Merge pull request #4177 from martin-frbg/issue4176

Fix ZAXPY calls with INCX=0 on pre-AVX x86_64 and add utest
OpenMathLib · Aug 4, 2023 · ef23240 · ef23240
2 parents c2f4bdb + 862d06a
commit ef23240
Show file tree

Hide file tree

Showing 2 changed files with 41 additions and 2 deletions.
diff --git a/kernel/x86_64/zaxpy_sse2.S b/kernel/x86_64/zaxpy_sse2.S
@@ -1418,10 +1418,10 @@
 	movq	M,  %rax
 //If incx==0 || incy==0, avoid unloop and jump to end.
 	cmpq	$0, INCX
-	je  .L58
+	jne  	.L59
 	cmpq	$0, INCY
 	je	.L58
-
+.L59:
 	sarq	$3, %rax
 	jle	.L55
 

diff --git a/utest/test_axpy.c b/utest/test_axpy.c
@@ -74,6 +74,26 @@ CTEST(axpy,zaxpy_inc_0)
 		ASSERT_DBL_NEAR_TOL(y2[i], y1[i], DOUBLE_EPS);
 	}
 }
+
+CTEST(axpy,zaxpy_incx_0)
+{
+        blasint i;
+        blasint N=4,incX=0,incY=1;
+        double a[2]={0.25,0.5};
+        double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+        double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+        double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+        double y2[]={0.75,5.25,4.75,9.25,0.75,5.25,4.75,9.25};
+
+        //OpenBLAS
+        BLASFUNC(zaxpy)(&N,a,x1,&incX,y1,&incY);
+
+        for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x2[i], x1[i], DOUBLE_EPS);
+                ASSERT_DBL_NEAR_TOL(y2[i], y1[i], DOUBLE_EPS);
+        }
+}
+
 #endif
 
 #ifdef BUILD_SINGLE
@@ -116,5 +136,24 @@ CTEST(axpy,caxpy_inc_0)
 		ASSERT_DBL_NEAR_TOL(y2[i], y1[i], DOUBLE_EPS);
 	}
 }
+
+CTEST(axpy,caxpy_incx_0)
+{
+	blasint i;
+	blasint N=4,incX=0,incY=1;
+	float a[2]={0.25,0.5};
+	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+        double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+        double y2[]={0.75,5.25,4.75,9.25,0.75,5.25,4.75,9.25};
+
+	//OpenBLAS
+	BLASFUNC(caxpy)(&N,a,x1,&incX,y1,&incY);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x2[i], x1[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y2[i], y1[i], DOUBLE_EPS);
+	}
+}
 #endif