Add a single-pass csum_partial_copy_nocheck() for arm64 and KUnit
coverage (plus an optional throughput benchmark) in checksum_kunit.
On ARM Neoverse-V2 (3.0 GHz), this reaches 39.2-45.1 GB/s aligned
(1.8-1.9x) and 36.5-43.8 GB/s unaligned (2.0-2.2x) for >= 1368B buffers
vs generic memcpy() + csum_partial().
Changes in v3:
- Use a 64B carry-tree primitive unrolled 4x to 256B (~21 insns/64B,
pure leaf function, zero spills on GCC 15 and Clang 21).
- Align @src to 16B for len >= 272, stashing the odd-offset bit in
bit 62 of the carry counter.
- Add zero-length, 72/256/272B threshold, and dst[-1]/dst[len]
sentinel checks to checksum_kunit.
Changes in v2:
- Drop the NEON path and separate csum-copy* files, folding the scalar
implementation directly into arch/arm64/lib/csum.c (resolves the
<asm/neon-intrinsics.h> include issue noted by Nathan Chancellor).
- Switch the scalar loop to four __uint128_t accumulators (David Laight).
- Canonicalize 0xffff (-0) to 0x0000 (+0) after csum_fold() in KUnit to
fix x86_64 test failures on zero/all-zero inputs (Jakub Kicinski).
Link: https://lore.kernel.org/r/20261114045500.v2-1-demyansh@gmail.com
Demian Shulhan (2):
arm64: csum: Add fused copy and Internet checksum
lib/tests: checksum: Add KUnit tests for csum_partial_copy_nocheck()
arch/arm64/include/asm/checksum.h | 3 +
arch/arm64/lib/csum.c | 189 +++++++++++++
lib/Kconfig.debug | 10 +
lib/tests/checksum_kunit.c | 453 ++++++++++++++++++++++++++++++
4 files changed, 655 insertions(+)
base-commit: 97d73c8251226cf27cc6656b4c542869bb8c69fa
--
2.53.0