| 3 Apr 2024 |
 vero | calling for reviews on https://github.com/gfx-rs/wgpu/pull/5383 pls :) | 20:39:28 |
| vero | ive fixed everything and its been quite a journey | 20:39:47 |
| 5 Apr 2024 |
| vero | Broke out a couple small PRs for easier review: https://github.com/gfx-rs/wgpu/pull/5498 https://github.com/gfx-rs/wgpu/pull/5497 | 05:21:45 |
| 7 Apr 2024 |
|  mehmetoguzderin changed their profile picture. | 10:12:06 |
| vero | teoxoy is there anything i can do to make the 64 bit atomics PR easier to review? maybe split it into two PRs, one for ALL_OPS and one for MIN_MAX? | 20:07:04 |
| vero | should I just leave it alone | 20:07:21 |
| 8 Apr 2024 |
 Daniel McNab | What are the soundness implications of zero_initialize_workgroup_memory (and equivalents)? We're investigating this in Vello, and finding that it has a really bad impact on pipeline initialisation time (multiplication by a factor of 5 on Android, factor of 8 on DXC, factor of 15 on FXC)
The easiest solution from our perspective is to wire through disabling it, but I'm concerned that this needs documentation | 09:55:13 |
| Daniel McNab | Exact numbers can be found in:
https://xi.zulipchat.com/#narrow/stream/197075-gpu/topic/Android.20Startup.20Time.20Investigation/near/431913809
(I don't know if Matrix supports tables) | 09:55:45 |
| 9 Apr 2024 |
 teoxoy | In reply to @atlv:matrix.org
teoxoy is there anything i can do to make the 64 bit atomics PR easier to review? maybe split it into two PRs, one for ALL_OPS and one for MIN_MAX? Should be fine, I should be able to get to it in the coming days. | 14:11:09 |
| Daniel McNab | Wow is spir-v convoluted to work with... | 16:23:17 |
 cwfitzgerald | Lazy CI seems to be failing after pipeline constants landed | 21:00:30 |
| cwfitzgerald | The dota2 shaders use spec constants and the spirv frontend seems to be balking | 21:00:57 |
| 10 Apr 2024 |
| teoxoy | In reply to @cwfitzgerald:matrix.org
The dota2 shaders use spec constants and the spirv frontend seems to be balking I'll take a look. | 15:15:48 |
| 11 Apr 2024 |
| mehmetoguzderin changed their profile picture. | 11:09:50 |
| 12 Apr 2024 |
|  Douglas Dwyer joined the room. | 23:43:12 |
| Douglas Dwyer | Are there any plans to support the shader-f16 extension in Naga in the future? If no one is working on it, what steps would be necessary to implement the feature myself for the WGSL frontend and SPIR-V backend? | 23:50:04 |
| Douglas Dwyer | Ah, just found the meta-issue: https://github.com/gfx-rs/wgpu/issues/4384
I still would be interested in knowing more about the steps, specifically wgsl-in and spv-out | 23:58:21 |
| 13 Apr 2024 |
| vero | In reply to @douglasdwyer:matrix.org
Ah, just found the meta-issue: https://github.com/gfx-rs/wgpu/issues/4384
I still would be interested in knowing more about the steps, specifically wgsl-in and spv-out should be pretty similar to what I did in https://github.com/gfx-rs/wgpu/pull/5154 | 00:30:02 |
| vero | i recommend just searching for some existing type like f32 and going through all the usages in the codebase naively making sure each one also handles f16 | 00:31:03 |
| vero | then adding an f16 naga test and iterating until it works on all the backends you can | 00:31:52 |
| Douglas Dwyer | In reply to @atlv:matrix.org
should be pretty similar to what I did in https://github.com/gfx-rs/wgpu/pull/5154 Thanks, that looks like a great example PR to follow :) | 02:08:14 |
| 14 Apr 2024 |
| vero | i opened 3 prs that should be easy reviews and merges:
https://github.com/gfx-rs/wgpu/pull/5532
https://github.com/gfx-rs/wgpu/pull/5533
https://github.com/gfx-rs/wgpu/pull/5534 | 19:58:44 |
| 19 Apr 2024 |
|  @multisn8:matrix.org left the room. | 15:29:43 |
| 20 Apr 2024 |
| vero | I need help with this: https://github.com/gfx-rs/wgpu/issues/5569 | 11:27:35 |
 imbris | I'm getting HLSL output like:
float4 _expr2703 = float4(_expr2697.x, (float4)0.y, (float4)0.z, (float4)0.w);
which Fxc does seem to like: error X3000: syntax error: unexpected token 'y'
(float4)0 is written for an Expression::ZeroValue. Would the correct fix be to adjust the hlsl writer to parenthesize these expressions?
float4 _expr2703 = float4(_expr2697.x, ((float4)0).y, ((float4)0).z, ((float4)0).w);
This is using some SPIRV input that I haven't looked through, I tried to reproduce the issue with some simple wgsl like vec2<f32>().x but naga outputs (float)0 for that.
| 16:49:32 |
| imbris | * I'm getting HLSL output like:
float4 _expr2703 = float4(_expr2697.x, (float4)0.y, (float4)0.z, (float4)0.w);
which Fxc doesn't seem to like: error X3000: syntax error: unexpected token 'y'
(float4)0 is written for an Expression::ZeroValue. Would the correct fix be to adjust the hlsl writer to parenthesize these expressions?
float4 _expr2703 = float4(_expr2697.x, ((float4)0).y, ((float4)0).z, ((float4)0).w);
This is using some SPIRV input that I haven't looked through, I tried to reproduce the issue with some simple wgsl like vec2<f32>().x but naga outputs (float)0 for that.
| 16:50:58 |
| 21 Apr 2024 |
| cwfitzgerald | In reply to @imbris:matrix.org
I'm getting HLSL output like:
float4 _expr2703 = float4(_expr2697.x, (float4)0.y, (float4)0.z, (float4)0.w);
which Fxc doesn't seem to like: error X3000: syntax error: unexpected token 'y'
(float4)0 is written for an Expression::ZeroValue. Would the correct fix be to adjust the hlsl writer to parenthesize these expressions?
float4 _expr2703 = float4(_expr2697.x, ((float4)0).y, ((float4)0).z, ((float4)0).w);
This is using some SPIRV input that I haven't looked through, I tried to reproduce the issue with some simple wgsl like vec2<f32>().x but naga outputs (float)0 for that.
Yeah that looks like what needs to happen, though the construction is a bit comical | 20:05:20 |
| imbris | yeah 😂 | 20:38:58 |
| 23 Apr 2024 |
| vero | fix for https://github.com/gfx-rs/wgpu/issues/5569 :
https://github.com/gfx-rs/wgpu/pull/5590 | 17:33:27 |
| 25 Apr 2024 |
| imbris | I'm trying to figure out the approach for fixing https://github.com/gfx-rs/wgpu/issues/4514.
The linked spirv-cross issue seems to have been resolved by replacing the switch with a do {} while(false). However, I'm wondering if there could be issues with the original switch containing a continue statement for an outer loop. Is this possible? | 04:18:25 |
