Fix implementation of iamax_sse.S as reported in OpenMathLib#2116.

The was a typo in iamax_sse.S where one of the comparison
was cmpeqps instead of cmpeqss. That misdetected index
for sequences where the minimum value was 0.

kernel/x86_64/KERNEL

@@ -171,7 +171,7 @@ IXAMAXKERNEL = izamax.S
 endif
 
 ifndef ISAMINKERNEL
-ISAMINKERNEL = iamax.S
+ISAMINKERNEL = iamax_sse.S
 endif
 
 ifndef IDAMINKERNEL
@@ -207,7 +207,7 @@ IQMAXKERNEL = iamax.S
 endif
 
 ifndef ISMINKERNEL
-ISMINKERNEL = iamax.S
+ISMINKERNEL = iamax_sse.S
 endif
 
 ifndef IDMINKERNEL

kernel/x86_64/iamax_sse.S

@@ -36,10 +36,6 @@
 /* or implied, of The University of Texas at Austin.                 */
 /*********************************************************************/
 
-/* This kernel was found to give wrong results when used for ISMIN/ISAMIN
-   with increment != 1, although it appears to be correct for corresponding
-   MAX operations. See issue 2116 */
-
 #define ASSEMBLER
 #include "common.h"
 
@@ -863,7 +859,7 @@
 #ifdef USE_ABS
 	andps	 %xmm15, %xmm5
 #endif
-	cmpeqps	%xmm0, %xmm5
+	cmpeqss	%xmm0, %xmm5
 
 	movss	0 * SIZE(X), %xmm6
 	addq	INCX, X

utest/CMakeLists.txt

@@ -7,6 +7,7 @@ else ()
   set(OpenBLAS_utest_src
     utest_main.c
     test_amax.c
+    test_asmin.c
     test_rotmg.c
     test_rot.c
     test_axpy.c

utest/Makefile

@@ -11,7 +11,7 @@ UTESTBIN=openblas_utest
 
 include $(TOPDIR)/Makefile.system
 
-OBJS=utest_main.o test_amax.o test_rotmg.o test_axpy.o test_dotu.o test_dsdot.o test_swap.o test_rot.o
+OBJS=utest_main.o test_amax.o test_ismin.c test_rotmg.o test_axpy.o test_dotu.o test_dsdot.o test_swap.o test_rot.o
 #test_rot.o test_swap.o test_axpy.o test_dotu.o test_dsdot.o test_fork.o
 
 ifneq ($(NO_LAPACK), 1)