Vector TG meeting minutes for 2020/6/19

Krste Asanovic

Attached below. Also a reminder we'll be meeting again on Friday per
group's calendar info.


Date: 2020/6/19
Task Group: Vector Extension
Chair: Krste Asanovic
Co-Chair: Roger Espasa
Number of Attendees: ~12
Current issues on github:

Issues discussed:

#235 reciprocal and reciprocal square-root approximations

This was a long-standing issue regarding adding these floating-point
instructions. The group was in favor of proceeding. The main issues
are 1) what precision of approximation to provide, and 2) should the
instructions be bit-compatible across implementations.

The group wanted to better understand the cost/benefit of the
precision choice, in terms of number of iterations required for
half/single/double precise calculations. While generally it was
preferred that implementations be required to be bit-compatible, the
concern was raised that some implementations might want to implement
these using regular FP operations, which might not be possible if
bit-compatibility (instead of range constraint) was given. A possible
path was requiring an 8b bit-compatible approximation that is
relatively cheap to implement, with possibly more accurate versions

#499 tight possible fractional LMUL settings

The issue was raised that current spec has a one-sided constraint that
implementations must provide fractional LMUL for all supported SEW,
where LMUL >= SEW/ELEN, but that it also implies that implementations
may implement smaller fractional LMUL. This could cause some code to
operate correctly on one implementation that supported a smaller
fractional LMUL, but to fail on a compliant implementation that only
implemented the mandated fractional LMULs. The resolution was to deem
the setting of LMUL < SEW/ELEN to be a "reserved" operation, where it
is legal but not required to set vill. Assemblers that are aware of
ELEN should raise a warning (but should not error) if an LMUL setting
is less than SEW/ELEN.

#487 Implementations with minimum SEW>8

Some partial implementations might not want to implement all element
widths, including omitting byte operations for example. The
description of most instructions still applies, but the whole register
move instructions were defined in terms of SEW=8.

If the register layout issue is resolved with a mandate to keep memory
layout byte order visible in registers, then this issue is moot as SEW
does not affect the visible operation of the whole register move

IF the register layout issue does not mandate to memory layout byte
order visible in registers, then we can reword the whole register move
specification to operate in terms of the minimum supported element
width. This will not affect the standard vector profiles.