Scatter-updates to global memory and false sharing

[Björne2]

I have an array of unique indexes. Let's call it I. So e.g I = {3, 9, 15, 22, 23, 24, ...}. I want to update all elements of a global float buffer G matching those indexes. Like this:

for (int i = 0; i < n; i++) {

    G[I[i]] += 123.4;

}

If all indexes in I are unique, is it ok to use #pragma ivdep on the loop, or will that lead to false sharing problems? The reason I ask is because as the attached screenshot shows the loop has an II of 338 which is terrible. I suspect that #pragma ivdep is not ok here, but I'm unsure of why. If #pragma ivdep is not ok are there other ways to avoid the terrible high II?

[Björne2]

Just to clarify what I mean by false sharing. If I have the loop iterations G[3] += 123.4;  and G[4] += 123.4; they "should" be independent. But perhaps if the indexes 3 and 4 refer to the same page these updates could nevertheless interfere with each other.

[aikeu]

Hi Björne2,

Are your question related to OneAPI?

Thanks.

Regards,

Aik Eu

[Björne2]

Maybe a little but it is more about how writes to DDR are implemented on Intel FPGAs.

[aikeu]

Hi Björne2,

Probably this link below may help to answer some of your question.

https://www.youtube.com/watch?v=1zGpN28mXN4

Thanks.

Regards,

Aik Eu

[aikeu]

Hi Björne2,

I will close the thread if no further question.

Thanks.

Regards,

Aik Eu

[aikeu]

Hi Björne2,

I am closing the thread for now as we do not receive any response from you on the previous question/reply/answer that we have provided. Please login to ‘https://supporttickets.intel.com’, view details of the desire request, and post a feed/response within the next 15 days to allow me to continue to support you. After 15 days, this thread will be transitioned to community support. The community users will be able to help you on your follow-up questions.

Thanks.

Regards,

Aik Eu