Faster CRC32 implementation (slicing-by-4 for 32-bit platforms and slicing-by-8 for 64-bit)

git-svn-id: https://svn.eduke32.com/eduke32@4625 1a8010ca-5511-0410-912e-c29ae57300e0

polymer/eduke32/build/include/crc32.h

@@ -7,30 +7,16 @@
 extern "C" {
 #endif
 
-extern uint32_t crc32table[256];
+#define POLY 0xEDB88320
 
-void initcrc32table(void);
+#ifdef BITNESS64
+extern uint32_t crc32table[8][256];
+#else
+extern uint32_t crc32table[4][256];
+#endif
 
-uint32_t crc32once(uint8_t *blk, uint32_t len);
-
-static inline void crc32init(uint32_t *crcvar)
-{
-    if (!crcvar) return;
-    *crcvar = 0xffffffffl;
-}
-
-static inline void crc32block(uint32_t *crcvar, uint8_t *blk, uint32_t len)
-{
-    uint32_t crc = *crcvar;
-    while (len--) crc = crc32table[(crc ^ *(blk++)) & 0xffl] ^(crc >> 8);
-    *crcvar = crc;
-}
-
-static inline uint32_t crc32finish(uint32_t *crcvar)
-{
-    *crcvar = *crcvar ^ 0xffffffffl;
-    return *crcvar;
-}
+extern uint32_t crc32(const void* data, size_t length, uint32_t crc);
+extern void initcrc32table(void);
 
 #ifdef EXTERNC
 }

polymer/eduke32/build/src/crc32.c

@@ -1,26 +1,102 @@
+// based on http://create.stephan-brumme.com/crc32/Crc32.cpp, zlib license
+
+#include "compat.h"
 #include "crc32.h"
 
-uint32_t crc32table[256];
+uint32_t crc32(const void* data, size_t length, uint32_t crc)
+{
+    const uint32_t* current = (const uint32_t*) data;
+    uint8_t *currentChar;
+    crc = ~crc;
+
+#ifdef BITNESS64
+    // process eight bytes at once (Slicing-by-8)
+    while (length >= 8)
+    {
+#if B_BIG_ENDIAN != 0
+        uint32_t one = *current ^ B_SWAP32(crc);
+        uint32_t two = *(current+1);
+        crc  = crc32table[0][two       & 0xFF] ^
+               crc32table[1][(two>> 8) & 0xFF] ^
+               crc32table[2][(two>>16) & 0xFF] ^
+               crc32table[3][(two>>24) & 0xFF] ^
+               crc32table[4][one       & 0xFF] ^
+               crc32table[5][(one>> 8) & 0xFF] ^
+               crc32table[6][(one>>16) & 0xFF] ^
+               crc32table[7][(one>>24) & 0xFF];
+#else
+        uint32_t one = *current ^ crc;
+        uint32_t two = *(current+1);
+        crc  = crc32table[0][(two>>24) & 0xFF] ^
+               crc32table[1][(two>>16) & 0xFF] ^
+               crc32table[2][(two>> 8) & 0xFF] ^
+               crc32table[3][two       & 0xFF] ^
+               crc32table[4][(one>>24) & 0xFF] ^
+               crc32table[5][(one>>16) & 0xFF] ^
+               crc32table[6][(one>> 8) & 0xFF] ^
+               crc32table[7][one       & 0xFF];
+#endif
+        current += 2;
+        length -= 8;
+    }
+#else
+    // process four bytes at once (Slicing-by-4)
+    while (length >= 4)
+    {
+#if B_BIG_ENDIAN != 0
+        uint32_t one = *current++ ^ B_SWAP32(crc);
+        crc  = crc32table[0][one       & 0xFF] ^
+               crc32table[1][(one>> 8) & 0xFF] ^
+               crc32table[2][(one>>16) & 0xFF] ^
+               crc32table[3][(one>>24) & 0xFF];
+#else
+        uint32_t one = *current++ ^ crc;
+        crc  = crc32table[0][(one>>24) & 0xFF] ^
+               crc32table[1][(one>>16) & 0xFF] ^
+               crc32table[2][(one>> 8) & 0xFF] ^
+               crc32table[3][one       & 0xFF];
+#endif
+
+        length -= 4;
+    }
+#endif
+
+    currentChar = (uint8_t*) current;
+    // remaining 1 to 7 bytes (standard algorithm)
+    while (length-- > 0)
+        crc = (crc >> 8) ^ crc32table[0][(crc & 0xFF) ^ *currentChar++];
+
+    return ~crc;
+}
+
+#ifdef BITNESS64
+uint32_t crc32table[8][256];
+#else
+uint32_t crc32table[4][256];
+#endif
 
 void initcrc32table(void)
 {
-    uint32_t i,j,k;
-
-    // algorithm and polynomial same as that used by infozip's zip
-    for (i=0; i<256; i++)
+    int i;
+    for (i = 0; i <= 0xFF; i++)
     {
-        j = i;
-        for (k=8; k; k--)
-            j = (j&1) ? (0xedb88320L^(j>>1)) : (j>>1);
-        crc32table[i] = j;
+      uint32_t j, crc = i;
+      for (j = 0; j < 8; j++)
+        crc = (crc >> 1) ^ ((crc & 1) * POLY);
+      crc32table[0][i] = crc;
+    }
+
+    for (i = 0; i <= 0xFF; i++)
+    {
+      crc32table[1][i] = (crc32table[0][i] >> 8) ^ crc32table[0][crc32table[0][i] & 0xFF];
+      crc32table[2][i] = (crc32table[1][i] >> 8) ^ crc32table[0][crc32table[1][i] & 0xFF];
+      crc32table[3][i] = (crc32table[2][i] >> 8) ^ crc32table[0][crc32table[2][i] & 0xFF];
+
+#ifdef BITNESS64
+      crc32table[4][i] = (crc32table[3][i] >> 8) ^ crc32table[0][crc32table[3][i] & 0xFF];
+      crc32table[5][i] = (crc32table[4][i] >> 8) ^ crc32table[0][crc32table[4][i] & 0xFF];
+      crc32table[6][i] = (crc32table[5][i] >> 8) ^ crc32table[0][crc32table[5][i] & 0xFF];
+      crc32table[7][i] = (crc32table[6][i] >> 8) ^ crc32table[0][crc32table[6][i] & 0xFF];
+#endif
     }
 }
-
-uint32_t crc32once(uint8_t *blk, uint32_t len)
-{
-    uint32_t crc;
-
-    crc32init(&crc);
-    crc32block(&crc, blk, len);
-    return crc32finish(&crc);
-}

polymer/eduke32/build/src/engine.c

@@ -7923,22 +7923,22 @@ static int32_t loadtables(void)
             reciptable[i] = divscale30(2048, i+2048);
 
         for (i=0; i<=512; i++)
-            sintable[i] = (int16_t)(16384*sin(i*BANG2RAD));
+            sintable[i] = (int16_t)(16384.f * sinf((float)i * BANG2RAD));
         for (i=513; i<1024; i++)
             sintable[i] = sintable[1024-i];
         for (i=1024; i<2048; i++)
             sintable[i] = -sintable[i-1024];
 
         for (i=0; i<640; i++)
-            radarang[i] = (int16_t)(-64*atan((640-0.5-i)/160)/BANG2RAD);
+            radarang[i] = (int16_t)(atanf(((float)(640-i)-0.5f) * (1.f/160.f)) * (-64.f * (1.f/BANG2RAD)));
         for (i=0; i<640; i++)
             radarang[1279-i] = -radarang[i];
 
 #ifdef B_LITTLE_ENDIAN
         i = 0;
-        if (crc32once((uint8_t *)sintable, sizeof(sintable)) != 0xee1e7aba)
+        if (crc32((uint8_t *)sintable, sizeof(sintable), 0) != 0xee1e7aba)
             i |= 1;
-        if (crc32once((uint8_t *)radarang, 640*sizeof(radarang[0])) != 0xee893d92)
+        if (crc32((uint8_t *)radarang, 640*sizeof(radarang[0]), 0) != 0xee893d92)
             i |= 2;
 
         if (i != 0)
@@ -8155,7 +8155,7 @@ static int32_t loadpalette(void)
 #endif
 
     // If Duke3D 1.5 GRP or LameDuke, ...
-    if (crc32once((uint8_t *)transluc, 65536)==0x94a1fac6 || lamedukep)
+    if (crc32((uint8_t *)transluc, 65536, 0)==0x94a1fac6 || lamedukep)
     {
         int32_t i;
         // ... fix up translucency table so that transluc(255,x)

polymer/eduke32/build/src/osd.c

@@ -340,14 +340,12 @@ static int32_t _internal_osdfunc_fileinfo(const osdfuncparm_t *parm)
     length = kfilelength(i);
 
     crctime = getticks();
-    crc32init(&crc);
     do
     {
         j = kread(i,buf,256);
-        crc32block(&crc,(uint8_t *)buf,j);
+        crc = crc32((uint8_t *)buf,j,crc);
     }
     while (j == 256);
-    crc32finish(&crc);
     crctime = getticks() - crctime;
 
     klseek(i, 0, BSEEK_SET);

polymer/eduke32/source/grpscan.c

@@ -384,21 +384,19 @@ int32_t ScanGroups(void)
 
         {
             int32_t b, fh;
-            int32_t crcval;
+            int32_t crcval = 0;
 
             fh = openfrompath(sidx->name, BO_RDONLY|BO_BINARY, BS_IREAD);
             if (fh < 0) continue;
             if (Bfstat(fh, &st)) continue;
 
             initprintf(" Checksumming %s...", sidx->name);
-            crc32init((uint32_t *)&crcval);
             do
             {
                 b = read(fh, buf, BUFFER_SIZE);
-                if (b > 0) crc32block((uint32_t *)&crcval, (uint8_t *)buf, b);
+                if (b > 0) crcval = crc32((uint8_t *)buf, b, crcval);
             }
             while (b == BUFFER_SIZE);
-            crc32finish((uint32_t *)&crcval);
             close(fh);
             initprintf(" Done\n");
 

polymer/eduke32/source/net.c

@@ -766,7 +766,7 @@ void Net_SendChallenge()
     tempnetbuf[0] = PACKET_AUTH;
     B_BUF16(&tempnetbuf[1], BYTEVERSION);
     B_BUF16(&tempnetbuf[3], NETVERSION);
-    B_BUF32(&tempnetbuf[5], crc32once((uint8_t *)g_netPassword, Bstrlen(g_netPassword)));
+    B_BUF32(&tempnetbuf[5], crc32((uint8_t *)g_netPassword, Bstrlen(g_netPassword), 0));
     tempnetbuf[9] = myconnectindex;
 
     enet_peer_send(g_netClientPeer, CHAN_GAMESTATE, enet_packet_create(&tempnetbuf[0], 10, ENET_PACKET_FLAG_RELIABLE));
@@ -786,7 +786,7 @@ void Net_ReceiveChallenge(uint8_t *pbuf, int32_t packbufleng, ENetEvent *event)
         initprintf("Bad client protocol: version %u.%u\n", byteVersion, netVersion);
         return;
     }
-    if (crc != crc32once((uint8_t *)g_netPassword, Bstrlen(g_netPassword)))
+    if (crc != crc32((uint8_t *)g_netPassword, Bstrlen(g_netPassword), 0))
     {
         enet_peer_disconnect_later(event->peer, DISC_BAD_PASSWORD);
         initprintf("Bad password from client.\n");