using args-mutation annotation to support inplace op in torch-blade pipline #1209
Conversation
Yancey1989
changed the title
| @@ -765,7 +765,7 @@ LogicalResult ConvertAtenOp<AtenSizeIntOp>::matchAndRewrite( | |||
| OpAdaptor adaptor, | |||
| ConversionPatternRewriter& rewriter) const { | |||
| // Not a tensor type. | |||
| auto selfType = adaptor.getSelf().getType().dyn_cast<TensorType>(); | |||
| auto selfType = adaptor.getSelf().getType().cast<RankedTensorType>(); | |||
There was a problem hiding this comment.
cast will raise error if it's not possible to cast? so the following if won't get reached?
There was a problem hiding this comment.
you're right, will polish this in the next PR.
| @@ -2725,6 +2726,8 @@ class ShapePropagator : public PropertyPropBase { | |||
| int64_t end = endOptional.value() != c10::nullopt | |||
| ? node->get<int64_t>(attr::end).value() | |||
| : INT64_MAX; | |||
| start = | |||
There was a problem hiding this comment.
nit: So we could not handle negative index before ... Orz..
| func.walk([&](Operation* op) { | ||
| if (isa<OverwriteTensorContentsOp>(op)) { | ||
| rewriter.setInsertionPoint(op); | ||
| auto nonValueTensor = rewriter.create<CopyToNonValueTensorOp>( |
There was a problem hiding this comment.
Why do we need to convert OverwriteTensorContentsOp to CopyToNonValueTensorOp and then to disc.ArgsMutation? Is it possible to skip CopyToNonValueTensorOp?
There was a problem hiding this comment.
Maybe CopyToNonValueTensorOp is required, the function reduceTensorConversions(func) at L79 removs all Tensor conversations and convert all !tensor to !vtensor, but OverwriteTensorContentsOp operand requires !tensor , so we need to insert CopyToNonValueTensorOp to make OverwriteTensorContentsOp validation after reduceTensorConversions.
fixed #1218