how to use non-temporal (streaming) store instructions to store/load a self-defined struct?

[Emeth Obando]

Hi all, 

I just start to use non-temporal store instructions to store some kinds of data to the memory (could be DRAM or NVM). 

I check out the Intel Intrinsics Guide for such storing functions and I find functions like _mm_stream_si32, _mm_stream_si18, _mm_stream_si256 etc. It seems that these kinds of functions can only be applied to some kinds of integers. 

My question is that if I self-define a certain type of struct and its size may be 1KB, 2KB ...... How can I perform non-temporal (streaming) stores to store such kinds of structs to my memory (or vice versa, load from memory). 

For now, I can only figure out one way, to cast my struct into a chunk of integers, and apply non-temporal/streaming store/load for each of the casted integers one-by-one.

 I think this method is somewhat inefficient, is there a more efficient way of coding to achieve my goal?

Also, if I want to store a large number of such self-defined struct, is it necessary to issue a sfence after every non-temporal store? I am not sure about that and wonder that if I could remove the sfence instruction or just issue one sfence instruction after performing all non-temporal stores?

Moreover, I found that the number of non-temporal streaming (load) functions is very limited. I only found one function, _mm_stream_load_si128, are there any other functions for loading?

Summary: 

• How to use non-temporal (streaming) store instructions to store/load a self-defined struct?

• If one self-defines a certain type of struct and its size may be 1KB, 2KB, how can I perform non-temporal (streaming) stores to store such kinds of structs to memory (or vice versa, load from memory)?
  
  
• Is it necessary to cast a struct into a chunk of integers, and apply non-temporal/streaming store/load for each of the casted integers one-by-one?

• If one wants to store a large number of such self-defined struct, is it necessary to issue a sfence after every non-temporal store? Is it possible to remove the sfence instruction or just issue one sfence instruction after performing all non-temporal stores?

• The number of non-temporal streaming (load) functions is very limited, only one function, _mm_stream_load_si128. Are there any other functions for loading?

Many thanks for the help :) 

[ppbb...@gmail.com]

Hi,

I would recommend against rolling your own low-level mechanisms like that. Instead, you can leverage nontemporal memcpy implemented in libpmem/libpmem2.

Here's an example that shows its usage:

https://github.com/pmem/pmdk/blob/master/src/examples/libpmem2/ringbuf/ringbuf.c#L289

This way you don't have to deal with all the details you've listed in your question.

Alternatively, you can find a lot of useful information (including answers to your questions) in the x86 Architectures Optimization Manual:

https://software.intel.com/content/www/us/en/develop/download/intel-64-and-ia-32-architectures-optimization-reference-manual.html

Look at sections 9.6 and 15.16.

Piotr

[Emeth Obando]

Hi Piotr and all, 

Thank you for the information. 

I am wondering if I should issue a `fence` instruction whenever I perform one such non-temporal `memcpy` (or non-temporal functions like, _mm_stream_si32, _mm_stream_load_si128, etc )?  

[ppbb...@gmail.com]

The implementation found in libpmem/libpmem2 performs all the necessary operations to make sure everything works.

If you want to use the intrinsics manually - as I said, I recommend familiarizing yourself with the relevant parts of the manual linked above. But yes, you need to use appropriate fencing operations with non-temporal stores. This is explained in more detail in section 9.4.1.1.

Piotr