Test L1 cache in raspi 4.

test_cpucache.c

#include <unistd.h>
#include <time.h>
#include <stdlib.h>
#include <stdio.h>

#define L1_CACHE_SZ (32 * 1024)
#define L2_CACHE_SZ (1024 * 1024)

#define TESTOBJ sx

#define VLEN ((L2_CACHE_SZ / sizeof(TESTOBJ)) * 2)

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

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

void test_stack()
{
    TESTOBJ *arr = malloc(sizeof(TESTOBJ) * VLEN);

	for (int i = 0; i < VLEN; i++) {
        TESTOBJ o = { .a = 0, .b = 0.0 };
        arr[i] = o;
	}

	int itmp;
	float ftmp;
    struct timespec start, end;

    clock_gettime(CLOCK_MONOTONIC_RAW, &start);
	for (int i = 0; i < VLEN; i++) {
		itmp = arr[i].a;
		ftmp = arr[i].b;
	}
    clock_gettime(CLOCK_MONOTONIC_RAW, &end);
    unsigned long elapsed_ns = (end.tv_sec - start.tv_sec) * 1000 * 1000 * 1000 + (end.tv_nsec - start.tv_nsec);
    printf("stack total: %lu\n", elapsed_ns);
}

void test_heap()
{
    TESTOBJ **arr = malloc(sizeof(TESTOBJ*) * VLEN);

	for (int i = 0; i < VLEN; i++) {
        TESTOBJ *xptr = malloc(sizeof(TESTOBJ));
        arr[i] = xptr;
	}

	int itmp;
	float ftmp;
    struct timespec start, end;

    clock_gettime(CLOCK_MONOTONIC_RAW, &start);
	for (int i = 0; i < VLEN; i++) {
		itmp = arr[i]->a;
		ftmp = arr[i]->b;
	}
    clock_gettime(CLOCK_MONOTONIC_RAW, &end);
    unsigned long elapsed_ns = (end.tv_sec - start.tv_sec) * 1000 * 1000 * 1000 + (end.tv_nsec - start.tv_nsec);
    printf("heap total: %lu\n", elapsed_ns);
}

int main() {
	test_stack();
	test_heap();
}

test output (rpi 4):

stack total: 2498611
heap total: 3569907

test output (macOS x86_64):

stack total: 655048
heap total: 2216167