Confusion whilst copying from host to device

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hello there,

I found an interesting error whilst attempting to copy data from a host array to a device array.

If the host and device arrays are declared to be of different extents i.e.

   integer,dimension(1000,2000)::host
   integer,allocatable,device,dimension(:,:)::device

   allocate(device(10,31))

and then an attempt is made to copy the relevant host array elements to the device array i.e.

   device(1:10,1:31)=host(1:10,1:31)

it seems that the cpu is incapable of putting the correct array elements in the correct device array elements on the gpu. All the array elements get muddled up resulting in a segfault or WORSE!

Is there some restriction here that I’m missing or is this a compiler problem? Presumably it’s because of the conversion to C when the CUDA memcpy function is called in the background which makes me think there might be a way around it???

Thanks in advance,
Crip_crop

2

I tried to recreate your example in a small test program, however, I was never able to get the example code to fail. I used a number of different compiler(11.1 - 12.5) and they all worked correctly. If you have a failing case that you can post, and what version of the compiler used, that would be very helpful.

3

Hi sorry for the late reply. I managed to fix the problem… definitely my fault. I had a very well disguised mapping bug but it’s fixed now and the copy from host to device works a treat.

Cheers anyways,
Crip_crop