xref: /aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/gpu/all_reduce_blueconnect.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2021 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_XLA_SERVICE_GPU_ALL_REDUCE_BLUECONNECT_H_
17 #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_ALL_REDUCE_BLUECONNECT_H_
18 
19 #include "tensorflow/compiler/xla/service/hlo_module.h"
20 #include "tensorflow/compiler/xla/service/hlo_pass_interface.h"
21 #include "tensorflow/compiler/xla/statusor.h"
22 
23 namespace xla {
24 
25 // Decomposes all-reduce operations using the BlueConnect algorithm.
26 //
27 // Paper: "BLUECONNECT: DECOMPOSING ALL-REDUCE FOR DEEP LEARNING ON
28 // HETEROGENEOUS NETWORK HIERARCHY"
29 // https://mlsys.org/Conferences/2019/doc/2019/130.pdf
30 //
31 // This algorithm attempts to minimize the number of levels of network hierarchy
32 // traversed for as much data transfer as possible. This implementation assumes
33 // that host IDs are ordered corresponding to network hierarchy.
34 class AllReduceBlueConnect : public HloModulePass {
35  public:
AllReduceBlueConnect(size_t num_devices_per_host)36   explicit AllReduceBlueConnect(size_t num_devices_per_host)
37       : num_devices_per_host_(num_devices_per_host) {}
38 
name()39   absl::string_view name() const override { return "all-reduce-blueconnect"; }
40 
41   using HloPassInterface::Run;
42   StatusOr<bool> Run(
43       HloModule* module,
44       const absl::flat_hash_set<absl::string_view>& execution_threads) override;
45 
46  private:
47   size_t num_devices_per_host_;
48 };
49 
50 }  // namespace xla
51 
52 #endif  // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_ALL_REDUCE_BLUECONNECT_H_
53