ublas::prod and viennacl::prod report incorrect results when 'typedef ScalarType long(or)int;' in examples/tutorial/blas3.cu

[Sanjay]

I'm looking at using viennacl for integer matrices, especially for matrix multiplication. I've been using the 1.7.1 release.

Changing the ScalarType to int or long in examples/tutorials/blas3.cu results in both ublas::prod and viennacl::prod to compute the wrong answer.

Surprisingly, the random number generator used to initialise the matrices generates numbers between 0 and 1, is static_cast to an int or long. This fills the matrices with zeros. I'm perplexed as to what's causing ublas's and viennacl's to report non-zero values in the result.

This has been happening for A=28x16384 B=16384x28 and A=1024x1024 B=1024x1024 (after making the changes included in myPatch.diff, with minimal changes to blas3.cu)

Any thoughts on why this is the case ?

Thanks,
Sanjay

[Karl Rupp]

Hi Sanjay,
let me have a look at this - will take 2-3 days.
Best regards,
Karli

[Sanjay]

Hello Karl,

viennacl::linalg::cuda::prod does work for int and long. I had to change the iteration bounds of the GPU-CPU result check to iterate over ublas_C, instead of ublas_A.

-Sanjay

[Karl Rupp]

Thanks, Sanjay, I used your suggestions to improve the blas3 example:
https://github.com/viennacl/viennacl-dev/commit/d38dcb39519ae49bf5398ec4906e5f5ee9e4c8cb