Implemented +sycl for slate (#39927)

* Implemented +sycl for slate

* style

* style

* add slate +sycl to ci

* slate +sycl: explicitly specify +mpi

* comment out slate+sycl+mpi in e4s oneapi stack

* removing requirement of intel-oneapi-mpi

* Added slate+sycl to e4s oneapi stack

* Removing obsolete comment

---------

Co-authored-by: eugeneswalker <eugenesunsetwalker@gmail.com>

share/spack/gitlab/cloud_pipelines/stacks/e4s-oneapi/spack.yaml

@@ -231,6 +231,7 @@ spack:
   - cabana +sycl ^kokkos +sycl +openmp cxxstd=17 +tests +examples
   - kokkos +sycl +openmp cxxstd=17 +tests +examples
   - kokkos-kernels build_type=Release %oneapi ^kokkos +sycl +openmp cxxstd=17 +tests +examples
+  - slate +sycl
   # --
   # - ginkgo +oneapi                    # InstallError: Ginkgo's oneAPI backend requires theDPC++ compiler as main CXX compiler.
   # - hpctoolkit +level_zero            # intel-tbb: icpx: error: unknown argument: '-flifetime-dse=1'

var/spack/repos/builtin/packages/blaspp/package.py

@@ -46,11 +46,14 @@ class Blaspp(CMakePackage, CudaPackage, ROCmPackage):
 
     variant("openmp", default=True, description="Use OpenMP internally.")
     variant("shared", default=True, description="Build shared libraries")
+    variant("sycl", default=False, description="Build support for the SYCL backend")
 
     depends_on("cmake@3.15.0:", type="build")
     depends_on("blas")
     depends_on("llvm-openmp", when="%apple-clang +openmp")
     depends_on("rocblas", when="+rocm")
+    depends_on("intel-oneapi-mkl", when="+sycl")
+    depends_on("intel-oneapi-mkl threads=openmp", when="+sycl")
 
     # only supported with clingo solver: virtual dependency preferences
     # depends_on('openblas threads=openmp', when='+openmp ^openblas')
@@ -63,7 +66,21 @@ class Blaspp(CMakePackage, CudaPackage, ROCmPackage):
     conflicts(
         "+rocm", when="@:2020.10.02", msg="ROCm support requires BLAS++ 2021.04.00 or greater"
     )
-    conflicts("+rocm", when="+cuda", msg="BLAS++ can only support one GPU backend at a time")
+    backend_msg = "BLAS++ supports only one GPU backend at a time"
+    conflicts("+rocm", when="+cuda", msg=backend_msg)
+    conflicts("+rocm", when="+sycl", msg=backend_msg)
+    conflicts("+cuda", when="+sycl", msg=backend_msg)
+    conflicts("+sycl", when="@:2023.06.00", msg="SYCL support requires BLAS++ version 2023.08.25")
+
+    # TODO: +sycl requires use of the intel-oneapi compiler, but we cannot express that directly.
+    #       For now, add conflicts for other compilers instead.
+    for __compiler in spack.compilers.supported_compilers():
+        if __compiler != "oneapi":
+            conflicts(
+                "%{0}".format(__compiler),
+                when="+sycl",
+                msg="blaspp+sycl must be compiled with %oneapi",
+            )
 
     def cmake_args(self):
         spec = self.spec
@@ -74,6 +91,8 @@ def cmake_args(self):
                 backend = "cuda"
             if "+rocm" in spec:
                 backend = "hip"
+            if "+sycl" in spec:
+                backend = "sycl"
             backend_config = "-Dgpu_backend=%s" % backend
 
         args = [

var/spack/repos/builtin/packages/lapackpp/package.py

@@ -59,6 +59,7 @@ class Lapackpp(CMakePackage, CudaPackage, ROCmPackage):
     )
 
     variant("shared", default=True, description="Build shared library")
+    variant("sycl", default=False, description="Build support for the SYCL backend")
 
     # Match each LAPACK++ version to a specific BLAS++ version
     for lpp_ver, bpp_ver in _versions:
@@ -66,6 +67,8 @@ class Lapackpp(CMakePackage, CudaPackage, ROCmPackage):
 
     depends_on("blaspp ~cuda", when="~cuda")
     depends_on("blaspp +cuda", when="+cuda")
+    depends_on("blaspp ~sycl", when="~sycl")
+    depends_on("blaspp +sycl", when="+sycl")
     depends_on("blaspp ~rocm", when="~rocm")
     for val in ROCmPackage.amdgpu_targets:
         depends_on("blaspp +rocm amdgpu_target=%s" % val, when="amdgpu_target=%s" % val)
@@ -74,8 +77,23 @@ class Lapackpp(CMakePackage, CudaPackage, ROCmPackage):
     depends_on("lapack")
     depends_on("rocblas", when="+rocm")
     depends_on("rocsolver", when="+rocm")
+    depends_on("intel-oneapi-mkl threads=openmp", when="+sycl")
 
-    conflicts("+rocm", when="+cuda", msg="LAPACK++ can only support one GPU backend at a time")
+    backend_msg = "LAPACK++ supports only one GPU backend at a time"
+    conflicts("+rocm", when="+cuda", msg=backend_msg)
+    conflicts("+rocm", when="+sycl", msg=backend_msg)
+    conflicts("+cuda", when="+sycl", msg=backend_msg)
+    conflicts("+sycl", when="@:2023.06.00", msg="+sycl requires LAPACK++ version 2023.08.25")
+
+    # TODO: +sycl requires use of the intel-oneapi compiler, but we cannot express that directly.
+    #       For now, add conflicts for other compilers instead.
+    for __compiler in spack.compilers.supported_compilers():
+        if __compiler != "oneapi":
+            conflicts(
+                "%{0}".format(__compiler),
+                when="+sycl",
+                msg="lapackpp+sycl must be compiled with %oneapi",
+            )
 
     def cmake_args(self):
         spec = self.spec
@@ -86,6 +104,8 @@ def cmake_args(self):
                 backend = "cuda"
             if "+rocm" in spec:
                 backend = "hip"
+            if "+sycl" in spec:
+                backend = "sycl"
 
         args = [
             "-DBUILD_SHARED_LIBS=%s" % ("+shared" in spec),

var/spack/repos/builtin/packages/slate/package.py

@@ -51,17 +51,23 @@ class Slate(CMakePackage, CudaPackage, ROCmPackage):
     )
     variant("openmp", default=True, description="Build with OpenMP support.")
     variant("shared", default=True, description="Build shared library")
+    variant("sycl", default=False, description="Build with SYCL backend")
 
     # The runtime dependency on cmake is needed by the stand-alone tests (spack test).
     depends_on("cmake", type="run")
 
     depends_on("mpi", when="+mpi")
+    depends_on("intel-oneapi-mkl threads=openmp", when="+sycl")
     depends_on("blas")
     depends_on("blaspp ~cuda", when="~cuda")
     depends_on("blaspp +cuda", when="+cuda")
+    depends_on("blaspp ~sycl", when="~sycl")
+    depends_on("blaspp +sycl", when="+sycl")
     depends_on("blaspp ~rocm", when="~rocm")
     depends_on("lapackpp ~cuda", when="~cuda")
     depends_on("lapackpp +cuda", when="+cuda")
+    depends_on("lapackpp ~sycl", when="~sycl")
+    depends_on("lapackpp +sycl", when="+sycl")
     depends_on("lapackpp ~rocm", when="~rocm")
     for val in CudaPackage.cuda_arch_values:
         depends_on("blaspp +cuda cuda_arch=%s" % val, when="cuda_arch=%s" % val)
@@ -78,6 +84,16 @@ class Slate(CMakePackage, CudaPackage, ROCmPackage):
     depends_on("scalapack", type="test")
     depends_on("hipify-clang", when="@:2021.05.02 +rocm ^hip@5:")
 
+    # TODO: +sycl requires use of the intel-oneapi compiler, but we cannot express that directly.
+    #       For now, add conflicts for other compilers instead.
+    for __compiler in spack.compilers.supported_compilers():
+        if __compiler != "oneapi":
+            conflicts(
+                "%{0}".format(__compiler),
+                when="+sycl",
+                msg="slate+sycl must be compiled with %oneapi",
+            )
+
     cpp_17_msg = "Requires C++17 compiler support"
     conflicts("%gcc@:5", msg=cpp_17_msg)
     conflicts("%xl", msg=cpp_17_msg)
@@ -86,7 +102,11 @@ class Slate(CMakePackage, CudaPackage, ROCmPackage):
     conflicts(
         "+rocm", when="@:2020.10.00", msg="ROCm support requires SLATE 2021.05.01 or greater"
     )
-    conflicts("+rocm", when="+cuda", msg="SLATE only supports one GPU backend at a time")
+    backend_msg = "SLATE supports only one GPU backend at a time"
+    conflicts("+rocm", when="+cuda", msg=backend_msg)
+    conflicts("+rocm", when="+sycl", msg=backend_msg)
+    conflicts("+cuda", when="+sycl", msg=backend_msg)
+    conflicts("+sycl", when="@:2022.07.00", msg="SYCL support requires SLATE version 2023.08.25")
 
     def cmake_args(self):
         spec = self.spec
@@ -97,6 +117,8 @@ def cmake_args(self):
                 backend = "cuda"
             if "+rocm" in spec:
                 backend = "hip"
+            if "+sycl" in spec:
+                backend = "sycl"
             backend_config = "-Dgpu_backend=%s" % backend
 
         config = [