Respect for guys who testing things as librarys etc

Python 3.11 did not add just in time compilation. 3.11 is an improvement of the Python VM in how data is managed, especially object types. The bytecode is still interpreted though.

@CoderSpaceChannel

Жыл бұрын

Thanks, I was wrong. But just in time compilation for small code fragments is planned to be implemented in 3.12, and for large ones in 3.13

@spidertyler

Жыл бұрын

@@CoderSpaceChannel oh wow, thats awesome. I didn't know they planned on doing this. I kept seeing them saying they wanted to try to avoid JIT for as long as possible in search of other solutions.

It's just the comparison I was looking for, I think it's great that you did it, thank you very much!

Found that Arcade performance can be increased by 25% by replacing the line *self.center_x, self.center_y = self.x, self.y* in the SpriteUnit class with *self.set_position(self.x, self.y)* Is this a bug or a feature, I don't know how to interpret it. And you can achieve even better performance if you don't care about collisions between sprites: SpriteUnit: super().__init__(hit_box_algorithm=None)

@DarrenEberly

Жыл бұрын

Hi, great video! I'm one of Arcade's maintainers. The difference in setting center_x and center_y vs using set_position is that internally the position in Sprite is represented as a tuple, and the center_x and center_y properties are mostly helper properties to make it easier to set/access just one of the values. So setting these both actually updates the position twice, whereas set_position only changes it internally one time. I'm also curious if this is tested with Arcade 2.6? We have an upcoming 3.0 release that is approximately a year of development ahead of 2.6 and should offer many speedups, as well as alternative classes like BasicSprite which strips out a lot of the unnecessary weight of Sprite so you can build faster Sprites if you don't need some of the extra niceties of the full class. We're also experimenting with a package to provide speedups via Rust, but that is still very early stages and mostly focused on speeding up collision detection at the moment.

@CoderSpaceChannel

Жыл бұрын

@@DarrenEberly Thank you for your contribution to the development of Arcade, this is a very promising and convenient library for creating 2D games.

great episode. i've always wanted to see a comparison between libs. good work!

Your tutorial are amazing, really impressive the update speed after upgrade to 3.11 !!!

AWESOME video, congrats, if you can do more videos like this, pleassse. Also, please make more 3D rendering things on python

great video, and would like to see more with raylib!!!!

nice trick of the cached rotations! I learned something today, if I haven't watched this video I'd have never think of it, though I'm familiar with cache.

Wow! pygame-gpu performance is unbelievable... on my MacBook Air M1 I reached 17401 sprites and holding 60 FPS.

@skilz8098

9 ай бұрын

And there's probably a lot of other tricks on could implement to even improve upon that...

Amazing test thanks a lot for it!

Amazing video, thank you for sharing

I understand this is a Python comparison, but can't help to wonder what Raylib in C++ could've achieved.

@cookiepolice2086

Жыл бұрын

I would imagine not much better because raylib python is calling c/c++ functions

@ensarija

Жыл бұрын

I have seen test somewhere else, simple setup, but it shows 3-5x increase in speed when using C++ in raylib. Which is expected, since there is no interpreter.

@weirddan455

Жыл бұрын

I was curious as well so I coded it up in C with Raylib (I'm more comfortable with straight C than C++). The answer is: a lot faster. I downloaded his Python code to compare. Python could render 3,400 sprites at 60 FPS. C could render 44,500 sprites. At 3,400 sprites C was getting ~800 FPS. About a 13x performance difference.

@samdavepollard

Жыл бұрын

@@weirddan455 nice; thanks for sharing

@smanzoli

Жыл бұрын

Have just tried it here: I tried 2 different ways: Using sprites (textures) and using shapes (drawing filled circles), using both Pygame and C Raylib. (Standard Python 3.7 interpreter, all with same algorithm (no sprite rotation, just borders collision), no numpy, no jit, nothing). As a bonus, also used p5JS for shapes. First using circle shape: Pygame reached 60 fps for 7.5K circles C Raylib reached 60fps for 1.7K circles p5JS reached 60 fps for 1.7K circles So, Pygame, for circle shape drawing, was 4x faster than C Raylib or p5JS. Then I changed for sprite (40x40 png, no rotation): Pygame reached 60 fps for 3K sprites (performance dropped 50% from shapes to sprite, yes, I used convert_alpha()) C Raylib reached 60 fps with 95K sprites p5JS not tested. So, Pygame got half of previous performance and C Raylib improved 53x, being 31x faster that Pygame using textures. For some reason, textures are easy for C Raylib and a pain for Python... and shapes are a bit easy for Pygame and a real pain for C Raylib. Here are the codes (simply to change comments to swap between textures and shapes). The shape I used was a 40x40 png, a ball with transparent background. Environment: old 7 year old laptop: Win10 64, i7 6 gen, nvidia 980m, 16GB DRAM **************************** C Raylib **************************** #include "raylib.h" #include typedef struct Ball { Vector2 position; Vector2 speed; Color color; int radius; } Ball; int main() { const int W= 1600; const int H = 800; //SetConfigFlags(FLAG_VSYNC_HINT); InitWindow(W, H, "Sander"); SetTargetFPS(240); const int N = 1700; //Texture2D textura = LoadTexture("ball2.png"); Ball *balls = (Ball *)malloc(N*sizeof(Ball)); for (int i = 0; i { balls[i].position.x = (int)GetRandomValue(50, W-50); balls[i].position.y = (int)GetRandomValue(50, H-50); balls[i].speed.x = (float)GetRandomValue(-4.0, 4.0); balls[i].speed.y = (float)GetRandomValue(-4.0, 4.0); balls[i].color = (Color){ GetRandomValue(50, 240), GetRandomValue(80, 240), GetRandomValue(100, 240), 255 }; balls[i].radius = (int)GetRandomValue(5, 10); } while(!WindowShouldClose()) { BeginDrawing(); ClearBackground(RAYWHITE); for (int i = 0; i { balls[i].position.x += balls[i].speed.x; balls[i].position.y += balls[i].speed.y; if (((balls[i].position.x + 30) > W) || ((balls[i].position.x - 30) if (((balls[i].position.y + 30) > H) || ((balls[i].position.y - 30) DrawCircle((int)balls[i].position.x, (int)balls[i].position.y, (int)balls[i].radius, balls[i].color); //DrawTexture(textura, (int)balls[i].position.x, (int)balls[i].position.y, balls[i].color); } DrawFPS(50,50); EndDrawing(); } free(balls); //UnloadTexture(textura); CloseWindow(); return 0; } **************************** Pygame **************************** import pygame import math import random pygame.init() w, h = 1600, 800 window = pygame.display.set_mode((w, h)) font = pygame.font.SysFont("Arial", 30) clock = pygame.time.Clock() #imagem = pygame.image.load("ball2.png").convert_alpha() def run(): run = True q = 7500 x = [] y = [] vy = [] vx = [] r = [] cor = [] for i in range(q): r.append(random.randint(5, 10)) x.append(random.randint(30, w-30)) y.append(random.randint(30, h-30)) cor.append((random.randint(10, 255), random.randint(10, 255), random.randint(10, 255))) vx.append(random.randint(-4, 4)) vy.append(random.randint(-4, 4)) if vy[i] == 0: vy[i] = 1 while run: window.fill((0, 0, 250)) for event in pygame.event.get(): if event.type == pygame.QUIT: run = False break for i in range(q): if y[i] > h-r[i] or y[i] vy[i] *= -1 if x[i] w-r[i]: vx[i] *= -1 x[i] += vx[i]; y[i] += vy[i]; pygame.draw.circle(window,cor[i],(x[i],y[i]),r[i]) #window.blit(imagem,(x[i]+20, y[i]+20)) clock.tick() fp = int(clock.get_fps()) txtsurf = font.render("FPS: " + str(fp), True, (0,0,0)) window.blit(txtsurf, (25,45)) pygame.display.update() pygame.quit() if __name__ == "__main__": run() ******************************************** Anyone could explain the huge jump from C Raylib when using textures???

I'm curious if you could look at Kivy game programming as pygame isn't supported on Android or iOS, there is pygame subset 4 android but that seems out of date and still doesn't support iOS. Mobile platforms are a major market and Kivy seems to be the most recommended even for games. Also what IDE are you using? Is it PyCharm? Still a Great Video! Great Job! :)

@skycloud4802

Жыл бұрын

I think Ren'py works for Android but it's not very optimised for that in my opinion, and very visual novel orientated. I think Kivy is still probably the best bet for for python on Android right now

Thanks, it was interesting to watch.

Hi! great job. Do you have any information about CPU and GPU ram consumption?

You're wok is great! Keep it up!

I've always meant to ask: What voice synthesizer do you use? It's so good :)

@CoderSpaceChannel

Жыл бұрын

cloud.google.com/text-to-speech

How technic you use to typewriter constantily and fast?

Awesome tutorial, I tried pygame cpu and cpu cache using a sprite sheet instead of separate files and the performance is surprisingly the same, spritesheets are better for use with programs like Tiled. Although I have one question, how would the numba and taichi libraries be used in this type of test?

@leobozkir5425

Жыл бұрын

using spritesheets only improves loading time, not performance while the program is running

This was unexpected. I love pygame but I thought it would be the worst. Anyway amazing video. And one more thing how did you install pygame on python 3.11 shouldn't it be not supported yet. I always get an error when using pip

@CoderSpaceChannel

Жыл бұрын

pip install pygame==2.1.3.dev8

How do I use the gpu with pygame?

what are the results for python 3.12 and what are they for the new 3.13 beta?

Is there some good tutorial how to use pygame sdl2?

😎 , cool video!

It would be interesting to also see some significant game logic running. These tests feel like profiling different c library wrappers.

Any updates on this SDL 2 Video?

How do you like the idea of creating Pacman on pygame?

Correct me if I'm wrong but can't the performance of pygame be improved by using the .convert_alpha() method on the loaded images

@CoderSpaceChannel

6 ай бұрын

.convert_alpha() is used for surfaces in software (CPU) rendering. When using the GPU to create a texture from an image, .convert_alpha() will result in an error

Damn! The pygame GPU is CRAZY! Do you know when this stuff was implemented? There seems to be essentially no information on this anywhere, but seems to be such a magical way to speed up pygame rendering yet no one is talking about it?

@CoderSpaceChannel

Жыл бұрын

this is still under development and the final API may change

@danub5551

Жыл бұрын

@@CoderSpaceChannel yeah, but it seems to be too useful of a feature to miss out on right now

@thinkingspace3438

Жыл бұрын

@@danub5551 a big missing feature is loading textures with the gpu, but you can use the sdl2 functions with ctypes actually

@distantforest2481

Жыл бұрын

They say the api is subject to change but it hasn't changed in a year so I'd say it's safe to atleast test out. At worst you can modify your code to any changes they make

@michaelpalmer2143

9 ай бұрын

@@distantforest2481 I hasn't changed since it was first implemented 3 or 4 years ago. I was the first to complete a game with it I think (Flyboy, trailer on this youtube account) and discussed possible enhancements with the dev team back then.

Do you recommend me using Raylib for a commerical game? I am considering Godot, but I am a Python engineer and I'd find it more pleasurable to keep everything in Python. I have 1 year experience with PyGame, but Raylib seems more lightweight and more performant.

I wonder how that compares to something like Bevy or naylib (Nim)

when are you going to make a video of raylib with angle vs pygame

Would PyGame GPU have any performance benefits from caching the same way PyGame CPU did?

@michaelpalmer2143

9 ай бұрын

No, rotation and scaling is done by the GPU and is free. The only optimization is to make sure the draw calls are properly batched by using as few separate Textures as possible and handling the draws with as few texture switches as possible. And being Python, limiting the number of separate iterations and function calls helps. But realistically, the performance is excellent. The games you can see on this KZread account run at about 2% CPU usage, and I've even gotten them running perfectly on a low powered Linux retro handheld (my Anbernic RG351p)

man, I am very goood with Python, but never worked with game engines. Your work is really good! Where can I donate some money?

@CoderSpaceChannel

Жыл бұрын

Thanks Button 👍

Are the sprites not being batched? I made sprite renderer in C and it can do 100K without breaking a sweat.

@DarrenEberly

Жыл бұрын

The problem is moving sprites, not drawing them. Arcade for example can easily draw several hundred thousand static sprites(albeit with a high load time), but moving them requires updating the data in the OpenGL buffers, pygame pulls ahead when using the GPU here because it is doing parts of this in C, whereas Arcade is 100% python.

@michaelpalmer2143

9 ай бұрын

@@DarrenEberly Actually the GPU renderer is a prototype and the only part of the engine written in cython instead of C. But it still gets a decent jump for that. But yes it's iteration and transforms in python, not the GPU rendering that limit performance here. I suspect my recommend transform_and_draw() enhancement might eventually be added, or if not, I'll write it in C just as soon as the GPU api is ported to C.

Very Nice .

Thank you

Quick question - is it better to choose Pyglet or Arcade for game development?

@CoderSpaceChannel

8 ай бұрын

If we talk about 2D games, then Arcade. But this is only for desktops - Windows, Linux, macOS. Arcade is not suitable for Android, you need something based on OpenGL ES or Pygame

You pay a lot just for the call overhead itself in python. By simplifying the code making less methods you can gain quite a bit of extra speed, but I think the classes you made are probably closer to what people make in real life. Maybe the separate translate and rotate method is overkill.

@skilz8098

9 ай бұрын

That's were a 4x4 matrix class with all of its operators and functions would come in handy.

Raylib is a pure C library with over 50 language bindings. I chose it in part for this reason. After developing a game in Python, you have the option of moving to another language for learning or performance reasons, while still keeping your knowledge of the library and all the same calls.

@alexale5488

2 ай бұрын

That's so true! I plan to build a game engine for 2D games in Raylib. Python and Raylib works too well. Also, Raylib is even simpler than PyGame and has a better implementation. For example, if you want to get the key inputs only once in PyGame, you're dependent on pygame's own event loop. But in Raylib you have a function call for that and you can associate it with any class.

@Unavailable8923

2 ай бұрын

@@alexale5488 I'm probably switching to Godot. Raylib was fun, but Godot 4.x just has more features out of the box including easy exporting of your game to different platforms.

did you consider pygame-ce ?

In arcade you can cheat and hookup up a compute or transform shader wirting to the position and rotation buffer in the spritelist if you want to get millions of sprites. It's probably cheating.. but you can do it :D

don't know why, but I have python 3.11 and the Arcade test went a little better than the pygame + gpu version, while the raylib was the same as pygame + gpu

@hillybankok

9 ай бұрын

its probably the hardware that you are using

I think you should use a sprite sheet instead of individual sprite images. Should make everything faster.

@CoderSpaceChannel

Жыл бұрын

thanks, helpful advice 👍

@thinkingspace3438

Жыл бұрын

it makes stuff faster only for stuff using the gpu i guess, using the same texture means less processing

@yds6268

Жыл бұрын

@ThinkingSpace doesn't rotation and movement use GPU? Changing texture coordinates could be done in vertex shader

@thinkingspace3438

Жыл бұрын

@@yds6268 gpu can rotate on the go yea, but in cpu cache mode the surface also already exists so it's the same

@skilz8098

9 ай бұрын

@@thinkingspace3438 Not exactly... If the information is in CPU cache sure the CPU could probably calculate it faster than what the GPU for small data sets could, but when you have large data sets of the same type, the GPU is much faster because of the scale of all of its cores. You are limited by the bus transfer speed from the CPU to the GPU. So sending batches of similar large data to the GPU so that the GPU can do those calculations is much faster than having to go back and forth from the CPU to the GPU and back, etc... hundreds or even thousands of times per frame. This causes a major bottleneck. It is not the same.

should test with pygame-ce (pygame community edition)

Yeah, other Python game engines are also important.

pls a tutorial about pygame.sdl2_video

But how about ray lib with C? surly that is a lot faster?

@michaelpalmer2143

9 ай бұрын

20 times faster probably, but also 10 times slower to develop in. Realistically there is no bottleneck for 2d games in pygame. And Python is an excellent language for learning about game engine and system design. You can always write performance dependent parts in C if you need to. I really think python is the best scripting langauge for game design, and should be default in Godot, Unity, and Unreal.

@PySnek

2 ай бұрын

Go with Nim (naylib) if you want Python-like syntax and speed

Your speed is like 7×, can you please slower it down a bit I would be very thankful, and can you please make a game where we will have two ships, one being in our control and the other one will be automatically controlled by computer

In your Raylib class, every call to update() and draw() creates garbage: a list that is never used. An imperative "for" loop would be more efficient.

you should add pygame2

Arcade is not better in everything though. I tried An nbody simulation with about 500 particles(each only 1 pixel) pygame did perfectly fine with Around 50-60 FPS. But arcade was like 2-5 FPS

you should compare it with pypy!

@vladimirkraus1438

Жыл бұрын

is any of these packages available for pypy?

@user-kn6uu8cc6l

Жыл бұрын

@@vladimirkraus1438 pygame and raylib yes, arcade maybe, because its pyglet based and it have pypy support

amongASS

Its definitively not a benchmark test if you use different hardware to get the results. All of them should be benchmarked on GPU or CPU and comparee objectively, otherwise you are fudging the results in favor of Pygame, because you kept reimplementing it just so it would out perform the others, while failing to actually ensure GPUs were used for the others as well. Biased data is useless data.

By just changing self.center_x, self.center_y = self.x, self.y to self.postion = self.x, self.y you can increase arcade test count from 3k to 4k (on my machine). You moving each sprite twice per frame and looks like moving is expensive operation in arcade. It sounds stupid I know but it really works. You can easily achieve 6k on arcade by skipping position changing processing for each sprite. To skip arcade position change processing for each sprite: add to SpriteUnit init: self.position = [self.x, self.y] # use list for sprite coords SpriteUnit def update(self): self.rotate() self.translate() self.position[:] = self.x, self.y # skip setter SpriteHandler: def update(self): self.sprites.update() for s in self.sprites: self.sprites.update_position(s) # update gl buffer. Looks like use sprite to just drawing is overkill. It has too many logic in python.

@CoderSpaceChannel

Жыл бұрын

Indeed, it matters. self.set_position(self.x, self.y) i think it's ok to skip setter when you don't need the collision functionality

@itnabigator

Жыл бұрын

​@@CoderSpaceChannel set_postion is just self.position = (center_x, center_y) it still include the same overhead. PS. I found how to get 5k without "hacking" just need to disable hit_box in SpriteUnit: super().__init__(hit_box_algorithm=None)

@CoderSpaceChannel

Жыл бұрын

@@itnabigator Thanks, that's pretty useful information when developing with Arcade

@itnabigator

Жыл бұрын

@@CoderSpaceChannel Thanks for your videos $) Without it I will not even try to investigate that.

Last time i checked, Raylib didn't have any python support (and was almost strictly C++, so that was long ago...), so this news bring me joy. You could also try to use @lru_chache from funktools for sprite operations. Could be helpful. Or not, I didn't checked. (or even better, just @cache because it's the same as @lru_cache(maximize = none)... And as somebody before already pointed out -> JIT is not here yet, but some data management is...