perf(pathfinder): fast-path exact directory searches

[KRRT7]

Description

Optimize exact directory discovery in cuda.pathfinder by joining the complete candidate path and checking it once.

Previously, exact paths were traversed one component at a time, resulting in one filesystem check per component. Wildcard paths retain the existing traversal behavior.

This affects the shared discovery path used for CUDA headers, binaries, static libraries, bitcode libraries, and dynamic-library locations.

A new cross-platform pyperf benchmark suite under benchmarks/cuda_pathfinder/ covers:

• exact-path hits and misses;
• wildcard traversal;
• cached lookup;
• cold public API discovery for headers, binaries, and static libraries;
• realistic one-root and three-root site-packages layouts.

Performance

Rigorous A/B measurements compared pristine main against this branch using identical environments and fixtures:

Public workload	Before	After	Improvement
Header discovery, 1 root	142 us	70.8 us	2.01x
Header discovery, 3 roots	202 us	141 us	1.44x
Binary discovery, 1 root	149 us	71.8 us	2.07x
Binary discovery, 3 roots	206 us	144 us	1.43x
Static-library discovery, 3 roots	219 us	141 us	1.55x

Geometric mean: 1.68x faster.

The benchmark fixtures use NVIDIA wheel layouts such as:

• nvidia/cuda_runtime/include
• nvidia/cuda_nvcc/bin
• nvidia/cuda_runtime/lib

Notes

• GPU execution performance is unchanged.
• This improves CPU-side cold discovery and startup latency.
• Wildcard search behavior and result ordering are unchanged.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.