topic Re:Issue when multiplying 0 to large arrays in Intel® Integrated Performance Primitives

Issue when multiplying 0 to large arrays

Animesh — Mon, 26 Jun 2023 10:43:22 GMT

We came to find that multiplication of an array (F32, F64, I32) with a scalar of 0 leaves (most of) the array untouched. For F32, this behavior starts with array sizes of 2^29 and up; for F64, it starts at 2^28 already. For I32 also, the behavior starts from 2^29. We have also tried CF32 and CF64 but they seem to work fine (checked until 2^30).

Description	IPP Function	Number of elements at which issue occurs
F32 in place	ippsMulC_32f_I	2^29
F64 in place	ippsMulC_64f_I	2^28
I32 in place	ippsMulC_32s_Isfs	2^29
CF32 in place	ippsMulC_32fc_I	no issue observed up to 2^29
CF64 in place	ippsMulC_64fc_I	no issue observed up to 2^29
F32 out of place	ippsMulC_32f	2^29
F64 out of place	ippsMulC_64f	2^28
I32 out of place	ippsMulC_32s_Sfs	2^29
CF32 out of place	ippsMulC_32fc	no issue observed up to 2^30
CF64 out of place	ippsMulC_64fc	no issue observed up to 2^30

I have attached a standalone C++ code which reproduces the behavior.
The behavior was checked on IPP 2021.8.0.

The example code creates an array of fixed size. It runs correctly for F32 (until size 2^29-1), but will show the error for F64. Any larger array size will show the error for both.

Re: Issue when multiplying 0 to large arrays

Wobo — Tue, 27 Jun 2023 08:34:03 GMT

Ok, I am no Intel staff. But I assume it has to do with that:

https://www.intel.com/content/www/us/en/docs/ipp/developer-guide-oneapi/2022-2/ipp-using-platform-aware-functions.html

This burden of the 32-bit size limitation of many many IPP functions is quite often dangerous for our 64-bit applications, where we have to deal with huge data sets. Creation of some platform aware 64-bit enabled functions are a good start by Intel, but my wish is make all IPP function properly usable in a 64-bit environment.

Kind Regards

Re: Issue when multiplying 0 to large arrays

Gennady_F_Intel — Tue, 27 Jun 2023 11:30:58 GMT

Yes, that's correct that not all IPP functions are available for platform - aware extensions ( to support of 64-bit integer size). In particularly, this ippsMulC_*.* is not available for platform-aware capabilities. We are continue to extend the list of such functions from case by case.

Re:Issue when multiplying 0 to large arrays

ShanmukhS_Intel — Wed, 28 Jun 2023 12:55:54 GMT

Hi Animesh Baranawal,

Thanks for posting in Intel communities.

Thanks for reporting this issue. We were able to reproduce the same at our end and we have informed the development team about it. We will get back to you with an update soon.

Best Regards,

Shanmukh.SS