1. Introduction

Data analytics is a growing application of high-performance computing. Many advanced data analytics problems can be couched as graph problems. In turn, many of the common graph problems today can be couched as sparse linear algebra. This is the motivation for nvGRAPH, new in NVIDIA® CUDA™ 8.0, which harnesses the power of GPUs for linear algebra to handle the largest graph analytics and big data analytics problems.

This release provides graph construction and manipulation primitives, and a set of useful graph algorithms optimized for the GPU. The core functionality is a SPMV (sparse matrix vector product) using a semi-ring model with automatic load balancing for any sparsity pattern. For more information on semi-rings and their uses, we recommend the book "Graph Algorithms in the Language of Linear Algebra", by Jeremy Kepner and John Gilbert.

To use nvGRAPH you should be sure the nvGRAPH library is in the environment (PATH on Windows, LD_LIBRARY_PATH on Linux), "#include nvgraph.h" to your source files referencing the nvGRAPH API, and link your code using -lnvgraph on the command line, or add libnvgraph to your library dependencies. We have tested nvGRAPH using GCC 4.8 and higher on Linux, Visual Studio 2012 and Visual Studio 2014 on Windows.

A typical workflow for using nvGRAPH is to begin by calling nvgraphCreate() to initialize the library. Next the user can proceed to upload graph data to the library through nvGRAPH's API; if there is already a graph loaded in device memory then you just need a pointer to the data arrays for the graph. Graphs may be uploaded using the CSR (compressed sparse row) format and the CSC(compressed column storage) format, using nvgraphCreateGraphDescr(). This creates an opaque handle to the graph object, called the "graph descriptor", which represents the graph topology and its data. Graph data can be attached to vertices and/or edges of the graph, using nvgraphSetVertexData() and nvgraphSetEdgeData() respectively. Multiple values for data can co-exist on each edge or vertex at the same time, each is accessed by an index into the array of data sets. Then the user can execute graph algorithms on the data, extract a subgraph from the data, or reformat the data using the nvGRAPH API. The user can download the results back to host, or copy them to another location on the device, and once all calculations are done the user should call nvgraphDestroy() to free resources used by nvGRAPH.

nvGRAPH depends on features only present in CUDA capability 3.0 and higher architectures. This means that nvGRAPH will only run on Kepler generation or newer cards. This choice was made to provide the best performance possible.

We recommend the user start by inspecting the example codes provided, and adapt from there for their own use.

2. nvGRAPH API Reference

This chapter specifies the behavior of the nvGRAPH library functions by describing their input/output parameters, data types, and error codes.

2.1. Return value nvgraphStatus_t

All nvGRAPH Library return values except for NVGRAPH_STATUS_SUCCESS indicate that the current API call failed and the user should reconfigure to correct the problem. The possible return values are defined as follows:

Return Values
NVGRAPH_STATUS_SUCCESS nvGRAPH operation was successful
NVGRAPH_STATUS_NOT_INITIALIZED

The nvGRAPH library was not initialized. This is usually caused by the lack of a prior call, an error in the CUDA Runtime API called by the nvGRAPH routine, or an error in the hardware setup.

To correct: call nvgraphCreate() prior to the function call; and check that the hardware, an appropriate version of the driver, and the nvGRAPH library are correctly installed.

NVGRAPH_STATUS_ALLOC_FAILED

Resource allocation failed inside the nvGRAPH library. This is usually caused by a cudaMalloc() failure.

NVGRAPH_STATUS_INVALID_VALUE

An unsupported value or parameter was passed to the function

To correct: ensure that all the parameters being passed have valid values.

NVGRAPH_STATUS_ARCH_MISMATCH

The function requires a feature absent from the device architecture.

To correct: compile and run the application on a device with appropriate compute capability.

NVGRAPH_STATUS_MAPPING_ERROR

An access to GPU memory space failed.

NVGRAPH_STATUS_EXECUTION_FAILED

The GPU program failed to execute. This is often caused by a launch failure of the kernel on the GPU, which can be caused by multiple reasons.

To correct: check that the hardware, an appropriate version of the driver, and the nvGRAPH library are correctly installed.

NVGRAPH_STATUS_INTERNAL_ERROR

An internal nvGRAPH operation failed.

To correct: check that the hardware, an appropriate version of the driver, and the nvGRAPH library are correctly installed. Also, check that the memory passed as a parameter to the routine is not being deallocated prior to the routine’s completion.

NVGRAPH_STATUS_TYPE_NOT_SUPPORTED

The type is not supported by this function. This is usually caused by passing an invalid graph descriptor to the function.

NVGRAPH_STATUS_NOT_CONVERGED

An algorithm failed to converge.

To correct: ensure that all the parameters being passed have valid values for this algorithm, increase the maximum number of iteration and/or the tolerance.

2.2. nvGRAPH graph topology types

nvGRAPH separates the topology (connectivity) of a graph from the values. To make specifying a topology easier, nvGRAPH supports three topology types. Each topology type defines its own storage format, which have benefits for some operations but detriments for others. Some algorithms can work only with specific topology types, see the algorithms descriptions for the list of supported topologies.

typedef enum
{
   NVGRAPH_CSR_32 = 0,
   NVGRAPH_CSC_32 = 1,
   NVGRAPH_COO_32 = 2
} nvgraphTopologyType_t;
Topology types
NVGRAPH_CSR_32 Compressed Sparse Row format (row major format). Used in SrSPMV algorithm. Use nvgraphCSRTopology32I_t topology structure for this format.
NVGRAPH_CSC_32 Compressed Sparse Column format (column major format). Used in SSSP, WidestPath and Pagerank algorithms. Use nvgraphCSCTopology32I_t topology structure for this format.
NVGRAPH_COO_32 Coordinate list format with source or destination major. Not used in any algorithm and provided for data storage only. Use nvgraphCOOTopology32I_t topology structure for this format.

nvGRAPH topology structure types

Graph topology structures are used to set or retrieve topology data. Users should use the structure that corresponds to the chosen topology type.

nvgraphCSRTopology32I_t

Used for NVGRAPH_CSR_32 topology type

struct nvgraphCSRTopology32I_st {
  int nvertices;
  int nedges;
  int *source_offsets;
  int *destination_indices;
};
typedef struct nvgraphCSRTopology32I_st *nvgraphCSRTopology32I_t;
Structure fields
nvertices Number of vertices in the graph.
nedges Number of edges in the graph.
source_offsets Array of size nvertices+1, where i element equals to the number of the first edge for this vertex in the list of all outgoing edges in the destination_indices array. Last element stores total number of edges
destination_indices Array of size nedges, where each value designates destanation vertex for an edge.

nvgraphCSCTopology32I_t

Used for NVGRAPH_CSC_32 topology type

struct nvgraphCSCTopology32I_st {
  int nvertices;
  int nedges;
  int *destination_offsets;
  int *source_indices;
};
typedef struct nvgraphCSCTopology32I_st *nvgraphCSCTopology32I_t;
Structure fields
nvertices Number of vertices in the graph.
nedges Number of edges in the graph.
destination_offsets Array of size nvertices+1, where i element equals to the number of the first edge for this vertex in the list of all incoming edges in the source_indices array. Last element stores total number of edges
source_indices Array of size nedges, where each value designates source vertex for an edge.

nvgraphCOOTopology32I_t

Used for NVGRAPH_COO_32 topology type

struct nvgraphCOOTopology32I_st {
  int nvertices;
  int nedges;
  int *source_indices;
  int *destination_indices;
  nvgraphTag_t tag;
};
typedef struct nvgraphCOOTopology32I_st *nvgraphCOOTopology32I_t;
Structure fields
nvertices Number of vertices in the graph.
nedges Number of edges in the graph.
source_indices Array of size nedges, where each value designates the source vertex for an edge.
destination_indices Array of size nedges, where each value designates the destination vertex for an edge.
tag One of the values of NVGRAPH_UNSORTED, NVGRAPH_SORT_BY_SOURCE or NVGRAPH_SORT_BY_DESTINATION to indicate topology order.

2.4. Function nvgraphGetProperty()

nvgraphStatus_t
   nvgraphGetProperty(libraryPropertyType type, int *value);
      

Returns property value of a library, such as version number.

Input
type Identifier of a library property, such as MAJOR_VERSION, MINOR_VERSION or PATCH_LEVEL
Output
value Value of a requested property

Return Values
NVGRAPH_STATUS_SUCCESS nvGRAPH successfully created the handle.
NVGRAPH_STATUS_INVALID_VALUE Invalid property requested.

2.5. Function nvgraphCreate()

nvgraphStatus_t
   nvgraphCreate(nvgraphHandle_t *handle);
      

Creates only an opaque handle, and allocates small data structures on the host. This handle is used in all of the nvGRAPH functions, so this function should be called first, before any other calls are made to the library.

Input/Output
handle Pointer to a nvgraphHandle_t object
Return Values
NVGRAPH_STATUS_SUCCESS nvGRAPH successfully created the handle.
NVGRAPH_STATUS_ALLOC_FAILED The allocation of resources for the handle failed.
NVGRAPH_STATUS_INTERNAL_ERROR An internal driver error was detected.

2.6. Function nvgraphDestroy()

nvgraphStatus_t
   nvgraphDestroy(nvgraphHandle_t handle);
      

Destroys a handle created with nvgraphCreate(). This will automatically release any allocated memory objects created with this handle, for example any graphs and their vertices' and edges' data. Any subsequent usage of this handle after calling nvgraphDestroy() will be invalid. Any calls to the nvGRAPH API after nvgraphDestroy() is called will return 'NVGRAPH_UNINITIALIZED' errors.

Input
handle The nvgraphHandle_t object of the handle to be destroyed.
Return Values
NVGRAPH_STATUS_SUCCESS nvGRAPH successfully destroyed the handle.
NVGRAPH_STATUS_INVALID_VALUE The handle parameter is not a valid handle.

2.7. Function nvgraphCreateGraphDescr()

nvgraphStatus_t
   nvgraphCreateGraphDescr(nvgraphHandle_t handle, nvgraphGraphDescr_t *descrG);
      

Creates opaque handle for a graph structure. This handle is required for any operation on the graph.

Input
handle nvGRAPH library handle
Input/Output
descrG Pointer to the empty nvgraphGraphDescr_t structure object.
Return Values
NVGRAPH_STATUS_SUCCESS Success
NVGRAPH_STATUS_INVALID_VALUE Bad library handle is provided
NVGRAPH_STATUS_ALLOC_FAILED Cannot allocate graph descriptor.

2.8. Function nvgraphDestroyGraphDescr()

nvgraphStatus_t
   nvgraphDestroyGraphDescr(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG);
      

Destroys a graph handle created with nvgraphCreateGraphDescr(). This won't release any memory allocated for this graph until the nvGRAPH library handle is destroyed. Calls to manipulate destroyed graphs will return NVGRAPH_STATUS_INVALID_VALUE.

Input
handle nvGRAPH library handle
descrG Graph descriptor to be released
Return Values
NVGRAPH_STATUS_SUCCESS Successful release of the graph descriptor
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Graph is stored with unknown type of data
NVGRAPH_STATUS_INVALID_VALUE Invalid library handle or graph descriptor handle

2.9. Function nvgraphSetGraphStructure()

nvgraphStatus_t
   nvgraphSetGraphStructure( nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      void* topologyData, nvgraphTopologyType_t TType);
      

This call sets both topology data and topology type for the given graph descriptor. Graph topology should be set only once. Users should choose one of the supported topologies, fill in the corresponding structure for the graph structure initialization and provide a pointer to this structure. The topology data and type are given in parameters topologyData and TType. Typically graph topology data includes a number of vertices, number of edges and connectivity information. Look at the description of the corresponding topology structures for details.

Input
handle nvGRAPH library handle
topologyData Pointer to a filled structure of one of the types {nvgraphCSRTopology32I_t, nvgraphCSCTopology32I_t}. The particular type to be used is defined by parameter TType.
TType Graph topology type. This value should be equal to one of the possible values of the enum nvgraphTopologyType_t. This defines what data structure should be provided by the topologyData parameter.
Input/Output
descrG Graph descriptor. Must not have topology previously defined.
Return Values
NVGRAPH_STATUS_SUCCESS Success
NVGRAPH_STATUS_INVALID_VALUE Invalid library handle, topology data structure pointer or topology values, or graph topology was already set
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Provided topology type is not supported
NVGRAPH_STATUS_INTERNAL_ERROR Unknown internal error was caught

2.10. Function nvgraphGetGraphStructure()

nvgraphStatus_t
   nvgraphGetGraphStructure( nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      void* topologyData, nvgraphTopologyType_t* TType);
      

This function allows users to retrieve a given graph's topology and topology data such as the number of vertices and edges in the graph. Users must provide a graph descriptor as well as an empty topology structure, where this information will be stored.

Input
handle nvGRAPH library handle
descrG Graph descriptor. This graph should have its topology structure previously set.
Input/Output
topologyData Pointer to a structure of one of the types {nvgraphCSRTopology32I_t, nvgraphCSCTopology32I_t} matching the graph topology. If the field is NULL it will be ignored and only graph topology will be returned. The user can point source and destination fields in the structure to a host or device memory to retrive connectivity information, if one of the fields is NULL it will be ignored.
TType Pointer to nvgraphTopologyType_t where graph topology will be returned. If the field is NULL it will be ignored and only topology data will be returned.
Return Values
NVGRAPH_STATUS_SUCCESS Success
NVGRAPH_STATUS_INVALID_VALUE Invalid library handle, graph descriptor or topology for the graph is not set.
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Unsupported topology or graph's topology doesn't match provided parameter.

2.11. Function nvgraphConvertTopology()

nvgraphStatus_t
   nvgraphConvertTopology(nvgraphHandle_t handle, 
      nvgraphTopologyType_t srcTType, void *srcTopology, void *srcEdgeData,
      cudaDataType_t *dataType,
      nvgraphTopologyType_t dstTType, void *dstTopology, void *dstEdgeData);
      

Convert one of the supported topologys to another. In case the source and destination topologies are the same, the function will perform a straight memory copy.

This function assumes that source and destination arrays within the topologies and edge data reside in GPU (device) memory. It is the user's responsibility to allocate memory and copy data between GPU memory and CPU memory using standard CUDA runtime API routines, such as cudaMalloc(), cudaFree(), cudaMemcpy(), and cudaMemcpyAsync().

If the destination topology is nvgraphCOOTopology32I_st, the tag field needs to be set to one of the values of NVGRAPH_UNSORTED, NVGRAPH_SORT_BY_SOURCE or NVGRAPH_SORT_BY_DESTINATION to tell the library how the user wants entries sorted (or not).

The function requires extra buffer storage in the device memory for some of the conversion operations, the extra storage is proportional to the topology size. The function is executed asynchronously with respect to the host and may return control to the application on the host before the result is ready.

Input
handle nvGRAPH library handle.
srcTType Source topology type. This value should be equal to one of the possible values of the enum nvgraphTopologyType_t. This defines what data structure type should be provided by the srcTopology parameter.
srcTopology Pointer to a structure of one of the types {nvgraphCSRTopology32I_t, nvgraphCSCTopology32I_t, nvgraphCOOTopology32I_st}. The particular type to be used is defined by parameter srcTType. The function assumes that source and destination arrays within the structure reside in device memory.
srcEdgeData Pointer to the user memoryspace where edge data are stored. Must be device memory.
dataType Edge data type. This value should be equal to one of CUDA_R_32F or CUDA_R_64F. This defines what data type is provided by the srcEdgeData and dstEdgeData parameters.
dstTType Destination topology type. This value should be equal to one of the possible values of the enum nvgraphTopologyType_t. This defines what data structure type should be provided by the dstTopology and parameter.
Input/Output
dstTopology Pointer to a structure of one of the types {nvgraphCSRTopology32I_t, nvgraphCSCTopology32I_t, nvgraphCOOTopology32I_st} where conversion results will be stored. The particular type to be used is defined by parameter dstTType. The function assumes that source and destination arrays within the structure are pre-allocated by the user with at least the number of bytes needed by the topology. Source and destination arrays are assumed to reside in device memory.
dstEdgeData Pointer to the user memoryspace where converted edge data will be stored. Must be device memory and have at least number_of_vertices*sizeof(dataType) bytes.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED An internal operation failed.
NVGRAPH_STATUS_INTERNAL_ERROR The type of at least one topology or edge set is not supported. Currently we support float and double type values.

2.12. Function nvgraphConvertGraph()

nvgraphStatus_t
   nvgraphConvertGraph(nvgraphHandle_t handle,
      nvgraphGraphDescr_t srcDescrG, nvgraphGraphDescr_t dstDescrG,
      nvgraphTopologyType_t dstTType);
      

Convert one of the supported graph types to another. The function will allocate the necessary memory for the destination graph. It is recommended to use this function over nvgraphConvertTopology when converting a large data set.

In addition to the destination graph memory, the function requires extra buffer storage in the device memory for some of the conversion operation, the extra storage is proportional to the topology size. The function is executed asynchronously with respect to the host and may return control to the application on the host before the result is ready.

Input
handle nvGRAPH library handle.
srcDescrG Source graph descriptor. This graph should have its topology structure previously set, and optionally vertex and edge data set.
dstTType Destination topology type. This value should be equal to one of the possible values of {nvgraphCSRTopology32I_t, nvgraphCSCTopology32I_t, nvgraphCOOTopology32I_t}. This defines what data structure type should be provided by the dstTopology and parameter.
Input/Output
dstDescrG Destination graph descriptor. Must be an empty descriptor created with nvgraphCreateGraphDescr and not have a topology set.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED An internal operation failed.
NVGRAPH_STATUS_INTERNAL_ERROR The type of at least one topology or edge set is not supported. Currently we support float and double type values.

2.13. Function nvgraphAllocateEdgeData()

nvgraphStatus_t
   nvgraphAllocateEdgeData(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      size_t numsets, cudaDataType_t *settypes);
      

Allocates one or more storages for the data associated with graph edges. Number of allocated storages is specified by the numsets parameter. Types for each of the allocated storages should be provided in the array settypes of size numsets. Right now nvGRAPH graphs are limited to have data storages to have same type and same size - all elements of settypes array should be the same and all of those storages will have number of elements equal to the number of edges in the graph. Vertices data allocated with nvgraphAllocateVerticesData() function should have the same datatype as edge data. These storages could later be used in other functions by indices from 0 to numsets-1. This function could be called successfully only once for each graph.

Input
handle nvGRAPH library handle
numsets Number of datasets to allocate for the edges. Should be more than 0.
settypes Array of size numsets that specifies types of allocated datasets. All values in this array should be the same and match graph's datasets data type, if exists.
Input/Output
descrG Descriptor of the graph for which edge data is allocated. Should not have previously allocated edge data and have it's topology properly initialized.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Invalid function parameters, inconsistent types in the type array, types doesn't match graph's type or graph is not initialized for data allocation.
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Types provided in parameter are not supported.

2.14. Function nvgraphSetEdgeData()

nvgraphStatus_t
   nvgraphSetEdgeData(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      void *edgeData, size_t setnum);
      

Update a specific edge value set (weights) of the graph with the user's provided values.

Input
handle nvGRAPH library handle.
descrG nvGRAPH graph descriptor, should contain the connectivity information and the edge set setnum
*edgeData Pointer to the data to load into the edge value set. This entry expects to read one value for each edge. Conversions are not supported so the user's type before the void* cast should be equivalent to the one specified in nvgraphAllocateEdgeData
setnum The identifier of the set to update. This assumes that setnum is one of the the edge set allocated in the past using nvgraphAllocateEdgeData. Sets have 0-based index
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_INTERNAL_ERROR An internal operation failed.

2.15. Function nvgraphGetEdgeData()

nvgraphStatus_t
   nvgraphGetEdgeData(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      void *edgeData, size_t setnum);
      

Downloads one dataset associated with graph edges using setnum index to the user memoryspace. edgeData could point to either host or device memoryspace. Size of the data transfer depends on the edges number of the graph and graph's data type.

Input
handle nvGRAPH library handle.
descrG Graph descriptor. Graph should contain at least one data set associated with it's vertices
setnum Index of the source data set of the graph edge data. Value should be between 0 and edge_dataset_number-1
Output
edgeData Pointer to the user memoryspace where edge data will be stored. Could be either host or device memory and have at least number_of_edges*sizeof(graph_data_type) bytes.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Incorrect function parameter, graph has no associated edge data sets or topology type doesn't match.
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Graph datatype or topology type is not supported.
NVGRAPH_STATUS_INTERNAL_ERROR Memory copy failed.

2.16. Function nvgraphAllocateVertexData()

nvgraphStatus_t
   nvgraphAllocateVertexData(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      size_t numsets, cudaDataType_t *settypes);
      

Allocates one or more storages for the data associated with graph vertices. Number of allocated storages is specified by the numsets parameter. Types for each of the allocated storages should be provided in the array settypes of size numsets. Right now nvGRAPH graphs are limited to have data storages to have same type and same size - all elements of settypes array should be the same and all of those storages will have number of elements equal to the number of vertices in the graph. Edge data allocated with nvgraphAllocateEdgeData() function should have the same datatype as vertex data. These storages could later be used in other functions by indices from 0 to numsets-1. This function could be called successfully only once for each graph.

Input
handle nvGRAPH library handle
numsets Number of datasets to allocate for the vertices. Should be more than 0.
settypes Array of size numsets that specifies types of allocated datasets. All values in this array should be the same and match graph's datasets data type, if exists.
Input/Output
descrG Descriptor of the graph for which edge data is allocated. Should not have previously allocated edge data and have it's topology properly initialized.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Invalid function parameters, inconsistent types in the type array, types doesn't match graph's type or graph is not initialized for data allocation.
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Types provided in parameter are not supported.

2.17. Function nvgraphSetVertexData()

nvgraphStatus_t
   nvgraphSetVertexData(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      void *vertexData, size_t setnum);
      

Update a specific vertex value set of the graph with the user's provided values.

Input
handle nvGRAPH library handle.
descrG nvGRAPH graph descriptor, should contain the connectivity information and vertex set setnum.
*vertexData Pointer to the data to load into the vertex value set. This entry expects to read one value for each vertex. Conversions are not supported so the user's type before the void* cast should be equivalent to the one specified in nvgraphAllocateVertexData
setnum The identifier of the set to update. This assumes that setnum is one of the the vertex set allocated in the past using nvgraphAllocateVertexData. Sets have 0-based index
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_INTERNAL_ERROR An internal operation failed.

2.18. Function nvgraphGetVertexData()

nvgraphStatus_t
   nvgraphGetVertexData(nvgraphHandle_t handle, nvgraphGraphDescr_t descrG,
      void *vertexData, size_t setnum);
      

Downloads one dataset associated with graph vertices using setnum index to the user memoryspace. vertexData could point to either host or device memoryspace. Size of the data transfer depends on the vertex number of the graph and graph's data type.

Input
handle nvGRAPH library handle.
descrG Graph descriptor. Graph should contain at least one data set associated with it's vertices
setnum Index of the source data set of the graph vertex data. Value should be between 0 and vertex_dataset_number-1
Output
edgeData Pointer to the user memoryspace where vertex data will be stored. Could be either host or device memory and have at least number_of_vertices*sizeof(graph_data_type) bytes.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Incorrect function parameter, graph has no associated vertex data sets or topology type doesn't match.
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED Graph datatype or topology type is not supported.
NVGRAPH_STATUS_INTERNAL_ERROR Memory copy failed.

2.19. Function nvgraphExtractSubgraphByVertex()

nvgraphStatus_t
   nvgraphExtractSubgraphByVertex(nvgraphHandle_t handle, 
      nvgraphGraphDescr_t descrG, nvgraphGraphDescr_t subdescrG, 
      int *subvertices, size_t numvertices);
      

Create a new graph by extracting a subgraph given an array of vertices, consisting of row indices in the graph incidence matrix; array must be: (i) free of duplicates; (ii) sorted in ascending order; (iii) must consist of indices with values between 0 and graph_nvertices-1.

Input
handle nvGRAPH handle of the source graph (original graph)
descrG nvGRAPH descriptor of the source graph (original graph)
subvertices array containing vertex indices (row indices in graph incidence matrix) of the subgraph to be extracted; array must be: (i) free of duplicates; (ii) sorted in ascending order; (iii) must consist of indices with values between 0 and graph_nvertices-1.
numvertices the size of subvertices[] array. Should be more than 0 and less or equal to the number of graph's vertices.
Output
subdescrG nvGRAPH graph descriptor of the target graph (subgraph)
Return Values
NVGRAPH_STATUS_SUCCESS nvGRAPH target (subgraph) was created succesfully.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED The type of specified nvGRAPH is not supported.

2.20. Function nvgraphExtractSubgraphByEdge()

nvgraphStatus_t
   nvgraphExtractSubgraphByEdge(nvgraphHandle_t handle, 
      nvgraphGraphDescr_t descrG, nvgraphGraphDescr_t subdescrG,
      int *subedges, size_t numedges);
      

Create a new graph by extracting a subgraph given an array of edges, consisting of indices in the col_ind[] array of the the graph incidence matrix CSR representation); the array of edges must be: (i) free of duplicates; (ii) sorted in ascending order; (iii) must consist of indices with values between 0 and graph_nedges-1.

Input
handle nvGRAPH handle of the source graph (original graph)
descrG nvGRAPH descriptor of the source graph (original graph)
subedges array containing edge indices (indices in the col_ind[] array of the the graph incidence matrix CSR representation) of the subgraph to be extracted; array must be: (i) free of duplicates; (ii) sorted in ascending order; (iii) must consist of indices with values between 0 and graph_nedges-1
numedges the size of subedges[] array, should be more than 0 and less or equal to number of graph's edges
Output
subdescrG nvGRAPH graph descriptor of the target graph (subgraph)
Return Values
NVGRAPH_STATUS_SUCCESS nvGRAPH target (subgraph) was created succesfully.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED The type of specified nvGRAPH is not supported.

2.21. Function nvgraphWidestPath()

nvgraphStatus_t
   nvgraphWidestPath(nvgraphHandle_t handle, const nvgraphGraphDescr_t descrG,
      const size_t weight_index, const int *source_vert, 
      const size_t widest_path_index);
      

Find the widest path from the vertex at source_index to every other vertex; this problem is also known as 'the bottleneck path problem' or 'the maximum capacity path problem'.

If some vertices are unreachable, the widest path to those vertices is -∞. In limited-precision arithmetic, this corresponds to -FLT_MAX or -DBL_MAX depending on the value type of the set (CUDA_R_32F or CUDA_R_64F respectively).

Input
handle nvGRAPH library handle.
descrG nvGRAPH graph descriptor, should contain the connectivity information in NVGRAPH_CSC_32, at least 1 edge set (the capacity ) and 1 vertex set (to store the result).
weight_index Index of the edge set for the weights.
*source_vert Index of the source, using 0-based indexes.
Output
widest_path_index The values strored inside the vertex set at widest_path_index (VertexData[widest_path_index]) are the widest path values. VertexData[widest_path_index][i] is the length of the widest path between source_vert and vertex i. If vertex i is not reachable from source_vert, VertexData[widest_path_index][i] = -∞.

Users can get a copy of the result using nvgraphGetVertexData
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED The type of at least one vetex or edge set is not supported.
NVGRAPH_STATUS_INTERNAL_ERROR An internal operation failed.

2.22. Function nvgraphSssp()

nvgraphStatus_t
   nvgraphSssp(nvgraphHandle_t handle, const nvgraphGraphDescr_t descrG,
      const size_t weight_index, const int *source_vert,
      const size_t sssp_index);
      

The Single Source Shortest Path (SSSP) algorithm calculates the shortest path distance from a single vertex in the graph to all other vertices.

If some vertices are unreachable, the shortest path to those vertices is ∞. In limited-precision arithmetic, that corresponds to FLT_MAX or DBL_MAX depending on the value type of the set (CUDA_R_32F or CUDA_R_64F respectively).

Input
handle nvGRAPH library handle.
descrG nvGRAPH graph descriptor, should contain the connectivity information in NVGRAPH_CSC_32, at least 1 edge set (distances) and 1 vertex set (the shortest path lengths).
weight_index Index of the edge set for the weights. The default value is 0, meaning the first edge set.
*source_vert Index of the source, using 0-based indexes.
Output
sssp_index The values stored inside the vertex set at sssp_index (VertexData[sssp_index]) are the shortest path values. VertexData[sssp_index][i] is the length of the shortest path between source_vert and vertex i. If vertex i is not reachable from source_vert, VertexData[sssp_index][i] = ∞.

User can get a copy of the result using nvgraphGetVertexData
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED The type of at least one vetex or edge set is not supported.
NVGRAPH_STATUS_INTERNAL_ERROR An internal operation failed.

2.23. Function nvgraphSrSpmv()

nvgraphStatus_t
   nvgraphSrSpmv(nvgraphHandle_t handle, const nvgraphGraphDescr_t descrG,
      const size_t weight_index, const void *alpha, const size_t x_index,
      const void *beta, const size_t y_index, const nvgraphSemiring_t SR);
      

The Semi-Ring Sparse Matrix Vector multiplication is an operation of the type y = α * A * x + β y. Where :

- A is a weighted graph seen as a compressed sparse matrix in CSR, x and y are vectors, α and β are scalars

- (*,+) is a set of two binary operators operating on real values and satisfying semi-ring properties.



In nvGRAPH all semi-rings operate on a set (R) with two binary operators: + and * that satisfies:

- (R, +) is associative, commutative with additive identity (additive_identity + a = a)

- (R, *) is associative with multiplicative identity (multiplicative_identity * a = a)

- Left and Right multiplication is distributive over addition

- Additive identity = multiplicative null operator ( null_operator * a = a * null_operator = null_operator).



nvGRAPH's approach for sparse matrix vector multiplication on the GPU is based on the CSRMV merge-path algorithm from Duane Merill. It is designed to handle arbitrary sparsity patterns in an efficient way by offering a good workload balance. As a result, this operation delivers consistent good performance even for networks with a power-law distribution of connections.

nvGRAPH has pre-defined useful semi-ring for graphs in nvgraphSemiring_t, so the user can select them directly.

Semi rings
Semiring Set Plus Times Add_ident Mult_ident
NVGRAPH_PLUS_TIMES_SR R + * 0 1
NVGRAPH_MIN_PLUS_SR R U {-∞,+∞} min + 0
NVGRAPH_MAX_MIN_SR R U {-∞,+∞} max min -∞ +∞
NVGRAPH_OR_AND_SR {0.0, 1,0} OR AND 0 1
Input
handle nvGRAPH library handle.
descrG nvGRAPH graph descriptor, should contain the connectivity information in NVGRAPH_CSR_32, at least 1 edge set (weights) and 2 vertex sets (input vector and output vector).
weight_index Index of the edge set for the weights.
*alpha Scalar used for multiplication
x_index Index of the vertex set for used for multiplication
*beta Scalar used for multiplication. If beta is zero, the vertex set at y_index does not have to be a valid input.
y_index (optional) Index of the vertex set for used for the addition.
SR The semi-ring type nvgraphSemiring_t which can be NVGRAPH_PLUS_TIMES_SR, NVGRAPH_MIN_PLUS_SR, NVGRAPH_MAX_MIN_SR, NVGRAPH_OR_AND_SR.
Output
Values at y_index The values stored inside the set at y_index (VertexData[y_index]) are the result of the operation.

User can get a copy of the result using nvgraphGetVertexData
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED The type of at least one vetex or edge set is not supported.
NVGRAPH_STATUS_INTERNAL_ERROR An internal operation failed.

2.24. Function nvgraphPagerank()

nvgraphStatus_t
   nvgraphPagerank(nvgraphHandle_t handle, const nvgraphGraphDescr_t descrG,
      const size_t weight_index, const void *alpha, 
      const size_t bookmark_index,
      const int has_guess, const size_t pagerank_index,
      const float tolerance, const int max_iter);
      

Find the PageRank vertex values for a graph with a given transition matrix (Markov chain), a bookmark vector of dangling vertices, and the damping factor. The transition matrix is sub-stochastic (ie. each row sums to 0 or 1) and has to be provided in column major order (ie. in CSC, which is equivalent to the transposed of the sub-stochastic matrix in CSR). The bookmark vector flags vertices without outgoing edges (also called dangling vertices).

This is equivalent to an eigenvalue problem where we compute the dominant eigenpair. By construction, the maximum eigenvalue is 1, only the eigenvector is interesting. nvGRAPH computes an approximation of the Pagerank eigenvector using the power method. The number of iterations depends on the properties of the network itself; it increases when the tolerance descreases and/or alpha increases toward the limiting value of 1.

The user is free to use default values or to provide inputs for the initial guess, tolerance and maximum number of iterations.

Input
handle nvGRAPH library handle.
descrG nvGRAPH graph descriptor, should contain the connectivity information in NVGRAPH_CSC_32, at least 1 edge set and 2 vertex sets.
weight_index Index of the edge set for the transition probability.
*alpha The damping factor alpha represents the probability to follow an outgoing edge, standard value is 0.85. Thus 1.0-alphais the probability to “teleport” to a random node. alphashould be greater than 0.0 and strictly lower than 1.0.
bookmark_index Index of the vertex set for the bookmark of dangling nodes (VertexData[bookmark_index][i] = 1.0 if i is a dangling node, 0.0 otherwise).
has_guess This parameter is used to notify nvGRAPH if it should use a user-provided initial guess. 0 means the user doesn't have a guess, in this case nvGRAPH will use a uniform vector set to 1/V. If the value is 1 nvGRAPH will read VertexData[pagerank_index] and use this as initial guess. The initial guess must not be the vector of 0s. Any value other than 1 or 0 is treated as an invalid value.
pagerank_index (optional) Index of the vertex set for the initial guess if has_guess=1
tolerance Set the tolerance the approximation, this parameter should be a small magnitude value. The lower the tolerance the better the approximation. If this value is 0.0f, nvGRAPH will use the default value which is 1.0E-6. Setting too small a tolerance (less than 1.0E-6 typically) can lead to non-convergence due to numerical roundoff. Usually 0.01 and 0.0001 are acceptable.
max_iter The maximum number of iterations before an answer is returned. This can be used to limit the execution time and do an early exit before the solver reaches the convergence tolerance. If this value is lower or equal to 0 nvGRAPH will use the default value, which is 500.
Output
Values at pagerank_index The values stored inside the vertex set at pagerank_index (VertexData[pagerank_index]) are the PageRank values. VertexData[pagerank_index][i] is the PageRank of vertex i.

Users can get a copy of the result using nvgraphGetVertexData
Return Values
NVGRAPH_STATUS_SUCCESS The Pagerank iteration reached the desired tolerance in less than max_iter iterations
NVGRAPH_STATUS_NOT_CONVERGED The Pagerank iteration did not reach the desired tolerance after max_iter iterations
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s).
NVGRAPH_STATUS_TYPE_NOT_SUPPORTED The type of at least one vertex or edge set is not supported. Currently we support float and double type values.

2.25. Function nvgraphTriangleCount()

nvgraphStatus_t
   nvgraphTriangleCount(nvgraphHandle_t handle, 
      const nvgraphGraphDescr_t graph_descr, 
      uint64_t* result);
      

The triangles counting algorithm calculates number of unique triangles formed by graph edges. Algorithm works on the undirected graphs and graph structure in parameter should store only lower triangular of the adjacency matrix (no diagonal or self loops for vertices).

Input
handle nvGRAPH library handle.
graph_descr nvGRAPH graph descriptor, should contain graph connectivity information in NVGRAPH_CSR_32 or NVGRAPH_CSC_32 format which should contain only lower triangular of adjacency matrix (no diagonal or self loops for vertices).
Output
result Number of triangles for provided graph will be stored in this parameter.
Return Values
NVGRAPH_STATUS_SUCCESS Success.
NVGRAPH_STATUS_INVALID_VALUE Bad parameter(s), for example unsupported topology or topology is not stored in the graph.
NVGRAPH_STATUS_MAPPING_ERROR Incorrect graph handle parameter.
NVGRAPH_STATUS_INTERNAL_ERROR An internal operation failed.

2.26. Function nvgraphStatusGetString()

const char* 
   nvgraphStatusGetString(nvgraphStatus_t status);
      

Gets string description for the nvGRAPH C API statuses.

Input
status Status returned from one of the C API functions
Return Values
Pointer to the string with the text description of the C API status.

3. nvGRAPH Code Examples

This chapter provides simple examples.

3.1. nvGRAPH convert topology example

void check(nvgraphStatus_t status) {
    if (status != NVGRAPH_STATUS_SUCCESS) {
        printf("ERROR : %d\n",status);
        exit(0);
    }
}
int main(int argc, char **argv) {
    size_t  n = 6, nnz = 10;
    // nvgraph variables
    nvgraphHandle_t handle;
    nvgraphCSCTopology32I_t CSC_input;
    nvgraphCSRTopology32I_t CSR_output;
    float *src_weights_d, *dst_weights_d;
    cudaDataType_t edge_dimT = CUDA_R_32F;
    // Allocate source data
    CSC_input = (nvgraphCSCTopology32I_t) malloc(sizeof(struct nvgraphCSCTopology32I_st));
    CSC_input->nvertices = n; CSC_input->nedges = nnz;
    cudaMalloc( (void**)&(CSC_input->destination_offsets), (n+1)*sizeof(int));
    cudaMalloc( (void**)&(CSC_input->source_indices), nnz*sizeof(int));
    cudaMalloc( (void**)&src_weights_d, nnz*sizeof(float));
    // Copy source data
    float src_weights_h[] = {0.333333f, 0.5f, 0.333333f, 0.5f, 0.5f, 1.0f, 0.333333f, 0.5f, 0.5f, 0.5f};
    int destination_offsets_h[] = {0, 1, 3, 4, 6, 8, 10};
    int source_indices_h[] = {2, 0, 2, 0, 4, 5, 2, 3, 3, 4};
    cudaMemcpy(CSC_input->destination_offsets, destination_offsets_h, (n+1)*sizeof(int), cudaMemcpyDefault);
    cudaMemcpy(CSC_input->source_indices, source_indices_h, nnz*sizeof(int), cudaMemcpyDefault);
    cudaMemcpy(src_weights_d, src_weights_h, nnz*sizeof(float), cudaMemcpyDefault);
    // Allocate destination data
    CSR_output = (nvgraphCSRTopology32I_t) malloc(sizeof(struct nvgraphCSRTopology32I_st));
    cudaMalloc( (void**)&(CSR_output->source_offsets), (n+1)*sizeof(int));
    cudaMalloc( (void**)&(CSR_output->destination_indices), nnz*sizeof(int));
    cudaMalloc( (void**)&dst_weights_d, nnz*sizeof(float));
    // Starting nvgraph and convert
    check(nvgraphCreate (&handle));
    check(nvgraphConvertTopology(handle, NVGRAPH_CSC_32, CSC_input, src_weights_d,
        &edge_dimT, NVGRAPH_CSR_32, CSR_output, dst_weights_d));
    // Free memory
    check(nvgraphDestroy(handle));
    cudaFree(CSC_input->destination_offsets);
    cudaFree(CSC_input->source_indices);
    cudaFree(CSR_output->source_offsets);
    cudaFree(CSR_output->destination_indices);
    cudaFree(src_weights_d);
    cudaFree(dst_weights_d);
    free(CSC_input);
    free(CSR_output);
    return 0;
}

3.2. nvGRAPH convert graph example

void check(nvgraphStatus_t status) {
    if (status != NVGRAPH_STATUS_SUCCESS) {
        printf("ERROR : %d\n",status);
        exit(0);
    }
}
int main(int argc, char **argv) {
    size_t  n = 6, nnz = 10, vert_sets = 2, edge_sets = 1;
    // nvgraph variables
    nvgraphHandle_t handle; nvgraphGraphDescr_t src_csc_graph;
    nvgraphCSCTopology32I_t CSC_input;
    cudaDataType_t edge_dimT = CUDA_R_32F;
    cudaDataType_t* vertex_dimT;
    // Allocate host data
    float *pr_1 = (float*)malloc(n*sizeof(float));
    void **vertex_dim = (void**)malloc(vert_sets*sizeof(void*));
    vertex_dimT = (cudaDataType_t*)malloc(vert_sets*sizeof(cudaDataType_t));
    CSC_input = (nvgraphCSCTopology32I_t) malloc(sizeof(struct nvgraphCSCTopology32I_st));
    // Initialize host data
    float weights_h[] = {0.333333f, 0.5f, 0.333333f, 0.5f, 0.5f, 1.0f, 0.333333f, 0.5f, 0.5f, 0.5f};
    int destination_offsets_h[] = {0, 1, 3, 4, 6, 8, 10};
    int source_indices_h[] = {2, 0, 2, 0, 4, 5, 2, 3, 3, 4};
    float bookmark_h[] = {0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f};
    vertex_dim[0] = (void*)bookmark_h; vertex_dim[1]= (void*)pr_1;
    vertex_dimT[0] = CUDA_R_32F; vertex_dimT[1]= CUDA_R_32F, vertex_dimT[2]= CUDA_R_32F;
    // Starting nvgraph
    check(nvgraphCreate (&handle));
    check(nvgraphCreateGraphDescr (handle, &src_csc_graph));
    CSC_input->nvertices = n; CSC_input->nedges = nnz;
    CSC_input->destination_offsets = destination_offsets_h;
    CSC_input->source_indices = source_indices_h;
    // Set graph connectivity and properties (tranfers)
    check(nvgraphSetGraphStructure(handle, src_csc_graph, (void*)CSC_input, NVGRAPH_CSC_32));
    check(nvgraphAllocateVertexData(handle, src_csc_graph, vert_sets, vertex_dimT));
    check(nvgraphAllocateEdgeData  (handle, src_csc_graph, edge_sets, &edge_dimT));
    for (int i = 0; i < 2; ++i)
        check(nvgraphSetVertexData(handle, src_csc_graph, vertex_dim[i], i));
    check(nvgraphSetEdgeData(handle, src_csc_graph, (void*)weights_h, 0));
    // Convert to CSR graph
    nvgraphGraphDescr_t dst_csr_graph;
    check(nvgraphCreateGraphDescr (handle, &dst_csr_graph));
    check(nvgraphConvertGraph(handle, src_csc_graph, dst_csr_graph, NVGRAPH_CSR_32));
    check(nvgraphDestroyGraphDescr(handle, src_csc_graph));
    check(nvgraphDestroyGraphDescr(handle, dst_csr_graph));
    check(nvgraphDestroy(handle));
    free(pr_1); free(vertex_dim); free(vertex_dimT);
    free(CSC_input);
    return 0;
}

3.3. nvGRAPH pagerank example

void check(nvgraphStatus_t status) {
    if (status != NVGRAPH_STATUS_SUCCESS) {
        printf("ERROR : %d\n",status);
        exit(0);
    }
}
int main(int argc, char **argv) {
    size_t  n = 6, nnz = 10, vert_sets = 2, edge_sets = 1;
    float alpha1 = 0.9f; void *alpha1_p = (void *) &alpha1;
    // nvgraph variables
    nvgraphHandle_t handle; nvgraphGraphDescr_t graph;
    nvgraphCSCTopology32I_t CSC_input;
    cudaDataType_t edge_dimT = CUDA_R_32F;
    cudaDataType_t* vertex_dimT;
    // Allocate host data
    float *pr_1 = (float*)malloc(n*sizeof(float));
    void **vertex_dim = (void**)malloc(vert_sets*sizeof(void*));
    vertex_dimT = (cudaDataType_t*)malloc(vert_sets*sizeof(cudaDataType_t));
    CSC_input = (nvgraphCSCTopology32I_t) malloc(sizeof(struct nvgraphCSCTopology32I_st));
    // Initialize host data
    float weights_h[] = {0.333333f, 0.5f, 0.333333f, 0.5f, 0.5f, 1.0f, 0.333333f, 0.5f, 0.5f, 0.5f};
    int destination_offsets_h[] = {0, 1, 3, 4, 6, 8, 10};
    int source_indices_h[] = {2, 0, 2, 0, 4, 5, 2, 3, 3, 4};
    float bookmark_h[] = {0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f};
    vertex_dim[0] = (void*)bookmark_h; vertex_dim[1]= (void*)pr_1;
    vertex_dimT[0] = CUDA_R_32F; vertex_dimT[1]= CUDA_R_32F, vertex_dimT[2]= CUDA_R_32F;
    // Starting nvgraph
    check(nvgraphCreate (&handle));
    check(nvgraphCreateGraphDescr (handle, &graph));
    CSC_input->nvertices = n; CSC_input->nedges = nnz;
    CSC_input->destination_offsets = destination_offsets_h;
    CSC_input->source_indices = source_indices_h;
    // Set graph connectivity and properties (tranfers)
    check(nvgraphSetGraphStructure(handle, graph, (void*)CSC_input, NVGRAPH_CSC_32));
    check(nvgraphAllocateVertexData(handle, graph, vert_sets, vertex_dimT));
    check(nvgraphAllocateEdgeData  (handle, graph, edge_sets, &edge_dimT));
    for (int i = 0; i < 2; ++i)
        check(nvgraphSetVertexData(handle, graph, vertex_dim[i], i));
    check(nvgraphSetEdgeData(handle, graph, (void*)weights_h, 0));
    
    check(nvgraphPagerank(handle, graph, 0, alpha1_p, 0, 0, 1, 0.0f, 0));
    // Get result
    check(nvgraphGetVertexData(handle, graph, vertex_dim[1], 1));
    check(nvgraphDestroyGraphDescr(handle, graph));
    check(nvgraphDestroy(handle));
    free(pr_1); free(vertex_dim); free(vertex_dimT);
    free(CSC_input);
    return 0;
}

3.4. nvGRAPH SSSP example

void check(nvgraphStatus_t status) {
    if (status != NVGRAPH_STATUS_SUCCESS) {
        printf("ERROR : %d\n",status);
        exit(0);
    }
}
int main(int argc, char **argv) {
    const size_t  n = 6, nnz = 10, vertex_numsets = 1, edge_numsets = 1;
    float *sssp_1_h;
    void** vertex_dim;
    // nvgraph variables
    nvgraphStatus_t status; nvgraphHandle_t handle;
    nvgraphGraphDescr_t graph;
    nvgraphCSCTopology32I_t CSC_input;
    cudaDataType_t edge_dimT = CUDA_R_32F;
    cudaDataType_t* vertex_dimT;
    // Init host data
    sssp_1_h = (float*)malloc(n*sizeof(float));
    vertex_dim  = (void**)malloc(vertex_numsets*sizeof(void*));
    vertex_dimT = (cudaDataType_t*)malloc(vertex_numsets*sizeof(cudaDataType_t));
    CSC_input = (nvgraphCSCTopology32I_t) malloc(sizeof(struct nvgraphCSCTopology32I_st));
    vertex_dim[0]= (void*)sssp_1_h; vertex_dimT[0] = CUDA_R_32F;
    float weights_h[] = {0.333333, 0.5, 0.333333, 0.5, 0.5, 1.0, 0.333333, 0.5, 0.5, 0.5};
    int destination_offsets_h[] = {0, 1, 3, 4, 6, 8, 10};
    int source_indices_h[] = {2, 0, 2, 0, 4, 5, 2, 3, 3, 4};
    check(nvgraphCreate(&handle));
    check(nvgraphCreateGraphDescr (handle, &graph));
    CSC_input->nvertices = n; CSC_input->nedges = nnz;
    CSC_input->destination_offsets = destination_offsets_h;
    CSC_input->source_indices = source_indices_h;
    // Set graph connectivity and properties (tranfers)
    check(nvgraphSetGraphStructure(handle, graph, (void*)CSC_input, NVGRAPH_CSC_32));
    check(nvgraphAllocateVertexData(handle, graph, vertex_numsets, vertex_dimT));
    check(nvgraphAllocateEdgeData  (handle, graph, edge_numsets, &edge_dimT));
    check(nvgraphSetEdgeData(handle, graph, (void*)weights_h, 0));
    // Solve
    int source_vert = 0;
    check(nvgraphSssp(handle, graph, 0,  &source_vert, 0));
    // Get and print result
    check(nvgraphGetVertexData(handle, graph, (void*)sssp_1_h, 0));
    //Clean 
    free(sssp_1_h); free(vertex_dim);
    free(vertex_dimT); free(CSC_input);
    check(nvgraphDestroyGraphDescr(handle, graph));
    check(nvgraphDestroy(handle));
    return 0;
}

3.5. nvGRAPH Semi-Ring SPMV example

void check(nvgraphStatus_t status) {
    if (status != NVGRAPH_STATUS_SUCCESS) {
        printf("ERROR : %d\n",status);
        exit(0);
    }
}
int main(int argc, char **argv) {
    size_t  n = 5, nnz = 10, vertex_numsets = 2, edge_numsets = 1;
    float alpha = 1.0, beta = 0.0;
    void *alpha_p = (void *)&alpha, *beta_p = (void *)&beta;
    void** vertex_dim;
    cudaDataType_t edge_dimT = CUDA_R_32F;
    cudaDataType_t* vertex_dimT;
    // nvgraph variables
    nvgraphStatus_t status; nvgraphHandle_t handle;
    nvgraphGraphDescr_t graph;
    nvgraphCSRTopology32I_t CSR_input;
    // Init host data
    vertex_dim  = (void**)malloc(vertex_numsets*sizeof(void*));
    vertex_dimT = (cudaDataType_t*)malloc(vertex_numsets*sizeof(cudaDataType_t));
    CSR_input = (nvgraphCSRTopology32I_t) malloc(sizeof(struct nvgraphCSRTopology32I_st));
    float x_h[] = {1.1f, 2.2f, 3.3f, 4.4f, 5.5f};
    float y_h[] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
    vertex_dim[0]= (void*)x_h; vertex_dim[1]= (void*)y_h;
    vertex_dimT[0] = CUDA_R_32F; vertex_dimT[1]= CUDA_R_32F;
    float weights_h[] = {1.0f, 4.0f, 2.0f, 3.0f, 5.0f, 7.0f, 8.0f, 9.0f, 6.0f, 1.5f};
    int source_offsets_h[] = {0, 2, 4, 7, 9, 10};
    int destination_indices_h[] = {0, 1, 1, 2, 0, 3, 4, 2, 4, 2};
    check(nvgraphCreate(&handle));
    check(nvgraphCreateGraphDescr(handle, &graph));
    CSR_input->nvertices = n; CSR_input->nedges = nnz;
    CSR_input->source_offsets = source_offsets_h;
    CSR_input->destination_indices = destination_indices_h;
    // Set graph connectivity and properties (tranfers)
    check(nvgraphSetGraphStructure(handle, graph, (void*)CSR_input, NVGRAPH_CSR_32));
    check(nvgraphAllocateVertexData(handle, graph, vertex_numsets, vertex_dimT));
    for (int i = 0; i < vertex_numsets; ++i)
       check(nvgraphSetVertexData(handle, graph, vertex_dim[i], i));
    check(nvgraphAllocateEdgeData  (handle, graph, edge_numsets, &edge_dimT));
    check(nvgraphSetEdgeData(handle, graph, (void*)weights_h, 0));
    // Solve
    check(nvgraphSrSpmv(handle, graph, 0, alpha_p, 0, beta_p, 1, NVGRAPH_PLUS_TIMES_SR));
    //Get result
    check(nvgraphGetVertexData(handle, graph, (void*)y_h, 1));
    //Clean 
    check(nvgraphDestroyGraphDescr(handle, graph));
    check(nvgraphDestroy(handle));
    free(vertex_dim); free(vertex_dimT); free(CSR_input);
    return 0;
}

3.6. nvGRAPH Triangles Counting example

#include "stdlib.h"
#include "inttypes.h"
#include "stdio.h"

#include "nvgraph.h"

#define check( a ) \
{\
    nvgraphStatus_t status = (a);\
    if ( (status) != NVGRAPH_STATUS_SUCCESS) {\
        printf("ERROR : %d in %s : %d\n", status, __FILE__ , __LINE__ );\
        exit(0);\
    }\
}

int main(int argc, char **argv) 
{
    // nvgraph variables
    nvgraphHandle_t handle;
    nvgraphGraphDescr_t graph;
    nvgraphCSRTopology32I_t CSR_input;

    // Init host data
    CSR_input = (nvgraphCSRTopology32I_t) malloc(sizeof(struct nvgraphCSRTopology32I_st));

    // Undirected graph:
    // 0       2-------4       
    //  \     / \     / \
    //   \   /   \   /   \
    //    \ /     \ /     \
    //     1-------3-------5
    // 3 triangles
    // CSR of lower triangular of adjacency matrix:
    const size_t n = 6, nnz = 8;
    int source_offsets[] = {0, 0, 1, 2, 4, 6, 8};
    int destination_indices[] = {0, 1, 1, 2, 2, 3, 3, 4};

    check(nvgraphCreate(&handle));
    check(nvgraphCreateGraphDescr (handle, &graph));
    CSR_input->nvertices = n; 
    CSR_input->nedges = nnz;
    CSR_input->source_offsets = source_offsets;
    CSR_input->destination_indices = destination_indices;
    // Set graph connectivity
    check(nvgraphSetGraphStructure(handle, graph, (void*)CSR_input, NVGRAPH_CSR_32));

    uint64_t trcount = 0;
    check(nvgraphTriangleCount(handle, graph, &trcount));
    printf("Triangles count: %" PRIu64 "\n", trcount);

    free(CSR_input);
    check(nvgraphDestroyGraphDescr(handle, graph));
    check(nvgraphDestroy(handle));
    return 0;
}

Notices

Notice

ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, "MATERIALS") ARE BEING PROVIDED "AS IS." NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.

Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation.

Trademarks

NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.