UIF 1.2 release

[varunsh-xilinx]

Summary of Changes

• Update VAI libraries to 3.5 versions
• Update ROCm to 5.6.1
• Fix links in README
• Add better support for V70* (see notes)

Motivation

VAI 3.5 is the latest release and this PR updates the tools to the same version. Similarly, ROCm 5.6.1 is the latest. These tools correspond to the UIF 1.2 release.

Implementation

I updated the vcpkg libraries to point to the latest SHAs.

Initially, we were observing some errors with V70 and running the default resnet50 example. I traced the problem to a few issues related to tensor size:

• VART is using the shape of the input tensor and asserting against it. Before, we were sending a flat shape which was failing the assertion. That's been addressed with an optional shape parameter when constructing the request in Python to set the shape explicitly.
• The XModel backend was also creating tensors with a leading 1 for the batch size which was failing assertions in VART. That's been modified to be zero and then the memory pool omits the external batch size when its value is zero. Then, the first index is correctly getting set to the batch size.

Notes

The container should be rebuilt to reflect the updated XRT package if Vitis is enabled. The host XRT version should also be updated to match v2.14.418.

*The V70 fix is removed from this PR and will be added in a patch release 0.4.1 for consistency with the released images for UIF 1.2.

[gbuildx]

Build failed!

[gbuildx]

Build failed!

[gbuildx]

Build failed!

[gbuildx]

Build failed!

[varunsh-xilinx]

The CI pipeline needs to be updated with these tool versions.