FFmpeg
Data Structures | Macros | Functions
dnn_backend_common.h File Reference
#include "queue.h"
#include "../dnn_interface.h"
#include "libavutil/thread.h"

Go to the source code of this file.

Data Structures

struct  TaskItem
 
struct  LastLevelTaskItem
 
struct  DNNAsyncExecModule
 Common Async Execution Mechanism for the DNN Backends. More...
 

Macros

#define DNN_BACKEND_COMMON_OPTIONS
 

Functions

int ff_check_exec_params (void *ctx, DNNBackendType backend, DNNFunctionType func_type, DNNExecBaseParams *exec_params)
 
int ff_dnn_fill_task (TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int async, int do_ioproc)
 Fill the Task for Backend Execution. More...
 
int ff_dnn_async_module_cleanup (DNNAsyncExecModule *async_module)
 Join the Async Execution thread and set module pointers to NULL. More...
 
int ff_dnn_start_inference_async (void *ctx, DNNAsyncExecModule *async_module)
 Start asynchronous inference routine for the TensorFlow model on a detached thread. More...
 
DNNAsyncStatusType ff_dnn_get_result_common (Queue *task_queue, AVFrame **in, AVFrame **out)
 Extract input and output frame from the Task Queue after asynchronous inference. More...
 
int ff_dnn_fill_gettingoutput_task (TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int input_height, int input_width, void *ctx)
 Allocate input and output frames and fill the Task with execution parameters. More...
 

Detailed Description

DNN common functions different backends.

Definition in file dnn_backend_common.h.

Macro Definition Documentation

◆ DNN_BACKEND_COMMON_OPTIONS

#define DNN_BACKEND_COMMON_OPTIONS
Value:
{ "nireq", "number of request", OFFSET(options.nireq), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, FLAGS }, \
{ "async", "use DNN async inference", OFFSET(options.async), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, FLAGS },

Definition at line 31 of file dnn_backend_common.h.

Function Documentation

◆ ff_check_exec_params()

int ff_check_exec_params ( void *  ctx,
DNNBackendType  backend,
DNNFunctionType  func_type,
DNNExecBaseParams exec_params 
)

Definition at line 29 of file dnn_backend_common.c.

Referenced by dnn_execute_model_tf(), and get_output_ov().

◆ ff_dnn_fill_task()

int ff_dnn_fill_task ( TaskItem task,
DNNExecBaseParams exec_params,
void *  backend_model,
int  async,
int  do_ioproc 
)

Fill the Task for Backend Execution.

It should be called after checking execution parameters using ff_check_exec_params.

Parameters
taskpointer to the allocated task
exec_parampointer to execution parameters
backend_modelvoid pointer to the backend model
asyncflag for async execution. Must be 0 or 1
do_ioprocflag for IO processing. Must be 0 or 1
Returns
0 if successful or error code otherwise.

Definition at line 49 of file dnn_backend_common.c.

Referenced by dnn_execute_model_tf(), ff_dnn_fill_gettingoutput_task(), and get_output_ov().

◆ ff_dnn_async_module_cleanup()

int ff_dnn_async_module_cleanup ( DNNAsyncExecModule async_module)

Join the Async Execution thread and set module pointers to NULL.

Parameters
async_modulepointer to DNNAsyncExecModule module
Returns
0 if successful or error code otherwise.

Definition at line 85 of file dnn_backend_common.c.

Referenced by destroy_request_item().

◆ ff_dnn_start_inference_async()

int ff_dnn_start_inference_async ( void *  ctx,
DNNAsyncExecModule async_module 
)

Start asynchronous inference routine for the TensorFlow model on a detached thread.

It calls the completion callback after the inference completes. Completion callback and inference function must be set before calling this function.

If POSIX threads aren't supported, the execution rolls back to synchronous mode, calling completion callback after inference.

Parameters
ctxpointer to the backend context
async_modulepointer to DNNAsyncExecModule module
Returns
0 on the start of async inference or error code otherwise.

Definition at line 104 of file dnn_backend_common.c.

Referenced by dnn_flush_tf(), and execute_model_tf().

◆ ff_dnn_get_result_common()

DNNAsyncStatusType ff_dnn_get_result_common ( Queue task_queue,
AVFrame **  in,
AVFrame **  out 
)

Extract input and output frame from the Task Queue after asynchronous inference.

Parameters
task_queuepointer to the task queue of the backend
indouble pointer to the input frame
outdouble pointer to the output frame
Return values
DAST_EMPTY_QUEUEif task queue is empty
DAST_NOT_READYif inference not completed yet.
DAST_SUCCESSif result successfully extracted

Definition at line 135 of file dnn_backend_common.c.

Referenced by dnn_get_result_tf(), and get_output_ov().

◆ ff_dnn_fill_gettingoutput_task()

int ff_dnn_fill_gettingoutput_task ( TaskItem task,
DNNExecBaseParams exec_params,
void *  backend_model,
int  input_height,
int  input_width,
void *  ctx 
)

Allocate input and output frames and fill the Task with execution parameters.

Parameters
taskpointer to the allocated task
exec_paramspointer to execution parameters
backend_modelvoid pointer to the backend model
input_heightheight of input frame
input_widthwidth of input frame
ctxpointer to the backend context
Returns
0 if successful or error code otherwise.

Definition at line 155 of file dnn_backend_common.c.

Referenced by get_output_ov(), and get_output_tf().

FLAGS
#define FLAGS
Definition: cmdutils.c:584
options
const OptionDef options[]
OFFSET
it s the only field you need to keep assuming you have a context There is some magic you don t need to care about around this just let it vf default minimum maximum flags name is the option keep it simple and lowercase description are in without and describe what they for example set the foo of the bar offset is the offset of the field in your see the OFFSET() macro
AV_OPT_TYPE_INT
@ AV_OPT_TYPE_INT
Definition: opt.h:235
AV_OPT_TYPE_BOOL
@ AV_OPT_TYPE_BOOL
Definition: opt.h:251