Supply model sizes through ABI for all provider paths by Tsilimidos · Pull Request #33 · MIT-PSAAP-IV/Exasim

Tsilimidos · 2026-07-03T01:10:24Z

Summary

Remove the requirement to specify model dimension sizes (ncu, ncv, ncw, nsca, nvec, nten, nsurf, nvqoi) in pdeapp.txt by having every ABI provider path supply them at compile time.

Changes

Core solver (`ExasimSolver.cpp`)

Remove the builtinmodelID > 0 guard so the ABI size query runs unconditionally for every provider.

Provider paths — all 6 now supply compile-time sizes

Provider	Mechanism
KokkosKernel (reference)	`PdeModel::ncu` etc. from `kokkoskernels.hpp`
BuiltInLibrary	`builtinGetModelSizes()` callback (unchanged)
UserDefined	`PdeModel::ncu` etc. from `my_model.hpp`
Text2codeGenerated	`PdeModel::ncu` etc. from generated `my_model.hpp`
SharedLibrary (libt2cmodel)	`PdeModel::ncu` etc. from generated `my_model.hpp`
FrontendGenerated	`model_sizes.hpp` via `__has_include` guard

Frontend generators — emit `model_sizes.hpp`

Python (Gencode.py): writes model_sizes.hpp into staging dir before sync
MATLAB (kkgencode.m): same
Julia (Gencode.jl): same

The file is synced alongside kernel .cpp files to kernels/ and is picked up by frontendprovider.cpp via __has_include("model_sizes.hpp"). When absent (e.g. exported bundles without re-running codegen), sizes default to 0 and the solver falls back to pdeapp.txt values.

Templates & examples

cmake/frontend-app/frontendprovider.cpp.in: __has_include guard
examples/exasimfe/frontendprovider.cpp: same

Verification

All 15 ctest tests pass (100%)
Manually verified SharedLibrary path with ncu/ncw stripped from pdeapp.txt — solver gets sizes from the .so ABI, correct QoI
consumer_builtin_cpu already tests the BuiltInLibrary path without sizes in pdeapp.txt

Relevant files

 backend/Main/ExasimSolver.cpp                          |  6 +--
 backend/Model/FrontendGenerated/frontendprovider.cpp   | 22 ++++++++++
 backend/Model/Text2codeGenerated/libt2cmodel.cpp       | 17 ++++++++
 backend/Model/Text2codeGenerated/text2codeprovider.cpp | 17 ++++++++
 backend/Model/UserDefined/userdefinedprovider.cpp      | 18 ++++++++-
 cmake/frontend-app/frontendprovider.cpp.in             | 22 ++++++++++
 examples/exasimfe/frontendprovider.cpp                 | 22 ++++++++++
 frontends/Julia/Exasim/src/Gencode/Gencode.jl          | 20 ++++++++++
 frontends/Matlab/Gencode/kkgencode.m                   | 27 ++++++++++++++
 frontends/Python/exasim/Gencode/Gencode.py             | 32 ++++++++++++++++
 10 files changed, 199 insertions(+), 4 deletions(-)

ModelDefaults: add nsca/nvec/nten/nsurf/nvqoi as static constexpr (default 0). ExasimDriverABI: add ncu/nco/ncw/nsca/nvec/nten/nsurf/nvqoi fields; bump version to 2. modelprovider.hpp + kokkoskernelprovider.hpp: populate ABI size fields from PdeModel.

PDE struct: change ncu default from 1 to 0 (unset). validateBuiltInAppMetadata: accept ncu >= 0 (allow 0 = not yet specified). Add applyBuiltInABIMetadata(PDE&, const ExasimDriverABI&) to fill sizes from ABI. ExasimSolver::ParseInputs(): query SelectExasimDriverABI() for missing dimension sizes; use programmatic CPreprocessing constructor so the ABI-filled PDE values flow into writeBinaryFiles.

ncu/ncv/ncw/nsca/nvec/nten/nsurf/nvqoi removed from pdeapp.txt — they are now provided at runtime by the compiled model's ABI. Add corresponding static constexpr to PdeModel struct so the ABI can supply them during preprocessing.

pdeapp.txt (53 files): remove ncu/ncv/ncw/nsca/nvec/nten/nsurf/nvqoi lines my_model.hpp (11 files): keep only non-zero nsca/nvec/nsurf/nvqoi decls (zero-valued inherit from ModelDefaults) ExasimSolver.cpp: add missing nsurf_/nvqoi_ propagation from pde text2code CodeGenerator.cpp: emit vis/QoI constexprs from parsed specs

…d by pdemodel.txt output_size(s)=4)

Backward-compatible size resolution: the KokkosKernel provider ABI populates size fields from PdeModel::ncu etc., while the BuiltInLibrary ABI cannot (it returns 0 for all model-dependent sizes). The >0 guard on all ABI fields prevents BuiltInLibrary's zeros from overwriting the defaults, while still allowing KokkosKernel's real values through.

Add nsca/nvec/nten/nsurf/nvqoi fields to ExasimDriverABI (version 3) with ModelSizes struct and GetModelSizes function pointer. KokkosKernel and BuiltInLibrary providers populate these from compile-time constexpr values; ExternalModel provider dispatches via per-model size TUs. Remove the 8 size parameters (ncu/ncv/ncw/nsca/nvec/nten/nsurf/nvqoi) from all 53 pdeapp.txt files — they are now sourced from the ABI when missing, with explicit pdeapp.txt values still taking precedence. BuiltInLibrary: auto-generate builtin_model_sizes.cpp at CMake time by parsing pdeapp*.txt + pdemodel*.txt; provides per-model dispatch via builtinGetModelSizes(int modelnumber). ExternalModel: Each model gets its own _model_sizes_.cpp TU via file(GENERATE) to avoid #pragma once collisions; extGetModelSizes dispatches via extern "C" per-model helpers. Fix shared library provider (builtinmodelID == 0): parse pdemodel.txt during programmatic preprocessing so applyParsedSpecMetadata extracts vis/QoI sizes from function output_size annotations. Fix struct_size checks: builtinlibprovider uses == (provider and consumer compiled together); sharedlibprovider/t2clibprovider use >= (cross-library compilation).

- Remove builtinmodelID guard in ExasimSolver.cpp:600 so ABI size query runs unconditionally for every provider. - UserDefined, Text2codeGenerated, SharedLibrary (libt2cmodel): set ABI size fields from PdeModel:: compile-time constants. - FrontendGenerated (frontendprovider.cpp, CMake template, example): read sizes from model_sizes.hpp via __has_include guard. - Frontend generators (Python, MATLAB, Julia): emit model_sizes.hpp into staging dir before sync, so FrontendGenerated provider has compile-time size constants at build time.

The modern frontend path (Python/Matlab/Julia -> cmakecompile.py -> ExternalModelProvider.cpp.in) was not consuming model_sizes.hpp, making the generated file dead code on the most common user path. - Add __has_include("model_sizes.hpp") guard - Populate kFrontendModelSizes constant and extGetModelSizes() fallback for model IDs >= 100 (frontend-generated range) - Set direct ABI size fields for redundancy The check is placed inside the #else branch of @EXT_MODEL_HAS_MY_MODEL@ so it is never compiled for the text2code (PDEMODEL) path.

…(shared library models, sizes come from .so ABI)

…ibility" This reverts commit a31101e.

Replace fragile CMake-time regex parsing of pdeapp.txt for ncu/nco/ncw and vis/QoI sizes. text2code now emits a namespaced model_sizes.hpp per model with compile-time constants extracted from the full pdemodel parser (vectors uhat, v, w and functions VisScalars, VisVectors, ...). CMakeLists.txt changes: - Removed ~90 lines of configure-time size parsing and fallback logic. - builtin_model_sizes.cpp now wraps each per-model #include "modelN/model_sizes.hpp" in namespace builtin_model_N {}, using the folder name to infer the namespace. - Removed unused EXASIM_BM_SIZES property. text2code changes: - CodeGenerator::generateModelSizesHpp follows the same pattern as generateEmpty*Cpp methods: writes a simple header from parsed spec vectors and function output sizes. - Called from text2code.cpp alongside the existing codegen call. pdeapp{1..15}.txt: removed ncu=, ncw= lines (now sourced from pdemodel.txt via text2code).

so text2code always generates a fresh file.

…e stamp GCC 14.3.0 deduplicates #pragma once headers by content fingerprint, not inode/path. Three groups of models (1/8, 2/3/15, 10/14) have identical model_sizes.hpp content, causing GCC to silently skip the second include and leave builtin_model_N::ncu etc. undefined for N in {3,8,14,15}. Fix: drop #pragma once from generateModelSizesHpp output. Each file is included exactly once inside a unique namespace block in builtin_model_sizes.cpp, so no include guard is needed. Also remove .text2code.stamp alongside model_sizes.hpp at configure time so regeneration is triggered on the next build after a stale-file cleanup.

Tsilimidos added 19 commits July 2, 2026 09:45

update gitignore

7e08251

Fix nsmach8 nsca: 5 -> 4 (vis_scalars writes 4 outputs)

2d775a9

Fix isoq/sharpb2 nsca: 7 -> 4 (vis_scalars writes 4 outputs, confirme…

dfb8b3f

…d by pdemodel.txt output_size(s)=4)

merge

02e4e58

remove orphaned nco key from reactingsharpb2 pdeapp.txt

6fdb79c

remove redundant ncu/ncv/ncw from isoq2d_sweep and isoq3d pdeapp.txt …

b6b427a

…(shared library models, sizes come from .so ABI)

add --break-system-packages to pip install for PEP 668 compatibility

a31101e

Revert "add --break-system-packages to pip install for PEP 668 compat…

7ebada7

…ibility" This reverts commit a31101e.

Remove model_sizes.hpp at configure time before each text2code run,

9d5e0ea

so text2code always generates a fresh file.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Supply model sizes through ABI for all provider paths#33

Supply model sizes through ABI for all provider paths#33
Tsilimidos wants to merge 19 commits into
masterfrom
remove-size-params-from-pdeapp

Tsilimidos commented Jul 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

Tsilimidos commented Jul 3, 2026

Summary

Changes

Core solver (ExasimSolver.cpp)

Provider paths — all 6 now supply compile-time sizes

Frontend generators — emit model_sizes.hpp

Templates & examples

Verification

Relevant files

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Core solver (`ExasimSolver.cpp`)

Frontend generators — emit `model_sizes.hpp`