int
_papi_libpfm_ntv_code_to_name( unsigned int EventCode, char *ntv_name, int len )
{
int ret;
unsigned int event, umask;
pfmlib_event_t gete;
memset( &gete, 0, sizeof ( gete ) );
if ( _pfm_decode_native_event( EventCode, &event, &umask ) != PAPI_OK )
return ( PAPI_ENOEVNT );
gete.event = event;
gete.num_masks = ( unsigned int ) prepare_umask( umask, gete.unit_masks );
if ( gete.num_masks == 0 )
ret = pfm_get_event_name( gete.event, ntv_name, ( size_t ) len );
else
ret = pfm_get_full_event_name( &gete, ntv_name, ( size_t ) len );
if ( ret != PFMLIB_SUCCESS ) {
char tmp[PAPI_2MAX_STR_LEN];
pfm_get_event_name( gete.event, tmp, sizeof ( tmp ) );
/* Skip error message if event is not supported by host cpu;
* we don't need to give this info away for papi_native_avail util */
if ( ret != PFMLIB_ERR_BADHOST )
PAPIERROR
( "pfm_get_full_event_name(%p(event %d,%s,%d masks),%p,%d): %d -- %s",
&gete, gete.event, tmp, gete.num_masks, ntv_name, len, ret,
pfm_strerror( ret ) );
if ( ret == PFMLIB_ERR_FULL ) {
return PAPI_EBUF;
}
return PAPI_EMISC;
}
return PAPI_OK;
}
static int
show_info(regex_t *preg)
{
unsigned int i, count = 0, match = 0;
int ret;
pfm_get_num_events(&count);
for(i=0; i < count; i++) {
ret = pfm_get_event_name(i, name, max_len+1);
/* skip unsupported events */
if (ret != PFMLIB_SUCCESS)
continue;
if (regexec(preg, name, 0, NULL, 0) == 0) {
show_event_info(name, i);
match++;
}
}
return match;
}
static int
show_info_sorted(regex_t *preg)
{
unsigned int i, n, count = 0, match = 0;
int code, ret;
code_info_t *codes = NULL;
pfm_get_num_events(&count);
codes = malloc(count * sizeof(*codes));
if (!codes)
fatal_error("cannot allocate memory\n");
for(i=0, n = 0; i < count; i++, n++) {
ret = pfm_get_event_code(i, &code);
/* skip unsupported events */
if (ret != PFMLIB_SUCCESS)
continue;
codes[n].idx = i;
codes[n].code = code;
}
qsort(codes, n, sizeof(*codes), compare_codes);
for(i=0; i < n; i++) {
ret = pfm_get_event_name(codes[i].idx, name, max_len+1);
/* skip unsupported events */
if (ret != PFMLIB_SUCCESS)
continue;
if (regexec(preg, name, 0, NULL, 0) == 0) {
show_event_info(name, codes[i].idx);
match++;
}
}
free(codes);
return match;
}
static int
pfmon_ita_print_header(FILE *fp)
{
pfmon_event_set_t *set;
pfmlib_ita_input_param_t *mod_in;
unsigned int i, k, l;
int isn;
char *name;
size_t len;
static const char *insn_str[]={
"ia32/ia64",
"ia32",
"ia64"
};
len = 1 + options.max_event_name_len;
name = malloc(len);
if (!name)
fatal_error("cannot allocate string buffer");
for(k=0, set = options.sets; set; k++, set = set->next) {
mod_in = (pfmlib_ita_input_param_t *)set->mod_inp;
fprintf(fp, "#\n#\n# instruction sets for set%u:\n", k);
for(i=0; i < set->event_count; i++) {
pfm_get_event_name(set->inp.pfp_events[i].event, name, len);
isn =mod_in->pfp_ita_counters[i].ism;
fprintf(fp, "#\tPMD%d: %-*s = %s\n",
set->outp.pfp_pmcs[i].reg_num,
(int)options.max_event_name_len,
name,
insn_str[isn]);
l--;
}
fprintf(fp, "#\n");
}
free(name);
return 0;
}
/*
* This function checks the configuration to verify
* that the user does not try to combine features with
* events that are incompatible.The library does this also
* but it's hard to then detail the cause of the error.
*/
static void
check_ita_event_combinations(pfmon_event_set_t *set)
{
unsigned int i, use_opcm, ev;
pfmlib_ita_input_param_t *param = set->mod_inp;
pfmon_ita_args_t *ita_args;
char *name;
ita_args = set->mod_args;
name = options.ev_name1;
use_opcm = param->pfp_ita_pmc8.opcm_used || param->pfp_ita_pmc9.opcm_used;
for (i=0; i < set->event_count; i++) {
ev = set->inp.pfp_events[i].event;
pfm_get_event_name(ev, name, options.max_event_name_len+1);
if (use_opcm && pfm_ita_support_opcm(ev) == 0)
fatal_error("event %s does not support opcode matching\n", name);
if (param->pfp_ita_irange.rr_used && pfm_ita_support_iarr(ev) == 0)
fatal_error("event %s does not support instruction address range restrictions\n", name);
if (param->pfp_ita_drange.rr_used && pfm_ita_support_darr(ev) == 0)
fatal_error("event %s does not support data address range restrictions\n", name);
if (ita_args->opt_ia32 && ita_args->opt_ia64 == 0 && pfm_ita_is_btb(ev))
fatal_error("cannot use BTB event (%s) when only monitoring IA-32 execution\n", name);
}
/*
* we do not call check_counter_conflict() because Itanium does not have events
* which can only be measured on one counter, therefore this routine would not
* catch anything at all.
*/
}
int
main(int argc, char **argv)
{
char **p;
int i, ret;
pid_t pid = getpid();
pfmlib_param_t evt;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
printf("Can't initialize library\n");
exit(1);
}
/*
* check that the user did not specify too many events
*/
if (argc-1 > pfm_get_num_counters()) {
printf("Too many events specified\n");
exit(1);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
/*
* prepare parameters to library. we don't use any Itanium
* specific features here. so the pfp_model is NULL.
*/
memset(&evt,0, sizeof(evt));
/*
* be nice to user!
*/
p = argc > 1 ? argv+1 : event_list;
for (i=0; *p ; i++, p++) {
if (pfm_find_event(*p, &evt.pfp_events[i].event) != PFMLIB_SUCCESS) {
fatal_error("Cannot find %s event\n", *p);
}
}
/*
* set the default privilege mode for all counters:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = i;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we have decided not to get notified
* on counter overflows and the monitoring is not to be inherited
* in derived tasks.
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, PFM_CREATE_CONTEXT, ctx, 1) == -1 ) {
if (errno == ENOSYS) {
fatal_error("Your kernel does not have performance monitoring support!\n");
}
fatal_error("Can't create PFM context %s\n", strerror(errno));
}
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* Now prepare the argument to initialize the PMDs.
* the memset(pd) initialized the entire array to zero already, so
* we just have to fill in the register numbers from the pc[] array.
*/
for (i=0; i < evt.pfp_event_count; i++) {
pd[i].reg_num = evt.pfp_pc[i].reg_num;
}
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, evt.pfp_event_count) == -1) {
{int i; for(i=0; i < evt.pfp_event_count; i++) printf("pmd%d: 0x%x\n", i, pd[i].reg_flags);}
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now
*/
pfm_start();
noploop(10000000);
pfm_stop();
/*
* now read the results
*/
if (perfmonctl(pid, PFM_READ_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error( "perfmonctl error READ_PMDS errno %d\n",errno);
return -1;
}
/*
* print the results
*
* It is important to realize, that the first event we specified may not
* be in PMD4. Not all events can be measured by any monitor. That's why
* we need to use the pc[] array to figure out where event i was allocated.
*
*/
for (i=0; i < evt.pfp_event_count; i++) {
char *name;
pfm_get_event_name(evt.pfp_events[i].event, &name);
printf("PMD%u %20lu %s\n",
pd[i].reg_num,
pd[i].reg_value,
name);
}
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error( "child: perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
int
main(void)
{
int ret;
int type = 0;
char *name;
pid_t pid = getpid();
pfmlib_param_t evt;
pfmlib_ita2_param_t ita2_param;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char *model;
pfm_get_pmu_name(&model);
fatal_error("this program does not work with the %s PMU\n", model);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for verbose */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
memset(&evt,0, sizeof(evt));
memset(&ita2_param,0, sizeof(ita2_param));
/*
* because we use a model specific feature, we must initialize the
* model specific pfmlib parameter structure and link it to the
* common structure.
* The magic number is a simple mechanism used by the library to check
* that the model specific data structure is decent. You must set it manually
* otherwise the model specific feature won't work.
*/
ita2_param.pfp_magic = PFMLIB_ITA2_PARAM_MAGIC;
evt.pfp_model = &ita2_param;
/*
* We indicate that we are using the PMC8 opcode matcher. This is required
* otherwise the library add PMC8 to the list of PMC to pogram during
* pfm_dispatch_events().
*/
ita2_param.pfp_ita2_pmc8.opcm_used = 1;
/*
* We want to match all the br.cloop in our test function.
* This branch is an IP-relative branch for which the major
* opcode (bits [40-37]=4) and the btype field is 5 (which represents
* bits[6-8]) so it is included in the match/mask fields of PMC8.
* It is necessarily in a B slot.
*
* We don't care which operands are used with br.cloop therefore
* the mask field of pmc8 is set such that only the 4 bits of the
* opcode and 3 bits of btype must match exactly. This is accomplished by
* clearing the top 4 bits and bits [6-8] of the mask field and setting the
* remaining bits. Similarly, the match field only has the opcode value and btype
* set according to the encoding of br.cloop, the
* remaining bits are zero. Bit 60 of PMC8 is set to indicate
* that we look only in B slots (this is the only possibility for
* this instruction anyway).
*
* So the binary representation of the value for PMC8 is as follows:
*
* 6666555555555544444444443333333333222222222211111111110000000000
* 3210987654321098765432109876543210987654321098765432109876543210
* ----------------------------------------------------------------
* 0001010000000000000000101000000000000011111111111111000111111000
*
* which yields a value of 0x1400028003fff1f8.
*
* Depending on the level of optimization to compile this code, it may
* be that the count reported could be zero, if the compiler uses a br.cond
* instead of br.cloop.
*
*
* The 0x1 sets the ig_ad field to make sure we ignore any range restriction.
* Also bit 2 must always be set
*/
ita2_param.pfp_ita2_pmc8.pmc_val = 0x1400028003fff1fa;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_event_byname("IA64_TAGGED_INST_RETIRED_IBRP0_PMC8", &evt.pfp_events[0].event) != PFMLIB_SUCCESS) {
fatal_error("cannot find event IA64_TAGGED_INST_RETIRED_IBRP0_PMC8\n");
}
/*
* set the privilege mode:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we have decided not to get notified
* on counter overflows and the monitoring is not to be inherited
* in derived tasks
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, PFM_CREATE_CONTEXT, ctx, 1) == -1 ) {
if (errno == ENOSYS) {
fatal_error("Your kernel does not have performance monitoring support!\n");
}
fatal_error("Can't create PFM context %s\n", strerror(errno));
}
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* Now prepare the argument to initialize the PMD.
*/
pd[0].reg_num = evt.pfp_pc[0].reg_num;
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now.
*/
pfm_start();
do_test(100UL);
pfm_stop();
/*
* now read the results
*/
if (perfmonctl(pid, PFM_READ_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error("perfmonctl error READ_PMDS errno %d\n",errno);
}
/*
* print the results
*/
pfm_get_event_name(evt.pfp_events[0].event, &name);
printf("PMD%u %20lu %s\n",
pd[0].reg_num,
pd[0].reg_value,
name);
if (pd[0].reg_value != 0)
printf("compiler used br.cloop\n");
else
printf("compiler did not use br.cloop\n");
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
