Junk values return while doing inclusive_scan for pair

[GoogleCodeExporter]

Iam trying to do a basic operation between 2 pairs. Iam getting junk results 
sometimes.

INPUT PAIR             OUTPUT PAIR
data  d_vec            data     d_vec
0       3.4e+38         0       3.4e+38 
2       1.16667 2       1.16667
2       3.4e+38 2       3.4e+38
1       3.5             1       3.5
1       3.4e+38 1       3.4e+38
2       5.5             2       5.5
2       3.4e+38 2       3.4e+38
1       7.5             1       7.5
1       3.4e+38 1       3.4e+38
2       9.5                 2       9.5
0       3.4e+38       0       5.5 -------------------
2       3.4e+38       2       3.4e+38
2       3.4e+38       2       3.4e+38

struct parallel_scan
{
  template <typename Pair1, typename Pair2>
  __host__ __device__
    Pair1 operator()(const Pair1 &x, const Pair2 &y)
  {

     if(( y.second == 0) && (x.second == 2))
        return thrust::make_pair(x.first,y.second);
     else
        return thrust::make_pair(y.first,y.second);

   } // end operator()
}; // end add_pairs

From main I call the following:

typedef thrust::pair<float,int>P ;
thrust::device_vector<P> d_result(d_vec.size());
thrust::device_vector<P> final_pair(d_vec.size());
thrust::transform(d_vec.begin(), d_vec.end(), data.begin(), 
d_result.begin(),make_pair_functor());
thrust::inclusive_scan(d_result.begin(), d_result.end(), final_pair.begin(), 
parallel_scan());

My concerns:
1) For the underlined  I should never get 5.5 .
In any case i should get either 9.5 or 3.4e+38. But Iam getting  a value that 
is passed long back. And now iam getting juk value as my result.

Original issue reported on code.google.com by vinitha....@gmail.com on 18 Nov 2010 at 7:45

[GoogleCodeExporter]

Hi Vinitha, thanks for the report.

If you could post a complete, self-contained code sample which reproduces the 
error, we'll be happy to look into it.

Original comment by jaredhoberock on 18 Nov 2010 at 7:49

[GoogleCodeExporter]

Hi Attaching the complete file. You just need to compile.

Original comment by vinitha....@gmail.com on 18 Nov 2010 at 8:32

Attachments:

• Output_ScanPair.o2584
• scanPair.cu

[GoogleCodeExporter]

Hi , Do you find a problem in there????

Original comment by vinitha....@gmail.com on 19 Nov 2010 at 1:47

[GoogleCodeExporter]

Hi Vinitha,

Actually, this result is correct, although not intuitive.  To parallelize scan 
we require that the binary operation be associative.  In this case the scan 
function is computing the partial sums
  P[0] = A[0];
  P[1] = P[0] op A[1];
  P[2] = P[1] op A[2];
  ...
  P[5] = P[4] op A[5];
and
  P[6] = A[6];
  P[7] = P[6] op A[7];
  P[8] = P[7] op A[8];
  ...
  P[11] = P[10] op A[9];
By associativity the final values are then computed as
  B[0] = P[0]
  B[1] = P[1]
  ...
  B[5] = P[5]
  B[6] = P[5] + P[6]
  B[7] = P[5] + P[7]
  ...
  B[10] = P[5] + P[10]
  B[11] = P[5] + P[11]
This is why the value of A[5] shows up in the result of B[10].  

The underlying problem is that your functor is not associative.  For example
  F( F({X,2}, {Y,1}), {Z,0}) = F({Y,1}, {Z,0}) = {Z,0}
but
  F( {X,2}, F({Y,1}, {Z,0})) = F({X,2}, {Z,0}) = {X,0}

Original comment by wnbell on 21 Nov 2010 at 5:42

• Changed state: Invalid
• Added labels: Milestone-Release1.4