Re: [eigen] Componentwise Operations on an Arbitrary Number of Tensors

On Dec 28, 2016 4:49 PM, "Gael Guennebaud" <gael.guennebaud@xxxxxxxxx> wrote:

Hi,

it seems that what you're looking for is a mean to merge multiple evaluation loops of the same size into a single one (the fact that they run on the GPU is not really important here). Actually, this needs already shows up for stuff like:

a = vec.minCoeff();
b = vec.maxCoeff();

that currently requires two loops. I remember that we already talked about that with Benoit S., and I don't think there is a general solution implemented in the Tensor module yet.

Technically, I don't think that's very difficult though. The main difficulty is perhaps on the API side. We could imagine something like:

auto E1 = (R1.deferred() = expr1);
auto E2 = (R2.deferred() = expr2);
...
merged_eval(E1, E2, ...);

that would essentially generate:

(parallel/GPU/whatever) for loop {
R1[i] = expr1.coeffl(i);
R2[i] = expr2.coeffl(i);
...
}

In Eigen/Core, "R.deferred().operator=(expr)" would return an Eigen::internal::Assignment _expression_ (without calling run) that would be merged by the merged_eval function.

gael

On Wed, Dec 28, 2016 at 3:22 PM, Graham Neubig <gneubig@xxxxxxxxxxx> wrote:
Hi Eigen Folks,

First, thanks for the great library. We're using it in our machine learning library DyNet to great success.

I had a quick question about something that seems like it should be possible, but I haven't found a reference. I currently have code here:
https://github.com/clab/dynet/blob/master/dynet/training.cc#L280

That implements the "Adam" update rule for stochastic gradient descent found in this paper:
https://arxiv.org/abs/1412.6980

Here, all places with "tvec()" are Eigen one-dimensional Tensors. The thing that bugs me here is that I'm calling 4 different operations, which results in 4 different GPU kernel launches, for an operation that is inherently componentwise. If possible, I'd like to be able to basically create a single functor that takes 4 floats, and modifies them appropriately, then pass this in a single GPU operation.

I know this is possible using binaryExpr() for binary expressions, but I couldn't find it for operations with a larger number of arguments. Is there any chance that there is an elegant way to do this within Eigen (i.e. without writing my own kernel)?

Graham