Problem with Scatter/Gather operations

shiva_rama_krishna_b — Sat, 10 May 2014 11:08:49 GMT

Hi, i am vectorizing some code using the MIC intrinsics. But i am getting segmentation fault on MIC (offload error: process on the device 0 was terminated by signal 11 (SIGSEGV) ). Some one please tell me where am i doing mistake. It seems like a simple error. but i am stuck here.

Normal code:

for(curIdx = 0; curIdx < index; curIdx++)
       {
       outreal[cidxarray[curIdx]] += inrealvalue*valuearray[curIdx];
       }

Vectorized code:

int remain = index %16;
int indexSize = index-remain;
v1 = _mm512_set1_ps(inrealvalue);

for(int curIdx = 0; curIdx < indexSize; curIdx = curIdx+16)

{

   v1i = _mm512_load_epi32(cidxarray+curIdx);
   v2 = _mm512_load_ps((void*)(valuearray+curIdx));

   v3 = _mm512_mul_ps(v1,v2); // doing inrealvalue*valuearray[curIdx];
   v4 = _mm512_i32gather_ps(v1i,(void*)outreal,sizeof(float)); // access values from outreal and store in v4;
   v5 = _mm512_add_ps(v3,v4);   // doing outreal[cidxarray[curIdx]] += inrealvalue*valuearray[curIdx] and stores in v5
   _mm512_i32scatter_ps((void*)outreal,v1i,v5,sizeof(float)); // storing it back into outreal array

}
  for(curIdx = indexSize; curIdx < index; curIdx++)
       {
       outreal[cidxarray[curIdx]] += inrealvalue*valuearray[curIdx];
       }

Thanks

sivaramakrishna

Hard to tell. Insert some

jimdempseyatthecove — Sat, 10 May 2014 12:23:28 GMT

Hard to tell. Insert some ASSERTs to assert that the arrays are cache line aligned.

If this doesn't expose something then insert some trace code to see at which point it fails (though this may alter the behavior).

Jim Dempsey

Are you saying you don't want

TimP — Sun, 11 May 2014 17:02:21 GMT

Are you saying you don't want to vectorize the C (e.g. with directives or CEAN)? Doing so would give you something to compare against.

topic Are you saying you don't want in Software Archive

Problem with Scatter/Gather operations

Hard to tell. Insert some

Are you saying you don't want