TDA4VM: Limiting MSMC

Adding more details:

1. you should change the value MSMCSIZE_KB to 2048?

With this change, you do not need to change anything on the board itself for runtime measurements, and only change required is in the CNN model conversion process. Even though the board has 8 MB MSMC memory size, the inference solution will use only 2 MB from the available memory.

2. In order to simulate AM62A runtime, you need to do below changes along with limiting MSMC size to 2 MB

• "advanced_options:high_resolution_optimization":1,
•  "tensor_bits": 16,

With these changes, you can get very close approximation of runtime to be expected in AM62A using tidl-tools version 9.0 (expected at July). The number what you get with these settings can be slightly poor (~20%) compared to what you will get on AM62A, but it should give you good reference and these settings can also be reliably used for relative comparison between 2 models

Thanks

Pramod

Hi Pramod
We changed
"MSMCSIZE_KB            = 7968" to "MSMCSIZE_KB            = 2048" in the device_config.cfg in tidl_tools and did 3 conversion runs (tests):

------------------ Network Compiler Traces -----------------------------
*** stack smashing detected ***: <unknown> terminated
Aborted (core dumped)

------------------ Network Compiler Traces -----------------------------
*** stack smashing detected ***: <unknown> terminated
Aborted (core dumped)

3.
    "tensor_bits": 8,
    "advanced_options:high_resolution_optimization": 0,
successful compilation.

I will attach the log for experiment (2) with debug level 2 in email

Sure please attach the logs or one sample network if you can share.

Thanks,

With Regards,

Pramod

Most likely it is because of network resolution or intermediate tensor size being very high, can you please tell the intermediate tensor size (worst case)?

Pramod

Hi Pramod
Since e2e is open to the general public I cannot share the logs here, I attached them in mail
Also, we tried it on a very simple (one convolution) model and it converted successfully...       due to time-constraints it is not ideal to do trial-error runs until we get a sample network that crashes

Thanks,
 - Yuval

biggest intermediate tensor dimension size is 8847360 (tensor size would be twice that, since it's 16bit 17,694,720)

Next steps:

1. Share some of the logs from the folder of executables (edgeai-tidl-tools). The logs should be with files having extension of .csv, .txt, .log

Thanks

Pramod

Hi Yuval/Hani,

I am sharing a updated ti_cnnperfsim.out (with debug mode and more traces enabled) at below link, you can use this to generate additional logs. Please replace existing file with same in edgeai-tidl-tols with this file

https://tidrive.ext.ti.com/u/3iCgaoKgoGDMIKBn/45342fc8-52ca-4322-a881-88c43f6fc5a9?l

I will be sharing access code via email

Thanks,

with Regards,

Pramod

replied with files via email