[BYOC][ACL] Add maximum support for float32 (apache#6506)

* ACL integration: add maximum support for float32.

* Added the code generation flow in arm_compute_lib.py
* Added the runtime calls in acl_runtime.cc

Change-Id: I69c5522f05a46c1dd235da5d57fe499134de0425

* Add maximum to the list of supported functions

Change-Id: Ia49087756be4c3ac92a3dc76fe03fb00de468f8d

docs/deploy/arm_compute_lib.rst

@@ -234,10 +234,6 @@ Operator support
 +----------------------+-------------------------------------------------------------------------+
 | maximum              | fp32                                                                    |
 +----------------------+-------------------------------------------------------------------------+
-| add                  | fp32                                                                    |
-+----------------------+-------------------------------------------------------------------------+
-| qnn.add              | uint8                                                                   |
-+----------------------+-------------------------------------------------------------------------+
 
 .. note::
     A composite operator is a series of operators that map to a single Arm Compute Library operator. You can view this

python/tvm/relay/op/contrib/arm_compute_lib.py

@@ -346,23 +346,3 @@ def maximum(attrs, args):
     type_a = args[0].checked_type
     type_b = args[0].checked_type
     return (type_a.dtype == "float32") and (type_b.dtype == "float32")
-
-
-@tvm.ir.register_op_attr("add", "target.arm_compute_lib")
-def add(attrs, args):
-    """Check if the external ACL codegen for add should be used."""
-    for typ in [args[0].checked_type, args[1].checked_type]:
-        if typ.dtype != "float32":
-            return False
-
-    return True
-
-
-@tvm.ir.register_op_attr("qnn.add", "target.arm_compute_lib")
-def qnn_add(attrs, args):
-    """Check if the external ACL codegen for add should be used."""
-    for typ in [args[0].checked_type, args[1].checked_type]:
-        if typ.dtype != "uint8":
-            return False
-
-    return True

src/runtime/contrib/arm_compute_lib/acl_runtime.cc

@@ -143,8 +143,6 @@ class ACLRuntime : public JSONRuntimeBase {
           CreateReshapeLayer(&layer_, node);
         } else if ("maximum" == op_name) {
           CreateMaximumLayer(&layer_, node);
-        } else if ("add" == op_name || "qnn.add" == op_name) {
-          CreateAddLayer(&layer_, node);
         } else {
           LOG(FATAL) << "Unsupported op: " << op_name;
         }
@@ -420,36 +418,6 @@ class ACLRuntime : public JSONRuntimeBase {
     function->configure(&layer->inputs[0], &layer->inputs[1], &layer->outputs[0]);
     layer->function = function;
   }
-  /*!
-   * \brief Creates an add/qnn.add layer
-   *
-   * \param layer The ACL layer to build. Containing inputs, outputs and the ACL function.
-   * \param node  The JSON representation of the operator.
-   */
-  void CreateAddLayer(CachedLayer* layer, const JSONGraphNode& node) {
-    auto op_name = node.GetOpName();
-    if ("add" == op_name) {
-      layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[0]));
-      layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[1]));
-      layer->outputs.push_back(MakeACLTensorFromJSONNode(node));
-    } else if ("qnn.add" == op_name) {
-      layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[0], &node.GetInputs()[2],
-                                                         &node.GetInputs()[3]));
-      layer->inputs.push_back(MakeACLTensorFromJSONEntry(node.GetInputs()[1], &node.GetInputs()[4],
-                                                         &node.GetInputs()[5]));
-      layer->outputs.push_back(
-          MakeACLTensorFromJSONNode(node, &node.GetInputs()[6], &node.GetInputs()[7]));
-    } else {
-      throw std::runtime_error("Unsupported form of add op: " + op_name);
-    }
-
-    auto f = std::make_shared<arm_compute::NEArithmeticAddition>();
-
-    // SATURATE is used as add_QASYMM8_QASYMM8_QASYMM8 always saturates result
-    f->configure(&layer->inputs[0], &layer->inputs[1], &layer->outputs[0],
-                 arm_compute::ConvertPolicy::SATURATE);
-    layer->function = f;
-  }
 
   /*! \brief Allow ACL functions to request auxiliary memory from TVM. */
   ACLAllocator allocator_;