diff options
Diffstat (limited to 'test/validation/api/ml/ml.c')
| -rw-r--r-- | test/validation/api/ml/ml.c | 572 |
1 files changed, 572 insertions, 0 deletions
diff --git a/test/validation/api/ml/ml.c b/test/validation/api/ml/ml.c new file mode 100644 index 000000000..5f8be1b64 --- /dev/null +++ b/test/validation/api/ml/ml.c @@ -0,0 +1,572 @@ +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright (c) 2023 Nokia + */ + +#include <odp_api.h> +#include <odp/helper/odph_api.h> +#include "odp_cunit_common.h" + +#define UAREA 0xaa +#define NUM_COMPL 10u +#define COMPL_POOL_NAME "ML compl pool" + +typedef struct global_t { + int disabled; + uint32_t num_compl; + odp_ml_capability_t ml_capa; +} global_t; + +typedef struct { + uint32_t count; + uint8_t mark[NUM_COMPL]; +} uarea_init_t; + +static global_t global; + +static int ml_suite_init(void) +{ + memset(&global, 0, sizeof(global_t)); + + if (odp_ml_capability(&global.ml_capa)) { + ODPH_ERR("ML capability failed\n"); + return -1; + } + + if (global.ml_capa.max_models == 0) { + global.disabled = 1; + ODPH_DBG("ML test disabled\n"); + return 0; + } + + global.num_compl = ODPH_MIN(NUM_COMPL, global.ml_capa.pool.max_num); + + return 0; +} + +static int check_ml_support(void) +{ + if (global.disabled) + return ODP_TEST_INACTIVE; + + return ODP_TEST_ACTIVE; +} + +static void test_ml_capability(void) +{ + odp_ml_capability_t ml_capa; + + memset(&ml_capa, 0, sizeof(odp_ml_capability_t)); + CU_ASSERT(odp_ml_capability(&ml_capa) == 0); + + if (ml_capa.max_models == 0) + return; + + CU_ASSERT(ml_capa.max_model_size > 0); + CU_ASSERT(ml_capa.max_models_loaded > 0); + CU_ASSERT(ml_capa.max_inputs > 0); + CU_ASSERT(ml_capa.max_outputs > 0); + CU_ASSERT(ml_capa.max_segs_per_input > 0); + CU_ASSERT(ml_capa.max_segs_per_output > 0); + CU_ASSERT(ml_capa.min_input_align > 0); + CU_ASSERT(ml_capa.min_output_align > 0); + + if ((ml_capa.load.compl_mode_mask | ml_capa.run.compl_mode_mask) & + ODP_ML_COMPL_MODE_EVENT) { + odp_pool_capability_t pool_capa; + + CU_ASSERT_FATAL(odp_pool_capability(&pool_capa) == 0); + + CU_ASSERT(ml_capa.pool.max_pools > 0); + CU_ASSERT(ml_capa.pool.max_pools <= pool_capa.max_pools); + CU_ASSERT(ml_capa.pool.max_num > 0); + CU_ASSERT(ml_capa.pool.max_cache_size >= ml_capa.pool.min_cache_size); + } + + CU_ASSERT(ml_capa.load.compl_mode_mask); + CU_ASSERT(ml_capa.run.compl_mode_mask); + + if (ml_capa.load.compl_mode_mask & ODP_ML_COMPL_MODE_EVENT) + CU_ASSERT(ml_capa.load.compl_queue_plain || ml_capa.load.compl_queue_sched); + + if (ml_capa.run.compl_mode_mask & ODP_ML_COMPL_MODE_EVENT) + CU_ASSERT(ml_capa.run.compl_queue_plain || ml_capa.run.compl_queue_sched); +} + +static void test_ml_param(uint8_t fill) +{ + odp_ml_config_t config; + odp_ml_model_param_t model_param; + odp_ml_compl_pool_param_t pool_param; + odp_ml_compl_param_t compl_param; + odp_ml_run_param_t run_param; + + memset(&config, fill, sizeof(config)); + odp_ml_config_init(&config); + CU_ASSERT(config.max_models_created == 1); + CU_ASSERT(config.max_models_loaded == 1); + CU_ASSERT(config.load_mode_mask == 0); + CU_ASSERT(config.run_mode_mask == 0); + + memset(&model_param, fill, sizeof(model_param)); + odp_ml_model_param_init(&model_param); + CU_ASSERT(model_param.max_compl_id == 0); + CU_ASSERT(!model_param.extra_stat_enable); + CU_ASSERT(model_param.extra_param == NULL); + CU_ASSERT(model_param.extra_info.num_inputs == 0); + CU_ASSERT(model_param.extra_info.num_outputs == 0); + CU_ASSERT(model_param.extra_info.input_format == NULL); + CU_ASSERT(model_param.extra_info.output_format == NULL); + + memset(&pool_param, fill, sizeof(pool_param)); + odp_ml_compl_pool_param_init(&pool_param); + CU_ASSERT(pool_param.uarea_size == 0); + CU_ASSERT(pool_param.uarea_init.args == NULL); + CU_ASSERT(pool_param.uarea_init.init_fn == NULL); + CU_ASSERT(pool_param.cache_size <= global.ml_capa.pool.max_cache_size); + CU_ASSERT(pool_param.cache_size >= global.ml_capa.pool.min_cache_size); + + memset(&compl_param, fill, sizeof(compl_param)); + odp_ml_compl_param_init(&compl_param); + CU_ASSERT(compl_param.user_ptr == NULL); + + memset(&run_param, fill, sizeof(run_param)); + odp_ml_run_param_init(&run_param); + CU_ASSERT(run_param.batch_size == 0); + CU_ASSERT(run_param.result == NULL); +} + +static void test_ml_param_init(void) +{ + test_ml_param(0x00); + test_ml_param(0xff); +} + +static void test_ml_debug(void) +{ + odp_ml_print(); +} + +static void ml_compl_pool_create_max_pools(void) +{ + int ret; + uint32_t i, j; + odp_ml_compl_pool_param_t ml_pool_param; + uint32_t max_pools = global.ml_capa.pool.max_pools; + odp_pool_t compl_pools[max_pools]; + + odp_ml_compl_pool_param_init(&ml_pool_param); + ml_pool_param.num = global.num_compl; + for (i = 0; i < max_pools; i++) { + compl_pools[i] = odp_ml_compl_pool_create(NULL, &ml_pool_param); + + if (compl_pools[i] == ODP_POOL_INVALID) + break; + } + + CU_ASSERT(i == max_pools); + + /* Destroy the created valid pools */ + for (j = 0; j < i; j++) { + ret = odp_pool_destroy(compl_pools[j]); + CU_ASSERT(ret == 0); + + if (ret == -1) + ODPH_ERR("ML completion pool destroy failed: %u / %u\n", j, i); + } +} + +static void compl_pool_info(void) +{ + odp_pool_t pool; + odp_pool_t compl_pool; + odp_pool_info_t pool_info; + odp_ml_compl_pool_param_t pool_param; + + /* Create an ML job completion pool */ + odp_ml_compl_pool_param_init(&pool_param); + pool_param.num = global.num_compl; + + compl_pool = odp_ml_compl_pool_create(COMPL_POOL_NAME, &pool_param); + CU_ASSERT_FATAL(compl_pool != ODP_POOL_INVALID); + + /* Verify info about the created ML completion pool compl_pool */ + pool = odp_pool_lookup(COMPL_POOL_NAME); + CU_ASSERT(pool == compl_pool); + + memset(&pool_info, 0x66, sizeof(odp_pool_info_t)); + CU_ASSERT_FATAL(odp_pool_info(compl_pool, &pool_info) == 0); + + CU_ASSERT(!strcmp(pool_info.name, COMPL_POOL_NAME)); + CU_ASSERT(pool_info.pool_ext == 0); + CU_ASSERT(pool_info.type == ODP_POOL_ML_COMPL); + CU_ASSERT(pool_info.ml_pool_param.num == NUM_COMPL); + CU_ASSERT(pool_info.ml_pool_param.uarea_size == 0); + CU_ASSERT(pool_info.ml_pool_param.cache_size == pool_param.cache_size); + + CU_ASSERT_FATAL(odp_pool_destroy(compl_pool) == 0); +} + +static void compl_alloc_max(void) +{ + uint64_t u64; + odp_event_t event; + odp_pool_t compl_pool; + odp_ml_compl_pool_param_t pool_param; + const int num = ODPH_MIN(global.ml_capa.pool.max_num, 1000000u); + odp_ml_compl_t *compl = malloc(num * sizeof(odp_ml_compl_t)); + + CU_ASSERT_FATAL(compl != NULL); + + /* Create an ML job completion pool */ + odp_ml_compl_pool_param_init(&pool_param); + pool_param.num = num; + + compl_pool = odp_ml_compl_pool_create(COMPL_POOL_NAME, &pool_param); + CU_ASSERT_FATAL(compl_pool != ODP_POOL_INVALID); + + for (int i = 0; i < num; i++) { + compl[i] = odp_ml_compl_alloc(compl_pool); + CU_ASSERT_FATAL(compl[i] != ODP_ML_COMPL_INVALID); + + u64 = odp_ml_compl_to_u64(compl[i]); + CU_ASSERT(u64 != odp_ml_compl_to_u64(ODP_ML_COMPL_INVALID)); + + event = odp_ml_compl_to_event(compl[i]); + CU_ASSERT(odp_event_type(event) == ODP_EVENT_ML_COMPL); + } + + for (int i = 0; i < num; i++) + odp_ml_compl_free(compl[i]); + + free(compl); + + CU_ASSERT_FATAL(odp_pool_destroy(compl_pool) == 0); +} + +static void compl_pool_lookup(void) +{ + odp_pool_t pool, pool_a, pool_b; + odp_ml_compl_pool_param_t pool_param; + + /* Create an ML job completion pool */ + odp_ml_compl_pool_param_init(&pool_param); + pool_param.num = global.num_compl; + + pool_a = odp_ml_compl_pool_create(COMPL_POOL_NAME, &pool_param); + CU_ASSERT_FATAL(pool_a != ODP_POOL_INVALID); + + pool = odp_pool_lookup(COMPL_POOL_NAME); + CU_ASSERT(pool == pool_a); + + /* Second pool with the same name */ + pool_b = odp_ml_compl_pool_create(COMPL_POOL_NAME, &pool_param); + CU_ASSERT_FATAL(pool_b != ODP_POOL_INVALID); + + pool = odp_pool_lookup(COMPL_POOL_NAME); + CU_ASSERT(pool == pool_a || pool == pool_b); + + CU_ASSERT(odp_pool_destroy(pool_a) == 0); + CU_ASSERT(odp_pool_destroy(pool_b) == 0); +} + +static int check_event_user_area(void) +{ + if (global.disabled) + return ODP_TEST_INACTIVE; + + if (((global.ml_capa.load.compl_mode_mask | global.ml_capa.run.compl_mode_mask) & + ODP_ML_COMPL_MODE_EVENT) && + (global.ml_capa.pool.max_uarea_size > 0)) + return ODP_TEST_ACTIVE; + + return ODP_TEST_INACTIVE; +} + +static void test_ml_compl_user_area(void) +{ + uint32_t i; + void *addr; + void *prev; + odp_pool_t pool; + odp_ml_compl_pool_param_t pool_param; + uint32_t size = global.ml_capa.pool.max_uarea_size; + uint32_t num = global.num_compl; + odp_ml_compl_t compl_evs[num]; + + odp_ml_compl_pool_param_init(&pool_param); + pool_param.num = num; + pool_param.uarea_size = size; + pool = odp_ml_compl_pool_create(NULL, &pool_param); + CU_ASSERT_FATAL(pool != ODP_POOL_INVALID); + + prev = NULL; + for (i = 0; i < num; i++) { + odp_event_t ev; + int flag = 0; + + compl_evs[i] = odp_ml_compl_alloc(pool); + + if (compl_evs[i] == ODP_ML_COMPL_INVALID) + break; + + addr = odp_ml_compl_user_area(compl_evs[i]); + + CU_ASSERT_FATAL(addr != NULL); + CU_ASSERT(prev != addr); + + memset(addr, 0, size); + + ev = odp_ml_compl_to_event(compl_evs[i]); + CU_ASSERT(odp_event_user_area(ev) == addr); + CU_ASSERT(odp_event_user_area_and_flag(ev, &flag) == addr); + CU_ASSERT(flag < 0); + + prev = addr; + } + CU_ASSERT(i == num); + + for (uint32_t j = 0; j < i; j++) + odp_ml_compl_free(compl_evs[j]); + + CU_ASSERT(odp_pool_destroy(pool) == 0); +} + +static int check_event_user_area_init(void) +{ + if (global.disabled) + return ODP_TEST_INACTIVE; + + if (global.ml_capa.pool.max_uarea_size > 0 && global.ml_capa.pool.uarea_persistence) + return ODP_TEST_ACTIVE; + + return ODP_TEST_INACTIVE; +} + +static void init_event_uarea(void *uarea, uint32_t size, void *args, uint32_t index) +{ + uarea_init_t *data = args; + + data->count++; + data->mark[index] = 1; + memset(uarea, UAREA, size); +} + +static void test_ml_compl_user_area_init(void) +{ + odp_ml_compl_pool_param_t pool_param; + uint32_t num = global.num_compl, i; + odp_pool_t pool; + uarea_init_t data; + odp_ml_compl_t compl_evs[num]; + uint8_t *uarea; + + memset(&data, 0, sizeof(uarea_init_t)); + odp_ml_compl_pool_param_init(&pool_param); + pool_param.uarea_init.init_fn = init_event_uarea; + pool_param.uarea_init.args = &data; + pool_param.num = num; + pool_param.uarea_size = 1; + pool = odp_ml_compl_pool_create(NULL, &pool_param); + + CU_ASSERT_FATAL(pool != ODP_POOL_INVALID); + CU_ASSERT(data.count == num); + + for (i = 0; i < num; i++) { + CU_ASSERT(data.mark[i] == 1); + + compl_evs[i] = odp_ml_compl_alloc(pool); + + CU_ASSERT(compl_evs[i] != ODP_ML_COMPL_INVALID); + + if (compl_evs[i] == ODP_ML_COMPL_INVALID) + break; + + uarea = odp_ml_compl_user_area(compl_evs[i]); + + CU_ASSERT(*uarea == UAREA); + } + + for (uint32_t j = 0; j < i; j++) + odp_ml_compl_free(compl_evs[j]); + + odp_pool_destroy(pool); +} + +static void test_ml_fp32_to_uint8(void) +{ + uint8_t u8[8]; + float fp[8] = {-20.f, -16.4f, -14.6f, -12.5f, 0, 31.4f, 80.f, 96.3f}; + uint8_t expected[8] = {0, 0, 4, 10, 43, 127, 255, 255}; + float scale = 0.3746f; + uint8_t zero_point = 43; + + odp_ml_fp32_to_uint8(u8, fp, 8, scale, zero_point); + for (uint32_t i = 0; i < 8; i++) + CU_ASSERT(u8[i] == expected[i]); +} + +static void test_ml_fp32_from_uint8(void) +{ + float fp[4]; + float scale = 0.4f; + uint8_t zero_point = 43; + uint8_t u8[4] = {0, 43, 145, 255}; + float expected[4] = {-17.2f, 0.0f, 40.8f, 84.8f}; + + odp_ml_fp32_from_uint8(fp, u8, 4, scale, zero_point); + for (uint32_t i = 0; i < 4; i++) + CU_ASSERT(fp[i] == expected[i]); +} + +static void test_ml_fp32_to_int8(void) +{ + int8_t i8[5]; + float scale = 0.0223f; + int8_t zero_point = 0; + float fp32[5] = {-3.4f, -2.5f, 0, 1.4f, 2.9f}; + int8_t i8_expected[5] = {-127, -112, 0, 63, 127}; + + odp_ml_fp32_to_int8(i8, fp32, 5, scale, zero_point); + + for (uint32_t i = 0; i < 5; i++) + CU_ASSERT(i8[i] == i8_expected[i]); +} + +static void test_ml_fp32_to_int8_positive_zp(void) +{ + int8_t i8[6]; + float scale = 0.0223f; + int8_t zero_point = 56; + float fp32[6] = {-4.1f, -3.4f, -2.5f, 0, 1.4f, 2.9f}; + int8_t i8_expected[6] = {-127, -96, -56, 56, 119, 127}; + + odp_ml_fp32_to_int8(i8, fp32, 6, scale, zero_point); + + for (uint32_t i = 0; i < 6; i++) + CU_ASSERT(i8[i] == i8_expected[i]); +} + +static void test_ml_fp32_to_int8_negative_zp(void) +{ + int8_t i8[6]; + float scale = 0.0223f; + int8_t zero_point = -56; + float fp32[6] = {-3.4f, -2.5f, 0, 1.4f, 2.9f, 4.1f}; + int8_t i8_expected[6] = {-127, -127, -56, 7, 74, 127}; + + odp_ml_fp32_to_int8(i8, fp32, 6, scale, zero_point); + + for (uint32_t i = 0; i < 6; i++) + CU_ASSERT(i8[i] == i8_expected[i]); +} + +static void test_ml_fp32_from_int8(void) +{ + float fp32[6]; + float scale = 0.05f; + int8_t zero_point = 56; + int8_t i8[6] = {-128, 46, 0, 56, 85, 127}; + float fp32_expected[6] = {-9.2f, -0.5f, -2.8f, 0.0f, 1.45f, 3.55f}; + + odp_ml_fp32_from_int8(fp32, i8, 6, scale, zero_point); + + for (uint32_t i = 0; i < 6; i++) + CU_ASSERT(fp32[i] == fp32_expected[i]); +} + +static int approx_equal(double a, double b) +{ + const double tolerance = .01; + + if (a < 0 && b < 0) { + a = -a; + b = -b; + } + + if (a < b) { + double tmp = a; + + a = b; + b = tmp; + } + + return (a * (1 - tolerance) < b && a * (1 + tolerance) > b); +} + +static void test_ml_fp32_fp16(void) +{ + float fp32[4]; + uint16_t fp16[4]; + float big = 1, small = 1; + const float phi = 1.618033988749; + + fp32[0] = 0.0; + fp32[1] = -0.0; + memset(fp16, 1, sizeof(fp16)); + odp_ml_fp32_to_fp16(fp16, fp32, 2); + memset(fp32, 1, sizeof(fp32)); + odp_ml_fp32_from_fp16(fp32, fp16, 2); + CU_ASSERT(fp32[0] == 0); + CU_ASSERT(fp32[1] == 0); + + /* + * 65504 is the largest normal number for fp16 with 5 exponent bits (IEEE 754-2008). + */ + while (big < 65504 / 2) { + fp32[0] = big; + fp32[1] = -big; + fp32[2] = small; + fp32[3] = -small; + memset(fp16, 0, sizeof(fp16)); + odp_ml_fp32_to_fp16(fp16, fp32, 4); + memset(fp32, 0, sizeof(fp32)); + odp_ml_fp32_from_fp16(fp32, fp16, 4); + CU_ASSERT(approx_equal(fp32[0], big)); + CU_ASSERT(approx_equal(fp32[1], -big)); + CU_ASSERT(approx_equal(fp32[2], small)); + CU_ASSERT(approx_equal(fp32[3], -small)); + big *= phi; + small /= phi; + } +} + +odp_testinfo_t ml_suite[] = { + ODP_TEST_INFO(test_ml_capability), + ODP_TEST_INFO_CONDITIONAL(test_ml_param_init, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_debug, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(ml_compl_pool_create_max_pools, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(compl_pool_info, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(compl_alloc_max, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(compl_pool_lookup, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_compl_user_area, check_event_user_area), + ODP_TEST_INFO_CONDITIONAL(test_ml_compl_user_area_init, check_event_user_area_init), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_to_uint8, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_from_uint8, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_to_int8, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_to_int8_positive_zp, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_to_int8_negative_zp, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_from_int8, check_ml_support), + ODP_TEST_INFO_CONDITIONAL(test_ml_fp32_fp16, check_ml_support), + ODP_TEST_INFO_NULL +}; + +odp_suiteinfo_t ml_suites[] = { + {"ML", ml_suite_init, NULL, ml_suite}, + ODP_SUITE_INFO_NULL +}; + +int main(int argc, char *argv[]) +{ + int ret; + + /* parse common options: */ + if (odp_cunit_parse_options(&argc, argv)) + return -1; + + ret = odp_cunit_register(ml_suites); + + if (ret == 0) + ret = odp_cunit_run(); + + return ret; +} |