Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

enhance Clang easyblock to add support for building with AMDGPU offload #2684

Merged
merged 2 commits into from
May 25, 2022
Merged

enhance Clang easyblock to add support for building with AMDGPU offload #2684

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
459ffb4
Add AMDGPU offload support to Clang
Feb 21, 2022
bd85142
Fix file checks based on correct versions and reorder extra vars
May 12, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
83 changes: 66 additions & 17 deletions easybuild/easyblocks/c/clang.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@
from easybuild.tools.environment import setvar

# List of all possible build targets for Clang
CLANG_TARGETS = ["all", "AArch64", "ARM", "CppBackend", "Hexagon", "Mips",
CLANG_TARGETS = ["all", "AArch64", "AMDGPU", "ARM", "CppBackend", "Hexagon", "Mips",
"MBlaze", "MSP430", "NVPTX", "PowerPC", "R600", "Sparc",
"SystemZ", "X86", "XCore"]

Expand All @@ -64,6 +64,11 @@
X86_64: ['X86'],
}

# List of all possible AMDGPU gfx targets supported by LLVM
AMDGPU_GFX_SUPPORT = ['gfx700', 'gfx701', 'gfx801', 'gfx803', 'gfx900',
'gfx902', 'gfx906', 'gfx908', 'gfx90a', 'gfx90c',
'gfx1010', 'gfx1030', 'gfx1031']


class EB_Clang(CMakeMake):
"""Support for bootstrapping Clang."""
Expand All @@ -87,6 +92,8 @@ def extra_options():
'skip_sanitizer_tests': [True, "Do not run the sanitizer tests", CUSTOM],
'default_cuda_capability': [None, "Default CUDA capability specified for clang, e.g. '7.5'", CUSTOM],
'build_extra_clang_tools': [False, "Build extra Clang tools", CUSTOM],
'amd_gfx_list': [None, "List of AMDGPU targets to build for. Possible values: " +
nordmoen marked this conversation as resolved.
Show resolved Hide resolved
', '.join(AMDGPU_GFX_SUPPORT), CUSTOM],
})
# disable regular out-of-source build, too simplistic for Clang to work
extra_vars['separate_build_dir'][0] = False
Expand Down Expand Up @@ -195,9 +202,15 @@ def prepare_step(self, *args, **kwargs):
arch = get_cpu_architecture()
try:
default_targets = DEFAULT_TARGETS_MAP[arch][:]
# If CUDA is included as a dep, add NVPTX as a target (could also support AMDGPU if we knew how)
# If CUDA is included as a dep, add NVPTX as a target
if get_software_root("CUDA"):
default_targets += ["NVPTX"]
# For AMDGPU support we need ROCR-Runtime and
# ROCT-Thunk-Interface, however, since ROCT is a dependency of
# ROCR we only check for the ROCR-Runtime here
# https://openmp.llvm.org/SupportAndFAQ.html#q-how-to-build-an-openmp-amdgpu-offload-capable-compiler
if get_software_root("ROCR-Runtime"):
default_targets += ["AMDGPU"]
self.cfg['build_targets'] = build_targets = default_targets
self.log.debug("Using %s as default build targets for CPU/GPU architecture %s.", default_targets, arch)
except KeyError:
Expand Down Expand Up @@ -339,6 +352,16 @@ def configure_step(self):
# whether CUDA was found
if not get_software_root('CUDA'):
setvar('CUDA_NVCC_EXECUTABLE', 'IGNORE')
# If 'AMDGPU' is in the build targets we assume the user would like OpenMP offload support for AMD
if 'AMDGPU' in build_targets:
if not get_software_root('ROCR-Runtime'):
raise EasyBuildError("Can't build Clang with AMDGPU support "
"without dependency 'ROCR-Runtime'")
ec_amdgfx = self.cfg['amd_gfx_list']
if not ec_amdgfx:
raise EasyBuildError("Can't build Clang with AMDGPU support "
"without specifying 'amd_gfx_list'")
self.cfg.update('configopts', '-DLIBOMPTARGET_AMDGCN_GFXLIST=%s' % ' '.join(ec_amdgfx))

self.log.info("Configuring")
super(EB_Clang, self).configure_step(srcdir=self.llvm_src_dir)
Expand Down Expand Up @@ -491,22 +514,48 @@ def sanity_check_step(self):
if LooseVersion(self.version) >= LooseVersion('3.8'):
custom_paths['files'].extend(["lib/libomp.%s" % shlib_ext, "lib/clang/%s/include/omp.h" % self.version])

# Detect OpenMP support for CPU architecture
arch = get_cpu_architecture()
# Check architecture explicitly since Clang uses potentially
# different names
if arch == X86_64:
arch = 'x86_64'
elif arch == POWER:
arch = 'ppc64'
elif arch == AARCH64:
arch = 'aarch64'
else:
print_warning("Unknown CPU architecture (%s) for OpenMP library check!" % arch)
custom_paths['files'].extend(["lib/libomptarget.%s" % shlib_ext,
"lib/libomptarget.rtl.%s.%s" % (arch, shlib_ext)])
# If building for CUDA check that OpenMP target library was created
Comment on lines +517 to +531
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should have been version restricted as it is causing sanity checks to fail with Clang-11.0.1-GCCcore-10.2.0.eb

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I said that because of https://gist.github.com/boegelbot/952bd8e34ed36a074a6e1fb7ad4f0ad9 but waiting on other test reports to confirm

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@ocaisa Worked fine with Clang-11.0.1-gcccuda-2020b.eb though in my test report?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, being investigated, may have just been a fluke

if 'NVPTX' in self.cfg['build_targets']:
arch = get_cpu_architecture()
# Check architecture explicitly since Clang uses potentially
# different names
if arch == X86_64:
arch = 'x86_64'
elif arch == POWER:
arch = 'ppc64'
elif arch == AARCH64:
arch = 'aarch64'
else:
print_warning("Unknown CPU architecture (%s) for OpenMP offloading!" % arch)
custom_paths['files'].extend(["lib/libomptarget.%s" % shlib_ext,
"lib/libomptarget-nvptx.a",
boegel marked this conversation as resolved.
Show resolved Hide resolved
"lib/libomptarget.rtl.cuda.%s" % shlib_ext,
"lib/libomptarget.rtl.%s.%s" % (arch, shlib_ext)])
custom_paths['files'].append("lib/libomptarget.rtl.cuda.%s" % shlib_ext)
# From version 13, and hopefully onwards, the naming of the CUDA
# '.bc' files became a bit simpler and now we don't need to take
# into account the CUDA version Clang was compiled with, making it
# easier to check for the bitcode files we expect
if LooseVersion(self.version) >= LooseVersion('13.0.1'):
ec_cuda_cc = self.cfg['cuda_compute_capabilities']
cfg_cuda_cc = build_option('cuda_compute_capabilities')
cuda_cc = cfg_cuda_cc or ec_cuda_cc or []
# We need the CUDA capability in the form of '75' and not '7.5'
cuda_cc = [cc.replace('.', '') for cc in cuda_cc]
custom_paths['files'].extend(["lib/libomptarget-nvptx-sm_%s.bc" % cc
for cc in cuda_cc])
custom_paths['files'].extend(["lib/libomptarget-new-nvptx-sm_%s.bc" % cc
for cc in cuda_cc])
# If building for AMDGPU check that OpenMP target library was created
if 'AMDGPU' in self.cfg['build_targets']:
custom_paths['files'].append("lib/libLLVMAMDGPUCodeGen.a")
# OpenMP offloading support to AMDGPU was not added until version
# 13, however, building for the AMDGPU target predates this and so
# doesn't necessarily mean that the AMDGPU target failed
if LooseVersion(self.version) >= LooseVersion('13.0.0'):
custom_paths['files'].append("lib/libomptarget.rtl.amdgpu.%s" % shlib_ext)
custom_paths['files'].extend(["lib/libomptarget-amdgcn-%s.bc" % gfx
for gfx in self.cfg['amd_gfx_list']])
custom_paths['files'].append("bin/amdgpu-arch")

custom_commands = ['clang --help', 'clang++ --help', 'llvm-config --cxxflags']
super(EB_Clang, self).sanity_check_step(custom_paths=custom_paths, custom_commands=custom_commands)
Expand Down