Open JoelMathewC opened 1 year ago
Since there are spaces in GPMA we are not benefiting from the coalesced memory access and hence we can maybe benefit from moving to shared memory as shown here: https://www.tutorialspoint.com/cuda/cuda_memory_considerations.htm
Is edge or node parallelism better for the count_sort portion of building the reverse graph.
We will need to think about how seastar can be used for evaluation. So no backdrop
deepcopy
(CORRECTION) GPMA Node labelling should start from 1
The node labelling in GPMA should start from 1, the reason is because when GPMA is initialized by default a sentinel value is inserted for every node. That sentinel value is
(src,0)
as a 64-bit element.We will probably have to perform a relabelling in preprocessing or find a better way to deal with this.
CORRECTION (CLOSED)
(src,0xFFFFFFFF)
and hence we don't really have to worry about node id 0 being ignored.HENCE NO CHANGE IS REQUIRED