[ets-public]

Requested by Adam · project thread

What

PR 7a of native numpy: float32 and int32 arrays and scalars.

• Dtypes: dtype=np.float32 / np.int32 (also "f4", "i4", np.single, np.intc) work in constructors, astype, arange and linspace. np.float32(x) and np.int32(x) work too.
• Promotion: NEP 50 rules. A Python int that does not fit int32 raises numpy's OverflowError.
• float32 arithmetic is computed in double and rounded once per operation, like numpy's float32 loops. // and % use float32 divmod.
• int32 wraps around and gives numpy's overflow warnings.
• Reductions: float32 sum, mean and var are pairwise in float32. int32 sum and prod give int64, as in numpy.
• Printing: shortest float32 digits, the 1e6 cutoff for scientific format, np.float32(...) scalar repr, and dtype= in array repr.
• Not native yet (gated): @/dot (sgemm), np.linalg, ufunc reduce/accumulate, and float32 exp/log/trig/powers, where numpy uses SVML. A build gate sends these files to CPython's numpy, and a runtime refusal catches what reaches them anyway.

Fixes found by fuzzing against numpy

240 random programs mixing float32, int32, float64 and int64 now match numpy byte for byte; the only differences are explicit refusals. Several of these bugs were already present for float64:

• Dragon4 breaks ties half-to-even; Rust's shortest form rounds ties up. This affected numpy printing, and also Python repr of floats in typed and dynamic builds (-254.62014770507812).
• When padding printed digits, numpy uses the value's exact digits, not zeros.
• np.rint of an int array now returns float64.
• A sum of -0.0 values gives 0.0 (numpy's add.reduce starts from the identity).
• An array ** 0.5 / ** -1 / ** 2 with a scalar exponent runs numpy's sqrt/reciprocal/square. This is numpy's fast_scalar_power, so (-0.0) ** 0.5 is -0.0, and the warnings name the right ufunc.
• The repr of an array with more than 1,000 items now shows shape=(...).
• int() of a numpy float that is nan or inf now raises; a large value stays exact.
• np.arange with a float step and an int dtype now fills as numpy does.
• int64 MIN // -1, and -/abs of the minimum int scalar, now warn.
• numpy's default sort is unstable quicksort, and its order differs by CPU (checked with NPY_DISABLE_CPU_FEATURES). Without kind="stable", np.argsort of repeated values and sorting 0.0 with -0.0 now raise NotImplementedError instead of silently giving a stable order. kind= is now accepted by np.sort.
• CPython bridge fixes: np.where was lexed as where_, and x ** -1 called float() on the array.

Tests

• New tests/numpy/narrow.py. Its output is the same with numpy's AVX2 dispatch and with OPENBLAS_CORETYPE=Haswell.
• The full tests/check.sh passes.
• The numpy corpus is unchanged: 37/40 native, all match.
• sh tools/ci_local.sh: all checks passed (2003 baseline solutions, 0 regressions).

---

Generated by Claude Code