Need to check against Device().properties.memory_pools_supported to decide if we use DeviceMemoryResource or _SynchronousMemoryResource
Any chance we can turn this into a wrapper function, or just have DeviceMemoryResource() return the appropriate type when constructed?
Originally posted by @shwina in #717 (comment)
We could consider adding a helper function under the cuda.core.experimental.utils module. Later we might reimplement DeviceMemoryResource using cudax::device_memory_resource or cuda::device_memory_pool, so if we make DeviceMemoryResource to wrap both cudaMallocFromPoolAsync and cudaMalloc now, the re-implementation would become harder later.
Originally posted by @shwina in #717 (comment)
We could consider adding a helper function under the
cuda.core.experimental.utilsmodule. Later we might reimplementDeviceMemoryResourceusingcudax::device_memory_resourceorcuda::device_memory_pool, so if we makeDeviceMemoryResourceto wrap bothcudaMallocFromPoolAsyncandcudaMallocnow, the re-implementation would become harder later.