{OpenFOAM,PETSc,petsc4Foam}: single-precision (HIP) OpenFOAM GPU-solver chain for foss/2025a by hmeiland · Pull Request #26578 · easybuilders/easybuild-easyconfigs

hmeiland · 2026-07-31T15:56:43Z

What

Adds the single-precision (SP) HIP / AMD GPU OpenFOAM external-solver chain for the foss/2025a toolchain:

OpenFOAM-v2506-foss-2025a-SP.eb — single-precision OpenFOAM v2506, built via the new precision='SP' easyconfig parameter.
PETSc-3.23.5-foss-2025a-HIP-fp32.eb — PETSc with --with-hip and --with-precision=single.
petsc4Foam-v2506-foss-2025a-HIP-fp32.eb — the petsc4Foam v2506 adapter (Bundle) linking the SP OpenFOAM to the single-precision PETSc.

Why single precision

The petsc4Foam adapter requires PetscScalar == Foam::solveScalar == float. That means PETSc must be built --with-precision=single and OpenFOAM must be a single-precision build — hence the -SP OpenFOAM paired with the -HIP-fp32 PETSc.

Dependency on the easyblocks repo

OpenFOAM-v2506-foss-2025a-SP.eb uses the new precision easyconfig parameter added to the OpenFOAM easyblock in easybuilders/easybuild-easyblocks#4189 — that PR must merge first for this easyconfig to build.

Target

AMD gfx1101, ROCm base deps resolved from the EESSI index.

Status

Draft — depends on easyblocks#4189, and build/test reports are still pending. Opening for early feedback on the SP packaging approach.

…solver chain for foss/2025a Add the single-precision (SP) OpenFOAM v2506 external-solver chain for the foss/2025a toolchain, targeting AMD GPUs (HIP, gfx1101): - OpenFOAM-v2506-foss-2025a-SP: SP OpenFOAM via the new 'precision' easyblock param (precision='SP'); requires easyblocks PR easybuilders#4189. - PETSc-3.23.5-foss-2025a-HIP-fp32: PETSc with --with-hip and --with-precision=single, matching PetscScalar to OpenFOAM's float solveScalar. - petsc4Foam-v2506-foss-2025a-HIP-fp32: Bundle adapter linking the SP OpenFOAM to the single-precision PETSc. petsc4Foam requires PetscScalar == Foam::solveScalar == float, hence the SP OpenFOAM + single-precision PETSc pairing.

github-actions · 2026-07-31T15:57:33Z

Updated software `OpenFOAM-v2506-foss-2025a-SP.eb`

Diff against OpenFOAM-v2406-foss-2023b-int64.eb

easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2406-foss-2023b-int64.eb

diff --git a/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2406-foss-2023b-int64.eb b/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025a-SP.eb
index df587365df..39ff4fb94e 100644
--- a/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2406-foss-2023b-int64.eb
+++ b/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025a-SP.eb
@@ -1,19 +1,26 @@
 name = 'OpenFOAM'
-version = 'v2406'
-_v = '64'
-versionsuffix = '-int%s' % _v
+version = 'v2506'
+versionsuffix = '-SP'
 
 homepage = 'https://www.openfoam.com/'
 description = """OpenFOAM is a free, open source CFD software package.
  OpenFOAM has an extensive range of features to solve anything from complex fluid flows
  involving chemical reactions, turbulence and heat transfer,
- to solid dynamics and electromagnetics."""
+ to solid dynamics and electromagnetics.
+ This is a single-precision (WM_PRECISION_OPTION=SP) build, required by the fp32
+ petsc4Foam adapter (PetscScalar == Foam::solveScalar == float)."""
 
-toolchain = {'name': 'foss', 'version': '2023b'}
+toolchain = {'name': 'foss', 'version': '2025a'}
 # Users have found that vectorizion caused OpenFOAM to produce some very incorrect results.
 # Disabling vectorize was confirmed to fix the the known issues.
 # With no test suite, sticking to known working toolchain options until proven otherwise.
-toolchainopts = {'cstd': 'c++14', 'vectorize': False, 'i8': True}
+toolchainopts = {'cstd': 'c++17', 'vectorize': False}
+
+# Single-precision build. Handled by the OpenFOAM easyblock 'precision' parameter (patched
+# in this repo's easyblocks/openfoam.py, deployed via --include-easyblocks). etc/bashrc sets
+# WM_PRECISION_OPTION=DP unconditionally, so the easyblock writes etc/prefs.sh (the sanctioned
+# override, sourced after the DP default and before WM_OPTIONS is computed) => linux64GccSPInt32Opt.
+precision = 'SP'
 
 source_urls = ['https://sourceforge.net/projects/openfoam/files/%(version)s/']
 sources = [
@@ -28,30 +35,53 @@ patches = [
     ('OpenFOAM-v2212-wmake-OpenMPI.patch', 1),
 ]
 checksums = [
-    {'OpenFOAM-v2406.tgz': '8d1450fb89eec1e7cecc55c3bb7bc486ccbf63d069379d1d5d7518fa16a4686a'},
-    {'OpenFOAM-plugins-v2406.tgz': '1d008f86fad06a4a568d194c6e3d5ab52be2b20c83a3b9b1b0230e2de2d0558b'},
+    {'OpenFOAM-v2506.tgz': '63d26f48ae7ee9a7806a0ceb339ef8a0ba485a4714d54fbfb31e78e1a4849965'},
+    {'OpenFOAM-plugins-v2506.tgz': '79309a0e2e798bd533dc60384542fdfd916b3f5c131600c5a7696b51d4120845'},
     {'OpenFOAM-v2406-cleanup.patch': '3abff48a517fb63719ad57fa32af746bc379a1e80c72d3e5852aa17cd13cf03e'},
     {'OpenFOAM-v2212-wmake-OpenMPI.patch': '241dc4898c22aab0cbd10c1ea931a07a786508ee03462d45dbc1c202fee3ebe8'},
 ]
 
 builddependencies = [
     ('Bison', '3.8.2'),
-    ('CMake', '3.27.6'),
+    ('CMake', '3.31.3'),
     ('flex', '2.6.4'),
 ]
 
 dependencies = [
     ('libreadline', '8.2'),
-    ('ncurses', '6.4'),
+    ('ncurses', '6.5'),
     # OpenFOAM requires 64 bit METIS using 32 bit indexes (array indexes)
     ('METIS', '5.1.0'),
-    ('SCOTCH', '7.0.4', versionsuffix),
-    ('KaHIP', '3.16'),
-    ('CGAL', '5.6.1'),
+    ('SCOTCH', '7.0.8'),
+    ('KaHIP', '3.19'),
+    ('CGAL', '6.0.1'),
     ('GMP', '6.3.0'),
-    ('MPFR', '4.2.1'),
-    ('ParaView', '5.12.0'),
-    ('gnuplot', '6.0.1'),
+    ('MPFR', '4.2.2'),
+    ('ParaView', '6.0.1'),
+    ('gnuplot', '6.0.3'),
 ]
 
+# --- EESSI-local fix: disable OpenFOAM readline auto-detection ---
+# On EESSI, OpenFOAM's wmake have_readline script defaulted READLINE_ARCH_PATH to /usr,
+# injecting `-DHAVE_LIBREADLINE -I/usr/include` into the setSet.C compile. On this system
+# /usr/include resolves to the EESSI compat layer
+# (/cvmfs/software.eessi.io/.../compat/linux/x86_64/usr/include/stdlib.h), whose __COLD
+# attribute breaks the GCCcore-14.2.0 libstdc++ <cstdlib> (error: expected initializer
+# before '__COLD'; 'abort' has not been declared). That aborts the mesh utilities build
+# (make Error 2), leaving 10 solvers unbuilt and failing the EasyBuild sanity check.
+# readline only provides interactive line-editing in setSet, which is irrelevant to the
+# solvers and to petsc4Foam. Disable it cleanly: overwrite the config.sh/readline snippet
+# with READLINE_ARCH_PATH=none (have_readline treats 'none' as "not found", so it emits no
+# -DHAVE_LIBREADLINE / -I/usr/include at all) and belt-and-suspenders export it in the env.
+# NOTE: prebuildopts must contain NO '%' characters. The OpenFOAM easyblock inlines
+# prebuildopts into the build command and then runs `cmd_tmpl % cmd` (a second %-format
+# pass), so a literal %s here raises "not enough arguments for format string". Use echo,
+# not printf, to write the config snippet.
+prebuildopts = (
+    'mkdir -p "$WM_PROJECT_DIR/etc/easybuild/config.sh" && '
+    'echo "export READLINE_ARCH_PATH=none" > "$WM_PROJECT_DIR/etc/easybuild/config.sh/readline" && '
+    'export FOAM_CONFIG_ETC=etc/easybuild && '
+    'export READLINE_ARCH_PATH=none && '
+)
+
 moduleclass = 'cae'

Diff against OpenFOAM-v2412_250814-foss-2024a.eb

easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2412_250814-foss-2024a.eb

diff --git a/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2412_250814-foss-2024a.eb b/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025a-SP.eb
index 0436ef701c..39ff4fb94e 100644
--- a/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2412_250814-foss-2024a.eb
+++ b/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025a-SP.eb
@@ -1,33 +1,49 @@
 name = 'OpenFOAM'
-version = 'v2412_250814'
+version = 'v2506'
+versionsuffix = '-SP'
 
 homepage = 'https://www.openfoam.com/'
 description = """OpenFOAM is a free, open source CFD software package.
  OpenFOAM has an extensive range of features to solve anything from complex fluid flows
  involving chemical reactions, turbulence and heat transfer,
- to solid dynamics and electromagnetics."""
+ to solid dynamics and electromagnetics.
+ This is a single-precision (WM_PRECISION_OPTION=SP) build, required by the fp32
+ petsc4Foam adapter (PetscScalar == Foam::solveScalar == float)."""
 
-toolchain = {'name': 'foss', 'version': '2024a'}
+toolchain = {'name': 'foss', 'version': '2025a'}
 # Users have found that vectorizion caused OpenFOAM to produce some very incorrect results.
 # Disabling vectorize was confirmed to fix the the known issues.
 # With no test suite, sticking to known working toolchain options until proven otherwise.
-toolchainopts = {'vectorize': False}
+toolchainopts = {'cstd': 'c++17', 'vectorize': False}
 
-source_urls = ['https://sourceforge.net/projects/openfoam/files/%s/' % version.split("_")[0]]
-sources = [SOURCE_TGZ]
+# Single-precision build. Handled by the OpenFOAM easyblock 'precision' parameter (patched
+# in this repo's easyblocks/openfoam.py, deployed via --include-easyblocks). etc/bashrc sets
+# WM_PRECISION_OPTION=DP unconditionally, so the easyblock writes etc/prefs.sh (the sanctioned
+# override, sourced after the DP default and before WM_OPTIONS is computed) => linux64GccSPInt32Opt.
+precision = 'SP'
+
+source_urls = ['https://sourceforge.net/projects/openfoam/files/%(version)s/']
+sources = [
+    SOURCE_TGZ,
+    {
+        'filename': '%(name)s-plugins-%(version)s.tgz',
+        'extract_cmd': 'tar --strip-components=1 -C %(installdir)s/%(name)s-%(version)s/ -xzf %s'
+    }
+]
 patches = [
     ('OpenFOAM-v2406-cleanup.patch', 1),
     ('OpenFOAM-v2212-wmake-OpenMPI.patch', 1),
 ]
 checksums = [
-    {'OpenFOAM-v2412_250814.tgz': '04af881e3a91be8e7bc6916c63a9e69d8ee98a3f9a5445c8c1da8cbaafbb99cd'},
+    {'OpenFOAM-v2506.tgz': '63d26f48ae7ee9a7806a0ceb339ef8a0ba485a4714d54fbfb31e78e1a4849965'},
+    {'OpenFOAM-plugins-v2506.tgz': '79309a0e2e798bd533dc60384542fdfd916b3f5c131600c5a7696b51d4120845'},
     {'OpenFOAM-v2406-cleanup.patch': '3abff48a517fb63719ad57fa32af746bc379a1e80c72d3e5852aa17cd13cf03e'},
     {'OpenFOAM-v2212-wmake-OpenMPI.patch': '241dc4898c22aab0cbd10c1ea931a07a786508ee03462d45dbc1c202fee3ebe8'},
 ]
 
 builddependencies = [
     ('Bison', '3.8.2'),
-    ('CMake', '3.29.3'),
+    ('CMake', '3.31.3'),
     ('flex', '2.6.4'),
 ]
 
@@ -36,13 +52,36 @@ dependencies = [
     ('ncurses', '6.5'),
     # OpenFOAM requires 64 bit METIS using 32 bit indexes (array indexes)
     ('METIS', '5.1.0'),
-    ('SCOTCH', '7.0.6'),
+    ('SCOTCH', '7.0.8'),
     ('KaHIP', '3.19'),
-    ('CGAL', '5.6.1'),
+    ('CGAL', '6.0.1'),
     ('GMP', '6.3.0'),
-    ('MPFR', '4.2.1'),
-    ('ParaView', '5.13.2'),
-    ('gnuplot', '6.0.1'),
+    ('MPFR', '4.2.2'),
+    ('ParaView', '6.0.1'),
+    ('gnuplot', '6.0.3'),
 ]
 
+# --- EESSI-local fix: disable OpenFOAM readline auto-detection ---
+# On EESSI, OpenFOAM's wmake have_readline script defaulted READLINE_ARCH_PATH to /usr,
+# injecting `-DHAVE_LIBREADLINE -I/usr/include` into the setSet.C compile. On this system
+# /usr/include resolves to the EESSI compat layer
+# (/cvmfs/software.eessi.io/.../compat/linux/x86_64/usr/include/stdlib.h), whose __COLD
+# attribute breaks the GCCcore-14.2.0 libstdc++ <cstdlib> (error: expected initializer
+# before '__COLD'; 'abort' has not been declared). That aborts the mesh utilities build
+# (make Error 2), leaving 10 solvers unbuilt and failing the EasyBuild sanity check.
+# readline only provides interactive line-editing in setSet, which is irrelevant to the
+# solvers and to petsc4Foam. Disable it cleanly: overwrite the config.sh/readline snippet
+# with READLINE_ARCH_PATH=none (have_readline treats 'none' as "not found", so it emits no
+# -DHAVE_LIBREADLINE / -I/usr/include at all) and belt-and-suspenders export it in the env.
+# NOTE: prebuildopts must contain NO '%' characters. The OpenFOAM easyblock inlines
+# prebuildopts into the build command and then runs `cmd_tmpl % cmd` (a second %-format
+# pass), so a literal %s here raises "not enough arguments for format string". Use echo,
+# not printf, to write the config snippet.
+prebuildopts = (
+    'mkdir -p "$WM_PROJECT_DIR/etc/easybuild/config.sh" && '
+    'echo "export READLINE_ARCH_PATH=none" > "$WM_PROJECT_DIR/etc/easybuild/config.sh/readline" && '
+    'export FOAM_CONFIG_ETC=etc/easybuild && '
+    'export READLINE_ARCH_PATH=none && '
+)
+
 moduleclass = 'cae'

Diff against OpenFOAM-v2506-foss-2025b.eb

easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025b.eb

diff --git a/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025b.eb b/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025a-SP.eb
index b06faf61f1..39ff4fb94e 100644
--- a/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025b.eb
+++ b/easybuild/easyconfigs/o/OpenFOAM/OpenFOAM-v2506-foss-2025a-SP.eb
@@ -1,18 +1,27 @@
 name = 'OpenFOAM'
 version = 'v2506'
+versionsuffix = '-SP'
 
 homepage = 'https://www.openfoam.com/'
 description = """OpenFOAM is a free, open source CFD software package.
  OpenFOAM has an extensive range of features to solve anything from complex fluid flows
  involving chemical reactions, turbulence and heat transfer,
- to solid dynamics and electromagnetics."""
+ to solid dynamics and electromagnetics.
+ This is a single-precision (WM_PRECISION_OPTION=SP) build, required by the fp32
+ petsc4Foam adapter (PetscScalar == Foam::solveScalar == float)."""
 
-toolchain = {'name': 'foss', 'version': '2025b'}
+toolchain = {'name': 'foss', 'version': '2025a'}
 # Users have found that vectorizion caused OpenFOAM to produce some very incorrect results.
 # Disabling vectorize was confirmed to fix the the known issues.
 # With no test suite, sticking to known working toolchain options until proven otherwise.
 toolchainopts = {'cstd': 'c++17', 'vectorize': False}
 
+# Single-precision build. Handled by the OpenFOAM easyblock 'precision' parameter (patched
+# in this repo's easyblocks/openfoam.py, deployed via --include-easyblocks). etc/bashrc sets
+# WM_PRECISION_OPTION=DP unconditionally, so the easyblock writes etc/prefs.sh (the sanctioned
+# override, sourced after the DP default and before WM_OPTIONS is computed) => linux64GccSPInt32Opt.
+precision = 'SP'
+
 source_urls = ['https://sourceforge.net/projects/openfoam/files/%(version)s/']
 sources = [
     SOURCE_TGZ,
@@ -24,19 +33,17 @@ sources = [
 patches = [
     ('OpenFOAM-v2406-cleanup.patch', 1),
     ('OpenFOAM-v2212-wmake-OpenMPI.patch', 1),
-    ('OpenFOAM-v2506-qt6_plugins.patch', 1),
 ]
 checksums = [
     {'OpenFOAM-v2506.tgz': '63d26f48ae7ee9a7806a0ceb339ef8a0ba485a4714d54fbfb31e78e1a4849965'},
     {'OpenFOAM-plugins-v2506.tgz': '79309a0e2e798bd533dc60384542fdfd916b3f5c131600c5a7696b51d4120845'},
     {'OpenFOAM-v2406-cleanup.patch': '3abff48a517fb63719ad57fa32af746bc379a1e80c72d3e5852aa17cd13cf03e'},
     {'OpenFOAM-v2212-wmake-OpenMPI.patch': '241dc4898c22aab0cbd10c1ea931a07a786508ee03462d45dbc1c202fee3ebe8'},
-    {'OpenFOAM-v2506-qt6_plugins.patch': '8844575a3dbed9baa2e4eb658a85ad52cd54ab9fe08489cdc26e0e4ed75f67db'},
 ]
 
 builddependencies = [
     ('Bison', '3.8.2'),
-    ('CMake', '4.0.3'),
+    ('CMake', '3.31.3'),
     ('flex', '2.6.4'),
 ]
 
@@ -45,7 +52,7 @@ dependencies = [
     ('ncurses', '6.5'),
     # OpenFOAM requires 64 bit METIS using 32 bit indexes (array indexes)
     ('METIS', '5.1.0'),
-    ('SCOTCH', '7.0.10'),
+    ('SCOTCH', '7.0.8'),
     ('KaHIP', '3.19'),
     ('CGAL', '6.0.1'),
     ('GMP', '6.3.0'),
@@ -54,4 +61,27 @@ dependencies = [
     ('gnuplot', '6.0.3'),
 ]
 
+# --- EESSI-local fix: disable OpenFOAM readline auto-detection ---
+# On EESSI, OpenFOAM's wmake have_readline script defaulted READLINE_ARCH_PATH to /usr,
+# injecting `-DHAVE_LIBREADLINE -I/usr/include` into the setSet.C compile. On this system
+# /usr/include resolves to the EESSI compat layer
+# (/cvmfs/software.eessi.io/.../compat/linux/x86_64/usr/include/stdlib.h), whose __COLD
+# attribute breaks the GCCcore-14.2.0 libstdc++ <cstdlib> (error: expected initializer
+# before '__COLD'; 'abort' has not been declared). That aborts the mesh utilities build
+# (make Error 2), leaving 10 solvers unbuilt and failing the EasyBuild sanity check.
+# readline only provides interactive line-editing in setSet, which is irrelevant to the
+# solvers and to petsc4Foam. Disable it cleanly: overwrite the config.sh/readline snippet
+# with READLINE_ARCH_PATH=none (have_readline treats 'none' as "not found", so it emits no
+# -DHAVE_LIBREADLINE / -I/usr/include at all) and belt-and-suspenders export it in the env.
+# NOTE: prebuildopts must contain NO '%' characters. The OpenFOAM easyblock inlines
+# prebuildopts into the build command and then runs `cmd_tmpl % cmd` (a second %-format
+# pass), so a literal %s here raises "not enough arguments for format string". Use echo,
+# not printf, to write the config snippet.
+prebuildopts = (
+    'mkdir -p "$WM_PROJECT_DIR/etc/easybuild/config.sh" && '
+    'echo "export READLINE_ARCH_PATH=none" > "$WM_PROJECT_DIR/etc/easybuild/config.sh/readline" && '
+    'export FOAM_CONFIG_ETC=etc/easybuild && '
+    'export READLINE_ARCH_PATH=none && '
+)
+
 moduleclass = 'cae'

Updated software `PETSc-3.23.5-foss-2025a-HIP-fp32.eb`

Diff against PETSc-3.24.0-lfoss-2025b.eb

easybuild/easyconfigs/p/PETSc/PETSc-3.24.0-lfoss-2025b.eb

diff --git a/easybuild/easyconfigs/p/PETSc/PETSc-3.24.0-lfoss-2025b.eb b/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a-HIP-fp32.eb
index cf6ce384bd..2a97030e73 100644
--- a/easybuild/easyconfigs/p/PETSc/PETSc-3.24.0-lfoss-2025b.eb
+++ b/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a-HIP-fp32.eb
@@ -1,52 +1,125 @@
-# HDF5-support added
-# Author: J. Saßmannshausen (Imperial College London/UK)
+# EasyBuild easyconfig: PETSc with AMD-GPU (HIP) backend — SINGLE PRECISION (fp32)
+#
+# =========================================================================
+#  fp32 VARIANT of PETSc-3.23.5-foss-2025a-HIP.eb.
+#  IDENTICAL to the (default, double-precision / fp64) HIP easyconfig in
+#  every respect EXCEPT:
+#    * versionsuffix   '-HIP'  ->  '-HIP-fp32'
+#    * configopts adds  --with-precision=single
+#  See PETSc-3.23.5-foss-2025a-HIP.eb for the full rationale on the GPU-arch
+#  parameterisation and the ROCm dependency reconciliation — all of that is
+#  unchanged here. Only the scalar precision differs.
+#
+#  WHY A SEPARATE MODULE:
+#    PETSc's scalar precision (double vs single) is fixed at configure time and
+#    baked into libpetsc + the installed headers. fp32 and fp64 cannot coexist
+#    in one install, so the fp32 build gets its own module
+#    (PETSc/3.23.5-foss-2025a-HIP-fp32) and its own petsc4Foam adapter linked
+#    against it (petsc4Foam/v2506-foss-2025a-HIP-fp32).
+# =========================================================================
+
+easyblock = 'EB_PETSc'
 
 name = 'PETSc'
-version = '3.24.0'
+version = '3.23.5'
+# fp32 single-precision HIP variant. Arch (gfxNNN) still lives in the EESSI accel
+# path, not the suffix; the suffix marks (HIP backend + single precision).
+versionsuffix = '-HIP-fp32'
 
 homepage = 'https://www.mcs.anl.gov/petsc'
 description = """PETSc, pronounced PET-see (the S is silent), is a suite of data structures and routines for the
- scalable (parallel) solution of scientific applications modeled by partial differential equations."""
+ scalable (parallel) solution of scientific applications modeled by partial differential equations.
+
+This variant enables the AMD-GPU (HIP) backend with hipBLAS and hipSPARSE, built in SINGLE precision
+(--with-precision=single, fp32), so GPU-accelerated Krylov solvers/preconditioners run in 32-bit
+floating point (e.g. for the petsc4Foam / OpenFOAM external-solver adapter). The target GPU
+architecture is taken from the amdgcn_capabilities parameter (default gfx1101, AMD Radeon Pro V710),
+overridable via --amdgcn-capabilities."""
 
-toolchain = {'name': 'lfoss', 'version': '2025b'}
+toolchain = {'name': 'foss', 'version': '2025a'}
 toolchainopts = {'openmp': True, 'usempi': True, 'pic': True}
 
 source_urls = [
     'https://web.cels.anl.gov/projects/petsc/download/release-snapshots',
 ]
 sources = [SOURCELOWER_TAR_GZ]
-checksums = ['cc9063d80cae3ca87dd34586a92bac49613818a0689d9eac1bd91a799c5d0983']
+checksums = ['b0bb614dfbf36c286c8cad30912fe77359dccbf6b65a5edd1dde82af293f21fc']
 
-builddependencies = [
-    ('CMake', '3.31.8'),
-    ('Bison', '3.8.2'),
-]
+builddependencies = [('CMake', '3.31.3')]
+
+# --- GPU architecture (first-class framework param; NOT hardcoded) -------------
+# Default target is gfx1101 (V710). EESSI overrides per accel path via
+# --amdgcn-capabilities (EASYBUILD_AMDGCN_CAPABILITIES); the value flows into
+# configopts through the %(amdgcn_cc_space_sep)s template below.
+amdgcn_capabilities = ['gfx1101']
 
 dependencies = [
-    ('Python', '3.13.5'),
-    ('SciPy-bundle', '2025.07'),
+    ('Python', '3.13.1'),
+    ('SciPy-bundle', '2025.06'),
     ('Boost', '1.88.0'),
     ('METIS', '5.1.0'),
-    ('SCOTCH', '7.0.10'),
+    ('SCOTCH', '7.0.8'),
     ('MUMPS', '5.8.1', '-metis'),
-    ('SuiteSparse', '7.11.0'),
-    ('Hypre', '2.33.0'),
+    # NOTE (fp32-only delta): SuiteSparse (7.10.3) and Hypre (2.33.0) are DROPPED
+    # from the single-precision build because both are double-precision-only:
+    #   * SuiteSparse ships no single-precision UMFPACK/CHOLMOD/etc. PETSc aborts
+    #     configure with "Cannot use SuiteSparse with single, it is not available
+    #     in this precision".
+    #   * Hypre's HYPRE_config.h has HYPRE_SINGLE #undef (built for HYPRE_Real=double),
+    #     so PETSc single cannot link it.
+    # The EB_PETSc easyblock auto-injects each listed dependency via get_software_root()
+    # (--with-<dep>=1 --with-<dep>-dir=<root>), so the only clean, reviewable way to keep
+    # them out of an fp32 configure is to not list them here. MUMPS (libsmumps) and
+    # SuperLU_DIST (psgssvx symbols present) DO provide single precision, so they stay.
     ('ParMETIS', '4.0.3'),
     ('SuperLU_DIST', '9.1.0'),
     ('mpi4py', '4.1.0'),
-    ('HDF5', '1.14.6'),
-    ('hwloc', '2.12.1'),
-    ('netCDF', '4.9.3'),
-    ('PnetCDF', '1.14.1'),
-    ('zlib', '1.3.1'),
+    # --- AMD-GPU (HIP) stack, as shipped by EESSI 2025.06 ---
+    # These are built with the rocm-compilers/19.0.0 toolchain (NOT foss), so the
+    # dep toolchain MUST be given explicitly (4-element form), otherwise EasyBuild
+    # applies PETSc's foss-2025a toolchain and looks for a non-existent module.
+    # Resulting module names: HIP/6.4.1-rocm-compilers-19.0.0-ROCm-6.4.1, etc.
+    ('HIP', '6.4.1', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    ('hipBLAS', '2.4.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    ('hipSPARSE', '3.2.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # PETSc's hip.py liblist requires libhipsolver (grouped with rocrand) whenever
+    # the HIP backend is on; without it configure fails at 'ld: cannot find -lhipsolver'.
+    ('hipSOLVER', '2.4.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # rocRAND is grouped WITH hipsolver in PETSc's liblist; it must be a direct dep so
+    # its lib dir is on LIBRARY_PATH, else configure fails at 'ld: cannot find -lrocrand'.
+    ('rocRAND', '3.3.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # rocm-core provides <rocm-core/rocm_version.h>, which PETSc's hip.py version
+    # conftest #includes to emit PETSC_PKG_HIP_VERSION_* macros. Without it on CPATH,
+    # version detection silently fails and petscdevice_hip.h errors on an undefined
+    # PETSC_PKG_HIP_VERSION_GE, falling through to a non-existent bare <hipblas.h>.
+    # NOTE: built with the GCCcore-14.2.0 subtoolchain (NOT rocm-compilers), so the
+    # 3-element form resolves rocm-core/6.4.1-GCCcore-14.2.0-ROCm-6.4.1.
+    ('rocm-core', '6.4.1', '-ROCm-6.4.1'),
+    # rocThrust provides <thrust/complex.h>, hard-required by petscsystypes.h when the
+    # HIP device backend is on (compiling src/sys/memory/hip/mhiphost for gfx1101).
+    # EESSI 2025.06 ships no rocThrust module, so we build it (header-only, deps rocPRIM).
+    ('rocThrust', '3.3.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
 ]
 
+# Base (CPU) configure knobs from upstream PETSc-3.23.5-foss-2025a.eb ...
 configopts = '--LIBS="$LIBS -lrt" --with-mpi4py=0 '
+# ... single precision (fp32): scalars are 32-bit. This is the ONLY functional
+# delta from the default (fp64) HIP easyconfig.
+configopts += '--with-precision=single '
+# PETSc/foss auto-enables FFTW from the toolchain (foss bundles FFTW), but PETSc
+# refuses to use FFTW in single precision ("Cannot use fftw with single, it is not
+# available in this precision"), which aborts configure. FFTW is not needed by the
+# petsc4Foam linear-solver path, so disable it for the fp32 build.
+configopts += '--with-fftw=0 '
+# ... plus the AMD-GPU (HIP) delta. --with-hip-arch is fed the framework template
+# so the EESSI/CLI amdgcn override propagates (expands to "gfx1101" here). PETSc
+# locates hipcc/ROCm from the loaded HIP module (ROCM_PATH/PATH), not /opt/rocm.
+configopts += '--with-hip=1 --with-hip-arch=%(amdgcn_cc_space_sep)s '
+configopts += '--with-hipblas=1 --with-hipsparse=1 '
 
-shared_libs = 1
+# Match the proven -O3, no-debugging production settings from the manual build.
+buildopts = 'COPTFLAGS="-O3" CXXOPTFLAGS="-O3" HIPOPTFLAGS="-O3"'
 
-# only required when building PETSc in a SLURM job environment
-# configopts += '--with-batch=1 --known-mpi-shared-libraries=1 --known-64-bit-blas-indices=0 '
-# prebuildopts = "srun ./conftest-arch-linux2-c-opt && ./reconfigure-arch-linux2-c-opt.py && "
+shared_libs = 1
 
 moduleclass = 'numlib'

Diff against PETSc-3.24.0-foss-2025b.eb

easybuild/easyconfigs/p/PETSc/PETSc-3.24.0-foss-2025b.eb

diff --git a/easybuild/easyconfigs/p/PETSc/PETSc-3.24.0-foss-2025b.eb b/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a-HIP-fp32.eb
index d268dbc92c..2a97030e73 100644
--- a/easybuild/easyconfigs/p/PETSc/PETSc-3.24.0-foss-2025b.eb
+++ b/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a-HIP-fp32.eb
@@ -1,52 +1,125 @@
-# HDF5-support added
-# Author: J. Saßmannshausen (Imperial College London/UK)
+# EasyBuild easyconfig: PETSc with AMD-GPU (HIP) backend — SINGLE PRECISION (fp32)
+#
+# =========================================================================
+#  fp32 VARIANT of PETSc-3.23.5-foss-2025a-HIP.eb.
+#  IDENTICAL to the (default, double-precision / fp64) HIP easyconfig in
+#  every respect EXCEPT:
+#    * versionsuffix   '-HIP'  ->  '-HIP-fp32'
+#    * configopts adds  --with-precision=single
+#  See PETSc-3.23.5-foss-2025a-HIP.eb for the full rationale on the GPU-arch
+#  parameterisation and the ROCm dependency reconciliation — all of that is
+#  unchanged here. Only the scalar precision differs.
+#
+#  WHY A SEPARATE MODULE:
+#    PETSc's scalar precision (double vs single) is fixed at configure time and
+#    baked into libpetsc + the installed headers. fp32 and fp64 cannot coexist
+#    in one install, so the fp32 build gets its own module
+#    (PETSc/3.23.5-foss-2025a-HIP-fp32) and its own petsc4Foam adapter linked
+#    against it (petsc4Foam/v2506-foss-2025a-HIP-fp32).
+# =========================================================================
+
+easyblock = 'EB_PETSc'
 
 name = 'PETSc'
-version = '3.24.0'
+version = '3.23.5'
+# fp32 single-precision HIP variant. Arch (gfxNNN) still lives in the EESSI accel
+# path, not the suffix; the suffix marks (HIP backend + single precision).
+versionsuffix = '-HIP-fp32'
 
 homepage = 'https://www.mcs.anl.gov/petsc'
 description = """PETSc, pronounced PET-see (the S is silent), is a suite of data structures and routines for the
- scalable (parallel) solution of scientific applications modeled by partial differential equations."""
+ scalable (parallel) solution of scientific applications modeled by partial differential equations.
+
+This variant enables the AMD-GPU (HIP) backend with hipBLAS and hipSPARSE, built in SINGLE precision
+(--with-precision=single, fp32), so GPU-accelerated Krylov solvers/preconditioners run in 32-bit
+floating point (e.g. for the petsc4Foam / OpenFOAM external-solver adapter). The target GPU
+architecture is taken from the amdgcn_capabilities parameter (default gfx1101, AMD Radeon Pro V710),
+overridable via --amdgcn-capabilities."""
 
-toolchain = {'name': 'foss', 'version': '2025b'}
+toolchain = {'name': 'foss', 'version': '2025a'}
 toolchainopts = {'openmp': True, 'usempi': True, 'pic': True}
 
 source_urls = [
     'https://web.cels.anl.gov/projects/petsc/download/release-snapshots',
 ]
 sources = [SOURCELOWER_TAR_GZ]
-checksums = ['cc9063d80cae3ca87dd34586a92bac49613818a0689d9eac1bd91a799c5d0983']
+checksums = ['b0bb614dfbf36c286c8cad30912fe77359dccbf6b65a5edd1dde82af293f21fc']
 
-builddependencies = [
-    ('CMake', '3.31.8'),
-    ('Bison', '3.8.2'),
-]
+builddependencies = [('CMake', '3.31.3')]
+
+# --- GPU architecture (first-class framework param; NOT hardcoded) -------------
+# Default target is gfx1101 (V710). EESSI overrides per accel path via
+# --amdgcn-capabilities (EASYBUILD_AMDGCN_CAPABILITIES); the value flows into
+# configopts through the %(amdgcn_cc_space_sep)s template below.
+amdgcn_capabilities = ['gfx1101']
 
 dependencies = [
-    ('Python', '3.13.5'),
-    ('SciPy-bundle', '2025.07'),
+    ('Python', '3.13.1'),
+    ('SciPy-bundle', '2025.06'),
     ('Boost', '1.88.0'),
     ('METIS', '5.1.0'),
-    ('SCOTCH', '7.0.10'),
+    ('SCOTCH', '7.0.8'),
     ('MUMPS', '5.8.1', '-metis'),
-    ('SuiteSparse', '7.11.0'),
-    ('Hypre', '2.33.0'),
+    # NOTE (fp32-only delta): SuiteSparse (7.10.3) and Hypre (2.33.0) are DROPPED
+    # from the single-precision build because both are double-precision-only:
+    #   * SuiteSparse ships no single-precision UMFPACK/CHOLMOD/etc. PETSc aborts
+    #     configure with "Cannot use SuiteSparse with single, it is not available
+    #     in this precision".
+    #   * Hypre's HYPRE_config.h has HYPRE_SINGLE #undef (built for HYPRE_Real=double),
+    #     so PETSc single cannot link it.
+    # The EB_PETSc easyblock auto-injects each listed dependency via get_software_root()
+    # (--with-<dep>=1 --with-<dep>-dir=<root>), so the only clean, reviewable way to keep
+    # them out of an fp32 configure is to not list them here. MUMPS (libsmumps) and
+    # SuperLU_DIST (psgssvx symbols present) DO provide single precision, so they stay.
     ('ParMETIS', '4.0.3'),
     ('SuperLU_DIST', '9.1.0'),
     ('mpi4py', '4.1.0'),
-    ('HDF5', '1.14.6'),
-    ('hwloc', '2.12.1'),
-    ('netCDF', '4.9.3'),
-    ('PnetCDF', '1.14.1'),
-    ('zlib', '1.3.1'),
+    # --- AMD-GPU (HIP) stack, as shipped by EESSI 2025.06 ---
+    # These are built with the rocm-compilers/19.0.0 toolchain (NOT foss), so the
+    # dep toolchain MUST be given explicitly (4-element form), otherwise EasyBuild
+    # applies PETSc's foss-2025a toolchain and looks for a non-existent module.
+    # Resulting module names: HIP/6.4.1-rocm-compilers-19.0.0-ROCm-6.4.1, etc.
+    ('HIP', '6.4.1', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    ('hipBLAS', '2.4.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    ('hipSPARSE', '3.2.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # PETSc's hip.py liblist requires libhipsolver (grouped with rocrand) whenever
+    # the HIP backend is on; without it configure fails at 'ld: cannot find -lhipsolver'.
+    ('hipSOLVER', '2.4.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # rocRAND is grouped WITH hipsolver in PETSc's liblist; it must be a direct dep so
+    # its lib dir is on LIBRARY_PATH, else configure fails at 'ld: cannot find -lrocrand'.
+    ('rocRAND', '3.3.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # rocm-core provides <rocm-core/rocm_version.h>, which PETSc's hip.py version
+    # conftest #includes to emit PETSC_PKG_HIP_VERSION_* macros. Without it on CPATH,
+    # version detection silently fails and petscdevice_hip.h errors on an undefined
+    # PETSC_PKG_HIP_VERSION_GE, falling through to a non-existent bare <hipblas.h>.
+    # NOTE: built with the GCCcore-14.2.0 subtoolchain (NOT rocm-compilers), so the
+    # 3-element form resolves rocm-core/6.4.1-GCCcore-14.2.0-ROCm-6.4.1.
+    ('rocm-core', '6.4.1', '-ROCm-6.4.1'),
+    # rocThrust provides <thrust/complex.h>, hard-required by petscsystypes.h when the
+    # HIP device backend is on (compiling src/sys/memory/hip/mhiphost for gfx1101).
+    # EESSI 2025.06 ships no rocThrust module, so we build it (header-only, deps rocPRIM).
+    ('rocThrust', '3.3.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
 ]
 
+# Base (CPU) configure knobs from upstream PETSc-3.23.5-foss-2025a.eb ...
 configopts = '--LIBS="$LIBS -lrt" --with-mpi4py=0 '
+# ... single precision (fp32): scalars are 32-bit. This is the ONLY functional
+# delta from the default (fp64) HIP easyconfig.
+configopts += '--with-precision=single '
+# PETSc/foss auto-enables FFTW from the toolchain (foss bundles FFTW), but PETSc
+# refuses to use FFTW in single precision ("Cannot use fftw with single, it is not
+# available in this precision"), which aborts configure. FFTW is not needed by the
+# petsc4Foam linear-solver path, so disable it for the fp32 build.
+configopts += '--with-fftw=0 '
+# ... plus the AMD-GPU (HIP) delta. --with-hip-arch is fed the framework template
+# so the EESSI/CLI amdgcn override propagates (expands to "gfx1101" here). PETSc
+# locates hipcc/ROCm from the loaded HIP module (ROCM_PATH/PATH), not /opt/rocm.
+configopts += '--with-hip=1 --with-hip-arch=%(amdgcn_cc_space_sep)s '
+configopts += '--with-hipblas=1 --with-hipsparse=1 '
 
-shared_libs = 1
+# Match the proven -O3, no-debugging production settings from the manual build.
+buildopts = 'COPTFLAGS="-O3" CXXOPTFLAGS="-O3" HIPOPTFLAGS="-O3"'
 
-# only required when building PETSc in a SLURM job environment
-# configopts += '--with-batch=1 --known-mpi-shared-libraries=1 --known-64-bit-blas-indices=0 '
-# prebuildopts = "srun ./conftest-arch-linux2-c-opt && ./reconfigure-arch-linux2-c-opt.py && "
+shared_libs = 1
 
 moduleclass = 'numlib'

Diff against PETSc-3.23.5-foss-2025a.eb

easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a.eb

diff --git a/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a.eb b/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a-HIP-fp32.eb
index bd28197d86..2a97030e73 100644
--- a/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a.eb
+++ b/easybuild/easyconfigs/p/PETSc/PETSc-3.23.5-foss-2025a-HIP-fp32.eb
@@ -1,9 +1,40 @@
+# EasyBuild easyconfig: PETSc with AMD-GPU (HIP) backend — SINGLE PRECISION (fp32)
+#
+# =========================================================================
+#  fp32 VARIANT of PETSc-3.23.5-foss-2025a-HIP.eb.
+#  IDENTICAL to the (default, double-precision / fp64) HIP easyconfig in
+#  every respect EXCEPT:
+#    * versionsuffix   '-HIP'  ->  '-HIP-fp32'
+#    * configopts adds  --with-precision=single
+#  See PETSc-3.23.5-foss-2025a-HIP.eb for the full rationale on the GPU-arch
+#  parameterisation and the ROCm dependency reconciliation — all of that is
+#  unchanged here. Only the scalar precision differs.
+#
+#  WHY A SEPARATE MODULE:
+#    PETSc's scalar precision (double vs single) is fixed at configure time and
+#    baked into libpetsc + the installed headers. fp32 and fp64 cannot coexist
+#    in one install, so the fp32 build gets its own module
+#    (PETSc/3.23.5-foss-2025a-HIP-fp32) and its own petsc4Foam adapter linked
+#    against it (petsc4Foam/v2506-foss-2025a-HIP-fp32).
+# =========================================================================
+
+easyblock = 'EB_PETSc'
+
 name = 'PETSc'
 version = '3.23.5'
+# fp32 single-precision HIP variant. Arch (gfxNNN) still lives in the EESSI accel
+# path, not the suffix; the suffix marks (HIP backend + single precision).
+versionsuffix = '-HIP-fp32'
 
 homepage = 'https://www.mcs.anl.gov/petsc'
 description = """PETSc, pronounced PET-see (the S is silent), is a suite of data structures and routines for the
- scalable (parallel) solution of scientific applications modeled by partial differential equations."""
+ scalable (parallel) solution of scientific applications modeled by partial differential equations.
+
+This variant enables the AMD-GPU (HIP) backend with hipBLAS and hipSPARSE, built in SINGLE precision
+(--with-precision=single, fp32), so GPU-accelerated Krylov solvers/preconditioners run in 32-bit
+floating point (e.g. for the petsc4Foam / OpenFOAM external-solver adapter). The target GPU
+architecture is taken from the amdgcn_capabilities parameter (default gfx1101, AMD Radeon Pro V710),
+overridable via --amdgcn-capabilities."""
 
 toolchain = {'name': 'foss', 'version': '2025a'}
 toolchainopts = {'openmp': True, 'usempi': True, 'pic': True}
@@ -16,6 +47,12 @@ checksums = ['b0bb614dfbf36c286c8cad30912fe77359dccbf6b65a5edd1dde82af293f21fc']
 
 builddependencies = [('CMake', '3.31.3')]
 
+# --- GPU architecture (first-class framework param; NOT hardcoded) -------------
+# Default target is gfx1101 (V710). EESSI overrides per accel path via
+# --amdgcn-capabilities (EASYBUILD_AMDGCN_CAPABILITIES); the value flows into
+# configopts through the %(amdgcn_cc_space_sep)s template below.
+amdgcn_capabilities = ['gfx1101']
+
 dependencies = [
     ('Python', '3.13.1'),
     ('SciPy-bundle', '2025.06'),
@@ -23,19 +60,66 @@ dependencies = [
     ('METIS', '5.1.0'),
     ('SCOTCH', '7.0.8'),
     ('MUMPS', '5.8.1', '-metis'),
-    ('SuiteSparse', '7.10.3'),
-    ('Hypre', '2.33.0'),
+    # NOTE (fp32-only delta): SuiteSparse (7.10.3) and Hypre (2.33.0) are DROPPED
+    # from the single-precision build because both are double-precision-only:
+    #   * SuiteSparse ships no single-precision UMFPACK/CHOLMOD/etc. PETSc aborts
+    #     configure with "Cannot use SuiteSparse with single, it is not available
+    #     in this precision".
+    #   * Hypre's HYPRE_config.h has HYPRE_SINGLE #undef (built for HYPRE_Real=double),
+    #     so PETSc single cannot link it.
+    # The EB_PETSc easyblock auto-injects each listed dependency via get_software_root()
+    # (--with-<dep>=1 --with-<dep>-dir=<root>), so the only clean, reviewable way to keep
+    # them out of an fp32 configure is to not list them here. MUMPS (libsmumps) and
+    # SuperLU_DIST (psgssvx symbols present) DO provide single precision, so they stay.
     ('ParMETIS', '4.0.3'),
     ('SuperLU_DIST', '9.1.0'),
     ('mpi4py', '4.1.0'),
+    # --- AMD-GPU (HIP) stack, as shipped by EESSI 2025.06 ---
+    # These are built with the rocm-compilers/19.0.0 toolchain (NOT foss), so the
+    # dep toolchain MUST be given explicitly (4-element form), otherwise EasyBuild
+    # applies PETSc's foss-2025a toolchain and looks for a non-existent module.
+    # Resulting module names: HIP/6.4.1-rocm-compilers-19.0.0-ROCm-6.4.1, etc.
+    ('HIP', '6.4.1', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    ('hipBLAS', '2.4.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    ('hipSPARSE', '3.2.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # PETSc's hip.py liblist requires libhipsolver (grouped with rocrand) whenever
+    # the HIP backend is on; without it configure fails at 'ld: cannot find -lhipsolver'.
+    ('hipSOLVER', '2.4.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # rocRAND is grouped WITH hipsolver in PETSc's liblist; it must be a direct dep so
+    # its lib dir is on LIBRARY_PATH, else configure fails at 'ld: cannot find -lrocrand'.
+    ('rocRAND', '3.3.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
+    # rocm-core provides <rocm-core/rocm_version.h>, which PETSc's hip.py version
+    # conftest #includes to emit PETSC_PKG_HIP_VERSION_* macros. Without it on CPATH,
+    # version detection silently fails and petscdevice_hip.h errors on an undefined
+    # PETSC_PKG_HIP_VERSION_GE, falling through to a non-existent bare <hipblas.h>.
+    # NOTE: built with the GCCcore-14.2.0 subtoolchain (NOT rocm-compilers), so the
+    # 3-element form resolves rocm-core/6.4.1-GCCcore-14.2.0-ROCm-6.4.1.
+    ('rocm-core', '6.4.1', '-ROCm-6.4.1'),
+    # rocThrust provides <thrust/complex.h>, hard-required by petscsystypes.h when the
+    # HIP device backend is on (compiling src/sys/memory/hip/mhiphost for gfx1101).
+    # EESSI 2025.06 ships no rocThrust module, so we build it (header-only, deps rocPRIM).
+    ('rocThrust', '3.3.0', '-ROCm-6.4.1', ('rocm-compilers', '19.0.0')),
 ]
 
+# Base (CPU) configure knobs from upstream PETSc-3.23.5-foss-2025a.eb ...
 configopts = '--LIBS="$LIBS -lrt" --with-mpi4py=0 '
+# ... single precision (fp32): scalars are 32-bit. This is the ONLY functional
+# delta from the default (fp64) HIP easyconfig.
+configopts += '--with-precision=single '
+# PETSc/foss auto-enables FFTW from the toolchain (foss bundles FFTW), but PETSc
+# refuses to use FFTW in single precision ("Cannot use fftw with single, it is not
+# available in this precision"), which aborts configure. FFTW is not needed by the
+# petsc4Foam linear-solver path, so disable it for the fp32 build.
+configopts += '--with-fftw=0 '
+# ... plus the AMD-GPU (HIP) delta. --with-hip-arch is fed the framework template
+# so the EESSI/CLI amdgcn override propagates (expands to "gfx1101" here). PETSc
+# locates hipcc/ROCm from the loaded HIP module (ROCM_PATH/PATH), not /opt/rocm.
+configopts += '--with-hip=1 --with-hip-arch=%(amdgcn_cc_space_sep)s '
+configopts += '--with-hipblas=1 --with-hipsparse=1 '
 
-shared_libs = 1
+# Match the proven -O3, no-debugging production settings from the manual build.
+buildopts = 'COPTFLAGS="-O3" CXXOPTFLAGS="-O3" HIPOPTFLAGS="-O3"'
 
-# only required when building PETSc in a SLURM job environment
-# configopts += '--with-batch=1 --known-mpi-shared-libraries=1 --known-64-bit-blas-indices=0 '
-# prebuildopts = "srun ./conftest-arch-linux2-c-opt && ./reconfigure-arch-linux2-c-opt.py && "
+shared_libs = 1
 
 moduleclass = 'numlib'

github-actions Bot added new update 2025a issues & PRs related to 2025a common toolchains labels Jul 31, 2026

Navigation Menu

Uh oh!

{OpenFOAM,PETSc,petsc4Foam}: single-precision (HIP) OpenFOAM GPU-solver chain for foss/2025a - #26578

{OpenFOAM,PETSc,petsc4Foam}: single-precision (HIP) OpenFOAM GPU-solver chain for foss/2025a#26578
hmeiland wants to merge 1 commit into
easybuilders:developfrom
hmeiland:petsc4foam-hip-sp

hmeiland commented Jul 31, 2026

Uh oh!

github-actions Bot commented Jul 31, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

hmeiland commented Jul 31, 2026

What

Why single precision

Dependency on the easyblocks repo

Target

Status

Uh oh!

github-actions Bot commented Jul 31, 2026

Updated software OpenFOAM-v2506-foss-2025a-SP.eb

Updated software PETSc-3.23.5-foss-2025a-HIP-fp32.eb

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Updated software `OpenFOAM-v2506-foss-2025a-SP.eb`

Updated software `PETSc-3.23.5-foss-2025a-HIP-fp32.eb`