#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include "papi.h"

#define add_event(es, e) { \
    int ret; \
    if ((ret = PAPI_add_event(es, e)) != PAPI_OK) \
	error("PAPI_add_event", ret); \
}
#define remove_event(es, e) { \
    int ret; \
    if ((ret = PAPI_remove_event(es, e)) != PAPI_OK) \
	error("PAPI_remove_event", ret); \
}
#define error(s, ret) { \
    char buffer[1024]; \
    PAPI_perror(ret, buffer, sizeof(buffer)); \
    fprintf(stderr, "%s: %s\n", s, buffer); \
    exit(1); \
}

#if 0
#  define matmul_ijk_	matmul_ijk__
#  define matmul_ikj_	matmul_ikj__
#  define matmul_jik_	matmul_jik__
#  define matmul_jki_	matmul_jki__
#  define matmul_kij_	matmul_kij__
#  define matmul_kji_	matmul_kji__
#  define matmul_blas_	matmul_blas__
#  define matmul0_papi_	matmul0_papi__
#endif

void matmul_ijk_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);
void matmul_ikj_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);
void matmul_jik_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);
void matmul_jki_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);
void matmul_kij_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);
void matmul_kji_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);
void matmul_blas_(int *n, double *a, double *b, double *c, 
		int *peak, int *rep, int *prt);

typedef void (*F_func_t)(int *n, double *a, double *b, double *c, 
			 int *peak, int *nrep, int *prt);
static F_func_t funcs[] = { matmul_ijk_, matmul_jki_, matmul_kij_,
			    matmul_jik_, matmul_ikj_, matmul_kji_, 
			    matmul_blas_ };

static int events;
static long long counters[128];

static void
cleanup(void)
{
    if (PAPI_is_initialized()) PAPI_shutdown();
}

void
matmul0_papi_(int *n, double *a, double *b, double *c, int *peak, int *rep)
{
    int i, zero = 0, one = 1;

    if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) {
	fprintf(stderr, "PAPI_library_init failed.\n");
	exit(1);
    }

    atexit(cleanup);

    PAPI_create_eventset(&events);

    for (i = 0; i < sizeof(funcs)/sizeof(*funcs); i++) {
	long long misses1, misses2, total1 = 0, total2 = 0;

	add_event(events, PAPI_L1_TCM);
	add_event(events, PAPI_L2_TCM);
	PAPI_start(events);
	PAPI_reset(events);
	funcs[i](n, a, b, c, peak, rep, &one);
	PAPI_stop(events, counters);
	misses1 = counters[0];
	misses2 = counters[1];
	PAPI_cleanup_eventset(events);

	if (PAPI_add_event(events, PAPI_L1_TCA) == PAPI_OK &&
	    PAPI_add_event(events, PAPI_L2_TCA) == PAPI_OK) {
	    PAPI_start(events);
	    PAPI_reset(events);
	    funcs[i](n, a, b, c, peak, rep, &zero);
	    PAPI_stop(events, counters);
	    total1 = counters[0];
	    total2 = counters[1];
	}
	PAPI_cleanup_eventset(events);

	if (total1 > 0 && total2 > 0) {
	    printf("\t%10lld L1-misses (%0.2lf), %10lld L2-misses (%0.2lf).\n",
			misses1, misses1/(double)total1, 
			misses2, misses2/(double)total2);
	} else {
	    printf("\t%10lld L1-misses, %10lld L2-misses.\n", misses1, misses2);
	}
    }

    return;
}