Update documentation for adding new ttnn operation #9841
Conversation
arakhmati
requested review from
eyonland,
cfjchu,
xanderchin,
ayerofieiev-tt,
tt-rkim,
TT-BrianLiu and
mo-tenstorrent
as code owners
arakhmati
force-pushed
the
ttnn-8835-add-documentation
branch
2 times, most recently
from
0023da1 to
7f41f47
Compare
arakhmati
changed the title
arakhmati
force-pushed
the
ttnn-8835-add-documentation
branch
from
7f41f47 to
c03faf6
Compare
arakhmati
force-pushed
the
ttnn-8835-add-documentation
branch
from
c03faf6 to
c4f8469
Compare
| dtype, | ||
| operation::DEFAULT_OUTPUT_MEMORY_CONFIG, | ||
| optional_output_tensor, | ||
| activations); |
There was a problem hiding this comment.
Did you want to do this on this PR so that the optional_output_tensor is properly handled?
// TODO: add support for optional_output_tensors
// auto optional_output_tensors = extract_args_to_vector(std::forward<args_t>(args)...,
// std::optional<ttnn::Tensor>);
```
There was a problem hiding this comment.
will do in next PR
|
|
||
| A ttnn operation is a function that takes in one or more input tensors and produces one or more output tensors. It is implemented in C++ and can be called from Python. |
There was a problem hiding this comment.
function but not really :D
|
|
||
| `ttnn/cpp/ttnn/operations/<category>/<operation_name>/device/<operation_name>_device_operation.hpp` |
There was a problem hiding this comment.
I am a little concerned about category part. People often mess up with categorizaton, specially when it is not 100% clear.. I am ok to keep it like that, just want to raise a Q for discussion.
|
|
||
| `ttnn/cpp/ttnn/operations/<category>/<operation_name>/device/<operation_name>_device_operation.hpp` | ||
| `ttnn/cpp/ttnn/operations/<category>/<operation_name>/device/<operation_name>_device_operation.cpp` |
There was a problem hiding this comment.
_device_operation is what we called <operation_name>_op.hpp/.cpp before?
|
|
||
| from tests.ttnn.utils_for_testing import assert_with_pcc | ||
| In order to add a new operation, add the following file: |
There was a problem hiding this comment.
Lets add a brief description of this file.
It basically provides a structure and registers the operation.
The cool thing about the structure is that it allows to provide multiple overloads.
The cool thing about registering that it allows to instrument calls to an operation with goodies like "...list...".
|
|
||
| .. code-block:: python | ||
| A concrete example: |
There was a problem hiding this comment.
maybe lets add some notes about consistency? not sure if here or in general.
we want ops to be consistent.
e.g.
kw_only memory_config = None
kw_only queue_id in python and overload in c++
kw_only optional output tensor/s
kw_only dtype
Specific names, specific defaults EXCEPT when not possible and something else is needed.
There was a problem hiding this comment.
will do in the next PR
| @@ -385,7 +385,13 @@ inline std::string op_meta_data_serialized_json( | |||
|
|
|||
| j["optional_input_tensors"] = std::vector<json>{}; | |||
|
|
|||
| auto perfModel = operation_t::create_op_performance_model(operation_attributes, tensor_args, tensor_return_value); | |||
There was a problem hiding this comment.
Do we want to write some more info on "how to perf model?"
arakhmati
force-pushed
the
ttnn-8835-add-documentation
branch
from
c4f8469 to
45889a5
Compare
Ticket
Link to Github Issue
Problem description
Documentation for adding a new ttnn operation is out of data
What's changed
Add an example of how to add a new operation and update the documentation using the example
Checklist