xref: /aosp_15_r20/external/tensorflow/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #ifndef TENSORFLOW_COMPILER_MLIR_TENSORFLOW_UTILS_XLA_SHARDING_UTIL_H_
17 #define TENSORFLOW_COMPILER_MLIR_TENSORFLOW_UTILS_XLA_SHARDING_UTIL_H_
18 
19 #include "absl/strings/string_view.h"
20 #include "llvm/ADT/ArrayRef.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "mlir/IR/Builders.h"  // from @llvm-project
23 #include "mlir/IR/Operation.h"  // from @llvm-project
24 #include "mlir/IR/Types.h"  // from @llvm-project
25 #include "mlir/IR/Value.h"  // from @llvm-project
26 #include "mlir/Support/LogicalResult.h"  // from @llvm-project
27 #include "tensorflow/compiler/mlir/tensorflow/ir/tf_device.h"
28 #include "tensorflow/compiler/xla/xla_data.pb.h"
29 #include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h"
30 
31 namespace tensorflow {
32 
33 inline constexpr absl::string_view kInputShardingAttr =
34     "input_sharding_configuration";
35 inline constexpr absl::string_view kOutputShardingAttr =
36     "output_sharding_configuration";
37 
38 // Parses "input_sharding_configuration" attribute and returns a list where i-th
39 // element is a list of mlir::Value's which represent inputs for the TPU
40 // computation correponding to i-th logical device. If the attribute does not
41 // exist, the all inputs are placed on logical core 0.
42 mlir::LogicalResult ExtractInputsForLogicalDevices(
43     const int num_cores_per_replica,
44     mlir::tf_device::ClusterFuncOp cluster_func, mlir::OpBuilder* builder,
45     llvm::SmallVectorImpl<llvm::SmallVector<mlir::Value, 4>>* input_list);
46 
47 // Extracts a list of OpSharding that represent output sharding configuration of
48 // `tf_device.cluster`.
49 mlir::LogicalResult ParseAndValidateOutputSharding(
50     const int num_cores_per_replica,
51     mlir::tf_device::ClusterFuncOp cluster_func,
52     mlir::SmallVector<xla::OpSharding, 4>* output_sharding_list);
53 
54 // Retrieves output types for TPUExecute op representing execution for provided
55 // logical device id. TPUExecute op for different logical device may have
56 // different outputs depending on the output sharding configuration.
57 mlir::LogicalResult GetOutputTypesForLogicalDeviceComputation(
58     const int core_id, llvm::ArrayRef<xla::OpSharding> output_sharding_config,
59     mlir::tf_device::ClusterFuncOp cluster_func,
60     llvm::SmallVectorImpl<mlir::Type>* output_types,
61     llvm::SmallVectorImpl<int>* cluster_to_core_index);
62 
63 // Remaps outputs of `new_parallel_execute` op that represent concurrent
64 // execution of the `tf_device.cluster_func` at index `cluster_idx` of
65 // `old_parallel_execute` with its users.
66 mlir::LogicalResult RemapOutputsFromLogicalDevices(
67     const mlir::Location& location,
68     llvm::ArrayRef<xla::OpSharding> output_sharding_config,
69     llvm::SmallVector<llvm::SmallVector<int, 4>, 4> cluster_to_core_index,
70     mlir::tf_device::ParallelExecuteOp old_parallel_execute, int cluster_idx,
71     mlir::tf_device::ParallelExecuteOp new_parallel_execute,
72     mlir::OpBuilder* builder);
73 
74 // Determines each logical core argument to metadata argument index mapping,
75 // based on sharding. The return value is indexed first by logical core then by
76 // argument index.
77 llvm::SmallVector<llvm::SmallVector<int64_t, 4>, 4> GetMetadataArgumentMapping(
78     const tpu::TPUCompileMetadataProto& metadata);
79 
80 }  // namespace tensorflow
81 
82 #endif  // TENSORFLOW_COMPILER_MLIR_TENSORFLOW_UTILS_XLA_SHARDING_UTIL_H_
83