AM69A: MMA and C71x utilisation

Akhilesh Gangwar

Part Number: AM69A

Hi,

I want to know the resource utilization for am69a hw while running a DL model. I found this link where it prints the same.

https://software-dl.ti.com/jacinto7/esd/processor-sdk-linux-edgeai/AM69A/latest/exports/docs/common/measure_perf.html

I am using perf-stats tools. It prints ARM core processor and C71x, something like this-

Summary of CPU load,
====================
CPU: mpu1_0: TOTAL LOAD = 43.81 % ( HWI =  0.74 %, SWI =  0.24 % )
CPU: c7x_1: TOTAL LOAD = 12. 0 % ( HWI =  00 %, SWI =  0. 0 % )
CPU: c7x_2: TOTAL LOAD = 0. 0 % ( HWI = 0. 0 %, SWI = 0. 0 % )
CPU: c7x_3: TOTAL LOAD = 0. 0 % ( HWI = 0. 0 %, SWI = 0. 0 % )
CPU: c7x_4: TOTAL LOAD = 0. 0 % ( HWI = 0. 0 %, SWI = 0. 0 % )

HWA performance statistics,
===========================


DDR performance statistics,
===========================

DDR: READ BW: AVG =  1455 MB/s, PEAK =  6140 MB/s
DDR: WRITE BW: AVG =   332 MB/s, PEAK =  2138 MB/s
DDR: TOTAL BW: AVG =  1787 MB/s, PEAK =  8278 MB/s

But where do we have results for MMA?

Also, if you can highlight how MMA and C71x are bsing coupled and when C71x is being used and when MMA is being used, that will be great.

Thanks

Akhilesh

over 2 years ago

0 Suman Anna over 2 years ago

TI__Guru*** 125155 points

Hi Akhilesh,

Please expect a response early next week once the assigned engineer is back in office after a TI regional holiday.

regards

Suman

0 Pratik Kedar over 2 years ago in reply to Suman Anna

TI__Mastermind 24041 points

Hi,

Firstly MMA load is not visible as part of per stat tool.

MMA is tightly coupled with C7x, and thus the loading profile of HW Accelerator MMA is not available.

0 Akhilesh Gangwar over 2 years ago in reply to Pratik Kedar

Intellectual 320 points

Thanks Pratik for the reply.

Is there distribution profile available when MMA is being used and when c7x is being used? Or, which operator is being mapped on mma and which on c7x?

Let me know.

Thanks

0 Pratik Kedar over 2 years ago in reply to Akhilesh Gangwar

TI__Mastermind 24041 points

The internal APIs(C7x optimized operators/kernels) usually makes call to the MMA specific operators.

Typically operations like Convolutions, Matrix's Mults ect are being delegated to MMA HW Accelerators, with best of my knowledge we dont have detailed info about which operators of TIDL consumers HW Accelerator available in public domain.

But I will double check the same with team for final confirmation.

0 Akhilesh Gangwar over 2 years ago in reply to Pratik Kedar

Intellectual 320 points

Sure.

Thanks Pratik

0 Pratik Kedar over 2 years ago in reply to Akhilesh Gangwar

TI__Mastermind 24041 points

TI internal,

Putting this thread under WFC

0 Pratik Kedar over 2 years ago in reply to Pratik Kedar

TI__Mastermind 24041 points

Hi Team,

Thanks for the patience.

I have checked internally with the team, we dont have this information disclosed to the customers public domain.

However based on your interest am sharing generic data points.

The C7x Kernel internally calls MMA based APIs, these are typically Convolutions, Matrix Arithmetic functions (Muls).

Thank You

Processors

Processors forum

AM69A: MMA and C71x utilisation