QML sandbox running in Firefox

Experimental Qt and QML in the browser

I was amazed when I read Morten Johan Sørvig’s blog post about bringing Qt to Native Client (NaCl) and Emscripten in September last year. I have been following multiple efforts to bring Qt to the web for some time. Mostly because I wanted to make my own Qt projects easily available to others, but also because I find the whole concept of running C++ code in a browser fascinating.

This is definitely not the first time someone has worked on bringing Qt to the browser. Qt4 was ported to Emscripten by Simon St James and published with a lot of examples. However, it seems the port was discontinued when Qt5 came along due to the changes that came with that transition. In the meantime, Lauri Paimen, Anton Kreuzkamp and many others have contributed to qmlweb, a QML interpreter written from scratch in JavaScript. That project does however not make it easy to use any C++ classes you may have defined. And most of my projects are QML/C++ hybrids.

These projects brought something important to the table and took us one step closer to writing QML for the browser. Personally I find QML very enjoyable because of its clean and intuitive structure in comparison to other options like XML. And the ability to easily bind GUI elements to high-performance C++ code has made Qt my favorite framework for many projects.

I tried to port Qt5 to Emscripten myself about a year ago. That turned out to be a real challenge. I based my work on the changes Simon St James made to Qt4 and tried to figure out how to make the same changes in Qt5. After a couple of weeks, I got to the point where I had QtCore running and QtDeclarative (QML) compiling, but other things caught up with me and I had to leave that project behind. The biggest disappointment was that I never got around to making any demos.

I decided to pick up the ball a couple of months ago This time I knew about Morten Johan Sørvig’s efforts and asked him to give me some info on the status of his port. At that point he had support for compiling to NaCl ready (that means Chrome only) and he had already started porting to Emscripten using pepper.js as glue between Emscripten and NaCl. There was still a bit left to do before any examples could be up and running in Firefox. I decided to give it a go and see if I could contribute with something.

A couple of weeks later, I managed to get Qt running in Firefox with a number of QML examples. The process was a lot of fun as I learned much about what happens inside Qt’s sources. It was interesting to see how often I could write many lines of code and rebuild multiple times only to arrive at a one-line change that fixed the issue at hand. I was really happy when I finally could send a patch to Qt’s NaCl branch. The main changes were to Qt’s event loop, which would hang if control wasn’t returned to the browser. I had to add a couple of hacks to keep Qt’s mutexes from locking up everything (this needs to be properly fixed later, perhaps by creating pthread stubs). And OpenGL needed to be explicitly set to OpenGL ES2 in addition to a bunch of smaller modifications.

QML sandbox running in Firefox
QML sandbox running in Firefox. See the links below for a live demo.

Below are two resulting demos of QML running in the browser. I’m amazed with the performance both in Firefox and Chrome. Just be aware that the size of the examples is quite big because no optimizations have been done to reduce the size of the binary JavaScript. Further, there may be lots of bugs and some browsers where the examples don’t work. Remember that this is still very experimental and only a small taste of what may be possible with Qt in the browser in the future:

What amazes me the most is the fact that we are now running Qt’s own JavaScript engine, which is written in C++ and compiled to JavaScript with Emscripten, inside the browser’s own JavaScript engine! I really want to see if we can get QML’s WebEngineView running inside itself inside Firefox. That’s a step on the way to true inception, which reminds me of a great talk by Gary Bernhardt titled “The Birth & Death of JavaScript”.

Using Blender and make-na to create DNA art

In CINPLA we are developing teaching material for a new course in biology and programming. One topic we’re working on is that of bioinformatics and DNA. I figured that an artistic rendering of the DNA double helix would be nice to have and decided to try and make one in Blender. At first I considered building this from the ground up using its design tools, but I quickly figured that it would be too time consuming and hard to get the proportions right. After all, I’d like it to be as close to the real structure of DNA as possible.

I then remembered that the RCSB Protein Data Bank (PDB) hosts a large number of molecular structures that have been described by researchers over the past decades. The structures can be downloaded freely as long as they are cited properly. I went looking for files with the DNA molecule and found a couple, such as 1BNA and 1D66, but they turned out to be a bit too short or too hard to untangle from the other molecules.

Then I found make-na, a fantastically simple open source tool that will generate a PDB file from any DNA sequence. Just open the make-na server and type in your favorite combination of the letters A, T, C and G. Out comes a PDB file with the corresponding molecular structure of DNA.

To open the PDB file in Blender, click File > User Preferences > Add-ons and search for PDB. Enable this add-on by clicking the check box next to it:

Screenshot from 2016-02-14 17-40-24(You may have to restart Blender after enabling the add-on).

Click File > Import > Protein Data Bank (.pdb):

Screenshot from 2016-02-14 17-43-40Voila! You should now have a DNA molecule ready for rendering in Blender.

I tweaked the material settings, enabled Cycles Render and added a couple of lights to make the image you see below:

duplex14Here is the .blend file used to create the image, if you’d like to modify it to make your own version:

For more technical figures, I’m thinking about using the cartoon style in JSmol or PyMOL which produces ribbon diagrams of proteins. Those make the DNA strands look more like this:


The next step is to load this image into Inkscape and start annotating the different parts of DNA.

The above images and the .blend file are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

An upcoming C++ library for HDF5 files

I recently started working on a new C++ library for reading and writing HDF5 files. I got the idea when I was working on a few files in Python and C++ at the same time. The Python library h5py is just way more comfortable than the HDF5 C++ API. But with the new features in C++11 and C++14, I figured it should be possible to make a C++ library that is just as easy to use as h5py. And I must admit that I believe I’m on the way to making it even easier.

One design goal I have set for this project is to make the library forgiving. This means that I assume you know what you are doing and try to make it possible, even though there might be some side effects.

One example of such a side effect is that of creating a dataset in a HDF5 file and later changing its size. Because of limitations in the HDF5 standard, the original dataset will still take up space in the file, although it’s no longer in use. I believe the HDF Group designed it this way because it is hard to free already allocated space in a file without truncating it first. So they optimized for performance rather than flexibility, which is a fair choice to make. However, I assume you want flexibility and that you’d rather see that extra space taken up than seeing your program crash for trying to reuse a dataset. So this code is perfectly fine within my library:

#include <armadillo>
#include <elegant/hdf5>

using namespace elegant::hdf5;
using namespace arma;

int main() {
    File myFile("myfile.h5");
    myFile["my_dataset"] = zeros(10, 15);
    myFile["my_dataset"] = zeros(20, 25);
    return 0;

This opens the myfile.h5 for reading and writing and sets my_dataset to a 10×15 matrix of zeros before resetting it to a 20×25 matrix. This will, however, leave space taken up by the 10×15 matrix used in the file, even though it’s no longer accessible. Oh, and did I mention that the Armadillo library is already supported?

On the contrary, if you try to do the same in h5py, you will get a RuntimeError:

from h5py import *
from pylab import *
my_file = File("myfile.h5")
my_file["my_dataset"] = zeros((10, 15))
my_file["my_dataset"] = zeros((20, 25))
# output:
RuntimeError: Unable to create link (Name already exists)

As you can see, the syntax is pretty much the same, but the C++ library will be slightly more forgiving than h5py. That, in addition to a range of other nice features, is what I hope will make this new HDF5 C++ library attractive.

An experimental version will hopefully be released soon.

Getting started with unit tests in Qt Creator and Catch

I have written about unit testing in Qt Creator on multiple occasions earlier. Since then, a new testing framework seems to have become a new leader in the field, and that is Catch. It is similar to UnitTest++ and Google’s gtest and brings the best of both worlds. It is header-only and appears to reduce the amount of code that must be written in comparison to the other two.

This is an adopted version my previous posts that shows how to use Catch in a Qt Creator project. It doesn’t any longer integrate the testing output into Qt Creator’s issue pane. Instead I’ve started to rely on Travis CI or Jenkins to run the tests and notify me about any newly introduced errors.

An example project using this code structure has been posted on Github by Filip Sund (thanks, Filip!) and further adapted by us with the move to Catch.

The structure of the project is as follows:

├─ MyProject.pro
├─ defaults.pri
├─ app/
│  ├─ app.pro
│  └─ main.cpp
├─ src/
│  ├─ src.pro
│  └─ myclass.cpp
└─ tests/
   ├─ tests.pro
   └─ main.cpp

The main project file, MyProject.pro will now be based on a subdirs template, and may look like this:

TEMPLATE = subdirs
    src \
    app \
app.depends = src
tests.depends = src

The app.depends and tests.depends statements makes sure that the src project is compiled before the application and tests. This is because the src directory contains the library that will be used by both the app and the tests.

CONFIG+=ordered is needed because depends doesn’t affect the order during make install.


Each of the other .pro files will include defaults.pri to have all the headers available. defaults.pri contains the following:


If the library, main program and tests use common libraries, it is very useful to have the defaults.pri define these dependencies too.


In the src folder, I have myclass.cpp, which is the class that I want to use and test. The src.pro needs to compile to a library, so that it may be used both by app and tests, and could look something like this:

CONFIG -= qt

TARGET = myapp

SOURCES += myclass.cpp
HEADERS += myclass.h

What this class does is not so interesting, it could be anything. A simple example could be this header file:

#ifndef MYCLASS_H
#define MYCLASS_H

class MyClass {
    double addition(double a, double b);

#endif // MYCLASS_H

With this accompaning source file:

#include "myclass.h"

double MyClass::addition(double a, double b) {
    return a * b;


I only have a main.cpp file in app, because the app is basically just something that uses everything in the src folder. It will depend on the shared compiled library from src, and app.pro would look something like this:


CONFIG += console
CONFIG -= app_bundle
CONFIG -= qt

SOURCES += main.cpp

LIBS += -L../src -lmyapp

The main.cpp file could be a simple program that uses MyClass:

#include <myclass.h>
#include <iostream>

using namespace std;

int main()
    MyClass adder;
    cout << adder.addition(10, 20) << endl;
    return 0;


In the tests folder I have simply added a main.cpp which will run the tests. Then tests.pro has the following contents:


CONFIG   += console
CONFIG   -= app_bundle
CONFIG   -= qt

SOURCES += main.cpp

LIBS += -L../src -lmyapp

Which now links to the myapp library in addition to the unit tests.

The main.cpp in tests file which could contain the following, if we were to use UnitTest++ as our testing library:

#define CATCH_CONFIG_MAIN  // This tells Catch to provide a main() - only do this in one cpp file
#include "catch.hpp"
#include <myclass.h>

TEST_CASE( "MyMath", "[mymath]" ) {
    SECTION("Addition") {
        MyClass my;
        REQUIRE(my.addition(3,4) == 7);

This test will fail because my implementation of MyClass::addition is completely wrong:

class MyClass {
    double addition(double a, double b) {
        return a * b;

Note that I’m including MyClass by through <myclass.h> which is possible because of the INCLUDEPATH variable in defaults.pri.

This should help you define a project that compiles a library, as well as tests and an application using the library.

Why I won’t be starting my next project in Rust

I have been inspired to learn Rust and Julia lately. The idea was to use these promising languages in all of my new projects. I wanted to learn Rust because it is a safe and fast replacement for C++. I wanted to learn Julia because it is a language tailored for scientific use that might someday replace Python.

I quickly realized that I would be learning both for the wrong reason. I already know C++ and Python well, and should be starting important new projects in these languages.

One article that changed my mind was “Criticizing the Rust Language, and Why C/C++ Will Never Die” by Eax Melanhovich. Not all the points Melanhovich makes are longer valid. For instance, the speed of Rust has improved significantly over just the past months. However, he is right in that the number and quality of development tools for Rust is way lower than that of C++. And it is hard to argue with the number of vacant positions for  C++ programmers in comparison to those of Rust and Julia.

This does of course not say anything about the quality of these languages. For all I know, Rust may be (or become) a much safer, faster and elegant language than C++. But I understand that many of the benefits I would get from Rust are already available in the new C++11 and C++14 standards. And from using C++ in a modern way.

I found Rust attractive because of its promise of memory safety. It is designed in a way so that you won’t be able to shoot yourself in the foot as easily as you can in C++. However, I’m currently using pointers in C++ the old-fashioned way. I still work with raw pointers, new and delete when I should be using smart pointers. Sometimes I even use pointers when references and values would be the right choice.

I realize now that I need to start a project or two where I use only modern C++. That should hopefully teach me how to steer clear of those nagging segfaults sometime in the future. I’ll be reporting back here about my experiences and will make a list of some recommendations for others who are trying to do the same.

Today I called Python from QML

I always use QML with Qt Quick for GUI programming. It’s incredible both for prototyping and larger applications. I find it easy to express myself in QML, because it is so flexible. It’s declarative. You can bind a button’s position to the value of a slider in just one line. Not that you’d ever want to do that, but it shows how easy it is to connect objects together. I really wish the web was written in QML and not HTML.

Typically, I’m working with both QML and C++. I write high performance and visualization code in C++ and define GUI elements in QML. But today I wanted to use QML and Python together because I’m working on an experiment browser for our lab. Not an experimental browser, but an application that lists all experiments and makes it easy to export data for further analysis.

We have decided to organize the experiments as HDF5 files. To read these files I want to use the Python HDF5 package. I could use the C++ HDF5 library, but using Python should hopefully make the application more easily maintainable in the future, also for non-C++ coders in our lab. To do this, I figured there were two possibilities: PyQt and PyOtherSide.

In brief (and a bit simplified), PyQt calls Qt code from Python, while PyOtherSide calls Python code from QML. The difference isn’t really that big, so it just boils down to where the business logic resides. I figured that PyOtherSide would be the better option for us, because it allows everyone to help out with the Python code without learning anything about Qt. PyQt would on the other hand require everyone to have at least some understanding of the Qt framework to make changes in the code. More of the business logic will have to take place in QML, though.

PyOtherSide is really simple in use. You just define a Python object in QML and this loads modules and files from the Qt resource file (qrc). Once loaded, Python functions can be called from QML and their results are automatically converted from Python types to Qt types:

import io.thp.pyotherside 1.3

Python {
    property bool ready: false
    function loadData() {
        if(!ready) {
        call("hdf5_loader.read_experiments", [], parseData)

    function parseData(result) {
        for(var i in result) {
            var element = result[i]
    Component.onCompleted: {
        importModule("hdf5_loader", function() {
            ready = true

As you can see, all Python calls are asynchronous. You can also make synchronous Python calls, but this is not recommended because it could cause the QML GUI to stall while waiting for the results.

The Python code in this case is just a simple function that returns a dictionary:

def read_experiments():
    return {{"experimenter": "test", 
             "date": "2015-10-26"}, 
            {"experimenter": "test2", 
             "date": "2015-10-25"}}

Now I just need to make this code read and parse real HDF5 files.

Straight from the source: NEURON’s incredible backwards compatibility

NEURON is a neural simulator, but it’s not just another neural network application. NEURON has been around much longer than your fancy deep learning software. It was first released in 1994, although some of it’s early code and concepts seem to stem from the 70’s and 80’s. In fact, NEURON is not a machine learning library at all. It simulates real neurons. It uses real morphologies and ion channel densities to reproduce experiments. Some labs even use machine learning algorithms to fit the parameters of their NEURON models. NEURON is one of the workhorses of the computational branch of the Human Brain Project and the core engine in the Blue Brain Project.

But the age of NEURON is starting to show in its source code. Because some things are a bit awkward when working with this simulator I decided to peek at its sources. Just to see what this old engine looks like on the inside. And it turned out to be an interesting journey. Deep inside NEURON’s plotting library, I found this:

#if VT125
case VT:
vtplot(mode, x, y);

Now, I didn’t know what VT125 was when I first saw this, but a quick search on the web reminded me that I’m still a young software developer. I present to you, the VT100:

VT100 terminal
I couldn’t find a good picture of the VT125. I apologize if it looks way more modern than its older cousin.

Now that’s what I call backwards compatibility!

Some of you may want to lecture me about how modern terminals actually emulate VT100s and that this code might be in there to allow some fancy plotting inside xterm or its siblings. But I would argue that there are other parts of this code that give away its impressive attempt at supporting older platforms. Such as the rest of the above switch statement:

		switch (graphdev)
		case SSUN:
			hoc_sunplot(&amp;amp;amp;amp;text, mode, x, y);
#if NRNOC_X11
#if NeXTstep
		case NX:
#if TEK
		case ADM:
		case SEL:
		case TEK4014:
			tplot(mode, x, y);
#if VT125
		case VT:
			vtplot(mode, x, y);

Now, we all love nested preprocessor if-statements inside switch blocks, but let’s look aside from that and at what’s supported here.

There’s the NRNOC_X11, which I believe introduces the only part of this block that might actually be executed nowadays. In addition we have SUNCORE, which might be Solaris, but I’d bet this supports something that existed long before Oracle acquired Sun back in 2010. There’s TEK, which may refer to something like the Tektronix 4010:

This baby was made in 1972, but it still runs NEURON’s plotting functions. Just download the 5 MB zip-file of NEURON’s source code and … oh, wait.

And then there’s NeXTstep, which Apple acquired in 1997 and used to replace its own Mac OS with Mac OS X. Considering that NeXTstep made its last official release in 1995, I think its fair to say that this piece of code was written in the previous century. It could, of course, just be that no one has deared to search-replace NeXTstep with OS X, but I doubt it.

I should finish this with a final caveat: It could be that this code isn’t at all in use by NEURON anymore. After all, I found all of the above in the plot.c file in the src/oc folder of NEURON’s source code. The above could be remainders of NEURON’s interpreter language, hoc, which was originally made as a demonstration of how to use the parser generator Yacc in the The Unix Programming Environment book. As far as I know, NEURON is the only software out there that’s still using hoc, but that’s a story for another day.

Image Credits: Wikipedia User:ClickRick, Wikipedia User:Rees11.

Hva er nevrale nettverk? (Del 1)

Kunstige nevrale nettverk er en etterligning av nervecellene våre i hjernen. De skal brukes til å gjøre datamaskiner like flinke til å lære som mennesker og dyr. Dette gjøres ved å herme etter hvordan hjernen er bygd opp.

Nevrale nettverk er noe som har blitt forsket på helt siden 1940-tallet, men det er først de siste årene at bruken av disse har skutt fart. Årsaken er en kombinasjon av økt datakraft, behovet for å analysere enorme mengder informasjon på internett og at mer avanserte nettverk har blitt oppfunnet. Senest i fjor annonserte Facebook at de ved hjelp av nevrale nettverk hadde klart å lage et dataprogram som var bedre til å gjenkjenne ansikt i bilder enn et gjennomsnittlig menneske.

Hjernen består av mange forskjellige celler. De mest spennende er nervecellene som kan lede elektriske signaler over store avstander. Når signalene skal overføres mellom nervecellene dusjer de hverandre med kjemikalier som igjen danner elektriske signaler i mottakercellene.

I synapsene dusjer nervecellene hverandre med kjemikalier for å overføre signaler.

Det er kommunikasjonen mellom nervecellene vi forsøker å etterligne når vi lager kunstige nevrale nettverk. Det kan gjøres på mange forskjellige måter, avhengig av hvor mye informasjon vi er ute etter. I forskningen jeg gjør med kollegaene mine forsøker vi å få med veldig mange av detaljene i kommunikasjonen mellom nerveceller. Slik som hvilke ionekanaler som er avgjørende for å få overført de rette signalene. Og hva formen på nervecellene har å si for hvordan de fungerer. Når Facebook, Google og Microsoft prøver å skape kunstig intelligens er disse tingene derimot ikke fullt så viktige. Da er det eneste målet at nettverket som helhet gjør den samme jobben som nervecellene i hjernen. Nemlig at de er i stand til å lære – og huske det de har lært.Det betyr at kunstige nevrale nettverk kan være svært enkle i forhold til de vi finner i våre egne hjerner. En fordel med enkle nettverk er at datamaskinen kan gjøre beregninger mye raskere enn om vi tok med alle detaljene.I de enkleste modellene av nevrale nettverk ser vi på hver nervecelle som et punkt som kan få input fra andre nerveceller og sende signaler videre. Vi forenkler med andre ord det som egentlig er en svært detaljert struktur i bildet til venstre, til en veldig enkel modell uten detaljer:

VI forenkler de ellers så detaljerte nervecellene til punkter som kan motta og sende signaler.

Hver nervecelle kan ha flere koblinger til andre nerveceller og sammen danner disse koblingene et nettverk:

Nevralt nettverk
Et nettverk av forenklede nerveceller.

Nettverket over er bare koblet sammen lagvis: Det er tre lag i dette nettverket, og hvert lag er kun koblet til det neste. I hjernen er også nervecellene inndelt lagvis, men det finnes også tilbakekoblinger mellom lagene. Ved å kun tillate lagvise koblinger er det likevel mye enklere å trene opp nettverket på en datamaskin. Men det brukes også tilbakekoblinger i kunstige nettverk, spesielt til problemer der disse har en større nytte, slik som når man skal lære ett nettverk opp til å forstå ting som endres over tid.

Hva gjør de enkelte nervecellene?

Dette er det første i en serie av innlegg jeg kommer til å skrive om nevrale nettverk. Jeg har så langt bare sagt litt om hva ideen er bak nettverkene og litt kort om veien fra detaljerte nerveceller til punktene vi bruker i våre modeller. I de neste innleggene vil jeg si litt mer om hvordan vi tenker oss at disse enkle nervecellene skal fungere og hvordan dette kan hjelpe oss med å skape en kunstig intelligens og en datamaskin som kan lære av seg selv.

Nerveceller med toon-shader

Da var det på tide med et nytt Blender-prosjekt igjen. Etter å ha lest om mulighetene til å lage tegneserieaktige materialer bestemte jeg meg for å teste dette ut på en 3D-figur av en nervecelle:


Jeg har for øyeblikket ingen anelse om hva det kan brukes til, men det er alltid gøy å teste ut ulike effekter i Blender. Kanskje dukker den opp på en fremtidig poster.

Last ned Blender prosjektet her.

Ny CINPLA-logo som helgeprosjekt

Da jeg fikk jobben som stipendiat ble jeg med i en gruppe ved navn CINPLA. Gruppen er dannet for å kombinere eksperimenter og simuleringer innen hjerneforskning.

I helgen bestemte jeg meg for å teste ut en idé til en CINPLA-logo som vi har diskutert på kontoret den siste tiden:


Tanken bak logoen er at venstre del representerer den biologiske delen av arbeidet som gjøres i gruppa. Den høyre delen representerer simuleringer. Sistnevnte er den typen arbeid jeg selv er involvert i. Logoen skal brukes på plakater og i foredrag som holdes av personer i gruppa, i tillegg til nettsider og denslags.

Logoen er laget i Inkscape, som forøvrig er et herlig verktøy til denne typen grafisk arbeid. Det er fri programvare og fungerer utmerket til vektorgrafikk i Ubuntu.

Logoen kommer i flere formater på CINPLA sin presseside snart. Den skal bare ut på prøve der den vurderes av gruppa og et par andre først.


Når jeg var ferdig kunne jeg selvfølgelig ikke dy meg, og åpnet i samme slengen 3D-programmet Blender for å se om jeg kunne gjøre noe artig med logoen der. Ved bruk av SVG-importering kunne jeg laste inn det nye bildet jeg hadde laget, men jeg ville også se hvordan det ville være med en skikkelig modell av en hjerne. En slik modell fant jeg på brainder.org, hvor A.M. Winkler har vært så grei at han har lagt ut resultatet av en MR-avbildning av en hjerne i åpne formater.

Jeg lastet denne 3D-modellen inn i Blender, plasserte ting på riktig plass og animerte kameraet slik at jeg endte opp med følgende video:

Jeg aner ikke hva vi skal bruke en slik video til, men kanskje CINPLA begynner med spillefilmer en dag. Da vil det sikkert passe fint å kjøre denne videoen som en intro.

Siden 3D-modellen har en er delt med en Creative Commons Attribution-ShareAlike 3.0 Unported License er naturlig nok denne videoen det også.