16:12 <jekstrand> I love how long a full panvk CTS run takes...

16:32 <jekstrand> I'm gonna land some panvk patches... \o/

16:36 <HdkR> \o/

16:36 <jekstrand> They're all 1-2 liners but it's a start.

16:36 <jekstrand> And compute should land soonish

16:54 <jekstrand> bbrezillon: Is there anything especially hard about texture buffers on panfrost? They should be mostly like 1D textures, right?

16:54 <jekstrand> I might hook up VkBufferView next.

16:58 <jekstrand> Also, I should look at secondaries again. Sorry.

17:00 <jekstrand> Also, index buffer support would be good to hook up at some point. I've seen a lot of crashes around that.

17:00 <jekstrand> In case anyone randomly wants to write panvk patches. :)

17:02 <HdkR> I look forward to all the features getting written in panvk so I can run Half-Life: Alyx on it :)

17:02 <jekstrand> lol

17:03 <jekstrand> That's gonna be a great VR experience. Make sure you have a bucket handy.

17:04 <HdkR> hehe

17:06 <jekstrand> I should really figure out some good exclude lists so we can cut down these CTS runtimes. 37 hours is nuts

17:11 <HdkR> Those are some spicy CTS runs

17:23 <jekstrand> Lots of crashes means lots of restarting dEQP. It's a bit of a pig.

18:00 <bbrezillon> jekstrand: yep, buffer == 1D texture

18:04 <bbrezillon> jekstrand: don't know which of the 'tmp' commit implements it, but it's there https://gitlab.freedesktop.org/bbrezillon/mesa/-/blob/panvk/src/panfrost/vulkan/panvk_vX_cmd_buffer.c#L1040 :-)

18:08 <jekstrand> bbrezillon: Oh, ok. I'll maybe try to find that one and land it then.

18:12 <bbrezillon> jekstrand: Re: sec-cmdbuf, that'd be great!

18:14 <jekstrand> bbrezillon: Going to try to finish reviewing compute. Then read the IMG uAPI and then secondaries. That's my list for today.

19:02 <jekstrand> bbrezillon: Why is ATTRIBUTE_BUFFER_CONTINUATION_3D not part of ATTRIBUTE_BUFFER? Are there cases where you only need ATTRIBUTE_BUFFER and you only need the second one when you're dealing with more complicated stuff?

19:03 <bbrezillon> yep, some descriptors require 2 slots

19:03 <jekstrand> Kinda neat, actually.

19:03 <jekstrand> Intel could learn a thing or two from that.

19:19 <jekstrand> bbrezillon: What's up with the separation between emit_compute_job and CmdDispatch?

19:19 <jekstrand> bbrezillon: At first brush, it doesn't seem like that's really saving anyone anything to copy things into a temp struct and call a thing in a different file.

19:19 <jekstrand> I mean, it works, but it seems unnecessary

19:21 <jekstrand> I guess the prepare -> dispatch.foo = pattern is kind of nice

19:43 <icecream95> jekstrand: About texture buffers, you *did* notice my MR changing them to be 2D textures, didn't you?

19:44 <jekstrand> icecream95: Nope

19:44 <jekstrand> icecream95: I'm not following especially closely.

19:44 <jekstrand> icecream95: I did write that MR for Intel once, though. (-:

19:44 <jekstrand> icecream95: Did you make them Nx1 2D textures or "real" 2D textures?

19:45 <icecream95> nir_unpack_32_2x16

19:45 <icecream95> (to split the coordinates to make them 2D)

19:46 <jekstrand> icecream95: Why? To increase the texture size limit?

19:46 <icecream95> Yup

19:46 <jekstrand> Makes sense. How did you handle bounds checking?

19:47 <icecream95> Loading the texture size and doing umin..

19:47 <jekstrand> But the texture size is now 2D

19:47 <icecream95> I load the texture size while it is still 1D :)

19:47 <jekstrand> Oh

19:47 <jekstrand> this sounds like some fun trickery

19:48 <jekstrand> Is the dimensionality of the texture part of the shader then? On Intel, it's part of the descriptor.

19:49 <icecream95> It's set in the shader.. which caused a problem with Nine as it tried using a 1D texture instruction with a 2D texture, which causes GPU faults

19:51 <jekstrand> Sure

19:51 <jekstrand> Trade-offs

19:51 <jekstrand> But it does mean tricks like that are possible which is really neat

19:54 <icecream95> On the topic of tricks, I'm still trying to find a use for the 3 bits-per-pixel (no alignment!) texture support

20:16 <jekstrand> Oh, that's wonderful. By which I mean I'm going to try to forget you ever said that!

21:03 <icecream95> kernel hacks were successful, I can now read the CSF registers from userspace..