Member tensorflow::tfcompile::Main (const MainFlags &flags)

3. Generate output (object, header, etc)

Member tensorflow::XlaCompiler::CompileGraph (const CompileOptions &options, string const &name, std::unique_ptr< Graph > graph, const std::vector< Argument > &args, CompilationResult *result)

Rest of the code

Class xla::AlgebraicSimplifier

Contains lots of code including broadcast sementics

Member xla::anonymous_namespace{dot_decomposer.cc}::DecomposeBatchDot (HloInstruction *dot)

Opinion: It seems like decompose high-rank dot operation to low-rank. Is based on mathematical approach, and can be applied for computer as well.

Class xla::BatchNormExpander

See how the function work

Member xla::cpu::CpuCompiler::RunHloPasses (HloModule *module, bool is_aot_compile)

See what those invariant checker and pass do

unknown argument

Class xla::cpu::CpuInstructionFusion

See what this pass do

Member xla::cpu::IrEmitter::EmitComputation (HloComputation *computation, const string &function_name_prefix, bool is_top_level_computation, std::vector< const HloInstruction *> *instruction_order)

See what it does

Member xla::CreateMemoryMinimizingSequence (const HloModule &module, const LogicalBuffer::SizeFunction &size_function, const MemorySchedulerAlgorithm &algorithm)

Trace who decide who should be fuse

Class xla::HloVerifier

Make sure what it check. For now it seems like it checks for HLO shape

Member xla::TuplePointsToAnalysis::Run (const HloModule *module)

Too deep. Abort.