handle hip_runtime.h uniformly

common/cuda_hip/base/runtime.hpp

@@ -0,0 +1,14 @@
+// SPDX-FileCopyrightText: 2017 - 2024 The Ginkgo authors
+//
+// SPDX-License-Identifier: BSD-3-Clause
+
+#ifndef GKO_COMMON_CUDA_HIP_BASE_RUNTIME_HPP_
+#define GKO_COMMON_CUDA_HIP_BASE_RUNTIME_HPP_
+
+
+#ifdef GKO_COMPILING_HIP
+#include <hip/hip_runtime.h>
+#endif
+
+
+#endif  // GKO_COMMON_CUDA_HIP_BASE_RUNTIME_HPP_

cuda/base/batch_multi_vector_kernels.cu

@@ -13,6 +13,7 @@
 #include <ginkgo/core/base/range_accessors.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "cuda/base/batch_struct.hpp"
 #include "cuda/base/config.hpp"

cuda/base/kernel_launch.cuh

@@ -12,6 +12,7 @@
 
 
 #include "accessor/cuda_helper.hpp"
+#include "common/cuda_hip/base/runtime.hpp"
 #include "cuda/base/types.hpp"
 #include "cuda/components/thread_ids.cuh"
 

cuda/base/kernel_launch_solver.cuh

@@ -8,6 +8,9 @@
 #endif
 
 
+#include "common/cuda_hip/base/runtime.hpp"
+
+
 namespace gko {
 namespace kernels {
 namespace cuda {

cuda/components/reduction.cuh

@@ -13,6 +13,7 @@
 #include <ginkgo/core/base/executor.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/array_access.hpp"
 #include "cuda/base/config.hpp"
 #include "cuda/base/types.hpp"

cuda/factorization/factorization_kernels.cu

@@ -8,6 +8,7 @@
 #include <ginkgo/core/base/array.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/array_access.hpp"
 #include "core/components/prefix_sum_kernels.hpp"
 #include "core/matrix/csr_builder.hpp"

cuda/factorization/par_ilu_kernels.cu

@@ -9,6 +9,7 @@
 #include <ginkgo/core/matrix/coo.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "cuda/base/math.hpp"
 #include "cuda/base/types.hpp"
 #include "cuda/components/memory.cuh"

cuda/matrix/batch_csr_kernels.cu

@@ -13,6 +13,7 @@
 #include <ginkgo/core/matrix/batch_csr.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "cuda/base/batch_struct.hpp"

cuda/matrix/batch_dense_kernels.cu

@@ -13,6 +13,7 @@
 #include <ginkgo/core/matrix/batch_dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "cuda/base/batch_struct.hpp"

cuda/matrix/batch_ell_kernels.cu

@@ -13,6 +13,7 @@
 #include <ginkgo/core/matrix/batch_ell.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "cuda/base/batch_struct.hpp"

cuda/matrix/coo_kernels.cu

@@ -12,6 +12,7 @@
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/matrix/dense_kernels.hpp"
 #include "cuda/base/config.hpp"
 #include "cuda/base/cusparse_bindings.hpp"

cuda/matrix/dense_kernels.cu

@@ -17,6 +17,7 @@
 #include <ginkgo/core/matrix/sparsity_csr.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/utils.hpp"
 #include "core/components/prefix_sum_kernels.hpp"
 #include "cuda/base/config.hpp"

cuda/matrix/diagonal_kernels.cu

@@ -9,6 +9,7 @@
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "cuda/base/config.hpp"
 #include "cuda/base/types.hpp"
 #include "cuda/components/cooperative_groups.cuh"

cuda/matrix/ell_kernels.cu

@@ -17,6 +17,7 @@
 
 #include "accessor/cuda_helper.hpp"
 #include "accessor/reduced_row_major.hpp"
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/mixed_precision_types.hpp"
 #include "core/components/fill_array_kernels.hpp"
 #include "core/components/prefix_sum_kernels.hpp"

cuda/matrix/fbcsr_kernels.template.cu

@@ -24,6 +24,7 @@
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "common/unified/base/kernel_launch.hpp"
 #include "core/base/array_access.hpp"
 #include "core/base/block_sizes.hpp"

cuda/matrix/sellp_kernels.cu

@@ -12,6 +12,7 @@
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/components/prefix_sum_kernels.hpp"
 #include "cuda/base/config.hpp"
 #include "cuda/base/cusparse_bindings.hpp"

cuda/matrix/sparsity_csr_kernels.cu

@@ -13,6 +13,7 @@
 
 #include "accessor/cuda_helper.hpp"
 #include "accessor/reduced_row_major.hpp"
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/mixed_precision_types.hpp"
 #include "core/components/fill_array_kernels.hpp"
 #include "core/components/format_conversion_kernels.hpp"

cuda/preconditioner/isai_kernels.cu

@@ -10,6 +10,7 @@
 #include <ginkgo/core/matrix/csr.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/components/prefix_sum_kernels.hpp"
 #include "core/matrix/csr_builder.hpp"
 #include "cuda/base/config.hpp"

cuda/preconditioner/jacobi_kernels.cu

@@ -8,6 +8,7 @@
 #include <ginkgo/core/base/exception_helpers.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/extended_float.hpp"
 #include "core/preconditioner/jacobi_utils.hpp"
 #include "core/synthesizer/implementation_selection.hpp"

cuda/solver/batch_bicgstab_kernels.cu

@@ -13,6 +13,7 @@
 #include <ginkgo/core/base/math.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "core/solver/batch_dispatch.hpp"

cuda/solver/idr_kernels.cu

@@ -13,6 +13,7 @@
 #include <ginkgo/core/base/math.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/components/fill_array_kernels.hpp"
 #include "cuda/base/config.hpp"
 #include "cuda/base/cublas_bindings.hpp"

cuda/solver/multigrid_kernels.cu

@@ -11,6 +11,7 @@
 #include <ginkgo/core/base/types.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/array_access.hpp"
 #include "core/components/fill_array_kernels.hpp"
 #include "cuda/base/types.hpp"

hip/base/batch_multi_vector_kernels.hip.cpp

@@ -5,7 +5,6 @@
 #include "core/base/batch_multi_vector_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
 #include <thrust/functional.h>
 #include <thrust/transform.h>
 
@@ -14,6 +13,7 @@
 #include <ginkgo/core/base/range_accessors.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "hip/base/batch_struct.hip.hpp"
 #include "hip/base/config.hip.hpp"

hip/base/kernel_launch.hip.hpp

@@ -8,11 +8,11 @@
 #endif
 
 
-#include <hip/hip_runtime.h>
 #include <thrust/tuple.h>
 
 
 #include "accessor/hip_helper.hpp"
+#include "common/cuda_hip/base/runtime.hpp"
 #include "hip/base/types.hip.hpp"
 #include "hip/components/thread_ids.hip.hpp"
 

hip/base/kernel_launch_solver.hip.hpp

@@ -8,7 +8,7 @@
 #endif
 
 
-#include <hip/hip_runtime.h>
+#include "common/cuda_hip/base/runtime.hpp"
 
 
 namespace gko {

hip/components/reduction.hip.hpp

@@ -9,13 +9,11 @@
 #include <type_traits>
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/base/array.hpp>
 #include <ginkgo/core/base/executor.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/array_access.hpp"
 #include "hip/base/config.hip.hpp"
 #include "hip/base/types.hip.hpp"

hip/factorization/factorization_kernels.hip.cpp

@@ -5,12 +5,10 @@
 #include "core/factorization/factorization_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/base/array.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/array_access.hpp"
 #include "core/components/prefix_sum_kernels.hpp"
 #include "core/matrix/csr_builder.hpp"

hip/factorization/par_ilu_kernels.hip.cpp

@@ -5,13 +5,11 @@
 #include "core/factorization/par_ilu_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/base/std_extensions.hpp>
 #include <ginkgo/core/matrix/coo.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "hip/base/math.hip.hpp"
 #include "hip/base/types.hip.hpp"
 #include "hip/components/memory.hip.hpp"

hip/matrix/batch_csr_kernels.hip.cpp

@@ -5,7 +5,6 @@
 #include "core/matrix/batch_csr_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
 #include <thrust/functional.h>
 
 
@@ -14,6 +13,7 @@
 #include <ginkgo/core/matrix/batch_csr.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "hip/base/batch_struct.hip.hpp"

hip/matrix/batch_dense_kernels.hip.cpp

@@ -5,7 +5,6 @@
 #include "core/matrix/batch_dense_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
 #include <thrust/functional.h>
 
 
@@ -14,6 +13,7 @@
 #include <ginkgo/core/matrix/batch_dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "hip/base/batch_struct.hip.hpp"

hip/matrix/batch_ell_kernels.hip.cpp

@@ -5,7 +5,6 @@
 #include "core/matrix/batch_ell_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
 #include <thrust/functional.h>
 
 
@@ -14,6 +13,7 @@
 #include <ginkgo/core/matrix/batch_ell.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/batch_struct.hpp"
 #include "core/matrix/batch_struct.hpp"
 #include "hip/base/batch_struct.hip.hpp"

hip/matrix/coo_kernels.hip.cpp

@@ -5,16 +5,14 @@
 #include "core/matrix/coo_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/base/exception_helpers.hpp>
 #include <ginkgo/core/base/math.hpp>
 #include <ginkgo/core/base/types.hpp>
 #include <ginkgo/core/matrix/csr.hpp>
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/matrix/dense_kernels.hpp"
 #include "hip/base/config.hip.hpp"
 #include "hip/base/hipsparse_bindings.hip.hpp"

hip/matrix/dense_kernels.hip.cpp

@@ -5,9 +5,6 @@
 #include "core/matrix/dense_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/base/math.hpp>
 #include <ginkgo/core/base/range_accessors.hpp>
 #include <ginkgo/core/matrix/coo.hpp>
@@ -20,6 +17,7 @@
 #include <ginkgo/core/matrix/sparsity_csr.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/utils.hpp"
 #include "core/components/prefix_sum_kernels.hpp"
 #include "hip/base/config.hip.hpp"

hip/matrix/diagonal_kernels.hip.cpp

@@ -5,13 +5,11 @@
 #include "core/matrix/diagonal_kernels.hpp"
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/matrix/csr.hpp>
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "hip/base/config.hip.hpp"
 #include "hip/base/types.hip.hpp"
 #include "hip/components/cooperative_groups.hip.hpp"

hip/matrix/ell_kernels.hip.cpp

@@ -8,9 +8,6 @@
 #include <array>
 
 
-#include <hip/hip_runtime.h>
-
-
 #include <ginkgo/core/base/exception_helpers.hpp>
 #include <ginkgo/core/base/math.hpp>
 #include <ginkgo/core/base/types.hpp>
@@ -20,6 +17,7 @@
 
 #include "accessor/hip_helper.hpp"
 #include "accessor/reduced_row_major.hpp"
+#include "common/cuda_hip/base/runtime.hpp"
 #include "core/base/mixed_precision_types.hpp"
 #include "core/components/fill_array_kernels.hpp"
 #include "core/components/prefix_sum_kernels.hpp"

hip/matrix/fbcsr_kernels.template.hip.cpp

@@ -8,7 +8,6 @@
 #include <algorithm>
 
 
-#include <hip/hip_runtime.h>
 #include <thrust/copy.h>
 #include <thrust/count.h>
 #include <thrust/device_ptr.h>
@@ -25,6 +24,7 @@
 #include <ginkgo/core/matrix/dense.hpp>
 
 
+#include "common/cuda_hip/base/runtime.hpp"
 #include "common/unified/base/kernel_launch.hpp"
 #include "core/base/array_access.hpp"
 #include "core/base/block_sizes.hpp"