optimized dgemm for POWER8

kernel/power/KERNEL.POWER8

@@ -21,12 +21,12 @@ SGEMMOTCOPYOBJ =  sgemm_otcopy.o
 DGEMMKERNEL    =  dgemm_kernel_16x4_power8.S
 DGEMMINCOPY    = ../generic/gemm_ncopy_16.c
 DGEMMITCOPY    =  dgemm_tcopy_16_power8.S
-DGEMMONCOPY    =  gemm_ncopy_4.S
-DGEMMOTCOPY    =  gemm_tcopy_4.S
-DGEMMINCOPYOBJ = dgemm_incopy.o
-DGEMMITCOPYOBJ = dgemm_itcopy.o
-DGEMMONCOPYOBJ = dgemm_oncopy.o
-DGEMMOTCOPYOBJ = dgemm_otcopy.o
+DGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMINCOPYOBJ =  dgemm_incopy.o
+DGEMMITCOPYOBJ =  dgemm_itcopy.o
+DGEMMONCOPYOBJ =  dgemm_oncopy.o
+DGEMMOTCOPYOBJ =  dgemm_otcopy.o
 
 CGEMMKERNEL    = cgemm_kernel_8x4_power8.S
 CGEMMINCOPY    = ../generic/zgemm_ncopy_8.c

kernel/power/dgemm_kernel_16x4_power8.S

@@ -131,6 +131,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define o0	0
 
+#define T4	r12
+#define T3	r11
+
 #define o8	r15
 #define o24	r16
 #define ALPHA	r17
@@ -265,7 +268,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi	ALPHA, SP, 224
 #endif
 
-	li	PRE, 256 
+	li	PRE, 384 
 	li	o8 , 8
 	li	o16, 16
 	li	o24, 24

kernel/power/dgemm_macros_16x4_power8.S

@@ -431,6 +431,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	mr		T1,	CO
 	addi		T2,	T1,	64
+	add		T3,	T1,	LDC
+	addi		T4,	T3,	64
 
 #ifndef TRMMKERNEL
 	lxvd2x		vs0,	0,	T1
@@ -442,6 +444,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	lxvd2x		vs5,	o16,	T2
 	lxvd2x		vs6,	o32,	T2
 	lxvd2x		vs7,	o48,	T2
+
+	lxvd2x		vs8,	0,	T3
+	lxvd2x		vs9,	o16,	T3
+	lxvd2x		vs10,	o32,	T3
+	lxvd2x		vs11,	o48,	T3
+
+	lxvd2x		vs12,	0,	T4
+	lxvd2x		vs13,	o16,	T4
+	lxvd2x		vs14,	o32,	T4
+	lxvd2x		vs15,	o48,	T4
 #endif
 
 #ifndef TRMMKERNEL
@@ -453,45 +465,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xvmaddadp	vs5,	vs37,	alpha_r
 	xvmaddadp	vs6,	vs38,	alpha_r
 	xvmaddadp	vs7,	vs39,	alpha_r
-#else
-	xvmuldp		vs0,	vs32,	alpha_r
-	xvmuldp		vs1,	vs33,	alpha_r
-	xvmuldp		vs2,	vs34,	alpha_r
-	xvmuldp		vs3,	vs35,	alpha_r
-	xvmuldp		vs4,	vs36,	alpha_r
-	xvmuldp		vs5,	vs37,	alpha_r
-	xvmuldp		vs6,	vs38,	alpha_r
-	xvmuldp		vs7,	vs39,	alpha_r
-#endif
-
-	stxvd2x		vs0,	0,	T1
-	stxvd2x		vs1,	o16,	T1
-	stxvd2x		vs2,	o32,	T1
-	stxvd2x		vs3,	o48,	T1
-
-	dcbt		T1, PRE
-
-	stxvd2x		vs4,	0,	T2
-	stxvd2x		vs5,	o16,	T2
-	stxvd2x		vs6,	o32,	T2
-	stxvd2x		vs7,	o48,	T2
-
-	add		T1,	T1,	LDC
-	add		T2,	T2,	LDC
-
-#ifndef TRMMKERNEL
-	lxvd2x		vs8,	0,	T1
-	lxvd2x		vs9,	o16,	T1
-	lxvd2x		vs10,	o32,	T1
-	lxvd2x		vs11,	o48,	T1
-
-	lxvd2x		vs12,	0,	T2
-	lxvd2x		vs13,	o16,	T2
-	lxvd2x		vs14,	o32,	T2
-	lxvd2x		vs15,	o48,	T2
-#endif
-
-#ifndef TRMMKERNEL
 	xvmaddadp	vs8,	vs40,	alpha_r
 	xvmaddadp	vs9,	vs41,	alpha_r
 	xvmaddadp	vs10,	vs42,	alpha_r
@@ -501,6 +474,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xvmaddadp	vs14,	vs46,	alpha_r
 	xvmaddadp	vs15,	vs47,	alpha_r
 #else
+	xvmuldp		vs0,	vs32,	alpha_r
+	xvmuldp		vs1,	vs33,	alpha_r
+	xvmuldp		vs2,	vs34,	alpha_r
+	xvmuldp		vs3,	vs35,	alpha_r
+	xvmuldp		vs4,	vs36,	alpha_r
+	xvmuldp		vs5,	vs37,	alpha_r
+	xvmuldp		vs6,	vs38,	alpha_r
+	xvmuldp		vs7,	vs39,	alpha_r
 	xvmuldp		vs8,	vs40,	alpha_r
 	xvmuldp		vs9,	vs41,	alpha_r
 	xvmuldp		vs10,	vs42,	alpha_r
@@ -511,20 +492,31 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xvmuldp		vs15,	vs47,	alpha_r
 #endif
 
-	stxvd2x		vs8,	0,	T1
-	stxvd2x		vs9,	o16,	T1
-	stxvd2x		vs10,	o32,	T1
-	stxvd2x		vs11,	o48,	T1
+	stxvd2x		vs0,	0,	T1
+	stxvd2x		vs1,	o16,	T1
+	stxvd2x		vs2,	o32,	T1
+	stxvd2x		vs3,	o48,	T1
 
-	dcbt		T1, PRE
+	stxvd2x		vs4,	0,	T2
+	stxvd2x		vs5,	o16,	T2
+	stxvd2x		vs6,	o32,	T2
+	stxvd2x		vs7,	o48,	T2
 
-	stxvd2x		vs12,	0,	T2
-	stxvd2x		vs13,	o16,	T2
-	stxvd2x		vs14,	o32,	T2
-	stxvd2x		vs15,	o48,	T2
+	stxvd2x		vs8,	0,	T3
+	stxvd2x		vs9,	o16,	T3
+	stxvd2x		vs10,	o32,	T3
+	stxvd2x		vs11,	o48,	T3
 
-	add		T1,	T1,	LDC
-	add		T2,	T2,	LDC
+	stxvd2x		vs12,	0,	T4
+	stxvd2x		vs13,	o16,	T4
+	stxvd2x		vs14,	o32,	T4
+	stxvd2x		vs15,	o48,	T4
+
+	slwi		T4,	LDC,	1
+	add		T1,	T1,	T4
+	add		T3,	T3,	T4
+	addi		T2,	T1,	64
+	addi		T4,	T3,	64
 
 #ifndef TRMMKERNEL
 	lxvd2x		vs0,	0,	T1
@@ -536,6 +528,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	lxvd2x		vs5,	o16,	T2
 	lxvd2x		vs6,	o32,	T2
 	lxvd2x		vs7,	o48,	T2
+
+	lxvd2x		vs8,	0,	T3
+	lxvd2x		vs9,	o16,	T3
+	lxvd2x		vs10,	o32,	T3
+	lxvd2x		vs11,	o48,	T3
+
+	lxvd2x		vs12,	0,	T4
+	lxvd2x		vs13,	o16,	T4
+	lxvd2x		vs14,	o32,	T4
+	lxvd2x		vs15,	o48,	T4
 #endif
 
 #ifndef TRMMKERNEL
@@ -547,45 +549,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xvmaddadp	vs5,	vs53,	alpha_r
 	xvmaddadp	vs6,	vs54,	alpha_r
 	xvmaddadp	vs7,	vs55,	alpha_r
-#else
-	xvmuldp		vs0,	vs48,	alpha_r
-	xvmuldp		vs1,	vs49,	alpha_r
-	xvmuldp		vs2,	vs50,	alpha_r
-	xvmuldp		vs3,	vs51,	alpha_r
-	xvmuldp		vs4,	vs52,	alpha_r
-	xvmuldp		vs5,	vs53,	alpha_r
-	xvmuldp		vs6,	vs54,	alpha_r
-	xvmuldp		vs7,	vs55,	alpha_r
-#endif
-
-	stxvd2x		vs0,	0,	T1
-	stxvd2x		vs1,	o16,	T1
-	stxvd2x		vs2,	o32,	T1
-	stxvd2x		vs3,	o48,	T1
-
-	dcbt		T1, PRE
-
-	stxvd2x		vs4,	0,	T2
-	stxvd2x		vs5,	o16,	T2
-	stxvd2x		vs6,	o32,	T2
-	stxvd2x		vs7,	o48,	T2
-
-	add		T1,	T1,	LDC
-	add		T2,	T2,	LDC
-
-#ifndef TRMMKERNEL
-	lxvd2x		vs8,	0,	T1
-	lxvd2x		vs9,	o16,	T1
-	lxvd2x		vs10,	o32,	T1
-	lxvd2x		vs11,	o48,	T1
-
-	lxvd2x		vs12,	0,	T2
-	lxvd2x		vs13,	o16,	T2
-	lxvd2x		vs14,	o32,	T2
-	lxvd2x		vs15,	o48,	T2
-#endif
-
-#ifndef TRMMKERNEL
 	xvmaddadp	vs8,	vs56,	alpha_r
 	xvmaddadp	vs9,	vs57,	alpha_r
 	xvmaddadp	vs10,	vs58,	alpha_r
@@ -595,6 +558,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xvmaddadp	vs14,	vs62,	alpha_r
 	xvmaddadp	vs15,	vs63,	alpha_r
 #else
+	xvmuldp		vs0,	vs48,	alpha_r
+	xvmuldp		vs1,	vs49,	alpha_r
+	xvmuldp		vs2,	vs50,	alpha_r
+	xvmuldp		vs3,	vs51,	alpha_r
+	xvmuldp		vs4,	vs52,	alpha_r
+	xvmuldp		vs5,	vs53,	alpha_r
+	xvmuldp		vs6,	vs54,	alpha_r
+	xvmuldp		vs7,	vs55,	alpha_r
 	xvmuldp		vs8,	vs56,	alpha_r
 	xvmuldp		vs9,	vs57,	alpha_r
 	xvmuldp		vs10,	vs58,	alpha_r
@@ -605,17 +576,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xvmuldp		vs15,	vs63,	alpha_r
 #endif
 
-	stxvd2x		vs8,	0,	T1
-	stxvd2x		vs9,	o16,	T1
-	stxvd2x		vs10,	o32,	T1
-	stxvd2x		vs11,	o48,	T1
+	stxvd2x		vs0,	0,	T1
+	stxvd2x		vs1,	o16,	T1
+	stxvd2x		vs2,	o32,	T1
+	stxvd2x		vs3,	o48,	T1
 
-	dcbt		T1, PRE
+	stxvd2x		vs4,	0,	T2
+	stxvd2x		vs5,	o16,	T2
+	stxvd2x		vs6,	o32,	T2
+	stxvd2x		vs7,	o48,	T2
 
-	stxvd2x		vs12,	0,	T2
-	stxvd2x		vs13,	o16,	T2
-	stxvd2x		vs14,	o32,	T2
-	stxvd2x		vs15,	o48,	T2
+	stxvd2x		vs8,	0,	T3
+	stxvd2x		vs9,	o16,	T3
+	stxvd2x		vs10,	o32,	T3
+	stxvd2x		vs11,	o48,	T3
+
+	stxvd2x		vs12,	0,	T4
+	stxvd2x		vs13,	o16,	T4
+	stxvd2x		vs14,	o32,	T4
+	stxvd2x		vs15,	o48,	T4
 
 	addi		CO,	CO,	128
 

kernel/power/dgemm_tcopy_16_power8.S

@@ -170,7 +170,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	B2, B2, B
 	add	B1, B1, B
 
-	li	PREA,  768
+	li	PREA,  256
 	addi	PREB,  M16, 128
 
 	li	o8,	8

kernel/power/dgemm_tcopy_logic_16_power8.S

@@ -57,16 +57,20 @@ DCOPYT_L4_BEGIN:
 
 DCOPYT_L4x16_LOOP:
 
+/*
 	addi	T1,	PREB,	128
 	addi	T2,	PREB,	256
+*/
 	dcbt	A0,	PREA
 	dcbt	A1,	PREA
 	dcbt	A2,	PREA
 	dcbt	A3,	PREA
+/*
 	dcbtst  BO,	M16
 	dcbtst  BO,	PREB
 	dcbtst  BO,	T1
 	dcbtst  BO,	T2
+*/
 	COPY_4x16
 
 	add		BO,	BO,	M16

kernel/power/dtrmm_kernel_16x4_power8.S

@@ -152,7 +152,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define PRE	r30
 #define T2	r31
 
-#include "dgemm_macros_16x4_power8.S"
+#include "dtrmm_macros_16x4_power8.S"
 
 
 #ifndef NEEDPARAM