Merge 76d675bd55 into d6a5174e9c

kernel/arm/iamax.c

@@ -56,13 +56,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG max=0;
 
 	if (n <= 0 || inc_x <= 0) return(max);
-
+        if (n==1) return(1);
+        if (x[0]!=x[0]) return(1);
 	maxf=ABS(x[0]);
 	ix += inc_x;
 	i++;
 
 	while(i < n)
 	{
+        	if (x[ix]!=x[ix]) return(i+1);
 		if( ABS(x[ix]) > maxf )
 		{
 			max = i;

kernel/arm/iamin.c

@@ -56,13 +56,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG min=0;
 
 	if (n <= 0 || inc_x <= 0) return(min);
-
+        if (n==1) return(1);      
-	minf=ABS(x[0]);
+        if (x[0]!=x[0]) return(1);
+        minf=ABS(x[0]);
 	ix += inc_x;
 	i++;
 
 	while(i < n)
 	{
+        	if (x[ix]!=x[ix]) return(i+1);
 		if( ABS(x[ix]) < ABS(minf) )
 		{
 			min = i;

kernel/mips/iamax.c

@@ -1,5 +1,5 @@
 /***************************************************************************
-Copyright (c) 2016, The OpenBLAS Project
+Copyright (c) 2013, The OpenBLAS Project
 All rights reserved.
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions are
@@ -25,6 +25,15 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
+/**************************************************************************************
+* 2013/09/14 Saar
+*	 BLASTEST float		: NoTest
+* 	 BLASTEST double	: NoTest
+* 	 CTEST			: OK
+* 	 TEST			: OK
+*
+**************************************************************************************/
+
 #include "common.h"
 #include <math.h>
 
@@ -47,13 +56,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG max=0;
 
 	if (n <= 0 || inc_x <= 0) return(max);
-
+        if (n==1) return(1);
+        if (x[0]!=x[0]) return(1);
 	maxf=ABS(x[0]);
 	ix += inc_x;
 	i++;
 
 	while(i < n)
 	{
+        	if (x[ix]!=x[ix]) return(i+1);
 		if( ABS(x[ix]) > maxf )
 		{
 			max = i;

kernel/mips/iamin.c

@@ -1,5 +1,5 @@
 /***************************************************************************
-Copyright (c) 2016, The OpenBLAS Project
+Copyright (c) 2013, The OpenBLAS Project
 All rights reserved.
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions are
@@ -25,6 +25,15 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
+/**************************************************************************************
+* 2013/09/14 Saar
+*	 BLASTEST float		: NoTest
+* 	 BLASTEST double	: NoTest
+* 	 CTEST			: NoTest
+* 	 TEST			: NoTest
+*
+**************************************************************************************/
+
 #include "common.h"
 #include <math.h>
 
@@ -47,13 +56,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG min=0;
 
 	if (n <= 0 || inc_x <= 0) return(min);
-
+        if (n==1) return(1);      
-	minf=ABS(x[0]);
+        if (x[0]!=x[0]) return(1);
+        minf=ABS(x[0]);
 	ix += inc_x;
 	i++;
 
 	while(i < n)
 	{
+        	if (x[ix]!=x[ix]) return(i+1);
 		if( ABS(x[ix]) < ABS(minf) )
 		{
 			min = i;

kernel/power/iamax.S

@@ -520,6 +520,19 @@ LL(1000):
 	.align 4
 
 LL(1010):
+	addi	RET, RET, 1
+	fcmpu	cr0, f24, f24
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f25, f25
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f26, f26
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f27, f27
+	bun	cr0, LL(9999)
+
 	fabs	f8,  f24
 	fabs	f9,  f25
 	fabs	f10, f26
@@ -530,6 +543,20 @@ LL(1010):
 	LFD	f26,  10 * SIZE(XX)
 	LFD	f27,  11 * SIZE(XX)
 	
+	addi	RET, RET, 1
+	fcmpu	cr0, f24, f24
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f25, f25
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f26, f26
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f27, f27
+	bun	cr0, LL(9999)
+	subi    RET, RET, 8
+
 	fabs	f12, f28
 	fabs	f13, f29
 	fabs	f14, f30
@@ -577,6 +604,32 @@ LL(1010):
 	.align 4
 
 LL(1020):
+	addi	RET, RET, 1
+	fcmpu	cr0, f24, f24
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f25, f25
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f26, f26
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f27, f27
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f28, f28
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f29, f29
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f30, f30
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f31, f31
+	bun	cr0, LL(9999)
+	subi	RET, RET, 8
+
 	fabs	f8,  f24
 	fabs	f9,  f25
 	fabs	f10, f26
@@ -631,8 +684,12 @@ LL(1050):
 LL(1060):
 	LFD	f8,  0 * SIZE(XX)
 	addi	XX, XX,  1 * SIZE
+	addi    RET, RET, 1
+	fcmpu   cr0, f8, f8
+	bun	cru, LL(9999)
+	
 	fabs	f8, f8
-	addi	RET, RET, 1
+	//addi	RET, RET, 1
 	fcmpu	cr0, f1, f8
 	beq	cr0, LL(9999)
 	bdnz	LL(1060)
@@ -658,6 +715,18 @@ LL(1100):
 	.align 4
 
 LL(1110):
+	addi	RET, RET, 1
+	fcmpu	cr0, f24, f24
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f25, f25
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f26, f26
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f27, f27
+	bun	cr0, LL(9999)
 	fabs	f8,  f24
 	fabs	f9,  f25
 	fabs	f10, f26
@@ -667,7 +736,19 @@ LL(1110):
 	LFDUX	f25,   XX, INCX
 	LFDUX	f26,   XX, INCX
 	LFDUX	f27,   XX, INCX
-
+	addi	RET, RET, 1
+	fcmpu	cr0, f24, f24
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f25, f25
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f26, f26
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f27, f27
+	bun	cr0, LL(9999)
+	subi    RET, RET, 8
 	fabs	f12, f28
 	fabs	f13, f29
 	fabs	f14, f30
@@ -714,6 +795,30 @@ LL(1110):
 	.align 4
 
 LL(1120):
+	addi	RET, RET, 1
+	fcmpu	cr0, f24, f24
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f25, f25
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f26, f26
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f27, f27
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f28, f28
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f29, f29
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f30, f30
+	bun	cr0, LL(9999)
+	addi	RET, RET, 1
+	fcmpu	cr0, f31, f31
+	subi	RET, RET, 8
 	fabs	f8,  f24
 	fabs	f9,  f25
 	fabs	f10, f26
@@ -765,8 +870,11 @@ LL(1150):
 
 LL(1160):
 	LFDUX	f8,    XX, INCX
-	fabs	f8, f8
 	addi	RET, RET, 1
+	fcmpu	cr0, f8, f8
+	bun	LL(9999)
+	fabs	f8, f8
+//	addi	RET, RET, 1
 	fcmpu	cr0, f1, f8
 	beq	cr0, LL(9999)
 	bdnz	LL(1160)

kernel/power/idamax.c

@@ -327,6 +327,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
     BLASLONG max = 0;
 
     if (n <= 0 || inc_x <= 0) return (max);
+    if (n == 1) return(1);
 
     if (inc_x == 1) {
 
@@ -335,7 +336,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
 
         BLASLONG n1 = n & -32;
 	if (n1 > 0) {
-
+	    for (int ii=i;ii<i+32;ii++) if (x[ii]!=x[ii]) return(ii+1);
             max = diamax_kernel_32(n1, x, &maxf);
 
             i = n1;
@@ -343,6 +344,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
 #endif	
 #endif	
         while (i < n) {
+	    if (x[i] != x[i]) return(i+1);	
             if (ABS(x[i]) > maxf) {
                 max = i;
                 maxf = ABS(x[i]);
@@ -356,6 +358,10 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
         BLASLONG n1 = n & -4;
         while (j < n1) {
 
+	    if (x[i] != x[i]) return(i+1);	
+	    if (x[i+inc_x] != x[i+inc_x]) return(j+1);	
+	    if (x[i+2*inc_x] != x[i+2*inc_x]) return(j+2);	
+	    if (x[i+3*inc_x] != x[i+3*inc_x]) return(j+3);	
             if (ABS(x[i]) > maxf) {
                 max = j;
                 maxf = ABS(x[i]);
@@ -381,6 +387,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
 
 
         while (j < n) {
+	    if (x[i] != x[i]) return(j+1);	
             if (ABS(x[i]) > maxf) {
                 max = j;
                 maxf = ABS(x[i]);

kernel/power/isamax.c

@@ -58,6 +58,78 @@ static BLASLONG siamax_kernel_64(BLASLONG n, FLOAT *x, FLOAT *maxf) {
     register __vector float quadruple_values={0,0,0,0};
     register __vector float * v_ptrx=(__vector float *)x;
     for(; i<n; i+=64){
+	    if (vec_any_nan(v_ptrx[0])) {
+	       float d=vec_extract(v_ptrx[0],0);
+	       if (d!=d) return(i+0);
+	       d=vec_extract(v_ptrx[0],1);
+	       if (d!=d) return(i+1);
+	       d=vec_extract(v_ptrx[0],2);
+	       if (d!=d) return(i+2);
+	       return(i+3);
+	    }
+	    if (vec_any_nan(v_ptrx[1])) {
+	       float d=vec_extract(v_ptrx[1],0);
+	       if (d!=d) return(i+4+0);
+	       d=vec_extract(v_ptrx[1],1);
+	       if (d!=d) return(i+4+1);
+	       d=vec_extract(v_ptrx[1],2);
+	       if (d!=d) return(i+4+2);
+	       return(i+4+3);
+	    }
+	    if (vec_any_nan(v_ptrx[2])) {
+	       float d=vec_extract(v_ptrx[2],0);
+	       if (d!=d) return(i+8+0);
+	       d=vec_extract(v_ptrx[2],1);
+	       if (d!=d) return(i+8+1);
+	       d=vec_extract(v_ptrx[2],2);
+	       if (d!=d) return(i+8+2);
+	       return(i+8+3);
+	    }
+	    if (vec_any_nan(v_ptrx[3])) {
+	       float d=vec_extract(v_ptrx[3],0);
+	       if (d!=d) return(i+12+0);
+	       d=vec_extract(v_ptrx[3],1);
+	       if (d!=d) return(i+12+1);
+	       d=vec_extract(v_ptrx[3],2);
+	       if (d!=d) return(i+12+2);
+	       return(i+12+3);
+	    }
+	    if (vec_any_nan(v_ptrx[4])) {
+	       float d=vec_extract(v_ptrx[4],0);
+	       if (d!=d) return(i+16+0);
+	       d=vec_extract(v_ptrx[4],1);
+	       if (d!=d) return(i+16+1);
+	       d=vec_extract(v_ptrx[4],2);
+	       if (d!=d) return(i+16+2);
+	       return(i+16+3);
+	    }
+	    if (vec_any_nan(v_ptrx[5])) {
+	       float d=vec_extract(v_ptrx[5],0);
+	       if (d!=d) return(i+20+0);
+	       d=vec_extract(v_ptrx[5],1);
+	       if (d!=d) return(i+20+1);
+	       d=vec_extract(v_ptrx[5],2);
+	       if (d!=d) return(i+20+2);
+	       return(i+20+3);
+	    }
+	    if (vec_any_nan(v_ptrx[6])) {
+	       float d=vec_extract(v_ptrx[6],0);
+	       if (d!=d) return(i+24+0);
+	       d=vec_extract(v_ptrx[6],1);
+	       if (d!=d) return(i+24+1);
+	       d=vec_extract(v_ptrx[6],2);
+	       if (d!=d) return(i+24+2);
+	       return(i+24+3);
+	    }
+	    if (vec_any_nan(v_ptrx[7])) {
+	       float d=vec_extract(v_ptrx[7],0);
+	       if (d!=d) return(i+28+0);
+	       d=vec_extract(v_ptrx[7],1);
+	       if (d!=d) return(i+28+1);
+	       d=vec_extract(v_ptrx[7],2);
+	       if (d!=d) return(i+28+2);
+	       return(i+28+3);
+	    }
        //absolute temporary vectors
        register __vector float v0=vec_abs(v_ptrx[0]);
        register __vector float v1=vec_abs(v_ptrx[1]);
@@ -226,7 +298,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
     BLASLONG max = 0;
 
     if (n <= 0 || inc_x <= 0) return (max);
-
+    if (x[0] != x[0]) return(1);
     if (inc_x == 1) {
 
         BLASLONG n1 = n & -64;
@@ -238,6 +310,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
         }
 
         while (i < n) {
+	    if (x[i] != x[i]) return(i+1);
             if (ABS(x[i]) > maxf) {
                 max = i;
                 maxf = ABS(x[i]);
@@ -251,18 +324,22 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
         BLASLONG n1 = n & -4;
         while (j < n1) {
 
+	    if (x[i] != x[i]) return(j+1);
             if (ABS(x[i]) > maxf) {
                 max = j;
                 maxf = ABS(x[i]);
             }
+	    if (x[i+inc_x] != x[i+inc_x]) return(j+1);
             if (ABS(x[i + inc_x]) > maxf) {
                 max = j + 1;
                 maxf = ABS(x[i + inc_x]);
             }
+	    if (x[i+2*inc_x] != x[i+2*inc_x]) return(j+2);
             if (ABS(x[i + 2 * inc_x]) > maxf) {
                 max = j + 2;
                 maxf = ABS(x[i + 2 * inc_x]);
             }
+	    if (x[i+3*inc_x] != x[i+3*inc_x]) return(j+3);
             if (ABS(x[i + 3 * inc_x]) > maxf) {
                 max = j + 3;
                 maxf = ABS(x[i + 3 * inc_x]);
@@ -276,6 +353,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
 
 
         while (j < n) {
+	    if (x[i] != x[i]) return(j+1);
             if (ABS(x[i]) > maxf) {
                 max = j;
                 maxf = ABS(x[i]);

kernel/riscv64/iamax.c

@@ -56,13 +56,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG max=0;
 
 	if (n <= 0 || inc_x <= 0) return(max);
-
+        if (n==1) return(1);
+        if (x[0]!=x[0]) return(1);
 	maxf=ABS(x[0]);
 	ix += inc_x;
 	i++;
 
 	while(i < n)
 	{
+        	if (x[ix]!=x[ix]) return(i+1);
 		if( ABS(x[ix]) > maxf )
 		{
 			max = i;

kernel/riscv64/iamin.c

@@ -56,13 +56,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG min=0;
 
 	if (n <= 0 || inc_x <= 0) return(min);
-
+        if (n==1) return(1);      
-	minf=ABS(x[0]);
+        if (x[0]!=x[0]) return(1);
+        minf=ABS(x[0]);
 	ix += inc_x;
 	i++;
 
 	while(i < n)
 	{
+        	if (x[ix]!=x[ix]) return(i+1);
 		if( ABS(x[ix]) < ABS(minf) )
 		{
 			min = i;

kernel/x86/iamax.S

@@ -100,6 +100,8 @@
 	movl	$1, RET
 
 	FLD	(X)
+	fcomi	%st, %st
+	jp	.L999
 #ifdef USE_ABS
 	fabs
 #endif
@@ -121,6 +123,8 @@
 #endif
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -131,6 +135,8 @@
 	incl	NUM
 
 	FLD	1 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -141,6 +147,8 @@
 	incl	NUM
 
 	FLD	2 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -151,6 +159,8 @@
 	incl	NUM
 
 	FLD	3 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -161,6 +171,8 @@
 	incl	NUM
 
 	FLD	4 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -171,6 +183,8 @@
 	incl	NUM
 
 	FLD	5 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -181,6 +195,8 @@
 	incl	NUM
 
 	FLD	6 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -191,6 +207,8 @@
 	incl	NUM
 
 	FLD	7 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -215,6 +233,8 @@
 
 .L21:
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -238,6 +258,8 @@
 
 .L50:
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -249,6 +271,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -260,6 +284,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -271,6 +297,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -282,6 +310,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -293,6 +323,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -304,6 +336,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -315,6 +349,8 @@
 	incl	NUM
 
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 	addl	INCX, X
 #ifdef USE_ABS
 	fabs
@@ -338,6 +374,8 @@
 
 .L61:
 	FLD	0 * SIZE(X)
+	fucomi
+	jp	.L998
 #ifdef USE_ABS
 	fabs
 #endif
@@ -361,4 +399,7 @@
 	popl	%ebp
 	ret
 
+.L998:	mov NUM, RET
+	jmp .L999
+
 	EPILOGUE

kernel/x86_64/iamax_sse.S

@@ -93,6 +93,10 @@
 	addq	INCX, X
 	decq	M
 	shufps	$0, %xmm0, %xmm0
+	incq RET
+	comiss %xmm0, %xmm0
+	jp .L999
+	decq RET
 #ifdef USE_ABS
 	andps	%xmm15, %xmm0
 #endif
@@ -254,6 +258,10 @@
 	decq	M
 	addq	$SIZE, X
 
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	decq RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 #endif
@@ -268,6 +276,14 @@
 
 	movss	0 * SIZE(X), %xmm1
 	movss	1 * SIZE(X), %xmm2
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	decq RET
+	decq RET
 
 	subq	$2, M
 	addq	$2 * SIZE, X
@@ -332,6 +348,31 @@
 	movss	5 * SIZE(X), %xmm6
 	movss	6 * SIZE(X), %xmm7
 	movss	7 * SIZE(X), %xmm8
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	incq RET
+	comiss %xmm3, %xmm3
+	jp .L998
+	incq RET
+	comiss %xmm4, %xmm4
+	jp .L998
+	incq RET
+	comiss %xmm5, %xmm5
+	jp .L998
+	incq RET
+	comiss %xmm6, %xmm6
+	jp .L998
+	incq RET
+	comiss %xmm7, %xmm7
+	jp .L998
+	incq RET
+	comiss %xmm8, %xmm8
+	jp .L998
+	subq $8, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -378,6 +419,19 @@
 	movss	1 * SIZE(X), %xmm2
 	movss	2 * SIZE(X), %xmm3
 	movss	3 * SIZE(X), %xmm4
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	incq RET
+	comiss %xmm3, %xmm3
+	jp .L998
+	incq RET
+	comiss %xmm4, %xmm4
+	jp .L998
+	subq $4, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -405,6 +459,13 @@
 
 	movss	0 * SIZE(X), %xmm1
 	movss	1 * SIZE(X), %xmm2
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	subq $2, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -593,6 +654,31 @@
 	movss	5 * SIZE(X), %xmm6
 	movss	6 * SIZE(X), %xmm7
 	movss	7 * SIZE(X), %xmm8
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	incq RET
+	comiss %xmm3, %xmm3
+	jp .L998
+	incq RET
+	comiss %xmm4, %xmm4
+	jp .L998
+	incq RET
+	comiss %xmm5, %xmm5
+	jp .L998
+	incq RET
+	comiss %xmm6, %xmm6
+	jp .L998
+	incq RET
+	comiss %xmm7, %xmm7
+	jp .L998
+	incq RET
+	comiss %xmm8, %xmm8
+	jp .L998
+	subq $8, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -639,6 +725,19 @@
 	movss	1 * SIZE(X), %xmm2
 	movss	2 * SIZE(X), %xmm3
 	movss	3 * SIZE(X), %xmm4
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	incq RET
+	comiss %xmm3, %xmm3
+	jp .L998
+	incq RET
+	comiss %xmm4, %xmm4
+	jp .L998
+	subq $4, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -666,6 +765,13 @@
 
 	movss	0 * SIZE(X), %xmm1
 	movss	1 * SIZE(X), %xmm2
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	subq $2, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -885,6 +991,31 @@
 	movss	0 * SIZE(X), %xmm2
 	subq	INCX, X
 	movss	0 * SIZE(X), %xmm1
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	incq RET
+	comiss %xmm3, %xmm3
+	jp .L998
+	incq RET
+	comiss %xmm4, %xmm4
+	jp .L998
+	incq RET
+	comiss %xmm5, %xmm5
+	jp .L998
+	incq RET
+	comiss %xmm6, %xmm6
+	jp .L998
+	incq RET
+	comiss %xmm7, %xmm7
+	jp .L998
+	incq RET
+	comiss %xmm8, %xmm8
+	jp .L998
+	subq $8, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -932,7 +1063,19 @@
 	addq	INCX, X
 	movss	0 * SIZE(X), %xmm4
 	addq	INCX, X
-
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	incq RET
+	comiss %xmm3, %xmm3
+	jp .L998
+	incq RET
+	comiss %xmm4, %xmm4
+	jp .L998
+	subq $4, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -961,6 +1104,13 @@
 	addq	INCX, X
 	movss	0 * SIZE(X), %xmm2
 	addq	INCX, X
+	incq RET
+	comiss %xmm1, %xmm1
+	jp .L998
+	incq RET
+	comiss %xmm2, %xmm2
+	jp .L998
+	subq $2, RET
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm1
 	andps	 %xmm15, %xmm2
@@ -982,5 +1132,9 @@
 
 	ret
 
+.L998:
+//	incq RET
+	jmp .L999
+
 	EPILOGUE
 

kernel/x86_64/iamax_sse2.S

@@ -79,6 +79,8 @@
 	movsd	(X), %xmm0
 	addq	INCX, X
 	decq	M
+	comisd  %xmm0, %xmm0
+	jp      .L987
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm0
 #endif
@@ -269,6 +271,11 @@
 	je	.L21
 
 	movsd	0 * SIZE(X), %xmm1
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	decq RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 #endif
@@ -342,6 +349,32 @@
 	movsd	5 * SIZE(X), %xmm6
 	movsd	6 * SIZE(X), %xmm7
 	movsd	7 * SIZE(X), %xmm8
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	incq RET
+	comisd  %xmm3, %xmm3
+	jp      .L987
+	incq RET
+	comisd  %xmm4, %xmm4
+	jp      .L987
+	incq RET
+	comisd  %xmm5, %xmm5
+	jp      .L987
+	incq RET
+	comisd  %xmm6, %xmm6
+	jp      .L987
+	incq RET
+	comisd  %xmm7, %xmm7
+	jp      .L987
+	incq RET
+	comisd  %xmm8, %xmm8
+	jp      .L987
+	subq $8, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -374,9 +407,9 @@
 	comisd	%xmm0, %xmm6
 	je	.L999
 	incq	RET
-	comisd	%xmm0, %xmm7
+	comisd  %xmm0, %xmm7
-	je	.L999
+	je      .L999
-	incq	RET
+	incq    RET
 	jmp	.L999
 	ALIGN_3
 
@@ -388,6 +421,21 @@
 	movsd	1 * SIZE(X), %xmm2
 	movsd	2 * SIZE(X), %xmm3
 	movsd	3 * SIZE(X), %xmm4
+
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	incq RET
+	comisd  %xmm3, %xmm3
+	jp      .L987
+	incq RET
+	comisd  %xmm4, %xmm4
+	jp      .L987
+	subq $4, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -415,6 +463,14 @@
 
 	movsd	0 * SIZE(X), %xmm1
 	movsd	1 * SIZE(X), %xmm2
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+        subq $2, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -670,6 +726,32 @@
 	movsd	5 * SIZE(X), %xmm6
 	movsd	6 * SIZE(X), %xmm7
 	movsd	7 * SIZE(X), %xmm8
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	incq RET
+	comisd  %xmm3, %xmm3
+	jp      .L987
+	incq RET
+	comisd  %xmm4, %xmm4
+	jp      .L987
+	incq RET
+	comisd  %xmm5, %xmm5
+	jp      .L987
+	incq RET
+	comisd  %xmm6, %xmm6
+	jp      .L987
+	incq RET
+	comisd  %xmm7, %xmm7
+	jp      .L987
+	incq RET
+	comisd  %xmm8, %xmm8
+	jp      .L987
+	subq $8, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -716,6 +798,20 @@
 	movsd	1 * SIZE(X), %xmm2
 	movsd	2 * SIZE(X), %xmm3
 	movsd	3 * SIZE(X), %xmm4
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	incq RET
+	comisd  %xmm3, %xmm3
+	jp      .L987
+	incq RET
+	comisd  %xmm4, %xmm4
+	jp      .L987
+	subq $4, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -743,11 +839,21 @@
 
 	movsd	0 * SIZE(X), %xmm1
 	movsd	1 * SIZE(X), %xmm2
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	subq $2, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
 #endif
 	addq	$2 * SIZE, X
+	comisd	%xmm0, %xmm0
+	jp	.L987
 	incq	RET
 	comisd	%xmm0, %xmm1
 	je	.L999
@@ -962,6 +1068,7 @@
 	ALIGN_4
 
 .L92:
+
 	movsd	0 * SIZE(X), %xmm1
 	addq	INCX, X
 	movhpd	0 * SIZE(X), %xmm1
@@ -1033,6 +1140,32 @@
 	movsd	0 * SIZE(X), %xmm2
 	subq	INCX, X
 	movsd	0 * SIZE(X), %xmm1
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	incq RET
+	comisd  %xmm3, %xmm3
+	jp      .L987
+	incq RET
+	comisd  %xmm4, %xmm4
+	jp      .L987
+	incq RET
+	comisd  %xmm5, %xmm5
+	jp      .L987
+	incq RET
+	comisd  %xmm6, %xmm6
+	jp      .L987
+	incq RET
+	comisd  %xmm7, %xmm7
+	jp      .L987
+	incq RET
+	comisd  %xmm8, %xmm8
+	jp      .L987
+	subq $8, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -1083,6 +1216,20 @@
 	addq	INCX, X
 	movsd	0 * SIZE(X), %xmm4
 	addq	INCX, X
+	incq RET
+	comisd  %xmm1, %xmm1
+	jp      .L987
+	incq RET
+	comisd  %xmm2, %xmm2
+	jp      .L987
+	incq RET
+	comisd  %xmm3, %xmm3
+	jp      .L987
+	incq RET
+	comisd  %xmm4, %xmm4
+	jp      .L987
+	subq $4, RET
+
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -1111,6 +1258,14 @@
 	addq	INCX, X
 	movsd	0 * SIZE(X), %xmm2
 	addq	INCX, X
+	incq RET
+	comisd	%xmm1, %xmm1
+	jp	.L987
+	incq RET
+	comisd	%xmm2, %xmm2
+	jp	.L987
+	decq RET
+	decq RET
 #ifdef USE_ABS
 	andpd	 %xmm15, %xmm1
 	andpd	 %xmm15, %xmm2
@@ -1122,7 +1277,6 @@
 	comisd	%xmm0, %xmm2
 	je	.L999
 	ALIGN_3
-
 .L98:
 	incq	RET
 	ALIGN_3
@@ -1132,5 +1286,8 @@
 
 	ret
 
-	EPILOGUE
+.L987:
+	incq RET //count from xmm0
+	jmp .L999
 
+	EPILOGUE

utest/test_amax.c

@@ -58,3 +58,29 @@ CTEST(amax, damax){
 }
 #endif
 
+#ifdef BUILD_SINGLE
+CTEST(amax, isamax_nan){
+  blasint N=3, inc=1;
+  int te_max=0, tr_max=2;
+  float x[]={1., 0./0., 0./0. };
+  //float x[]={ 0./0., 2., 3. };
+
+
+  te_max=BLASFUNC(isamax)(&N, x, &inc);
+
+  ASSERT_EQUAL(tr_max, te_max);
+}
+#endif
+
+#ifdef BUILD_DOUBLE
+CTEST(amax, idamax_nan){
+  blasint N=4, inc=1;
+  int te_max=0, tr_max=1;
+  //float x[]={1., 0./0., 0./0. };
+  double x[]={ 0./0., 1.,2. ,3.};
+
+  te_max=BLASFUNC(idamax)(&N, x, &inc);
+
+  ASSERT_EQUAL(tr_max, te_max);
+}
+#endif