Hyperloglog ARM NEON SIMD optimization

[xbasel]

Add ARM NEON optimization for HyperLogLog

• Implement two NEON optmized functions for converting between raw and
  dense representations in HyperLogLog:

  1. hllMergeDenseNEON
  2. hllDenseCompressNEON
     These functions process 16 registers in each iteration.

• Utilize existing SIMD test in hyperloglog.tcl (previously added for
  AVX2 optimization) to validate NEON implementation

Test:
valkey-benchmark -n 1000000 --dbnum 9 -p 21111 PFMERGE z hll1{t} hll2{t}

+-------------------+-----------+-----------+---------------+
|      Metric       |  Before   |   After   | Improvement % |
+-------------------+-----------+-----------+---------------+
| Throughput (k rps)|    7.42   |   76.98   |    937.47%    |
+-------------------+-----------+-----------+---------------+
| Latency (msec)    |           |           |               |
|   avg             |   6.686   |   0.595   |     91.10%    |
|   min             |   0.520   |   0.152   |     70.77%    |
|   p50             |   7.799   |   0.599   |     92.32%    |
|   p95             |   8.039   |   0.767   |     90.46%    |
|   p99             |   8.111   |   0.807   |     90.05%    |
|   max             |   9.263   |   1.463   |     84.21%    |
+-------------------+-----------+-----------+---------------+

Hardware:

CPU: Graviton 3
Architecture:           aarch64
  CPU op-mode(s):       32-bit, 64-bit
  Byte Order:           Little Endian
CPU(s):                 64
  On-line CPU(s) list:  0-63
NUMA:
  NUMA node(s):         1
  NUMA node0 CPU(s):    0-63
Memory: 256 GB

Command stats:
Before:

cmdstat_pfmerge:calls=1000002,usec=126327984,**usec_per_call=126.33**,rejected_calls=0,failed_calls=0

After:

cmdstat_pfmerge:calls=1000002,usec=8588205,**usec_per_call=8.59**,rejected_calls=0,failed_calls=0

Improved by ~14.7x.

Functional testing command:

./runtest --single unit/hyperloglog --only "PFMERGE results with simd"  --loops 10000  --fastfail

The SIMD test randomizes input and comapres scalar vs simd results.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 71.03%. Comparing base (aa88453) to head (4c45315).

Additional details and impacted files

@@             Coverage Diff              @@
##           unstable    #1859      +/-   ##
============================================
- Coverage     71.09%   71.03%   -0.07%     
============================================
  Files           123      123              
  Lines         65671    65671              
============================================
- Hits          46692    46649      -43     
- Misses        18979    19022      +43     

Files with missing lines	Coverage Δ
src/hyperloglog.c	92.23% <100.00%> (ø)

... and 13 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[zuiderkwast]

10 times faster is a pretty good improvement. :)

I didn't read the NEON code carefully because I'm not familiar with it. Is the logic basically the same as the one for AVX2?

[zuiderkwast]

#if defined(HAVE_AVX) || defined(__ARM_NEON)

or we can consider some macro like HLL_HAVE_SIMD to avoid this repetition, also to avoid defining the variable twice.

[xbasel]

10 times faster is a pretty good improvement. :)

I didn't read the NEON code carefully because I'm not familiar with it. Is the logic basically the same as the one for AVX2?

It is similar. NEON vectors are 128 bit, AVX2 is 256 bit. The padding and lookup is a bit different in AVX2.
The execution time of pfmerge ~14.7x faster. The end to end is ~10x faster.

[xbasel]

May crash on ARMv7-A if the address isn't 16-byte aligned.

TODO