Auto-vectorization benchmark

We currently use fastlanes to bitpack wide integers into small integers: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/liquid_array/raw/bit_pack_array.rs

It has some test cases, to get an idea of what it does.

Step 1

But we don't have a benchmark yet. We want to study the performance of from_primitive: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/liquid_array/raw/bit_pack_array.rs#L63, encoding path

to_primitive: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/liquid_array/raw/bit_pack_array.rs#L125

We want to benchmark these two functions:

1. varying the bitwidth -- value range from 2^{1-32}
2. varying the number of integers in the array

Array size of 8192, multiple of 8192.
Datatype of UInt8 -> UInt64

Context: PrimitiveArray<T> is data type from Arrow: https://docs.rs/arrow/latest/arrow/array/struct.PrimitiveArray.html

Probably want study how arrow-rs represent primitive array: https://arrow.apache.org/docs/format/Columnar.html#fixed-size-primitive-layout, and how we represent bit_packed_array: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/liquid_array/raw/bit_pack_array.rs#L8

To benchmark, we prefer https://github.com/bheisler/criterion.rs, it will look like one of those benchmarks: https://github.com/apache/arrow-rs/blob/main/arrow/benches/arithmetic_kernels.rs

After getting benchmark results, we want to know the throughput, e.g., 1GB/s. Both encode and decode.

Step 2

We want to know which functions are been auto-vectorized. We probably need to look at assembly, which can use in this tool: https://github.com/pacak/cargo-show-asm

Bonus

FSST: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/liquid_array/raw/fsst_array.rs
string compression, which we also hope to have auto-vectorization.

cc @jp-reddy

Start with src/liquid_parquet/src/liquid_array/raw/bit_pack_array.rs and the from_primitive and to_primitive paths. Add Criterion benchmarks varying bitwidth, array size, and UInt8–UInt64 types, then use cargo-show-asm to inspect auto-vectorization. Done means benchmark results report encode/decode throughput and identify which functions are auto-vectorized.