Enable ConvReplaceWithQLinear when using ACL (#22823)

### Description Enable the ConvReplaceWithQLinear graph optimization when using the ACL execution provider. ### Motivation and Context Fixes an issue where quantized Conv nodes followed by ReLU don't get converted to QLinearConv, so ACL sees the weights as mutable and therefore cannot run the Conv node. Signed-off-by: Michael Tyler <[email protected]>
microsoft · Nov 14, 2024 · dd99e34 · dd99e34
1 parent 8268120
commit dd99e34
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/...ntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc b/...ntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc
@@ -236,7 +236,7 @@ void ConvQDQRules(SelectorActionRegistry& qdq_selector_action_registry, bool is_
 
 #if !defined(ORT_MINIMAL_BUILD)
   // TODO: Enable 16-bit types in selector when QLinearConv supports 16-bit.
-  std::vector<const char*> providers = {kCpuExecutionProvider, kDmlExecutionProvider};
+  std::vector<const char*> providers = {kCpuExecutionProvider, kDmlExecutionProvider, kAclExecutionProvider};
   std::unique_ptr<NodeSelector> selector = std::make_unique<QDQ::ConvSelector>(is_int8_allowed,
                                                                                false,
                                                                                false,