Improve inner_product performance

[boostorg]

Improve inner_product performance

[ddemidov]

I think PERF_N (1024) is too small to provide an adequate performance comparison for the inner product. For the vectors this small the performance is dominated by OpenCL overheads. Here the boost from OpenCL/CUDA was only apparent starting from vectors longer than 10e5-10e6.

[kylelutz]

Yeah, performance shouldn't be judged with such small data sizes. @roshanr95 At what data sizes did you see the performance differences?

I'll look into bumping up the default number. In my usual testing I usually set PERF_N to 16000000 (by specifying it on the command line). Also @roshanr95, I assume you found the perf.py script which will run the algorithm with a range of data sizes and also display a graph if you run it with the --plot-time or --plot-rate arguments.

[roshanr95]

Yeah, I used the python script. The performance differences are for size of 33554432 elements.

[kylelutz]

Fixed. See issue #73 and the discussion in issue #69.