<div dir="ltr">Hi Adam,<div><br></div><div>First, I sincerely apologize for stalling this thread.</div><div><br></div><div>I understand your point about loop nest and reference groups. Both loop nest and the reference groups form the key to cache the analysis result. In this case, isn't it better not to cache anything and just compute the cost dynamically; based on loop nest and references provided as input by the calling transformation (fission/fusion/etc)?</div><div><br></div><div>Thank you</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Jul 5, 2016 at 11:39 PM, Adam Nemet <span dir="ltr"><<a href="mailto:anemet@apple.com" target="_blank">anemet@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word"><br><div><span class=""><blockquote type="cite"><div>On Jun 23, 2016, at 12:06 PM, Vikram TV <<a href="mailto:vikram.tarikere@gmail.com" target="_blank">vikram.tarikere@gmail.com</a>> wrote:</div><br><div><div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Jun 23, 2016 at 11:34 PM, Adam Nemet <span dir="ltr"><<a href="mailto:anemet@apple.com" target="_blank">anemet@apple.com</a>></span> wrote:<br><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"><br><div><div><div><blockquote type="cite"><div><br></div><div><div class="gmail_quote" 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"><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"><div><span><div><br></div></span><div>Hi Vikram,</div><div><br></div><div>Is the analysis result specific to a loop nest or to a loop nest together with a set of reference groups?</div></div></div></blockquote><div>The result is specific to each loop in the loop nest and the calculations are based on the references in the loop nest.</div></div></div></blockquote><div><br></div></div></div><div>Sorry can you please rephrase/elaborate, I don’t understand what you mean. Analysis results are retained unless a transformation invalidates them. My question is how the reference groups affect all this.</div></div></div></blockquote><div> </div><div>Sorry, I now understood that you meant llvm's analysis result (I thought how the analysis calculates the result). The analysis result is specific to each loop in its loop nest. Reference groups are necessary during cost computation only.</div></div></div></div></div></blockquote><div><br></div></span><div>Hi Vikram,</div><div><br></div><div>Let me ask you differently. This analysis can be thought of as having two inputs, (1) a loop or loops (?) and (2) the reference group. If a transformation pass requires some analysis, the analysis result would have to be computed unless it has already been computed and no transformation has invalidated it since the last use.</div><div><br></div><div>My question how does this apply to this pass. Consider this sequence of passes:</div><div><br></div><div>1. loop fusion of loops L1 and L2</div><div><br></div></div><blockquote style="margin:0 0 0 40px;border:none;padding:0px"><div><div>Computes Loop Cost for L1 and L2 with the reference group that includes accesses from the both groups.</div><div><br></div><div>Let’s assume here that the transformation does not occur at the end so the analysis result is still valid at this point.</div></div></blockquote><div><div><br></div><div>2. loop distribution/fission for loop L1</div><div><br></div></div><blockquote style="margin:0 0 0 40px;border:none;padding:0px"><div><div>Does this have to recompute the Loop Cost for L1 because now the reference group only includes accesses from L1?</div><div><br></div></div></blockquote>To put it different my question what the tag/key is when caching the analysis result. Is it the loop only or also the reference group.<span class="HOEnZb"><font color="#888888"><div><br></div></font></span><div><div><span class="HOEnZb"><font color="#888888">Adam</font></span><span class=""><br><div><br><blockquote type="cite"><div><div dir="ltr"><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"><div><div><br></div><div>You could probably describe how you envision this analysis would be used with something like Loop Fusion.</div></div></div></blockquote><div> </div><div>For Loop Interchange, the cost calculated for each loop can provide a desired ordering of the loops. Loop interchange can start interchanging the loops to match the desired order iff interchange legality check passes.<br></div><div>Eg: For matrix multiplication case,</div><div> for (i = 0 to 5000)<br> for (j = 0 to 5000)<br> for (k = 0 to 5000)</div><div> C[i,j] = C[i,j] + A[i,k]* B[k,j]</div><div> Here, based on cache line size and the cache penalties for the references (please refer to first mail of the thread or the reference paper for different penalties), following costs result:</div><div> Loop i cost: <span style="white-space:pre-wrap">2.501750e+11</span></div><div><span style="white-space:pre-wrap"> Loop j cost: </span><span style="white-space:pre-wrap">6.256252e+10</span></div><div><span style="white-space:pre-wrap"> Loop k cost: </span><span style="white-space:pre-wrap">1.563688e+11</span></div><div><font><span style="white-space:pre-wrap">Loop cost here is nothing but the total cache penalities due to all references, including penalties due to outer loops. So lower costing loop is a better fit for innermost loop because it has better cache locality. This would result in the desired loop order: i, k, j; and LoopInterchange can interchange 'j' and 'k' loops given the legality check passes (a nearest ordering for which legality is correct is also an optimal loop ordering).</span></font></div><div><font><span style="white-space:pre-wrap"><br></span></font></div><div><font><span style="white-space:pre-wrap">For fusion/fission, loop costs due to cache can be calculated by assigning cache penalties for references with respect to loops before and after fusion/fission. The costs will be a profitability measure for fusion/fission.</span></font></div><div><font><span style="white-space:pre-wrap"><br></span></font></div><div><font><span style="white-space:pre-wrap">I think use case description was very brief in a previous mail. So I have elaborated this time. More details can also be found in <a href="http://www.cs.utexas.edu/users/mckinley/papers/asplos-1994.pdf" target="_blank">http://www.cs.utexas.edu/<wbr>users/mckinley/papers/asplos-<wbr>1994.pdf</a>.</span></font></div><div><font><span style="white-space:pre-wrap"><br></span></font></div><div><font><span style="white-space:pre-wrap">Thanks</span></font></div></div><div data-smartmail="gmail_signature"><div dir="ltr"><div><div dir="ltr"><div><br></div><div>Good time...</div><div>Vikram TV</div><div>CompilerTree Technologies</div><div>Mysore, Karnataka, INDIA</div></div></div></div></div>
</div></div>
</div></blockquote></div><br></span></div></div></div></blockquote></div><br><br clear="all"><div><br></div>-- <br><div class="gmail_signature" data-smartmail="gmail_signature"><div dir="ltr"><div><div dir="ltr"><div><br></div><div>Good time...</div><div>Vikram TV</div><div>CompilerTree Technologies</div><div>Mysore, Karnataka, INDIA</div></div></div></div></div>
</div>