[optimization] Add AVX-512 VPCLMULQDQ accelerated CRC32-C for TXSQL 8.0.30

Summary

This patch introduces an AVX-512 + VPCLMULQDQ accelerated CRC32-C implementation for InnoDB's page checksum and redo log verification, significantly improving throughput on modern x86-64 CPUs like the Hygon C86-4G.

Background & Motivation

In high-concurrency OLTP workloads (e.g., BenchmarkSQL TPC-C), InnoDB spends a non-trivial amount of CPU time computing CRC32-C checksums for:

• 16 KB buffer pool pages
• Variable-length redo log blocks
• Doublewrite buffer pages

The existing crc32_using_pclmul (SSE4.2 + PCLMULQDQ, 128-bit folding) is efficient for small buffers but becomes throughput-bound when processing large contiguous blocks, as it only utilizes one 128-bit lane per fold step.

AVX-512 VPCLMULQDQ extends this to four parallel 128-bit lanes (512-bit ZMM registers), enabling:

• 4× polynomial folding throughput for large buffers (≥256 bytes)
• Dual/quad-stream pipelining to hide latency via register-level parallelism
• Alignment-aware dispatch (VMOVDQA64 vs VMOVDQU64) for InnoDB's naturally aligned 16 KB pages

Technical Details

1. Runtime CPU Detection (can_use_avx512_vpclmul)

Implements the full Intel/AMD specification for AVX-512 feature detection:

• CPUID leaf 1: OSXSAVE[27] + AVX[28] (prerequisites for XGETBV)
• XCR0 state mask: validates OS-managed ZMM register save/restore (0xE6 = SSE|AVX|OPMASK|ZMM_Hi256|Hi16_ZMM)
• CPUID leaf 7: AVX512F[16] + AVX512DQ[17] + AVX512BW[30] + AVX512VL[31] + VPCLMULQDQ[10]

2. Algorithm: 512-bit Polynomial Folding

Large buffer (≥256 bytes): 4-stream folding with b2048 constants
→ 2-stream folding with b1024
→ 1-stream 512-bit loop with b512
→ 512→128 bit reduction with b384
→ 128-bit folding loop with b128
→ 128→64 bit reduction with b64
→ Barrett reduction to 32-bit CRC

Key optimizations:

• vpternlogd ternary logic: replaces 2× XOR + AND sequences with single-uop operations (e.g., AVX512_TL_XOR3, AVX512_TL_XOR2_AND)
• Template-based alignment dispatch: crc32c_avx512_impl<true/false> generates zero-overhead aligned/unaligned load paths at compile time
• _mm256_zeroupper(): eliminates AVX-SSE transition penalty after ZMM execution

3. Compiler Compatibility

Compiler	Minimum Version	Target Attribute
GCC	11.0	pclmul,avx512f,avx512dq,avx512bw,avx512vl,vpclmulqdq
GCC	15.0	pclmul,avx10.1,vpclmulqdq (unified ISA)
Clang	9.0	pclmul,avx512f,avx512dq,avx512bw,avx512vl,vpclmulqdq

MSVC is excluded from __attribute__((target(...))) and would require a separate /arch:AVX512 compilation unit (not included in this patch).

4. Integration Points

• New flag: ut_crc32_avx512_enabled (exposed in ut0crc32.h)
• Priority: AVX-512 path > SSE4.2+PCLMUL > SSE4.2-only > software fallback
• Logging: startup message indicates active path for observability

Performance Impact (Expected)

Based on standalone benchmarks and crc32c_x86.cc reference data:

<style> </style>

缓冲区大小	Software (GB/s)	SSE4.2  (GB/s)	AVX-512 (GB/s)	加速比 (AVX/SW)	加速比 (AVX/SSE)
64 B	1.40	2.35	2.56	1.8×	1.1×
256 B	1.76	4.84	8.45	4.8×	1.7×
512 B	1.80	5.88	15.89	8.8×	2.7×
1 KB	1.80	6.30	23.22	12.9×	3.7×
4 KB	1.83	7.09	40.62	22.2×	5.7×
16 KB	1.83	7.33	45.05	24.6×	6.1×
64 KB	1.83	7.40	46.17	25.2×	6.2×
256 KB	1.83	7.40	46.56	25.5×	6.3×
1 MB	1.82	7.40	45.87	25.2×	6.2×
4 MB	1.83	7.39	45.88	25.1×	6.2×

Note: Actual TXSQL TPC-C improvement depends on checksum CPU fraction; typically 2–5% overall throughput gain at high concurrency.

Compatibility & Risk Assessment

• x86-64 only: ARM64 path unchanged; no impact on ARM deployments
• Graceful degradation: if AVX-512 is unavailable, falls back to existing PCLMUL or SSE4.2 paths
• OS requirement: Linux kernel ≥ 4.0 (for proper XSAVE state management) or any modern distribution
• No ABI changes: ut_crc32_func_t signature unchanged

Patch Scope

storage/innobase/include/ut0crc32.h | 8 ++++++++
storage/innobase/ut/crc32.cc | 647 +++++++++++++++++++++++++-
2 files changed, 653 insertions(+), 2 deletions(-)

Checklist

• Implementation compiles with GCC 11+ / Clang 9+
• Runtime CPU detection covers all AVX-512 sub-features
• Alignment-aware dispatch for InnoDB page boundaries
• VZEROUPPER penalty mitigation
• BenchmarkSQL TPC-C validation on Intel Sapphire Rapids / Hygon C86-4G
• Cross-validation against existing crc32_using_pclmul output (bit-exact)
• Review for MSVC /arch:AVX512 separate compilation unit (future)

0001-.add-avx512_crc32-for-TXSQL8.0.30.patch

References

• Intel: "Fast CRC Computation Using PCLMULQDQ Instruction" (white paper)
• MariaDB: storage/innobase/ut/crc32c_x86.cc (AVX-512 reference implementation)
• Linux kernel: arch/x86/kernel/fpu/xstate.c (XSAVE/AVX-512 state management)

Start by inspecting the existing crc32_using_pclmul implementation and the ut0crc32.h integration point, then compare the crc32c_x86.cc reference data. Implement and integrate the AVX-512 VPCLMULQDQ path with runtime feature detection and fallback priority, and verify checksum and redo-log behavior across supported CPU and compiler paths.