Wednesday, March 21, 2018

Microsoft's Imagine Cup Challenge - Team Heuristic Clinic

A month ago I prepared a quick team and a project for consideration at Microsoft's Imagine Cup. It was an intriguing experience to document details, to present in a video, and to polish a previously developed prototype - a wireless motion/gesture capturing app and a set of sensors.
I understood there was a slim chance to actually win the Cup, so I instead focused on setting up this project for possible future applications elsewhere.

The team is called Heuristic Clinic, the project is Applied Neural Diagnosis. Brief description can be found here: Heuristic Clinic, and the resulting certificate can be viewed here.
They were kind to offer $500 worth of Azure cloud, but I had to decline because the biggest expense is funding client-side development. Appreciated.


©

2018 Big Idea Challenge

Top 50 - 2018 Big Idea Challenge

This Certificate is Presented To:

alexey l

Team Heuristic Clinic

Congratulations and thank you for your hard work and dedication as a competitor in Imagine Cup 2018. You are now part of an elite international community of students who have shown remarkable creativity and innovation to push technology forward.
©
Pablo Veramendi
Imagine Cup Competition Manager
Microsoft Corporation

Sunday, June 25, 2017

Martian sandstorm

Now a fully interactive volumetric sand storm, probably somewhere on Mars.
This marks an intermediate milestone following months of research into volume rendering. It uses finite differences solver for a fluid dynamics equation, some additional research into graphics pipeline on a Mac and iOS.
Note the shadows, they depend on volume density, and also note the unbounded nature of volumes. Other methods would produce visible banding artifacts.


Saturday, June 17, 2017

Unexpected volumetric clouds

I was playing with code that I am testing to produce fluid dynamic effects, and I came across a weird cloud generating setup.
The advection part is failing at some point here and causes gas density to freeze in some regions of the map. It remains static and appears as a cloud, and I think this can be further improved to produce realistic looking real cloud. What's missing is a better phase function, and shading based on anisotropic scattering of different wavelenths.




Monday, May 29, 2017

Fun with self-shadowing volumetrics

As a preparation to a full-fledged interactive volume rendering, here is the first functional volumetric composition. It uses a few techniques to achieve a self-shadowing appearance, as you would expect from a real-life volume of transparent matter.


Wednesday, May 24, 2017

Volumetric hints.

A few volume rendering hints.

1. Cannot discard writing into render-targets selectively in an MRT frame-buffer. This arises, for example, when two targets are used to store two custom denormalized depth buffers. A frame-buffer has a depth attachment as well, but it is used in a regular sense of a depth buffer.

2. Volume rendering can use analytic volume buffer or a set of depth textures. To limit volume to a custom 3D shape requires preliminary step to render that shape into two depth attachments, one with inverted normals and inverted depth test:
// Enable depth test, inverted, front to back.
glDepthMask(GL_TRUE);
glEnable(GL_DEPTH_TEST);
glDepthFunc(GL_GREATER);

// Render only back faces.
glEnable(GL_CULL_FACE);
glCullFace(GL_FRONT);

// Clear the buffers.
// Note that the most distant shape will overwrite all others.
glClearDepth(0.0);
glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);

//... Render the shape of the volume, i.e. a cloud.

// Enable depth test, back to front.
glDepthMask(GL_TRUE);
glEnable(GL_DEPTH_TEST);
glDepthFunc(GL_LESS);

// Render only front faces.
glEnable(GL_CULL_FACE);
glCullFace(GL_BACK);

// Clear the buffers.
glClearDepth(1.0);
glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);

//... Render the same shape of the volume.


3. In a fragment shader calculate the distance to a rasterized point. Do it in fragment instead of in a vertex, hoping for speedup due to hardware interpolation. If two vertices spread out evenly in a view space, their depths will be equal and much bigger than the distance to a front clipping plane. This results in artifacts when moving towards a volume and crossing it.

4. A depth buffer attachment can now be shared between with other frame-buffers. It works for simple convex shapes as volumes that can be rendered from inside and outside. The real work is to be done elsewhere to modulate this shape with other detail, i.e. noise.

Saturday, May 13, 2017

Volumetric rendering with custom volume shapes


It took me a week to figure out a problem with a vertex shader not able to interpolate a single float output.
I started with a multi-pass stage that produced a volume from a shape, then fed that volume to a volumetric pass and observed unexplainable artefacts. When flying through a scene and entering a volume the buffer that was clipped by a view frustum's near plane would mangle a default buffer value (set by glClear). Attempts to fix that by playing with glEnable(GL_DEPTH_CLAMP), glDepthRange(-1.0, 1.0) etc did not help.
What did work was switching from a float output to a whole vec3, and calculating depth in a fragment shader. Inefficient, and annoying.

Friday, April 14, 2017

Real-time atmospheric effects

This demonstrates a few atmospheric effects that I've added to a custom engine. It is real-time, using stages in shader pipeline, and very customizable (fog, lighting, composition etc).
In addition, the demo is showing Fourier transform to simulate ocean waves.
More to follow.

Wednesday, May 25, 2016

On UI composition, embedded view controllers, weak vars, multi-cast delegates.

Interface Builder is great for composing plain interfaces, where each view controller is represented by a single instance of view in a view hierarchy.

Embedded view controllers are different.

Even though you can create multiple Embed segues connecting different places in a storyboard to the same embedded view controller - during app initialization the framework will create as many instances of that controller as there are segues.

This is counter-intuitive at first. A collection of views is an ordered tree, where each node has a reference to its single parent. But in a storyboard the view controllers can be connected by Embed segues arbitrarily and can form cycles.

It appears Apple gets away with this by forcing each Embed segue to create a new instance of a destination view controller.

For example, the following storyboard with embedded view controllers (part of a project in a coursera.org course) has many parent-child Embed relations:




Note the three parent Container views, and two of them share a view controller.

A parent view controller will have outlets to instances of child view controllers. Those outlets are declared weak and fished out of the storyboard in an override of prepareForSegue() method. If there are multiple instances of the same embedded view controller (as in the screenshot) - you’d want them stored in an array of weak references. But the challenge with Swift 2 is that it does not support collections of weak references.

In a simple case - having separate weak refs is ok. The prepareForSegue() implementation might look like this:

    // MARK: - Navigation

    override func prepareForSegue(segue: UIStoryboardSegue, sender: AnyObject?) {

        // !!! Embed segues are processed before viewDidLoad().
        switch segue.destinationViewController {
        case let vc as MyHobbiesCollectionViewController:
            // Note that this embedded VC has multiple instances,
            // need additional id check (i.e. by segue.identifier).
            switch segue.identifier {
                case .Some("My1stHobbiesEmbedSegue"):
                    my1stHobbiesCVC = vc;
            default:
                my2ndHobbiesCVC = vc;
            }
           
            // Also prepare non-embed segues.
        case let vc as EditHobbiesViewController:
            break;
        default: break;
        }

    }


In other more advanced cases, a recommended solution is to declare a weak ref in a structure and then have an array of those structures. Hopefully Apple is working to fix this minor annoyance.

A few words about how a hierarchy of view controllers should communicate with the rest of the tree.

KVO and NSNotificationCenter are ok in general case, but they lack type-safety of messages, and require not very nice symmetry of subscrbe/unsubscribe calls. Instead, I prefer declaring protocols for each child view controller, then a parent would implement those protocols, and assign itself to weak ref delegates of its children.

Events of the child controllers will first travel up a hierarchy, where a parent will route them to peers.

This messaging system could make use of a collection of weak ref delegates, which would be referred to as multi-cast delegates. But this will have to wait till a version of Swift that implements weak ref collections natively.

Tuesday, May 10, 2016

Quick Auto Layout rig for a responsive tab bar control

Let’s say I need to create a tab bar with 3 tabs, each having a button and an indicator. It needs to be responsive (autoresizeable) and each tab equally sized with the others.





Using Auto Layout here is how I do it.

1. Select a view that renders a tab (here called “TabView”). Copy-paste as many as needed. Customize each copy.





2. Add missing constraints manually. Select each tab view, uncheck “Constrain to margins”, and add space constraints. The first tab needs 4 constraints, the rest - only 3.







3. Add size constraints (width or height) by selecting all tab views, then Equal Widths.







4. Finally, update all frames in a parent view.





Embed in a parent controller, declare protocols to handle taps, etc.

Saturday, February 20, 2016

Slide-out menu (with hamburger).

This short piece is about interactive slide-out menu in iOS. The demo is using an embedded container view for a menu view controller, and that controller handles animations and interaction.

The following implementation will be used for interaction-animation control where view hierarchy remains mostly static.




1. Menu view consists of two container views - a full screen dim view and an overlapping menu view. Interaction is registered by a pan gesture recognizer and is converted to a normalized “progress” property value. Setting this property on a menu view controller updates both the alpha property of the dim view and an origin of a menu view (via a horizontal leading constraint). 


import UIKit


class RootViewController: UIViewController, UIGestureRecognizerDelegate{

    @IBOutlet weak var menuContainerView: UIView!
    @IBOutlet weak var mainContainerView: UIView!
    
    weak var menuVC: MenuViewController!;

    override func viewDidLoad() {
        super.viewDidLoad()

        // Do any additional setup after loading the view.
        
        // Instead of fishing for an embedded controller in a prepareForSegue.
        for vc in childViewControllers {
            if let mvc = vc as? MenuViewController {
                menuVC = mvc;
            }
        }
    }

    
    // MARK: - Gesture Recognizer
    
    var initialProgress: Float = 0.0;
    var ignoreGesture = false;

    @IBAction func handlePanGesture(recognizer: UIPanGestureRecognizer) {
        let vel = recognizer.velocityInView(view);
        let velThreshold = menuVC.velocityThreshold;

        switch(recognizer.state) {
        case .Began:
            ignoreGesture = abs(vel.y) > abs(vel.x);
            if (ignoreGesture) {
                return;
            }
            initialProgress = menuVC.progress;
            recognizer.setTranslation(CGPointZero, inView: view)
            
        case .Changed:
            if (ignoreGesture) {
                return;
            }
            let translation = recognizer.translationInView(view);
            let tX = Float(translation.x);
            let progress = initialProgress + tX / velThreshold;
            // The following can be used to prevent menu progress overshot (not a good "responsive UI").
            // progress = min(1.0, max(0.0, progress));
            menuVC.progress = progress;
            
        case .Ended:
            if (ignoreGesture) {
                return;
            }
            let velX = Float(vel.x);
            let expand = ((menuVC.progress > 0.5 && velX > -velThreshold) || (menuVC.progress < 0.5 && velX > velThreshold));
            menuVC.animateMenu(expand);
            
        default:
            break
        }
    }
    
}



2. Progress property setter does a few things differently if the value is to be animated. Because menu works in conjunction with the embedding superview, the progress will enable/disable the superview as needed so that collapsed menu stays out of way when user interacts with the main content. To animate this property from its current value - there is a function animateMenu:.


import UIKit

class MenuViewController: UIViewController {
    @IBOutlet weak var dimView: UIView!
    @IBOutlet weak var menuView: UIView!
    @IBOutlet weak var menuLeadingConstraint: NSLayoutConstraint!

    var _useAnimation = false;
    var _progress: Float = 0.0;
    
    // To animate to a value set the _useAnimation true.
    var progress: Float {
        get { return _progress; }
        set {
            if (abs(newValue) > 1.0) {
                _progress = min(1.5, pow(abs(newValue), 0.2) * (newValue >= 0.0 ? 1.0 : -1.0));
            }
            else {
                _progress = newValue;
            }
            if (newValue > 1e-5 && !menuEnabled) {
                // Don't just set to condition via assignment, "false" will break animation (0.0 will collapse the menu before animation begins).
                menuEnabled = true;
            }
            
            let animations = {
                self.dimView.alpha = CGFloat(min(1.0, self._progress) / 2);
                let menuOriginX = CGFloat(self.velocityThreshold * self._progress) - self.menuView.frame.width;
                self.menuLeadingConstraint.constant = menuOriginX;
                self.view.layoutIfNeeded();
            };
            let completion: (Bool -> Void) = { finished in
                if (newValue <= 1e-5) {
                    self.menuEnabled = false;
                }
                self._useAnimation = false;
            };

            if (_useAnimation) {
                UIView.animateWithDuration(0.25, delay: 0.0, options: .CurveEaseOut, animations: animations, completion: completion);
            } else {
                animations();
                completion(true);
            }
        }
    }
    
    private(set) var velocityThreshold: Float = 100.0;
    
    // Enables/disables container view.
    var menuEnabled: Bool {
        get {
            return !view.superview!.hidden;
        }
        set {
            view.superview!.hidden = !newValue;
        }
    }

    func animateMenu(expand: Bool) {
        _useAnimation = true;
        
        if (expand) {
            progress = 1;
        } else {
            progress = 0;
        }
    }
    
    
    override func viewDidLoad() {
        super.viewDidLoad()

        // Do any additional setup after loading the view.
        
        // Initial menu view offset is a design-time position for an "expanded" state.
        velocityThreshold = Float(menuView.frame.width + menuView.frame.origin.x);
    }
    
    override func viewDidAppear(animated: Bool) {
        menuEnabled = true;
        progress = 1.0;
        animateMenu(false);
    }

    @IBAction func tapDimView(sender: AnyObject) {
        animateMenu(false);
    }
}




Other ways to implement this, such as using UIScrollView or custom transition animators, are more appropriate if standard transitions (on navigation controller) are desired. I find them either lacking in brevity of code or having tight dependence on standard view controllers.