voidstarhq / voidstar Goto Github PK

Add some flag to toggle distortion by stacking this shader at the end of the render pipe:
OpenHMD/OpenHMD#83

Main idea is for CI to verify that a file gets properly rendered.

This should depend on #31 & #32
Maybe use on xvfb?
Test miners again some files already in the repo.
A first step would be to count the non-background pixels in the image.

When asked to open a file named -, read data from standard input.

https://www.reddit.com/r/EngineeringPorn/comments/m28nox/giant_ping_pong_rgb_led_cube/

https://docs.travis-ci.com/user/languages/objective-c/
https://gist.github.com/johanneswuerbach/5559514
https://www.objc.io/issues/6-build-tools/travis-ci/
https://www.buddybuild.com/pricing/#starter

Right now a very static vertex buffer is allocated on host memory, this can probably be done directly on the GPU with geometry shaders (see layout). By default this buffer is full and of size 256 * 256 * 256 * 3 * 4, or about 200MB.

Also, it is of note that the colour buffer is mostly empty most of the time, while it is of the same size as the vertex buffer.
To fix this I am thinking of first processing the input file, turning it into a texture, then moving the sliding window on this texture.

We draw 256**3 pixels when we could be skipping all the empty ones.

16M pixels is a lot for small graphic chipsets.

The only things not compiled into the binary right now are the shader files.
I really like the simplicity of a single binary file right now,
so this issue is to request that the shaders be compiled into the binary.

Using something like https://github.com/tmpvar/voxviz/blob/fca77c586fc4b682d9a66bb271ea920c0d71963c/shaders/convert.js (but just using a shell script) should do it.

...or any open & easily readable 3D asset format. (Something that can be read in the browser).
Not sure if OBJ supports color?

Format should also support adding metadata (we might want to add sound / velocity info).

It can be done with shaders: freeciv/freeciv-web@2d88ebb

Impl https://github.com/voidstarHQ/voidstar/blob/6d797f96f938391805a10ae43439cd0c3dd93243/voidstar/algos_3d/cube_conti_frebet.cc RGB for other color spaces e.g. YCbCr, ...

Use arrow keys to enlarge and move a sliding window on the inspected file's contents.
Probably add a key to start/stop an animation.

Miners was the first thing that came to mind after "starting" minepi (which itself came from https://github.com/philipl/pifs).

It's not a great name, or more exactly: I've found a much more fitting name!
void * "void star" is (appart from both a Star Trek & Star Wars reference) a void pointer in the C language.
It is used to cast any kind of data into whichever pointer type you want. Pretty fitting!

For reasons the binary name should have to change to voidstar.

Here's an excerpt:

./miners -a conti miners
OpenGL version: 4.1 INTEL-10.24.45
GLSL version: 4.10
Vendor: Intel Inc.
Renderer: Intel HD Graphics 4000 OpenGL Engine
Loading file miners
aspect: 1.33333 (800x600)
read 639264 bytes
#points: 45871
aspect: 1.33333 (800x600)
OpenGL Error 1282
OpenGL Error 1282
OpenGL Error 1282
...SNIP

Using glDrawArrays works without issue but is very far from optimal.

Problem has been happening ever since 9868865...607a0cf.

and potentially OSX too

When reading data, it may be useful to convert plaintext to binary:

• from hex plaintext this way for instance
• from URL encoded plaintext

CLI flag: --input-transform=TRANSFORM

Some CLI option (also maybe a key) should trigger a "screen capture" of the current render.
CLI needs an option to specify the file to write to (or in case of directory, generate a timestamp-based file name).
In the future this path could also be a URL to upload the file to.
This issue probably doesn't need a config file (yet).

Output path should default to $HOME.

t-sne is a non-parametric unsupervised data projection algorithm that could be interesting to run on chunks of pre-specified size of the input data.

When starting the program

• sometimes the render will be right in the middle of the screen, visible
• most of the times it will not & user will need to "look around" to find it...

Starting with the --fullscreen CLI option on Mac creates a window as big as the viewport (full screen)
But nothing shows up. The screen is black.
Maybe related to #32 ?

Reported by @rickying

Note: hitting F on an open window resizes the window to the maximum possible though.

The time it takes to transmit data to GPU is non negligible.

Right now what we do is

1. mmap the file to RAM
2. create RAM buffers for vertices & colors & the indices of the elements that are worth rendering
3. transfer these buffers to GPU memory
4. draw
5. repeat

Instead we could

1. load the file into a texture (mmap on GPU?)
2. create vertices & their colors with shaders (tricky part*)
3. draw
4. repeat

This would make the host code much simpler.
Thus porting to Web or Android would be much simpler.

*The tricky part is that to do something like Algo3DCubeFrebet with shaders, one needs to somehow:

1. read multiple bytes from the file-texture
2. write a color to the output texture but at different coordinates
   I believe that (x1,y1) --> (x2,y2,z) transformation is hard for a GPU.
   Maybe compute shaders can help here?

@icehess Ideas?

Work with valgrind to clean up after GLFW.

Simple enough to reproduce leaks:

valgrind --leak-check=full --show-leak-kinds=all ./voidstar -a 4col ../data/*

and

valgrind --leak-check=full --show-leak-kinds=all ./voidstar ../data/*

Note: in my testing just calling glfwDestroyWindow in the GlfwManager destructor is not enough (and actually shows more leaks).

Tested on #48 (so tag v1.4.*)

Use something like https://github.com/lightsofapollo/x-recorder
to add videos to the readme.

Might help for #33