Actual CMO Operations

The spreadsheet CMOs.xlsx is a list of some of the desired CMO operation. It is by no means a complete list.

The version uploaded as of 2020-04-30_08.04.31 (TBD: provide link to GitHub version) counts these. These counts suggest the regular format In the next section

Regular format

por	bits	name	description
n?	1	LG	0=>local, 1=> global
y	3	scope	e.g. cache level to flush to although sometimes not strictly a level (8 encodings used)
y	4	cmop	operation type ?? encodings used
n	1	sec	security related, 0=no, 1=> flush predictors and prefetchers

Issue:

• LG: should we just assume that all CMOs must be "shoot down", applicabe to all of a coherence domain?
  • saves one bit, at the cost of performance for some "advanced" cases (like some supercomputers)
• sec: do we need the sec bit for address range CMO.VAR, or only for "whole cache"?
  • this saves one bit for the most expensive CMO.VAR instruction format

Bottom line: we can fit into 7 bits by making compromises, 8 bits fairly easily, although 9 bits is all rows abocve, and I would prefer 10 bits.

Not orthogonal: a very few operations require write permission, but not enough to warrant an orthogonal bit.

Scope encodings

We get away with "only" eight encodings, three bits, by overloading - using the same encoding to indicate slightly different things for outbound operations (pulls/flushes) and inbound operations.

for push CMOs	for pull CMOs (prefetchesd)
to pou(I$,D$)	to I$
to pou coherent processor caches	to L1 D$
to pou non-coherent processor caches	to L2$ pou(I,D )
to pou non-coherent I/O	to L3$
to ordinary DRAM	from NVRAM to DRAM
to battery backed up DRAM
to NVRAM ( first point of persistence)
to all NVRAM (full persistence)

we can of course argue about details, to try to reduce the count

• do we need to have two points of NVRAM persistence, first and all?
  • e.g. Keith Packard
  • e.g. HP "Machine" (TBD: ref)
• do we need to distinguish DRAM from battery backed DRAM
  • there are existence proofs, but we don't necessarily need to order them
• do we need to distinguish processor coherence from I/O coherence?
• could I/O coherence be just DRAM

But at the very least, I am sure that most people agree that we need at least four scopes, and probably more. => 3 bits. My biggest concern is that we should probably provide four bits rather than three.

NOT HANDLED:

• Prefetch operations might want to "skip" certain cache levels
  • e.g. fetch into L1 but no other levels
  • e.g. fetch into L1 and L3 but not L2
• Prefetch operations that want to stop - may want only to prefetch from into L1 from L2 or L3, but not from DRAM if missing L3 (to avoid saturating DRAM bus)
• CMOs that specify remote caches
  • e.g. P1 executes a CMO to prefetch/flush into some other processor P2's cache
    • like ARM stashing

CMO operation list for encodings

Placing this into a separate wiki page to make the table easier to edit.

• too hard to edit in long page
• would use section editing, except that GitHub wiki does not have that
• would use transclusion, except GitHub wiki does not have that

The table uses B+x? syntax to indicate priority classes

Count	Priority / Extension
5	Base
1	+xIO
2	+ xD+
1	+xLRU
1	+xPE
2	+xL
2	+xxLP
1	+xA
2	+xZ
1	+xW

Bottom line: 5 base CMO types => 3 bits.

17 with all of the above => 5 bits (i.e. more than 4)

B+xZ+xL+xLRU gets us to 9 encodings => highly likely that we will need more than 3 bits.

Excel spreadsheet "CMOS.xlsx"

Thursday, April 9, 2020-04-09:

• in GitHub repo at https://github.com/AndyGlew/Ri5-stuff/blob/master/CMOs.xlsx
• (probably has more recent copies elsewhere, e.g. personal machine or cloud Drive) is a "list" of CMOs. Not exactly a list, more like a table from which the actual list can be generated. Many rows of the table can be expanded into several different CMO operations with different privilege requirements, caches affected, etc.

TBD: actually generate a "flat" list. Preferably by script, so that I can automatically go back between the expanded list and a compact form that is folded with common sub expressions that is easier to understand.

Why CMOs.xlsx was written in Excel

OLD, Obsolete

Quick and dirty list of Actual CMOs