<html><head><meta http-equiv="Content-Type" content="text/html charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class="">Hi,<div class=""><br class=""></div><div class="">What is the performance gain you expect? Can you share modified source(s) or IR? It might be good to have such 'test cases’ that can show potential performance suites eg. as a separate test suite. And this might be a good start.</div><div class=""><br class=""></div><div class="">Since you are looking at a loop transformation you might find some inspiration from Adam’s loop distribution pass <a href="http://reviews.llvm.org/D8831" class="">http://reviews.llvm.org/D8831</a></div><div class=""><br class=""></div><div class="">Good luck!</div><div class=""><br class=""></div><div class="">Gerolf</div><div class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Feb 15, 2016, at 6:44 AM, RCU via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div class=""> Hello, Michael.<br class=""> I come back to this older email. Sorry if you receive it again.<br class=""><br class=""> I am trying to implement coalescing/collapsing of nested loops. This would be clearly beneficial for the loop vectorizer, also.<br class=""> I'm normally planning to start modifying the LLVM loop vectorizer to add loop coalescing of the LLVM language.<br class=""><br class=""> Are you aware of a similar effort on loop coalescing in LLVM (maybe even a different LLVM pass, not related to the LLVM loop vectorizer)?<br class=""><br class=""> Thank you,<br class=""> Alex<br class=""><br class="">On 7/9/2015 10:38 AM, RCU wrote:<br class=""><blockquote type="cite" class=""><br class=""><br class="">With best regards,<br class=""> Alex Susu<br class=""><br class="">On 7/8/2015 9:17 PM, Michael Zolotukhin wrote:<br class=""><blockquote type="cite" class="">Hi Alex,<br class=""><br class="">Example from the link you provided looks like this:<br class=""><br class="">|for (i=0; i<M; i++ ){<br class=""> z[i]=0;<br class=""> for (ckey=row_ptr[i]; ckey<row_ptr[i+1]; ckey++) {<br class=""> z[i] += data[ckey]*x[colind[ckey]];<br class=""> }<br class=""> }|<br class=""><br class="">Is it the loop you are trying to vectorize? I don’t see any ‘if’ inside the innermost loop.<br class=""></blockquote>I tried to simplify this code in the hope the loop vectorizer can take care of it better:<br class="">I linearized...<br class=""><br class=""><blockquote type="cite" class="">But anyway, here vectorizer might have following troubles:<br class="">1) iteration count of the innermost loop is unknown.<br class="">2) Gather accesses ( a[b[i]] ). With AVX512 set of instructions it’s possible to generate<br class="">efficient code for such case, but a) I think it’s not supported yet, b) if this ISA isn’t<br class="">available, then vectorized code would need to ‘manually’ gather scalar values to vector,<br class="">which might be slow (and thus, vectorizer might decide to leave the code scalar).<br class=""><br class="">And here is a list of papers vectorizer is based on:<br class="">// The reduction-variable vectorization is based on the paper:<br class="">// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.<br class="">//<br class="">// Variable uniformity checks are inspired by:<br class="">// Karrenberg, R. and Hack, S. Whole Function Vectorization.<br class="">//<br class="">// The interleaved access vectorization is based on the paper:<br class="">// Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved<br class="">// Data for SIMD<br class="">//<br class="">// Other ideas/concepts are from:<br class="">// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.<br class="">//<br class="">// S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of<br class="">// Vectorizing Compilers.<br class="">And probably, some of the parts are written from scratch with no reference to a paper.<br class=""><br class="">The presentations you found are a good starting point, but while they’re still good from<br class="">getting basics of the vectorizer, they are a bit outdated now in a sense that a lot of new<br class="">features has been added since then (and bugs fixed:) ). Also, I’d recommend trying a newer<br class="">LLVM version - I don’t think it’ll handle the example above, but it would be much more<br class="">convenient to investigate why the loop isn’t vectorized and fix vectorizer if we figure<br class="">out how.<br class=""><br class="">Best regards,<br class="">Michael<br class=""><br class=""></blockquote><br class=""> Thanks for the papers - these appear to be written in the header of the file<br class="">implementing the loop vect. tranformation (found at<br class="">"where-you-want-llvm-to-live"/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp ).<br class=""><br class=""><blockquote type="cite" class=""><blockquote type="cite" class="">On Jul 8, 2015, at 10:01 AM, RCU <<a href="mailto:alex.e.susu@gmail.com" class="">alex.e.susu@gmail.com</a> <<a href="mailto:alex.e.susu@gmail.com" class="">mailto:alex.e.susu@gmail.com</a>>><br class="">wrote:<br class=""><br class=""> Hello.<br class=""> I am trying to vectorize a CSR SpMV (sparse matrix vector multiplication) procedure<br class="">but the LLVM loop vectorizer is not able to handle such code.<br class=""> I am using cland and llvm version 3.4 (on Ubuntu 12.10). I use the -fvectorize option<br class="">with clang and -loop-vectorize with opt-3.4 .<br class=""> The CSR SpMV function is inspired from<br class=""><a href="http://stackoverflow.com/questions/13636464/slow-sparse-matrix-vector-product-csr-using-open-mp" class="">http://stackoverflow.com/questions/13636464/slow-sparse-matrix-vector-product-csr-using-open-mp</a><br class=""><br class="">(I can provide the exact code samples used).<br class=""><br class=""> Basically the problem is the loop vectorizer does NOT work with if inside loop (be it<br class="">2 nested loops or a modification of SpMV I did with just 1 loop - I can provide the<br class="">exact code) changing the value of the accumulator z. I can sort of understand why LLVM<br class="">isn't able to vectorize the code.<br class=""> However, at http://llvm.org/docs/Vectorizers.html#if-conversion it is written:<br class=""> <<The Loop Vectorizer is able to "flatten" the IF statement in the code and<br class="">generate a single stream of instructions.<br class=""> The Loop Vectorizer supports any control flow in the innermost loop.<br class=""> The innermost loop may contain complex nesting of IFs, ELSEs and even<br class="">GOTOs.>><br class=""> Could you please tell me what are these lines exactly trying to say.<br class=""><br class=""> Could you please tell me what algorithm is the LLVM loop vectorizer using (maybe the<br class="">algorithm is described in a paper) - I currently found only 2 presentations on this<br class="">topic: http://llvm.org/devmtg/2013-11/slides/Rotem-Vectorization.pdf and<br class="">https://archive.fosdem.org/2014/schedule/event/llvmautovec/attachments/audio/321/export/events/attachments/llvmautovec/audio/321/AutoVectorizationLLVM.pdf<br class=""><br class="">.<br class=""><br class=""> Thank you very much,<br class=""> Alex<br class="">_______________________________________________<br class="">LLVM Developers mailing list<br class="">LLVMdev@cs.uiuc.edu <mailto:LLVMdev@cs.uiuc.edu> http://llvm.cs.uiuc.edu<br class="">http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev<br class=""></blockquote><br class=""></blockquote></blockquote>_______________________________________________<br class="">LLVM Developers mailing list<br class=""><a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a><br class="">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev<br class=""></div></div></blockquote></div><br class=""></div></body></html>