cexa-project / shaman Goto Github PK

Shaman requires std::is_arithmeticwhich is only fulfilled by std::bfloat16_t since C++23 and more annoyingly cannot be changed on the user side.

https://en.cppreference.com/w/cpp/types/is_floating_point

Checks whether T is a floating-point type. Provides the member constant value which is equal to true, if T is the type float, double, long double, or any extended floating-point types (std::float16_t, std::float32_t, std::float64_t, std::float128_t, or std::bfloat16_t)(since C++23), including any cv-qualified variants. Otherwise, value is equal to false.

The behavior of a program that adds specializations for std::is_floating_point or std::is_floating_point_v is undefined.

Instead of replacing standard floating point type by Shamman version, make the Shamman version a type alias that refers to either standard floating point or Shamman floating point. The actual nature of the alias should be selected through a single macro.

The idea is to let client code switch between standard and shamman type by changing its code at a single point.

Propose a way to ensure that IEEE floating point semantic is strictly followed by compiler.

Old Kokkos helpers do not work with current versions of Kokkos.

'Kokkos_ArithTraits.hpp' file not found

Related to #2

Add Kokkos macros to make Shaman type portable accross Kokkos backends

• Shaman types have to use Kokkos:: namespace for mathematical functions

Mandatory Checks:

• Compilation on CPU / gcc
• Compilation on CPU / clang
• Compilation on GPU / Nvidia
• Compilation on GPU / AMD
• Compatibility with Kokkos Views

Should be tested with at least examples from #1

Shaman should stay in a Shaman:: namespace.

We may want to translate the SummerSchool example to Kokkos.

• With basic arithmetic
• With mathematical functions
• With mathematical functions from the recommended Kokkos:: namespace

https://en.cppreference.com/w/cpp/types/is_floating_point

The behavior of a program that adds specializations for std::is_floating_point or std::is_floating_point_v is undefined.