<html><head><meta http-equiv="Content-Type" content="text/html charset=us-ascii"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Feb 29, 2016, at 1:50 PM, Alina Sbirlea <<a href="mailto:alina.sbirlea@gmail.com" class="">alina.sbirlea@gmail.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div dir="ltr" class=""><br class=""><div class="gmail_extra"><br class=""><div class="gmail_quote">On Mon, Feb 29, 2016 at 12:18 PM, Mehdi Amini <span dir="ltr" class=""><<a href="mailto:mehdi.amini@apple.com" target="_blank" class="">mehdi.amini@apple.com</a>></span> wrote:<br class=""><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word" class=""><br class=""><div class=""><span class=""><blockquote type="cite" class=""><div class="">On Feb 29, 2016, at 11:40 AM, Mehdi Amini via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank" class="">llvm-dev@lists.llvm.org</a>> wrote:</div><br class=""><div class=""><div style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><blockquote type="cite" class=""><div class=""><br class="">On Feb 29, 2016, at 11:16 AM, Alina Sbirlea via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank" class="">llvm-dev@lists.llvm.org</a>> wrote:</div><br class=""><div class=""><div dir="ltr" class="">All,<div class=""><br class=""></div><div class="">To get the discussion going in a focused manner, here is an initial patch with a running test. The test is from the Halide suite and is checking the correctness of several simd operations.</div><div class="">(Notes: the patch is large due to the number of operations being tested; </div><div class="">I expect a lot of changes before actually landing it, this is simply to continue the discussion using a concrete example.)<br class=""></div><div class=""><a href="http://reviews.llvm.org/D17726" target="_blank" class="">http://reviews.llvm.org/D17726</a><br class=""></div></div></div></blockquote><div class=""><br class=""></div><div class="">I can't figure how to download the patch *with the bitcode files* from Phabricator. Can you push this on github (or somewhere else)? (or if I missed how to proceed...).</div></div></div></blockquote><div class=""><br class=""></div></span>I was able to figure how get them "one by one", it would still be more convenient to have an archive or a repo to clone somewhere.</div><div class=""><span class=""><br class=""><blockquote type="cite" class=""><div class=""><div style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="">A few questions/todos to start the discussion:</div><div class="">1. What is a good location for these tests? They are in a separate Bitcode directory atm, but using the llvm_multisource. This may change to more closely model the approach for external tests (see next item). </div></div></div></blockquote></div></div></blockquote><div class=""><br class=""></div></span><div class="">A good location would be their own external repository IMO :)</div><span class=""><div class=""><br class=""></div><blockquote type="cite" class=""><div class=""><div style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="">2. There is a single .cpp file testing all operations provided by individual bitcode files. I expect this to change. Instead of using llvm_multisource to have the same test run with specific arguments, each run testing a single operation.</div><div class="">3. The building approach I took is to first link all bitcode files into a single one, then obtain the assembly for it, which cmake knows to take as an input source. </div></div></div></blockquote></div></div></blockquote><div class=""><br class=""></div></span><div class="">Yeah, so I'd rather have a split-build model, with a split execution model. Having a gigantic bitcode file to debug an issue is not friendly.</div><div class="">I'd expect to have a .cpp file that contains the main and the logic to run test, and then every test that is linked-in to be executed, a bit like gtests is doing (there are multiple registering mechanisms that would avoid to declare explicitly a test in the header).</div><div class="">-> filters.h and filter_headers.h should just go away.</div></div></div></blockquote><div class=""><br class=""></div><div class="">I agree, this is related to point 2. The plan here is to update the current test .cpp file to test each operation individually. In this model it will be enough to link with a single bitcode file per test.</div><div class=""> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word" class=""><div class=""><div class=""><br class=""></div><div class="">Also on the test in general: we should have an idea for each test what it is doing and how.</div><div class="">I was expecting your tests to be on the pattern of having an implementation in C++ and an implementation in Halide bitcode of a filters (or whatever) and run both on random data and verifies that the result is matching.</div><div class="">Unfortunately from what I can see you are feeding the tests with random data, and the tests are "blackboxes" that set an error flag if they detect an error. </div><div class="">This is not super robust: the compiler can mess with the error checking and eliminate it for instance, making any error undetected. </div></div></div></blockquote><div class=""><br class=""></div><div class="">The Halide bitcode filters compare the result of vectorized operations vs scalar runs of the same code. The error code against which we compare the output will be set to loose tolerance - it is currently 0. We're interested in codegen bugs that return the wrong value entirely, not accuracy differences (especially for floating point tests).</div><div class="">With the new error threshold, the data fed into may be random or read from provided input files, I can do either.</div><div class="">The filters will still look somewhat like blackboxes, though the name of the filter says what operation it's being tested and the disassembled bitcode files are reasonably readable.</div><div class="">Using your suggestion, the driver .cpp file will test one operation at a time (argvs set accordingly) and return right away once an error is found. Sound about right?</div></div></div></div></div></blockquote><div><br class=""></div><div>All of this is great. </div><div>The part that is not clear to me is why isn't it to have (what does it buy us over, or why is it better for us compared to) a possible a C/C++ reference implementation of the filter, and hoisting (and refactor) all the logic to feed the tests and validate the output *out* of the filters. A filter would be just the mathematical function performed and nothing more (separation of concerns, more robust framework, easier debugging when things-go-wrong, etc.).</div><div><br class=""></div><blockquote type="cite" class=""><div dir="ltr" class=""><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word" class=""><div class=""><div class="">Also, just looking quickly at one IR I'm surprised by things like:</div><div class=""><br class=""></div><div class="">"assert succeeded165": ; preds = %"assert succeeded146"<br class=""> %buf_host181 = getelementptr inbounds %struct.buffer_t, %struct.buffer_t* %error_op_pcmpeqq_272.buffer, i64 0, i32 1<br class=""> %23 = bitcast i8** %buf_host181 to double**<br class=""> %error_op_pcmpeqq_272.host226227232 = load double*, double** %23, align 8<br class=""> %24 = icmp eq %struct.buffer_t* %error_op_pcmpeqq_272.buffer, null<br class=""> br i1 %24, label %"assert failed183", label %"assert succeeded184", !prof !4<br class=""><br class=""></div><div class="">Here you have as check for nullptr at %24, but you already loaded %error_op_pcmpeqq_272.host226227232 from this pointer just before!</div></div></div></blockquote><div class=""><br class=""></div><div class="">It's checking that the host value loaded from buffer_t is not null. I don't see what's wrong with this. What am I missing?</div></div></div></div></blockquote><div><br class=""></div><div>I may be misreading it, my impression when skimming through the code was that it seems equivalent to:</div><div><br class=""></div><div>foo(buffer_t *out) {</div><div> auto value = out->host;</div><div> if (!out) {</div><div> error("nullptr");</div><div> }</div><div>}</div><div><br class=""></div><div><br class=""></div><div>In case I haven't been clear: I think this work is valuable for the project, and thank you for putting some effort into it :)</div><div><br class=""></div><div>-- </div><div>Mehdi</div><div><br class=""></div><div><br class=""></div><div><br class=""></div><div><br class=""></div><div><br class=""></div><div><br class=""></div><br class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="gmail_extra"><div class="gmail_quote"><div class=""><br class=""></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word" class=""><div class=""><span class=""><div class=""><br class=""></div><div class=""><br class=""></div><blockquote type="cite" class=""><div class=""><div style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="">A separate discussion is on reading metadata (mcpu and mattr) in llc. I added a script to work around that for now.</div></div></div></blockquote></div></div></blockquote><div class=""><br class=""></div></span><div class="">The generic way of doing it in llvm is (I think) to use function attributes:</div><div class=""><br class=""></div><div class="">attributes #0 = { "target-cpu"="x86-64" "target-features"="+avx2" }<br class=""><br class=""></div><div class="">You shouldn't need it on the command line I think?</div></div></div></blockquote><div class=""><br class=""></div><div class="">Yes, I believe so too. Currently these are set in mcpu and mattr by Halide and not read in by llc, hence the need for feeding them as parameters. It's a separate issue that we'll need to go into in depth, but I don't want it to interfere with getting feedback on how to best publish these tests.</div><div class=""> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word" class=""><div class=""><span class=""><font color="#888888" class=""><div class=""><br class=""></div><div class="">-- </div><div class="">Mehdi</div></font></span><span class=""><div class=""><br class=""></div><div class=""><br class=""></div><br class=""><blockquote type="cite" class=""><div class=""><div style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class=""><br class=""></div><div class="">Looking forward to your feedback!</div><div class=""><br class=""></div><div class="">Thanks,<br class=""></div><div class="">Alina</div><div class=""><br class=""></div><div class=""><br class=""></div></div><div class="gmail_extra"><br class=""><div class="gmail_quote">On Fri, Feb 19, 2016 at 6:50 AM, Kristof Beyls<span class=""> </span><span dir="ltr" class=""><<a href="mailto:kristof.beyls@arm.com" target="_blank" class="">kristof.beyls@arm.com</a>></span><span class=""> </span>wrote:<br class=""><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div bgcolor="#FFFFFF" text="#000000" class=""><span class=""><br class=""><br class=""><div class="">On 18/02/2016 19:12, Alina Sbirlea via llvm-dev wrote:<br class=""></div><blockquote type="cite" class=""><div dir="ltr" class=""><br class=""><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><br class="">I have more questions for Alina. What kind of tests do you have:<br class=""><br class="">- "the compiler takes the bitcode and generates code without crashing"<br class="">- "the compiled test runs without crashing"<br class="">- "the compiled test will produce an output that be checked against a reference"<br class="">- "the compiled test is meaningful as a benchmarks"<br class=""></blockquote><div class=""><br class=""></div><div class="">We have all 4 kinds of tests in Halide. The bitcode files for the first category is already available and I'm working on building the ones for the next 3. We'd like to include all incrementally.</div><div class=""> </div><br class=""></div></div></div></blockquote></span>It seems to me that the first category ("the compiler takes the bitcode and generates code without crashing") are tests that should be added to the "make check-all" tests in the LLVM subproject, rather than the test-suite subproject?<br class="">Or if these tests currently don't crash the compiler anymore, the bugs must have been fixed, and there should already be equivalent tests?<br class=""></div></blockquote></div><br class=""></div>_______________________________________________<br class="">LLVM Developers mailing list<br class=""><a href="mailto:llvm-dev@lists.llvm.org" target="_blank" class="">llvm-dev@lists.llvm.org</a><br class=""><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank" class="">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br class=""></div></blockquote></div><br style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><span style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;float:none;display:inline!important" class="">_______________________________________________</span><br style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><span style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;float:none;display:inline!important" class="">LLVM Developers mailing list</span><br style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><a href="mailto:llvm-dev@lists.llvm.org" style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" target="_blank" class="">llvm-dev@lists.llvm.org</a><br style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" class=""><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px" target="_blank" class="">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a></div></blockquote></span></div><br class=""></div></blockquote></div><br class=""></div></div>
</div></blockquote></div><br class=""></body></html>