TDA4VM: Improving the processing speed of custom avp4 demo

Daigo Kikuoka

Prodigy 130 points

Part Number: TDA4VM

Tool/software:

My SDK version: RTOS 10.01.00.04

Hi TI experts,

I'm creating a custom app based on "app_tidl_avp4", and I've made the following two changes:

Change display output to SRV
Change TIDL from Seg to OD

This change has caused a performance drop across the graph, it's now around 20fps.

GRAPH: custom_app_graph (#nodes =  11, #executions =  39722)
 NODE:       CAPTURE1:             capture_node: avg =   1622 usecs, min/max =    154 /  61714 usecs, #executions =      39722
 NODE:     VPAC_VISS1:                viss_node: avg =  18505 usecs, min/max =  17735 /  24338 usecs, #executions =      39722
 NODE:         MCU2-0:                aewb_node: avg =    455 usecs, min/max =    154 /  65242 usecs, #executions =      39722
 NODE:      VPAC_LDC1:                 ldc_node: avg =  16601 usecs, min/max =  13899 /  24016 usecs, #executions =      39722
 NODE:      VPAC_MSC2:              scaler_node: avg =  26454 usecs, min/max =  21934 /  33801 usecs, #executions =      39722
 NODE:          DSP-1:              PreProcNode: avg =   3554 usecs, min/max =   3328 /   4447 usecs, #executions =      39722
 NODE:       DSP_C7-1:                tidl_node: avg =  49816 usecs, min/max =  47580 /  54819 usecs, #executions =      39722
 NODE:          DSP-2:    DrawBoxDetectionsNode: avg =  14447 usecs, min/max =   8209 /  19117 usecs, #executions =      39722
 NODE:      VPAC_MSC1:              mosaic_node: avg =  16820 usecs, min/max =   3871 /  84152 usecs, #executions =      39722
 NODE:          MPU-0:          OpenGL_SRV_Node: avg =  15226 usecs, min/max =  11878 /  46502 usecs, #executions =      39722
 NODE:       DISPLAY1:              DisplayNode: avg =   8845 usecs, min/max =    106 /  33199 usecs, #executions =      39729

 PERF:           FILEIO: avg =      0 usecs, min/max = 4294967295 /      0 usecs, #executions =          0
 PERF:            TOTAL: avg =  49885 usecs, min/max =  32556 /  74620 usecs, #executions =      17449

 PERF:            TOTAL:   20. 4 FPS

I thought that by splitting the graphs into two, one for the SRV side and one for the OD side, I could improve performance.(The red and blue areas in the figure below)

But splitting the graph into two did not improve the performance.

GRAPH: custom_app_graph1 (#nodes =   5, #executions =    296)
 NODE:       CAPTURE1:             capture_node: avg =  13394 usecs, min/max =    166 /  59766 usecs, #executions =        296
 NODE:     VPAC_VISS1:                viss_node: avg =  19128 usecs, min/max =  17962 /  20964 usecs, #executions =        296
 NODE:         MCU2-0:                aewb_node: avg =    310 usecs, min/max =    156 /   1371 usecs, #executions =        296
 NODE:          MPU-0:          OpenGL_SRV_Node: avg =  11470 usecs, min/max =  10603 /  44925 usecs, #executions =        296
 NODE:       DISPLAY1:              DisplayNode: avg =   9912 usecs, min/max =     89 /  16342 usecs, #executions =        296

GRAPH: custom_app_graph2 (#nodes =   6, #executions =    148)
 NODE:      VPAC_LDC1:                 ldc_node: avg =  14735 usecs, min/max =  13905 /  15450 usecs, #executions =        148
 NODE:      VPAC_MSC2:              scaler_node: avg =  23128 usecs, min/max =  20792 /  24294 usecs, #executions =        148
 NODE:          DSP-1:              PreProcNode: avg =   3243 usecs, min/max =   3108 /   3427 usecs, #executions =        148
 NODE:       DSP_C7-1:                tidl_node: avg =  45332 usecs, min/max =  41988 /  48336 usecs, #executions =        148
 NODE:          DSP-2:    DrawBoxDetectionsNode: avg =   7428 usecs, min/max =   7023 /  13685 usecs, #executions =        151
 NODE:      VPAC_MSC1:              mosaic_node: avg =   4696 usecs, min/max =   3857 /  29426 usecs, #executions =        151

 PERF:           FILEIO: avg =      0 usecs, min/max = 4294967295 /      0 usecs, #executions =          0
 PERF:            TOTAL: avg =  50167 usecs, min/max =     20 / 188239 usecs, #executions =        304

 PERF:            TOTAL:   19.93 FPS

My goal is to have SRV running at 30fps and OD running at 15fps. Could you please give me some advice how I can achieve this?

Regards,

Daigo

5 months ago

0 Christina Kuruvilla 5 months ago

TI__Expert 5790 points

Hello Daigo,

Could you share your model that you used for your case? Also, could you share all the changes you made with your model (as well as the SRV) so I can recreate it on my end?

Warm regards,

Christina

0 Daigo Kikuoka 5 months ago in reply to Christina Kuruvilla

Prodigy 130 points

Hi Christina,

I use the model included in SDK. In other words, the model I used is ported from "app_tidl_od_cam", so the cfg file says to use:

# location of config
tidl_config   /opt/vision_apps/test_data/psdkra/tidl_models/tidl_io_peele_300_1.bin

# location of network
tidl_network  /opt/vision_apps/test_data/psdkra/tidl_models/tidl_net_peele_300.bin

Please be patient as I will prepare all my changes as a patch.

Regards

Daigo

0 Daigo Kikuoka 5 months ago in reply to Daigo Kikuoka

Prodigy 130 points

Hi Christina,

I attached a patch of my custom avp4 demo. I'm not familiar with OpenVX or how to use graph parameters, so maybe the dual graph scheduling is wrong. What I want to achieve most is to keep SRV at 30fps, and I want detection to run asynchronously at a lower frequency.

Best regards,

Daigo

1780.custom_avp4.zip

0 Christina Kuruvilla 5 months ago in reply to Daigo Kikuoka

TI__Expert 5790 points

Understood Daigo. Thanks for sharing the changes and information. I will investigate and keep you updated as I move forward.

Warm regards,

Christina

0 Gokul S 5 months ago in reply to Christina Kuruvilla

TI__Genius 13490 points

Hi Daigo,

Daigo Kikuoka said:
My goal is to have SRV running at 30fps and OD running at 15fps. Could you please give me some advice how I can achieve this?

I will create a patch for this in avp4 demo and share it to you on or before Thursday.

Regards,
Gokul

0 Daigo Kikuoka 5 months ago in reply to Gokul S

Prodigy 130 points

Hi Gokul,

I hope for a good example of 2 graphs where SRV and detection run asynchronously with each other. I think it will be a great help for my custom app creation.

Thank you for your cooperation.

Regards,

Daigo

0 Gokul S 5 months ago in reply to Daigo Kikuoka

TI__Genius 13490 points

Hi Daigo,

Yes, I understand you usecase but we cannot make the graph run asynchronously, alternate way would be enqueuing alternate frames which would result in 15 fps, but if you want different fps for the second graph then the application should have the logic to enqueue/dequeue the appropriate frames to achieve the fps.

Anyway once you get the patch for 15 fps you would get an idea for the rest.

Regards,
Gokul

0 Daigo Kikuoka 5 months ago in reply to Gokul S

Prodigy 130 points

Hi Gokul,

> but we cannot make the graph run asynchronously

I'm not sure if I understand correctly, but is this a constraint imposed by the enqueue/dequeue logic? I want SRV to maintain at least 30fps regardless of OD performance.

Either way, I'll wait for your update. Thank you for your cooperation.

Regards,

Daigo

+1 Gokul S 5 months ago in reply to Daigo Kikuoka

TI__Genius 13490 points

Hi Daigo,

apply the patch in the vision_apps folder,

https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/791/avp4_5F00_15_5F00_30_5F00_fps.patch

this will run srv at 30 fps and the segmentation model at 15 fps.

Regards,
Gokul

0 Daigo Kikuoka 5 months ago in reply to Gokul S

Prodigy 130 points

Hi Gokul,

Thank you for your patch!

Certainly your patch is working fine and I have a better understanding of the enqueue/dequeue logic of dual graphs.

Based on this, we will begin to implement a custom application. Thank you for your generous support!

Regards,

Daigo

0 Daigo Kikuoka 4 months ago in reply to Gokul S

Prodigy 130 points

Hi Gokul,

Sorry for additional question,

Not directly related to this patch, but why is there no queue depth set for the SRV node?

Regards,

Daigo

0 Gokul S 4 months ago in reply to Daigo Kikuoka

TI__Genius 13490 points

Hi Daigo,

In the demo the output of srv node is not connected so the buffer depth is not set.

if you are using its output to display node you can set the buffer depth by calling this function,

tivxSetNodeParameterNumBufByIndex(obj->srvObj.node, 4, APP_BUFFER_Q_DEPTH);

Even if you do not set buffer depth the framework will set buffer depth to 2 by default.

Regards,
Gokul

0 Daigo Kikuoka 4 months ago in reply to Gokul S

Prodigy 130 points

Hi Gokul,

Thank you for your answer. I understand the reason.

I am a little confused about the pipeline depth.

Q1. How should I design/customize the queue depth? (Number of nodes? Processing time?)

Q2. Is my understanding correct that the deeper the queue, the later the processing?

Regards,

Daigo

0 Gokul S 4 months ago in reply to Daigo Kikuoka

TI__Genius 13490 points

Hi Daigo,

Can you please raise another thread for this.

Regards,
Gokul

0 Daigo Kikuoka 4 months ago in reply to Gokul S

Prodigy 130 points

Hi Gokul,

I created new thread.

(+) TDA4VM: How to adjust graph queue depth - Processors forum - Processors - TI E2E support forums

Reagrds,

daigo