confirm k1om support via test

[ghost]

greetings @edanor
Please consider implementation of benchmark test of non-vector code vs fully optimized code for k1om.
k1om is flooding market, we need to see (in seconds) how this library rocks! [I already know from usage].

Thanks,
coast

[edanor]

Hi,

I'm glad that you find this library useful! Hope it works well for you.

Unfortunately the reason why KNC is flooding market is that it has been EOL'ed ( see: https://ark.intel.com/products/series/92649/Intel-Xeon-Phi-x100-Product-Family ). This is the reason why computer centers are getting rid of it.

Because of AVX512 already released, the whole k1om ISA is unlikely to re-appear. In other words: making any code developments in that direction will only create a long term burden on the library. Because the library only covers code generation, you will still have to work hard on your code to make sure it works on KNC. For instance all loads have to be aligned or else you would pay a HUGE performance penalty. So even if it is cheap hardware, long term you might save some money (and a lot of time) on software development if you choose AVX512 machine or CUDA-enabled coprocessor.

Another, more practical problems are that I don't have the access to hardware to verify whether the code works as expected and I would have to spend my private time on developing this code, which I personally don't need (I mean the benchmarks + KNC plugin, not the whole library;]).

Since this is a free software you have two choices:

1. write the code on your own and submit a pull request. I am more than willing to review and accept most of modifications and guide you through the development process if needed.
2. make a donation so that I can pay somebody competent for making this modification for you. Send me a message at [email protected] if you are interested so that we can discuss any details.

Cheers,
Przemek.

[ghost]

Greetings Przemek, I appreciate the explanation. Perhaps it is time to abandon the KNC and move on CUDA/OpenCL. Thanks!

…

On Thu, Nov 16, 2017 at 9:23 AM, Przemysław Karpiński < ***@***.***> wrote: Hi, I'm glad that you find this library useful! Hope it works well for you. Unfortunately the reason why KNC is flooding market is that it has been EOL'ed ( see: https://ark.intel.com/products/series/92649/Intel- Xeon-Phi-x100-Product-Family ). This is the reason why computer centers are getting rid of it. Because of AVX512 already released, the whole k1om ISA is unlikely to re-appear. In other words: making any code developments in that direction will only create a long term burden on the library. Because the library only covers code generation, you will still have to work hard on your code to make sure it works on KNC. For instance all loads have to be aligned or else you would pay a HUGE performance penalty. So even if it is cheap hardware, long term you might save some money (and a lot of time) on software development if you choose AVX512 machine or CUDA-enabled coprocessor. Another, more practical problems are that I don't have the access to hardware to verify whether the code works as expected and I would have to spend my private time on developing this code, which I personally don't need (I mean the benchmarks + KNC plugin, not the whole library;]). Since this is a free software you have two choices: 1. write the code on your own and submit a pull request. I am more than willing to review and accept most of modifications and guide you through the development process if needed. 2. make a donation so that I can pay somebody competent for making this modification for you. Send me a message at ***@***.*** ***@***.***>* if you are interested so that we can discuss any details. Cheers, Przemek. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#73 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGtt7fal3Fjq0IamTRTEqBm0offGbgCDks5s3EVrgaJpZM4QVgrt> .