brintos

brintos / llvm-project-archived public Read only

0
0
Text · 4.9 KiB · d8b8039 Raw
102 lines · plain
1// RUN: mlir-opt %s \2// RUN:   --pass-pipeline="builtin.module(transform-interpreter{ \3// RUN:        debug-bind-trailing-args=linalg.matmul,linalg.elementwise},\4// RUN:        canonicalize,cse,symbol-dce)" \5// RUN:   --split-input-file --verify-diagnostics6 7// ****************************** IMPORTANT NOTE ******************************8//9// If you are changing this file, you may also need to change10// mlir/docs/Tutorials/Transform accordingly.11//12// ****************************************************************************13 14module attributes {transform.with_named_sequence} {15  transform.named_sequence @__transform_main(16      %arg0: !transform.any_op,17      // expected-note @below {{handle to invalidated ops}}18      %arg1: !transform.op<"linalg.matmul">,19      %arg2: !transform.op<"linalg.elementwise">) {20    // The actual tiling transformation takes tile sizes as attributes.21    // expected-note @below {{invalidated by this transform op that consumes its operand #0 and invalidates all handles to payload IR entities associated with this operand and entities nested in them}}22    %tiled, %loop = transform.structured.tile_using_forall %arg1 tile_sizes [4, 32]23        : (!transform.op<"linalg.matmul">) -> (!transform.any_op, !transform.any_op)24 25    // This is trying to use an invalidated handle leading to undefined behavior.26    // expected-error @below {{uses a handle invalidated by a previously executed transform op}}27    transform.debug.emit_remark_at %arg1, "remark" : !transform.op<"linalg.matmul">28    transform.yield29  }30}31 32// Original function to optimize.33func.func @fc_relu(%lhs: tensor<512x512xf32>, %rhs: tensor<512x512xf32>,34                   %bias: tensor<512x512xf32>, %output: tensor<512x512xf32>)35                   -> tensor<512x512xf32> {36  // Matrix-matrix multiplication.37  // expected-note @below {{payload op}}38  %matmul = linalg.matmul ins(%lhs, %rhs: tensor<512x512xf32>, tensor<512x512xf32>)39                          outs(%output: tensor<512x512xf32>) -> tensor<512x512xf32>40 41  // Elementwise addition.42  %biased = linalg.elementwise kind=#linalg.elementwise_kind<add>43    ins(%matmul, %bias : tensor<512x512xf32>, tensor<512x512xf32>)44    outs(%output : tensor<512x512xf32>) -> tensor<512x512xf32>45 46  // Elementwise max with 0 (ReLU).47  %c0f = arith.constant dense<0.0> : tensor<512x512xf32>48  %relued = linalg.elementwise kind=#linalg.elementwise_kind<max_signed>49    ins(%biased, %c0f : tensor<512x512xf32>, tensor<512x512xf32>)50    outs(%output : tensor<512x512xf32>) -> tensor<512x512xf32>51  func.return %relued : tensor<512x512xf32>52}53 54// -----55 56module attributes {transform.with_named_sequence} {57  transform.named_sequence @__transform_main(58      %arg0: !transform.any_op,59      %arg1: !transform.op<"linalg.matmul">,60      %arg2: !transform.op<"linalg.elementwise">) {61    // We can cast one type to another as long as operations are compatible62    // with both types. This creates "aliasing" handles.63    // expected-note @below {{handle to invalidated ops}}64    %casted = transform.cast %arg1 : !transform.op<"linalg.matmul"> to65        !transform.any_op66 67    // The actual tiling transformation takes tile sizes as attributes.68    // expected-note @below {{invalidated by this transform op that consumes its operand #0 and invalidates all handles to payload IR entities associated with this operand and entities nested in them}}69    %tiled, %loop = transform.structured.tile_using_forall %arg1 tile_sizes [4, 32]70      : (!transform.op<"linalg.matmul">) -> (!transform.any_op, !transform.any_op)71 72    // Consuming an operand invalidates the consumed handle and any other handle that is73    // associated with the same payload operations, or payload operations nested in them.74    // expected-error @below {{uses a handle invalidated by a previously executed transform op}}75    transform.debug.emit_remark_at %casted, "remark"76      : !transform.any_op77    transform.yield78  }79}80 81// Original function to optimize.82func.func @fc_relu(%lhs: tensor<512x512xf32>, %rhs: tensor<512x512xf32>,83                   %bias: tensor<512x512xf32>, %output: tensor<512x512xf32>)84                   -> tensor<512x512xf32> {85  // Matrix-matrix multiplication.86  // expected-note @below {{payload op}}87  %matmul = linalg.matmul ins(%lhs, %rhs: tensor<512x512xf32>, tensor<512x512xf32>)88                          outs(%output: tensor<512x512xf32>) -> tensor<512x512xf32>89 90  // Elementwise addition.91  %biased = linalg.elementwise kind=#linalg.elementwise_kind<add>92    ins(%matmul, %bias : tensor<512x512xf32>, tensor<512x512xf32>)93    outs(%output : tensor<512x512xf32>) -> tensor<512x512xf32>94 95  // Elementwise max with 0 (ReLU).96  %c0f = arith.constant dense<0.0> : tensor<512x512xf32>97  %relued = linalg.elementwise kind=#linalg.elementwise_kind<max_signed>98    ins(%biased, %c0f : tensor<512x512xf32>, tensor<512x512xf32>)99    outs(%output : tensor<512x512xf32>) -> tensor<512x512xf32>100  func.return %relued : tensor<512x512xf32>101}102