Add scatter operation to direct bindings

[rdspring1]

PR List

• Add threshold and clamp operations to direct_bindings #4876
• Add addcmul operation to direct_bindings #4877
• Add slice operation to direct bindings #4878
• Add Iota operation to direct bindings #4879
• Add full operation to direct bindings #4880
• Add topk operation to direct bindings #4881
• Support alias outputs in direct bindings. #4882
• Add welford operation to direct bindings #4883
• Add scatter operation to direct bindings #4884 << This PR.

[github-actions]

Review updated until commit febeff6

Description

• Added scatter operation in direct bindings with error handling

• Mapped ScatterOp to Python frontend for direct binding usage

• Enabled scatter operation testing with supports_direct_bindings=True

---

Changes walkthrough 📝

	Relevant files
Enhancement	ops.cpp Add scatter function with error handling                                  python/python_direct/ops.cpp Added scatter function to bindIndexingOps with parameters arg1, index, src, and dim Implemented dimension count validation via NVF_CHECK Added dimension range validation for dim parameter +69/-20  python_translate.cpp Map ScatterOp to Python frontend                                                  python/python_direct/python_translate.cpp Added handle method for ScatterOp class Connected C++ ScatterOp to Python frontend via fd.ops.scatter Used generateKwargsOperation for Python binding generation +14/-0
Tests	opinfos.py Enable scatter operation testing                                                  tests/python/opinfo/opinfos.py Added scatter_opinfo to shape_ops list Enabled direct bindings support for scatter tests Updated scatter_wrapper with symbolic parameter declarations +1/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Possible Issue The error message in NVF_CHECK for dimension validation uses incorrect format specifiers when printing tensor dimensions. The variables arg1->nDims(), index->nDims(), and src->nDims() are of type size_t, but the message uses %d which may cause undefined behavior on 64-bit systems. NVF_CHECK( arg1->nDims() == index->nDims() && arg1->nDims() == src->nDims(), "Tensor arguments have different dimensions ", arg1->nDims(), ", ", index->nDims(), " and ", src->nDims()); auto num_dims = (int64_t)arg1->nDims(); NVF_CHECK( dim >= -num_dims && dim < num_dims, "Tensor arguments have dimension ", num_dims, " so dim argument must satisfy ", -num_dims, " <= dim < ", num_dims, ", but received ", dim); Missing Test The PR adds a new scatter operation but does not include corresponding test cases to validate its correctness or performance. The opinfo in tests/python/opinfo/opinfos.py only enables direct bindings without concrete test implementations. ops.def( "scatter", [](TensorView* arg1, TensorView* index, TensorView* src, int64_t dim) -> TensorView* { NVF_CHECK( arg1->nDims() == index->nDims() && arg1->nDims() == src->nDims(), "Tensor arguments have different dimensions ", arg1->nDims(), ", ", index->nDims(), " and ", src->nDims()); auto num_dims = (int64_t)arg1->nDims(); NVF_CHECK( dim >= -num_dims && dim < num_dims, "Tensor arguments have dimension ", num_dims, " so dim argument must satisfy ", -num_dims, " <= dim < ", num_dims, ", but received ", dim); return scatter(arg1, dim, index, src); }, py::arg("arg1"), py::arg("index"), py::arg("src"), py::arg("dim"), R"( Scatter a tensor. Parameters ---------- arg1 : TensorView The tensor to scatter into. index : TensorView The tensor containing the indices. src : TensorView The source tensor to scatter from. dim : int The dimension to scatter along. Returns ------- TensorView The scattered tensor. )", py::return_value_policy::reference);

[jjsjann123]

LGTM overall.

[jjsjann123]

nitpick: we don't have a functional scatter at this moment. @naoyam is adding limited support, which might not handle all the scatter cases in opinfo. #4742

[rdspring1]

!test