Skip to content

Latest commit

 

History

History
67 lines (46 loc) · 3.26 KB

TRANSITIONGUIDE.md

File metadata and controls

67 lines (46 loc) · 3.26 KB
Raw

0.12

Python API

Loading an edge list

Renumbering is now enabled by default in from_cudf_edgelist. The renumbering feature allows us to hide the fact that we need vertices to be integers starting at 0. The auto-renumbering feature converts the data into the proper data type required by the underlying implementation. All algorithms accepting vertex identifiers (like the souce vertex for SSSP) now automatically accept user's notation too. On output, it will transparently un-renumber results, basically convert the internal IDs back.

C++ API

Pagerank, BFS, and SSSP have dropped the gdf_column dependency in favor of basic types and templates

Example :

// 0.11 API 
cugraph::pagerank(cugraph::Graph, gdf_column *pagerank, ...) 
// 0.12 API 
cugraph::pagerank<int,float>(cugraph::Graph, float *pagerank ...)

0.11

Python API

This release introduces new concepts in the API and improves user experience through more automation and better DataFrame support. Python users are encouraged to review these changes and potentially upgrade their code before using 0.11 version.

New graph types

The Python API now has Graph (undirected) and DiGraph (directed) types. This is closer to NetworkX's API.

In the past, directed graphs were stored using the Graph type. Starting in 0.11, DiGraph should be used for directed graphs instead. Graph only refers to undirected graphs now.

The Multi(Di)Graph types were added and more support for this new structure will be added in the next releases (more details in issue #604).

Undirected graphs

cuGraph currently automatically "symmetrize" undirected inputs: each undirected edge (u,v) is stored as two directed edges (u,v) and (v,u).

When viewing the graph or requesting the number of edges, cuGraph will currently return this symmetrized view. This is an implementation detail that will be hidden to the user in 0.12 (more details in issue #603).

Loading an edge list

Users are encouraged to use from_cudf_edgelist instead of add_edge_list.

This new API supports cuDF DataFrame. Users can now ask for an automatic renumbering of the edge list at the time it is loaded. In this case, all analytics outputs are automatically un-renumbered before being returned.

C++ API

This release is the first step toward converting the former C-like API into a C++ API. Major changes have been made, C++ users should review these changes and upgrade their code before using 0.11 version.

  • All C++ API functions are now encapsulated in namespace cugraph
  • All functions are now processing errors through exceptions. As a result, gdf_error is no longer used as a return type and void is used instead.
  • gdf_graph is now Graph. The content of the structure is still the same (will be upgraded in future releases).
  • gdf_ prefix has been removed from all C++ API functions.

Example :

// < 0.11 API 
gdf_error err = gdf_pagerank(<gdf_graph>, ...) 
// >= 0.11 API 
cugraph::pagerank(<cugraph::Graph>, ...)

The C++ API provides functions that efficiently convert between data formats and access to the efficient CUDA algorithms. In 0.11, all automatic conversions and decision making were removed from the C++ layer.

Directory Structure and File Naming

The c_ prefix of all .pxd files has been removed.