| | <<<Back 1 day (to 2011/08/30) | 2011/08/31 |
| cryptopsy | a tiny camera with a big lens: domain SWEEThome | 00:08.04 |
| | whoops | 00:08.06 |
| | http://images.4chan.org/p/src/1314735259705.jpg | 00:08.31 |
| | /p/ is the photography channel on 4chan | 00:08.37 |
| mvrhel2 | darn my sse implementation for the table interpolation in lcms is not any faster than the old code. at least it is giving the correct output. I will poke around at it a bit more. one thing that I did learn in this excercise is that we should be able to get a big improvement by doing this in transparency blending where we can work on 16 pixels at a time for compositing | 00:11.22 |
| | I think the issue in lcms is simply that the set up time is taking too much time | 00:11.47 |
| | loading the table data in particular into the simdi registers is not fast due to how it is organized | 00:12.16 |
| | anyway. bbiaw. need to see about getting the kids some new ski equipment for the winter | 00:12.45 |
| | good night | 05:00.56 |
| chrisl | Phew, looks like cust 532 is (temporarily) happy again - that was a bit of a "curve ball" problem they pitched there! | 07:33.44 |
| kens | Lucky you, my customer is just getting sillier. | 07:34.00 |
| chrisl | Yes, I saw that. I guess the sooner I get the direct referencing of CIDFont file from cidfmap working fully, the better! | 07:34.42 |
| kens | Gien that today's screw-up is using an '=' with -I (eg -I+<path>) that won't save him. | 07:35.13 |
| | His QA guy got it right, but he didn't think it was a valid test...... | 07:35.32 |
| chrisl | Doh! I'm surprised at how many people (developers!) thing that the "=" is needed | 07:36.03 |
| kens | Especially when he's been sent numerous comand line examples that don't use it, and his QA engineer manages to get it right. | 07:36.30 |
| | If it was me, I would have tried to reproduce the environment (in fact that *is* what I've just done). He just plows on using his own and then bitches when he messes up. | 07:37.24 |
| chrisl | par for the course :-( | 07:37.45 |
| kens | In this case, yes, he's already shown he can't follow simple instructions. | 07:38.04 |
| | So now I'm sending him a complete simple environment, using his own paths, and including a test.bat file so he can't arse up the command line. | 07:38.38 |
| | I wonder what he'll manage to get wrong this time.... | 07:39.12 |
| chrisl | It'll still end up with: "ah, but when I do it this (wrong) way, it doesn't work......." | 07:39.46 |
| kens | He'll have to modify the batch file first :-) | 07:40.02 |
| | I suspect he'll fail to rename the .zpi file to a .zip file (bloody GMail) | 07:40.28 |
| chrisl | Remember the axiom about idiot proof software? ;-) | 07:40.49 |
| kens | Well I think we may have a better class here. Of course, when you look at the other bug reports from this customer, they are obviously breeding for it. | 07:41.23 |
| | I can't believe how much time this non-bug has cost. | 07:41.49 |
| chrisl | It does seem crazy - are they using a specific font, or would the proposed general fallback CIDFont have made it "just work" for them? | 07:42.45 |
| kens | The general fallback would work, they are using MSMincho to substitute for a font with a Japanese name. That was the previous problem, they were using a PDF_escsaped font name. | 07:43.28 |
| | Which doens't work in PostScript of course. | 07:43.47 |
| chrisl | Well, it just confirms what I already thought: I really need to get the fallback CIDFont in place for 9.05 - at least for PDF files. | 07:45.43 |
| kens | It would really help.... | 07:46.00 |
| chrisl | Well, falling back to a font like the DroidSansFallback.ttf was pretty easy to do, but I'm a little hazy on what might be needed to handle different glyph orderings. | 07:52.50 |
| kens | Surely just define each fallback (Eg Adobe-Japan1) as droid sans with the correct Decoding (Japan1) ? | 07:53.37 |
| chrisl | Like I said, not sure - I got distracted by something else important before I could experiment much. I'm hoping it is just that simple. In fact, I'd like to avoid defining all the combinations, and create it on the fly. | 07:55.40 |
| kens | That was kind of what I had in mind ;-) | 07:55.54 |
| | Great, the customer's email bounced my message | 07:56.24 |
| chrisl | I'd probably have finished it, if I hadn't had to add handling of the extra cmap table type - even though it was fairly simple, in the end | 07:56.55 |
| | So I guess their mail system doesn't much like binary attachments at all! | 07:57.28 |
| kens | Probably too big or soemthing. Not that it stopped them sending us large attachments. | 07:57.54 |
| chrisl | Well, you have your public_html are on casper, you could stick it on there, and give them the link | 07:58.25 |
| | s/are/area | 07:58.34 |
| kens | I think I'll leave the problem with them, I'm not feeling generously inclined today | 07:58.52 |
| Robin_Watts | Damn. More pattern problems in the planar stuff. | 12:38.14 |
| kens | Can join my pattern problems in pdfwrite | 12:38.31 |
| Robin_Watts | Hopefully I can apply the same stuff that mvrhel2 did for the transparent case. | 12:38.46 |
| | lunchtime | 12:39.32 |
| ray_laptop | First day of school :-) (at least for 2 of the 3 kids) | 13:51.26 |
| kens | Melanie doesn't go back until next Tuesday | 13:51.59 |
| ray_laptop | that's when our 5 yr old starts | 13:54.36 |
| | chrisl: good job supporting kanazawa-san | 13:55.23 |
| chrisl | ray_laptop: thanks - as I said, that was bit a of a curve ball - but a fairly sensible thing for a printer | 13:56.23 |
| ray_laptop | chrisl: can you send me the download for that version of the sim | 13:56.35 |
| | chrisl: I suppose so. As long as the paper output looks OK, who needs the fonts | 13:56.59 |
| kens | ray_laptop : can you watch out for Craig replying later today please ? I'm off in good time for a pizza, and won't see his response. I'm pretty certain the problem is that his command line has -I=... | 13:57.57 |
| | I did send a zip file with a complete system, which got cc'ed to support, but his email rejected it. | 13:58.22 |
| kens | is hoping Craig doesn't trim support of the CC list this time. | 13:58.47 |
| chrisl | ray_laptop: the sim is the same one we got back in June - it's on their "DOX" site, if you need the details for login, I'll send them through. | 13:59.02 |
| ray_laptop | kens: OK. I'm reviewing the email thread now. BTW, I usually suggest (rather forcefully) that they use / instead of \ | 13:59.07 |
| kens | I did suggest that, and also that he didn't use relative paths, but both do axctually work OK for me | 13:59.30 |
| ray_laptop | chrisl: OK. Thanks. I'll check which version I have | 13:59.46 |
| chrisl | ray_laptop: it does worry me slightly - I *know* Len and his team have changes that are not in the archives that Kanazawa-san gives us. | 14:00.52 |
| ray_laptop | chrisl: I'll touch base with Len today and will mention that. | 14:01.28 |
| | kens: to send large files to customers, you can always put them up on your public_html/ path on casper (they would access that using http://ghostscript.com/~kens/ ). If you want thnigs hidden there, just put a dummy index.htm and it (apache) won't do a directory listing, so you can just send the full URL including the filename | 14:06.18 |
| kens | ray_laptop : yes chrisl suggested that, but as I said, I wasn't feeling inclined to be generous by then. | 14:06.44 |
| ray_laptop | kens: e.g. http://ghostscript.com/~kens/forcraig.zip | 14:06.59 |
| | kens: I'll upload the attachment to my public_html, then so you don't have to bother with it, and I'll let Craig know that I'll follow up during today. | 14:08.18 |
| kens | OK, thanks ray | 14:08.33 |
| | I do think fixing the -I will solve ti for him though | 14:08.43 |
| ray_laptop | kens: I hope so. | 14:08.59 |
| henrys | if anyone has something that can view cgm files let me know. | 14:49.05 |
| | nvm I think I found something. | 14:52.06 |
| ray_laptop | cgm is sort of ancient isn't it ? | 14:53.26 |
| kens | very I think./ | 14:54.02 |
| henrys | yes I guess I should just remove the device. | 14:54.04 |
| ray_laptop | of course, HP-GL is even older. | 14:55.51 |
| henrys | it isn't really the age that matters, I think folks moved on to svg. | 14:59.48 |
| | kens2:if you want me to poke at that pattern problem a bit more I can - I don't know if you have higher priority stuff to look at. | 15:01.28 |
| kens2 | henrys I have a solution, but regression test showed zillions of diffs, including many not in pdfwrite | 15:02.22 |
| | Which isn't possible. | 15:02.31 |
| | Going to do a clean regression test (without my changes) then test with my changes. | 15:02.48 |
| | I think some of the code I updated to has caused lots of diffs. | 15:03.01 |
| | I think my fix will be OK. | 15:04.20 |
| | Drat, can't start next run until current is finished, oh weell. | 15:16.55 |
| henrys | bbiaw | 15:31.34 |
| Robin_Watts | kens2: export CLUSTER_USER=robin then git cluster and your job should be scheduled as mine. | 15:38.35 |
| kens2 | its OK the second job is running now, but thanks | 15:38.52 |
| | I won't get my result before I go off tomight, assuming all is OK I'll commit hte change tomorrow. | 15:42.15 |
| aditsu | hi, can gs write a pdf over another pdf (with the same number of pages)? | 15:47.43 |
| kens2 | depends what you mean by 'over'. | 15:48.04 |
| aditsu | or write text on a pdf | 15:48.13 |
| Robin_Watts | aditsu: You mean merge the page contents of 2 sets of PDFs ? | 15:48.19 |
| kens2 | I'm guessing you mean combine 2 PDFs into one, in which case the answer is no. | 15:48.28 |
| aditsu | yes, add the contents of the 2nd pdf into the pages of the first file | 15:49.01 |
| kens2 | Assuming you mena overlay the contents of page 1 from file with with page 1 form file 2 etc. | 15:49.05 |
| | You can add them together to get one PDF with twice as many pages | 15:49.31 |
| Robin_Watts | You'd have to open the pdfwrite device, then 'play' page 1 from the first file, and then 'play' page 1 from the second file. | 15:49.52 |
| | It's certainly not something trivial. | 15:50.07 |
| kens2 | The problem is the page erasure at the start of each PDF page. | 15:50.27 |
| | Also I don't think the PDF interpreter has provision to have 2 PDF files open at the same time, but I may be wrong there | 15:50.52 |
| aditsu | I know how to split/concatenate pdf pages, but I want to overlap them instead | 15:50.54 |
| Robin_Watts | Ah yes, hadn't considered that :) | 15:50.55 |
| aditsu | alternatively, I need a way to add text on a pdf (given the string, font parameters and coordinates) | 15:51.42 |
| kens2 | All the work needs to be done in the PDF itnerpreter, so I'm sure its possible to do it, its also a lot of work I think. | 15:51.46 |
| Robin_Watts | aditsu: That may be easier. | 15:51.50 |
| kens2 | Hmmm, maybe. Need an opinion from Alex. | 15:52.15 |
| | You want to inject content after 'showpage', so you need to redefine showpage, | 15:52.37 |
| | The PDF interpreter uses the systemdict showpage, so you need to redefine that. | 15:52.54 |
| Robin_Watts | You'd need to run the pdf through the pdfwrite device, and somehow find a hook to 'add' new page contents just before the end of a page. | 15:53.04 |
| kens2 | Or you could possibly use a custom EndPage procedure. | 15:53.09 |
| Robin_Watts | what ken said. | 15:53.12 |
| kens2 | Depends how complex the requirement is. | 15:53.31 |
| | If its the same for every page, tehn a custom EndPage would be fine | 15:53.41 |
| Robin_Watts | aditsu: Are you looking for watermarking or something ? | 15:53.57 |
| aditsu | more like form filling, but not necessarily for fillable pdf forms | 15:54.45 |
| Robin_Watts | aditsu: Isn't this what PDF annotations are for ? | 15:55.04 |
| kens2 | Even injecting pdfmark annotations into an existing PDF file would be tricky | 15:55.35 |
| aditsu | Robin_Watts: I don't know, is it? I thought they were for adding comments that are not integral part of the pdf | 15:55.41 |
| kens2 | Annotations are part of the PDF. | 15:55.55 |
| Robin_Watts | aditsu: Well, annotations can be text in arbitrary positions on the page. sounds like what you want to me. | 15:56.06 |
| aditsu | adobe reader has an option to hide all comments | 15:57.46 |
| Robin_Watts | A quick google suggests: http://www.tracker-software.com/product/pdf-xchange-viewer?act%5b69%5d=scr_shot | 15:57.53 |
| | oh, right, that watermarks PDFs. | 15:58.39 |
| kens2 | Heading off, goodnight all | 16:02.19 |
| aditsu | heh, I could use pdf2svg and insert the stuff into the svg | 16:03.23 |
| | ah, the pdfbox library seems to have the features I need | 16:32.36 |
| ray_laptop | what a mess. Took my kids to school and discovered that they had been put into 8th grade instead of 7th. Took most of the first period for them to get straightened out and get their schedule set up. What a way to start at a new school that is about 100 times bigger ! | 16:35.51 |
| aditsu | ray_laptop: maybe they thought your kids were too smart for 7th grade | 17:04.44 |
| mvrhel2 | reading IRC logs: ray_laptop: what a mess | 17:09.33 |
| henrys | ray_laptop:never heard of anything like that, wow. | 17:11.27 |
| ray_laptop | aditsu: maybe -- probably since they were just entering the public system they may have just focused on their SAT test scores. | 17:15.06 |
| | mvrhel2: henrys: the kids took it in stride, but I had to keep my wife from making them nervous :-) | 17:15.52 |
| mvrhel2 | :) | 17:16.03 |
| ray_laptop | I had to keep her from going into the classroom with them and explaining to the teacher all about it (not cool for pre-teens to have Mom there) ;-) | 17:17.01 |
| henrys | it's just a bad time for it to happen - that is a difficult transition as it is, without screwups | 17:17.05 |
| ray_laptop | henrys: well, now other things (like getting lost between classes and being a few minutes late) won't bother them in comparison. | 17:17.59 |
| henrys | you could have just left them and them out of the house a year earlier ;-) | 17:18.46 |
| ray_laptop | at least the guidance counselor and principal got to know them right off | 17:19.09 |
| | henrys: I tried to convince them to just skip 7th, but they thought they might have to work too hard. (see, they _are_ smart) | 17:19.55 |
| | never knew it would be so easy to have them skip a grade, tho' | 17:20.27 |
| henrys | I can't remember learning anything earth shattering in 7th anyway. | 17:20.41 |
| | so you're going from private to public? | 17:21.40 |
| ray_laptop | henrys: the main thing I learned was how to scramble from class to class | 17:21.46 |
| | henrys: yep. tiny private school (4th, 5th and 6th combined class of 21 -- only 4 6th graders) to 6th, 7th 8th middle school of 1,100 | 17:22.49 |
| | I think there are about 400 7th graders, so two orders of magnitude | 17:23.10 |
| | a bit of a change ;-) | 17:23.43 |
| | back to work, now.... | 17:24.39 |
| Robin_Watts | mvrhel2: Morning. | 17:30.47 |
| mvrhel2 | good morning Robin_Watts | 17:30.55 |
| Robin_Watts | So far I've found 3 files that still fail in the plank/pamcmyk4 comparison in a worrying way. | 17:31.08 |
| | I think I've fixed one of them locally. | 17:31.20 |
| mvrhel2 | great | 17:31.27 |
| Robin_Watts | (Maybe 2) | 17:31.31 |
| henrys | did the test include pcl or just pdf and postscript? | 17:31.54 |
| Robin_Watts | At least one of the others is a problem in your new planar simple transparent tiling code. | 17:32.12 |
| | henrys: just pdf and ps, I believe. | 17:32.19 |
| | I'll try and simplify the file down to something easier. | 17:32.38 |
| mvrhel2 | Robin_Watts: ok | 17:32.48 |
| henrys | that's might be an issue as the rop and transparency stuff won't be tested. | 17:32.51 |
| Robin_Watts | When I say "I've found 3", that means "in the files I've looked at so far, I've found 3". There may well be more. | 17:33.09 |
| | But many are fine. | 17:33.14 |
| | and I haven't looked at banding vs non-banding yet. | 17:33.32 |
| | henrys: Sure, but there is no point marcosw running larger tests until we sort these problems out. | 17:33.53 |
| henrys | okay makes sense. | 17:34.22 |
| Robin_Watts | OK that fixes 2 of the 3. | 17:47.00 |
| | Let's cut the file down for the remaining problem. | 17:47.10 |
| mvrhel2 | Robin_Watts: so the last problem is a pattern issue? | 17:47.43 |
| Robin_Watts | yes. | 17:47.51 |
| | It looks like you're calculating the 'plane_span' to step over to find the next plane, and there isn't one there. | 17:48.15 |
| mvrhel2 | there isnt a next plane? | 17:51.26 |
| Robin_Watts | indeed. | 17:52.22 |
| | so maybe it's not really a planar buffer? | 17:52.34 |
| | I'll know more in a bit. | 17:52.46 |
| mvrhel2 | or it only has one plane... | 17:52.49 |
| | henrys: are you around? | 18:44.49 |
| henrys | yup | 18:44.57 |
| mvrhel2 | I am going to check in my SSE2 function for color table interpolation for lcms so that I don't lose it. I am stuck at the same level of performance as the non-sse2 code and I need to take a break from this | 18:45.58 |
| | It will be disabled and I may talk with Robin_Watts a bit about it at the meeting | 18:46.18 |
| henrys | okay fine by me. | 18:46.44 |
| Robin_Watts | mvrhel2: Cool. I'd love to see what you're doing with it. | 18:46.55 |
| mvrhel2 | After going through this excercise though, I am pretty certain I could readily speed up our transparency blending though | 18:47.16 |
| Robin_Watts | mvrhel2: I do wonder if we should be putting more work in moving to lcms2. | 18:47.19 |
| henrys | was this motivated by 692323? | 18:47.26 |
| Robin_Watts | lcms1 is pretty dreadful w.r.t tidying up after memory problems. | 18:47.43 |
| mvrhel2 | Robin_Watts: yes. I agree on lcms2. I looked at lcms2 before I dug into this and luckily they use the same interpolation so it will port over | 18:47.56 |
| Robin_Watts | mvrhel2: Ah, cool. | 18:48.05 |
| mvrhel2 | otherwise I would not have done it | 18:48.14 |
| Robin_Watts | There is a bug open about a file that lcms2 gets wrong and lcms1 gets right. | 18:48.23 |
| | If we can fix that, then it would be worth getting marcos to do an lcms1 vs lcms2 run. | 18:48.45 |
| | (maybe it's worth getting that done anyway?) | 18:49.14 |
| mvrhel2 | henrys: in general anything with images this is a bottle neck for us. that is why I was looking at it. bug 692323 would likely improve with what I am thinking of doing with SSE and transparency | 18:49.44 |
| | Robin_Watts: yes, I will take a look at lcms2 this week | 18:50.13 |
| henrys | if we can't get the call counts down - smarter higher up I don't see how it would have much effect. | 18:50.17 |
| | but I imagine it is something worth doing for other jobs. | 18:50.59 |
| mvrhel2 | well the interpolation itself is readily parrallelized (spelling) but the issue is that the table data itself is not so well organized to get things set up to do this | 18:52.37 |
| | anyway, in transparency, we have long strips of data that we can do blending on in parallel. With 16 values at once | 18:53.25 |
| | for blending | 18:53.28 |
| | need to eat lunch. bbiab | 18:53.46 |
| Robin_Watts | mvrhel2: I haven't looked at this, so may be talking rubbish, but... is there scppe for us changing lcms (or our use of it) so that rather than calling 'interpolate this value', we can call 'interpolate this buffer of values' ? | 18:54.02 |
| | mvhrel2: In my home directory on casper, there is out13.pdf | 18:56.52 |
| | gswin32c.exe -r300 -sDEVICE=plank -o out.ppm out13.pdf SEGVs due to the pattern problem I was talking about earlier. | 18:57.26 |
| | I'll keep bashing, but will be called for dinner in a bit. | 18:58.10 |
| | It's to do with clists. clist-playback_band calls gx_dc_pattern_read which calls gx_dc_pattern_read_raster and that allocates a single plane. | 19:15.47 |
| | It's only reading one plane out of the clist, I think. I am not qualified to comment on how many it is writing in. | 19:18.57 |
| mvrhel2 | Robin_Watts: lcms has a call into it to "interpolate a buffer of values" but internally, I don't know if it really takes advantage of this | 20:11.44 |
| henrys | I wonder if we have some windows configuration issue - the job in 692323 renders in 6 minutes for me on a macbook 2.53G and alexcher got 19 minutes. alexcher what is your platform? | 20:39.38 |
| | oh I assume alex was using a linux box (gprof output) | 20:42.01 |
| mvrhel2 | hmmm. I do see an issue in cmsTetrahedralInterp16 where there is a whole decision cascade inside a loop where the decision is not dependent upon the loop index | 20:43.15 |
| alexcher | henrys: 64-biy Linux on AMD Sempron(tm) Processor LE-1300, 1994.92 bogomips | 20:47.26 |
| henrys | mvhrel2: I have a line by line profile of that code if you want it. | 20:48.36 |
| | I wish I could publish it with html or something - it's trapped in shark | 20:49.31 |
| mvrhel2 | I am pulling the decision out of the loop and will run a profile comparison on windoze | 20:51.01 |
| henrys | most of the time is spent on line 997 | 20:51.07 |
| | 30% | 20:51.31 |
| mvrhel2 | oh my line number are different | 20:51.47 |
| | is that the rescale line? division by ffff | 20:52.01 |
| henrys | yes | 20:52.26 |
| | right after // There is a lot of math hidden ... | 20:52.47 |
| mvrhel2 | yes | 20:53.15 |
| henrys | and it report 17% as the procedure is entered suggesting too much stack I don't know how large LPL16PARAMS is... | 20:53.58 |
| | perhaps naive - are we caching these calls? | 20:54.58 |
| mvrhel2 | that is not large | 20:55.23 |
| | but luttable can be large | 20:55.30 |
| | and we end up having to access across a fairly large part if the table is large | 20:56.10 |
| | a table be 33x33x33 * 2 bytes * num outputs | 20:56.45 |
| | but usually not that large | 20:57.09 |
| henrys | alexcher that seems odd to me that our times are so different I guess I should upload my profile. | 20:57.19 |
| alexcher | henrys: My computer is rather slow. I think the results are just right. | 20:59.55 |
| henrys | okay | 21:00.44 |
| | just over 10% setting fx, fy, and fz right at the beginning | 21:03.03 |
| mvrhel2 | henrys: so I went from about 13% of time spent in the function to 11% time spent in it when I moved the decision tree out from the loop | 21:04.41 |
| | this is on a different file though | 21:04.50 |
| | let me test the file from the bug | 21:04.54 |
| | the set-up stuff can be sped up a bit through the use of some SSE2 commands that I have | 21:05.40 |
| | but let me test on the bug file | 21:05.58 |
| | what resolution are you running? | 21:06.40 |
| henrys | 1200 | 21:07.06 |
| | I copy pasted comment 1's command | 21:07.31 |
| mvrhel2 | ok | 21:07.48 |
| | see how long this takes to profile... | 21:08.25 |
| | at 1200 | 21:08.28 |
| | I am thinking also that we should probably do the color management before interpolation for this situation | 21:09.08 |
| | Robin_Watts mentioned that earlier. | 21:09.19 |
| | when we are enlarging we should interpolate in the destination color space | 21:09.46 |
| | taking too long... | 21:15.27 |
| | my poor machine is getting over heated... | 21:16.57 |
| | ok. so at 600 dpi, I have about 30 percent is zoom_y and 16 precent is the lcms interpolation | 21:17.45 |
| henrys | mine has more in lcms - let me try 600 | 21:19.32 |
| mvrhel2 | with my change I have 30% is zoom_y and 14.8 precent is the lcms interpolation. Nothing to great | 21:19.51 |
| henrys | oh okay this is with your change. | 21:20.21 |
| | processor? | 21:20.28 |
| mvrhel2 | Intel 2 Duo | 21:21.04 |
| | that is at 600dpi. | 21:21.13 |
| henrys | I get about 40 sec. 600 dpi | 21:21.35 |
| mvrhel2 | the numbers above are percent time from the profiler. Let me see what the absolute time is | 21:22.36 |
| henrys | it sounds like you've made a good speedup though - before lcms was the bottleneck. | 21:22.40 |
| mvrhel2 | with release version | 21:22.41 |
| | do you not see zoom_y as the biggest issue? | 21:23.34 |
| | that is what I am seeing | 21:23.37 |
| | with and without my change | 21:23.48 |
| | i.e. 30% of the time is spent in zoom_y | 21:24.13 |
| henrys | nope 40% lcms and 30% interp - that's why I was asking if I am seeing different code here on the mac - | 21:24.27 |
| | something seems peculiar with the users and alexcher's numbers. | 21:24.45 |
| mvrhel2 | with 15 percent of the time on line 389 in zoom_y | 21:25.02 |
| henrys | yes shark say please convert this to sse. | 21:25.47 |
| | s/say/says | 21:25.55 |
| mvrhel2 | at line line 389? | 21:26.22 |
| henrys | and it whines about mixing singles/doubles do we need doubles here? | 21:26.34 |
| mvrhel2 | I don't think we need doubles there | 21:26.45 |
| henrys | right weigth += ... | 21:26.51 |
| mvrhel2 | yes | 21:26.56 |
| | let me change that to float | 21:27.22 |
| | for sec | 21:27.24 |
| | whoa | 21:28.35 |
| | that made a difference | 21:28.40 |
| henrys | as far as I know we should not be using gs_memset? alexcher? should go to the system memset which on linux uses something fancy with 64 bit registers. | 21:28.48 |
| mvrhel2 | zoom_y went from 30 percent down to 16.5 percent | 21:29.27 |
| | by changing the double to float | 21:29.33 |
| henrys | did you check the quality? | 21:29.40 |
| mvrhel2 | hehe no | 21:29.44 |
| | well it looks OK. | 21:30.18 |
| | let me compare with and without | 21:30.24 |
| henrys | it can be all doubles too and will probably be close to what is now. mixed is the problem. | 21:32.12 |
| mvrhel2 | right. I dont see any diff between the mixed case and float only | 21:33.37 |
| henrys | clusterpush and commit? | 21:34.41 |
| mvrhel2 | yes. let me go the float route and see what the clusterpush shows | 21:35.00 |
| henrys | we'll see where the bottleneck goes next ;-) | 21:35.15 |
| mvrhel2 | yes | 21:35.21 |
| | I think it will be lcms now | 21:35.29 |
| | but if we do the color management where ever it will be doing less data we will make an improvement | 21:36.01 |
| henrys | alexcher can you post a profile with the command line in comment 1? | 21:37.46 |
| | ie don't disable anything. | 21:38.02 |
| alexcher | henrys: gs_memset is used only by pg build. | 21:38.23 |
| henrys | oh that makes sense. | 21:38.37 |
| | I would buy zoom and put the dinosaur away ;-) | 21:39.03 |
| alexcher | henrys: The original profile has been posted first. | 21:39.06 |
| mvrhel2 | crap. need to reboot | 21:39.25 |
| | brb | 21:39.27 |
| henrys | oh i'm sorry alexcher I didn't see comment #2 | 21:40.12 |
| mvrhel2 | henrys: OK. I am pushing the double/float fix. need to run and errand. I will check this when I get back | 21:58.24 |
| | an errand... | 21:58.47 |
| henrys | easy change and good progress | 21:59.20 |
| | can't beat that. | 21:59.26 |
| | alexcher were you using -O2 I used tree-vectorize and the time went to 12 minutes from 6. | 22:04.47 |
| | I think that is on with -O3 | 22:05.13 |
| | that was wrong sorry ignore my babblings | 22:08.44 |
| | I accidentally ran the debug build | 22:09.04 |
| alexcher | henrys: I was using the default release build to do timing. | 22:16.53 |
| jdeisenberg | I am getting the infamous rangecheck error, last OS Error: 11 under strange (to me) circumstances. | 22:44.54 |
| | I have file A, which displays just fine in ghostview. I then use the "poster" utility to convert it to a larger file B of several pages. | 22:45.20 |
| | "poster" works by wrapping the original page in a bind def, and then calling that definition to tile the pages. | 22:45.38 |
| | File B gives me the error. | 22:45.43 |
| | I eliminated all the "cruft" around the bind def in file B, and did a diff with the contents of file A -- they are identical. | 22:46.04 |
| alexcher | jdeisenberg: Not every PS file can be converted to a procedure. | 22:48.08 |
| jdeisenberg | what kind of things should I look for in the PS that would cause it to go boom? | 22:48.35 |
| alexcher | jdeisenberg: inline data | 22:49.09 |
| jdeisenberg | an example, please? (I am not a PS expert) | 22:49.29 |
| alexcher | Something like //SYSIN DD * | 22:50.13 |
| | Joke, please post the file somewhere. | 22:50.38 |
| jdeisenberg | my gosh that looks like JCL. | 22:50.40 |
| | I'm old enough to recognize that :) | 22:50.52 |
| | post the original and the revised? | 22:51.08 |
| alexcher | Any one. You can also email the file to alex.cherepanov@artifex.com | 22:52.37 |
| jdeisenberg | http://evc-cit.info/map/rangecheck.ps | 22:52.45 |
| | original (good) file is at http://evc-cit.info/map/goodfile.eps | 22:54.03 |
| | I have been able to successfully convert other files. I'm using ghostview version 3.7.1 on Linux, if that is useful info. | 22:58.12 |
| alexcher | jdeisenberg: I think, the file is just too big. Let be do a few tests . | 23:00.07 |
| jdeisenberg | pity; it didn't *seem* like that much data. | 23:02.25 |
| | Thanks very much, alexcher! | 23:03.13 |
| alexcher | jdeisenberg: wait | 23:03.21 |
| jdeisenberg | OK; thanking you for at least being willing to take a look at it :) | 23:03.45 |
| alexcher | jdeisenberg: PS arrays are limited to 64k-1 elements. The file you have has about 70K elements. | 23:04.26 |
| jdeisenberg | Where is the extra data coming from between the good file and the bad file? | 23:04.56 |
| | never mind that question; I wasn't thinking. | 23:05.26 |
| alexcher | jdeisenberg: You need to split the procedure into 2 smaller ones /foo {...} exec {...} exec} def | 23:05.29 |
| | jdeisenberg: I've tested this approach and it works. | 23:06.05 |
| jdeisenberg | that would be the print_complete_page_0 procedure, yes? | 23:07.07 |
| alexcher | jdeisenberg: Yes | 23:07.33 |
| jdeisenberg | OK; that gives me a good starting point. | 23:07.59 |
| | is exec nested in foo? | 23:08.33 |
| alexcher | jdeisenberg: Sorry, /foo {{...} exec {...} exec} def | 23:09.12 |
| jdeisenberg | alexcher: thank you; that solves it nicely. | 23:16.56 |
| Robin_Watts | mvrhel2: You here ? | 23:28.44 |
| jdeisenberg | bye for now all. | 23:31.25 |
| Robin_Watts | mvrhel2: Going to bed now. Wondered if you had any insights into the pattern clist stuff I mentioned above. | 23:34.50 |
| | I'll check logs in the morning. | 23:35.03 |
| | Forward 1 day (to 2011/09/01)>>> | |
IRC Logs
