[CUDA] Make PTXAS optimisation default to -O3

[hdelan]

Previously the PTX optimization defaulted to -O0.

The ptxjitcompiler defaults to -O3, so this change makes the optimization levels of ahead of time and JIT ptxas compilation the same.

[AGindinson]

The changes LGTM as such; that said, would it be possible to commit the patch directly to LLORG?

[hdelan]

Review requested in LLORG https://reviews.llvm.org/D116583

[hdelan]

@tra in LLORG has said:

I think this would be contrary to the expectation that lack of -O in clang means - do not optimize and it generally implies the whole compilation chain, including assembler. Matching whatever nvidia tools do is an insufficient reason for breaking this assumption, IMO.

I have provided our rationale for changing it in the review, but it seems that this needs a bit more discussion if we want the change to be made in LLORG. Let me know your thoughts. Feel free to give your opinion in https://reviews.llvm.org/D116583

[bader]

@tra in LLORG has said:

I think this would be contrary to the expectation that lack of -O in clang means - do not optimize and it generally implies the whole compilation chain, including assembler. Matching whatever nvidia tools do is an insufficient reason for breaking this assumption, IMO.

I have provided our rationale for changing it in the review, but it seems that this needs a bit more discussion if we want the change to be made in LLORG. Let me know your thoughts. Feel free to give your opinion in https://reviews.llvm.org/D116583

@intel/dpcpp-clang-driver-reviewers, what is your opinion?
I don't understand the issues caused by optimization levels mismatch good enough to give any suggestions.

[mdtoguchi]

@tra in LLORG has said:

I think this would be contrary to the expectation that lack of -O in clang means - do not optimize and it generally implies the whole compilation chain, including assembler. Matching whatever nvidia tools do is an insufficient reason for breaking this assumption, IMO.

I have provided our rationale for changing it in the review, but it seems that this needs a bit more discussion if we want the change to be made in LLORG. Let me know your thoughts. Feel free to give your opinion in https://reviews.llvm.org/D116583

@intel/dpcpp-clang-driver-reviewers, what is your opinion? I don't understand the issues caused by optimization levels mismatch good enough to give any suggestions.

In general, I'm of the opinion that default optimization levels can be set to whatever we feel is proper for our product. Use of any disabling option like -O0 should inherently disable any default optimization levels and even use of a different level like -O1 or -O2 should override defaults as well.

[hdelan]

Use of any disabling option like -O0 should inherently disable any default optimization levels and even use of a different level like -O1 or -O2 should override defaults as well.

This behavior is unchanged by this PR. The only difference is the default level of ptxas when no -O level is specified.

The reason we thought this change would be beneficial is because of a few bugs we encountered in ptxas/ptxjitcompiler for different opt levels. This caused JIT errors but not offline ptxas errors when no opt value was provided. These kind of errors are more common than you'd think, and having different opt levels for ptxjitcompiler and ptxas makes them harder to track down.