一些有效的快速嵌汇编数字计算函数

查询CPU支持属性函数,输入查询码,返回结果到dwRet中,本人原创.

关于cpuid指令详见VC2005: ms-help://MS.VSCC.v80/MS.MSDN.v80/MS.VisualStudio.v80.chs/dv_vclang/html/f8c344d3-91bf-405f-8622-cb0e337a6bdc.htm

inline void DoCPUID(DWORD nType, DWORD dwRet[4]) { __asm { mov eax, nType _emit 0x0F // cpuid 指令号为 0FA2h，并不是所有 CPU 都支持 cpuid 指令。 _emit 0xA2 mov edi, dword ptr [dwRet] mov dword ptr[edi], eax add edi, 4 mov dword ptr[edi], ebx add edi, 4 mov dword ptr[edi], ecx add edi, 4 mov dword ptr[edi], edx mov eax,0 mov ebx,0 mov ecx,0 mov edx,0 mov edi,0 } }

 

快速开方函数,此函数需要CPU支持SEE指令集,来源Intel的SEE指令集文档

inline float Sqrt_SSE(float x) { float ret = 0.f; _asm { sqrtss xmm0, x movss ret, xmm0 } return ret; }

 

AMD内存复制函数,来源AMD

inline void Memcpy(void* Dst, const void* Src, INT Count) { __asm { mov ecx, [Count] ; number of bytes to copy mov edi, [Dst] ; destination mov esi, [Src] ; source mov ebx, ecx ; keep a copy of count cld cmp ecx, TINY_BLOCK_COPY jb $memcpy_ic_3 ; tiny? skip mmx copy cmp ecx, 32*1024 ; don't align between 32k-64k because jbe $memcpy_do_align ; it appears to be slower cmp ecx, 64*1024 jbe $memcpy_align_done $memcpy_do_align: mov ecx, 8 ; a trick that's faster than rep movsb... sub ecx, edi ; align destination to qword and ecx, 111b ; get the low bits sub ebx, ecx ; update copy count neg ecx ; set up to jump into the array add ecx, offset $memcpy_align_done jmp ecx ; jump to array of movsb's align 4 movsb movsb movsb movsb movsb movsb movsb movsb $memcpy_align_done: ; destination is dword aligned mov ecx, ebx ; number of bytes left to copy shr ecx, 6 ; get 64-byte block count jz $memcpy_ic_2 ; finish the last few bytes cmp ecx, IN_CACHE_COPY/64 ; too big 4 cache? use uncached copy jae $memcpy_uc_test // This is small block copy that uses the MMX registers to copy 8 bytes // at a time. It uses the "unrolled loop" optimization, and also uses // the software prefetch instruction to get the data into the cache. align 16 $memcpy_ic_1: ; 64-byte block copies, in-cache copy prefetchnta [esi + (200*64/34+192)] ; start reading ahead movq mm0, [esi+0] ; read 64 bits movq mm1, [esi+8] movq [edi+0], mm0 ; write 64 bits movq [edi+8], mm1 ; note: the normal movq writes the movq mm2, [esi+16] ; data to cache; a cache line will be movq mm3, [esi+24] ; allocated as needed, to store the data movq [edi+16], mm2 movq [edi+24], mm3 movq mm0, [esi+32] movq mm1, [esi+40] movq [edi+32], mm0 movq [edi+40], mm1 movq mm2, [esi+48] movq mm3, [esi+56] movq [edi+48], mm2 movq [edi+56], mm3 add esi, 64 ; update source pointer add edi, 64 ; update destination pointer dec ecx ; count down jnz $memcpy_ic_1 ; last 64-byte block? $memcpy_ic_2: mov ecx, ebx ; has valid low 6 bits of the byte count $memcpy_ic_3: shr ecx, 2 ; dword count and ecx, 1111b ; only look at the "remainder" bits neg ecx ; set up to jump into the array add ecx, offset $memcpy_last_few jmp ecx ; jump to array of movsd's $memcpy_uc_test: cmp ecx, UNCACHED_COPY/64 ; big enough? use block prefetch copy jae $memcpy_bp_1 $memcpy_64_test: or ecx, ecx ; tail end of block prefetch will jump here jz $memcpy_ic_2 ; no more 64-byte blocks left // For larger blocks, which will spill beyond the cache, it's faster to // use the Streaming Store instruction MOVNTQ. This write instruction // bypasses the cache and writes straight to main memory. This code also // uses the software prefetch instruction to pre-read the data. align 16 $memcpy_uc_1: ; 64-byte blocks, uncached copy prefetchnta [esi + (200*64/34+192)] ; start reading ahead movq mm0,[esi+0] ; read 64 bits add edi,64 ; update destination pointer movq mm1,[esi+8] add esi,64 ; update source pointer movq mm2,[esi-48] movntq [edi-64], mm0 ; write 64 bits, bypassing the cache movq mm0,[esi-40] ; note: movntq also prevents the CPU movntq [edi-56], mm1 ; from READING the destination address movq mm1,[esi-32] ; into the cache, only to be over-written movntq [edi-48], mm2 ; so that also helps performance movq mm2,[esi-24] movntq [edi-40], mm0 movq mm0,[esi-16] movntq [edi-32], mm1 movq mm1,[esi-8] movntq [edi-24], mm2 movntq [edi-16], mm0 dec ecx movntq [edi-8], mm1 jnz $memcpy_uc_1 ; last 64-byte block? jmp $memcpy_ic_2 ; almost done // For the largest size blocks, a special technique called Block Prefetch // can be used to accelerate the read operations. Block Prefetch reads // one address per cache line, for a series of cache lines, in a short loop. // This is faster than using software prefetch. The technique is great for // getting maximum read bandwidth, especially in DDR memory systems. $memcpy_bp_1: ; large blocks, block prefetch copy cmp ecx, CACHEBLOCK ; big enough to run another prefetch loop? jl $memcpy_64_test ; no, back to regular uncached copy mov eax, CACHEBLOCK / 2 ; block prefetch loop, unrolled 2X add esi, CACHEBLOCK * 64 ; move to the top of the block align 16 $memcpy_bp_2: mov edx, [esi-64] ; grab one address per cache line mov edx, [esi-128] ; grab one address per cache line sub esi, 128 ; go reverse order to suppress HW prefetcher dec eax ; count down the cache lines jnz $memcpy_bp_2 ; keep grabbing more lines into cache mov eax, CACHEBLOCK ; now that it's in cache, do the copy align 16 $memcpy_bp_3: movq mm0, [esi ] ; read 64 bits movq mm1, [esi+ 8] movq mm2, [esi+16] movq mm3, [esi+24] movq mm4, [esi+32] movq mm5, [esi+40] movq mm6, [esi+48] movq mm7, [esi+56] add esi, 64 ; update source pointer movntq [edi ], mm0 ; write 64 bits, bypassing cache movntq [edi+ 8], mm1 ; note: movntq also prevents the CPU movntq [edi+16], mm2 ; from READING the destination address movntq [edi+24], mm3 ; into the cache, only to be over-written, movntq [edi+32], mm4 ; so that also helps performance movntq [edi+40], mm5 movntq [edi+48], mm6 movntq [edi+56], mm7 add edi, 64 ; update dest pointer dec eax ; count down jnz $memcpy_bp_3 ; keep copying sub ecx, CACHEBLOCK ; update the 64-byte block count jmp $memcpy_bp_1 ; keep processing chunks // The smallest copy uses the X86 "movsd" instruction, in an optimized // form which is an "unrolled loop". Then it handles the last few bytes. align 4 movsd movsd ; perform last 1-15 dword copies movsd movsd movsd movsd movsd movsd movsd movsd ; perform last 1-7 dword copies movsd movsd movsd movsd movsd movsd $memcpy_last_few: ; dword aligned from before movsd's mov ecx, ebx ; has valid low 2 bits of the byte count and ecx, 11b ; the last few cows must come home jz $memcpy_final ; no more, let's leave rep movsb ; the last 1, 2, or 3 bytes $memcpy_final: emms ; clean up the MMX state sfence ; flush the write buffer //mov eax, [dest] ; ret value = destination pointer } }

 

内存清零函数,来源Unreal 3 Engine

inline void Memzero(void* Dest, int Count) { __asm { mov ecx, [Count] mov edi, [Dest] xor eax, eax mov edx, ecx shr ecx, 2 and edx, 3 rep stosd mov ecx, edx rep stosb } }

 

快速Hash函数,来源Lua

inline DWORD Hash(const TCHAR* szStr) { size_t nLength = _tcslen(szStr); DWORD HashVL = (DWORD)nLength; size_t step = (nLength>>5) + 1; for (size_t i = nLength; i >= step; i -= step) HashVL = HashVL ^ ((HashVL<<5) + (HashVL>>2) + (DWORD)szStr[i - 1]); return HashVL; }

 

测试CPU速度函数,来源Unreal 3 Engine

inline double GetCPUSpeed() { timeBeginPeriod(1); double GSecondsPerCycle = 1.f; for (int i = 0; i < 3; i++) { LARGE_INTEGER StartCycles, EndCycles; volatile double DoubleDummy = 0.0; volatile int IntDummy = 0; __asm { rdtsc mov StartCycles.HighPart, edx mov StartCycles.LowPart, eax } DWORD StartMsec = timeGetTime(), EndMsec = StartMsec; while( EndMsec-StartMsec < 150 ) { DoubleDummy += (double)FastSqrt((float)DoubleDummy) + 3.14; IntDummy *= ((int)(DoubleDummy) + 555) / 13; EndMsec = timeGetTime(); } __asm { rdtsc mov EndCycles.HighPart, edx mov EndCycles.LowPart, eax } double C1 = (double)(signed __int64)(((unsigned __int64)StartCycles.LowPart) + ((unsigned __int64)StartCycles.HighPart<<32)), C2 = (double)(signed __int64)(((unsigned __int64)EndCycles.LowPart) + ((unsigned __int64)EndCycles.HighPart<<32)); double fRet = 1.0 / (1000.f * ( C2 - C1 ) / (EndMsec - StartMsec)); GSecondsPerCycle = GSecondsPerCycle <= fRet ? GSecondsPerCycle : fRet; } timeEndPeriod(1); return 0.000001 / GSecondsPerCycle; }

 

关于DoCPUID的用法如下

// 取CPU制造商，注意是ebx,edx,ecx char szCPUVender[13]; DWORD dwCPUIDVender[4]; DoCPUID(0x00000000, dwCPUIDVender); Memcpy(szCPUVender + 0, dwCPUIDVender + 1, 4); Memcpy(szCPUVender + 4, dwCPUIDVender + 3, 4); Memcpy(szCPUVender + 8, dwCPUIDVender + 2, 4); // 取CPU支持特性 DWORD dwCPUSpecialityEDX; DWORD dwCPUSpecialityECX; DWORD dwCPUIDSupport[4]; DoCPUID(0x00000001, dwCPUIDSupport); dwCPUSpecialityECX = dwCPUIDSupport[2]; dwCPUSpecialityEDX = dwCPUIDSupport[3]; // 取CPU商标信息 char szCPUBrand[49]; DWORD dwCPUIDBrand[4 * 3]; DoCPUID(0x80000002, dwCPUIDBrand + 0); DoCPUID(0x80000003, dwCPUIDBrand + 4); DoCPUID(0x80000004, dwCPUIDBrand + 8); Memcpy(szCPUBrand, dwCPUIDBrand, 4 * 3 * 4);