Make Shape working with forwardDims()

src/operator/Shape.cpp

@@ -20,14 +20,8 @@
 #include "aidge/utils/Types.h"
 
 void Aidge::Shape_OpImpl::forward() {
-    const Shape_Op& op = dynamic_cast<const Shape_Op&>(mOp);
-    const auto start = op.start();
-    const auto end = op.end();
-
-    op.getOutput(0)->getImpl()->copyCast(std::next(op.getInput(0)->dims().data(),
-                                                   start),
-                                         DataType::UInt64,
-                                         end - start + 1);
+    // Do nothing...
+    // Output is already valid after forwardDims()
 }
 
 ///////////////////////////////////////////////
@@ -75,6 +69,11 @@ bool Aidge::Shape_Op::forwardDims(bool /*allowDataDependency*/) {
         AIDGE_ASSERT(roi> 1, "Invalid ROI for Shape");
 
         mOutputs[0]->resize({roi});
+        // Ensure the output of this operator is valid after forwardDims():
+        mOutputs[0]->getImpl()->copyCast(std::next(getInput(0)->dims().data(),
+                                                    start),
+                                            DataType::UInt64,
+                                            end - start + 1);
         return true;
     }
 

[Cyril Moineau]

@olivierbichler this commit is annoying.

Indeed this force Shape to have an implementation in order to run forward_dims, otherwise mOutputs[0]->getImpl() -> 0x0.

Is it possible to revert it? If so I would like ot do it in !291 (merged)

[Olivier BICHLER]

@cmoineau But Shape is supposed to have a default implementation, in which case this is an issue?

[Cyril Moineau]

In case we do the import of an ONNX followed by a forward dims followed by an ONNX export. In this case we don't have access to any backend of intermediate tensor.