hipblaslt: new package (#43846)

* initial commit to add hipblaslt package
* remove master and update patch
* add docstring comment

2 files changed, 135 insertions, 0 deletions

diff --git a/var/spack/repos/builtin/packages/hipblaslt/001_Set_LLVM_Paths_And_Add_Includes.patch b/var/spack/repos/builtin/packages/hipblaslt/001_Set_LLVM_Paths_And_Add_Includes.patch
new file mode 100644
index 0000000000..27eb39effd
--- /dev/null
+++ b/var/spack/repos/builtin/packages/hipblaslt/001_Set_LLVM_Paths_And_Add_Includes.patch
@@ -0,0 +1,70 @@
+diff --git a/clients/gtest/CMakeLists.txt b/clients/gtest/CMakeLists.txt
+index 825bdca..f817e12 100644
+--- a/clients/gtest/CMakeLists.txt
++++ b/clients/gtest/CMakeLists.txt
+@@ -53,6 +53,7 @@ target_include_directories( hipblaslt-test
+     $<BUILD_INTERFACE:${BLAS_INCLUDE_DIR}>
+     $<BUILD_INTERFACE:${BLIS_INCLUDE_DIR}> # may be blank if not used
+     $<BUILD_INTERFACE:${GTEST_INCLUDE_DIRS}>
++    $<BUILD_INTERFACE:${HIPBLAS_INCLUDE_DIRS}>
+ )
+ message("BLIS_INCLUDE_DIR=" ${BLIS_INCLUDE_DIR})
+ target_link_libraries( hipblaslt-test PRIVATE ${BLAS_LIBRARY} ${GTEST_BOTH_LIBRARIES} roc::hipblaslt )
+diff --git a/library/CMakeLists.txt b/library/CMakeLists.txt
+index 3252da0..1b8d628 100644
+--- a/library/CMakeLists.txt
++++ b/library/CMakeLists.txt
+@@ -72,6 +72,8 @@ include(src/CMakeLists.txt)
+ # Create hipBLASLt library
+ add_library(hipblaslt ${hipblaslt_source} ${hipblaslt_headers_public})
+ add_library(roc::hipblaslt ALIAS hipblaslt)
++target_include_directories( hipblaslt PRIVATE ${HIPBLAS_INCLUDE_DIRS} )
++target_include_directories( hipblaslt PRIVATE ${MSGPACK_DIR}/include )
+
+ # Target compile definitions
+ if(NOT BUILD_CUDA)
+diff --git a/tensilelite/Tensile/Common.py b/tensilelite/Tensile/Common.py
+index 0f862b8..0baeb55 100644
+--- a/tensilelite/Tensile/Common.py
++++ b/tensilelite/Tensile/Common.py
+@@ -270,6 +270,8 @@ globalParameters["LazyLibraryLoading"] = False # Load library and code object fi
+
+ globalParameters["UseUserArgs"] = False
+
++globalParameters["LLVMPath"] = os.environ.get("LLVM_PATH")
++
+ # Save a copy - since pytest doesn't re-run this initialization code and YAML files can override global settings - odd things can happen
+ defaultGlobalParameters = deepcopy(globalParameters)
+
+@@ -1471,7 +1473,7 @@ def assignGlobalParameters( config ):
+     if os.name == "nt":
+       globalParameters["AssemblerPath"] = locateExe(globalParameters["ROCmBinPath"], "clang++.exe")
+     else:
+-      globalParameters["AssemblerPath"] = locateExe(os.path.join(globalParameters["ROCmPath"], "llvm/bin"), "clang++")
++      globalParameters["AssemblerPath"] = locateExe(os.path.join(globalParameters["LLVMPath"], "bin"), "clang++")
+
+   globalParameters["ROCmSMIPath"] = locateExe(globalParameters["ROCmBinPath"], "rocm-smi")
+
+@@ -1483,7 +1485,7 @@ def assignGlobalParameters( config ):
+     if os.name == "nt":
+       globalParameters["ClangOffloadBundlerPath"] = locateExe(globalParameters["ROCmBinPath"], "clang-offload-bundler.exe")
+     else:
+-      globalParameters["ClangOffloadBundlerPath"] = locateExe(os.path.join(globalParameters["ROCmPath"], "llvm/bin"), "clang-offload-bundler")
++      globalParameters["ClangOffloadBundlerPath"] = locateExe(os.path.join(globalParameters["LLVMPath"], "bin"), "clang-offload-bundler")
+
+   if "ROCmAgentEnumeratorPath" in config:
+     globalParameters["ROCmAgentEnumeratorPath"] = config["ROCmAgentEnumeratorPath"]
+diff --git a/tensilelite/Tensile/Ops/gen_assembly.sh b/tensilelite/Tensile/Ops/gen_assembly.sh
+index 7b16069..1392cab 100644
+--- a/tensilelite/Tensile/Ops/gen_assembly.sh
++++ b/tensilelite/Tensile/Ops/gen_assembly.sh
+@@ -32,7 +32,9 @@ if ! [ -z ${ROCM_PATH+x} ]; then
+     rocm_path=${ROCM_PATH}
+ fi
+
+-toolchain=${rocm_path}/llvm/bin/clang++
++$LLVM_PATH=$ENV{'LLVM_PATH'};
++
++toolchain=${LLVM_PATH}/bin/clang++
+
+ . ${venv}/bin/activate
diff --git a/var/spack/repos/builtin/packages/hipblaslt/package.py b/var/spack/repos/builtin/packages/hipblaslt/package.py
new file mode 100644
index 0000000000..36df9a0ca8
--- /dev/null
+++ b/var/spack/repos/builtin/packages/hipblaslt/package.py
@@ -0,0 +1,65 @@
+# Copyright 2013-2024 Lawrence Livermore National Security, LLC and other
+# Spack Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+from spack.package import *
+
+
+class Hipblaslt(CMakePackage):
+    """hipBLASLt is a library that provides general matrix-matrix operations with a flexible API
+    and extends functionalities beyond a traditional BLAS library"""
+
+    homepage = "https://github.com/ROCm/hipBLASLt"
+    url = "https://github.com/ROCm/hipBLASLt/archive/refs/tags/rocm-6.0.2.tar.gz"
+    git = "https://github.com/ROCm/hipBLASLt.git"
+
+    maintainers("srekolam", "afzpatel", "renjithravindrankannath")
+
+    license("MIT")
+
+    version("6.0.2", sha256="e281a1a7760fab8c3e0baafe17950cf43c422184e3226e3c14eb06e50c69d421")
+    version("6.0.0", sha256="6451b6fdf7f24787628190bbe8f2208c929546b68b692d8355d2f18bea7ca7db")
+
+    amdgpu_targets = ROCmPackage.amdgpu_targets
+
+    variant(
+        "amdgpu_target",
+        description="AMD GPU architecture",
+        values=auto_or_any_combination_of(*amdgpu_targets),
+        sticky=True,
+    )
+    variant("asan", default=False, description="Build with address-sanitizer enabled or disabled")
+
+    for ver in ["6.0.0", "6.0.2"]:
+        depends_on(f"hip@{ver}", when=f"@{ver}")
+        depends_on(f"hipblas@{ver}", when=f"@{ver}")
+        depends_on(f"rocm-openmp-extras@{ver}", type="test", when=f"@{ver}")
+
+    depends_on("msgpack-c")
+    depends_on("py-joblib")
+    depends_on("googletest@1.10.0:", type="test")
+    depends_on("netlib-lapack@3.7.1:", type="test")
+    depends_on("py-pyyaml", type="test")
+
+    # Sets the proper for clang++ and clang-offload-blunder.
+    # Also adds hipblas and msgpack include directories
+    patch("001_Set_LLVM_Paths_And_Add_Includes.patch")
+
+    def setup_build_environment(self, env):
+        env.set("CXX", self.spec["hip"].hipcc)
+
+    def cmake_args(self):
+        args = [
+            self.define("Tensile_CODE_OBJECT_VERSION", "V3"),
+            self.define("MSGPACK_DIR", self.spec["msgpack-c"].prefix),
+            self.define_from_variant("ADDRESS_SANITIZER", "asan"),
+            self.define("BUILD_CLIENTS_TESTS", self.run_tests),
+        ]
+        if "auto" not in self.spec.variants["amdgpu_target"]:
+            args.append(self.define_from_variant("AMDGPU_TARGETS", "amdgpu_target"))
+        if self.run_tests:
+            args.append(
+                self.define("ROCM_OPENMP_EXTRAS_DIR", self.spec["rocm-openmp-extras"].prefix)
+            )
+        return args