experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
STT-tensorflow/tensorflow/compiler/xla/service/interpreter
History
Tim Shen b0e5853cf0 [XLA] Pack some parameters in xla::Compiler member functions into a struct. This is for later expanding for more parameters, like a thread pool for parallel compilation. 
PiperOrigin-RevId: 346276092
Change-Id: Ic0028cd5fb366f5d6441d1913398ae5b76f6df37
2020-12-08 01:54:36 -08:00
..
BUILD
Remove dependencies on aliases in tensorflow/core/BUILD
2020-10-11 21:03:09 -07:00
compiler.cc
[XLA] Pack some parameters in xla::Compiler member functions into a struct. This is for later expanding for more parameters, like a thread pool for parallel compilation.
2020-12-08 01:54:36 -08:00
compiler.h
[XLA] Pack some parameters in xla::Compiler member functions into a struct. This is for later expanding for more parameters, like a thread pool for parallel compilation.
2020-12-08 01:54:36 -08:00
executable_base.cc
Remove platform field from shaped buffer.
2020-10-30 14:26:55 -07:00
executable_base.h
Pass in RunOptions when Evaluating an Interpreter executable.
2020-08-17 12:01:09 -07:00
executable.cc
Pass in RunOptions when Evaluating an Interpreter executable.
2020-08-17 12:01:09 -07:00
executable.h
Pass in RunOptions when Evaluating an Interpreter executable.
2020-08-17 12:01:09 -07:00
executor.cc
Allocate input/output buffers at the correct memory space.
2020-06-22 17:31:14 -07:00
executor.h
Remove SharedMemoryConfig since it is not used anywhere.
2020-08-11 20:01:04 -07:00
interpreter_transfer_manager.cc
…
interpreter_transfer_manager.h
…
platform_id.cc
…
platform_id.h
…
platform.cc
…
platform.h
…
README.md
…

README.md

XLA Interpreter Backend

The XLA Interpreter backend operates at HLO-level by ingesting a HloModule and evaluating the result of the HLO graph directly with HloEvaluator, without lowering it further (to LLVM IR for example) before execution as other backends (CPU and GPU for example) do.

Its key components are:

  • [InterpreterCompiler] despite the inherited naming of "compiler", all InterpreterCompiler really does is the following:
    1. Runs certain HLO optimization passes on the given HLO graph.
    2. Generates an InterpreterExecutable from the optimized HLO graph.
    3. Registers itself in the global compiler factory registry.
  • [InterpreterExecutable]: responsible for running input HLO graph through the HloEvaluator, allocating output buffer and finally copying evaluated Literal result over.
  • [HloEvaluator]: traverses a HLO graph and evaluates each node in DFS ordering along the way.