2014-09-15 15:13 GMT+02:00 phil@highoctane.be <phil@highoctane.be>:

On Mon, Sep 15, 2014 at 2:33 PM, Thierry Goubier <thierry.goubier@gmail.com> wrote:

Hum, I would more like SSE and AVX code done this way: matrix multiplications, bitmap processing, heavily used code, SciSmalltalk stuff on very large datasets.

Indeed! I'd live to generate some CUDA code from Pharo. But the toolchain goes C code which gets compiled through some "black-hole toolchain" in the NVidia SDK with host and GPU blocks.��

You may be interested to contact Bernard Pottier. His NetGen toolkit does CUDA code generation. NetGen is VisualWorks only for the moment, even if I know someone who would like to have it in Pharo (well, two guys at least).

��

That's too much to handle for my tastes, so the best is for me to have an engine (or for a better metaphor) and using Pharo to drive/steer/harness it.

Yes. I intended to do things around that, but didn't found the time.

��

Indeed! CUDA =-)��

OpenCL, SPIR code generation and link to Cog.

The ability to generate better code than when targetting C.

True too. But aren't those compilers smart enough already? With current CPUs, well, it takes a strong man to do better things than the compiler it seems. Or am I mistaken?

You are. A lot may be done if your input language is not C, but this is not allways relevant to the Pharo situation.

I should know a lot more about LLVM-IR generation in a few months time. I'll say if I see easy gains when doing Smalltalk code.

��

��

Ah, I should have taken the research career :-)

Hum, you know that researchers are allways on the lookout for SMEs to collaborate on projects :-)

Duly noted :-) Feel free to tickle me :-)��

Noted :)

Thierry