e62681e70a
tensor.empty op elimination is an optimization that brings IR in a more bufferization-friendly form. E.g.:
```
%0 = tensor.empty()
%1 = linalg.fill(%cst, %0) {inplace = [true]}
%2 = tensor.insert_slice %1 into %t[10][20][1]
```
Is rewritten to:
```
%0 = tensor.extract_slice %t[10][20][1]
%1 = linalg.fill(%cst, %0) {inplace = [true]}
%2 = tensor.insert_slice %1 into %t[10][20][1]
```
This optimization used to operate on bufferization.alloc_tensor ops. This is not correct because the documentation of bufferization.alloc_tensor says that it always bufferizes to an allocation. Instead, this optimization should operate on tensor.empty ops, which can then be lowered to bufferization.alloc_tensor ops (if they don't get eliminated).
Differential Revision: https://reviews.llvm.org/D137162
50 lines
2.3 KiB
MLIR
50 lines
2.3 KiB
MLIR
// RUN: mlir-opt %s -eliminate-empty-tensors -empty-tensor-to-alloc-tensor -one-shot-bufferize="bufferize-function-boundaries test-analysis-only allow-return-allocs" -split-input-file | FileCheck %s
|
|
|
|
// CHECK-LABEL: func @buffer_forwarding_conflict
|
|
func.func @buffer_forwarding_conflict(%arg0: tensor<?xf32> {bufferization.writable = true}, %arg1: index) -> (tensor<?xf32>, tensor<?xf32>) {
|
|
%cst = arith.constant 0.000000e+00 : f32
|
|
// CHECK: tensor.extract_slice
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["false", "none"]
|
|
// Instead of allocating, share buffer with some inplace bufferization?
|
|
%0 = tensor.empty(%arg1) : tensor<?xf32>
|
|
|
|
// CHECK: linalg.fill
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["none", "true"]
|
|
%1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?xf32>) -> tensor<?xf32>
|
|
|
|
// CHECK: tensor.insert_slice
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["true", "false", "none"]
|
|
%2 = tensor.insert_slice %1 into %arg0[0] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
|
|
|
|
// CHECK: tensor.insert_slice
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["true", "true", "none"]
|
|
%3 = tensor.insert_slice %1 into %arg0[42] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
|
|
|
|
// CHECK: return
|
|
// CHECK-SAME: __equivalent_func_args__ = [-1, 0]
|
|
return %2, %3 : tensor<?xf32>, tensor<?xf32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func @buffer_forwarding_no_conflict
|
|
func.func @buffer_forwarding_no_conflict(%arg0: tensor<?xf32> {bufferization.writable = true}, %arg1: index) -> (tensor<?xf32>, tensor<?xf32>) {
|
|
%cst = arith.constant 0.000000e+00 : f32
|
|
// CHECK: tensor.extract_slice
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["true", "none"]
|
|
// Instead of allocating, share buffer with some inplace bufferization?
|
|
%0 = tensor.empty(%arg1) : tensor<?xf32>
|
|
|
|
// CHECK: linalg.fill
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["none", "true"]
|
|
%1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?xf32>) -> tensor<?xf32>
|
|
|
|
// CHECK: tensor.insert_slice
|
|
// CHECK-SAME: {__inplace_operands_attr__ = ["true", "true", "none"]
|
|
%2 = tensor.insert_slice %1 into %arg0[42] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
|
|
|
|
// CHECK: return
|
|
// CHECK-SAME: __equivalent_func_args__ = [0, 0]
|
|
return %2, %2 : tensor<?xf32>, tensor<?xf32>
|
|
}
|