rdivide and minus operation runs faster on GPU than rdivide alone.

I am experimenting with GPU and the runtime is interesting for two test functions.

Input in both cases:

 d = rand(1,100000000,'single','gpuArray');
 b = rand(1,1,'single','gpuArray');

First function:

 function d = gputest1(d,b)
 tic
 for i=1:10000
     d=d./(d-b);
 end
 wait(gpuDevice)
 toc
 end

Second:

 function d = gputest2(d,b)
 tic
 for i=1:10000
     d=d./b;
 end
 wait(gpuDevice)
 toc
 end

I expect longer runtime for gputest1 because it has to do two operations in one iteration, but the measured runtime is 12 s for gputest1, and 27 s for gputest2. Does anyone have an explanation for this?

Tests are performed on a GTX 1060 6GB (CPU: i7-7700, RAM: 32 GB).

[Joss Knight]

I can reproduce this, and it is indeed odd. It is probably explained by some special kind of optimisation, or perhaps that division by the same scalar value is slow due to i/o issues on the device. However, I'll look into it further.

[Prabhakar]

Which version of MATLAB are you running?

[Laszlo Kormoczi]

R2017b

Answer 1

[Prabhakar]

Try varying the size of the inputs. Reducing the size of the input from 1e8 to 1e5 shows expected behavior. (ie gpuTest2 being faster than gpuTest1.)

Performance is probably being dictated by the amount of memory being fetched vs the amount of work being carried out in each kernel.