CORRECTION, minutes from 2020/07/10 meeting
I managed to swap the file names around, and sent earlier week's
minutes instead of last week's. Here's the correct minutes from last
week (also fixed in repo).
Task Group: Vector Extension
Chair: Krste Asanovic
Co-Chair: Roger Espasa
Number of Attendees: ~22
Current issues on github: https://github.com/riscv/riscv-v-spec
#235 reciprocal/square root approximation
Group discussed the new proposal, and will study accuracy and
performance issues on the mailing list.
#533 Mandating VLEN>=128 on application processor profiles
The proposal would provide a guarantee on minimum VLEN to simplify
code that used short vectors. This also ensures some instructions
that rely on more than one element in a vector operate correctly in
all cases. The discussion contemplated a larger value of 256, but
consensus seemed that VLEN>=128 would include common design points.
#392/#520 vrgather index size
The group discussed the problem that vrgather indices when SEW=8 would
not allow byte permutation of even relatively short vector register
groups (at LMUL=8, VLEN=256 is the largest size fully supported). The
proposal was to add a variant vrgatherei16 that would fix EEW of the
indices to 16b.
#529/516 Alignment of whole-register load/store instructions
The EEW given as a hint in whole-register load/store instructions is
currently ignored, but it would simplify implementations to be able to
report misalignment exceptions when the effective base address was not
aligned to an element boundary. Much of the discussion was around
what alignment the code using the instruction could rely on, but the
point was made that the whole-register alignment was usually very well
controlled and the main point of this relaxation was to simplify
#365 should vsetvli x0, x0, imm set vill if vl changes?
Discussion over the behavior in this case focused on whether the cases
when vl changes could ever be useful, in which case the vill setting
would remove functionality. Also, on when errors could be reported.
# Reserving custom space for encodings/vtype
A question was whether 1.0 spec should reserve vector custom space
within current encoding. For now, any unused encoding is reserved.
Discussion will continue on list.
# Categorizing issues
The group was to continue categorizing issues that must be resolved
for V 1.0, versus those for a future extension.