Ben Garney

Last post, we got our networking stack up and running, figured out how to work around firewalls, and saw how our codec performs over a real network link. This motivated us to revisit our compression schemes, which is what we’ll do in today’s post.

We’ll start with some algorithmic improvements to reduce the amount of data we need, then implement an entropy coding technique to reduce our data rate even further.

YUV & Chroma Subsampling

Up to now, we have skipped a key part of JPEG – conversion to YUV and chroma subsampling. DCT works on any kind of data – RGB, YUV, monochrome, etc. The DCT we have been using is for 8×8 blocks. So up to now we have been sending 4 sets of 3 blocks – one each for the red, green, and blue channels in each of the four 8×8 regions in a macroblock. This is a lot of data!

What JPEG and a lot of other formats do is take advantage of limitations in the human visual system to save bandwidth. Specifically, people are much better at detecting changes in brightness then they are in color. So you can send color data at half resolution, while you keep brightness at full resolution. To do this, you must first separate brightness from color, which is accomplished by converting to the YUV color space.

Doing this saves about 50% bandwidth with a very minimal reduction in quality (6 blocks instead of 12 blocks). It brings us down to 4 bits per pixel.

Flat Blocks

Right now the shortest block we can send is around 18 bytes (8 coefficients at 2 bytes each plus 2 more bytes to encode zeroes for the remaining 56 coefficients). So a macroblock will be at least 108 bytes plus overhead. This works out to around 0.4 bits per pixel for even a very simple macroblock.

However, during high motion frames, we need to prioritize quantity over quality. Sending fully detailed macroblocks is costly in CPU and bandwidth for pixels that are blurry and rapidly changing. Further, macroblocks that aren’t updated are very obvious in this sort of situation – so we want to prefer a full frame of updates rather than part of a frame.

To this end, we create a new macroblock encoding type called “flat.” We add a single bit at the start of the macroblock indicating if it’s DCT or flat. If flat, then the format is as follows:

 6 bits - red   color quantized by 4
 6 bits - green color quantized by 4
 6 bits - blue  color quantized by 4

We select this type when the RMS error of the flat color is less than some threshold. The threshold is based on the current total error of all macroblocks, so when we have lots of motion causing lots of error, we are more likely to encode flat macroblocks. Even a very low bandwidth link is capable of sending full frames where every macroblock is flat.

Flat mode has a flat cost of 0.07 bits per macroblock, which is very low. It improves the responsiveness of video feeds quite a bit. There are a whole family of techniques along these lines used by H.264 and other codecs (32×32 or larger macroblocks, special gradient blocks, etc.) that can make a big difference in high motion or scenes with out of focus regions. These techniques do tend to make high motion frames that can’t fit in available bandwidth “blur out” but this is a lot less offensive than partial updates.

Entropy Coding with RANS

We’ve gotten pretty far – down to 4 bits per pixel on an average feed. But there’s one important piece of the JPEG approach we’ve neglected up till now: entropy coding. Right now, all coefficients are sent as 2 byte signed shorts. We’re doing a lot of work to make them close to zero, but it only reduces size if we get adjacent coefficients to zero and RLE gives us a win.

JPEG uses a Huffman coder, which uses fewer bits for numbers closer to zero. Common +1 and -1 coefficients get sent in only a couple of bits, while rare, larger coefficients such as 1000 cost more bits. The balance is a significant reduction in size. JPEG also stores some coefficients relative to previous coefficients to help reduce the average magnitude and thus the bit count.

Because I’ve worked with Huffman codes before, I decided to try something new to me – arithmetic coders. I’ve read about them but never used one directly. I had been wanting to try the RANS coder from @rygorous (found at https://github.com/rygorous/ryg_rans), so I started implementing one based on his code.

Getting it fully integrated and working was a beast! I lost a couple of days to getting the RANS coder up and running with no issues. A lot of this was my learning curve, as the algorithm is subtle. Because the operation of an arithmetic coder is fairly opaque, I had to take a black box approach most of the time- I could tell via sentinels if I got right or wrong values, but it was difficult to determine the cause when I got a wrong value out of the decoder. Eventually, I ended up implementing a set of unit tests (using catch.hpp) which helped me constantly check each step of the encoding and decoding process so I could quickly spot any issues when I made a change.

Arithmetic coding has two parts – the actual coder (which is clever and small) and the context. The context models what values are likely at each point in the stream, and the arithmetic coder ensures that likely values are encoded with fewer bits than unlikely values. Good contexts make a huge difference in the effectiveness of compression.

In preparation for all of the following techniques, I captured eleven megabytes of RLE encoded DCT data at various quantizations to use as a training set. For convenience I regenerate my contexts at application launch, but in the real world we would use a much larger training set and only ship the final statistics. This data could easily be included in a header file, and would add just kilobytes to the executable footprint.

Order 0 Context

Contexts are classified by how many symbols back they consider for their prediction. An order 0 context looks only at the current symbol state to determine likely and unlikely symbols. Typically this boils down to a single static list of symbol frequencies.

Order 0 context was easy to implement (Ryg conveniently provides such an implementation in his ryg_rans project!) and I found it saved us about 1.9 bits per pixel.

Order 1 Context

I next tried to implement an order 1 context that looks at the previous symbol and the current symbol to determine its odds. However, I ran into huge issues – I just could not round trip data properly. The reason why is interesting.

RANS-style arithmetic coding is as a deterministic, symmetric process. Think of it like braiding a rope. To unbraid the rope, you do the exact same things but in reverse. Similarly, you run RANS backwards to generate the encoded data, and forward to get the decoded data.

When you are working with an order 0 context, this isn’t a big deal. You get the same statistics out whether you submit symbols forwards or backwards. However, more complex contexts aren’t symmetric (ie, odds of A after B are different than B after A). So no matter what I tried I could not get my order 1 context to work!

Subsequently, I got some advice from the excellent and wise Fabian Giesen, who pointed out that I should buffer symbols for encoding so I can process them forwards and write them backwards. This unblocked progress, but unfortunately I ran out of time to implement the order 1 context – I had already moved on to something else.

Dual Order 0 Context

Since the order 1 context was initially a bust, I tried an alternate approach. I wanted to take advantage of the structured nature of the RLE encoded data, which is short aligned. So I built two order 0 contexts, one for the first byte of all the shorts and the other for the second byte. This was simple and worked pretty well – and it was symmetric, so it bypassed the RANS order concerns. This took me down another 0.3 bits per pixel.

Extended Quantization

JPEG only has quantization levels from 1 to 22. However, we can tolerate much worse quality in video than we can in a still image – motion hides artifacts. We can extend the quantization range, lose some quality, and drop up to another 0.9 bits per pixel.

Block State Format (v3)

If you will recall how we encoded our blocks in our packets, you might have thought it wasn’t as efficient as possible. Specifically, a few things cost us in overhead. We re-send the compression type and level for every macroblock, when we don’t change them very often and could probably send them once per packet. This gains us 0.03 bits per pixel.

We also have a flag to indicate if there is another macroblock coming. This is cheap for small numbers of macroblocks, but now we have good compression so we can send a lot in a single packet. If we have 100 macroblocks, we burned 100 bits. It would be a lot cheaper to use 8 or 9 bits to encode the macroblock count. This gains us 0.003 bits per pixel. Not a huge win but it dovetails with the next change well.

Ryg kindly pointed out this win to me via Twitter. Every time you flush RANS, you have to write out 4 bytes. When we flush after every block, then we burn ~32 bits per block. If we modify our packet to have one continuous RANS stream for all the macroblocks, we only have to flush once for perhaps a hundred macroblocks. This gains us around 0.12 bits per pixel.

One cool thing about RANS is you can insert uncompressed bits directly without any additional overhead. (It’s symmetrical and deterministic, remember? So you just stuff bits in/out and if you are consistent it works itself out.) We use this bypass mode to handle special cases like the flat encoding mode data and the flag that determines if we’re a flat or a DCT block. This allows us to put all our macroblock data through RANS with no hassle.

We could arithmetically code this data based on the likelihood of flat vs. non-flat blocks, or based on common RGB values, but since flat mode is already so much cheaper than DCT, I skipped it. So, no change in bits per pixel, but it keeps our implementation straightforward.

Now the packet look like this:

4 bits - mode (raw, zip, lzo, dct) for packet
6 bits - quality level (if dct)
8 bits - block count
[RANS bytes follow to end of packet]

There is one gotcha. Because RANS pops out bytes at irregular intervals, we can only know if our packet is full by compressing all our pending data after each block is added, and checking to see if we ran out of space. Currently, I do this after every block and as a result I spend about two thirds of my encoding time re-compressing the same data over and over. Because RANS operates in reverse, I can’t easily “checkpoint” my encoding state and only do work for the new content – I have to process the whole packet’s worth of data every time.

I believe this could be substantially improved by heuristically checking more or less often based on the available space. So I might check only once every 10 blocks at first, then more and more regularly as I get towards the end of the packet. This would give a substantial reduction in CPU load.

A Word on Performance

(Notes for self: I tested commit c8e59a2, on a Core i7-2600k 3.4ghz quad core. Program is single threaded.)

Overall, the system has no problem maintaining 60hz performance in release build. Deblocking or very high bandwidth usage can cause it to miss this target, but there are clear paths for optimization. Playback is surprisingly lightweight.

Memory footprint is around 75MB and does not vary substantially. Higher latency connections will cause slightly higher usage, and a lower latency connection slightly less (as we store per-packet information until the packet is received or dropped).

% of Frame Time	Description
25%	Speculative compression
12%	Block serialization
25%	GUI rendering
5%	Decompression
3%	Update error metric

Overall, we see about 40% of frame time spent encoding, and 5% on decoding.

The system is designed to be easy to parallelize. Specifically, since frames are often sent with multiple packets, we would recommend divvying up pending blocks and encoding packets in worker threads. Decoding could also be done with workers for a good speedup. On a fast desktop CPU this is less of an issue, but when trying to achieve low latency on embedded hardware, it could make a big difference.

We can also move to SIMD versions of hot path code (RANS, DCT, and color conversion are all eligible). Our bitblt is also naive, and contributes to about half of the GUI rendering time. A SIMDized blitter would run much faster.

We estimate an optimized decoder implementation on the same hardware would take around 1% frametime, and an optimized encoder implementation would take around 5% frametime. This puts screensharing of 1080p on a single core well within reach – we would expect to see 7% frametime decoding and 40% frametime encoding.

Closing Thoughts

Writing my own video chat app and video codec was a great experience. Much like writing a toy compiler, language, or operating system, it is a learning experience more than a practical application. It gave me a deeper appreciation for the excellent video codecs available on the market today, and a better understanding of the issues facing things like Skype and WebRTC.

Ultimately, for someone often frustrated by video chat applications, it was cathartic to finally take a swing at the problem. And now I have something to tinker with when Skype breaks on me.

Final Statistics

Avg bits per pixel: 0.6 bits for low quality, 1.0 bits for good quality
Max usable sustained packet loss: 25%
Max survivable sustained packet loss: 75%

Thanks for reading. My team and I solve hard problems in the streaming video, AR/VR, scalable backend, gaming, and IoT spaces. If we can help you, contact us at ben AT theengine DOT co.

Last time, I got angry at Skype and metaphorically flipped the table by starting to write my own video conference application. We got a skeleton of a chat app and a video codec up, only to find it had a ludicrous 221 megabit bandwidth requirement. We added some simple compression but didn’t get much return from our efforts.

We definitely need better compression before we implement networking. There’s no point trying to stream hundreds of megabits over a real world link. We need to slim things way down.

It turns out that most video codecs use some variant of the lossy compression techniques from the JPEG format. I’ve always wanted to learn the details of JPEG. With the excellent mango library from my friend Jukka Liimatta as a reference, I started down the well trodden road of abusing JPEG for my own purposes.

DCT

The biggest building block of JPEG is the Discrete Cosine Transform. The DCT is related to the Fourier Transform. It takes a signal (in this case, a 8 by 8 array of pixels) and breaks it down into frequency components. Later, we can take the frequency components and run them through the Inverse Discrete Cosine Transform to get our original signal back.

Lower frequencies live in the top left of the output matrix, while higher frequencies live in the bottom right. Larger numbers indicate more of that frequency, smaller numbers less. The following image shows the frequencies that each of the 64 coefficients you get from a DCT transform encode. By combining them in varying amounts, you can encode any 8×8 pixel image.

(A word on blocks and macroblocks. You will notice that this is an 8×8 DCT but we are working with 16×16 macroblocks. Macroblocks are so named because they are made up of a combination of smaller blocks. At this point in the game, we are breaking our 16×16 RGB pixel macroblock into 12 8×8 DCT blocks. However, we will see later how we can do much better than that. Thanks for reading, now back to your regularly scheduled programming.)

Lossy compression is all about removing data that doesn’t contribute much to the final output. By working in frequency space, we can start dropping components that are less important. For instance, imagine a macroblock with a smooth gradient. If we are working with RGB values directly, it might be difficult to efficiently detect or encode the gradient. We would be easily confused by noise in the image. By running it through a DCT, we’d immediately see that the largest components were low frequency elements, and the smallest high frequency. It is much simpler to drop the smallest components in a grid!

JPEG is built around making the best use of DCT. One of several tricks it uses is zigzag tables to reorder the DCT components for better encoding efficiency. In adapting the JPEG code from Mango, one issue that got me was that the encoding and decoding order aren’t symmetric. That is, if you zigzagged an 8×8 matrix with the encoding table, then run it through the corresponding decoding table, you get a transposed version of the matrix. This manifested as blown out colors in the output image, and I’ll discuss why in the next section…

Quantization

The biggest trick JPEG uses for better encoding efficiency is quantization. This is where things get lossy – DCT, like FFT, is fully reversible if you preserve the full precision of its output. By rounding the values from the DCT – also known as quantizing them – you can control how much data is lost. Frequencies that don’t contribute much tend to have small components in the DCT, so they rapidly become zero as they are quantized by larger amounts. Frequencies that matter a great deal have larger coefficients and will not be removed or will be removed very late in the process.

The following image shows similar video encoded with varying quantization level, going from high quantization (which looks blocky) to low quantization (which looks smooth and detailed).

To quantize, JPEG divides each DCT component by a rounding factor derived from the quality level setting. In fact, it generates a whole quantization matrix – one quantization factor for each of the 64 values of DCT output – that biases towards preserving component values or low frequencies. Each value is rounded a little bit differently.

This is how I was getting blown out colors (as mentioned in the previous section). Imagine a partial DCT output matrix consisting of frequency coefficients [40, 30, 20, 10]. We generate a (partial) quantization matrix [1, 2, 3, 4]. We divide and get quantized values of [40, 15, 7, 3] which are then written out.

Originally I was not taking the zigzag’s transposition into account, so I was scaling up by a rotated version of the quantization matrix. This meant that some values were being scaled up with a different factor than they were scaled down. If we represent rotation of the full quantization matrix by scrambling our small demo matrix, it would be like dequantizing by [1, 4, 3, 2]. We would end up with [40, 60, 21, 6] at the end. In compression speak, we failed to preserve the overall “energy” of the block – we ended up with bigger coefficient factors than we started, causing the output to be brighter! The effects vary depending on specifically how you get it wrong – here’s a shot where the luminance quantization coefficients were wrong:

Great – so now we’ve gotten the coefficients for unimportant parts of the DCT to be smaller, and in many cases zero. We can adjust how much we lose by varying a quality level factor. But we didn’t actually reduce the amount of data we transfer at all! Wait what?!

RLE

The reason why JPEG tries to get everything closer to zero is simple. Zeroes are cheap to send!

Why? It’s not just because of the convenient hole in the middle of a 0. The next step in JPEG is to collapse all those zeroes with run length encoding. We walk through the array, and replace adjacent zeroes with a special value indicating how many zeroes there were. Suppose our quantization makes the last 32 coefficients zero. We would go from 128 bytes (JPEG DCT output is one signed short per coefficient) to 66 bytes (32 populated coefficients plus one special value indicating 32 zeroes). That’s 50% savings with a pretty simple compression algorithm!

It sounds super simple, and once written it is, but getting RLE encoding to work reliably took several important debugging techniques. In addition to the old fashioned “stare at it in the debugger” technique, I added sentinels – known values at specific locations in the data stream – to help catch any misalignments. If I didn’t see sentinels exactly when expected, the code would instantly assert. I also put some asserts into the RLE decoding routines to check that we always got 64 values out and that a run would not cause us to over flow the buffer (a common sign of bad data would be runs longer than the remaining space in the block).

Because any bugs in later stages of processing would corrupt the RLE data, these asserts were usually my first sign that I had gotten something wrong in later parts of the compression pipeline.

Block State Format (v1)

At this point, although the app isn’t network capable, it starts making sense to describe the format for macroblock updates. Again, keeping things simple, we have the following structure:

 1 bit - is there another block update? if false, stop processing.
 4 bits - compression type (raw, zip, lzo, dct).
 6 bits - only present for zip/dct, indicating quality level.
 6 bits - macroblock x index
 6 bits - macroblock y index
 10 bits - size of compressed data
 size * 8 bits - output from the compression algorithm

Note that the macroblock x and y index are counting macroblocks, not pixels. So in this version of the protocol we can have images up to 1024 16px macroblocks or 16,384px on a side. We could shave a few bits if needed by taking into account the actual image size.

As you can see, this incurs a fixed overhead of 33 bits per macroblock, or 0.51 bits per pixel.

We see an average data rate of 8 bits per pixel with the subset of JPEG implemented above at average quality (ie, quality level 50). Lowest quality setting ran at 6 bits per pixel, and highest quality setting was 20 bits per pixel.

Now we’re down to only 40-60 megabits per second. It’s progress! Our video codec is starting to be feasible over very fast broadband connections, but we can do a lot better.

Next Time

Now that we can transfer little enough data to run the codec over a network, we should implement some networking and test it for real. And that’s exactly what we’ll do in Part 3!

Networking expert Glenn Fielder’s article on network serialization techniques is a great read. It discusses a number of scenarios including serializing a sparse set of items from an array. This blog post discusses Glenn’s implementation of sparse array serialization and my thoughts on it. To be sure, Glenn’s code has the virtue of simplicity, and the concepts I’m going to discuss are more complex. Don’t assume the space or time trade offs of different serialization techniques without measuring them in practice.

Without further ado, the original code:

template <typename Stream> 
bool serialize_object_index_internal( Stream & stream, 
                                      int & previous, 
                                      int & current )
{
    uint32_t difference;
    if ( Stream::IsWriting )
    {
        assert( previous < current );
        difference = current - previous;
        assert( difference > 0 );
    }

    // +1 (1 bit)
    bool plusOne;
    if ( Stream::IsWriting )
       plusOne = difference == 1;
    serialize_bool( stream, plusOne );
    if ( plusOne )
    {
        if ( Stream::IsReading )
            current = previous + 1;
        previous = current;
        return true;
    }

    // [+2,5] -> [0,3] (2 bits)
    bool twoBits;
    if ( Stream::IsWriting )
        twoBits = difference <= 5;
    serialize_bool( stream, twoBits );
    if ( twoBits )
    {
        serialize_int( stream, difference, 2, 5 );
        if ( Stream::IsReading )
            current = previous + difference;
        previous = current;
        return true;
    }

    // [6,13] -> [0,7] (3 bits)
    bool threeBits;
    if ( Stream::IsWriting )
        threeBits = difference <= 13;
    serialize_bool( stream, threeBits );
    if ( threeBits )
    {
        serialize_int( stream, difference, 6, 13 );
        if ( Stream::IsReading )
            current = previous + difference;
        previous = current;
        return true;
    }

    // [14,29] -> [0,15] (4 bits)
    bool fourBits;
    if ( Stream::IsWriting )
        fourBits = difference <= 29;
    serialize_bool( stream, fourBits );
    if ( fourBits )
    {
        serialize_int( stream, difference, 14, 29 );
        if ( Stream::IsReading )
            current = previous + difference;
        previous = current;
        return true;
    }

    // [30,61] -> [0,31] (5 bits)
    bool fiveBits;
    if ( Stream::IsWriting )
        fiveBits = difference <= 61;
    serialize_bool( stream, fiveBits );
    if ( fiveBits )
    {
        serialize_int( stream, difference, 30, 61 );
        if ( Stream::IsReading )
            current = previous + difference;
        previous = current;
        return true;
    }

    // [62,125] -> [0,63] (6 bits)
    bool sixBits;
    if ( Stream::IsWriting )
        sixBits = difference <= 125;
    serialize_bool( stream, sixBits );
    if ( sixBits )
    {
        serialize_int( stream, difference, 62, 125 );
        if ( Stream::IsReading )
            current = previous + difference;
        previous = current;
        return true;
    }

    // [126,MaxObjects+1] 
    serialize_int( stream, difference, 126, MaxObjects + 1 );
    if ( Stream::IsReading )
        current = previous + difference;
    previous = current;
    return true;
}

Intuition and Rules of Thumb

Let’s talk about why this code immediately stood out to me – it triggered an intuition: code structure correlates with entropy.  Entropy in the information theory sense means that repetitive data has lower entropy, and unique data has high entropy. This intuition cues me to think about refactoring along DRY principles or moving to a more data driven approach. These kinds of changes take low entropy code and “compress” it so it has higher entropy – in other words, code that is less repetitive to accomplish the same results.

Consider the following example:

// Choose largest of provided parameters.
int pickHighest(int a, int b, int c)
{
   if(a > b)
      if(a > c)
         return a;
      else
         return c;
   else
      if (b > c)
         return b;
      else
         return c;
}

The repetitive tree structure of the code immediately suggests that it could be written more concisely, especially as you have more arguments from which to pick. A sort algorithm is an obvious choice, e.g.

int pickHighest(vector<int> items)
{
   items.sort(); // Sort items in descending order.
   return items[0]; // Pick the highest item.
}

This version is much shorter, it can handle any number of items, and gets right to the point.

(Caveats: Of course, the astute reader will note that there are situations where for small argument counts the tree can be a faster choice. Rewriting it with ternary operators doesn’t count – we’re concerned with execution structure here, not syntactic structure. Finally, there are plenty of loose ends to tune in the new implementation for correctness and speed.)

Huffman Codes

How does this apply to Glenn’s implementation above? Well, we see it has a repetitive structure. It’s attempting to encode more common numbers in fewer bits, and rarer numbers with more bits. That smells like a Huffman code to me! A Huffman code is a tree-like encoding structure that assigns unique bit strings of varying lengths to different values based on their expected frequency. So the most common value is encoded with a couple of bits, while the least common value might take ten or twenty bits. This lets you get a big win assuming that your assumptions about what’s common are right. As an example, algorithms like deflate/zip rely on Huffman codes. There are also techniques like arithmetic coding that are more powerful and complex expressions of this concept.

A canonical Huffman code is one which generates N codes where the first value is expected to be most frequent, and the last value is least frequent. Implicitly, this means the first value’s code is shortest and the last value’s code is longest. Fortunately, this is exactly the situation we want to deal with! If you knew more about the actual distribution of values you could generate your own tree based on the observed values and get better efficiency in your serialization.

(In the following code, I’m going to assume you have a serialize function to read/write a canonical Huffman code that can represent values between 0 and 127. This is definitely sweeping complexity under the rug, although a Huffman coder is a handy thing to have in general. If I was implementing this, I’d start with the HuffmanProcessor from Torque, although there are surely more efficient implementations.)

Here’s the implementation using Huffman coding:

template  
bool serialize_object_index_internal( Stream & stream, 
                                      int & previous, 
                                      int & current )
{
    uint32_t difference;
    uint32_t cappedDifference;
    if ( Stream::IsWriting )
    {
        assert( previous < current );
        difference = current - previous;
        assert( difference > 0 );

        // Store a capped difference so we know when to 
        // write overflow values.
        cappedDifference = difference > 127 ? 127 : difference;
    }

    // Write huffman encoded delta.
    serialize_huffman127(stream, cappedDifference);

    if(cappedDifference == 127)
    {
        // Highest value keys us to code large gaps.
        serialize_int( stream, difference, 126, MaxObjects + 1 );
    }
    else
    {
        // The serialized valus is the actual value.
        difference = cappedDifference;
    }

    if ( Stream::IsWriting )
        current = previous + difference;
    return true;
}

Notice that there is now no repetitive structure to the code. We have to deal with a couple of details (what if it’s a number bigger than we can Huffman code?), but the basic flow is much simpler. And it achieves a very similar result.

Conclusions

What are the trade offs of this approach? Well, first off, it may not be as bit efficient in the “every item is consecutive” case with the canonical Huffman code because the canonical Huffman code won’t assign a single bit to that most common case. You could add a special case for that (and spend an extra bit per serialization in the non-consecutive case).

Second, Huffman codes are more complex to read and write. If you are doing a LOT of serialization this might affect CPU usage. Most likely it will not be a significant difference – I’ve rarely seen any game be bottleneck on network serialization.

Third, tuning is more indirect. With Glenn’s implementation, you can modify the code to directly affect the bit cost of various scenarios. With this approach, your main tuning option is to regenerate the Huffman code based on actual network traffic. You can also add special cases e.g. for the consecutive item case. Regardless of the approach you use, you always want to check your changes by running your game and measuring actual update sizes in various scenarios to confirm you are really getting a win.

Fourth, there are better coding techniques than Huffman – in practice you probably want a variant of arithmetic coding. Compressing the entire packet with a standard compression algorithm like LZMA is also an option. This article is just a piece in the puzzle of the ultimate network protocol.

Is the Huffman approach a slam dunk win over the manual approach? Unclear. But it gives a more formal perspective on the smart serialization technique Glenn uses, and is a good entree into a larger discussion on coding techniques for network serialization. I hope you found this article interesting!

(Thanks to  @glennfiedler for prompting me to write this.)

I was recently advising a fellow TEC-er on setting up their Mac, and they suggested I write it up. These are the tools I use daily for development! Here we go…

1. Alfred is an awesome tool. Like the Windows key on Win8, you can hit a keyboard shortcut (cmd-space by default) and type an app name or other command. Super handy.
2. iTerm2 is a replacement for the terminal. The best thing about OS X is that it has a great UNIX command line alongside a nice GUI, and iTerm2 makes the terminal even better. Don’t forget to configure and use the global terminal hotkey. So handy!
3. Spectacles lets you control your windows with the keyboard. So handy, especially on my Air with limited screenspace.
4. Sublime Text 3 is my go to text editor on every platform, and especially Mac. Favorite features include cmd-d (to select multiple matching text ranges and edit them simultaneously) and using the subl command line tool to open folders/files from the terminal.
5. Oh My ZSH! is a great upgrade for your terminal. Maintaining your own custom config is even better… but as a default setup it’s pretty good.
6. RVM is a good way to get assorted Rubies install on your system. Ruby is handy but it suffers from versioning hell. RVM can help, sometimes.
7. The GitHub app is also mega handy. No replacement for command line but it simplifies auth and basic commit/branch switching.

There are also some miscellaneous things you probably need to do: get the latest git, possibly install brew and ports (but I find more and more I prefer to compile and install from source), get all the latest updates from the App Store, and install XCode 5 (from the App Store these days). You might need CMake for Loom builds, along with Android SDK and Android NDK. We like to use the HipChat app for communication amongst our team.

As you can see, I really like a keyboard oriented workflow. Perhaps it is the inevitable outcome of programming for so long. I’ve trended more and more towards it. I haven’t quite gotten to the point of using tmux and emacs in a fullscreen terminal, but who knows…