CUDA malloc() minimum and typical actual alignment

Question

The allocated memory is suitably aligned for any kind of variable.

But...

What affects the actual aligment? Compute capability? CUDA driver version? The specific kind of card? The allocation size?
Can I determine the minimum / typical allocation alignment as a function of these parameters?

@DmitriBudnikov: It seems to me the allocation quantum is significantly larger than `alignof(long long)`... isn't it? — einpoklum, Apr 11 '16 at 06:47
@DmitriBudnikov: Sorry, I meant aligned by a larger quantum. I notice quite a few 0's at the lower bits of allocated areas. — einpoklum, Apr 11 '16 at 07:39
Relevant: http://stackoverflow.com/a/14083295/2386951. Also consider @njuffa's comment on that. — Farzad, Apr 12 '16 at 05:58

einpoklum · Answer 1 · 2017-08-04T10:42:30.220

You can un-safely assume it's at least cudaDeviceProp::textureAlignment
(i.e. 256 on Fermi, 512 on Kepler, Maxwell).

@sgarizvi reports that, from his experimentation on multiple devices, that the alignment of allocated blocks of device memory is no less than the texture alignment field of the device properties ( cudaDeviceProp::textureAlignment). For Kepler and Maxwell devices this is 512 bytes.

Of course, like @talonmies notes, this is not actually guaranteed nor is it documented.

CUDA malloc() minimum and typical actual alignment

1 Answers1