Adding support for ROCm for the LBANN software stack. (#21716)

* Also removed LBANN CUDA CMake flags that are set by the
version of Hydrogen that is compiled against.

* Updated recipes to use HWLOC 2.3 with ROCm to enable
topology awareness.

Co-authored-by: Harmen Stoppels <harmenstoppels@gmail.com>

4 files changed, 92 insertions, 8 deletions

diff --git a/var/spack/repos/builtin/packages/aluminum/package.py b/var/spack/repos/builtin/packages/aluminum/package.py
index 247ac56d2f..b835019f7a 100644
--- a/var/spack/repos/builtin/packages/aluminum/package.py
+++ b/var/spack/repos/builtin/packages/aluminum/package.py
@@ -7,7 +7,7 @@ import os
 from spack import *
 
 
-class Aluminum(CMakePackage, CudaPackage):
+class Aluminum(CMakePackage, CudaPackage, ROCmPackage):
     """Aluminum provides a generic interface to high-performance
     communication libraries, with a focus on allreduce
     algorithms. Blocking and non-blocking algorithms and GPU-aware
@@ -38,13 +38,19 @@ class Aluminum(CMakePackage, CudaPackage):
             ' communication of accelerator data')
     variant('cuda_rma', default=False, description='Builds with support for CUDA intra-node '
             ' Put/Get and IPC RMA functionality')
+    variant('rccl', default=False, description='Builds with support for NCCL communication lib')
 
     depends_on('cmake@3.17.0:', type='build')
     depends_on('mpi')
     depends_on('nccl', when='+nccl')
     depends_on('hwloc@1.11:')
     depends_on('hwloc +cuda +nvml', when='+cuda')
+    depends_on('hwloc@2.3.0:', when='+rocm')
     depends_on('cub', when='@:0.1,0.6.0: +cuda ^cuda@:10.99')
+    depends_on('hipcub', when='@:0.1,0.6.0: +rocm')
+
+    conflicts('~cuda', when='+cuda_rma', msg='CUDA RMA support requires CUDA')
+    conflicts('+cuda', when='+rocm', msg='CUDA and ROCm support are mutually exclusive')
 
     generator = 'Ninja'
     depends_on('ninja', type='build')
@@ -54,7 +60,8 @@ class Aluminum(CMakePackage, CudaPackage):
         args = [
             '-DCMAKE_CXX_STANDARD=14',
             '-DALUMINUM_ENABLE_CUDA:BOOL=%s' % ('+cuda' in spec),
-            '-DALUMINUM_ENABLE_NCCL:BOOL=%s' % ('+nccl' in spec)]
+            '-DALUMINUM_ENABLE_NCCL:BOOL=%s' % ('+nccl' in spec or '+rccl' in spec),
+            '-DALUMINUM_ENABLE_ROCM:BOOL=%s' % ('+rocm' in spec)]
 
         if '+cuda' in spec:
             args.append('-DCMAKE_CUDA_STANDARD=14')
@@ -82,4 +89,16 @@ class Aluminum(CMakePackage, CudaPackage):
             args.extend([
                 '-DOpenMP_DIR={0}'.format(clang_root)])
 
+        if '+rocm' in spec:
+            args.extend([
+                '-DHIP_ROOT_DIR={0}'.format(spec['hip'].prefix),
+                '-DHIP_CXX_COMPILER={0}'.format(self.spec['hip'].hipcc)])
+            archs = self.spec.variants['amdgpu_target'].value
+            if archs != 'none':
+                arch_str = ",".join(archs)
+                args.append(
+                    '-DHIP_HIPCC_FLAGS=--amdgpu-target={0}'
+                    ' -g -fsized-deallocation -fPIC'.format(arch_str)
+                )
+
         return args
diff --git a/var/spack/repos/builtin/packages/dihydrogen/package.py b/var/spack/repos/builtin/packages/dihydrogen/package.py
index 0591eaaafb..9f4b74ec35 100644
--- a/var/spack/repos/builtin/packages/dihydrogen/package.py
+++ b/var/spack/repos/builtin/packages/dihydrogen/package.py
@@ -7,7 +7,7 @@ import os
 from spack import *
 
 
-class Dihydrogen(CMakePackage, CudaPackage):
+class Dihydrogen(CMakePackage, CudaPackage, ROCmPackage):
     """DiHydrogen is the second version of the Hydrogen fork of the
        well-known distributed linear algebra library,
        Elemental. DiHydrogen aims to be a basic distributed
@@ -77,10 +77,16 @@ class Dihydrogen(CMakePackage, CudaPackage):
 
     # Add Aluminum variants
     depends_on('aluminum +cuda +nccl +ht +cuda_rma', when='+al +cuda')
+    depends_on('aluminum +rocm +rccl +ht +cuda_rma', when='+al +rocm')
 
     for arch in CudaPackage.cuda_arch_values:
         depends_on('aluminum cuda_arch=%s' % arch, when='+al +cuda cuda_arch=%s' % arch)
 
+    # variants +rocm and amdgpu_targets are not automatically passed to
+    # dependencies, so do it manually.
+    for val in ROCmPackage.amdgpu_targets:
+        depends_on('aluminum amdgpu_target=%s' % val, when='amdgpu_target=%s' % val)
+
     depends_on('cuda', when=('+cuda' or '+legacy'))
     depends_on('cudnn', when=('+cuda' or '+legacy'))
     depends_on('cub', when='^cuda@:10.99')
@@ -190,6 +196,18 @@ class Dihydrogen(CMakePackage, CudaPackage):
                 '-DOpenMP_libomp_LIBRARY={0}/lib/libomp.dylib'.format(
                     clang_root)])
 
+        if '+rocm' in spec:
+            args.extend([
+                '-DHIP_ROOT_DIR={0}'.format(spec['hip'].prefix),
+                '-DHIP_CXX_COMPILER={0}'.format(self.spec['hip'].hipcc)])
+            archs = self.spec.variants['amdgpu_target'].value
+            if archs != 'none':
+                arch_str = ",".join(archs)
+                args.append(
+                    '-DHIP_HIPCC_FLAGS=--amdgpu-target={0}'
+                    ' -g -fsized-deallocation -fPIC'.format(arch_str)
+                )
+
         return args
 
     def setup_build_environment(self, env):
diff --git a/var/spack/repos/builtin/packages/hydrogen/package.py b/var/spack/repos/builtin/packages/hydrogen/package.py
index 72ce07f188..2998aa1342 100644
--- a/var/spack/repos/builtin/packages/hydrogen/package.py
+++ b/var/spack/repos/builtin/packages/hydrogen/package.py
@@ -7,7 +7,7 @@ import os
 from spack import *
 
 
-class Hydrogen(CMakePackage, CudaPackage):
+class Hydrogen(CMakePackage, CudaPackage, ROCmPackage):
     """Hydrogen: Distributed-memory dense and sparse-direct linear algebra
        and optimization library. Based on the Elemental library."""
 
@@ -64,11 +64,13 @@ class Hydrogen(CMakePackage, CudaPackage):
             description='Builds with support for FP16 precision data types')
 
     conflicts('~openmp', when='+omp_taskloops')
+    conflicts('+cuda', when='+rocm', msg='CUDA and ROCm support are mutually exclusive')
 
     depends_on('cmake@3.17.0:', type='build')
     depends_on('mpi')
     depends_on('hwloc@1.11:')
     depends_on('hwloc +cuda +nvml', when='+cuda')
+    depends_on('hwloc@2.3.0:', when='+rocm')
 
     # Note that #1712 forces us to enumerate the different blas variants
     depends_on('openblas', when='blas=openblas')
@@ -96,10 +98,16 @@ class Hydrogen(CMakePackage, CudaPackage):
 
     # Add Aluminum variants
     depends_on('aluminum +cuda +nccl +ht +cuda_rma', when='+al +cuda')
+    depends_on('aluminum +rocm +rccl +ht', when='+al +rocm')
 
     for arch in CudaPackage.cuda_arch_values:
         depends_on('aluminum cuda_arch=%s' % arch, when='+al +cuda cuda_arch=%s' % arch)
 
+    # variants +rocm and amdgpu_targets are not automatically passed to
+    # dependencies, so do it manually.
+    for val in ROCmPackage.amdgpu_targets:
+        depends_on('aluminum amdgpu_target=%s' % val, when='+al +rocm amdgpu_target=%s' % val)
+
     # Note that this forces us to use OpenBLAS until #1712 is fixed
     depends_on('lapack', when='blas=openblas ~openmp_blas')
 
@@ -110,6 +118,7 @@ class Hydrogen(CMakePackage, CudaPackage):
 
     depends_on('cuda', when='+cuda')
     depends_on('cub', when='^cuda@:10.99')
+    depends_on('hipcub', when='+rocm')
     depends_on('half', when='+half')
 
     depends_on('llvm-openmp', when='%apple-clang +openmp')
@@ -143,8 +152,9 @@ class Hydrogen(CMakePackage, CudaPackage):
             '-DHydrogen_ENABLE_MPC:BOOL=%s'        % ('+mpfr' in spec),
             '-DHydrogen_GENERAL_LAPACK_FALLBACK=ON',
             '-DHydrogen_ENABLE_ALUMINUM=%s' % ('+al' in spec),
-            '-DHydrogen_ENABLE_CUB=%s' % ('+cuda' in spec),
+            '-DHydrogen_ENABLE_CUB=%s' % ('+cuda' in spec or '+rocm' in spec),
             '-DHydrogen_ENABLE_CUDA=%s' % ('+cuda' in spec),
+            '-DHydrogen_ENABLE_ROCM=%s' % ('+rocm' in spec),
             '-DHydrogen_ENABLE_TESTING=%s' % ('+test' in spec),
             '-DHydrogen_ENABLE_HALF=%s' % ('+half' in spec),
             '-DHydrogen_ENABLE_GPU_FP16=%s' % enable_gpu_fp16,
@@ -153,6 +163,18 @@ class Hydrogen(CMakePackage, CudaPackage):
         if '+cuda' in spec:
             args.append('-DCMAKE_CUDA_STANDARD=14')
 
+        if '+rocm' in spec:
+            args.extend([
+                '-DHIP_ROOT_DIR={0}'.format(spec['hip'].prefix),
+                '-DHIP_CXX_COMPILER={0}'.format(self.spec['hip'].hipcc)])
+            archs = self.spec.variants['amdgpu_target'].value
+            if archs != 'none':
+                arch_str = ",".join(archs)
+                args.append(
+                    '-DHIP_HIPCC_FLAGS=--amdgpu-target={0}'
+                    ' -g -fsized-deallocation -fPIC'.format(arch_str)
+                )
+
         # Add support for OS X to find OpenMP (LLVM installed via brew)
         if self.spec.satisfies('%clang +openmp platform=darwin'):
             clang = self.compiler.cc
diff --git a/var/spack/repos/builtin/packages/lbann/package.py b/var/spack/repos/builtin/packages/lbann/package.py
index fdc4b4a777..241fc433b6 100644
--- a/var/spack/repos/builtin/packages/lbann/package.py
+++ b/var/spack/repos/builtin/packages/lbann/package.py
@@ -7,7 +7,7 @@ import os
 from spack import *
 
 
-class Lbann(CMakePackage, CudaPackage):
+class Lbann(CMakePackage, CudaPackage, ROCmPackage):
     """LBANN: Livermore Big Artificial Neural Network Toolkit.  A distributed
     memory, HPC-optimized, model and data parallel training toolkit for deep
     neural networks."""
@@ -73,6 +73,7 @@ class Lbann(CMakePackage, CudaPackage):
     conflicts('~cuda', when='+nvprof')
     conflicts('~hwloc', when='+al')
     conflicts('~cuda', when='+nvshmem')
+    conflicts('+cuda', when='+rocm', msg='CUDA and ROCm support are mutually exclusive')
 
     depends_on('cmake@3.17.0:', type='build')
 
@@ -89,6 +90,8 @@ class Lbann(CMakePackage, CudaPackage):
     depends_on('hydrogen +cuda', when='+cuda')
     depends_on('hydrogen ~half', when='~half')
     depends_on('hydrogen +half', when='+half')
+    depends_on('hydrogen ~rocm', when='~rocm')
+    depends_on('hydrogen +rocm', when='+rocm')
     depends_on('hydrogen build_type=Debug', when='build_type=Debug')
 
     # Older versions depended on Elemental not Hydrogen
@@ -103,6 +106,7 @@ class Lbann(CMakePackage, CudaPackage):
 
     # Add Aluminum variants
     depends_on('aluminum +cuda +nccl +ht +cuda_rma', when='+al +cuda')
+    depends_on('aluminum +rocm +rccl +ht', when='+al +rocm')
 
     depends_on('dihydrogen +openmp', when='+dihydrogen')
     depends_on('dihydrogen ~cuda', when='+dihydrogen ~cuda')
@@ -114,6 +118,8 @@ class Lbann(CMakePackage, CudaPackage):
     depends_on('dihydrogen +half', when='+dihydrogen +half')
     depends_on('dihydrogen ~nvshmem', when='+dihydrogen ~nvshmem')
     depends_on('dihydrogen +nvshmem', when='+dihydrogen +nvshmem')
+    depends_on('dihydrogen ~rocm', when='+dihydrogen ~rocm')
+    depends_on('dihydrogen +rocm', when='+dihydrogen +rocm')
     depends_on('dihydrogen@0.1', when='@0.101:0.101.99 +dihydrogen')
     depends_on('dihydrogen@:0.0,0.2:', when='@:0.90,0.102: +dihydrogen')
     conflicts('~dihydrogen', when='+distconv')
@@ -124,13 +130,22 @@ class Lbann(CMakePackage, CudaPackage):
         depends_on('dihydrogen cuda_arch=%s' % arch, when='+dihydrogen cuda_arch=%s' % arch)
         depends_on('nccl cuda_arch=%s' % arch, when='+cuda cuda_arch=%s' % arch)
 
+    # variants +rocm and amdgpu_targets are not automatically passed to
+    # dependencies, so do it manually.
+    for val in ROCmPackage.amdgpu_targets:
+        depends_on('hydrogen amdgpu_target=%s' % val, when='amdgpu_target=%s' % val)
+        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('cudnn', when='@0.90:0.100.99 +cuda')
     depends_on('cudnn@8.0.2:', when='@:0.90,0.101: +cuda')
     depends_on('cub', when='@0.94:0.98.2 +cuda ^cuda@:10.99')
+    depends_on('hipcub', when='+rocm')
     depends_on('mpi')
     depends_on('hwloc@1.11:', when='@:0.90,0.102: +hwloc')
     depends_on('hwloc@1.11:1.11.99', when='@0.95:0.101.99 +hwloc')
     depends_on('hwloc +cuda +nvml', when='+cuda')
+    depends_on('hwloc@2.3.0:', when='+rocm')
 
     depends_on('half', when='+half')
 
@@ -236,8 +251,6 @@ class Lbann(CMakePackage, CudaPackage):
             '-DLBANN_WITH_HWLOC=%s' % ('+hwloc' in spec),
             '-DLBANN_WITH_ALUMINUM:BOOL=%s' % ('+al' in spec),
             '-DLBANN_WITH_CONDUIT:BOOL=%s' % ('+conduit' in spec),
-            '-DLBANN_WITH_CUDA:BOOL=%s' % ('+cuda' in spec),
-            '-DLBANN_WITH_CUDNN:BOOL=%s' % ('+cuda' in spec),
             '-DLBANN_WITH_NVSHMEM:BOOL=%s' % ('+nvshmem' in spec),
             '-DLBANN_WITH_FFT:BOOL=%s' % ('+fft' in spec),
             '-DLBANN_WITH_ONEDNN:BOOL=%s' % ('+onednn' in spec),
@@ -322,6 +335,18 @@ class Lbann(CMakePackage, CudaPackage):
             args.append(
                 '-DLBANN_WITH_DISTCONV:BOOL=%s' % ('+distconv' in spec))
 
+        if '+rocm' in spec:
+            args.extend([
+                '-DHIP_ROOT_DIR={0}'.format(spec['hip'].prefix),
+                '-DHIP_CXX_COMPILER={0}'.format(self.spec['hip'].hipcc)])
+            archs = self.spec.variants['amdgpu_target'].value
+            if archs != 'none':
+                arch_str = ",".join(archs)
+                args.append(
+                    '-DHIP_HIPCC_FLAGS=--amdgpu-target={0}'
+                    ' -g -fsized-deallocation -fPIC -std=c++17'.format(arch_str)
+                )
+
         return args
 
     @when('@0.91:0.93')