Added 30xk support for packm ukernels.

Details:
- Updated bli_kernel_*_macro_defs.h headers to include default
  definitions for 30xk packm kernels.
- Extended function pointer arrays in bli_packm_cxk_*() out to 31 and
  included 30xk kernels.
- Addex 30xk kernels to frame/1m/packm/ukernels/bli_packm_ref_cxk_*.c.
This commit is contained in:
Field G. Van Zee
2014-09-22 16:06:15 -05:00
parent b6d4bd792e
commit 4a7df04e8a
16 changed files with 1356 additions and 67 deletions

View File

@@ -45,7 +45,7 @@ typedef void (*FUNCPTR_T)(
);
#undef FUNCPTR_ARRAY_LENGTH
#define FUNCPTR_ARRAY_LENGTH 18
#define FUNCPTR_ARRAY_LENGTH 32
static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
{
@@ -59,10 +59,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 2 */
{
NULL,
BLIS_CPACKM_2XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_2XK_3M_KERNEL,
NULL, BLIS_CPACKM_2XK_3M_KERNEL,
NULL, BLIS_ZPACKM_2XK_3M_KERNEL,
},
/* micro-panel width = 3 */
{
@@ -70,10 +68,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 4 */
{
NULL,
BLIS_CPACKM_4XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_4XK_3M_KERNEL,
NULL, BLIS_CPACKM_4XK_3M_KERNEL,
NULL, BLIS_ZPACKM_4XK_3M_KERNEL,
},
/* micro-panel width = 5 */
{
@@ -81,10 +77,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 6 */
{
NULL,
BLIS_CPACKM_6XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_6XK_3M_KERNEL,
NULL, BLIS_CPACKM_6XK_3M_KERNEL,
NULL, BLIS_ZPACKM_6XK_3M_KERNEL,
},
/* micro-panel width = 7 */
{
@@ -92,10 +86,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 8 */
{
NULL,
BLIS_CPACKM_8XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_8XK_3M_KERNEL,
NULL, BLIS_CPACKM_8XK_3M_KERNEL,
NULL, BLIS_ZPACKM_8XK_3M_KERNEL,
},
/* micro-panel width = 9 */
{
@@ -103,10 +95,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 10 */
{
NULL,
BLIS_CPACKM_10XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_10XK_3M_KERNEL,
NULL, BLIS_CPACKM_10XK_3M_KERNEL,
NULL, BLIS_ZPACKM_10XK_3M_KERNEL,
},
/* micro-panel width = 11 */
{
@@ -114,10 +104,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 12 */
{
NULL,
BLIS_CPACKM_12XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_12XK_3M_KERNEL,
NULL, BLIS_CPACKM_12XK_3M_KERNEL,
NULL, BLIS_ZPACKM_12XK_3M_KERNEL,
},
/* micro-panel width = 13 */
{
@@ -125,10 +113,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 14 */
{
NULL,
BLIS_CPACKM_14XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_14XK_3M_KERNEL,
NULL, BLIS_CPACKM_14XK_3M_KERNEL,
NULL, BLIS_ZPACKM_14XK_3M_KERNEL,
},
/* micro-panel width = 15 */
{
@@ -136,15 +122,72 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 16 */
{
NULL,
BLIS_CPACKM_16XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_16XK_3M_KERNEL,
NULL, BLIS_CPACKM_16XK_3M_KERNEL,
NULL, BLIS_ZPACKM_16XK_3M_KERNEL,
},
/* micro-panel width = 17 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 18 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 19 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 20 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 21 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 22 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 23 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 24 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 25 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 26 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 27 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 28 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 29 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 30 */
{
NULL,
BLIS_CPACKM_30XK_3M_KERNEL,
NULL,
BLIS_ZPACKM_30XK_3M_KERNEL,
},
/* micro-panel width = 31 */
{
NULL, NULL, NULL, NULL,
},
};

View File

@@ -45,7 +45,7 @@ typedef void (*FUNCPTR_T)(
);
#undef FUNCPTR_ARRAY_LENGTH
#define FUNCPTR_ARRAY_LENGTH 18
#define FUNCPTR_ARRAY_LENGTH 32
static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
{
@@ -59,10 +59,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 2 */
{
NULL,
BLIS_CPACKM_2XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_2XK_4M_KERNEL,
NULL, BLIS_CPACKM_2XK_4M_KERNEL,
NULL, BLIS_ZPACKM_2XK_4M_KERNEL,
},
/* micro-panel width = 3 */
{
@@ -70,10 +68,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 4 */
{
NULL,
BLIS_CPACKM_4XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_4XK_4M_KERNEL,
NULL, BLIS_CPACKM_4XK_4M_KERNEL,
NULL, BLIS_ZPACKM_4XK_4M_KERNEL,
},
/* micro-panel width = 5 */
{
@@ -81,10 +77,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 6 */
{
NULL,
BLIS_CPACKM_6XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_6XK_4M_KERNEL,
NULL, BLIS_CPACKM_6XK_4M_KERNEL,
NULL, BLIS_ZPACKM_6XK_4M_KERNEL,
},
/* micro-panel width = 7 */
{
@@ -92,10 +86,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 8 */
{
NULL,
BLIS_CPACKM_8XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_8XK_4M_KERNEL,
NULL, BLIS_CPACKM_8XK_4M_KERNEL,
NULL, BLIS_ZPACKM_8XK_4M_KERNEL,
},
/* micro-panel width = 9 */
{
@@ -103,10 +95,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 10 */
{
NULL,
BLIS_CPACKM_10XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_10XK_4M_KERNEL,
NULL, BLIS_CPACKM_10XK_4M_KERNEL,
NULL, BLIS_ZPACKM_10XK_4M_KERNEL,
},
/* micro-panel width = 11 */
{
@@ -114,10 +104,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 12 */
{
NULL,
BLIS_CPACKM_12XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_12XK_4M_KERNEL,
NULL, BLIS_CPACKM_12XK_4M_KERNEL,
NULL, BLIS_ZPACKM_12XK_4M_KERNEL,
},
/* micro-panel width = 13 */
{
@@ -125,10 +113,8 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 14 */
{
NULL,
BLIS_CPACKM_14XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_14XK_4M_KERNEL,
NULL, BLIS_CPACKM_14XK_4M_KERNEL,
NULL, BLIS_ZPACKM_14XK_4M_KERNEL,
},
/* micro-panel width = 15 */
{
@@ -136,15 +122,72 @@ static FUNCPTR_T ftypes[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
},
/* micro-panel width = 16 */
{
NULL,
BLIS_CPACKM_16XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_16XK_4M_KERNEL,
NULL, BLIS_CPACKM_16XK_4M_KERNEL,
NULL, BLIS_ZPACKM_16XK_4M_KERNEL,
},
/* micro-panel width = 17 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 18 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 19 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 20 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 21 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 22 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 23 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 24 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 25 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 26 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 27 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 28 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 29 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 30 */
{
NULL,
BLIS_CPACKM_30XK_4M_KERNEL,
NULL,
BLIS_ZPACKM_30XK_4M_KERNEL,
},
/* micro-panel width = 31 */
{
NULL, NULL, NULL, NULL,
},
};

View File

@@ -46,7 +46,7 @@ typedef void (*FUNCPTR_T)(
);
#undef FUNCPTR_ARRAY_LENGTH
#define FUNCPTR_ARRAY_LENGTH 18
#define FUNCPTR_ARRAY_LENGTH 32
static FUNCPTR_T ftypes_rih[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
{
@@ -130,6 +130,63 @@ static FUNCPTR_T ftypes_rih[FUNCPTR_ARRAY_LENGTH][BLIS_NUM_FP_TYPES] =
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 18 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 19 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 20 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 21 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 22 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 23 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 24 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 25 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 26 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 27 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 28 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 29 */
{
NULL, NULL, NULL, NULL,
},
/* micro-panel width = 30 */
{
NULL, BLIS_CPACKM_30XK_RIH_KERNEL,
NULL, BLIS_ZPACKM_30XK_RIH_KERNEL,
},
/* micro-panel width = 31 */
{
NULL, NULL, NULL, NULL,
},
};

View File

@@ -824,3 +824,186 @@ void PASTEMAC(ch,varname)( \
INSERT_GENTFUNC_BASIC0( packm_ref_16xk )
#undef GENTFUNC
#define GENTFUNC( ctype, ch, varname ) \
\
void PASTEMAC(ch,varname)( \
conj_t conja, \
dim_t n, \
void* kappa, \
void* a, inc_t inca, inc_t lda, \
void* p, inc_t ldp \
) \
{ \
ctype* restrict kappa_cast = kappa; \
ctype* restrict alpha1 = a; \
ctype* restrict pi1 = p; \
\
if ( PASTEMAC(ch,eq1)( *kappa_cast ) ) \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,copyjs)( *(alpha1 + 0*inca), *(pi1 + 0) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 1*inca), *(pi1 + 1) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 2*inca), *(pi1 + 2) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 3*inca), *(pi1 + 3) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 4*inca), *(pi1 + 4) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 5*inca), *(pi1 + 5) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 6*inca), *(pi1 + 6) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 7*inca), *(pi1 + 7) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 8*inca), *(pi1 + 8) ); \
PASTEMAC(ch,copyjs)( *(alpha1 + 9*inca), *(pi1 + 9) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +10*inca), *(pi1 +10) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +11*inca), *(pi1 +11) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +12*inca), *(pi1 +12) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +13*inca), *(pi1 +13) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +14*inca), *(pi1 +14) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +15*inca), *(pi1 +15) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +16*inca), *(pi1 +16) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +17*inca), *(pi1 +17) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +18*inca), *(pi1 +18) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +19*inca), *(pi1 +19) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +20*inca), *(pi1 +20) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +21*inca), *(pi1 +21) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +22*inca), *(pi1 +22) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +23*inca), *(pi1 +23) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +24*inca), *(pi1 +24) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +25*inca), *(pi1 +25) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +26*inca), *(pi1 +26) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +27*inca), *(pi1 +27) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +28*inca), *(pi1 +28) ); \
PASTEMAC(ch,copyjs)( *(alpha1 +29*inca), *(pi1 +29) ); \
\
alpha1 += lda; \
pi1 += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,copys)( *(alpha1 + 0*inca), *(pi1 + 0) ); \
PASTEMAC(ch,copys)( *(alpha1 + 1*inca), *(pi1 + 1) ); \
PASTEMAC(ch,copys)( *(alpha1 + 2*inca), *(pi1 + 2) ); \
PASTEMAC(ch,copys)( *(alpha1 + 3*inca), *(pi1 + 3) ); \
PASTEMAC(ch,copys)( *(alpha1 + 4*inca), *(pi1 + 4) ); \
PASTEMAC(ch,copys)( *(alpha1 + 5*inca), *(pi1 + 5) ); \
PASTEMAC(ch,copys)( *(alpha1 + 6*inca), *(pi1 + 6) ); \
PASTEMAC(ch,copys)( *(alpha1 + 7*inca), *(pi1 + 7) ); \
PASTEMAC(ch,copys)( *(alpha1 + 8*inca), *(pi1 + 8) ); \
PASTEMAC(ch,copys)( *(alpha1 + 9*inca), *(pi1 + 9) ); \
PASTEMAC(ch,copys)( *(alpha1 +10*inca), *(pi1 +10) ); \
PASTEMAC(ch,copys)( *(alpha1 +11*inca), *(pi1 +11) ); \
PASTEMAC(ch,copys)( *(alpha1 +12*inca), *(pi1 +12) ); \
PASTEMAC(ch,copys)( *(alpha1 +13*inca), *(pi1 +13) ); \
PASTEMAC(ch,copys)( *(alpha1 +14*inca), *(pi1 +14) ); \
PASTEMAC(ch,copys)( *(alpha1 +15*inca), *(pi1 +15) ); \
PASTEMAC(ch,copys)( *(alpha1 +16*inca), *(pi1 +16) ); \
PASTEMAC(ch,copys)( *(alpha1 +17*inca), *(pi1 +17) ); \
PASTEMAC(ch,copys)( *(alpha1 +18*inca), *(pi1 +18) ); \
PASTEMAC(ch,copys)( *(alpha1 +19*inca), *(pi1 +19) ); \
PASTEMAC(ch,copys)( *(alpha1 +20*inca), *(pi1 +20) ); \
PASTEMAC(ch,copys)( *(alpha1 +21*inca), *(pi1 +21) ); \
PASTEMAC(ch,copys)( *(alpha1 +22*inca), *(pi1 +22) ); \
PASTEMAC(ch,copys)( *(alpha1 +23*inca), *(pi1 +23) ); \
PASTEMAC(ch,copys)( *(alpha1 +24*inca), *(pi1 +24) ); \
PASTEMAC(ch,copys)( *(alpha1 +25*inca), *(pi1 +25) ); \
PASTEMAC(ch,copys)( *(alpha1 +26*inca), *(pi1 +26) ); \
PASTEMAC(ch,copys)( *(alpha1 +27*inca), *(pi1 +27) ); \
PASTEMAC(ch,copys)( *(alpha1 +28*inca), *(pi1 +28) ); \
PASTEMAC(ch,copys)( *(alpha1 +29*inca), *(pi1 +29) ); \
\
alpha1 += lda; \
pi1 += ldp; \
} \
} \
} \
else \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 0*inca), *(pi1 + 0) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 1*inca), *(pi1 + 1) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 2*inca), *(pi1 + 2) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 3*inca), *(pi1 + 3) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 4*inca), *(pi1 + 4) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 5*inca), *(pi1 + 5) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 6*inca), *(pi1 + 6) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 7*inca), *(pi1 + 7) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 8*inca), *(pi1 + 8) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 + 9*inca), *(pi1 + 9) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +10*inca), *(pi1 +10) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +11*inca), *(pi1 +11) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +12*inca), *(pi1 +12) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +13*inca), *(pi1 +13) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +14*inca), *(pi1 +14) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +15*inca), *(pi1 +15) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +16*inca), *(pi1 +16) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +17*inca), *(pi1 +17) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +18*inca), *(pi1 +18) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +19*inca), *(pi1 +19) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +20*inca), *(pi1 +20) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +21*inca), *(pi1 +21) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +22*inca), *(pi1 +22) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +23*inca), *(pi1 +23) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +24*inca), *(pi1 +24) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +25*inca), *(pi1 +25) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +26*inca), *(pi1 +26) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +27*inca), *(pi1 +27) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +28*inca), *(pi1 +28) ); \
PASTEMAC(ch,scal2js)( *kappa_cast, *(alpha1 +29*inca), *(pi1 +29) ); \
\
alpha1 += lda; \
pi1 += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 0*inca), *(pi1 + 0) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 1*inca), *(pi1 + 1) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 2*inca), *(pi1 + 2) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 3*inca), *(pi1 + 3) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 4*inca), *(pi1 + 4) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 5*inca), *(pi1 + 5) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 6*inca), *(pi1 + 6) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 7*inca), *(pi1 + 7) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 8*inca), *(pi1 + 8) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 + 9*inca), *(pi1 + 9) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +10*inca), *(pi1 +10) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +11*inca), *(pi1 +11) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +12*inca), *(pi1 +12) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +13*inca), *(pi1 +13) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +14*inca), *(pi1 +14) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +15*inca), *(pi1 +15) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +16*inca), *(pi1 +16) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +17*inca), *(pi1 +17) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +18*inca), *(pi1 +18) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +19*inca), *(pi1 +19) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +20*inca), *(pi1 +20) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +21*inca), *(pi1 +21) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +22*inca), *(pi1 +22) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +23*inca), *(pi1 +23) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +24*inca), *(pi1 +24) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +25*inca), *(pi1 +25) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +26*inca), *(pi1 +26) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +27*inca), *(pi1 +27) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +28*inca), *(pi1 +28) ); \
PASTEMAC(ch,scal2s)( *kappa_cast, *(alpha1 +29*inca), *(pi1 +29) ); \
\
alpha1 += lda; \
pi1 += ldp; \
} \
} \
} \
}
INSERT_GENTFUNC_BASIC0( packm_ref_30xk )

View File

@@ -51,4 +51,5 @@ INSERT_GENTPROT_BASIC( packm_ref_10xk )
INSERT_GENTPROT_BASIC( packm_ref_12xk )
INSERT_GENTPROT_BASIC( packm_ref_14xk )
INSERT_GENTPROT_BASIC( packm_ref_16xk )
INSERT_GENTPROT_BASIC( packm_ref_30xk )

View File

@@ -984,3 +984,207 @@ void PASTEMAC(ch,varname)( \
INSERT_GENTFUNCCO_BASIC0( packm_ref_16xk_3m )
#undef GENTFUNCCO
#define GENTFUNCCO( ctype, ctype_r, ch, chr, varname ) \
\
void PASTEMAC(ch,varname)( \
conj_t conja, \
dim_t n, \
void* kappa, \
void* a, inc_t inca, inc_t lda, \
void* p, inc_t is_p, inc_t ldp \
) \
{ \
const inc_t inca2 = 2 * inca; \
const inc_t lda2 = 2 * lda; \
\
ctype* kappa_cast = kappa; \
ctype_r* restrict kappa_r = ( ctype_r* )kappa; \
ctype_r* restrict kappa_i = ( ctype_r* )kappa + 1; \
ctype_r* restrict alpha1_r = ( ctype_r* )a; \
ctype_r* restrict alpha1_i = ( ctype_r* )a + 1; \
ctype_r* restrict pi1_r = ( ctype_r* )p; \
ctype_r* restrict pi1_i = ( ctype_r* )p + is_p; \
ctype_r* restrict pi1_rpi = ( ctype_r* )p + 2*is_p; \
\
if ( PASTEMAC(ch,eq1)( *kappa_cast ) ) \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0), *(pi1_rpi + 0) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1), *(pi1_rpi + 1) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2), *(pi1_rpi + 2) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3), *(pi1_rpi + 3) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4), *(pi1_rpi + 4) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5), *(pi1_rpi + 5) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6), *(pi1_rpi + 6) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7), *(pi1_rpi + 7) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8), *(pi1_rpi + 8) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9), *(pi1_rpi + 9) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10), *(pi1_rpi +10) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11), *(pi1_rpi +11) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12), *(pi1_rpi +12) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13), *(pi1_rpi +13) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14), *(pi1_rpi +14) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15), *(pi1_rpi +15) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16), *(pi1_rpi +16) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17), *(pi1_rpi +17) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18), *(pi1_rpi +18) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19), *(pi1_rpi +19) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20), *(pi1_rpi +20) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21), *(pi1_rpi +21) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22), *(pi1_rpi +22) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23), *(pi1_rpi +23) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24), *(pi1_rpi +24) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25), *(pi1_rpi +25) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26), *(pi1_rpi +26) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27), *(pi1_rpi +27) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28), *(pi1_rpi +28) ); \
PASTEMAC(ch,copyjri3s)( *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29), *(pi1_rpi +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
pi1_rpi += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0), *(pi1_rpi + 0) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1), *(pi1_rpi + 1) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2), *(pi1_rpi + 2) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3), *(pi1_rpi + 3) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4), *(pi1_rpi + 4) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5), *(pi1_rpi + 5) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6), *(pi1_rpi + 6) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7), *(pi1_rpi + 7) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8), *(pi1_rpi + 8) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9), *(pi1_rpi + 9) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10), *(pi1_rpi +10) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11), *(pi1_rpi +11) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12), *(pi1_rpi +12) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13), *(pi1_rpi +13) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14), *(pi1_rpi +14) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15), *(pi1_rpi +15) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16), *(pi1_rpi +16) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17), *(pi1_rpi +17) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18), *(pi1_rpi +18) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19), *(pi1_rpi +19) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20), *(pi1_rpi +20) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21), *(pi1_rpi +21) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22), *(pi1_rpi +22) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23), *(pi1_rpi +23) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24), *(pi1_rpi +24) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25), *(pi1_rpi +25) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26), *(pi1_rpi +26) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27), *(pi1_rpi +27) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28), *(pi1_rpi +28) ); \
PASTEMAC(ch,copyri3s)( *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29), *(pi1_rpi +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
pi1_rpi += ldp; \
} \
} \
} \
else \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0), *(pi1_rpi + 0) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1), *(pi1_rpi + 1) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2), *(pi1_rpi + 2) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3), *(pi1_rpi + 3) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4), *(pi1_rpi + 4) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5), *(pi1_rpi + 5) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6), *(pi1_rpi + 6) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7), *(pi1_rpi + 7) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8), *(pi1_rpi + 8) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9), *(pi1_rpi + 9) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10), *(pi1_rpi +10) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11), *(pi1_rpi +11) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12), *(pi1_rpi +12) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13), *(pi1_rpi +13) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14), *(pi1_rpi +14) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15), *(pi1_rpi +15) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16), *(pi1_rpi +16) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17), *(pi1_rpi +17) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18), *(pi1_rpi +18) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19), *(pi1_rpi +19) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20), *(pi1_rpi +20) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21), *(pi1_rpi +21) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22), *(pi1_rpi +22) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23), *(pi1_rpi +23) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24), *(pi1_rpi +24) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25), *(pi1_rpi +25) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26), *(pi1_rpi +26) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27), *(pi1_rpi +27) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28), *(pi1_rpi +28) ); \
PASTEMAC(ch,scal2jri3s)( *kappa_r, *kappa_i, *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29), *(pi1_rpi +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
pi1_rpi += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0), *(pi1_rpi + 0) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1), *(pi1_rpi + 1) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2), *(pi1_rpi + 2) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3), *(pi1_rpi + 3) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4), *(pi1_rpi + 4) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5), *(pi1_rpi + 5) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6), *(pi1_rpi + 6) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7), *(pi1_rpi + 7) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8), *(pi1_rpi + 8) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9), *(pi1_rpi + 9) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10), *(pi1_rpi +10) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11), *(pi1_rpi +11) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12), *(pi1_rpi +12) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13), *(pi1_rpi +13) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14), *(pi1_rpi +14) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15), *(pi1_rpi +15) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16), *(pi1_rpi +16) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17), *(pi1_rpi +17) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18), *(pi1_rpi +18) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19), *(pi1_rpi +19) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20), *(pi1_rpi +20) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21), *(pi1_rpi +21) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22), *(pi1_rpi +22) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23), *(pi1_rpi +23) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24), *(pi1_rpi +24) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25), *(pi1_rpi +25) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26), *(pi1_rpi +26) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27), *(pi1_rpi +27) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28), *(pi1_rpi +28) ); \
PASTEMAC(ch,scal2ri3s)( *kappa_r, *kappa_i, *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29), *(pi1_rpi +29) ); \
\
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
pi1_rpi += ldp; \
} \
} \
} \
}
INSERT_GENTFUNCCO_BASIC0( packm_ref_30xk_3m )

View File

@@ -51,4 +51,5 @@ INSERT_GENTPROT_BASIC( packm_ref_10xk_3m )
INSERT_GENTPROT_BASIC( packm_ref_12xk_3m )
INSERT_GENTPROT_BASIC( packm_ref_14xk_3m )
INSERT_GENTPROT_BASIC( packm_ref_16xk_3m )
INSERT_GENTPROT_BASIC( packm_ref_30xk_3m )

View File

@@ -944,3 +944,201 @@ void PASTEMAC(ch,varname)( \
INSERT_GENTFUNCCO_BASIC0( packm_ref_16xk_4m )
#undef GENTFUNCCO
#define GENTFUNCCO( ctype, ctype_r, ch, chr, varname ) \
\
void PASTEMAC(ch,varname)( \
conj_t conja, \
dim_t n, \
void* kappa, \
void* a, inc_t inca, inc_t lda, \
void* p, inc_t is_p, inc_t ldp \
) \
{ \
const inc_t inca2 = 2 * inca; \
const inc_t lda2 = 2 * lda; \
\
ctype* kappa_cast = kappa; \
ctype_r* restrict kappa_r = ( ctype_r* )kappa; \
ctype_r* restrict kappa_i = ( ctype_r* )kappa + 1; \
ctype_r* restrict alpha1_r = ( ctype_r* )a; \
ctype_r* restrict alpha1_i = ( ctype_r* )a + 1; \
ctype_r* restrict pi1_r = ( ctype_r* )p; \
ctype_r* restrict pi1_i = ( ctype_r* )p + is_p; \
\
if ( PASTEMAC(ch,eq1)( *kappa_cast ) ) \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,copyjris)( *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28) ); \
PASTEMAC(ch,copyjris)( *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,copyris)( *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8) ); \
PASTEMAC(ch,copyris)( *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28) ); \
PASTEMAC(ch,copyris)( *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
} \
} \
} \
else \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28) ); \
PASTEMAC(ch,scal2jris)( *kappa_r, *kappa_i, *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0), *(pi1_i + 0) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1), *(pi1_i + 1) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2), *(pi1_i + 2) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3), *(pi1_i + 3) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4), *(pi1_i + 4) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5), *(pi1_i + 5) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6), *(pi1_i + 6) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7), *(pi1_i + 7) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8), *(pi1_i + 8) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9), *(pi1_i + 9) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10), *(pi1_i +10) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11), *(pi1_i +11) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12), *(pi1_i +12) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13), *(pi1_i +13) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14), *(pi1_i +14) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15), *(pi1_i +15) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16), *(pi1_i +16) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17), *(pi1_i +17) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18), *(pi1_i +18) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19), *(pi1_i +19) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20), *(pi1_i +20) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21), *(pi1_i +21) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22), *(pi1_i +22) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23), *(pi1_i +23) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24), *(pi1_i +24) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25), *(pi1_i +25) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26), *(pi1_i +26) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27), *(pi1_i +27) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28), *(pi1_i +28) ); \
PASTEMAC(ch,scal2ris)( *kappa_r, *kappa_i, *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29), *(pi1_i +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
pi1_i += ldp; \
} \
} \
} \
}
INSERT_GENTFUNCCO_BASIC0( packm_ref_30xk_4m )

View File

@@ -51,4 +51,5 @@ INSERT_GENTPROT_BASIC( packm_ref_10xk_4m )
INSERT_GENTPROT_BASIC( packm_ref_12xk_4m )
INSERT_GENTPROT_BASIC( packm_ref_14xk_4m )
INSERT_GENTPROT_BASIC( packm_ref_16xk_4m )
INSERT_GENTPROT_BASIC( packm_ref_30xk_4m )

View File

@@ -2080,3 +2080,490 @@ void PASTEMAC(ch,varname)( \
INSERT_GENTFUNCCO_BASIC0( packm_ref_16xk_rih )
#undef GENTFUNCCO
#define GENTFUNCCO( ctype, ctype_r, ch, chr, varname ) \
\
void PASTEMAC(ch,varname)( \
conj_t conja, \
pack_t schema, \
dim_t n, \
void* kappa, \
void* a, inc_t inca, inc_t lda, \
void* p, inc_t ldp \
) \
{ \
const inc_t inca2 = 2 * inca; \
const inc_t lda2 = 2 * lda; \
\
ctype* kappa_cast = kappa; \
ctype* restrict alpha1 = a; \
ctype_r* restrict alpha1_r = ( ctype_r* )a; \
ctype_r* restrict alpha1_i = ( ctype_r* )a + 1; \
ctype_r* restrict pi1_r = ( ctype_r* )p; \
\
\
if ( bli_is_ro_packed( schema ) ) \
{ \
if ( PASTEMAC(ch,eq1)( *kappa_cast ) ) \
{ \
/* This works regardless of conja since we are only copying
the real part. */ \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(chr,copys)( *(alpha1_r + 0*inca2), *(pi1_r + 0) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 1*inca2), *(pi1_r + 1) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 2*inca2), *(pi1_r + 2) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 3*inca2), *(pi1_r + 3) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 4*inca2), *(pi1_r + 4) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 5*inca2), *(pi1_r + 5) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 6*inca2), *(pi1_r + 6) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 7*inca2), *(pi1_r + 7) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 8*inca2), *(pi1_r + 8) ); \
PASTEMAC(chr,copys)( *(alpha1_r + 9*inca2), *(pi1_r + 9) ); \
PASTEMAC(chr,copys)( *(alpha1_r +10*inca2), *(pi1_r +10) ); \
PASTEMAC(chr,copys)( *(alpha1_r +11*inca2), *(pi1_r +11) ); \
PASTEMAC(chr,copys)( *(alpha1_r +12*inca2), *(pi1_r +12) ); \
PASTEMAC(chr,copys)( *(alpha1_r +13*inca2), *(pi1_r +13) ); \
PASTEMAC(chr,copys)( *(alpha1_r +14*inca2), *(pi1_r +14) ); \
PASTEMAC(chr,copys)( *(alpha1_r +15*inca2), *(pi1_r +15) ); \
PASTEMAC(chr,copys)( *(alpha1_r +16*inca2), *(pi1_r +16) ); \
PASTEMAC(chr,copys)( *(alpha1_r +17*inca2), *(pi1_r +17) ); \
PASTEMAC(chr,copys)( *(alpha1_r +18*inca2), *(pi1_r +18) ); \
PASTEMAC(chr,copys)( *(alpha1_r +19*inca2), *(pi1_r +19) ); \
PASTEMAC(chr,copys)( *(alpha1_r +20*inca2), *(pi1_r +20) ); \
PASTEMAC(chr,copys)( *(alpha1_r +21*inca2), *(pi1_r +21) ); \
PASTEMAC(chr,copys)( *(alpha1_r +22*inca2), *(pi1_r +22) ); \
PASTEMAC(chr,copys)( *(alpha1_r +23*inca2), *(pi1_r +23) ); \
PASTEMAC(chr,copys)( *(alpha1_r +24*inca2), *(pi1_r +24) ); \
PASTEMAC(chr,copys)( *(alpha1_r +25*inca2), *(pi1_r +25) ); \
PASTEMAC(chr,copys)( *(alpha1_r +26*inca2), *(pi1_r +26) ); \
PASTEMAC(chr,copys)( *(alpha1_r +27*inca2), *(pi1_r +27) ); \
PASTEMAC(chr,copys)( *(alpha1_r +28*inca2), *(pi1_r +28) ); \
PASTEMAC(chr,copys)( *(alpha1_r +29*inca2), *(pi1_r +29) ); \
\
alpha1_r += lda2; \
pi1_r += ldp; \
} \
} \
} \
else \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 0*inca), *(pi1_r + 0) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 1*inca), *(pi1_r + 1) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 2*inca), *(pi1_r + 2) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 3*inca), *(pi1_r + 3) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 4*inca), *(pi1_r + 4) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 5*inca), *(pi1_r + 5) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 6*inca), *(pi1_r + 6) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 7*inca), *(pi1_r + 7) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 8*inca), *(pi1_r + 8) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 + 9*inca), *(pi1_r + 9) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +10*inca), *(pi1_r +10) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +11*inca), *(pi1_r +11) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +12*inca), *(pi1_r +12) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +13*inca), *(pi1_r +13) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +14*inca), *(pi1_r +14) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +15*inca), *(pi1_r +15) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +16*inca), *(pi1_r +16) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +17*inca), *(pi1_r +17) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +18*inca), *(pi1_r +18) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +19*inca), *(pi1_r +19) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +20*inca), *(pi1_r +20) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +21*inca), *(pi1_r +21) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +22*inca), *(pi1_r +22) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +23*inca), *(pi1_r +23) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +24*inca), *(pi1_r +24) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +25*inca), *(pi1_r +25) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +26*inca), *(pi1_r +26) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +27*inca), *(pi1_r +27) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +28*inca), *(pi1_r +28) ); \
PASTEMAC(ch,scal2jros)( *kappa_cast, *(alpha1 +29*inca), *(pi1_r +29) ); \
\
alpha1 += lda; \
pi1_r += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 0*inca), *(pi1_r + 0) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 1*inca), *(pi1_r + 1) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 2*inca), *(pi1_r + 2) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 3*inca), *(pi1_r + 3) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 4*inca), *(pi1_r + 4) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 5*inca), *(pi1_r + 5) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 6*inca), *(pi1_r + 6) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 7*inca), *(pi1_r + 7) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 8*inca), *(pi1_r + 8) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 + 9*inca), *(pi1_r + 9) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +10*inca), *(pi1_r +10) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +11*inca), *(pi1_r +11) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +12*inca), *(pi1_r +12) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +13*inca), *(pi1_r +13) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +14*inca), *(pi1_r +14) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +15*inca), *(pi1_r +15) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +16*inca), *(pi1_r +16) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +17*inca), *(pi1_r +17) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +18*inca), *(pi1_r +18) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +19*inca), *(pi1_r +19) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +20*inca), *(pi1_r +20) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +21*inca), *(pi1_r +21) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +22*inca), *(pi1_r +22) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +23*inca), *(pi1_r +23) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +24*inca), *(pi1_r +24) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +25*inca), *(pi1_r +25) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +26*inca), *(pi1_r +26) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +27*inca), *(pi1_r +27) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +28*inca), *(pi1_r +28) ); \
PASTEMAC(ch,scal2ros)( *kappa_cast, *(alpha1 +29*inca), *(pi1_r +29) ); \
\
alpha1 += lda; \
pi1_r += ldp; \
} \
} \
} \
} \
else if ( bli_is_io_packed( schema ) ) \
{ \
if ( PASTEMAC(ch,eq1)( *kappa_cast ) ) \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(chr,copys)( -*(alpha1_i + 0*inca2), *(pi1_r + 0) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 1*inca2), *(pi1_r + 1) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 2*inca2), *(pi1_r + 2) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 3*inca2), *(pi1_r + 3) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 4*inca2), *(pi1_r + 4) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 5*inca2), *(pi1_r + 5) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 6*inca2), *(pi1_r + 6) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 7*inca2), *(pi1_r + 7) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 8*inca2), *(pi1_r + 8) ); \
PASTEMAC(chr,copys)( -*(alpha1_i + 9*inca2), *(pi1_r + 9) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +10*inca2), *(pi1_r +10) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +11*inca2), *(pi1_r +11) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +12*inca2), *(pi1_r +12) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +13*inca2), *(pi1_r +13) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +14*inca2), *(pi1_r +14) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +15*inca2), *(pi1_r +15) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +16*inca2), *(pi1_r +16) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +17*inca2), *(pi1_r +17) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +18*inca2), *(pi1_r +18) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +19*inca2), *(pi1_r +19) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +20*inca2), *(pi1_r +20) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +21*inca2), *(pi1_r +21) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +22*inca2), *(pi1_r +22) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +23*inca2), *(pi1_r +23) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +24*inca2), *(pi1_r +24) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +25*inca2), *(pi1_r +25) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +26*inca2), *(pi1_r +26) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +27*inca2), *(pi1_r +27) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +28*inca2), *(pi1_r +28) ); \
PASTEMAC(chr,copys)( -*(alpha1_i +29*inca2), *(pi1_r +29) ); \
\
alpha1_i += lda2; \
pi1_r += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(chr,copys)( *(alpha1_i + 0*inca2), *(pi1_r + 0) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 1*inca2), *(pi1_r + 1) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 2*inca2), *(pi1_r + 2) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 3*inca2), *(pi1_r + 3) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 4*inca2), *(pi1_r + 4) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 5*inca2), *(pi1_r + 5) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 6*inca2), *(pi1_r + 6) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 7*inca2), *(pi1_r + 7) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 8*inca2), *(pi1_r + 8) ); \
PASTEMAC(chr,copys)( *(alpha1_i + 9*inca2), *(pi1_r + 9) ); \
PASTEMAC(chr,copys)( *(alpha1_i +10*inca2), *(pi1_r +10) ); \
PASTEMAC(chr,copys)( *(alpha1_i +11*inca2), *(pi1_r +11) ); \
PASTEMAC(chr,copys)( *(alpha1_i +12*inca2), *(pi1_r +12) ); \
PASTEMAC(chr,copys)( *(alpha1_i +13*inca2), *(pi1_r +13) ); \
PASTEMAC(chr,copys)( *(alpha1_i +14*inca2), *(pi1_r +14) ); \
PASTEMAC(chr,copys)( *(alpha1_i +15*inca2), *(pi1_r +15) ); \
PASTEMAC(chr,copys)( *(alpha1_i +16*inca2), *(pi1_r +16) ); \
PASTEMAC(chr,copys)( *(alpha1_i +17*inca2), *(pi1_r +17) ); \
PASTEMAC(chr,copys)( *(alpha1_i +18*inca2), *(pi1_r +18) ); \
PASTEMAC(chr,copys)( *(alpha1_i +19*inca2), *(pi1_r +19) ); \
PASTEMAC(chr,copys)( *(alpha1_i +20*inca2), *(pi1_r +20) ); \
PASTEMAC(chr,copys)( *(alpha1_i +21*inca2), *(pi1_r +21) ); \
PASTEMAC(chr,copys)( *(alpha1_i +22*inca2), *(pi1_r +22) ); \
PASTEMAC(chr,copys)( *(alpha1_i +23*inca2), *(pi1_r +23) ); \
PASTEMAC(chr,copys)( *(alpha1_i +24*inca2), *(pi1_r +24) ); \
PASTEMAC(chr,copys)( *(alpha1_i +25*inca2), *(pi1_r +25) ); \
PASTEMAC(chr,copys)( *(alpha1_i +26*inca2), *(pi1_r +26) ); \
PASTEMAC(chr,copys)( *(alpha1_i +27*inca2), *(pi1_r +27) ); \
PASTEMAC(chr,copys)( *(alpha1_i +28*inca2), *(pi1_r +28) ); \
PASTEMAC(chr,copys)( *(alpha1_i +29*inca2), *(pi1_r +29) ); \
\
alpha1_i += lda2; \
pi1_r += ldp; \
} \
} \
} \
else \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 0*inca), *(pi1_r + 0) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 1*inca), *(pi1_r + 1) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 2*inca), *(pi1_r + 2) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 3*inca), *(pi1_r + 3) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 4*inca), *(pi1_r + 4) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 5*inca), *(pi1_r + 5) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 6*inca), *(pi1_r + 6) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 7*inca), *(pi1_r + 7) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 8*inca), *(pi1_r + 8) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 + 9*inca), *(pi1_r + 9) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +10*inca), *(pi1_r +10) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +11*inca), *(pi1_r +11) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +12*inca), *(pi1_r +12) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +13*inca), *(pi1_r +13) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +14*inca), *(pi1_r +14) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +15*inca), *(pi1_r +15) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +16*inca), *(pi1_r +16) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +17*inca), *(pi1_r +17) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +18*inca), *(pi1_r +18) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +19*inca), *(pi1_r +19) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +20*inca), *(pi1_r +20) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +21*inca), *(pi1_r +21) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +22*inca), *(pi1_r +22) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +23*inca), *(pi1_r +23) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +24*inca), *(pi1_r +24) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +25*inca), *(pi1_r +25) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +26*inca), *(pi1_r +26) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +27*inca), *(pi1_r +27) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +28*inca), *(pi1_r +28) ); \
PASTEMAC(ch,scal2jios)( *kappa_cast, *(alpha1 +29*inca), *(pi1_r +29) ); \
\
alpha1 += lda; \
pi1_r += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 0*inca), *(pi1_r + 0) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 1*inca), *(pi1_r + 1) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 2*inca), *(pi1_r + 2) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 3*inca), *(pi1_r + 3) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 4*inca), *(pi1_r + 4) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 5*inca), *(pi1_r + 5) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 6*inca), *(pi1_r + 6) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 7*inca), *(pi1_r + 7) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 8*inca), *(pi1_r + 8) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 + 9*inca), *(pi1_r + 9) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +10*inca), *(pi1_r +10) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +11*inca), *(pi1_r +11) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +12*inca), *(pi1_r +12) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +13*inca), *(pi1_r +13) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +14*inca), *(pi1_r +14) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +15*inca), *(pi1_r +15) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +16*inca), *(pi1_r +16) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +17*inca), *(pi1_r +17) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +18*inca), *(pi1_r +18) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +19*inca), *(pi1_r +19) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +20*inca), *(pi1_r +20) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +21*inca), *(pi1_r +21) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +22*inca), *(pi1_r +22) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +23*inca), *(pi1_r +23) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +24*inca), *(pi1_r +24) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +25*inca), *(pi1_r +25) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +26*inca), *(pi1_r +26) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +27*inca), *(pi1_r +27) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +28*inca), *(pi1_r +28) ); \
PASTEMAC(ch,scal2ios)( *kappa_cast, *(alpha1 +29*inca), *(pi1_r +29) ); \
\
alpha1 += lda; \
pi1_r += ldp; \
} \
} \
} \
} \
else /* if ( bli_is_rpi_packed( schema ) ) */ \
{ \
if ( PASTEMAC(ch,eq1)( *kappa_cast ) ) \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(chr,add3s)( *(alpha1_r + 0*inca2), -*(alpha1_i + 0*inca2), *(pi1_r + 0) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 1*inca2), -*(alpha1_i + 1*inca2), *(pi1_r + 1) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 2*inca2), -*(alpha1_i + 2*inca2), *(pi1_r + 2) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 3*inca2), -*(alpha1_i + 3*inca2), *(pi1_r + 3) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 4*inca2), -*(alpha1_i + 4*inca2), *(pi1_r + 4) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 5*inca2), -*(alpha1_i + 5*inca2), *(pi1_r + 5) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 6*inca2), -*(alpha1_i + 6*inca2), *(pi1_r + 6) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 7*inca2), -*(alpha1_i + 7*inca2), *(pi1_r + 7) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 8*inca2), -*(alpha1_i + 8*inca2), *(pi1_r + 8) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 9*inca2), -*(alpha1_i + 9*inca2), *(pi1_r + 9) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +10*inca2), -*(alpha1_i +10*inca2), *(pi1_r +10) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +11*inca2), -*(alpha1_i +11*inca2), *(pi1_r +11) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +12*inca2), -*(alpha1_i +12*inca2), *(pi1_r +12) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +13*inca2), -*(alpha1_i +13*inca2), *(pi1_r +13) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +14*inca2), -*(alpha1_i +14*inca2), *(pi1_r +14) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +15*inca2), -*(alpha1_i +15*inca2), *(pi1_r +15) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +16*inca2), -*(alpha1_i +16*inca2), *(pi1_r +16) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +17*inca2), -*(alpha1_i +17*inca2), *(pi1_r +17) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +18*inca2), -*(alpha1_i +18*inca2), *(pi1_r +18) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +19*inca2), -*(alpha1_i +19*inca2), *(pi1_r +19) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +20*inca2), -*(alpha1_i +20*inca2), *(pi1_r +20) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +21*inca2), -*(alpha1_i +21*inca2), *(pi1_r +21) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +22*inca2), -*(alpha1_i +22*inca2), *(pi1_r +22) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +23*inca2), -*(alpha1_i +23*inca2), *(pi1_r +23) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +24*inca2), -*(alpha1_i +24*inca2), *(pi1_r +24) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +25*inca2), -*(alpha1_i +25*inca2), *(pi1_r +25) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +26*inca2), -*(alpha1_i +26*inca2), *(pi1_r +26) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +27*inca2), -*(alpha1_i +27*inca2), *(pi1_r +27) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +28*inca2), -*(alpha1_i +28*inca2), *(pi1_r +28) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +29*inca2), -*(alpha1_i +29*inca2), *(pi1_r +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(chr,add3s)( *(alpha1_r + 0*inca2), *(alpha1_i + 0*inca2), *(pi1_r + 0) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 1*inca2), *(alpha1_i + 1*inca2), *(pi1_r + 1) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 2*inca2), *(alpha1_i + 2*inca2), *(pi1_r + 2) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 3*inca2), *(alpha1_i + 3*inca2), *(pi1_r + 3) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 4*inca2), *(alpha1_i + 4*inca2), *(pi1_r + 4) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 5*inca2), *(alpha1_i + 5*inca2), *(pi1_r + 5) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 6*inca2), *(alpha1_i + 6*inca2), *(pi1_r + 6) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 7*inca2), *(alpha1_i + 7*inca2), *(pi1_r + 7) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 8*inca2), *(alpha1_i + 8*inca2), *(pi1_r + 8) ); \
PASTEMAC(chr,add3s)( *(alpha1_r + 9*inca2), *(alpha1_i + 9*inca2), *(pi1_r + 9) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +10*inca2), *(alpha1_i +10*inca2), *(pi1_r +10) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +11*inca2), *(alpha1_i +11*inca2), *(pi1_r +11) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +12*inca2), *(alpha1_i +12*inca2), *(pi1_r +12) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +13*inca2), *(alpha1_i +13*inca2), *(pi1_r +13) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +14*inca2), *(alpha1_i +14*inca2), *(pi1_r +14) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +15*inca2), *(alpha1_i +15*inca2), *(pi1_r +15) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +16*inca2), *(alpha1_i +16*inca2), *(pi1_r +16) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +17*inca2), *(alpha1_i +17*inca2), *(pi1_r +17) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +18*inca2), *(alpha1_i +18*inca2), *(pi1_r +18) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +19*inca2), *(alpha1_i +19*inca2), *(pi1_r +19) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +20*inca2), *(alpha1_i +20*inca2), *(pi1_r +20) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +21*inca2), *(alpha1_i +21*inca2), *(pi1_r +21) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +22*inca2), *(alpha1_i +22*inca2), *(pi1_r +22) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +23*inca2), *(alpha1_i +23*inca2), *(pi1_r +23) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +24*inca2), *(alpha1_i +24*inca2), *(pi1_r +24) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +25*inca2), *(alpha1_i +25*inca2), *(pi1_r +25) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +26*inca2), *(alpha1_i +26*inca2), *(pi1_r +26) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +27*inca2), *(alpha1_i +27*inca2), *(pi1_r +27) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +28*inca2), *(alpha1_i +28*inca2), *(pi1_r +28) ); \
PASTEMAC(chr,add3s)( *(alpha1_r +29*inca2), *(alpha1_i +29*inca2), *(pi1_r +29) ); \
\
alpha1_r += lda2; \
alpha1_i += lda2; \
pi1_r += ldp; \
} \
} \
} \
else \
{ \
if ( bli_is_conj( conja ) ) \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 0*inca), *(pi1_r + 0) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 1*inca), *(pi1_r + 1) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 2*inca), *(pi1_r + 2) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 3*inca), *(pi1_r + 3) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 4*inca), *(pi1_r + 4) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 5*inca), *(pi1_r + 5) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 6*inca), *(pi1_r + 6) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 7*inca), *(pi1_r + 7) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 8*inca), *(pi1_r + 8) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 + 9*inca), *(pi1_r + 9) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +10*inca), *(pi1_r +10) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +11*inca), *(pi1_r +11) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +12*inca), *(pi1_r +12) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +13*inca), *(pi1_r +13) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +14*inca), *(pi1_r +14) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +15*inca), *(pi1_r +15) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +16*inca), *(pi1_r +16) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +17*inca), *(pi1_r +17) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +18*inca), *(pi1_r +18) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +19*inca), *(pi1_r +19) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +20*inca), *(pi1_r +20) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +21*inca), *(pi1_r +21) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +22*inca), *(pi1_r +22) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +23*inca), *(pi1_r +23) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +24*inca), *(pi1_r +24) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +25*inca), *(pi1_r +25) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +26*inca), *(pi1_r +26) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +27*inca), *(pi1_r +27) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +28*inca), *(pi1_r +28) ); \
PASTEMAC(ch,scal2jrpis)( *kappa_cast, *(alpha1 +29*inca), *(pi1_r +29) ); \
\
alpha1 += lda; \
pi1_r += ldp; \
} \
} \
else \
{ \
for ( ; n != 0; --n ) \
{ \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 0*inca), *(pi1_r + 0) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 1*inca), *(pi1_r + 1) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 2*inca), *(pi1_r + 2) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 3*inca), *(pi1_r + 3) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 4*inca), *(pi1_r + 4) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 5*inca), *(pi1_r + 5) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 6*inca), *(pi1_r + 6) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 7*inca), *(pi1_r + 7) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 8*inca), *(pi1_r + 8) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 + 9*inca), *(pi1_r + 9) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +10*inca), *(pi1_r +10) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +11*inca), *(pi1_r +11) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +12*inca), *(pi1_r +12) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +13*inca), *(pi1_r +13) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +14*inca), *(pi1_r +14) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +15*inca), *(pi1_r +15) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +16*inca), *(pi1_r +16) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +17*inca), *(pi1_r +17) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +18*inca), *(pi1_r +18) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +19*inca), *(pi1_r +19) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +20*inca), *(pi1_r +20) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +21*inca), *(pi1_r +21) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +22*inca), *(pi1_r +22) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +23*inca), *(pi1_r +23) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +24*inca), *(pi1_r +24) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +25*inca), *(pi1_r +25) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +26*inca), *(pi1_r +26) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +27*inca), *(pi1_r +27) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +28*inca), *(pi1_r +28) ); \
PASTEMAC(ch,scal2rpis)( *kappa_cast, *(alpha1 +29*inca), *(pi1_r +29) ); \
\
alpha1 += lda; \
pi1_r += ldp; \
} \
} \
} \
} \
}
INSERT_GENTFUNCCO_BASIC0( packm_ref_30xk_rih )

View File

@@ -52,4 +52,5 @@ INSERT_GENTPROT_BASIC( packm_ref_10xk_rih )
INSERT_GENTPROT_BASIC( packm_ref_12xk_rih )
INSERT_GENTPROT_BASIC( packm_ref_14xk_rih )
INSERT_GENTPROT_BASIC( packm_ref_16xk_rih )
INSERT_GENTPROT_BASIC( packm_ref_30xk_rih )

View File

@@ -186,6 +186,16 @@
#define BLIS_ZPACKM_16XK_3M_KERNEL BLIS_ZPACKM_16XK_3M_KERNEL_REF
#endif
// packm_30xk_3m kernels
#ifndef BLIS_CPACKM_30XK_3M_KERNEL
#define BLIS_CPACKM_30XK_3M_KERNEL BLIS_CPACKM_30XK_3M_KERNEL_REF
#endif
#ifndef BLIS_ZPACKM_30XK_3M_KERNEL
#define BLIS_ZPACKM_30XK_3M_KERNEL BLIS_ZPACKM_30XK_3M_KERNEL_REF
#endif
#endif

View File

@@ -186,6 +186,16 @@
#define BLIS_ZPACKM_16XK_4M_KERNEL BLIS_ZPACKM_16XK_4M_KERNEL_REF
#endif
// packm_30xk_4m kernels
#ifndef BLIS_CPACKM_30XK_4M_KERNEL
#define BLIS_CPACKM_30XK_4M_KERNEL BLIS_CPACKM_30XK_4M_KERNEL_REF
#endif
#ifndef BLIS_ZPACKM_30XK_4M_KERNEL
#define BLIS_ZPACKM_30XK_4M_KERNEL BLIS_ZPACKM_30XK_4M_KERNEL_REF
#endif
#endif

View File

@@ -333,6 +333,24 @@
#define BLIS_ZPACKM_16XK_KERNEL BLIS_ZPACKM_16XK_KERNEL_REF
#endif
// packm_30xk kernels
#ifndef BLIS_SPACKM_30XK_KERNEL
#define BLIS_SPACKM_30XK_KERNEL BLIS_SPACKM_30XK_KERNEL_REF
#endif
#ifndef BLIS_DPACKM_30XK_KERNEL
#define BLIS_DPACKM_30XK_KERNEL BLIS_DPACKM_30XK_KERNEL_REF
#endif
#ifndef BLIS_CPACKM_30XK_KERNEL
#define BLIS_CPACKM_30XK_KERNEL BLIS_CPACKM_30XK_KERNEL_REF
#endif
#ifndef BLIS_ZPACKM_30XK_KERNEL
#define BLIS_ZPACKM_30XK_KERNEL BLIS_ZPACKM_30XK_KERNEL_REF
#endif
// unpackm_2xk kernels
#ifndef BLIS_SUNPACKM_2XK_KERNEL

View File

@@ -212,6 +212,13 @@
#define BLIS_CPACKM_16XK_KERNEL_REF bli_cpackm_ref_16xk
#define BLIS_ZPACKM_16XK_KERNEL_REF bli_zpackm_ref_16xk
// packm_30xk kernels
#define BLIS_SPACKM_30XK_KERNEL_REF bli_spackm_ref_30xk
#define BLIS_DPACKM_30XK_KERNEL_REF bli_dpackm_ref_30xk
#define BLIS_CPACKM_30XK_KERNEL_REF bli_cpackm_ref_30xk
#define BLIS_ZPACKM_30XK_KERNEL_REF bli_zpackm_ref_30xk
// packm_2xk_4m kernels
#define BLIS_CPACKM_2XK_4M_KERNEL_REF bli_cpackm_ref_2xk_4m
@@ -252,6 +259,11 @@
#define BLIS_CPACKM_16XK_4M_KERNEL_REF bli_cpackm_ref_16xk_4m
#define BLIS_ZPACKM_16XK_4M_KERNEL_REF bli_zpackm_ref_16xk_4m
// packm_30xk_4m kernels
#define BLIS_CPACKM_30XK_4M_KERNEL_REF bli_cpackm_ref_30xk_4m
#define BLIS_ZPACKM_30XK_4M_KERNEL_REF bli_zpackm_ref_30xk_4m
// packm_2xk_3m kernels
#define BLIS_CPACKM_2XK_3M_KERNEL_REF bli_cpackm_ref_2xk_3m
@@ -292,6 +304,11 @@
#define BLIS_CPACKM_16XK_3M_KERNEL_REF bli_cpackm_ref_16xk_3m
#define BLIS_ZPACKM_16XK_3M_KERNEL_REF bli_zpackm_ref_16xk_3m
// packm_30xk_3m kernels
#define BLIS_CPACKM_30XK_3M_KERNEL_REF bli_cpackm_ref_30xk_3m
#define BLIS_ZPACKM_30XK_3M_KERNEL_REF bli_zpackm_ref_30xk_3m
// packm_2xk_rih kernels
#define BLIS_CPACKM_2XK_RIH_KERNEL_REF bli_cpackm_ref_2xk_rih
@@ -332,6 +349,11 @@
#define BLIS_CPACKM_16XK_RIH_KERNEL_REF bli_cpackm_ref_16xk_rih
#define BLIS_ZPACKM_16XK_RIH_KERNEL_REF bli_zpackm_ref_16xk_rih
// packm_30xk_rih kernels
#define BLIS_CPACKM_30XK_RIH_KERNEL_REF bli_cpackm_ref_30xk_rih
#define BLIS_ZPACKM_30XK_RIH_KERNEL_REF bli_zpackm_ref_30xk_rih
// unpack_2xk kernels
#define BLIS_SUNPACKM_2XK_KERNEL_REF bli_sunpackm_ref_2xk

View File

@@ -163,6 +163,16 @@
#define BLIS_ZPACKM_16XK_RIH_KERNEL BLIS_ZPACKM_16XK_RIH_KERNEL_REF
#endif
// packm_30xk_rih kernels
#ifndef BLIS_CPACKM_30XK_RIH_KERNEL
#define BLIS_CPACKM_30XK_RIH_KERNEL BLIS_CPACKM_30XK_RIH_KERNEL_REF
#endif
#ifndef BLIS_ZPACKM_30XK_RIH_KERNEL
#define BLIS_ZPACKM_30XK_RIH_KERNEL BLIS_ZPACKM_30XK_RIH_KERNEL_REF
#endif
#endif