Will "hoist array store" be supported later?

[were]

Consider the simple convolution.

for (int i = 0; i < 10; ++i) {
    sum[i] = 0;
}
for (int i = 0; i < 10; ++i) {
    for (int j = 0; j < 3; ++i) {
        sum[i] = sum[i] + a[i+j] * k[j];
    }
}
for (int i = 0; i < 10; ++i) {
    conv[i] = sum[i]
}

In the second stage of computation, the Store of array sum can be hoisted.

for (int i = 0; i < 10; ++i) {
    int _sum = sum[i];
    for (int j = 0; j < 3; ++i) {
        _sum = _sum + a[i+j] * k[j];
    }
    sum[i] = _sum;
}

[abadams]

This is a different loop nest structure (the three outer loops have been fused), so you get at it using scheduling. We never change the loop structure specified by the schedule, so this optimization is not something Halide should do automatically. You want:

sum.compute_at(conv, i)

If there's no wrapper function to schedule it within, and it's called directly, you can use Func::in to make a wrapper:

sum.compute_at(sum.in(), i);

[were]

It a new scheduling function, right?
I haven't updated my Halide for a long time.
Sorry.

[abadams]

Func::in is new. compute_at has existed a long time.

…

On Mon, Dec 5, 2016 at 4:25 PM, Jian Weng ***@***.***> wrote: It a new scheduling function, right? I haven't updated my Halide for a long time. Sorry. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <https://github.com/halide/Halide/issues/1661#issuecomment-265022881>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAfdRr-VFVnkO5-riwXvJYAeoC6iUD9Xks5rFKuVgaJpZM4LEnof> .