<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=""><br class=""><div><blockquote type="cite" class=""><div class="">On Aug 31, 2015, at 4:08 PM, Bjarke Roune <<a href="mailto:broune@google.com" class="">broune@google.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div dir="ltr" class=""><div class="gmail_extra"><br class=""><div class="gmail_quote">On Mon, Aug 31, 2015 at 10:42 AM, Owen Anderson <span dir="ltr" class=""><<a href="mailto:resistor@mac.com" target="_blank" class="">resistor@mac.com</a>></span> wrote:<br class=""><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word" class=""><span class=""><br class=""><div class=""><blockquote type="cite" class=""><div class="">On Aug 31, 2015, at 10:01 AM, Bjarke Roune <<a href="mailto:broune@google.com" target="_blank" class="">broune@google.com</a>> wrote:</div><br class=""><div 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="">I'm not sure if you're saying that there should be no limitations as the use cases you had mind for convergent are always OK with such duplication, or that we should be figuring this stuff out on a case-by-case basis, or something else?</span></div></blockquote></div><br class=""></span><div class="">I designed the semantics of convergent to meet the needs of texture and gradient operations, without a lot of consideration for barriers. IMO, It would be a nice end result if it turns out be just what barriers need as well, but I don’t want to accidentally hose the use cases I intended it for in pursuit of making it work for barriers.</div><div class=""><br class=""></div><div class="">Everything you’ve identified re: full vs partial unrolling, unstitching, etc. seems fine to me so far.</div><span class="HOEnZb"><font color="#888888" class=""><div class=""><br class=""></div></font></span></div></blockquote></div>So texture and gradient operations also requires these restrictions for unswitching and partial unrolling, or is it just that imposing such restrictions even if unnecessary is fine?</div><div class="gmail_extra"><br class=""></div><div class="gmail_extra">Would you say that we need the langref to say something about when convergent operations can be duplicated? I think it's hard to reason about whether unrolling, loop unswitching and other passes are doing the right thing without that.</div></div></div></blockquote><br class=""></div><div>I really want to avoid talking about duplication as part of the semantics of convergent, because I fundamentally believe the focus on duplication is a red herring. It’s not the duplication that’s the problem, it’s the introduction of conditionals.</div><div><br class=""></div><div>Within a SPMD model, the constraint that textures and gradients, and to some extent barriers, must preserve is that the set of executing threads must be the same at the location where they end up as the place where the user wrote them. To phrase that without talking about SPMD models or threads, it suffices to say that they must be executed in under the same set of execution predicates as where the user wrote them, aka in a control-equivalent location.</div><div><br class=""></div><div>—Owen</div></body></html>