API Description

Machine Learning (ML) offload.

ML API call sequence

Before ML offload can be used, it must be configured with an odp_ml_config() call. An application fills in configuration parameters to describe its intended ML offload usage. The parameter values may help ODP implementation to optimize memory and other HW resource usage. The application may use odp_ml_capability() to check ML capabilities both before and after the configuration step.

After configuration, an ML model binary is passed with other parameters to odp_ml_model_create() call which checks and prepares the model for usage. The application may use odp_ml_model_info(), odp_ml_model_input_info() and odp_ml_model_output_info() calls to check model input and output data formats. Before the application can use the model for inference, it loads the model with an odp_ml_model_load() or odp_ml_model_load_start() call. After a successful load, the application may use e.g. odp_ml_run() or odp_ml_run_start() to perform inferences.

When all previously started inference operations are complete, application uses odp_ml_model_unload() or odp_ml_model_unload_start() to unload the model. After a successful unload, the model may be destroyed with an odp_ml_model_destroy() call, or loaded again.

Completion identifiers

Completion identifiers are used with ML operations in asynchronous poll mode (ODP_ML_COMPL_MODE_POLL). Application declares the maximum identifier value it will use per model with odp_ml_model_param_t.max_compl_id parameter. It cannot exceed the implementation capability of odp_ml_capability_t.max_compl_id. Completion identifier values are model specific. The same value can be used simultaneously with two different models, but cannot be used simultaneously in two ML operations on the same model. A value may be reused for the next ML operation (on the same model) only after the previous operation is complete. Within those limitations, application may use/reuse completion identifier values from 0 to max_compl_id range freely.

Data Structures
struct	odp_ml_compl_pool_capability_t
	ML completion event pool capabilities. More...

struct	odp_ml_compl_pool_param_t
	ML completion event pool parameters. More...

struct	odp_ml_capability_t
	Machine learning capabilities. More...

struct	odp_ml_config_t
	Machine learning configuration parameters. More...

struct	odp_ml_shape_info_t
	Model input / output data shape information. More...

struct	odp_ml_input_info_t
	Model input information. More...

struct	odp_ml_output_info_t
	Model output information. More...

struct	odp_ml_model_info_t
	Model information. More...

struct	odp_ml_data_format_t
	Model input / output data format. More...

struct	odp_ml_model_param_t
	Machine learning model parameters. More...

struct	odp_ml_run_result_t
	Results of model run operation. More...

struct	odp_ml_load_result_t
	Result of model load / unload operation. More...

struct	odp_ml_compl_param_t
	ML completion parameters. More...

struct	odp_ml_data_seg_t
	Model input / output data segment. More...

struct	odp_ml_data_t
	Model input / output data for a model inference run. More...

struct	odp_ml_run_param_t
	Parameters for model run. More...

struct	odp_ml_extra_stat_info_t
	ML extra statistics counter information. More...

Macros
#define	ODP_ML_MODEL_INVALID ((odp_ml_model_t)0)
	Invalid ML model.

#define	ODP_ML_COMPL_INVALID ((odp_ml_compl_t)0)
	Invalid ML completion event.

#define	ODP_ML_MODEL_NAME_LEN 64
	Maximum length of model name, including the null character.

#define	ODP_ML_MODEL_IO_NAME_LEN 64
	Maximum length of model input/output name, including the null character.

#define	ODP_ML_SHAPE_NAME_LEN 16
	Maximum length of data dimension name, including the null character.

#define	ODP_ML_EXTRA_STAT_NAME_LEN 64
	Maximum length of extra statistics counter name, including the null character.

#define	ODP_ML_MAX_DIMS 8
	Maximum number of dimensions in input / output data shape.

#define	ODP_ML_DIM_DYNAMIC 0
	Dimension size is dynamic.

#define	ODP_ML_COMPL_MODE_SYNC 0x1u
	Synchronous operation.

#define	ODP_ML_COMPL_MODE_POLL 0x2u
	Asynchronous poll mode operation. More...

#define	ODP_ML_COMPL_MODE_EVENT 0x4u
	Asynchronous event mode operation. More...

Typedefs
typedef _odp_abi_ml_model_t *	odp_ml_model_t
	ODP ML model handle.

typedef _odp_abi_ml_compl_t *	odp_ml_compl_t
	ML completion event.

typedef struct _odp_ml_model_extra_param_t	odp_ml_model_extra_param_t
	ODP implementation specific extra parameters for model creation.

typedef uint32_t	odp_ml_compl_mode_t
	ML completion mode.

typedef struct odp_ml_compl_pool_capability_t	odp_ml_compl_pool_capability_t
	ML completion event pool capabilities. More...

typedef struct odp_ml_compl_pool_param_t	odp_ml_compl_pool_param_t
	ML completion event pool parameters. More...

typedef struct odp_ml_capability_t	odp_ml_capability_t
	Machine learning capabilities.

typedef struct odp_ml_config_t	odp_ml_config_t
	Machine learning configuration parameters.

typedef struct odp_ml_shape_info_t	odp_ml_shape_info_t
	Model input / output data shape information.

typedef struct odp_ml_input_info_t	odp_ml_input_info_t
	Model input information.

typedef struct odp_ml_output_info_t	odp_ml_output_info_t
	Model output information.

typedef struct odp_ml_model_info_t	odp_ml_model_info_t
	Model information.

typedef struct odp_ml_data_format_t	odp_ml_data_format_t
	Model input / output data format.

typedef struct odp_ml_model_param_t	odp_ml_model_param_t
	Machine learning model parameters. More...

typedef struct odp_ml_run_result_t	odp_ml_run_result_t
	Results of model run operation.

typedef struct odp_ml_load_result_t	odp_ml_load_result_t
	Result of model load / unload operation.

typedef struct odp_ml_compl_param_t	odp_ml_compl_param_t
	ML completion parameters. More...

typedef struct odp_ml_data_seg_t	odp_ml_data_seg_t
	Model input / output data segment.

typedef struct odp_ml_data_t	odp_ml_data_t
	Model input / output data for a model inference run.

typedef struct odp_ml_run_param_t	odp_ml_run_param_t
	Parameters for model run. More...

typedef struct odp_ml_extra_stat_info_t	odp_ml_extra_stat_info_t
	ML extra statistics counter information.

Enumerations
enum	odp_ml_data_type_t { ODP_ML_DATA_TYPE_NONE = 0 , ODP_ML_DATA_TYPE_INT8 , ODP_ML_DATA_TYPE_UINT8 , ODP_ML_DATA_TYPE_INT16 , ODP_ML_DATA_TYPE_UINT16 , ODP_ML_DATA_TYPE_INT24 , ODP_ML_DATA_TYPE_UINT24 , ODP_ML_DATA_TYPE_INT32 , ODP_ML_DATA_TYPE_UINT32 , ODP_ML_DATA_TYPE_INT64 , ODP_ML_DATA_TYPE_UINT64 , ODP_ML_DATA_TYPE_FP16 , ODP_ML_DATA_TYPE_BFP16 , ODP_ML_DATA_TYPE_FP32 , ODP_ML_DATA_TYPE_FP64 }
	Model input / output data type enumeration. More...

enum	odp_ml_shape_type_t { ODP_ML_SHAPE_NONE = 0 , ODP_ML_SHAPE_STATIC , ODP_ML_SHAPE_BATCH }
	Model input / output data shape type. More...

Functions
int	odp_ml_capability (odp_ml_capability_t *capa)
	Query ML capabilities. More...

void	odp_ml_config_init (odp_ml_config_t *config)
	Initialize ML configuration parameters. More...

int	odp_ml_config (const odp_ml_config_t *config)
	Configure ML offload. More...

void	odp_ml_model_param_init (odp_ml_model_param_t *param)
	Initialize ML model parameters. More...

odp_ml_model_t	odp_ml_model_create (const char name, const odp_ml_model_param_t param)
	Create an ML model. More...

int	odp_ml_model_destroy (odp_ml_model_t model)
	Destroy an ML model. More...

odp_ml_model_t	odp_ml_model_lookup (const char *name)
	Find a model by name. More...

int	odp_ml_model_load (odp_ml_model_t model, odp_ml_load_result_t *result)
	Load ML model. More...

int	odp_ml_model_load_start (odp_ml_model_t model, const odp_ml_compl_param_t *compl_param)
	Start asynchronous model load. More...

int	odp_ml_model_load_status (odp_ml_model_t model, uint32_t compl_id, odp_ml_load_result_t *result)
	Check model load completion. More...

int	odp_ml_model_unload (odp_ml_model_t model, odp_ml_load_result_t *result)
	Unload ML model. More...

int	odp_ml_model_unload_start (odp_ml_model_t model, const odp_ml_compl_param_t *compl_param)
	Start asynchronous model unload. More...

int	odp_ml_model_unload_status (odp_ml_model_t model, uint32_t compl_id, odp_ml_load_result_t *result)
	Check model unload completion. More...

void	odp_ml_run_param_init (odp_ml_run_param_t *param)
	Initialize model run parameters. More...

int	odp_ml_run (odp_ml_model_t model, const odp_ml_data_t data, const odp_ml_run_param_t param)
	Run the model in synchronous mode. More...

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)
	Run the model multiple times in synchronous mode. More...

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)
	Start model run in asynchronous mode. More...

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)
	Start multiple model runs in asynchronous mode. More...

int	odp_ml_run_status (odp_ml_model_t model, uint32_t compl_id, odp_ml_run_result_t *result)
	Check model run completion. More...

void	odp_ml_compl_pool_param_init (odp_ml_compl_pool_param_t *param)
	Initialize ML completion event pool parameters. More...

odp_pool_t	odp_ml_compl_pool_create (const char name, const odp_ml_compl_pool_param_t param)
	Create ML completion event pool. More...

odp_ml_compl_t	odp_ml_compl_alloc (odp_pool_t pool)
	Allocate ML completion event. More...

void	odp_ml_compl_free (odp_ml_compl_t ml_compl)
	Free ML completion event. More...

int	odp_ml_compl_run_result (odp_ml_compl_t ml_compl, odp_ml_run_result_t *result)
	Check ML model run results from completion event. More...

int	odp_ml_compl_load_result (odp_ml_compl_t ml_compl, odp_ml_load_result_t *result)
	Check ML model load / unload results from completion event. More...

void *	odp_ml_compl_user_area (odp_ml_compl_t ml_compl)
	ML completion event user area. More...

odp_ml_compl_t	odp_ml_compl_from_event (odp_event_t event)
	Convert event to ML completion event. More...

odp_event_t	odp_ml_compl_to_event (odp_ml_compl_t ml_compl)
	Convert ML completion event to event. More...

uint64_t	odp_ml_compl_to_u64 (odp_ml_compl_t ml_compl)
	Convert ML completion event handle to a uint64_t value for debugging. More...

void	odp_ml_compl_param_init (odp_ml_compl_param_t *param)
	Initialize ML completion parameters. More...

int	odp_ml_model_info (odp_ml_model_t model, odp_ml_model_info_t *info)
	Retrieve model information. More...

uint32_t	odp_ml_model_input_info (odp_ml_model_t model, odp_ml_input_info_t info[], uint32_t num)
	Retrieve model input information. More...

uint32_t	odp_ml_model_output_info (odp_ml_model_t model, odp_ml_output_info_t info[], uint32_t num)
	Retrieve model output information. More...

uint64_t	odp_ml_model_to_u64 (odp_ml_model_t model)
	Convert ML model handle to a uint64_t value for debugging. More...

void	odp_ml_model_print (odp_ml_model_t model)
	Print debug information about the model. More...

void	odp_ml_print (void)
	Print ML debug information. More...

int	odp_ml_model_extra_stat_info (odp_ml_model_t model, odp_ml_extra_stat_info_t info[], int num)
	Extra statistics counter information. More...

int	odp_ml_model_extra_stats (odp_ml_model_t model, uint64_t stats[], int num)
	Read extra statistics counter values. More...

void	odp_ml_fp32_to_uint8 (uint8_t dst_u8, const float src_fp32, uint32_t num, float scale, uint8_t zerop)
	Quantize 32-bit float to uint8_t. More...

void	odp_ml_fp32_from_uint8 (float dst_fp32, const uint8_t src_u8, uint32_t num, float scale, uint8_t zerop)
	De-quantize 32-bit float from uint8_t. More...

void	odp_ml_fp32_to_int8 (int8_t dst_i8, const float src_fp32, uint32_t num, float scale, int8_t zerop)
	Quantize 32-bit float to int8_t. More...

void	odp_ml_fp32_from_int8 (float dst_fp32, const int8_t src_i8, uint32_t num, float scale, int8_t zerop)
	De-quantize 32-bit float from int8_t. More...

void	odp_ml_fp32_to_uint16 (uint16_t dst_u16, const float src_fp32, uint32_t num, float scale, uint16_t zerop)
	Quantize 32-bit float to uint16_t. More...

void	odp_ml_fp32_from_uint16 (float dst_fp32, const uint16_t src_u16, uint32_t num, float scale, uint16_t zerop)
	De-quantize 32-bit float from uint16_t. More...

void	odp_ml_fp32_to_int16 (int16_t dst_i16, const float src_fp32, uint32_t num, float scale, int16_t zerop)
	Quantize 32-bit float to int16_t. More...

void	odp_ml_fp32_from_int16 (float dst_fp32, const int16_t src_i16, uint32_t num, float scale, int16_t zerop)
	De-quantize 32-bit float from int16_t. More...

void	odp_ml_fp32_to_fp16 (uint16_t dst_fp16, const float src_fp32, uint32_t num)
	Quantize 32-bit float to 16-bit float. More...

void	odp_ml_fp32_from_fp16 (float dst_fp32, const uint16_t src_fp16, uint32_t num)
	De-quantize 32-bit float from 16-bit float. More...

Macro Definition Documentation

◆ ODP_ML_COMPL_MODE_POLL

#define ODP_ML_COMPL_MODE_POLL 0x2u

Asynchronous poll mode operation.

A function call starts an operation and a status function call indicates when the operation has finished.

Definition at line 88 of file api/spec/ml_types.h.

◆ ODP_ML_COMPL_MODE_EVENT

#define ODP_ML_COMPL_MODE_EVENT 0x4u

Asynchronous event mode operation.

A function call starts an operation and a completion event indicates when the operation has finished.

Definition at line 96 of file api/spec/ml_types.h.

Typedef Documentation

◆ odp_ml_compl_pool_capability_t

typedef struct odp_ml_compl_pool_capability_t odp_ml_compl_pool_capability_t

ML completion event pool capabilities.

Pool statistics are not supported with ML completion event pools.

◆ odp_ml_compl_pool_param_t

typedef struct odp_ml_compl_pool_param_t odp_ml_compl_pool_param_t

ML completion event pool parameters.

Use odp_ml_compl_pool_param_init() to initialize the structure to its default values.

◆ odp_ml_model_param_t

typedef struct odp_ml_model_param_t odp_ml_model_param_t

Machine learning model parameters.

Use odp_ml_model_param_init() to initialize the structure to its default values.

◆ odp_ml_compl_param_t

typedef struct odp_ml_compl_param_t odp_ml_compl_param_t

ML completion parameters.

Use odp_ml_compl_param_init() to initialize the structure to its default values.

◆ odp_ml_run_param_t

typedef struct odp_ml_run_param_t odp_ml_run_param_t

Parameters for model run.

Use odp_ml_run_param_init() to initialize the structure to its default values.

Enumeration Type Documentation

◆ odp_ml_data_type_t

enum odp_ml_data_type_t

Model input / output data type enumeration.

Enumerator
ODP_ML_DATA_TYPE_NONE	Data type is not defined.
ODP_ML_DATA_TYPE_INT8	8-bit integer
ODP_ML_DATA_TYPE_UINT8	8-bit unsigned integer
ODP_ML_DATA_TYPE_INT16	16-bit integer
ODP_ML_DATA_TYPE_UINT16	16-bit unsigned integer
ODP_ML_DATA_TYPE_INT24	24-bit integer
ODP_ML_DATA_TYPE_UINT24	24-bit unsigned integer
ODP_ML_DATA_TYPE_INT32	32-bit integer
ODP_ML_DATA_TYPE_UINT32	32-bit unsigned integer
ODP_ML_DATA_TYPE_INT64	64-bit integer
ODP_ML_DATA_TYPE_UINT64	64-bit unsigned integer
ODP_ML_DATA_TYPE_FP16	16-bit floating point number
ODP_ML_DATA_TYPE_BFP16	16-bit brain floating point (bfloat16) number
ODP_ML_DATA_TYPE_FP32	32-bit floating point number
ODP_ML_DATA_TYPE_FP64	64-bit floating point number

Definition at line 378 of file api/spec/ml_types.h.

◆ odp_ml_shape_type_t

enum odp_ml_shape_type_t

Model input / output data shape type.

Enumerator

ODP_ML_SHAPE_NONE

Type of shape is not defined.

ODP_ML_SHAPE_STATIC

Static shape of data.

Shape is static when all dimensions have fixed sizes.

ODP_ML_SHAPE_BATCH

Dynamic batch size.

Shape that has only one dynamic dimension, and the dimension is used as batch size of
input / output data. The same batch size is applied for all inputs and outputs of
the model.

Definition at line 427 of file api/spec/ml_types.h.

Function Documentation

◆ odp_ml_capability()

int odp_ml_capability ( odp_ml_capability_t * capa )

Query ML capabilities.

Outputs ML capabilities on success. Use this capability call to check ML offload implementation limits and its support of various ML API features. When ML offload is not available, odp_ml_capability_t.max_models is zero.

Parameters

[out] capa Pointer to a capability structure for output

Return values

0	on success
<0	on failure

Examples: odp_ml_perf.c, odp_ml_run.c, and odp_sysinfo.c.

◆ odp_ml_config_init()

void odp_ml_config_init ( odp_ml_config_t * config )

Initialize ML configuration parameters.

Initialize an odp_ml_config_t to its default values.

Parameters

[out] config Configuration structure to be initialized

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_config()

int odp_ml_config ( const odp_ml_config_t * config )

Configure ML offload.

Initializes and configures ML offload according to the configuration parameters. This function must be called only once and before any ML resources are created. Use odp_ml_capability() to query configuration capabilities and odp_ml_config_init() to initialize configuration parameters into their default values.

Parameters

config ML configuration parameters

Return values

0	on success
<0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_param_init()

void odp_ml_model_param_init ( odp_ml_model_param_t * param )

Initialize ML model parameters.

Initialize an odp_ml_model_param_t to its default values.

Parameters

[out] param Parameters structure to be initialized

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_create()

odp_ml_model_t odp_ml_model_create	(	const char *	name,
		const odp_ml_model_param_t *	param
	)

Create an ML model.

Creates an ML model according to the parameters. Use odp_ml_model_param_init() to initialize parameters into their default values. The use of model name is optional. Unique names are not required. However, odp_ml_model_lookup() returns only a single matching model.

The call copies the model binary and prepares it for loading. Application may free memory buffers pointed by the parameters when the call returns. Use odp_ml_model_load() or odp_ml_model_load_start() to load the model. A model is ready for inference runs (see e.g. odp_ml_run()) after it has been loaded successfully.

When model metadata misses some details of model input / output data format, user can pass those with odp_ml_model_param_t.extra_info. Some ODP implementations may define implementation specific extra parameters (e.g. hints about HW resource usage), user can pass those with odp_ml_model_param_t.extra_param when applicable.

Parameters

name	Name of the model, or NULL. Maximum string length is ODP_ML_MODEL_NAME_LEN, including the null character.
param	ML model parameters

Returns: ML model handle on success

Return values

ODP_ML_MODEL_INVALID on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_destroy()

int odp_ml_model_destroy ( odp_ml_model_t model )

Destroy an ML model.

Destroys a model and releases the resources reserved for it. If the model has been loaded, it must be unloaded (see odp_ml_model_unload() or odp_ml_model_unload_start()) prior to calling this function.

Parameters

model ML model to be destroyed

Return values

0	on success
<0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_lookup()

odp_ml_model_t odp_ml_model_lookup ( const char * name )

Find a model by name.

Parameters

name	Name of the model

Returns: Handle of the first matching ML model

Return values

ODP_ML_MODEL_INVALID Model could not be found

◆ odp_ml_model_load()

int odp_ml_model_load	(	odp_ml_model_t	model,
		odp_ml_load_result_t *	result
	)

Load ML model.

Loads ML model in synchronous mode. When the call returns, load is complete and the model is ready for inference requests. A loaded model must be unloaded before it can be destroyed. The same model can be loaded and unloaded multiple times before being destroyed.

The call optionally outputs results. Use NULL as 'result' pointer if results are not required.

Application should not try to keep loaded more than configured number of models (odp_ml_config_t.max_models_loaded). Check ML capabilities for maximum number of loaded models (odp_ml_capability_t.max_models_loaded) and support of load completion modes (odp_ml_capability_t.load).

Parameters

	model	ML model to be loaded
[out]	result	Pointer to load result structure for output, or NULL

Return values

0	Model load was successful
<0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_load_start()

int odp_ml_model_load_start	(	odp_ml_model_t	model,
		const odp_ml_compl_param_t *	compl_param
	)

Start asynchronous model load.

Otherwise like odp_ml_model_load(), but loads the model asynchronously. A successful call requests the model to be loaded, but does not wait for load completion. Completion parameters are used to select if load completion is reported in poll (ODP_ML_COMPL_MODE_POLL) or event (ODP_ML_COMPL_MODE_EVENT) mode. For poll mode, odp_ml_model_load_status() is called to check for completion. For event mode, ML offload sends the completion event into the completion queue when the load is complete. Use odp_ml_compl_param_init() to initialize completion parameters into their default values.

Parameters

model	ML model to be loaded
compl_param	Completion parameters for load

Return values

0	Model load started successfully
<0	on failure

◆ odp_ml_model_load_status()

int odp_ml_model_load_status	(	odp_ml_model_t	model,
		uint32_t	compl_id,
		odp_ml_load_result_t *	result
	)

Check model load completion.

Checks if a previously started model load (in ODP_ML_COMPL_MODE_POLL mode) has completed. The completion identifier value from load operation completion parameters (odp_ml_compl_param_t.compl_id) is passed as a parameter. It specifies the load operation to be checked. Initially 0 is returned for all configured (but unused) completion identifier values. An odp_ml_model_load_start() call clears the previous completion status of an identifier, and this function returns 0 while the load is in progress. When the load is successfully complete, >0 is returned. If the load completed with a failure, -1 is returned. The same value is returned until the next start operation that reuses the identifier (with the same model). The completion identifier may be reused only after >0 or -1 is returned.

Optionally, outputs more detailed operation results into odp_ml_load_result_t structure. Use NULL as 'result' pointer if these results are not required.

Parameters

	model	ML model being loaded
	compl_id	Completion identifier that was used in load start
[out]	result	Pointer to load result structure for output, or NULL

Return values

>0	Model load was successful
0	Model load has not finished
-1	Model load failed
<-1	Failed to read completion status (e.g. bad handle)

◆ odp_ml_model_unload()

int odp_ml_model_unload	(	odp_ml_model_t	model,
		odp_ml_load_result_t *	result
	)

Unload ML model.

Unloads ML model in synchronous mode. All previously started inference operations must have been completed before model unload is attempted. When the call returns, unload is complete and the model is ready to be destroyed or loaded again.

The call optionally outputs results. Use NULL as 'result' pointer if results are not required.

Parameters

	model	ML model to be unloaded
[out]	result	Pointer to load result structure for output, or NULL

Return values

0	Model unload was successful
<0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_unload_start()

int odp_ml_model_unload_start	(	odp_ml_model_t	model,
		const odp_ml_compl_param_t *	compl_param
	)

Start asynchronous model unload.

Otherwise like odp_ml_model_unload(), but unloads the model asynchronously. A successful call requests the model to be unloaded, but does not wait for unload completion. Completion parameters are used to select if unload completion is reported in poll (ODP_ML_COMPL_MODE_POLL) or event (ODP_ML_COMPL_MODE_EVENT) mode. For poll mode, odp_ml_model_unload_status() is called to check for completion. For event mode, ML offload sends the completion event into the completion queue when the unload is complete. Use odp_ml_compl_param_init() to initialize completion parameters into their default values.

Parameters

model	ML model to be unloaded
compl_param	Completion parameters for unload

Return values

0	Model unload started successfully
<0	on failure

◆ odp_ml_model_unload_status()

int odp_ml_model_unload_status	(	odp_ml_model_t	model,
		uint32_t	compl_id,
		odp_ml_load_result_t *	result
	)

Check model unload completion.

Checks if a previously started model unload (in ODP_ML_COMPL_MODE_POLL mode) has completed. The completion identifier value from unload operation completion parameters (odp_ml_compl_param_t.compl_id) is passed as a parameter. It specifies the unload operation to be checked. Initially 0 is returned for all configured (but unused) completion identifier values. An odp_ml_model_unload_start() call clears the previous completion status of an identifier, and this function returns 0 while the unload is in progress. When the unload is successfully complete, >0 is returned. If the unload completed with a failure, -1 is returned. The same value is returned until the next start operation that reuses the identifier (with the same model). The completion identifier may be reused only after >0 or -1 is returned.

Optionally, outputs more detailed operation results into odp_ml_load_result_t structure. Use NULL as 'result' pointer if these results are not required.

Parameters

	model	ML model being unloaded
	compl_id	Completion identifier that was used in unload start
[out]	result	Pointer to load result structure for output, or NULL

Return values

>0	Model unload was successful
0	Model unload has not finished
-1	Model unload failed
<-1	Failed to read completion status (e.g. bad handle)

◆ odp_ml_run_param_init()

void odp_ml_run_param_init ( odp_ml_run_param_t * param )

Initialize model run parameters.

Initialize an odp_ml_run_param_t to its default values.

Parameters

[out] param Model run parameters structure to be initialized

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_run()

int odp_ml_run	(	odp_ml_model_t	model,
		const odp_ml_data_t *	data,
		const odp_ml_run_param_t *	param
	)

Run the model in synchronous mode.

Performs an ML inference operation using the model and input data pointed by the data descriptor. A successful operation writes inference output data into memory buffers pointed by the data descriptor. Input/output data buffers are described as an array of segment descriptors. Each segment descriptor specifies a memory buffer used with only one model input/output. Multiple subsequent descriptors may be used to specify segmented data for the same input/output. When the model has multiple inputs/outputs, descriptor order in the array follows the model input/output order reported by odp_ml_model_input_info() and odp_ml_model_output_info() calls. All memory buffers for the first input/output are specified before any buffers for the second input/output, and so on.

When some model inputs/outputs have ODP_ML_SHAPE_BATCH shape type, the batch size is specified in run parameters (odp_ml_run_param_t.batch_size). The same batch size is used for all such inputs/outputs. Application may request additional operation results by setting 'result' pointer in run parameters. Use odp_ml_run_param_init() to initialize run parameters into their default values. Default run parameter values are used when 'param' is NULL.

Returns 1 when model run completed successfully. Returns 0 when the operation was not performed due to ML offload resources being temporarily busy. Returns <0 on failure.

Parameters

model	ML model to be run
data	Model input/output data descriptor
param	Model run parameters, or NULL

Return values

1	Model run completed successfully
0	Resources are busy and model was not run
<0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_run_multi()

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
	)

Run the model multiple times in synchronous mode.

Otherwise like odp_ml_run(), but runs the model 'num' times with different input/output data buffers. Output data buffers of one ML inference operation must not overlap with input/output data buffers of another one.

Returns number of model runs successfully completed. When return value is less than 'num', the remaining runs were not performed due to ML offload resources being temporarily busy. Returns <0 on failure.

Parameters

model	ML model to be run
data	Array of model input/output data descriptors. The array has 'num' elements.
param	Array of model run parameters, or NULL. The array has 'num' elements when used.
num	Number of model runs to perform

Returns: Number of model runs completed successfully (1 ... num)

Return values

0	Resources are busy and model was not run
<0	on failure

◆ odp_ml_run_start()

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
	)

Start model run in asynchronous mode.

Otherwise like odp_ml_run(), but runs the model asynchronously. A successful call requests the model to be run, but does not wait for run completion. Completion parameters select if run completion is reported in poll (ODP_ML_COMPL_MODE_POLL) or event (ODP_ML_COMPL_MODE_EVENT) mode. For poll mode, odp_ml_run_status() is called to check for completion. For event mode, ML offload sends the completion event into the completion queue when the run is complete. Use odp_ml_compl_param_init() to initialize completion parameters into their default values.

Additional operation results (odp_ml_run_result_t) are available through the status call (odp_ml_run_status()) or completion event (odp_ml_compl_run_result()). Results are not output through the run parameters structure (i.e. odp_ml_run_param_t.result is ignored).

Returns 1 when model run was started successfully. Returns 0 when model run was not started due to ML offload resources being temporarily busy. Returns <0 on failure.

Parameters

model	ML model to be run
data	Model input/output data descriptor
compl_param	Completion parameters
run_param	Model run parameters, or NULL

Return values

1	Model run started successfully
0	Resources are busy and model run was not started
<0	on failure

◆ odp_ml_run_start_multi()

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
	)

Start multiple model runs in asynchronous mode.

Otherwise like odp_ml_run_start(), but starts 'num' model runs with different input/output data buffers. Output data buffers of one ML inference operation must not overlap with input/output data buffers of another one.

Returns number of model runs started successfully. When return value is less than 'num', the remaining runs were not started due to ML offload resources being temporarily busy. Returns <0 on failure.

Parameters

model	ML model to be run
data	Array of model input/output data descriptors. The array has 'num' elements.
compl_param	Array of completion parameters. The array has 'num' elements.
run_param	Array of model run parameters, or NULL. The array has 'num' elements when used.
num	Number of model runs to start

Returns: Number of model runs started successfully (1 ... num)

Return values

0	Resources are busy and model runs were not started
<0	on failure

◆ odp_ml_run_status()

int odp_ml_run_status	(	odp_ml_model_t	model,
		uint32_t	compl_id,
		odp_ml_run_result_t *	result
	)

Check model run completion.

Checks if a previously started model run (in ODP_ML_COMPL_MODE_POLL mode) has completed. The completion identifier value from run operation completion parameters (odp_ml_compl_param_t.compl_id) is passed as a parameter. It specifies the run operation to be checked. Initially 0 is returned for all configured (but unused) completion identifier values. An odp_ml_run_start() call clears the previous completion status of an identifier, and this function returns 0 while the run is in progress. When the run is successfully complete, >0 is returned. If the run completed with a failure, -1 is returned. The same value is returned until the next start operation that reuses the identifier (with the same model). The completion identifier may be reused only after >0 or -1 is returned.

Optionally, outputs more detailed operation results into odp_ml_run_result_t structure. Use NULL as 'result' pointer if these results are not required.

Parameters

	model	ML model running
	compl_id	Completion identifier that was used in run start
[out]	result	Pointer to run result structure for output, or NULL

Return values

>0	Model run was successful
0	Model run has not finished
-1	Model run failed
<-1	Failed to read completion status (e.g. bad handle)

◆ odp_ml_compl_pool_param_init()

void odp_ml_compl_pool_param_init ( odp_ml_compl_pool_param_t * param )

Initialize ML completion event pool parameters.

Initialize an odp_ml_compl_pool_param_t to its default values.

Parameters

[out] param Parameter structure to be initialized

◆ odp_ml_compl_pool_create()

odp_pool_t odp_ml_compl_pool_create	(	const char *	name,
		const odp_ml_compl_pool_param_t *	param
	)

Create ML completion event pool.

Creates a pool of ML completion events (ODP_EVENT_ML_COMPL). Pool type is ODP_POOL_ML_COMPL. The use of pool name is optional. Unique names are not required. However, odp_pool_lookup() returns only a single matching pool. Use odp_ml_compl_pool_param_init() to initialize pool parameters into their default values. Parameters values must not exceed pool capabilities (see odp_ml_compl_pool_capability_t).

Parameters

name	Name of the pool or NULL. Maximum string length is ODP_POOL_NAME_LEN, including the null character.
param	Pool parameters

Returns: Pool handle on success

Return values

ODP_POOL_INVALID on failure

◆ odp_ml_compl_alloc()

odp_ml_compl_t odp_ml_compl_alloc ( odp_pool_t pool )

Allocate ML completion event.

Allocates an ML completion event from a pool. The pool must have been created with odp_ml_compl_pool_create() call. All completion event metadata are set to their default values.

Parameters

pool	ML completion event pool

Returns: ML completion event handle

Return values

ODP_ML_COMPL_INVALID Completion event could not be allocated

◆ odp_ml_compl_free()

void odp_ml_compl_free ( odp_ml_compl_t ml_compl )

Free ML completion event.

Frees an ML completion event into the pool it was allocated from.

Parameters

ml_compl ML completion event handle

◆ odp_ml_compl_run_result()

int odp_ml_compl_run_result	(	odp_ml_compl_t	ml_compl,
		odp_ml_run_result_t *	result
	)

Check ML model run results from completion event.

Reads model run results from an ML completion event (ODP_EVENT_ML_COMPL). The event indicates completion of a previously started inference operation. Subtype of the completion event must be ODP_EVENT_ML_COMPL_RUN. Function return value indicates if the model run succeeded or failed. Additionally, outputs more detailed results into the provided odp_ml_run_result_t structure. Use NULL as 'result' pointer if those results are not required.

Parameters

	ml_compl	ML completion event (subtype ODP_EVENT_ML_COMPL_RUN)
[out]	result	Pointer to ML run result structure for output, or NULL.

Return values

0	Model run was successful
-1	Model run failed
<-1	Failed to read results from the event (e.g. bad handle)

◆ odp_ml_compl_load_result()

int odp_ml_compl_load_result	(	odp_ml_compl_t	ml_compl,
		odp_ml_load_result_t *	result
	)

Check ML model load / unload results from completion event.

Reads model load / unload results from an ML completion event (ODP_EVENT_ML_COMPL). The event indicates completion of a previously started model load / unload operation. Subtype of the completion event must be ODP_EVENT_ML_COMPL_LOAD. Function return value indicates if the model load / unload succeeded or failed. Additionally, outputs more detailed results into the provided odp_ml_load_result_t structure. Use NULL as 'result' pointer if those results are not required.

Parameters

	ml_compl	ML completion event (subtype ODP_EVENT_ML_COMPL_LOAD)
[out]	result	Pointer to model load / unload result structure for output, or NULL.

Return values

0	Model load / unload was successful
-1	Model load / unload failed
<-1	Failed to read results from the event (e.g. bad handle)

◆ odp_ml_compl_user_area()

void* odp_ml_compl_user_area ( odp_ml_compl_t ml_compl )

ML completion event user area.

Returns pointer to the user area associated with the completion event. Size of the area is fixed and defined in pool parameters.

Parameters

ml_compl ML completion event

Returns: Pointer to the user area of the completion event

Return values

NULL	The completion event does not have user area

◆ odp_ml_compl_from_event()

odp_ml_compl_t odp_ml_compl_from_event ( odp_event_t event )

Convert event to ML completion event.

Converts an ODP_EVENT_ML_COMPL type event to an ML completion event.

Parameters

event Event handle

Returns: ML completion event handle

◆ odp_ml_compl_to_event()

odp_event_t odp_ml_compl_to_event ( odp_ml_compl_t ml_compl )

Convert ML completion event to event.

Parameters

ml_compl ML completion event handle

Returns: Event handle

◆ odp_ml_compl_to_u64()

uint64_t odp_ml_compl_to_u64 ( odp_ml_compl_t ml_compl )

Convert ML completion event handle to a uint64_t value for debugging.

Parameters

ml_compl ML completion event handle to be converted

Returns: uint64_t value that can be used for debugging (e.g. printed)

◆ odp_ml_compl_param_init()

void odp_ml_compl_param_init ( odp_ml_compl_param_t * param )

Initialize ML completion parameters.

Initialize an odp_ml_compl_param_t to its default values.

Parameters

[out] param Address of parameters structure to be initialized

◆ odp_ml_model_info()

int odp_ml_model_info	(	odp_ml_model_t	model,
		odp_ml_model_info_t *	info
	)

Retrieve model information.

Retrieve information about the model. Model information includes e.g. version numbers and number of model inputs/outputs. Information about each input and output can be retrieved with odp_ml_model_input_info() and odp_ml_model_output_info() calls.

Parameters

	model	ML model handle
[out]	info	Pointer to model information structure for output

Return values

0	on success
<0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_input_info()

uint32_t odp_ml_model_input_info	(	odp_ml_model_t	model,
		odp_ml_input_info_t	info[],
		uint32_t	num
	)

Retrieve model input information.

Writes information about each model input into the array. If there are more inputs than array elements, writes only 'num' elements. Returns the number of model inputs on success, and zero on failure. When 'num' is zero, ignores value of 'info' and returns normally.

Parameters

	model	ML model handle
[out]	info	Pointer to model input information array for output
	num	Number of elements in the array

Returns: Number of model inputs

Return values

0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_output_info()

uint32_t odp_ml_model_output_info	(	odp_ml_model_t	model,
		odp_ml_output_info_t	info[],
		uint32_t	num
	)

Retrieve model output information.

Writes information about each model output into the array. If there are more outputs than array elements, writes only 'num' elements. Returns the number of model outputs on success, and zero on failure. When 'num' is zero, ignores value of 'info' and returns normally.

Parameters

	model	ML model handle
[out]	info	Pointer to model output information array for output
	num	Number of elements in the array

Returns: Number of model outputs

Return values

0	on failure

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_model_to_u64()

uint64_t odp_ml_model_to_u64 ( odp_ml_model_t model )

Convert ML model handle to a uint64_t value for debugging.

Parameters

model ML model handle

Returns: uint64_t value that can be used for debugging (e.g. printed)

◆ odp_ml_model_print()

void odp_ml_model_print ( odp_ml_model_t model )

Print debug information about the model.

Print implementation defined information about ML model to the ODP log. The information is intended to be used for debugging.

Parameters

model ML model handle

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_print()

void odp_ml_print ( void )

Print ML debug information.

Print implementation defined information about ML offload to the ODP log. The information is intended to be used for debugging.

◆ odp_ml_model_extra_stat_info()

int odp_ml_model_extra_stat_info	(	odp_ml_model_t	model,
		odp_ml_extra_stat_info_t	info[],
		int	num
	)

Extra statistics counter information.

Returns the number of extra statistics counters supported by the ML offload, and outputs information (e.g. name) about those. Counters are implementation specific and maintained per model. Statistics counting is enabled through model create parameters.

When 'info' pointer is not NULL, fills in up to 'num' counter info structures. If the return value is larger than 'num', there are more counters than the function was allowed to output. If the return value N is less than 'num' (on success), only first N structures have been written.

Info array elements are filled in the same order than odp_ml_model_extra_stats() outputs counter values.

Parameters

	model	ML model
[out]	info	Pointer to extra statistics counter information array for output. NULL may be used to query only the number of counters.
	num	Number of elements in the array

Returns: Number of extra statistics counters

Return values

<0	on failure

◆ odp_ml_model_extra_stats()

int odp_ml_model_extra_stats	(	odp_ml_model_t	model,
		uint64_t	stats[],
		int	num
	)

Read extra statistics counter values.

Reads extra statistics counter values and returns the number of supported counters. Outputs up to 'num' counter values into 'stats' array. If the return value is larger than 'num', there are more counters than the function was allowed to output. If the return value N is less than 'num' (on success), only first N counters have been written. The order of counters in the array matches the counter information array order on odp_ml_model_extra_stat_info() output.

Parameters

	model	ML model
[out]	stats	Pointer to extra statistics counter array for output
	num	Number of elements in the array

Returns: Number of extra statistics counters

Return values

<0	on failure

◆ odp_ml_fp32_to_uint8()

void odp_ml_fp32_to_uint8	(	uint8_t *	dst_u8,
		const float *	src_fp32,
		uint32_t	num,
		float	scale,
		uint8_t	zerop
	)

Quantize 32-bit float to uint8_t.

Quantizes 'num' 32-bit floating point values to uint8_t values using the provided scale and zero point.

dst_u8 = (src_fp32 / scale) + zerop

Parameters

[out]	dst_u8	Destination address for quantized values
	src_fp32	Source address of values to be quantized
	num	Number of values
	scale	Scale for quantization
	zerop	Zero point for quantization

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_fp32_from_uint8()

void odp_ml_fp32_from_uint8	(	float *	dst_fp32,
		const uint8_t *	src_u8,
		uint32_t	num,
		float	scale,
		uint8_t	zerop
	)

De-quantize 32-bit float from uint8_t.

De-quantizes 'num' 32-bit floating point values from uint8_t values using the provided scale and zero point.

dst_fp32 = (src_u8 - zerop) * scale

Parameters

[out]	dst_fp32	Destination address for de-quantized values
	src_u8	Source address of values to be de-quantized
	num	Number of values
	scale	Scale for de-quantization
	zerop	Zero point for de-quantization

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_fp32_to_int8()

void odp_ml_fp32_to_int8	(	int8_t *	dst_i8,
		const float *	src_fp32,
		uint32_t	num,
		float	scale,
		int8_t	zerop
	)

Quantize 32-bit float to int8_t.

Quantizes 'num' 32-bit floating point values to int8_t values using the provided scale and zero point.

dst_i8 = (src_fp32 / scale) + zerop

Parameters

[out]	dst_i8	Destination address for quantized values
	src_fp32	Source address of values to be quantized
	num	Number of values
	scale	Scale for quantization
	zerop	Zero point for quantization

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_fp32_from_int8()

void odp_ml_fp32_from_int8	(	float *	dst_fp32,
		const int8_t *	src_i8,
		uint32_t	num,
		float	scale,
		int8_t	zerop
	)

De-quantize 32-bit float from int8_t.

De-quantizes 'num' 32-bit floating point values from int8_t values using the provided scale and zero point.

dst_fp32 = (src_i8 - zerop) * scale

Parameters

[out]	dst_fp32	Destination address for de-quantized values
	src_i8	Source address of values to be de-quantized
	num	Number of values
	scale	Scale for de-quantization
	zerop	Zero point for de-quantization

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_fp32_to_uint16()

void odp_ml_fp32_to_uint16	(	uint16_t *	dst_u16,
		const float *	src_fp32,
		uint32_t	num,
		float	scale,
		uint16_t	zerop
	)

Quantize 32-bit float to uint16_t.

Quantizes 'num' 32-bit floating point values to uint16_t values using the provided scale and zero point.

dst_u16 = (src_fp32 / scale) + zerop

Parameters

[out]	dst_u16	Destination address for quantized values
	src_fp32	Source address of values to be quantized
	num	Number of values
	scale	Scale for quantization
	zerop	Zero point for quantization

◆ odp_ml_fp32_from_uint16()

void odp_ml_fp32_from_uint16	(	float *	dst_fp32,
		const uint16_t *	src_u16,
		uint32_t	num,
		float	scale,
		uint16_t	zerop
	)

De-quantize 32-bit float from uint16_t.

De-quantizes 'num' 32-bit floating point values from uint16_t values using the provided scale and zero point.

dst_fp32 = (src_u16 - zerop) * scale

Parameters

[out]	dst_fp32	Destination address for de-quantized values
	src_u16	Source address of values to be de-quantized
	num	Number of values
	scale	Scale for de-quantization
	zerop	Zero point for de-quantization

◆ odp_ml_fp32_to_int16()

void odp_ml_fp32_to_int16	(	int16_t *	dst_i16,
		const float *	src_fp32,
		uint32_t	num,
		float	scale,
		int16_t	zerop
	)

Quantize 32-bit float to int16_t.

Quantizes 'num' 32-bit floating point values to int16_t values using the provided scale and zero point.

dst_i16 = (src_fp32 / scale) + zerop

Parameters

[out]	dst_i16	Destination address for quantized values
	src_fp32	Source address of values to be quantized
	num	Number of values
	scale	Scale for quantization
	zerop	Zero point for quantization

◆ odp_ml_fp32_from_int16()

void odp_ml_fp32_from_int16	(	float *	dst_fp32,
		const int16_t *	src_i16,
		uint32_t	num,
		float	scale,
		int16_t	zerop
	)

De-quantize 32-bit float from int16_t.

De-quantizes 'num' 32-bit floating point values from int16_t values using the provided scale and zero point.

dst_fp32 = (src_i16 - zerop) * scale

Parameters

[out]	dst_fp32	Destination address for de-quantized values
	src_i16	Source address of values to be de-quantized
	num	Number of values
	scale	Scale for de-quantization
	zerop	Zero point for de-quantization

◆ odp_ml_fp32_to_fp16()

void odp_ml_fp32_to_fp16	(	uint16_t *	dst_fp16,
		const float *	src_fp32,
		uint32_t	num
	)

Quantize 32-bit float to 16-bit float.

Quantizes 'num' 32-bit floating point values to 16-bit floating point values.

Parameters

[out]	dst_fp16	Destination address for quantized values
	src_fp32	Source address of values to be quantized
	num	Number of values

Examples: odp_ml_perf.c, and odp_ml_run.c.

◆ odp_ml_fp32_from_fp16()

void odp_ml_fp32_from_fp16	(	float *	dst_fp32,
		const uint16_t *	src_fp16,
		uint32_t	num
	)

De-quantize 32-bit float from 16-bit float.

De-quantizes 'num' 32-bit floating point values from 16-bit floating point values.

Parameters

[out]	dst_fp32	Destination address for de-quantized values
	src_fp16	Source address of values to be de-quantized
	num	Number of values

Examples: odp_ml_perf.c, and odp_ml_run.c.

API Description

Data Structures

Macros

Typedefs

Enumerations

Functions

Macro Definition Documentation

◆ ODP_ML_COMPL_MODE_POLL

◆ ODP_ML_COMPL_MODE_EVENT

Typedef Documentation

◆ odp_ml_compl_pool_capability_t

◆ odp_ml_compl_pool_param_t

◆ odp_ml_model_param_t

◆ odp_ml_compl_param_t

◆ odp_ml_run_param_t

Enumeration Type Documentation

◆ odp_ml_data_type_t

◆ odp_ml_shape_type_t

Function Documentation

◆ odp_ml_capability()

◆ odp_ml_config_init()

◆ odp_ml_config()

◆ odp_ml_model_param_init()

◆ odp_ml_model_create()

◆ odp_ml_model_destroy()

◆ odp_ml_model_lookup()

◆ odp_ml_model_load()

◆ odp_ml_model_load_start()

◆ odp_ml_model_load_status()

◆ odp_ml_model_unload()

◆ odp_ml_model_unload_start()

◆ odp_ml_model_unload_status()

◆ odp_ml_run_param_init()

◆ odp_ml_run()

◆ odp_ml_run_multi()

◆ odp_ml_run_start()

◆ odp_ml_run_start_multi()

◆ odp_ml_run_status()

◆ odp_ml_compl_pool_param_init()

◆ odp_ml_compl_pool_create()

◆ odp_ml_compl_alloc()

◆ odp_ml_compl_free()

◆ odp_ml_compl_run_result()

◆ odp_ml_compl_load_result()

◆ odp_ml_compl_user_area()

◆ odp_ml_compl_from_event()

◆ odp_ml_compl_to_event()

◆ odp_ml_compl_to_u64()

◆ odp_ml_compl_param_init()

◆ odp_ml_model_info()

◆ odp_ml_model_input_info()

◆ odp_ml_model_output_info()

◆ odp_ml_model_to_u64()

◆ odp_ml_model_print()

◆ odp_ml_print()

◆ odp_ml_model_extra_stat_info()

◆ odp_ml_model_extra_stats()

◆ odp_ml_fp32_to_uint8()

◆ odp_ml_fp32_from_uint8()

◆ odp_ml_fp32_to_int8()

◆ odp_ml_fp32_from_int8()

◆ odp_ml_fp32_to_uint16()

◆ odp_ml_fp32_from_uint16()

◆ odp_ml_fp32_to_int16()

◆ odp_ml_fp32_from_int16()

◆ odp_ml_fp32_to_fp16()

◆ odp_ml_fp32_from_fp16()