multiply behave differently for sparse and array2d

[Lifeng Jin]

Hi, I have been using CUSP and noticed the following behavior. The following example gives the correct output only when B is used in multiply. When A is used, the result is all zeros. Any ideas?

int main()

{

  int state_size = 15;

  int batch_size = 2;

  // allocate storage for (10,10) matrix with 5 nonzeros

  cusp::csr_matrix<int,float,cusp::device_memory> A(state_size, state_size, 5);

  // initialize matrix entries on host

  A.row_offsets[0] = 0;  // first offset is always zero

  A.row_offsets[1] = 0;

  A.row_offsets[2] = 1;

  A.row_offsets[3] = 2;

  A.row_offsets[4] = 3;

  A.row_offsets[5] = 4;

  A.row_offsets[6] = 5; A.row_offsets[7] = 5; A.row_offsets[8] = 5; A.row_offsets[9] = 5; A.row_offsets[10] = 5; A.row_offsets[11] = 5;

  A.row_offsets[12] = 5; A.row_offsets[13] = 5; A.row_offsets[14] = 5; A.row_offsets[15] = 5;

  

  A.column_indices[0] = 0; A.values[0] = 0.19f;

  A.column_indices[1] = 0; A.values[1] = 0.2f;

  A.column_indices[2] = 0; A.values[2] = 0.21f;

  A.column_indices[3] = 0; A.values[3] = 0.22f;

  A.column_indices[4] = 0; A.values[4] = 0.18f;

  

  cusp::array2d<float,cusp::device_memory> B(A);

  cusp::print(B);

  

  cusp::array2d<float, cusp::device_memory> prev_mat(state_size, batch_size, 0.0f);

  cusp::array2d<float, cusp::device_memory> next_mat(state_size, batch_size, 0.0f);

  for(int i = 0; i < batch_size; i++){

      prev_mat(0,i) = 1.0f;

  }

  

  multiply(A, prev_mat, next_mat);  // this one gives the wrong answer

  // multiply(B, prev_mat, next_mat);   // this is fine.

  cout << "Previous matrix:" << endl;

  cusp::print(prev_mat);

  cout << "Next matrix:" << endl;

  cusp::print(next_mat);

}

[Steven Dalton]

Thanks, there was a shared memory bug in the CUDA implementation, verify the update I committed fixes your issue. CSR * DENSE is still somewhat experimental. If you run into additional errors please file a bug report on Github.

Steve

--
You received this message because you are subscribed to the Google Groups "cusp-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cusp-users+unsubscribe@googlegroups.com.
To post to this group, send email to cusp-...@googlegroups.com.
Visit this group at https://groups.google.com/group/cusp-users.
For more options, visit https://groups.google.com/d/optout.

[Lifeng Jin]

Thanks, it works now. I was not entirely sure if it was a bug or I was not understanding something correctly. One more question, for generalized_spgemm, all three matrices need to be of the same dense/sparse type to work right?

在 2016年11月19日星期六 UTC-5上午3:24:40，Steve写道：

To unsubscribe from this group and stop receiving emails from it, send an email to cusp-users+...@googlegroups.com.