hpc$\_$helpers.hpp code:

#ifndef HPC$\_$HELPERS$\_$HPP

#define HPC$\_$HELPERS$\_$HPP

#include

#include

#ifndef $\_\_$CUDACC$\_\_$

#include

#endif

#ifndef $\_\_$CUDACC$\_\_$

#define TIMERSTART$($label$)\backslash$

std::chrono::time$\_$point a##label, b##label; $\backslash$

a##label $=$ std::chrono::system$\_$clock::now$()$;

#else

#define TIMERSTART$($label$)\backslash$

cudaEvent$\_$t start##label, stop##label; $\backslash$

float time##label; $\backslash$

cudaEventCreate$($&start##label$)$; $\backslash$

cudaEventCreate$($&stop##label$)$; $\backslash$

cudaEventRecord$($start##label, $0)$;

#endif

#ifndef $\_\_$CUDACC$\_\_$

#define TIMERSTOP$($label$)\backslash$

b##label $=$ std::chrono::system$\_$clock::now$()$; $\backslash$

std::chrono::duration delta##label $=$ b##label$-$a##label; $\backslash$

std::cout # elapsed time $("$ #label : $"\backslash$

$$ delta##label.counts$"$ std::endl;

#else

#define TIMERSTOP$($label$)\backslash$

cudaEventRecord$($stop##label, $0)$; $\backslash$

cudaEventSynchronize$($stop##label$)$; $\backslash$

cudaEventElapsedTime$($&time##label, start##label, stop##label$)$; $\backslash$

std::cout $$ "TIMING: $"$ time##label ms $("$ #label

$$ std::endl;

#endif

#ifdef $\_\_$CUDACC$\_\_$

#define CUERR $\{\backslash$

cudaError$\_$t err; $\backslash$

if $(($err $=$ cudaGetLastError$())!=$ cudaSuccess$)\{\backslash$

std::cout $$ "CUDA error: $"$ cudaGetErrorString$($err : $"\backslash$

FILE line LINE$\_\_$ std::endl; $\backslash$

exit$(1)$; $\backslash$

$\}\backslash$

$\}$

$//$ transfer constants

#define H$2$D $($cudaMemcpyHostToDevice$)$

#define D$2$H $($cudaMemcpyDeviceToHost$)$

#define H$2$H $($cudaMemcpyHostToHost$)$

#define D$2$D $($cudaMemcpyDeviceToDevice$)$

#endif

$//$ safe division

#define SDIV$($x$,$y$)((($x$)+($y$)-1)/($y$))$

$//$ no$\_$init$\_$t

#include

template

class no$\_$init$\_$t $\{$

public:

static$\_$assert$($std::is$\_$fundamental::value &&

std::is$\_$arithmetic::value,

"wrapped type must be a fundamental, numeric type"$)$;

$//$do nothing

constexpr no$\_$init$\_$t$()$ noexcept $\{\}$

$//$convertible from a T

constexpr no$\_$init$\_$t$($T value$)$ noexcept: v$\_($value$)\{\}$

$//$act as a T in all conversion contexts

constexpr operator T $()$ const noexcept $\{$ return v$\_$; $\}$

$//$ negation on value and bit level

constexpr no$\_$init$\_$t& operator $-()$ noexcept $\{$ v$\_=-$v$\_$; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator ~ $()$ noexcept $\{$ v$\_=$ ~v$\_$; return $*$this; $\}$

$//$ prefix increment$/$decrement operators

constexpr no$\_$init$\_$t& operator $++()$ noexcept $\{$ v$\_++$; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $--()$ noexcept $\{$ v$\_--$; return $*$this; $\}$

$//$ postfix increment$/$decrement operators

constexpr no$\_$init$\_$t operator $++($int$)$ noexcept $\{$

auto old$(*$this$)$;

v$\_++$;

return old;

$\}$

constexpr no$\_$init$\_$t operator $--($int$)$ noexcept $\{$

auto old$(*$this$)$;

v$\_--$;

return old;

$\}$

$//$ assignment operators

constexpr no$\_$init$\_$t& operator $+=($T v$)$ noexcept $\{$ v$\_+=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $-=($T v$)$ noexcept $\{$ v$\_-=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $*=($T v$)$ noexcept $\{$ v$\_*=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $/=($T v$)$ noexcept $\{$ v$\_/=$ v; return $*$this; $\}$

$//$ bit$-$wise operators

constexpr no$\_$init$\_$t& operator &$=($T v$)$ noexcept $\{$ v$\_$ &$=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $|=($T v$)$ noexcept $\{$ v$\_|=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $^=($T v$)$ noexcept $\{$ v$\_^=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator $>>=($T v$)$ noexcept $\{$ v$\_>>=$ v; return $*$this; $\}$

constexpr no$\_$init$\_$t& operator T v$)$ noexcept $\{$ v v; return $*$this; $\}$

private:

T v$\_$;

$\}$;

#endif