COMPMID-1995: Prepare Graph to support different input/output quantization info
- Added support for different input/output qinfo in ActivationLayer and DepthwiseConv
- Added support for different input/output qinfo in ConcatenateLayer introducing ConcatDescriptor
- Added reshape validate
- Allow OutputLayer to return a specific connection index from the input
- Not run Inplace and Depth mutator when input/output quantization info are different
Change-Id: I03f5e416fc43ddd284e1501887202a3145f76d8a
Signed-off-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-on: https://review.mlplatform.org/c/852
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/graph/mutators/InPlaceOperationMutator.cpp b/src/graph/mutators/InPlaceOperationMutator.cpp
index 31921b3..1c2985d 100644
--- a/src/graph/mutators/InPlaceOperationMutator.cpp
+++ b/src/graph/mutators/InPlaceOperationMutator.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -56,8 +56,8 @@
ARM_COMPUTE_ERROR_ON(current_output_tensor == nullptr || new_output_tensor == nullptr);
- // Prevent in-place operation if there is an accessor bound to the in-place tensor
- if(new_output_tensor->accessor() == nullptr)
+ // Prevent in-place operation if there is an accessor bound to the in-place tensor or quantization info are different
+ if(new_output_tensor->accessor() == nullptr || current_output_tensor->desc().quant_info == new_output_tensor->desc().quant_info)
{
ARM_COMPUTE_LOG_GRAPH_VERBOSE("Switching to in-place computation for the node with ID : "
<< node->id() << " and name : " << node->name() << std::endl);