[libre-riscv-dev] texture opcodes

Luke Kenneth Casson Leighton lkcl at lkcl.net
Wed Jun 5 01:21:42 BST 2019


On Wed, Jun 5, 2019 at 1:14 AM Mitchalsup <mitchalsup at aol.com> wrote:
>
> Luke,
>
> I think you should really consider creating texture reference instructions. This would
> dump a crap-load (tm) functionality into a new-different FU where you would throw
> FP operands, use them as addresses, and deliver Quads of interpolated results (RGBA}.
>
> Texture basically takes FP operands, uses them to index multiple levels of "texture memory"
> (the FP int parts index) and interpolate between the levels, the FP fraction parts interpolate.
>
> There are lots of special cases that are best handled by a HW FU.That HW-FU will be able to
> recognize the special nature(s) of texture memory requests, and both minimize the instruction
> count but also optimize memory efficiency, and avoid CPU<->cache memory ordering issues
> "when you don't know".

hmmm, it makes a lot of sense.  it would mean creating special
instruction opcodes specifically for covering them, which, if we're
going to "optimise" that area with special texturising accelerated
instructions, that has to be done anyway, so is no big deal.

jacob what's your thoughts?

l.



More information about the libre-riscv-dev mailing list