!FZyQrssSlHEZqrYcOb:matrix.org

wgpu

90 Members
Rust implementation of WebGPU5 Servers

Load older messages


Timestamp Message
27 Feb 2020
20:57:06@m4b:matrix.orgm4bIf your file above is compiled with x86-apple-ios then I’m not sure what your issue is
20:57:28@m4b:matrix.orgm4bYou are on Catalina yes ?
20:57:55@waha.06x36:matrix.orgwaha.06x36Yes.
20:58:02@waha.06x36:matrix.orgwaha.06x36I can simulate Metal just fine,.
20:58:28@waha.06x36:matrix.orgwaha.06x36And it's built with whatever cargo-lipo uses, which I think is x86_64-apple-ios?
20:58:36@m4b:matrix.orgm4bHmmm ... well you may be the first person ever running wgpu/rust code in simulator on Catalina :)
20:59:01@m4b:matrix.orgm4bI don’t use cargo lipo either, it has a lot of bugs. I just have makefile that calls lipo manually
20:59:32@waha.06x36:matrix.orgwaha.06x36Yeah, that's what I planned to do but I have a patched version and it seemed to the right thing.
20:59:45@m4b:matrix.orgm4bLet me dink around try to get a wgpu example running in simulator
21:05:16@waha.06x36:matrix.orgwaha.06x36 kvark: I tried using the memory graph tool, and I found a leak but it's on the CPU... sigh.
21:52:52@waha.06x36:matrix.orgwaha.06x36Hmm, this time I ran out of memory before even filling up my 100MB buffer and even starting to try to free things. Surely there can't be that little memory available?
21:53:52@waha.06x36:matrix.orgwaha.06x36I found the CaptureMTLBuffer objects that seem to correspond to buffers, and there aren't many more of them than I expect.
21:53:57@kvark:matrix.orgkvarkstrange!
21:55:18@waha.06x36:matrix.orgwaha.06x36 Very. Getting thread '<unnamed>' panicked at 'calledResult::unwrap()on anErrvalue: AllocationError(OutOfMemory(Device))', src/libcore/result.rs:1051:5
21:56:02@waha.06x36:matrix.orgwaha.06x36Some day I'll work on a project and only have regular, everyday problems instead of mystery ones.
22:10:53@waha.06x36:matrix.orgwaha.06x36Hmm... I use wgpu_device_create_buffer_mapped and unmap immediately, but, could I be somehow running out of virtual address space? But no, iOS is all 64 bits these days isn't it...
22:12:16@kvark:matrix.orgkvarkI use tons of these in vange-rs, and it's fine for that case. Although, they are relatively small there.
22:13:19@waha.06x36:matrix.orgwaha.06x36Mine are all a few tens of ks, and I have 2-3 thousand alive at any moment.
22:14:28@m4b:matrix.orgm4bYa I can repro your issue waha.06x36. It’s crashing in enumerate_adapters in gfx metal backend. In particular internal.rs:496
22:15:05@kvark:matrix.orgkvark wait, what does this have to do with enumerate_adapters? thought you are talking about OOM?
22:15:22@waha.06x36:matrix.orgwaha.06x36 kvark: This was the earlier issue about not working on simulator.
22:16:10@m4b:matrix.orgm4b Ya, about simulator. Sorry should have clarified was following up on my last comment
22:29:42@m4b:matrix.orgm4bSo I think you now need the x86_64-apple-ios-macabi target on catalina. But rust up currently does not ship the std for that target, so I think you’re out of luck for now (if I’m right about the new target being required)
23:07:13@waha.06x36:matrix.orgwaha.06x36 m4b: Do you have a way to test it?
23:08:57@m4b:matrix.orgm4bI’d have to build stdlib with the target and then rustup link it and I’m too lazy/busy. It’s a lot of tediousness. Why do you want to run in simulator ? It’s just going to use host gpu anyway. So you aren’t testing really how it performs / works on iOS real device.
23:18:19@m4b:matrix.orgm4b(Other than it being fun/more convenient for testing iOS ui control flow/application start logic)
23:20:13@waha.06x36:matrix.orgwaha.06x36Basically just convenience yeah. I have a macOS build I can use for most testing.
23:22:40@m4b:matrix.orgm4bYea that’s what I do. Every now and then load on iOS device to check everything is ok. Honestly wgpu is so cross platform and “just sort of works” (except for some corner cases in shaders) it’s been refreshing just developing on whatever machine is easiest and then checking in every now and then. It’s pretty bodacious :)
23:26:04@m4b:matrix.orgm4b I became effectively convinced of this after building on windows amd discrete and it just worked identically. Contrast this with Linux + macOS with Vulkan + moltenvk and it was awful. Every other moltenvk has some regression, or was busted on real iOS, was quite stressful.
23:27:26@m4b:matrix.orgm4bNot even to mention just getting the tool chain working, libraries building and linking properly for moltenvk was really unpleasant; all gone now.... ❤️

There are no newer messages yet.


Back to Room List