{SYCL][PI][L0] - Eliminate std::string construction/destruction overh…

[kbsmith-intel]

…ead.

This change eliminates std::string construction/destruction overhead
for calls to ZeCall::doCall, and similar entry points. The changes have
no observable functional effect but reduce the overall size of the
built library by 8-9%, and eliminate the memory allocation and
std::string construction ond destruction overhead, thus lowering
overall execution time on every level zero API call that uses the
macros ZE_CALL and ZE_CALL_NOCHECK.

[keryell]

That looks like a good motivation.
But I wonder whether this is not a good use case for some std::string_view...

[AGindinson]

That looks like a good motivation.
But I wonder whether this is not a good use case for some std::string_view...

AFAIU, C++17 features are not available in the current library/runtime build.

[kbsmith-intel]

That looks like a good motivation.
But I wonder whether this is not a good use case for some std::string_view...

AFAIU, C++17 features are not available in the current library/runtime build.

I did a quick check where I used std::string_view instead of the const char * for the params of doCall. This is almost as good with respect to the overhead as the const char *, but not quite as good. In particular, for every usage of the macro ZE_CALL or ZE_CALL_NOCHECK, with string_view, there is extra overhead for construction of the string_view. With const char *, it just loads the address of the string into a register (for each of the two string parameters), while for string_view ir loads both the address of the string, and the strings length. So, this causes a bit more overhead, and this really is unnecessary since all uses in doCall really just use the string itself.

So, my preference is to keep this implementation using const char *.

[keryell]

I did a quick check where I used std::string_view instead of the const char * for the params of doCall. This is almost as good with respect to the overhead as the const char *, but not quite as good. In particular, for every usage of the macro ZE_CALL or ZE_CALL_NOCHECK, with string_view, there is extra overhead for construction of the string_view. With const char *, it just loads the address of the string into a register (for each of the two string parameters), while for string_view ir loads both the address of the string, and the strings length. So, this causes a bit more overhead, and this really is unnecessary since all uses in doCall really just use the string itself.

So, my preference is to keep this implementation using const char *.

Thanks for the feedback after experimenting!
If at the end the L0 API takes a const char * anyway, there is probably no benefit to try to avoid counting the characters in the L0 API by providing the length upfront...