diff options
Diffstat (limited to 'platform/linux-generic/odp_ml.c')
| -rw-r--r-- | platform/linux-generic/odp_ml.c | 2651 |
1 files changed, 2651 insertions, 0 deletions
diff --git a/platform/linux-generic/odp_ml.c b/platform/linux-generic/odp_ml.c new file mode 100644 index 000000000..a5d055f4f --- /dev/null +++ b/platform/linux-generic/odp_ml.c @@ -0,0 +1,2651 @@ +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright (c) 2023 Nokia + */ + +#include <odp/autoheader_external.h> + +#include <odp/api/atomic.h> +#include <odp/api/buffer.h> +#include <odp/api/event.h> +#include <odp/api/hints.h> +#include <odp/api/ml.h> +#include <odp/api/pool.h> +#include <odp/api/queue.h> +#include <odp/api/shared_memory.h> +#include <odp/api/std_types.h> +#include <odp/api/ticketlock.h> + +#include <odp/api/plat/event_inline_types.h> +#include <odp/api/plat/strong_types.h> + +#include <odp_buffer_internal.h> +#include <odp_config_internal.h> +#include <odp_debug_internal.h> +#include <odp_global_data.h> +#include <odp_init_internal.h> +#include <odp_libconfig_internal.h> +#include <odp_macros_internal.h> +#include <odp_pool_internal.h> +#include <odp_string_internal.h> + +#include <onnxruntime_c_api.h> + +#include <inttypes.h> +#include <stdint.h> +#include <stdio.h> +#include <string.h> + +#define ML_MAX_IO_SEGS UINT32_MAX +#define ML_MAX_COMPL_ID 32 +#define ML_MAX_CONFIG_STR_LEN 65 +#define ML_MAX_MODEL_SIZE (1024 * 1024 * 1024) +#define ML_MAX_MODELS_CREATED CONFIG_ML_MAX_MODELS +#define ML_MAX_MODELS_LOADED CONFIG_ML_MAX_MODELS + +/* Error codes */ +enum { + /* Feature not supported */ + ML_FEATURE_NOT_SUPPORTED = 1, + + /* Model is not created */ + ML_NOT_CREATED, + + /* Model was not loaded */ + ML_NOT_LOADED, + + /* Model has already loaded */ + ML_LOADED, + + /* Bad input */ + ML_BAD_INPUT, + + /* Fail from underlying library onnxruntime */ + ML_LIB_FAILED, + + /* Bad output */ + ML_BAD_OUTPUT, + + /* Bad handle */ + ML_BAD_HDL +}; + +typedef struct ort_run_opts_t { + int enable_profiling; + + ExecutionMode execution_mode; + + int inter_op_num_threads; + + int intra_op_num_threads; + + GraphOptimizationLevel graph_opt_level; + + char opt_model_filepath[ML_MAX_CONFIG_STR_LEN]; +} ort_run_opts_t; + +typedef struct ml_input_t { + /* Combined input start address */ + void *addr; + /* Data size in bytes */ + uint64_t size; +} ml_input_t; + +/* Onnxruntime model info */ +typedef struct ml_model_t { + /* Guards state, which must be accessed atomically */ + odp_ticketlock_t lock; + + enum { + ML_STATE_FREE = 0, /* Not allocated */ + ML_STATE_CREATED, /* Model is created */ + ML_STATE_LOADED, /* Model is loaded */ + ML_STATE_INFERENCING, /* Model is inferencing */ + } state; + + OrtSession *session; + OrtSessionOptions *session_opts; + uint32_t max_compl_id; + odp_atomic_u32_t compl_status[ML_MAX_COMPL_ID]; + + odp_ml_model_info_t info; + odp_ml_input_info_t input_info[CONFIG_ML_MAX_INPUTS]; + uint64_t input_sizes[CONFIG_ML_MAX_INPUTS]; + odp_ml_output_info_t output_info[CONFIG_ML_MAX_OUTPUTS]; + uint64_t output_sizes[CONFIG_ML_MAX_OUTPUTS]; + + struct { + void *user_ptr; + } result[ML_MAX_COMPL_ID]; +} ml_model_t; + +typedef struct ml_global_t { + odp_shm_t shm; + + odp_ml_capability_t capa; + odp_ml_config_t ml_config; + + odp_pool_param_t pool_param; + + const OrtApi *ort_api; + OrtEnv *env; + ort_run_opts_t ort_run_opts; + + ml_model_t models[ML_MAX_MODELS_CREATED]; + +} ml_global_t; + +static ml_global_t *_odp_ml_glb; + +static inline ml_model_t *ml_model_from_handle(odp_ml_model_t model) +{ + return (ml_model_t *)(uintptr_t)model; +} + +int odp_ml_capability(odp_ml_capability_t *capa) +{ + odp_pool_capability_t pool_capa; + + memset(capa, 0, sizeof(odp_ml_capability_t)); + + if (odp_global_ro.disable.ml) { + _ODP_PRINT("ML is disabled\n"); + return 0; + } + + capa->max_model_size = ML_MAX_MODEL_SIZE; + capa->max_models = ML_MAX_MODELS_CREATED; + capa->max_models_loaded = ML_MAX_MODELS_LOADED; + capa->max_compl_id = ML_MAX_COMPL_ID; + capa->max_inputs = CONFIG_ML_MAX_INPUTS; + capa->max_outputs = CONFIG_ML_MAX_OUTPUTS; + capa->max_segs_per_input = ML_MAX_IO_SEGS; + capa->max_segs_per_output = ML_MAX_IO_SEGS; + capa->min_input_align = 1; + capa->min_output_align = 1; + + capa->load.compl_mode_mask = ODP_ML_COMPL_MODE_SYNC | + ODP_ML_COMPL_MODE_POLL | + ODP_ML_COMPL_MODE_EVENT; + capa->load.compl_queue_plain = 1; + capa->load.compl_queue_sched = 1; + + capa->run.compl_mode_mask = ODP_ML_COMPL_MODE_SYNC | + ODP_ML_COMPL_MODE_POLL | + ODP_ML_COMPL_MODE_EVENT; + capa->run.compl_queue_plain = 1; + capa->run.compl_queue_sched = 1; + + if (odp_pool_capability(&pool_capa)) { + _ODP_ERR("Pool capability failed\n"); + return -1; + } + + capa->pool.max_pools = pool_capa.buf.max_pools; + capa->pool.max_num = pool_capa.buf.max_num; + capa->pool.max_uarea_size = pool_capa.buf.max_uarea_size; + capa->pool.uarea_persistence = pool_capa.buf.uarea_persistence; + capa->pool.max_cache_size = pool_capa.buf.max_cache_size; + capa->pool.min_cache_size = pool_capa.buf.min_cache_size; + + return 0; +} + +void odp_ml_config_init(odp_ml_config_t *config) +{ + memset(config, 0, sizeof(odp_ml_config_t)); + config->max_models_created = 1; + config->max_models_loaded = 1; +} + +int odp_ml_config(const odp_ml_config_t *config) +{ + if (!config) { + _ODP_ERR("Error: config must not be NULL\n"); + return -1; + } + + if (config->max_model_size == 0 || config->max_models_created == 0 || + config->max_models_loaded == 0) { + _ODP_ERR("Error: max_model_size, max_models_created and max_models_loaded" + " must be bigger than 0\n"); + return -1; + } + + if (config->max_models_loaded > config->max_models_created) { + _ODP_ERR("Error: max_models_loaded %d exceeds max_models_created %d\n", + config->max_models_loaded, config->max_models_created); + return -1; + } + + if (config->max_models_created > ML_MAX_MODELS_CREATED) { + _ODP_ERR("Error: max_models_created %d exceeds maximum number" + " of models that can be created in this driver %d\n", + config->max_models_created, ML_MAX_MODELS_CREATED); + return -1; + } + + if (config->max_models_loaded > ML_MAX_MODELS_LOADED) { + _ODP_ERR("Error: max_models_loaded %d exceeds maximum number" + " of models that can be loaded in this driver %d\n", + config->max_models_loaded, ML_MAX_MODELS_LOADED); + return -1; + } + + if (config->max_model_size > ML_MAX_MODEL_SIZE) { + _ODP_ERR("max_model_size %" PRIu64 " exceeds supported maximum model size %d\n", + config->max_model_size, ML_MAX_MODEL_SIZE); + return -1; + } + + _odp_ml_glb->ml_config = *config; + return 0; +} + +void odp_ml_model_param_init(odp_ml_model_param_t *param) +{ + memset(param, 0, sizeof(odp_ml_model_param_t)); +} + +static int check_ortstatus(OrtStatus * const status) +{ + if (status != NULL) { + const char *msg = _odp_ml_glb->ort_api->GetErrorMessage(status); + + _ODP_ERR("%s\n", msg); + _odp_ml_glb->ort_api->ReleaseStatus(status); + return -1; + } + + return 0; +} + +/* Get model input and output count */ +static int get_model_io_count(OrtSession *model, uint32_t *num_inputs, uint32_t *num_outputs) +{ + size_t num = 0; + OrtStatus *status = NULL; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + status = ort_api->SessionGetInputCount(model, &num); + if (check_ortstatus(status)) { + _ODP_ERR("Get model input count failed\n"); + return -1; + } + + *num_inputs = num; + _ODP_DBG("num_inputs: %u\n", *num_inputs); + + status = ort_api->SessionGetOutputCount(model, &num); + if (check_ortstatus(status)) { + _ODP_ERR("Get model output count failed\n"); + return -1; + } + + *num_outputs = num; + _ODP_DBG("num_outputs: %u\n", *num_outputs); + + return 0; +} + +static odp_ml_data_type_t onnx_dtype_to_odp_dtype(ONNXTensorElementDataType onnx_dtype) +{ + switch (onnx_dtype) { + case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT: + return ODP_ML_DATA_TYPE_FP32; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8: + return ODP_ML_DATA_TYPE_UINT8; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8: + return ODP_ML_DATA_TYPE_INT8; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16: + return ODP_ML_DATA_TYPE_UINT16; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16: + return ODP_ML_DATA_TYPE_INT16; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32: + return ODP_ML_DATA_TYPE_INT32; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32: + return ODP_ML_DATA_TYPE_UINT32; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64: + return ODP_ML_DATA_TYPE_INT64; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64: + return ODP_ML_DATA_TYPE_UINT64; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16: + return ODP_ML_DATA_TYPE_FP16; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16: + return ODP_ML_DATA_TYPE_BFP16; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE: + return ODP_ML_DATA_TYPE_FP64; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL: + /* Fall through */ + case ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX64: + /* Fall through */ + case ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX128: + /* Fall through */ + default: + _ODP_ERR("onnx_dtype %d not supported by odp_ml\n", onnx_dtype); + return ODP_ML_DATA_TYPE_NONE; + } +} + +/* Get the size of given odp_ml_data_type_t in bytes */ +static uint32_t size_of_odp_ml_data_type(odp_ml_data_type_t data_type) +{ + switch (data_type) { + case ODP_ML_DATA_TYPE_NONE: + return 0; + + case ODP_ML_DATA_TYPE_INT8: + /* Fall through */ + case ODP_ML_DATA_TYPE_UINT8: + return 1; + + case ODP_ML_DATA_TYPE_INT16: + /* Fall through */ + case ODP_ML_DATA_TYPE_UINT16: + /* Fall through */ + case ODP_ML_DATA_TYPE_FP16: + /* Fall through */ + case ODP_ML_DATA_TYPE_BFP16: + return 2; + + case ODP_ML_DATA_TYPE_INT24: + /* Fall through */ + case ODP_ML_DATA_TYPE_UINT24: + return 3; + + case ODP_ML_DATA_TYPE_INT32: + /* Fall through */ + case ODP_ML_DATA_TYPE_UINT32: + /* Fall through */ + case ODP_ML_DATA_TYPE_FP32: + return 4; + + case ODP_ML_DATA_TYPE_INT64: + /* Fall through */ + case ODP_ML_DATA_TYPE_UINT64: + /* Fall through */ + case ODP_ML_DATA_TYPE_FP64: + return 8; + + default: + return 0; + } +} + +static int get_shape(int64_t dims[], odp_ml_shape_info_t *shape) +{ + uint32_t dyn_cnt = 0; + + for (uint32_t i = 0; i < shape->num_dim; i++) { + if (dims[i] == 0) { + _ODP_ERR("Dimension value: %" PRId64 " must be at least 1\n", dims[i]); + return -1; + } else if (dims[i] == -1) { /* Symbolic dimension */ + dyn_cnt++; + shape->dim[i] = ODP_ML_DIM_DYNAMIC; + shape->dim_min[i] = 0; /*unknown*/ + shape->dim_max[i] = 0; /*unknown*/ + } else if (dims[i] > 0 && dims[i] < UINT32_MAX) { + shape->dim[i] = dims[i]; + shape->dim_min[i] = dims[i]; + shape->dim_max[i] = dims[i]; + } else { + _ODP_ERR("Dimension value: %" PRId64 " invalid\n", dims[i]); + return -1; + } + } + + if (dyn_cnt == 0) { + shape->type = ODP_ML_SHAPE_STATIC; + } else if (dyn_cnt == 1) { + shape->type = ODP_ML_SHAPE_BATCH; + } else { + _ODP_ERR("Data shape type not supported by ODP\n"); + return -1; + } + + return 0; +} + +static inline void calculate_model_io_size(const odp_ml_shape_info_t *shape, uint64_t *size) +{ + /* Calculate the data size in bytes of this tensor, 0 for tensors with + * dynamic batch sizes */ + for (size_t i = 0; i < shape->num_dim; i++) { + /* Skip dynamic dimension size */ + if (shape->dim[i] == ODP_ML_DIM_DYNAMIC) { + *size = 0; + break; + } + (*size) *= shape->dim[i]; + } +} + +static int get_model_io_type_shape_size(OrtTypeInfo *type_info, odp_ml_shape_info_t *shape, + odp_ml_data_type_t *data_type, uint32_t *data_type_size, + uint64_t *size) +{ + ONNXTensorElementDataType tensor_type; + const OrtTensorTypeAndShapeInfo *tensor_info; + size_t num_dim = 0; + OrtStatus *status = NULL; + int64_t dims[ODP_ML_MAX_DIMS] = {0}; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + status = ort_api->CastTypeInfoToTensorInfo(type_info, &tensor_info); + if (check_ortstatus(status)) { + _ODP_ERR("CastTypeInfoToTensorInfo failed\n"); + return -1; + } + + status = ort_api->GetTensorElementType(tensor_info, &tensor_type); + if (check_ortstatus(status)) { + _ODP_ERR("GetTensorElementType failed\n"); + return -1; + } + + *data_type = onnx_dtype_to_odp_dtype(tensor_type); + if (*data_type == ODP_ML_DATA_TYPE_NONE) /* Type not supported by odp */ + return -1; + + status = ort_api->GetDimensionsCount(tensor_info, &num_dim); + if (check_ortstatus(status)) { + _ODP_ERR("GetDimensionsCount failed\n"); + return -1; + } + + if (num_dim > ODP_ML_MAX_DIMS) { + _ODP_ERR("Number of dimensions: %zu exceeds supported maximum number" + " of dimensions: %d\n", num_dim, ODP_ML_MAX_DIMS); + return -1; + } + shape->num_dim = num_dim; + + status = ort_api->GetDimensions(tensor_info, dims, num_dim); + if (check_ortstatus(status)) { + _ODP_ERR("GetDimensions failed\n"); + return -1; + } + + if (get_shape(dims, shape)) + return -1; + + *data_type_size = size_of_odp_ml_data_type(*data_type); + + *size = *data_type_size; + calculate_model_io_size(shape, size); + + return 0; +} + +/* Get model input and output info */ +static int get_model_io_info(OrtSession *session, ml_model_t *mdl, + const odp_ml_model_param_t *param) +{ + char *name; + OrtTypeInfo *type_info; + const odp_ml_data_format_t *data_format; + OrtStatus *status = NULL; + OrtAllocator *allocator = NULL; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + odp_ml_input_info_t *input_info = mdl->input_info; + odp_ml_output_info_t *output_info = mdl->output_info; + + status = ort_api->GetAllocatorWithDefaultOptions(&allocator); + if (check_ortstatus(status)) { + _ODP_ERR("GetAllocatorWithDefaultOptions failed\n"); + return -1; + } + + /* Retrieve info about input array. */ + memset(input_info, 0, sizeof(mdl->input_info)); + for (uint32_t i = 0; i < mdl->info.num_inputs; i++) { + name = NULL; + status = ort_api->SessionGetInputName(session, i, allocator, &name); + if (check_ortstatus(status)) { + _ODP_ERR("Get %uth input name failed\n", i); + return -1; + } + + _odp_strcpy(input_info[i].name, name, ODP_ML_MODEL_IO_NAME_LEN); + + /* Free memory allocated by SessionGetInputName */ + status = ort_api->AllocatorFree(allocator, name); + if (check_ortstatus(status)) { + _ODP_ERR("AllocatorFree %uth input_name failed\n", i); + return -1; + } + + if (param->extra_info.num_inputs) { + data_format = ¶m->extra_info.input_format[i]; + + input_info[i].shape = data_format->shape; + input_info[i].data_type = data_format->data_type; + input_info[i].data_type_size = data_format->data_type_size; + + mdl->input_sizes[i] = input_info[i].data_type_size; + calculate_model_io_size(&data_format->shape, &mdl->input_sizes[i]); + continue; + } + + type_info = NULL; + status = ort_api->SessionGetInputTypeInfo(session, i, &type_info); + if (check_ortstatus(status)) { + _ODP_ERR("SessionGetInputTypeInfo failed\n"); + return -1; + } + + if (get_model_io_type_shape_size(type_info, &input_info[i].shape, + &input_info[i].data_type, + &input_info[i].data_type_size, + &mdl->input_sizes[i])) { + _ODP_ERR("get_model_io_type_shape_size() for input failed\n"); + ort_api->ReleaseTypeInfo(type_info); + return -1; + } + + ort_api->ReleaseTypeInfo(type_info); + } + + /* Retrieve info about output array. */ + memset(output_info, 0, sizeof(mdl->output_info)); + for (uint32_t i = 0; i < mdl->info.num_outputs; i++) { + name = NULL; + status = ort_api->SessionGetOutputName(session, i, allocator, &name); + if (check_ortstatus(status)) { + _ODP_ERR("Get %uth output name failed\n", i); + return -1; + } + + _odp_strcpy(output_info[i].name, name, ODP_ML_MODEL_IO_NAME_LEN); + + /* Free memory allocated by SessionGetOutputName */ + status = ort_api->AllocatorFree(allocator, name); + if (check_ortstatus(status)) { + _ODP_ERR("AllocatorFree %uth output_name failed\n", i); + return -1; + } + + if (param->extra_info.num_outputs) { + data_format = ¶m->extra_info.output_format[i]; + + output_info[i].shape = data_format->shape; + output_info[i].data_type = data_format->data_type; + output_info[i].data_type_size = data_format->data_type_size; + + mdl->output_sizes[i] = output_info[i].data_type_size; + calculate_model_io_size(&data_format->shape, &mdl->output_sizes[i]); + continue; + } + + type_info = NULL; + status = ort_api->SessionGetOutputTypeInfo(session, i, &type_info); + if (check_ortstatus(status)) { + _ODP_ERR("SessionGetOutputTypeInfo failed\n"); + return -1; + } + + if (get_model_io_type_shape_size(type_info, &output_info[i].shape, + &output_info[i].data_type, + &output_info[i].data_type_size, + &mdl->output_sizes[i])) { + _ODP_ERR("get_model_io_type_shape_size() for output failed\n"); + ort_api->ReleaseTypeInfo(type_info); + return -1; + } + + ort_api->ReleaseTypeInfo(type_info); + } + + return 0; +} + +static inline int check_model_io_num(const odp_ml_model_param_t *param, + uint32_t num_inputs, uint32_t num_outputs) +{ + /* Make sure the number of inputs/outputs not exceeding the supported + * model max inputs/outputs */ + if (num_inputs > CONFIG_ML_MAX_INPUTS) { + _ODP_ERR("The model's number of inputs %u exceeds the maximum " + "number of inputs supported in a model %u\n", + num_inputs, CONFIG_ML_MAX_INPUTS); + return -1; + } + + if (num_outputs > CONFIG_ML_MAX_OUTPUTS) { + _ODP_ERR("The model's number of outputs %u exceeds the maximum " + "number of outputs supported in a model %u\n", + num_outputs, CONFIG_ML_MAX_OUTPUTS); + + return -1; + } + + /* Make sure the numbers of inputs/outputs provided in the extra_info of + * param match the numbers defined in model metadata. */ + if (param->extra_info.num_inputs && + param->extra_info.num_inputs != num_inputs) { + _ODP_ERR("Provided param->extra_info.num_inputs %u does not match the" + " number of inputs defined in model metadata: %u\n", + param->extra_info.num_inputs, num_inputs); + return -1; + } + + if (param->extra_info.num_outputs && param->extra_info.num_outputs != num_outputs) { + _ODP_ERR("Provided param->extra_info.num_outputs %u does not match the" + " number of outputs defined in model metadata: %u\n", + param->extra_info.num_outputs, num_outputs); + return -1; + } + + if (param->extra_info.num_inputs && !param->extra_info.input_format) { + _ODP_ERR("num_inputs is provided but not input_format in param->extra_info\n"); + return -1; + } + + if (param->extra_info.num_outputs && !param->extra_info.output_format) { + _ODP_ERR("num_outputs is provided but not output_format in param->extra_info\n"); + return -1; + } + + return 0; +} + +static int create_ort_model(const odp_ml_model_param_t *param, OrtSession **session, + ml_model_t *mdl, OrtSessionOptions *session_opts) +{ + OrtStatus *status; + int64_t model_version; + uint32_t num_inputs = 0; + uint32_t num_outputs = 0; + OrtModelMetadata *metadata = {0}; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + status = ort_api->CreateSessionFromArray(_odp_ml_glb->env, + param->model, + param->size, + session_opts, + session); + if (check_ortstatus(status) || !(*session)) { + _ODP_ERR("CreateSessionFromArray failed\n"); + return -1; + } + + if (get_model_io_count(*session, &num_inputs, &num_outputs)) { + _ODP_ERR("get_model_io_count() failed\n"); + ort_api->ReleaseSession(*session); + return -1; + } + + if (check_model_io_num(param, num_inputs, num_outputs)) { + ort_api->ReleaseSession(*session); + return -1; + } + + mdl->max_compl_id = param->max_compl_id; + mdl->info.num_inputs = num_inputs; + mdl->info.num_outputs = num_outputs; + + /* Get metadata */ + status = ort_api->SessionGetModelMetadata(*session, &metadata); + if (check_ortstatus(status) || !metadata) { + _ODP_ERR("SessionGetModelMetadata failed\n"); + ort_api->ReleaseSession(*session); + return -1; + } + + /* Get model version */ + status = ort_api->ModelMetadataGetVersion(metadata, &model_version); + if (check_ortstatus(status)) { + _ODP_ERR("ModelMetadataGetVersion failed\n"); + ort_api->ReleaseModelMetadata(metadata); + ort_api->ReleaseSession(*session); + return -1; + } + mdl->info.model_version = model_version; + mdl->info.interface_version = 0; + + if (get_model_io_info(*session, mdl, param)) { + _ODP_ERR("get_model_io_info() failed\n"); + ort_api->ReleaseModelMetadata(metadata); + ort_api->ReleaseSession(*session); + return -1; + } + + ort_api->ReleaseModelMetadata(metadata); + return 0; +} + +static int set_ort_run_opts(const char *name, OrtSessionOptions *se_opts) +{ + OrtStatus *status; + ort_run_opts_t *opts = &_odp_ml_glb->ort_run_opts; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + if (opts->enable_profiling) { + status = ort_api->EnableProfiling(se_opts, name); + if (check_ortstatus(status)) { + _ODP_ERR("Enable profiling failed\n"); + return -1; + } + } + + status = ort_api->SetSessionExecutionMode(se_opts, opts->execution_mode); + if (check_ortstatus(status)) { + _ODP_ERR("SetSessionExecutionMode failed\n"); + return -1; + } + + if (opts->intra_op_num_threads) { + status = ort_api->SetIntraOpNumThreads(se_opts, opts->intra_op_num_threads); + if (check_ortstatus(status)) { + _ODP_ERR("SetIntraOpNumThreads failed\n"); + return -1; + } + } + + if (opts->inter_op_num_threads) { + status = ort_api->SetInterOpNumThreads(se_opts, opts->inter_op_num_threads); + if (check_ortstatus(status)) { + _ODP_ERR("SetInterOpNumThreads failed\n"); + return -1; + } + } + + status = ort_api->SetSessionGraphOptimizationLevel(se_opts, opts->graph_opt_level); + if (check_ortstatus(status)) { + _ODP_ERR("SetSessionGraphOptimizationLevel failed\n"); + return -1; + } + + /* Optimized model file path is not provided */ + if (opts->opt_model_filepath[0] == '\0') + return 0; + + status = ort_api->SetOptimizedModelFilePath(se_opts, opts->opt_model_filepath); + if (check_ortstatus(status)) { + _ODP_ERR("SetOptimizedModelFilePath failed\n"); + return -1; + } + + return 0; +} + +static inline void reset_mdl_info_sizes(ml_model_t *mdl) +{ + memset(&mdl->info, 0, sizeof(odp_ml_model_info_t)); + memset(mdl->input_info, 0, sizeof(mdl->input_info)); + memset(mdl->output_info, 0, sizeof(mdl->output_info)); + memset(mdl->input_sizes, 0, sizeof(mdl->input_sizes)); + memset(mdl->output_sizes, 0, sizeof(mdl->output_sizes)); +} + +static int check_io_shape(ml_model_t *mdl) +{ + odp_ml_shape_info_t *shape; + + for (uint32_t i = 0; i < mdl->info.num_inputs; i++) { + int num_dyn = 0; + + shape = &mdl->input_info[i].shape; + + switch (shape->type) { + case ODP_ML_SHAPE_STATIC: + break; + + case ODP_ML_SHAPE_BATCH: + for (uint32_t j = 0; j < shape->num_dim; j++) + if (shape->dim[j] == ODP_ML_DIM_DYNAMIC) + num_dyn++; + + if (num_dyn != 1) { + _ODP_ERR("Incorrect number (%d) of dynamic dimensions " + "for model input[%u]\n", + num_dyn, i); + return -1; + } + + break; + + default: + _ODP_ERR("Undefined shape type for model input[%u]\n", i); + return -1; + } + } + + for (uint32_t i = 0; i < mdl->info.num_outputs; i++) { + if (mdl->output_info[i].shape.type == ODP_ML_SHAPE_NONE) { + _ODP_ERR("Undefined shape type for model output[%u]\n", i); + return -1; + } + } + + return 0; +} + +odp_ml_model_t odp_ml_model_create(const char *name, const odp_ml_model_param_t *param) +{ + OrtStatus *status; + odp_ml_model_info_t *info; + OrtSessionOptions *session_opts; + uint32_t i = 0; + ml_model_t *mdl = NULL; + OrtSession *session = NULL; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + if (odp_unlikely(odp_global_ro.disable.ml)) { + _ODP_ERR("ML is disabled\n"); + return ODP_ML_MODEL_INVALID; + } + + if (odp_unlikely(param->size > _odp_ml_glb->ml_config.max_model_size)) { + _ODP_ERR("Model size %" PRIu64 " exceeds maximum model size configured %" PRIu64 "\n", + param->size, _odp_ml_glb->ml_config.max_model_size); + return ODP_ML_MODEL_INVALID; + } + + if (odp_unlikely(!param->size || !param->model)) { + _ODP_ERR("Invalid model param: param->model: %p, param->size: %" PRIu64 "\n", + param->model, param->size); + return ODP_ML_MODEL_INVALID; + } + + if (odp_unlikely(param->max_compl_id > ML_MAX_COMPL_ID)) { + _ODP_ERR("param->max_compl_id: %u exceeds maximum completion id supported: %d\n", + param->max_compl_id, ML_MAX_COMPL_ID); + return ODP_ML_MODEL_INVALID; + } + + /* Find an emtpy slot to store the new model */ + for (i = 0; i < ML_MAX_MODELS_CREATED; i++) { + if (_odp_ml_glb->models[i].state) + continue; + + odp_ticketlock_lock(&_odp_ml_glb->models[i].lock); + + if (_odp_ml_glb->models[i].state) { + odp_ticketlock_unlock(&_odp_ml_glb->models[i].lock); + continue; + } + + mdl = &_odp_ml_glb->models[i]; + break; + } + + if (i == ML_MAX_MODELS_CREATED) { + _ODP_ERR("Maximum number of models has already been created!\n"); + return ODP_ML_MODEL_INVALID; + } + + /* Free model entry was found and is now locked */ + mdl->state = ML_STATE_CREATED; + + status = ort_api->CreateSessionOptions(&session_opts); + if (check_ortstatus(status) || !session_opts) { + _ODP_ERR("Error: CreateSessionOptions failed.\n"); + mdl->state = ML_STATE_FREE; + odp_ticketlock_unlock(&mdl->lock); + return ODP_ML_MODEL_INVALID; + } + + if (set_ort_run_opts(name, session_opts)) { + _odp_ml_glb->ort_api->ReleaseSessionOptions(session_opts); + mdl->state = ML_STATE_FREE; + odp_ticketlock_unlock(&mdl->lock); + return ODP_ML_MODEL_INVALID; + } + + /* Store model info */ + info = &mdl->info; + memset(info, 0, sizeof(odp_ml_model_info_t)); + + if (create_ort_model(param, &session, mdl, session_opts)) { + mdl->state = ML_STATE_FREE; + + /* Initialize info back to 0 when some fields have been filled + * while later failed */ + reset_mdl_info_sizes(mdl); + odp_ticketlock_unlock(&mdl->lock); + + _odp_ml_glb->ort_api->ReleaseSessionOptions(session_opts); + _ODP_ERR("create_ort_model() failed\n"); + return ODP_ML_MODEL_INVALID; + } + + if (check_io_shape(mdl)) { + mdl->state = ML_STATE_FREE; + reset_mdl_info_sizes(mdl); + odp_ticketlock_unlock(&mdl->lock); + + ort_api->ReleaseSession(session); + _odp_ml_glb->ort_api->ReleaseSessionOptions(session_opts); + return ODP_ML_MODEL_INVALID; + } + + mdl->session = session; + mdl->session_opts = session_opts; + info->index = i; + + if (name) + _odp_strcpy(info->name, name, ODP_ML_MODEL_NAME_LEN); + + mdl->max_compl_id = param->max_compl_id; + for (uint32_t j = 0; j < ML_MAX_COMPL_ID; j++) + odp_atomic_init_u32(&mdl->compl_status[j], 1); + + odp_ticketlock_unlock(&mdl->lock); + return (odp_ml_model_t)mdl; +} + +int odp_ml_model_destroy(odp_ml_model_t model) +{ + ml_model_t *mdl = ml_model_from_handle(model); + + if (model == ODP_ML_MODEL_INVALID) { + _ODP_ERR("Bad ML model handle\n"); + return -1; + } + + odp_ticketlock_lock(&mdl->lock); + + if (mdl->state != ML_STATE_CREATED) { + _ODP_ERR("Model not created\n"); + odp_ticketlock_unlock(&mdl->lock); + return -1; + } + + _odp_ml_glb->ort_api->ReleaseSessionOptions(mdl->session_opts); + _odp_ml_glb->ort_api->ReleaseSession(mdl->session); + mdl->state = ML_STATE_FREE; + mdl->session = NULL; + odp_ticketlock_unlock(&mdl->lock); + + return 0; +} + +int odp_ml_model_info(odp_ml_model_t model, odp_ml_model_info_t *info) +{ + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return -1; + } + + if (odp_unlikely(!info)) { + _ODP_ERR("info must not be NULL\n"); + return -1; + } + + odp_ticketlock_lock(&mdl->lock); + if (odp_unlikely(mdl->state == ML_STATE_FREE)) { + _ODP_ERR("Model not created\n"); + odp_ticketlock_unlock(&mdl->lock); + return -1; + } + + *info = mdl->info; + + odp_ticketlock_unlock(&mdl->lock); + return 0; +} + +uint32_t odp_ml_model_input_info(odp_ml_model_t model, odp_ml_input_info_t info[], uint32_t num) +{ + uint32_t num_model_inputs; + uint32_t num_written; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return 0; + } + + odp_ticketlock_lock(&mdl->lock); + num_model_inputs = mdl->info.num_inputs; + num_written = num_model_inputs >= num ? num : num_model_inputs; + + if (num == 0) { + odp_ticketlock_unlock(&mdl->lock); + return num_model_inputs; + } + + for (uint32_t i = 0; i < num_written; i++) + info[i] = mdl->input_info[i]; + + odp_ticketlock_unlock(&mdl->lock); + return num_model_inputs; +} + +uint32_t odp_ml_model_output_info(odp_ml_model_t model, odp_ml_output_info_t info[], uint32_t num) +{ + uint32_t num_model_outputs; + uint32_t num_written; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return 0; + } + + odp_ticketlock_lock(&mdl->lock); + num_model_outputs = mdl->info.num_outputs; + num_written = num_model_outputs >= num ? num : num_model_outputs; + + if (num == 0) { + odp_ticketlock_unlock(&mdl->lock); + return num_model_outputs; + } + + for (uint32_t i = 0; i < num_written; i++) + info[i] = mdl->output_info[i]; + + odp_ticketlock_unlock(&mdl->lock); + return num_model_outputs; +} + +odp_ml_model_t odp_ml_model_lookup(const char *name) +{ + uint32_t i; + ml_model_t *mdl; + + for (i = 0; i < ML_MAX_MODELS_CREATED; i++) { + mdl = &_odp_ml_glb->models[i]; + + odp_ticketlock_lock(&mdl->lock); + + if (mdl->state == ML_STATE_FREE) { + odp_ticketlock_unlock(&mdl->lock); + continue; + } + + if (!strcmp(mdl->info.name, name)) { + /* found it */ + odp_ticketlock_unlock(&mdl->lock); + return (odp_ml_model_t)mdl; + } + odp_ticketlock_unlock(&mdl->lock); + } + + return ODP_ML_MODEL_INVALID; +} + +uint64_t odp_ml_model_to_u64(odp_ml_model_t model) +{ + return _odp_pri(model); +} + +static const char *data_type_str(odp_ml_data_type_t data_type) +{ + switch (data_type) { + case ODP_ML_DATA_TYPE_INT8: + return "int8"; + case ODP_ML_DATA_TYPE_UINT8: + return "uint8"; + case ODP_ML_DATA_TYPE_UINT16: + return "uint16"; + case ODP_ML_DATA_TYPE_INT16: + return "int16"; + case ODP_ML_DATA_TYPE_INT32: + return "int32"; + case ODP_ML_DATA_TYPE_UINT32: + return "uint32"; + case ODP_ML_DATA_TYPE_INT64: + return "int64"; + case ODP_ML_DATA_TYPE_UINT64: + return "uint64"; + case ODP_ML_DATA_TYPE_FP16: + return "fp16"; + case ODP_ML_DATA_TYPE_FP32: + return "fp32"; + case ODP_ML_DATA_TYPE_BFP16: + return "bfp16"; + default: + return "unknown"; + } +} + +static const char *shape_type_str(odp_ml_shape_type_t shape_type) +{ + switch (shape_type) { + case ODP_ML_SHAPE_NONE: + return "none"; + case ODP_ML_SHAPE_STATIC: + return "static"; + case ODP_ML_SHAPE_BATCH: + return "batch"; + default: + return "Unknown"; + } +} + +static void print_shape(const odp_ml_shape_info_t *shape) +{ + /* Print shape */ + _ODP_PRINT("Shape: %s [", shape_type_str(shape->type)); + + for (uint32_t i = 0; i < shape->num_dim; i++) { + if (shape->dim[i] == ODP_ML_DIM_DYNAMIC) + _ODP_PRINT("Dyn"); + else + _ODP_PRINT("%" PRIu32, shape->dim[i]); + + if (i == (shape->num_dim - 1)) + _ODP_PRINT("]\n"); + else + _ODP_PRINT(", "); + } + + /* The number of dimensions for a scalar input is 0, in which case did not + * go into above for loop */ + if (shape->num_dim == 0) + _ODP_PRINT("]\n"); +} + +void odp_ml_model_print(odp_ml_model_t model) +{ + ml_model_t *mdl = ml_model_from_handle(model); + const odp_ml_model_info_t * const info = &mdl->info; + const odp_ml_input_info_t * const input_info = mdl->input_info; + const odp_ml_output_info_t * const output_info = mdl->output_info; + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return; + } + + odp_ticketlock_lock(&mdl->lock); + if (odp_unlikely(mdl->state == ML_STATE_FREE)) { + odp_ticketlock_unlock(&mdl->lock); + _ODP_ERR("Model not created\n"); + return; + } + + _ODP_PRINT("\nModel info\n"); + _ODP_PRINT("----------\n"); + _ODP_PRINT(" Model handle: 0x%" PRIx64 "\n", odp_ml_model_to_u64(model)); + _ODP_PRINT(" Name: %s\n", info->name); + _ODP_PRINT(" Model version: %" PRIu64 "\n", info->model_version); + _ODP_PRINT(" Model interface version: %" PRIu64 "\n", info->interface_version); + _ODP_PRINT(" Index: %u\n", info->index); + _ODP_PRINT(" Number of inputs: %u\n", info->num_inputs); + + for (uint32_t i = 0; i < info->num_inputs; i++) { + _ODP_PRINT(" Input[%u]: ", i); + _ODP_PRINT("Name: %s, ", input_info[i].name); + _ODP_PRINT("Data_type: %s, ", data_type_str(input_info[i].data_type)); + print_shape(&input_info[i].shape); + } + + _ODP_PRINT(" Number of outputs: %u\n", info->num_outputs); + for (uint32_t i = 0; i < info->num_outputs; i++) { + _ODP_PRINT(" Output[%u]: ", i); + _ODP_PRINT("Name: %s, ", output_info[i].name); + _ODP_PRINT("Data_type: %s, ", data_type_str(output_info[i].data_type)); + print_shape(&output_info[i].shape); + } + + odp_ticketlock_unlock(&mdl->lock); + + _ODP_PRINT("\n"); +} + +static inline void mode_print(odp_ml_compl_mode_t compl_mode_mask) +{ + if (compl_mode_mask & ODP_ML_COMPL_MODE_SYNC) + _ODP_PRINT(" syn"); + + if (compl_mode_mask & ODP_ML_COMPL_MODE_POLL) + _ODP_PRINT(" poll"); + + if (compl_mode_mask & ODP_ML_COMPL_MODE_EVENT) + _ODP_PRINT(" event"); +} + +void odp_ml_print(void) +{ + _ODP_PRINT("\nML info\n"); + _ODP_PRINT("-----------\n"); + _ODP_PRINT(" max_model_size: %u\n", ML_MAX_MODEL_SIZE); + _ODP_PRINT(" max_compl_id: %u\n", ML_MAX_COMPL_ID); + _ODP_PRINT(" max_models_created: %u\n", ML_MAX_MODELS_CREATED); + _ODP_PRINT(" max_models_loaded: %u\n", ML_MAX_MODELS_LOADED); + _ODP_PRINT(" model_max_inputs: %u\n", CONFIG_ML_MAX_INPUTS); + _ODP_PRINT(" model_max_outputs: %u\n", CONFIG_ML_MAX_OUTPUTS); + + _ODP_PRINT(" load:\n"); + _ODP_PRINT(" completion mode: "); + mode_print(_odp_ml_glb->capa.load.compl_mode_mask); + _ODP_PRINT(", plain queue: %c, schedule queue: %c\n", + _odp_ml_glb->capa.load.compl_queue_plain ? 'Y' : 'N', + _odp_ml_glb->capa.load.compl_queue_sched ? 'Y' : 'N'); + + _ODP_PRINT(" run:\n"); + _ODP_PRINT(" completion mode:"); + mode_print(_odp_ml_glb->capa.run.compl_mode_mask); + _ODP_PRINT(", plain queue: %c, schedule queue: %c\n", + _odp_ml_glb->capa.run.compl_queue_plain ? 'Y' : 'N', + _odp_ml_glb->capa.run.compl_queue_sched ? 'Y' : 'N'); + _ODP_PRINT("\n"); +} + +int odp_ml_model_extra_stat_info(odp_ml_model_t model, + odp_ml_extra_stat_info_t info[] ODP_UNUSED, + int num ODP_UNUSED) +{ + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return -1; + } + + return 0; +} + +int odp_ml_model_extra_stats(odp_ml_model_t model, uint64_t stats[] ODP_UNUSED, int num ODP_UNUSED) +{ + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return -1; + } + + return 0; +} + +void odp_ml_compl_pool_param_init(odp_ml_compl_pool_param_t *pool_param) +{ + if (odp_unlikely(!pool_param)) { + _ODP_ERR("Param 'pool_param' must not NULL\n"); + return; + } + + memset(pool_param, 0, sizeof(odp_ml_compl_pool_param_t)); + + pool_param->cache_size = _odp_ml_glb->pool_param.buf.cache_size; +} + +odp_pool_t odp_ml_compl_pool_create(const char *name, const odp_ml_compl_pool_param_t *pool_param) +{ + odp_pool_t pool; + odp_pool_param_t ml_pool_param; + uint32_t num = pool_param->num; + uint32_t uarea_size = pool_param->uarea_size; + uint32_t cache_size = pool_param->cache_size; + uint32_t buf_size = _ODP_MAX(sizeof(odp_ml_run_result_t), + sizeof(odp_ml_load_result_t)); + + if (num > _odp_ml_glb->capa.pool.max_num) { + _ODP_ERR("Too many ML completion events: %u\n", num); + return ODP_POOL_INVALID; + } + + if (uarea_size > _odp_ml_glb->capa.pool.max_uarea_size) { + _ODP_ERR("Bad uarea size: %u\n", uarea_size); + return ODP_POOL_INVALID; + } + + if (cache_size < _odp_ml_glb->capa.pool.min_cache_size || + cache_size > _odp_ml_glb->capa.pool.max_cache_size) { + _ODP_ERR("Bad cache size: %u\n", cache_size); + return ODP_POOL_INVALID; + } + + odp_pool_param_init(&ml_pool_param); + ml_pool_param.type = ODP_POOL_BUFFER; + ml_pool_param.uarea_init.init_fn = pool_param->uarea_init.init_fn; + ml_pool_param.uarea_init.args = pool_param->uarea_init.args; + ml_pool_param.buf.num = num; + ml_pool_param.buf.cache_size = cache_size; + ml_pool_param.buf.size = buf_size; + ml_pool_param.buf.uarea_size = uarea_size; + + pool = _odp_pool_create(name, &ml_pool_param, ODP_POOL_ML_COMPL); + + return pool; +} + +odp_ml_compl_t odp_ml_compl_alloc(odp_pool_t pool) +{ + odp_buffer_t buf; + odp_event_t ev; + odp_ml_run_result_t *result; + uint32_t buf_size = _ODP_MAX(sizeof(odp_ml_run_result_t), + sizeof(odp_ml_load_result_t)); + + buf = odp_buffer_alloc(pool); + + if (odp_unlikely(buf == ODP_BUFFER_INVALID)) + return ODP_ML_COMPL_INVALID; + + result = odp_buffer_addr(buf); + memset(result, 0, buf_size); + + ev = odp_buffer_to_event(buf); + _odp_event_type_set(ev, ODP_EVENT_ML_COMPL); + + return (odp_ml_compl_t)(uintptr_t)buf; +} + +void odp_ml_compl_free(odp_ml_compl_t ml_compl) +{ + odp_event_t ev; + odp_buffer_t buf = (odp_buffer_t)(uintptr_t)ml_compl; + + if (odp_unlikely(ml_compl == ODP_ML_COMPL_INVALID)) { + _ODP_ERR("Bad ML job completion handle\n"); + return; + } + + ev = odp_buffer_to_event(buf); + _odp_event_type_set(ev, ODP_EVENT_BUFFER); + + odp_buffer_free(buf); +} + +int odp_ml_compl_run_result(odp_ml_compl_t ml_compl, odp_ml_run_result_t *result) +{ + odp_event_subtype_t subtype; + odp_ml_run_result_t *run_result; + odp_buffer_t buf = (odp_buffer_t)(uintptr_t)ml_compl; + odp_event_t ev = odp_buffer_to_event(buf); + + if (odp_unlikely(ml_compl == ODP_ML_COMPL_INVALID)) { + _ODP_ERR("Given ML completion event is invalid\n"); + return -2; + } + + if (odp_event_types(ev, &subtype) != ODP_EVENT_ML_COMPL || + subtype != ODP_EVENT_ML_COMPL_RUN) { + _ODP_ERR("Given completion event has wrong event type or subtype\n"); + return -2; + } + + run_result = odp_buffer_addr(buf); + if (result) + *result = *run_result; + + return run_result->error_code ? -1 : 0; +} + +int odp_ml_compl_load_result(odp_ml_compl_t ml_compl, odp_ml_load_result_t *result) +{ + odp_event_subtype_t subtype; + odp_ml_load_result_t *load_result; + odp_buffer_t buf = (odp_buffer_t)(uintptr_t)ml_compl; + odp_event_t ev = odp_buffer_to_event(buf); + + if (odp_unlikely(ml_compl == ODP_ML_COMPL_INVALID)) { + _ODP_ERR("Given ML completion event is invalid\n"); + return -2; + } + + if (odp_event_types(ev, &subtype) != ODP_EVENT_ML_COMPL || + subtype != ODP_EVENT_ML_COMPL_LOAD) { + _ODP_ERR("Given completion event has wrong event type or subtype\n"); + return -2; + } + + load_result = odp_buffer_addr(buf); + if (result) + *result = *load_result; + + return load_result->error_code ? -1 : 0; +} + +void *odp_ml_compl_user_area(odp_ml_compl_t ml_compl) +{ + return odp_buffer_user_area((odp_buffer_t)(uintptr_t)ml_compl); +} + +odp_ml_compl_t odp_ml_compl_from_event(odp_event_t event) +{ + _ODP_ASSERT(_odp_event_hdr_field(event, int8_t, event_type) == ODP_EVENT_ML_COMPL); + + return (odp_ml_compl_t)(uintptr_t)event; +} + +odp_event_t odp_ml_compl_to_event(odp_ml_compl_t ml_compl) +{ + return (odp_event_t)(uintptr_t)ml_compl; +} + +uint64_t odp_ml_compl_to_u64(odp_ml_compl_t ml_compl) +{ + return (uint64_t)(uintptr_t)ml_compl; +} + +void odp_ml_compl_param_init(odp_ml_compl_param_t *compl_param) +{ + memset(compl_param, 0, sizeof(odp_ml_compl_param_t)); + + compl_param->queue = ODP_QUEUE_INVALID; + compl_param->event = ODP_EVENT_INVALID; +} + +int odp_ml_model_load(odp_ml_model_t model, odp_ml_load_result_t *result) +{ + odp_ml_load_result_t result_local; + int ret = -1; + ml_model_t *mdl = ml_model_from_handle(model); + + memset(&result_local, 0, sizeof(result_local)); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + result_local.error_code = ML_BAD_HDL; + goto load_fail; + } + + odp_ticketlock_lock(&mdl->lock); + if (odp_unlikely(mdl->state != ML_STATE_CREATED)) { + _ODP_ERR("Model has not been created yet or is already loaded\n"); + odp_ticketlock_unlock(&mdl->lock); + result_local.error_code = ML_NOT_CREATED; + goto load_fail; + } + + mdl->state = ML_STATE_LOADED; + odp_ticketlock_unlock(&mdl->lock); + ret = 0; + +load_fail: + if (result) + *result = result_local; + + return ret; +} + +static inline int check_compl_param(const odp_ml_compl_param_t *compl_param, + uint32_t max_compl_id, odp_bool_t is_load) +{ + odp_ml_config_t *config = &_odp_ml_glb->ml_config; + + switch (compl_param->mode) { + case ODP_ML_COMPL_MODE_POLL: + if (is_load && !(config->load_mode_mask & ODP_ML_COMPL_MODE_POLL)) { + _ODP_ERR("Poll mode loading/unloading is not configured\n"); + return -1; + } + + if (!is_load && !(config->run_mode_mask & ODP_ML_COMPL_MODE_POLL)) { + _ODP_ERR("Poll mode run is not configured\n"); + return -1; + } + + if (compl_param->compl_id > max_compl_id) { + _ODP_ERR("Bad compl_id: %u, exceeding model max completion id %u\n", + compl_param->compl_id, max_compl_id); + return -1; + } + break; + case ODP_ML_COMPL_MODE_EVENT: + if (is_load && !(config->load_mode_mask & ODP_ML_COMPL_MODE_EVENT)) { + _ODP_ERR("Event mode loading/unloading is not configured\n"); + return -1; + } + + if (!is_load && !(config->run_mode_mask & ODP_ML_COMPL_MODE_EVENT)) { + _ODP_ERR("Event mode run is not configured\n"); + return -1; + } + + if (compl_param->event == ODP_EVENT_INVALID || + compl_param->queue == ODP_QUEUE_INVALID) { + _ODP_ERR("Bad event or queue\n"); + return -1; + } + + if (odp_event_type(compl_param->event) != ODP_EVENT_ML_COMPL) { + _ODP_ERR("Bad completion event type\n"); + return -1; + } + break; + default: + /* Including ODP_ML_COMPL_MODE_SYNC, which is not supported by + * asynchrous functions (e.g. *_start()) either. + */ + _ODP_ERR("Invalid completion mode %u\n", compl_param->mode); + return -1; + } + + return 0; +} + +int odp_ml_model_load_start(odp_ml_model_t model, const odp_ml_compl_param_t *compl_param) +{ + int ret; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad model handle\n"); + return -1; + } + + if (odp_unlikely(check_compl_param(compl_param, mdl->max_compl_id, true))) + return -1; + + if (compl_param->mode == ODP_ML_COMPL_MODE_POLL) + odp_atomic_store_rel_u32(&mdl->compl_status[compl_param->compl_id], 0); + + ret = odp_ml_model_load(model, NULL); + + if (odp_unlikely(ret)) + return -1; + + /* Send a completion event to the given queue */ + if (compl_param->mode == ODP_ML_COMPL_MODE_EVENT) { + odp_ml_load_result_t *result; + odp_buffer_t buf = (odp_buffer_t)(uintptr_t)compl_param->event; + + _odp_buffer_subtype_set(buf, ODP_EVENT_ML_COMPL_LOAD); + + result = odp_buffer_addr(buf); + result->error_code = 0; + result->user_ptr = compl_param->user_ptr; + + if (odp_unlikely(odp_queue_enq(compl_param->queue, compl_param->event))) { + _ODP_ERR("Completion event enqueue failed %" PRIu64 "\n", + odp_queue_to_u64(compl_param->queue)); + if (odp_ml_model_unload(model, NULL)) + _ODP_ERR("Failed to unload model\n"); + return -1; + } + + return 0; + } + + mdl->result[compl_param->compl_id].user_ptr = compl_param->user_ptr; + odp_atomic_store_rel_u32(&mdl->compl_status[compl_param->compl_id], 1); + return 0; +} + +int odp_ml_model_load_status(odp_ml_model_t model, uint32_t compl_id, odp_ml_load_result_t *result) +{ + int ret; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID || compl_id > mdl->max_compl_id)) { + _ODP_ERR("Invalid model or compl_id: %u\n", compl_id); + return -2; + } + + ret = odp_atomic_load_acq_u32(&mdl->compl_status[compl_id]); + + if (ret && result) { + result->error_code = 0; + result->user_ptr = mdl->result[compl_id].user_ptr; + } + + return ret; +} + +int odp_ml_model_unload(odp_ml_model_t model, odp_ml_load_result_t *result) +{ + odp_ml_load_result_t result_local; + int ret = -1; + ml_model_t *mdl = ml_model_from_handle(model); + + memset(&result_local, 0, sizeof(result_local)); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + result_local.error_code = ML_BAD_HDL; + _ODP_ERR("Bad ML model handle\n"); + goto unload_fail; + } + + odp_ticketlock_lock(&mdl->lock); + /* mdl->state == ML_STATE_FREE, ML_STATE_CREATED, ML_STATE_INFERENCING */ + if (odp_unlikely(mdl->state != ML_STATE_LOADED)) { + _ODP_ERR("Model has not been created/loaded or inferencing has not finished yet\n"); + odp_ticketlock_unlock(&mdl->lock); + result_local.error_code = ML_NOT_LOADED; + goto unload_fail; + } + + mdl->state = ML_STATE_CREATED; + odp_ticketlock_unlock(&mdl->lock); + + ret = 0; + +unload_fail: + if (result) + *result = result_local; + + return ret; +} + +int odp_ml_model_unload_start(odp_ml_model_t model, const odp_ml_compl_param_t *compl_param) +{ + int ret; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad model handle\n"); + return -1; + } + + if (odp_unlikely(check_compl_param(compl_param, mdl->max_compl_id, true))) + return -1; + + if (compl_param->mode == ODP_ML_COMPL_MODE_POLL) + odp_atomic_store_rel_u32(&mdl->compl_status[compl_param->compl_id], 0); + + ret = odp_ml_model_unload(model, NULL); + + if (odp_unlikely(ret)) + return -1; + + /* Upon successful unloading, send a completion event to the given queue */ + if (compl_param->mode == ODP_ML_COMPL_MODE_EVENT) { + odp_ml_load_result_t *result; + odp_buffer_t buf = (odp_buffer_t)(uintptr_t)compl_param->event; + + _odp_buffer_subtype_set(buf, ODP_EVENT_ML_COMPL_LOAD); + + result = odp_buffer_addr(buf); + result->error_code = 0; + result->user_ptr = compl_param->user_ptr; + + if (odp_unlikely(odp_queue_enq(compl_param->queue, compl_param->event))) { + _ODP_ERR("Completion event enqueue failed %" PRIu64 "\n", + odp_queue_to_u64(compl_param->queue)); + return -1; + } + + return 0; + } + + mdl->result[compl_param->compl_id].user_ptr = compl_param->user_ptr; + odp_atomic_store_rel_u32(&mdl->compl_status[compl_param->compl_id], 1); + return 0; +} + +int odp_ml_model_unload_status(odp_ml_model_t model, uint32_t compl_id, + odp_ml_load_result_t *result) +{ + return odp_ml_model_load_status(model, compl_id, result); +} + +void odp_ml_run_param_init(odp_ml_run_param_t *param) +{ + memset(param, 0, sizeof(odp_ml_run_param_t)); +} + +static void ml_shape_to_int64(const odp_ml_shape_info_t *shape, uint32_t batch_size, int64_t *array) +{ + for (uint32_t i = 0; i < shape->num_dim; i++) { + /* Replace dynamic dimension size with provided batch_size */ + if (shape->dim[i] == ODP_ML_DIM_DYNAMIC) + array[i] = batch_size; + else + array[i] = shape->dim[i]; + } +} + +/* Get the number of elements in given shape */ +static inline uint64_t get_num_elem(uint32_t batch_size, const odp_ml_shape_info_t *shape) +{ + uint64_t num_elements = 1; + int64_t dim[ODP_ML_MAX_DIMS] = {0}; + + ml_shape_to_int64(shape, batch_size, dim); + + for (uint32_t i = 0; i < shape->num_dim; i++) + num_elements *= (uint64_t)dim[i]; + + return num_elements; +} + +static inline uint32_t dyn_io_size(const odp_ml_shape_info_t *shape, uint32_t data_type_size, + const odp_ml_run_param_t *param) +{ + uint32_t size; + + if (!param || !param->batch_size) { + _ODP_ERR("Parameter 'param' must not be NULL and batch_size must be " + "provided when a input/output has dynamic dimension size\n"); + return 0; + } + + size = get_num_elem(param->batch_size, shape); + size *= data_type_size; + + return size; +} + +static int verify_run_params(odp_ml_model_t model, const odp_ml_data_t *data, + const odp_ml_run_param_t *param) +{ + const ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad ML model handle\n"); + return -1; + } + + if (odp_unlikely(!data)) { + _ODP_ERR("Parameter 'data' must not be NULL\n"); + return -1; + } + + /* Make sure that the number of input data segments equals or bigger than + * the number of model inputs. */ + if (mdl->info.num_inputs > data->num_input_seg) { + _ODP_ERR("The num of input data segments %u must not less than " + "the number of model inputs %u\n", data->num_input_seg, + mdl->info.num_inputs); + return -1; + } + + if (mdl->info.num_outputs > data->num_output_seg) { + _ODP_ERR("The num of output data segments %u must not less than " + "the number of model outputs %u\n", data->num_output_seg, + mdl->info.num_outputs); + return -1; + } + + if (data->num_input_seg > mdl->info.num_inputs && + (_odp_ml_glb->capa.max_segs_per_input == 1)) { + _ODP_ERR("Segmented input data is not supported\n"); + return -1; + } + + if (data->num_output_seg > mdl->info.num_outputs && + (_odp_ml_glb->capa.max_segs_per_output == 1)) { + _ODP_ERR("Segmented output data is not supported"); + return -1; + } + + uint32_t size = 0; + uint32_t input_index = 0; + uint32_t seg_size_sum = 0; + odp_bool_t index_new = true; + uint32_t segs_per_input = 1; + + for (uint32_t i = 0; i < data->num_input_seg; i++) { + if (data->input_seg[i].addr == NULL) { + _ODP_ERR("data->input_seg[%u].addr must not NULL\n", i); + return -1; + }; + + if (index_new) { + if (input_index > mdl->info.num_inputs - 1) { + _ODP_ERR("Too much number of input segments given\n"); + return -1; + } + + /* Input with dynamic batch size */ + if (mdl->input_info[input_index].shape.type == ODP_ML_SHAPE_BATCH) + size = dyn_io_size(&mdl->input_info[input_index].shape, + mdl->input_info[input_index].data_type_size, + param); + else + size = mdl->input_sizes[input_index]; + + if (!size) { + _ODP_ERR("Size for %uth input is 0\n", input_index); + return -1; + } + } + + seg_size_sum += data->input_seg[i].size; + + if (seg_size_sum > size) { + _ODP_ERR("Sum of segment sizes %u exceeds %uth input data size %u\n", + seg_size_sum, input_index, size); + return -1; + } + + if (seg_size_sum == size) { + if (segs_per_input > _odp_ml_glb->capa.max_segs_per_input) { + _ODP_ERR("Number of segments %u for input[%u] exceeds maximum" + " number of data segments per model input %u\n", + segs_per_input, input_index, + _odp_ml_glb->capa.max_segs_per_input); + return -1; + } + input_index++; + index_new = true; + seg_size_sum = 0; + segs_per_input = 1; + } else { + segs_per_input++; + index_new = false; + } + } + + if (input_index != mdl->info.num_inputs) { + _ODP_ERR("Data is not provided for all model inputs\n"); + return -1; + } + + seg_size_sum = 0; + index_new = true; + uint32_t output_index = 0; + uint32_t segs_per_output = 1; + + for (uint32_t i = 0; i < data->num_output_seg; i++) { + if (data->output_seg[i].addr == NULL) { + _ODP_ERR("data->output_seg[%u].addr must not NULL\n", i); + return -1; + } + + if (index_new) { + if (output_index > mdl->info.num_outputs - 1) { + _ODP_ERR("Too much number of output segments given\n"); + return -1; + } + + /* Output with dynamic batch size */ + if (mdl->output_info[output_index].shape.type == ODP_ML_SHAPE_BATCH) + size = dyn_io_size(&mdl->output_info[output_index].shape, + mdl->output_info[output_index].data_type_size, + param); + else + size = mdl->output_sizes[output_index]; + + if (!size) { + _ODP_ERR("Size for %uth output is 0\n", output_index); + return -1; + } + } + + seg_size_sum += data->output_seg[i].size; + + if (seg_size_sum > size) { + _ODP_ERR("Sum of segment sizes %u exceeds %uth output data size %u\n", + seg_size_sum, output_index, size); + return -1; + } + + if (seg_size_sum >= size) { + if (segs_per_output > _odp_ml_glb->capa.max_segs_per_output) { + _ODP_ERR("Number of segments %u for output[%u] exceeds maximum" + " number of data segments per model output %u\n", + segs_per_output, output_index, + _odp_ml_glb->capa.max_segs_per_output); + return -1; + } + output_index++; + index_new = true; + seg_size_sum = 0; + segs_per_output = 1; + } else { + segs_per_output++; + index_new = false; + } + } + + if (output_index != mdl->info.num_outputs) { + _ODP_ERR("Not enough output_segs to hold all output data\n"); + return -1; + } + + return 0; +} + +static ONNXTensorElementDataType onnx_dtype_from_odp_dtype(odp_ml_data_type_t data_type) +{ + switch (data_type) { + case ODP_ML_DATA_TYPE_NONE: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED; + case ODP_ML_DATA_TYPE_INT8: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8; + case ODP_ML_DATA_TYPE_UINT8: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8; + case ODP_ML_DATA_TYPE_INT16: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16; + case ODP_ML_DATA_TYPE_UINT16: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16; + case ODP_ML_DATA_TYPE_INT24: + /* Fall through*/ + case ODP_ML_DATA_TYPE_UINT24: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED; + case ODP_ML_DATA_TYPE_FP64: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE; + case ODP_ML_DATA_TYPE_INT32: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32; + case ODP_ML_DATA_TYPE_UINT32: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32; + case ODP_ML_DATA_TYPE_INT64: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64; + case ODP_ML_DATA_TYPE_UINT64: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64; + case ODP_ML_DATA_TYPE_FP16: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16; + case ODP_ML_DATA_TYPE_FP32: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT; + case ODP_ML_DATA_TYPE_BFP16: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16; + default: + return ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED; + } +} + +static int verify_tensor(const OrtValue *tensor, odp_ml_data_type_t expected_type, + const odp_ml_shape_info_t *expected_shape, uint32_t batch_size) +{ + OrtTensorTypeAndShapeInfo *tensor_info; + ONNXTensorElementDataType tensor_type; + size_t dim_count; + OrtStatus *status = NULL; + int64_t dims[ODP_ML_MAX_DIMS] = {0}; + int64_t shape_arr[ODP_ML_MAX_DIMS] = {0}; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + status = ort_api->GetTensorTypeAndShape(tensor, &tensor_info); + if (check_ortstatus(status)) { + _ODP_ERR("GetTensorTypeAndShape() failed\n"); + return -1; + } + + status = ort_api->GetTensorElementType(tensor_info, &tensor_type); + if (check_ortstatus(status)) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("GetTensorElementType() failed\n"); + return -1; + } + + if (onnx_dtype_to_odp_dtype(tensor_type) != expected_type) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("Tensor type does not match model type\n"); + return -1; + } + + status = ort_api->GetDimensionsCount(tensor_info, &dim_count); + if (check_ortstatus(status)) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("GetDimensionsCount() failed\n"); + return -1; + } + + if (dim_count != expected_shape->num_dim) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("Tensor dimension does not match shape_dim\n"); + return -1; + } + + status = ort_api->GetDimensions(tensor_info, dims, dim_count); + if (check_ortstatus(status)) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("GetDimensions() failed\n"); + return -1; + } + + ml_shape_to_int64(expected_shape, batch_size, shape_arr); + + for (uint32_t i = 0; i < dim_count; i++) { + if (dims[i] != shape_arr[i]) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("Shape[%u]: %" PRIu64 " does not match expected: %" PRIu64 "\n", + i, dims[i], shape_arr[i]); + return -1; + } + } + + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + return 0; +} + +static int input_data_to_tensor(const odp_ml_input_info_t *input_info, uint32_t num_seg, + const odp_ml_data_seg_t *input_seg, uint32_t *seg_idx, + uint32_t batch_size, OrtValue **input_tensor) +{ + int is_tensor; + uint64_t input_size; + OrtAllocator *allocator; + void *data = NULL; + OrtStatus *status = NULL; + int64_t shape[ODP_ML_MAX_DIMS] = {0}; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + ONNXTensorElementDataType onnx_dtype = ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED; + + ml_shape_to_int64(&input_info->shape, batch_size, shape); + + onnx_dtype = onnx_dtype_from_odp_dtype(input_info->data_type); + _ODP_ASSERT(onnx_dtype != ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED); + + status = ort_api->GetAllocatorWithDefaultOptions(&allocator); + if (check_ortstatus(status)) { + _ODP_ERR("GetAllocatorWithDefaultOptions() failed\n"); + return -1; + } + + status = ort_api->CreateTensorAsOrtValue(allocator, + shape, + input_info->shape.num_dim, + onnx_dtype, + input_tensor); + if (check_ortstatus(status) || !input_tensor[0]) { + _ODP_ERR("CreateTensorWithDataAsOrtValue() failed\n"); + return -1; + } + + input_size = input_info->data_type_size * get_num_elem(batch_size, &input_info->shape); + + status = ort_api->GetTensorMutableData(input_tensor[0], &data); + if (check_ortstatus(status) || !data) { + _ODP_ERR("GetTensorMutableData() failed\n"); + return -1; + } + + for (uint64_t i = 0; i < input_size; ) { + if (*seg_idx >= num_seg) { + _ODP_ERR("Insufficient input data\n"); + return -1; + } + + uint64_t seg_size = input_seg[*seg_idx].size; + + if (i + seg_size > input_size) { + _ODP_ERR("Excess input data in segment %" PRIu32 "\n", *seg_idx); + return -1; + } + + memcpy((uint8_t *)data + i, input_seg[(*seg_idx)++].addr, seg_size); + i += seg_size; + } + + if (!ODP_DEBUG) + return 0; + + status = ort_api->IsTensor(input_tensor[0], &is_tensor); + if (check_ortstatus(status) || !is_tensor) { + _ODP_ERR("input_tensor IsTensor failed\n"); + return -1; + } + + /* Make sure tensor shape matches input_shape */ + if (verify_tensor(input_tensor[0], input_info->data_type, + &input_info->shape, batch_size)) { + _ODP_ERR("Verify input_tensor failed\n"); + return -1; + } + + return 0; +} + +static int verify_output_tensor(OrtValue *output_tensor, odp_ml_data_type_t expected_type, + const odp_ml_shape_info_t *expected_shape, uint32_t batch_size) +{ + int is_tensor = 0; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + OrtStatus *status = ort_api->IsTensor(output_tensor, &is_tensor); + + if (check_ortstatus(status) || !is_tensor) { + _ODP_ERR("output_tensor IsTensor failed\n"); + return -1; + } + + /* Make sure tensor shape matches output_shape */ + if (verify_tensor(output_tensor, expected_type, expected_shape, batch_size)) { + _ODP_ERR("Verify output_tensor failed\n"); + return -1; + } + + return 0; +} + +static int get_tensor_data_size(OrtValue *tensor, uint32_t *size, uint32_t data_type_size) +{ + size_t num_elem; + OrtStatus *status; + OrtTensorTypeAndShapeInfo *tensor_info; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + + status = ort_api->GetTensorTypeAndShape(tensor, &tensor_info); + if (check_ortstatus(status)) { + _ODP_ERR("GetTensorTypeAndShape() failed\n"); + return -1; + } + + status = ort_api->GetTensorShapeElementCount(tensor_info, &num_elem); + if (check_ortstatus(status)) { + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + _ODP_ERR("GetTensorShapeElementCount() failed\n"); + return -1; + } + *size = data_type_size * num_elem; + + ort_api->ReleaseTensorTypeAndShapeInfo(tensor_info); + return 0; +} + +static int check_output_size(odp_bool_t is_segmented, uint32_t output_idx, uint32_t seg_idx, + uint64_t out_tensor_data_size, const odp_ml_data_t data[]) +{ + uint64_t output_size = 0; + + /* Output is not segmented */ + if (!is_segmented) { + /* Make sure tensor data size does not exceed size allocated for + * data->output_seg[seg_idx].addr */ + if (out_tensor_data_size > data->output_seg[seg_idx].size) { + _ODP_ERR("Malloc at least %" PRIu64 " bytes for %dth output tensor\n", + out_tensor_data_size, output_idx); + return -1; + } + + return 0; + } + + /* Output is segmented, first calculate total size for one tensor */ + for (; seg_idx < data->num_output_seg; seg_idx++) { + output_size += data->output_seg[seg_idx].size; + if (output_size >= out_tensor_data_size) + break; + } + + if (0 == output_size) { + _ODP_ERR("No output data segments for %uth output tensor\n", output_idx); + return -1; + } + + if (out_tensor_data_size > output_size) { + _ODP_ERR("Output segments (%" PRIu64 " bytes in total) for %uth output" + " is expected to be at least %" PRIu64 " bytes\n", + output_size, output_idx, out_tensor_data_size); + return -1; + } + + return 0; +} + +static int output_tensors_to_data(OrtValue **output_tensors, + uint32_t model_num_outputs, + const odp_ml_run_param_t *param, + const odp_ml_output_info_t *output_info, + const odp_ml_data_t *data, + odp_ml_run_result_t *result_local) +{ + uint32_t seg_idx; + uint64_t seg_size; + uint64_t cpy_size; + uint64_t left_size; + uint64_t output_val_offset; + uint32_t out_tensor_data_size; + void *output_val = NULL; /* Pointer to store one raw output value */ + OrtStatus *status = NULL; + uint32_t batch_size = (param && param->batch_size) ? param->batch_size : 0; + const OrtApi *ort_api = _odp_ml_glb->ort_api; + odp_bool_t is_segmented = (data->num_output_seg != model_num_outputs); + + seg_idx = 0; + for (uint32_t i = 0; i < model_num_outputs; i++) { + if (ODP_DEBUG && + verify_output_tensor(output_tensors[i], output_info[i].data_type, + &output_info[i].shape, batch_size)){ + result_local->error_code = ML_BAD_OUTPUT; + return -1; + } + + /* Get tensor data size */ + if (get_tensor_data_size(output_tensors[i], &out_tensor_data_size, + output_info[i].data_type_size)) { + result_local->error_code = ML_LIB_FAILED; + return -1; + } + + /* When output_tensor is an empty tensor [], skip getting data */ + if (out_tensor_data_size == 0) + continue; + + if (ODP_DEBUG && check_output_size(is_segmented, i, seg_idx, + out_tensor_data_size, data)) { + result_local->error_code = ML_BAD_OUTPUT; + return -1; + } + + /* Following assumes param and data->output_seg are valid */ + /* Get tensor data */ + output_val = NULL; + status = ort_api->GetTensorMutableData(output_tensors[i], &output_val); + if (check_ortstatus(status) || !output_val) { + result_local->error_code = ML_LIB_FAILED; + return -1; + } + + /* Output is not segmented */ + if (!is_segmented) { + /* Store output data to data->output_seg[i].addr */ + memcpy(data->output_seg[i].addr, output_val, out_tensor_data_size); + seg_idx++; + continue; + } + + /* Output is segmented */ + output_val_offset = 0; + left_size = out_tensor_data_size; + for (; seg_idx < data->num_output_seg; seg_idx++) { + seg_size = data->output_seg[seg_idx].size; + cpy_size = left_size > seg_size ? seg_size : left_size; + memcpy(data->output_seg[seg_idx].addr, + ((char *)output_val) + output_val_offset, cpy_size); + + output_val_offset += cpy_size; + left_size = out_tensor_data_size - output_val_offset; + + if (!left_size) { + seg_idx++; + break; + } + } + } + + return 0; +} + +int odp_ml_run(odp_ml_model_t model, const odp_ml_data_t *data, const odp_ml_run_param_t *param) +{ + odp_ml_run_result_t result_local; + + int retval = -1; /* Return value of this function */ + int ret = 0; + OrtStatus *status = NULL; + uint32_t batch_size = 0; + + OrtValue *input_tensor[CONFIG_ML_MAX_INPUTS] = {0}; + OrtValue *output_tensors[CONFIG_ML_MAX_OUTPUTS] = {0}; + const char *input_names[CONFIG_ML_MAX_INPUTS] = {0}; + const char *output_names[CONFIG_ML_MAX_OUTPUTS] = {0}; + + const OrtApi *ort_api = _odp_ml_glb->ort_api; + ml_model_t *mdl = ml_model_from_handle(model); + const odp_ml_model_info_t *ml_info = &mdl->info; + const odp_ml_input_info_t *input_info = mdl->input_info; + const odp_ml_output_info_t *output_info = mdl->output_info; + OrtSession *session = mdl->session; + + odp_ticketlock_lock(&mdl->lock); + if (odp_unlikely(mdl->state == ML_STATE_INFERENCING)) { + odp_ticketlock_unlock(&mdl->lock); + return 0; + } + if (odp_unlikely(mdl->state != ML_STATE_LOADED)) { + _ODP_ERR("Wrong model state: not created or not loaded\n"); + odp_ticketlock_unlock(&mdl->lock); + return -1; + } + mdl->state = ML_STATE_INFERENCING; + odp_ticketlock_unlock(&mdl->lock); + + memset(&result_local, 0, sizeof(result_local)); + + if (ODP_DEBUG && verify_run_params(model, data, param)) { + result_local.error_code = ML_BAD_INPUT; + goto init_fail; + } + + if (param && param->batch_size) + batch_size = param->batch_size; + + uint32_t seg_idx = 0; + + /* Transfer input data to tensor */ + for (uint32_t i = 0; i < ml_info->num_inputs; i++) { + ret = input_data_to_tensor(&input_info[i], + data->num_input_seg, + data->input_seg, + &seg_idx, + batch_size, + &input_tensor[i]); + if (ret) { + _ODP_ERR("%uth input data to tensor failed\n", i); + result_local.error_code = ML_LIB_FAILED; + goto release_input_tensors; + } + + _ODP_DBG("input_tensor[%u]: %p\n", i, input_tensor[i]); + + /* Model input names */ + input_names[i] = input_info[i].name; + } + + if (seg_idx < data->num_input_seg) { + _ODP_ERR("Excess input segments\n"); + ret = -1; + } + + for (uint32_t i = 0; i < ml_info->num_outputs; i++) + output_names[i] = output_info[i].name; + + /* Run inference */ + status = ort_api->Run(session, + NULL, + (const char * const *)input_names, + (const OrtValue * const*)input_tensor, + ml_info->num_inputs, + (const char * const *)output_names, + ml_info->num_outputs, + output_tensors); + + if (check_ortstatus(status)) { + _ODP_ERR("Run inference failed\n"); + result_local.error_code = ML_LIB_FAILED; + goto release_all_tensors; + } + + /* Verify output tensors and store them to output */ + if (output_tensors_to_data(output_tensors, ml_info->num_outputs, param, + output_info, data, &result_local)) { + _ODP_ERR("Output tensors to data failed\n"); + goto release_all_tensors; + } + + retval = 1; + +release_all_tensors: + for (uint32_t i = 0; i < ml_info->num_outputs; i++) + ort_api->ReleaseValue(output_tensors[i]); + +release_input_tensors: + for (uint32_t i = 0; i < ml_info->num_inputs; i++) + ort_api->ReleaseValue(input_tensor[i]); + +init_fail: + if (param && param->result) + *param->result = result_local; + + odp_ticketlock_lock(&mdl->lock); + mdl->state = ML_STATE_LOADED; + odp_ticketlock_unlock(&mdl->lock); + + return retval; +} + +int odp_ml_run_multi(odp_ml_model_t model, const odp_ml_data_t data[], + const odp_ml_run_param_t param[], int num) +{ + int i; + int ret; + + if (odp_unlikely(num < 1)) { + _ODP_ERR("Bad number of runs\n"); + return -1; + } + + for (i = 0; i < num; i++) { + if (param) + ret = odp_ml_run(model, &data[i], ¶m[i]); + else + ret = odp_ml_run(model, &data[i], NULL); + + if (odp_unlikely(ret != 1)) + break; + } + + if (odp_unlikely(i == 0)) + return ret; + + return i; +} + +int odp_ml_run_start(odp_ml_model_t model, const odp_ml_data_t *data, + const odp_ml_compl_param_t *compl_param, + const odp_ml_run_param_t *run_param) +{ + int ret; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID)) { + _ODP_ERR("Bad model handle\n"); + return -1; + } + + if (odp_unlikely(!compl_param)) { + _ODP_ERR("Completion parameter is NULL\n"); + return -1; + } + + /* Check completion mode */ + if (odp_unlikely(check_compl_param(compl_param, mdl->max_compl_id, false))) { + _ODP_ERR("Bad ML job completion parameter\n"); + return -1; + } + + if (compl_param->mode == ODP_ML_COMPL_MODE_POLL) + odp_atomic_store_rel_u32(&mdl->compl_status[compl_param->compl_id], 0); + + ret = odp_ml_run(model, data, run_param); + + if (odp_unlikely(ret < 1)) + return ret; + + /* Send a completion event to the given queue */ + if (compl_param->mode == ODP_ML_COMPL_MODE_EVENT) { + odp_ml_run_result_t *result; + odp_buffer_t buf = (odp_buffer_t)(uintptr_t)compl_param->event; + + _odp_buffer_subtype_set(buf, ODP_EVENT_ML_COMPL_RUN); + + result = odp_buffer_addr(buf); + result->error_code = 0; + result->user_ptr = compl_param->user_ptr; + + if (odp_unlikely(odp_queue_enq(compl_param->queue, compl_param->event))) { + _ODP_ERR("Completion event enqueue failed %" PRIu64 "\n", + odp_queue_to_u64(compl_param->queue)); + return -1; + } + + return 1; + } + + /* compl_param->mode == ODP_ML_COMPL_MODE_POLL */ + mdl->result[compl_param->compl_id].user_ptr = compl_param->user_ptr; + odp_atomic_store_rel_u32(&mdl->compl_status[compl_param->compl_id], 1); + + return 1; +} + +int odp_ml_run_start_multi(odp_ml_model_t model, const odp_ml_data_t data[], + const odp_ml_compl_param_t compl_param[], + const odp_ml_run_param_t run_param[], int num) +{ + int i; + int ret = 0; + + if (odp_unlikely(num < 1)) { + _ODP_ERR("Bad number of runs\n"); + return -1; + } + + for (i = 0; i < num; i++) { + if (run_param) + ret = odp_ml_run_start(model, &data[i], &compl_param[i], &run_param[i]); + else + ret = odp_ml_run_start(model, &data[i], &compl_param[i], NULL); + + if (odp_unlikely(ret != 1)) + break; + } + + if (odp_unlikely(i == 0)) + return ret; + + return i; +} + +int odp_ml_run_status(odp_ml_model_t model, uint32_t compl_id, odp_ml_run_result_t *result) +{ + int ret; + ml_model_t *mdl = ml_model_from_handle(model); + + if (odp_unlikely(model == ODP_ML_MODEL_INVALID || + compl_id > mdl->max_compl_id)) { + _ODP_ERR("Invalid model handle or completion id: %u\n", compl_id); + return -2; + } + + ret = odp_atomic_load_acq_u32(&mdl->compl_status[compl_id]); + + if (result) { + result->error_code = 0; + result->user_ptr = mdl->result[compl_id].user_ptr; + } + + return ret; +} + +static int opt_level_from_str(const char *level_str, GraphOptimizationLevel *level) +{ + if (strcmp(level_str, "DISABLE_ALL") == 0) + *level = ORT_DISABLE_ALL; + else if (strcmp(level_str, "ENABLE_BASIC") == 0) + *level = ORT_ENABLE_BASIC; + else if (strcmp(level_str, "ENABLE_EXTENDED") == 0) + *level = ORT_ENABLE_EXTENDED; + else if (strcmp(level_str, "ENABLE_ALL") == 0) + *level = ORT_ENABLE_ALL; + else + return -1; + + return 0; +} + +static int execution_mode_from_str(const char *mode_str, ExecutionMode *mode) +{ + if (strcmp(mode_str, "SEQUENTIAL") == 0) + *mode = ORT_SEQUENTIAL; + else if (strcmp(mode_str, "PARALLEL") == 0) + *mode = ORT_PARALLEL; + else + return -1; + + return 0; +} + +static int read_config_file(ort_run_opts_t *opts) +{ + const char *conf_str; + char mode_str[ML_MAX_CONFIG_STR_LEN]; + char opt_level_str[ML_MAX_CONFIG_STR_LEN]; + + _ODP_PRINT("ML config:\n"); + + conf_str = "ml.enable_profiling"; + if (!_odp_libconfig_lookup_int(conf_str, &opts->enable_profiling)) { + _ODP_ERR("Config option '%s' not found.\n", conf_str); + return -1; + } + _ODP_PRINT(" %s: %i\n", conf_str, opts->enable_profiling); + + conf_str = "ml.execution_mode"; + if (_odp_libconfig_lookup_str(conf_str, mode_str, ML_MAX_CONFIG_STR_LEN - 1) < 0) { + _ODP_ERR("Config option '%s' not found.\n", conf_str); + return -1; + } + + if (execution_mode_from_str(mode_str, &opts->execution_mode)) { + _ODP_ERR("Unsupported execution mode: %s\n", mode_str); + return -1; + } + _ODP_PRINT(" %s: %s\n", conf_str, mode_str); + + conf_str = "ml.inter_op_num_threads"; + if (!_odp_libconfig_lookup_int(conf_str, &opts->inter_op_num_threads)) { + _ODP_ERR("Config option '%s' not found.\n", conf_str); + return -1; + } + _ODP_PRINT(" %s: %i\n", conf_str, opts->inter_op_num_threads); + + conf_str = "ml.intra_op_num_threads"; + if (!_odp_libconfig_lookup_int(conf_str, &opts->intra_op_num_threads)) { + _ODP_ERR("Config option '%s' not found.\n", conf_str); + return -1; + } + _ODP_PRINT(" %s: %i\n", conf_str, opts->intra_op_num_threads); + + conf_str = "ml.graph_optimization_level"; + if (_odp_libconfig_lookup_str(conf_str, opt_level_str, + ML_MAX_CONFIG_STR_LEN - 1) < 0) { + _ODP_ERR("Config option '%s' not found.\n", conf_str); + return -1; + } + + if (opt_level_from_str(opt_level_str, &opts->graph_opt_level)) { + _ODP_ERR("Graph optimize level %s not supported\n", opt_level_str); + return -1; + } + _ODP_PRINT(" %s: %s\n", conf_str, opt_level_str); + + conf_str = "ml.optimized_model_filepath"; + if (_odp_libconfig_lookup_str(conf_str, opts->opt_model_filepath, + ML_MAX_CONFIG_STR_LEN - 1) < 0) { + _ODP_ERR("Config option '%s' not found.\n", conf_str); + return -1; + } + _ODP_PRINT(" %s: %s\n", conf_str, opts->opt_model_filepath); + + return 0; +} + +int _odp_ml_init_global(void) +{ + int i; + OrtEnv *env; + odp_shm_t shm; + OrtStatus *status; + const OrtApi *ort_api; + + if (odp_global_ro.disable.ml) { + _ODP_ERR("ML is disabled\n"); + return 0; + } + + shm = odp_shm_reserve("_odp_ml_global", sizeof(ml_global_t), ODP_CACHE_LINE_SIZE, 0); + _odp_ml_glb = odp_shm_addr(shm); + + if (_odp_ml_glb == NULL) { + _ODP_ERR("SHM reserve failed for odp_ml\n"); + return -1; + } + + memset(_odp_ml_glb, 0, sizeof(ml_global_t)); + _odp_ml_glb->shm = shm; + + if (odp_ml_capability(&_odp_ml_glb->capa)) { + _ODP_ERR("ML capability failed\n"); + return -1; + } + + odp_pool_param_init(&_odp_ml_glb->pool_param); + + if (read_config_file(&_odp_ml_glb->ort_run_opts)) + return -1; + + ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION); + if (!ort_api) { + _ODP_ERR("Failed to init ONNX Runtime engine.\n"); + return -1; + } + _odp_ml_glb->ort_api = ort_api; + + status = ort_api->CreateEnv(ORT_LOGGING_LEVEL_WARNING, "Default", &env); + if (check_ortstatus(status) || !env) { + _ODP_ERR("ort_api->CreateEnv() failed.\n"); + return -1; + } + _odp_ml_glb->env = env; + + for (i = 0; i < ML_MAX_MODELS_CREATED; i++) + odp_ticketlock_init(&_odp_ml_glb->models[i].lock); + + return 0; +} + +int _odp_ml_term_global(void) +{ + if (odp_global_ro.disable.ml) + return 0; + + if (_odp_ml_glb == NULL) + return 0; + + if (_odp_ml_glb->env) + _odp_ml_glb->ort_api->ReleaseEnv(_odp_ml_glb->env); + + if (odp_shm_free(_odp_ml_glb->shm)) { + _ODP_ERR("Shm free failed for odp_ml\n"); + return -1; + } + + return 0; +} |