<html><body><div style="color:#000; background-color:#fff; font-family:arial, helvetica, sans-serif;font-size:12pt"><div><span>Hi Andrew,</span></div><div><br></div><div>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:</div><div><br></div><div>(1) LLVM2.9+ourScheduler+preciseLatency <br></div><div>against <br></div><div>(2) LLVM2.9+ourScheduler+LLVM's-rough-1-10-latency-model <br></div><div><br></div><div>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.<br></div><div><br></div><div>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? <br></div><div><br></div><div>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. <br></div><div><br></div><div>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?</div><div><br></div><div>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.</div><div><br></div><div>Regards</div><div>-Ghassan<br></div><div><br></div><div><br></div><div style="font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div style="font-family: times new roman,new york,times,serif; font-size: 12pt;"><font face="Arial" size="2"><hr size="1"><b><span style="font-weight: bold;">From:</span></b> Andrew Trick <atrick@apple.com><br><b><span style="font-weight: bold;">To:</span></b> Ghassan Shobaki <ghassan_shobaki@yahoo.com><br><b><span style="font-weight: bold;">Cc:</span></b> "llvmdev@cs.uiuc.edu" <llvmdev@cs.uiuc.edu><br><b><span style="font-weight: bold;">Sent:</span></b> Wednesday, September 21, 2011 6:54 AM<br><b><span style="font-weight: bold;">Subject:</span></b> Re: [LLVMdev] Pre-Allocation Schedulers in LLVM<br></font><br><div id="yiv937279449"><div><div>On Sep 17, 2011, at 10:07 AM, Ghassan
Shobaki wrote:</div><blockquote type="cite"><div><div style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div>Hi,</div><div><br></div><div>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:<br></div><div><br> </div><div>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. <br></div></div></div></blockquote><div><br></div><div>Yes. For those wondering how to find out, grep for 'setSchedulingPreference'.</div><div><br></div><blockquote type="cite"><div><div style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div>
My questions are:</div><div><br></div><div>-Are there any references (such as published research) that describe each/any of these scheduling algorithms? </div></div></div></blockquote><div><br></div><div>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.</div><div><br></div><blockquote type="cite"><div><div style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div>-
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?</div></div></div></blockquote><div><br></div><div>The x86 LLVM target does not have a machine pipeline model (instruction itinerary).</div><div><br></div><blockquote type="cite"><div><div style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div>- 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?</div></div></div></blockquote><div><br></div><div>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.</div><div><br></div><div>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.</div><div><br></div><div>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.</div><div><br></div><blockquote type="cite"><div><div style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div>-
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?</div></div></div></blockquote><div><br></div><div>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.</div><br><blockquote type="cite"><div><div style="color: rgb(0, 0, 0); background-color: rgb(255, 255, 255); font-family: arial,helvetica,sans-serif; font-size: 12pt;"><div>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?</div></div></div></blockquote><div><br></div><div>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.</div><div><br></div><div>Thanks,</div><div>-Andy</div></div></div><br><br></div></div></div></body></html>