PERF: Use ABI-guaranteed SIMD baselines for redistribution-safe FFTW builds

[hjmjohnson]

Summary

Addresses seanm's review of PR #6006: `check_c_source_runs` probed the build host's CPU at configure time, making distributed FFTW binaries unsafe on machines that lack the detected SIMD extensions (SIGILL crash). This is incompatible with conda, pip/PyPI, and manylinux redistribution workflows.

Detection policy (compile-time only, never runtime)

Platform	Extension	Default	Rationale
x86_64 / AMD64	SSE, SSE2	ON	Mandated by the AMD64 ABI — every 64-bit x86 CPU has them
aarch64 / arm64	NEON	ON	Mandated by the AArch64 ABI — every arm64 CPU has it
x86_64	AVX	OFF by default	Auto-ON when compiler already targets AVX (e.g. `-march=native`, `/arch:AVX`)
x86_64	AVX2	OFF by default	Auto-ON when compiler already targets AVX2 (e.g. `-mavx2`, `/arch:AVX2`)
macOS universal	all SIMD	OFF	Single configure cannot target both arm64 and x86_64 simultaneously

AVX/AVX2 detection uses `check_c_source_compiles` (not `_runs`) against the predefined macros `AVX`/`AVX2`, which are only defined when the compiler is actually generating those instructions. Cache variables are unset before each probe so detection re-runs every configure when compiler flags change.

Impact on redistribution platforms

• conda/pip x86_64: SSE+SSE2 always ON (previously OFF without runtime probe passing)
• conda/pip arm64: NEON always ON (unchanged)
• AVX2 on build host: ON only when compiler targets it (previously ON for all build hosts with AVX2 hardware)

Follows up on / closes #6006.

Test plan

• Configure on x86_64 Linux without extra flags → SSE/SSE2=ON, AVX/AVX2=OFF
• Configure on x86_64 with `-DCMAKE_C_FLAGS=-mavx2` → AVX/AVX2=ON
• Configure on arm64 → NEON=ON
• Configure with `CMAKE_OSX_ARCHITECTURES="arm64;x86_64"` → all SIMD=OFF
• Re-configure after adding `-march=native` — AVX detection re-runs (not stale)

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR replaces a runtime CPU probe (check_c_source_runs + __builtin_cpu_supports) with a redistribution-safe detection policy: SSE/SSE2 are unconditionally ON for x86_64 (ABI-mandated), NEON is unconditionally ON for aarch64, and AVX/AVX2 are enabled only when the compiler's pre-defined __AVX__/__AVX2__ macros are present. A macOS universal-binary guard disables all SIMD defaults when CMAKE_OSX_ARCHITECTURES names more than one arch.

• The check_c_source_compiles result for _fftw_compiler_targets_avx is cached as a CMake cache entry. If a user first configures without -march=native (caching it as 0) and later adds the flag, the stale cached value prevents re-detection, and option(FFTW_ENABLE_AVX) will not update either since it is already in the cache — effectively making the advertised "auto-enable" a one-shot event. Adding unset(_fftw_compiler_targets_avx CACHE) before each probe (or a clear warning comment) would make the behaviour match user expectations.

Confidence Score: 4/5

Safe to merge with a known UX rough edge: the auto-detection of AVX/AVX2 is a one-shot operation due to CMake check-variable caching, which may surprise users who change CMAKE_C_FLAGS on an existing build tree.

The core redistribution-safety goal (replacing _runs with _compiles / ABI baselines) is correct and well-reasoned. One P1 finding remains: the stale-cache double-lock means the auto-enable AVX behaviour silently does nothing on a reconfigure, contrary to the stated semantics. This does not produce incorrect FFTW binaries but fails to upgrade them when the user's intent changes.

CMake/itkExternal_FFTW.cmake — the two check_c_source_compiles blocks for AVX/AVX2 and their interaction with option() caching.

Important Files Changed

Filename	Overview
CMake/itkExternal_FFTW.cmake	Replaces runtime CPU-probing (check_c_source_runs) with ABI-guaranteed SSE/SSE2/NEON defaults and compile-time AVX/AVX2 detection via check_c_source_compiles; adds macOS universal-binary guard. Correct for redistribution safety, with one stale-cache edge case for users who change CMAKE_C_FLAGS between configure runs.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[cmake configure] --> B{ITK_USE_SYSTEM_FFTW?}
    B -- yes --> Z[find_package FFTW]
    B -- no --> C{APPLE AND CMAKE_OSX_ARCHITECTURES count > 1?}
    C -- yes universal2 --> D[All SIMD OFF + status message]
    C -- no --> E{CMAKE_SYSTEM_PROCESSOR}
    E -- aarch64/arm64/ARM64 --> F[NEON=ON, SSE/AVX=OFF]
    E -- x86_64/AMD64 --> G[SSE=ON, SSE2=ON]
    G --> H[check_c_source_compiles __AVX__ macro]
    H -- defined --> I[_fftw_default_avx=ON]
    H -- not defined --> J[_fftw_default_avx=OFF]
    I --> K[check_c_source_compiles __AVX2__ macro]
    J --> K
    K -- defined --> L[_fftw_default_avx2=ON]
    K -- not defined --> M[_fftw_default_avx2=OFF]
    E -- i686/i386 --> N[All SIMD OFF]
    E -- other --> N
    F --> O[option FFTW_ENABLE_* cached defaults]
    L --> O
    M --> O
    D --> O
    N --> O
    O --> P[ExternalProject_Add fftwf/fftwd with -DENABLE_* flags]

Loading

_{Reviews (1): Last reviewed commit: "PERF: Use ABI-guaranteed SIMD baselines ..." | Re-trigger Greptile}

[hjmjohnson]

Addressed the greptile P1 finding: added unset(_fftw_compiler_targets_avx CACHE) and unset(_fftw_compiler_targets_avx2 CACHE) before each check_c_source_compiles probe.

This ensures the AVX/AVX2 compiler capability check re-runs on every cmake configure, correctly reflecting the current CMAKE_C_FLAGS. Previously, a user who configured without -march=native (caching the result as 0) and then added the flag on a subsequent configure would silently get the stale detection.

The option() caching semantics are intentionally preserved: FFTW_ENABLE_AVX / FFTW_ENABLE_AVX2 are only auto-set from the detected default when not already present in the CMake cache, so explicit user overrides (-DFFTW_ENABLE_AVX2=ON) survive reconfigures. The commit message documents this for users who need to force re-evaluation.

[dzenanz]

Looks good on a glance. Somebody else should review too.