[platelett]

./mfc.sh build --mixed (half-precision storage stp, single-precision working wp) does not compile on master with NVHPC. Once it does, runs on two or more ranks with parallel_io = 'T' write corrupted initial-condition and restart files. This PR fixes both.

Non-mixed builds are unaffected. There stp == wp, mpi_io_p == mpi_p and mpi_io_type == 1, so every change below is a no-op. The Cray --mixed failure in #1965 happens earlier, in m_derived_types, and is not addressed here.

1. The build

On master, ./mfc.sh build --mixed --no-gpu with NVHPC 26.3 fails in simulation:

NVFORTRAN-S-0155-Illegal POINTER assignment - type mismatch           (m_igr.fpp)
NVFORTRAN-S-0450-Argument number 1 to f_compute_shear_rate_from_components: kind mismatch   (m_viscous.fpp)

--gpu acc also fails on f_bulk_modulus and f_advance_step. The fixes:

• real(stp) field values passed to real(wp) dummies.
  • Where: m_viscous, m_bubbles_EE, m_acoustic_src, and post_process/m_start_up (s_phase_temperature).
  • Fix: they are converted with real(..., kind=wp) at the call site. The dummies are intent(in), so nothing is written back.
  • Operands first: where an argument combines two field values (the off-diagonal shear rates, and the phase density alpha_rho/max(alpha, sgm_eps)), each operand is converted first. Otherwise the sum, quotient and max would run in half precision, and max would mix stp with the wp constant sgm_eps.
• jac in m_igr is declared real(wp), but it is exposed through a scalar_field pointer, which is real(stp). Every assignment to jac already converts to stp, so it is now declared real(stp). jac_rhs and jac_old stay wp.
• integer_field stays a default integer. It used to drop to integer(kind=1) under --mixed (since d290396), but its users need a default integer:
  • s_decode_patch_periodicity takes a default integer. This is the error left in m_ibm.fpp once the real-kind fixes are in.
  • The IB marker MPI-IO view is built from MPI_INTEGER.
  • ib_markers is updated with a GPU_ATOMIC max.
  • An encoded periodic IB id, gbl_id + (num_gbl_ibs+1)*(xp + 3*yp + 9*zp), exceeds 127 with only a few bodies.

2. Parallel I/O

Under --mixed a stored cell is 2 bytes (mpi_io_p = MPI_BYTE, mpi_io_type = 2). The field file views did not follow that:

• s_initialize_mpi_data and s_initialize_mpi_data_ds built each field's subarray from mpi_p, which is 4 bytes.
• The MPI_FILE_SET_VIEW calls in pre_process/m_start_up, simulation/m_data_output and post_process/m_data_input used mpi_p as the etype, while the transfer right after uses mpi_io_p. This breaks the Common Pitfalls rule in contributing.md: "mpi_io_p must match stp".
• pre_process/m_start_up read data_size elements of mpi_p into a buffer of data_size 2-byte cells.

On one rank the layout still comes out right by accident. On two or more, each rank's block lands at the wrong offset.

The fix:

• Subarrays are now built from a 2-byte contiguous cell type, in s_create_mpi_field_subarray, a private helper in m_mpi_common.
• Every view uses mpi_io_p as its etype.
• The pre_process reader reads data_size*mpi_io_type elements of mpi_io_p.

The remaining mpi_p views (Lagrange bubbles and the post_process output) describe real(wp) buffers and are correct as they are.

toolchain/mfc/test_mixed_mpi_views.py:

• What it does: compiles the production s_initialize_mpi_data and every production field SET_VIEW/transfer pair against 2-byte opaque cells, so it needs no REAL(2) support. It writes from 2 ranks and compares the result with a serially written file.
• Results: it fails on master and passes with this PR.
• When it skips: without an MPI compiler, and under MFC_SKIP_COMPILER_TESTS=1, like the other compiler tests.

Side effects

• Non-mixed builds: no change. The conversions are no-ops, and the helper builds exactly the old subarray. No goldens change.
• --mixed builds:
  • ib_markers and bc_type take 4 bytes per cell instead of 1.
  • IB-marker files written by an older --mixed build cannot be read back. NVHPC could not produce such a build on master.
  • jac (IGR) is stored in stp, which is what every assignment to it already produced.
• Overlap with SCC 26: bug: fix parallel I/O failures and performance overhead #2085 (%sf → %sf(:,:,:) in the same MPI-IO calls), in pre_process/m_start_up, simulation/m_data_output and post_process/m_data_input:
  • The conflict is textual, and the resolution keeps both changes.
  • On that merge the new test passes. On SCC 26: bug: fix parallel I/O failures and performance overhead #2085 alone it fails, with 774 mismatched bytes.
  • The test accepts either spelling.
• Overlap with SCC26: Use nonblocking MPI for halo exchange #2090: one private :: line in m_mpi_common, keep both.

Verification

Hardware: Intel Core i7-14700F, RTX 4060 Ti (sm_89). Compiler: NVHPC 26.3 with its bundled HPC-X MPI, built in clean build directories.

Build:

Build	master	this PR
--mixed --no-gpu, pre/sim/post	simulation fails (above)	builds
--mixed --gpu acc, pre/sim/post	simulation fails (also f_bulk_modulus, f_advance_step)	builds

Parallel I/O:

• Case: examples/2D_riemann_test at 64², 20 steps, parallel_io = 'T', --mixed.
• Comparison: the 1-rank and 2-rank restart_data/lustre_<step>.dat files, byte for byte. Each is 40,960 bytes (64·64·5 fields·2 bytes).

	lustre_0 (initial condition)	lustre_10	lustre_20
build fixes only, without the view fix: 1 rank vs 2 ranks	differ: 3,272 of 20,480 values; density off by up to 0.97 (range 0.14–1.5)	differ	differ
this PR, CPU: 1 rank vs 2 ranks	identical	identical	identical
this PR, GPU (--gpu acc): 1 rank vs 2 ranks	identical	—	identical

Density at step 20, 2 ranks, against a double-precision run of the same case:

--mixed	Max relative difference
this PR, CPU	1.1e-3 (half-precision storage)
this PR, GPU	1.1e-3
build fixes only, without the view fix, CPU	6.5e-1

The final code writes restart files byte-identical to the runs above, on CPU and GPU. This inviscid case does not reach the operand-first conversions, which were added after those runs.

The operand-first conversions:

Check	Result
post_process temperature (T_wrt = 'T' on the case above), --mixed, CPU and GPU, against double	max relative difference 1.0e-3 (initial condition) and 2.5e-3 (step 20), the same level as density and pressure
Same case, 1 rank vs 2 ranks, CPU and GPU	restart files identical
examples/2D_poiseuille_nn (non-Newtonian), --mixed, CPU, 2,730 steps	runs, all fields finite. Its velocities (~1e-5) are below half precision's normal range, so it cannot be compared meaningfully with double
simulation, --mixed --gpu acc --unified (the UVM jac_host path)	builds

Tests:

• test_mixed_mpi_views.py: fails on master. Passes with this PR on Open MPI 4.1.6 + gfortran 13.3 and on HPC-X + nvfortran 26.3. Also passes merged with SCC 26: bug: fix parallel I/O failures and performance overhead #2085.
• Full regression suite, non-mixed: final code on current master ee99151b, GNU 13.3 + Open MPI 4.1.6, CPU: 725 passed, 0 failed (39 convergence cases skipped by default). No golden changes.
• Lint gate: ./mfc.sh lint, including the compiler tests, passes.

Contribution Policy

All contributions are expected to demonstrate:

• A clear understanding of the codebase
• Alignment with product direction
• Thoughtful reasoning behind changes
• Evidence of real-world usage or hands-on experience with the problem

If these expectations are not met, we would prefer to implement the changes ourselves rather than spend time reviewing low-effort submissions.

Acknowledgement

• I confirm this PR meets the above expectations and reflects my own understanding and real-world context.

---

This PR is from Diablo (Tsinghua University) participating in the SC26 Student Cluster Competition.

[github-actions]

Lines of Code

File	Lines	Diff
src/common/m_mpi_common.fpp	1503	+9
src/common/m_derived_types.fpp	479	-4
src/simulation/m_viscous.fpp	1023	+2
src/simulation/m_igr.fpp	2290	+1

Directory	Lines	Diff
common	10541	+5
simulation	28566	+3
total	47640	+8

[github-actions]

Claude Code Review

Head SHA: 910f07b

Files changed:

• 11
• src/common/m_derived_types.fpp
• src/common/m_mpi_common.fpp
• src/post_process/m_data_input.f90
• src/post_process/m_start_up.fpp
• src/pre_process/m_start_up.fpp
• src/simulation/m_acoustic_src.fpp
• src/simulation/m_bubbles_EE.fpp
• src/simulation/m_data_output.fpp
• src/simulation/m_igr.fpp
• src/simulation/m_viscous.fpp

Findings:
No issues found. Checked the changed hunks in m_mpi_common, m_data_input, both m_start_up files, m_acoustic_src, m_bubbles_EE, m_viscous and m_derived_types. The m_igr jac declaration change could not be checked against the other jac uses, because m_igr has no file in the context directory.

[codecov]

Codecov Report

❌ Patch coverage is 65.21739% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 61.96%. Comparing base (ee99151) to head (910f07b).

Files with missing lines	Patch %	Lines
src/pre_process/m_start_up.fpp	0.00%	4 Missing ⚠️
src/simulation/m_viscous.fpp	0.00%	2 Missing ⚠️
src/common/m_mpi_common.fpp	87.50%	1 Missing ⚠️
src/post_process/m_start_up.fpp	0.00%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2110   +/-   ##
=======================================
  Coverage   61.95%   61.96%           
=======================================
  Files          86       86           
  Lines       22906    22908    +2     
  Branches     3378     3378           
=======================================
+ Hits        14192    14195    +3     
+ Misses       6206     6205    -1     
  Partials     2508     2508           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.