-
Type:
Task
-
Resolution: Done
-
Affects Version/s: None
-
Component/s: None
There is some reason to believe hardware checksums are faster than our current checksums, although I'm not convinced our checksum functions are worth worrying about (for example, the default is to turn off checksums if block compression is configured, so we're not checksumming a significant amount of data in any application configured for block compression).
There is also a sourceforge project that's implementing the "Slicing by 8" algorithm we implemented, it might be worth review.
The following is (I'm told) the checksum functionality from LevelDB, plugged into WiredTiger.
diff --git a/src/support/cksum.c b/src/support/cksum.c index 7e9befe..b924db7 100644 --- a/src/support/cksum.c +++ b/src/support/cksum.c @@ -27,6 +27,13 @@ #include "wt_internal.h" +#if defined(__amd64) || defined(__x86_64) +#define USE_HARDWARE_CRC32 1 +#else +#undef USE_HARDWARE_CRC32 +#endif + +#ifdef USE_HARDWARE_CRC32 static const uint32_t g_crc_slicing[8][256] = { #ifdef WORDS_BIGENDIAN /* @@ -1078,6 +1085,7 @@ static const uint32_t g_crc_slicing[8][256] = { } #endif }; +#endif /* USE_HARDWARE_CRC32 */ /* * __wt_cksum -- @@ -1106,15 +1114,29 @@ __wt_cksum(const void *chunk, size_t len) /* Checksum one byte at a time to the first 4B boundary. */ for (p = chunk; ((uintptr_t)p & (sizeof(uint32_t) - 1)) != 0 && - len > 0; ++p, --len) + len > 0; ++p, --len) { +#ifdef USE_HARDWARE_CRC32 + __asm__ __volatile__( + ".byte 0xF2, 0x0F, 0x38, 0xF0, 0xF1" + : "=S" (crc) + : "0" (crc), "c" (*p)); +#else #ifdef WORDS_BIGENDIAN crc = g_crc_slicing[0][((crc >> 24) ^ *p) & 0xFF] ^ (crc << 8); #else crc = g_crc_slicing[0][(crc ^ *p) & 0xFF] ^ (crc >> 8); #endif +#endif + } /* Checksum in 8B chunks. */ for (nqwords = len / sizeof(uint64_t); nqwords; nqwords--) { +#ifdef USE_HARDWARE_CRC32 + __asm__ __volatile__ ( + ".byte 0xf2, 0x48, 0x0f, 0x38, 0xf0, 0xf1;" + : "=S"(crc) + : "S"(crc), "c"(*p)); +#else crc ^= *(uint32_t *)p; p += sizeof(uint32_t); next = *(uint32_t *)p; @@ -1139,22 +1161,32 @@ __wt_cksum(const void *chunk, size_t len) g_crc_slicing[1][(next >> 16) & 0xFF] ^ g_crc_slicing[0][(next >> 24)]; #endif +#endif } /* Checksum trailing bytes one byte at a time. */ + for (len &= 0x7; len > 0; ++p, len--) { +#ifdef USE_HARDWARE_CRC32 + __asm__ __volatile__( + ".byte 0xF2, 0x0F, 0x38, 0xF0, 0xF1" + : "=S" (crc) + : "0" (crc), "c" (*p)); +#else #ifdef WORDS_BIGENDIAN - for (len &= 0x7; len > 0; ++p, len--) crc = g_crc_slicing[0][((crc >> 24) ^ *p) & 0xFF] ^ (crc << 8); +#else + crc = g_crc_slicing[0][(crc ^ *p) & 0xFF] ^ (crc >> 8); +#endif +#endif + } +#ifdef WORDS_BIGENDIAN /* Do final byte swap to produce a result identical to little endian */ crc = ((crc << 24) & 0xFF000000) | ((crc << 8) & 0x00FF0000) | ((crc >> 8) & 0x0000FF00) | ((crc >> 24) & 0x000000FF); -#else - for (len &= 0x7; len > 0; ++p, len--) - crc = g_crc_slicing[0][(crc ^ *p) & 0xFF] ^ (crc >> 8); #endif return (~crc); }
- related to
-
- Closed