %[[INIT_RESHAPE:.+]] = tensor.collapse_shape %[[INIT]] {{\[}}[0], [1, 2, 3], [4], [5, 6], [7]{{\]}}861//      CHECK:   %[[COLLAPSED_OP:.+]] = linalg.generic862// CHECK-SAME:       indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]], #[[MAP3]]]863// CHECK-SAME:       iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]864// CHECK-SAME:       ins(%[[ARG0_RESHAPE]], %[[ARG1_RESHAPE]], %[[ARG2_RESHAPE]] :865// CHECK-SAME:       outs(%[[INIT_RESHAPE]] :866//      CHECK:   return %[[COLLAPSED_OP]]867 868//      CONTROL: func @fuse_by_collapsing_change_reshape_order_bubblecollapse(869// CONTROL-SAME:   %[[ARG0:.+]]: tensor<9x7x8x2x3x4x5x6xi32>870// CONTROL-SAME:   %[[ARG1:.+]]: tensor<7x8x2xi32>871// CONTROL-SAME:   %[[ARG2:.+]]: tensor<6x3x4x5xi32>872//      CONTROL:   %[[GENERIC:.+]] = linalg.generic873// CONTROL-SAME:       ins(%[[ARG0]],874//      CONTROL:   %[[COLLAPSE:.+]] = tensor.collapse_shape %[[GENERIC]]875//      CONTROL:   return %[[COLLAPSE]]876 877// -----878 879// Check that new ops are inserted at `%0` because `%0` is also used by `tensor.dim`.880#map0 = affine_map<(d0, d1) -> (d0, d1)>881func.func @fuse_by_collapsing_correct_insertion(%arg0 : tensor<?x?xf32>,882    %sz0: index, %sz1: index) -> (tensor<?xf32>, index) {883  %c0 = arith.constant 0 : index884  %init = tensor.empty(%sz1, %sz0) : tensor<?x?xf32>885  %0 = linalg.generic {886      indexing_maps = [#map0, #map0],887      iterator_types = ["parallel", "parallel"]}888      ins(%arg0 : tensor<?x?xf32>)889      outs(%init : tensor<?x?xf32>) {890        ^bb0(%b0 : f32, %b1 : f32):891          %out = arith.negf %b0 : f32892          linalg.yield %out : f32893      } -> tensor<?x?xf32>894  %dim = tensor.dim %0, %c0 : tensor<?x?xf32>895  %1 = tensor.collapse_shape %0 [[0, 1]] : tensor<?x?xf32> into tensor<?xf32>896  return %1, %dim : tensor<?xf32>, index897}898 899// CHECK-LABEL: func @fuse_by_collapsing_correct_insertion900// CHECK-SAME: %[[ARG0:.+]]: tensor<?x?xf32>901// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index902// CHECK:     %[[COLLAPSE:.+]] = tensor.collapse_shape %[[ARG0]]903// CHECK:     %[[OUT:.+]] = linalg.generic904// CHECK-SAME:   ins(%[[COLLAPSE]] : tensor<?xf32>)905// CHECK:     %[[EXPANDED:.+]] = tensor.expand_shape %[[OUT]]906// CHECK:     %[[DIM:.+]] = tensor.dim %[[EXPANDED]], %[[C0]]907// CHECK:      return %[[OUT]], %[[DIM]]908 909// -----910 911#map = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>912#map1 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d3, d4, d1, d2)>913func.func @partial_fuse_by_collapsing(%arg0: tensor<4x?x32x128x192xf16>, %arg1: tensor<4x128x192x?x32xf32>) -> tensor<512x192x?xf32> {914  %0 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<4x?x32x128x192xf16>) outs(%arg1 : tensor<4x128x192x?x32xf32>) {915  ^bb0(%in: f16, %out: f32):916    linalg.yield %out : f32917  } -> tensor<4x128x192x?x32xf32>918  %collapsed = tensor.collapse_shape %0 [[0, 1], [2], [3, 4]] : tensor<4x128x192x?x32xf32> into tensor<512x192x?xf32>919  return %collapsed : tensor<512x192x?xf32>920}921// CHECK-LABEL: func @partial_fuse_by_collapsing922//  CHECK-SAME:  %[[ARG0:.+]]: tensor<4x?x32x128x192xf16>923//  CHECK-SAME:  %[[ARG1:.+]]: tensor<4x128x192x?x32xf32>924//   CHECK-DAG:   %[[COLLAPSED0:.+]] = tensor.collapse_shape %[[ARG0]]925//  CHECK-SAME:     tensor<4x?x32x128x192xf16> into tensor<4x?x128x192xf16>926//   CHECK-DAG:   %[[COLLAPSED1:.+]] = tensor.collapse_shape %[[ARG1]]927//  CHECK-SAME:     tensor<4x128x192x?x32xf32> into tensor<4x128x192x?xf32>928//       CHECK:   %[[GENERIC:.+]] = linalg.generic929//  CHECK-SAME:     ins(%[[COLLAPSED0]]930//  CHECK-SAME:     outs(%[[COLLAPSED1]]931//       CHECK:   %[[EXPANDED:.+]] = tensor.expand_shape %[[GENERIC]]932//  CHECK-SAME:     tensor<4x128x192x?xf32> into tensor<4x128x192x?x32xf32>933//       CHECK:   %[[COLLAPSED:.+]] = tensor.collapse_shape %[[EXPANDED]]934//  CHECK-SAME:     tensor<4x128x192x?x32xf32> into tensor<512x192x?xf32>935//       CHECK:   return %[[COLLAPSED]] : tensor<512x192x?xf32>936