Example Library

This directory stores manually prepared CUDA and Ascend C SIMT example pairs for reference.

Purpose

Use these examples as a pattern library:

• to understand how one CUDA operator was migrated into Ascend C SIMT
• to compare host-side launch flow, build setup, and basic implementation structure
• to borrow layout and organization ideas for new migrations

Layout

The current example library is organized by delivery shape:

• cuda-sample/
  • standalone CUDA samples
• simt-sample/
  • standalone Ascend C SIMT samples
• cuda-torch/
  • CUDA-side PyTorch custom operator examples
• simt-torch/
  • Ascend C SIMT-side PyTorch custom operator examples

Current example entries:

• cuda-sample/vector-add/
• simt-sample/vector-add/
• cuda-torch/muladd/
• simt-torch/muladd/

Example notes:

• simt-torch/muladd/ is the finalized Ascend C SIMT torch extension reference for this operator pair.
• The SIMT side keeps the original extension_cpp/ project layout as the primary comparison target.
• CUDA-side examples may use different registration styles, but the SIMT-side migration target should follow the SIMT torch registration style used by this example.
• Use this example to study minimal structural changes from the CUDA extension_cpp/ project into a standalone Ascend C SIMT torch-style project.
• The Python wrapper and fake/meta registration may remain intentionally close to the CUDA-side wrapper when that preserves the user-facing Python API. torch.library.register_fake, FakeTensor/meta kernels, and wrapper code kept only for torch.compile compatibility are not native JIT compilation or loading paths unless they compile or load native code at runtime.
• Native JIT compilation, runtime compilation, extension JIT loading, and nvrtc-backed paths remain unsupported migration targets and should be removed or reported according to reference/constraint_rules.yaml.

Limits

Do not treat these examples as the source of truth for API compatibility.

For actual migration decisions, always prioritize:

1. reference/api-mapping/*.yaml
2. reference/constraints.md
3. reference/grammar.md
4. official local documentation

Maintenance

When adding new examples:

• keep each operator self-contained
• keep CUDA and SIMT examples easy to compare
• prefer one operator per directory if multiple examples are added later
• document any important simplifications or assumptions near the example
• place the example under the directory that matches its delivery shape