Path: csiph.com!eternal-september.org!reader02.eternal-september.org!.POSTED!not-for-mail
From: Tim Rentsch <tr.17687@z991.linuxsc.com>
Newsgroups: comp.lang.c++
Subject: Re: A better benchmark
Date: Mon, 07 Feb 2022 06:57:31 -0800
Organization: A noiseless patient Spider
Lines: 182
Message-ID: <86y22miv1g.fsf@linuxsc.com>
References: <stj7m4$bim$1@dont-email.me> <stlq6s$8pc$1@dont-email.me> <stok2f$eg7$1@dont-email.me>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Injection-Info: reader02.eternal-september.org; posting-host="792c2a82c28252afd7e8a94f82466189"; logging-data="23609"; mail-complaints-to="abuse@eternal-september.org";	posting-account="U2FsdGVkX18ISFOcZ1PgtQyOkanfJ7mgS/lZgNb6GFY="
User-Agent: Gnus/5.11 (Gnus v5.11) Emacs/22.4 (gnu/linux)
Cancel-Lock: sha1:1P8OyX1Qr46NnGcVx7d/sGOH1ik= sha1:siHTdJ61hHVNVOJLINNVQM5gECI=
Xref: csiph.com comp.lang.c++:82956

Bonita Montero <Bonita.Montero@gmail.com> writes:

[.. some minor editing to fix line-wrapping problems ..]

> Now I have an even distribution from 1 to 20 digits.
>
> #include <iostream>
> #include <concepts>
> #include <array>
> #include <cstdio>
> #include <cstring>
> #include <chrono>
> #include <vector>
> #include <random>
>
> using namespace std;
> using namespace chrono;
>
> #if defined(_MSC_VER)
>     #pragma warning(disable : 4996) // disable deprecated sprintf
>     #pragma warning(disable : 6200) // array out of range
> #endif
>
> template<typename char_t>
>     requires same_as<char_t, char> || same_as<char_t, wchar_t>
> char_t const *formatClockCyclesMine( uint64_t clockCycles );
> char const *formatClockCyclesTim( uint64_t clockCycles );
> char const *formatClockCyclesBen( uint64_t clockCycles );
>
> int main()
> {
>     auto bench = []<typename FccFn>( FccFn fn )
>         requires
>             requires( FccFn fn, uint64_t clockCycles ) {
>                 { *fn( clockCycles ) } -> convertible_to<char>;
              }
>     {
>         static size_t const N_VALUES = 10'000, N = 1'000;
>         vector<uint64_t> values( N_VALUES );
>         mt19937_64 mt;
>         uniform_int_distribution<size_t> uidLength( 1, 20 );
>         uniform_int_distribution<unsigned> uidDigit( 0, 9 );
>         auto getNumber = [&]() -> uint64_t
>         {
>             uint64_t number = 0, nextNumber;
>             size_t initialLength = uidLength( mt ), length = initialLength;
>             unsigned digit;
>             do
>                 if( 
>                     (nextNumber = number * 10 + (digit = uidDigit( mt )))
>                         >= number
>                 )
>                     number = nextNumber;
>                 else
>                     number = 0,
>                     length = initialLength + 1;
>             while( --length );
>             return number;
>         };
>         for( uint64_t &v : values )
>             v = getNumber();
>         char volatile sum = 0;
>         auto start = high_resolution_clock::now();
>         for( size_t n = N; n--; )
>             for( uint64_t v : values )
>                 sum += *fn( v );
>         return
>             (double)(int64_t)duration_cast<nanoseconds>(
>                  high_resolution_clock::now() - start
>              ).count() / ((double)N * (double)N_VALUES);
>     };
>     cout << "mine: " << bench( formatClockCyclesMine<char> ) << endl;
>     cout << "Tim:  " << bench( formatClockCyclesTim ) << endl;
>     cout << "Ben:  " << bench( formatClockCyclesBen ) << endl;
> }
>
> #if defined(_MSC_VER)
>     #define NOINLINE __declspec(noinline)
> #elif defined(__GNUC__) || defined(__llvm__)
>     #define NOINLINE __attribute__((noinline))
> #elif
>     #define NOINLINE
> #endif
>
> template<typename char_t>
>     requires same_as<char_t, char> || same_as<char_t, wchar_t>
> NOINLINE
> char_t const *formatClockCyclesMine( uint64_t clockCycles )
> {
>     using dig_arr_t = array<uint8_t, 20>; // max 20 digits for 64 bit
>     using dig_arr_it = typename dig_arr_t::iterator;
>     dig_arr_t reverseDigits;
>     dig_arr_it rDigitsEnd = reverseDigits.begin();
>     do
>         *rDigitsEnd++ = clockCycles % 10,
>         clockCycles /= 10;
>     while( clockCycles );
>     size_t
>         digitsLen = rDigitsEnd - reverseDigits.begin(),
>         groups = (digitsLen - 1) / 3;
>     using str_arr_t = array<char_t, 20 + (20 - 1) / 3 + 1>;
>     using str_arr_it = typename str_arr_t::iterator;
>     thread_local str_arr_t str;
>     str_arr_it wrt = str.begin() + digitsLen + groups;
>     *wrt = '\0';
>     dig_arr_it digit = reverseDigits.begin();
>     for(
>         dig_arr_it groupsEnd = reverseDigits.begin() + groups * 3;
>         digit != groupsEnd;
>         wrt -= 4, digit += 3
>      )
>         wrt[-1] = digit[0] + '0',
>         wrt[-2] = digit[1] + '0',
>         wrt[-3] = digit[2] + '0',
>         wrt[-4] = '.';
>     do
>         *--wrt = *digit++ + '0';
>     while( digit != rDigitsEnd );
>     return &*wrt;
> }
>
> NOINLINE
> char const *formatClockCyclesTim( uint64_t clockCycles )
> {
>     thread_local char result[20 + (20 - 1) / 3 + 1];
>     size_t n =
>         (unsigned)sprintf( result, "%llu",
>             (unsigned long long) clockCycles );
>     char *ep = (&result)[1];
>     *--ep = 0;
>     while( n > 3 )
>         memmove( ep -= 3, &result[n -= 3], 3 ),
>         *--ep = '.';
>     do
>         *--ep = result[--n];
>     while( n );
>     return ep;
> }
>
> NOINLINE
> char const *formatClockCyclesBen( uint64_t clockCycles )
> {
>     thread_local char result[20 + (20 - 1) / 3 + 1];
>     char *ep = (&result)[1];
>     do
>     {
>         // initially with a missing cast, actually would have
>         // run only on little endian machines
>         sprintf( ep - 4, "%03lu", (unsigned long)(clockCycles % 1000) );
>         if( ep != (&result)[1] )
>             ep[-1] = '.';
>         ep -= 4;
>     } while( clockCycles /= 1000 );
>     for( ; *ep == '0' && ep[1]; ++ep );
>     return ep;
> }
>
> These are the resuls:
>
> MSVC 2022:
>     mine: 29.8306
>     Tim:  109.97
>     Ben:  320.64
>
> g++ 11.1.0:
>     mine: 28.3386
>     Tim:  110.586
>     Ben:  317.135
>
> clang++ 10.0.0:
>     mine: 29.7089
>     Tim:  100.721
>     Ben:  311.768

The code that I posted was meant only to show an improvement on
the code in the posting to which I was responding.  It wasn't
meant as any kind of comparison with any code you posted earlier.

After seeing the posting above, it wasn't hard to write a
simple routine that is both shorter and faster than the
formatClockCyclesMine() function shown above (using the code
shown above in main() to do the performance testing).