CRC-32C (Castagnoli) 校验算法优化(Delphi SSE优化算法应用之三)

CRC-32C (Castagnoli)  算法是 iSCSI 和SCTP 数据校验的算法，和常用CRC-32-IEEE 802.3算法所不同的是多项式常数CRC32C是0x1EDC6F41 ,CRC32是0x04C11DB7 ,也就是说由此生成的CRC表不同外算法是一模一样.

CRC32 常规算法如下：

function _CRC32CX86(Data: PByte; aLength: Integer): DWORD; const _CRC32CTable: array[0..255] of DWORD = ( $00000000, $F26B8303, $E13B70F7, $1350F3F4, //CRC32C Table $C79A971F, $35F1141C, $26A1E7E8, $D4CA64EB, $8AD958CF, $78B2DBCC, $6BE22838, $9989AB3B, $4D43CFD0, $BF284CD3, $AC78BF27, $5E133C24, $105EC76F, $E235446C, $F165B798, $030E349B, $D7C45070, $25AFD373, $36FF2087, $C494A384, $9A879FA0, $68EC1CA3, $7BBCEF57, $89D76C54, $5D1D08BF, $AF768BBC, $BC267848, $4E4DFB4B, $20BD8EDE, $D2D60DDD, $C186FE29, $33ED7D2A, $E72719C1, $154C9AC2, $061C6936, $F477EA35, $AA64D611, $580F5512, $4B5FA6E6, $B93425E5, $6DFE410E, $9F95C20D, $8CC531F9, $7EAEB2FA, $30E349B1, $C288CAB2, $D1D83946, $23B3BA45, $F779DEAE, $05125DAD, $1642AE59, $E4292D5A, $BA3A117E, $4851927D, $5B016189, $A96AE28A, $7DA08661, $8FCB0562, $9C9BF696, $6EF07595, $417B1DBC, $B3109EBF, $A0406D4B, $522BEE48, $86E18AA3, $748A09A0, $67DAFA54, $95B17957, $CBA24573, $39C9C670, $2A993584, $D8F2B687, $0C38D26C, $FE53516F, $ED03A29B, $1F682198, $5125DAD3, $A34E59D0, $B01EAA24, $42752927, $96BF4DCC, $64D4CECF, $77843D3B, $85EFBE38, $DBFC821C, $2997011F, $3AC7F2EB, $C8AC71E8, $1C661503, $EE0D9600, $FD5D65F4, $0F36E6F7, $61C69362, $93AD1061, $80FDE395, $72966096, $A65C047D, $5437877E, $4767748A, $B50CF789, $EB1FCBAD, $197448AE, $0A24BB5A, $F84F3859, $2C855CB2, $DEEEDFB1, $CDBE2C45, $3FD5AF46, $7198540D, $83F3D70E, $90A324FA, $62C8A7F9, $B602C312, $44694011, $5739B3E5, $A55230E6, $FB410CC2, $092A8FC1, $1A7A7C35, $E811FF36, $3CDB9BDD, $CEB018DE, $DDE0EB2A, $2F8B6829, $82F63B78, $709DB87B, $63CD4B8F, $91A6C88C, $456CAC67, $B7072F64, $A457DC90, $563C5F93, $082F63B7, $FA44E0B4, $E9141340, $1B7F9043, $CFB5F4A8, $3DDE77AB, $2E8E845F, $DCE5075C, $92A8FC17, $60C37F14, $73938CE0, $81F80FE3, $55326B08, $A759E80B, $B4091BFF, $466298FC, $1871A4D8, $EA1A27DB, $F94AD42F, $0B21572C, $DFEB33C7, $2D80B0C4, $3ED04330, $CCBBC033, $A24BB5A6, $502036A5, $4370C551, $B11B4652, $65D122B9, $97BAA1BA, $84EA524E, $7681D14D, $2892ED69, $DAF96E6A, $C9A99D9E, $3BC21E9D, $EF087A76, $1D63F975, $0E330A81, $FC588982, $B21572C9, $407EF1CA, $532E023E, $A145813D, $758FE5D6, $87E466D5, $94B49521, $66DF1622, $38CC2A06, $CAA7A905, $D9F75AF1, $2B9CD9F2, $FF56BD19, $0D3D3E1A, $1E6DCDEE, $EC064EED, $C38D26C4, $31E6A5C7, $22B65633, $D0DDD530, $0417B1DB, $F67C32D8, $E52CC12C, $1747422F, $49547E0B, $BB3FFD08, $A86F0EFC, $5A048DFF, $8ECEE914, $7CA56A17, $6FF599E3, $9D9E1AE0, $D3D3E1AB, $21B862A8, $32E8915C, $C083125F, $144976B4, $E622F5B7, $F5720643, $07198540, $590AB964, $AB613A67, $B831C993, $4A5A4A90, $9E902E7B, $6CFBAD78, $7FAB5E8C, $8DC0DD8F, $E330A81A, $115B2B19, $020BD8ED, $F0605BEE, $24AA3F05, $D6C1BC06, $C5914FF2, $37FACCF1, $69E9F0D5, $9B8273D6, $88D28022, $7AB90321, $AE7367CA, $5C18E4C9, $4F48173D, $BD23943E, $F36E6F75, $0105EC76, $12551F82, $E03E9C81, $34F4F86A, $C69F7B69, $D5CF889D, $27A40B9E, $79B737BA, $8BDCB4B9, $988C474D, $6AE7C44E, $BE2DA0A5, $4C4623A6, $5F16D052, $AD7D5351); var i: Integer; begin Result := $FFFFFFFF; for I := 0 to aLength - 1 do begin Result := (Result shr 8) xor _CRC32CTable[(Result and $FF) xor Data^]; Inc(Data); end; Result := not Result; end;  

 

CRC32C使用SSE4.2硬件指令优化算法部分代码如下：

function _CRC32CSSE(Data: PByte; aLength: Integer): DWORD; asm push esi push edx push ecx mov esi,eax mov eax,$FFFFFFFF test edx,edx jz @Exit test esi,esi jz @Exit mov ecx,edx shr ecx, 2 test ecx,ecx jz @Exit xor edx,edx @Alignment: crc32 eax,[edx*4+esi] inc edx cmp edx,ecx jb @Alignment @Exit: not eax pop ecx pop edx pop esi end;

 

以上2个不同实现方式在Intel Core i7 720QM 1.60GHz CPU上测试成绩如下：

 (数据采用随机算法生成，1M*100表示使用1M数据进行100次重复计算，数据量相当于100M)

 

--------------------------------------------------------------------------------------

|     数据量    |         常规算法时间      |     优化算法时间         |  快出百分比 |

--------------------------------------------------------------------------------------

| 1M *100    |  X86 Time:390ms     |   SSE Time:32ms     |  1218%    |

--------------------------------------------------------------------------------------

| 4M *100    |  X86 Time:1575ms   |   SSE Time:156ms   |  1009%    |

--------------------------------------------------------------------------------------

| 8M *100    |  X86 Time:3136ms   |   SSE Time:280ms   |  1120%    |

--------------------------------------------------------------------------------------

| 32M *100  |  X86 Time:12542ms   |   SSE Time:1092ms | 1148%   |

--------------------------------------------------------------------------------------

 

通过对比可以清楚的看到使用SSE4.2中的新指令crc32可以比常规CRC32C算法要快出最少10倍的效率，Intel新增的指令确实对常规某些算法提供了高效的解决方案，使用好它们将对我们在以后的开发中得到质的提升。

 

本例测试DEMO： http://download.csdn.net/source/3361729

 

*References   intel sse4 programming reference (D91561-001)

 http://software.intel.com/file/18187/