sum.cpp

  // Compute the sum of f in [0, extent_x) x [0, extent_y)
  Func s("s");
  RDom r(0, extent_x, 0, extent_y);
  s() += f(r.x, r.y);

  // Schedule: rfactor the reduction in to vector_width x 4 tiles.
  // This gives 4-way instruction level parallelism with SIMD.
  // Requires that the extents are a multiple of the tile size.
  RVar rxo, rxi, ryo, ryi;
  Var x, y;
  s.compute_root().update(0)
      .tile(r.x, r.y, rxo, ryo, rxi, ryi,
            target.natural_vector_size<float>(), 4)
      .rfactor({{rxi, x}, {ryi, y}})
      .compute_at(s, Var::outermost())
      .vectorize(x)
      .update(0)
      .vectorize(x)
      .unroll(y);