How much NEON state should be saved in a context switch? (xilinx zynq)
Stephen Tether
tether at slac.stanford.edu
Fri Jan 16 23:22:47 UTC 2015
We have an RTEMS application that makes heavy use of the SIMD features
of the NEON on a Cortex-A9; without them the speed degrades by a factor
of 15. This is hand-written assembler code using integer and logical
vector operations, not compiler-generated floating-point code.
So far only one coder is using the NEON in one task. Others would like
to use the NEON, for example in networking code, but can only do so
safely if context switches handle NEON state correctly.
I see that RTEMS can be built with NEON/VFP3-D32 support but when I look
at the context switching function I find that it saves and restores only
registers D8-D15. That doesn't seem like enough to me; wouldn't we need
to save and restore all 32 registers if a context switch can occur at
any time and if code using NEON makes use of all the registers?
I know that D8-D15 are the registers that must be preserved across
function calls according to the ABI, but saving/restoring just those
registers on a context switch would seem to work only if all tasks were
non-preemptable. In that case I believe that context switches would
occur only inside functions such as yield(), sleep(), blocking I/O, etc.
- Steve Tether
More information about the users
mailing list