First Draft EyeLike CPU OP9 #121
Conversation
| 9, | ||
| KernelDefBuilder().TypeConstraint("T1", | ||
| std::vector<MLDataType>{ | ||
| DataTypeImpl::GetTensorType<float>(), |
There was a problem hiding this comment.
this is the entire list of supported types as per ONNX spec... should I add all of them?
tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)
There was a problem hiding this comment.
we could add the types step by step based on requirement. Is there any models require other types?
There was a problem hiding this comment.
I don't know if any model actually needs this operator right now. I am adding this as part of OP9 support. Do we maintain this info somewhere? How do I check?
There was a problem hiding this comment.
By default, we start with tensor(float) and add other types if need be.
pranavsharma
left a comment
pranavsharma
left a comment
There was a problem hiding this comment.
Did you look into Eigen? They support Nd matrices and operations on them.
| public: | ||
| EyeLike(const OpKernelInfo& info) : OpKernel(info) { | ||
| if (!info.GetAttr("k", &k_).IsOK()) { | ||
| ONNXRUNTIME_THROW("Missing 'k' attribute value"); |
There was a problem hiding this comment.
Looks like 'k' is optional based on the spec in which case we don't have to throw. #Resolved
| 9, | ||
| KernelDefBuilder().TypeConstraint("T1", | ||
| std::vector<MLDataType>{ | ||
| DataTypeImpl::GetTensorType<float>(), |
There was a problem hiding this comment.
we could add the types step by step based on requirement. Is there any models require other types?
| input_dims[1]); | ||
| output_mat.setZero(); | ||
|
|
||
| if ((k_ >= 0 && k_ >= input_dims[1]) || (k_ < 0 && std::abs(k_) >= input_dims[0])) { |
There was a problem hiding this comment.
is this an error condition?
There was a problem hiding this comment.
Not really an error. If K is greater than row(for lower diag) or col (in case of upper diag) then we terminate early and return a tensor with all zeroes.
* Add null check for node arguments. * Add null check to Shape property of node arg.
Adding implementation for new op-9 operator eyelike