TDA4VH-Q1: TIDL initialization or process exceptionon when use mixed precision on 8.6 SDK

GUANG REN

Intellectual 350 points

Part Number: TDA4VH-Q1

Hi, All:

Errors often occur when using mixed-precision features for multi-branched networks.

When I don't use OTF optimization ，the error info is "OUT_FEAT_MAP READ MSMC not enough in tidl_conv2d_mma"

and When I use OTF optimization, the error info is "MMALIB_CNN_convolve_row_ixX_ixX_oxX_init: FAILED

The sdk viersion is ti-processor-sdk-rtos-j784s4-evm-08_06_00_14

over 2 years ago

0 Varun Tripathi over 2 years ago

TI__Genius 10155 points

Hi,
Can you let me know what the parameters for the layer 42 are? (Kernel Width, Height, .. etc.)
Additionally is this the layer marked as 16-bit by you?

Regards,
Varun

0 GUANG REN over 2 years ago in reply to Varun Tripathi

Intellectual 350 points

the parameters info for the layer 42 is this (kernel shape(1,1) stride (1,1) pad (0,0))

and I set the concat layer as 16 bit . The red box is the layer 42

0 Varun Tripathi over 2 years ago in reply to GUANG REN

TI__Genius 10155 points

Hi,
Is it possible for you to share a dummy model which reflects this issue? I tried creating a unit test case with this configuration but I didn't see the same issue.

Regards,
Varun

0 Tony Li over 2 years ago in reply to Varun Tripathi

TI__Intellectual 2250 points

Hi, Varun

Thank you for your help! I send the associated material to you via e-mail. You may try to reproduce this issue by referring to the attachment!

Best Regards

Tony

Processors

Processors forum

TDA4VH-Q1: TIDL initialization or process exceptionon when use mixed precision on 8.6 SDK