Mill Computing, Inc. › Forums › The Mill › Architecture › Pipelining › Reply To: Pipelining
Yes, latencies are member dependent, so order is too. Yes, same latency drops in slot order, and ops with multiple results drop in result-number order. Yes, the specializer deals with all that in the scheduling.
As for level of presentation, it’s real hard to balance the needs of the newbies against the impatience of the old hands, and fit it all into a talk of reasonable length. I try to cut things into reasonably complete topics with adequate background, mindful of the physical audience; at FaceBook only half the audience had prior exposure to notions like the Belt or Nones, and I don’t want to blow off the other half.
In many talks I can speak faster, but in this one the animations largely paced the talk; they can’t be too fast for the audience either. I was forced to leave out quite a few niceties; the drop-order issue for example, and things like extended scratchpad – I’m a little surprised that no one here has asked about what you do when you run out of scratchpad. And the whole question of epilogues.
The new Wiki will be up on our site soon, initially very light on content but I’m hoping you folks will help populate it quickly with searchable details. Those talk-viewers who need more can always find it here on the Forum, unless it is still NYF. There are at least six more talks to go.