Disclaimer

The content of this blog is my personal opinion only. Although I am an employee - currently of Nvidia, in the past of other companies such as Iagination Technologies, MIPS, Intellectual Ventures, Intel, AMD, Motorola, and Gould - I reveal this only so that the reader may account for any possible bias I may have towards my employer's products. The statements I make here in no way represent my employer's position, nor am I authorized to speak on behalf of my employer. In fact, this posting may not even represent my personal opinion, since occasionally I play devil's advocate.

See http://docs.google.com/View?id=dcxddbtr_23cg5thdfj for photo credits.

Sunday, February 06, 2011

Managing branch prediction history: copying short history versus pointing to large history

http://semipublic.comp-arch.net/wiki/Managing_branch_prediction_history:_copying_short_history_versus_pointing_to_large_history

When [[branch prediction history]] was short, 8-12 bits, and global,
then it was not unreasonable to manage the history by copying it.

E.g. in one simulator I actually arranged so that a branch [[uop]]
wrote  back, as its result
* an indication of whether it was mispredicted or not
* the taken [[target IP]]
* the branch predictor history to be restored on a branch misprediction.

I.e. the branch prediction history was propagated, in this simulator, from the [[instruction fetch front end]]
across the scheduler to execution, and back again.

While simple, this involves a lot of unnecessary data movement - both for the history, but also for the [[target IP]].
Most machines of my acquaintance create a [[branch information table (BIT)]], holding information for branches in flight.
This avoids copying the history from the front end to execution and back again,
but nevertheless may involve making copies of  the history.


    I confess an ulterior motive for copying the history and other branch information around: this naturally leads to the number of branches in flight scaling with the window size. Many early OOO designs were crippled by supported too few branches in flight.


Making copies of the history seems silly if, as in a TNT history, they differ only by 2 bits:
 new_history := (old_history << 1) | new_branch_taken_or_not


    (Note: when shifting a branch history, we will often say h<


But making copies seems to be required if we want to be able to restore to any mispredicted branch point, i.e. if we want to do [[instantaneous versus incremental branch misprediction repair]].

Making copies works well enough if the branch prediction history is small enough - 8 bits, etc.
But by the time we are talking about 16 bit histories and 32 or 64 bit IPs, we are talking about quite a few bits.

Furthermore, in the  late 1990s and early 2000s branch predictors arose with  hitherto unseen long histories,
such as Seznec's OGEHL predictor (with multiple history lengths 9, 2, 4, 8, 16, 32, 64, 128, ...).
Copying around a 128 bit history, even to a [[BIT]], is wasteful;
copying around the even larger 1000+ bit histories that have been proposed is even worse.

Hence the interest in pointing to a position in a branch prediction history,
rather than copying the entire history.
On a branch misprediction one would restore the pointer, rather than overwriting the  history with a savedcopy.

This would  be straightforward for a long [[TNT]] history, since only 1 bit depends on any branch.
You would simply keep a history of total length TL=PHL+BIF, the sum of predictor history length plus branches in flight.
On a misprediction you would restore the pointer in this circular buffer.
(Or equivalently shift the buffer- I suspect that shifting is too power hungry.)

This could scale to almost any length of history, potentially thousands of bits.

Unfortunately, modern [[stew]] histories are more complicated than [[TNT]] histories.
The [[branch IP]], or even both [[from IP]] and [[to IP]], may be [[hashed, e.g. XORed]] into the stew.
This means that several of the youngest bits in the history may change on every branch.
Simply restoring a pointer will not suffice.

TBD: explain stew management in more detail.

Simple strategy:  constrain the stew to have only the N youngest bits affected by the most recent branch.  Bits TL..N are unaffected  by the most recent branch, except for shifting.
One can then keep a copy of the parts of the history affected by recent branches, the N youngest bits, and a pointer that locates the older bits.

= See Also =

* [[How to use a really long predictor history]]

Branch Prediction Stew

{{Terminology Term}}
http://semipublic.comp-arch.net/wiki/Stew

The [[stew]] is a form of history used by certain branch predictors.

See, for example, US patent 7143273,
Method and apparatus for dynamic branch prediction utilizing multiple stew algorithms for indexing a global history,
Mile, Slade, and Jourdan,
filed March 31, 2003,
assignee Intel.

A simple [[branch predictor history]] might be  a simple [[TNT]] history,
with 0s corresponding to non-taken and 1s corresponding to taken.
Such a simple TNT history cannot distinguish some convergeing paths,
and indirect branches.

US7143273
describes one embodiment of a stew as
 stew = ((stew << 1)|new_bit ^  ip)

A stew formed  in this way can distinguish converging paths such as  ("if true" means a condition that evaluates as true, not an unconditional branch):
   L1: if true got L2
   L2: if true goto L99
   L10: if true goto L11
   L11: if true  goto L
   L:  if ?? goto L99

However, it does not distinguish multiple  branch targets and paths out of an indirect branch, such as
   L1: Reg:= IL1; if true goto L2
   L2: if true goto L99
   L10: Reg:=IL2; if true goto L11
   L11: if true  goto L
   L:  if ?? goto [Reg]
   IL1: ...
   IL2: ...
It can be seen that mixing in arc information as well as node information remedies this situation,
and distinguishes different paths so long as the hashes do not collide:

 stew <<= number_of_bits_to_discard
 stew = hash( stew, from_IP, to_IP, taken/not_taken, ...)

Issue: how many bits to use?  Which may vary as a function of the type of branch: e.g. a direct conditional branch
may not need as many to_ip bits to be hashed in
as a completely random indirect branch.
Similarly, indirect calls and returns may be handled separately.

(TBD-IP)

Monday, January 31, 2011

Slow :-(

Minor bitching: writing in the comp-arch wiki is so slow.

Hard to get big chunks of time.

Hard to write big chunks when I don't have big  chunks of time.

But at least my present employer allows me to write this.  And does not forbid me, like my past employer.

I wish that I could average 1 new page a day.  Instead, I am averaging less than half a page a day, and probably less than 1 substantial page a week.

OOO_versus_Runahead_versus_Log-based_replay_for_runahead

http://semipublic.comp-arch.net/wiki/OOO_versus_Runahead_versus_Log-based_replay_for_runahead

In the era of Willamette I suffered a moral dilemma - or, rather, a dilemma of morale.
I had evangelized [[OOO execution]] with P6.
But I had been introduced to runahead.
And runahead execution seemed to be asymptotically as good as OOO execution,
as memory latencies got longer - i.e. as we entered the era of [[MLP (memory level parallelism)]], which I was then starting to evangelize.

OOO execution is ... well, I'll assume you know what OOO is, or can read about it [[OOO|elsewhere]].

= What Runahead Is ==

[[Runahead]] consists essentially of
* executing
* when an unpredictable, long-latency, event such as a cache miss occurs
** take a checkpoint
** mark the result invalid
* continue executing past the checkpoint in a speculative manner
** saving speculative values into registers
** saving speculative values into a store buffer, or similar [[memory speculation datastructure]]
* when the cache miss returns
** discard all speculative state, registers and memory
** restore from the checkpoint.

= Runahead and OOO are asymptotically equal =

Runahead is a comparatively simple action.
And, in some sense, it gets nearly all of the benefit of OOO.
I.e. as memory latencies increase to infinity and execution latencies becomeinfinitesimal by comparison,
runahead execution will approach OOO execution in performance.

Normally, barring funny effects, runahead  will always be lower performance than OOO.
For the same size of instruction window.
But runahead is much cheaper in hardware, and can potentially runahead much farther than OOO for the same hardware cost.
Both runahead and OOO are limited by the [[memory speculation datastructure]];
but runahead need build no large physical register file, etc., save only for a number of RF  checkpoints.

= [[log-based]] =

Confounded by OOO/run-ahead, I looked for something better than runahead in some asymptotic sense.
Hence  [[log-based]] execution - specifically [[log-based verify re-execution]].

Although subsequently I applied [[log-based]] to [[speculative multithreading]],
at first I thought about it only wrt the same [[single sequencer microarchitecture]]
as OOO or runahead:

Imagine that you are executing in a runahead manner.
However, instead of re-executing all of the instructions after a cache miss returns and a checkpoint is restored,
instead you (1) record all of the results of runahead execution in a log,
and (2) you execute instructions out of the log.

This raises several questions:
Why would this be any faster than runahead?
Why would it be any "larger" than OOO or as large as runahead?
Why would it be as cheap as runahead?

Let;'s answer them out of order.

;Why would this be as cheap as runahead?

Well, if you had a dedicated log [[hardware  datastructure]] it would not be.
However, the log is accessed sequentially during [[verify re-execution]].
Sequential access patterns can be easily prefetched.
therefore, you could put the log in slow memory - potentially in main memory.
you might need some fast hardware log to "start off" the verify re-execution,
sufficient to tolerate the latency to main memory.
Equivalently, you might manage cache replacement to ensure that the oldest entries in the log are not displaced from a cache of main memory,
as they normally would be in LRU.

;Why would it be any "larger" than OOO or as large as runahead?

As explained above, the log need not be a small size hardware datastructure.  It could potentially scale in size as main memory.

;Why would this be any faster than runahead?

Two reasons.

First, [[verify re-execution]] is inherently more parallel than normal execution.

I like to explain this using repeatedly increment a register as an example:

 r9 := load(cache misss)
 INC r9
 INC r9
 ...
 INC r9

If INC is unit latency, then runhead re-execution would take N cycles to re-execute N such INCrements of a register.

However, in [[log-based verify re-execution]],
the example code above would be rewritten in the log as

 r9 := load(cache misss)
 assert(r9 = 1042)
 INC r9 // assert(r9=1042); r9:= 1043
 INC r9 // assert(r9=1043); r9:= 1044
 INC r9 // assert(r9=1044); r9:= 1045
 ...
 INC r9 // assert(r9=1042+N); r9:= 1042+N+1

i.e. the instructions rewritten for storage in the log,
and for [[verify re-execution]] out of the  log, consist of nothing except moves of a cionstant to a register.
Rather than N cycles, these would take N/IF cycles, where IF is the instruction width.

Furthermore, there are many optimizations that can be applied to the log:
removing redundant asserts and redundant overwrites, etc.
In general, for a block of register based code,
[[log-based verify re-execution]] has  one assert for every live-in,
and one constant to register move for every live-out.
Assuming the assertions are met.
(And if the assertions are not met, fall back to normal (re-)execution.

The second reason for log-based verify re-execution being faster is this:

Normal runahead re-execution is limited by cache associativity.
If you have N independent memory references that were cachemisses prefetched by the runahead speculative execution epoch,
they are normally independent, and all prefetched.
EXCEPT when they happen to collide in the cache, e.g. in the limited associativity cache.
In which case, another runahead/checkpoint/re-execution would begin.

Log-based verify re-execution suffers this not at all.
If a cache missing load is thrashed out of the cache,
its value is still present in the log,
and dependent instructions can still be executed - or, rather, verify re-executed.
You will need to pay the memory latency to fetch it and verify, but that can be overlapped.


= Conclusion =

Log-based verify re-execution, or`replay, for runahead
is asymptotically faster than ordinary runahead in two senses:
non-cache miss execution latency,
and cache thrashing.

Ultimately, the fashion or fad of runahead petered out.
Runahead did not replace OOO, although it caused much heartburn.

Runahead did not replace OOO because, although asymptotically better than OOO, we aren't at the asymptote.
Execution unit latencies are not 0, and cache misses infinite.
OOO handles both.

Also, runahead may consume more power than OOO.  This is somewhat debatable:
OOO's hardware datastructures may waste power, e.g. leakage,
whereas runahead may consume static power.

Probably the biggest reason why runahead did not supplant OOO in the late 1990s and early 2000s is this:
runhead is only as good as OOO.
It doesn't do anything new.
OOO was an already established and well understood technology by the time runahead appeared on the scene.
Although runahead has putative benefits, they were not enough to warrant a switch-over.

Be this as it may, log-based verify re-execution is a technique that rose out of the [[Runahead versus OOO]] debate.
It is better than runahead in some asymptotic senses.  And it has application to other areas, such as [[SpMT]].

= Hybrids =

Of course it is possible  to create hybrids of OOO, runahead, and log-based.

For example, one need not record all instruction results in the log.  One can re-execute that which is fast,
and only record in the log that which is slow.

Similarly, an OOO machine need not record all speculative results in the instruction window.

Onur Mutlu and others have proposed hybrids that fill an OOO instruction window, and then use runahead.

Saturday, January 29, 2011

Cygwin X multidisplay bug

Today I figured out a problem that has been plaguing me on Cygwin for a year or two, since  things broke in an update.   (I probably figured it out before, but neglected to record or remember it.)

When I start up Xwin -multiwindow, using  windows manipulated by MS Windows rather than a desktop in an MS Window window (try saying that fast), the initial xterm was always invisible.

Eventually I realized that if I maximized it from the MS Windows task bar, I could see it.

Apparently, this window was being created at location (0,0) in a bounding box for my multi-display configuration.   (That's multiple physical displays, not multiple X  displays.)  However, location (0,0) was invisible - since my multi-display configuration has a big tall portrait mode window in the middle for reading PDFs, with flat landscape displays on either side for webpages. The portrait display rising above landscape displays.

By the way, it is actually +1+1 asthe  default location, not +0+0. Minor  difference.

I need to play around  and see if I can find a guaranteed visible place.

Removing the geometry spec seems to work okay for me.

Friday, January 28, 2011

Why_full-speed_denormal_handling_makes_more_sense_for_CPUs_than_CPUs

http://semipublic.comp-arch.net/wiki/Why_full-speed_denormal_handling_makes_more_sense_for_CPUs_than_CPUs

o a collector of computer architecture trivia like me,
it was quite remarkable
that the Nvidia Fermi GPU
provides full-speed support for denormal operands and results,
whereas Intel and AMD CPUs, at the time of writing, do not.

By the way, terminology: [[full-speed support for denormals]] is not quite the same as [[hardware support for denormals]].

The first question is whether you provide hardware support,
or whether it is necessary to [[trap to microcode or software]], to provide the denormal support.

The second question is,
if you do provide hardware support for denormals and do not need to trap to microcode or software.
Even if denormals are performed in hardware, there may be a performance cost:
it may cost latency, e.g. 5 cycles for an FADD  rather than 4  cycles;
or it may cost bandwidth.
The latency and bandwidth impacts are related but may be decoupled:
e.g. it is possible that a scheduler can arrange so that throughput is not reduced even if latency is increased,
so long as there is sufficient [[ILP]].

By [[full-speed denorms]] we mean mainly that throughput or bandwidth is not affected.
E.g. you may arrange to add to the  latency of all FP ops,
to avoid [[writeback port collisions for operations with different latencies]].
Since GPUs are throughput machines, this is usually a reasonable tradeoff;
on some GPUs even integer arithmetic is stretched out to 40 cycles.

So, why would a GPU like Nvidia Fermi provide full-speed denorms,
whereas x86 CPUs from Intel and  AMD do not?

Let's skip the possibility of a marketing bullet.

Remember [[GPU-style SIMD or SIMT coherent threading]]?

If a GPU takes a trap, or otherwise takes a hiccup, to handle denorms, then not only is the current thread impacted.
All 16 or 64 threads the [[wavefront or warp]] are impacted.
And if only one of the [[spatial threads]] in the [[warp]] is  taking the trap,
the efficiency of the machine decreases by at least 16-64X in that period.
Let alone the possibility that the denorm handling code is also not well suited to a GPU.

I.e. the relative cost of [[denorm handling via trapping]] is much higher on a GPU than on a CPU.
Even denorm handling in hardware, in a manner that impacts latency or throughput,
is relatively more expensive and/or harder to deal with.

Hence: there are good technical reasons to do [[full-speed denorm handling]] in GPUs.
These reasons are not quite so compelling for CPUs
- although I predict that ultimately CPUs will be compelled to follow.

--


    Ancedote: on one of my few trips to Pixar, I talked to Bruce Perens about denorm handling. Apparently Pixar would be happily rendering a movie at high speed, and then Bam! they would have a scene like a night sky, black, full of denorms, and performance would fall off a cliff. We talked about flush-to-zero modes, which Intel x86 CPUs did not have at that time. I suggested biasing the numbers, e.g. adding 1 or similar, to prevent getting into denorm range. Now that x86 has flush-to-zero the need for such [[kluge]]s is greatly reduced. But, denorms exist for a reason. flush-to-zero can introduce artifacts, and always introduces [[FUD (Fear, Uncretainty, and Doubt)]]. [[Full-speed support for denorms]] may just be the coming thing, a place where GPUs lead the way for CPUs.


Thursday, January 27, 2011

ISO mediawaiki macro

I'm still using mediawiki for my comp-arch wiki.

I want a "macro", something like

   {{acronym_and_term SIMT "simultaneous multithreading"}

that will create pages

   SMT
   simultaneous mulrtithreading
   SMT (simultaneous multithreading)

and redirect them all to
   simultaneous multithreading (SMT)
  

Come to think of it, a primitive
    {{create_redirect_to_current_page  Redirect_From}
would be nice.

I imagine that it would first see if the Redirect_From page already exists.  If not, it would create
it, redirecting to the current page.  If so, it would probably not do anything - which might be a minor lossage if
the Redirect_From page doesn't link to the current page, but which allows things like disambiguation pages.



Perfunctory attempts to find this on mediawiki fail; I'm not sure where to go ask (and I confess to not liking the wikimedia community).