Andrew Adams
Andrew Adams
The general scattering problem is really no worse than simple indirection on the LHS. Of all the computed values you could put there, a load is the least well-behaved. It's...
Ah, indeed. Scattering to affine or constant addresses is way easier.
The bulk of the work in the Halide autodiff was handling non-trivial indexing in ways that preserve parallelism. E.g. consider: A(i) = 2 * B(some nasty expression in i) You...
Equation solvers exist once you get down to Halide IR or polyhedral IR. It would be silly to implement yet another one at the TC-front-end-IR level. I'd just allow complex...
Sorry, snowed under with Halide work I neglected during the TC release.
Note that the Halide simplifier will convert all the simpler bit operations to multiplication (x > 5), and modulo (x & 0x3f), so the only bit operations that reach isl...
Note that this would also enable the following formulation of a concat: out(0, i) = a(i) out(1, i) = b(i)
Currently blocked by the many places that assume the bounds on tensors are affine. Scattering makes them at least piecewise affine, which we could probably handle, but in general makes...
What are the contexts in which TEST_TMPDIR is not set? Is it just running a test manually in the shell? When I'm doing that, I like output files to go...
Actually I guess I'm not sure what TEST_TMPDIR is used for. Is it used for intermediate diagnostic outputs for things like simd op check that a human may read? Or...