Skip to content

Commit

Permalink
test: performance: use common test runner function
Browse files Browse the repository at this point in the history 
Use common test runner function bench_run() for all odp_bench_*
applications. Reduces the amount of duplicate code.

A separate warm-up round has been removed from odp_bench_packet, as
bench_run() already performs one.

Signed-off-by: Matias Elo <matias.elo@nokia.com>
Reviewed-by: Petri Savolainen <petri.savolainen@nokia.com>
  • Loading branch information
@MatiasElo
MatiasElo committed Sep 20, 2023
1 parent 7674c0c commit 6d5c269
Show file tree
Hide file tree
Showing 6 changed files with 232 additions and 421 deletions.
104 changes: 104 additions & 0 deletions test/performance/bench_common.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,18 @@

#include "bench_common.h"

#include <inttypes.h>
#include <stdint.h>
#include <string.h>

void bench_suite_init(bench_suite_t *suite)
{
memset(suite, 0, sizeof(bench_suite_t));

suite->measure_time = true;

odp_atomic_init_u32(&suite->exit_worker, 0);
}

void bench_run_indef(bench_info_t *info, odp_atomic_u32_t *exit_thread)
{
Expand All @@ -32,3 +43,96 @@ void bench_run_indef(bench_info_t *info, odp_atomic_u32_t *exit_thread)
ODPH_ABORT("Benchmark %s failed\n", desc);
}
}

int bench_run(void *arg)
{
uint64_t c1, c2;
odp_time_t t1, t2;
bench_suite_t *suite = arg;
const uint64_t repeat_count = suite->repeat_count;
const odp_bool_t meas_time = suite->measure_time;

printf("\nAverage %s per function call\n", meas_time ? "time (nsec)" : "CPU cycles");
printf("-------------------------------------------------\n");

/* Run each test twice. Results from the first warm-up round are ignored. */
for (int i = 0; i < 2; i++) {
uint64_t total = 0;
uint64_t round = 1;

for (int j = 0; j < suite->num_bench; round++) {
int ret;
const char *desc;
const bench_info_t *bench = &suite->bench[j];
uint64_t max_rounds = suite->rounds;

if (bench->max_rounds && bench->max_rounds < max_rounds)
max_rounds = bench->max_rounds;

/* Run selected test indefinitely */
if (suite->indef_idx) {
if ((j + 1) != suite->indef_idx) {
j++;
continue;
}
bench_run_indef(&suite->bench[j], &suite->exit_worker);
return 0;
}

desc = bench->desc != NULL ? bench->desc : bench->name;

if (bench->init != NULL)
bench->init();

if (meas_time)
t1 = odp_time_local_strict();
else
c1 = odp_cpu_cycles();

ret = bench->run();

if (meas_time)
t2 = odp_time_local_strict();
else
c2 = odp_cpu_cycles();

if (bench->term != NULL)
bench->term();

if (!ret) {
ODPH_ERR("Benchmark odp_%s failed\n", desc);
suite->retval = -1;
return -1;
}

if (meas_time)
total += odp_time_diff_ns(t2, t1);
else
total += odp_cpu_cycles_diff(c2, c1);

if (round >= max_rounds) {
double result;

/* Each benchmark runs internally 'repeat_count' times. */
result = ((double)total) / (max_rounds * repeat_count);

/* No print or results from warm-up round */
if (i > 0) {
printf("[%02d] odp_%-26s: %12.2f\n", j + 1, desc, result);

if (suite->result)
suite->result[j] = result;
}
j++;
total = 0;
round = 1;
}
}
}
printf("\n");
/* Print dummy result to prevent compiler to optimize it away*/
if (suite->dummy)
printf("(dummy result: 0x%" PRIx64 ")\n\n", suite->dummy);

return 0;
}
44 changes: 44 additions & 0 deletions test/performance/bench_common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -63,11 +63,55 @@ typedef struct {

} bench_info_t;

/* Benchmark suite data */
typedef struct {
/* Array of benchmark functions */
bench_info_t *bench;

/* Number of benchmark functions */
int num_bench;

/* Optional benchmark index to run indefinitely (1...num_bench) */
int indef_idx;

/* Suite exit value output */
int retval;

/* Measure time vs. CPU cycles */
odp_bool_t measure_time;

/* Break worker loop if set to 1 */
odp_atomic_u32_t exit_worker;

/* Number of API function calls per test case */
uint64_t repeat_count;

/* Number of rounds per test case */
uint64_t rounds;

/* Dummy test result output */
uint64_t dummy;

/* Optional test result output array */
double *result;

} bench_suite_t;

/**
* Initialize benchmark suite parameters
*/
void bench_suite_init(bench_suite_t *suite);

/**
* Run selected test indefinitely
*/
void bench_run_indef(bench_info_t *info, odp_atomic_u32_t *exit_thread);

/**
* Run tests suite and print results
*/
int bench_run(void *arg);

#ifdef __cplusplus
}
#endif
Expand Down
113 changes: 14 additions & 99 deletions test/performance/odp_bench_buffer.c
View file
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/* Copyright (c) 2017-2018, Linaro Limited
* Copyright (c) 2022, Nokia
* Copyright (c) 2022-2023, Nokia
* All rights reserved.
*
* SPDX-License-Identifier: BSD-3-Clause
Expand Down Expand Up @@ -61,20 +61,16 @@ typedef struct {
typedef struct {
/** Application (parsed) arguments */
appl_args_t appl;
/** Common benchmark suite data */
bench_suite_t suite;
/** Buffer pool */
odp_pool_t pool;
/** Benchmark functions */
bench_info_t *bench;
/** Number of benchmark functions */
int num_bench;
/** Buffer size */
uint32_t buf_size;
/** Buffer user area size */
uint32_t uarea_size;
/** Max flow id */
uint32_t max_flow_id;
/** Break worker loop if set to 1 */
odp_atomic_u32_t exit_thread;
/** Array for storing test buffers */
odp_buffer_t buf_tbl[TEST_REPEAT_COUNT * TEST_MAX_BURST];
/** Array for storing test event */
Expand All @@ -87,8 +83,6 @@ typedef struct {
odp_event_type_t event_type_tbl[TEST_REPEAT_COUNT];
/** Array for storing test event subtypes */
odp_event_subtype_t event_subtype_tbl[TEST_REPEAT_COUNT];
/** Benchmark run failed */
uint8_t bench_failed;
/** CPU mask as string */
char cpumask_str[ODP_CPUMASK_STR_SIZE];
} args_t;
Expand All @@ -100,89 +94,7 @@ static void sig_handler(int signo ODP_UNUSED)
{
if (gbl_args == NULL)
return;
odp_atomic_store_u32(&gbl_args->exit_thread, 1);
}

static int run_benchmarks(void *arg)
{
int i, j, k;
args_t *args = arg;

printf("\nAverage CPU cycles per function call\n"
"---------------------------------------------\n");

/* Run each test twice. Results from the first warm-up round are ignored. */
for (i = 0; i < 2; i++) {
uint64_t tot_cycles = 0;

for (j = 0, k = 1; j < gbl_args->num_bench; k++) {
int ret;
uint64_t c1, c2;
const char *desc;

/* Run selected test indefinitely */
if (args->appl.bench_idx &&
(j + 1) != args->appl.bench_idx) {
j++;
continue;
} else if (args->appl.bench_idx &&
(j + 1) == args->appl.bench_idx) {
bench_run_indef(&args->bench[j], &gbl_args->exit_thread);
return 0;
}

desc = args->bench[j].desc != NULL ?
args->bench[j].desc :
args->bench[j].name;

/* Skip unsupported tests */
if (args->bench[j].cond != NULL && !args->bench[j].cond()) {
j++;
k = 1;
if (i > 0)
printf("[%02d] odp_%-26s: n/a\n", j, desc);
continue;
}

if (args->bench[j].init != NULL)
args->bench[j].init();

c1 = odp_cpu_cycles();
ret = args->bench[j].run();
c2 = odp_cpu_cycles();

if (args->bench[j].term != NULL)
args->bench[j].term();

if (!ret) {
ODPH_ERR("Benchmark odp_%s failed\n", desc);
args->bench_failed = 1;
return -1;
}

tot_cycles += odp_cpu_cycles_diff(c2, c1);

if (k >= args->appl.test_cycles) {
double cycles;

/** Each benchmark runs internally TEST_REPEAT_COUNT times. */
cycles = ((double)tot_cycles) /
(args->appl.test_cycles *
TEST_REPEAT_COUNT);

/* No print from warm-up round */
if (i > 0)
printf("[%02d] odp_%-26s: %8.1f\n", j + 1, desc, cycles);

j++;
k = 1;
tot_cycles = 0;
}
}
}
printf("\n");

return 0;
odp_atomic_store_u32(&gbl_args->suite.exit_worker, 1);
}

static void allocate_test_buffers(odp_buffer_t buf[], int num)
Expand Down Expand Up @@ -732,14 +644,17 @@ int main(int argc, char *argv[])
}

memset(gbl_args, 0, sizeof(args_t));
odp_atomic_init_u32(&gbl_args->exit_thread, 0);

gbl_args->bench = test_suite;
gbl_args->num_bench = sizeof(test_suite) / sizeof(test_suite[0]);

/* Parse and store the application arguments */
parse_args(argc, argv, &gbl_args->appl);

bench_suite_init(&gbl_args->suite);
gbl_args->suite.bench = test_suite;
gbl_args->suite.num_bench = sizeof(test_suite) / sizeof(test_suite[0]);
gbl_args->suite.indef_idx = gbl_args->appl.bench_idx;
gbl_args->suite.rounds = gbl_args->appl.test_cycles;
gbl_args->suite.repeat_count = TEST_REPEAT_COUNT;

/* Get default worker cpumask */
if (odp_cpumask_default_worker(&default_mask, 1) != 1) {
ODPH_ERR("Error: unable to allocate worker thread\n");
Expand Down Expand Up @@ -825,15 +740,15 @@ int main(int argc, char *argv[])
thr_common.share_param = 1;

odph_thread_param_init(&thr_param);
thr_param.start = run_benchmarks;
thr_param.arg = gbl_args;
thr_param.start = bench_run;
thr_param.arg = &gbl_args->suite;
thr_param.thr_type = ODP_THREAD_WORKER;

odph_thread_create(&worker_thread, &thr_common, &thr_param, 1);

odph_thread_join(&worker_thread, 1);

ret = gbl_args->bench_failed;
ret = gbl_args->suite.retval;

if (odp_pool_destroy(gbl_args->pool)) {
ODPH_ERR("Error: pool destroy\n");
Expand Down
Loading

0 comments on commit 6d5c269

Please sign in to comment.