2 ms·
"The problem is not figuring out what sizes things are, but the routing" Aha! Thanks. "Given the instructions are disjoint and they only need synchronization
by Someone 11y ago
"The problem is not figuring out what sizes things are, but the routing"
Aha! Thanks.
"Given the instructions are disjoint and they only need synchronization at jumps"
But they are (AFAIK) executing a single thread of execution, so creating those disjoint streams in general will be far from easy, won't it? if it were possible to do that, we would see a 2x speed up of single-threaded code on CPUs with two cores.
- renox 11y agoUh? That's quite different AFAIK: two CPU cores doesn't share their registers, those disjoint streams do.
- Someone 11y agoI know this architecture tries to innovate everywhere, but disjoint streams that share registers, but not caches? I haven't read all documentation on the Mill, but I would expect that that sharing part means that both streams can read or one can write each register.
- willvarfar 11y ago(Mill team) The Mill is Havard architecture. Each Mill core has two separate physically separated instruction caches that each feed a dedicated decoder, which decodes for a distinct set of pipelines.
- renox 11y ago> which decodes for a distinct set of pipelines I apologize for my mistake, I hope that it hasn't induced too many confusions (I take it that 'distinct set of pipelines' means distinct operation and 'belt' so the registers aren't shared).
- willvarfar 11y agoNo need to apologise :) The pipelines are distinct, but the set of ops that any particular pipeline supports is arbitrary (as in picked by the HW guys who have to fit the FUs in) and the belt is shared. The belt is where the two sides join. We tend to get good entropy putting the flow on one side and exec on the other, and that's nice for humans to model too, but there is no technical requirement. The thing about the Mill is its easily customisable. If you have some new FU (lets say you want a particular op that assists in say encrpytion) you can choose where you put it, and how many pipelines you put it in etc.
- renox 11y agoEspecially since if the belt is shared then in fact I remembered correctly the presentation.
- Veedrac 11y agoWell, I was actually trying to say the decoders work independently and produce disjoint subsets of the instruction stream. How the calculations actually get done is harder, but should be largely independent of how the instructions are encoded. One great thing about the Mill is that instructions within a bundle are mostly disjoint. The Mill has no data hazards and each instruction takes from the belt, which is not changed until the bundle is finished. There are "gangs" that allow combining multiple instructions, but I don't know how they are handled. In general, this means a Mill has more instruction interdependence than an x86. Although an x86 could have a split-stream instruction cache, data dependence prevents a traditional architecture from parallelizing itself as much as the Mill and hazards prevent the kind of speculation the Mill can do. Since instruction decode isn't a big bottleneck on fixed-width architechtures, there's no real pressure to innovate there.