Skip to content

GitLab

Commit 2d2bd148 authored by Andreas Marek's avatar Andreas Marek
Browse files
Options

Better error messages

parent 173beacd
Hide whitespace changes
Inline Side-by-side
Showing with 35 additions and 32 deletions
src/elpa1_tridiag_template.X90
View file @ 2d2bd148
......@@ -314,23 +314,23 @@
if (useGPU) then
successCUDA = cuda_malloc(v_row_dev, max_local_rows * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: v_row_dev", successCUDA)
successCUDA = cuda_malloc(u_row_dev, max_local_rows * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: u_row_dev", successCUDA)
successCUDA = cuda_malloc(v_col_dev, max_local_cols * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: v_col_dev", successCUDA)
successCUDA = cuda_malloc(u_col_dev, max_local_cols * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: u_col_dev", successCUDA)
successCUDA = cuda_malloc(vu_stored_rows_dev, max_local_rows * 2 * max_stored_uv * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: vu_stored_rows_dev", successCUDA)
successCUDA = cuda_malloc(uv_stored_cols_dev, max_local_cols * 2 * max_stored_uv * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: vu_stored_rows_dev", successCUDA)
endif !useGPU
......@@ -349,10 +349,10 @@
! allocate memmory for matrix A on the device and than copy the matrix
successCUDA = cuda_malloc(a_dev, lda * matrixCols * size_of_datatype)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: a_dev", successCUDA)
successCUDA = cuda_memcpy(a_dev, loc(a_mat(1,1)), lda * matrixCols * size_of_datatype, cudaMemcpyHostToDevice)
check_alloc_cuda("tridiag", successCUDA)
check_alloc_cuda("tridiag: a_dev", successCUDA)
endif
! main cycle of tridiagonalization
......@@ -378,7 +378,7 @@
! we use v_row on the host at the moment! successCUDA = cuda_memcpy(v_row_dev, a_dev + a_offset, (l_rows)*size_of_PRECISION_real, cudaMemcpyDeviceToDevice)
successCUDA = cuda_memcpy(loc(v_row(1)), a_dev + a_offset, (l_rows)* size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag a_dev 1", successCUDA)
else
v_row(1:l_rows) = a_mat(1:l_rows,l_cols+1)
endif
......@@ -487,17 +487,17 @@
if (l_rows > 0 .and. l_cols> 0 ) then
if(useGPU) then
successCUDA = cuda_memset(u_col_dev, 0, l_cols * size_of_datatype)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: u_col_dev", successCUDA)
successCUDA = cuda_memset(u_row_dev, 0, l_rows * size_of_datatype)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: u_row_dev", successCUDA)
successCUDA = cuda_memcpy(v_col_dev, loc(v_col(1)), l_cols * size_of_datatype, cudaMemcpyHostToDevice)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: v_col_dev", successCUDA)
successCUDA = cuda_memcpy(v_row_dev, loc(v_row(1)), l_rows * size_of_datatype, cudaMemcpyHostToDevice)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: v_row_dev", successCUDA)
endif ! useGU
#if REALCASE == 1
......@@ -634,10 +634,10 @@
if (useGPU) then
successCUDA = cuda_memcpy(loc(u_col(1)), u_col_dev, l_cols * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: u_col_dev 1", successCUDA)
successCUDA = cuda_memcpy(loc(u_row(1)), u_row_dev, l_rows * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: u_row_dev 1", successCUDA)
endif
! call PRECISION_SYMV('U', l_cols, &
......@@ -786,12 +786,12 @@
successCUDA = cuda_memcpy(vu_stored_rows_dev, loc(vu_stored_rows(1,1)), &
max_local_rows * 2 * max_stored_uv * &
size_of_datatype, cudaMemcpyHostToDevice)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: vu_stored_rows_dev", successCUDA)
successCUDA = cuda_memcpy(uv_stored_cols_dev, loc(uv_stored_cols(1,1)), &
max_local_cols * 2 * max_stored_uv * &
size_of_datatype, cudaMemcpyHostToDevice)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: uv_stored_cols_dev", successCUDA)
endif
do i = 0, (istep-2)/tile_size
......@@ -845,7 +845,7 @@
successCUDA = cuda_memcpy(loc(a_mat(l_rows, l_cols)), a_dev + a_offset, &
1 * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: a_dev 3", successCUDA)
endif
if (n_stored_vecs > 0) then
......@@ -856,9 +856,12 @@
if (useGPU) then
!a_dev(l_rows,l_cols) = a_mat(l_rows,l_cols)
successCUDA = cuda_memcpy(a_dev + a_offset, loc(a_mat(l_rows, l_cols)), &
1 * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
!successCUDA = cuda_threadsynchronize()
!check_memcpy_cuda("tridiag: a_dev 4a5a", successCUDA)
successCUDA = cuda_memcpy(a_dev + a_offset, int(loc(a_mat(l_rows, l_cols)),kind=c_size_t), &
int(1 * size_of_datatype, kind=c_size_t), cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag: a_dev 4", successCUDA)
endif
endif
......@@ -873,7 +876,7 @@
if(useGPU) then
successCUDA = cuda_memcpy(loc(aux3(1)), a_dev + (lda * (l_cols - 1)) * size_of_datatype, &
1 * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: a_dev 5", successCUDA)
vrl = aux3(1)
else !useGPU
vrl = a_mat(1,l_cols)
......@@ -908,7 +911,7 @@
if(useGPU) then
successCUDA = cuda_memcpy(loc(aux3(1)), a_dev, &
1 * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: a_dev 6", successCUDA)
d_vec(1) = PRECISION_REAL(aux3(1))
else !useGPU
d_vec(1) = PRECISION_REAL(a_mat(1,1))
......@@ -924,7 +927,7 @@
if(useGPU) then
successCUDA = cuda_memcpy(loc(e_vec(1)), a_dev + (lda * (l_cols - 1)) * size_of_datatype, &
1 * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: a_dev 7", successCUDA)
else !useGPU
e_vec(1) = a_mat(1,l_cols) ! use last l_cols value of loop above
endif !useGPU
......@@ -934,7 +937,7 @@
if (my_prow==prow(1, nblk, np_rows) .and. my_pcol==pcol(1, nblk, np_cols)) then
if(useGPU) then
successCUDA = cuda_memcpy(loc(d_vec(1)), a_dev, 1 * size_of_datatype, cudaMemcpyDeviceToHost)
check_memcpy_cuda("tridiag", successCUDA)
check_memcpy_cuda("tridiag: a_dev 8", successCUDA)
else !useGPU
d_vec(1) = a_mat(1,1)
endif !useGPU
......@@ -955,25 +958,25 @@
if (useGPU) then
! todo: should we leave a_mat on the device for further use?
successCUDA = cuda_free(a_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag: a_dev 9", successCUDA)
successCUDA = cuda_free(v_row_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag: v_row_dev", successCUDA)
successCUDA = cuda_free(u_row_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag: (u_row_dev", successCUDA)
successCUDA = cuda_free(v_col_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag: v_col_dev", successCUDA)
successCUDA = cuda_free(u_col_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag: u_col_dev ", successCUDA)
successCUDA = cuda_free(vu_stored_rows_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag: vu_stored_rows_dev ", successCUDA)
successCUDA = cuda_free(uv_stored_cols_dev)
check_dealloc_cuda("tridiag", successCUDA)
check_dealloc_cuda("tridiag:uv_stored_cols_dev ", successCUDA)
endif
! distribute the arrays d_vec and e_vec to all processors
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment