Test for L1 cache misses in raspi 4.

test_cpucache.cpp

#include <unistd.h>
#include <iostream>
#include <chrono>
#include <vector>

typedef struct x {
	int a;
	float b;
} sx;

typedef struct y {
	int a;
	float b;
	double c[8];
// 8 + 8 * 8 = 72 bytes
} sy;

template <typename T>
void test_stack(long vlen)
{
	std::vector<T> v;
	for (int i = 0; i < vlen; i++) {
		v.push_back({ 0,0.0 });
	}

	int itmp;
	float ftmp;

	std::chrono::steady_clock::time_point ts = std::chrono::steady_clock::now();
	for (int i = 0; i < vlen; i++) {
		//std::chrono::steady_clock::time_point t1 = std::chrono::steady_clock::now();
		itmp = v[i].a;
		ftmp = v[i].b;
		//std::chrono::steady_clock::time_point t2 = std::chrono::steady_clock::now();
		//std::cout << i << ": " << std::chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count() << std::endl;
	}
	std::chrono::steady_clock::time_point tf = std::chrono::steady_clock::now();
	std::cout << "stack total : " << std::chrono::duration_cast<std::chrono::microseconds>(tf - ts).count() << std::endl;
}

template <typename T>
void test_heap(long vlen)
{
	std::vector<T*> v;
	for (int i = 0; i < vlen; i++) {
		T *xptr = (T*)malloc(sizeof(T));
		v.push_back(xptr);
	}

	int itmp;
	float ftmp;

	std::chrono::steady_clock::time_point ts = std::chrono::steady_clock::now();
	for (int i = 0; i < vlen; i++) {
		//std::chrono::steady_clock::time_point t1 = std::chrono::steady_clock::now();
		itmp = v[i]->a;
		ftmp = v[i]->b;
		//std::chrono::steady_clock::time_point t2 = std::chrono::steady_clock::now();
		//std::cout << i << ": " << std::chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count() << std::endl;
	}
	std::chrono::steady_clock::time_point tf = std::chrono::steady_clock::now();
	std::cout << "heap total : " << std::chrono::duration_cast<std::chrono::microseconds>(tf - ts).count() << std::endl;
}

template <typename T>
void test()
{
	long l1cachesz = 32 * 1024;
	long l2cachesz = 1024 * 1024;
	long cachelinesz = sysconf(_SC_LEVEL1_DCACHE_LINESIZE);

	std::cout << "L1 cache size: " << l1cachesz << std::endl;
	std::cout << "L1 cache line size: " << cachelinesz << std::endl;
	std::cout << "L2 cache size: " << l2cachesz << std::endl;

	long vlen = (l2cachesz / sizeof(T)) * 2;
	test_stack<T>(vlen);
	test_heap<T>(vlen);

}

int main() {
	test<sx>();
}

test output:

stack total : 2980
heap total : 4724