Skip to content

[FEA] RMM should not pad allocations #865

Closed
#1278
Closed
[FEA] RMM should not pad allocations#865
#1278
Assignees
vyasr
Labels
0 - BacklogIn queue waiting for assignmentfeature requestNew feature or requesttech debtdebt Internal clean up and improvements to reduce maintenance and technical debt in general
@harrism

Description

@harrism
harrism
opened on Sep 8, 2021

Is your feature request related to a problem? Please describe.
Currently, RMM pads every allocation that goes through the rmm::device_memory_resource interface to a multiple of 8 bytes in size.

rmm/include/rmm/mr/device/device_memory_resource.hpp

Lines 102 to 105 in 8527317

void* allocate(std::size_t bytes, cuda_stream_view stream = cuda_stream_view{})
{
return do_allocate(rmm::detail::align_up(bytes, 8), stream);
}

This was originally added in #165, which doesn't provide explanation of why. I believe it was to allow accessing structures via aliases in cuIO in a way that would otherwise be UB.

Describe the solution you'd like
We should not pad allocations unnecessarily. RMM allocation should not have surprising behavior like this.

Metadata

Metadata

Assignees

Labels

0 - BacklogIn queue waiting for assignmentfeature requestNew feature or requesttech debtdebt Internal clean up and improvements to reduce maintenance and technical debt in general

Type

No type

Projects

RMM Project Board

Status

Done

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions