Add CI test, adjust the compiler version to check the lowest requirement, fix the corresponding error

.gitlab-ci.yml

@@ -530,12 +530,12 @@ build/cuda114/nompi/gcc/cuda/debug/shared:
     EXTRA_CMAKE_FLAGS: "-DCMAKE_CUDA_FLAGS=-diag-suppress=177"
     CUDA_ARCH: 61
 
-# HIP AMD
-build/amd/nompi/gcc/hip/debug/shared:
+# ROCm 3.5 and friends
+build/amd/nompi/gcc/rocm35/debug/shared:
   <<: *default_build_with_test
   extends:
     - .quick_test_condition
-    - .use_gko-amd-openmpi-gnu8-llvm7
+    - .use_gko-rocm35-openmpi-gnu5-llvm50
   variables:
     <<: *default_variables
     BUILD_OMP: "ON"
@@ -544,11 +544,11 @@ build/amd/nompi/gcc/hip/debug/shared:
     BUILD_TYPE: "Debug"
     FAST_TESTS: "ON"
 
-build/amd/openmpi/clang/hip/release/static:
+build/amd/openmpi/clang/rocm35/release/static:
   <<: *default_build_with_test
   extends:
-    - .quick_test_condition
-    - .use_gko-amd-openmpi-gnu8-llvm7
+    - .full_test_condition
+    - .use_gko-rocm35-openmpi-gnu5-llvm50
   variables:
     <<: *default_variables
     C_COMPILER: "clang"
@@ -560,17 +560,73 @@ build/amd/openmpi/clang/hip/release/static:
     BUILD_TYPE: "Release"
     BUILD_SHARED_LIBS: "OFF"
 
-# Build HIP AMD without omp
-build/amd/nompi/clang/hip_wo_omp/release/shared:
+# ROCm 4.5 and friends
+build/amd/mvapich2/gcc/rocm45/release/shared:
+  <<: *default_build_with_test
+  extends:
+    - .quick_test_condition
+    - .use_gko-rocm45-mvapich2-gnu8-llvm8
+  variables:
+    <<: *default_variables
+    BUILD_OMP: "ON"
+    BUILD_HIP: "ON"
+    RUN_EXAMPLES: "ON"
+    BUILD_TYPE: "Release"
+
+build/amd/mvapich2/clang/rocm45/debug/shared:
+  <<: *default_build_with_test
+  extends:
+    - .quick_test_condition
+    - .use_gko-rocm45-mvapich2-gnu8-llvm8
+  variables:
+    <<: *default_variables
+    C_COMPILER: "clang"
+    CXX_COMPILER: "clang++"
+    BUILD_OMP: "ON"
+    BUILD_HIP: "ON"
+    RUN_EXAMPLES: "ON"
+    BUILD_TYPE: "Debug"
+
+# ROCm 5.0.2 and friends
+build/amd/openmpi/gcc/rocm502/debug/static:
   <<: *default_build_with_test
   extends:
     - .full_test_condition
-    - .use_gko-amd-openmpi-gnu8-llvm7
+    - .use_gko-rocm502-openmpi-gnu11-llvm11
+  variables:
+    <<: *default_variables
+    BUILD_OMP: "ON"
+    BUILD_HIP: "ON"
+    RUN_EXAMPLES: "ON"
+    BUILD_TYPE: "Debug"
+    BUILD_SHARED_LIBS: "OFF"
+
+build/amd/openmpi/clang/rocm502/release/shared:
+  <<: *default_build_with_test
+  extends:
+    - .quick_test_condition
+    - .use_gko-rocm502-openmpi-gnu11-llvm11
   variables:
     <<: *default_variables
     C_COMPILER: "clang"
     CXX_COMPILER: "clang++"
+    BUILD_OMP: "ON"
+    BUILD_HIP: "ON"
+    RUN_EXAMPLES: "ON"
+    BUILD_TYPE: "Release"
+
+# without omp
+build/amd/nompi/gcc/rocm502_wo_omp/release/shared:
+  <<: *default_build_with_test
+  extends:
+    - .full_test_condition
+    - .use_gko-rocm502-openmpi-gnu11-llvm11
+  variables:
+    <<: *default_variables
+    BUILD_OMP: "OFF"
+    BUILD_MPI: "OFF"
     BUILD_HIP: "ON"
+    RUN_EXAMPLES: "ON"
     BUILD_TYPE: "Release"
 
 # no cuda but latest gcc and clang
@@ -597,18 +653,6 @@ build/nocuda/nompi/clang/core/release/shared:
     CXX_COMPILER: "clang++"
     BUILD_TYPE: "Release"
 
-build/nocuda/nompi/intel/core/debug/shared:
-  <<: *default_build_with_test
-  extends:
-    - .quick_test_condition
-    - .use_gko-nocuda-mvapich2-gnu9-llvm8-intel
-  variables:
-    <<: *default_variables
-    C_COMPILER: "icc"
-    CXX_COMPILER: "icpc"
-    BUILD_TYPE: "Debug"
-    FAST_TESTS: "ON"
-
 build/nocuda/nompi/gcc/omp/release/shared:
   <<: *default_build_with_test
   extends:
@@ -635,19 +679,46 @@ build/nocuda/openmpi/clang/omp/debug/static:
     FAST_TESTS: "ON"
     BUILD_SHARED_LIBS: "OFF"
 
-build/nocuda/nompi/intel/omp/release/static:
+# nocuda with the oldest supported compiler
+build/nocuda/nompi/intel/core/debug/shared:
   <<: *default_build_with_test
   extends:
     - .quick_test_condition
-    - .use_gko-nocuda-mvapich2-gnu9-llvm8-intel
+    - .use_gko-nocuda-mvapich2-gnu5-llvm39-intel2018
   variables:
     <<: *default_variables
+    # intel with old gcc without include path leads error: identifier "____m128d" is undefined
+    CXX_FLAGS: "-I /opt/intel/include/icc"
     C_COMPILER: "icc"
     CXX_COMPILER: "icpc"
+    BUILD_TYPE: "Debug"
+    FAST_TESTS: "ON"
+
+build/nocuda/nompi/gcc/omp/release/static:
+  <<: *default_build_with_test
+  extends:
+    - .quick_test_condition
+    - .use_gko-nocuda-mvapich2-gnu5-llvm39-intel2018
+  variables:
+    <<: *default_variables
     BUILD_OMP: "ON"
     BUILD_TYPE: "Release"
     BUILD_SHARED_LIBS: "OFF"
 
+build/nocuda-nomixed/nompi/clang/omp/release/static:
+  <<: *default_build_with_test
+  extends:
+    - .full_test_condition
+    - .use_gko-nocuda-mvapich2-gnu5-llvm39-intel2018
+  variables:
+    <<: *default_variables
+    C_COMPILER: "clang"
+    CXX_COMPILER: "clang++"
+    BUILD_OMP: "ON"
+    BUILD_TYPE: "Release"
+    BUILD_SHARED_LIBS: "OFF"
+    MIXED_PRECISION: "OFF"
+
 build/nocuda-nomixed/openmpi/gcc/omp/release/shared:
   <<: *default_build_with_test
   extends:
@@ -675,20 +746,6 @@ build/nocuda-nomixed/nompi/clang/omp/debug/static:
     BUILD_SHARED_LIBS: "OFF"
     MIXED_PRECISION: "OFF"
 
-build/nocuda-nomixed/nompi/intel/omp/release/static:
-  <<: *default_build_with_test
-  extends:
-    - .full_test_condition
-    - .use_gko-nocuda-mvapich2-gnu9-llvm8-intel
-  variables:
-    <<: *default_variables
-    C_COMPILER: "icc"
-    CXX_COMPILER: "icpc"
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "OFF"
-    MIXED_PRECISION: "OFF"
-
 build/dpcpp/cpu/release/static:
   <<: *default_build_with_test
   extends:
@@ -1021,7 +1078,7 @@ cudamemcheck:
 fineci-benchmark-build:
   stage: benchmark-build
   extends:
-    - .use_gko-nocuda-mvapich2-gnu9-llvm8-intel
+    - .use_gko-nocuda-mvapich2-gnu5-llvm39-intel2018
   variables:
     <<: *default_variables
     BENCHMARK_SERVER: FINECI
@@ -1099,7 +1156,7 @@ fineci-benchmark-build:
 fineci-benchmark-cuda:
   stage: benchmark-cuda
   extends:
-    - .use_gko-nocuda-mvapich2-gnu9-llvm8-intel
+    - .use_gko-nocuda-mvapich2-gnu5-llvm39-intel2018
   variables:
     <<: *default_variables
     BENCHMARK_SERVER: FINECI

.gitlab/image.yml

@@ -5,8 +5,8 @@
     - cpu
     - amdci
 
-.use_gko-nocuda-mvapich2-gnu9-llvm8-intel:
-  image: ginkgohub/cpu:mvapich2-gnu9-llvm8-intel2020
+.use_gko-nocuda-mvapich2-gnu5-llvm39-intel2018:
+  image: ginkgohub/cpu:mvapich2-gnu5-llvm39-intel2018
   tags:
     - private_ci
     - cpu
@@ -72,8 +72,22 @@
     - private_ci
     - nvidia-gpu
 
-.use_gko-amd-openmpi-gnu8-llvm7:
-  image: ginkgohub/rocm:gnu8-openmpi-llvm7
+.use_gko-rocm35-openmpi-gnu5-llvm50:
+  image: ginkgohub/rocm:35-openmpi-gnu5-llvm50
+  tags:
+    - private_ci
+    - amdci
+    - gpu
+
+.use_gko-rocm45-mvapich2-gnu8-llvm8:
+  image: ginkgohub/rocm:45-mvapich2-gnu8-llvm8
+  tags:
+    - private_ci
+    - amdci
+    - gpu
+
+.use_gko-rocm502-openmpi-gnu11-llvm11:
+  image: ginkgohub/rocm:502-openmpi-gnu11-llvm11
   tags:
     - private_ci
     - amdci

CMakeLists.txt

@@ -380,6 +380,7 @@ add_custom_target(test_install
     -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
     -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
     -DCMAKE_CUDA_COMPILER=${CMAKE_CUDA_COMPILER}
+    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
     # `--config cfg` is ignored by single-configuration generator.
     # `$<CONFIG>` is always be the same as `CMAKE_BUILD_TYPE` in
     # single-configuration generator.
@@ -400,6 +401,7 @@ add_custom_target(test_exportbuild
     -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
     -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
     -DCMAKE_CUDA_COMPILER=${CMAKE_CUDA_COMPILER}
+    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
     # `--config cfg` is ignored by single-configuration generator.
     # `$<CONFIG>` is always be the same as `CMAKE_BUILD_TYPE` in
     # single-configuration generator.

core/base/iterator_factory.hpp

@@ -84,7 +84,11 @@ class zip_iterator_reference
     template <std::size_t... idxs>
     value_type cast_impl(std::index_sequence<idxs...>) const
     {
-        return {std::get<idxs>(*this)...};
+        // gcc 5 throws error as using unintialized array
+        // std::tuple<int, char> t = { 1, '2' }; is not allowed.
+        // converting to 'std::tuple<...>' from initializer list would use
+        // explicit constructor
+        return value_type(std::get<idxs>(*this)...);
     }
 
     template <std::size_t... idxs>

hip/CMakeLists.txt

@@ -1,3 +1,7 @@
+if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.21)
+    set(CMAKE_HIP_ARCHITECTURES OFF)
+endif()
+
 if (GINKGO_HIP_PLATFORM MATCHES "${HIP_PLATFORM_NVIDIA_REGEX}"
     AND GINKGO_BUILD_CUDA AND CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 9.2)
     message(FATAL_ERROR "Ginkgo HIP backend requires CUDA >= 9.2.")
@@ -275,7 +279,12 @@ if(GINKGO_HIP_AMDGPU)
     endforeach()
 endif()
 
-set(GINKGO_HIPCC_OPTIONS ${GINKGO_HIP_COMPILER_FLAGS} "-std=c++14 -DGKO_COMPILING_HIP")
+if(GINKGO_STATIC_OR_SHARED MATCHES "STATIC")
+    # Debug Static: Hip requires PIC flags
+    set(GINKGO_HIPCC_OPTIONS ${GINKGO_HIP_COMPILER_FLAGS} "-std=c++14 -DGKO_COMPILING_HIP $<$<CONFIG:Debug>:-fPIC>")
+else()
+    set(GINKGO_HIPCC_OPTIONS ${GINKGO_HIP_COMPILER_FLAGS} "-std=c++14 -DGKO_COMPILING_HIP")
+endif()
 set(GINKGO_HIP_NVCC_OPTIONS ${GINKGO_HIP_NVCC_COMPILER_FLAGS} ${GINKGO_HIP_NVCC_ARCH} ${GINKGO_HIP_NVCC_ADDITIONAL_FLAGS})
 set(GINKGO_HIP_CLANG_OPTIONS ${GINKGO_HIP_CLANG_COMPILER_FLAGS} ${GINKGO_AMD_ARCH_FLAGS})
 

hip/matrix/fft_kernels_stub.hip.cpp

@@ -39,6 +39,14 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 namespace gko {
+
+
+std::string HipfftError::get_error(int64 error_code)
+{
+    return "Does not contain hipfft.";
+}
+
+
 namespace kernels {
 namespace hip {
 /**

test/base/device_matrix_data_kernels.cpp

@@ -30,7 +30,7 @@ THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 ******************************<GINKGO LICENSE>*******************************/
 
-#include "core/base/device_matrix_data_kernels.hpp"
+#include <ginkgo/core/base/device_matrix_data.hpp>
 
 
 #include <memory>
@@ -41,11 +41,11 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 #include <ginkgo/core/base/array.hpp>
-#include <ginkgo/core/base/device_matrix_data.hpp>
 #include <ginkgo/core/base/executor.hpp>
 #include <ginkgo/core/base/matrix_data.hpp>
 
 
+#include "core/base/device_matrix_data_kernels.hpp"
 #include "core/test/utils.hpp"
 #include "core/test/utils/assertions.hpp"
 #include "test/utils/executor.hpp"
@@ -365,7 +365,8 @@ TYPED_TEST(DeviceMatrixData, SumsDuplicates)
             max_error, std::abs(arrays.values.get_const_data()[i] -
                                 ref_arrays.values.get_const_data()[i]));
     }
-    ASSERT_LT(max_error, r<value_type>::value);
+    // when Hip with GNU < 7, it will give a little difference.
+    ASSERT_LT(max_error, 2 * r<value_type>::value);
 }
 
 

test/test_install/CMakeLists.txt

@@ -57,6 +57,8 @@ if(GINKGO_BUILD_HIP)
     if (GINKGO_BUILD_SHARED_LIBS)
         set (GINKGO_PIC_OPTION "-fPIC")
         set (GINKGO_CUDA_PIC_OPTION "-Xcompiler '-fPIC'")
+    else()
+        set (GINKGO_PIC_OPTION "$<$<CONFIG:Debug>:-fPIC>")
     endif()
     if (CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
         set(TESTINSTALL_CUDA_ARCH_FLAGS "${GINKGO_CUDA_ARCH_FLAGS}")