xref: /aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/reduce_decomposer.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2022 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_REDUCE_DECOMPOSER_H_
17 #define TENSORFLOW_COMPILER_XLA_SERVICE_REDUCE_DECOMPOSER_H_
18 
19 #include <functional>
20 
21 #include "tensorflow/compiler/xla/service/hlo_pass_interface.h"
22 
23 namespace xla {
24 
25 // For each reduction R(I), ensures the postcondition:
26 //
27 //    !custom_layout_allowed(R)
28 //      =>
29 //    layout(R) == layout(I) # modulo removed dimensions
30 //
31 // To achieve that, decomposes layout-mutating reductions which do not satisfy
32 // `custom_layout_allowed` into a reduction and a copy.
33 //
34 // For a singular reduction:
35 //
36 //   -> reduce ->
37 //
38 // Gets turned into:
39 //
40 //    -> reduce -> copy ->
41 //
42 // For a variadic recuction, the layout assignment guarantees that the layout
43 // is the same for all outputs. This pass will transpose the variadic reduction
44 // inputs which have different physical layout to the first operand.
45 //
46 //   A{L} \
47 //   B{L} -> reduce{L'} ->
48 //   C{L} /
49 //
50 // Get turned into:
51 //
52 //   A{L} \                 / GTE(1) -> copy{L'} \
53 //   B{L} -> reduce{E(L)} --- GTE(2) -> copy{L'} - Tuple{L'}
54 //   C{L} /                 \ GTE(3) -> copy{L'} /
55 //
56 //   Where E(L) is expected layout of a reduction (original layout with reduce
57 //   dimensions dropped).
58 //
59 // PRECONDITION:
60 //  In variadic reduction, all outputs have the same layout
61 //  (enforced by layout assignment).
62 class ReduceDecomposer : public HloModulePass {
63  public:
64   explicit ReduceDecomposer(HloPredicate custom_layout_allowed = nullptr)
custom_layout_allowed_(custom_layout_allowed)65       : custom_layout_allowed_(custom_layout_allowed) {}
66 
name()67   absl::string_view name() const override { return "reduce-decomposer"; }
68 
69   using HloPassInterface::Run;
70   StatusOr<bool> Run(
71       HloModule* module,
72       const absl::flat_hash_set<absl::string_view>& execution_threads) override;
73 
74  private:
75   HloPredicate custom_layout_allowed_;
76 };
77 
78 }  // namespace xla
79 
80 #endif  // TENSORFLOW_COMPILER_XLA_SERVICE_REDUCE_DECOMPOSER_H_
81