ez complete all of the callback implementations for Nx.Backend

[ ] iota
[ ] random_uniform
[ ] random_normal
[ ] to_batched_list
[x] to_binary
[ ] backend_deallocate
[ ] backend_transfer
[ ] tensor
[x] inspect
[x] from_binary
[ ] as_type
[ ] reshape
[ ] squeeze
[ ] broadcast
[ ] transpose
[ ] pad
[ ] reverse
[ ] dot
[ ] clip
[ ] slice
[ ] concatenate
[ ] select
[ ] conv
[ ] all?
[ ] any?
[ ] sum
[ ] product
[ ] reduce_max
[ ] reduce_min
[ ] argmax
[ ] argmin
[ ] reduce
[ ] reduce_window
[ ] window_sum
[ ] window_product
[ ] window_max
[ ] window_min
[ ] map
[ ] sort
[ ] cholesky
[x] add
[ ] subtract
[x] multiply
[ ] power
[ ] remainder
[ ] divide
[ ] arctan2
[ ] min
[ ] max
[ ] quotient
[ ] bitwise_and
[ ] bitwise_or
[ ] bitwise_xor
[ ] left_shift
[ ] right_shift
[ ] equal
[ ] not_equal
[ ] greater
[ ] less
[ ] greater_equal
[ ] less_equal
[ ] logical_and
[ ] logical_or
[ ] logical_xor
[ ] outer

Feb 17 '21 20:02 ityonemo

Can you pick one and write how to complete one of them?

Feb 17 '21 20:02 fire

the starter ones are the binary tensor ops, the implementation lives here:

https://github.com/ityonemo/ez/blob/af2dd139d953f42d0f0a3bd4695907ad7bc69db2/lib/ez/zig_slice_backend.ex#L31-L64

tests are here: https://github.com/ityonemo/ez/blob/af2dd139d953f42d0f0a3bd4695907ad7bc69db2/test/ez_test.exs#L15-L61

most of the other tensor ops should be straightforward, but will need a different harness (since they emit a different type, e.g. the boolean ones).

Not sure if reduce or map are possible since those require calling out to elixir or an nx-rewritten ast sort of a thing, but I think it's ultimately doable

Feb 17 '21 20:02 ityonemo

related: https://github.com/ityonemo/ez/issues/6

Feb 17 '21 20:02 ityonemo