[Launcher] Add packed-launch fast path (Fix E)

[NathanVoldman]

Adds an optional "packed launch" path to the XPU launcher that resolves pointer arguments once in Python and passes them to C as a flat void*[] buffer, bypassing the per-arg PyObject_GetAttr + PyObject_CallNoArgs chain in extractPointer.

C-side additions (driver.c):

• buildRawKernelArgIndices(): walks arg_annotations to map signature slots to raw-args-tuple indices; sets packable=false on ARG_TUPLE.
• build_pack_layout(): computes pointer_raw_indices + pack_buffer_size for a given (kernel_signature, arg_annotations) pair. Returns packable=false when annotations contain tuple-nested entries or when there are no pointer args (packing has no upside).
• launch_packed(): counterpart to launch() that reads pointer values from a caller-provided y*-buffer instead of extracting them from Python objects one at a time. Scalar args still go through their normal extractor.

Python-side additions (driver.py):

• SpirvUtils splits DLL symbols into _REQUIRED_METHODS and _OPTIONAL_METHODS. launch_packed / build_pack_layout are optional and gated by has_packed_launch; older / stripped builds transparently fall back to classic launch().
• XPULauncher.init probes eligibility once at specialization time: when the kernel has pointer args and no tensordesc / tuple annotations, it pre-allocates a bytearray pack buffer and a memoryview cast to native pointer size, and selects launch_packed as the launcher.
• XPULauncher.call takes a packed fast path when enabled: iterate _pointer_arg_indices, resolve each arg via numbers.Integral / .data_ptr(), write into the pack buffer, then hand C the buffer + args tuple.
• Escape hatch: TRITON_INTEL_DISABLE_PACKED_LAUNCH=1 forces the classic launch() path.

Fix E is independent of Fix A (KernelInfo cache) and Fix B (data_ptr intern) — this branch contains only Fix E on top of origin/main.

New contributor declaration

• I am not making a trivial change, such as fixing a typo in a comment.

• I have written a PR description following these
  rules.

• I have run pre-commit run --from-ref origin/main --to-ref HEAD.

• Select one of the following.

  • I have added tests.
    • /test for lit tests
    • /unittest for C++ tests
    • /python/test for end-to-end tests
  • This PR does not need a test because flow run in every launch.

• Select one of the following.

  • I have not added any lit tests.
  • The lit tests I have added follow these best practices,
    including the "tests should be minimal" section. (Usually running Python code
    and using the instructions it generates is not minimal.)

[Copilot]

Pull request overview

Adds an optional packed-pointer fast path to reduce XPU kernel launch overhead.

Changes:

• Adds packed layout construction and native packed launching.
• Selects packed launching for eligible kernels with compatibility fallbacks.
• Adds an environment-variable escape hatch.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
third_party/intel/backend/driver.py	Selects, configures, and invokes packed launches.
third_party/intel/backend/driver.c	Builds pack layouts and launches using packed pointers.

Suppressed comments (1)

third_party/intel/backend/driver.c:1403

• The exported helper assumes the signature length equals the number of ARG_KERNEL annotations, then uses the unchecked PyList_GET_ITEM macro. A mismatched direct call can read outside the list and crash Python. Check the lengths and raise before entering this loop.

    for (Py_ssize_t i = 0; i < num_sig_args; ++i) {
      if (sig[i] == EXTRACTOR_POINTER_INDEX) {
        PyObject *raw_idx =
            PyList_GET_ITEM(kernel_arg_raw_indices, i); // borrowed

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.