generic: sycl: resampling: Avoid using the cudnn resampling to use the sycl impl

[kala855]

Description

We think that the problem reported in issue #1728 is because of the accuracy of the generated grid. To make it work we should make the grid double precision regardless of the input/output types. Otherwise, the grid affects the accuracy of the results too much. Effectively the grid is specified in the range [-1,1], but most of the bits in these values are used for encoding which values to use to calculate the result instead of what we need - how much of these values to use.

To sum up, this approach doesn’t quite fit our needs and leads to poor precision or performance. We think we should consider using the SYCL implementation instead. In this PR we removed the cudnn resampling calls and rely on using the SYCL kernels instead.

Fixes #1728

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

[dzarukin]

The suggestion is to introduce a generic resampling kernel first before purging this implementation as it will be a functionality regression if the generic kernel won't make it in v3.6.

[kala855]

The suggestion is to introduce a generic resampling kernel first before purging this implementation as it will be a functionality regression if the generic kernel won't make it in v3.6.

The SYCL implementation will be used as the generic kernel in these cases. The cudnn kernel covered 177 tests (of which 56 were failing) of a total of 12551 benchdnn tests. These cases are also passed by the SYCL implementation. I may be missing something else here. Let me know if you think something different needs to be handled here. Thank you.

[dzarukin]

The suggestion is to introduce a generic resampling kernel first before purging this implementation as it will be a functionality regression if the generic kernel won't make it in v3.6.

The SYCL implementation will be used as the generic kernel in these cases. The cudnn kernel covered 177 tests (of which 56 were failing) of a total of 12551 benchdnn tests. These cases are also passed by the SYCL implementation. I may be missing something else here. Let me know if you think something different needs to be handled here. Thank you.

Sorry, I missed the fact that resampling is already there. No concerns then.

[kala855]

The suggestion is to introduce a generic resampling kernel first before purging this implementation as it will be a functionality regression if the generic kernel won't make it in v3.6.

The SYCL implementation will be used as the generic kernel in these cases. The cudnn kernel covered 177 tests (of which 56 were failing) of a total of 12551 benchdnn tests. These cases are also passed by the SYCL implementation. I may be missing something else here. Let me know if you think something different needs to be handled here. Thank you.

Sorry, I missed the fact that resampling is already there. No concerns then.

@dzarukin If no concerns. And if there is nothing else to take into account. Is it possible to accept this PR? Thank you.