50804 2010-12-10 00:38:57 -0800 [chromium] Reduce canvas.toDataURL("image/jpeg") run-time cost by 10% 2010-12-17 01:49:09 -0800 1 1 1 Unclassified WebKit Canvas 528+ (Nightly build) PC Windows 7 RESOLVED FIXED P2 Normal --- 1 noel.gordon webkit-unassigned hbono levin mdelaney7 mikelawther senorblanco webkit.review.bot oldest_to_newest 320005 0 noel.gordon 2010-12-10 00:38:57 -0800 Reduce canvas.toDataURL("image/jpeg") run-time cost by 10% 320010 1 76173 noel.gordon 2010-12-10 00:58:34 -0800 Created attachment 76173 patch 320011 2 webkit.review.bot 2010-12-10 01:00:16 -0800 Attachment 76173 did not pass style-queue: Failed to run "['WebKitTools/Scripts/check-webkit-style', '--diff-files', u'WebCore/ChangeLog', u'WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp']" exit_code: 1 WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:86: One or more unexpected \r (^M) found; better to use only a \n [whitespace/carriage_return] [1] Suppressing further [whitespace/carriage_return] reports for this file. WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:37: Alphabetical sorting problem. [build/include_order] [4] WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:90: Tests for true/false, null/non-null, and zero/non-zero should all be done without equality comparisons. [readability/comparison_to_zero] [5] Total errors found: 12 in 2 files If any of these errors are false positives, please file a bug against check-webkit-style. 320013 3 76174 noel.gordon 2010-12-10 01:06:04 -0800 Created attachment 76174 patch style: can't avoid the non-alphabetical include ordering nit. 320014 4 noel.gordon 2010-12-10 01:07:48 -0800 so I expected to nitted again. 320015 5 webkit.review.bot 2010-12-10 01:08:49 -0800 Attachment 76174 did not pass style-queue: Failed to run "['WebKitTools/Scripts/check-webkit-style', '--diff-files', u'WebCore/ChangeLog', u'WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp']" exit_code: 1 WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:37: Alphabetical sorting problem. [build/include_order] [4] WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:90: Tests for true/false, null/non-null, and zero/non-zero should all be done without equality comparisons. [readability/comparison_to_zero] [5] Total errors found: 2 in 2 files If any of these errors are false positives, please file a bug against check-webkit-style. 320026 6 76174 eric 2010-12-10 01:38:23 -0800 Comment on attachment 76174 patch You should consider using webkit-patch upload to upload your patches so that they get marked properly as patch files. 320027 7 76174 eric 2010-12-10 01:40:51 -0800 Comment on attachment 76174 patch View in context: https://bugs.webkit.org/attachment.cgi?id=76174&action=review It's a little surprising that this unroll makes things faster, but OK. Are you certain in the stability of your numbers? > WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:90 > + if ((alpha != 0) && (alpha != 255)) { We generally use just "alpha" instead of "alpha != 0", but it might be OK here. 320911 8 noel.gordon 2010-12-12 17:40:43 -0800 (In reply to comment #7) > (From update of attachment 76174 [details]) > View in context: https://bugs.webkit.org/attachment.cgi?id=76174&action=review > > It's a little surprising that this unroll makes things faster, but OK. Are you certain in the stability of your numbers? Test image is 1800 x 1350 px JPEG. I load it into a <canvas> and then I do back-to-back canvas.toDataURL("image/jpeg") calls in js, logging the time as I go. My results are: Safari 5.0.2 win32 (7533.18.5): 12:15:45:062 extract image/jpeg 145 ms length: 667763 bytes 12:15:45:214 extract image/jpeg 146 ms length: 667763 bytes 12:15:45:366 extract image/jpeg 146 ms length: 667763 bytes 12:15:45:517 extract image/jpeg 145 ms length: 667763 bytes 12:15:45:668 extract image/jpeg 145 ms length: 667763 bytes 12:15:45:819 extract image/jpeg 144 ms length: 667763 bytes 12:15:45:970 extract image/jpeg 144 ms length: 667763 bytes 12:15:46:122 extract image/jpeg 145 ms length: 667763 bytes 12:15:46:273 extract image/jpeg 145 ms length: 667763 bytes 12:15:46:424 extract image/jpeg 145 ms length: 667763 bytes 12:15:46:575 extract image/jpeg 144 ms length: 667763 bytes Chrome ToT win32: 12:09:16:144 extract image/jpeg 129 ms length: 630179 bytes 12:09:16:301 extract image/jpeg 130 ms length: 630179 bytes 12:09:16:458 extract image/jpeg 130 ms length: 630179 bytes 12:09:16:617 extract image/jpeg 130 ms length: 630179 bytes 12:09:16:752 extract image/jpeg 130 ms length: 630179 bytes 12:09:16:909 extract image/jpeg 129 ms length: 630179 bytes 12:09:17:044 extract image/jpeg 129 ms length: 630179 bytes 12:09:17:203 extract image/jpeg 130 ms length: 630179 bytes 12:09:17:339 extract image/jpeg 130 ms length: 630179 bytes 12:09:17:474 extract image/jpeg 130 ms length: 630179 bytes 12:09:17:632 extract image/jpeg 130 ms length: 630179 bytes Chrome ToT win32 with this change: 12:13:14:649 extract image/jpeg 115 ms length: 630179 bytes 12:13:14:769 extract image/jpeg 115 ms length: 630179 bytes 12:13:14:891 extract image/jpeg 114 ms length: 630179 bytes 12:13:15:011 extract image/jpeg 114 ms length: 630179 bytes 12:13:15:131 extract image/jpeg 115 ms length: 630179 bytes 12:13:15:253 extract image/jpeg 115 ms length: 630179 bytes 12:13:15:373 extract image/jpeg 115 ms length: 630179 bytes 12:13:15:494 extract image/jpeg 115 ms length: 630179 bytes 12:13:15:617 extract image/jpeg 115 ms length: 630179 bytes 12:13:15:738 extract image/jpeg 115 ms length: 630179 bytes 12:13:15:857 extract image/jpeg 114 ms length: 630179 bytes worth it? > > WebCore/platform/image-encoders/skia/JPEGImageEncoder.cpp:90 > > + if ((alpha != 0) && (alpha != 255)) { > > We generally use just "alpha" instead of "alpha != 0", but it might be OK here. agree, OK here. 320952 9 76174 eric 2010-12-13 00:20:34 -0800 Comment on attachment 76174 patch Those are definitely small numbers. :) But 10% is still 10%. Woud be even better if we had a benchmark for this sort of thing, but 115ms is pretty small already. :) WebCore/benchmarks is where such a benchmark would go. 320973 10 76174 webkit.review.bot 2010-12-13 01:12:15 -0800 Comment on attachment 76174 patch Clearing flags on attachment: 76174 Committed r73890: <http://trac.webkit.org/changeset/73890> 320975 11 webkit.review.bot 2010-12-13 01:12:21 -0800 All reviewed patches have been landed. Closing bug. 321115 12 76174 senorblanco 2010-12-13 07:46:02 -0800 Comment on attachment 76174 patch I'm fine with this change, but I'm surprised that putting the conditional in the loop doesn't make performance worse. Perhaps your tests are mostly alpha 255 or 0, and are benefiting disproportionately from branch prediction? You could also look at writing an int at a time to the output, rather than a char at a time. Of course, this would require unrolling 3X or so in order to write RGBR GBRG BRGB, kind of a pain, but usually nets a good speedup. If this continues to be a bottleneck, it might be worth looking at an SSE2 version as well. 321633 13 noel.gordon 2010-12-13 20:02:07 -0800 > Those are definitely small numbers. :) But 10% is still 10%. Woud be even better if we had a benchmark for this sort of thing, but 115ms is pretty small already. :) WebCore/benchmarks is where such a benchmark would go. Happy to take the 10%, we are about to get faster still. hbono is is working on bug 50054, which should make the encoding phase 3x faster. 321640 14 noel.gordon 2010-12-13 20:37:32 -0800 (In reply to comment #12) > I'm fine with this change, but I'm surprised that putting the conditional in the loop doesn't make performance worse. > Perhaps your tests are mostly alpha 255 or 0, and are benefiting disproportionately from branch prediction? Apply the first statement to canvas.getImageData and canvas.putImageData :) But yes, I believe the else branch benefits from alpha 0|255 and that it also exposes the cost of the existing "lookup" table. This is not new idea; see the PNG ncoder implementation in the skia repository for example. > You could also look at writing an int at a time to the output, rather than a char at a time. Of course, this would > require unrolling 3X or so in order to write RGBR GBRG BRGB, kind of a pain, but usually nets a good speedup. Yes a pain. Would it be better to build premultipication/unpremultiplication tables? http://mxr.mozilla.org/mozilla2.0/source/content/canvas/src/nsCanvasRenderingContext2D.cpp#3917 http://mxr.mozilla.org/mozilla2.0/source/content/canvas/src/nsCanvasRenderingContext2D.cpp#3996 > If this continues to be a bottleneck, it might be worth looking at an SSE2 version as well. Indeed, hbono suggested exactly that approach, in particular that he could implement the (un)premultiplication via SSE or MMX. I'm awaiting the result of his bug 50054 (I'm not synced to it yet) to measure the expected encoding phase speed-up (3x), and to consider if the unpremultiplication is a "bottleneck" worth more optimization ... 322423 15 hbono 2010-12-14 18:21:02 -0800 (In reply to comment #13) > Happy to take the 10%, we are about to get faster still. hbono is is working > on bug 50054, which should make the encoding phase 3x faster. This is just for your information. I have been integrating premultiply/unpremultiply operations into libjpeg-turbo so we can directly use skia pixels, i.e. we don't need Vector<JSAMPLE>. For those who are interested in it, I have uploaded a part of this change (only unpremultiply operations for SSE) to Rietveld: <http://codereview.chromium.org/5862001/>. Regards, 322498 16 noel.gordon 2010-12-14 21:44:02 -0800 (In reply to comment #14) > (In reply to comment #12) > > I'm fine with this change, but I'm surprised that putting the conditional in the loop doesn't make performance worse. > > Perhaps your tests are mostly alpha 255 or 0, and are benefiting disproportionately from branch prediction? > > Apply the first statement to canvas.getImageData and canvas.putImageData :) correction, those are templated. 322522 17 noel.gordon 2010-12-14 23:59:43 -0800 (In reply to comment #15) @hbono thanks for http://codereview.chromium.org/5862001, sounds good. I should note, be aware of http://www.w3.org/Bugs/Public/show_bug.cgi?id=11431 323661 18 noel.gordon 2010-12-17 01:49:09 -0800 http://www.w3.org/Bugs/Public/show_bug.cgi?id=11431 is easy to resolve - the spec is correct. Opened bug 51237 to correct chromium. It's related to this bug, and we maintain the speedup mentioned here, but there's no need to unpremultiply RGB. See bug 51237 for details. 76173 2010-12-10 00:58:34 -0800 2010-12-10 01:06:04 -0800 patch git-diff.patch text/plain 2522 noel.gordon ZGlmZiAtLWdpdCBhL1dlYkNvcmUvQ2hhbmdlTG9nIGIvV2ViQ29yZS9DaGFuZ2VMb2cKaW5kZXgg YjZlMjMwMC4uZThiYTM4YSAxMDA2NDQKLS0tIGEvV2ViQ29yZS9DaGFuZ2VMb2cKKysrIGIvV2Vi Q29yZS9DaGFuZ2VMb2cKQEAgLTEsMyArMSwxOSBAQAorMjAxMC0xMi0xMCAgTm9lbCBHb3Jkb24g IDxub2VsLmdvcmRvbkBnbWFpbC5jb20+CisKKyAgICAgICAgUmV2aWV3ZWQgYnkgTk9CT0RZIChP T1BTISkuCisKKyAgICAgICAgW2Nocm9taXVtXSBSZWR1Y2UgY2FudmFzLnRvRGF0YVVSTCgiaW1h Z2UvanBlZyIpIHJ1bi10aW1lIGNvc3QgYnkgMTAlCisgICAgICAgIGh0dHBzOi8vYnVncy53ZWJr aXQub3JnL3Nob3dfYnVnLmNnaT9pZD01MDgwNAorCisgICAgICAgIEZvbGxvdyBvbiBmcm9tIHI3 MzE3MywgdW5yb2xsIHRoZSBTa1VuUHJlTXVsdGlwbHk6OlBNQ29sb3JUb0NvbG9yKCkgY2FsbAor ICAgICAgICB2aXouLCBjb21wdXRlIHRoZSB1bnByZW11bHRpcGxhdGNpb24gaW4tcGxhY2UuIFRo aXMgcmVkdWNlcyB0aGUgcnVuLXRpbWUKKyAgICAgICAgY29zdCBvZiBqcGVnIGVuY29kaW5nIGJ5 IDEwLTE1JSBmb3IgbXkgaW1hZ2UgdGVzdCBzZXQuCisKKyAgICAgICAgTm8gbmV3IHRlc3RzOiBj YW52YXMudG9EYXRhVVJMKCkgaXMgY292ZXJlZCBieSBleGlzdGluZyB0ZXN0cy4KKworICAgICAg ICAqIHBsYXRmb3JtL2ltYWdlLWVuY29kZXJzL3NraWEvSlBFR0ltYWdlRW5jb2Rlci5jcHA6Cisg ICAgICAgIChXZWJDb3JlOjpwcmVNdWx0aXBsaWVkQkdSQXRvUkdCKToKKwogMjAxMC0xMi0wOCAg U2FtIFdlaW5pZyAgPHNhbUB3ZWJraXQub3JnPgogCiAgICAgICAgIFJldmlld2VkIGJ5IERhcmlu IEFkbGVyLgpkaWZmIC0tZ2l0IGEvV2ViQ29yZS9wbGF0Zm9ybS9pbWFnZS1lbmNvZGVycy9za2lh L0pQRUdJbWFnZUVuY29kZXIuY3BwIGIvV2ViQ29yZS9wbGF0Zm9ybS9pbWFnZS1lbmNvZGVycy9z a2lhL0pQRUdJbWFnZUVuY29kZXIuY3BwCmluZGV4IGMwNDAxOWQuLmQ3ZGFlNTMgMTAwNjQ0Ci0t LSBhL1dlYkNvcmUvcGxhdGZvcm0vaW1hZ2UtZW5jb2RlcnMvc2tpYS9KUEVHSW1hZ2VFbmNvZGVy LmNwcAorKysgYi9XZWJDb3JlL3BsYXRmb3JtL2ltYWdlLWVuY29kZXJzL3NraWEvSlBFR0ltYWdl RW5jb2Rlci5jcHAKQEAgLTM0LDYgKzM0LDcgQEAKICNpbmNsdWRlICJJbnRTaXplLmgiCiAjaW5j bHVkZSAiU2tCaXRtYXAuaCIKICNpbmNsdWRlICJTa1VuUHJlTXVsdGlwbHkuaCIKKyNpbmNsdWRl ICJTa0NvbG9yUHJpdi5oIgogZXh0ZXJuICJDIiB7CiAjaW5jbHVkZSA8c3RkaW8uaD4gLy8ganBl Z2xpYi5oIG5lZWRzIHN0ZGlvLmggRklMRQogI2luY2x1ZGUgImpwZWdsaWIuaCIKQEAgLTgyLDEx ICs4MywxOSBAQCBzdGF0aWMgdm9pZCBoYW5kbGVFcnJvcihqX2NvbW1vbl9wdHIgY29tbW9uKQog Ly8gYmUgaWdub3JlZD8gU2VlIGJ1ZyBodHRwOi8vd2Via2l0Lm9yZy9iLzQwMTQ3Lgogdm9pZCBw cmVNdWx0aXBsaWVkQkdSQXRvUkdCKGNvbnN0IFNrUE1Db2xvciogaW5wdXQsIHVuc2lnbmVkIGlu dCBwaXhlbHMsIHVuc2lnbmVkIGNoYXIqIG91dHB1dCkKIHsKLSAgICB3aGlsZSAocGl4ZWxzLS0g PiAwKSB7Ci0gICAgICAgIFNrQ29sb3IgdW5tdWx0aXBsaWVkID0gU2tVblByZU11bHRpcGx5OjpQ TUNvbG9yVG9Db2xvcigqaW5wdXQrKyk7Ci0gICAgICAgICpvdXRwdXQrKyA9IFNrQ29sb3JHZXRS KHVubXVsdGlwbGllZCk7Ci0gICAgICAgICpvdXRwdXQrKyA9IFNrQ29sb3JHZXRHKHVubXVsdGlw bGllZCk7Ci0gICAgICAgICpvdXRwdXQrKyA9IFNrQ29sb3JHZXRCKHVubXVsdGlwbGllZCk7Cisg ICAgc3RhdGljIGNvbnN0IFNrVW5QcmVNdWx0aXBseTo6U2NhbGUqIHNjYWxlID0gU2tVblByZU11 bHRpcGx5OjpHZXRTY2FsZVRhYmxlKCk7DQorDQorICAgIGZvciAoOyBwaXhlbHMtLSA+IDA7ICsr aW5wdXQpIHsNCisgICAgICAgIGNvbnN0IHVuc2lnbmVkIGFscGhhID0gU2tHZXRQYWNrZWRBMzIo KmlucHV0KTsNCisgICAgICAgIGlmICgoYWxwaGEgIT0gMCkgJiYgKGFscGhhICE9IDI1NSkpIHsN CisgICAgICAgICAgICAqb3V0cHV0KysgPSBTa1VuUHJlTXVsdGlwbHk6OkFwcGx5U2NhbGUoc2Nh bGVbYWxwaGFdLCBTa0dldFBhY2tlZFIzMigqaW5wdXQpKTsNCisgICAgICAgICAgICAqb3V0cHV0 KysgPSBTa1VuUHJlTXVsdGlwbHk6OkFwcGx5U2NhbGUoc2NhbGVbYWxwaGFdLCBTa0dldFBhY2tl ZEczMigqaW5wdXQpKTsNCisgICAgICAgICAgICAqb3V0cHV0KysgPSBTa1VuUHJlTXVsdGlwbHk6 OkFwcGx5U2NhbGUoc2NhbGVbYWxwaGFdLCBTa0dldFBhY2tlZEIzMigqaW5wdXQpKTsNCisgICAg ICAgIH0gZWxzZSB7DQorICAgICAgICAgICAgKm91dHB1dCsrID0gU2tHZXRQYWNrZWRSMzIoKmlu cHV0KTsKKyAgICAgICAgICAgICpvdXRwdXQrKyA9IFNrR2V0UGFja2VkRzMyKCppbnB1dCk7Cisg ICAgICAgICAgICAqb3V0cHV0KysgPSBTa0dldFBhY2tlZEIzMigqaW5wdXQpOworICAgICAgICB9 DQogICAgIH0KIH0KIAo= 76174 2010-12-10 01:06:04 -0800 2010-12-13 07:46:01 -0800 patch git-diff.patch text/plain 2512 noel.gordon ZGlmZiAtLWdpdCBhL1dlYkNvcmUvQ2hhbmdlTG9nIGIvV2ViQ29yZS9DaGFuZ2VMb2cKaW5kZXgg YjZlMjMwMC4uZThiYTM4YSAxMDA2NDQKLS0tIGEvV2ViQ29yZS9DaGFuZ2VMb2cKKysrIGIvV2Vi Q29yZS9DaGFuZ2VMb2cKQEAgLTEsMyArMSwxOSBAQAorMjAxMC0xMi0xMCAgTm9lbCBHb3Jkb24g IDxub2VsLmdvcmRvbkBnbWFpbC5jb20+CisKKyAgICAgICAgUmV2aWV3ZWQgYnkgTk9CT0RZIChP T1BTISkuCisKKyAgICAgICAgW2Nocm9taXVtXSBSZWR1Y2UgY2FudmFzLnRvRGF0YVVSTCgiaW1h Z2UvanBlZyIpIHJ1bi10aW1lIGNvc3QgYnkgMTAlCisgICAgICAgIGh0dHBzOi8vYnVncy53ZWJr aXQub3JnL3Nob3dfYnVnLmNnaT9pZD01MDgwNAorCisgICAgICAgIEZvbGxvdyBvbiBmcm9tIHI3 MzE3MywgdW5yb2xsIHRoZSBTa1VuUHJlTXVsdGlwbHk6OlBNQ29sb3JUb0NvbG9yKCkgY2FsbAor ICAgICAgICB2aXouLCBjb21wdXRlIHRoZSB1bnByZW11bHRpcGxhdGNpb24gaW4tcGxhY2UuIFRo aXMgcmVkdWNlcyB0aGUgcnVuLXRpbWUKKyAgICAgICAgY29zdCBvZiBqcGVnIGVuY29kaW5nIGJ5 IDEwLTE1JSBmb3IgbXkgaW1hZ2UgdGVzdCBzZXQuCisKKyAgICAgICAgTm8gbmV3IHRlc3RzOiBj YW52YXMudG9EYXRhVVJMKCkgaXMgY292ZXJlZCBieSBleGlzdGluZyB0ZXN0cy4KKworICAgICAg ICAqIHBsYXRmb3JtL2ltYWdlLWVuY29kZXJzL3NraWEvSlBFR0ltYWdlRW5jb2Rlci5jcHA6Cisg ICAgICAgIChXZWJDb3JlOjpwcmVNdWx0aXBsaWVkQkdSQXRvUkdCKToKKwogMjAxMC0xMi0wOCAg U2FtIFdlaW5pZyAgPHNhbUB3ZWJraXQub3JnPgogCiAgICAgICAgIFJldmlld2VkIGJ5IERhcmlu IEFkbGVyLgpkaWZmIC0tZ2l0IGEvV2ViQ29yZS9wbGF0Zm9ybS9pbWFnZS1lbmNvZGVycy9za2lh L0pQRUdJbWFnZUVuY29kZXIuY3BwIGIvV2ViQ29yZS9wbGF0Zm9ybS9pbWFnZS1lbmNvZGVycy9z a2lhL0pQRUdJbWFnZUVuY29kZXIuY3BwCmluZGV4IGMwNDAxOWQuLmQ0ZTE0ODEgMTAwNjQ0Ci0t LSBhL1dlYkNvcmUvcGxhdGZvcm0vaW1hZ2UtZW5jb2RlcnMvc2tpYS9KUEVHSW1hZ2VFbmNvZGVy LmNwcAorKysgYi9XZWJDb3JlL3BsYXRmb3JtL2ltYWdlLWVuY29kZXJzL3NraWEvSlBFR0ltYWdl RW5jb2Rlci5jcHAKQEAgLTM0LDYgKzM0LDcgQEAKICNpbmNsdWRlICJJbnRTaXplLmgiCiAjaW5j bHVkZSAiU2tCaXRtYXAuaCIKICNpbmNsdWRlICJTa1VuUHJlTXVsdGlwbHkuaCIKKyNpbmNsdWRl ICJTa0NvbG9yUHJpdi5oIgogZXh0ZXJuICJDIiB7CiAjaW5jbHVkZSA8c3RkaW8uaD4gLy8ganBl Z2xpYi5oIG5lZWRzIHN0ZGlvLmggRklMRQogI2luY2x1ZGUgImpwZWdsaWIuaCIKQEAgLTgyLDEx ICs4MywxOSBAQCBzdGF0aWMgdm9pZCBoYW5kbGVFcnJvcihqX2NvbW1vbl9wdHIgY29tbW9uKQog Ly8gYmUgaWdub3JlZD8gU2VlIGJ1ZyBodHRwOi8vd2Via2l0Lm9yZy9iLzQwMTQ3Lgogdm9pZCBw cmVNdWx0aXBsaWVkQkdSQXRvUkdCKGNvbnN0IFNrUE1Db2xvciogaW5wdXQsIHVuc2lnbmVkIGlu dCBwaXhlbHMsIHVuc2lnbmVkIGNoYXIqIG91dHB1dCkKIHsKLSAgICB3aGlsZSAocGl4ZWxzLS0g PiAwKSB7Ci0gICAgICAgIFNrQ29sb3IgdW5tdWx0aXBsaWVkID0gU2tVblByZU11bHRpcGx5OjpQ TUNvbG9yVG9Db2xvcigqaW5wdXQrKyk7Ci0gICAgICAgICpvdXRwdXQrKyA9IFNrQ29sb3JHZXRS KHVubXVsdGlwbGllZCk7Ci0gICAgICAgICpvdXRwdXQrKyA9IFNrQ29sb3JHZXRHKHVubXVsdGlw bGllZCk7Ci0gICAgICAgICpvdXRwdXQrKyA9IFNrQ29sb3JHZXRCKHVubXVsdGlwbGllZCk7Cisg ICAgc3RhdGljIGNvbnN0IFNrVW5QcmVNdWx0aXBseTo6U2NhbGUqIHNjYWxlID0gU2tVblByZU11 bHRpcGx5OjpHZXRTY2FsZVRhYmxlKCk7CisKKyAgICBmb3IgKDsgcGl4ZWxzLS0gPiAwOyArK2lu cHV0KSB7CisgICAgICAgIGNvbnN0IHVuc2lnbmVkIGFscGhhID0gU2tHZXRQYWNrZWRBMzIoKmlu cHV0KTsKKyAgICAgICAgaWYgKChhbHBoYSAhPSAwKSAmJiAoYWxwaGEgIT0gMjU1KSkgeworICAg ICAgICAgICAgKm91dHB1dCsrID0gU2tVblByZU11bHRpcGx5OjpBcHBseVNjYWxlKHNjYWxlW2Fs cGhhXSwgU2tHZXRQYWNrZWRSMzIoKmlucHV0KSk7CisgICAgICAgICAgICAqb3V0cHV0KysgPSBT a1VuUHJlTXVsdGlwbHk6OkFwcGx5U2NhbGUoc2NhbGVbYWxwaGFdLCBTa0dldFBhY2tlZEczMigq aW5wdXQpKTsKKyAgICAgICAgICAgICpvdXRwdXQrKyA9IFNrVW5QcmVNdWx0aXBseTo6QXBwbHlT Y2FsZShzY2FsZVthbHBoYV0sIFNrR2V0UGFja2VkQjMyKCppbnB1dCkpOworICAgICAgICB9IGVs c2UgeworICAgICAgICAgICAgKm91dHB1dCsrID0gU2tHZXRQYWNrZWRSMzIoKmlucHV0KTsKKyAg ICAgICAgICAgICpvdXRwdXQrKyA9IFNrR2V0UGFja2VkRzMyKCppbnB1dCk7CisgICAgICAgICAg ICAqb3V0cHV0KysgPSBTa0dldFBhY2tlZEIzMigqaW5wdXQpOworICAgICAgICB9CiAgICAgfQog fQogCg==