Use radix sort for all fixed-width-types in cudf::sort

[davidwendt]

Description

Refactors internal dispatch function to use cub radix sort for all fixed-width types in cudf::sort
Also adds a benchmark for cudf::sort

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

Performance improvement similar to those for sort_ordered. The integer ones are unchanged since thrust was calling radix sort for us. The float and timestamp now use radix sort as well.

## [0] NVIDIA RTX A6000

|    T    |  stable  |  nulls  |  num_rows  |  cols  |   Ref Time |   Cmp Time |          Diff |   %Diff |
|---------|----------|---------|------------|--------|------------|------------|---------------|---------|
|   I32   |    0     |    0    |   262144   |    1   |  83.939 us |  79.200 us |     -4.740 us |  -5.65% |
|   I32   |    1     |    0    |   262144   |    1   |  84.394 us |  78.708 us |     -5.686 us |  -6.74% |
|   I32   |    0     |    0    |  2097152   |    1   | 218.597 us | 213.289 us |     -5.309 us |  -2.43% |
|   I32   |    1     |    0    |  2097152   |    1   | 218.998 us | 213.994 us |     -5.004 us |  -2.28% |
|   I32   |    0     |    0    |  16777216  |    1   |   1.198 ms |   1.184 ms |    -14.235 us |  -1.19% |
|   I32   |    1     |    0    |  16777216  |    1   |   1.186 ms |   1.184 ms |     -2.359 us |  -0.20% |
|   I32   |    0     |    0    |  67108864  |    1   |   4.537 ms |   4.540 ms |      2.572 us |   0.06% |
|   I32   |    1     |    0    |  67108864  |    1   |   4.534 ms |   4.540 ms |      6.389 us |   0.14% |
|   F32   |    0     |    0    |   262144   |    1   | 207.022 us | 153.256 us |    -53.767 us | -25.97% |
|   F32   |    1     |    0    |   262144   |    1   | 207.423 us | 153.565 us |    -53.857 us | -25.96% |
|   F32   |    0     |    0    |  2097152   |    1   | 961.763 us | 655.231 us |   -306.531 us | -31.87% |
|   F32   |    1     |    0    |  2097152   |    1   | 962.108 us | 655.494 us |   -306.615 us | -31.87% |
|   F32   |    0     |    0    |  16777216  |    1   |   6.700 ms |   4.394 ms |  -2306.049 us | -34.42% |
|   F32   |    1     |    0    |  16777216  |    1   |   6.709 ms |   4.395 ms |  -2314.116 us | -34.49% |
|   F32   |    0     |    0    |  67108864  |    1   |  26.987 ms |  17.252 ms |  -9734.893 us | -36.07% |
|   F32   |    1     |    0    |  67108864  |    1   |  26.956 ms |  17.256 ms |  -9700.540 us | -35.99% |
| time_ms |    0     |    0    |   262144   |    1   | 189.160 us | 128.811 us |    -60.348 us | -31.90% |
| time_ms |    1     |    0    |   262144   |    1   | 186.618 us | 129.041 us |    -57.577 us | -30.85% |
| time_ms |    0     |    0    |  2097152   |    1   |   2.021 ms | 566.269 us |  -1455.118 us | -71.99% |
| time_ms |    1     |    0    |  2097152   |    1   |   2.017 ms | 565.946 us |  -1451.536 us | -71.95% |
| time_ms |    0     |    0    |  16777216  |    1   |  18.572 ms |   3.961 ms | -14610.808 us | -78.67% |
| time_ms |    1     |    0    |  16777216  |    1   |  18.566 ms |   3.962 ms | -14604.384 us | -78.66% |
| time_ms |    0     |    0    |  67108864  |    1   |  83.250 ms |  15.612 ms | -67637.807 us | -81.25% |
| time_ms |    1     |    0    |  67108864  |    1   |  83.240 ms |  15.612 ms | -67627.840 us | -81.24% |

[bdice]

Really nice, so glad to see this. No concerns from my side aside from naming. I'd like everything labeled "faster sort" to be labeled as "radix sort" or a variation thereof. I think it's a lot clearer to readers.

[bdice]

Maybe we can rename this and related files to radix_sort or sort_radix? I think the word "faster" isn't especially helpful to readers.

[davidwendt]

Ok. I'm working on one more follow on PR for this and I'm trying not to collide the names too much.
I like sort_radix too. Mainly because it puts the file closer alphabetically to the other sort source files otherwise the rank files get in between them.

[jrhemstad]

This is cool. I'm impressed at how much faster the radix sort path is even with the extra materialization to initialize the float_pair array and convert it back.

[mhaseeb123]

LGTM. Some minor comments

[ttnghia]

This makes me think that alongside with column and table now we have some "radix" data structure to sort 😃

[davidwendt]

/merge