njtynhj

// 改自英文維基 http://en.wikipedia.org/wiki/Vector_processor
//SSE simd function for vectorized multiplication of 2 arrays with single-precision floatingpoint numbers
//1st param pointer on source/destination array, 2nd param 2. source array, 3rd param number of floats per array
void mul_asm(float* out, float* in, unsigned int leng)
 unsigned int count, rest;

 //compute if array is big enough for vector operation
 rest = (leng*4)%16;
 count = (leng*4)-rest;

 // vectorized part; 4 floats per loop iteration
 if (count>0)
 __asm __volatile__ (".intel_syntax noprefixnt"
 "loop: nt"
 "sub ecx,16 nt" // decrease address pointer by 4 floats
 "movups xmm0,[ebx+ecx] nt" // loads 4 floats in first register (xmm0)
 "movups xmm1,[eax+ecx] nt" // loads 4 floats in second register (xmm1)
 "mulps xmm0,xmm1 nt" // multiplies both vector registers
 "movups [eax+ecx],xmm0 nt" // write back the result to memory
 "jnz loop nt"
 ".att_syntax prefix nt"
 : : "a" (out), "b" (in), "c"(count), "d"(rest): "xmm0","xmm1");
 

 // scalar part; 1 float per loop iteration
 if (rest!=0)
 
 __asm __volatile__ (".intel_syntax noprefixnt"
 "add eax,ecx nt"
 "add ebx,ecx nt"

 "rest: nt"
 "sub edx,4 nt"
 "movss xmm0,[ebx+edx] nt" // load 1 float in first register (xmm0)
 "movss xmm1,[eax+edx] nt" // load 1 float in second register (xmm1)
 "mulss xmm0,xmm1 nt" // multiplies both scalar parts of registers
 "movss [eax+edx],xmm0 nt" // write back the resultnt"
 "jnz rest nt"
 ".att_syntax prefix nt"
 : : "a" (out), "b" (in), "c"(count), "d"(rest): "xmm0","xmm1");
 
 return;