constexpr and performance result

Mac1ek · 03-13-2024, 04:38 AM

Hi Why constexpr in function giving extremely high performance when i compare to plain function ?

example:

Code:

constexpr auto fib_constexpr(unsigned int n) -> unsigned int {
	return (n <= 1) ? n : fib_constexpr(n - 1) + fib_constexpr(n - 2); 
}

auto fib_normal(unsigned int n) -> unsigned int {
	return (n <= 1) ? n : fib_normal(n - 1) + fib_normal(n - 2);
}

this constexpr function giving whole different result in performance

Thank You for answer.

pan64 · 03-13-2024, 05:18 AM

would be nice to demonstrate it somehow. Or at least explain how did you measure it?

Mac1ek · 03-13-2024, 05:57 AM

Quote:

Originally Posted by pan64

would be nice to demonstrate it somehow. Or at least explain how did you measure it?

Il write simple benchmark for this, i not shure is correct:

Code:

#include <iostream>
#include <chrono>

constexpr auto fib_constexpr(unsigned int n) -> unsigned int {
	return (n <= 1) ? n : fib_constexpr(n - 1) + fib_constexpr(n - 2); 
}

auto fib_normal(unsigned int n) -> unsigned int {
	return (n <= 1) ? n : fib_normal(n - 1) + fib_normal(n - 2);
}

auto benchmark_fib_constexpr() -> void {

	auto start = std::chrono::system_clock::now();
	auto end = std::chrono::system_clock::now();

	unsigned int count = 0;
	constexpr unsigned int n = 20;

	for(unsigned int i = 0; i < 10;) {
		fib_constexpr(n);
		++count;
		end = std::chrono::system_clock::now();
		if(std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() >= 1000) {
			std::cout<<"In second operation count is: "<<count<<std::endl;
			count = 0;
			++i;
			start = std::chrono::system_clock::now();
		}

	}	
}

auto benchmark_fib_normal() -> void {

	auto start = std::chrono::system_clock::now();
	auto end = std::chrono::system_clock::now();

	unsigned int count = 0;
	unsigned int n = 20;

	for(unsigned int i = 0; i < 10;) {
		fib_normal(n);
		++count;
		end = std::chrono::system_clock::now();
		if(std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() >= 1000) {
			std::cout<<"In second operations count is: "<<count<<std::endl;
			count = 0;
			++i;
			start = std::chrono::system_clock::now();
		}
	}
}


auto main() -> int {
	std::cout<<"CONSTEXPR FIB BENCHMARK"<<std::endl;
	benchmark_fib_constexpr();
	std::cout<<"NORMAL FIB BENCHMARK"<<std::endl;
	benchmark_fib_normal();
	return (0);
}

important is compile with -O2 flag:

Code:

g++ -std=c++11 -O2 test.cpp

then in terminal i have result:

Code:

CONSTEXPR FIB BENCHMARK
In second operation count is: 49968765
In second operation count is: 49981539
In second operation count is: 50005027
In second operation count is: 49964753
In second operation count is: 49948262
In second operation count is: 49992233
In second operation count is: 49954760
In second operation count is: 49976978
In second operation count is: 49989060
In second operation count is: 49987518
NORMAL FIB BENCHMARK
In second operations count is: 131435
In second operations count is: 131457
In second operations count is: 131451
In second operations count is: 130226
In second operations count is: 130350
In second operations count is: 130034
In second operations count is: 129069
In second operations count is: 129363
In second operations count is: 120881
In second operations count is: 126952

NevemTeve · 03-13-2024, 06:17 AM

That could have been entirely optimized away. The standard benchark method is calculating fib(largeint) modulo 1000000009

pan64 · 03-13-2024, 07:11 AM

yes, n=20 is not really useful, try other numbers too (like 1234567890).

sundialsvcs · 03-13-2024, 08:43 AM

Since you defined it as a "constant expression," the compiler doesn't have to generate a function at all.

pan64 · 03-13-2024, 11:06 AM

Quote:

Originally Posted by sundialsvcs

Since you defined it as a "constant expression," the compiler doesn't have to generate a function at all.

As it was mentioned (and I guess) compiler can optimize it, so I'm not really sure this is the real reason. Furthermore I don't think an additional function call can cause this difference.
Would be nice to see the object code to check it.