Luc Verhaegen (libv) wrote,
Luc Verhaegen
libv

  • Location:
  • Mood:
  • Music:

FOSDEM and early Tamil driver work.

FOSDEM 2015

It was another great FOSDEM this year. Even with their 5-10.000 attendants, the formula of being absolutely free, with limited sponsorship, and while only making small changes each year is an absolute winner. There is just no conference which comes even close to FOSDEM.

For those on ICE14 on Friday, the highspeed train from Frankfurt to Brussels south at 14:00, who were so nice to participate in my ad-hoc visitor count: 66. I counted 66 people, but i might have skipped a few as people were sometimes too engrossed in their laptops to hear me over their headphones. On a ~400 seat train, that's a pretty high number, and i never see the same level of geekiness on the Frankfurt to Brussels trains as on the Friday before FOSDEM. If it didn't sound like an organizational nightmare, it might have been a good idea to talk to DB and get a whole carriage reserved especially for FOSDEM goers.

With the Graphics DevRoom we returned to the K building this year, and i absolutely love the cozy 80 people classrooms we have there. With good airflow, freely movable tables, and an easy way to put in my powersockets, i have an easy time as a devroom organizer. While some speakers probably prefer bigger rooms to have a higher number of attendants, there is nothing like the more direct interaction of the rooms in the K buildings. With us sitting on the top floor, we also only had people who were very actively interested in the topics of the graphics devroom.

Despite the fact that FOSDEM has no equal, and anyone who does anything with open source software in the European Union should attend, i always have a very difficult time recruiting a full set of speakers for the graphics DevRoom. Perhaps the biggest reason for this is the fact that it is a free conference, and it lacks the elitarian status of a paid-for conference. Everyone can attend your talk, even people who do not work on the kernel or on graphics drivers, and the potential speaker might feel as if he needs to waste his time on people who are below his own perceived station. Another reason may be that it is harder to convince the beancounters to sponsor a visit to a free conference. In that case, if you live in the European Union and when you are paid to do open source software, you should be able to afford going to the must-visit FOSDEM by yourself.

As for next year, i am not sure whether there will be a graphics devroom again. Speaker count really was too low and perhaps it is time for another hiatus. Perhaps it is once again time to show people that talking in a devroom at FOSDEM truly is a privilege, and not to be taken for granted.

Tamil "Driver" talk.

My talk this year, or rather, my incoherent mumble finished off with a demo, was about showing my work on the ARM Mali Midgard GPUs. For those who had to endure it, my apologies for my ill-preparedness; i poured all my efforts into the demo (which was finished on Wednesday), and spent too much time doing devroom stuff (which ate Thursday) and of course in drinking up, ahem, the event that is FOSDEM (which ate the rest of the weekend). I will try to make up for it now in this blog post.

Current Tamil Status.

As some might remember, in September and October 2013, i did some preliminary work on the Mali T-series. I spent about 3 to 3.5 weeks building the infrastructure to capture the command stream and replay it. At the same time I also dug deep into the Mali binary driver to expose the binary shader compiler. These two feats gave me all the prerequisites for doing the job of reverse engineering the command stream of the Mali Midgard.

During the Christmas holidays of 2014 and in January 2015, i spent my time building a command stream parser. This is a huge improvement over my work on lima, where i only ended doing so later on in the process (while bringing up Q3A). As i built up the capabilities of this parser, i threw ever more complex GLES2 programs at it. A week before FOSDEM, my parser was correctly handling multiple draws and frames, uniforms, attributes, varyings and textures. Instead of having raw blobs of memory, i had C structs and tables, allowing me to easily see the differences between streams.

I then took the parsed result of my most complex test and slowly turned that into actual C code, using the shader binary produced by the binary compiler, and adding a trivial sequential memory allocator. I then added rotation into the mix, and this is the demo as seen on FOSDEM (and now uploaded to youtube).

All the big things are known.

For textures. I only have simple texture support at this time, no mipmapping nor cubemapping yet, and only RGB565 and RGBA32 are supported at this time. I also still have not figured out how to disable swizzling, instead i re-use the texture swizzling code from lima, the only place where I was able to re-use code in tamil. This missing knowledge is just some busywork, and a bit of coding away.

As for programs, while both the Mali Utgard (M-series) and Midgard (T-series) binary compilers output in a format called MBS (Mali Binary Shader), the contents of each file is significantly different. I had no option but to rewrite the MBS parser for tamil.

Instead of rewriting the vertex shaders binaries like ARMs binary driver does, i reorder the entries in the attribute descriptor table to match the order as described by the shader compiler output. This avoids adding a whole lot of logic to handle this correctly, even though MBS now describes which bits to alter in the binary. I still lay uniforms, attributes and varyings out by hand though, i similarly have only limited knowledge of typing at this point. This mostly is a bit of busywork of writing up the actual logic, and trying out a few different things.

I know only very few things about most of the GL state. Again, mostly busywork with a bit of testing and coding up the results. And while many values left and right are still magic, nothing big is hiding any more.

Unlike lima, i am refraining from building up more infrastructure (than necessary to show the demo) outside of Mesa. The next step really is writing up a mesa driver. Since my lima driver for mesa was already pretty advanced, i should be able to re-use a lot of the knowledge gained there, and perhaps some code.

The demo

The demo was shown on a Samsung ARM Chromebook, with a kernel and linux installation from september 2013 (when i brought up cs capture and exposed the shader compiler). The exynos KMS driver on this 3.4.0 kernel is terrible. It only accepts a few fixed resolutions (as if I never existed and modesetting wasn't a thing), and then panics when you even look at it. Try to leave X while using HDMI: panic. Try to use a KMS plane to display the resulting render: panic.

In the end, i used fbdev and memcpy the rendered frame over to the console. On this kernel, i cannot even disable the console, so some of the visible flashing is the console either being overwritten by or overwriting the copied render.

The youtube video shows a capture of the Chromebooks built in LCD, at 1280x720 on a 1366x768 display. At FOSDEM, i was lucky that the projector accepted the 1280x720 mode the exynos hdmi driver produced. My dumb HDMI->VGA converter (which makes the image darker) was willing to pass this through directly. I have a more intelligent HDMI->VGA adapter which also does scaling and which keeps colours nice, but that one just refused the output of the exynos driver. The video that was captured in our devroom probably does not show the demo correctly, as that should've been at 1024x768.

The demo shows 3 cubes rotating in front of the milky way. It is 4 different draws, using 3 different textures, and 3 different programs. These cubes currently rotate at 47fps, with the memcpy. During the talk, the chromebook slowed down progressively down to 26fps and even 21fps at one point, but i have not seen that behaviour before or since. I do know of an issue that makes the demo fail at frame 79530, which is 100% reproducible. I still need to track this down, it probably is an issue with my job handling code.

linux-exynos.org

With Lima and Tamil i am in a very unique position. Unlike on adreno, tegra or Videocore, i have to deal with many different SoCs. Apart from the difference in kernel GPU drivers, i also have to deal with differences in display drivers and run into a whole new world of hurts every time i move over to a new target device. The information for doing a proper linux installation on an android or chrome device is usually dispersed, not up to date, and not too good, and i get to do a lot of the legwork for myself every time, knowing full well that a lot of others have done so already but couldn't be bothered to document things (hence my role in the linux-sunxi community).

The ARM chromebook and its broken kms driver is much of the same. Last FOSDEM i complained how badly supported and documented the Samsung ARM chromebook is, despite its popularity, and appealed for more linux-sunxi style, SoC specific communities, especially since I, as a graphics driver developer, cannot go and spend as much time in each and every of the SoC projects as i have done with sunxi.

During the questions round of this talk, one guy in the audience asked what needed to be done to fix the SoC pain. At first i completely missed the question, upon which he went and rephrased his question. My answer was: provide the infrastructure, make some honest noise and people will come. Usually, when some asks such a question, nothing ever comes from it. But Merlijn "Wizzup" Wajer and his buddy S.J.R. "Swabbles" van Schaik really came through.

Today there is the linux-exynos.org wiki, the linux-exynos mailinglist, and the #linux-exynos irc channel. While the community is not as large as linux-sunxi, it is steadily growing. So if you own exynos based hardware, or if your company is basing a product on the exynos chipset, head to linux-exynos.org and help these guys out. Linux-exynos deserves your support.
Tags: allwinner, arm, exynos, fosdem, lima, mali, opengl, reverse engineering, samsung, soc, sunxi, tamil
Subscribe
  • Post a new comment

    Error

    default userpic

    Your reply will be screened

    When you submit the form an invisible reCAPTCHA check will be performed.
    You must follow the Privacy Policy and Google Terms of use.
  • 1 comment