TDA4VM: how to send frame to two thread by CaptureNode

Hi,

We've got the issue and escalated to e2e, please expect the response.

Thanks!

Hi,

In order to understand the issue could you please address my below queries:

1. Are you running multiple graphs or a single graph?

2. Could you brief your setup?

3. Are both the threads run on same graph?

Thanks,

Cherry

Hello

First, I run multiple graphs

Second, my setup for a thread is avm, it's as same as /ti-processor-sdk-rtos-j7-evm-07_03_00_05/vision_apps/apps/basic_demos/app_aitronx_basic/app_aitronx_main.c 

the second thread is run deeplearning, I input image to deeplearning Node and get the result. It's already can be work.

Third, two threads run two graph, first graph is avm graph, second is deeplearning graph

Thanks

best regard

Hi,

May we know the source of the file "/ti-processor-sdk-rtos-j7-evm-07_03_00_05/vision_apps/apps/basic_demos/app_aitronx_basic/app_aitronx_main.c"... This is not part of SDK 7.3 release. 
It would be great if we could get this file to see which implementation you are referring.

Is the capture node part of the first graph or second graph?

It would be better if the capture node is kept in a separate graph and the frames are dequeued from this graph and enqueued in avm graph and deeplearning graph.

Currently what is the issue that you are facing when 2 graphs are running dequeue and enqueue functions?

Thanks!

Hello

In my project, the capture node's code is like this

First init

#define NUM_CAPT_INST                       (1U)
capture_cfg->num_cap_channels  = 4

vx_status tsovx_init_capture(vx_context context, CaptureConfig *capture_cfg, CaptureObject *capture_obj)
{
    vx_status status = VX_FAILURE;

	if (NULL != capture_cfg && NULL != capture_obj) {
		if (vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_CAPTURE1))
			status = VX_SUCCESS;
		
		if(VX_SUCCESS == status) {
			tivx_capture_params_init(&capture_obj->capture_params);
			capture_obj->capture_params.numInst = NUM_CAPT_INST;
			capture_obj->capture_params.numCh = (capture_cfg->num_cap_channels * NUM_CAPT_INST);
			uint32_t chIdx = 0U;
			for (uint32_t instIdx = 0U ; instIdx < NUM_CAPT_INST ; instIdx++)
			{
				capture_obj->capture_params.instId[instIdx] = instIdx;
				capture_obj->capture_params.instCfg[instIdx].enableCsiv2p0Support = (uint32_t)vx_true_e;
				capture_obj->capture_params.instCfg[instIdx].numDataLanes         = 4;
				for (uint32_t loop_id = 0U; loop_id < capture_obj->capture_params.instCfg[0U].numDataLanes ; loop_id++)
				{
					capture_obj->capture_params.instCfg[instIdx].dataLanesMap[loop_id] = (loop_id + 1u);
				}
				for (uint32_t loop_id = 0U; loop_id < capture_cfg->num_cap_channels; loop_id++)
				{
					capture_obj->capture_params.chVcNum[chIdx]   = loop_id;
					capture_obj->capture_params.chInstMap[chIdx] = instIdx;
					chIdx++;
				}
			}
		   
			capture_obj->capture_params_obj = vxCreateUserDataObject(context, "tivx_capture_params_t" , sizeof(tivx_capture_params_t), &capture_obj->capture_params);
			status = vxGetStatus((vx_reference)capture_obj->capture_params_obj);
			if(VX_SUCCESS == status) {
				tsovx_init_capture_buf(context, capture_cfg, capture_obj);
			}
		}
	}
	
    return status;
}

then create capture node
the frame is uyvy

vx_status tsovx_create_capture_node(vx_graph graph, CaptureObject *capture_obj, vx_object_array frame)
{
    capture_obj->node = NULL;
    vx_status status = VX_FAILURE;
    if (vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_CAPTURE1)) {
        capture_obj->node = tivxCaptureNode(graph, capture_obj->capture_params_obj, frame);
        vxSetNodeTarget(capture_obj->node, VX_TARGET_STRING, TIVX_TARGET_CAPTURE1);
        status = vxGetStatus((vx_reference)capture_obj->node);
    }
    return status;
}

set the dequeue&&enqueue param is like this

obj->pipeline_dep = 7

void add_graph_parameter_by_node_index(vx_graph graph, vx_node node, vx_uint32 node_parameter_index)
{
    vx_parameter parameter = vxGetParameterByIndex(node, node_parameter_index);

    vxAddParameterToGraph(graph, parameter);
    vxReleaseParameter(&parameter);
}

add_graph_parameter_by_node_index(obj->avm_graph, obj->capture_obj.node, 1);

vx_graph_parameter_queue_params_t graph_parameters_queue_params_list[2];
graph_parameters_queue_params_list[0].graph_parameter_index = 0;
graph_parameters_queue_params_list[0].refs_list_size = obj->num_buf;
graph_parameters_queue_params_list[0].refs_list = (vx_reference*)&obj->frames[0];

vxSetGraphScheduleConfig(obj->avm_graph,
                VX_GRAPH_SCHEDULE_MODE_QUEUE_AUTO,
                1,
                graph_parameters_queue_params_list
                );
tivxSetGraphPipelineDepth(obj->avm_graph, obj->pipeline_dep);

at last, get the frame is by these code

obj->num_buf = 4
vx_status tsovx_avm_apa_run_graph(AVMObject *obj)
{
    uint32_t num_refs, loop_id = 0;
    for(uint32_t buf_id=0; buf_id<obj->num_buf-2; buf_id++) {
        tivxGraphParameterEnqueueReadyRef(obj->avm_graph, 0, (vx_reference*)&obj->frames[buf_id], 1, TIVX_GRAPH_PARAMETER_ENQUEUE_FLAG_PIPEUP);
    }
    vxGraphParameterEnqueueReadyRef(obj->avm_graph, 0, (vx_reference*)&obj->frames[obj->num_buf-2], 1);

    /* After first trigger, start the sensor */
    vx_status status = appRemoteServiceRun(APP_IPC_CPU_MCU2_0,
        APP_REMOTE_SERVICE_SENSOR_NAME,
        APP_REMOTE_SERVICE_SENSOR_CMD_CONFIG_IMX390, &obj->cmdPrms, sizeof(obj->cmdPrms), 0);

    if(VX_SUCCESS!=status) {
        ZF_LOGE("appRemoteServiceRun is error");
        return status;
    }

    /* Need to trigger again since display holds on to a buffer */
    vxGraphParameterEnqueueReadyRef(obj->avm_graph, 0, (vx_reference*)&obj->frames[obj->num_buf-1], 1);
    if (obj->encode_conf.option != 0)
        vxGraphParameterEnqueueReadyRef(obj->avm_graph, 1, (vx_reference*)&obj->encode_obj.bitstream_obj, 1);

    while(1) {
        vx_object_array frame;
        /* Get output reference, waits until a frame is available */
        vxGraphParameterDequeueDoneRef(obj->avm_graph, 0, (vx_reference*)&frame, 1, &num_refs);

        /* Recycles dequeued input and output refs */
        /* input and output can be enqueued in any order */
        vxGraphParameterEnqueueReadyRef(obj->avm_graph, 0, (vx_reference*)&frame, 1);

    }
    /* ensure all graph processing is complete */
    vxWaitGraph(obj->avm_graph);
    return VX_SUCCESS;
}

The CaptureNode is run in first graph

My project need to run two thread, both of them need the capture node's result, I want to use dequeue&&enqueue to get them
Because I think if two thread run dequeue&&enqueue by the same index in vxGraphParameterDequeueDoneRef is not right.
I want to know how to set capture node result to different pipeline or I just need to use dequeue&&enqueue in different thread, and there is no problem, no lost frame.

Thank you

Best regard

Hi,

Please refer the below FAQ which consists of the implementation of single cam demo in 2 graphs where output of capture node is fed to the secondary graph.
https://e2e.ti.com/support/processors/f/791/t/1078116 

There wouldn't be any lost frame in this scenario.

We would suggest to keep the capture node is kept in a separate graph and the frames are dequeued from this graph and enqueued in avm graph and deeplearning graph.

Thanks.

Hello,

I see this FAQ, but it is created by RTOS 8.1. For company reason, I have to use RTOS 7.3, so is it can be used in this version?

And it is single camera, but I need to use 4 camera, is there any special thing I should to focus on?

Thanks

ok, I will try to run this one

Thank you, it's ok now