Forgot to set rocBlas handle

With this commit ELPA does run correctly on AMD GPUs.
Functionality and correctness tests have been carried out
on AMD MI100

configure.ac

@@ -1532,22 +1532,22 @@ if test x"${use_nvidia_gpu}" = x"yes" ; then
   use_nvidia_complex_gpu=yes
 fi
 
-AC_MSG_CHECKING(whether INTEL GPU version should be used)
-AC_ARG_ENABLE([intel-gpu],
-                AS_HELP_STRING([--enable-intel-gpu],
-                               [do use INTEL GPU version]),
-                               [if test x"$enableval" = x"yes"; then
-                                  use_intel_gpu=yes
-                                else
-                                  use_intel_gpu=no
-                                fi],
-              [use_intel_gpu=no])
-AC_MSG_RESULT([${use_intel_gpu}])
-if test x"${use_intel_gpu}" = x"yes" ; then
-  need_intel_gpu=yes
-  use_real_intel_gpu=yes
-  use_complex_intel_gpu=yes
-fi
+#AC_MSG_CHECKING(whether INTEL GPU version should be used)
+#AC_ARG_ENABLE([intel-gpu],
+#                AS_HELP_STRING([--enable-intel-gpu],
+#                               [do use INTEL GPU version]),
+#                               [if test x"$enableval" = x"yes"; then
+#                                  use_intel_gpu=yes
+#                                else
+#                                  use_intel_gpu=no
+#                                fi],
+#              [use_intel_gpu=no])
+#AC_MSG_RESULT([${use_intel_gpu}])
+#if test x"${use_intel_gpu}" = x"yes" ; then
+#  need_intel_gpu=yes
+#  use_real_intel_gpu=yes
+#  use_complex_intel_gpu=yes
+#fi
 
 AC_MSG_CHECKING(whether AMD-GPU version should be used)
 AC_ARG_ENABLE([AMD-gpu],
@@ -2094,7 +2094,7 @@ AC_CONFIG_FILES([
 m4_include([m4/ax_fc_check_define.m4])
 AC_MSG_CHECKING([if workaround for broken preprocessor is needed])
 
-need_manual_cpp=yes
+need_manual_cpp=no
 AX_FC_CHECK_DEFINE([__INTEL_COMPILER],[is_intel=yes],[])
 AX_FC_CHECK_DEFINE([__PGI],[is_pgi=yes],[])
 ACTUAL_FC="$FC"

src/GPU/check_for_gpu.F90

@@ -146,9 +146,10 @@ module mod_check_for_gpu
 #endif
           stop 1
         endif
+
       else
 
-#ifdef WITH_NVIDIA_GPU_VERSION
+#if defined(WITH_NVIDIA_GPU_VERSION) || !defined(WITH_AMD_GPU_VERSION)
         if (cublasHandle .ne. -1) then
 #endif
 #ifdef WITH_AMD_GPU_VERSION

src/elpa1/elpa1_trans_ev_template.F90

@@ -453,11 +453,11 @@ subroutine trans_ev_&
             call obj%timer%stop("mkl_offload")
 
           else
-            call obj%timer%start("cublas")
+            call obj%timer%start("gpublas")
             call gpublas_PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N',   &
                                      nstor, l_cols, l_rows, ONE, hvm_dev, hvm_ubnd,  &
                                      q_dev, ldq, ZERO, tmp_dev, nstor)
-            call obj%timer%stop("cublas")
+            call obj%timer%stop("gpublas")
           endif
         else ! useGPU
 
@@ -559,7 +559,7 @@ subroutine trans_ev_&
             call obj%timer%stop("mkl_offload")
 #endif /* WITH_MPI */
           else
-            call obj%timer%start("cublas")
+            call obj%timer%start("gpublas")
             call gpublas_PRECISION_TRMM('L', 'L', 'N', 'N',     &
                                      nstor, l_cols, ONE, tmat_dev, max_stored_rows,  &
                                      tmp_dev, nstor)
@@ -567,7 +567,7 @@ subroutine trans_ev_&
             call gpublas_PRECISION_GEMM('N', 'N' ,l_rows ,l_cols ,nstor,  &
                                      -ONE, hvm_dev, hvm_ubnd, tmp_dev, nstor,   &
                                      ONE, q_dev, ldq)
-            call obj%timer%stop("cublas")
+            call obj%timer%stop("gpublas")
           endif
         else !useGPU
 #ifdef WITH_MPI

src/elpa1/elpa1_tridiag_template.F90

@@ -782,7 +782,7 @@ subroutine tridiag_&
                 ! Unlike for CPU, we (for each MPI thread) do just one large mat-vec multiplication
                 ! this requires altering of the algorithm when later explicitly updating the matrix
                 ! after max_stored_uv is reached : we need to update all tiles, not only those above diagonal
-                if (wantDebug) call obj%timer%start("cublas")
+                if (wantDebug) call obj%timer%start("gpublas")
                 call gpublas_PRECISION_GEMV(BLAS_TRANS_OR_CONJ, l_rows,l_cols,  &
                                           ONE, a_dev, matrixRows,                   &
                                           v_row_dev , 1,                          &
@@ -797,7 +797,7 @@ subroutine tridiag_&
 !                                             ONE, u_row_dev + (l_row_beg - 1) *                 &
 !                                             size_of_datatype, 1)
 !                 endif
-                if (wantDebug) call obj%timer%stop("cublas")
+                if (wantDebug) call obj%timer%stop("gpublas")
               endif
             else  ! mat_vec_as_one_block
               !perform multiplication by stripes - it is faster than by blocks, since we call cublas with
@@ -1096,7 +1096,7 @@ subroutine tridiag_&
                else
                  ! if using mat-vec multiply by stripes, it is enough to update tiles above (or on) the diagonal only
                  ! we than use the same calls as for CPU version
-                 if (wantDebug) call obj%timer%start("cublas")
+                 if (wantDebug) call obj%timer%start("gpublas")
                  call gpublas_PRECISION_GEMM('N', BLAS_TRANS_OR_CONJ,     &
                                          l_row_end-l_row_beg+1, l_col_end-l_col_beg+1, 2*n_stored_vecs,                      &
                                          ONE, vu_stored_rows_dev + (l_row_beg - 1) *                                         &
@@ -1105,7 +1105,7 @@ subroutine tridiag_&
                                          size_of_datatype,  &
                                          max_local_cols, ONE, a_dev + ((l_row_beg - 1) + (l_col_beg - 1) * matrixRows) *     &
                                          size_of_datatype , matrixRows)
-                 if (wantDebug) call obj%timer%stop("cublas")
+                 if (wantDebug) call obj%timer%stop("gpublas")
                endif
              endif
            else !useGPU
@@ -1147,12 +1147,12 @@ subroutine tridiag_&
              else
                !update whole (remaining) part of matrix, including tiles below diagonal
                !we can do that in one large cublas call
-               if (wantDebug) call obj%timer%start("cublas")
+               if (wantDebug) call obj%timer%start("gpublas")
                call gpublas_PRECISION_GEMM('N', BLAS_TRANS_OR_CONJ, l_rows, l_cols, 2*n_stored_vecs,   &
                                          ONE, vu_stored_rows_dev, max_local_rows, &
                                          uv_stored_cols_dev, max_local_cols,  &
                                          ONE, a_dev, matrixRows)
-               if (wantDebug) call obj%timer%stop("cublas")
+               if (wantDebug) call obj%timer%stop("gpublas")
              endif
            endif
          endif

src/elpa1/elpa_multiply_a_b.F90

@@ -360,11 +360,11 @@
               aux_off = (lrs-1)*size_of_datatype
               b_off = ((lcs-1)*ldb+lrs-1)*size_of_datatype
 
-              call obj%timer%start("cublas")
+              call obj%timer%start("gpublas")
               call gpublas_PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N', nstor, lce-lcs+1, &
                    lre-lrs+1, ONE, aux_dev+aux_off, l_rows, b_dev+b_off, ldb, ZERO, &
                    tmp1_dev, nstor)
-              call obj%timer%stop("cublas")
+              call obj%timer%stop("gpublas")
 
               num = nstor*(lce-lcs+1)*size_of_datatype
               successGPU = gpu_memcpy(int(loc(tmp1),kind=c_intptr_t), &

src/elpa2/GPU/ROCm/interface_c_hip_kernel.F90

@@ -330,8 +330,8 @@ module hip_c_kernel
       integer(kind=c_intptr_t) :: a_dev
       integer(kind=c_intptr_t) :: row_group_dev
 #ifdef WITH_AMD_GPU_VERSION
-      call launch_my_pack_c_hip_kernel_real_double(row_count, n_offset, max_idx,stripe_width, a_dim2, stripe_count, l_nev, a_dev, &
-           row_group_dev)
+      call launch_my_pack_c_hip_kernel_real_double(row_count, n_offset, max_idx,stripe_width, a_dim2, stripe_count, l_nev, &
+              a_dev, row_group_dev)
 #endif
     end subroutine
 
@@ -344,8 +344,8 @@ module hip_c_kernel
       integer(kind=c_intptr_t) :: a_dev
       integer(kind=c_intptr_t) :: row_group_dev
 #ifdef WITH_AMD_GPU_VERSION
-      call launch_my_pack_c_hip_kernel_real_single(row_count, n_offset, max_idx,stripe_width, a_dim2, stripe_count, l_nev, a_dev, &
-           row_group_dev)
+      call launch_my_pack_c_hip_kernel_real_single(row_count, n_offset, max_idx,stripe_width, a_dim2, stripe_count, l_nev, &
+              a_dev, row_group_dev)
 #endif
     end subroutine
 #endif
@@ -426,8 +426,8 @@ module hip_c_kernel
       integer(kind=c_intptr_t) :: a_dev
       integer(kind=c_intptr_t) :: row_group_dev
 #ifdef WITH_AMD_GPU_VERSION
-      call launch_my_pack_c_hip_kernel_complex_single(row_count, n_offset, max_idx,stripe_width,a_dim2, stripe_count, l_nev, a_dev, &
-           row_group_dev)
+      call launch_my_pack_c_hip_kernel_complex_single(row_count, n_offset, max_idx,stripe_width,a_dim2, stripe_count, l_nev, &
+              a_dev, row_group_dev)
 #endif
     end subroutine
 #endif

src/elpa2/elpa2_bandred_template.F90

@@ -1187,7 +1187,7 @@ max_threads)
               call obj%timer%stop("mkl_offload")
 
             else
-              call obj%timer%start("cublas")
+              call obj%timer%start("gpublas")
               call gpublas_PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N',                   &
                                          lce-lcs+1, n_cols, lre,     &
                                          ONE, (a_dev + ((lcs-1)*lda* &
@@ -1197,10 +1197,10 @@ max_threads)
                                              size_of_datatype),      &
                                          cur_l_cols)
 
-              call obj%timer%stop("cublas")
+              call obj%timer%stop("gpublas")
 
               if(i==0) cycle
-              call obj%timer%start("cublas")
+              call obj%timer%start("gpublas")
 
               lre = min(l_rows,i*l_rows_tile)
               if (isSkewsymmetric) then
@@ -1222,7 +1222,7 @@ max_threads)
                                             size_of_datatype),              &
                                               cur_l_rows)
               endif
-              call obj%timer%stop("cublas")
+              call obj%timer%stop("gpublas")
             endif
           else ! useGPU
 
@@ -1419,13 +1419,13 @@ max_threads)
                       nbw*nbw*size_of_datatype,gpuMemcpyHostToDevice)
         check_memcpy_gpu("bandred: tmat -> tmat_dev ", successGPU)
 
-        call obj%timer%start("cublas")
+        call obj%timer%start("gpublas")
         call gpublas_PRECISION_TRMM('Right', 'Upper', BLAS_TRANS_OR_CONJ, 'Nonunit',  &
                             l_cols, n_cols, ONE, tmat_dev, nbw, umc_dev, cur_l_cols)
-        call obj%timer%stop("cublas")
+        call obj%timer%stop("gpublas")
 
         ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T
-        call obj%timer%start("cublas")
+        call obj%timer%start("gpublas")
         call gpublas_PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N',             &
                                  n_cols, n_cols, l_cols, ONE, umc_dev, cur_l_cols, &
                                  (umc_dev+(cur_l_cols * n_cols )*size_of_datatype),cur_l_cols, &
@@ -1433,7 +1433,7 @@ max_threads)
 
         call gpublas_PRECISION_TRMM('Right', 'Upper', BLAS_TRANS_OR_CONJ, 'Nonunit',    &
            n_cols, n_cols, ONE, tmat_dev, nbw, vav_dev, nbw)
-        call obj%timer%stop("cublas")
+        call obj%timer%stop("gpublas")
 
         successGPU = gpu_memcpy(int(loc(vav),kind=c_intptr_t), &
                     vav_dev, nbw*nbw*size_of_datatype, gpuMemcpyDeviceToHost)
@@ -1538,7 +1538,7 @@ max_threads)
       endif
 
       else
-        call obj%timer%start("cublas")
+        call obj%timer%start("gpublas")
         if (isSkewsymmetric) then
           call gpublas_PRECISION_GEMM('N', 'N', l_cols, n_cols, n_cols,&
 #if REALCASE == 1
@@ -1564,7 +1564,7 @@ max_threads)
                                    cur_l_cols, vav_dev,nbw,        &
                                    ONE, umc_dev, cur_l_cols)
         endif
-        call obj%timer%stop("cublas")
+        call obj%timer%stop("gpublas")
 
         successGPU = gpu_memcpy(int(loc(umcGPU(1)),kind=c_intptr_t), &
                     umc_dev, umc_size*size_of_datatype, gpuMemcpyDeviceToHost)
@@ -1689,7 +1689,7 @@ max_threads)
             print *,"this should never happen"
             stop
           endif
-          call obj%timer%start("cublas")
+          call obj%timer%start("gpublas")
 
           call gpublas_PRECISION_GEMM('N', BLAS_TRANS_OR_CONJ, myend-mystart+1,    &
                                    lce-lcs+1, 2*n_cols, -ONE, &
@@ -1697,7 +1697,7 @@ max_threads)
                                    size_of_datatype), &
                                    cur_l_cols, ONE, (a_dev+(lcs-1)*lda* &
                                    size_of_datatype), lda)
-          call obj%timer%stop("cublas")
+          call obj%timer%stop("gpublas")
         endif
       else
         call obj%timer%start("blas")
@@ -1731,7 +1731,7 @@ max_threads)
           call obj%timer%stop("mkl_offload")
 
         else
-          call obj%timer%start("cublas")
+          call obj%timer%start("gpublas")
 
           call gpublas_PRECISION_GEMM('N', BLAS_TRANS_OR_CONJ,     &
                                      lre, lce-lcs+1, 2*n_cols, -ONE, &
@@ -1739,7 +1739,7 @@ max_threads)
                                      size_of_datatype), &
                                      cur_l_cols, ONE, (a_dev+(lcs-1)*lda* &
                                      size_of_datatype), lda)
-          call obj%timer%stop("cublas")
+          call obj%timer%stop("gpublas")
         endif
       else ! useGPU
 

src/elpa2/elpa2_trans_ev_band_to_full_template.F90

@@ -491,11 +491,11 @@ subroutine trans_ev_band_to_full_&
                           max_local_rows*cwy_blocking*size_of_datatype, gpuMemcpyHostToDevice)
           check_memcpy_gpu("trans_ev_band_to_full: hvm -> hvm_dev", successGPU)
 
-          call obj%timer%start("cublas")
+          call obj%timer%start("gpublas")
           call gpublas_PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N', &
                                        n_cols, l_cols, l_rows, ONE, hvm_dev, max_local_rows, &
                                        q_dev, ldq , ZERO, tmp_dev, n_cols)
-          call obj%timer%stop("cublas")
+          call obj%timer%stop("gpublas")
 
 #ifdef WITH_MPI
           ! copy data from device to host for a later MPI_ALLREDUCE
@@ -544,12 +544,12 @@ subroutine trans_ev_band_to_full_&
                         cwy_blocking*cwy_blocking*size_of_datatype, gpuMemcpyHostToDevice)
           check_memcpy_gpu("trans_ev_band_to_full: tmat_complete -> tmat_dev", successGPU)
 
-          call obj%timer%start("cublas")
+          call obj%timer%start("gpublas")
           call gpublas_PRECISION_TRMM('L', 'U', BLAS_TRANS_OR_CONJ, 'N', &
                                    n_cols, l_cols, ONE, tmat_dev, cwy_blocking, tmp_dev, n_cols)
           call gpublas_PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm_dev, max_local_rows, tmp_dev, &
                                      n_cols, ONE, q_dev, ldq)
-          call obj%timer%stop("cublas")
+          call obj%timer%stop("gpublas")
         endif
       else
         call obj%timer%start("blas")
@@ -593,13 +593,13 @@ subroutine trans_ev_band_to_full_&
                         cwy_blocking*cwy_blocking*size_of_datatype, gpuMemcpyHostToDevice)
           check_memcpy_gpu("trans_ev_band_to_full: tmat_complete -> tmat_dev", successGPU)
 
-          call obj%timer%start("cublas")
+          call obj%timer%start("gpublas")
           call gpublas_PRECISION_TRMM('L', 'U', BLAS_TRANS_OR_CONJ, 'N', &
                                      n_cols, l_cols, ONE, tmat_dev, cwy_blocking, &
                                      tmp_dev, n_cols)
           call gpublas_PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, &
                                       -ONE, hvm_dev, max_local_rows, tmp_dev, n_cols, ONE, q_dev, ldq)
-          call obj%timer%stop("cublas")
+          call obj%timer%stop("gpublas")
         endif
       else
         call obj%timer%start("blas")

src/solve_tridi/merge_systems_template.F90

@@ -836,12 +836,12 @@
 #endif
                   call obj%timer%stop("mkl_offload")
                 else
-                  call obj%timer%start("cublas")
+                  call obj%timer%start("gpublas")
                   call gpublas_PRECISION_GEMM('N', 'N', l_rnm, ncnt, nnzu,   &
                                       1.0_rk, qtmp1_dev, ubound(qtmp1,dim=1),    &
                                       ev_dev, ubound(ev,dim=1), &
                                       1.0_rk, qtmp2_dev, ubound(qtmp2,dim=1))
-                  call obj%timer%stop("cublas")
+                  call obj%timer%stop("gpublas")
                 endif
               else
                 call obj%timer%start("blas")
@@ -898,12 +898,12 @@
                   call obj%timer%stop("mkl_offload")
 
                 else
-                  call obj%timer%start("cublas")
+                  call obj%timer%start("gpublas")
                   call gpublas_PRECISION_GEMM('N', 'N', l_rows-l_rnm, ncnt, nnzl,   &
                                       1.0_rk, qtmp1_dev + l_rnm * size_of_datatype, ubound(qtmp1,dim=1),    &
                                       ev_dev, ubound(ev,dim=1), &
                                       1.0_rk, qtmp2_dev + l_rnm * size_of_datatype, ubound(qtmp2,dim=1))
-                  call obj%timer%stop("cublas")
+                  call obj%timer%stop("gpublas")
                 endif
               else
                 call obj%timer%start("blas")