Sort improve

[sundy-li]

Related #245

Changes:

• Add to_vec function to MergeSortSlices, this makes iterator reuseable.
• Add buffer_from_range to Index trait
• Fix some bad codes in benches.
• Change cmp(a,b).reverse() to cmp(b,a)

[sundy-li]

Seems #![feature(step_trait)] is not acceptable.

Without this trait, the performance reduces by 16%.

sort-limit 2^14 f32     time:   [53.572 us 53.758 us 53.967 us]                                
                        change: [+15.351% +16.099% +16.909%] (p = 0.00 < 0.05)
                        Performance has regressed.

[jorgecarleitao]

Thanks! Left some comments around.

[sundy-li]

The only way currently I can find to reduce hot path from_usize().unwrap() is to add a method buffer_from_range in trait Index.

[codecov]

Codecov Report

Merging #246 (22d58cf) into main (eea4d33) will decrease coverage by 0.12%.
The diff coverage is 52.32%.

@@            Coverage Diff             @@
##             main     #246      +/-   ##
==========================================
- Coverage   76.81%   76.68%   -0.13%     
==========================================
  Files         229      229              
  Lines       19617    19675      +58     
==========================================
+ Hits        15068    15087      +19     
- Misses       4549     4588      +39     

Impacted Files	Coverage Δ
src/array/specification.rs	41.97% <0.00%> (-30.37%)	⬇️
src/compute/sort/mod.rs	53.55% <0.00%> (ø)
src/compute/sort/utf8.rs	100.00% <ø> (ø)
src/compute/sort/boolean.rs	89.47% <50.00%> (-4.98%)	⬇️
src/buffer/mutable.rs	91.51% <75.00%> (-0.25%)	⬇️
src/compute/merge_sort/mod.rs	93.16% <88.46%> (-0.93%)	⬇️
src/compute/sort/common.rs	92.64% <91.66%> (-0.78%)	⬇️
src/compute/sort/lex_sort.rs	82.14% <100.00%> (+0.16%)	⬆️
src/compute/sort/primitive/indices.rs	100.00% <100.00%> (ø)
src/compute/sort/primitive/sort.rs	94.93% <100.00%> (+0.13%)	⬆️
... and 11 more

---

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 eea4d33...22d58cf. Read the comment docs.

[jorgecarleitao]

I think that this misses the last entries: if the last slice is past limit, imo it should be modified to be equal to limit - current_len, so that we do not lose the last items within the slice

[jorgecarleitao]

I can see that there is some limitations with stable Rust wrt to the ranges and indices. I will take a stab at addressing this here, to see if we can get the remaining performance out of this without the from_usize around.

[jorgecarleitao]

I think that #247 addresses the limitations; could you take a look and see what do you think?

[jorgecarleitao]

I have also created https://users.rust-lang.org/t/idiomatic-way-to-populate-vector-from-branched-loop/63152 to see if we get good ideas.

[sundy-li]

Update:

Add shrink_to_fit() to fix #249

[sundy-li]

I think that #247 addresses the limitations; could you take a look and see what do you think?

Ok, I'll take look at that.

[jorgecarleitao]

I will go ahead and merge this, and clear up the API in a follow-up PR. Thanks a lot, @sundy-li !