Home | History | Annotate | Download | only in service
      1 /* Copyright 2017 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_COPY_INSERTION_H_
     17 #define TENSORFLOW_COMPILER_XLA_SERVICE_COPY_INSERTION_H_
     18 
     19 #include "tensorflow/compiler/xla/service/buffer_liveness.h"
     20 #include "tensorflow/compiler/xla/service/hlo_computation.h"
     21 #include "tensorflow/compiler/xla/service/hlo_instruction.h"
     22 #include "tensorflow/compiler/xla/service/hlo_module.h"
     23 #include "tensorflow/compiler/xla/service/hlo_pass_interface.h"
     24 
     25 namespace xla {
     26 
     27 // Copy insertion is a legalization HLO pass which inserts copies (kCopy
     28 // instructions) to eliminate several kinds of problems in the HLO module.
     29 //
     30 //   (1) Entry parameter or a constant live out of the entry computation.  Entry
     31 //       computation arguments and constants have different lifetimes than the
     32 //       computation result and cannot share the same allocation. Parameters and
     33 //       constants live out of non-entry computations do not need copies.
     34 //
     35 //   (2) Different values which are simultaneously live and which must be held
     36 //       in the same buffer. This can occur in while bodies. Specifically, the
     37 //       while loop state (the arguments to the while instruction) is updated
     38 //       in-place and the update may clobber the value from the previous
     39 //       iteration before the previous value is dead. Computations called from
     40 //       kCall instructions do not need such copies because kCall has no update
     41 //       in-place semantics.
     42 //
     43 //   (3) The buffer set of the root instruction of the entry computation must be
     44 //       unambiguous and distinct. That is, InstructionAliasSet::IsAmbiguous and
     45 //       InstructionAliasSet::IsDistinct return true.
     46 class CopyInsertion : public HloModulePass {
     47  public:
     48   absl::string_view name() const override { return "copy-insertion"; }
     49 
     50   // fusion_can_share_buffer: backend specific function that decides whether a
     51   // fusion can share buffer with its operand.
     52   //
     53   // TODO(b/80315712): Find a better way to tell whether a fusion can share
     54   // buffer.
     55   CopyInsertion(const HloDataflowAnalysis::FusionCanShareBufferFunction&
     56                     fusion_can_share_buffer = nullptr)
     57       : fusion_can_share_buffer_(fusion_can_share_buffer) {}
     58 
     59   // Run the pass on the given module. Returns whether the module was changed
     60   // (copies were inserted).
     61   StatusOr<bool> Run(HloModule* module) override;
     62 
     63   // The CPU and GPU backend need additional copies added due to deficiencies in
     64   // buffer assignment. Specifically, copies are needed for constants live-out
     65   // of computations, and for values which are live-in and live-out of the same
     66   // computation. These copies are needed because buffer-assignment uses a
     67   // computation-scoped analyis (TuplePointsToAnalysis) and has limited
     68   // visibility across computation boundaries. This method adds these necessary
     69   // copies. Returns whether the module was modified.
     70   //
     71   // TODO(b/62548313): Remove this when buffer assignment is module-scoped.
     72   static StatusOr<bool> AddCopiesForBufferAssignment(HloModule* module);
     73 
     74   // Try to remove as many copies from the module as possible without
     75   // introducing live range interference. Only copy instructions that are
     76   // eligible for copy elision are considered for removal.
     77   Status RemoveUnnecessaryCopies(const HloOrdering& ordering,
     78                                  HloModule* module);
     79 
     80   // Add copies to address special constraints on the roots of computations not
     81   // related to live range interference:
     82   //
     83   //    (1) Entry computation root must be unambiguous and distinct.
     84   //
     85   //    (2) Any computation called by a kCall instruction must have an
     86   //        unambiguous root.
     87   //
     88   //    (3) Constants and parameters cannot be live out of the entry computation
     89   //
     90   Status AddSpecialCaseCopies(HloModule* module);
     91 
     92   // Verifies that no HLO values have interfering live ranges using the given
     93   // ordering.
     94   Status VerifyNoLiveRangeInterference(const HloOrdering& ordering,
     95                                        HloModule* module);
     96 
     97  protected:
     98   // Override which requires the caller to pass in a call graph.
     99   virtual Status AddSpecialCaseCopies(const CallGraph& call_graph,
    100                                       HloModule* module);
    101 
    102  private:
    103   Status AddCopiesToResolveInterference(HloModule* module);
    104 
    105   // Backend specific function that decides whether a fusion can share buffer
    106   // with its operand.
    107   HloDataflowAnalysis::FusionCanShareBufferFunction fusion_can_share_buffer_;
    108 };
    109 
    110 }  // namespace xla
    111 
    112 #endif  // TENSORFLOW_COMPILER_XLA_SERVICE_COPY_INSERTION_H_
    113