The CUDA Driver implementation of srad_v2 from rodinia gives a 700 error when copying the J_cuda array from device to J on the host memory

[Abhishekghosh1998]

CUDA Version: 10.1 and Driver Version: 418.226.00
GPU: Tesla P40

The issue I am facing is for the following benchmark: https://github.com/utcs-scea/ava-benchmarks/tree/master/rodinia/cuda/srad_v2

When I try to run the program as is (without remoting on the bare metal system with the setup written above), the program fails with the following error:

cuMemcpyDtoH failed: res = 700

The line that is printing the error is the following line (Note that the printf has a typo. It should be cuMemcpyHtoD instead of cuMemcpyDtoH):

ava-benchmarks/rodinia/cuda/srad_v2/srad_cuda.c

Line 329 in eefd1c3

printf("cuMemcpyHtoD failed: res = %u\n", res);

The run file used by me is the basic one:

./srad2 2048 2048 0 127 0 127 0.5 2

But this problem disappears when I slightly reduce this 2048 to a lower value, for example, 2032 (making sure to keep it a multiple of 16) [or even increase the value from 2048 to 2064 or even 4096]

I don't understand what is causing the issue. But if I comment out the first kernel launch (call to srad)here (leaving the call to srad2 as it is):

ava-benchmarks/rodinia/cuda/srad_v2/srad_cuda.c

Line 317 in eefd1c3

res = srad_launch(mod, gdx, gdy, bdx, bdy, E_C, W_C, N_C, S_C,

, the mentioned issue goes away.

Any pointers on what might be causing the issue?