Differential D20398

math/fftw3: Add arm64 performance counter and enable neon
AbandonedPublic
Actions

Authored by diizzy on May 24 2019, 3:59 PM.

Details

Reviewers

jhale
manu

Summary

Enable cycle counter via the PMCCNTR_EL0 register on arm64 for all variants
Enable NEON instructions for -float flavor

Test Plan

Runtime tested with bench utility in tests directory (./work/fftw-3.3.8/tests/bench if compiling ports) on Orange Pi PC running 13.0-CURRENT r347967

Diff Detail

Repository

rP FreeBSD ports repository

Lint

Lint Skipped

Unit

Tests Skipped

Event Timeline

diizzy created this revision.May 24 2019, 3:59 PM

Herald added a subscriber: mat. · View Herald TranscriptMay 24 2019, 3:59 PM

val_packett.cool added a subscriber: val_packett.cool.May 25 2019, 8:39 PM

LGFM

This revision is now accepted and ready to land.Jul 2 2019, 8:30 PM

mikael added a subscriber: mikael.Jul 3 2019, 7:06 AM

mikael added inline comments.

math/fftw3/Makefile

missing space after #

65–67

You said 'Enable cycle counter via the PMCCNTR_EL0' but you used enable-armv8-cntvct-el0, can you fix the comment or the configure knob?

--enable-armv8-pmccntr-el0
                        enable the cycle counter on ARMv8 via the
                        PMCCNTR_EL0 register (see README-perfcounters for
                        details and mandatory instructions)
--enable-armv8-cntvct-el0
                        enable the cycle counter on ARMv8 via the CNTVCT_EL0
                        register (see README-perfcounters for details and
                        mandatory instructions)

diizzy marked 2 inline comments as done.Jul 3 2019, 9:37 AM

diizzy added inline comments.

math/fftw3/Makefile
65	That would be inconsistent with the rest of the Makefile https://svnweb.freebsd.org/ports/head/math/fftw3/Makefile?revision=487272&view=markup
65–67	Ahh, silly cut 'n paste mistake. Thanks for catching it.