From today's meeting, Aig and James helped me to solve the problem that the completion time in simpleStreams is much higher than the expected time. And we also came up with some equations to model the completion time. There should be some diagram to explain but I have not drown it yet.
The equations to calculate the total time of memory copy are:
TS = n(t + l), for synchronous memory copy
TA = (n/s)t + ((n/s) + s - 1)l, for asynchronous memory copy.
n = number of data
s = number of streams
l = latency or overhead
t = copy time
Another idea that James raised in the meeting is about data scheduling, to keep the GPU occupied while the data are being copied from the CPU.
Thank Aig and James for your help. ;)
No comments:
Post a Comment