[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 16 Sep 2020 15:49:28 -0500
From: <sidneym@...eaurora.org>
To: "'Szabolcs Nagy'" <nsz@...t70.net>
Cc: "'Rich Felker'" <dalias@...c.org>,
<musl@...ts.openwall.com>
Subject: RE: Hexagon DSP support
> -----Original Message-----
> From: sidneym@...eaurora.org <sidneym@...eaurora.org>
> Sent: Friday, July 24, 2020 12:50 PM
> To: 'Szabolcs Nagy' <nsz@...t70.net>
> Cc: 'Rich Felker' <dalias@...c.org>; 'musl@...ts.openwall.com'
> <musl@...ts.openwall.com>
> Subject: RE: [musl] Hexagon DSP support
>
>
>
> > -----Original Message-----
> > From: Szabolcs Nagy <nsz@...t70.net>
> > Sent: Thursday, July 23, 2020 4:56 PM
> > To: sidneym@...eaurora.org
> > Cc: 'Rich Felker' <dalias@...c.org>; musl@...ts.openwall.com
> > Subject: Re: [musl] Hexagon DSP support
> >
> > * sidneym@...eaurora.org <sidneym@...eaurora.org> [2020-07-20
> > 16:26:58 -0500]:
> > > I removed fma/fmal/fmax/fmin/fabs from compiler-rt-builtins,
> > > https://reviews.llvm.org/D82263
> > > The comparison with musl can be found here:
> > > https://github.com/quic/musl/compare/hexagon but I've also attached
> > > the patch.
> > >
> > > An assert in clang when building both musl and libc-test for hexagon
> > > was fixed by, https://reviews.llvm.org/D80952 prior to this change
> > > -frounding-math had to be used.
> > >
> > > The test-results are also attached. Everything is built with the
> > > tip-of-tree llvm so sometimes results vary but these are the results
> > > I got from this morning's clone. The only notable difference in the
> > > results would be that both fma and fmal fail and this is because of
> > > the compiler-rt change. I didn't add fma to musl because it require
> > > more complex assembly, along the lines you saw in an earlier version
> > > with
> > sqrt.
> >
> >
> > the fma and sqrt failures are still not fully explained, e.g. this looks
wrong:
> >
> > src/math/special/fma.h:42: RN fma(0x1p+0,0x1p+0,-0x1p-1074) want
> > 0x1p+0 got -0x1.fffffp-43 ulperr -4503599627370496.000 = -0x1p+52 +
> > 0x0p+0
> >
> > the only target specific bit in fma is a_clz_64 so i would check that.
> >
> > e.g. a_clz_64(1ULL << 42) should give 21 (this computation happens
> > during the fma test case above).
>
> Hexagon didn't have a_clz_64 implemented however I added this morning it
> and noticed no differences. I will update the patch with that routine
> included.
>
> I did notice a compiler regression in how it compiled fma and have asked a
> compiler person to take a look. An older version of our internally
maintained
> compiler does produce the expected results for the values I used but later
> versions do not. Unfortunately changing optimization levels will produce
> different results as well.
I've attached updated test results and patch. The patch doesn't change much
other than adding the above mentioned a_clz_64. The only other change was
an update to pthread_arch.h for an api update so hexagon conforms with the
rest of musl.
Between updates to llvm and musl both fma and sqrt issues are resolved
provided I compile the library without optimization enabled. No new tests
fail.
I guess I also need to know what the thoughts are about adding hexagon to
the mainline base. There are no issues adding from this end.
Thanks,
Download attachment "add-hexagon.diff" of type "application/octet-stream" (41835 bytes)
Download attachment "libc_test_REPORT" of type "application/octet-stream" (41794 bytes)
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
