AM4378: Double precision floating point questions

Colombo Carlo

Mastermind 24070 points

Part Number: AM4378

Hi everybody ,

on AM437x we have Neon ( single precision ) and VFPv3 ( double precision ) ( while on AM57xx we have VFPv4 ) .

so please some questions :

1- you need to invoke VFPvx to get HW double precision support in A8-A9-A15 ( 32 bits architecture ) , am I right ?

2-do you have any benchmark /numbers showing VFP capabilities in handle double precision ? I m very interested in particular in filtering

3- do you have any comparison between VFPv3 vs VFPv4 ?

thank you very much

regards

Carlo

over 7 years ago

0 Biser Gatchev-XID over 7 years ago

TI__Guru**** 393215 points

The factory team have been notified. They will respond here.

0 jian35385 over 7 years ago

TI__Mastermind 23125 points

Carlo,

You are correct that AM335x and AM437x devices implemented the ARMv7-A architecture WITH the VFPv3-D16 FPU. The AM57xx devices implemented the VFPv4-D16 FPU. They all support double precision floating-point operations.

Main enhancement of VFPv4 vs. VFPv3 are the half-precision support and fused multiply-add instructions. See:
developer.arm.com/.../floating-point
for further details.

Regarding to performance benchmark, there are some general benchmarks from independent sources, to name a few below:
www.youtube.com/watch
www.anandtech.com/.../exploring-the-floating-point-performance-of-modern-arm-processors

Also note that compiler options must be specified for given VFP configurations. For example, the gcc compiler would need to specify -mfpu to specify. See details on:
gcc.gnu.org/.../ARM-Options.html
The -mfpu=name specifies what floating-point hardware (or hardware emulation) is available on the target. Permissible names are: ‘vfp’, ‘vfpv3’, ‘vfpv3-fp16’, ‘vfpv3-d16’, ‘vfpv3-d16-fp16’, ‘vfpv3xd’, ‘vfpv3xd-fp16’, ‘neon’, ‘neon-fp16’, ‘vfpv4’, ‘vfpv4-d16’, ‘fpv4-sp-d16’, ‘neon-vfpv4’, ‘fp-armv8’, ‘neon-fp-armv8’, and ‘crypto-neon-fp-armv8’.
If -msoft-float is specified this specifies the format of floating-point values.

Let us know if you need further clarifications.
Jian

0 Thomas Smith48 over 7 years ago in reply to jian35385

Prodigy 90 points

Please forgive me, but I would like to ask a question to further my understanding for the fpu use. If I want to build a user space application, what settings do I need so that the app will use the fpu on the AM335x, AM437x and the AM57xx. Meaning, if I set the toolchain to build the app with --with-arch=armv7-a --with-tune=generic-armv7-a --with-fpu=vfpv3-d16 --with-float=hard will the app run on all of the devices without any issues? Is the vfpv4-d16 backward compatible to vfpv3-d16?

Thanks for the help,

Ben

0 jian35385 over 7 years ago in reply to Thomas Smith48

TI__Mastermind 23125 points

Ben let me check with our team on specific options needed for TI tools.
Sorry for the delayed response I was sidetracked.
Jian

0 jian35385 over 7 years ago in reply to jian35385

TI__Mastermind 23125 points

Ben,

The TI Linux SDK points to uses gcc, that have built-in options. you can turn-on the "-v" when compile, to report all options. See below from our SDK team:

> I did this for GCC v7.2 and I got this:
> $ arm-linux-gnueabihf-gcc -v -c -o hello.o hello.c
> [lots of stuff]
> COLLECT_GCC_OPTIONS='-v' '-c' '-o' 'hello.o' '-march=armv7-a'
> '-mtune=cortex-a9' '-mfloat-abi=hard' '-mfpu=vfpv3-d16' '-mthumb'
> '-mtls-dialect=gnu'

to your second question, yes if you build for --with-fpu=vfpv3-d16 that should run on VFPv4 devices.

regards
Jian

0 jian35385 over 7 years ago in reply to jian35385

TI__Mastermind 23125 points

Typo:

The TI Linux SDK points to uses Linero tools.

Jian

0 jian35385 over 7 years ago in reply to jian35385

TI__Mastermind 23125 points

i will close the ticket per our internal tracking. please reopen if you additional questions.
jian

Processors

Processors forum

AM4378: Double precision floating point questions