IS updates about npb-gpu HOT 9 OPEN

Hello.

The verification routine is failing because an overflow error is occurring with class D (int type is not large enough for this class).

We will fix it as soon as possible.

Thank you for using the NPB-GPU, and thank you for reporting this bug.

from npb-gpu.

Thank you for explaining the error.

from npb-gpu.

Is there a upper bound to the value of MAX_ITERATION ? The default value is 10.

from npb-gpu.

Hello.

Yes, in the practice, there is an upper bound. However, the NPB documentation does not explain which is the maximum amount of iterations supported (it may vary according to the hardware).

Our implementation followed the original NPB version, and the default value of MAX_ITERATIONS set by NASA is 10.

On my machine, I tested other values, and the benchmark worked with up to 24 iterations. With an amount of iterations larger than 24, the IS benchmark started failing in the correctness verification, even when using long int type.

from npb-gpu.

Using long int type implies the class D problem. Have you updated your program to support the class D problem ?

Thanks

from npb-gpu.

We haven't had time to fix this issue yet.

from npb-gpu.

Hello.

We fixed the IS benchmark issue with class D.

However, class D consumes about 19 gigabytes of GPU memory. Thus, a GPU needs at least 19 gigabytes of memory to work correctly with this workload.

from npb-gpu.

A memory check may be added to the CUDA program so that the test is skipped without enough memory. Thanks.

from npb-gpu.

A memory check may be added to the CUDA program so that the test is skipped without enough memory. Thanks.

It is a good suggestion. We will add this feature soon.

from npb-gpu.