Huffman on GPU

[Bulat Ziganshin]

While very slow, it may be interesting: http://habrahabr.ru/blogs/hi/124210/

Another, faster implementation: https://github.com/Kentzo/phuffman, "cuda" branch (the say that it's 20% faster than fast ari = www.cipr.rpi.edu/~said/FastAC.html )

[Shelwien]

Comparing C/GPU vs C#/CPU on compression algos - that's really fresh imho.
10MB/s static huffman becomes fast.
Java would be better though - I've seen many reports that native java on linux works faster than C++ windows exe under wine.
Still not sure whether wine is that slow, process time is that wrong, some java engine stays after program quits to finish the cached
writes, or something else.

[Surfer]

AFAIK java is slower than C#.

[m^2]

AFAIK java is slower than C#.

On Linux? Definitely no.
On Windows? I'd like to see a pointer to some good benchmarks, I haven't seen any yet.

[Surfer]

On Linux? Definitely no.

No, and I think it just because Mono slower than .Net.

On Windows? I'd like to see a pointer to some good benchmarks, I haven't seen any yet.

Yes, I've seen previously some article in russian, but can't found it now.

[Shelwien]

http://reverseblade.blogspot.com/200...rformance.html

[m^2]

http://reverseblade.blogspot.com/200...rformance.html

Thanks. Too bad there are so few details...I went to the site mentioned as source, but the thing is not there. No hardware info. No compiler versions. And I find it hard to believe that in a good test gcc is on average 34.8% slower than VC.
I'm doing some comparisons right now on my filesystem benchmark. It's clearly not a general-purpose comparison, but gcc 4.5.2 beats VC8 9:2 (tied in 5 tests). In that benchmark it always lost. So while it looked really interesting at first, now I have doubts that it's reliable.

EDIT:
There is a mistake in the benchmark. It shows that GCC is 34.8% slower than VS, but numbers (and chart) show that it's actually faster by ~1/3.
It looks much better now...though I would still like to know why is Java on Linux so much faster than on Windows. Was it even tested on the same box?

[Piotr Tarsa]

Well, that charts shows that Java on Linux was faster than anything on Windows Bear in mind that some tested programs uses C/ C++ libraries like GMP.

Some benchmarks like binary-trees are non representative, as C/ C++ versions use object pools but Java versions use normal allocation through 'new' operator. If we compare allocation through new and then garbage collection or manual deallocation, Java wins over C/ C++.

If you want actual numbers, look at: http://shootout.alioth.debian.org/u6...re-fastest.php
Today, on Linux at least, Haskell on GHC is almost as fast as C# on Mono and Lisp on SBCL is faster than both.

Comparing the medians:
1.05 for C
1.08 for C++
1.45 for warmed up Java 7
1.70 for cold Java 7

Overall it seems that for complex HPC Linux is the way to go (definitely not Windows) and the performance gains of C++ over Java do not offset the productivity deficits often. On the other hand Java still lacks unsigned numbers (WTF?) so for many tasks this can be too troublesome (as was in case of my QRotSort). And in case of relatively simple algorithms (eg below 100 KiB of raw C++ code) C++ is rather bearable and provides noticeable performance gains.

[Karhunen]

While looking for something else, I found this CUDA/MP library for those with access to such architectures @Google code

[mo_]

I just created a huffman decoder that runs on the GPU (Metal) for iOS devices, have a look at the full source code on github: https://github.com/mdejong/MetalHuffman