CUDA allocators

Using allocators with pool_allocator

Allocators are needed when working with pool_allocator. There are 1D, 2D and 3D allocators for (pinned) host memory and device memory - the latter is pitched when using a 2D or 3D memory layout. They are also not interchangeable; a 1D allocator will never be able to allocate 2D memory and vice versa.

To declare them, pass the desired type and the memory layout as template parameters:

#include <glados/cuda/memory.h>

using host_allocator_1D = glados::cuda::host_allocator<float, glados::memory_layout::pointer_1D>;
using device_allocator_2D = glados::cuda::device_allocator<float, glados::memory_layout::pointer_2D>;

using pool_allocator_host_1D = glados::pool_allocator<float, glados::memory_layout::pointer_1D, host_allocator_1D>;
using pool_allocator_device_2D = glados::pool_allocator<float, glados::memory_layout::pointer_2D, device_allocator_2D>;

Other use cases

GLADOS' allocators are designed to operate with GLADOS' pool_allocator - they have never been tested as a replacement for the STL's allocator. For almost all cases users will never need to call their member functions directly - the CUDA memory management section of this wiki will suffice. DO NOT CONTINUE TO FOLLOW THIS PATH unless you know exactly what you are doing. Using GLADOS' allocators in a program is error-prone and likely to cause memory leaks. You have been warned.

However, their interface is very similar to std::allocator while offering (template) specializations for 2D and 3D objects. If you know std::allocator you will easily grasp the concept of these allocators (just add the extra dimensions as needed).