Improve crc32c() performance on lengths >= 512 bytes by using
crc32_lsb_vpclmul_avx512() instead of crc32c_x86_3way(), when the CPU
supports VPCLMULQDQ and has a "good" implementation of AVX-512. For now
that means AMD Zen 4 and later, and Intel Sapphire Rapids and later.
Pass crc32_lsb_vpclmul_avx512() the table of constants needed to make it
use the CRC-32C polynomial.
Rationale: VPCLMULQDQ performance has improved on newer CPUs, making
crc32_lsb_vpclmul_avx512() faster than crc32c_x86_3way(), even though
crc32_lsb_vpclmul_avx512() is designed for generic 32-bit CRCs and does
not utilize x86_64's dedicated CRC-32C instructions.
Performance results for len=4096 using crc_kunit:
CPU Before (MB/s) After (MB/s)
====================== ============= ============
AMD Zen 4 (Genoa) 19868 28618
AMD Zen 5 (Ryzen AI 9 365) 24080 46940
AMD Zen 5 (Turin) 29566 58468
Intel Sapphire Rapids 22340 73794
Intel Emerald Rapids 24696 78666
Performance results for len=512 using crc_kunit:
CPU Before (MB/s) After (MB/s)
====================== ============= ============
AMD Zen 4 (Genoa) 7251 7758
AMD Zen 5 (Ryzen AI 9 365) 17481 19135
AMD Zen 5 (Turin) 21332 25424
Intel Sapphire Rapids 18886 29312
Intel Emerald Rapids 19675 29045
That being said, in the above benchmarks the ZMM registers are "warm",
so they don't quite tell the whole story. While significantly improved
from older Intel CPUs, Intel still has ~2000 ns of ZMM warm-up time
where 512-bit instructions execute 4 times more slowly than they
normally do. In contrast, AMD does better and has virtually zero ZMM
warm-up time (at most ~60 ns). Thus, while this change is always
beneficial on AMD, strictly speaking there are cases in which it is not
beneficial on Intel, e.g. a small number of 512-byte messages with
"cold" ZMM registers. But typically, it is beneficial even on Intel.
Note that on AMD Zen 3--5, crc32c() performance could be further
improved with implementations that interleave crc32q and VPCLMULQDQ
instructions. Unfortunately, it appears that a different such
implementation would be optimal on *each* of these microarchitectures.
Such improvements are left for future work. This commit just improves
the way that we choose the implementations we already have.
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20250719224938.126512-3-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Reorganize the crc-pclmul static_call initialization to place more of
the logic in the *_mod_init_arch() functions instead of in the
INIT_CRC_PCLMUL macro. This provides the flexibility to do more than a
single static_call update for each CPU feature check. Right away,
optimize crc64_mod_init_arch() to check the CPU features just once
instead of twice, doing both the crc64_msb and crc64_lsb static_call
updates together. A later commit will also use this to initialize an
additional static_key when crc32_lsb_vpclmul_avx512() is enabled.
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20250719224938.126512-2-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Move the x86-optimized CRC code from arch/x86/lib/crc* into its new
location in lib/crc/x86/, and wire it up in the new way. This new way
of organizing the CRC code eliminates the need to artificially split the
code for each CRC variant into separate arch and generic modules,
enabling better inlining and dead code elimination. For more details,
see "lib/crc: Prepare for arch-optimized code in subdirs of lib/crc/".
Reviewed-by: "Martin K. Petersen" <martin.petersen@oracle.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Acked-by: "Jason A. Donenfeld" <Jason@zx2c4.com>
Link: https://lore.kernel.org/r/20250607200454.73587-12-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
