Add support for int32 parameter axes

[tgujar]

NVBench integer axes return int64 values. Projects using 32-bit size types must repeatedly check that an axis value fits before casting it.

Could NVBench provide a checked int32 accessor, for example:

auto value = state.get_int32("num_rows");

The axis can remain int64 internally. The accessor should return an int32 value when it fits and report an error when it does not. This would also safely handle values supplied through command-line axis overrides.

[oleksandr-pavlyk]

I think templated getters may be useful for C++ users:

• template <typename IntegralT> auto get_integral(const char * name) -> IntegralT
• template <typename FloatingPointT> auto get_floating_point(const char * name) -> FloatingPointT

The integral getter can perform runtime bounds checking and then do return static_cast<IntegralT>(value_i64);, the floating point getter would just do a static assertion followed by return static_cast<FloatingPointT>(value_f64);.

[lktr-wq]

Hi @oleksandr-pavlyk, thanks for the suggestion. I built the current main branch locally and all 46 tests passed. I’d like to work on this. Just to confirm, should I implement both get_integral and get_floating_point as part of this issue?

[bernhardmgruber]

Would cuda::narrow<int>(state.get_int64("num_rows")) work for you?