Merge pull request #408 from Logan007/revPearson

added aes-ni support to Pearson hashing

include/pearson.h

@@ -16,6 +16,18 @@
  *
  */
 
+
+#include <stddef.h>
+#include <stdint.h>
+
+
+#if defined (__SSSE3__) && defined (__AES__) // AES-NI & SSSE3 -----------------------------
+
+#include <immintrin.h>
+
+#endif // AES-NI & SSSE3 -------------------------------------------------------------------
+
+
 void pearson_hash_256 (uint8_t *out, const uint8_t *in, size_t len);
 
 void pearson_hash_128 (uint8_t *out, const uint8_t *in, size_t len);

src/pearson.c

@@ -19,15 +19,9 @@
 // taken from https://github.com/Logan007/pearson
 // This is free and unencumbered software released into the public domain.
 
-#include <stddef.h>
-#include <stdint.h>
 
 #include "pearson.h"
 
-// compile with 'LOW_MEM_FOOTPRINT' defined to make use of 256 byte look-up tabe only
-// otherwise, a 16-bit look-up table is used which allows considerably faster hashing
-// however, it needs to be generated by once calling pearson_hash_init() upfront
-// #define LOW_MEM_FOOTPRINT
 
 // AES S-Box table -- allows for eventually supported hardware accelerated look-up
 static const uint8_t t[256] ={
@@ -71,9 +65,120 @@ static const uint8_t t[256] ={
 */
 
 
-#ifndef LOW_MEM_FOOTPRINT
+#if defined (__AES__) && defined (__SSSE3__) // AES-NI & SSSE3 ----------------------------
+
+
+void pearson_hash_256 (uint8_t *out, const uint8_t *in, size_t len) {
+
+	size_t i;
+
+        uint8_t upper[8] = { 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+        uint8_t lower[8] = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00 };
+
+        uint64_t upper_hash_mask = *(uint64_t*)&upper;
+        uint64_t lower_hash_mask = *(uint64_t*)&lower;
+
+        __m128i tmp = _mm_set1_epi8(0x10);
+
+        __m128i hash_mask = _mm_set_epi64 ((__m64)lower_hash_mask, (__m64)upper_hash_mask);
+        __m128i high_hash_mask = _mm_xor_si128 (tmp, hash_mask);
+        __m128i hash= _mm_setzero_si128();
+        __m128i high_hash= _mm_setzero_si128(); hash;
+
+	__m128i ZERO = _mm_setzero_si128();
+	__m128i ISOLATE_SBOX_MASK = _mm_set_epi32(0x0306090C, 0x0F020508, 0x0B0E0104, 0x070A0D00);
+
+        for (i = 0; i < len; i++) {
+                // broadcast the character, xor into hash, make them different permutations
+                __m128i cc = _mm_set1_epi8 (in[i]);
+                hash = _mm_xor_si128 (hash, cc);
+                high_hash = _mm_xor_si128 (high_hash, cc);
+                hash = _mm_xor_si128 (hash, hash_mask);
+                high_hash = _mm_xor_si128 (high_hash, high_hash_mask);
+                // table lookup
+                hash = _mm_shuffle_epi8(hash, ISOLATE_SBOX_MASK);           // re-order along AES round
+                high_hash = _mm_shuffle_epi8(high_hash, ISOLATE_SBOX_MASK); // re-order along AES round
+                hash = _mm_aesenclast_si128(hash, ZERO);
+                high_hash = _mm_aesenclast_si128(high_hash, ZERO);
+        }
+        // store output
+        _mm_store_si128 ((__m128i*)out , high_hash);
+        _mm_store_si128 ((__m128i*)&out[16] , hash);
+}
+
+
+void pearson_hash_128 (uint8_t *out, const uint8_t *in, size_t len) {
+
+        size_t i;
+
+        uint8_t upper[8] = { 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+        uint8_t lower[8] = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00 };
+
+        uint64_t upper_hash_mask = *(uint64_t*)&upper;
+        uint64_t lower_hash_mask = *(uint64_t*)&lower;
+
+        __m128i hash_mask = _mm_set_epi64 ((__m64)lower_hash_mask, (__m64)upper_hash_mask);
+        __m128i hash = _mm_setzero_si128(); // _mm_xor_si128 (hash, hash);
+
+        __m128i ZERO = _mm_setzero_si128();
+        __m128i ISOLATE_SBOX_MASK = _mm_set_epi32(0x0306090C, 0x0F020508, 0x0B0E0104, 0x070A0D00);
+
+        for (i = 0; i < len; i++) {
+                // broadcast the character, xor into hash, make them different permutations
+                __m128i cc = _mm_set1_epi8 (in[i]);
+                hash = _mm_xor_si128 (hash, cc);
+                hash = _mm_xor_si128 (hash, hash_mask);
+                // table lookup
+                hash = _mm_shuffle_epi8(hash, ISOLATE_SBOX_MASK); // re-order along AES round
+                hash = _mm_aesenclast_si128(hash, ZERO);
+        }
+        // store output
+        _mm_store_si128 ((__m128i*)out , hash);
+}
+
+
+// so far, only used internally
+static uint64_t pearson_hash_64 (const uint8_t *in, size_t len) {
+
+        size_t i;
+
+        __m128i hash_mask = _mm_cvtsi64_si128(0x0706050403020100);
+
+        __m128i hash = _mm_setzero_si128();
+
+        __m128i ZERO = _mm_setzero_si128();
+        __m128i ISOLATE_SBOX_MASK = _mm_set_epi32(0x0306090C, 0x0F020508, 0x0B0E0104, 0x070A0D00);
+
+        for (i = 0; i < len; i++) {
+                // broadcast the character, xor into hash, make them different permutations
+                __m128i cc = _mm_set1_epi8 (in[i]);
+                hash = _mm_xor_si128 (hash, cc);
+                hash = _mm_xor_si128 (hash, hash_mask);
+                // table lookup
+                hash = _mm_shuffle_epi8(hash, ISOLATE_SBOX_MASK); // re-order along AES round
+                hash = _mm_aesenclast_si128(hash, ZERO);
+        }
+        // store output
+	return _mm_cvtsi128_si64 (hash);
+}
+
+
+uint32_t pearson_hash_32 (const uint8_t *in, size_t len) {
+
+  return pearson_hash_64(in, len);
+}
+
+
+uint16_t pearson_hash_16 (const uint8_t *in, size_t len) {
+
+  return pearson_hash_64(in, len);
+}
+
+
+#else // plain C --------------------------------------------------------------------------
+
+
 static uint16_t t16[65536]; // 16-bit look-up table
-#endif
 
 #define ROR64(x,r) (((x)>>(r))|((x)<<(64-(r))))
 #define ROR32(x,r) (((x)>>(r))|((x)<<(32-(r))))
@@ -112,51 +217,6 @@ void pearson_hash_256 (uint8_t *out, const uint8_t *in, size_t len) {
 
     // table lookup
     uint64_t h = 0;
-#ifdef LOW_MEM_FOOTPRINT // 256 byte look-up table ----------
-    uint8_t x;
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    upper_hash = h;
-
-    h = 0;
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    lower_hash = h;
-
-    h = 0;
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_upper_hash; x = t[x]; high_upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    high_upper_hash = h;
-
-    h = 0;
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = high_lower_hash; x = t[x]; high_lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    high_lower_hash = h;
-#else // 16-bit look-up table -------------------------------
     uint16_t x;
     x = upper_hash; x = t16[x]; upper_hash >>= 16; h |= x; h=ROR64(h,16);
     x = upper_hash; x = t16[x]; upper_hash >>= 16; h |= x; h=ROR64(h,16);
@@ -184,7 +244,6 @@ void pearson_hash_256 (uint8_t *out, const uint8_t *in, size_t len) {
     x = high_lower_hash; x = t16[x]; high_lower_hash >>= 16; h |= x; h=ROR64(h,16);
     x = high_lower_hash; x = t16[x]; high_lower_hash >>= 16; h |= x; h=ROR64(h,16);
     high_lower_hash = h;
-#endif // LOW_MEM_FOOTPRINT ------
   }
   // store output
   uint64_t *o;
@@ -225,29 +284,6 @@ void pearson_hash_128 (uint8_t *out, const uint8_t *in, size_t len) {
     lower_hash ^= c ^ lower_hash_mask;
     // table lookup
     uint64_t h = 0;
-#ifdef LOW_MEM_FOOTPRINT // 256 byte look-up table ----------
-    uint8_t x;
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = upper_hash; x = t[x]; upper_hash >>= 8; h |= x; h=ROR64(h,8);
-    upper_hash = h;
-
-    h = 0;
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    x = lower_hash; x = t[x]; lower_hash >>= 8; h |= x; h=ROR64(h,8);
-    lower_hash= h;
-#else // 16-bit look-up table -------------------------------
     uint16_t x;
     x = upper_hash; x = t16[x]; upper_hash >>= 16; h |= x; h=ROR64(h,16);
     x = upper_hash; x = t16[x]; upper_hash >>= 16; h |= x; h=ROR64(h,16);
@@ -261,7 +297,6 @@ void pearson_hash_128 (uint8_t *out, const uint8_t *in, size_t len) {
     x = lower_hash; x = t16[x]; lower_hash >>= 16; h |= x; h=ROR64(h,16);
     x = lower_hash; x = t16[x]; lower_hash >>= 16; h |= x; h=ROR64(h,16);
     lower_hash = h;
-#endif // LOW_MEM_FOOTPRINT ------
   }
   // store output
   uint64_t *o;
@@ -287,7 +322,6 @@ uint32_t pearson_hash_32 (const uint8_t *in, size_t len) {
     c |= c << 16;
     hash ^= c ^ hash_mask;
     // table lookup
-#ifdef LOW_MEM_FOOTPRINT
     uint32_t h = 0;
     uint8_t x;
     x = hash; x = t[x]; hash >>= 8; h |= x; h=ROR32(h,8);
@@ -295,9 +329,6 @@ uint32_t pearson_hash_32 (const uint8_t *in, size_t len) {
     x = hash; x = t[x]; hash >>= 8; h |= x; h=ROR32(h,8);
     x = hash; x = t[x]; hash >>= 8; h |= x; h=ROR32(h,8);
     hash = h;
-#else
-    hash = (t16[hash >> 16] << 16) + t16[(uint16_t)hash];
-#endif
   }
   // output
   return hash;
@@ -317,20 +348,21 @@ uint16_t pearson_hash_16 (const uint8_t *in, size_t len) {
     c |= c <<  8;
     hash ^= c ^ hash_mask;
     // table lookup
-#ifdef LOW_MEM_FOOTPRINT
     hash = t[(uint8_t)hash] + (t[hash >> 8] << 8);
-#else
-    hash = t16[hash];
-#endif
   }
   // output
   return hash;
 }
 
 
+#endif // AES-NI & SSSE3, plain C ---------------------------------------------------------
+
+
 void pearson_hash_init () {
 
-#ifndef LOW_MEM_FOOTPRINT
+#if defined (__AES__) && (__SSSE3__)
+
+#else
   size_t i;
 
   for (i = 0; i < 65536; i++)