#asahi-gpu · 2024-02-11 21:09
i509vcbLLaMA models are probably not going to use OpenGL or Vulkan compute however#asahi-gpu · 2024-02-11 20:07
tertl8the gpu to cpu memory exchange makes a difference#asahi-gpu · 2024-02-11 20:07
tertl8those mac chips have really show to be great for the local llama stuff#asahi-gpu · 2024-02-11 20:06
tertl8nice#asahi-gpu · 2024-02-11 19:59
i509vcbVulkan is a whole other story since the rest of the driver needs to be implemented still, but I did get compute working in some of the dEQP tests there#asahi-gpu · 2024-02-11 19:58
i509vcbOpenCL no idea#asahi-gpu · 2024-02-11 19:58
i509vcbAnd GL_ARB_compute_shader is advertised on the OpenGL side#asahi-gpu · 2024-02-11 19:57
i509vcbI have ES 3.1 on M2 so if you just need GLES compute that would work fine#asahi-gpu · 2024-02-11 19:53
tertl8should i start saving for a mac studio 64GB?#asahi-gpu · 2024-02-11 19:53
tertl8how is the gpu compute coming guys?#asahi-gpu · 2024-02-07 19:27
kosherbombshttps://appealforassistance.notion.site/Appeal-for-Assistance-abcca31346e944b38e09ebabb4208152#asahi-gpu · 2024-02-06 21:31
alyssayeah, Lina is away due to IRL#asahi-gpu · 2024-02-06 20:46
j`eyalthough lina isnt around rn and she knows them the best..#asahi-gpu · 2024-02-06 20:44
svenyes#asahi-gpu · 2024-02-06 20:29
Lyudespecifically the DRM bindings#asahi-gpu · 2024-02-06 20:29
LyudeI assume this is the right place to ask if I have any questions around Asahi's rust bindings for the kernel?#asahi-gpu · 2024-01-21 01:15
i509vcbErr more of an imply#asahi-gpu · 2024-01-21 01:13
i509vcbVUID-VkCopyImageInfo2-dstImage-04444 seems to state that I should be using the depth and dstOffset for layerCount and baseArrayLayer in the 2D Array view of the destination 3d view#asahi-gpu · 2024-01-21 00:14
alyssain a regular fragment shader, the layer ID needs to be passed as a varying from the vertex shader#asahi-gpu · 2024-01-21 00:14
alyssai509vcb: layer_id_sr doesn't work in fragment shaders, only background&EOT#asahi-gpu · 2024-01-20 23:53
i509vcbFrom what I can tell, image views add the result of ail_get_layer_offset_B to the base address of the image to get a view that starts at some layer #asahi-gpu · 2024-01-20 23:46
i509vcbalthough it's all layer 0 writes still#asahi-gpu · 2024-01-20 23:46
i509vcbI do tell the fragment shader to always have a layered tib layout and lower with layer_id_sr#asahi-gpu · 2024-01-20 23:45
i509vcb4. This sounds like forcing 2D array for everything? I've tried this a few times but it ends in breaking everything#asahi-gpu · 2024-01-20 23:45
i509vcb2. Seems like vk-meta does do this and creates 2d views of the 3d image for single layer copies, although why these still fail is still in the air#asahi-gpu · 2024-01-20 20:27
alyssai509vcb: ^^#asahi-gpu · 2024-01-20 20:27
alyssa(This is not the only correct way to implement this. I don't remember what Metal does exactly.)#asahi-gpu · 2024-01-20 20:26
alyssabut key the EOT's tib layout to the specialized thing#asahi-gpu · 2024-01-20 20:26
alyssa4. current GL driver behaviour should be implementable for VK: force the fragment shader tib layout to always be layered (and use a 2D Array view even of a single layer if applicable for spilled render targets)#asahi-gpu · 2024-01-20 20:25
alyssa3. 2d arrays can be linear, hence the Depth (linear) fielld. 3D images cannot be linear. #asahi-gpu · 2024-01-20 20:24
alyssafor rendering to a single layer of an array, the GL driver uses a non-layered EOT shader with a non-array view (doing the required math to produce a 2d view of one layer of a 2d array. ail has a helper for this. see also: VK_EXT_image_2d_view_of_3d )#asahi-gpu · 2024-01-20 20:23
alyssa2. EOT shader can be keyed to layerCount > 1 for the render pass itself, but not for the underlying image#asahi-gpu · 2024-01-20 20:22
alyssa1. correct#asahi-gpu · 2024-01-20 19:25
i509vcb(Although the problem with forcing the tib layout to always be layered is that everything kind of breaks lol)#asahi-gpu · 2024-01-20 19:25
i509vcbbut assuming always layered means that the issue of Arrays not being allowed to be linear becomes a problem#asahi-gpu · 2024-01-20 19:24
i509vcbI did notice the EOT shader was never initialized as layered for some reason, I guess that checking if the layerCount > 1 is pretty much wrong#asahi-gpu · 2024-01-20 19:23
i509vcbIf the texture layout is linear, Depth is ignored? (And Depth(Linear) is used instead)?#asahi-gpu · 2024-01-17 15:04
alyssahttps://rosenzweig.io/a.txt#asahi-gpu · 2024-01-17 15:03
alyssaer wait#asahi-gpu · 2024-01-17 15:03
alyssanot on macos rn sorry#asahi-gpu · 2024-01-17 14:49
azsxdcz@alyssa: can you share your meson.build with me?#asahi-gpu · 2024-01-17 14:28
alyssanah, I've just hacked up the meson.build to work locally#asahi-gpu · 2024-01-17 14:16
azsxdcz@alyssa: Oh I understand... you don't have any makefile left by any chance to build wrap as a standalone binary outside of mesa?#asahi-gpu · 2024-01-17 12:54
alyssanever saw that film#asahi-gpu · 2024-01-17 12:53
TheLink3Don't go into the light!!#asahi-gpu · 2024-01-17 12:51
alyssahttps://www.youtube.com/watch?v=c8a3K0HIKdA#asahi-gpu · 2024-01-17 12:23
eric_engestromyeah, I was trying to think about it and it's all tangled indeed#asahi-gpu · 2024-01-17 12:22
alyssait's all bad#asahi-gpu · 2024-01-17 12:22
alyssaIDK#asahi-gpu · 2024-01-17 12:20
alyssaand the higher level NIR relevant stuff is... maybe a third folder, or maybe compilerized, or something I don't know if I knew I would've just done that git mv already