Every kernel takes plain pointers, so any CPU-accessible memory works: mmap, pinned buffers, CUDA unified memory, custom arenas. C++ views wrap any pointer without ownership. Owning containers accept any C++ Allocator.
template <typename T>
struct cuda_allocator {
using value_type = T;
T *allocate(std::size_t n) { T *p;
cudaMallocManaged(&p, n * sizeof(T), cudaMemAttachGlobal);
return p; }
void deallocate(T *p, std::size_t) noexcept { cudaFree(p); }
};
nk_dot_f32_best(cuda_managed_ptr, cuda_managed_ptr, 1024, &dot, capabilities, NULL); // C ABI, any pointer
auto view = nk::tensor_view<nk::f32_t>(mmap_ptr, rows, columns); // non-owning view
auto v = nk::vector<float, cuda_allocator<float>>::zeros(1024); // allocator-aware owningaligned_allocatorCache-aligned allocator with non-throwing allocation.template <typename value_type_,
std::size_t alignment_ = nk_default_alignment_k>
struct aligned_allocatorAllocation returns nullptr on failure instead of throwing. It is stateless and always compares equal.
alignment | std::size_t |
|---|
allocatorAdapts a C allocation policy and caller-owned stream to the container allocator interface.template <typename value_type_>
struct allocatorpolicy | nk_allocator_t | |
|---|---|---|
stream | nk_stream_t |