hi Jin -- the idea of using unified memory is interesting. My initial thinking was to make the semantics of host vs. device memory explicit, but it's hard to tell until we see how more GPU users will be using the Arrow libraries. As an initial first step, we could create a Buffer implementation (like CudaBuffer) that uses unified memory. From there we could explore what is most useful in general in various downstream applications.
best Wes On Sat, Dec 23, 2017 at 10:23 PM, Jin Hai <mildwolf...@sina.com> wrote: > > Hi there, > > I found in arrow, there are some GPU memory buffer related code and also use > malloc for the buffer allocate. Have you considered using the unified memory > instead of cudamemcpy and default malloc? > > Best Regards > Jin Hai
