Support ROCm backing in DiHydrogen (#33563)

* Added support for building the DiHydrogen package and LBANN extensions
to DiHydrogen with ROCm libraries.

Fixed a bug on Cray systems where CMake didn't try hard enough to find
an MPI-compatible compiler wrapper.  Make it look more.

Added support for the roctracer package when using ROCm libraries.

* Fixed how ROCm support is defined for pre-v0.3 versions.

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

@@ -83,9 +83,9 @@ class Dihydrogen(CMakePackage, CudaPackage, ROCmPackage):
     for val in ROCmPackage.amdgpu_targets:
         depends_on("aluminum amdgpu_target=%s" % val, when="amdgpu_target=%s" % val)
 
-    for when in ["+cuda", "+distconv"]:
+    depends_on("roctracer-dev", when="+rocm +distconv")
-        depends_on("cuda", when=when)
+
-        depends_on("cudnn", when=when)
+    depends_on("cudnn", when="+cuda +distconv")
     depends_on("cub", when="^cuda@:10")
 
     # Note that #1712 forces us to enumerate the different blas variants
@@ -108,8 +108,8 @@ class Dihydrogen(CMakePackage, CudaPackage, ROCmPackage):
     depends_on("cray-libsci", when="blas=libsci")
     depends_on("cray-libsci +openmp", when="blas=libsci +openmp_blas")
 
-    # Distconv builds require cuda
+    # Distconv builds require cuda or rocm
-    conflicts("~cuda", when="+distconv")
+    conflicts("+distconv", when="~cuda ~rocm")
 
     conflicts("+distconv", when="+half")
     conflicts("+rocm", when="+half")
@@ -120,6 +120,8 @@ class Dihydrogen(CMakePackage, CudaPackage, ROCmPackage):
     depends_on("ninja", type="build")
     depends_on("cmake@3.17.0:", type="build")
 
+    depends_on("spdlog", when="@:0.1,0.2:")
+
     depends_on("llvm-openmp", when="%apple-clang +openmp")
 
     # TODO: Debug linker errors when NVSHMEM is built with UCX
@@ -155,10 +157,14 @@ def cmake_args(self):
             "-DH2_ENABLE_DISTCONV_LEGACY=%s" % ("+distconv" in spec),
             "-DH2_ENABLE_OPENMP=%s" % ("+openmp" in spec),
             "-DH2_ENABLE_FP16=%s" % ("+half" in spec),
-            "-DH2_ENABLE_HIP_ROCM=%s" % ("+rocm" in spec),
             "-DH2_DEVELOPER_BUILD=%s" % ("+developer" in spec),
         ]
 
+        if spec.version < Version("0.3"):
+            args.append("-DH2_ENABLE_HIP_ROCM=%s" % ("+rocm" in spec))
+        else:
+            args.append("-DH2_ENABLE_ROCM=%s" % ("+rocm" in spec))
+
         if not spec.satisfies("^cmake@3.23.0"):
             # There is a bug with using Ninja generator in this version
             # of CMake
@@ -181,7 +187,7 @@ def cmake_args(self):
             if spec.satisfies("%cce") and spec.satisfies("^cuda+allow-unsupported-compilers"):
                 args.append("-DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler")
 
-        if "+cuda" in spec or "+distconv" in spec:
+        if "+cuda" in spec:
             args.append("-DcuDNN_DIR={0}".format(spec["cudnn"].prefix))
 
         if spec.satisfies("^cuda@:10"):
@@ -209,6 +215,12 @@ def cmake_args(self):
                     "-DHIP_CXX_COMPILER={0}".format(self.spec["hip"].hipcc),
                 ]
             )
+            if "platform=cray" in spec:
+                args.extend(
+                    [
+                        "-DMPI_ASSUME_NO_BUILTIN_MPI=ON",
+                    ]
+                )
             archs = self.spec.variants["amdgpu_target"].value
             if archs != "none":
                 arch_str = ",".join(archs)

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

@@ -167,7 +167,8 @@ class Lbann(CMakePackage, CudaPackage, ROCmPackage):
     depends_on("dihydrogen +cuda", when="+dihydrogen +cuda")
     depends_on("dihydrogen ~al", when="+dihydrogen ~al")
     depends_on("dihydrogen +al", when="+dihydrogen +al")
-    depends_on("dihydrogen +distconv +cuda", when="+distconv")
+    depends_on("dihydrogen +distconv +cuda", when="+distconv +cuda")
+    depends_on("dihydrogen +distconv +rocm", when="+distconv +rocm")
     depends_on("dihydrogen ~half", when="+dihydrogen ~half")
     depends_on("dihydrogen +half", when="+dihydrogen +half")
     depends_on("dihydrogen ~nvshmem", when="+dihydrogen ~nvshmem")
@@ -191,6 +192,8 @@ class Lbann(CMakePackage, CudaPackage, ROCmPackage):
         depends_on("aluminum amdgpu_target=%s" % val, when="+al amdgpu_target=%s" % val)
         depends_on("dihydrogen amdgpu_target=%s" % val, when="+dihydrogen amdgpu_target=%s" % val)
 
+    depends_on("roctracer-dev", when="+rocm +distconv")
+
     depends_on("cudnn", when="@0.90:0.100 +cuda")
     depends_on("cudnn@8.0.2:", when="@:0.90,0.101: +cuda")
     depends_on("cub", when="@0.94:0.98.2 +cuda ^cuda@:10")
@@ -334,6 +337,7 @@ def cmake_args(self):
                 "-DLBANN_WITH_ONNX:BOOL=%s" % ("+onnx" in spec),
                 "-DLBANN_WITH_EMBEDDED_PYTHON:BOOL=%s" % ("+python" in spec),
                 "-DLBANN_WITH_PYTHON_FRONTEND:BOOL=%s" % ("+pfe" in spec),
+                "-DLBANN_WITH_ROCTRACER:BOOL=%s" % ("+rocm +distconv" in spec),
                 "-DLBANN_WITH_TBINF=OFF",
                 "-DLBANN_WITH_UNIT_TESTING:BOOL=%s" % ("+unit_tests" in spec),
                 "-DLBANN_WITH_VISION:BOOL=%s" % ("+vision" in spec),
@@ -424,6 +428,12 @@ def cmake_args(self):
                     "-DHIP_CXX_COMPILER={0}".format(self.spec["hip"].hipcc),
                 ]
             )
+            if "platform=cray" in spec:
+                args.extend(
+                    [
+                        "-DMPI_ASSUME_NO_BUILTIN_MPI=ON",
+                    ]
+                )
             archs = self.spec.variants["amdgpu_target"].value
             if archs != "none":
                 arch_str = ",".join(archs)