enabling autotuning of the CPU/GPU parameters

src/elpa_index.c

@@ -172,23 +172,23 @@ static const elpa_index_int_entry_t int_entries[] = {
         INT_ANY_ENTRY("blacs_context", "BLACS context"),
         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),
-        INT_ENTRY("gpu", "Use GPU acceleration", 0, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu", "Use GPU acceleration", 0, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, gpu_is_valid, NULL),
         //default of gpu ussage for individual phases is 1. However, it is only evaluated, if GPU is used at all, which first has to be determined
         //by the parameter gpu and presence of the device
-        INT_ENTRY("gpu_tridiag", "Use GPU acceleration for ELPA1 tridiagonalization", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_tridiag", "Use GPU acceleration for ELPA1 tridiagonalization", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa1, NULL),
-        INT_ENTRY("gpu_solve_tridi", "Use GPU acceleration for ELPA solve tridi", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_solve_tridi", "Use GPU acceleration for ELPA solve tridi", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu, NULL),
-        INT_ENTRY("gpu_trans_ev", "Use GPU acceleration for ELPA1 trans ev", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_trans_ev", "Use GPU acceleration for ELPA1 trans ev", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa1, NULL),
-        INT_ENTRY("gpu_bandred", "Use GPU acceleration for ELPA2 band reduction", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_bandred", "Use GPU acceleration for ELPA2 band reduction", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa2, NULL),
-        INT_ENTRY("gpu_tridiag_band", "Use GPU acceleration for ELPA2 tridiagonalization", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_tridiag_band", "Use GPU acceleration for ELPA2 tridiagonalization", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa2, NULL),
-        INT_ENTRY("gpu_trans_ev_tridi_to_band", "Use GPU acceleration for ELPA2 trans_ev_tridi_to_band", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_trans_ev_tridi_to_band", "Use GPU acceleration for ELPA2 trans_ev_tridi_to_band", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa2, NULL),
-        INT_ENTRY("gpu_trans_ev_band_to_full", "Use GPU acceleration for ELPA2 trans_ev_band_to_full", 1, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
+        INT_ENTRY("gpu_trans_ev_band_to_full", "Use GPU acceleration for ELPA2 trans_ev_band_to_full", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa2, NULL),
         INT_ENTRY("real_kernel", "Real kernel to use if 'solver' is set to ELPA_SOLVER_2STAGE", ELPA_2STAGE_REAL_DEFAULT, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_REAL, \
                         number_of_real_kernels, real_kernel_enumerate, real_kernel_is_valid, real_kernel_name),

test/Fortran/test_autotune.F90

@@ -206,7 +206,7 @@ program test
 
    if (myid == 0) print *, ""
 
-   tune_state => e%autotune_setup(ELPA_AUTOTUNE_FAST, AUTOTUNE_DOMAIN, error)
+   tune_state => e%autotune_setup(ELPA_AUTOTUNE_MEDIUM, AUTOTUNE_DOMAIN, error)
    assert_elpa_ok(error)
 
    iter=0