Description
Bellman-Ford algorithm is a well-known solution to “the single-source shortest
path(SSSP)” problem. It is slower than Dijkstra’s algorithm, but more versatile, as it is
capable of handling graphs in which some of the edge weights are negative numbers.
The input graph G(V, E) for this assignment is connected, directed and may contain
negative weights. The algorithm finds a shortest path from a specified vertex (the
‘source vertex’) to every other vertex in the graph. If there is a negative cycle (a cycle
on which the sum of weights is negative) in the graph, there will be no shortest path. In
this case, the algorithm will find no result.
In this assignment, you will implement an CUDA version of Bellman-Ford algorithm.
The input file will be in following format:
1. The first line is an integer N, the number of vertices in the input graph.
2. The following lines are an N*N adjacency matrix mat, one line per row. The
entry in row v and column w, mat[v][w], is the distance (weight) from vertex
v to vertex w. All distances are integers. If there is no edge joining vertex v and
w, mat[v][w] will be 1000000 to represent infinity.
The vertex labels are non-negative, consecutive integers, for an input graph with N
vertices, the vertices will be labeled by 0, 1, 2, …, N-1. We always use vertex 0
as the source vertex.
The output file of your program consists the distances from vertex 0 to all vertices, in
the increasing order of the vertex label (vertex 0, 1, 2, … and so on), one distance per
line. If there are at least one negative cycle (the sum of the weights of the cycle is
negative in the graph), your program will set variable has_negative_cycle to true
and print “FOUND NEGATIVE CYCLE!” as there will be no shortest path.
Here are two examples input/output for your reference:
The code skeleton cuda_bellman_ford.cpp is provided. You task is to complete the
following function in the code:
void bellman_ford
(int blocksPerGrid, int threadsPerBlock, int n, int *mat, int *dist, bool *has_negative_cycle)
The description of the parameters is as follows:
Parameter Description
int blocksPerGrid Number of blocks for each grid.
int threadsPerBlock Number of threads for each block.
int n Number of vertices.
int *mat Adjacency matrix (stored in one dimension), N * N elements
int *dist The result array storing the final distance from the source for each
vertex, N elements
bool
*has_negative_cycle
set it to true if there is negative weight cycle. Otherwise, set it to
false
The element mat[v * N + w] stores distance(weight) from vertex v to vertex w.
Example 1:
Input:
3
0 3 2
1000000 0 -2
1000000 2 0
Output:
0
3
1
Example 2:
Input:
4
0 100 100 100
100 0 100 -1
100 -1 0 100
100 100 -1 0
Output:
FOUND NEGATIVE CYCLE!
Note 1: The sequential algorithm of Bellman-Ford is provided for your reference. Your
parallel version can follow the same logic flow of the sequential version, but you will
need to parallelize it in CUDA.
Note 2: You can add helper/kernel functions and variables as you wish, but keep the
existing code skeleton unchanged.
Note 3: We will use different input files, possibly with negative weights and cycles and
specify different numbers of blocksPerGrid/threadsPerBlock to test your program
(4<=blocksPerGrid<=32, 32<=threadsPerBlock<=1024 & threadsPerBlock is power
of 2).
Note 4: The running time and speedup of your program will be considered in grading.
