[LLVMdev] Pre-Allocation Schedulers in LLVM

[Ghassan Shobaki]

Hi Andy,

Please see my in-line answers below.

Regards

-Ghassan



________________________________
From: Andrew Trick <atrick at apple.com>
To: Ghassan Shobaki <ghassan_shobaki at yahoo.com>
Cc: "llvmdev at cs.uiuc.edu" <llvmdev at cs.uiuc.edu>
Sent: Friday, September 23, 2011 8:02 PM
Subject: Re: [LLVMdev] Pre-Allocation Schedulers in LLVM




On Sep 23, 2011, at 6:16 AM, Ghassan Shobaki wrote:

Hi Andrew,
>
>
>What we have is not a patch to any of LLVM's schedulers. We have implemented our own scheduler and integrated it into LLVM 2.9 as yet-another scheduler. Our scheduler uses a combinatorial optimization approach to balance ILP and register pressure. In one experiment, we added more precise latency information for most common x86 instructions to our scheduler and noticed a 10% performance improvement on one FP2006 benchmark, namely gromacs. More precisely, we compared:
>
>
>(1) LLVM2.9+ourScheduler+preciseLatency 
>
>against 
>
>(2) LLVM2.9+ourScheduler+LLVM's-rough-1-10-latency-model 
>
>
>
>And (1) was faster than (2) by 10%. We concluded that adding precise latency information may significantly improve the performance of programs like gromacs, which have a high degree of ILP. However, we did not do any performance analysis to verify that the performance improvement is indeed due to improving ILP scheduling (reducing pipeline stalls) not due to some other factor, such as memory performance, which may have gotten coincidentally improved by the reordering.

Thanks for the clear explanation. How much of LLVM's existing scheduler framework do you use? Do you schedule the selection DAG? Are  you using the same scheduling DAG built by the existing scheduler?

Ghassan: We are using LLVM's DAG but we convert it to our own format. Our scheduler has been originally developed as a standalone scheduler that reads its input DAGs from a text file. We currently have a little software layer that converts LLVM's DAG into our DAG data structure. Note that what we have now is a research prototype. The work may be productized later if it gives significantly better performance with an acceptable increase in compile time and if there is mutual interest in productizing it.    


We are not currently running any performance analysis tools, but if you are interested, we can identify the code section whose reordering has led to this performance improvement and send you the generated code in both cases. We can probably narrow it down to one basic block fairly easily, because our scheduler only operates on the hottest 2 functions in that benchmark and those have a total of 18 basic blocks. BTW, do you guys run SPEC CPU2006? If not, which benchmarks do you use for evaluating LLVM's performance? 
>
I run all the benchmarks in llvm.org/svn/llvm-project/test-suite/trunk. You can see that the "External" tests invoke the major SPEC CPU benchmark suites.


BTW, our scheduler's performance is about 3% faster than LLVM's ILP scheduler on that benchmark when it uses LLVM's 1-10 latency model. So, with the precise latency info, our scheduler is about 13% faster than LLVM's ILP scheduler on that benchmark.

Great. I'll be interested to see the details in your paper. We have not specifically tuned for SPEC CPU 2006.
I'm more interested in any insights you've gained while evaluating your scheduler's performance. It's one thing to measure speedup and another to understand it. And for me, the source of scheduling impact on x86 has been elusive. Are you using a machine-level profiling tool such as VTune?

Ghassan: We are currently looking into the gromacs case trying to identify the cause of the improved performance. We have been able to narrow it down to one basic block, but it is a relatively large one (~300 instructions). We are not currently using any performance analysis tool, but this test case may give us the motivation to start doing that.

Our work on pre-allocation instruction scheduling with register pressure minimization will be published soon, but if you would like to get an idea about my previous work on scheduling using branch-and-bound enumeration, you can find it on our school's research web page:

http://www.psut.edu.jo/research/index.html?fs=publications/drpub/ghassanshobakic.htm
  
Since the approach is fundamentally different from the heuristic approach used in LLVM's scheduler (and probably all schedulers found in production compilers), it will be hard to gain an interesting insight through a top-down analysis. In a heuristic approach, you are trying to come up with a prioritized set of rules for *guessing* the best choice, while in a branch-and-bound approach, you are trying to come up with a set of pruning techniques to make your exhaustive search as efficient as possible. However, a bottom-up analysis that examines particular test cases (like what we are currently trying to do with gromacs) may give some interesting insights. BTW, there are two other test cases for which our scheduler found significantly better schedules than LLVM's scheduler using LLVM's latencies (lbm on x86-64 (21%) and milc on x86-32 (13%)).  I will keep you updated with any findings that we may come up with.  
 

What may be more interesting for you though is finding out whether adding more precise latency information improves the performance of LLVM's ILP scheduler. We are interested in answering that question too, and are willing to add an x86 machine model to LLVM if that task does not turn out to be too complicated. So, how easy will it be to add an x86 itinerary? Can you point me to the file or files that have to be added or changed? Where can I find an example of an existing itinerary for some other target?

ARM itineraries are in lib/Target/ARM/ARMScheduleV?.td. A description of the itinerary format is in include/llvm/Target/TargetSchedule.td.


It's easy to add a new itinerary, but hard to specify the behavior of x86 microarchitecture to put it mildly.The itinerary format is quite detailed and verbose as a result of being target independent. It's reasonably powerful, but you have to play tricks such as defining fake functional units to express constraints. You may run into some difficulty with micro-ops, but let's worry about that later.

Fortunately, latency is modeled independent of pipeline and functional units and much simpler to deal with. You won't be able to precisely specify the latency of the bypass network, but I wouldn't be concerned with that level of detail.

BTW, In the above described experiment, we have only added latency information; we did not add functional-unit information. However, we are interested in adding both, but we would like to do it in two steps (first latency only then functional unit info) and measure the impact of each step.
Good.

Thanks,
-Andy


From: Andrew Trick <atrick at apple.com>
>To: Ghassan Shobaki <ghassan_shobaki at yahoo.com>
>Cc: "llvmdev at cs.uiuc.edu" <llvmdev at cs.uiuc.edu>
>Sent: Wednesday, September 21, 2011 6:54 AM
>Subject: Re: [LLVMdev] Pre-Allocation Schedulers in LLVM
>
>
>On Sep 17, 2011, at 10:07 AM, Ghassan Shobaki wrote:
>Hi,
>>
>>
>>I am currently writing a paper 
documenting a research project that we have done on pre-allocation 
instruction scheduling to balance ILP and register pressure. In the 
paper we compare the pre-allocation scheduler that we have developed to 
LLVM's default schedulers for two targets: x86-64 and x86-32. We would 
like to include in our paper some brief descriptions of the two LLVM 
schedulers that we are comparing against and some information about the 
machine model that they are scheduling for.  So, it would be great if 
you could confirm or correct the following information and answer my 
questions below:
>>
>>
>>
>>The default scheduler for the x86-32 target is the bottom-up register-pressure reduction (BURR) 
scheduler, while for the x86-64 target it is the ILP Scheduler. 
According to the brief documentation in the 
source file ScheduleDAGRRList, the BURR is a register pressure reduction scheduler, while the ILP is a register-pressure aware scheduler that 
tries to balance ILP and register pressure. 
>>
>
>
>Yes. For those wondering how to find out, grep for 'setSchedulingPreference'.
>
>
>My questions are:
>>
>>
>>-Are there any references (such as published research) that describe each/any of these scheduling algorithms? 
>
>
>The LLVM ILP scheduler is a mix of heuristics, some of which may be standard practice while others are LLVM inventions. The combination of heuristics can only be described as ad hoc.
>
>
>- By examining the source code, it appears that neither scheduler has a 
machine model describing the functional units and the mapping of 
instructions to functional units on the target x86 machine. Is that 
right?
>
>
>The x86 LLVM target does not have a machine pipeline model (instruction itinerary).
>
>
>- Based on the test cases that I have 
analyzed, it looks that the BURR scheduler sets all latencies to 1, 
which essentially eliminates any scheduling for ILP and makes scheduling for register pressure reduction the only objective of this scheduler. 
Can you please confirm or correct this?
>
>
>You're obviously wondering why this is called an "ILP" scheduler, which is understandable. I would even argue that the term "scheduler" is not entirely accurate.
>
>
>At any rate, this scheduler does not directly increase IPC by avoiding pipeline stalls as you might expect. What we mean by ILP scheduling is increasing the number of "in-flight" instructions. This is often the opposite of register pressure. The key to this is carefully tracking register pressure (locally) and dependence height. The heuristics decide whether an instruction is likely to increase register pressure. If not, then we attempt to overlap chains of dependent instructions.
>
>
>Note that LLVM's "hybrid" scheduler works much the same as the ILP scheduler but uses a mix of heuristics that happens to work better for targets with an intruction itinerary.
>
>
>- Again based on analyzing test cases, it appears that the ILP scheduler 
sets the latencies of DIV and SQRT (both INT and FP) to 10, while the 
latencies of all other instructions are set to 10. Can you please 
confirm or correct this observation?
>
>
>The scheduler is aware that certain operations are "very high latency", but frankly that knowledge isn't crucial. 10 is used, because we want a number much greater than 1, and within reasonable expected latency. Without a pipeline model, it just doesn't make much difference.
>
>Apparently, the developers 
of the ILP scheduler assumed that this rough latency model would be 
sufficient to do ILP scheduling on the x86 target, because the x86 
hardware has a good dynamic scheduler. Our testing, however, shows that 
this is the case for most but not all programs. For one particular 
benchmark with a high-degree of ILP, using more precise latency info 
significantly improved performance. Will the LLVM developers be 
interested in adding more precise latency info for the x86 target?
>
>
>Free performance? Sure. If you're talking about adding an instruction itinerary for X86, that should be no problem. A major change to the scheduler may be harder to swallow. Either way, it would be nice to see the patch and benchmark data or specific examples of code sequences that improve.
>
>
>Thanks,
>-Andy
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20110926/6ab0bae3/attachment.html>