08:25 <tomeu> austriancoder: thanks for pushing your img load/store RE!

08:26 <tomeu> something I have noticed is that right after the texture address, stride and dimensions, there is in the uniform area in the cmdstream other values that seem to be related to the sampler

08:31 <tomeu> eg. https://paste.debian.net/1268506/

08:32 <tomeu> there, we have the address, stride and dimensions at u1, but at u2, u3 and u4 we have stuff that seems to be related to how the pixels are fetched

09:47 <austriancoder> tomeu: but .. they are not used in the shader . I have seen these extra uniforms with img_load_3d and img_store_3d. I think that the binary blob could be smarter here and clean up some unused uniforms.

09:55 <tomeu> austriancoder: yeah, I have wondered if the silicon can keep reading the next uniforms even if they aren't explicitly referenced in the instruction encoding

09:56 <austriancoder> tomeu: never seen that

09:56 <tomeu> at some point it looked as if changing the contents of those extra uniform values changed how the sampling was done

09:56 <tomeu> but everything is quitefragile right now, so it could have been something else

09:56 <tomeu> austriancoder: did you find out what texture formats those instructions support?

09:57 <tomeu> specially, if any tiling modifier is supported/expected

10:09 <austriancoder> tomeu: not out of my head. sadly I have no access to my home lab from dayjob

10:10 <tomeu> oh, np, I can try to figure it out myself

10:10 <tomeu> just trying to avoid duplicating work

10:11 <austriancoder> tomeu: thanks .. feel free to RE it :)

10:19 <tomeu> btw, src0.w seems to be related to the format

10:20 <tomeu> at least, it changes based on the format of the texture

10:20 <tomeu> - 0x321011c2, /* SH.HALTI5_UNIFORMS[7] := 0x321011c2 */

10:20 <tomeu> + 0x301211c2, /* SH.HALTI5_UNIFORMS[7] := 0x301211c2 */

10:20 <tomeu> RGBA->BGRA

10:21 <austriancoder> great

10:21 <austriancoder> what are your plans regarding the open opencl MR?

10:22 <tomeu> italove has been busy with some panfrost stuff, but hopefully will be able to get back to it

10:22 <tomeu> personally, my main goal right now is to scope better all the features needed to run tensorflow models

10:24 <austriancoder> thanks

12:21 <tomeu> austriancoder: btw, this is the direction I'm going towards: https://gitlab.freedesktop.org/tomeu/etna_viv/-/commit/575dc6e8a5594c3477ebc312a7c2475b693a4871

12:26 <tomeu> austriancoder: btw, how does one update the rnndb headers in mesa? if I run rnndb/gen_headers.sh in my etna_viv checkout, I get a ton of unrelated changes

12:39 <austriancoder> tomeu: the rnndb change needs to land in the github repo .. gen_headers.sh and then copy most generated files to mesa tree .. create a commit like: https://cgit.freedesktop.org/mesa/mesa/commit/?id=b00fc1a08b1bc44f00af3aa2cbd63ec56673716d

14:56 <tomeu> it's working nicely :)

14:57 <tomeu> well, for coords (0,0), anything else samples zeroes

14:57 <tomeu> probably some sampler config missing

14:57 <tomeu> wonder fi it can be configured to interpret src1.xy as floats

15:01 <tomeu> or maybe the instruction reads in a tiling format and I'm feeding it in linear...