ADR-1495: icx and icpx builds link glibc's libm, not Intel's libimf¶

Status: Accepted
Date: 2026-10-03
Deciders: lusoris
Tags: build, numerics, icx, sycl, testing, rc3, fork-local

Context¶

The CPU extractors of a build made with Intel's compilers (CC=icx CXX=icpx: every SYCL build, the vmaf of the dev image, the Linux Intel LLVM CI job) returned other scores than a GCC build in the last digits (T-ICX-LIBIMF-HOST-MATH-2026-10-01). After ADR-1415 and ADR-1461 removed the contraction differences, what was left was the math library. Measured on ryzen-4090-arc with icx 2026.0 and GCC 16.2.1, glibc 2.44, --backend cpu --precision max --threads 4, on the Netflix 576x324 pair (48 frames), both 1080p checkerboard pairs, BBB 1280x720 (48 frames; the clip testdata/generate.sh makes, whose scores equal testdata/scores_cpu_720.json on 720 of 720 values) and BBB 3840x2160 (200 frames): 268 of 13288 values differed. integer_adm_scale1 by 7.9e-8 on BBB 1280x720 frame 26, float_adm's adm_scale3 with adm_f1s3=2.25:adm_f2s0=0.3 by 7.5e-8 on Netflix frame 27, ciede2000 on 191 of 200 BBB 3840x2160 frames (6.3e-12), psnr, psnr_hvs and their chroma outputs by up to 1.4e-14, and the default model's vmaf by up to 1.3e-7.

How the Intel library got in, read off the driver (icx -###, the same in 2026.0 and 2026.1.1) and the linked binaries:

The driver appends -lsvml -lirng -limf -lm to every link it runs, and rewrites a -lm it is given into -limf -lm. Meson's math_lib is such a -lm, so libvmaf.so needed libimf.so, did not need libm.so.6, and imported log10, pow, powf, log2f, exp, log and 24 other math functions (atan2, cbrt, tgamma, sqrt, ...) without a glibc symbol version. Putting -lm first cannot work.
An executable takes the Intel libraries statically. The installed man page (/opt/intel/oneapi/compiler/2026.0/share/man/man1/icx.1), on -static-intel: "Default: ON. Intel(R) libraries are linked in statically with the following exceptions: [...] The Intel(R) libraries are linked in dynamically when you specify option shared." The icx-built vmaf therefore carried libimf's copies of all 30 math functions libvmaf.so imports and exported them; with LD_DEBUG=bindings the loader binds libvmaf.so.3's exp, log, pow, ... to tools/vmaf. A program built with GCC binds the same library to glibc first: the dev image's ffmpeg binds the icx-built libvmaf.so.3's six functions to libm.so.6. Which library an icx build's extractors used depended on the program that loaded them.

Decision¶

Every C and C++ link whose linker is Intel LLVM (cc.get_id() or cxx.get_id() is intel-llvm) gets -no-intel-lib=libimf, as Meson project link arguments declared in core/src/meson.build between the markers BEGIN / END VMAF host math library link policy, directly after the strict FP policy and above the first target. The library, the tools and the tests then take the math functions from glibc's libm, as a GCC or clang build does.

The option, as the installed version documents it. icx --help of 2026.0 and of 2026.1.1: "-no-intel-lib=<value> Restrict linking of Intel specific libraries. Valid arguments are libirc, libimf, libirng, libsvml". The man page of 2026.0: "-no-intel-lib[=library] (L*X only) Disables linking to specified Intel(R) libraries, or to all Intel(R) libraries. [...] libimf Disables linking to the Intel(R) oneAPI DPC++/C++ Compiler Math library. [...] NOTE: This option only applies to host compilation. When offloading is enabled, it does not impact device-specific compilation." (The 2026.1.1 image carries no man page; its --help lists the option with the same text, and its driver drops -limf the same way.)

libsvml, libirc and libintlc stay linked. The man page, on -fimf-use-svml: "Default: false. Math library functions are implemented using the Intel(R) oneAPI DPC++/C++ Compiler Math Library, though other compiler options may give the compiler the flexibility to implement math library functions with either LIBM or SVML." With icx's default fast model the vectoriser turns a loop over powf / log2f / exp / log10 into __svml_* calls; under vmaf_strict_fp_args (-fp-model=precise -ffp-contract=off, ADR-1461) it emits none, at -O3 and with -mavx2 -mfma or -mavx512f too. Neither the library of a CPU icx build nor that of a SYCL build imports an __svml_ symbol, before or after this change.

core/test/test_icx_system_libm.py (suite fast) reads the build it runs in: libvmaf.so needs libm.so.6 and not libimf.so, imports the six functions with glibc versions and no __svml_ routine, vmaf defines none of them, and vmaf --version under LD_BIND_NOW=1 LD_DEBUG=bindings binds every reference to them to libm.so.6. On a build whose compilers are not Intel LLVM the binary checks skip with the reason. test_strict_fp_compiler_args executes the policy block for seven compiler pairs and pins the two link argument lines.

Alternatives considered¶

Option	Pros	Cons	Why not chosen
`-lm` ahead of the driver's libraries (link order)	No Intel option	The driver rewrites a given `-lm` into `-limf -lm` (measured, 2026.0 and 2026.1.1)	Does not work
`-shared-intel`	Stops the static copies in the executables	`libimf.so` stays the first math library of the link: the library and the executables bind to it	Changes how libimf is linked, not whether
`-nodefaultlibs` or `-Wl,-lm` around the driver	Bypasses the driver	Drops libc, libgcc and the SYCL runtime libraries the driver adds, or hides the intent in a linker flag the driver does not know about	Fragile
`-no-intel-lib` without a value (all four libraries)	Nothing Intel-specific on the host link	Not needed for the scores; `libirc` holds `_intel_fast_memcpy`, which icx-compiled objects call	Wider than the defect
Also pass the option on the compile lines	The compiler would no longer be told libimf is available (`-fintel-libimf-allowed` leaves the cc1 line)	The 12 math-heavy translation units measured (`adm_tools.c`, `ciede.c`, `psnr_hvs.c`, `psnr.c`, `integer_adm.c`, `float_adm.c`, `speed.c`, `ssimulacra2.c`, `cambi.c`, `vif_tools.c`, `svm.cpp`, `predict.c`) compile to the same machine code with and without it; a call into a libimf-only entry point would fail the link (`--no-undefined`) rather than pass silently	No effect; noise on every compile line
Shared routines with defined rounding for the affected calls (as `ssimulacra2_math.h` does for `cbrtf`)	Independent of any C library	Every CPU extractor that calls `pow`, `log10`, `exp`, ...; changes the scores of the GCC build, which the GPU twins and the snapshots are measured against	The difference is a link property; the reference stays glibc
`-no-intel-lib=libimf` on every Intel LLVM link (chosen)	One documented driver option, host only; the library and every executable bind to glibc whatever program loads them	The host link loses Intel's math library, which was not faster here (measured below)	Chosen

Consequences¶

Positive: an icx build returns a GCC build's CPU scores. Same measurement as above, after: 13288 of 13288 values identical (psnr, psnr_hvs, ciede, integer adm, float_adm plain and with adm_f1s3=2.25:adm_f2s0=0.3, and the default model, on the five fixtures). In the dev container (icx 2026.1.1 against GCC 15.2.0, glibc 2.43, BBB 3840x2160 cut to 20 frames): 5368 of 5368. The CPU half of a SYCL build is the GCC build too, so a SYCL twin is now exact against the GCC build's CPU extractor, not only against its own build's.
Negative: none measured; the icx build got faster. Time per frame on a 48-frame 1920x1080 clip (.corpus/bench-data/ref_1920x1080.yuv, local corpus), --threads 4 pinned to four cores, median of three interleaved runs, load average 11 to 13 from other work on the host: the default model 4.35 ms before and 3.97 ms after; psnr, psnr_hvs, ciede, adm and float_adm together 105.45 ms before and 63.71 ms after (the GCC build: 70.76 ms). All of the change is ciede (95.63 ms before, 54.31 ms after, alone): its math calls are faster in glibc 2.44 than through libimf here; psnr_hvs, psnr, adm and float_adm alone moved by less than 2 %.
Neutral / follow-ups:
SYCL device code is not touched. The compile commands of an icx build are the same before and after (1599 of 1599 entries of compile_commands.json); only the link arguments gain the option. An icpx link of a separately compiled SYCL object (spir64_gen for dg2-g11 and spir64, the precision pair) runs 20 driver jobs, and only the host ld line differs, by -limf: the device link that builds the SPIR-V image is the same. On the Arc A380 a SYCL build with this change (icx 2026.0, AOT for dg2-g11) passes the parity gate over all 24 features on the Netflix pair, both checkerboards, BBB 1280x720 and 20 frames of BBB 3840x2160: the cells of the 23 exact twins read 0; ciede, a bounded cell (1e-9, ADR-1436), reads 1.1e-12 at 3840x2160 (3.5e-12 with the build before this change). The CPU half of that build equals the GCC build on 10126 of 10126 values of those runs, and the twins equal the GCC build's CPU extractor on 9987 of 10004; the 17 others are ciede at 3840x2160. test_sycl_exact_twins, test_sycl_ms_ssim_parity, test_sycl_float_psnr_parity and test_sycl_ciede_parity pass.
The published images pick this up when they are rebuilt; the Containerfile needs no change. The Linux Intel LLVM job runs the whole meson test suite, so it runs the new test; the two SYCL lanes of libvmaf-build-matrix.yml run it as a step of their own.
Windows icx-cl (intel-llvm-cl) links another set of libraries and is not covered: no Windows host was available to measure it (T-ICX-CL-WINDOWS-HOST-MATH-2026-10-03).
The golden gate's build profile (ADR-1317) still refuses icx builds; this change removes one of the reasons, not the decision.
testdata/scores_sycl_a380_720.json holds frame 26's vmaf from an icx build with libimf (88.435637 against the CPU snapshot's 88.435634). The same recording (--backend sycl, vmaf_v0.6.1, A380) from the host SYCL build with this change equals testdata/scores_cpu_720.json on 720 of 720 values, from the build before it on 719. The file is re-recorded in the rebuilt image, not here (ADR-1102; T-SYCL-SNAPSHOTS-STALE-2026-10-02).

References¶

Q (maintainer popup, 2026-10-03, on T-ICX-LIBIMF-HOST-MATH-2026-10-01): "Fix in RC3 (Recommended)".
/opt/intel/oneapi/compiler/2026.0/share/man/man1/icx.1 (-no-intel-lib, -static-intel, -shared-intel, -fimf-use-svml); icx --help of oneAPI 2026.0.0 and 2026.1.1.
ADR-1461 (the strict FP project argument), ADR-1415, ADR-1367 (the SYCL device FP policy), ADR-1475 (the integer ADM frame), ADR-1434, ADR-1317.