void Date::add_m() {
++mth;
if (mth >= mpy) {
mth = 0;
add_y();
}
}
void PlotAgent::add_point(double x, double y, const string& v)
{
assert(ndim == 3);
plot_os << x << '\t' << y << '\t' << v << '\n';
add_x(x);
add_y(y);
add_v(atof(v.chars()));
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
BLASLONG j;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *ap[4];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG n2;
FLOAT xbuffer[4],*ybuffer;
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
n1 = n / 4 ;
n2 = n % 4 ;
m1 = m - ( m % 16 );
m2 = (m % NBMAX) - (m % 16) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
x_ptr = x;
zero_y(NB,ybuffer);
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = alpha * x_ptr[0];
x_ptr += inc_x;
xbuffer[1] = alpha * x_ptr[0];
x_ptr += inc_x;
xbuffer[2] = alpha * x_ptr[0];
x_ptr += inc_x;
xbuffer[3] = alpha * x_ptr[0];
x_ptr += inc_x;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
sgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
a_ptr += 4 * lda;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = alpha * x_ptr[0];
x_ptr += inc_x;
sgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
add_y(NB,ybuffer,y_ptr,inc_y);
a += NB;
y_ptr += NB * inc_y;
}
j=0;
while ( j < (m % 16))
{
a_ptr = a;
x_ptr = x;
FLOAT temp = 0.0;
for( i = 0; i < n; i++ )
{
temp += a_ptr[0] * x_ptr[0];
a_ptr += lda;
x_ptr += inc_x;
}
y_ptr[0] += alpha * temp;
y_ptr += inc_y;
a++;
j++;
}
return(0);
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *ap[4];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG m3;
BLASLONG n2;
BLASLONG lda4;
FLOAT xbuffer[8],*ybuffer;
#if 0
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
#endif
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
lda4 = 4 * lda;
n1 = n / 4 ;
n2 = n % 4 ;
m3 = m % 4;
m1 = m - ( m % 4 );
m2 = (m % NBMAX) - (m % 4) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
x_ptr = x;
//zero_y(NB,ybuffer);
memset(ybuffer,0,NB*16);
if ( inc_x == 2 )
{
for( i = 0; i < n1 ; i++)
{
zgemv_kernel_4x4(NB,ap,x_ptr,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
x_ptr += 8;
}
if ( n2 & 2 )
{
zgemv_kernel_4x2(NB,ap,x_ptr,ybuffer);
x_ptr += 4;
a_ptr += 2 * lda;
}
if ( n2 & 1 )
{
zgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer);
/* x_ptr += 2;
a_ptr += lda; */
}
}
else
{
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;
zgemv_kernel_4x4(NB,ap,xbuffer,ybuffer);
ap[0] += lda4;
ap[1] += lda4;
ap[2] += lda4;
ap[3] += lda4;
a_ptr += lda4;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
zgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
}
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}
if ( m3 == 0 ) return(0);
if ( m3 == 1 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
if ( lda == 2 && inc_x == 2 )
{
for( i=0 ; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] - a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] + a_ptr[3] * x_ptr[2];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r += a_ptr[2] * x_ptr[2] + a_ptr[3] * x_ptr[3];
temp_i += a_ptr[2] * x_ptr[3] - a_ptr[3] * x_ptr[2];
#endif
a_ptr += 4;
x_ptr += 4;
}
for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += 2;
x_ptr += 2;
}
}
else
{
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
return(0);
}
if ( m3 == 2 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
if ( lda == 4 && inc_x == 2 )
{
for( i = 0; i < (n & -2); i+=2 )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r0 += a_ptr[4] * x_ptr[2] - a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] + a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] - a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] + a_ptr[7] * x_ptr[2];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r0 += a_ptr[4] * x_ptr[2] + a_ptr[5] * x_ptr[3];
temp_i0 += a_ptr[4] * x_ptr[3] - a_ptr[5] * x_ptr[2];
temp_r1 += a_ptr[6] * x_ptr[2] + a_ptr[7] * x_ptr[3];
temp_i1 += a_ptr[6] * x_ptr[3] - a_ptr[7] * x_ptr[2];
#endif
a_ptr += 8;
x_ptr += 4;
}
for( ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif
a_ptr += 4;
x_ptr += 2;
}
}
else
{
for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
#endif
return(0);
}
if ( m3 == 3 )
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r0 = 0.0;
FLOAT temp_i0 = 0.0;
FLOAT temp_r1 = 0.0;
FLOAT temp_i1 = 0.0;
FLOAT temp_r2 = 0.0;
FLOAT temp_i2 = 0.0;
if ( lda == 6 && inc_x == 2 )
{
for( i=0 ; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif
a_ptr += 6;
x_ptr += 2;
}
}
else
{
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
#else
temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 - alpha_i * temp_i2;
y_ptr[1] += alpha_r * temp_i2 + alpha_i * temp_r2;
#else
y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
y_ptr += inc_y;
y_ptr[0] += alpha_r * temp_r2 + alpha_i * temp_i2;
y_ptr[1] -= alpha_r * temp_i2 - alpha_i * temp_r2;
#endif
return(0);
}
return(0);
}
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
{
BLASLONG i;
BLASLONG j;
FLOAT *a_ptr;
FLOAT *x_ptr;
FLOAT *y_ptr;
FLOAT *ap[4];
BLASLONG n1;
BLASLONG m1;
BLASLONG m2;
BLASLONG n2;
FLOAT xbuffer[8],*ybuffer;
#if 0
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
#endif
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
ybuffer = buffer;
inc_x *= 2;
inc_y *= 2;
lda *= 2;
n1 = n / 4 ;
n2 = n % 4 ;
m1 = m - ( m % 16 );
m2 = (m % NBMAX) - (m % 16) ;
y_ptr = y;
BLASLONG NB = NBMAX;
while ( NB == NBMAX )
{
m1 -= NB;
if ( m1 < 0)
{
if ( m2 == 0 ) break;
NB = m2;
}
a_ptr = a;
x_ptr = x;
zero_y(NB,ybuffer);
for( i = 0; i < n1 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
xbuffer[2] = x_ptr[0];
xbuffer[3] = x_ptr[1];
x_ptr += inc_x;
xbuffer[4] = x_ptr[0];
xbuffer[5] = x_ptr[1];
x_ptr += inc_x;
xbuffer[6] = x_ptr[0];
xbuffer[7] = x_ptr[1];
x_ptr += inc_x;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
a_ptr += 4 * lda;
}
for( i = 0; i < n2 ; i++)
{
xbuffer[0] = x_ptr[0];
xbuffer[1] = x_ptr[1];
x_ptr += inc_x;
zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
}
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
a += 2 * NB;
y_ptr += NB * inc_y;
}
j=0;
while ( j < (m % 16))
{
a_ptr = a;
x_ptr = x;
FLOAT temp_r = 0.0;
FLOAT temp_i = 0.0;
for( i = 0; i < n; i++ )
{
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
#else
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
#endif
a_ptr += lda;
x_ptr += inc_x;
}
#if !defined(XCONJ)
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
#else
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
#endif
y_ptr += inc_y;
a+=2;
j++;
}
return(0);
}
