Skip to content

@neoblizz neoblizz released this Jun 14, 2019 · 126 commits to master since this release

(documentation: https://gunrock.github.io/docs/)

Release v1.0 is a major API refactor release, along with some feature updates and new applications/primitives. An important note from this release is that moving forward, v1.x will be incompatible with v0.5.x Gunrock, some applications are still pending to be ported over to the new and improved API.

  • New operator interfaces
  • New graph representations
  • New frontier structure
  • New test driver
  • Restructured enactor routines
  • New parameter handling
  • New operators
  • Code restructuring
  • Optional Boost dependency

v1.0 Changelog

Operators (i.e. advance, filter)

  • Take in lambda functions for per-element operations, instead of static
    functions in a structure. <algo>_functor.cuh is merged into
    <algo>_enactor.cuh
  • Use OprtrParameters structure to keep inputs, except for the graph,
    input / output frontiers, and the lambdas
  • KernelPolicy is defined within each operator, instead of in the enactor
  • Templatized options (Idempotence, mark-preds, advance types, reduce ops,
    reduce types, etc.) are provided as a combined 32bit OprtrFlag
  • Queue index and selector are automatically changed by the operator when
    needed

Graph representation

  • A single structure encloses all graph related data
  • Different representations (CSR, CSC, COO, etc.) can be selected based
    on algorithmic needs
  • New graph representations could be added without changing other parts of
    Gunrock, except operator implementation that handles how to traverse such
    new representation
  • CPU, GPU and sub-graphs use the same graph data structure, no more
    GraphSlice and GRGraph

Frontier

  • A single structure gunrock/app/frontier.cuh:Frontier
    encloses all frontier related data

Test driver

  • Allows multiple graph types (64bit-VertexT, 64bit-SizeT,
    64bit-ValueT, directed vs. undirected) and multiple parameters
    combinations to run in a single execution
  • Allows result validation for each run, instead of only the last run
  • Result validation without reference for BFS and SSSP
  • Moved common functions into gunrock/app/test_base.cuh
  • Moved CPU reference code and result validation into
    gunrock/app/<algo>/<algo>_test.cuh

Enactor

  • Common functions moved into gunrock/app/enactor_base.cuh
  • Use OpenMP to maintain controlling threads on CPU
  • Use instances of Iteration instead of static access to its functions

Command line parameters

  • A dedicated Parameters struct to store all running parameters
  • Need to define parameters via. Use function before using them
  • Command line is parsed by get_opt
  • Set to set parameter values
  • Get to get parameter values
  • Handles vectors as parameter values

1D operators for Array1D

  • Per-element operations, e.g. ForAll and ForEach
  • Vector-Vector operations, e.g. Add, Minus, Mul, Div, Mad, Set
  • Vector-Scalar operations
  • Sort

Code restructuring

  • Partitioners moved from gunrock/app to gunrock/partitioner
  • LB operator moved from gunrock/oprtr/edge_map_partitioned_forward to
    gunrock/oprtr/LB_advance
  • TWC operator moved from 'gunrock/oprtr/edge_map_forward' to
    gunrock/oprtr/TWC_advance

Optional Boost dependency

  • Utility functions changed to C++11 or implemented
  • CPU references implemented for BFS and SSSP, and will be called when BOOST
    is not available
  • info will use RapidJson-based implementation, when Boost is not available

Known Issues

  • Multi-GPU framework not fully tested,
  • Operators have decreased performance, due to more than 32 registers used by
    a single thread in the kernels,
  • RGG and GRMAT generators not ported yet,
  • SSSP may have incorrect predecessors, due to data racing in marking the
    predecessors within the operator kernels
Assets 2
You can’t perform that action at this time.