Only commutative reductions can be parallelized

[Elarnon]

Because parallelization changes the order of computation within the reduction, parallelizing associative but non-commutative reductions can result in (non-deterministically) incorrect results in the same way reordering them can.

For instance Halide currently accepts the following code, but generates non-deterministic outputs on GPU. On CPU with .parallel(r.x), OpenMP rejects the generated code (correctly) stating that the #pragma omp atomic is invalid for the same reasons.

#include <stdio.h>

#include "Halide.h"

using namespace Halide;

int main(int argc, char **argv) {
        Halide::Func A("A"), B("B");
        Halide::Var i("i");

        A(i) = i;
        B() = -1;
        Halide::RDom r(0, 1024);
        B() = A(r.x);

        A.compute_root();
        B.update().atomic().gpu_blocks(r.x);

        B.compile_jit(get_host_target().with_feature(Target::CUDA));
        Halide::Buffer<int32_t> b = B.realize();
        printf("%d\n", b());

        return 0;
}

[abadams]

I believe this is correct. Thanks for the fix.