Fixed error in grad_chooser (for e.g., max) when dtype is not numpy float64

[kswersky]

Grads of e.g., sum(anp.max(x, 1)) fail when x is of dtype float32. The issue is that numpy implicitly casts to float64 when dividing a float32 array by an int64 array, which happens in grad_chooser. This fixes the issue by casting appropriately in grad_chooser.

[alexbw]

FYI, I have taken some autograd gradients for a side project, and found float64 casting performance bugs scattered around.

[j-towns]

@alexbw is this when you're using float32 arrays?

It should be straightforward to add float32 tests. I think adding something to check_fun_and_grads/check_grads should cover almost everything.

[alexbw]

Yes, when using float32s. Sorry I don't anything more specific than that, I no longer depend on autograd grads, so am not spending more bandwidth on testing them right now. But, if you're interested, there's a possible 2x speed gain by avoiding the casts. Sorry to be unhelpful, but I figured having the info out there would be more useful than silence.

…

On Mon, Mar 20, 2017 at 6:00 PM Jamie Townsend ***@***.***> wrote: @alexbw <https://github.com/alexbw> is this when you're using float32 arrays? It should be straightforward to add float32 tests. I think adding something to check_fun_and_grads/check_grads should cover almost everything. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#199 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAJ4j29e-4XUqnGZeAMQzxjHxa-gFO4mks5rnvb-gaJpZM4MdhXB> .

[j-towns]

Yeah it's even better than 2x for some things. I made this pr a little while ago but I'll make another one to add testing for float32 dtype support accross all the primitives.