Optimize GpuAdvancedIncSubtensor1_dev20 for mixed dtype

[nouiz]

See: dc7f7a4

The cast should be inside the kernel, not inside the graph.

[lamblin]

I think @abergeron disagreed with that in #5747 (comment)

[nouiz]

Here is @abergeron comment: """ I would prefer not to multiply the possible C code b y including a cartesian product of all the types. Also, if we do it this way it gives the opportunity to merge the cast with another elemwise operation. Finally since we have float16 support, there would need to be special-case handling of casting to and from that type which would complicate the kernel code. """ There isn't always elemwise. We could skip not supporting all cases, like float16. I don't know how frequently that will help. Mostly, it depend how frequently this happen in real user code. But there is still theoretical case where this could be good to do. I have no problem to wait to see it in real code before working on that.

…

On Fri, Jun 2, 2017 at 1:41 PM Pascal Lamblin ***@***.***> wrote: I think @abergeron <https://github.com/abergeron> disagreed with that in #5747 (comment) <#5747 (comment)> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5999 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALC-9-lRgFTY-LDk06QOcJ1CGmzstmfks5sAElOgaJpZM4Ns5L9> .