Hardware acceleration with Hastlayer - consultant wanted!

[Piedone]

Hi everyone,

we at Lombiq run a pretty unique project called Hastlayer: Hastlayer can automatically transform compatible .NET code into a hardware implementation, making the result much faster and consume much less power (along the lines of 5-10x in suitable cases). We'd like to implement a Hastlayer-accelerated ImageSharp PoC and are looking for a developer experienced with the internals of ImageSharp to help us with it. Would you like to work with us? (And I mean paid consulting of course.)

The PoC, if successful, would serve as a demonstration in web apps (like Orchard Core apps, something that we're actually focused on as a company), as well as advanced scenarios like running image processing tasks for space industry use-cases, including ground segment and on board of a satellite (yes, we actually want to put Hastlayer on a satellite, and it's not as far-fetched as you may think).

Hastlayer is not a silver bullet and can only help if the algorithm in question is massively parallelizable (using hundreds of Tasks) and compute or memory-bound, see details here. Also, the code needs to be compatible with what Hastlayer supports, which currently is elaborated here. There is a simple but well-performing image processing sample already.

BTW all of this happens with FPGAs which are reconfigurable chips that can mimic other logic hardware. You can think of them as kind of an accelerator card in this use-case. So no need to actually manufacture the hardware generated by Hastlayer though that would also be a possibility :). FPGAs are available in the cloud too, like in AWS and Azure.

What do you think? If you're interested, please let me know here or send me an e-mail to my address under my profile.

[antonfirsov]

Hey @Piedone nice to see you here again! :)

The hardest part is that the vast majority of our codebase doesn't play well with those limitations, the most painful being the lack of support for IEEE-754 float-s. When we last discussed this topic, the most realistic POC idea we were able to pick was to to implement a a very basic variant of ResizeProcessor on top of Hastlayer. I think this still stands.

As of my personal involvement, I really struggle to find time these days, even for the basic ImageSharp contributions, so I can't promise anything 😞

@Sergio0694 likes to try crazy things, maybe you can convince him to help you 😉

[Piedone]

Yeah, that's some good feedback, thank you. On certain boards, IEEE-754 floating-point is actually available in hard DSPs (otherwise you have to configure a portion of the FPGA fabric to serve as a floating-point ALU, which is possible but inefficient) which we could make available. IEEE-754 floats are not guaranteed to provide the exact same results between different hardware platforms though so that's a challenge.

[antonfirsov]

IEEE-754 floats are not guaranteed to provide the exact same results between different hardware platforms though so that's a challenge.

A lesson every graphics programmer learns very early 😉

I would put this into the "not good, not terrible" category. ImageSharp does not produce identical results across platforms, 99.99% of our users are fine with this, unless results look visually good.

[Piedone]

So anybody interested? @Sergio0694? Or @brianpopow, @tocsoft (taking a look at recent top contributors)? And obviously @JimBobSquarePants but I guess you have enough on your plate already.

[brianpopow]

@Piedone i do find the idea interesting, but i barely find time to do my own tasks for ImageSharp at the moment. So taking another task on top of that is just not feasible for me.

[Piedone]

I see, sorry to hear that.