brintos

brintos / llvm-project-archived public Read only

0
0
Text · 21.1 KiB · 34aaf05 Raw
302 lines · plain
1// RUN: mlir-opt --pass-pipeline="builtin.module(func.func(sharding-propagation,cse))" %s | FileCheck %s2 3shard.grid @grid_2(shape = 2)4shard.grid @grid_1d(shape = ?)5shard.grid @grid_2d(shape = 2x4)6shard.grid @grid_3d(shape = ?x?x?)7 8// CHECK-LABEL: func.func @element_wise_empty_sharding_info9func.func @element_wise_empty_sharding_info(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {10  // CHECK-NEXT: tosa.sigmoid11  %0 = tosa.sigmoid %arg0 : (tensor<8x16xf32>) -> tensor<8x16xf32>12  // CHECK-NEXT: return13  return %0 : tensor<8x16xf32>14}15 16// CHECK-LABEL: func.func @element_wise_on_def17// CHECK-SAME:    %[[ARG:.*]]: tensor<8x16xf32>18func.func @element_wise_on_def(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {19  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0], [1]] : !shard.sharding20  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG]] to %[[S0]] annotate_for_users  : tensor<8x16xf32>21  // CHECK-NEXT:  %[[V1:.*]] = tosa.sigmoid %[[V0]]22  %0 = tosa.sigmoid %arg0 : (tensor<8x16xf32>) -> tensor<8x16xf32>23  // CHECK-NEXT:  %[[V2:.*]] = shard.shard %[[V1]] to %[[S0]]  : tensor<8x16xf32>24  %s1 = shard.sharding @grid_2d split_axes = [[0], [1]] : !shard.sharding25  %1 = shard.shard %0 to %s1  : tensor<8x16xf32>26  // CHECK-NEXT:  return %[[V2]]27  return %1 : tensor<8x16xf32>28}29 30// CHECK-LABEL: func.func @element_wise_on_use31// CHECK-SAME:    %[[ARG:.*]]: tensor<8x16xf32>32func.func @element_wise_on_use(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {33  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0], [1]] : !shard.sharding34  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG]] to %[[S0]] annotate_for_users  : tensor<8x16xf32>35  %s0 = shard.sharding @grid_2d split_axes = [[0], [1]] : !shard.sharding36  %0 = shard.shard %arg0 to %s0 annotate_for_users  : tensor<8x16xf32>37  // CHECK-NEXT:  %[[V1:.*]] = tosa.sigmoid %[[V0]]38  %1 = tosa.sigmoid %0 : (tensor<8x16xf32>) -> tensor<8x16xf32>39  // CHECK-NEXT:  %[[V2:.*]] = shard.shard %[[V1]] to %[[S0]]  : tensor<8x16xf32>40  // CHECK-NEXT:  return %[[V2]]41  return %1 : tensor<8x16xf32>42}43 44// CHECK-LABEL: func.func @element_wise_on_graph_output45// CHECK-SAME:    %[[ARG:.*]]: tensor<8x16xf32>46func.func @element_wise_on_graph_output(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {47  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0], [1]] : !shard.sharding48  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG]] to %[[S0]] annotate_for_users  : tensor<8x16xf32>49  // CHECK-NEXT:  %[[V1:.*]] = tosa.sigmoid %[[V0]]50  %0 = tosa.sigmoid %arg0 : (tensor<8x16xf32>) -> tensor<8x16xf32>51  // CHECK-NEXT:  %[[V2:.*]] = shard.shard %[[V1]] to %[[S0]]  : tensor<8x16xf32>52  // CHECK-NEXT:  %[[V3:.*]] = shard.shard %[[V2]] to %[[S0]] annotate_for_users  : tensor<8x16xf32>53  %s1 = shard.sharding @grid_2d split_axes = [[0], [1]] : !shard.sharding54  %1 = shard.shard %0 to %s1 annotate_for_users  : tensor<8x16xf32>55  // CHECK-NEXT:  return %[[V3]]56  return %1 : tensor<8x16xf32>57}58 59// CHECK-LABEL: func.func @element_wise_on_graph_input60// CHECK-SAME:    %[[ARG:.*]]: tensor<8x16xf32>61func.func @element_wise_on_graph_input(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {62  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0], [1]] : !shard.sharding63  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG]] to %[[S0]]  : tensor<8x16xf32>64  // CHECK-NEXT:  %[[V1:.*]] = shard.shard %[[V0]] to %[[S0]] annotate_for_users  : tensor<8x16xf32>65  %s0 = shard.sharding @grid_2d split_axes = [[0], [1]] : !shard.sharding66  %0 = shard.shard %arg0 to %s0  : tensor<8x16xf32>67  // CHECK-NEXT:  %[[V2:.*]] = tosa.sigmoid %[[V1]]68  %1 = tosa.sigmoid %0 : (tensor<8x16xf32>) -> tensor<8x16xf32>69  // CHECK-NEXT:  %[[V3:.*]] = shard.shard %[[V2]] to %[[S0]]  : tensor<8x16xf32>70  // CHECK-NEXT:  return %[[V3]]71  return %1 : tensor<8x16xf32>72}73 74// CHECK-LABEL: func.func @arrow_structure75// CHECK-SAME:    %[[ARG:.*]]: tensor<8x16xf32>76func.func @arrow_structure(%arg0: tensor<8x16xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> (tensor<8x16xf32>, tensor<8x16xf32>) {77  // CHECK-NEXT:  %[[S1:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0], [1]] : !shard.sharding78  // CHECK-NEXT:  %[[V1:.*]] = shard.shard %[[ARG]] to %[[S1]] annotate_for_users  : tensor<8x16xf32>79  // CHECK-NEXT:  %[[V2:.*]] = tosa.tanh %[[V1]]80  // CHECK-NEXT:  %[[V3:.*]] = shard.shard %[[V2]] to %[[S1]]  : tensor<8x16xf32>81  %0 = tosa.tanh %arg0 : (tensor<8x16xf32>) -> tensor<8x16xf32>82  // CHECK-NEXT:  %[[V4:.*]] = shard.shard %[[V3]] to %[[S1]] annotate_for_users  : tensor<8x16xf32>83  // CHECK-NEXT:  %[[V5:.*]] = tosa.abs %[[V4]]84  // CHECK-NEXT:  %[[V6:.*]] = shard.shard %[[V5]] to %[[S1]]  : tensor<8x16xf32>85  %1 = tosa.abs %0: (tensor<8x16xf32>) -> tensor<8x16xf32>86  // CHECK-NEXT:  %[[S3:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}]] : !shard.sharding87  // CHECK-NEXT:  %[[ZP1:.*]] = shard.shard %arg1 to %[[S3]] annotate_for_users : tensor<1xf32>88  // CHECK-NEXT:  %[[ZP2:.*]] = shard.shard %arg2 to %[[S3]] annotate_for_users : tensor<1xf32>89  // CHECK-NEXT:  %[[V7:.*]] = tosa.negate %[[V4]], %[[ZP1]], %[[ZP2]]90  // CHECK-NEXT:  %[[V8:.*]] = shard.shard %[[V7]] to %[[S1]]  : tensor<8x16xf32>91  %2 = tosa.negate %0, %arg1, %arg2 : (tensor<8x16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<8x16xf32>92  %s3 = shard.sharding @grid_2d split_axes = [[0], [1]] : !shard.sharding93  %3 = shard.shard %2 to %s3  : tensor<8x16xf32>94  // CHECK-NEXT: return %[[V6]], %[[V8]]95  return %1, %3 : tensor<8x16xf32>, tensor<8x16xf32>96}97 98// CHECK-LABEL: func.func @matmul_on_def_shard_batch_and_m99// CHECK-SAME:     %[[ARG0:.*]]: tensor<2x16x8xf32>, %[[ARG1:.*]]: tensor<2x8x32xf32>, %[[ARG2:.*]]: tensor<1xf32>100func.func @matmul_on_def_shard_batch_and_m(%arg0: tensor<2x16x8xf32>, %arg1: tensor<2x8x32xf32>, %arg2: tensor<1xf32>) -> tensor<2x16x32xf32> {101  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0], [1]] : !shard.sharding102  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG0]] to %[[S0]] annotate_for_users  : tensor<2x16x8xf32>103  // CHECK-NEXT:  %[[S1:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}0]] : !shard.sharding104  // CHECK-NEXT:  %[[V1:.*]] = shard.shard %[[ARG1]] to %[[S1]] annotate_for_users  : tensor<2x8x32xf32>105  // CHECK-NEXT:  %[[S2:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}]] : !shard.sharding106  // CHECK-NEXT:  %[[ZP:.*]] = shard.shard %[[ARG2]] to %[[S2]] annotate_for_users  : tensor<1xf32>107  // CHECK-NEXT:  %[[V2:.*]] = tosa.matmul %[[V0]], %[[V1]], %[[ZP]], %[[ZP]]108  %0 = tosa.matmul %arg0, %arg1, %arg2, %arg2 : (tensor<2x16x8xf32>, tensor<2x8x32xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x16x32xf32>109  // CHECK-NEXT:  %[[V3:.*]] = shard.shard %[[V2]] to %[[S0]]  : tensor<2x16x32xf32>110  %s1 = shard.sharding @grid_2d split_axes = [[0], [1]] : !shard.sharding111  %1 = shard.shard %0 to %s1  : tensor<2x16x32xf32>112  // CHECK-NEXT:  return %[[V3]]113  return %1 : tensor<2x16x32xf32>114}115 116// CHECK-LABEL: func.func @matmul_on_def_shard_m_and_n117// CHECK-SAME:     [[varg0:%.*]]: tensor<2x16x8xf32>, [[varg1:%.*]]: tensor<2x8x32xf32>, [[varg2:%.*]]: tensor<1xf32>118func.func @matmul_on_def_shard_m_and_n(%arg0: tensor<2x16x8xf32>, %arg1: tensor<2x8x32xf32>, %arg2: tensor<1xf32>) -> tensor<2x16x32xf32> {119  // CHECK: [[vsharding:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [0]] : !shard.sharding120  // CHECK: [[vsharded:%.*]] = shard.shard [[varg0]] to [[vsharding]] annotate_for_users : tensor<2x16x8xf32>121  // CHECK: [[vsharding_0:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [], [1]] : !shard.sharding122  // CHECK: [[vsharded_1:%.*]] = shard.shard [[varg1]] to [[vsharding_0]] annotate_for_users : tensor<2x8x32xf32>123  // CHECK: [[vsharding_2:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}]] : !shard.sharding124  // CHECK: [[vsharded_3:%.*]] = shard.shard [[varg2]] to [[vsharding_2]] annotate_for_users : tensor<1xf32>125  // CHECK: [[v0:%.*]] = tosa.matmul126  %0 = tosa.matmul %arg0, %arg1, %arg2, %arg2 : (tensor<2x16x8xf32>, tensor<2x8x32xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x16x32xf32>127  // CHECK: [[vsharding_4:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [0], [1]] : !shard.sharding128  // CHECK: [[vsharded_5:%.*]] = shard.shard [[v0]] to [[vsharding_4]] : tensor<2x16x32xf32>129  %s1 = shard.sharding @grid_2d split_axes = [[], [0], [1]] : !shard.sharding130  %1 = shard.shard %0 to %s1  : tensor<2x16x32xf32>131  // CHECK-NEXT:  return [[vsharded_5]]132  return %1 : tensor<2x16x32xf32>133}134 135// CHECK-LABEL: func.func @matmul_on_use_shard_m_and_k136// CHECK-SAME:     [[varg0:%.*]]: tensor<2x16x8xf32>, [[varg1:%.*]]: tensor<2x8x32xf32>, [[varg2:%.*]]: tensor<1xf32>137func.func @matmul_on_use_shard_m_and_k(%arg0: tensor<2x16x8xf32>, %arg1: tensor<2x8x32xf32>, %arg2: tensor<1xf32>) -> tensor<2x16x32xf32> {138  // CHECK: [[vsharding:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [0], [1]] : !shard.sharding139  %s0 = shard.sharding @grid_2d split_axes = [[], [0], [1]] : !shard.sharding140  // CHECK: [[vsharded:%.*]] = shard.shard [[varg0]] to [[vsharding]] : tensor<2x16x8xf32>141  %arg0_s = shard.shard %arg0 to %s0 : tensor<2x16x8xf32>142  // CHECK: [[vsharded_0:%.*]] = shard.shard [[vsharded]] to [[vsharding]] annotate_for_users : tensor<2x16x8xf32>143  // CHECK: [[vsharding_1:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [1]] : !shard.sharding144  // CHECK: [[vsharded_2:%.*]] = shard.shard [[varg1]] to [[vsharding_1]] annotate_for_users : tensor<2x8x32xf32>145  // CHECK: [[vsharding_3:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}]] : !shard.sharding146  // CHECK: [[vsharded_4:%.*]] = shard.shard [[varg2]] to [[vsharding_3]] annotate_for_users : tensor<1xf32>147  // CHECK: [[v0:%.*]] = tosa.matmul148  // CHECK: [[vsharding_5:%.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [0]] : !shard.sharding149  // CHECK: [[vsharded_6:%.*]] = shard.shard [[v0]] to [[vsharding_5]] : tensor<2x16x32xf32>150  %0 = tosa.matmul %arg0_s, %arg1, %arg2, %arg2 : (tensor<2x16x8xf32>, tensor<2x8x32xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x16x32xf32>151  // CHECK: return [[vsharded_6]]152  return %0 : tensor<2x16x32xf32>153}154 155// CHECK-LABEL: func.func @matmul_on_use_shard_m_and_duplicted_k156// CHECK-SAME:     %[[ARG0:.*]]: tensor<2x16x8xf32>, %[[ARG1:.*]]: tensor<2x8x32xf32>, %[[ARG2:.*]]: tensor<1xf32>157func.func @matmul_on_use_shard_m_and_duplicted_k(%arg0: tensor<2x16x8xf32>, %arg1: tensor<2x8x32xf32>, %arg2: tensor<1xf32>) -> tensor<2x16x32xf32> {158  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [1], [0]] : !shard.sharding159  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG0]] to %[[S0]] annotate_for_users  : tensor<2x16x8xf32>160  %s0 = shard.sharding @grid_2d split_axes = [[], [1], [0]] : !shard.sharding161  %0 = shard.shard %arg0 to %s0 annotate_for_users  : tensor<2x16x8xf32>162  // CHECK-NEXT:  %[[S1:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [0]] : !shard.sharding163  // CHECK-NEXT:  %[[V1:.*]] = shard.shard %[[ARG1]] to %[[S1]] annotate_for_users  : tensor<2x8x32xf32>164  %s1 = shard.sharding @grid_2d split_axes = [[], [0]] : !shard.sharding165  %1 = shard.shard %arg1 to %s1 annotate_for_users  : tensor<2x8x32xf32>166  // CHECK-NEXT:  %[[S2:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}]] : !shard.sharding167  // CHECK-NEXT:  %[[ZP:.*]] = shard.shard %[[ARG2]] to %[[S2]] annotate_for_users  : tensor<1xf32>168  // CHECK-NEXT:  %[[V2:.*]] = tosa.matmul %[[V0]], %[[V1]], %[[ZP]], %[[ZP]]169  %2 = tosa.matmul %0, %1, %arg2, %arg2 : (tensor<2x16x8xf32>, tensor<2x8x32xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x16x32xf32>170  // CHECK-NEXT:  %[[S3:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}], [1]] : !shard.sharding171  // CHECK-NEXT:  %[[V3:.*]] = shard.shard %[[V2]] to %[[S3]]  : tensor<2x16x32xf32>172  // CHECK-NEXT:  return %[[V3]]173  return %2 : tensor<2x16x32xf32>174}175 176// CHECK-LABEL: func.func @resolve_conflicting_annotations177func.func @resolve_conflicting_annotations(178  // CHECK-SAME: %[[IN1:.*]]: tensor<2x3xf32>,179  %arg0: tensor<2x3xf32>,180  // CHECK-SAME: %[[IN2:.*]]: tensor<3x2xf32>,181  %arg1: tensor<3x2xf32>,182  // CHECK-SAME: %[[OUT_DPS:.*]]: tensor<2x2xf32>183  %out_dps: tensor<2x2xf32>184// CHECK-SAME: ) -> tensor<2x2xf32> {185) -> tensor<2x2xf32> {186  // CHECK: %[[SIN1_SHARDED1:.*]] = shard.sharding @grid_2 split_axes = {{\[\[}}0]] : !shard.sharding187  // CHECK-NEXT:  %[[IN1_SHARDED1:.*]] = shard.shard %[[IN1]] to %[[SIN1_SHARDED1]]  : tensor<2x3xf32>188  // CHECK: %[[SIN2_SHARDED:.*]] = shard.sharding @grid_2 split_axes = {{\[\[}}]] : !shard.sharding189  // CHECK-NEXT:  %[[IN1_SHARDED2:.*]] = shard.shard %[[IN1_SHARDED1]] to %[[SIN2_SHARDED]] annotate_for_users  : tensor<2x3xf32>190  // CHECK-NEXT:  %[[IN2_SHARDED:.*]] = shard.shard %[[IN2]] to %[[SIN2_SHARDED]] annotate_for_users  : tensor<3x2xf32>191  // CHECK-NEXT:  %[[OUT_DPS_SHARDED:.*]] = shard.shard %[[OUT_DPS]] to %[[SIN2_SHARDED]] annotate_for_users  : tensor<2x2xf32>192  %sarg0_sharded = shard.sharding @grid_2 split_axes = [[0]] : !shard.sharding193  %arg0_sharded = shard.shard %arg0 to %sarg0_sharded  : tensor<2x3xf32>194  // CHECK: %[[MATMUL:.*]] = linalg.matmul ins(%[[IN1_SHARDED2]], %[[IN2_SHARDED]] : tensor<2x3xf32>, tensor<3x2xf32>)195  // CHECK-SAME: outs(%[[OUT_DPS_SHARDED]] : tensor<2x2xf32>) -> tensor<2x2xf32>196  %res = linalg.matmul ins(%arg0_sharded, %arg1 : tensor<2x3xf32>, tensor<3x2xf32>)197    outs(%out_dps : tensor<2x2xf32>) -> tensor<2x2xf32>198  // CHECK-NEXT: %[[RES:.*]] = shard.shard %[[MATMUL]] to %[[SIN2_SHARDED]] : tensor<2x2xf32>199  %sres_sharded = shard.sharding @grid_2 split_axes = [[]] : !shard.sharding200  %res_sharded = shard.shard %res to %sres_sharded  : tensor<2x2xf32>201  // CHECK: return %[[RES]] : tensor<2x2xf32>202  return %res_sharded : tensor<2x2xf32>203}204 205// https://arxiv.org/abs/2211.05102 Figure 2(a)206// The sharding propagation results in unnecessary reshards,207//   an optimization pass should be able to remove them.208// CHECK-LABEL: func.func @mlp_1d_weight_stationary209// CHECK-SAME:     [[varg0:%.*]]: tensor<2x4x8xf32>, [[varg1:%.*]]: tensor<2x8x32xf32>, [[varg2:%.*]]: tensor<2x32x8xf32>, [[varg3:%.*]]: tensor<1xf32>210func.func @mlp_1d_weight_stationary(%arg0: tensor<2x4x8xf32>, %arg1: tensor<2x8x32xf32>, %arg2: tensor<2x32x8xf32>, %arg3: tensor<1xf32>) -> tensor<2x4x8xf32> {211  %s0 = shard.sharding @grid_1d split_axes = [[], [], [0, 1, 2]] : !shard.sharding212  %sharded0 = shard.shard %arg0 to %s0 : tensor<2x4x8xf32>213  %sharded1 = shard.shard %arg1 to %s0 : tensor<2x8x32xf32>214  // CHECK: [[vsharding:%.*]] = shard.sharding @grid_1d split_axes = {{\[\[}}], [], [0, 1, 2]] : !shard.sharding215  // CHECK: [[vsharded:%.*]] = shard.shard [[varg0]] to [[vsharding]] : tensor<2x4x8xf32>216  // CHECK: [[vsharded_0:%.*]] = shard.shard [[varg1]] to [[vsharding]] : tensor<2x8x32xf32>217  // CHECK: [[vsharded_1:%.*]] = shard.shard [[vsharded]] to [[vsharding]] annotate_for_users : tensor<2x4x8xf32>218  // CHECK: [[vsharding_2:%.*]] = shard.sharding @grid_1d split_axes = {{\[\[}}], [0, 1, 2]] : !shard.sharding219  // CHECK: [[vsharded_3:%.*]] = shard.shard [[vsharded_0]] to [[vsharding_2]] annotate_for_users : tensor<2x8x32xf32>220  // CHECK: [[vsharding_4:%.*]] = shard.sharding @grid_1d split_axes = {{\[\[}}]] : !shard.sharding221  // CHECK: [[vsharded_5:%.*]] = shard.shard [[varg3]] to [[vsharding_4]] annotate_for_users : tensor<1xf32>222  // CHECK: [[v0:%.*]] = tosa.matmul223  %1 = tosa.matmul %sharded0, %sharded1, %arg3, %arg3 : (tensor<2x4x8xf32>, tensor<2x8x32xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x4x32xf32>224  // CHECK: [[vsharded_6:%.*]] = shard.shard [[v0]] to [[vsharding_4]] : tensor<2x4x32xf32>225  // CHECK: [[vsharded_7:%.*]] = shard.shard [[vsharded_6]] to [[vsharding_4]] annotate_for_users : tensor<2x4x32xf32>226  // CHECK: [[v1:%.*]] = tosa.sigmoid [[vsharded_7]] : (tensor<2x4x32xf32>) -> tensor<2x4x32xf32>227  // CHECK: [[vsharded_8:%.*]] = shard.shard [[v1]] to [[vsharding_4]] : tensor<2x4x32xf32>228  %2 = tosa.sigmoid %1 : (tensor<2x4x32xf32>) -> tensor<2x4x32xf32>229  %sharding = shard.sharding @grid_1d split_axes = [[], [0, 1, 2]] : !shard.sharding230  // CHECK: [[vsharded_9:%.*]] = shard.shard [[varg2]] to [[vsharding_2]] : tensor<2x32x8xf32>231  %sharded2 = shard.shard %arg2 to %sharding  : tensor<2x32x8xf32>232  // CHECK: [[vsharded_10:%.*]] = shard.shard [[vsharded_8]] to [[vsharding_4]] annotate_for_users : tensor<2x4x32xf32>233  // CHECK: [[vsharded_11:%.*]] = shard.shard [[vsharded_9]] to [[vsharding]] annotate_for_users : tensor<2x32x8xf32>234  // CHECK: [[v2:%.*]] = tosa.matmul235  %3 = tosa.matmul %2, %sharded2, %arg3, %arg3 : (tensor<2x4x32xf32>, tensor<2x32x8xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x4x8xf32>236  // CHECK: [[vsharded_12:%.*]] = shard.shard [[v2]] to [[vsharding]] : tensor<2x4x8xf32>237  %s4 = shard.sharding @grid_1d split_axes = [[], [], [0, 1, 2]] : !shard.sharding238  %4 = shard.shard %3 to %s4  : tensor<2x4x8xf32>239  // CHECK: return [[vsharded_12]]240  return %4 : tensor<2x4x8xf32>241}242 243// https://arxiv.org/abs/2211.05102 Figure 2(b)244// The sharding propagation results in unnecessary reshards,245//   an optimization pass should be able to remove them.246// CHECK-LABEL: func.func @mlp_2d_weight_stationary247// CHECK-SAME:     [[varg0:%.*]]: tensor<2x4x8xf32>, [[varg1:%.*]]: tensor<2x8x32xf32>, [[varg2:%.*]]: tensor<2x32x8xf32>, [[varg3:%.*]]: tensor<1xf32>248func.func @mlp_2d_weight_stationary(%arg0: tensor<2x4x8xf32>, %arg1: tensor<2x8x32xf32>, %arg2: tensor<2x32x8xf32>, %arg3: tensor<1xf32>) -> tensor<2x4x8xf32> {249    // CHECK: [[vsharding:%.*]] = shard.sharding @grid_3d split_axes = {{\[\[}}], [], [0, 1, 2]] : !shard.sharding250  %s0 = shard.sharding @grid_3d split_axes = [[], [], [0, 1, 2]] : !shard.sharding251    // CHECK: [[vsharded:%.*]] = shard.shard [[varg0]] to [[vsharding]] : tensor<2x4x8xf32>252  %arg0_s = shard.shard %arg0 to %s0  : tensor<2x4x8xf32>253    // CHECK: [[vsharding_0:%.*]] = shard.sharding @grid_3d split_axes = {{\[\[}}], [0], [1, 2]] : !shard.sharding254  %s1 = shard.sharding @grid_3d split_axes = [[], [0], [1, 2]] : !shard.sharding255    // CHECK: [[vsharded_1:%.*]] = shard.shard [[varg1]] to [[vsharding_0]] : tensor<2x8x32xf32>256  %arg1_s = shard.shard %arg1 to %s1  : tensor<2x8x32xf32>257    // CHECK: [[vsharding_2:%.*]] = shard.sharding @grid_3d split_axes = {{\[\[}}]] : !shard.sharding258    // CHECK: [[vsharded_3:%.*]] = shard.shard [[vsharded]] to [[vsharding_2]] annotate_for_users : tensor<2x4x8xf32>259    // CHECK: [[vsharded_4:%.*]] = shard.shard [[vsharded_1]] to [[vsharding]] annotate_for_users : tensor<2x8x32xf32>260    // CHECK: [[vsharded_5:%.*]] = shard.shard [[varg3]] to [[vsharding_2]] annotate_for_users : tensor<1xf32>261    // CHECK: [[v0:%.*]] = tosa.matmul262  %1 = tosa.matmul %arg0_s, %arg1_s, %arg3, %arg3 : (tensor<2x4x8xf32>, tensor<2x8x32xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x4x32xf32>263    // CHECK: [[vsharded_6:%.*]] = shard.shard [[v0]] to [[vsharding]] : tensor<2x4x32xf32>264  %2 = shard.shard %1 to %s0  : tensor<2x4x32xf32>265    // CHECK: [[vsharded_7:%.*]] = shard.shard [[vsharded_6]] to [[vsharding]] annotate_for_users : tensor<2x4x32xf32>266    // CHECK: [[v1:%.*]] = tosa.sigmoid 267    // CHECK: [[vsharded_8:%.*]] = shard.shard [[v1]] to [[vsharding]] : tensor<2x4x32xf32>268  %3 = tosa.sigmoid %2 : (tensor<2x4x32xf32>) -> tensor<2x4x32xf32>269    // CHECK: [[vsharding_9:%.*]] = shard.sharding @grid_3d split_axes = {{\[\[}}], [1, 2], [0]] : !shard.sharding270  %s2 = shard.sharding @grid_3d split_axes = [[], [1, 2], [0]] : !shard.sharding271    // CHECK: [[vsharded_10:%.*]] = shard.shard [[varg2]] to [[vsharding_9]] : tensor<2x32x8xf32>272  %arg2_s = shard.shard %arg2 to %s2  : tensor<2x32x8xf32>273    // CHECK: [[vsharded_11:%.*]] = shard.shard [[vsharded_8]] to [[vsharding_2]] annotate_for_users : tensor<2x4x32xf32>274    // CHECK: [[vsharded_12:%.*]] = shard.shard [[vsharded_10]] to [[vsharding]] annotate_for_users : tensor<2x32x8xf32>275    // CHECK: [[v2:%.*]] = tosa.matmul276  %4 = tosa.matmul %3, %arg2_s, %arg3, %arg3 : (tensor<2x4x32xf32>, tensor<2x32x8xf32>, tensor<1xf32>, tensor<1xf32>)  -> tensor<2x4x8xf32>277    // CHECK: [[vsharded_13:%.*]] = shard.shard [[v2]] to [[vsharding]] : tensor<2x4x8xf32>278  %5 = shard.shard %4 to %s0  : tensor<2x4x8xf32>279    // CHECK: [[vsharded_14:%.*]] = shard.shard [[vsharded_13]] to [[vsharding]] annotate_for_users : tensor<2x4x8xf32>280  %6 = shard.shard %5 to %s0 annotate_for_users  : tensor<2x4x8xf32>281    // CHECK: return [[vsharded_14]]282  return %6 : tensor<2x4x8xf32>283}284 285// CHECK-LABEL: func.func @elementwise_duplicated_chain286// CHECK-SAME:    %[[ARG:.*]]: tensor<8x16xf32>287func.func @elementwise_duplicated_chain(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {288  // CHECK-NEXT:  %[[S0:.*]] = shard.sharding @grid_2d split_axes = {{\[\[}}]] : !shard.sharding289  // CHECK-NEXT:  %[[V0:.*]] = shard.shard %[[ARG]] to %[[S0]] annotate_for_users  : tensor<8x16xf32>290  // CHECK-NEXT:  %[[V1:.*]] = tosa.sigmoid %[[V0]]291  %0 = tosa.sigmoid %arg0 : (tensor<8x16xf32>) -> tensor<8x16xf32>292  // CHECK-NEXT:  %[[V2:.*]] = shard.shard %[[V1]] to %[[S0]] : tensor<8x16xf32>293  // CHECK-NEXT:  %[[V3:.*]] = shard.shard %[[V2]] to %[[S0]] annotate_for_users : tensor<8x16xf32>294  // CHECK-NEXT:  %[[V4:.*]] = tosa.sigmoid %[[V3]]295  %1 = tosa.sigmoid %0 : (tensor<8x16xf32>) -> tensor<8x16xf32>296  // CHECK-NEXT:  %[[V5:.*]] = shard.shard %[[V4]] to %[[S0]]  : tensor<8x16xf32>297  %s0 = shard.sharding @grid_2d split_axes = [[]] : !shard.sharding298  %2 = shard.shard %1 to %s0 : tensor<8x16xf32>299  // CHECK-NEXT:  return %[[V5]]300  return %2 : tensor<8x16xf32>301}302