PERF: unroll take_1d/take_2d inner loops with contig fast path

[jbrockmendel]

Summary

Adds a contiguous-stride fast path to take_1d, take_2d_axis0, take_2d_axis1, and take_2d_multi in pandas/_libs/algos_take_helper.pxi.in. When all input arrays have strides == sizeof(dtype) along the relevant axes (the common case), the kernel walks raw typed pointers instead of memoryview byte-offsets — this lets the C compiler drop the runtime stride multiply. The inner gather loop is then manually unrolled 4× so the out-of-order pipeline can overlap independent indexed loads (aarch64 has no SIMD gather; this is the main available lever).

The existing same-dtype 2D axis0 memcpy fast path is preserved.

Motivation

Inspecting the pre-patch disassembly of take_1d_int64_int64 showed the hot loop was 7 scalar instructions per element with a per-iteration mul for the stride multiply — the memoryview indexing expansion *(data + i * strides[0]) defeats auto-vectorization and auto-unrolling. The patched loop is 14 instructions per 4 elements (3.5/elem) with 4 independent indexed loads exposed to the pipeline.

Benchmarks

Kernel microbenchmarks (aarch64, M-series Mac, n=1M unless noted):

Kernel	Case	Speedup
take_1d_i64_i64	no-fill monotonic	1.24×
take_1d_i64_i64	fill monotonic	1.58×
take_1d_i64_i64	random	1.27–1.35×
take_2d_axis0 f64→f64	narrow k (< 32 elem)	3.0×
take_2d_axis0 different-dtype (e.g. int32→int64, float32→float64)		1.9–2.0×
take_2d_axis1 f64→f64	random	1.2×
take_2d_multi f64→f64		1.2–1.3×

End-to-end (back-to-back against main):

Operation	Baseline	Patched	Speedup
Series[float64].take mono n=1M	1255 µs	982 µs	+22%
DataFrame[100k×10].take mono	656 µs	439 µs	+33%
DataFrame[10k×100].take mono	537 µs	374 µs	+30%
DataFrame[1k×1k].take random	523 µs	389 µs	+25%
Series[bool_mask] n=1M	4581 µs	4083 µs	+11%
DataFrame[bool_mask] n=1M	4271 µs	3902 µs	+9%
Series.reindex n=1M	3215 µs	2729 µs	+15%

Test plan

• pandas/tests/test_take.py (86 tests)
• pandas/tests/series/indexing/test_take.py + pandas/tests/frame/indexing/test_take.py
• pandas/tests/series/methods/test_reindex.py + pandas/tests/frame/methods/test_reindex.py
• pandas/tests/indexing/ full sweep (4950 passed)
• x86 CI — aarch64 has no SIMD gather; on x86 with AVX2+ the same pattern could additionally emit vpgatherdq. Worth confirming wins translate.
• Wheel size delta (unroll inflates per-dtype instantiation by ~40 lines × ~20 type pairs)

TODO

• whatsnew entry in v3.1.0.rst (will add after PR number is known)

🤖 Generated with Claude Code