| | <<<Back 1 day (to 2012/11/21) | 2012/11/22 |
| paulgardiner | Robin_Watts: ping | 12:29.26 |
| Robin_Watts | pong | 12:29.31 |
| paulgardiner | Found the problem with saving. MuPDF can't save back to the same file... perhaps unsurprisingly... although it surprised me. :-) | 12:30.12 |
| Robin_Watts | paulgardiner: Ah, well, it probably can save back to the same file, but only if you've run through and loaded every object into memory first :) | 12:30.58 |
| paulgardiner | I wondered that. | 12:31.16 |
| Robin_Watts | which is a non-starter I feel. | 12:31.26 |
| paulgardiner | I'm guessing that's probably not a good idea, and a tmp file would be better | 12:31.30 |
| Robin_Watts | yeah. | 12:31.42 |
| paulgardiner | ok, ta | 12:31.50 |
| tor8 | paulgardiner: Robin_Watts: oh no. don't tell me you're doing fopen("output.pdf", "w") and clobbering, instead of the safe juggling of creating to a temp file and renaming.... | 13:07.01 |
| Robin_Watts | tor8: I think he's doing the latter now :) | 13:07.22 |
| tor8 | we certainly don't want a crash, ^C or other error to destroy the user's file! | 13:07.31 |
| paulgardiner | Yes, I suppos the direct output to the original file wasn't the brightest of ideas, even if MuPDF didn't need access to the original during writing | 13:10.41 |
| | But I'm sure there are still many ways we'll find to clobber the users file even without that. :-) | 13:12.04 |
| Robin_Watts | So a large percentage of time in mupdf is spent calling the divide function. | 13:15.27 |
| | Can I figure out where in mupdf that's from though? can I hell. | 13:15.39 |
| tor8 | fz_idiv? | 13:16.27 |
| Robin_Watts | paulgardiner: Did you have a funky trick for dividing by 2^n -1 | 13:16.40 |
| | ? | 13:16.42 |
| | tor8: no, __aeabi_idiv | 13:17.01 |
| tor8 | Robin_Watts: right. because we have a fz_idiv used in the scan converter | 13:17.18 |
| Robin_Watts | yeah. | 13:17.23 |
| paulgardiner | Robin_Watts: yes. I'll see if I can remember it | 13:17.44 |
| Robin_Watts | It's possible that fz_idiv ends up calling __aeabi_idiv... | 13:17.46 |
| paulgardiner | Robin_Watts: I sort of remember it, but exact details depend on the range of the thing being divided and the bit-size of the result required | 13:23.09 |
| Robin_Watts | http://www.hackersdelight.org/divcMore.pdf ? | 13:23.28 |
| | oh, ass, this is / ((2^n)+1) | 13:23.57 |
| paulgardiner | Something similar may work. The series then alternates sign | 13:25.15 |
| | 1/(2^n + 1) = 1/2^n * (1 - 1/2^n + 1/2^2n - 1/2^3n + 1/2^4n ... ) | 13:27.04 |
| Robin_Watts | 1/17 is the one I'm looking at now. | 13:27.36 |
| | That's alternating 4 bits on, 4 bits off. | 13:28.18 |
| | (x>>4)-(x>>8)+(x>>12)-(x>>16) etc | 13:28.36 |
| paulgardiner | yep | 13:28.51 |
| | Same trick | 13:29.08 |
| | There was some sort of extra bit concerning working out if there was an infinite carry, but I can't remember how it went now. | 13:30.24 |
| Robin_Watts | tor8, paulgardiner: Are we ever planning to allow the android app to change the amount of antialiasing we use? | 14:46.35 |
| | If not, we can predefine AA_BITS to 8 and save some time. | 14:46.51 |
| tor8 | Robin_Watts: I doubt it; that's just for mudraw really? | 14:47.51 |
| Robin_Watts | Cool. | 14:47.59 |
| tor8 | at least that's the only place I see it as useful. | 14:48.13 |
| | even eink devices have grayscale screens | 14:48.21 |
| Robin_Watts | For grayscale screens, we'd presumably predefine it to 4 or something. | 14:49.49 |
| tor8 | Robin_Watts: we may want to do that anyway, since it's just used for vector art | 15:57.20 |
| | and we don't do gamma correct alpha blending anyway... | 15:57.32 |
| | which is something we really should be doing for text at least | 15:57.42 |
| | (i.e. do the transparency blending and compositing in a linear color space) | 15:58.12 |
| Robin_Watts | predefining it to a fixed value enables the compiler to optimise out lots of divides, so they don't really show up in the profiles. | 15:58.14 |
| tor8 | Robin_Watts: yeah. predefining it is definitely the way to go on non-desktop apps. | 15:58.33 |
| | should probably do the same for iOS | 15:58.38 |
| Robin_Watts | So the slowest file I have has the top of the profile that looks like: | 15:58.48 |
| | 22.67 4.18 4.18 3446 1.21 2.17 fz_scan_convert_aa | 15:59.06 |
| | 19.14 7.71 3.53 912 3.87 3.87 fz_subsample_pixmap_ARM | 15:59.08 |
| | 12.91 10.09 2.38 94070 0.03 0.03 scale_row_to_temp1 | 15:59.10 |
| | 7.65 11.50 1.41 457 3.09 4.86 read_faxd | 15:59.12 |
| | 7.10 12.81 1.31 22478167 0.00 0.00 non_zero_winding_aa | 15:59.13 |
| | 6.56 14.02 1.21 22636515 0.00 0.00 add_span_aa | 15:59.15 |
| | 6.07 15.14 1.12 459 2.44 2.44 fz_unpack_tile | 15:59.17 |
| | 3.96 15.87 0.73 1281069 0.00 0.00 fz_paint_affine_color_N_near | 15:59.18 |
| | 3.85 16.58 0.71 1513292 0.00 0.00 fz_paint_span_with_color_4 2.82 17.10 0.52 2845493 0.00 0.00 find_changing | 15:59.20 |
| | 2.06 17.48 0.38 459 0.83 12.71 decomp_image_from_stream | 15:59.21 |
| | 1.19 17.70 0.22 50402 0.00 0.00 scale_row_from_temp | 15:59.23 |
| tor8 | a substantial amount of line-art on that file? | 15:59.48 |
| Robin_Watts | Now, that page has 160 fax images on it, each 15571 wide and 100 or so tall. | 15:59.50 |
| | I'm not aware of there being *any* line art on that page. | 16:00.18 |
| | I mean, presumably there must be, but... | 16:00.41 |
| chrisl | Behing the images? I've seen that before | 16:01.23 |
| | s/Behing/Behind | 16:01.35 |
| Robin_Watts | chrisl: I'm thinking so. Just taking the file to bits now. | 16:02.13 |
| chrisl | I had one a few years ago that had a page full of transparency, then the entire page covered by a scanned, opaque image! | 16:03.17 |
| tor8 | Robin_Watts: do we still use the scan converter to antialias the image edges? | 16:03.46 |
| Robin_Watts | As far as I can tell, it's just images. | 16:03.59 |
| kens2 | why do you think it has linework ? | 16:04.16 |
| Robin_Watts | with clipping paths. | 16:04.28 |
| | kens2: see the profile I posted. | 16:04.34 |
| kens2 | You can apply clip to images | 16:04.39 |
| Robin_Watts | fz_scan_convert_aa etc. | 16:04.41 |
| | kens2: Yes, I am aware of that. | 16:05.00 |
| kens2 | SO that's scan convertting as path ? | 16:05.08 |
| Robin_Watts | kens2: That's the point at discussion. | 16:05.19 |
| | sorry, fz_scan_convert_aa is the code that scan converts a path, with antialiasing, yes. | 16:05.40 |
| tor8 | Robin_Watts: if clips are rectangular and unrotated, they should short circuit to scissor operations in the draw device | 16:05.49 |
| Robin_Watts | The ctm is 90 degrees, and there are rectangle for paths. | 16:06.25 |
| | Ah, right. Each image is not accompanied with a clip path. | 16:27.39 |
| | each image first has the area it will fill cleared by having a white rectangle drawn there first. | 16:27.57 |
| kens2 | Ah, stupid creator | 16:28.12 |
| Robin_Watts | But I can't believe that scan converting a rectangular box should take that long. | 16:31.26 |
| kens2 | Packing up time, goodnight everyone | 17:10.13 |
| Robin_Watts | tor8: Scan converter optimisation commit on my repo for your consideration. | 20:01.18 |
| henrys` | powerline networking experts: if I get say this http://www.amazon.com/Netgear-XAVB5004-Powerline-Network-Adapter/dp/B004PA9PBQ I'll be able to plug 2 cluster machines into it and we'll be able to ssh to it and all that stuff like a regular cluster server? It won't work if this bridge (or actually a switch) won't allow incoming ssh connections to pass through. The main use of these devices is for home media stuff where incoming | 20:05.58 |
| | connections like we want would be undesirable. | 20:05.58 |
| Robin_Watts | henrys`: Powerline network adapters work at the ethernet level. | 20:07.54 |
| | They read packets in, retransmit across the powerline, and play them out again. | 20:08.11 |
| | They work at a lower level than ssh etc. | 20:08.46 |
| | That seems expensive. | 20:09.20 |
| henrys` | this is more than an adapter it's a bridge. | 20:09.41 |
| Robin_Watts | It's gigabit, which is good. | 20:09.58 |
| henrys` | which I assume can firewall incoming stuff | 20:10.13 |
| Robin_Watts | but basically all you need is 2 plug units and a switch, right ? | 20:10.14 |
| | No, no firewall. | 20:10.33 |
| | Why do you need a firewall? | 20:11.07 |
| | Your home router that connects to the net will firewall your network from the net. | 20:11.31 |
| henrys` | The product that I sent the link for is a bridge and an adapter, I assume the adapter is for my router. | 20:11.32 |
| | I am concerned this netgear switch is going to firewall incoming connections, I don't want it to do that. | 20:12.20 |
| Robin_Watts | Right, I'm looking at the amazon page. | 20:12.26 |
| henrys` | if it works it will be fantastic. I'll move the cluster machines and a network printer to the basement... wonderful. | 20:13.22 |
| Robin_Watts | http://www.netgear.com/home/products/powerline-and-coax/high-performance/xavb5004.aspx | 20:13.24 |
| | So, basically you plug the black box into the basement, and that gives you 4 ethernet sockets. | 20:13.49 |
| | You plug the white thing into a socket next to you router, and use ethernet to connect from router to the white thing. | 20:14.18 |
| | The only firewall is the one in your router. | 20:14.36 |
| henrys` | okay than it should work great. | 20:15.04 |
| Robin_Watts | yeah, looks ideal. | 20:16.55 |
| | Forward 1 day (to 2012/11/23)>>> | |
IRC Logs