fix complex abs( inf + inf*i ) in CUDA and HIP

icl-utk-edu · Jul 7, 2024 · b740a66 · b740a66
1 parent 20d8dd2
commit b740a66
Show file tree

Hide file tree

Showing 3 changed files with 7 additions and 7 deletions.
diff --git a/src/cuda/device_util.cuh b/src/cuda/device_util.cuh
@@ -173,7 +173,7 @@ inline float abs(cuFloatComplex x)
         b = fabsf(b);
         w = max(a, b);
         z = min(a, b);
-        if (z == 0) {
+        if (z == 0 || isinf( w )) {
             t = w;
         }
         else {
@@ -194,7 +194,7 @@ inline double abs(cuDoubleComplex x)
     // CUDA has a good implementation.
     return cuCabs(x);
 #else
-    // For HIP, use our implementation that scales per LAPACK.
+    // For HIP, use our implementation that scales per LAPACK lapy2.
     double a = real( x );
     double b = imag( x );
     double z, w, t;
@@ -209,7 +209,7 @@ inline double abs(cuDoubleComplex x)
         b = fabs(b);
         w = max(a, b);
         z = min(a, b);
-        if (z == 0) {
+        if (z == 0 || isinf( w )) {
             t = w;
         }
         else {

diff --git a/src/hip/device_util.hip.hh b/src/hip/device_util.hip.hh
@@ -173,7 +173,7 @@ inline float abs(rocblas_float_complex x)
         b = fabsf(b);
         w = max(a, b);
         z = min(a, b);
-        if (z == 0) {
+        if (z == 0 || isinf( w )) {
             t = w;
         }
         else {
@@ -194,7 +194,7 @@ inline double abs(rocblas_double_complex x)
     // CUDA has a good implementation.
     return hipCabs(x);
 #else
-    // For HIP, use our implementation that scales per LAPACK.
+    // For HIP, use our implementation that scales per LAPACK lapy2.
     double a = real( x );
     double b = imag( x );
     double z, w, t;
@@ -209,7 +209,7 @@ inline double abs(rocblas_double_complex x)
         b = fabs(b);
         w = max(a, b);
         z = min(a, b);
-        if (z == 0) {
+        if (z == 0 || isinf( w )) {
             t = w;
         }
         else {

diff --git a/src/hip/device_util.hip.hh.dep b/src/hip/device_util.hip.hh.dep
@@ -1 +1 @@
-a403f5b1c0195b503252fa643c722495  src/cuda/device_util.cuh
+3a2509f9538de94f6212c4e678b236a6  src/cuda/device_util.cuh
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		a403f5b1c0195b503252fa643c722495 src/cuda/device_util.cuh
		3a2509f9538de94f6212c4e678b236a6 src/cuda/device_util.cuh