COMPMID-909: Enabling in-place computation for batchnormalization and activation at graph level
Change-Id: I84d4a212629b21794451ab5fb5c5b187b5e28f98
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/120127
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/src/graph/Graph.cpp b/src/graph/Graph.cpp
index 7af313a..98d9590 100644
--- a/src/graph/Graph.cpp
+++ b/src/graph/Graph.cpp
@@ -131,6 +131,11 @@
_previous_hints = _current_hints; // For the first node just assume the previous node was of the same type as this one
}
+ if(_current_node->supports_in_place())
+ {
+ _current_output = _current_input;
+ }
+
//Automatic output configuration ?
if(_current_output == nullptr)
{
@@ -152,8 +157,12 @@
_ctx.hints() = _current_hints;
std::unique_ptr<arm_compute::IFunction> func = _current_node->instantiate_node(_ctx, _current_input, _current_output);
- // Allocate current input
- _current_input->allocate();
+ // If the operation is done in-place, do not allocate or it will prevent following layers from performing the configuration
+ if(!_current_node->supports_in_place())
+ {
+ // Allocate current input
+ _current_input->allocate();
+ }
// Map input if needed
if(_current_input->target() == TargetHint::OPENCL)