[Cuis-dev] OMeta working again for Cuis 6.3

[Michał Olszewski]

Hi Juan, Hi All,

In the upcoming weeks I'm gonna try to draft up a document about 
preliminary design to help myself with overall direction it's heading to 
and to show what I have in my mind.

How about a backend for my language that auto generates VM plugins with 
selected, optimized routines hehe.

Cheers,
Michał

On 14.04.2024 15:37, Juan Vuletich wrote:
> Hi Luciano, Michal,
>
> Michal, I look forward to know more about your project. Sounds like 
> useful, interesting, and fun!
>
> On 4/14/2024 12:14 AM, Luciano Notarfrancesco via Cuis-dev wrote:
>> Hi Michal,
>> That’s very interesting. I have a very concrete application for this. 
>> I need to compute with tuples, matrices and polinomials over integers 
>> modulo a small modulus (say, it fits in a register, or half 
>> register). These tuples, polynomials and matrices are stored in 
>> arrays (for example WordArrays, 32 bits per entry). The most critical 
>> operation is matrix multiplication, since a lot of the rest can 
>> reduce to matrix multiplication. I’ve been experimenting with a small 
>> external library written in C that converts to floating point, does 
>> matrix multiplication with a BLAS library, and converts back to 
>> modular integers, and i call it via FFI. Of course I’d prefer to do 
>> all in Cuis… so i’m looking forward for your project!
>>
>> Regards,
>> Luciano
>>
>
> Still, Luciano, I'd do it differently. I'd write a VM plugin using  
> gcc SIMD extensions if available. If coded properly, it should give 
> similar performance to BLAS for big matrices, and much better for 
> small ones (because calling plugin primitives is way faster than 
> calling FFI).
>
> If it is only about best performance for big matrices, I'd write it in 
> OpenCL, and maybe use the GPU (if it outperforms CPU).
>
> Cheers,
>