Package org.apache.sysds.hops.ipa

package org.apache.sysds.hops.ipa
  • Classes
    Class
    Description
    FunctionCallGraph
     
    FunctionCallSizeInfo
    Auxiliary data structure to hold function call summaries in terms of information about number of function calls, consistent dimensions, consistent sparsity, and dimension-preserving functions.
    InterProceduralAnalysis
    This Inter Procedural Analysis (IPA) serves two major purposes: 1) Inter-Procedure Analysis: propagate statistics from calling program into functions and back into main program.
    IPAPass
    Base class for all IPA passes.
    IPAPassApplyStaticAndDynamicHopRewrites
    This rewrite applies static hop dag and statement block rewrites such as constant folding and branch removal in order to simplify statistic propagation.
    IPAPassCompressionWorkloadAnalysis
    This rewrite obtains workload summaries for all hops candidates amenable for compression as a basis for workload-aware compression planning.
    IPAPassEliminateDeadCode
    This rewrite eliminates unnecessary sub-DAGs that produce transient write outputs which are never consumed subsequently.
    IPAPassFlagFunctionsRecompileOnce
    This rewrite marks functions with loops as recompile once in order to reduce recompilation overhead.
    IPAPassFlagLoopsRecompileOnce
    This rewrite marks loops in the main program as recompile once in order to reduce recompilation overhead.
    IPAPassFlagNonDeterminism
     
    IPAPassForwardFunctionCalls
    This rewrite forwards a function call to a function with a simple function call that only consumes function parameters and literals into the original call location.
    IPAPassInjectOOCTee
    Applies OOC tee injection after static/dynamic rewrites in IPA.
    IPAPassInlineFunctions
    This rewrite inlines single statement block functions, which have fewer operations than an internal threshold.
    IPAPassPropagateReplaceLiterals
    This rewrite propagates and replaces literals into functions in order to enable subsequent rewrites such as branch removal.
    IPAPassPruneUnreachableHops
    Prune stale parent links by keeping only parent references reachable from the statement block roots/predicates.
    IPAPassRemoveConstantBinaryOps
    This rewrite identifies binary operations with constant matrices such as X * ones, where ones might be created as a vector of ones before a loop.
    IPAPassRemoveUnnecessaryCheckpoints
    This rewrite identifies and removes unnecessary checkpoints, i.e., persisting of Spark RDDs into a given storage level.
    IPAPassRemoveUnusedFunctions
    This rewrite identifies and removes unused functions in order to reduce compilation overhead and other overheads such as parfor worker creation, where we construct function copies.
    IPAPassReplaceEvalFunctionCalls
    This rewrite applies static hop dag and statement block rewrites such as constant folding and branch removal in order to simplify statistic propagation.
    IPAPassRewriteFederatedPlan
    This rewrite generates a federated execution plan by estimating and setting costs and the FederatedOutput values of all relevant hops in the DML program.