diff --git a/azure-pipelines.yml b/azure-pipelines.yml
index 2c1fadd8a..d72baabe1 100644
--- a/azure-pipelines.yml
+++ b/azure-pipelines.yml
@@ -136,7 +136,7 @@ jobs:
       cmake -G "Ninja" -DCMAKE_C_COMPILER=cl -DCMAKE_Fortran_COMPILER=flang-new -DC_LAPACK=1 -DCMAKE_MT=mt -DCMAKE_BUILD_TYPE=Release -DMSVC_STATIC_CRT=ON ..
       cmake --build . --config Release
       ctest
-
+      ctest --rerun-failed --output-on-failure
 
 
 - job: OSX_OpenMP
diff --git a/kernel/arm/scal.c b/kernel/arm/scal.c
index 4ef49e293..4455d7643 100644
--- a/kernel/arm/scal.c
+++ b/kernel/arm/scal.c
@@ -48,7 +48,11 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 	{
 
 		if ( da == 0.0 )
+			if (!isnan(x[i]) && !isinf(x[i])) {
 			x[i]=0.0;
+			} else {
+			x[i]=NAN;
+			}
 		else
 			x[i] = da * x[i] ;
 
diff --git a/kernel/power/sbgemm_ncopy_16_power10.c b/kernel/power/sbgemm_ncopy_16_power10.c
index c6b633011..595edfdad 100644
--- a/kernel/power/sbgemm_ncopy_16_power10.c
+++ b/kernel/power/sbgemm_ncopy_16_power10.c
@@ -37,8 +37,11 @@
 /*********************************************************************/
 
 #include <stdio.h>
+#include <altivec.h>
 #include "common.h"
 
+typedef uint32_t vec_bf16x2 __attribute__ ((vector_size (16)));
+
 int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
   BLASLONG i, j;
 
@@ -82,7 +85,84 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
       aoffset16 = aoffset15 + lda;
       aoffset += 16 * lda;
 
-      i = (m >> 1);
+      i = (m >> 3);
+      if (i > 0) {
+        do {
+          vec_bf16x2 vtemp01 = *(vec_bf16x2 *)(aoffset1);
+          vec_bf16x2 vtemp02 = *(vec_bf16x2 *)(aoffset2);
+          vec_bf16x2 vtemp03 = *(vec_bf16x2 *)(aoffset3);
+          vec_bf16x2 vtemp04 = *(vec_bf16x2 *)(aoffset4);
+          vec_bf16x2 vtemp05 = *(vec_bf16x2 *)(aoffset5);
+          vec_bf16x2 vtemp06 = *(vec_bf16x2 *)(aoffset6);
+          vec_bf16x2 vtemp07 = *(vec_bf16x2 *)(aoffset7);
+          vec_bf16x2 vtemp08 = *(vec_bf16x2 *)(aoffset8);
+          vec_bf16x2 vtemp09 = *(vec_bf16x2 *)(aoffset9);
+          vec_bf16x2 vtemp10 = *(vec_bf16x2 *)(aoffset10);
+          vec_bf16x2 vtemp11 = *(vec_bf16x2 *)(aoffset11);
+          vec_bf16x2 vtemp12 = *(vec_bf16x2 *)(aoffset12);
+          vec_bf16x2 vtemp13 = *(vec_bf16x2 *)(aoffset13);
+          vec_bf16x2 vtemp14 = *(vec_bf16x2 *)(aoffset14);
+          vec_bf16x2 vtemp15 = *(vec_bf16x2 *)(aoffset15);
+          vec_bf16x2 vtemp16 = *(vec_bf16x2 *)(aoffset16);
+
+          vec_bf16x2 vtemp17 = vec_mergeh(vtemp01, vtemp03);
+          vec_bf16x2 vtemp18 = vec_mergel(vtemp01, vtemp03);
+          vec_bf16x2 vtemp19 = vec_mergeh(vtemp02, vtemp04);
+          vec_bf16x2 vtemp20 = vec_mergel(vtemp02, vtemp04);
+          vec_bf16x2 vtemp21 = vec_mergeh(vtemp05, vtemp07);
+          vec_bf16x2 vtemp22 = vec_mergel(vtemp05, vtemp07);
+          vec_bf16x2 vtemp23 = vec_mergeh(vtemp06, vtemp08);
+          vec_bf16x2 vtemp24 = vec_mergel(vtemp06, vtemp08);
+          vec_bf16x2 vtemp25 = vec_mergeh(vtemp09, vtemp11);
+          vec_bf16x2 vtemp26 = vec_mergel(vtemp09, vtemp11);
+          vec_bf16x2 vtemp27 = vec_mergeh(vtemp10, vtemp12);
+          vec_bf16x2 vtemp28 = vec_mergel(vtemp10, vtemp12);
+          vec_bf16x2 vtemp29 = vec_mergeh(vtemp13, vtemp15);
+          vec_bf16x2 vtemp30 = vec_mergel(vtemp13, vtemp15);
+          vec_bf16x2 vtemp31 = vec_mergeh(vtemp14, vtemp16);
+          vec_bf16x2 vtemp32 = vec_mergel(vtemp14, vtemp16);
+
+          *(vec_bf16x2 *)(boffset +   0) = vec_mergeh(vtemp17, vtemp19);
+          *(vec_bf16x2 *)(boffset +   8) = vec_mergeh(vtemp21, vtemp23);
+          *(vec_bf16x2 *)(boffset +  16) = vec_mergeh(vtemp25, vtemp27);
+          *(vec_bf16x2 *)(boffset +  24) = vec_mergeh(vtemp29, vtemp31);
+          *(vec_bf16x2 *)(boffset +  32) = vec_mergel(vtemp17, vtemp19);
+          *(vec_bf16x2 *)(boffset +  40) = vec_mergel(vtemp21, vtemp23);
+          *(vec_bf16x2 *)(boffset +  48) = vec_mergel(vtemp25, vtemp27);
+          *(vec_bf16x2 *)(boffset +  56) = vec_mergel(vtemp29, vtemp31);
+          *(vec_bf16x2 *)(boffset +  64) = vec_mergeh(vtemp18, vtemp20);
+          *(vec_bf16x2 *)(boffset +  72) = vec_mergeh(vtemp22, vtemp24);
+          *(vec_bf16x2 *)(boffset +  80) = vec_mergeh(vtemp26, vtemp28);
+          *(vec_bf16x2 *)(boffset +  88) = vec_mergeh(vtemp30, vtemp32);
+          *(vec_bf16x2 *)(boffset +  96) = vec_mergel(vtemp18, vtemp20);
+          *(vec_bf16x2 *)(boffset + 104) = vec_mergel(vtemp22, vtemp24);
+          *(vec_bf16x2 *)(boffset + 112) = vec_mergel(vtemp26, vtemp28);
+          *(vec_bf16x2 *)(boffset + 120) = vec_mergel(vtemp30, vtemp32);
+
+          aoffset1  += 8;
+          aoffset2  += 8;
+          aoffset3  += 8;
+          aoffset4  += 8;
+          aoffset5  += 8;
+          aoffset6  += 8;
+          aoffset7  += 8;
+          aoffset8  += 8;
+          aoffset9  += 8;
+          aoffset10 += 8;
+          aoffset11 += 8;
+          aoffset12 += 8;
+          aoffset13 += 8;
+          aoffset14 += 8;
+          aoffset15 += 8;
+          aoffset16 += 8;
+
+          boffset   += 128;
+
+          i--;
+        } while (i > 0);
+      }
+
+      i = (m & 7) >> 1;
       if (i > 0){
 	do{
 	  ctemp01 = *(aoffset1 +  0);
diff --git a/test/compare_sgemm_sbgemm.c b/test/compare_sgemm_sbgemm.c
index bc74233ab..de589458b 100644
--- a/test/compare_sgemm_sbgemm.c
+++ b/test/compare_sgemm_sbgemm.c
@@ -86,14 +86,26 @@ main (int argc, char *argv[])
 {
   blasint m, n, k;
   int i, j, l;
-  blasint x;
+  blasint x, y;
   int ret = 0;
   int loop = 100;
   char transA = 'N', transB = 'N';
   float alpha = 1.0, beta = 0.0;
 
   for (x = 0; x <= loop; x++)
+  {
+    for (y = 0; y < 4; y++)
     {
+      if ((y == 0) || (y == 2)) {
+        transA = 'N';
+      } else {
+        transA = 'T';
+      }
+      if ((y == 0) || (y == 1)) {
+        transB = 'N';
+      } else {
+        transB = 'T';
+      }
       m = k = n = x;
       float A[m * k];
       float B[k * n];
@@ -104,43 +116,55 @@ main (int argc, char *argv[])
       blasint one=1;
 
       for (j = 0; j < m; j++)
-	{
-	  for (i = 0; i < m; i++)
-	    {
-	      A[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
-	      B[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
-	      C[j * k + i] = 0;
-	      sbstobf16_(&one, &A[j*k+i], &one, &atmp, &one);
-	      sbstobf16_(&one, &B[j*k+i], &one, &btmp, &one);
-	      AA[j * k + i].v = atmp;
-	      BB[j * k + i].v = btmp;
-	      CC[j * k + i] = 0;
-	      DD[j * k + i] = 0;
-	    }
-	}
+      {
+        for (i = 0; i < m; i++)
+        {
+          A[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
+          B[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
+          C[j * k + i] = 0;
+          sbstobf16_(&one, &A[j*k+i], &one, &atmp, &one);
+          sbstobf16_(&one, &B[j*k+i], &one, &btmp, &one);
+          AA[j * k + i].v = atmp;
+          BB[j * k + i].v = btmp;
+          CC[j * k + i] = 0;
+          DD[j * k + i] = 0;
+        }
+      }
       SGEMM (&transA, &transB, &m, &n, &k, &alpha, A,
-	     &m, B, &k, &beta, C, &m);
+        &m, B, &k, &beta, C, &m);
       SBGEMM (&transA, &transB, &m, &n, &k, &alpha, (bfloat16*) AA,
-	      &m, (bfloat16*)BB, &k, &beta, CC, &m);
+        &m, (bfloat16*)BB, &k, &beta, CC, &m);
       for (i = 0; i < n; i++)
-	for (j = 0; j < m; j++)
-	    if (fabs (CC[i * m + j] - C[i * m + j]) > 1.0)
-	      ret++;
-      if (transA == 'N' && transB == 'N')
-	{
-	  for (i = 0; i < n; i++)
-	    for (j = 0; j < m; j++)
-	      for (l = 0; l < k; l++)
-		{
-		  DD[i * m + j] +=
-		    float16to32 (AA[l * m + j]) * float16to32 (BB[l + k * i]);
-		}
-	  for (i = 0; i < n; i++)
-	    for (j = 0; j < m; j++)
-		if (CC[i * m + j] != DD[i * m + j])
-		  ret++;
-	}
+        for (j = 0; j < m; j++)
+          if (fabs (CC[i * m + j] - C[i * m + j]) > 1.0)
+            ret++;
+      for (i = 0; i < n; i++)
+        for (j = 0; j < m; j++)
+          for (l = 0; l < k; l++)
+            if (transA == 'N' && transB == 'N')
+            {
+              DD[i * m + j] +=
+                float16to32 (AA[l * m + j]) * float16to32 (BB[l + k * i]);
+            } else if (transA == 'T' && transB == 'N')
+            {
+              DD[i * m + j] +=
+                float16to32 (AA[k * j + l]) * float16to32 (BB[l + k * i]);
+            } else if (transA == 'N' && transB == 'T')
+            {
+              DD[i * m + j] +=
+                float16to32 (AA[l * m + j]) * float16to32 (BB[i + l * n]);
+            } else if (transA == 'T' && transB == 'T')
+            {
+              DD[i * m + j] +=
+                float16to32 (AA[k * j + l]) * float16to32 (BB[i + l * n]);
+            }
+      for (i = 0; i < n; i++)
+        for (j = 0; j < m; j++)
+          if (CC[i * m + j] != DD[i * m + j])
+            ret++;
     }
+  }
+
   if (ret != 0)
     fprintf (stderr, "FATAL ERROR SBGEMM - Return code: %d\n", ret);
   return ret;