Adding a "get_corpus" benchmark. #456
Conversation
There was a problem hiding this comment.
double bench(std::string filename, simdjson::padded_string& p) C++17 is value-based semantics oriented. It very much depends on the compiler and its std lib implementation, but it is better to use
double bench(std::string_view , simdjson::padded_string const & ) ;Above is fast and can be called with: std::string, const char * and std::string_view.
|
Copying the filename is trivial in this instance compared to the cost of loading up the file's content. |
There was a problem hiding this comment.
simdjson::get_corpus(filename).swap(p);This is also measuring the performance of the swap(), padded_string method.
This is a bit better
(void)simdjson::get_corpus(filename) ;Although I expect the fastest version would be C11 one, without exceptions. But not much faster than C++17 without exceptions
padded_string get_corpus ( std::errc & /* posix error code */, std::string_view /* flename */ )
That is true. That was a general remark for why simdjson should use |
|
When benchmarking, we have to worry that the compiler might optimize away much of the work. The swap prevents the compiler from getting too clever because the data escapes the function. |
|
this "new no exception" get_corpus has not made much performance gains |
We have a PR that proposes to streamline the get_corpus function, but we have no benchmark. Let us create one.