Benchmark custom getline implemention

[springmeyer]

@artemp added a new getline impl in be437eb that is able to cope with embedded newlines in csv lines. To feel comfortable going forward with this we need to benchmark compared to std::getline to ensure performance is not terrible. Our getline will do slightly more work so as long as it is not > 10% slower we should be okay.

[springmeyer]

Slower, but not by much:

$ ./benchmark/out/test_getline --iterations 10000000 --line "one          
two"
std::getline:                              t:0 i:10000000     285 milliseconds
csv_utils::getline_csv:                    t:0 i:10000000     320 milliseconds

[springmeyer]

Sometimes faster when multithreaded:

$ ./benchmark/run
Inheriting from mapnik-settings.env
std::getline:                              t:0 i:10000000     293 milliseconds
csv_utils::getline_csv:                    t:0 i:10000000     338 milliseconds
std::getline:                              t:20 i:500000      488 milliseconds
csv_utils::getline_csv:                    t:20 i:500000      371 milliseconds

[springmeyer]

Ugh: profiling shows that most time is being taken in locale locking due to the stream.widen called per loop (

mapnik/benchmark/test_getline.cpp

Line 47 in da054c8

std::getline(s,csv_line,s.widen(newline));

). So we're not actually isolating getline yet...

[springmeyer]

Before:

std::getline:                              t:20 i:500000      488 milliseconds
csv_utils::getline_csv:                    t:20 i:500000      371 milliseconds

After 336170c:

$ ./benchmark/out/test_getline --iterations 500000 --threads 20
std::getline:                              t:20 i:500000       32 milliseconds
csv_utils::getline_csv:                    t:20 i:500000       61 milliseconds

[springmeyer]

I'm comfortable with perf for now. I've modified the tests so that they pass with the new getline implementation. If getline shows up in profiling as a bottleneck in the future then we can optimize it perhaps by avoiding the sentry (which is currently the bottleneck) and using C-style code that does not use std::ios: http://stackoverflow.com/a/8123198

[springmeyer]

for the record, sentry and clear appear to be the bottlenecks in our getline impl currently:

[springmeyer]

hmm, github image uploads are now working right now...

[artemp]

@springmeyer - thanks for the bench 👍