|
void | bl1_sdotaxmyv2 (int n, float *alpha, float *beta, float *x, int inc_x, float *u, int inc_u, float *rho, float *y, int inc_y, float *z, int inc_z) |
|
void | bl1_ddotaxmyv2 (int n, double *alpha, double *beta, double *x, int inc_x, double *u, int inc_u, double *rho, double *y, int inc_y, double *z, int inc_z) |
|
| if (inc_x !=1||inc_u !=1||inc_y !=1||inc_z !=1) |
|
| for (i=0;i< n_run;++i) |
|
| if (n_left > 0) |
|
void | bl1_cdotaxmyv2 (int n, scomplex *alpha, scomplex *beta, scomplex *x, int inc_x, scomplex *u, int inc_u, scomplex *rho, scomplex *y, int inc_y, scomplex *z, int inc_z) |
|
void | bl1_zdotaxmyv2 (int n, dcomplex *alpha, dcomplex *beta, dcomplex *x, int inc_x, dcomplex *u, int inc_u, dcomplex *rho, dcomplex *y, int inc_y, dcomplex *z, int inc_z) |
|
void bl1_ddotaxmyv2 |
( |
int |
n, |
|
|
double * |
alpha, |
|
|
double * |
beta, |
|
|
double * |
x, |
|
|
int |
inc_x, |
|
|
double * |
u, |
|
|
int |
inc_u, |
|
|
double * |
rho, |
|
|
double * |
y, |
|
|
int |
inc_y, |
|
|
double * |
z, |
|
|
int |
inc_z |
|
) |
| |
44 double* restrict
chi1;
46 double* restrict
psi1;
47 double* restrict
zeta1;
65 if ( (
unsigned long ) z % 16 != 0 )
67 if ( (
unsigned long ) x % 16 == 0 ||
68 (
unsigned long ) u % 16 == 0 ||
69 (
unsigned long ) y % 16 == 0 )
bl1_abort();
88 double chi1_c = *
chi1;
91 rho_c += chi1_c * upsilon_c;
101 a1v.
v = _mm_loaddup_pd( (
double* )alpha );
102 b1v.
v = _mm_loaddup_pd( (
double* )beta );
104 rho1v.
v = _mm_setzero_pd();
108 x1v.
v = _mm_load_pd( (
double* )
chi1 );
109 u1v.
v = _mm_load_pd( (
double* )
upsilon1 );
110 y1v.
v = _mm_load_pd( (
double* )
psi1 );
111 z1v.
v = _mm_load_pd( (
double* )
zeta1 );
113 rho1v.
v += x1v.
v * u1v.
v;
114 y1v.
v -= a1v.
v * x1v.
v;
115 z1v.
v -= b1v.
v * x1v.
v;
117 _mm_store_pd( (
double* )
psi1, y1v.
v );
118 _mm_store_pd( (
double* )
zeta1, z1v.
v );
126 rho_c += rho1v.
d[0] + rho1v.
d[1];
135 double chi1_c = *
chi1;
138 rho_c += chi1_c * upsilon_c;
chi1
Definition: bl1_axmyv2.c:366
double beta_c
Definition: bl1_dotaxmyv2.c:158
double alpha_c
Definition: bl1_dotaxmyv2.c:157
int n_left
Definition: bl1_dotaxmyv2.c:164
double *restrict upsilon1
Definition: bl1_dotaxmyv2.c:152
int n_pre
Definition: bl1_dotaxmyv2.c:162
double rho_c
Definition: bl1_dotaxmyv2.c:159
double *restrict psi1
Definition: bl1_dotaxmyv2.c:155
int n_run
Definition: bl1_dotaxmyv2.c:163
* rho
Definition: bl1_dotaxmyv2.c:258
int i
Definition: bl1_dotaxmyv2.c:160
double *restrict zeta1
Definition: bl1_dotaxmyv2.c:156
Definition: blis_type_defs.h:117
double d[2]
Definition: blis_type_defs.h:119
__m128d v
Definition: blis_type_defs.h:118
References alpha_c, beta_c, bl1_abort(), chi1, v2df_t::d, i, n_left, n_pre, n_run, psi1, rho, rho_c, upsilon1, v2df_t::v, and zeta1.
Referenced by FLA_Fused_Uhu_Yhu_Zhu_opd_var1().