TDA4VM: Measuring performance on MMA with TIDL

Hi,

   C7x and MMA are not two separate IP's ( MMA is part of C7x). Can you tell us what exactly you are looking for?

Regards,

Anshu

Hi Anshu,

I've implemented an application by the dl_demo vision apps and tried one of our networks to run.

Despitre the network was measured to Total Giga Macs : 17.4663, it runs on TDA4 on approx. .99FPS.
I thought maybe the utilization had some issues around my configuration, and this is why I was curious about the MMA usage withing C7x.

CPU: mpu1_0: TOTAL LOAD =   6.69 % ( HWI =   0.41 %, SWI =   0.95 % )
CPU: mcu2_0: TOTAL LOAD =  16.66 % ( HWI =   4.93 %, SWI =   1.41 % )
CPU:  c6x_1: TOTAL LOAD =   0.58 % ( HWI =   0. 2 %, SWI =   0. 1 % )
CPU:  c6x_2: TOTAL LOAD =   0.56 % ( HWI =   0. 2 %, SWI =   0. 2 % )
CPU:  c7x_1: TOTAL LOAD =  52.62 % ( HWI =   0.16 %, SWI =   0. 4 % )

Hi Gergely,

     Can you enable the layer level performance traces to see which layer is running slow. Please refer the section Steps to Analyze Performance in below documentation to enable this :

http://downloads.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_02_00_09/ti_dl/docs/user_guide_html/md_tidl_fsg_steps_to_debug_mismatch.html

  Note that you can first try this with TIDL standalone test bench instead of running with Vision apps demo.

Regards,

Anshu

Dear Anshu,

Thank You, I ran it and I don't see any extreme/peaking value there.
Is there any other beast practices or measurement methodologies to figure out why our models are so slow?

BR,
Gergely

Hi Gergely,

     Is the sum of all the layer's cycles matching with what you are observing across the network?

Regards,

Anshu