Fix errors introduced with HIP

src/elpa2/compute_hh_trafo.F90

@@ -185,8 +185,8 @@ last_stripe_width, kernel)
   j = -99
 
   if (wantDebug) then
-    if (useGPU .and. &
 #ifdef WITH_NVIDIA_GPU_VERSION
+    if (useGPU .and. &
 #if REALCASE == 1
       ( kernel .ne. ELPA_2STAGE_REAL_NVIDIA_GPU)) then
 #endif
@@ -194,8 +194,11 @@ last_stripe_width, kernel)
       ( kernel .ne. ELPA_2STAGE_COMPLEX_NVIDIA_GPU)) then
 #endif
       print *,"ERROR: useGPU is set in compute_hh_trafo but not a NVIDIA GPU kernel!"
+      stop
+    endif
 #endif
 #ifdef WITH_AMD_GPU_VERSION
+    if (useGPU .and. &
 #if REALCASE == 1
       ( kernel .ne. ELPA_2STAGE_REAL_AMD_GPU)) then
 #endif
@@ -203,9 +206,9 @@ last_stripe_width, kernel)
       ( kernel .ne. ELPA_2STAGE_COMPLEX_AMD_GPU)) then
 #endif
       print *,"ERROR: useGPU is set in compute_hh_trafo but not a AMD GPU kernel!"
-#endif
       stop
     endif
+#endif
   endif
 
 #if REALCASE == 1

src/elpa2/elpa2_template.F90

@@ -207,12 +207,15 @@
 
 #if REALCASE == 1
 #undef GPU_KERNEL
+#define GPU_KERNEL ELPA_2STAGE_REAL_NVIDIA_GPU
 #undef GENERIC_KERNEL
 #undef KERNEL_STRING
 #ifdef WITH_NVIDIA_GPU_VERSION
+#undef GPU_KERNEL
 #define GPU_KERNEL ELPA_2STAGE_REAL_NVIDIA_GPU
 #endif
 #ifdef WITH_AMD_GPU_VERSION
+#undef GPU_KERNEL
 #define GPU_KERNEL ELPA_2STAGE_REAL_AMD_GPU
 #endif
 #define GENERIC_KERNEL ELPA_2STAGE_REAL_GENERIC
@@ -221,12 +224,15 @@
 
 #if COMPLEXCASE == 1
 #undef GPU_KERNEL
+#define GPU_KERNEL ELPA_2STAGE_COMPLEX_NVIDIA_GPU
 #undef GENERIC_KERNEL
 #undef KERNEL_STRING
 #ifdef WITH_NVIDIA_GPU_VERSION
-#define GPU_KERNEL ELPA_2STAGE_NVIDIA_COMPLEX_GPU
+#undef GPU_KERNEL
+#define GPU_KERNEL ELPA_2STAGE_COMPLEX_NVIDIA_GPU
 #endif
 #ifdef WITH_AMD_GPU_VERSION
+#undef GPU_KERNEL
 #define GPU_KERNEL ELPA_2STAGE_COMPLEX_AMD_GPU
 #endif
 #define GENERIC_KERNEL ELPA_2STAGE_COMPLEX_GENERIC

src/elpa_index.c

@@ -776,9 +776,12 @@ static int real_kernel_is_valid(elpa_index_t index, int n, int new_value) {
         switch(new_value) {
 #ifdef WITH_NVIDIA_GPU_VERSION
                 ELPA_FOR_ALL_2STAGE_REAL_KERNELS(VALID_CASE_3, REAL_NVIDIA_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
-#endif
+#else
 #ifdef WITH_AMD_GPU_VERSION
                 ELPA_FOR_ALL_2STAGE_REAL_KERNELS(VALID_CASE_3, REAL_AMD_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
+#else
+                ELPA_FOR_ALL_2STAGE_REAL_KERNELS(VALID_CASE_3, REAL_NVIDIA_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
+#endif
 #endif
                 default:
                         return 0;
@@ -822,9 +825,12 @@ static int complex_kernel_is_valid(elpa_index_t index, int n, int new_value) {
         switch(new_value) {
 #ifdef WITH_NVIDIA_GPU_VERISION
                 ELPA_FOR_ALL_2STAGE_COMPLEX_KERNELS(VALID_CASE_3, COMPLEX_NVIDIA_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
-#endif
+#else
 #ifdef WITH_AMD_GPU_VERISION
                 ELPA_FOR_ALL_2STAGE_COMPLEX_KERNELS(VALID_CASE_3, COMPLEX_AMD_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
+#else
+                ELPA_FOR_ALL_2STAGE_COMPLEX_KERNELS(VALID_CASE_3, COMPLEX_NVIDIA_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
+#endif
 #endif
                 default:
                         return 0;

test/Fortran/test.F90

@@ -199,7 +199,7 @@ program test
 #ifdef WITH_OPENMP_TRADITIONAL
    TEST_INT_TYPE      :: max_threads, threads_caller
 #endif
-#ifdef TEST_GPU_SET_ID
+#if  TEST_GPU_SET_ID == 1
    TEST_INT_TYPE      :: gpuID
 #endif
 #ifdef SPLIT_COMM_MYSELF
@@ -643,7 +643,7 @@ program test
    call e%set("nvidia-gpu", TEST_GPU, error_elpa)
    assert_elpa_ok(error_elpa)
 
-#ifdef TEST_GPU_SET_ID
+#if TEST_GPU_SET_ID == 1
    ! simple test
    ! Can (and should) fail often
    gpuID = mod(myid,2)