120813 2013-09-05 17:07:50 -0700 Use of Uint8ClampedArray makes filters twice as slow 2022-12-30 16:39:31 -0800 1 1 1 Unclassified WebKit Layout and Rendering 528+ (Nightly build) Unspecified Unspecified NEW InRadar P2 Normal --- 1 simon.fraser webkit-unassigned ahmad.saleem792 krit oliver savagobr simon.fraser thorton webkit-bug-importer oldest_to_newest 925294 0 simon.fraser 2013-09-05 17:07:50 -0700 I was curious about why blur filters on image backgrounds are incredibly slow. Sampling shows lots of time in GenericTypedArrayView::set(unsigned index, double value) const which is doing a safe double -> int cast with clamping. This comes out of a tight look in the blurring code: for (int x = 0; x < effectWidth; ++x) { int pixelByteOffset = line + x * stride + channel; dstPixelArray->set(pixelByteOffset, static_cast<unsigned char>(sum / dx)); It's crazy that we're wasting time doing double -> int conversion when the caller is passing in an unsigned char. We're converting double -> unsigned char -> double -> int32_t -> uint8_t Bypassing this crap by just filling a uint8_t[] and using dstPixelArray->setRange() makes blurring twice as fast. 925295 1 210684 simon.fraser 2013-09-05 17:08:10 -0700 Created attachment 210684 Testcase 925724 2 webkit-bug-importer 2013-09-06 11:08:52 -0700 <rdar://problem/14930799> 985658 3 savagobr 2014-02-28 11:34:49 -0800 Started looking into this. 1922336 4 ahmad.saleem792 2022-12-30 16:39:31 -0800 It is still taking several steps to blur completely in Safari 16.2 & STP160 but this does not seem to be working in Chrome Canary 111 and Firefox Nightly 110 for some reasons even after modifying to change pic URL to HTTPS and removing -webkit- prefix in following: URL - https://jsfiddle.net/unz3h758/2/show 210684 2013-09-05 17:08:10 -0700 2013-09-05 17:08:10 -0700 Testcase background-filters.html text/html 872 simon.fraser PCFET0NUWVBFIGh0bWw+Cgo8aHRtbD4KPGhlYWQ+CiAgICA8c3R5bGU+CiAgICAgICAgLmJveCB7 CiAgICAgICAgICB3aWR0aDogNDI3cHg7CiAgICAgICAgICBoZWlnaHQ6IDY0MHB4OwogICAgICAg ICAgYmFja2dyb3VuZC1pbWFnZTogLXdlYmtpdC1maWx0ZXIodXJsKGh0dHA6Ly9mYXJtNi5zdGF0 aWNmbGlja3IuY29tLzU1NDIvOTY2MjU5Njk4NF8yZTRjMjNjMDQxX3ouanBnKSwgYmx1cigwcHgp KTsKICAgICAgICAgIGJhY2tncm91bmQtc2l6ZTogMTAwJSAxMDAlOwogICAgICAgICAgdHJhbnNp dGlvbjogYmFja2dyb3VuZC1pbWFnZSA0czsKICAgICAgICB9CiAgICAgICAgCiAgICAgICAgLmJv eC5jaGFuZ2VkLCAuYm94OmhvdmVyIHsKICAgICAgICAgICAgYmFja2dyb3VuZC1pbWFnZTogLXdl YmtpdC1maWx0ZXIodXJsKGh0dHA6Ly9mYXJtNi5zdGF0aWNmbGlja3IuY29tLzU1NDIvOTY2MjU5 Njk4NF8yZTRjMjNjMDQxX3ouanBnKSwgYmx1cigxMHB4KSk7CiAgICAgICAgfQogICAgPC9zdHls ZT4KICAgIDxzY3JpcHQ+CiAgICAgICAgCiAgICAgICAgZnVuY3Rpb24gZG9UZXN0KCkKICAgICAg ICB7CiAgICAgICAgICAgIHdpbmRvdy5zZXRUaW1lb3V0KGZ1bmN0aW9uKCkgewogICAgICAgICAg ICAgICAgZG9jdW1lbnQuZ2V0RWxlbWVudEJ5SWQoJ3RhcmdldCcpLmNsYXNzTGlzdC5hZGQoJ2No YW5nZWQnKTsKICAgICAgICAgICAgfSwgMTAwMCk7CiAgICAgICAgfQogICAgICAgIHdpbmRvdy5h ZGRFdmVudExpc3RlbmVyKCdsb2FkJywgZG9UZXN0LCBmYWxzZSk7CiAgICA8L3NjcmlwdD4KPC9o ZWFkPgo8Ym9keT4KCiAgICA8ZGl2IGlkPSJ0YXJnZXQiIGNsYXNzPSJib3giPgogICAgPC9kaXY+ CjwvYm9keT4KPC9odG1sPgo=