Fixing d-tiler

[rajopadhye]

I was speaking to Campbell about the options for him to be able to get wavefront parallelized d-tiled code and the bug that was previously fixed in an ad-hoc way previously. He suggested that could try to fix d-tiler, and we had a brief exchange -- the fix is conceptually really simple, IMHO. Here's what I sent him

Hi Campbell,

Here’s the link to the pdf of his dissertation. You want to look specifically at Ch 8 (starting page 82). the error is in the two formulae for computing the first and last tile. What is needed is to:

1. Compute the vertices of the outset (not from the loop bounds, but rather, by an isl call — these are all parametric function of the program parameters and tile sizes.

2. Then, first and last are given by:

first = min_i \sum_j frac{V_i_j}{s_j}
last = max_i \sum_j frac{V_i_j}{s_j}

v_i_j is the j’th dimension of the ith vertex, and s_j is the tile size in dimension j.

Oe thing I forgot to ass is that the lower bounds have to be rounded up to the next multiple of the tile size in that dimension.