My Bachelor Thesis "Software Renderer Accelerated by CUDA Technology"

Congrats! Feels nice to get that out of the way, huh?

Yeah, definitely . I have been using OpenGL and Direct3D for a few years and I never knew how they actually work internally. Now I finally do

I miss the good old software renderer days

/nostalgia

Right. Today everything (in real-time graphics) goes around OpenGL/Direct3D and shaders. Sometimes I wish I had been born maybe 10 years earlier so I could start my graphics adventure with software graphics, when the whole graphics industry was emerging .

Congrats on getting that out of the way. Definitely feels nice to have that off your shoulders, doesn't it?

If you are going to use slides in the future, and you are certainly going to, almost regardless of what you'll end up doing professionally, you might want to include slide numbers on every slide. It helps a lot for people to go back to a specific slide with a question. Undergraduate classmates don't like doing this (either due to a general lack of interest, due to a desire to get the class over with early, or due to being afraid to ask something that you won't be able to answer and thus messing up your grade) but eventually someone will ask something


The paper itself does have a couple of flaws. First, even for a practical paper documenting an implementation, the thesis itself is a bit on the short side, especially the literature list. Your introduction does not do much to introduce the topic, instead diving straight into definitions. I would have started out with a somewhat expanded version of your abstract there. The implementation section could've done with a bit more explanation, also. The research itself looks good, I can see you put a lot of time into that. I would also include the full (or the critical parts if its really long) source code in an appendix, although that's more of a preference and depends on what your professor wants. It does tend to make the paper longer, which does wonders for when you print a bound copy

Second, you're not using citations. Is your professor cool with that? Mine would've killed me Every time you have some kind of statement that you did not come up with yourself, whether it's some equation you read in a book or some value judgement or whatever, you really ought to back it up by a citation, otherwise people will assume that it's your original thought, which leads down the unpleasant road of plagiarism charges (which can get you kicked out of university or make you fail the course, plus it'll follow your reputation for life, not a very pleasant thing). Especially potentially contentious statements such as the last sentence of the first paragraph on page 4 need citations.

Oh, and I see you're studying in Wroc?aw. I was there in 2008 on an IT law seminar. Nice place, especially that Japanese garden

Thanks for your suggestions stanirya.

As for the literature; I have included only those books/articles that actually used in my research.
The introduction is pretty short, yes. Computer graphics rookies might find my paper harder to read than people who have some background in OpenGL/Direct3D.
I also have not put enough attention into implementation section. I though there was no need to do so as the implementation is pretty straightforward (when you know the theory, which I mostly focused on). However, on page 35 there is a flow diagram that explains in pseudocode what the renderer does.

We have a Japenese garden? I guess I should take more routes than lodging <-> university

That's pretty cool! Good job. The performance comparisons in Test 2 is OpenGL with 2500 draw calls of cubes (25ms) and then CUDA with one draw call and 2500 instanced cubes (33ms)? The CUDA implementation requires 415ms in the case you do 2500 draw calls of cubes?

Thank you .

And yes, OpenGL looks not that good in comparisons to even the software renderer running without CUDA. The problem is of course neither vertex nor pixel processing but CPU-GPU communication. So the time of 22ms of OGL is actually the time needed to call glDrawElements 2500 times .
CUDA with one draw call (One-Call-Many-Triangles) uses quite a lot of additional global memory to store "processing information", so to speak. This way I avoided calling CUDA for every single triangle but for the cost of memory accesses, what has proven to be very expensive (Test 1). However, it is not that expensive when there is a lot of triangles to process, where One-Call-One-Triangle's performance significantly drops, and One-Call-Many-Triangles still behaves decently.

I have some loose further concepts on how to utilize CUDA to speed the rendering, with different strategies. I am sure I'll get back to the project when I have some spare time and more robust ideas

Hey folks,

After almost 1.5 years I got back to Vainmoinen. You can check the news on the website http://maxest.gct-game.net/vainmoinen/index.html

Cheers, Wojtek