TDA4VM: OpenVX nodes to multiply float matrices

FredC_LT

Part Number: TDA4VM

Hi,

I'm currently trying to use OpenVX nodes to accelerate the element-wise multiplication/addition/substraction of two float matrices on the TDA4.

My first thought was to use the Pixel-wise Multiplication node, but then I found out that float is not a supported type for vx_image.

After digging a bit, I found out about vx_tensor and the tensor multiply node of OpenVX 1.2, but TIOVX only supports OpenVX 1.1 (even though vx_tensors are used for the TIDL node).

So I'm currently thinking that I need to implement my own vx nodes that perform element-wise multiplication/addition of vx_tensors/float matrices. Before that, I wanted to verify:

1. What is TI's suggested method to multiply float/double matrices using the provided TIOVX API?

2. Is support for OpenVX 1.2 for the TDA4 on the roadmap?

Thank you,

Fred

over 3 years ago

+1 Lucas Weaver over 3 years ago

TI__Genius 13845 points

Hi Fred,

For #1, we don't have this included as a default kernel, but this can be added as a new custom target kernel using the below documentation as a reference for creating target kernels.

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tiovx/docs/user_guide/TIOVX_TARGET_KERNEL.html

For #2, no there is not a plan for this is the roadmap, although we do include the vx_tensor as a part of our offering which was only introduced in 1.2.

Regards,

Lucas

0 FredC_LT over 3 years ago in reply to Lucas Weaver

Intellectual 796 points

Thank you for the quick answer and reference.

Processors

Processors forum

TDA4VM: OpenVX nodes to multiply float matrices