Merge branch 'release-v0.1alpha2' into loongson3a

kernel/Makefile

@@ -53,6 +53,11 @@ SBLASOBJS	+= setparam$(TSUFFIX).$(SUFFIX)
 CCOMMON_OPT	+= -DTS=$(TSUFFIX)
 endif
 
+KERNEL_INTERFACE = ../common_level1.h ../common_level2.h ../common_level3.h 
+ifneq ($(NO_LAPACK), 1)
+KERNEL_INTERFACE += ../common_lapack.h 
+endif
+
 ifeq ($(ARCH), x86)
 COMMONOBJS	+= cpuid.$(SUFFIX)
 endif
@@ -88,9 +93,10 @@ setparam$(TSUFFIX).$(SUFFIX): setparam$(TSUFFIX).c kernel$(TSUFFIX).h
 setparam$(TSUFFIX).c : setparam-ref.c
 	sed 's/TS/$(TSUFFIX)/g' $< > $(@F)
 
-kernel$(TSUFFIX).h : ../common_level1.h ../common_level2.h ../common_level3.h ../common_lapack.h 
+kernel$(TSUFFIX).h : $(KERNEL_INTERFACE) 
 	sed 's/\ *(/$(TSUFFIX)(/g' $^ > $(@F)
 
+
 cpuid.$(SUFFIX): $(KERNELDIR)/cpuid.S
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
@@ -112,10 +118,10 @@ lsame.$(PSUFFIX): $(KERNELDIR)/$(LSAME_KERNEL)
 cpuid.$(PSUFFIX): $(KERNELDIR)/cpuid.S
 	$(CC) -c $(PFLAGS) $< -o $(@F)
 
-ifdef DYNAMIC_ARCH
+#ifdef DYNAMIC_ARCH
 clean ::
 	@rm -f setparam_*.c kernel_*.h setparam.h kernel.h
 
-endif
+#endif
 
 include $(TOPDIR)/Makefile.tail

kernel/Makefile.L1

@@ -668,7 +668,7 @@ $(KDIR)qdot_k$(TSUFFIX).$(SUFFIX) $(KDIR)qdot_k$(TPSUFFIX).$(PSUFFIX) : $(KERNEL
 	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE $< -o $@
 
 $(KDIR)dsdot_k$(TSUFFIX).$(SUFFIX) $(KDIR)dsdot_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SDOTKERNEL) 
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE -DDSDOT $< -o $@
 
 $(KDIR)sdot_k$(TSUFFIX).$(SUFFIX) $(KDIR)sdot_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SDOTKERNEL) 
 	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@

kernel/mips64/dot.S

@@ -300,7 +300,11 @@
 	.align 3
 
 .L999:
-	j	$31
 	ADD	s1, s1, s2
-
+#ifdef DSDOT
+	cvt.d.s s1, s1
+#endif
+	j	$31
+	NOP
+	
 	EPILOGUE

kernel/setparam-ref.c

@@ -101,7 +101,11 @@ gotoblas_t TABLE_NAME = {
 #endif
   ssymm_outcopyTS, ssymm_oltcopyTS,
 
+#ifndef NO_LAPACK
   sneg_tcopyTS, slaswp_ncopyTS,
+#else
+  NULL,NULL,
+#endif
 
   0, 0, 0,
   DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
@@ -147,7 +151,11 @@ gotoblas_t TABLE_NAME = {
 #endif
   dsymm_outcopyTS, dsymm_oltcopyTS,
 
+#ifndef NO_LAPACK
   dneg_tcopyTS, dlaswp_ncopyTS,
+#else
+  NULL, NULL,
+#endif
 
 #ifdef EXPRECISION
 
@@ -195,7 +203,11 @@ gotoblas_t TABLE_NAME = {
 #endif
   qsymm_outcopyTS, qsymm_oltcopyTS,
 
+#ifndef NO_LAPACK
   qneg_tcopyTS, qlaswp_ncopyTS,
+#else
+  NULL, NULL,
+#endif
 
 #endif
 
@@ -286,7 +298,11 @@ gotoblas_t TABLE_NAME = {
   chemm3m_oucopyrTS,  chemm3m_olcopyrTS,
   chemm3m_oucopyiTS,  chemm3m_olcopyiTS,
 
+#ifndef NO_LAPACK
   cneg_tcopyTS, claswp_ncopyTS,
+#else
+  NULL, NULL,
+#endif
 
   0, 0, 0,
   ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N, MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
@@ -375,7 +391,11 @@ gotoblas_t TABLE_NAME = {
   zhemm3m_oucopyrTS,  zhemm3m_olcopyrTS,
   zhemm3m_oucopyiTS,  zhemm3m_olcopyiTS,
 
+#ifndef NO_LAPACK
   zneg_tcopyTS, zlaswp_ncopyTS,
+#else
+  NULL, NULL,
+#endif
 
 #ifdef EXPRECISION
 
@@ -466,7 +486,11 @@ gotoblas_t TABLE_NAME = {
   xhemm3m_oucopyrTS,  xhemm3m_olcopyrTS,
   xhemm3m_oucopyiTS,  xhemm3m_olcopyiTS,
 
+#ifndef NO_LAPACK
   xneg_tcopyTS, xlaswp_ncopyTS,
+#else
+  NULL, NULL,
+#endif
 
 #endif
 

kernel/x86/zdot_sse2.S

@@ -1541,5 +1541,8 @@
 	popl	%ebx
 	popl	%esi
 	popl	%edi
+/*remove the hidden return value address from the stack.*/
+	popl	%ecx
+	xchgl	%ecx, 0(%esp)
 	ret
 	EPILOGUE

kernel/x86_64/dot_sse.S

@@ -1286,6 +1286,10 @@
 	haddps	%xmm0, %xmm0
 #endif
 
+#ifdef DSDOT
+	cvtss2sd %xmm0, %xmm0
+#endif
+
 	RESTOREREGISTERS
 
 	ret

kernel/x86_64/zgemm_kernel_1x4_nehalem.S

@@ -544,7 +544,7 @@
 	jg	.L11
 
 #if defined(TRMMKERNEL) && !defined(LEFT)
-	addq	$1, KK
+	addq	$4, KK
 #endif
 
 	leaq	(C, LDC, 4), C
@@ -594,7 +594,7 @@
 	jg	.L11
 
 #if defined(TRMMKERNEL) && !defined(LEFT)
-	addq	$1, KK
+	addq	$4, KK
 #endif
 
 	leaq	(C, LDC, 4), C