Game Development

Stephane Hockenhull

00:00

I think I just about broke even with my $300 3D printer 2 years later with all the little broken pieces of plastic I replaced on chairs, windows, bed, coat hangers, misc brackets, etc. Not even counting the puzzle-solving entertainment value of it.

DMGregory

@Bálint 20:02 Monday here (Toronto)

Bálint

@DMGregory Yep, it's tuesday here

DMGregory

Neat :) I hope the future is bright and full of potential. :D

Stephane Hockenhull

I know for a fact the future has candy.

Alexandre Vaillancourt

My cousin lives in Toronto :P

Been there a couple of times. It's a big city!

Stephane Hockenhull

00:21

@DMGregory okay, the guy said something false: he said the cache changes the result.
It doesn't. What will change the result between executions on the same machine is that some bad hardware drivers (eg: printer driver) changes the rounding mode.
Or if the compiler mixes x87 and SSE float operations in unrolled loops and one unroll uses x87 while the other uses SSE or a different mix of both.

DMGregory

Not being an expert on hardware myself, I'll take your word for it.

Stephane Hockenhull

Unless there's a really old machine I don't know/care about :P
At least on x86, ARM, PPC, and MIPS the cache doesn't affect the FPU results.

Alexandre Vaillancourt

@StephaneHockenhull I've been looking a little at it lately, do you know if it's possible to tell the CPU to make SIMD operations in certain contects?

contexts

Stephane Hockenhull

You have to tell the compiler that. the CPU only executes what the compiler wrote.

Bálint

Ok, this is very strange

Stephane Hockenhull

00:30

GCC will try to pack SIMD but last I checked was 4.x, I still have to look at GCC 5.x's compiler output.

Bálint

I have a list of integers, e.g. 1, 2, 3, 4

And I want create a number out of them, e.g. 1234

I tried doing

for (int l = 0; l < comb.size(); l++) {
    num += comb[l] * pow(10, l);
}

But this somehow returns one less if the length of the number is greater than 1

Stephane Hockenhull

int nums[] = {1, 2, 3, 4};
int result = 0;
for(int n : nums) { result = result * 10 + n; }

Alexandre Vaillancourt

@StephaneHockenhull I'm using VS. I'd like to improve vector3 operations if at all possible. I don't know if the compiler uses SSE(2) or not...

Bálint

@StephaneHockenhull That' a lot better

thanks

I shouldn't code at 1 AM

Stephane Hockenhull

@AlexandreVaillancourt What I have seen with SSE (I think it was SSE1 when I tested it) was that in order for it to load into an SSE register I had to store all vec3 as vec4 and then I had speed ups on very small datasets that fit in the L1 cache but anything larger (like skinning a mesh or particles) the penalty of the wasted 4 bytes (133% size) caused enough cache misses to not be worth it.

the memory/L3/L2 bandwidth cost was more than the SSE gain.

So unless the compilers got much smarter and can manager reading vec3s into SSE registers it wasn't worth it.

Bálint

00:37

@AlexandreVaillancourt

SIMD and Vectorization: Parallelism in C++ #1/3 (multitasking on single core)

►

This is a pretty nice overview on c++ SIMD

5:40 is where the actual coding begins

The Mattbat999

How do I link a tag in a post like shown here :gamedev.meta.stackexchange.com/questions/2550/…

Alexandre Vaillancourt

@StephaneHockenhull Ok, I have a nice test app at work with actual data I'll test on. Maybe in that specific case it will be a speed up, maybe not.

Bálint

@TheMattbat999 hello => [ tag: hello ]

without the spaces

The Mattbat999

@Bálint ok

Alexandre Vaillancourt

@Bálint Thanks! I'll take a look :)

I have the source of GML too which I can use.

Stephane Hockenhull

00:43

@AlexandreVaillancourt So the compiler will use SSE but in scalar form when working on vec3 which isn't too bad because the FPU is pipelined and will run all 3 scalar operations *almost* in parallel.
If you want to try getting the compiler to issue full SIMD operations you have to copy your vec3 into a vec4 first, then operate on that (eg vec4 = vec4 * mat4) then store the result back as vec3. which *might* be faster.

Alexandre Vaillancourt

According to the profiler, the dot product method is "hot"

Stephane Hockenhull

But most of the time in practice I've seen that the memory latency is such that neither have impact compared to the CPU's speed.

Can you inline it?

Stephane Hockenhull

00:55

SSE is not great for manipulating vec3s.
The trick with SSE if you have massive amount of vec3 is to rearrange your data so that all the Xs follow one another, Ys, and Zs, in different arrays. So you can process 4x X, then 4x Y, then 4x Z then write the result.

Alexandre Vaillancourt

@StephaneHockenhull It's a template function, so I guess it is. It's not explicitly inlined.

@StephaneHockenhull I don't think it's possible in this situation.

Stephane Hockenhull

const float *x1;
const float *y1;
const float *z1;
const float *x2;
const float *y2;
const float *z2;

float *dots;

for (int i = 0; i < size_div_by_4; ++i){
((vec4 *)dots)[i] =
((vec4 *)x1)[i] * ((vec4 *)x2)[i]
+ ((vec4 *)y1)[i] * ((vec4 *)y2)[i]
+ ((vec4 *)z1)[i] * ((vec4 *)z2)[i];
}

megafastlotsofSSEsupergigaflopdotproduct();

When they boast that SSE can to a lot of dotproducts per second that's how the data has to be arranged :P

So I'm doing 4 dot products at a time. That's not practical in normal game situations.
Might be doable in a particle system.

Alexandre Vaillancourt

Hmm, ok, that's an interesting take.

Stephane Hockenhull

because for SSE to really work the data has to be aligned to 16bytes. vec3s don't do that.

So SSE is great for photoshop dealing with vec4 RGBA floating point pixels for example

Alexandre Vaillancourt

I managed to cut off 10% in time of an algorithm by changing how the data was placed in memory, but the "rearrangement" overhead takes 10% of the time, so in the end, I did not gain anything. But it was worth try trying.

Stephane Hockenhull

01:04

that's pretty much the result I got every time I've tried to be SSE-friendly too

Alexandre Vaillancourt

Oh I was only trying to be cache friendly :)

(which worked, I had a lot less last level cache misses and references)

being SSE friendly would be the next step, but not now, I guess. I have other things to do at work.

The Mattbat999

01:18

Grrr.... post up for over an hour and no activity...

Alexandre Vaillancourt

@TheMattbat999 GDSE is slower than SO :)

The Mattbat999

@AlexandreVaillancourt what's SO?

Alexandre Vaillancourt

@TheMattbat999 Stackoverflow

The Mattbat999

@AlexandreVaillancourt oh. It seemed familiar... now I feel rather silly...

Alexandre Vaillancourt

No worries :)

The Mattbat999

01:27

@AlexandreVaillancourt I have also began work on a new project (has no name) but it is where you defend against an onslaught of space ships. You buy rotate speed, shoot speed, and health bonuses in the store witg coins obtained from baddies

Alexandre Vaillancourt

01:40

@TheMattbat999 Nice!

Is it fun?

The Mattbat999

@AlexandreVaillancourt ain't finished yet

Alexandre Vaillancourt

@TheMattbat999 Ah!

@StephaneHockenhull Maybe our dot product is hot because it's not aligned? But some of the calculation is made with data on the stack, so I not sure how it works there...

Stephane Hockenhull

If the loop is unrolled and the dot product inlined it shouldn't even show in the profiler unless it's so spread out that it confuses the profiler into thinking the entire loop is the dot product function.

The Mattbat999

@AlexandreVaillancourt almost though. Just gotta test gameplay and add the shop

DMGregory

@TheMattbat999 I've got an idea for the terrain thing. Let me give it a quick test to see if it works...

Alexandre Vaillancourt

01:53

@StephaneHockenhull I seem to be having trouble profiling...

DMGregory

Hmm, first off, @TheMattbat999, did you get an error about too many texture interpolators when using the terrain hole shader you linked?

Stephane Hockenhull

02:12

@AlexandreVaillancourt you want to profile with the compiler-optimized version, but also the compiler-optimized version makes inlined functions interleaved / blended into the calling function and that confuses the profiler. you may have to go up one on the callstack to figure out what's the hot function.

The Mattbat999

@DMGregory in all honesty, I haven't tried the hole shader thing because I didn't want to waste my time on it if I couldn't get the collider thing to do right :)

@AlexandreVaillancourt tested it. It is fun. Just gotta add the store....

Stephane Hockenhull

So the dot product may not be the hot part, it might just be the big loop that has the dot product blended into its code.

Alexandre Vaillancourt

@StephaneHockenhull yeah; for some reason, I don't get the little highlights VS used to give me. I remember the .dot method being redder than the others, even if they were inlined.

The dot product is not in a loop per se. The part of the code (from a particle system) computes the propagation of impulse between particles.

@TheMattbat999 Cool!

The Mattbat999

@AlexandreVaillancourt it is challenging though, since without upgrades you rotate somewhat slow

Alexandre Vaillancourt

@TheMattbat999 Yeah, that's the point of a game!

DMGregory

02:18

@TheMattbat999 Ah, cool. Personally I prefer the method of detecting the hole parts in the vertex shader and setting the position to NaNs. That chops out whole triangles without alpha blending whatnot, so you don't pay alpha blending/testing cost on the whole terrain. It leaves a jagged edge, but usually you've got meshes skirting the rim of the cave entrance anyway.

The Mattbat999

@DMGregory ummm... ? You said "chopping" out vertexes. How do you do that? (Please dumb it down to the lowest level possible. I am still a little noobish to some of the more complicated stuff)

@AlexandreVaillancourt gotta get the shop going, then I will put some screenshots up on here!

Alexandre Vaillancourt

@TheMattbat999 :)

DMGregory

You sample your hole texture (or generator function) in the vertex shader. If that returns "this should be a hole" then you overwrite your projected vertex position with 0.0f/0.0f

Alexandre Vaillancourt

@StephaneHockenhull But yeah, there are no mention in the 'functions' of the dot method. It's inlined for sure.

Evorlor

@DMGregory I cannot believe you aren't a mod already! 1st choice, hands down. I really appreciate all the help you have given the community, especially me

@AlexandreVaillancourt You're on my list too :)

DMGregory

02:23

Aw, I appreciate that! Being a mod isn't about answers though - I'll keep doing that either way.

And others here have more community management experience than I have, so definitely think carefully about your votes. :)

Evorlor

@DMGregory you have also helped me reform my questions in the past as well

Alexandre Vaillancourt

@Evorlor Thanks!

DMGregory

...The primitive assembly step can't make a triangle with one of its is NaN, so it aborts the whole triangle, and you don't pay any rasterization or blending cost in the hole. And you get to keep your shadows. :)

"when one of its vertices is NaN" I meant to say. Looks like setting the vertex to NaN aborted it right out of my sentence too! XD

The Mattbat999

@DMGregory ok.... I will have to look more into this.... tommorw however. I will probably be going to be in a few minutes and there is 0 point in trying to start on this tonight :)

DMGregory

Gotcha. I'll finish my test so with a little luck you'll have at least one option to check out the next time you're back.

Stephane Hockenhull

02:27

@DMGregory unless you're on a 3DS. Then the GPU just goes completely comatose and the game freezes waiting for the GPU to finish but it never does.

Alexandre Vaillancourt

Good night all!

Free Internet Points

3

Q: Why is there a white border around my app icon?

I am developing an Android app in Unity. I have two versions of the app on my phone, both with identical icons. The white border is only on the newer version of the app. I am using a Google Pixel. Why is this there, and how do I get rid of it? Is this tied to my phone or my app? If it is tied ...

Stephane Hockenhull

@DMGregory https://www.3dbrew.org/wiki/GPU/Internal_Registers#Overview
> Invalid GPU command parameters including NaN floats can cause the GPU to hang, which then causes the GSP module to hang as well.

The Mattbat999

@AlexandreVaillancourt night

Stephane Hockenhull

And then Stephane spends weeks debugging the 3DS port trying to figure out why the game is crashing and where do those NaN come from. (actual thing that happened)

/me twitches

DMGregory

02:31

Hahaha, good to know. Yeah, I think I've mainly seen this trick used in PC / non-portable-console titles.

Evorlor

@DMGregory out of curiosity, how often are you on se?

Stephane Hockenhull

I've seen another GPU treat NaN as zeros or infinities. But I can't remember which. But that meant polygons going weird places.

DMGregory

I'm usually scanning it at least a couple of times a day (more if I'm anxious - I like finding problems I can solve when I'm unsure of myself). I'm usually not that active on chat but I'm trying to be more present so I'm not out of touch with what's happening.

Good to know. It might then be an artifact of console dev, where we know with certainty how the GPUs we're targeting will react since they're limited to only a few distinct models.

Stephane Hockenhull

Speaking of NaNs, I'm still watching the guy's videos on posits (floats alternative)

no NaNs

He says "you just stop computing and throw an exception" ... yaaaa except that doesn't work with GPUs / GPGPU computing.

Noctrine

We are now in the election phase of the 2017 Moderator Election

4

DMGregory

03:02

Yeah, that was my thought on the video too.

As we know though, GPUs already do quirky things sometimes with operations that hypothetically should NaN, treating them as zeroes or infinities, so those might be acceptable workarounds to keep things flowing.

For a lot of purposes, an infinity gives a signal similar to a NaN ("something outside normal parameters happened"), especially since posits are less eager about rounding valid-but-extreme results to infinity than floats are.

But yeah, I still don't really expect posits to take off. The math geek in me finds them thrilling, but such a huge amount of hardware, software, documentation, and expertise is already committed to floats.

Stephane Hockenhull

And posits are very complex to implement in hardware. They got variable bit size for both the header bits part and mantissa size.

That is really difficult to implement in hardware.

I think it'll take many more transistors than regular floats

Like, sure, you can make a bunch of 256x256 tables for your 8bit x 8bit posit operations. But when you get to 32bit posits? A table? really?

DMGregory

03:26

He talks a bit about the die space issue, but I haven't gone through the steps to check his math. ;)

Stephane Hockenhull

He's got this top part with the repeating bits which you need to count the bits before the zero in order to figure out how many bits of mantissa there is. That is pretty heavy in transistors.

DMGregory

Again, not being a hardware expert, I'll defer to your judgement. I found the video fascinating mathematically but I'm certainly not trying to evangelize the concept. ;)

Stephane Hockenhull

I find it very interesting tho. The results look much better than floats

Especially that math part with the elongated triangle. That's an issue I've hit with floats a lot.

DMGregory

Point. For all their problems though, I do have to hand it to floats that they're "good enough" for an impressive range of problems. I more often encounter errors that are algorithmic, rather than intrinsic to the number representation, once basic precautions are taken like re-centering simulations that stray away from the origin.

Stephane Hockenhull

good enough is the key here :) And he's right about video games. 99.9999999% of the time floats do the job just fine for us :)

DMGregory

03:41

So far I've only hit two situations where I had to reach for a double.

And that's given that my current project is set in outer space. XD

Stephane Hockenhull

are you combining it with an integer sector offset vector?

DMGregory

I'm not doing that math myself, so I can't speak to how it's implemented under the hood. I'd suspect that's the crux of it, plus or minus some details.

Stephane Hockenhull

gotta go. getting late. cya

DMGregory

Have a good night!

person27