Merge #2003 Bump ROCm support 6.2.0+

This PR bumps ROCm support 6.2.0+

Related PR: #2003

Author: yhmtsai

.gitlab-ci.yml

@@ -226,88 +226,23 @@ build/cuda131/nompi/gcc/cuda/release/shared:
     BUILD_TYPE: "Release"
     MODULE_LOAD: "cmake/3.30.8 cuda/13.1.1 gcc/14.3.0"
 
-# ROCm 4.5 and friends
-build/amd/nompi/gcc/rocm45/release/shared:
+build/amd/openmpi/gcc/rocm620/debug/shared:
   extends:
-    - .build_and_test_template
+    - .build_and_test_tum_template
     - .default_variables
     - .quick_test_condition
-    - .use_gko-rocm45-nompi-gnu8-llvm8
+    - .use_tum-amd
   variables:
-    BUILD_OMP: "ON"
     BUILD_HIP: "ON"
-    NONDEFAULT_STREAM: "ON"
-    BUILD_TYPE: "Release"
     BUILD_HWLOC: "OFF"
-
-build/amd/nompi/clang/rocm45/release/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-rocm45-nompi-gnu8-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    BUILD_HIP: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "OFF"
-    ENABLE_HALF: "OFF"
-    ENABLE_BFLOAT16: "ON"
-
-build/amd/nompi/clang/rocm45/debug/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-rocm45-nompi-gnu8-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    BUILD_HIP: "ON"
-    BUILD_TYPE: "Debug"
-
-# ROCm 5.1.4 and friends
-build/amd/nompi/gcc/rocm514/debug/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-rocm514-nompi-gnu11-llvm11
-  variables:
+    BUILD_MPI: "ON"
     BUILD_OMP: "ON"
-    BUILD_HIP: "ON"
     BUILD_TYPE: "Debug"
-    BUILD_SHARED_LIBS: "OFF"
-
-build/amd/nompi/clang/rocm514/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-rocm514-nompi-gnu11-llvm11
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    BUILD_HIP: "ON"
-    BUILD_TYPE: "Release"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-
-# without omp
-build/amd/nompi/gcc/rocm514_wo_omp/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-rocm514-nompi-gnu11-llvm11
-  variables:
-    BUILD_OMP: "OFF"
-    BUILD_MPI: "OFF"
-    BUILD_HIP: "ON"
-    BUILD_TYPE: "Release"
+    BUILD_SHARED_LIBS: "ON"
+    MODULE_LOAD: "cmake/3.24.4 rocm/6.2.0 gcc/13.3.0 openmpi/5.0.7"
 
-build/amd/openmpi/gcc/rocm600/release/static:
+# rocm 6.3.x is the last version support MI50 officially
+build/amd/openmpi/gcc/rocm634_wo_omp/release/shared:
   extends:
     - .build_and_test_tum_template
     - .default_variables
@@ -318,66 +253,52 @@ build/amd/openmpi/gcc/rocm600/release/static:
     BUILD_HWLOC: "OFF"
     BUILD_MPI: "ON"
     BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "OFF"
-    MODULE_LOAD: "cmake/3.24.4 rocm/6.0.0 gcc/13.3.0 openmpi/5.0.7"
+    MODULE_LOAD: "cmake/3.29.6 rocm/6.3.4 gcc/13.3.0 openmpi/5.0.7"
 
-build/amd/nompi/gcc/rocm720/release/shared:
+# mi50 is not officially supported by ROCm >= 6.4.0
+build/amd/nompi/gcc/rocm644/debug/static:
   extends:
     - .build_and_test_tum_template
     - .default_variables
     - .full_test_condition
-    - .use_tum-amd
+    - .use_tum-amd-mi210
   variables:
     BUILD_HIP: "ON"
-    BUILD_OMP: "OFF"
     BUILD_HWLOC: "OFF"
-    BUILD_TYPE: "Release"
-    ENABLE_BFLOAT16: "OFF"
-    ENABLE_HALF: "OFF"
-    MODULE_LOAD: "cmake/3.25.3 rocm/7.2.0 gcc/12.4.0"
-
-build/amd/nompi/gcc/rocm573/debug/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-amd
-  variables:
-    BUILD_HIP: "ON"
+    BUILD_MPI: "OFF"
     BUILD_OMP: "OFF"
-    BUILD_HWLOC: "OFF"
     BUILD_TYPE: "Debug"
-    ENABLE_BFLOAT16: "ON"
-    ENABLE_HALF: "OFF"
-    MODULE_LOAD: "cmake/3.25.3 rocm/5.7.3 gcc/12.4.0"
+    BUILD_SHARED_LIBS: "OFF"
+    MODULE_LOAD: "cmake/3.29.6 rocm/6.4.4 gcc/14.3.0"
 
-build/amd/openmpi/gcc/rocm634_wo_omp/release/shared:
+build/amd/nompi/gcc/rocm710/release/shared:
   extends:
     - .build_and_test_tum_template
     - .default_variables
     - .full_test_condition
-    - .use_tum-amd
+    - .use_tum-amd-mi210
   variables:
     BUILD_HIP: "ON"
     BUILD_HWLOC: "OFF"
-    BUILD_MPI: "ON"
+    BUILD_MPI: "OFF"
+    BUILD_OMP: "OFF"
     BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.29.6 rocm/6.3.4 gcc/13.3.0 openmpi/5.0.7"
+    MODULE_LOAD: "cmake/3.29.6 rocm/7.1.0 gcc/14.3.0"
 
-# mi50 is not officially supported by ROCm >= 7
-build/amd/nompi/gcc/rocm710/release/shared:
+build/amd/nompi/gcc/rocm720/release/shared:
   extends:
     - .build_and_test_tum_template
     - .default_variables
     - .full_test_condition
-    - .use_tum-amd-mi210
+    - .use_tum-amd
   variables:
     BUILD_HIP: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "OFF"
     BUILD_OMP: "OFF"
+    BUILD_HWLOC: "OFF"
     BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.29.6 rocm/7.1.0 gcc/14.3.0"
+    ENABLE_BFLOAT16: "OFF"
+    ENABLE_HALF: "OFF"
+    MODULE_LOAD: "cmake/3.25.3 rocm/7.2.0 gcc/12.4.0"
 
 # no cuda but latest gcc and clang
 build/nocuda/nompi/gcc/core/debug/static:

CMakeLists.txt

@@ -226,23 +226,12 @@ if(GINKGO_BUILD_CUDA)
 endif()
 if(GINKGO_BUILD_HIP)
     include(cmake/hip.cmake)
-    if(GINKGO_HIP_PLATFORM_AMD AND GINKGO_HIP_VERSION VERSION_LESS 5.7)
-        # Hip allow custom namespace but does not fully make everything in the custom namespace before rocm-5.7
-        # more specific pr: https://github.com/ROCm/rocThrust/pull/286
-        message(
-            STATUS
-            "Disable custom thrust namespace for hip before 5.7 because hip does not fully support it before 5.7"
-        )
-        set(GINKGO_HIP_CUSTOM_THRUST_NAMESPACE OFF)
-    elseif(
-        GINKGO_HIP_PLATFORM_AMD
-        AND GINKGO_HIP_VERSION VERSION_GREATER_EQUAL 7.1
-    )
+    if(GINKGO_HIP_PLATFORM_AMD AND GINKGO_HIP_VERSION VERSION_GREATER_EQUAL 7.1)
         # track this issue by https://github.com/ROCm/rocm-libraries/issues/6309 
         # HIP VERSION does not use the exact version number as ROCm. Need to wait for ROCm 7.1.1 to set proper range for ROCm 7.1.0
         message(
             STATUS
-            "Disable custom thrust namespace for hip 7.1 because hip does not adapt the custom namespace fully."
+            "Disable custom thrust namespace after hip 7.1 because hip does not adapt the custom namespace fully."
         )
         set(GINKGO_HIP_CUSTOM_THRUST_NAMESPACE OFF)
     else()

README.md

@@ -54,7 +54,7 @@ The Ginkgo CUDA module has the following __additional__ requirements:
 
 The Ginkgo HIP module has the following __additional__ requirements:
 
-* _ROCm 4.5+_
+* _ROCm 6.2.0+_
 * the HIP, hipBLAS, hipSPARSE, hip/rocRAND and rocThrust packages compiled with the ROCm backend
 * if the hipFFT package is available, it is used to implement the FFT LinOps.
 * _cmake 3.21+_

accessor/hip_helper.hpp

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
+// SPDX-FileCopyrightText: 2017 - 2026 The Ginkgo authors
 //
 // SPDX-License-Identifier: BSD-3-Clause
 
@@ -74,11 +74,7 @@ struct hip_type<gko::half> {
 
 template <>
 struct hip_type<gko::bfloat16> {
-#if HIP_VERSION >= 60200000
     using type = __hip_bfloat16;
-#else
-    using type = hip_bfloat16;
-#endif
 };
 
 

cmake/autodetect_executors.cmake

@@ -35,9 +35,6 @@ if(NOT DEFINED GINKGO_BUILD_HIP)
     if(CMAKE_HIP_COMPILER)
         message(STATUS "Enabling HIP executor")
         set(GINKGO_HAS_HIP ON)
-    else()
-        include(cmake/hip_helpers.cmake)
-        ginkgo_check_hip_detection_issue()
     endif()
 endif()
 

cmake/hip.cmake

@@ -3,7 +3,6 @@ cmake_minimum_required(VERSION 3.21 FATAL_ERROR)
 include(cmake/hip_helpers.cmake)
 include(CheckLanguage)
 check_language(HIP)
-ginkgo_check_hip_detection_issue()
 
 enable_language(HIP)
 
@@ -34,7 +33,7 @@ find_package(rocrand REQUIRED)
 find_package(rocthrust REQUIRED)
 find_package(ROCTX)
 
-if(GINKGO_HIP_AMD_UNSAFE_ATOMIC AND GINKGO_HIP_VERSION VERSION_GREATER_EQUAL 5)
+if(GINKGO_HIP_AMD_UNSAFE_ATOMIC)
     set(CMAKE_HIP_FLAGS
         "${CMAKE_HIP_FLAGS} -munsafe-fp-atomics -Wno-unused-command-line-argument"
     )

cmake/hip_helpers.cmake

@@ -23,20 +23,3 @@ function(ginkgo_find_hip_version)
     )
     set(GINKGO_HIP_VERSION ${GINKGO_HIP_VERSION} PARENT_SCOPE)
 endfunction()
-
-# This function checks if ROCm might not be detected correctly.
-# ROCm < 5.7 has a faulty CMake setup that requires setting
-# CMAKE_PREFIX_PATH=$ROCM_PATH/lib/cmake, otherwise HIP will not be detected.
-function(ginkgo_check_hip_detection_issue)
-    if(NOT CMAKE_HIP_COMPILER)
-        ginkgo_find_hip_version()
-        if(GINKGO_HIP_VERSION AND GINKGO_HIP_VERSION VERSION_LESS 5.7)
-            message(
-                WARNING
-                "Could not find a HIP compiler, but HIP version ${GINKGO_HIP_VERSION} was detected through "
-                "hipconfig. Try setting the environment variable CMAKE_PREFIX_PATH=$ROCM_PATH/lib/cmake, or "
-                "update to ROCm >= 5.7."
-            )
-        endif()
-    endif()
-endfunction()

common/cuda_hip/base/bf16_alias.hpp

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2025 The Ginkgo authors
+// SPDX-FileCopyrightText: 2025 - 2026 The Ginkgo authors
 //
 // SPDX-License-Identifier: BSD-3-Clause
 
@@ -24,7 +24,6 @@ using vendor_bf16 = __nv_bfloat16;
 #elif defined(GKO_COMPILING_HIP)
 
 
-#if HIP_VERSION >= 60200000
 // HIP has __hip_bfloat16 after ROCM 5.6.0 but enough implementation for us
 // (conversion and operation overload) after ROCM 6.2.0 which provides more
 // native operations support.
@@ -39,23 +38,5 @@ using vendor_bf16 = __hip_bfloat16;
 }
 
 
-#else
-
-
-// HIP has hip_bfloat16 but only the type with the operation fallback to the
-// single precision
-#include <hip/hip_bfloat16.h>
-
-
-namespace gko {
-
-
-using vendor_bf16 = hip_bfloat16;
-
-
-}
-
-
-#endif
 #endif
 #endif  // GKO_COMMON_CUDA_HIP_BASE_BF16_ALIAS_HPP_