* Re: [PATCH v2] lib/crc: arm64: Add NEON PMULL+EOR3 crc64_nvme implementation
2026-11-14 17:29 [PATCH v2] lib/crc: arm64: Add NEON PMULL+EOR3 crc64_nvme implementation Shreesh Adiga
@ 2026-11-17 0:01 ` Eric Biggers
2026-11-17 0:45 ` Demian Shulhan
1 sibling, 0 replies; 3+ messages in thread
From: Eric Biggers @ 2026-11-17 0:01 UTC (permalink / raw)
To: Shreesh Adiga; +Cc: Ard Biesheuvel, Demian Shulhan, linux-crypto, linux-kernel
On Sat, Nov 14, 2026 at 10:59:51PM +0530, Shreesh Adiga wrote:
> Add a 4x-folding implementation for CPUs with FEAT_SHA3 which benefits
> newer ARMv9 CPUs like Cortex-X925.
> Usage of EOR3 from SHA3 extension improves performance by 20% compared
> to the regular PMULL without EOR3 on Cortex-X925 measure in userspace
> application via libbenchmark for 512 byte input size.
> Measured about 3x speedup for 512 byte input compared with the existing
> NEON single fold implementation on Cortex-X925 in userspace.
Was crc_kunit run? It already has a benchmark in it.
The commit message should also clarify that this benefits ARMv8.2 and
later processors with the sha3 extension, not just ARMv9. (E.g., Apple
M1 is an example of an ARMv8.4 processor that benefits.)
> +#ifdef CONFIG_ARM64
> +#define SHA3_TARGET __attribute__((target("arch=armv8.2-a+crypto+sha3")))
Doesn't this break the build with gcc 8.1? The kernel supports gcc 8.1
and later, but support for the sha3 target and intrinsics were added in
gcc 9.1. Either this code will need to check the gcc version, or inline
assembly will need to be used for eor3 like lib/raid/xor/ does.
> + v0 = fold_single_round(v0, v2, fold_consts_2x);
> + v1 = fold_single_round(v1, v3, fold_consts_2x);
> + v0 = fold_single_round(v0, v1, fold_consts);
> +
> + while (len >= 16) {
> + uint64x2_t t0 = vreinterpretq_u64_u8(vld1q_u8(p));
> +
> + v0 = fold_single_round(v0, t0, fold_consts);
> + p += 16;
> + len -= 16;
> + }
> +
> + /* Multiply the 128-bit value by x^64 and reduce it back to 128 bits. */
> + v0 = vextq_u64(v0, zero, 1) ^ pmull64_hi_lo(fold_consts, v0);
> +
> + /* Final Barrett reduction */
> + uint64x2_t bconsts = vld1q_u64(bconsts_val);
> + uint64x2_t final = pmull64(bconsts, v0);
> +
> + v0 = veor3q_u64(v0, vextq_u64(zero, final, 1),
> + pmull64_hi_lo(bconsts, final));
> +
> + return vgetq_lane_u64(v0, 1);
Note that this is missing several optimizations compared to the x86
version (lib/crc/x86/crc-pclmul-template.S), including using
fold_consts_2x for a tail of 'len & 32', efficiently handling lengths
that aren't a multiple of 16 bytes, and reducing to the final CRC in two
carryless multiplication instructions rather than three.
This is also just one of several CRC variants. This would have to be
done all over again for each variant if they are not going to be
generated by a template like x86 and riscv do.
So we can do this for now, but it isn't a very scalable solution.
- Eric
^ permalink raw reply [flat|nested] 3+ messages in thread* Re: [PATCH v2] lib/crc: arm64: Add NEON PMULL+EOR3 crc64_nvme implementation
2026-11-14 17:29 [PATCH v2] lib/crc: arm64: Add NEON PMULL+EOR3 crc64_nvme implementation Shreesh Adiga
2026-11-17 0:01 ` Eric Biggers
@ 2026-11-17 0:45 ` Demian Shulhan
1 sibling, 0 replies; 3+ messages in thread
From: Demian Shulhan @ 2026-11-17 0:45 UTC (permalink / raw)
To: Shreesh Adiga; +Cc: Eric Biggers, Ard Biesheuvel, linux-crypto, linux-kernel
Hi Shreesh,
> +#define SHA3_TARGET __attribute__((target("arch=armv8.2-a+crypto+sha3")))
> +
> +SHA3_TARGET
> +static inline uint64x2_t fold_single_round(uint64x2_t fold,
> + uint64x2_t temp,
> + uint64x2_t consts)
> +{
> + uint64x2_t temp1 = pmull64(consts, fold);
> + uint64x2_t temp2 = pmull64_high(consts, fold);
> +
> + return veor3q_u64(temp, temp1, temp2);
> +}
> +
> +SHA3_TARGET
> +u64 crc64_nvme_neon_pmull_eor3(u64 crc, const u8 *p, size_t len)
> +{
> + /* Assumes len >= 64. */
> + uint64x2_t fold_consts = vld1q_u64(fold_consts_val);
> + uint64x2_t fold_consts_4x = vld1q_u64(fold_consts_4x_val);
> + uint64x2_t fold_consts_2x = vld1q_u64(fold_consts_2x_val);
Scalar vmull_p64() makes compilers materialize the constants via mov/movk + fmov
(42/97 insns on Clang). SHA3_TARGET also makes codegen fragile: with
-march=armv8.5-a+crypto, GCC stops inlining pmull64*(). On Neoverse-V2, EOR3
contends with PMULL on pipe V0, whereas regular EOR runs on all SIMD pipes.
Using inline-asm pmull/pmull2 and a clean 4x unroll in crc64_nvme_neon() without
EOR3 should give better performance across all sizes and cleaner code.
/* clang -O2 */
24: mov x10, #0x45a
28: movk x10, #0xace5, lsl #16
2c: movk x10, #0x2240, lsl #32
30: movk x10, #0x6224, lsl #48
34: dup v4.2d, x10
38: mov x10, #0xa84a
3c: movk x10, #0x1e18, lsl #16
40: movk x10, #0xcdb3, lsl #32
44: movk x10, #0xc32, lsl #48
48: fmov d5, x10
Thanks,
Demian
сб, 14 лист. 2026 р. о 18:30 Shreesh Adiga <16567adigashreesh@gmail.com> пише:
>
> Add a 4x-folding implementation for CPUs with FEAT_SHA3 which benefits
> newer ARMv9 CPUs like Cortex-X925.
> Usage of EOR3 from SHA3 extension improves performance by 20% compared
> to the regular PMULL without EOR3 on Cortex-X925 measure in userspace
> application via libbenchmark for 512 byte input size.
> Measured about 3x speedup for 512 byte input compared with the existing
> NEON single fold implementation on Cortex-X925 in userspace.
>
> Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
> ---
> Changes in v2:
> Improved the 4x to 1x reduction by using x^255 and x^319 x2 reduction.
> Added static keys for the CPU feature checks instead of regular
> branches.
>
> lib/crc/arm64/crc64.h | 24 +++++++++++--
> lib/crc/crc64-neon.c | 83 +++++++++++++++++++++++++++++++++++++++++++
> 2 files changed, 104 insertions(+), 3 deletions(-)
>
> diff --git a/lib/crc/arm64/crc64.h b/lib/crc/arm64/crc64.h
> index c7a69e1f3d8f..e5be0ef21140 100644
> --- a/lib/crc/arm64/crc64.h
> +++ b/lib/crc/arm64/crc64.h
> @@ -9,20 +9,38 @@
> #include <linux/sizes.h>
>
> u64 crc64_nvme_neon(u64 crc, const u8 *p, size_t len);
> +u64 crc64_nvme_neon_pmull_eor3(u64 crc, const u8 *p, size_t len);
> +
> +static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_eor3);
> +static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_pmull);
>
> #define crc64_be_arch crc64_be_generic
>
> static inline u64 crc64_nvme_arch(u64 crc, const u8 *p, size_t len)
> {
> - if (len >= 128 && cpu_have_named_feature(PMULL) &&
> + if (len >= 128 && static_branch_likely(&have_pmull) &&
> likely(may_use_simd())) {
> size_t chunk = len & ~15;
>
> - scoped_ksimd()
> - crc = crc64_nvme_neon(crc, p, chunk);
> + scoped_ksimd() {
> + if (static_branch_likely(&have_eor3))
> + crc = crc64_nvme_neon_pmull_eor3(crc, p, chunk);
> + else
> + crc = crc64_nvme_neon(crc, p, chunk);
> + }
>
> p += chunk;
> len &= 15;
> }
> return crc64_nvme_generic(crc, p, len);
> }
> +
> +#define crc64_mod_init_arch crc64_mod_init_arch
> +static void crc64_mod_init_arch(void)
> +{
> + if (cpu_have_named_feature(PMULL)) {
> + static_branch_enable(&have_pmull);
> + if (cpu_have_named_feature(SHA3))
> + static_branch_enable(&have_eor3);
> + }
> +}
> diff --git a/lib/crc/crc64-neon.c b/lib/crc/crc64-neon.c
> index 4753fb94a4be..767e55edc0e0 100644
> --- a/lib/crc/crc64-neon.c
> +++ b/lib/crc/crc64-neon.c
> @@ -10,6 +10,17 @@
>
> u64 crc64_nvme_neon(u64 crc, const u8 *p, size_t len);
>
> +#ifdef CONFIG_ARM64
> +u64 crc64_nvme_neon_pmull_eor3(u64 crc, const u8 *p, size_t len);
> +
> +/* x^575 mod G, x^511 mod G */
> +static const u64 fold_consts_4x_val[2] = { 0x0c32cdb31e18a84aULL,
> + 0x62242240ace5045aULL };
> +/* x^319 mod G, x^255 mod G */
> +static const u64 fold_consts_2x_val[2] = { 0xb0bc2e589204f500ULL,
> + 0xe1e0bb9d45d7a44cULL };
> +#endif
> +
> /* x^191 mod G, x^127 mod G */
> static const u64 fold_consts_val[2] = { 0xeadc41fd2ba3d420ULL,
> 0x21e9761e252621acULL };
> @@ -45,3 +56,75 @@ u64 crc64_nvme_neon(u64 crc, const u8 *p, size_t len)
>
> return vgetq_lane_u64(v0, 1);
> }
> +
> +#ifdef CONFIG_ARM64
> +#define SHA3_TARGET __attribute__((target("arch=armv8.2-a+crypto+sha3")))
> +
> +SHA3_TARGET
> +static inline uint64x2_t fold_single_round(uint64x2_t fold,
> + uint64x2_t temp,
> + uint64x2_t consts)
> +{
> + uint64x2_t temp1 = pmull64(consts, fold);
> + uint64x2_t temp2 = pmull64_high(consts, fold);
> +
> + return veor3q_u64(temp, temp1, temp2);
> +}
> +
> +SHA3_TARGET
> +u64 crc64_nvme_neon_pmull_eor3(u64 crc, const u8 *p, size_t len)
> +{
> + /* Assumes len >= 64. */
> + uint64x2_t fold_consts = vld1q_u64(fold_consts_val);
> + uint64x2_t fold_consts_4x = vld1q_u64(fold_consts_4x_val);
> + uint64x2_t fold_consts_2x = vld1q_u64(fold_consts_2x_val);
> + uint64x2_t v0 = { crc, 0 };
> + uint64x2_t zero = { };
> + uint64x2_t v1, v2, v3;
> +
> + v0 ^= vreinterpretq_u64_u8(vld1q_u8(p));
> + v1 = vreinterpretq_u64_u8(vld1q_u8(p + 16));
> + v2 = vreinterpretq_u64_u8(vld1q_u8(p + 32));
> + v3 = vreinterpretq_u64_u8(vld1q_u8(p + 48));
> +
> + p += 64;
> + len -= 64;
> + while (len >= 64) {
> + uint64x2_t t0 = vreinterpretq_u64_u8(vld1q_u8(p + 0));
> + uint64x2_t t1 = vreinterpretq_u64_u8(vld1q_u8(p + 16));
> + uint64x2_t t2 = vreinterpretq_u64_u8(vld1q_u8(p + 32));
> + uint64x2_t t3 = vreinterpretq_u64_u8(vld1q_u8(p + 48));
> +
> + v0 = fold_single_round(v0, t0, fold_consts_4x);
> + v1 = fold_single_round(v1, t1, fold_consts_4x);
> + v2 = fold_single_round(v2, t2, fold_consts_4x);
> + v3 = fold_single_round(v3, t3, fold_consts_4x);
> + p += 64;
> + len -= 64;
> + }
> +
> + v0 = fold_single_round(v0, v2, fold_consts_2x);
> + v1 = fold_single_round(v1, v3, fold_consts_2x);
> + v0 = fold_single_round(v0, v1, fold_consts);
> +
> + while (len >= 16) {
> + uint64x2_t t0 = vreinterpretq_u64_u8(vld1q_u8(p));
> +
> + v0 = fold_single_round(v0, t0, fold_consts);
> + p += 16;
> + len -= 16;
> + }
> +
> + /* Multiply the 128-bit value by x^64 and reduce it back to 128 bits. */
> + v0 = vextq_u64(v0, zero, 1) ^ pmull64_hi_lo(fold_consts, v0);
> +
> + /* Final Barrett reduction */
> + uint64x2_t bconsts = vld1q_u64(bconsts_val);
> + uint64x2_t final = pmull64(bconsts, v0);
> +
> + v0 = veor3q_u64(v0, vextq_u64(zero, final, 1),
> + pmull64_hi_lo(bconsts, final));
> +
> + return vgetq_lane_u64(v0, 1);
> +}
> +#endif
> --
> 2.56.1
>
^ permalink raw reply [flat|nested] 3+ messages in thread