Tune Select and Partition on A100

[gevtushenko]

Description

Partially addresses #238

H100

This PR provides a few improvements (4-5%) for select and partition on H100.
It also improves three-way partition on H100:

H100 HBM3

Entropy	I64	F64
0	-18.12%	-17.48%
0.544	-17.82%	-17.37%
1	-17.42%	-16.90%

H100 HBM2e

Entropy	I64	F64
0	-23.94%	-22.25%
0.544	-20.63%	-21.96%
1	-20.33%	-22.02%

The main focus of this PR is tuning of select and partition on A100.

A100 SXM4

Three-Way Partition

Entropy	I8	I16	I32	I64	I128	F32	F64
0	0.02%	-20.05%	-14.96%	-12.78%	-18.44%	-15.37%	-17.35%
0.544	-0.04%	-19.94%	-15.22%	-12.45%	-18.33%	-15.62%	-16.51%
1	0.01%	-19.72%	-15.31%	-12.26%	-18.38%	-15.65%	-16.49%

Partition::If

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-39.48%	-35.47%	-23.12%	-20.40%	-24.30%	-23.16%	-20.33%
0.544	-35.51%	-31.32%	-22.54%	-21.15%	-24.30%	-22.69%	-20.86%
1	-39.48%	-35.45%	-23.06%	-20.35%	-24.29%	-22.71%	-20.28%

Partition::Flagged

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-33.28%	-28.39%	-14.78%	-21.24%	-21.53%	-14.76%	-21.24%
0.544	-30.66%	-25.00%	-15.17%	-22.62%	-21.13%	-15.14%	-22.63%
1	-33.28%	-28.63%	-14.86%	-21.22%	-21.50%	-14.80%	-21.22%

Select::If

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-46.80%	-39.65%	-41.89%	-42.23%	-47.01%	-41.58%	-42.22%
0.544	-39.13%	-31.71%	-30.17%	-28.33%	-31.21%	-40.76%	-37.80%
1	-39.94%	-32.06%	-21.70%	-20.99%	-21.44%	-21.04%	-20.82%

Select::Flagged

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-32.77%	-34.21%	-22.93%	-36.07%	-44.48%	-22.87%	-36.43%
0.544	-30.02%	-31.23%	-21.23%	-23.43%	-30.03%	-21.16%	-23.45%
1	-31.42%	-30.00%	-14.46%	-16.80%	-22.08%	-14.46%	-16.82%

A100 PCIe

Three-Way Partition

Entropy	I8	I16	I32	I64	I128	F32	F64
0	0.01%	-20.09%	-12.79%	-7.64%	-11.08%	-12.91%	-10.55%
0.544	-0.02%	-17.98%	-13.11%	-8.24%	-11.01%	-13.14%	-10.54%
1	-0.01%	-16.47%	-14.37%	-9.39%	-11.59%	-14.16%	-11.80%

Partition::If

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-38.02%	-31.86%	-18.12%	-12.86%	-14.24%	-18.13%	-12.76%
0.544	-33.25%	-27.55%	-18.24%	-14.22%	-14.94%	-18.37%	-14.01%
1	-37.74%	-31.67%	-18.34%	-13.06%	-14.28%	-17.93%	-12.77%

Partition::Flagged

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-32.21%	-23.91%	-11.92%	-14.51%	-12.63%	-11.83%	-14.56%
0.544	-29.10%	-20.90%	-12.62%	-15.90%	-13.01%	-12.74%	-16.11%
1	-32.17%	-23.74%	-11.79%	-14.34%	-12.56%	-12.14%	-14.42%

Select::If

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-46.97%	-39.52%	-37.61%	-34.74%	-38.57%	-37.74%	-34.68%
0.544	-39.56%	-29.54%	-25.10%	-20.99%	-21.70%	-34.68%	-31.73%
1	-39.08%	-29.03%	-15.97%	-13.54%	-12.40%	-15.22%	-13.60%

Select::Flagged

Entropy	I8	I16	I32	I64	I128	F32	F64
0	-32.37%	-32.54%	-21.85%	-31.53%	-36.33%	-22.14%	-31.70%
0.544	-28.58%	-28.06%	-18.37%	-17.87%	-20.89%	-18.25%	-18.08%
1	-29.50%	-26.29%	-11.65%	-10.77%	-12.81%	-11.52%	-11.00%

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[elstehle]

🚀