C++ Accelerator Libraries

In preparation for my C++Now talk entitled The Future of Accelerator Programming in C++ I am currently reviewing numerous C++ libraries. I put together a catalogue of questions for these reviews. The questions are intended to gauge scope, use-cases, performance, quality and level of abstraction of each library.

Iqra: Read, image by Farrukh
  1. Is concurrency supported?
    Accelerators are massive parallel devices, but due to memory transfer overhead, concurrency is a central aspect for many efficient programs.
  2. How is memory managed?
    This is a central question since simple and efficient management of distributed memory is not trivial.
  3. What parallel primitives are provided?
    Parallel primitives are essential building blocks for many accelerator-enabled programs.
  4. How is numerical analysis supported?
    Massive parallel accelerator architectures lend themselves well to numerical analysis.
  5. How can users specify custom accelerator functions?
    A useful accelerator library should allow users to specify custom functions.
  6. What is the intended use-case for the library? Who is the target audience?
    Is the library suitable for i.e. high performance computing, prototyping or signal processing?
  7. What are noteworthy features of the library?

This is a list of all libraries that I am reviewing:

Library CUDA OpenCL Other Type1
Thrust X OMP, TBB header
Bolt X2 TBB, C++ AMP link
VexCL X3 X header
Boost.Compute X header
C++ AMP X4 DX11 compiler
SyCL X5 compiler
ViennaCL X X OMP header
SkePU X X OMP, seq header
SkelCL X link
HPL X link
ArrayFire X X link
CLOGS X link
hemi X header
MTL4 X header
Kokkos X OMP, PTH, seq link
Aura6 X X header

If I missed a library, please let me know. I will add it immediately. I’m going to publish selected library reviews here on my blog. I’m hoping to discuss specific reviews with the original library authors. The conclusions of these reviews will be part of my talk at C++Now.

  1. either header-only library, link-library or library that requires compiler support
  2. custom AMD OpenCL Static C++ Kernel Language extension required
  3. CUDA SDK required at runtime
  4. Prototype implemenation available here
  5. only specification released so far
  6. disclaimer: library developed by the author

Migrating multiple repositories to Git

A few weeks ago I faced the challenge of migrating and merging multiple SVN and Git repositories into one single repository. The stackoverflow discussion “Merge two separate SVN repositories into a single Git repository” contains all the information required to solve this problem. This is a concise reproduction of all the bits an pieces presented in the article.

Migrating Birds, image by Emilian Robert Vicol

The plan is simple:

  1. clone the involved Git repositories
  2. migrate relevant SVN repositories to Git
  3. rewrite the repositories in case of overlaps or errors
  4. create new repository and add empty commit
  5. add remotes for all repositories
  6. fetch all remotes
  7. create a list of all commits of all repositories, sort it chronologically
  8. cherry-pick each commit in the list and apply it in the new repository

And here are the commands that implement the plan above. First clone and migrate Git and SVN repositories.

mkdir ~/delme
cd ~/delme/
git clone ~/dev/repo1
git clone ~/dev/repo2
git svn clone svn://server:/repo3/
git svn clone svn://server:/repo4/

If the repositories have the same file or folder names a history rewrite is necessary. Assuming repo1 overlaps with other repositories, it is a good idea to put the contents of repo1 in a subfolder in the target repository. To accomplish this, the history of the master branch of repo1 is rewritten and all its contents is moved to the folder “subfolder”.

cd repo1
git filter-branch --tree-filter 'mkdir -p subfolder; find -mindepth 1 -maxdepth 1 -not -name subfolder -exec mv {} $fname subfolder \;' master

In this step, it is also possible to completely remove files from a repository. The following command removes the file “invalidfile” in “subfolder” from the repository completely.

git filter-branch -f --index-filter 'git rm -r --cached --ignore-unmatch subfolder/invalidfile;' master

This can be repeated for other repositories as well if necessary or desired. In the next step, the target repository that should contain all merges is created. Remote repositories are added and fetched.

mkdir ~/newpreo
cd ~/newpreo
git init .
git commit --allow-empty -m'Initial commit (empty)'
git branch seed
git checkout seed

git remote add repo1 ~/delme/repo1
git remote add repo2 ~/delme/repo2
git remote add repo3 ~/delme/repo3
git remote add repo4 ~/delme/repo4

git fetch repo1
git fetch repo2
git fetch repo3
git fetch repo4

Finally, file containing lists are created for all commits from all repositories. The lists include the timestamp for each commit (seconds since 1/1/1970). The lists are then sorted and merged. The final result is stored in the file “ordered_commits”. This list is then iterated over and each entry is fed to the git cherry-pick command.

git --no-pager log --format='%at %H' repo1/master > reco1_commits
git --no-pager log --format='%at %H' repo2/master > reco2_commits
git --no-pager log --format='%at %H' repo3/master > reco3_commits
git --no-pager log --format='%at %H' repo4/master > reco4_commits

cat *_commits | sort | cut -d' ' -f2 > ordered_commits

cat ordered_commits | while read commit; do git cherry-pick $commit; done

The cherry-pick command prompts git to apply the commit to the current branch. This results in a repository containing all commits from all 4 repositories in a chronological order. That’s all there is to it.

Stop teaching Matlab

Many universities rely on Matlab for their mathematical and technical computation curriculum. This is because the syntax of the Matlab language is very intuitive and a perfect fit for numerical computation. Matlab also comes with a huge library of sophisticated math functions and excellent documentation. And universities are often equipped with campus-wide Matlab licenses. Professors as well as students can use Matlab for free.

Disorderly Conduct, image by Ken

Mathworks, the company behind Matlab, is pursuing an obvious plan with these generous campus licenses. Their strategy is aimed at selling software to mathematicians, engineers, physicists and computer scientists after they graduate. Since Matlab is often the only or most convenient tool these scientists get to know during their studies, Mathworks’ plan is very successful.

If for-profit companies decide to base their research and product development on Matlab I have no objections. The market will decide if it is the right decision.

But I find it appalling that a wide variety of todays scientific advances are based on a proprietary software1 product such as Matlab. Institutes that base their research on Matlab are at the mercy of a for-profit US company to sell them and renew licenses.

Scientific results based on Matlab are not free2. To reproduce, validate and build upon them, a Matlab software license is required. It is my opinion that, since science is largely paid for by the public, its results must also be available to the public. They must be free. It is thus a fatal mistake to train students and young scientists, the future creators of scientific knowledge, in using tools that restrict the freedom of their results.

There are many excellent free alternatives to Matlab. I would just like to point out two of them here: The long-established Matlab alternative is Python with the computing environment SciPy. The other alternative is new: Julia, a dynamic programming language designed to address the requirements of high-performance numerical and scientific computing. From a C++ developer’s perspective, Julia’s expressive type system and the excellent performance compared to compiled languages are very attractive.

There are numerous free software alternatives that allows researchers to do open and reproducible science. As of Spring 2014, the MIT linear algebra course suggests Julia as a Matlab alternative to solve homework problems. I hope teachers and professors will switch to free software and instruct the next generation of scientists how to produce free results.

  1. Proprietary software is software that does not give the user freedoms to study, modify and share the software, and threatens users with legal penalties if they do not conform to the terms of restrictive software licenses (source).
  2. free as in freedom, both negative (free of oppression or coercion) and positive liberty

Cutting off Google’s Tentacles

I just realized how easy it is to cut off one of Google’s tentacles throughout the web. This is a WordPress blog and I used the Ultimate Google Analytics plugin to keep track of the number of visitors, where they come from (referer, not geo-location), keywords and so forth.

Google Analytics is just one of the many tentacles, Google spreads throughout the web. There is the Google Fonts API, there is Google Hosted Libraries and probably numerous other things that I am not aware of. Visitors to websites that include any one of these Google services will always contact one or multiple Google servers, thus identifying themselves (to some degree).

So I uninstalled the Google Analytics plugin, deleted my Google Analytics account and installed the WordPress Statistics plugin instead. It works flawlessly so far. I can recommend it.

Mere users of the web can monitor and block these tentacles as they browse the web through a Firefox add-on called Disconnect. I can recommend this plugin as well.

Let’s work together to make the web a less centralized, more private place for everyone. Let’s try to exclude large corporations from our interactions between each other. It is probably none of their business and certainly should not be their business.

8 GPU GeForce Titan Tyan System

A box arrived a few days ago at work.

8 GPU Titan System box

The box contained a little supercomputer comprised of 8 GeForce Titan GPUs in a Tyan FT77A platform.

8 GPU Nvidia Titan System

The system fits nicely in our server room.

8 GPU Nvidia Titan System Installed

And the 8 GPUs light up too!

8 GPU Nvidia Titan System Glow

scikit-image – Image processing in Python

I just discovered scikit-image – an image processing toolbox for the python programming language. The project appears to be very active and under heavy development. I have been looking for a while for ways to replace my scripts that rely on the Matlaband this seems to do the trick. I’m especially excited about the libraries functionality to measure region properties. This will come in very handy. I have either been sleeping under a stone or this project is not very well known. Let’s change that. Go check out scikit-image!

Nvidia OpenCL Examples

It looks like Nvidia is slowly but steadily abandoning OpenCL – their OpenCL examples are not included in the 5.0 SDK anymore and the links on their OpenCL webpage are dead. It is not clear if this is an oversight or intentional. With the recent introduction of the Intel Xeon Phi, an accelerator that supports OpenCL, this could be a strategic move on Nvidia’s part.

I’ve created a github repository that contains all Nvidia OpenCL examples from CUDA version 4.2.9. Please note that it currently contains only the Linux examples. Feel free to fork and add the Windows and Mac OS examples.

C99: casting to variable-length arrays

C99 understands variable-length arrays. They look something like this:

int d1, d2, d3, d4;
// runtime-assign those variables
int vla[d1][d2][d3][d3];

Now the question is, how to properly cast a pointer to this type when passing it to a function that accepts such a type in a way that makes the compiler happy? The syntax is somewhat unusual:

void func(int d1, int d2, int d3, int d4, int vla[d1][d2][d3][d4]){}
// ...
int * x = malloc();
func(d1, d2, d3, d4, (int (*)[(int)(d2)][(int)(d3)][(int)(d4)])x);

Tested with gcc 4.6.1 and the –std=c99 compiler option. I could not find information about this anywhere on the web so I hope this will help others who wonder how it should be done.

Mozilla Thunderbird and Flowed Format

I like my e-mails to look a particular way. I send plain-text only e-mails with line wrapping at about 80 characters. This way I feel I can control to some degree what the e-mail will look like at the receivers end. And I want my e-mails to look like that even when I’m quoting someone.

Up until now I hated to use Thunderbird because it did not allow me to send e-mails like that. Since gmail looks ugly as hell now I came back to Thunderbird and set out to solve this issue once and for all. I searched for “thunderbird plain text wrap” and whatnot. A smart fellow on a forum had the answer:

There are two hidden preference settings which one can access through Edit > Preference > Advanced > General > Config Editor button:

  • mailnews.display.disable_format_flowed_support must be set to true to avoid rewarpping of messages received for which the sender allows flowed formatting
  • mailnews.send_plaintext_flowed must be set to “false” to avoid sending flowed format e-mails

The problem seems to be something called “flowed format” but I’m too ignorant to go into the details of what it does, what it means and why I would want it. I’m just happy the e-mails my Thunderbird sends look the way I want them to look. Finally. Thank you rsx11m!

Cryptography for the Masses

“The multiple human needs and desires that demand privacy among two or more people in the midst of social life must inevitably lead to cryptology wherever men thrive and wherever they write.” wrote David Kahn in his book “The Codebreakers”, chronicling the history of cryptography. The book was published in 1967. Almost 45 years later cryptography is seldom used to protect our privacy.

The information age spawned databases and networks capable of extracting and storing large amounts of private data. Those databases are often unknown to us and if we know of their existence we can not control them. They store personal information, communication and financial transactions. This gathering of private data happens against our will if we believe surveys that show that we actually do care about privacy. Skeptics and experts caution us but the majority of web users is forced to give in to the subtle but grave disintegration of privacy, pushed forward by industry and government. They are growing their databases steadily, expanding the records they keep on all of us.

Good question, image by Garrett Coakley

We can see the consequences of these uncontrollable, central databases today. In what is believed to be one of the largest data security breaches in history, attackers stole personally identifiable information of 77 million PlayStation Network users earlier this year.

Accidental exposure of personal data is another problem. It is very difficult to control who has access to which piece of information. People get fired for how they behave online because they confuse personal with public communication. The web does not forget. And ever since the uprisings in the Arab world it should be clear to everybody that what one posts online can have severe consequences, including imprisonment and torture.

There are a variety of interesting judicial and ethical approaches to cope with these issues. And there is cryptography – a technological means of preserving privacy. Cryptography enables anonymity, the concept of ‘publishing information while ones identity is publicly unknown’ as well as privacy, the ability to to ‘seclude oneself or information about oneself and reveal oneself selectively’.

But almost nobody uses cryptography. Asked if he encrypts his e-mail, Bruce Schneier, cryptographer and highly regarded computer security specialist answers “I do not, except for special circumstances”. He further argues that for more people to encrypt their communication, services like Gmail would have to do it by default. This will of course never happen, since those services draw their revenue from reading our messages.

It has to work out of the box

But the more important point Schneier makes is this: what has to happen to spread the use of cryptology? It has to work out of the box. No additional application should be required, no plug-in, no add-on and certainly no driver installation. There exists a concept that could potentially offer a transparent solution for everyone: browser based cryptography.

Browsers have evolved from being a mere presentation and navigation tool for the world wide web to a platform for collaboration and information sharing web applications. If browsers were able to do cryptography, every web users could potentially benefit from it. JavaScript engines have evolved rapidly to a point where they are efficient enough to handle the complex algorithms that cryptography entails. It is ironic that the same web applications that threaten our privacy are the main reason such powerful engines were developed in the first place.

The idea of browser based cryptography is simple: before users upload their personal data to application hosts they encrypt the data in the browser. The host only receives encrypted blobs of data and since users don’t share their key with the host the data is secure. If they decide to share their data with someone else they can provide them with means of decrypting the blobs. Users are in control at all times.

But JavaScript cryptography has many critics and there has been some discussion whether or not it is a viable solution. But the potential is vast and the issue of retaking privacy is too important to dismiss the technology right away. The discussion should not stop here. Solutions can be found to the given objections.

JavaScript Cryptography Criticism

The trust model certainly is a problem and seems inconsistent. The general assumption is that users don’t trust application providers with their data, thus the need for encryption. Modern web applications however download their code from the very same provider and consequently also download the code required for decryption and encryption. A contradiction: users don’t entrust providers with their data but they trust the provider to deliver the application and most importantly the correct cryptography code. Critics argue that users can decide to either trust or not trust whoever hosts an application. If they trust the host there is no need for encryption. If they don’t trust the host they should not share their data in the first place. If someone suspects a host has malicious intentions JavaScript cryptology is worthless.

The situation changes if the ‘honest-but-curious’ adversary model is taken as a basis. It assumes that the company providing a web services carries out the stated instructions and is not lying to the user. It is further assumed that it might do more than it promised such as storing the data for an unreasonably long time or even sharing data with third parties. In such a case JavaScript cryptography might be viable. Furthermore if the company is attacked and database dumps are stolen, the data is worthless. For an attacker to gain access to the data the web application source code has to be modified and users have to use the malicious application. To defeat such attacks, browsers would have to be able to validate web applications to make sure that they were not modified.

Dystopia 2, image by Hervé Girod

Today the browser is an environment not very well suited for cryptography. There is the threat of cross-site attacks: a web page loads content from many sources and all of those can potentially modify the cryptography code. Due to the dynamic nature of the JavaScript language, an attacker can replace correct code with a malicious version. Such an attack can only be discovered through tedious code analysis of all sources.

Critics also argue that browsers lack some crucial primitives important for cryptography such as a random number generation. Fortunately browser vendors see the need for such functions and are moving toward implementing them. Browsers further lack a secure key store, a crucial component in every crypto-system. It is also important to have the ability to securely erase secrets from memory once they are no longer needed. Since JavaScript engines employ garbage collection there is no way to control when objects are deleted and secrets are forgotten.

These issues are significant but they can all be addressed with care. Web developers, cryptographers and browser vendors will have to work together to find solutions to these shortcomings. Once this happens I expect to see secure JavaScript cryptography applications that satisfy even the skeptics. There is still a lot to be done but the potential is tremendous. So we should get to work.

Existing Implementations

Despite the criticism there are already some implementations that utilize JavaScript cryptography out there. They try to make do with the current state of browser support for de- and encryption. Some supplement browser capabilities with custom add-ons. This of course defeats the purpose of JavaScript cryptography but is a necessity at the moment since browser support is still in its infancy.

Aldo Cortesi’s is the most interesting project. It is an online list-manager that encrypts user data in the browser and sends only an encrypted blob to the host. It inspired people to think about what JavaScript cryptology can do and what it means to encrypt data in the browser. Encrypted user data is completely exposed as the application intentionally lacks authentication (except to prevent overwrites). Only the passphrase protects the data. Cortesi uses the expression ‘host-proof’ in his documentation of the project which is a dubious term, especially in the context of JavaScript cryptography. It was coined by Richard Schwartz and emphasizes that the host does not have to be trusted. This is a difficult claim. Users still have to trust the host because it grants the privilege of encrypting the data and it can revoke that privilege at any time. The ‘honest-but-curious’ adversary model again makes more sense.

More questionable applications are Lockify, Zero-Knowledge Box and Clipperz because they explicitly advertise the security of their products that is solely based on JavaScript cryptography. Critics argue that appearance of security is worse than no security at all. Their claim of increased security should indeed be taken with a grain of salt: data that otherwise would not be uploaded due to security considerations should not be stored with those hosts. Even so, the cryptology is a welcome addition to the security measures these companies take. It accomplishes the goal of preserving privacy.

Vintage Mail Boxes, image by Nathen Jantzen, all rights reserved

When asked to choose between a regular web application and a JavaScript cryptography enabled one, the latter is preferable due to privacy considerations since the user does not lose control of his data. Claiming that JavaScript cryptography enabled applications, with the current state of research and browser support, are more secure than others is debatable. We are not there yet.


There are a number of alternatives and especially the concept of storing encrypted data with a curious or even untrusted host is not new. Traditionally, host applications have been used to handle cryptographic operations. These tools must be installed and have to be properly set up by the user. Mobile platforms might be an ideal environment for these alternatives. Installing applications is hassle-free and very common on mobile devices. Due to the well defined platform, developers can keep user effort to configure these applications to a minimum.

Another promising alternative is a browser add-on called Cipherbox. Its developers recognize that it is unlikely that application providers will enable cryptography as well as the shortcomings of current JavaScript solutions. In their architecture they make a point of separating the interaction with the web content from the cryptology functionality. This could proof to be a significant security advantage over other solutions that give web content access to cryptology functions.

A colleague and I devised the idea of a cryptography enabled http proxy that is similar to the Cipherbox. The proxy is a trusted instance possibly hosted locally or connected via a VPN. All http traffic is sent through the proxy. It analyzes the traffic and encrypts and encrypts relevant parts like messages or images depending on its configuration. We implemented a prototype that is capable of transparently encrypting and decrypting Facebook messages using gpg. A proxy like this could run on a user’s FreedomBox and can in theory be extended to provide crypto-functionality for various platforms including for example Gmail.

A special form of cryptography called homomorphic encryption could enable users to take advantage of both cryptography and computing as a service at the same time. If encrypted data is sent to hosts, they usually can not process the data. If instead a homomorphic encryption scheme is in place, for certain algebraic functions on the plaintext, equivalent functions exist that can be applied to the ciphertext. Proponents of this technology argue that it could enable widespread use of cloud computing by ensuring the confidentiality of private data.


Controlling ones personal data is more difficult with every new database and network based innovation. At the same time privacy is more important than ever in a world that prepares to conglomerate health records, gathers and centralizes consumer behavior data and merges individual financial records into powerful profiles. Cryptography is an effective safeguard we must implement to prevent exploitation and discrimination based on our personal information. Every user must be enabled to use cryptology to control the data he wishes to share.

Browsers vendors must implement the building blocks required for cryptography including a secure key store that can be managed by the user. They should also include means of validating a running application against a checksum. Cryptographers and web developers must work together to implement correct and easy to use de- and encryption functionality for browser based applications.

More people must start thinking about this problem, more ideas are needed and should be carefully vetted by cryptographers and security experts. User interface specialists should work on making cryptography a transparent process. We need to get everyone involved and try to revert the damage that has already been done.

« Older Entries     Newer Entries »