Todo List

Member cuda::context_t::maximum_depth_of_child_grid_sync_calls () const

Is this really a feature of the context? Not of the device?

Member cuda::context_t::maximum_outstanding_kernel_launches () const

Is this really a feature of the context? Not of the device?

Member cuda::context_t::stack_size (context::limit_value_t new_value) const

Verify this is in bytes!

Member cuda::device::peer_to_peer::disable_access (const device_t &accessor, const device_t &peer)

Consider disabling this, given that access is context-specific

Member cuda::device::peer_to_peer::enable_access (const device_t &accessor, const device_t &peer)

Consider disabling this, given that access is context-specific

Member cuda::device_t::memory () const

Consider a scoped/unscoped dichotomy.

Member cuda::device_t::properties () const

get rid of this in favor of individual properties only.

Member cuda::device_t::reset () const

Determine whether this actually performs a hardware reset or not

Member cuda::grid::block_dimensions_t

Consider having both grid and block dims inhert from the same dimensions_t structure, but be incompatible, to prevent mis-casting one as the other.

Member cuda::link::wrap (device::id_t device_id, context::handle_t context_handle, link::handle_t handle, const link::options_t &options, bool take_ownership=false) noexcept

: Consider checking if the linking has already been completed!

Member cuda::memory::mapped::is_part_of_a_region_pair (const void *ptr)

What if it's a managed pointer?

Member cuda::memory::shared::size_t

consider using uint32_t.

Class cuda::memory::unique_region< Deleter >

: Should we really expose the region parent class? We could, instead, have a get() or region() method which provides lvalue-ref access to it.

: Should this be called a unique_region, a-la-unique_ptr? Or perhaps unique_ptr is a misnomer, and should have been called owned_ptr (as opposed to shared_ptr), and regardless, this should be called an owned_region?

Member cuda::module::load_from_file (const context_t &context, const char *path)

consider adding load_module methods to context_t

When switching to the C++17 standard, use string_view's instead of the const char* and ::std::string reference

Member cuda::rtc::compilation_options_base_t< Kind >::targets_

Use something less fancy than ::std::unordered_set, e.g. a vector-backed ordered set or a dynamic bit-vector for membership.

Member cuda::rtc::compilation_options_t< cuda_cpp >::additional_include_paths

In C++17, consider making the elements ::std::filesystem::path's.

Member cuda::rtc::compilation_options_t< cuda_cpp >::compile_extensible_whole_program

explain what that is.

Member cuda::rtc::compilation_options_t< cuda_cpp >::preinclude_files

In C++17, consider making the elements ::std::filesystem::path's.

Check how these strings are interpreted. Do they need quotation marks? brackets? full paths?

Class cuda::runtime_error

Consider renaming this to avoid confusion with the CUDA Runtime.

Member cuda::stream_t::has_work_remaining () const

What if there are incomplete operations, but they're all waiting on something on another queue? Should the queue count as "busy" then?

Class cuda::texture::descriptor_t

Could be expanded into a richer wrapper class allowing actual settings of the various fields.

Member cuda::version_numbers::driver ()

In future CUDA versions which support C++17 - return an optional

Member cuda::wait (const event_t &event)

figure out what happens if the event has not been recorded before this call is made.

File pointer.hpp

Consider allowing for storing attributes within the class, lazily (e.g. with an ::std::optional).

File template.hpp

Implement batch mem op insertion and param setting: cuGraphAddBatchMemOpNode, cuGraphBatchMemOpNodeGetParams, cuGraphBatchMemOpNodeSetParams