temporary buffer allocation about rocprim HOT 6 CLOSED

total_size += sizeof(int) * some_other_size * 3; - this might be the problem. Alignment might be wrong. hipcub_buffer should have the same alignment as if it was allocated directly with hipMalloc.

from rocprim.

Then it should work, if the hipcub_buffer is the first chunk of buffer data in the row, right?

from rocprim.

I guess. I don't see any other problem right now, but only looked at your code. If you are allocating multiple buffers with one hipMalloc I'd recommend aligning every buffer to 256 bytes. It's not much and it helps.

from rocprim.

Works, thanks!
Btw: Buffer size returned by rocprim is not a multiple of 256.
I think I remember that for some nvidia cards pointer alignment is not always 256, do you happen to have any information on that or am I wrong.

from rocprim.

Btw: Buffer size returned by rocprim is not a multiple of 256.

That's true. Temporary buffer size does not have to be a multiple of 256. However, its pointer may need to be aligned to 256 bytes.

I think I remember that for some nvidia cards pointer alignment is not always 256, do you happen to have any information on that or am I wrong.

From NVIDIA docs: "Any address of a variable residing in global memory or returned by one of the memory allocation routines from the driver or runtime API is always aligned to at least 256 bytes." It might be more on newer GPUs, but 256 bytes is good enough for everything to work correctly.

from rocprim.

Great, thanks a lot!

from rocprim.