[QST] Get a slice of the Tensor while keeping the dimension

[phantaurus]

What is your question?
Hello!

Would there be easy ways to get a slice of the Tensor while keeping the dimension?

I have a tensor with shape (16, 4, 16)

When doing

for(int idx = 0; idx < size<1>(tensor); ++idx) {
    slice = tensor(_, idx, _)...
}

This generates slice with shape (16, 16).
Would it be possible to get (16, 1, 16) instead?

Thanks!

[thakkarV]

Tile it first using logical or zipped divide that takes the modes you care about and then slice it. In this case your tiler can be (16,1,16)

[phantaurus]

Thank you so much for your reply!

I have experimented with partitions:

Tensor acc = make_tensor(make_layout(Shape<Shape<_2, _2>, _2, Int<16>>{}, Stride<Stride<_1, _2>, Int<4>, Int<8>>{}));
Layout acc_layout = acc.layout();
Layout acc_tiler = make_layout(get<0>(acc_layout), Layout<Shape<_1>>{}, get<2>(acc_layout));

Tensor partitioned_acc = logical_divide(acc, acc_tiler);

for(int idx = 0; idx < size<1>(partitioned_acc); ++idx) {
        Tensor acc_iter = partitioned_acc(make_coord(_,_,_), idx);
}

In this example,
acc has layout ((_2,_2),_2,_16):((_1,_2),_4,_8)
acc_tiler is ((_2,_2),(_1),_16):((_1,_2),(_0),_8)
partitioned_acc is o (((_2,_2),(_1),_16),_2):(((_1,_2),(_0),_8),_4)
acc_iter is o ((_2,_2),(_1),_16):((_1,_2),(_0),_8)

My first question is would it be possible to get rid of the parentiesis for (_1) in acc_tiler and make it ((_2,_2),_1,_16):((_1,_2),_0,_8)?
-- This is solved by defining

Layout acc_tiler = make_layout(get<0>(acc_layout), Layout<_1>{}, get<2>(acc_layout));

Second question is what's the meaning of the prefixing o in the layout of partitioned_acc and acc_iter?

Thanks!

[phantaurus]

Another question is

Layout acc_tiler_new = make_layout(get<0>(acc_layout), Layout<Shape<_1>>{},Layout<Shape<_1>>{});
Tensor partitioned_acc_new = logical_divide(acc, acc_tiler_new);

acc_tiler_new is ((_2,_2),(_1),(_1)):((_1,_2),(_0),(_0))
partitioned_acc_new is o (((_2,_2),(_1),(_1)),_32):(((_1,_2),(_0),(_0)),_4)

However, I would expect partitioned_acc_new to be o (((_2,_2),(_1),(_1)), (_2, _16)):(((_1,_2),(_0),(_0)), (_4, _8))
Do you know how to avoid logical_divide from automatically merging the _2 and _16 dimensions into one _32 dimension?

Thanks!

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.