[Libre-soc-dev] RV SIMD
Luke Kenneth Casson Leighton
lkcl at lkcl.net
Fri Feb 12 19:12:25 GMT 2021
so it begins - illustrating very starkly the O(N^6) opcode proliferation.
most significant word multiply, in 32 16 and 8 bit selection.
saturation and rounding, 10 opcodes per width (8, 16) per operation (add,
byte packing 40 opcodes.
signed crossed multiply (HI-LO LO-HI words)
the list goes on and on and on, a total nightmare.
SV *literally* covers the entire lot with swizzle, elwidth overrides,
saturation contexts and VL, with the addition of something like 5 or 6
instructions: average-add, min/max, abs-diff, i think that's it.
one good operation comes out of this: bitrev. this is an extract operation
except including reversal of the bits extracted.
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68
More information about the Libre-soc-dev