math.hpp File Reference

Scalar math helpers usable in constant expressions. More...

#include <cmath>
#include <tdls/core/macros.hpp>
#include <tdls/core/target.hpp>

Include dependency graph for math.hpp:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Functions
template<typename T >
TDLS_HOST_DEVICE TDLS_FORCE_INLINE_GPU constexpr T	tdls::detail::abs (const T x) noexcept
	Absolute value usable in constant expressions.
TDLS_HOST_DEVICE TDLS_FORCE_INLINE_GPU constexpr int	tdls::detail::runtime_dimension (int v) noexcept
	A dimension hidden from the optimizer of nvcc when it compiles device code for Blackwell (sm_100 and later), the identity elsewhere.

Detailed Description

Scalar math helpers usable in constant expressions.

Author

Tristan Chenaille

Copyright

Copyright (C) 2026 CEA. All rights reserved. This project is publicly released under the BSD 3-Clause License (see the LICENSE file). CEA may also distribute it under specific licensing conditions.

Function Documentation

abs()

template<typename T >

TDLS_HOST_DEVICE TDLS_FORCE_INLINE_GPU constexpr T tdls::detail::abs ( const T x )

nodiscardconstexprnoexcept

Absolute value usable in constant expressions.

std::fabs is not constexpr before C++23, so constant evaluation falls back to a ternary, indistinguishable from std::fabs in the magnitude comparisons of the solvers (signed zeros compare equal, NaNs propagate either way). At run time the function IS std::fabs: the naive ternary alone would forbid the single sign-bit instruction (it must preserve -0.0) and was measured 3% heavier in device code on the pivot search.

Template Parameters

T	scalar type

Parameters

[in]

x

value

Returns

the absolute value of x

runtime_dimension()

TDLS_HOST_DEVICE TDLS_FORCE_INLINE_GPU constexpr int tdls::detail::runtime_dimension ( int v )

constexprnoexcept

A dimension hidden from the optimizer of nvcc when it compiles device code for Blackwell (sm_100 and later), the identity elsewhere.

nvcc miscompiles TiledLUppSolverDynamic for these targets when a dimension passed by the caller is a compile-time constant and propagates into its loops (TDLS issue #16). Every entry point of that solver passes its dimensions through this function first, so that the solver is compiled for a runtime dimension, as it is meant to be. The runtime solvers of CooperativeLUpp and ArrowheadLUpp were found correct; ArrowheadLUpp reaches TiledLUpp through these entry points. nvc++ compiles the solver correctly for these targets (checked with -stdpar=gpu for cc120), so the statement is limited to nvcc. The empty assembly statement costs no instruction.

Parameters

[in]

v

dimension

Returns

v