Updated TensorRT to v8.2.1

[stemann]

Awaiting #4347 and #4602

[stemann]

This (and #4604 and #4605) fails due to CUDNN_jll v8.2.1 not having a mapping for CUDA 11.5. @maleadt Should CUDNN v8.2.1 actually have such a mapping?

TensorRT v8.2.1 - v8.2.3 is tested with CUDNN v8.2.1 - and is said to support CUDA 11.5: https://docs.nvidia.com/deeplearning/tensorrt/release-notes/tensorrt-8.html#rel-8-2-1

CUDNN v8.3.2 (and CUDA 11.6) is used for testing TensorRT starting from v8.4.0.

[maleadt]

CUDNN 8.2.1 only supports up to CUDA 11.3, I think. Or at least that was the supported release back when that version of CUDNN was released.

[stemann]

CUDNN 8.2.1 only supports up to CUDA 11.3, I think. Or at least that was the supported release back when that version of CUDNN was released.

Right: https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-824/support-matrix/index.html
v8.2.1 - CUDA 11.3: https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-821/support-matrix/index.html
v8.2.2 added CUDA 11.4: https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-822/support-matrix/index.html
v8.2.3: No archived support matrix.

So really our CUDNN v8.2.1 should not have built an 11.4, but only 11.3 and then TensorRT v8.2+ must use a newer CUDNN for build when building for CUDA 11.4+ platforms.

[ViralBShah]

Superseded by #4604 I suppose.

[stemann]

@ViralBShah I know that this and #4604 seemed like stale PR's, but this PR in particular is the only version to support CUDA 10.2 on aarch64-linux-gnu, so can we please reopen this?

I will just check that #4386 can work with this packaging of TensorRT and then get this, #4604 and #4605 merged.