linuxdc-dev

[linuxdc-dev]Floating point problems - Dreamcast 2.4.10-pre6 kernel

[Adrian M. <ad...@mc...>]

I have been working to make/port a decent mp3 player for the Dreamcast now 
that there is a sound driver.

I have built mpg123 with these options:

dream:
	$(MAKE) CC=sh4-linux-gcc LDFLAGS= \
	OBJECTS='decode.o dct64.o audio_oss.o' \
	CFLAGS='-DREAL_IS_FLOAT -DLINUX -Wall -O2 \
		-DOSS \
		-fomit-frame-pointer -funroll-all-loops \
		-finline-functions -ffast-math' \
		mpg123-make

but this always generates a floating point exception at run time.

I have been able to isolate this to the following function:

costab[k] = 1.0 / (2.0 * cos( 3.14159265358979323846 * ((double) k * 2.0 + 
1.0) / (double) divv));

This is looped 0 < k < kr With k as 0 it costab[k] is returned as NaN and 
with k = 1 the floating point exception is generated. kr is 16 and divv is 64.

My maths indicates costab[0] should be roughly 0.5.

Incuidentally with -DREAL_IS_FLOAT switched off then there is no fp exception 
but the program outputs static.

So: the questions are - is this a known problem, has it been fixed in later 
kernels or have I just got it wrong?


Adrian

[linuxdc-dev]Re: [linuxsh-dev] Floating point problems - Dreamcast 2.4.10-pre6 kernel

[Adrian M. <ad...@mc...>]

Additional point: I am using gcc 3.0.1 as the cross compiler

[linuxdc-dev][linuxsh-dev] Floating point problems - Dreamcast 2.4.10-pre6 kernel

[NIIBE Y. <gn...@m1...>]

Adrian McMenamin wrote:
 > costab[k] = 1.0 / (2.0 * cos( 3.14159265358979323846 * ((double) k * 2.0 + 
 > 1.0) / (double) divv));
[...]
 > So: the questions are - is this a known problem, has it been fixed in later 
 > kernels or have I just got it wrong?

I don't know if this is related, we have a FPU bug wrt floating point
to double conversion.  For example, compiling runtime of Java, GCC
fails because of exception.

I have experimental fix for this issue, I'll send this to linuxsh-dev
soon.
-- 

[linuxdc-dev]Re: [linuxsh-dev] Floating point problems - Dreamcast 2.4.10-pre6 kernel

[Adrian M. <ad...@mc...>]

On Friday 28 Dec 2001 2:54 am, NIIBE Yutaka wrote:
> Adrian McMenamin wrote:
>  > costab[k] = 1.0 / (2.0 * cos( 3.14159265358979323846 * ((double) k * 2.0
>  > + 1.0) / (double) divv));
>
> [...]
>
>  > So: the questions are - is this a known problem, has it been fixed in
>  > later kernels or have I just got it wrong?
>
> I don't know if this is related, we have a FPU bug wrt floating point
> to double conversion.  For example, compiling runtime of Java, GCC
> fails because of exception.
>
> I have experimental fix for this issue, I'll send this to linuxsh-dev
> soon.

Just to close this. I rebuilt my tools and upgraded to gcc 3.0.3 and now the 
programme compiles and runs fine.

Not sure what the problem was exactly - could have been a badly patched 
compiler, could have been a corrupted library - but it works now.

Adrian

[linuxdc-dev]New pre-release version of Dreamcast AICA sound driver (dsp and mixer) available for testing

[Adrian M. <ad...@mc...>]

A new, much improved, pre-release version of the open sound system (OSS) 
compatible AICA sounddriver for Linux on the Dreamcast is available for 
testing.

This version implements a new buffering mechanism that has meant continuous 
playing of PCM and MP3 samples. The driver needs to be tested in very high 
demand environments (eg MPEG video) as there is potential for synchronisation 
problems - I haven't ported any such applications, but I know others have and 
could test this.

Anyway:

The files are at:

http://www.mcmen.demon.co.uk/linuxdc/aica

You need aica.c and arm7.h to build the drivers

However, as this is an OSS driver you will also need a minimal OSS system on 
your DC. The files and a smaple makefile are also available at the above 
address to build this. Simply download, alter to fit your set up and install 
with insmod soundcore.o and insmod aica.o


Please feel free to post any questions to me or (preferably) the linuxdc-dev 
(and others if you feel it appropriate - but don't spam people) list.


Happy New Year to all,

Adrian

Re: [linuxdc-dev]Floating point problems - Dreamcast 2.4.10-pre6 kernel

[M. R. B. <mr...@0x...>]

* Adrian McMenamin <ad...@mc...> on Thu, Dec 27, 2001:

> I have been working to make/port a decent mp3 player for the Dreamcast no=
w=20
> that there is a sound driver.
>=20
> I have built mpg123 with these options:
>=20
> dream:
> 	$(MAKE) CC=3Dsh4-linux-gcc LDFLAGS=3D \
> 	OBJECTS=3D'decode.o dct64.o audio_oss.o' \
> 	CFLAGS=3D'-DREAL_IS_FLOAT -DLINUX -Wall -O2 \
> 		-DOSS \
> 		-fomit-frame-pointer -funroll-all-loops \
> 		-finline-functions -ffast-math' \
> 		mpg123-make
>=20
> but this always generates a floating point exception at run time.
>=20

I'm not to keen on how the flags function for the SH backend of GCC (maybe
NIIBE-san knows?) but I always thought -ffast-math was a i386-based (not
necessarily specific) optimization.  I don't think it's wise to use it here
=2E.. although it may be a no-op.  Only a read through GCC's source will
confirm that.

>=20
> Incuidentally with -DREAL_IS_FLOAT switched off then there is no fp excep=
tion=20
> but the program outputs static.
>=20

Have you tried with -m4-single-only?  It generates FP code conforming to
what its name implies.

> So: the questions are - is this a known problem, has it been fixed in lat=
er=20
> kernels or have I just got it wrong?
>=20

I don't see how it could be a kernel issue.

M. R.

Re: [linuxdc-dev]Floating point problems - Dreamcast 2.4.10-pre6 kernel

[M. R. B. <mr...@0x...>]

* Adrian McMenamin <ad...@mc...> on Thu, Dec 27, 2001:

> I have been working to make/port a decent mp3 player for the Dreamcast no=
w=20
> that there is a sound driver.
>=20
> I have built mpg123 with these options:
>=20
> dream:
> 	$(MAKE) CC=3Dsh4-linux-gcc LDFLAGS=3D \
> 	OBJECTS=3D'decode.o dct64.o audio_oss.o' \
> 	CFLAGS=3D'-DREAL_IS_FLOAT -DLINUX -Wall -O2 \
> 		-DOSS \
> 		-fomit-frame-pointer -funroll-all-loops \
> 		-finline-functions -ffast-math' \
> 		mpg123-make
>

Update: Using the above line in mpg123/Makefile (with dream changed to
linux-sh for consistency), I built mpg123 and executed it on the Dreamcast =
...
no problems whatsoever.

You must have a snag in your toolchain somewhere.  I'm using (all patched
with the latest patchset from ftp.m17n.org/pub/super-h/testing/) binutils
2.11.2, GCC 3.0.3, and glibc 2.2.4.  This was executed under the busybox
shell (the sh-like one) under kernel 2.4.16.

M. R.