On Fri, Sep 12, 2025 at 06:49:01PM +0000, [email protected]
wrote:
> Using simd.h does make it easier to maintain. Is there a plan to upgrade
> simd.h to use SSE4 or SSSE3 in the future? Since SSE2 is much older, it
> lacks some of the more specialized intrinsics. For example, vectorized
> table lookup can be implemented via [0], and it’s available in SSSE3 and
> later x86 instruction sets.
There have been a couple of discussions about the possibility of requiring
x86-64-v2 for Postgres, but I'm not aware of any serious efforts in that
area.
I've attached a new version of the patch with a simd.h version of
hex_decode(). Here are the numbers:
arm
buf | HEAD | patch | % diff
-------+-------+-------+--------
16 | 22 | 23 | -5
64 | 61 | 23 | 62
256 | 158 | 47 | 70
1024 | 542 | 122 | 77
4096 | 2103 | 429 | 80
16384 | 8548 | 1673 | 80
65536 | 34663 | 6738 | 81
x86
buf | HEAD | patch | % diff
-------+-------+-------+--------
16 | 13 | 14 | -8
64 | 42 | 15 | 64
256 | 126 | 42 | 67
1024 | 461 | 149 | 68
4096 | 1802 | 576 | 68
16384 | 7166 | 2280 | 68
65536 | 28625 | 9108 | 68
A couple of notes:
* For hex_decode(), we just give up on the SIMD path and fall back on the
scalar path as soon as we see anything outside [0-9A-Za-z]. I suspect
this might introduce a regression for inputs of ~32 to ~64 bytes that
include whitespace (which must be skipped) or invalid characters, but I
don't whether those inputs are common or whether we care.
* The code makes some assumptions about endianness that might not be true
everywhere, but I've yet to dig into this further.
--
nathan
>From 155535f92df1d9cdf97739465b26ba970ed97063 Mon Sep 17 00:00:00 2001
From: Nathan Bossart <[email protected]>
Date: Fri, 12 Sep 2025 15:51:55 -0500
Subject: [PATCH v9 1/1] Optimize hex_encode() and hex_decode() using SIMD.
---
src/backend/utils/adt/encode.c | 142 ++++++++++++++++++++++-
src/include/port/simd.h | 201 +++++++++++++++++++++++++++++++++
2 files changed, 339 insertions(+), 4 deletions(-)
diff --git a/src/backend/utils/adt/encode.c b/src/backend/utils/adt/encode.c
index 4ccaed815d1..13883c27680 100644
--- a/src/backend/utils/adt/encode.c
+++ b/src/backend/utils/adt/encode.c
@@ -16,6 +16,7 @@
#include <ctype.h>
#include "mb/pg_wchar.h"
+#include "port/simd.h"
#include "utils/builtins.h"
#include "utils/memutils.h"
#include "varatt.h"
@@ -177,8 +178,8 @@ static const int8 hexlookup[128] = {
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
};
-uint64
-hex_encode(const char *src, size_t len, char *dst)
+static inline uint64
+hex_encode_scalar(const char *src, size_t len, char *dst)
{
const char *end = src + len;
@@ -193,6 +194,57 @@ hex_encode(const char *src, size_t len, char *dst)
return (uint64) len * 2;
}
+uint64
+hex_encode(const char *src, size_t len, char *dst)
+{
+#ifdef USE_NO_SIMD
+ return hex_encode_scalar(src, len, dst);
+#else
+ const uint64 tail_idx = len & ~(sizeof(Vector8) - 1);
+ uint64 i;
+
+ /*
+ * This works by splitting the high and low nibbles of each byte into
+ * separate vectors, adding the vectors to a mask that converts the
+ * nibbles to their equivalent ASCII bytes, and interleaving those bytes
+ * back together to form the final hex-encoded string. It might be
+ * possible to squeeze out a little more gain by manually unrolling the
+ * loop, but for now we don't bother.
+ */
+ for (i = 0; i < tail_idx; i += sizeof(Vector8))
+ {
+ Vector8 srcv;
+ Vector8 lo;
+ Vector8 hi;
+ Vector8 mask;
+
+ vector8_load(&srcv, (const uint8 *) &src[i]);
+
+ lo = vector8_and(srcv, vector8_broadcast(0x0f));
+ mask = vector8_gt(lo, vector8_broadcast(0x9));
+ mask = vector8_and(mask, vector8_broadcast('a' - '0' - 10));
+ mask = vector8_add(mask, vector8_broadcast('0'));
+ lo = vector8_add(lo, mask);
+
+ hi = vector8_and(srcv, vector8_broadcast(0xf0));
+ hi = vector32_shift_right_nibble(hi);
+ mask = vector8_gt(hi, vector8_broadcast(0x9));
+ mask = vector8_and(mask, vector8_broadcast('a' - '0' - 10));
+ mask = vector8_add(mask, vector8_broadcast('0'));
+ hi = vector8_add(hi, mask);
+
+ vector8_store((uint8 *) &dst[i * 2],
+ vector8_interleave_low(hi, lo));
+ vector8_store((uint8 *) &dst[i * 2 + sizeof(Vector8)],
+ vector8_interleave_high(hi, lo));
+ }
+
+ (void) hex_encode_scalar(src + i, len - i, dst + i * 2);
+
+ return (uint64) len * 2;
+#endif
+}
+
static inline bool
get_hex(const char *cp, char *out)
{
@@ -213,8 +265,8 @@ hex_decode(const char *src, size_t len, char *dst)
return hex_decode_safe(src, len, dst, NULL);
}
-uint64
-hex_decode_safe(const char *src, size_t len, char *dst, Node *escontext)
+static inline uint64
+hex_decode_safe_scalar(const char *src, size_t len, char *dst, Node *escontext)
{
const char *s,
*srcend;
@@ -254,6 +306,88 @@ hex_decode_safe(const char *src, size_t len, char *dst,
Node *escontext)
return p - dst;
}
+/*
+ * This helper converts each byte to its binary-equivalent nibble by
+ * subtraction. Along the way, it verifies the input is within the expected
+ * range of ASCII values and returns false if not. Finally, it combines the
+ * generated nibbles to form the return bytes, which will be separated by zero
+ * bytes in the destination vector. If everything goes as planned, returns
+ * true.
+ */
+#ifndef USE_NO_SIMD
+static inline bool
+hex_decode_simd_helper(const Vector8 src, Vector8 *dst)
+{
+ Vector8 msk;
+ Vector8 sub;
+
+ msk = vector8_lt(src, vector8_broadcast('0'));
+ if (unlikely(vector8_is_highbit_set(msk)))
+ return false;
+
+ msk = vector8_lt(src, vector8_broadcast('A'));
+ sub = vector8_and(msk, vector8_broadcast('0'));
+ *dst = vector8_ssub(src, sub);
+
+ msk = vector8_and(*dst, msk);
+ msk = vector8_gt(msk, vector8_broadcast(0x9));
+ if (unlikely(vector8_is_highbit_set(msk)))
+ return false;
+
+ msk = vector8_gt(src, vector8_broadcast('a' - 1));
+ sub = vector8_and(msk, vector8_broadcast('a' - 10));
+ msk = vector8_xor(msk, vector8_gt(sub, vector8_broadcast('A' - 1)));
+ msk = vector8_and(msk, vector8_broadcast('A' - 10));
+ sub = vector8_or(sub, msk);
+ *dst = vector8_ssub(*dst, sub);
+
+ msk = vector8_gt(*dst, vector8_broadcast(0xf));
+ if (unlikely(vector8_is_highbit_set(msk)))
+ return false;
+
+ msk = vector8_and(*dst, vector32_broadcast(0xff00ff00));
+ msk = vector32_shift_right_byte(msk);
+ *dst = vector8_and(*dst, vector32_broadcast(0x00ff00ff));
+ *dst = vector32_shift_left_nibble(*dst);
+ *dst = vector8_or(*dst, msk);
+ return true;
+}
+#endif /* ! USE_NO_SIMD */
+
+uint64
+hex_decode_safe(const char *src, size_t len, char *dst, Node *escontext)
+{
+#ifdef USE_NO_SIMD
+ return hex_decode_safe_scalar(src, len, dst, escontext);
+#else
+ const uint64 tail_idx = len & ~(sizeof(Vector8) * 2 - 1);
+ uint64 i;
+
+ /*
+ * We must process 2 vectors at a time since the output will be half the
+ * length of the input.
+ */
+ for (i = 0; i < tail_idx; i += sizeof(Vector8) * 2)
+ {
+ Vector8 srcv;
+ Vector8 dstv1;
+ Vector8 dstv2;
+
+ vector8_load(&srcv, (const uint8 *) &src[i]);
+ if (unlikely(!hex_decode_simd_helper(srcv, &dstv1)))
+ break;
+
+ vector8_load(&srcv, (const uint8 *) &src[i + sizeof(Vector8)]);
+ if (unlikely(!hex_decode_simd_helper(srcv, &dstv2)))
+ break;
+
+ vector8_store((uint8 *) &dst[i / 2], vector8_pack_16(dstv1,
dstv2));
+ }
+
+ return i / 2 + hex_decode_safe_scalar(src + i, len - i, dst + i / 2,
escontext);
+#endif
+}
+
static uint64
hex_enc_len(const char *src, size_t srclen)
{
diff --git a/src/include/port/simd.h b/src/include/port/simd.h
index 97c5f353022..0a9805e8ef1 100644
--- a/src/include/port/simd.h
+++ b/src/include/port/simd.h
@@ -70,6 +70,7 @@ static inline void vector32_load(Vector32 *v, const uint32
*s);
static inline Vector8 vector8_broadcast(const uint8 c);
#ifndef USE_NO_SIMD
static inline Vector32 vector32_broadcast(const uint32 c);
+static inline void vector8_store(uint8 *s, Vector8 v);
#endif
/* element-wise comparisons to a scalar */
@@ -86,6 +87,9 @@ static inline uint32 vector8_highbit_mask(const Vector8 v);
static inline Vector8 vector8_or(const Vector8 v1, const Vector8 v2);
#ifndef USE_NO_SIMD
static inline Vector32 vector32_or(const Vector32 v1, const Vector32 v2);
+static inline Vector8 vector8_xor(const Vector8 v1, const Vector8 v2);
+static inline Vector8 vector8_and(const Vector8 v1, const Vector8 v2);
+static inline Vector8 vector8_add(const Vector8 v1, const Vector8 v2);
static inline Vector8 vector8_ssub(const Vector8 v1, const Vector8 v2);
#endif
@@ -99,6 +103,18 @@ static inline Vector8 vector8_ssub(const Vector8 v1, const
Vector8 v2);
static inline Vector8 vector8_eq(const Vector8 v1, const Vector8 v2);
static inline Vector8 vector8_min(const Vector8 v1, const Vector8 v2);
static inline Vector32 vector32_eq(const Vector32 v1, const Vector32 v2);
+static inline Vector8 vector8_lt(const Vector8 v1, const Vector8 v2);
+static inline Vector8 vector8_gt(const Vector8 v1, const Vector8 v2);
+#endif
+
+/* vector manipulation */
+#ifndef USE_NO_SIMD
+static inline Vector8 vector8_interleave_low(const Vector8 v1, const Vector8
v2);
+static inline Vector8 vector8_interleave_high(const Vector8 v1, const Vector8
v2);
+static inline Vector8 vector8_pack_16(const Vector8 v1, const Vector8 v2);
+static inline Vector32 vector32_shift_left_nibble(const Vector32 v1);
+static inline Vector32 vector32_shift_right_nibble(const Vector32 v1);
+static inline Vector32 vector32_shift_right_byte(const Vector32 v1);
#endif
/*
@@ -128,6 +144,21 @@ vector32_load(Vector32 *v, const uint32 *s)
}
#endif /* ! USE_NO_SIMD */
+/*
+ * Store a vector into the given memory address.
+ */
+#ifndef USE_NO_SIMD
+static inline void
+vector8_store(uint8 *s, Vector8 v)
+{
+#ifdef USE_SSE2
+ _mm_storeu_si128((Vector8 *) s, v);
+#elif defined(USE_NEON)
+ vst1q_u8(s, v);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
/*
* Create a vector with all elements set to the same value.
*/
@@ -358,6 +389,51 @@ vector32_or(const Vector32 v1, const Vector32 v2)
}
#endif /* ! USE_NO_SIMD */
+/*
+ * Return the bitwise XOR of the inputs.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_xor(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_xor_si128(v1, v2);
+#elif defined(USE_NEON)
+ return veorq_u8(v1, v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Return the bitwise AND of the inputs.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_and(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_and_si128(v1, v2);
+#elif defined(USE_NEON)
+ return vandq_u8(v1, v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Return the result of adding the respective elements of the input vectors.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_add(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_add_epi8(v1, v2);
+#elif defined(USE_NEON)
+ return vaddq_u8(v1, v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
/*
* Return the result of subtracting the respective elements of the input
* vectors using saturation (i.e., if the operation would yield a value less
@@ -404,6 +480,39 @@ vector32_eq(const Vector32 v1, const Vector32 v2)
}
#endif /* ! USE_NO_SIMD */
+/*
+ * Return a vector with all bits set for each lane of v1 that is less than the
+ * corresponding lane of v2. NB: The comparison treats the elements as signed.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_lt(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_cmplt_epi8(v1, v2);
+#elif defined(USE_NEON)
+ return vcltq_s8((int8x16_t) v1, (int8x16_t) v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Return a vector with all bits set for each lane of v1 that is greater than
+ * the corresponding lane of v2. NB: The comparison treats the elements as
+ * signed.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_gt(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_cmpgt_epi8(v1, v2);
+#elif defined(USE_NEON)
+ return vcgtq_s8((int8x16_t) v1, (int8x16_t) v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
/*
* Given two vectors, return a vector with the minimum element of each.
*/
@@ -419,4 +528,96 @@ vector8_min(const Vector8 v1, const Vector8 v2)
}
#endif /* ! USE_NO_SIMD */
+/*
+ * Interleave elements of low halves of given vectors.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_interleave_low(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_unpacklo_epi8(v1, v2);
+#elif defined(USE_NEON)
+ return vzip1q_u8(v1, v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Interleave elements of high halves of given vectors.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_interleave_high(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_unpackhi_epi8(v1, v2);
+#elif defined(USE_NEON)
+ return vzip2q_u8(v1, v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Pack 16-bit elements in the given vectors into a single vector of 8-bit
+ * elements. NB: The upper 8-bits of each 16-bit element must be zeros, else
+ * this will produce different results on different architectures.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector8
+vector8_pack_16(const Vector8 v1, const Vector8 v2)
+{
+#ifdef USE_SSE2
+ return _mm_packus_epi16(v1, v2);
+#elif defined(USE_NEON)
+ return vuzp1q_u8(v1, v2);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Unsigned shift left of each element in the vector by 4 bits.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector32
+vector32_shift_left_nibble(const Vector32 v1)
+{
+#ifdef USE_SSE2
+ return _mm_slli_epi32(v1, 4);
+#elif defined(USE_NEON)
+ return vshlq_n_u32(v1, 4);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Unsigned shift right of each element in the vector by 4 bits.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector32
+vector32_shift_right_nibble(const Vector32 v1)
+{
+#ifdef USE_SSE2
+ return _mm_srli_epi32(v1, 4);
+#elif defined(USE_NEON)
+ return vshrq_n_u32(v1, 4);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
+/*
+ * Unsigned shift right of each element in the vector by 1 byte.
+ */
+#ifndef USE_NO_SIMD
+static inline Vector32
+vector32_shift_right_byte(const Vector32 v1)
+{
+#ifdef USE_SSE2
+ return _mm_srli_epi32(v1, 8);
+#elif defined(USE_NEON)
+ return vshrq_n_u32(v1, 8);
+#endif
+}
+#endif /* ! USE_NO_SIMD */
+
#endif /* SIMD_H */
--
2.39.5 (Apple Git-154)
