Codecov Report

Merging #340 (a9251a9) into master (17bd1ed) will decrease coverage by 0.01%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #340      +/-   ##
==========================================
- Coverage   89.07%   89.06%   -0.02%     
==========================================
  Files          38       38              
  Lines        9275     9275              
==========================================
- Hits         8262     8261       -1     
- Misses       1013     1014       +1     

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 17bd1ed...a9251a9. Read the comment docs.

I think this looks good.
I'd just like to see an @inbounds on the sequential version and updated benchmarks.
If you don't have AVX512, I don't expect to see too much of an improvement from @inbounds because LLVM does not handle the conversion from Int64 to Float64 well.

Benchmarks ran on my i7-4770k, which (I'm pretty sure) has AVX instructions.

By the way, what did you use for generating the GFLOPS vs. size plots (like this one)?

Benchmarks ran on my i7-4770k, which (I'm pretty sure) has AVX instructions.

Yes, AVX, FMA, and AVX2.

By the way, what did you use for generating the GFLOPS vs. size plots (like this one)?

The code is in the benchmark directory. The script to run the benchmarks is in driver.jl
plotbenchmark.jl has the code calling plot and specifying colors to be consistant.

For actually counting flops, I just did that analytically.
Integer subtraction technically isn't a floating point operation, but you could just use 2*length(data) (subtract, divide/multiply) or 3*length(data) (also count the conversion) and call it something like "OPS" (leaving off the "F").

The code running the benchmarks and recording the flops is here:
https://github.com/JuliaSIMD/LoopVectorization.jl/blob/master/benchmark/benchmarkflops.jl
but it is a mess / not something I'd recommend emulating.

More generally, you could measure GFLOPS by removing @turbo and using https://github.com/triscale-innov/GFlops.jl (noting again that it probably won't count things like integer ops).

You're welcome to make another PR to add a GFLOPS plot, but I figured there was no reason to hold this one up.

Thanks for the contribution!