Don't emit atomics for thread-safe assignment patterns

[Infinoid]

This attempts to fix #316.

It prevents a #pragma omp atomic from being emitted for the assignment to y_vals[i] in this case:

$ bin/taco 'y(i) = A(i, j) * x(j)' -f=A:ds -s='parallelize(i, CPUThread, Atomics)'
// Generated by the Tensor Algebra Compiler (tensor-compiler.org)

int compute(taco_tensor_t *y, taco_tensor_t *A, taco_tensor_t *x) {
  int y1_dimension = (int)(y->dimensions[0]);
  double* restrict y_vals = (double*)(y->vals);
  int A1_dimension = (int)(A->dimensions[0]);
  int* restrict A2_pos = (int*)(A->indices[1][0]);
  int* restrict A2_crd = (int*)(A->indices[1][1]);
  double* restrict A_vals = (double*)(A->vals);
  int x1_dimension = (int)(x->dimensions[0]);
  double* restrict x_vals = (double*)(x->vals);

  #pragma omp parallel for schedule(runtime)
  for (int32_t i = 0; i < A1_dimension; i++) {
    double tjy_val = 0.0;
    for (int32_t jA = A2_pos[i]; jA < A2_pos[(i + 1)]; jA++) {
      int32_t j = A2_crd[jA];
      tjy_val += A_vals[jA] * x_vals[j];
    }
    y_vals[i] = tjy_val;
  }
  return 0;
}

Because each thread gets its own value of i, no other thread will assign to that location, and recent compilers treat an atomic pragma as invalid there. See #316 for details.

Are there any other cases this doesn't cover? All tests pass for me when configuring with -DOPENMP=ON -DPYTHON=ON.

[stephenchouca]

This patch seems to generate incorrect code for expressions like 'y(j) = A(i,j,k) * x(k)' -s='parallelize(j, CPUThread, Atomics)', which still require atomics even after scalar promotion (the generate code omits the required atomic pragma).

More fundamentally though, I wonder if we should instead just fix the lowerer so that it won't emit unnecessary atomic pragmas for (non-reduction) assignments. That way, the bug would be fixed even if, for instance, a user decides to manually invoke the lowerer with concrete index notation that has an assignment inside a Forall statement synchronized with atomics. This should only require modifying a few lines in lowerAssignment; I can take a stab at it if you'd like.

[Infinoid]

This should only require modifying a few lines in lowerAssignment; I can take a stab at it if you'd like.

Sure, please do. And I'll happily add your example as a test case.

[stephenchouca]

Sure, please do. And I'll happily add your example as a test case.

That'd be great, thanks!

[Infinoid]

Your patch works, I added the test case in #402. Thanks!

Canceling PR in favor of 0a48d17.