Reactivate GPU kernel

configure.ac

@@ -748,12 +748,12 @@ m4_define(elpa_m4_bgq_kernels, [
         complex_bgq
 ])
 
-#m4_define(elpa_m4_gpu_kernels, [
-#        real_gpu
-#        complex_gpu
-#])
+m4_define(elpa_m4_gpu_kernels, [
+        real_gpu
+        complex_gpu
+])
 
-m4_define(elpa_m4_kernel_types, [generic sparc64 neon_arch64 vsx sse sse_assembly avx avx2 avx512 bgp bgq])
+m4_define(elpa_m4_kernel_types, [generic sparc64 neon_arch64 vsx sse sse_assembly avx avx2 avx512 bgp bgq gpu])
 
 m4_define(elpa_m4_all_kernels,
           m4_foreach_w([elpa_m4_type],
@@ -794,7 +794,7 @@ ELPA_SELECT_KERNELS([sse_assembly],[enable])
 ELPA_SELECT_KERNELS([avx],[enable])
 ELPA_SELECT_KERNELS([avx2],[enable])
 ELPA_SELECT_KERNELS([avx512],[enable])
-#ELPA_SELECT_KERNELS([gpu],[disable])
+ELPA_SELECT_KERNELS([gpu],[disable])
 ELPA_SELECT_KERNELS([bgp],[disable])
 ELPA_SELECT_KERNELS([bgq],[disable])
 
@@ -836,16 +836,16 @@ m4_foreach_w([elpa_m4_kind],[real complex],[
         fi
 ])
 
-#AC_ARG_WITH(gpu-support-only, [AS_HELP_STRING([--with-gpu-support-only],
-#            [Compile and always use the GPU version])],
-#            [],[with_gpu_support_only=no])
-#if test x"$with_gpu_support_only" = x"yes" ; then
-#        m4_foreach_w([elpa_m4_kernel],elpa_m4_all_kernels,[
-#                      use_[]elpa_m4_kernel[]=no
-#        ])
-#        use_real_gpu=yes
-#        use_complex_gpu=yes
-#fi
+AC_ARG_WITH(gpu-support-only, [AS_HELP_STRING([--with-gpu-support-only],
+            [Compile and always use the GPU version])],
+            [],[with_gpu_support_only=no])
+if test x"$with_gpu_support_only" = x"yes" ; then
+        m4_foreach_w([elpa_m4_kernel],elpa_m4_all_kernels,[
+                      use_[]elpa_m4_kernel[]=no
+        ])
+        use_real_gpu=yes
+        use_complex_gpu=yes
+fi
 
 
 dnl
@@ -1307,7 +1307,7 @@ m4_foreach_w([elpa_m4_kernel],elpa_m4_all_kernels,[
 AM_CONDITIONAL([WITH_GPU_VERSION],[test x"$use_real_gpu" = x"yes" -o x"$use_complex_gpu" = x"yes"])
 if test x"$use_real_gpu" = x"yes" -o x"$use_complex_gpu" = x"yes" ; then
   AC_DEFINE([WITH_GPU_VERSION],[1],[enable GPU support])
-  #AC_DEFINE([WITH_GPU_KERNEL],[1],[GPU kernel should be build])
+  AC_DEFINE([WITH_GPU_KERNEL],[1],[GPU kernel should be build])
   ELPA_2STAGE_COMPLEX_GPU_COMPILED=1
   ELPA_2STAGE_REAL_GPU_COMPILED=1
 

src/elpa1/elpa1_tridiag_template.F90

@@ -276,9 +276,13 @@ call prmat(na, useGpu, a_mat, a_dev, matrixRows, matrixCols, nblk, my_prow, my_p
       &MATH_DATATYPE ", "tmp", istat, errorMessage)
 
       ! allocate v_row 1 element longer to allow store and broadcast tau together with it
-        allocate(uv_stored_cols(max_local_cols,2*max_stored_uv), stat=istat, errmsg=errorMessage)
-        call check_alloc("tridiag_&
-        &MATH_DATATYPE ", "uv_stored_cols", istat, errorMessage)
+      allocate(uv_stored_cols(max_local_cols,2*max_stored_uv), stat=istat, errmsg=errorMessage)
+      call check_alloc("tridiag_&
+           &MATH_DATATYPE ", "uv_stored_cols", istat, errorMessage)
+
+      allocate(vu_stored_rows(max_local_rows,2*max_stored_uv), stat=istat, errmsg=errorMessage)
+      call check_alloc("tridiag_&
+           &MATH_DATATYPE ", "vu_stored_rows", istat, errorMessage)
 
       if (useGPU) then
         num = (max_local_rows+1) * size_of_datatype
@@ -302,16 +306,11 @@ call prmat(na, useGpu, a_mat, a_dev, matrixRows, matrixCols, nblk, my_prow, my_p
         call c_f_pointer(u_row_host,u_row,(/num/))
 
         num = (max_local_rows * 2*max_stored_uv) * size_of_datatype
-        successCUDA = cuda_malloc_host(vu_stored_rows_host,num)
-        check_host_alloc_cuda("tridiag: vu_stored_rows_host", successCUDA)
-        call c_f_pointer(vu_stored_rows_host,vu_stored_rows,(/max_local_rows,2*max_stored_uv/))
+        successCUDA = cuda_host_register(int(loc(vu_stored_rows),kind=c_intptr_t),num,&
+                      cudaHostRegisterDefault)
+        check_host_register_cuda("tridiag: vu_stored_roes", successCUDA)
 
         num = (max_local_cols * 2*max_stored_uv) * size_of_datatype
-        !successCUDA = cuda_malloc_host(uv_stored_cols_host,num)
-        !check_alloc_cuda("tridiag: uv_stored_cols_host", successCUDA)
-
-        !call c_f_pointer(uv_stored_cols_host,uv_stored_cols,(/max_local_cols,2*max_stored_uv/))
-
         successCUDA = cuda_host_register(int(loc(uv_stored_cols),kind=c_intptr_t),num,&
                       cudaHostRegisterDefault)
         check_host_register_cuda("tridiag: uv_stored_cols", successCUDA)
@@ -350,10 +349,6 @@ call prmat(na, useGpu, a_mat, a_dev, matrixRows, matrixCols, nblk, my_prow, my_p
         allocate(u_row(max_local_rows), stat=istat, errmsg=errorMessage)
         call check_alloc("tridiag_&
         &MATH_DATATYPE ", "u_row", istat, errorMessage)
-
-        allocate(vu_stored_rows(max_local_rows,2*max_stored_uv), stat=istat, errmsg=errorMessage)
-        call check_alloc("tridiag_&
-        &MATH_DATATYPE ", "vu_stored_rows", istat, errorMessage)
         
       endif
 
@@ -1134,26 +1129,31 @@ call prmat(na, useGpu, a_mat, a_dev, matrixRows, matrixCols, nblk, my_prow, my_p
         check_host_dealloc_cuda("tridiag: u_row_host", successCUDA)
         nullify(u_row)
 
-        successCUDA = cuda_free_host(vu_stored_rows_host)
-        check_host_dealloc_cuda("tridiag: uv_stored_rows", successCUDA)
-        nullify(vu_stored_rows)
-
         successCUDA = cuda_host_unregister(int(loc(uv_stored_cols),kind=c_intptr_t))
         check_host_unregister_cuda("tridiag: uv_stored_cols", successCUDA)
 
+        successCUDA = cuda_host_unregister(int(loc(vu_stored_rows),kind=c_intptr_t))
+        check_host_unregister_cuda("tridiag: vu_stored_rows", successCUDA)
+
         successCUDA = cuda_host_unregister(int(loc(e_vec),kind=c_intptr_t))
         check_host_unregister_cuda("tridiag: e_vec", successCUDA)
 
         successCUDA = cuda_host_unregister(int(loc(d_vec),kind=c_intptr_t))
         check_host_unregister_cuda("tridiag: d_vec", successCUDA)
       else
-        deallocate(v_row, v_col, u_row, u_col, vu_stored_rows, uv_stored_cols, stat=istat, errmsg=errorMessage)
+        deallocate(v_row, v_col, u_row, u_col, stat=istat, errmsg=errorMessage)
         if (istat .ne. 0) then
           print *,"tridiag: error when deallocating "//errorMessage
           stop 1
         endif
       endif
 
+      deallocate(vu_stored_rows, uv_stored_cols, stat=istat, errmsg=errorMessage)
+      if (istat .ne. 0) then
+        print *,"tridiag: error when deallocating "//errorMessage
+        stop 1
+      endif
+
       call obj%timer%stop("tridiag_&
       &MATH_DATATYPE&
       &" // &

src/elpa2/elpa2_template.F90

@@ -386,7 +386,7 @@
     endif
 
     do_useGPU_bandred = do_useGPU
-    do_useGPU_tridiag_band = do_useGPU
+    do_useGPU_tridiag_band = .false.  ! not yet ported
     do_useGPU_solve_tridi = do_useGPU
     do_useGPU_trans_ev_tridi_to_band = do_useGPU
     do_useGPU_trans_ev_band_to_full = do_useGPU
@@ -403,12 +403,13 @@
       endif
       do_useGPU_bandred = (gpu == 1)
 
-      call obj%get("gpu_tridiag_band", gpu, error)
-      if (error .ne. ELPA_OK) then
-        print *,"Problem getting option for gpu_tridiag_band settings. Aborting..."
-        stop
-      endif
-      do_useGPU_tridiag_band = (gpu == 1)
+      ! not yet ported
+      !call obj%get("gpu_tridiag_band", gpu, error)
+      !if (error .ne. ELPA_OK) then
+      !  print *,"Problem getting option for gpu_tridiag_band settings. Aborting..."
+      !  stop
+      !endif
+      !do_useGPU_tridiag_band = (gpu == 1)
 
       call obj%get("gpu_solve_tridi", gpu, error)
       if (error .ne. ELPA_OK) then

src/elpa_generalized/cannon.c

@@ -86,6 +86,7 @@
 #ifdef WITH_MPI
 #include <mpi.h>
 
+#if 0
 #ifndef Add_
 #define numroc_ numroc
 #define dlacpy_ dlacpy
@@ -102,6 +103,7 @@
 #define pctranc_ pctranc
 #define pclacpy_ pclacpy
 #endif /* Add_ */
+#endif
 
 //***********************************************************************************************************
 

src/elpa_index.c

@@ -224,8 +224,9 @@ static const elpa_index_int_entry_t int_entries[] = {
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa1, NULL, PRINT_YES),
         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, PRINT_YES),
-        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, PRINT_YES),
+	//not yet ported to GPU
+        //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, PRINT_YES),
         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, PRINT_YES),
         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, \
@@ -704,9 +705,7 @@ static const char *real_kernel_name(int kernel) {
 }
 
 #define REAL_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE(kernel_number) \
-        kernel_number == ELPA_2STAGE_REAL_GPU ? 0 : 1
-// currently the GPU kernel is never valid
-// previously:       kernel_number == ELPA_2STAGE_REAL_GPU ? gpu_is_active : 1
+        kernel_number == ELPA_2STAGE_REAL_GPU ? gpu_is_active : 1
 
 static int real_kernel_is_valid(elpa_index_t index, int n, int new_value) {
         int solver = elpa_index_get_int_value(index, "solver", NULL);
@@ -745,9 +744,7 @@ static const char *complex_kernel_name(int kernel) {
 }
 
 #define COMPLEX_GPU_KERNEL_ONLY_WHEN_GPU_IS_ACTIVE(kernel_number) \
-        kernel_number == ELPA_2STAGE_COMPLEX_GPU ? 0 : 1
-// currenttly the GPU kernel is never valid
-// previously:       kernel_number == ELPA_2STAGE_COMPLEX_GPU ? gpu_is_active : 1
+        kernel_number == ELPA_2STAGE_COMPLEX_GPU ? gpu_is_active : 1
 
 static int complex_kernel_is_valid(elpa_index_t index, int n, int new_value) {
         int solver = elpa_index_get_int_value(index, "solver", NULL);

test/Fortran/test.F90

@@ -656,6 +656,14 @@ program test
 #endif
 
 #ifdef TEST_SOLVER_2STAGE
+#if TEST_GPU == 1
+#if defined TEST_REAL
+     kernel = ELPA_2STAGE_REAL_GPU
+#endif
+#if defined TEST_COMPLEX
+     kernel = ELPA_2STAGE_COMPLEX_GPU
+#endif
+#endif
      call e%set(KERNEL_KEY, kernel, error_elpa)
 #ifdef TEST_KERNEL
      assert_elpa_ok(error_elpa)