Compiler error with some MIC intrinsics

[Sebastian_v_]

I'm currently trying to use intrinsics to speed up my application. However I'm having problems with some intrinsics that won't compile. This is some test code:

#pragma offload_attribute(push, target(mic))
#include "immintrin.h"
void foo()
{
  //_mm_clevict(0, _MM_HINT_T0);
  //_mm_prefetch(0, _MM_HINT_ET1);
}
#pragma offload_attribute(pop)

int main()
{
  #pragma offload target(mic:0)
  {
    foo();
  }  
}

If I compile this with the commandline "icc test.cpp" and comment out the _mm_clevict line I'm getting a "undefined reference to `_mm_clevict'" error. If I comment out the _mm_prefetch I'm getting an "internal error: 04010002_1809". However if I compile this as a native application using "icc test.cpp -mmic" everything compiles fine. Also if I move the call to _mm_clevict or _mm_prefetch into the offload region in the main function everything compiles fine. How do I use these intrinsics correctly? My compiler version is: icc version 15.0.0 (gcc version 4.4.7 compatibility)

[Kevin_D_Intel]

I reproduced the internal error and will report that to Development (see internal tracking id below) since that should not happen regardless of the program correctness.

These intrinsics require using the __MIC__ predefine so they only participate in the target-side compilation for the offloaded code. Without this they participate in both the host and target compilations where they will be problematic on the host. So consider something like:

#ifdef __MIC__
  _mm_clevict(0, _MM_HINT_T0);
  _mm_prefetch(0, _MM_HINT_ET1);
#endif

(Internal tracking id: DPD200365512)

[pbkenned1]

Actually, it looks like the ICE is caused by not defining __MIC__, so perhaps that is a workaround. 

I made some slight modifications -- this works for offload or native:

[U538762]$ cat U538762-MIC.cpp
#include <omp.h>
#pragma offload_attribute(push, target(mic))
int global_nt;
#include "immintrin.h"
#include <stdio.h>
void foo()
{
#ifdef __MIC__
  _mm_clevict(0, _MM_HINT_T0);
  _mm_prefetch(0, _MM_HINT_ET1);
#pragma omp parallel
   global_nt = omp_get_num_threads();
  printf("\n Number of threads = % d \n",global_nt);
#endif
}
#pragma offload_attribute(pop)

int main()
{
  #pragma offload target(mic:0)
  {
    foo();
  }
}
[U538762]$

***OFFLOAD MODE****

[U538762]$ icc -V
Intel(R) C Intel(R) 64 Compiler XE for applications running on Intel(R) 64, Version 15.0.1.133 Build 20141023
Copyright (C) 1985-2014 Intel Corporation.  All rights reserved.

[U538762]$ icc -openmp U538762-MIC.cpp -o U538762-MIC-host.x
[U538762]$ ./U538762-MIC-host.x

 Number of threads =  224
[U538762]$

***NATIVE MODE****

[U538762]$ icc -openmp -mmic -wd161 U538762-MIC.cpp -o U538762-MIC-mic.x
[U538762]$
[root@dpdmic09-mic0 pbkenned]# ./U538762-MIC-mic.x

 Number of threads =  228
[root@dpdmic09-mic0 pbkenned]#

Patrick

[Sebastian_v_]

Hi,

thanks for your replies. Soon after posting my question I also realized that the errors were caused by the host compiler, so using #ifdef __MIC__ solves the problem. However I'm still wondering why this wasn't a problem with other intrinsics. Also the internal compiler error only occurs with this specific hint. Using _MM_HINT_ET0, _MM_HINT_T0 and _MM_HINT_T1 as hint for _mm_prefetch is also compiling fine. But I'll just use #ifdef __MIC__ in the future when working with MIC specific intrinsics.