Skip to contentSkip to conversation

Improve crc32c() performance on newer x86_64 CPUs

PATCH In review Eric Biggers 2 participants 4 messages Jul 19, 2025 – Jul 21, 2025

Subsystems (2)
lore
Share & tools
lore

Eric Biggers

author maintainer
#0 lore raw

cover Improve crc32c() performance on newer x86_64 CPUs

This series improves crc32c() performance on lengths >= 512 bytes on newer x86_64 CPUs by enabling the VPCLMULQDQ (vector carryless multiplication) optimized CRC code.

This series targets crc-next.

Eric Biggers (2):
lib/crc: x86: Reorganize crc-pclmul static_call initialization lib/crc: x86/crc32c: Enable VPCLMULQDQ optimization where beneficial

 lib/crc/x86/crc-pclmul-consts.h   | 47 +++++++++++++++++++++++++++++-
 lib/crc/x86/crc-pclmul-template.h | 31 +++++++++-----------
 lib/crc/x86/crc-t10dif.h          |  9 +++++-
 lib/crc/x86/crc32.h               | 48 +++++++++++++++++++++++++++++--
 lib/crc/x86/crc64.h               | 15 ++++++++--
 5 files changed, 127 insertions(+), 23 deletions(-)

base-commit: 9b0236f4efb889869f7d4f3f084f508cc0433ec9

Signature · 1 line
2.50.1
author maintainer
#1 lore raw

PATCH 1/2 lib/crc: x86: Reorganize crc-pclmul static_call initialization

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_lsb and crc64_msb static_call updates together. A later commit will also use this to initialize an additional static_key when crc32_lsb_vpclmul_avx512() is enabled.

---
 lib/crc/x86/crc-pclmul-template.h | 31 ++++++++++++++-----------------
 lib/crc/x86/crc-t10dif.h          |  9 ++++++++-
 lib/crc/x86/crc32.h               | 10 +++++++++-
 lib/crc/x86/crc64.h               | 15 +++++++++++++--
 4 files changed, 44 insertions(+), 21 deletions(-)
lib/crc/x86/crc-pclmul-template.h +14 -17
index 51cba520a7dbd..35c950d7010c2 100644
@@ -23,27 +23,24 @@ crc_t prefix##_vpclmul_avx2(crc_t crc, const u8 *p, size_t len,		\
 			    const void *consts_ptr);			\
 crc_t prefix##_vpclmul_avx512(crc_t crc, const u8 *p, size_t len,	\
 			      const void *consts_ptr);			\
 DEFINE_STATIC_CALL(prefix##_pclmul, prefix##_pclmul_sse)
 
-#define INIT_CRC_PCLMUL(prefix)						\
-do {									\
-	if (boot_cpu_has(X86_FEATURE_VPCLMULQDQ) &&			\
-	    boot_cpu_has(X86_FEATURE_AVX2) &&				\
-	    cpu_has_xfeatures(XFEATURE_MASK_YMM, NULL)) {		\
-		if (boot_cpu_has(X86_FEATURE_AVX512BW) &&		\
-		    boot_cpu_has(X86_FEATURE_AVX512VL) &&		\
-		    !boot_cpu_has(X86_FEATURE_PREFER_YMM) &&		\
-		    cpu_has_xfeatures(XFEATURE_MASK_AVX512, NULL)) {	\
-			static_call_update(prefix##_pclmul,		\
-					   prefix##_vpclmul_avx512);	\
-		} else {						\
-			static_call_update(prefix##_pclmul,		\
-					   prefix##_vpclmul_avx2);	\
-		}							\
-	}								\
-} while (0)
+static inline bool have_vpclmul(void)
+{
+	return boot_cpu_has(X86_FEATURE_VPCLMULQDQ) &&
+	       boot_cpu_has(X86_FEATURE_AVX2) &&
+	       cpu_has_xfeatures(XFEATURE_MASK_YMM, NULL);
+}
+
+static inline bool have_avx512(void)
+{
+	return boot_cpu_has(X86_FEATURE_AVX512BW) &&
+	       boot_cpu_has(X86_FEATURE_AVX512VL) &&
+	       !boot_cpu_has(X86_FEATURE_PREFER_YMM) &&
+	       cpu_has_xfeatures(XFEATURE_MASK_AVX512, NULL);
+}
 
 /*
  * Call a [V]PCLMULQDQ optimized CRC function if the data length is at least 16
  * bytes, the CPU has PCLMULQDQ support, and the current context may use SIMD.
  *
lib/crc/x86/crc-t10dif.h +8 -1
index eb1f23db4daa1..2a02a3026f3f8 100644
@@ -21,8 +21,15 @@ static inline u16 crc_t10dif_arch(u16 crc, const u8 *p, size_t len)
 #define crc_t10dif_mod_init_arch crc_t10dif_mod_init_arch
 static inline void crc_t10dif_mod_init_arch(void)
 {
 	if (boot_cpu_has(X86_FEATURE_PCLMULQDQ)) {
 		static_branch_enable(&have_pclmulqdq);
-		INIT_CRC_PCLMUL(crc16_msb);
+		if (have_vpclmul()) {
+			if (have_avx512())
+				static_call_update(crc16_msb_pclmul,
+						   crc16_msb_vpclmul_avx512);
+			else
+				static_call_update(crc16_msb_pclmul,
+						   crc16_msb_vpclmul_avx2);
+		}
 	}
 }
lib/crc/x86/crc32.h +9 -1
index 28451d5769c3a..ba4dacf23340c 100644
@@ -75,11 +75,19 @@ static inline void crc32_mod_init_arch(void)
 {
 	if (boot_cpu_has(X86_FEATURE_XMM4_2))
 		static_branch_enable(&have_crc32);
 	if (boot_cpu_has(X86_FEATURE_PCLMULQDQ)) {
 		static_branch_enable(&have_pclmulqdq);
-		INIT_CRC_PCLMUL(crc32_lsb);
+		if (have_vpclmul()) {
+			if (have_avx512()) {
+				static_call_update(crc32_lsb_pclmul,
+						   crc32_lsb_vpclmul_avx512);
+			} else {
+				static_call_update(crc32_lsb_pclmul,
+						   crc32_lsb_vpclmul_avx2);
+			}
+		}
 	}
 }
 
 static inline u32 crc32_optimizations_arch(void)
 {
lib/crc/x86/crc64.h +13 -2
index 54aca3a9475c9..aea64409e4193 100644
@@ -29,9 +29,20 @@ static inline u64 crc64_nvme_arch(u64 crc, const u8 *p, size_t len)
 #define crc64_mod_init_arch crc64_mod_init_arch
 static inline void crc64_mod_init_arch(void)
 {
 	if (boot_cpu_has(X86_FEATURE_PCLMULQDQ)) {
 		static_branch_enable(&have_pclmulqdq);
-		INIT_CRC_PCLMUL(crc64_msb);
-		INIT_CRC_PCLMUL(crc64_lsb);
+		if (have_vpclmul()) {
+			if (have_avx512()) {
+				static_call_update(crc64_lsb_pclmul,
+						   crc64_lsb_vpclmul_avx512);
+				static_call_update(crc64_msb_pclmul,
+						   crc64_msb_vpclmul_avx512);
+			} else {
+				static_call_update(crc64_lsb_pclmul,
+						   crc64_lsb_vpclmul_avx2);
+				static_call_update(crc64_msb_pclmul,
+						   crc64_msb_vpclmul_avx2);
+			}
+		}
 	}
 }

base-commit: 9b0236f4efb889869f7d4f3f084f508cc0433ec9

Signature · 1 line
2.50.1
author maintainer
#2 lore raw

PATCH 2/2 lib/crc: x86/crc32c: Enable VPCLMULQDQ optimization where beneficial

In reply to Eric Biggers

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 "hot", 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.

---
 lib/crc/x86/crc-pclmul-consts.h | 47 ++++++++++++++++++++++++++++++++-
 lib/crc/x86/crc32.h             | 38 +++++++++++++++++++++++++-
 2 files changed, 83 insertions(+), 2 deletions(-)
lib/crc/x86/crc-pclmul-consts.h +46 -1
index fcc63c0643330..6ae94158fca28 100644
@@ -1,10 +1,10 @@
 /* SPDX-License-Identifier: GPL-2.0-or-later */
 /*
  * CRC constants generated by:
  *
- *	./scripts/gen-crc-consts.py x86_pclmul crc16_msb_0x8bb7,crc32_lsb_0xedb88320,crc64_msb_0x42f0e1eba9ea3693,crc64_lsb_0x9a6c9329ac4bc9b5
+ *	./scripts/gen-crc-consts.py x86_pclmul crc16_msb_0x8bb7,crc32_lsb_0xedb88320,crc32_lsb_0x82f63b78,crc64_msb_0x42f0e1eba9ea3693,crc64_lsb_0x9a6c9329ac4bc9b5
  *
  * Do not edit manually.
  */
 
 /*
@@ -96,10 +96,55 @@ static const struct {
 		0xb4e5b025f7011641,	/* HI64_TERMS: floor(x^95 / G) */
 		0x00000001db710640,	/* LO64_TERMS: (G - x^32) * x^31 */
 	},
 };
 
+/*
+ * CRC folding constants generated for least-significant-bit-first CRC-32 using
+ * G(x) = x^32 + x^28 + x^27 + x^26 + x^25 + x^23 + x^22 + x^20 + x^19 + x^18 +
+ *        x^14 + x^13 + x^11 + x^10 + x^9 + x^8 + x^6 + x^0
+ */
+static const struct {
+	u64 fold_across_2048_bits_consts[2];
+	u64 fold_across_1024_bits_consts[2];
+	u64 fold_across_512_bits_consts[2];
+	u64 fold_across_256_bits_consts[2];
+	u64 fold_across_128_bits_consts[2];
+	u8 shuf_table[48];
+	u64 barrett_reduction_consts[2];
+} crc32_lsb_0x82f63b78_consts ____cacheline_aligned __maybe_unused = {
+	.fold_across_2048_bits_consts = {
+		0x00000000dcb17aa4,	/* HI64_TERMS: (x^2079 mod G) * x^32 */
+		0x00000000b9e02b86,	/* LO64_TERMS: (x^2015 mod G) * x^32 */
+	},
+	.fold_across_1024_bits_consts = {
+		0x000000006992cea2,	/* HI64_TERMS: (x^1055 mod G) * x^32 */
+		0x000000000d3b6092,	/* LO64_TERMS: (x^991 mod G) * x^32 */
+	},
+	.fold_across_512_bits_consts = {
+		0x00000000740eef02,	/* HI64_TERMS: (x^543 mod G) * x^32 */
+		0x000000009e4addf8,	/* LO64_TERMS: (x^479 mod G) * x^32 */
+	},
+	.fold_across_256_bits_consts = {
+		0x000000003da6d0cb,	/* HI64_TERMS: (x^287 mod G) * x^32 */
+		0x00000000ba4fc28e,	/* LO64_TERMS: (x^223 mod G) * x^32 */
+	},
+	.fold_across_128_bits_consts = {
+		0x00000000f20c0dfe,	/* HI64_TERMS: (x^159 mod G) * x^32 */
+		0x00000000493c7d27,	/* LO64_TERMS: (x^95 mod G) * x^32 */
+	},
+	.shuf_table = {
+		-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
+		 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+		-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
+	},
+	.barrett_reduction_consts = {
+		0x4869ec38dea713f1,	/* HI64_TERMS: floor(x^95 / G) */
+		0x0000000105ec76f0,	/* LO64_TERMS: (G - x^32) * x^31 */
+	},
+};
+
 /*
  * CRC folding constants generated for most-significant-bit-first CRC-64 using
  * G(x) = x^64 + x^62 + x^57 + x^55 + x^54 + x^53 + x^52 + x^47 + x^46 + x^45 +
  *        x^40 + x^39 + x^38 + x^37 + x^35 + x^33 + x^32 + x^31 + x^29 + x^27 +
  *        x^24 + x^23 + x^22 + x^21 + x^19 + x^17 + x^13 + x^12 + x^10 + x^9 +
lib/crc/x86/crc32.h +37 -1
index ba4dacf23340c..0c2f4e5cc86df 100644
@@ -9,10 +9,11 @@
 
 #include "crc-pclmul-template.h"
 
 static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_crc32);
 static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_pclmulqdq);
+static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_vpclmul_avx512);
 
 DECLARE_CRC_PCLMUL_FUNCS(crc32_lsb, u32);
 
 static inline u32 crc32_le_arch(u32 crc, const u8 *p, size_t len)
 {
@@ -42,16 +43,50 @@ static inline u32 crc32c_arch(u32 crc, const u8 *p, size_t len)
 	if (!static_branch_likely(&have_crc32))
 		return crc32c_base(crc, p, len);
 
 	if (IS_ENABLED(CONFIG_X86_64) && len >= CRC32C_PCLMUL_BREAKEVEN &&
 	    static_branch_likely(&have_pclmulqdq) && crypto_simd_usable()) {
+		/*
+		 * Long length, the vector registers are usable, and the CPU is
+		 * 64-bit and supports both CRC32 and PCLMULQDQ instructions.
+		 * It is worthwhile to divide the data into multiple streams,
+		 * CRC them independently, and combine them using PCLMULQDQ.
+		 * crc32c_x86_3way() does this using 3 streams, which is the
+		 * most that x86_64 CPUs have traditionally been capable of.
+		 *
+		 * However, due to improved VPCLMULQDQ performance on newer
+		 * CPUs, use crc32_lsb_vpclmul_avx512() instead of
+		 * crc32c_x86_3way() when the CPU supports VPCLMULQDQ and has a
+		 * "good" implementation of AVX-512.
+		 *
+		 * Future work: the optimal strategy on Zen 3--5 is actually to
+		 * use both crc32q and VPCLMULQDQ in parallel.  Unfortunately,
+		 * different numbers of streams and vector lengths are optimal
+		 * on each CPU microarchitecture, making taking advantage of
+		 * this challenging.  (Zen 5 even supports 7 parallel crc32q, a
+		 * major upgrade.)  For now, just choose between
+		 * crc32c_x86_3way() and crc32_lsb_vpclmul_avx512().  The latter
+		 * is needed anyway for crc32_le(), so we just reuse it here.
+		 */
 		kernel_fpu_begin();
-		crc = crc32c_x86_3way(crc, p, len);
+		if (static_branch_likely(&have_vpclmul_avx512))
+			crc = crc32_lsb_vpclmul_avx512(crc, p, len,
+				       crc32_lsb_0x82f63b78_consts.fold_across_128_bits_consts);
+		else
+			crc = crc32c_x86_3way(crc, p, len);
 		kernel_fpu_end();
 		return crc;
 	}
 
+	/*
+	 * Short length, XMM registers unusable, or the CPU is 32-bit; but the
+	 * CPU supports CRC32 instructions.  Just issue a single stream of CRC32
+	 * instructions inline.  While this doesn't use the CPU's CRC32
+	 * throughput very well, it avoids the need to combine streams.  Stream
+	 * combination would be inefficient here.
+	 */
+
 	for (num_longs = len / sizeof(unsigned long);
 	     num_longs != 0; num_longs--, p += sizeof(unsigned long))
 		asm(CRC32_INST : "+r" (crc) : ASM_INPUT_RM (*(unsigned long *)p));
 
 	if (sizeof(unsigned long) > 4 && (len & 4)) {
@@ -79,10 +114,11 @@ static inline void crc32_mod_init_arch(void)
 		static_branch_enable(&have_pclmulqdq);
 		if (have_vpclmul()) {
 			if (have_avx512()) {
 				static_call_update(crc32_lsb_pclmul,
 						   crc32_lsb_vpclmul_avx512);
+				static_branch_enable(&have_vpclmul_avx512);
 			} else {
 				static_call_update(crc32_lsb_pclmul,
 						   crc32_lsb_vpclmul_avx2);
 			}
 		}
Signature · 1 line
2.50.1
reviewer Acked-by on the v1 series
#3 lore raw
In reply to Eric Biggers
Quoted 11 lines · Eric Biggers

This series improves crc32c() performance on lengths >= 512 bytes on newer x86_64 CPUs by enabling the VPCLMULQDQ (vector carryless multiplication) optimized CRC code.

This series targets crc-next.

Eric Biggers (2):
lib/crc: x86: Reorganize crc-pclmul static_call initialization lib/crc: x86/crc32c: Enable VPCLMULQDQ optimization where beneficial