Post-merge review of #1419 "Universal tensor transform feature" and #1481...

[atamazov]

UPDATE: Leftovers:

• [Tensor reorder] Universal tensor transform feature, a fallback of batched transpose kernel #1419 (comment)
• [Tensor reorder] Universal tensor transform feature, a fallback of batched transpose kernel #1419 (comment)

@junliume Currently this is of normal_urgency

---

The 1st version of this ticket (outdated)

All review comments must be resolved (the most important ones are marked with 🔴).

• [Tensor reorder] Universal tensor transform feature, a fallback of batched transpose kernel #1419 (review)
• [Tensor reorder][Quality][#issue 1476] Improve naming style and CTest design #1481 (review)

/cc @aska-0096 @shaojiewang @DrizztDoUrden

@junliume Currently this is of high_urgency

[junliume]

@aska-0096 could you follow up with Artem's comments in #1419 asap?

[aska-0096]

@atamazov @junliume I send the PR #1481 to cover these review suggestions, hope it works.

[DrizztDoUrden]

Is there any specific reason for not basing this on TransposingSolver?https://github.com/ROCmSoftwarePlatform/MIOpen/blob/0c6bdeb3ff46505d65f69a8c6e72ec0eef29835a/src/include/miopen/utility/transposing_solver.hpp#L411
That removes a whole lot of duplications and doesn't require any modification to the transposed solver.
Example of usage: https://github.com/ROCmSoftwarePlatform/MIOpen/blob/0c6bdeb3ff46505d65f69a8c6e72ec0eef29835a/src/include/miopen/pooling/solvers.hpp#L140
If your kernel is universal, just replace this pseudo-solver: https://github.com/ROCmSoftwarePlatform/MIOpen/blob/0c6bdeb3ff46505d65f69a8c6e72ec0eef29835a/src/include/miopen/utility/transposing_solver.hpp#L164

[junliume]

@aska-0096 could you take a look and respond to @DrizztDoUrden 's inquiries above? Thanks!

[aska-0096]

@DrizztDoUrden Sorry for late reply. I'd like to give three main advantage in tensor reorder solver.

Performance:

1. In tensor reorder method, we re-utilize the exist batched-transpose kernel for higher performance. You can refer to performance data in PR# 1419.
2. All Tensor reorder kernel is compiled in compile-time, and chosen via specific reorder sequence(i.e. Reorder NCHW to NHWC is (0,2,3,1)). This kind of design will decrease the run-time workload.

Completeness:
1.We support various of datatype, int8, fp16, fp32 and fp64.

[junliume]

Can we close this one since #1481 is merged?

[atamazov]

@junliume This can be closed after resolving all comments in

• [Tensor reorder] Universal tensor transform feature, a fallback of batched transpose kernel #1419 (review)

[atamazov]

7 comments from #1419 (review) are not resolved yet; couple of them are red.

[atamazov]

Description updated with #1481 (review).

[aska-0096]

Updated in #1515. @atamazov

[atamazov]

Description updated. @junliume Now this is of normal_urgency.