New keywords for GPU in ELPA index

- Rename keyword "gpu" -> "nvidia-gpu"
- Add keyword "amd-gpu"

Makefile.am

@@ -139,7 +139,7 @@ libelpa@SUFFIX@_private_la_SOURCES += \
 endif
 
 if WITH_NVIDIA_GPU_VERSION
-  libelpa@SUFFIX@_private_la_SOURCES += src/GPU/CUDA/elpa_index_gpu.cu src/GPU/CUDA/cudaFunctions.cu src/GPU/CUDA/cuUtils.cu src/elpa2/GPU/CUDA/ev_tridi_band_gpu_real.cu src/elpa2/GPU/CUDA/ev_tridi_band_gpu_complex.cu
+  libelpa@SUFFIX@_private_la_SOURCES += src/GPU/CUDA/elpa_index_nvidia_gpu.cu src/GPU/CUDA/cudaFunctions.cu src/GPU/CUDA/cuUtils.cu src/elpa2/GPU/CUDA/ev_tridi_band_gpu_real.cu src/elpa2/GPU/CUDA/ev_tridi_band_gpu_complex.cu
 endif
 
 if WITH_AMD_GPU_VERSION

src/GPU/CUDA/elpa_index_gpu.cu → src/GPU/CUDA/elpa_index_nvidia_gpu.cu

 extern "C" {
-  int gpu_count() {
+  int nvidia_gpu_count() {
     int count;
     cudaError_t cuerr = cudaGetDeviceCount(&count);
     if (cuerr != cudaSuccess) {

src/elpa1/elpa1_template.F90

@@ -294,7 +294,7 @@ function elpa_solve_evp_&
      obj%eigenvalues_only = .true.
    endif
 
-   call obj%get("gpu",gpu,error)
+   call obj%get("nvidia-gpu",gpu,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option for gpu. Aborting..."
      stop

src/elpa1/elpa_multiply_a_b.F90

@@ -109,7 +109,7 @@
   success = .true.
 
   ! GPU settings
-  call obj%get("gpu", gpu,error)
+  call obj%get("nvidia-gpu", gpu,error)
   if (error .ne. ELPA_OK) then
     print *,"Problem getting option for gpu. Aborting..."
     stop

src/elpa2/elpa2_template.F90

@@ -365,7 +365,7 @@
     wantDebug = debug == 1
 
     ! GPU settings
-    call obj%get("gpu", gpu,error)
+    call obj%get("nvidia-gpu", gpu,error)
     if (error .ne. ELPA_OK) then
       print *,"Problem getting option gpu settings. Aborting..."
       stop

src/elpa_index.c

@@ -113,7 +113,10 @@ static int min_tile_size_enumerate(elpa_index_t index, int i);
 static int min_tile_size_is_valid(elpa_index_t index, int n, int new_value);
 
 #ifdef WITH_NVIDIA_GPU_VERSION
-int gpu_count();
+int nvidia_gpu_count();
+#endif
+#ifdef WITH_AMD_GPU_VERSION
+int amd_gpu_count();
 #endif
 
 static int use_gpu_id_cardinality(elpa_index_t index);
@@ -223,7 +226,9 @@ static const elpa_index_int_entry_t int_entries[] = {
                          number_of_matrix_layouts, matrix_layout_enumerate, matrix_layout_is_valid, elpa_matrix_layout_name, PRINT_YES), \
         INT_ENTRY("solver", "Solver to use", ELPA_SOLVER_1STAGE, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_ANY, \
                         number_of_solvers, solver_enumerate, solver_is_valid, elpa_solver_name, PRINT_YES),
-        INT_ENTRY("gpu", "Use GPU acceleration", 0, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY, \
+        INT_ENTRY("nvidia-gpu", "Use Nvidia GPU acceleration", 0, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY, \
+                        cardinality_bool, enumerate_identity, gpu_is_valid, NULL, PRINT_YES),
+        INT_ENTRY("amd-gpu", "Use AMD GPU acceleration", 0, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY, \
                         cardinality_bool, enumerate_identity, gpu_is_valid, NULL, PRINT_YES),
         INT_ENTRY("is_skewsymmetric", "Matrix is skewsymmetic", 0, ELPA_AUTOTUNE_NOT_TUNABLE, 0,
                         cardinality_bool, enumerate_identity, skewsymmetric_is_valid, NULL, PRINT_YES),
@@ -765,7 +770,7 @@ static int real_kernel_is_valid(elpa_index_t index, int n, int new_value) {
         if (solver == ELPA_SOLVER_1STAGE) {
                 return new_value == ELPA_2STAGE_REAL_DEFAULT;
         }
-        int gpu_is_active = elpa_index_get_int_value(index, "gpu", NULL);
+        int gpu_is_active = (elpa_index_get_int_value(index, "nvidia-gpu", NULL) || elpa_index_get_int_value(index, "amd-gpu", NULL));
         switch(new_value) {
                 ELPA_FOR_ALL_2STAGE_REAL_KERNELS(VALID_CASE_3, REAL_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
                 default:
@@ -804,7 +809,7 @@ static int complex_kernel_is_valid(elpa_index_t index, int n, int new_value) {
         if (solver == ELPA_SOLVER_1STAGE) {
                 return new_value == ELPA_2STAGE_COMPLEX_DEFAULT;
         }
-        int gpu_is_active = elpa_index_get_int_value(index, "gpu", NULL);
+        int gpu_is_active = (elpa_index_get_int_value(index, "nvidia-gpu", NULL) || elpa_index_get_int_value(index, "amd-gpu", NULL));
         switch(new_value) {
                 ELPA_FOR_ALL_2STAGE_COMPLEX_KERNELS(VALID_CASE_3, COMPLEX_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE)
                 default:
@@ -1038,7 +1043,7 @@ static int omp_threads_is_valid(elpa_index_t index, int n, int new_value) {
 
 
 static int valid_with_gpu(elpa_index_t index, int n, int new_value) {
-        int gpu_is_active = elpa_index_get_int_value(index, "gpu", NULL);
+        int gpu_is_active = (elpa_index_get_int_value(index, "nvidia-gpu", NULL) || elpa_index_get_int_value(index, "amd-gpu", NULL));
         if (gpu_is_active == 1) {
                 return ((new_value == 0 ) || (new_value == 1));
         }
@@ -1049,7 +1054,7 @@ static int valid_with_gpu(elpa_index_t index, int n, int new_value) {
 
 static int valid_with_gpu_elpa1(elpa_index_t index, int n, int new_value) {
         int solver = elpa_index_get_int_value(index, "solver", NULL);
-        int gpu_is_active = elpa_index_get_int_value(index, "gpu", NULL);
+        int gpu_is_active = (elpa_index_get_int_value(index, "nvidia-gpu", NULL) || elpa_index_get_int_value(index, "amd-gpu", NULL));
         if ((solver == ELPA_SOLVER_1STAGE) && (gpu_is_active == 1)) {
                 return ((new_value == 0 ) || (new_value == 1));
         }
@@ -1060,7 +1065,7 @@ static int valid_with_gpu_elpa1(elpa_index_t index, int n, int new_value) {
 
 static int valid_with_gpu_elpa2(elpa_index_t index, int n, int new_value) {
         int solver = elpa_index_get_int_value(index, "solver", NULL);
-        int gpu_is_active = elpa_index_get_int_value(index, "gpu", NULL);
+        int gpu_is_active = (elpa_index_get_int_value(index, "nvidia-gpu", NULL) || elpa_index_get_int_value(index, "amd-gpu", NULL));
         if ((solver == ELPA_SOLVER_2STAGE) && (gpu_is_active == 1)) {
                 return ((new_value == 0 ) || (new_value == 1));
         }
@@ -1106,12 +1111,14 @@ static int max_stored_rows_is_valid(elpa_index_t index, int n, int new_value) {
 static int use_gpu_id_cardinality(elpa_index_t index) {
 #ifdef WITH_NVIDIA_GPU_VERSION
 	int count;
-	count = gpu_count();
+	count = nvidia_gpu_count();
         if (count == -1000) {
           fprintf(stderr, "Querrying GPUs failed! Set GPU count = 0\n");
 	return 0;
         }
 	return count;
+#elif WITH_AMD_GPU_VERION
+	return 0;
 #else
 	return 0;
 #endif
@@ -1125,13 +1132,15 @@ static int use_gpu_id_enumerate(elpa_index_t index, int i) {
 static int use_gpu_id_is_valid(elpa_index_t index, int n, int new_value) {
 #ifdef WITH_NVIDIA_GPU_VERSION
 	int count;
-	count = gpu_count();
+	count = nvidia_gpu_count();
         if (count == -1000) {
           fprintf(stderr, "Querrying GPUs failed! Return with error\n");
 	  return 0 == 1 ;
 	} else {
           return (0 <= new_value) && (new_value <= count);
 	}
+#elif WITH_AMD_GPU_VERION
+	return 0 == 1;
 #else
 	return 0 == 0;
 #endif

test/C/test.c

@@ -273,7 +273,7 @@ int main(int argc, char** argv) {
 #endif
    assert_elpa_ok(error_elpa);
 
-   elpa_set(handle, "gpu", TEST_GPU, &error_elpa);
+   elpa_set(handle, "nvidia-gpu", TEST_GPU, &error_elpa);
    assert_elpa_ok(error_elpa);
 
 #if defined(TEST_SOLVE_2STAGE) && defined(TEST_KERNEL)

test/C/test_autotune.c

@@ -226,7 +226,7 @@ int main(int argc, char** argv) {
    /* Setup */
    assert_elpa_ok(elpa_setup(handle));
 
-   elpa_set(handle, "gpu", 0, &error_elpa);
+   elpa_set(handle, "nvidia-gpu", 0, &error_elpa);
    assert_elpa_ok(error_elpa);
  
    autotune_handle = elpa_autotune_setup(handle, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_REAL, &error_elpa);

test/C/test_multiple_objs.c

@@ -223,7 +223,7 @@ int main(int argc, char** argv) {
    /* Setup */
    assert_elpa_ok(elpa_setup(elpa_handle_1));
 
-   elpa_set(elpa_handle_1, "gpu", 0, &error_elpa);
+   elpa_set(elpa_handle_1, "nvidia-gpu", 0, &error_elpa);
    assert_elpa_ok(error_elpa);
 
    elpa_set(elpa_handle_1, "timings", 1, &error_elpa);
@@ -253,7 +253,7 @@ int main(int argc, char** argv) {
 
    elpa_load_settings(elpa_handle_2, "initial_parameters.txt", &error_elpa);
 
-   elpa_get(elpa_handle_2, "gpu", &gpu, &error_elpa);
+   elpa_get(elpa_handle_2, "nvidia-gpu", &gpu, &error_elpa);
    assert_elpa_ok(error_elpa);
 
    elpa_get(elpa_handle_2, "timings", &timings, &error_elpa);

test/Fortran/test.F90

@@ -640,7 +640,7 @@ program test
 #endif
    assert_elpa_ok(error_elpa)
 
-   call e%set("gpu", TEST_GPU, error_elpa)
+   call e%set("nvidia-gpu", TEST_GPU, error_elpa)
    assert_elpa_ok(error_elpa)
 
 #ifdef TEST_GPU_SET_ID

test/Fortran/test_autotune.F90

@@ -228,7 +228,7 @@ program test
 
    call e%set("debug",1, error_elpa)
    assert_elpa_ok(error_elpa)
-   call e%set("gpu", 0, error_elpa)
+   call e%set("nvidia-gpu", 0, error_elpa)
    assert_elpa_ok(error_elpa)
    !call e%set("max_stored_rows", 15, error_elpa)
 

test/Fortran/test_multiple_objs.F90

@@ -209,7 +209,7 @@ program test
    call e1%set("debug",1, error_elpa)
    assert_elpa_ok(error_elpa)
 
-   call e1%set("gpu", 0, error_elpa)
+   call e1%set("nvidia-gpu", 0, error_elpa)
    assert_elpa_ok(error_elpa)
    !call e1%set("max_stored_rows", 15, error_elpa)
 
@@ -238,7 +238,7 @@ program test
    assert_elpa_ok(error_elpa)
    call e2%get("debug", int(debug,kind=c_int), error_elpa)
    assert_elpa_ok(error_elpa)
-   call e2%get("gpu", int(gpu,kind=c_int), error_elpa)
+   call e2%get("nvidia-gpu", int(gpu,kind=c_int), error_elpa)
    assert_elpa_ok(error_elpa)
 
    if ((timings .ne. 1) .or. (debug .ne. 1) .or. (gpu .ne. 0)) then

test/Fortran/test_skewsymmetric.F90

@@ -233,7 +233,7 @@ program test
    call e_complex%set("timings",1, error_elpa)
 
    call e_complex%set("debug",1,error_elpa)
-   call e_complex%set("gpu", 0,error_elpa)
+   call e_complex%set("nvidia-gpu", 0,error_elpa)
    call e_complex%set("omp_threads", 8, error_elpa)
 
    assert_elpa_ok(e_complex%setup())
@@ -271,7 +271,7 @@ program test
    call e_skewsymmetric%set("timings",1, error_elpa)
 
    call e_skewsymmetric%set("debug",1,error_elpa)
-   call e_skewsymmetric%set("gpu", 0,error_elpa)
+   call e_skewsymmetric%set("nvidia-gpu", 0,error_elpa)
    call e_skewsymmetric%set("omp_threads",8, error_elpa)
 
    assert_elpa_ok(e_skewsymmetric%setup())

test/Fortran/test_split_comm.F90

@@ -264,7 +264,7 @@ program test
    call e%set("timings",1, error_elpa)
 
    call e%set("debug",1, error_elpa)
-   call e%set("gpu", 0, error_elpa)
+   call e%set("nvidia-gpu", 0, error_elpa)
    !call e%set("max_stored_rows", 15, error_elpa)
 
    assert_elpa_ok(e%setup())