228298 2021-07-26 13:52:45 -0700 powerPreference may not be working correctly in WebGL's Metal backend 2021-10-19 05:32:52 -0700 1 1 1 Unclassified WebKit WebGL WebKit Local Build Unspecified Unspecified RESOLVED FIXED https://bugs.webkit.org/show_bug.cgi?id=228698 https://bugs.webkit.org/show_bug.cgi?id=231012 P2 Normal --- 220843 227408 228576 1 kbr webkit-unassigned dino kbr kkinnunen kpiddington marcin.ignac oldest_to_newest 1779550 0 kbr 2021-07-26 13:52:45 -0700 Marcin Ignac indicates on Bug 220843 that in the current Safari Technology Preview (Release 128 (Safari 15.0, WebKit 16612.1.22.11.3)), on a dual-GPU system running macOS 11.5, the powerPreference implementation may not be working correctly. Attached is a test case. Switch which powerPreference line is picked in order to see the difference. The system does still perform a graphics mux (legacy behavior, from OpenGL) when "powerPreference: 'high-performance'" is chosen. The mux shouldn't be necessary with ANGLE's Metal backend. However, WebKit's unmasked renderer string doesn't itself provide enough information to know whether the discrete GPU is actually used on a dual-GPU system. 1779552 1 434233 kbr 2021-07-26 13:54:00 -0700 Created attachment 434233 Test case Simple test case allowing selection of low-power/high-performance GPUs 1779652 2 kpiddington 2021-07-26 17:36:55 -0700 Checked this by hand on a MBP, I'm seeing the correct GPU get selected based off of the LowPowerMode / highPowerMode flag. Should we respond to the unmasked renderer / GPU strings at all? 1779714 3 marcin.ignac 2021-07-27 02:55:14 -0700 I did manage to make it use high end gpu on my old laptop but can't make it work on MBP16. After upgrading macOS and Safari on my old laptop performance degraded. How to replicate: 1. Open http://projects.variable.io/webgl/shader-performance/ (noise based particle animation) 2. Open Activity Monitor and then GPU History MacBook Pro (16-inch, 2019), macOS BigSur 11.5 - Safari 14.1.2 -> Intel UHD Graphics 630, 10fps unresponsive UI - Chrome 92 -> AMD Radeon Pro 5600M, 60fps - Firefox 90 -> AMD Radeon Pro 5600M, 60fps Things like being plugged in and Automatic Graphic Switching on/off don't seem to make a difference. Also when i have Chrome using Radeon PRo 5600M and open Safari with the link above then Intel UHD kicks in anyway MacBook Pro (15-inch 2017), macOS BigSur 11.0.1 - Safari 14.0.1 → AMD Radeon Pro 560, 60fps - Chrome 92 -> AMD Radeon Pro 560, 60fps After upgrading to macOS BigSur 11.5.1 MacBook Pro (15-inch 2017), macOS BigSur 11.5 - Safari 14.1.2 → Intel HD GRaphics 630, 20fps, unresponsive UI - Chrome 92 -> AMD Radeon Pro 560, 60fps 1779849 4 kbr 2021-07-27 15:06:01 -0700 Let me reopen this because there's at least one bug where "powerPreference: high-performance" is still causing an OpenGL-based mux to the high-performance GPU which is unnecessary in the Metal backend. Marcin, can you please test with Safari Technology Preview (Safari 15.0) and tell us your results? There might be bugs in Safari 14 and selection of the high-performance GPU via WebGL, but they should be fixed in Safari 15 and Safari Technology Preview. 1779858 5 marcin.ignac 2021-07-27 15:30:02 -0700 Safari TP behaves the same: MacBook Pro (16-inch, 2019), macOS BigSur 11.5 Safari TP 128 (Safari 15.0, WebKit 16612.1.22.11.3) -> Intel UHD Graphics 630, 20-30fps, feels like 5fps, unresponsive UI 1779884 6 kbr 2021-07-27 16:31:35 -0700 I don't think I can reproduce this on my 2017 15" MacBook Pro. Could this issue be hardware dependent? macOS 11.5 Chrome 94.0.4588.0 (Official Build) canary (x86_64) -> 60 FPS STP Release 128 (Safari 15.0, WebKit 16612.1.22.11.3) -> 60 FPS Kyle, do you have a 16" MBP to test on? Can you reproduce this if so? 1780009 7 marcin.ignac 2021-07-28 03:24:40 -0700 Sorry guys. I can confirm that Safari TP Release 128 (Safari 15.0, WebKit 16612.1.22.11.3) fixes the issue on both MBP15 with AMD Radeon Pro 560, and MBP16 with AMD Radeon Pro 5600M. The reason why it was running at 10fps on MBP16 was that i had "GPU Process: WebGL" enabled. Resetting all the flags gives me now big gpu and 60fps on both computers when using Safari TP. 1780266 8 dino 2021-07-28 17:05:22 -0700 Yep! This was broken and then fixed recently. https://bugs.webkit.org/show_bug.cgi?id=227408 1780269 9 kbr 2021-07-28 17:09:08 -0700 Thanks Marcin for confirming. Have filed Bug 228576 about removing the unnecessary OpenGL-based GPU mux when using the Metal backend. 1781205 10 kkinnunen 2021-08-02 01:05:01 -0700 (In reply to Marcin Ignac from comment #7) > The reason why it was running at 10fps on MBP16 was that i had "GPU Process: > WebGL" enabled. Resetting all the flags gives me now big gpu and 60fps on > both computers when using Safari TP. Marcin, if you have a test case for this, feel free to link to it. I can file a bug about that. Thanks. 1781218 11 marcin.ignac 2021-08-02 03:40:53 -0700 (In reply to Kimmo Kinnunen from comment #10) > Marcin, if you have a test case for this, feel free to link to it. I can > file a bug about that. Thanks. It's the same as above http://projects.variable.io/webgl/shader-performance/ What does "GPU Process: WebGL" do under the hood that might have such a big impact on performance? 1781223 12 kkinnunen 2021-08-02 04:35:21 -0700 (In reply to Marcin Ignac from comment #11) > (In reply to Kimmo Kinnunen from comment #10) > > Marcin, if you have a test case for this, feel free to link to it. I can > > file a bug about that. Thanks. > > It's the same as above http://projects.variable.io/webgl/shader-performance/ Thanks, missed it. > What does "GPU Process: WebGL" do under the hood that might have such a big > impact on performance? It transfers the calls from web content process to GPU process, executes those in GPU process and then transfers the results back to web content process. 434233 2021-07-26 13:54:00 -0700 2021-07-26 13:54:00 -0700 Test case which-gpu.html text/html 912 kbr PGJvZHkgb25sb2FkPSJydW4oKSIKPGRpdiBpZD0ib3V0cHV0Ij4KPC9kaXY+CjxjYW52YXMgaWQ9 ImMiIHdpZHRoPSIxMjgiIGhlaWdodD0iMTI4Ij4KPC9jYW52YXM+CjxzY3JpcHQ+CmZ1bmN0aW9u IG91dHB1dChzdHIpIHsKICAgIGRvY3VtZW50LmdldEVsZW1lbnRCeUlkKCJvdXRwdXQiKS5pbm5l ckhUTUwgKz0gIjxicj4iICsgc3RyOwogICAgY29uc29sZS5sb2coc3RyKTsKfQoKZnVuY3Rpb24g cnVuKCkgewogICAgZ2wgPSBkb2N1bWVudC5nZXRFbGVtZW50QnlJZCgiYyIpLmdldENvbnRleHQo IndlYmdsIiwgewogICAgICAgIC8vIFN3aXRjaCBiZXR3ZWVuIHRoZXNlIHRvIHNlZSB0aGUgZGlm ZmVyZW5jZSBvbiBhIGR1YWwtR1BVIE1hY0Jvb2sgUHJvCiAgICAgICAgLy8gcG93ZXJQcmVmZXJl bmNlOiAiaGlnaC1wZXJmb3JtYW5jZSIKICAgICAgICBwb3dlclByZWZlcmVuY2U6ICJsb3ctcG93 ZXIiCiAgICB9KTsKICAgIG91dHB1dCgidmVuZG9yOiAiICsgZ2wuZ2V0UGFyYW1ldGVyKGdsLlZF TkRPUikpOwogICAgb3V0cHV0KCJyZW5kZXJlcjogIiArIGdsLmdldFBhcmFtZXRlcihnbC5SRU5E RVJFUikpOwogICAgZXh0ID0gZ2wuZ2V0RXh0ZW5zaW9uKCJXRUJHTF9kZWJ1Z19yZW5kZXJlcl9p bmZvIik7CiAgICBpZiAoIWV4dCkgewogICAgICAgIG91dHB1dCgiKG5vIFdFQkdMX2RlYnVnX3Jl bmRlcmVyX2luZm8gZXh0ZW5zaW9uKSIpOwogICAgICAgIHJldHVybjsKICAgIH0KICAgIG91dHB1 dCgidW5tYXNrZWQgdmVuZG9yOiAiICsgZ2wuZ2V0UGFyYW1ldGVyKGV4dC5VTk1BU0tFRF9WRU5E T1JfV0VCR0wpKTsKICAgIG91dHB1dCgidW5tYXNrZWQgcmVuZGVyZXI6ICIgKyBnbC5nZXRQYXJh bWV0ZXIoZXh0LlVOTUFTS0VEX1JFTkRFUkVSX1dFQkdMKSk7Cn0KPC9zY3JpcHQ+CjwvYm9keT4K