Menu
▾
▴
CVS: blob/src/lib crc32.c,1.2,1.3
|
From: Erik M. <er...@us...> - 2002-01-21 13:13:36
|
Update of /cvsroot/blob/blob/src/lib
In directory usw-pr-cvs1:/tmp/cvs-serv13855/src/lib
Modified Files:
crc32.c
Log Message:
Careful profiling and some hand loop unrolling makes crc32() faster than
the table implementation. Also added lots of comments.
Index: crc32.c
===================================================================
RCS file: /cvsroot/blob/blob/src/lib/crc32.c,v
retrieving revision 1.2
retrieving revision 1.3
diff -u -d -r1.2 -r1.3
--- crc32.c 2002/01/20 23:26:46 1.2
+++ crc32.c 2002/01/21 13:13:33 1.3
@@ -1,7 +1,8 @@
/*
* crc32.c: calculate CRC32
*
- * Copyright (C) 2002 Erik Mouw <J.A...@it...>
+ * Copyright (C) 2002 Erik Mouw <J.A...@it...> and
+ * Jan-Derk Bakker <J.D...@it...>
* based on linux/drivers/net/am79c961a.c
* Copyright (C) 1995-2001 Russell King <rm...@ar...>
*
@@ -22,15 +23,160 @@
*
*/
/*
- * Note: this implementation doesn't use the usual lookup table
- * implementation by Gary S. Brown. The lookup table is quite large
- * (1kB) and it's quite slow because we run with d-cache enabled. The
- * i-cache is enabled, so computing the crc will be a lot faster. And
- * of course this implementation is a lot smaller :)
- *
+ * There are many ways to calculate a CRC32. The most common is the
+ * table lookup variant by Gary S. Brown:
+ *
+ * const u32 crc32_table[256] = {
+ * 0x00000000L, 0x77073096L, 0xee0e612cL, 0x990951baL, 0x076dc419L,
+ * 0x706af48fL, 0xe963a535L, 0x9e6495a3L, 0x0edb8832L, 0x79dcb8a4L,
+ * 0xe0d5e91eL, 0x97d2d988L, 0x09b64c2bL, 0x7eb17cbdL, 0xe7b82d07L,
+ * 0x90bf1d91L, 0x1db71064L, 0x6ab020f2L, 0xf3b97148L, 0x84be41deL,
+ * 0x1adad47dL, 0x6ddde4ebL, 0xf4d4b551L, 0x83d385c7L, 0x136c9856L,
+ * 0x646ba8c0L, 0xfd62f97aL, 0x8a65c9ecL, 0x14015c4fL, 0x63066cd9L,
+ * 0xfa0f3d63L, 0x8d080df5L, 0x3b6e20c8L, 0x4c69105eL, 0xd56041e4L,
+ * 0xa2677172L, 0x3c03e4d1L, 0x4b04d447L, 0xd20d85fdL, 0xa50ab56bL,
+ * 0x35b5a8faL, 0x42b2986cL, 0xdbbbc9d6L, 0xacbcf940L, 0x32d86ce3L,
+ * 0x45df5c75L, 0xdcd60dcfL, 0xabd13d59L, 0x26d930acL, 0x51de003aL,
+ * 0xc8d75180L, 0xbfd06116L, 0x21b4f4b5L, 0x56b3c423L, 0xcfba9599L,
+ * 0xb8bda50fL, 0x2802b89eL, 0x5f058808L, 0xc60cd9b2L, 0xb10be924L,
+ * 0x2f6f7c87L, 0x58684c11L, 0xc1611dabL, 0xb6662d3dL, 0x76dc4190L,
+ * 0x01db7106L, 0x98d220bcL, 0xefd5102aL, 0x71b18589L, 0x06b6b51fL,
+ * 0x9fbfe4a5L, 0xe8b8d433L, 0x7807c9a2L, 0x0f00f934L, 0x9609a88eL,
+ * 0xe10e9818L, 0x7f6a0dbbL, 0x086d3d2dL, 0x91646c97L, 0xe6635c01L,
+ * 0x6b6b51f4L, 0x1c6c6162L, 0x856530d8L, 0xf262004eL, 0x6c0695edL,
+ * 0x1b01a57bL, 0x8208f4c1L, 0xf50fc457L, 0x65b0d9c6L, 0x12b7e950L,
+ * 0x8bbeb8eaL, 0xfcb9887cL, 0x62dd1ddfL, 0x15da2d49L, 0x8cd37cf3L,
+ * 0xfbd44c65L, 0x4db26158L, 0x3ab551ceL, 0xa3bc0074L, 0xd4bb30e2L,
+ * 0x4adfa541L, 0x3dd895d7L, 0xa4d1c46dL, 0xd3d6f4fbL, 0x4369e96aL,
+ * 0x346ed9fcL, 0xad678846L, 0xda60b8d0L, 0x44042d73L, 0x33031de5L,
+ * 0xaa0a4c5fL, 0xdd0d7cc9L, 0x5005713cL, 0x270241aaL, 0xbe0b1010L,
+ * 0xc90c2086L, 0x5768b525L, 0x206f85b3L, 0xb966d409L, 0xce61e49fL,
+ * 0x5edef90eL, 0x29d9c998L, 0xb0d09822L, 0xc7d7a8b4L, 0x59b33d17L,
+ * 0x2eb40d81L, 0xb7bd5c3bL, 0xc0ba6cadL, 0xedb88320L, 0x9abfb3b6L,
+ * 0x03b6e20cL, 0x74b1d29aL, 0xead54739L, 0x9dd277afL, 0x04db2615L,
+ * 0x73dc1683L, 0xe3630b12L, 0x94643b84L, 0x0d6d6a3eL, 0x7a6a5aa8L,
+ * 0xe40ecf0bL, 0x9309ff9dL, 0x0a00ae27L, 0x7d079eb1L, 0xf00f9344L,
+ * 0x8708a3d2L, 0x1e01f268L, 0x6906c2feL, 0xf762575dL, 0x806567cbL,
+ * 0x196c3671L, 0x6e6b06e7L, 0xfed41b76L, 0x89d32be0L, 0x10da7a5aL,
+ * 0x67dd4accL, 0xf9b9df6fL, 0x8ebeeff9L, 0x17b7be43L, 0x60b08ed5L,
+ * 0xd6d6a3e8L, 0xa1d1937eL, 0x38d8c2c4L, 0x4fdff252L, 0xd1bb67f1L,
+ * 0xa6bc5767L, 0x3fb506ddL, 0x48b2364bL, 0xd80d2bdaL, 0xaf0a1b4cL,
+ * 0x36034af6L, 0x41047a60L, 0xdf60efc3L, 0xa867df55L, 0x316e8eefL,
+ * 0x4669be79L, 0xcb61b38cL, 0xbc66831aL, 0x256fd2a0L, 0x5268e236L,
+ * 0xcc0c7795L, 0xbb0b4703L, 0x220216b9L, 0x5505262fL, 0xc5ba3bbeL,
+ * 0xb2bd0b28L, 0x2bb45a92L, 0x5cb36a04L, 0xc2d7ffa7L, 0xb5d0cf31L,
+ * 0x2cd99e8bL, 0x5bdeae1dL, 0x9b64c2b0L, 0xec63f226L, 0x756aa39cL,
+ * 0x026d930aL, 0x9c0906a9L, 0xeb0e363fL, 0x72076785L, 0x05005713L,
+ * 0x95bf4a82L, 0xe2b87a14L, 0x7bb12baeL, 0x0cb61b38L, 0x92d28e9bL,
+ * 0xe5d5be0dL, 0x7cdcefb7L, 0x0bdbdf21L, 0x86d3d2d4L, 0xf1d4e242L,
+ * 0x68ddb3f8L, 0x1fda836eL, 0x81be16cdL, 0xf6b9265bL, 0x6fb077e1L,
+ * 0x18b74777L, 0x88085ae6L, 0xff0f6a70L, 0x66063bcaL, 0x11010b5cL,
+ * 0x8f659effL, 0xf862ae69L, 0x616bffd3L, 0x166ccf45L, 0xa00ae278L,
+ * 0xd70dd2eeL, 0x4e048354L, 0x3903b3c2L, 0xa7672661L, 0xd06016f7L,
+ * 0x4969474dL, 0x3e6e77dbL, 0xaed16a4aL, 0xd9d65adcL, 0x40df0b66L,
+ * 0x37d83bf0L, 0xa9bcae53L, 0xdebb9ec5L, 0x47b2cf7fL, 0x30b5ffe9L,
+ * 0xbdbdf21cL, 0xcabac28aL, 0x53b39330L, 0x24b4a3a6L, 0xbad03605L,
+ * 0xcdd70693L, 0x54de5729L, 0x23d967bfL, 0xb3667a2eL, 0xc4614ab8L,
+ * 0x5d681b02L, 0x2a6f2b94L, 0xb40bbe37L, 0xc30c8ea1L, 0x5a05df1bL,
+ * 0x2d02ef8dL
+ * };
+ *
+ * u32 crc32_tbl(u32 val, const void *ss, int len)
+ * {
+ * const unsigned char *s = ss;
+ *
+ * while (--len >= 0)
+ * val = crc32_table[(val ^ *s++) & 0xff] ^ (val >> 8);
+ *
+ * return val;
+ * }
+ *
+ * The performance of this function on a StrongARM SA-1100 running at
+ * 221MHz with i-cache enabled and d-cache disabled is 1.29 s on a 3MB
+ * data block. The lookup table makes this function quite large (1068
+ * bytes), which is not what we want in blob.
+ *
+ * To get the code size smaller, we can take advantage of the fact
+ * that we run with i-cache enabled and d-cache disabled and compute
+ * the CRC on the fly. This function is by Russell King:
+ *
+ * static inline u32 update_crc(u32 crc, u8 byte)
+ * {
+ * int i;
+ * u32 polynome = POLYNOME;
+ *
+ * for (i = 8; i != 0; i--) {
+ * byte ^= crc & 1;
+ * crc >>= 1;
+ *
+ * if (byte & 1)
+ * crc ^= polynome;
+ *
+ * byte >>= 1;
+ * }
+ *
+ * return crc;
+ * }
+ *
+ * u32 crc32(u32 val, const void *data, int len)
+ * {
+ * u32 crc = val;
+ * u8 *tmp = (u8 *)data;
+ *
+ * while(len--)
+ * crc = update_crc(crc, *tmp++);
+ *
+ * return crc;
+ * }
+ *
+ * The function is a lot smaller: only 84 bytes, but unfortunately it
+ * is three times as slow on our benchmark: 3.13 seconds.
+ *
+ * Jan-Derk Bakker suggested to rewrite update_crc() like this:
+ *
+ * static inline u32 update_crc(u32 crc, u8 byte)
+ * {
+ * int i;
+ * u32 polynome = POLYNOME;
+ *
+ * crc ^= byte;
+ *
+ * for (i = 8; i != 0; i--) {
+ * if(crc & 1) {
+ * crc >>= 1;
+ * crc ^= polynome;
+ * }
+ * else
+ * crc >>= 1;
+ * }
+ *
+ * return crc;
+ * }
+ *
+ * The compiler generates a very tight inner loop for this:
+ *
+ * c020354: tst r0, #1 ; 0x1
+ * eorne r0, lr, r0, lsr #1
+ * moveq r0, r0, lsr #1
+ * subs r12, r12, #1 ; 0x1
+ * bne c020354
+ *
+ * Which really pays off: code size is 68 bytes and it scores 2.25 s
+ * on the benchmark.
+ *
+ * Unfortunately it is still slower than the table lookup variant, and
+ * some analysis shows that the for() loop is the culprit: the
+ * subtract costs 1 clock cycle, and the conditional branch another
+ * three, so we waste 32 cycles for each CRC byte in the loop.
+ *
+ * To get the code faster, we do loop unrolling by hand (like the
+ * function is right now). The code size almost doubles (124 bytes),
+ * but the speed almost halves: 1.23 s on the benchmark. Not a bad
+ * result at all: faster than the table implementation at only 12% of
+ * the size.
+ *
*/
-
#ident "$Id$"
#ifdef HAVE_CONFIG_H
@@ -47,47 +193,47 @@
-static inline u32 update_crc(u32 crc, u8 byte)
-{
- int i;
- u32 polynome = POLYNOME;
- /* the above might look stupid, but it is a cludge to get
- * better code. with the -Os flag, the compiler moves the
- * POLYNOME into the .rodata segment and reloads it on every
- * run. this kinda defeats the purpose of doing CRC32 without
- * a lookup table. by moving the polynome explicitly out of
- * the loop, the compiler will allocate a register for it
- * making the resulting code faster. -- Erik
- */
-
- for (i = 8; i != 0; i--) {
- byte ^= crc & 1;
- crc >>= 1;
-
- if (byte & 1)
- crc ^= polynome;
-
- byte >>= 1;
- }
-
- return crc;
-}
-
-
-
-
/* calculate CRC32 on len bytes pointed by data. the usual
* initialisation is to put 0xffffffff in val, but by supplying a
* previous CRC value into it it can be used to calculate the CRC on
* streams of data
*/
-u32 crc32(u32 val, const void *data, int len)
+u32 crc32(u32 crc, const void *data, int len)
{
- u32 crc = val;
u8 *tmp = (u8 *)data;
+ u32 polynome = POLYNOME;
+ /* The above might look stupid, but it is a cludge to get
+ * better code. With the -Os flag, the compiler Moves the
+ * POLYNOME into the .rodata segment and reloads it on every
+ * run. This kinda defeats the purpose of doing CRC32 without
+ * a lookup table. By moving the polynome explicitly out of
+ * the loop, the compiler will allocate a register for it
+ * making the resulting code faster. JDB commented that the
+ * CSE module of GCC is broken, but it works perfectly well
+ * with -O2. It's just that memory size is more important than
+ * speed with -Os. -- Erik
+ */
- while(len--)
- crc = update_crc(crc, *tmp++);
+#define DO_CRC \
+ if(crc & 1) { \
+ crc >>= 1; \
+ crc ^= polynome; \
+ } else { \
+ crc >>= 1; \
+ }
+
+ while(len--) {
+ crc ^= *tmp++;
+
+ DO_CRC;
+ DO_CRC;
+ DO_CRC;
+ DO_CRC;
+ DO_CRC;
+ DO_CRC;
+ DO_CRC;
+ DO_CRC;
+ }
return crc;
}
|
