Version 0.6.8: Launch config builder improvement, CUDA 12.x kernel-library support and launch attributes

[eyalroz]

Changes since v0.6.7:

Build process & build configuration changes

• Have all targets mark dependencies of CUDA library dependencies which NVIDIA ignores #583 Exported targets now make sure apps link against some system libraries CUDA depends on (to circumvent CMake bug 25665).
• Want alias for runtime-and-driver #567 The runtime-and-driver target now has driver-and-runtime as an alias

Launch configuration & launch config builder changes

• Full representation of launch configurations + launch-with-full-config support #564 Can now launch kernels with the full range of CUDA 12.x launch attributes, including remote memory sync domain, programmatic launch dependence and programmatic completion events (see descriptions in the CUDA Driver API documentation).
• Check overall dimensions are not too large to be expressed with our grid and block dimension limits #577, Improve launch configuration validation #582 More extensive validation of launch configurations when building them with the launch config builder gadget.
• Robustify dimension comparison operators #581 More robust comparison operators for dimension structures
• Support use_max_active_blocks_per_multiprocessor() on a config builder #580 Launch config builder can now be told to the "use the maximum number of active blocks per multiprocessor".
• config builder is missing a method for setting the device #579 User can now set a target device on a launch configuration target device without setting a contextualized kernel, in a launch config builder
• Use the launch config builder in more examples #578 Now using the launch config builder in more of the example programs
• Leftover warning regarding validate_all_dimensions_compatibility not being used #569 Took care of unused validation function which was triggering a warning with newer compilers

CUDA libraries and in-library, non-associated kernel support

• Support CUlibrary's and CUkernel's - CUDA driver low-level "library management" #565 Now supporting "CUDA libraries" - files or blocks of data in memory containing compiled kernels, which are not loaded immediately into modules within contexts; and contain device-and-context-independent compiled kernels. Both of these can now be represented and worked with.
• module proxy class doesn't need to hold its link options #576 A module no longer holds the link options it was created with; those are not essential to its use, and at times are impossible to (re)create when obtaining a module from a library (which also doesn't hold its link options ).

Other changes

• Move apriori_compiled_kernel_t into the kernel namespace #575 Moved apriori_compiled_kernel_t into the kernel namespace, yielding kernel::apriori_compiled_it
• Redundant definition of cuda::module::handle_t #570, Remove redundant inclusions of NVIDIA headers #573 Removed some redundant inclusions and definitions
• cuda::kernel_t::set_attribute is broken! #568 Fixed some breakage of kernel_t::set_attribute()
• Want ability to set function attributes on function pointers and handles #566 Can now properly get and set properties on kernels (raw functions and handles)

Compatibility

• Compatibility broken for CUDA 9.x #572 Fixed broken CUDA 9.x compatibility

---

• Want to help me with testing 0.7? Drop me a line... (it will have CUDA execution graph support)
• Have you tried this version and are satisfied with the changes? Thumb-up just below this line to let others know as well.

---

This discussion was created from the release Version 0.6.8: Launch config builder improvement, CUDA 12.x kernel-library support and launch attributes.