Skip to content

Commit 3662401

Browse files
hjmjohnsonclaude
hjmjohnson
and
claude
committed
PERF: Use ABI-guaranteed SIMD baselines for redistribution-safe FFTW builds
Addresses seanm's review of ITK PR #6006: the previous check_c_source_runs approach probed the BUILD HOST's CPU at configure time, producing FFTW binaries that require the build machine's exact CPU and SIGILL on any machine that lacks the detected SIMD extensions. This is unsafe for redistributed binary packages (conda, pip/PyPI, manylinux Docker images) where build and target machines differ. New detection policy (compile-time only, never runtime): x86_64 / AMD64: SSE and SSE2 are mandated by the AMD64 ABI -- every 64-bit x86 CPU supports them regardless of age. Both are enabled by default. Safe for all manylinux2014 / manylinux_2_28 / conda x86_64 builds. aarch64 / arm64: NEON is mandated by the AArch64 ABI -- every arm64 CPU has it. Enabled by default. Safe for all conda / manylinux aarch64 builds. AVX / AVX2 (Sandy Bridge 2011 / Haswell 2013 required): NOT universally available; default OFF for redistribution safety. Auto-enabled only when the compiler is already generating those instructions -- i.e. when the user passed -march=native, -mavx2, /arch:AVX2, or similar. Detected via check_c_source_compiles (not _runs) which tests what the compiler targets, not what the build host's CPU can execute. This implements seanm's recommended "the compiler knows what CPU it's compiling for" approach. The AVX/AVX2 cache variables are unset before each probe so that detection re-runs on every configure when compiler flags change (e.g. user later adds -march=native). macOS universal binary (CMAKE_OSX_ARCHITECTURES with >1 entry): SIMD defaults disabled; a single configure pass cannot produce correct per-slice codelets for both arm64 and x86_64. This change is a strict improvement on the previous behaviour for the most important redistribution platforms: - conda/pip on x86_64: SSE+SSE2 always ON (was OFF without runtime probe) - conda/pip on arm64: NEON always ON (unchanged) - AVX2 on build host: ON only when compiler targets it (was ON always) Closes #6006 (follow-up addressing seanm review) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1 parent b94cca5 commit 3662401

1 file changed

Lines changed: 117 additions & 52 deletions

File tree

‎CMake/itkExternal_FFTW.cmake‎

Lines changed: 117 additions & 52 deletions
Original file line numberDiff line numberDiff line change
@@ -1,32 +1,56 @@
11
#
22
# Encapsulates building FFTW as an External Project.
33
#
4-
# SIMD codelet selection
5-
# ----------------------
6-
# FFTW SIMD codelets are hand-written assembly routines baked into the
7-
# library at compile time. Passing -march=native to the ITK build does
8-
# NOT activate them; they must be requested explicitly via FFTW's own
9-
# CMake options (ENABLE_NEON, ENABLE_SSE, ENABLE_SSE2, ENABLE_AVX, ENABLE_AVX2).
4+
# SIMD codelet selection and binary redistribution policy
5+
# -------------------------------------------------------
6+
# FFTW SIMD codelets are hand-written assembly routines compiled INTO the
7+
# library at build time. Unlike -march=native on the ITK side, FFTW codelets
8+
# must be requested explicitly via FFTW's own CMake options
9+
# (ENABLE_NEON, ENABLE_SSE, ENABLE_SSE2, ENABLE_AVX, ENABLE_AVX2).
10+
#
11+
# For redistributable binary packages (conda, pip/PyPI, manylinux Docker
12+
# images, etc.) SIMD codelets must only be enabled when the resulting binary
13+
# will run correctly on ALL machines in the target distribution. The ISA
14+
# baseline mandated by each architecture ABI is universally safe:
15+
#
16+
# x86_64 / AMD64 : SSE and SSE2 are required by the AMD64 ABI. Every
17+
# 64-bit x86 CPU (including all manylinux2014 /
18+
# manylinux_2_28 targets) supports them. DEFAULT ON.
1019
#
11-
# This file detects appropriate defaults at cmake configure time:
20+
# aarch64 / arm64 : NEON is required by the AArch64 ABI. Every 64-bit ARM
21+
# CPU (Apple Silicon, all Linux aarch64 targets) supports
22+
# it. DEFAULT ON.
1223
#
13-
# Native builds (CMAKE_CROSSCOMPILING is false):
14-
# - ARM64 (aarch64/arm64/ARM64): NEON=ON (mandatory in ARMv8); x86 SIMD off.
15-
# - x86/x86_64 with GCC/Clang: each of SSE, SSE2, AVX, AVX2 is probed
16-
# individually via __builtin_cpu_supports() / CheckCSourceRuns so that
17-
# the detected flags match the actual build-host CPU. A pre-AVX
18-
# Sandy Bridge gets SSE+SSE2 only; a Haswell or later gets all four.
19-
# On MSVC the probes are skipped (intrinsic unavailable) and SIMD
20-
# defaults to off; users can override via FFTW_ENABLE_* options.
21-
# - Other architectures: all SIMD off (conservative fallback).
24+
# AVX / AVX2 : Not part of the baseline ABI; present only on Sandy
25+
# Bridge (2011) and Haswell (2013) and newer CPUs
26+
# respectively. Enabling them by default would produce
27+
# binaries that SIGILL on older (but spec-compliant)
28+
# x86_64 CPUs. DEFAULT OFF unless the compiler is already
29+
# targeting a micro-architecture that includes them.
2230
#
23-
# Cross-compiled builds (CMAKE_CROSSCOMPILING is true):
24-
# - ARM64: NEON=ON (mandatory); x86 SIMD off.
25-
# - x86_64: SSE+SSE2 only (baseline; AVX/AVX2 not assumed for target).
26-
# - Other: all SIMD off.
31+
# Opt-in to AVX / AVX2
32+
# ---------------------
33+
# If the user's toolchain is already generating AVX/AVX2 instructions
34+
# (because they passed -march=native, -mavx2, -march=haswell, or an
35+
# equivalent MSVC /arch: flag) the compiler pre-defines __AVX__ / __AVX2__.
36+
# This file detects those macros at cmake configure time via
37+
# check_c_source_compiles (compile-time, NOT runtime — no build-host CPU
38+
# probe is performed) and auto-enables the matching FFTW codelets so that
39+
# FFTW's generated code aligns with the rest of the ITK build.
40+
# Users who want AVX2 in a redistributed package can set:
41+
# cmake -DFFTW_ENABLE_AVX2=ON ...
42+
#
43+
# macOS universal binary
44+
# ----------------------
45+
# When CMAKE_OSX_ARCHITECTURES lists more than one value (e.g. "arm64;x86_64")
46+
# a single FFTW configure/build pass cannot correctly serve both slices.
47+
# SIMD defaults are set to OFF in this case; use ITK_USE_SYSTEM_FFTW with a
48+
# proper universal FFTW installation (e.g., built with lipo) if SIMD
49+
# performance is required in a macOS universal build.
2750
#
28-
# Every flag is an individually overridable cache option, e.g.:
29-
# cmake -DFFTW_ENABLE_AVX2=OFF ...
51+
# Every flag remains individually overridable, e.g.:
52+
# cmake -DFFTW_ENABLE_AVX2=ON # opt in to AVX2 for a non-redistributed build
53+
# cmake -DFFTW_ENABLE_SSE2=OFF # opt out of SSE2 (unusual)
3054
# Note: option() defaults are only applied on the first configure.
3155
# To re-detect after a toolchain change, delete the CMake cache or use
3256
# cmake --fresh, or pass explicit -DFFTW_ENABLE_*= overrides.
@@ -84,70 +108,111 @@ if(NOT ITK_USE_SYSTEM_FFTW)
84108
set(FFTW_STAGED_INSTALL_PREFIX "${ITK_BINARY_DIR}/fftw")
85109

86110
# Detect SIMD defaults (see file header for full policy description).
87-
# CheckCSourceRuns results are cached after the first cmake configure run.
88-
include(CheckCSourceRuns)
111+
#
112+
# Architecture-guaranteed ISA baselines (no runtime probe needed):
113+
# - x86_64 mandates SSE + SSE2 in the AMD64 ABI.
114+
# - arm64/aarch64 mandates NEON in the AArch64 ABI.
115+
#
116+
# AVX/AVX2 opt-in via compiler predefined macros:
117+
# check_c_source_compiles (not _runs) reflects what the compiler is
118+
# generating for the TARGET architecture, not what the BUILD HOST's CPU
119+
# can execute. This is safe for cross-compilation and redistribution.
120+
include(CheckCSourceCompiles)
89121

90122
set(_fftw_default_neon OFF)
91123
set(_fftw_default_sse OFF)
92124
set(_fftw_default_sse2 OFF)
93125
set(_fftw_default_avx OFF)
94126
set(_fftw_default_avx2 OFF)
95127

96-
if(NOT CMAKE_CROSSCOMPILING)
97-
if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
98-
# NEON is mandatory in ARMv8/AArch64 — every arm64 CPU has it.
99-
set(_fftw_default_neon ON)
100-
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64|i686")
101-
# Probe each x86 SIMD level individually via CPUID so the defaults
102-
# are accurate for the actual build-host CPU (e.g. pre-AVX Sandy Bridge
103-
# or pre-AVX2 Ivy Bridge get only the levels their hardware supports).
104-
# __builtin_cpu_supports is a GCC/Clang intrinsic; skip on MSVC.
105-
if(CMAKE_C_COMPILER_ID MATCHES "GNU|Clang|AppleClang")
106-
foreach(_fftw_simd IN ITEMS sse sse2 avx avx2)
107-
check_c_source_runs(
108-
"int main(void){return __builtin_cpu_supports(\"${_fftw_simd}\")?0:1;}"
109-
_fftw_cpu_has_${_fftw_simd}
110-
)
111-
if(_fftw_cpu_has_${_fftw_simd})
112-
set(_fftw_default_${_fftw_simd} ON)
113-
endif()
114-
endforeach()
115-
endif()
128+
# Detect macOS universal binary build: a single configure+build pass cannot
129+
# simultaneously produce correct SIMD for both arm64 and x86_64 slices.
130+
set(_fftw_is_universal FALSE)
131+
if(APPLE AND CMAKE_OSX_ARCHITECTURES)
132+
list(LENGTH CMAKE_OSX_ARCHITECTURES _fftw_arch_count)
133+
if(_fftw_arch_count GREATER 1)
134+
set(_fftw_is_universal TRUE)
135+
message(
136+
STATUS
137+
"FFTW: macOS universal binary (${CMAKE_OSX_ARCHITECTURES}): "
138+
"per-architecture SIMD defaults disabled. "
139+
"Use ITK_USE_SYSTEM_FFTW with a universal FFTW to enable SIMD."
140+
)
116141
endif()
117-
else()
118-
# Cross-compiling: conservative architecture-level fallback.
142+
endif()
143+
144+
if(NOT _fftw_is_universal)
119145
if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
146+
# NEON is mandatory in the AArch64 ABI — every arm64 CPU has it.
147+
# Safe for all conda/pip arm64 packages and manylinux aarch64.
120148
set(_fftw_default_neon ON)
121149
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64")
122-
# SSE/SSE2 are baseline on all 64-bit x86 CPUs; AVX/AVX2 not assumed.
150+
# SSE and SSE2 are required by the AMD64 ABI — universally present on
151+
# every 64-bit x86 CPU, including the oldest manylinux build targets.
152+
# Safe for all conda/pip x86_64 packages.
123153
set(_fftw_default_sse ON)
124154
set(_fftw_default_sse2 ON)
155+
# AVX and AVX2 are NOT part of the AMD64 baseline. Auto-enable them
156+
# only when the compiler is already producing those instructions — i.e.
157+
# when the user explicitly asked for a specific micro-architecture via
158+
# -march=native, -mavx2, /arch:AVX2, etc. This compile-time check
159+
# mirrors the approach recommended by seanm in ITK PR #6006:
160+
# "the compiler knows what CPU it's compiling for."
161+
#
162+
# check_c_source_compiles caches its result by variable name. Unset
163+
# the cache entry first so the probe always re-runs against the current
164+
# CMAKE_C_FLAGS; this ensures that adding -march=native on a subsequent
165+
# configure is correctly reflected in the auto-detected default.
166+
# Note: FFTW_ENABLE_AVX / FFTW_ENABLE_AVX2 follow standard option()
167+
# caching — they are only auto-set from the detected default when not
168+
# already present in the cache. To force re-evaluation of the option
169+
# after a FLAGS change, delete those entries from the CMake cache or
170+
# pass -DFFTW_ENABLE_AVX2=ON explicitly.
171+
unset(_fftw_compiler_targets_avx CACHE)
172+
check_c_source_compiles(
173+
"#if !defined(__AVX__) || !__AVX__\n#error AVX not enabled\n#endif\nint main(void){return 0;}"
174+
_fftw_compiler_targets_avx
175+
)
176+
if(_fftw_compiler_targets_avx)
177+
set(_fftw_default_avx ON)
178+
endif()
179+
unset(_fftw_compiler_targets_avx2 CACHE)
180+
check_c_source_compiles(
181+
"#if !defined(__AVX2__) || !__AVX2__\n#error AVX2 not enabled\n#endif\nint main(void){return 0;}"
182+
_fftw_compiler_targets_avx2
183+
)
184+
if(_fftw_compiler_targets_avx2)
185+
set(_fftw_default_avx2 ON)
186+
endif()
187+
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "i686|i386")
188+
# 32-bit x86 ABI does not mandate SSE/SSE2. Leave defaults OFF;
189+
# users may opt in explicitly if their minimum target CPU supports them.
125190
endif()
126191
endif()
127192

128193
option(
129194
FFTW_ENABLE_NEON
130-
"Enable FFTW NEON SIMD codelets (ARM64)"
195+
"Enable FFTW NEON SIMD codelets (ARM64; ON by default on aarch64/arm64)"
131196
${_fftw_default_neon}
132197
)
133198
option(
134199
FFTW_ENABLE_SSE
135-
"Enable FFTW SSE SIMD codelets (x86)"
200+
"Enable FFTW SSE SIMD codelets (x86; ON by default on x86_64 — required by AMD64 ABI)"
136201
${_fftw_default_sse}
137202
)
138203
option(
139204
FFTW_ENABLE_SSE2
140-
"Enable FFTW SSE2 SIMD codelets (x86)"
205+
"Enable FFTW SSE2 SIMD codelets (x86; ON by default on x86_64 — required by AMD64 ABI)"
141206
${_fftw_default_sse2}
142207
)
143208
option(
144209
FFTW_ENABLE_AVX
145-
"Enable FFTW AVX SIMD codelets (x86)"
210+
"Enable FFTW AVX SIMD codelets (Sandy Bridge+; OFF by default for redistribution safety)"
146211
${_fftw_default_avx}
147212
)
148213
option(
149214
FFTW_ENABLE_AVX2
150-
"Enable FFTW AVX2 SIMD codelets (x86)"
215+
"Enable FFTW AVX2 SIMD codelets (Haswell+; OFF by default for redistribution safety)"
151216
${_fftw_default_avx2}
152217
)
153218

0 commit comments

Comments
 (0)