emx-onnx-cgen 0.3.8__tar.gz → 0.4.2.dev0__tar.gz

emx_onnx_cgen-0.4.2.dev0/.gitattributes

@@ -0,0 +1 @@
+*.sh text eol=lf

emx_onnx_cgen-0.4.2.dev0/.github/workflows/linux-release.yml

@@ -0,0 +1,61 @@
+name: linux-release
+
+on:
+  workflow_dispatch:
+    inputs:
+      upload:
+        description: "Upload release asset"
+        required: false
+        default: "false"
+  release:
+    types: [published]
+
+jobs:
+  build-linux-bundle:
+    runs-on: ubuntu-latest
+    container:
+      image: python:3.11-slim
+    permissions:
+      contents: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        run: |
+          python --version
+          python -m pip --version
+
+      - name: Install system dependencies
+        run: |
+          apt-get update
+          apt-get install -y --no-install-recommends \
+            build-essential \
+          && rm -rf /var/lib/apt/lists/*
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install -r requirements-ci.txt
+          python -m pip install pyinstaller
+
+      - name: Build PyInstaller bundle
+        run: |
+          python tools/pyinstaller_build.py
+
+      - name: Package bundle
+        run: |
+          tar -czf emx-onnx-cgen-linux-amd64.tar.gz -C dist/emx-onnx-cgen .
+
+      - name: Upload workflow artifact
+        if: ${{ github.event_name == 'workflow_dispatch' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: emx-onnx-cgen-linux-amd64
+          path: emx-onnx-cgen-linux-amd64.tar.gz
+
+      - name: Upload release asset
+        if: ${{ github.event_name == 'release' }}
+        uses: softprops/action-gh-release@v2
+        with:
+          files: emx-onnx-cgen-linux-amd64.tar.gz

emx_onnx_cgen-0.4.2.dev0/.github/workflows/tests.yml

@@ -0,0 +1,53 @@
+name: tests
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+
+jobs:
+  pytest:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10"]
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+
+      - name: Install dependencies
+        run: |
+          git submodule sync --recursive
+          git submodule update --init --recursive
+          python -m pip install -r requirements-ci.txt
+
+      - name: Run tests
+        if: github.event_name == 'push'
+        run: pytest -n auto -q
+
+      - name: Run tests (update refs for PRs)
+        if: github.event_name == 'pull_request'
+        run: UPDATE_REFS=3 pytest -n auto -q
+
+      - name: Commit updated references
+        if: github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository
+        run: |
+          if git status --porcelain | grep -q .; then
+            git config user.name "github-actions[bot]"
+            git config user.email "github-actions[bot]@users.noreply.github.com"
+            git add -A
+            git commit -m "tests: update golden references"
+            git push origin "HEAD:${GITHUB_HEAD_REF}"
+          else
+            echo "No reference updates."
+          fi

emx_onnx_cgen-0.4.2.dev0/.github/workflows/windows-release.yml

@@ -0,0 +1,48 @@
+name: windows-release
+
+on:
+  workflow_dispatch:
+  release:
+    types: [published]
+
+jobs:
+  build-windows-bundle:
+    runs-on: windows-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+          cache: pip
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install -q -r requirements-ci.txt
+          python -m pip install pyinstaller
+          
+      - name: Build PyInstaller bundle
+        run: |
+          python tools/pyinstaller_build.py
+
+      - name: Zip bundle
+        run: |
+          Compress-Archive -Path dist/emx-onnx-cgen/* -DestinationPath emx-onnx-cgen-windows-amd64.zip
+
+      - name: Upload workflow artifact
+        if: ${{ github.event_name == 'workflow_dispatch' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: emx-onnx-cgen-windows-amd64
+          path: emx-onnx-cgen-windows-amd64.zip
+
+      - name: Upload release asset
+        if: ${{ github.event_name == 'release' }}
+        uses: softprops/action-gh-release@v2
+        with:
+          files: emx-onnx-cgen-windows-amd64.zip

emx_onnx_cgen-0.4.2.dev0/AGENTS.md

@@ -0,0 +1,380 @@
+# AGENTS.md
+
+This repository builds an **ONNX → C compiler** implemented in Python.
+
+This document describes the working conventions for humans and automated agents (e.g., code assistants) contributing to the project.
+
+## Project Goals
+
+- Compile ONNX models into **portable C code** with minimal runtime dependencies.
+- Produce **correct-by-comparison** outputs against ONNX Runtime (ORT).
+- Keep the compiler architecture **pass-based** (import → normalize → optimize → lower → emit).
+- Support reproducible builds and deterministic code generation.
+- Maintain high code quality: prioritize clean OO design, avoid code smells, and continuously refactor toward simpler abstractions.
+
+## Related Repositories / References
+
+The directories **`emx-pytorch2c-org/`** and **`onnx2c-org/`** serve as **reference implementations / knowledge bases** for ideas, operator coverage, and practical edge cases:
+
+- **`onnx2c-org/`** is a **C++ implementation** with similar goals (ONNX → C compilation, deterministic codegen, minimal runtime).
+  *Note:* It may generate **ugly C code**. Use it for semantic reference and operator handling, not as a style template.
+- **`emx-pytorch2c-org/`** is a **Python-based PyTorch → C compiler**.
+  *Note:* It may contain **poor design choices / architecture smells**. Use it for concepts, codegen patterns, and testing approaches, but do **not** replicate the design blindly.
+
+When making architectural decisions, prefer clean and maintainable design even if it diverges from these repositories.
+
+## Repository Structure (expected)
+
+> If folders do not exist yet, create them following these conventions.
+
+- `src/`  
+  Python package for the compiler.
+- `src/onnx2c/`  
+  Main package (or rename to your chosen package name).
+- `src/onnx2c/ir/`  
+  Internal IR definitions (graph, tensors, types, attributes).
+- `src/onnx2c/ir/ops/`  
+  IR op dataclasses + validation/inference hooks.
+- `src/onnx2c/passes/`  
+  Normalization + optimization passes.
+- `src/onnx2c/lowering/`  
+  Lowering from IR to C-level IR / operator kernels.
+- `src/onnx2c/codegen/`  
+  C emission (templates + emit logic).
+- `src/onnx2c/runtime/`  
+  Minimal C runtime support (allocator, tensor struct, operator kernels).
+- `templates/`  
+  Jinja2 templates for C sources/headers.
+- `tests/`  
+  Pytest-based tests.
+- `tests/models/`  
+  Small ONNX models + known-good test artifacts.
+- `tools/`  
+  Developer scripts (benchmarking, model zoo fetch, etc.).
+- `docs/`  
+  Design notes, decisions, and architecture docs.
+- `examples/`  
+  Example models and compilation pipelines.
+
+## Key Invariants
+
+- **Correctness first:** any optimization must preserve semantics.
+- **Deterministic output:** code generation must be stable across runs:
+  - stable node ordering
+  - stable symbol names
+  - stable formatting
+- **No hidden state:** passes should be pure functions over IR where possible.
+- **Explicit errors:** unsupported ops or ambiguous shapes must raise actionable errors.
+
+## How to Run
+
+### Unit tests
+
+```bash
+pytest -n auto -q --maxfail=10
+```
+
+- When reporting executed tests, include the test duration (if known) in your feedback.
+- Prefer running only targeted tests known to be affected by your changes.
+- Skip running tests for documentation-only changes unless explicitly requested.
+
+### Verification (single ONNX model)
+
+To reproduce verification for an ONNX model, use the expectation entry as the
+source of truth for CLI arguments:
+
+1. Locate the model's expected error JSON in `tests/expected_errors/` (file
+   names are the repo-relative path with `/` replaced by `__`).
+2. Read the `command_line` field to capture the exact CLI arguments.
+3. Run verification by invoking the CLI with the recorded command line:
+
+```bash
+PYTHONPATH=src python -m emx_onnx_cgen.cli <command_line from JSON>
+```
+
+### Golden reference updates
+
+Golden tests compare generated code against reference files. To refresh references
+after intentional changes, set `UPDATE_REFS=1` when running tests:
+
+```bash
+UPDATE_REFS=1 pytest -n auto -q --maxfail=10
+```
+
+### Test execution policy
+
+Prefer running targeted single tests while working (e.g., a specific test module or
+test case) to keep iteration fast. Run the full test suite with reference updates
+enabled when changes impact code generation or test expectations, but it is not
+required for every change:
+
+* Run tests only when files under `src/` or `tests/` change.
+* Prefer targeted tests over `pytest -q` for routine validation.
+* `pytest -n auto -q --maxfail=10` alone takes over 5 minutes; avoid running general tests without `-n auto`.
+
+```bash
+UPDATE_REFS=1 pytest -n auto -q --maxfail=10
+```
+
+## Compiler Pipeline (conceptual)
+
+1. **Load ONNX**
+
+   * Validate model
+   * Resolve initializers
+   * Run ONNX shape inference (best-effort)
+
+2. **Import to IR**
+
+   * Build internal graph representation
+   * Normalize attributes and types
+
+3. **Normalize passes**
+
+   * Constant folding
+   * Canonicalize patterns
+   * Eliminate no-ops / identities
+   * Ensure all required shapes/types exist
+
+4. **Optimize passes**
+
+   * Fuse patterns (e.g., Conv+Relu)
+   * Layout transforms (optional)
+   * Dead code elimination
+
+5. **Lowering**
+
+   * Map IR nodes to C-level kernels or intrinsic ops
+   * Produce explicit tensor views / memory model
+   * Decide storage (stack/static/global)
+
+6. **Validate/Infer**
+
+   * Validate lowered ops
+   * Infer dtypes and shapes
+
+7. **Code generation**
+
+   * Emit `model.c/.h`
+   * Emit kernels (or link runtime kernels)
+   * Emit weights
+
+8. **Verification**
+
+   * Run ORT and generated code, compare numerically
+   * Log max abs/rel error
+
+## Coding Standards
+
+### Python style
+
+* Format: `black`
+* Lint: `ruff`
+* Prefer type hints (`typing`, `dataclasses`)
+* Avoid global mutable state
+* Keep functions small and testable
+
+### Design & Architecture (OO)
+
+* Prefer clear, maintainable OO design over cleverness.
+* Keep responsibilities small:
+
+  * one module = one concern
+  * one class = one responsibility
+  * one function = one purpose
+* Use composition over inheritance unless inheritance is clearly justified.
+* Prefer immutable data structures for IR objects where practical.
+* Keep pass interfaces consistent:
+
+  * every pass has a clear input/output contract
+  * avoid side effects and implicit global state
+* Separate concerns strictly:
+
+  * ONNX parsing/import
+  * IR + transformations
+  * lowering decisions (memory model, kernels)
+  * code generation (pure emission)
+  * runtime kernels (C)
+* Prefer explicit "context" objects over global configuration:
+
+  * e.g. `CompilerOptions`, `Target`, `DiagnosticContext`
+
+### Code Smells to Avoid
+
+* Long methods/functions (split when complexity grows).
+* "God" classes or modules that combine unrelated responsibilities.
+* Hidden coupling via globals, singletons, or implicit module-level state.
+* Deep inheritance hierarchies.
+* Feature envy: methods that operate mostly on other objects’ data.
+* Duplicated logic (introduce helpers or shared utilities).
+* Unclear naming: prefer descriptive names over abbreviations.
+* Tight coupling between passes and codegen or runtime internals.
+
+### Error handling
+
+* Prefer explicit custom exceptions:
+
+  * `UnsupportedOpError`
+  * `ShapeInferenceError`
+  * `LoweringError`
+  * `CodegenError`
+* Error messages must include:
+
+  * op type
+  * node name (if any)
+  * input/output shapes (if known)
+  * hint to fix (e.g., “run onnxsim” or “export with static shapes”)
+
+### Logging
+
+* Use Python `logging` (no prints in library code).
+* Default level: INFO.
+* Provide debug dumps behind flags.
+
+## Code comments
+
+- Add comments only where they add durable value: intent/why, invariants/assumptions, non-obvious trade-offs, external constraints/workarounds.
+- Do not comment obvious code or restate what the code already shows. Keep comments brief and keep them updated when changing related code.
+
+## Testing Guidelines
+
+### Requirements for new features
+
+* Every new operator/kernel requires:
+
+  * at least one unit test
+  * at least one ORT comparison test
+  * edge cases (broadcasting, padding, axis, negative indices, etc.)
+
+### Test model sources
+
+* Keep test ONNX models small.
+* Prefer:
+
+  * hand-constructed graphs
+  * ONNX Model Zoo models only for integration tests (optional)
+
+## Performance Guidelines
+
+* Avoid Python-level tensor computations in hot paths except for:
+
+  * constant folding
+  * verification
+* In C runtime:
+
+  * prefer contiguous memory layouts
+  * avoid dynamic allocations unless explicitly enabled
+* Any performance optimization must include a benchmark or profiling note.
+
+## Documentation Expectations
+
+* Any major design change needs a short note in `docs/`:
+
+  * motivation
+  * alternatives considered
+  * decision
+  * consequences
+* Public-facing usage should be documented in `README.md`.
+
+## Contribution Workflow
+
+* Keep PRs focused and small.
+* For any change affecting output C code:
+
+  * include golden snapshot updates (if applicable)
+  * run ORT comparison tests
+* Prefer descriptive commit messages:
+
+  * `passes: fold constants for Add/Mul`
+  * `codegen: stable symbol naming`
+  * `runtime: add Conv2D kernel (NHWC)`
+
+## Agent Instructions (for automated assistants)
+
+When acting as an agent in this repo:
+
+1. **Do not introduce new dependencies** without a clear need.
+2. **Do not change code generation output formatting** unless necessary.
+3. Maintain determinism:
+
+   * stable ordering, stable naming
+4. If unsure about semantics, prefer correctness and add a TODO + test.
+5. Always propose refactorings when they improve:
+
+   * readability
+   * testability
+   * separation of concerns
+   * determinism
+6. Detect and call out common code smells (see above) and suggest concrete improvements.
+7. Prefer incremental refactorings:
+
+   * keep PRs small and mechanical
+   * preserve behavior and determinism
+   * add tests before/with refactoring when needed
+8. When adding new functionality, keep the design extensible:
+
+   * avoid hardcoding operator names and special cases
+   * prefer registries / dispatch tables for op lowering and kernels
+9. When adding a new operator, clarify the operator semantics first.
+10. If the architecture starts drifting, propose a short design note in `docs/` with:
+* problem
+* options
+* recommendation
+11. To find operator specifications within this repository, prefer:
+    * `onnx-org/docs/Operators.md` for the current consolidated operator docs.
+    * `onnx-org/docs/Changelog.md` for versioned operator specs (e.g., Slice-13).
+    * `onnx-org/onnx/defs/**/defs.cc` or `onnx-org/onnx/defs/**/old.cc` for the
+      schema source and historical documentation strings.
+    * Use ripgrep (`rg -n "<OpName>" onnx-org`) to locate the relevant sections.
+12. When implementing min/max selection in templates, use scalar min/max helpers
+    from the scalar function registry instead of `if`-based comparisons.
+13. Do not modify `onnx-org/` or any other submodule contents.
+
+## Adding an operator (checklist)
+
+When adding support for a new operator:
+
+1. **Specification:** read and understand the ONNX operator spec at
+   `onnx-org/docs/Operators.md`.
+2. **Lowering:** add a `register_lowering()` handler in `src/onnx2c/lowering/` that validates shapes/dtypes and returns a new op dataclass.
+3. **Codegen:** add a new op dataclass + rendering path in `src/onnx2c/codegen/c_emitter.py` and a matching template in `templates/`.
+4. **Runtime evaluator:** implement the op in `src/onnx2c/runtime/evaluator.py` for verification/constant folding.
+5. **Tests:** add at least one unit test and one ORT comparison test.
+6. **Reference updates:** refresh `tests/expected_errors/**.json` and regenerate `OFFICIAL_ONNX_FILE_SUPPORT*.md` via `UPDATE_REFS=1 pytest -n auto -q`.
+
+## Maintaining this document (AGENTS.md)
+
+This file is part of the project’s contract.
+
+Update `AGENTS.md` when:
+- the repository structure changes (new top-level folders, renamed packages)
+- the compiler pipeline changes (new major phases or invariants)
+- coding standards or tooling decisions change (`ruff`, `black`, test strategy, etc.)
+- new supported targets/constraints are added (dynamic shapes, quantization, new runtimes)
+- recurring confusion appears in PR reviews (add clarifying rules)
+- an agent repeatedly makes the same mistake (add an explicit guardrail)
+
+Rules for updating:
+- Keep changes small and specific.
+- Include the rationale in the commit/PR description.
+- Prefer adding explicit examples over vague guidelines.
+- Do not add rules that are not enforced or not useful.
+- If a new rule conflicts with existing rules, resolve the conflict in `AGENTS.md` as part of the same change.
+
+Agents:
+- If you encounter ambiguity not covered by this document, propose an update to `AGENTS.md`.
+- If your change introduces new conventions, update this file in the same PR.
+- If you need the operator reference implementation or schema details, look in
+  `onnx-org/onnx/defs/**/defs.cc` or `onnx-org/onnx/defs/**/old.cc`, and use
+  `rg -n "<OpName>" onnx-org` to locate the relevant operator definition.
+
+## Security / Safety
+
+* Do not execute untrusted ONNX models from unknown sources.
+* Avoid shelling out to external tools in the compilation path.
+* Generated C code should not include:
+
+  * filesystem I/O
+  * network I/O
+  * dynamic loading
+  * `memset` or `memcpy`