Geek Answers Handbook

Are int operations possible on the GPU in Theano?

So, I read that theano cannot perform gpu calculations using float64and store ints as shared variables in gpu, they must be initialized as shared data float32and then processed into ints (for example, in the “little hack” example in a logistic regression ). .. but after such a rebuild, can anano do gpu calculations in ints ? and is it a prerequisite for the calculation? In other words, are the following two scenarios possible?

Scenario 1. I want to make a point product on two large ints vectors. So I make them separable as I float32rewrite them in int before the point product, is this point product then on gpu (regardless of the type of int)?

Scenario 2. If scenario 1 is possible, will it be possible to perform calculations on gpu without first saving them as shared float32? (I understand that exchange variables can soften the gpu-cpu connection, but will a point product be possible as well? Is the repository a prerequisite for computing on gpu?)

+4
source share
1 answer

, - , float32.

:

import numpy
import theano
import theano.tensor as tt

x = theano.shared(numpy.arange(9 * 10).reshape((9, 10)).astype(numpy.float32))
y = theano.shared(numpy.arange(10 * 11).reshape((10, 11)).astype(numpy.float32))
z = theano.dot(tt.cast(x, 'int32'), tt.cast(y, 'int32'))
f = theano.function([], outputs=z)
theano.printing.debugprint(f)

:

dot [@A] ''   4
 |Elemwise{Cast{int32}} [@B] ''   3
 | |HostFromGpu [@C] ''   1
 |   |<CudaNdarrayType(float32, matrix)> [@D]
 |Elemwise{Cast{int32}} [@E] ''   2
   |HostFromGpu [@F] ''   0
     |<CudaNdarrayType(float32, matrix)> [@G]

, GPU ( CudaNdarrayType s), ( CPU/ ) ( HostFromGpu), , int dot.

,

HostFromGpu [@A] ''   1
 |GpuDot22 [@B] ''   0
   |<CudaNdarrayType(float32, matrix)> [@C]
   |<CudaNdarrayType(float32, matrix)> [@D]

, GPU ( GpuDot22), , .

+5

More articles:


All Articles