Go to advanced search

by mstorsjo
Thu May 12, 2016 5:53 am
Forum: Bare metal, Assembly language
Topic: Entering aarch64 execution state
Replies: 171
Views: 76468

Re: Entering aarch64 execution state

Probably related to /etc/udev/rules.d/70-persistent-net.rules That attempts to give the same "eth0"/"eth1"/etc names to the same ethernet adapters (identified by MAC address). I suspect that random MAC addresses are causing increasing "eth7"/"eth8" etc names to be used. I believe deleting /etc/udev...
by mstorsjo
Wed May 11, 2016 8:25 pm
Forum: Bare metal, Assembly language
Topic: Entering aarch64 execution state
Replies: 171
Views: 76468

Re: Entering aarch64 execution state

FWIW, I tried the prebuilt kernel/boot partition, and got it working fine with a normal 32 bit raspbian userland. (This allows me to run statically linked 64 bit binaries at least.) I ran into some issues with the networking though. One gotcha that I didn't see mentioned earlier in this thread, was ...
by mstorsjo
Thu Feb 05, 2015 12:34 pm
Forum: Raspbian
Topic: RPI 2 : Raspbian or Debian "armhf" ?
Replies: 34
Views: 21197

Re: RPI 2 : Raspbian or Debian "armhf" ?

Apps that can use NEON will in many cases already have runtime detection for it (e.g. on Android, the armeabi-v7a baseline does not include NEON so all such code needs to be runtime enabled) - in these cases, one single build will work just fine, and will give just about as good performance (as long...
by mstorsjo
Thu Aug 28, 2014 3:36 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

are there any interactions with "try to make it go fast" settings I have got turned on, like -mfpu=vfp -mfloat-abi=hard -ffast-math -ftree-vectorize? I will try dropping this in and turn off all my 'manually catch denormals' to see what happens. And thanks in advance for the snippet. There's not mu...
by mstorsjo
Thu Aug 28, 2014 3:20 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

Ijust found this post related to a problem I'm having. I'm apparently gettign denormals generated (my code does a lot of IIRs and these things will happen) and when it happens performance plummets - can anyone advise on how to get g++ to put the device into a mode where denormals are treated as zer...
by mstorsjo
Mon Feb 25, 2013 1:29 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

The full patch by Russell is now available at http://lists.arm.linux.org.uk/lurker/me ... 8c.en.html.
by mstorsjo
Mon Feb 25, 2013 1:07 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

Thanks for the deep investigations, especially for the very educative bare metal example by David. After looking ever closer on it in the linux setup, I noticed that the issue seems to be that the VFP exception is triggered much later. In the following code snippet: 8424: ee312ac1 vsub.f32 s4, s3, s...
by mstorsjo
Sun Feb 24, 2013 8:05 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

The kernel handlers do the same calculation using a SW implementation of it - that's the one that actually does the calculation 0x00012345 - 0x00000000 = 0x00012345. If enabling runfast mode (that disables using the kernel handler), this one ends up with just plain zero (but doesn't disturb the seco...
by mstorsjo
Sun Feb 24, 2013 7:14 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

FWIW, yet another condition for this error happening seems to be that none of the registers used in the first VFP instruction should be referenced in the second one. As soon as I change any of the registers in the second instruction (either one of the operand registers or the destination register) t...
by mstorsjo
Sun Feb 24, 2013 6:11 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

I've narrowed down the issue even further, with a pretty minimal reproduction example at http://albin.abo.fi/~mstorsjo/denormal-vfp.s (and a corresponding binary at http://albin.abo.fi/~mstorsjo/denormal-vfp). The meat of the example is: flds s0, [r4, #0] @ 0x00000000 flds s1, [r4, #4] @ 0x3f800000 ...
by mstorsjo
Sat Feb 23, 2013 10:55 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

I've managed to narrow the issue down even further, with an updated C version of the test app at http://albin.abo.fi/~mstorsjo/denormal-fft.c (only exhibiting the issue if built with -O2) and a hand-written assembly version at http://albin.abo.fi/~mstorsjo/denormal-fft-asm.s. Binaries of the same ar...
by mstorsjo
Fri Feb 22, 2013 3:43 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

What different versions of VFP would there be involved in this case? I can reproduce the same issue both using the toolchain from https://github.com/raspberrypi/tools and using the raspbian default compiler, both of them without using any compiler flags that alter what's targeted, only -O2. And yeah...
by mstorsjo
Fri Feb 22, 2013 2:06 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

Re: VFP issues with denormal numbers

Yes, normally I'd suspect the code itself as well, but this is odd since the same static binary works on a number of other ARM+VFP cpus, except for this one.

I'll see if I can simplify the test case further into fixed VFP instruction sequences that fail.
by mstorsjo
Thu Feb 21, 2013 11:00 pm
Forum: Bare metal, Assembly language
Topic: VFP issues with denormal numbers
Replies: 25
Views: 5363

VFP issues with denormal numbers

Hi, I've run into an issue where the Raspberry PI seems to calculate incorrect (sometimes totally incorrect, other times just slightly wrong) values, when doing floating point calculations using VFP involving denormal numbers. I ran into the issue in AAC decoder tests in libav, and I've managed to n...

Go to advanced search