Add onnx GridSample support for border padding mode #3819
Conversation
| Value zeroInt = | ||
| b.create<arith::ConstantOp>(loc, b.getIntegerAttr(int64type, 0)); |
There was a problem hiding this comment.
It might be better to keep this constant outside the lambda function, and just get it from the capture (otherwise you will re-write the constant multiple times in the IR).
There was a problem hiding this comment.
Since paddingMode is checked at compile time, I have removed zeroInt completely.
|
Thanks for the feedback! I'll work on the requested changes. |
Evaluate paddingMode at compile time
| @@ -2587,10 +2587,30 @@ class ConvertAtenGridSamplerOp : public OpConversionPattern<AtenGridSamplerOp> { | |||
| return res; | |||
| }; | |||
|
|
|||
| auto lambdaBorder = [&](OpBuilder &b, Location loc, Value x, | |||
| Value SizeSubOne) -> Value { | |||
There was a problem hiding this comment.
SizeSubOne -> sizeSubOne
| auto resultType = cast<RankedTensorType>( | ||
| getTypeConverter()->convertType(op.getResult().getType())); | ||
| Value alignCorners = adaptor.getAlignCorners(); | ||
| Value interMode = adaptor.getInterpolationMode(); | ||
|
|
||
| int64_t paddingModeInt; | ||
| matchPattern(op.getPaddingMode(), m_TorchConstantInt(&paddingModeInt)); |
There was a problem hiding this comment.
Why not put this inside an if and return failure?
|
I completely agree with you, but I accidentally removed the branch and the PR was automatically closed, and I can't seem to open it again, even though the branch is restored.... Is it due to the new one being open? |
I think that's happening probably because the link between your past and the restored one is lost. Anyway, you should open any of the PR which you are able to. |
|
Hey, I reopened the other one. Sorry for the inconvenience. |
Adds support for padding_mode="border"
Clamps the grid coordinates between 0 and size - 1 like this when using this padding mode: