You could try to build with the APU support - then you get
(single-precision only)
FP hardware support; it should work with normal context switching since
the APU
shares the registers with the integer processor.
HTH
- Till
On 06/06/2018 02:10 PM, Michael Davidsaver wrote:
Hi Matt,
The MVME3100 is based on the MPC8540 processor. This chip doesn't have an FPU.
Other boards do (eg. mvme270x, mvme6100, or mvme2500) and RTEMS supports this.
I can certainly confirm that lack of an FPU really hurts performance, even
something as (apparently) simple as integer to float.
Michael
On 06/06/2018 02:03 PM, Matt Rippa wrote:
It appears we've compiled our mvme3100 bsp without FPU support.
We're on the mvme3100 running EPICS 3.14.12.7 and RTEMS 4.10.2.
We're attempting to commission our astrometric kernel control system
and we've run into trouble meeting some of our thread deadlines. What
we've found is a transform routine that takes our mvme2700 ~200 us,
takes the mvme3100 ~8 ms. So not a surprise to see this with only
"-msoft-float"?
...
2018-05-31-tcs.log:May 31 11:41:15 E) PORT: tcs_vme, MSG: Welcome to
rtems-4.10.2(PowerPC/Generic (*no FPU*)/mvme3100)
Is FPU support possible with RTEMS 4.10.2? If so, what would we need to do?
It looks like there's a multilib on 4.11.2
https://docs.rtems.org/releases/rtems-docs-4.11.2/cpu-supplement/powerpc.html#multilibs(*Eg.,
see #13* )
Here's our bsp specs:
beatnik:
-specs bsp_specs -qrtems -fasm -c -mcpu=7400 -D__ppc_generic -DUNIX -O2 -g -g
-Wall
(mvme2700) mvme-2307:
-specs bsp_specs -qrtems -fasm -c -mcpu=604 -mmultiple -mstring -mstrict-align
-meabi -DUNIX -O2 -g -g -Wall
mvme-3100
-specs bsp_specs -qrtems -fasm -c -mcpu=powerpc *-msoft-float* -D__ppc_generic
-DUNIX -O2 -g -g -Wall
Thank you!
-Matt
_______________________________________________
users mailing list
users@rtems.org
http://lists.rtems.org/mailman/listinfo/users
