Iterator of ndarray has poor performance than Iterator of slice

[tuxzz]

I've made a simple performance test

i7-4790 3.60GHz + Windows 10 17763 + rustc 1.31.0-nightly (2bd5993ca 2018-10-02) x86_64-pc-windows-msvc

Test	Time(s)
ndarray_iter	2.684
slice_iter	0.875
slice_loop_safe_good	0.298

i7-4790 3.60GHz + Windows 10 17763 + rustc 1.31.0-nightly (2bd5993ca 2018-10-02) x86_64-pc-windows-msvc + Thin LTO

Test	Time(s)
ndarray_iter	1.310
slice_iter	0.776
slice_loop_safe_good	0.288
ndarray_loop_safe_good	0.133
slice_loop_unsafe_good	0.132

Iterate over a ArrayView is incredible slow than slice.
But for loop on ArrayView as fast as to get_unchecked on slice.
Maybe a bug?

[jturner314]

Thanks for reporting this. Out of curiosity, how does the fn_ndarray_method_2 approach perform?

I haven't taken a look at the assembly, but I have a guess why ndarray_iter is significantly slower than slice_iter. Will you please try rerunning the benchmarks using the iter-nth branch on my fork? (It should only affect the ndarray_iter benchmark.) Add this to your Cargo.toml:

[patch.crates-io]
ndarray = { git = "https://github.com/jturner314/ndarray.git", branch = "iter-nth" }

By the way, I recommend criterion for benchmarking. It makes setting up benchmarks easy, provides statistics, and produces useful charts.

[bluss]

A safe for loop is as good as on a slice? This sounds good. Can you point out which code is running for that slow ndarray_iter benchmark? If it's the iterator that I suspect, with step_by, I'm not surprised we have a performance problem there

[jturner314]

I suspect that this is fixed by #614.

[LukeMathWalker]

I repeated the benchmark using the current master:

Test	Time(s)
NDArray 1d_1d iterator	0.968
Slice 1d_1d iterator	0.816
NDArray 1d_1d high order	0.175
Slice 1d_1d safe loop_bad	0.651
NDArray 1d_1d safe loop_bad	0.819
Slice 1d_1d safe loop_good	0.233
NDArray 1d_1d safe loop_good	0.298
Slice 1d_1d unsafe loop_good	0.069
NDArray 1d_1d unsafe loop_good	0.166

The situation seems to have improved significantly (most likely thanks to #614) - should we close this?

[bluss]

Nice!