unified memory allocator and pointer classes?

[rnburn]

Could you add an allocator class that uses cudaMallocManaged? This would be very useful for those implementing custom container classes.

I would image it as having a distinct pointer object (e.g. ManagedPtr) that doesn't default to either host or device execution policy but is implicitly convertible to either a built-in pointer for host execution, some sort of ManagedDevicePtr object that indicates device execution, or could be used by explicitly supplying an execution policy to the thrust algorithm functions.

Also, MangedPtr could be made to fulfill all of the requirements for a RandomAccessIterator (including having an iterator_category that's convertible to random_access_iterator_tag) so that it could be used with STL algorithm functions as well.

[brycelelbach]

Doesn't device_allocator do this?

[jaredhoberock]

device_allocator uses cudaMalloc and allocates fancy device_ptrs.

A hypothetical managed_allocator [1,2] which used cudaMallocManaged might work differently.

[1] https://github.com/jaredhoberock/managed_allocator

[2] https://github.com/agency-library/agency/blob/master/agency/cuda/memory/allocator/managed_allocator.hpp

[brycelelbach]

Ah, yes, I misunderstood.

Yes, this is on my todo list :).

[sbudenkov]

Hi,
Is it still in the backlog?

[gevtushenko]

Hello @sbudenkov! The following PR seems to have added universal_allocator.