Switch to using SIMD-oriented Fast Mersenne Twister for random number generation.

In testing, reduced random number generation time by a factor of 8-10.
Drops support for processors older than Pentium 4.
Drop Mersenne Twister library; use a C++ SFMT library.

--HG--
branch : trunk

1 files changed, 286 insertions, 135 deletions

diff --git a/dep/SFMT/SFMT.h b/dep/SFMT/SFMT.h
index 7c8b35e9e95..03a7e853316 100644
--- a/dep/SFMT/SFMT.h
+++ b/dep/SFMT/SFMT.h
@@ -1,157 +1,308 @@
-/** 
- * @file SFMT.h 
- *
- * @brief SIMD oriented Fast Mersenne Twister(SFMT) pseudorandom
- * number generator
- *
- * @author Mutsuo Saito (Hiroshima University)
- * @author Makoto Matsumoto (Hiroshima University)
- *
- * Copyright (C) 2006, 2007 Mutsuo Saito, Makoto Matsumoto and Hiroshima
- * University. All rights reserved.
- *
- * The new BSD License is applied to this software.
- * see LICENSE.txt
- *
- * @note We assume that your system has inttypes.h.  If your system
- * doesn't have inttypes.h, you have to typedef uint32_t and uint64_t,
- * and you have to define PRIu64 and PRIx64 in this file as follows:
- * @verbatim
- typedef unsigned int uint32_t
- typedef unsigned long long uint64_t  
- #define PRIu64 "llu"
- #define PRIx64 "llx"
-@endverbatim
- * uint32_t must be exactly 32-bit unsigned integer type (no more, no
- * less), and uint64_t must be exactly 64-bit unsigned integer type.
- * PRIu64 and PRIx64 are used for printf function to print 64-bit
- * unsigned int and 64-bit unsigned int in hexadecimal format.
+/*
+ * Copyright notice
+ * ================
+ * GNU General Public License http://www.gnu.org/licenses/gpl.html
+ * This C++ implementation of SFMT contains parts of the original C code
+ * which was published under the following BSD license, which is therefore
+ * in effect in addition to the GNU General Public License.
+ * Copyright (c) 2006, 2007 by Mutsuo Saito, Makoto Matsumoto and Hiroshima University.
+ * Copyright (c) 2008 by Agner Fog.
+ * Copyright (c) 2010 Trinity Core
+ * 
+ *  BSD License:
+ *  Redistribution and use in source and binary forms, with or without 
+ * modification, are permitted provided that the following conditions are met:
+ *     > Redistributions of source code must retain the above copyright notice, 
+ *       this list of conditions and the following disclaimer.
+ *     > Redistributions in binary form must reproduce the above copyright notice, 
+ *       this list of conditions and the following disclaimer in the documentation
+ *       and/or other materials provided with the distribution.
+ *     > Neither the name of the Hiroshima University nor the names of its 
+ *       contributors may be used to endorse or promote products derived from 
+ *       this software without specific prior written permission.
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
 #ifndef SFMT_H
 #define SFMT_H
 
-#include <stdio.h>
-
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)
-  #include <inttypes.h>
-#elif defined(_MSC_VER) || defined(__BORLANDC__)
-  typedef unsigned int uint32_t;
-  typedef unsigned __int64 uint64_t;
-  #define inline __inline
-#else
-  #include <inttypes.h>
-  #if defined(__GNUC__)
-    #define inline __inline__
-  #endif
-#endif
+#include <emmintrin.h>                 // Define SSE2 intrinsics
+#include "randomc.h"                   // Define integer types etc
+#include <time.h>
 
-#ifndef PRIu64
-  #if defined(_MSC_VER) || defined(__BORLANDC__)
-    #define PRIu64 "I64u"
-    #define PRIx64 "I64x"
-  #else
-    #define PRIu64 "llu"
-    #define PRIx64 "llx"
-  #endif
-#endif
+// Choose one of the possible Mersenne exponents.
+// Higher values give longer cycle length and use more memory:
+//#define MEXP   607
+//#define MEXP  1279
+//#define MEXP  2281
+//#define MEXP  4253
+  #define MEXP 11213
+//#define MEXP 19937
+//#define MEXP 44497
 
-#if defined(__GNUC__)
-#define ALWAYSINLINE __attribute__((always_inline))
-#else
-#define ALWAYSINLINE
-#endif
+// Define constants for the selected Mersenne exponent:
+#if MEXP == 44497
+#define SFMT_N    348                  // Size of state vector
+#define SFMT_M    330                  // Position of intermediate feedback
+#define SFMT_SL1    5                  // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	  3                  // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1    9                  // Right shift of W[M], 32-bit words
+#define SFMT_SR2	  3                  // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	  0xeffffffb,0xdfbebfff,0xbfbf7bef,0x9ffd7bff // AND mask
+#define SFMT_PARITY 1,0,0xa3ac4000,0xecc1327a   // Period certification vector
+
+#elif MEXP == 19937
+#define SFMT_N    156                  // Size of state vector
+#define SFMT_M    122                  // Position of intermediate feedback
+#define SFMT_SL1   18                  // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	  1                  // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1   11                  // Right shift of W[M], 32-bit words
+#define SFMT_SR2	  1                  // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	  0xdfffffef,0xddfecb7f,0xbffaffff,0xbffffff6 // AND mask
+#define SFMT_PARITY 1,0,0,0x13c9e684   // Period certification vector
+
+#elif MEXP == 11213
+#define SFMT_N    88                   // Size of state vector
+#define SFMT_M    68                   // Position of intermediate feedback
+#define SFMT_SL1	14                   // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	 3                   // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1	 7                   // Right shift of W[M], 32-bit words
+#define SFMT_SR2	 3                   // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	 0xeffff7fb,0xffffffef,0xdfdfbfff,0x7fffdbfd // AND mask
+#define SFMT_PARITY 1,0,0xe8148000,0xd0c7afa3 // Period certification vector
+
+#elif MEXP == 4253
+#define SFMT_N    34                   // Size of state vector
+#define SFMT_M    17                   // Position of intermediate feedback
+#define SFMT_SL1	20                   // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	 1                   // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1	 7                   // Right shift of W[M], 32-bit words
+#define SFMT_SR2	 1                   // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	 0x9f7bffff, 0x9fffff5f, 0x3efffffb, 0xfffff7bb // AND mask
+#define SFMT_PARITY 0xa8000001, 0xaf5390a3, 0xb740b3f8, 0x6c11486d // Period certification vector
+
+#elif MEXP == 2281
+#define SFMT_N    18                   // Size of state vector
+#define SFMT_M    12                   // Position of intermediate feedback
+#define SFMT_SL1	19                   // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	 1                   // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1	 5                   // Right shift of W[M], 32-bit words
+#define SFMT_SR2	 1                   // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	 0xbff7ffbf, 0xfdfffffe, 0xf7ffef7f, 0xf2f7cbbf // AND mask
+#define SFMT_PARITY 0x00000001, 0x00000000, 0x00000000, 0x41dfa600  // Period certification vector
+
+#elif MEXP == 1279
+#define SFMT_N    10                   // Size of state vector
+#define SFMT_M     7                   // Position of intermediate feedback
+#define SFMT_SL1	14                   // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	 3                   // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1	 5                   // Right shift of W[M], 32-bit words
+#define SFMT_SR2	 1                   // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	  0xf7fefffd, 0x7fefcfff, 0xaff3ef3f, 0xb5ffff7f  // AND mask
+#define SFMT_PARITY 0x00000001, 0x00000000, 0x00000000, 0x20000000  // Period certification vector
 
-#if defined(_MSC_VER)
-  #if _MSC_VER >= 1200
-    #define PRE_ALWAYS __forceinline
-  #else
-    #define PRE_ALWAYS inline
-  #endif
-#else
-  #define PRE_ALWAYS inline
+#elif MEXP == 607
+#define SFMT_N     5                   // Size of state vector
+#define SFMT_M     2                   // Position of intermediate feedback
+#define SFMT_SL1	15                   // Left shift of W[N-1], 32-bit words
+#define SFMT_SL2	 3                   // Left shift of W[0], *8, 128-bit words
+#define SFMT_SR1	13                   // Right shift of W[M], 32-bit words
+#define SFMT_SR2	 3                   // Right shift of W[N-2], *8, 128-bit words
+#define SFMT_MASK	  0xfdff37ff, 0xef7f3f7d, 0xff777b7d, 0x7ff7fb2f  // AND mask
+#define SFMT_PARITY 0x00000001, 0x00000000, 0x00000000, 0x5986f054  // Period certification vector
 #endif
 
-uint32_t gen_rand32(void);
-uint64_t gen_rand64(void);
-void fill_array32(uint32_t *array, int size);
-void fill_array64(uint64_t *array, int size);
-void init_gen_rand(uint32_t seed);
-void init_by_array(uint32_t *init_key, int key_length);
-const char *get_idstring(void);
-int get_min_array_size32(void);
-int get_min_array_size64(void);
-
-/* These real versions are due to Isaku Wada */
-/** generates a random number on [0,1]-real-interval */
-inline static double to_real1(uint32_t v)
-{
-    return v * (1.0/4294967295.0); 
-    /* divided by 2^32-1 */ 
+// Functions used by SFMTRand::RandomInitByArray
+static uint32_t func1(uint32_t x) {
+    return (x ^ (x >> 27)) * 1664525U;
 }
 
-/** generates a random number on [0,1]-real-interval */
-inline static double genrand_real1(void)
-{
-    return to_real1(gen_rand32());
+static uint32_t func2(uint32_t x) {
+    return (x ^ (x >> 27)) * 1566083941U;
 }
 
-/** generates a random number on [0,1)-real-interval */
-inline static double to_real2(uint32_t v)
-{
-    return v * (1.0/4294967296.0); 
-    /* divided by 2^32 */
+// Subfunction for the sfmt algorithm
+static inline __m128i sfmt_recursion(__m128i const &a, __m128i const &b, 
+__m128i const &c, __m128i const &d, __m128i const &mask) {
+    __m128i a1, b1, c1, d1, z1, z2;
+    b1 = _mm_srli_epi32(b, SFMT_SR1);
+    a1 = _mm_slli_si128(a, SFMT_SL2);
+    c1 = _mm_srli_si128(c, SFMT_SR2);
+    d1 = _mm_slli_epi32(d, SFMT_SL1);
+    b1 = _mm_and_si128(b1, mask);
+    z1 = _mm_xor_si128(a, a1);
+    z2 = _mm_xor_si128(b1, d1);
+    z1 = _mm_xor_si128(z1, c1);
+    z2 = _mm_xor_si128(z1, z2);
+    return z2;
 }
 
-/** generates a random number on [0,1)-real-interval */
-inline static double genrand_real2(void)
-{
-    return to_real2(gen_rand32());
-}
+// Class for SFMT generator
+class SFMTRand {                              // Encapsulate random number generator
+public:
+    SFMTRand() { LastInterval = 0; RandomInit((int)(time(0))); } 
 
-/** generates a random number on (0,1)-real-interval */
-inline static double to_real3(uint32_t v)
-{
-    return (((double)v) + 0.5)*(1.0/4294967296.0); 
-    /* divided by 2^32 */
-}
+    void RandomInit(int seed)                     // Re-seed
+    {
+        // Re-seed
+        uint32_t i;                         // Loop counter
+        uint32_t y = seed;                  // Temporary
+        uint32_t statesize = SFMT_N*4;      // Size of state vector
 
-/** generates a random number on (0,1)-real-interval */
-inline static double genrand_real3(void)
-{
-    return to_real3(gen_rand32());
-}
-/** These real versions are due to Isaku Wada */
+        // Fill state vector with random numbers from seed
+        ((uint32_t*)state)[0] = y;
+        const uint32_t factor = 1812433253U;// Multiplication factor
 
-/** generates a random number on [0,1) with 53-bit resolution*/
-inline static double to_res53(uint64_t v) 
-{ 
-    return v * (1.0/18446744073709551616.0L);
-}
+        for (i = 1; i < statesize; i++) {
+            y = factor * (y ^ (y >> 30)) + i;
+            ((uint32_t*)state)[i] = y;
+        }
 
-/** generates a random number on [0,1) with 53-bit resolution from two
- * 32 bit integers */
-inline static double to_res53_mix(uint32_t x, uint32_t y) 
-{ 
-    return to_res53(x | ((uint64_t)y << 32));
-}
+        // Further initialization and period certification
+        Init2();
+    }
 
-/** generates a random number on [0,1) with 53-bit resolution
- */
-inline static double genrand_res53(void) 
-{ 
-    return to_res53(gen_rand64());
-} 
+    int32_t IRandom(int32_t min, int32_t max)     // Output random integer
+    {
+        // Output random integer in the interval min <= x <= max
+        // Slightly inaccurate if (max-min+1) is not a power of 2
+        if (max <= min) {
+            if (max == min) return min; else return 0x80000000;
+        }
+        // Assume 64 bit integers supported. Use multiply and shift method
+        uint32_t interval;                  // Length of interval
+        uint64_t longran;                   // Random bits * interval
+        uint32_t iran;                      // Longran / 2^32
 
-/** generates a random number on [0,1) with 53-bit resolution
-    using 32bit integer.
- */
-inline static double genrand_res53_mix(void) 
-{ 
-    uint32_t x, y;
-
-    x = gen_rand32();
-    y = gen_rand32();
-    return to_res53_mix(x, y);
-} 
-#endif
+        interval = (uint32_t)(max - min + 1);
+        longran  = (uint64_t)BRandom() * interval;
+        iran = (uint32_t)(longran >> 32);
+        // Convert back to signed and return result
+        return (int32_t)iran + min;
+    }
+
+    uint32_t URandom(uint32_t min, uint32_t max)
+    {
+        // Output random integer in the interval min <= x <= max
+        // Slightly inaccurate if (max-min+1) is not a power of 2
+        if (max <= min) {
+            if (max == min) return min; else return 0;
+        }
+        // Assume 64 bit integers supported. Use multiply and shift method
+        uint32_t interval;                  // Length of interval
+        uint64_t longran;                   // Random bits * interval
+        uint32_t iran;                      // Longran / 2^32
+
+        interval = (uint32_t)(max - min + 1);
+        longran  = (uint64_t)BRandom() * interval;
+        iran = (uint32_t)(longran >> 32);
+        // Convert back to signed and return result
+        return iran + min;
+    }
+
+    double Random()                               // Output random floating point number
+    {
+        // Output random floating point number
+        if (ix >= SFMT_N*4-1) {
+            // Make sure we have at least two 32-bit numbers
+            Generate();
+        }
+        uint64_t r = *(uint64_t*)((uint32_t*)state+ix);
+        ix += 2;
+        // 52 bits resolution for compatibility with assembly version:
+        return (int64_t)(r >> 12) * (1./(67108864.0*67108864.0));
+    }
+
+    uint32_t BRandom()                            // Output random bits
+    {
+        // Output 32 random bits
+        uint32_t y;
+
+        if (ix >= SFMT_N*4) {
+            Generate();
+        }
+        y = ((uint32_t*)state)[ix++];
+        return y;
+    }
+private:
+    void Init2()                                   // Various initializations and period certification
+    {
+        // Various initializations and period certification
+        uint32_t i, j, temp;
+    
+        // Initialize mask
+        static const uint32_t maskinit[4] = {SFMT_MASK};
+        mask = _mm_loadu_si128((__m128i*)maskinit);
+
+        // Period certification
+        // Define period certification vector
+        static const uint32_t parityvec[4] = {SFMT_PARITY};
+
+        // Check if parityvec & state[0] has odd parity
+        temp = 0;
+        for (i = 0; i < 4; i++)
+            temp ^= parityvec[i] & ((uint32_t*)state)[i];
+
+        for (i = 16; i > 0; i >>= 1) temp ^= temp >> i;
+        if (!(temp & 1)) {
+            // parity is even. Certification failed
+            // Find a nonzero bit in period certification vector
+            for (i = 0; i < 4; i++) {
+                if (parityvec[i]) {
+                    for (j = 1; j; j <<= 1) {
+                        if (parityvec[i] & j) {
+                            // Flip the corresponding bit in state[0] to change parity
+                            ((uint32_t*)state)[i] ^= j;
+                            // Done. Exit i and j loops
+                            i = 5;  break;
+                        }
+                    }
+                }
+            }
+        }
+
+        // Generate first random numbers and set ix = 0
+        Generate();
+    }
+
+    void Generate()                                // Fill state array with new random numbers
+    {
+        // Fill state array with new random numbers
+        int i;
+        __m128i r, r1, r2;
+
+        r1 = state[SFMT_N - 2];
+        r2 = state[SFMT_N - 1];
+        for (i = 0; i < SFMT_N - SFMT_M; i++) {
+            r = sfmt_recursion(state[i], state[i + SFMT_M], r1, r2, mask);
+            state[i] = r;
+            r1 = r2;
+            r2 = r;
+        }
+        for (; i < SFMT_N; i++) {
+            r = sfmt_recursion(state[i], state[i + SFMT_M - SFMT_N], r1, r2, mask);
+            state[i] = r;
+            r1 = r2;
+            r2 = r;
+        }
+        ix = 0;
+    }
+
+    uint32_t ix;                                  // Index into state array
+    uint32_t LastInterval;                        // Last interval length for IRandom
+    uint32_t RLimit;                              // Rejection limit used by IRandom
+    __m128i  mask;                                // AND mask
+    __m128i  state[SFMT_N];                       // State vector for SFMT generator
+};
+
+#endif // SFMT_H