topic However doing this fixed the in OpenCL* for CPU

Using lookup table on reduction kernel produces incorrect results on GPU

Joose_S_ — Wed, 01 Jun 2016 10:26:43 GMT

I have a kernel that takes in an array of integers and returns the index of the smallest element.

#ifndef LOCAL_SIZE
#define LOCAL_SIZE 8
#endif // LOCAL_SIZE

kernel void test( global int* in, global int* out )
{   
  int id = get_local_id(0);

  local int indx[LOCAL_SIZE];

  int temp = id;
  for (int i = id; i < 1024; i += LOCAL_SIZE)
  {
    temp = in < in[temp] ? i : temp;
  }
  indx[id] = temp;
  barrier(CLK_LOCAL_MEM_FENCE);

  for(int i = LOCAL_SIZE / 2; i!= 0; i>>=1)
  {
    if(id < i)
    {
      printf( "%4d: %3d, %4d: %3d\n", indx[id], in[indx[id]], indx[id + i], in[indx[id + i]] );
      indx[id] = in[indx[id]] < in[indx[id+ i]] ? indx[id] : indx[id + i];
    }
    barrier(CLK_LOCAL_MEM_FENCE);
    if(id == 0) printf("\n");
  }
  out[0] = indx[0];
}

Before the first barrier each work item finds its smallest value and places it into a local buffer. Everything works fine here.

In the for loop the results from each work item is reduced further to find the result. However the second to last iteration fails on GPU everytime: in[indx[id]] and in[indx[id + i]] both return the same value.

Operating system: Windows 7 Enterprise

Device Driver Version: 10.18.14.4280

Device: Intel HD 4600 & Processor Intel i5-4590

Works fine on CPU and Nvidia GTX 970

I've attached the kernel and host code to reproduce

I am afraid there might be in

Robert_I_Intel — Wed, 01 Jun 2016 21:50:00 GMT

I am afraid there might be in race in your code. What about this:

#ifndef LOCAL_SIZE
#define LOCAL_SIZE 8
#endif // LOCAL_SIZE

kernel void test( global int* in, global int* out )
{   
  int id = get_local_id(0);

  local int indx[LOCAL_SIZE];

  int temp = id;
  for (int i = id; i < 1024; i += LOCAL_SIZE)
  {
    temp = in < in[temp] ? i : temp;
  }
  indx[id] = temp;
  barrier(CLK_LOCAL_MEM_FENCE);

  for(int i = LOCAL_SIZE / 2; i!= 0; i>>=1)
  {
    int tmp = indx[id], tmpi = indx[id + i];
    int val = in[tmp], vali = in[tmpi];
    barrier(CLK_LOCAL_MEM_FENCE);
    if(id < i)
    {
      printf( "%4d: %3d, %4d: %3d\n", tmp, val, tmpi, vali );
      indx[id] = val < vali ? tmp : tmpi;
    }
    barrier(CLK_LOCAL_MEM_FENCE);
    if(id == 0) printf("\n");
  }
  out[0] = indx[0];
}

Quote:Robert I. (Intel) wrote

Joose_S_ — Thu, 02 Jun 2016 06:03:00 GMT

Robert I. (Intel) wrote:

I am afraid there might be in race in your code. What about this:

Nope, the issue still persist even with this change.

However doing this fixed the

Joose_S_ — Fri, 03 Jun 2016 10:48:09 GMT

However doing this fixed the problem:

#ifndef LOCAL_SIZE
#define LOCAL_SIZE 8
#endif // LOCAL_SIZE

kernel void test( global int* in, global int* out )
{   
  int id = get_local_id(0);

  local int indx[LOCAL_SIZE];

  int temp = id;
  for (int i = id; i < 1024; i += LOCAL_SIZE)
  {
    temp = in < in[temp] ? i : temp;
  }
  indx[id] = temp;
  barrier(CLK_LOCAL_MEM_FENCE);

  for(int i = LOCAL_SIZE / 2; i!= 0; i>>=1)
  {
    int tmp = indx[id]; 
    int val = in[tmp];
    barrier(CLK_LOCAL_MEM_FENCE);
    int tmpi = indx[id + i];
    int vali = in[tmpi];
    if(id < i)
    {
      printf( "%4d: %3d, %4d: %3d\n", tmp, val, tmpi, vali );
      indx[id] = val < vali ? tmp : tmpi;
    }
    barrier(CLK_LOCAL_MEM_FENCE);
    if(id == 0) printf("\n");
  }
  out[0] = indx[0];
}

I just don't quite understand why the program behaves the way it does.

Ok, looks like a compiler

Robert_I_Intel — Tue, 07 Jun 2016 01:30:39 GMT

Ok, looks like a compiler team needs to take a look at this one :)

Hi Robert,

Joose_S_ — Mon, 20 Jun 2016 06:59:46 GMT

Hi Robert, Have you found out anything regarding the issue?

Hi Joose,

Robert_I_Intel — Tue, 21 Jun 2016 14:42:06 GMT

Hi Joose,

I showed this to the compiler folks and they think this is a compiler bug. Will file a bug. Thanks!