Optimize sort strings (speedup in 2 - 8 times).

[AlexeyAB]

• For strings length is a multiple of two speedup in 2 - 2.5 times, for multiple of four in 3 - 3.5 times. (Speedup through optimize of comparison of static strings.)
• Fix bug in conversion between little endian and big endian, for speedup sort strings in 8 times for strings with lengthes: 1/2/4/8 bytes. (Speedup through use optimized for gpu radix_sort for fundamental types).

[antonmks]

Excellent, so now it makes sense to round up the possible lengths of varchars to 4 or 8 bytes.

[AlexeyAB]

Yes, you right! The general rule for my optimization - rounding up to a multiple of four the length:

Radix_sort very optimized algorithm on CUDA. Round up must be such as: 3 -> 4 5, 6, 7 -> 8

And CUDA is very sensitive to aligned data, like all SIMD processors! All strings with odd length -> to an even length, or better to length is a multiple of four. An example round up length 25 -> 28 etc.

I wrote a benchmark of different variants of sorting strings.

Results for variant which I added to Alenka, such: https://github.com/AlexeyAB/Benchmarks/tree/687d265ff8b38ed02dc7837b1d29d5825611d6ec/CUDA_sort_strings_bench Comparing speedup and perfomance for strings with different lengthes.

Speedup through round up to fundamental types:

• round up to 4 from: 3(in 9.3 times)
• round up to 8 from: 5(5.8X), 6(5X), 7(7.4X)

Speedup sort strings with length equal to fundamental types:

• 1 bytes in 7 times
• 2 bytes in 14 times
• 4 bytes in 11 times
• 8 bytes in 8.5 times

Speedup through optimize of comparison of static strings:

• length is a multiple of two speedup in 2 - 2.5 times
• length is a multiple of four speedup in 3 - 3.5 times