support RMM aligned resource adapter in JNI [skip ci] #8266
Conversation
rongou
added
3 - Ready for Review
java/src/main/native/src/RmmJni.cpp
Outdated
| Initialized_resource, allocation_alignment, alignment_threshold); | ||
| } | ||
|
|
||
| auto wrapped = make_tracking_adaptor(Initialized_resource.get(), RMM_ALLOC_SIZE_ALIGNMENT); |
There was a problem hiding this comment.
RMM_ALLOC_SIZE_ALIGNMENT seems wrong if we know we're using the aligned adapter and a different alignment. I think there needs to be a max(RMM_ALLOC_SIZE_ALIGNMENT, allocation_alignment) or something similar here.
There was a problem hiding this comment.
Hmm, the actual allocation size for aligned adapter is a bit complicated. Just curious, why are we tracking the total allocation size ourselves and not using the get_mem_info() method from the device memory resource?
There was a problem hiding this comment.
Because many device memory resource implementations do not implement get_mem_info(), the ARENA allocator apparently being one of them:
$ git grep supports_get_mem_info | grep false
benchmarks/utilities/simulated_memory_resource.hpp: bool supports_get_mem_info() const noexcept override { return false; }
include/rmm/mr/device/arena_memory_resource.hpp: bool supports_get_mem_info() const noexcept override { return false; }
include/rmm/mr/device/binning_memory_resource.hpp: bool supports_get_mem_info() const noexcept override { return false; }
include/rmm/mr/device/cuda_async_memory_resource.hpp: bool supports_get_mem_info() const noexcept override { return false; }
include/rmm/mr/device/fixed_size_memory_resource.hpp: bool supports_get_mem_info() const noexcept override { return false; }
include/rmm/mr/device/pool_memory_resource.hpp: bool supports_get_mem_info() const noexcept override { return false; }
There was a problem hiding this comment.
Yeah I guess that api is really only for cuda. Added the max of two alignment sizes.
jlowe
changed the title
|
@gpucibot merge |
Depends on rapidsai/rmm#768.