Final Linux Benchmarks Of Project Dirndl

[Biazzotto]

Dirndl is not "yet" a Hybrid Compiler

As you can see, Dirndl is the open-source EKOPath 4 compiler suite, but it does not have the HMPP "yet", that is included on PathScale ENZO Suite:

As Michael Larabel said here:

Beyond maintaining the EKOPath 4 compiler suite, PathScale also offers the ENZO 2011 and EKOPath Boost products. ENZO is a GPGPU/multi-core solution designed for NVIDIA Tesla hardware with CUDA support for compiling HMPP C, C++, and Fortran. EKOPath Boost is the Boost Libraries that are built with EKOPath. The licensing on ENZO 2011 and EKOPath Boost have not changed, but this announcement is just concerning EKOPath 4.

I said "yet" because HMPP, as I said before, will be released as an Open Standard, so anybody will be allowed to use it, and in the future it could be implemented on EKOPath compiler suite (or even GNU GCC).

[VinzC]

Hmmmwell, I was about to bet that ?magic? thing was AMD finally releasing all of their GPU driver under an Open License. Or something that was closely related to using GPU's mostly because the impact on most of the ?conventional? usages is said to be unnoticed. Now with that HMPP thing, if true, I'd love to see (and understand) the relation with the rest.

[rvdboom]

If this is true, then the timing with AMD's release of fusion APU is perfect.
There will be an open-source compiler that is able to automatically compile for their APUs, using the on-die GPU to speed up FPU code, from the very begining.

[kacperpl1]

How about a link to the project's homepage? So far it seems too good to be true - I don't believe anything until I've seen proof. Everyone seems to be assuming it's a compiler suite but is this for sure?

The thing is that Michael won't tell us anything more until the company releases it officially at the end of this week/next week or something like that.

[MalleRIM]

How about a link to the project's homepage? So far it seems too good to be true - I don't believe anything until I've seen proof. Everyone seems to be assuming it's a compiler suite but is this for sure?

[Biazzotto]

Dirndl is a Hybrid Compiler

PathScale and CAPS are about to announce the new HMPP Open Standard as you can see here and here.

Definitely I think PathScale open-sourced EKOPath 4 Compiler, as it has one of the best HMPP implementations.

A definition of HMPP are here:

"The HMPP (Hybrid Multicore Parallel Programming) open standard directives are an OpenMP-like approach for explicitly offloading portions of code to the GPU and scaling host code. This approach leverages the strength of the GPU as a hardware accelerator (HWAs) to replace traditional SIMD computing units. PathScale ENZO and CAPS HMPP Workbench compiler are radically different implementations, but both allow the programmer to write hardware independent applications that can significantly speed up C and Fortran code."

[spiritofreason]

@Michael:
Himeno: I'd call that a performance boost of 229%, or a performance boost to 329% of the base speed.
C-Ray: In my book, if the time it takes decreases by 43%, the performance has improved by 75%.

Please correct me if I'm mistaken, since English is not my native tongue.

Can't wait to see what will be behind this Dirndl.

If Michael was referring to speedup, then he has it right for Himeno. I don't think it fits in with normal English, but in computing, speedup is defined by the ratio of execution times. So a speedup of 100% means they have the same performance.

The C-Ray comparison is wrong no matter which jargon he's using.

[Michael]

Usually much better Bavarians and more fun in the Phoronix area of the Augustiner tent

[allquixotic]

[kacperpl1]

What about performance benefits on crappy hardware like Atom? Or even regular systems (something whose cpu doesn't cost more than a typical laptop)?

For low power/low end systems it wouldn't give much speed, however in mid range if only gpu has enough processing units it could quite well accelerate vector and matrix based calculations however those won't normally happen if you are using your notebook/desktop for standard office/internet/multimedia except gaming where it won't matter because games are already using the gpu acceleration.

It will matter mostly for servers and computing workstations.

Oh and btw - @Michael, if you have this dual quad-core opteron system, would you be that nice and check out if this new magic can speed up some mesa software rasterization on cpu's? Maybe it's a bit silly to do this but it would do a nice proof of concept if this can be used for accelerating some system specific tasks.