Improve performance of rem_scalar/div_scalar for integer types (4x-10x)

[ritchie46]

Already make this PR to start the discussion. I implemented the strength reduction algorithm discussed in #259 for division of u64 arrays and the performance increases are almost 10x!

I benchmarked a divsion by 4 (which is a single bit shift) and a divison by a prime number.

Gnuplot not found, using plotters backend
divide_scalar 4         time:   [41.871 us 41.883 us 41.899 us]                             
                        change: [-89.393% -89.388% -89.382%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 11 outliers among 100 measurements (11.00%)
  2 (2.00%) high mild
  9 (9.00%) high severe

divide_scalar prime     time:   [71.806 us 71.821 us 71.845 us]                                
                        change: [-82.082% -82.035% -81.993%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 8 outliers among 100 measurements (8.00%)
  1 (1.00%) low mild
  2 (2.00%) high mild
  5 (5.00%) high severe

@jorgecarleitao, I used some unsafe to convince the compiler of the proper types. How do you want to go about this. We could also create some extra traits to do the conversion.

[codecov]

Codecov Report

Merging #275 (7a6f439) into main (4c18c0a) will increase coverage by 0.10%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #275      +/-   ##
==========================================
+ Coverage   77.14%   77.24%   +0.10%     
==========================================
  Files         254      254              
  Lines       20648    20733      +85     
==========================================
+ Hits        15929    16016      +87     
+ Misses       4719     4717       -2     

Impacted Files	Coverage Δ
src/compute/arithmetics/mod.rs	51.87% <ø> (+0.75%)	⬆️
src/compute/arithmetics/basic/div.rs	96.52% <100.00%> (+2.00%)	⬆️
src/compute/arithmetics/basic/rem.rs	96.52% <100.00%> (+2.00%)	⬆️
src/io/ipc/read/read_basic.rs	88.50% <0.00%> (+0.13%)	⬆️
src/compute/nullif.rs	1.81% <0.00%> (+1.81%)	⬆️

---

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 4c18c0a...7a6f439. Read the comment docs.

[jorgecarleitao]

Overall looks great!

I think that for now we should live with the transmute. I think that the new scalar API will enable us to write this function as div_scalar(&dyn Array, &dyn Scalar) -> Box<dyn Array>, which will allow us to avoid this transmute and specialize directly based on DataType.

u16 and u32 is also supported by strength division.

[ritchie46]

I think that for now we should live with the transmute. I think that the new scalar API will enable us to write this function as div_scalar(&dyn Array, &dyn Scalar) -> Box, which will allow us to avoid this transmute and specialize directly based on DataType.

Great! than I will implement the other types as well.

[ritchie46]

@jorgecarleitao I updated the PR. It is now implemented for rem_scalar and div_scalar.

There is a bit repetition in the code. Because branches work on different concrete types it seems not possible to write a generic function for that. A macro would be possible but, but given the usage of unsafe, I chose for repetition as that's more readable.

[jorgecarleitao]

Do we have a test hitting the branch?

[ritchie46]

Do we have a test hitting the branch?

Will write that tomorrow.

[ritchie46]

@jorgecarleitao all branches are ran in tests now. I think its good to go.