From 3323e2018d6d736a25b15902bc85f559ea98adb5 Mon Sep 17 00:00:00 2001 From: Damjan Marion Date: Thu, 2 Dec 2021 11:28:57 +0100 Subject: vppinfra: add perf testing to test_vector_func Type: improvement Change-Id: I7aacd58d113c13036c15655817400032dd8d1932 Signed-off-by: Damjan Marion --- src/vppinfra/vector/test/test.c | 249 +++++++++++++++++++++++++++++++++++++++- 1 file changed, 243 insertions(+), 6 deletions(-) (limited to 'src/vppinfra/vector/test/test.c') diff --git a/src/vppinfra/vector/test/test.c b/src/vppinfra/vector/test/test.c index 1a8b9d6ea10..d098766481c 100644 --- a/src/vppinfra/vector/test/test.c +++ b/src/vppinfra/vector/test/test.c @@ -4,8 +4,9 @@ #include #include +#include -test_registration_t *test_registrations[CLIB_MARCH_TYPE_N_VARIANTS] = {}; +test_main_t test_main; int test_march_supported (clib_march_variant_type_t type) @@ -18,14 +19,12 @@ test_march_supported (clib_march_variant_type_t type) return 0; } -int -main (int argc, char *argv[]) +clib_error_t * +test_funct (test_main_t *tm) { - clib_mem_init (0, 64ULL << 20); - for (int i = 0; i < CLIB_MARCH_TYPE_N_VARIANTS; i++) { - test_registration_t *r = test_registrations[i]; + test_registration_t *r = tm->registrations[i]; if (r == 0 || test_march_supported (i) < 0) continue; @@ -51,3 +50,241 @@ main (int argc, char *argv[]) fformat (stdout, "\n"); return 0; } + +#define TEST_PERF_MAX_EVENTS 7 +typedef struct +{ + u64 config[TEST_PERF_MAX_EVENTS]; + u8 n_events; + format_function_t *format_fn; +} test_perf_event_bundle_t; + +static u8 * +format_test_perf_bundle_default (u8 *s, va_list *args) +{ + test_perf_event_bundle_t __clib_unused *b = + va_arg (*args, test_perf_event_bundle_t *); + test_perf_t *tp = va_arg (*args, test_perf_t *); + u64 *data = va_arg (*args, u64 *); + + if (data) + s = format (s, "%5.2f", (f64) data[1] / data[0]); + else + s = format (s, "%5s", "IPC"); + + if (data) + s = format (s, "%8.2f", (f64) data[0] / tp->n_ops); + else + s = format (s, "%8s", "Clks/Op"); + + if (data) + s = format (s, "%8.2f", (f64) data[1] / tp->n_ops); + else + s = format (s, "%8s", "Inst/Op"); + + if (data) + s = format (s, "%9.2f", (f64) data[2] / tp->n_ops); + else + s = format (s, "%9s", "Brnch/Op"); + + if (data) + s = format (s, "%10.2f", (f64) data[3] / tp->n_ops); + else + s = format (s, "%10s", "BrMiss/Op"); + return s; +} + +test_perf_event_bundle_t perf_bundles[] = { { + .config[0] = PERF_COUNT_HW_CPU_CYCLES, + .config[1] = PERF_COUNT_HW_INSTRUCTIONS, + .config[2] = PERF_COUNT_HW_BRANCH_INSTRUCTIONS, + .config[3] = PERF_COUNT_HW_BRANCH_MISSES, + .n_events = 4, + .format_fn = format_test_perf_bundle_default, +} }; + +#ifdef __linux__ +clib_error_t * +test_perf (test_main_t *tm) +{ + clib_error_t *err = 0; + test_perf_event_bundle_t *b = perf_bundles; + int group_fd = -1, fds[TEST_PERF_MAX_EVENTS]; + u64 count[TEST_PERF_MAX_EVENTS + 3] = {}; + struct perf_event_attr pe = { + .size = sizeof (struct perf_event_attr), + .type = PERF_TYPE_HARDWARE, + .disabled = 1, + .exclude_kernel = 1, + .exclude_hv = 1, + .pinned = 1, + .exclusive = 1, + .read_format = (PERF_FORMAT_GROUP | PERF_FORMAT_TOTAL_TIME_ENABLED | + PERF_FORMAT_TOTAL_TIME_RUNNING), + }; + + for (int i = 0; i < TEST_PERF_MAX_EVENTS; i++) + fds[i] = -1; + + for (int i = 0; i < b->n_events; i++) + { + pe.config = b->config[i]; + int fd = syscall (__NR_perf_event_open, &pe, /* pid */ 0, /* cpu */ -1, + /* group_fd */ group_fd, /* flags */ 0); + if (fd < 0) + { + err = clib_error_return_unix (0, "perf_event_open"); + goto done; + } + + if (group_fd == -1) + { + group_fd = fd; + pe.pinned = 0; + pe.exclusive = 0; + } + fds[i] = fd; + } + + for (int i = 0; i < CLIB_MARCH_TYPE_N_VARIANTS; i++) + { + test_registration_t *r = tm->registrations[i]; + + if (r == 0 || test_march_supported (i) < 0) + continue; + + fformat (stdout, "\nMultiarch Variant: %U\n", format_march_variant, i); + fformat (stdout, + "-------------------------------------------------------\n"); + while (r) + { + if (r->perf_tests) + { + test_perf_t *pt = r->perf_tests; + fformat (stdout, "%-22s%-12s%U\n", r->name, "OpType", + b->format_fn, b, pt, 0UL); + do + { + u32 read_size = (b->n_events + 3) * sizeof (u64); + for (int i = 0; i < tm->repeat; i++) + { + test_perf_event_reset (group_fd); + pt->fn (group_fd, pt); + if ((read (group_fd, &count, read_size) != read_size)) + { + err = clib_error_return_unix (0, "read"); + goto done; + } + if (count[1] != count[2]) + clib_warning ( + "perf counters were not running all the time." +#ifdef __x86_64__ + "\nConsider turning NMI watchdog off ('sysctl -w " + "kernel.nmi_watchdog=0')." +#endif + ); + fformat (stdout, " %-20s%-12s%U\n", pt->name, + pt->op_name ? pt->op_name : "", b->format_fn, b, + pt, count + 3); + } + } + while ((++pt)->fn); + } + r = r->next; + } + } + +done: + for (int i = 0; i < TEST_PERF_MAX_EVENTS; i++) + if (fds[i] != -1) + close (fds[i]); + return err; +} +#endif + +int +main (int argc, char *argv[]) +{ + test_main_t *tm = &test_main; + unformat_input_t _i = {}, *i = &_i; + clib_mem_init (0, 64ULL << 20); + clib_error_t *err; + int perf = 0; + + /* defaults */ + tm->repeat = 3; + + unformat_init_command_line (i, argv); + + while (unformat_check_input (i) != UNFORMAT_END_OF_INPUT) + { + if (unformat (i, "perf")) + perf = 1; + else if (unformat (i, "repeat %d", &tm->repeat)) + ; + else + { + clib_warning ("unknown input '%U'", format_unformat_error, i); + exit (1); + } + } + + if (perf) + err = test_perf (tm); + else + err = test_funct (tm); + + if (err) + { + clib_error_report (err); + fformat (stderr, "\n"); + return 1; + } + return 0; +} + +void * +test_mem_alloc (uword size) +{ + void *rv; + size = round_pow2 (size, CLIB_CACHE_LINE_BYTES); + rv = clib_mem_alloc_aligned (size, CLIB_CACHE_LINE_BYTES); + clib_memset_u8 (rv, 0, size); + return rv; +} + +void * +test_mem_alloc_and_fill_inc_u8 (uword size, u8 start, u8 mask) +{ + u8 *rv; + mask = mask ? mask : 0xff; + size = round_pow2 (size, CLIB_CACHE_LINE_BYTES); + rv = clib_mem_alloc_aligned (size, CLIB_CACHE_LINE_BYTES); + for (uword i = 0; i < size; i++) + rv[i] = ((u8) i + start) & mask; + return rv; +} + +void * +test_mem_alloc_and_splat (uword elt_size, uword n_elts, void *elt) +{ + u8 *rv, *e; + uword data_size = elt_size * n_elts; + uword alloc_size = round_pow2 (data_size, CLIB_CACHE_LINE_BYTES); + e = rv = clib_mem_alloc_aligned (alloc_size, CLIB_CACHE_LINE_BYTES); + while (e - rv < data_size) + { + clib_memcpy_fast (e, elt, elt_size); + e += elt_size; + } + + if (data_size < alloc_size) + clib_memset_u8 (e, 0, alloc_size - data_size); + return rv; +} + +void +test_mem_free (void *p) +{ + clib_mem_free (p); +} -- cgit 1.2.3-korg