my program for solving the lower triangular equations,i testthe serial version and the version with AVX, result is:

Matrix Order : 10000x4
Matrix Band : 5000x4
Data Layout : 4x4 columned
Data Type : double
CPU : Corei7 2600
memory size : 16G
platform : VS2010

serial version: 0.357240s ( 100 cycles averaged )

SMID(AVX) version : 0.338360s ( 100 cycles averaged )

my question is theversion with AVX nearly unavailable, why?

this my code:

inline void solveL_pivot( double* x, const double* L )

{

double e0=x[ 0 ];

double e1=x[ 1 ];

double e2=x[ 2 ];

double e3=x[ 3 ];

e0/=L[ 0 ]; e1-=e0*L[ 1 ]; e2-=e0*L[ 2 ]; e3-=e0*L[ 3 ];

e1/=L[ 5 ]; e2-=e1*L[ 6 ]; e3-=e1*L[ 7 ];

e2/=L[ 10 ]; e3-=e2*L[ 11 ];

e3/=L[ 15 ];

x[ 0 ]=e0;

x[ 1 ]=e1;

x[ 2 ]=e2;

x[ 3 ]=e3;

}




void vsolveL_band( double* o, const double* c, unsigned int order, unsigned int band )

{ 

double* xproxy;

double* Lproxy;

__m256d ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7;

double* x=o;

double* L=( double* )c;

unsigned int const stride=band<<4; 

unsigned int const d=order-band;

unsigned int n=band-1;

 
for( unsigned int i=0; id ){ --n; }

xproxy=x;

Lproxy=( double* )L;

solveL_pivot( xproxy, Lproxy );

ymm0=_mm256_broadcast_sd( &xproxy[ 0 ] );

ymm1=_mm256_broadcast_sd( &xproxy[ 1 ] );

ymm2=_mm256_broadcast_sd( &xproxy[ 2 ] );

ymm3=_mm256_broadcast_sd( &xproxy[ 3 ] );

for( unsigned int k=0; kd ){ --n; }

solveL_pivot( o, c );

solveL_update( o, c+16, n );

o+=4; c+=stride;

}

solveL_pivot( o, c );

}