blog tags:


I'm Dmitry Popov,
lead developer and director of Infognition.

Known in the interwebs as Dee Mon since 1997. You could see me as thedeemon on reddit or LiveJournal.

Articles Technology Blog News Company
Older posts

A video codec in JavaScript
November 25, 2017

Recently we released a pure JavaScript implementation of in-browser player for our lossless video codec ScreenPressor. It supersedes our old player made with Flash. Here's some anecdotal evidence of how fast current JavaScript interpreters in browsers are. The chart shows Frames Per Second:

Here we took a screen recording with resolution 1364x768 - 1M pixels, 4MB RGB32 data per frame - and measured how fast our players could decode 480 consequent delta-frames when seeking from 0 (a key frame) to frame 490 (while the next key frame is the 500th). When the player is at position 0 it decodes and buffers a few frames in advance, in this case 10, so when seeking then to frame 490 the number of frames to decode is 480. You can see the test video here (btw this sample shows ~300x lossless compression). "v2" and "v3" mean codec versions: we tested implementations of both 2nd and 3rd versions of ScreenPressor, they use different entropy coders and have other differences.

Of course the speed depends a lot on the video content, how complex and diverse it is, so on different videos numbers will be different. In typical screen videos there are not so many changes from frame to frame, most of the frame can be copied from previous one, this includes scrolling and window movement scenes. But anyway it's quite impressive how a JS-based player can produce 150-200 4MB frames a second.

In this case we can see how different JS speed is in different modern browsers. Firefox Quantum is incredibly fast and smooth when it comes to browsing sites but its JS engine in this test is slower than of MS Internet Explorer and Edge, while Chrome is king when it comes to JS speed.

All JavaScript engines here worked faster than Flash, this probably shows how they evolved since our previous test three years ago. However native code is still 2-4 times faster than JS, thanks to well optimizing vectorizing compiler, more compact data representation and no dynamic types with their runtime costs.

More information about both (JS and Flash) web players for ScreenPressor and more samples here.



Error-checking smart pointer
April 27, 2016

As you might know, in the world of Windows programming COM is still actively used, alive and kicking after all these years. It must be usable from plain C so for error handling they use return codes: most functions return HRESULT, an integer that is zero if all went well, negative in case of error and positive if it went "okay but...". Return codes should not be ignored or forgotten, they need to be handled, and in practice it is often the case that all local handling logic reduces to just stopping what we were doing and passing the error code up the stack, to the caller. In languages like Go it looks like this:

err := obj1.Action1(a, b)
if err != nil {
    return err
err := obj1.Action2(c, d)
if err != nil {
    return err
err := obj2.Action3(e)
if err != nil {
    return err

etc. Each meaningful line turns into four, in the best tradition of



Static introspection for message passing
April 23, 2016

There is a famous in PLT community paper on free theorems, giving praise to generics and parametricity allowing you to infer a lot about some generic function behavior just from its type, because since such generic function does not know anything about the type of value it gets, it cannot really do much with this value: mostly just pass it around, store somewhere or throw away. On the other end of generics spectrum (call it "paid theorems" if you will) we have languages like D where a generic function can ask different questions about the type it gets as argument and depending on the answers, i.e. properties of the passed type, do one thing or another. I think this end is more interesting and practical.



Dependent types in half of D
July 30, 2015

Recently I saw a talk by one of active Idris contributors: David Christiansen - Coding for Types: The Universe Pattern in Idris where he shows some uses of dependent types. I'd like to use some of those examples to demonstrate that we actually have dependent types in D, or, really, in part of it. First let me show some original examples in Idris, a full blown depentently typed language, and then we'll translate them to D closely.

In the talk David says that in dependently typed languages we have types as first class values: you can create them on the fly, return from functions, store in data structures and pass around, however you shouldn't pattern-match on them because if we allow pattern-matching on types we lose free theorems (id may start to return 42 for all natural arguments) and this prevents proper type erasure when compiling the programs. I'm not sure if matching on types is actually banned in Idris now, I remember a few versions ago it actually worked, at least in simple cases. But sometimes doing different things depending on type passed is what we do want, and for this cases David describes what he calls The Universe Pattern: you take the types you wish to match on, encode them in a simple data type (as in AST), match on this data type, and when you need actual types use a mapping function from this encoding to actual types.



From native code to browser: Flash, Haxe, Dart or asm.js?
November 17, 2014

If you developed your own video codec and wanted to watch the video in a browser what would you do? That is a question we faced a few years ago with ScreenPressor and at that time the answer was Flash. It was cross-platform, cross-browser, widely available and pretty fast if you use the right programming language, i.e. Haxe instead of ActionScript. So we implemented a decoder (and a player) in Flash back then. But now Flash is clearly on decline, supported only on desktop, meanwhile different browsers in a race for speed made JavaScript significantly faster than it was a few years ago. So I decided it's time to check how JS can compare to Flash when it comes to computation-intensive task such as a video codec. I really don't like the idea of writing in pure JavaScript 'cause the language is such a mess, I'd rather use something that at least gets closures, objects and modules right and has some static type checking. Since I already had working code in C++ for native code and in Haxe for Flash, obvious choices were using Emscripten to generate asm.js from C++ code and retargeting Haxe code to JS (just another target for Haxe compiler). Also, Dart is pretty close as a language (porting to Dart is simpler than rewriting in some Haskell or Lisp clones) and Dart VM is marketed as a faster and better replacement for JS engines, so I was curious to try it.

To test and compare different languages and compilers I decided to implement in them a small part of the codec, the most CPU consuming one: decompression of a key frame to RGB24. I'm going to show the results first and then follow with some notes on each language.

Here are the times for decompressing one particular 960x540 px frame on a laptop with a 2.4 GHz Core i3 CPU and Windows 8.1:



Optimizations in a multicore pipeline
November 6, 2014

This is a story that happened during the development of Video Enhancer a few minor versions ago. It is a video processing application that, when doing its work, shows two images: "before" and "after", i.e. part of original video frame and the same part after processing.

It uses DirectShow and has a graph where vertices (called filters) are things like file reader, audio/video streams splitter, decoders, encoders, a muxer, a file writer and a number of processing filters, and the graph edges are data streams. What usually happens is: a reader reads the source video file, splitter splits it in two streams (audio and video) and splits them by frames, decoder turns compressed frames into raw bitmaps, a part of bitmap is drawn on screen (the "before"), then processing filters turn them into stream of different bitmaps (in this case our Super Resolution filter increases resolution, making each frame bigger), then a part of processed frame is displayed on screen (the "after"), encoder compresses the frame and sends to AVI writer that collects frames from both video and audio streams and writes to an output AVI file.

Doing it in this order sequentially is not very effective because now we usually have multiple CPU cores and it would be better to use them all. In order to do it special Parallelizer filters were added to the filter chain. Such filter receives a frame, puts it into a queue and immediately returns. In another thread it takes frames from this queue and feeds them to downstream filter. In effect, as soon as the decoder decompressed a frame and gave it to parallelizer it can immediately start decoding the next frame, and the just decoded frame will be processed in parallel. Similarly, as soon as a frame is processed the processing filter can immediately start working on the next frame and the just processed frame will be encoded and written in parallel, on another core. A pipeline!

At some point I noticed this pipeline didn't work as well on my dual core laptop as on quad core desktop, so I decided to look closer what happens, when and where any unnecessary delays may be. I added some logging to different parts of the pipeline and, since in text form they weren't too clear, made a converter into SVG, gaining some interesting pictures.



The real problem with GC in D
September 29, 2014

About a year ago one blog post titled "Why mobile web apps are slow" made a lot of fuss in the internets. In the section about garbage collection it showed one interesting chart taken from a 2005 paper "Quantifying the Performance of Garbage Collection vs. Explicit Memory Management":

The intent was to frighten people and show how bad GCs were: they allegedly make your program super slow unless you give them 5-8x more memory than strictly necessary. You may notice that the author has chosen a chart for one particular program from the test set and you can rightfully guess that most other programs didn't behave that bad and average case picture is much less frightening.

However what I immediately saw in this graph is something different and probably more important.



Why D?
September 19, 2014

If you look at stuff on this site you'll notice a trend: newer apps tend to be written in D programming language. There's a photo editing app, disk space visualizer and analyzer, a tool for automatic video processing, a tool for reading DirectShow graph files as text, all in D (and all free and open source btw). We also have many small internal tools written in D, even this very blog is powered by a hundred lines of D now. One of our main products, a video processing app called Video Enhancer, is currently written in C++ but the next major version is going to be in D for most parts except for the heavy number crunching bit where we need the extremes of performance only reachable by C++ compilers today. So obviously I must like D, and here are my personal reasons for doing it:



Recursive algebraic types in D
September 15, 2014

One of the things D programming language seemingly lacks is support for algebraic data types and pattern matching on them. This is a very convenient kind of types which most functional languages have built-in (as well as modern imperative ones like Rust or Haxe). There were some attempts at making algebraic types on the library level in D (such as Algebraic template in std.variant module of D's stdlib) however they totally failed their name: they don't support recursion and hence are not algrabraic at all, they are sum types at best. What follows is a brief explanation of the topic and a proof of concept implementation of recursive algebraic types in D.



On Robin Hood hashing
September 13, 2014

Recently I played a bit with Robin Hood hashing, a variant of open addressing hash table. Open addressing means we store all data (key-value pairs) in one big array, no linked lists involved. To find a value by key we calculate hash for the key, map it onto the array index range, and from this initial position we start to wander. In the simplest case called Linear Probing we just walk right from the initial position until we find our key or an empty slot which means this key is absent from the table. If we want to add a new key-value pair we find the initial position similarly and walk right until an empty slot is found, here the key-value pair gets stored. Robin Hood hashing is usually described as follows. It's like linear probing, to add a key we walk right from its initial position (determined by key's hash) but during the walk we look at hash values of the keys stored in the table and if we find a "rich guy", a key whose Distance from Initial Bucket (DIB) a.k.a. Path from Starting Location (PSL) a.k.a. Probe Count is less than the distance we've just walked searching, then we swap our new key-value pair with this rich guy and continue the process of walking right with the intention to find a place for this guy we just robbed, using the same technique. This approach decreases number of "poors", i.e. keys living too far from their initial positions and hence requiring a lot of time to find them.

When our array is large and keys are scarce, and a good hashing function distributed them evenly so that each key lives where its hash suggests (i.e. at its initial position) then everything is trivial and there's no difference between linear probing and Robin Hood. Interesting things start to happen when several keys have the same initial position (due to hash collisions), and this starts to happen very quickly - see the Birthday Problem. Clusters start to appear - long runs of consecutive non-empty array slots. For the sake of simplicity let's take somewhat exaggerated and unrealistic case that however allows to get a good intuition of what's happening. Imagine that when mapping hashes to array index space 200 keys landed onto [0..100] interval while none landed onto [100..200]. Then in the case of linear probing it will look like this:



Older posts