libs/libm: Add fmaf, using fmadd.s on RISC-V targets with an FPU. by anatol · Pull Request #20553 · apache/nuttx

anatol · 2026-10-09T10:22:48Z

Summary

Add a portable fmaf() implementation for targets where double has a 53-bit mantissa. It uses a correction step to handle halfway and subnormal cases that a simple binary64 multiply and add can round incorrectly.
Add an optional RISC-V implementation using the fmadd.s instruction. Enable it in the canmv230 (K230) defconfigs.
This fixes link failures in K230 builds when GCC emits calls to fmaf().
The portable implementation is adapted from musl, with its license and upstream revision recorded in the source.
Related NuttX or NuttX Apps issue/PR: N/A.

Impact

New feature: YES. Adds fmaf() to libm on supported targets and an optional RISC-V FMA implementation.
User impact: YES. Applications can link against fmaf() where double has a 53-bit mantissa. The RISC-V optimization is enabled through RISCV_LIBM.
Build impact: YES. Adds the portable source to Make and CMake builds, with an architecture-specific source selected when LIBM_ARCH_FMAF is enabled.
Hardware impact: YES. Adds an optional RISC-V FPU optimization and enables it in six canmv230 defconfigs.
Documentation impact: YES. This PR adds libm functionality; confirm whether a user documentation update is needed.
Security impact: NO known impact.
Compatibility impact: NO intended impact. The portable implementation is omitted when double does not have a 53-bit mantissa.
Other: The RISC-V implementation uses fmadd.s and relies on the hardware to apply the active rounding mode and update exception flags.

Testing

Host: Linux x86_64, glibc reference.
Target: bare-metal RISC-V under qemu-system-riscv64 11.1.2 (virt, rv64imafdc, lp64d).
Build host/compiler: Linux x86_64; GCC 16.2.1 and Clang 23.1.1 for host regression; riscv64-elf-gcc 16.2.0 for NuttX builds.
Portable implementation regression: testing/fmaf/test_fmaf.c from NuttX Apps commit 298e3887a. Compared result bits and invalid, overflow, underflow, and inexact flags against host libm in all four rounding modes. All 4,072,672 cases
matched across the listed GCC and Clang optimization and sanitizer builds.
RISC-V comparison: portable and fmadd.s implementations matched in 1,072,672 cases across four rounding modes in a bare-metal QEMU harness.
NuttX builds: canmv230:nsh with RISCV_LIBM=y selected lib_fmaf_riscv.o; portable implementation builds were checked with RISCV_LIBM disabled and on rv-virt:nsh64, rv-virt:nsh with ARCH_FPU disabled, esp32c3-legacy- devkit:elf, and linum-stm32h753bi:buzzer. No warnings reported.
nxstyle passed on the new C files.

Testing logs before change:
Not available.

Testing logs after change:
Regression and build results are summarized above. Real-world hardware build and runtime logs still need to be added; the RISC-V runtime comparison used QEMU.

PR verification Self-Check

This PR introduces only one functional change.
I have updated all required description fields above.
My PR adheres to the Contributing Guidelines and Documentation.
My PR is still work in progress (not ready for review).
My PR is ready for review and can be safely merged into a codebase.

github-actions · 2026-10-09T14:00:38Z

MemBrowse Memory Report

qemu-armv8a

Code: .text.init_builtin_run +60 B (+0.0%, 347,480 B)

qemu-intel64

Code: .text +74 B (+0.0%, 8,682,514 B)
No memory changes detected for:
arduino-mega2560
esp32-devkitc
hifive1-revb
mirtoo
rx65n-rsk2mb
s698pm-dkit
stm32-nucleo-f103rb

xiaoxiang781216 · 2026-10-11T08:29:52Z

@anatol please squash your patch into one.

cederom

Thank you @anatol please update PR description according to https://github.com/apache/nuttx/blob/master/CONTRIBUTING.md :-)

anatol · 2026-10-11T09:26:16Z

Thank you @anatol please update PR description according to https://github.com/apache/nuttx/blob/master/CONTRIBUTING.md :-)

Hi @cederom thank you for the feedback. i updated the PR description as well. PTAL.

xiaoxiang781216 · 2026-10-11T10:36:10Z

@anatol please fix cmake style error:


cmake-format check failed, run following command to update the style:
  $ cmake-format <src> -o <dst>

ERROR __main__.py:618: Check failed: /home/runner/work/nuttx/nuttx/nuttx/libs/libm/libm/risc-v/CMakeLists.txt
Some checks failed. For contributing guidelines, see:
  https://github.com/apache/nuttx/blob/master/CONTRIBUTING.md

The public math header declares fmaf(), but the NuttX libm does not provide it. An unoptimized K230 Sigmoid build therefore fails to link when GCC leaves the fused operations as calls to the standard function. Add a portable, correctly rounded fmaf() adapted from musl (MIT). The license header and the pinned upstream revision are kept. It adds in binary64, where the binary32 product is exact, and rounds that sum to odd if it is a binary32 halfway or subnormal value, so that the second rounding to binary32 neither picks the wrong neighbour nor misses the underflow and inexact flags. It needs a binary64 double: it is not built where DBL_MANT_DIG is not 53, instead of rounding wrongly. On RISC-V targets with an FPU, add the LIBM_ARCH_FMAF option, selected by the new RISCV_LIBM ("Architecture specific FPU optimizations", which depends on ARCH_FPU and LIBM), to use the fmadd.s instruction instead, in both the Make and CMake builds. This follows ARMV7M_LIBM: it is a user option, enabled here in the canmv230 (K230) defconfigs. The host regression for the portable fmaf() is testing/fmaf/test_fmaf.c in nuttx-apps (298e3887a, "testing: Add host fmaf regression program"). It compares the portable fmaf() with the host libm: result bits and the invalid, overflow, underflow and inexact flags, in all four rounding modes. Besides boundary, tie and random inputs it has the double rounding traps, whose exact sum is just below or above a binary32 halfway value but rounds to it in binary64. Random inputs hit those with a probability near 2^-29. Testing Host, Linux x86_64 with glibc as the reference: test_fmaf.c built with gcc 16.2.1 (-O0, -O2, -O3) and clang 23.1.1 (-O0, -O3), and with both at -O2 with ASan and UBSan. All 4,072,672 cases match in the four rounding modes (boundary, tie and double rounding trap cases, and 1,000,000 random cases per mode). It fails for each of four broken fmaf() variants: the plain (float)((double)x * y + z), no correction at all, no halfway correction, and no subnormal correction. RISC-V, bare metal on qemu-system-riscv64 11.1.2 (virt, rv64imafdc, lp64d), with a throwaway harness that is not part of this commit: the fmadd.s version and the portable version return the same bits and exception flags in 1,072,672 cases, in the four rounding modes. The comparison also fails with the portable version without the halfway correction. Build, Make with riscv64-elf-gcc 16.2.0: canmv230:nsh (K230, with RISCV_LIBM=y) builds lib_fmaf_riscv.o, which is a single fmadd.s, and not lib_fmaf.o. The portable lib_fmaf.o is used when RISCV_LIBM is disabled and on rv-virt:nsh64 and rv-virt:nsh with ARCH_FPU disabled, esp32c3-legacy-devkit:elf and linum-stm32h753bi:buzzer, with no warnings. nxstyle passes on the new C files. Assisted-by: Claude Code:claude-sonnet-5-5 Signed-off-by: Anatol Pomozov <anatol.pomozov@gmail.com>

anatol · 2026-10-11T12:30:49Z

@xiaoxiang781216 Given that LIBM_ARCH_FMAF is disabled by default - should i at least enable it for k230 targets?

xiaoxiang781216 · 2026-10-11T14:38:53Z

@xiaoxiang781216 Given that LIBM_ARCH_FMAF is disabled by default - should i at least enable it for k230 targets?

yes, you can.

anatol · 2026-10-11T15:30:43Z

@xiaoxiang781216 Given that LIBM_ARCH_FMAF is disabled by default - should i at least enable it for k230 targets?

yes, you can.

i doublechecked the changes and i see that LIBM_ARCH_FMAF is automaticlly enabled if LIBM and FPU enabled. So for k230 it already y. No need to go with my suggestion.

fixed

cederom

Thank you @anatol :-)

anatol requested review from jerpelea and xiaoxiang781216 as code owners October 9, 2026 10:22

xiaoxiang781216 reviewed Oct 9, 2026

View reviewed changes

Comment thread tools/test_fmaf.c Outdated

Comment thread libs/libm/libm/Kconfig Outdated

Comment thread libs/libm/libm/Make.defs Outdated

Comment thread libs/libm/libm/Make.defs Outdated

github-actions Bot added Area: OS Components OS Components issues Size: M The size of the change in this PR is medium labels Oct 9, 2026

anatol force-pushed the master branch from fe48005 to 8f16507 Compare October 11, 2026 00:05

anatol requested review from anchao, eren-terzioglu, lupyuen and pussuw as code owners October 11, 2026 00:05

anatol force-pushed the master branch from 8f16507 to 21c67d9 Compare October 11, 2026 00:20

anatol mentioned this pull request Oct 11, 2026

testing: Add host fmaf regression program apache/nuttx-apps#3830

Open

xiaoxiang781216 reviewed Oct 11, 2026

View reviewed changes

Comment thread arch/risc-v/Kconfig Outdated

cederom previously requested changes Oct 11, 2026

View reviewed changes

anatol force-pushed the master branch from 21c67d9 to 929ffb8 Compare October 11, 2026 09:17

anatol requested a review from yf13 as a code owner October 11, 2026 09:17

anatol force-pushed the master branch from 929ffb8 to 487027f Compare October 11, 2026 09:18

github-actions Bot added the Board: risc-v label Oct 11, 2026

xiaoxiang781216 previously approved these changes Oct 11, 2026

View reviewed changes

anatol dismissed xiaoxiang781216’s stale review via ae3893b October 11, 2026 11:20

anatol force-pushed the master branch from 487027f to ae3893b Compare October 11, 2026 11:20

xiaoxiang781216 approved these changes Oct 11, 2026

View reviewed changes

cederom approved these changes Oct 11, 2026

View reviewed changes

xiaoxiang781216 merged commit f5999d9 into apache:master Oct 12, 2026
53 checks passed

Navigation Menu

Repository navigation

libs/libm: Add fmaf, using fmadd.s on RISC-V targets with an FPU. - #20553

libs/libm: Add fmaf, using fmadd.s on RISC-V targets with an FPU.#20553
xiaoxiang781216 merged 1 commit into
apache:masterfrom
anatol:master

anatol commented Oct 9, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Oct 9, 2026 •

edited

Loading

Uh oh!

Uh oh!

xiaoxiang781216 commented Oct 11, 2026

Uh oh!

cederom left a comment

Uh oh!

anatol commented Oct 11, 2026

Uh oh!

xiaoxiang781216 commented Oct 11, 2026

Uh oh!

anatol commented Oct 11, 2026 •

edited

Loading

Uh oh!

xiaoxiang781216 commented Oct 11, 2026

Uh oh!

anatol commented Oct 11, 2026

Uh oh!

cederom left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

anatol commented Oct 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Impact

Testing

PR verification Self-Check

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Oct 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

MemBrowse Memory Report

qemu-armv8a

qemu-intel64

Uh oh!

Uh oh!

xiaoxiang781216 commented Oct 11, 2026

Uh oh!

cederom left a comment

Choose a reason for hiding this comment

Uh oh!

anatol commented Oct 11, 2026

Uh oh!

xiaoxiang781216 commented Oct 11, 2026

Uh oh!

anatol commented Oct 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

xiaoxiang781216 commented Oct 11, 2026

Uh oh!

anatol commented Oct 11, 2026

Uh oh!

cederom left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

anatol commented Oct 9, 2026 •

edited

Loading

github-actions Bot commented Oct 9, 2026 •

edited

Loading

anatol commented Oct 11, 2026 •

edited

Loading