IntelPython / dpnp / 38446507311

83%

push

github

Fix kernel launches over more than INT_MAX work-items with DPC++ 2026.2 (#3089)

This PR fixes a `RuntimeError` raised by kernels launched over more than
`INT_MAX` work-items when dpnp is built with DPC++ compiler 2026.2:
```
RuntimeError: The kernel was compiled with -fsycl-id-queries-range=int, but the provided range/offset exceeds the maximum value storable in an int. Either reduce the range/offset or recompile the kernel with -fsycl-id-queries-range=[uint|size_t].
```

It affects operations on arrays with more than `2**31` elements that go
through a strided kernel, e.g. copying a broadcast view, `dpnp.take`,
advanced indexing, `dpnp.choose` and assignment through `__setitem__`.
Builds with DPC++ 2026.1 are not affected.

Starting with 2026.2, the range a kernel assumes for its work-item IDs
is recorded in the device image at the device link step, and the SYCL
runtime checks each launch against it. dpnp passed
`-fno-sycl-id-queries-fit-in-int` only at compile time, so the link step
recorded the default `int` range even though the device code was built
for `size_t`.

The existing option spelling is kept because both 2026.1 and 2026.2
accept it.

The third party tests for arrays with more than `2**31` elements are
updated:
- `TestHugeArrays`: new
`test_take_negative_index_at_int32_size_boundary`, which uses a
broadcast view, needs no real allocation and runs by default; reworked
`test_take_array` and `test_take_scalar`
- new `TestTriLargeDimensions` for `dpnp.tri` with more than `2**31`
rows or columns
- added tests that depend on implementation-internal kernels are marked
as skipped

Validated with DPC++ 2026.2 on Intel Data Center GPU Max 1100 (Level
Zero) and on CPU (OpenCL), running the slow tests with `-m slow`;
results are the same on both devices:

| Test | Before | After |
|---|---|---|
| `TestHugeArrays::test_advanced` | failed | passed |
| `TestHugeArrays::test_take_array` | failed | passed |
| `TestHugeArrays::test_take_scalar` | failed | ... (continued)

1608 of 3000 branches covered (53.6%)

Branch coverage included in aggregate %.

27914 of 34528 relevant lines covered (80.84%)

19617.16 hits per line