add custom annoatation for new model export (pytorch#12126)

Summary:

add  custom annoatation for new model export

Reviewed By: limintang

Differential Revision: D77569950

Author: billmguo

backends/qualcomm/quantizer/custom_annotation.py

@@ -227,6 +227,40 @@ def annotate_single_in_single_out(
             _annotated=True,
         )
 
+    def annotate_single_in_share_out(
+        node: Node, quantization_config: QuantizationConfig
+    ) -> None:
+
+        input_qspec_map = {}
+        input_act = node.args[0]
+        input_qspec_map[input_act] = quantization_config.input_activation
+
+        node.meta[QUANT_ANNOTATION_KEY] = QuantizationAnnotation(
+            input_qspec_map=input_qspec_map,
+            output_qspec=SharedQuantizationSpec((input_act, node)),
+            _annotated=True,
+        )
+
+    def annotate_stack(node: Node, quantization_config: QuantizationConfig) -> None:
+        input_nodes = node.args[0]
+
+        first_input_node = input_nodes[0]
+        input_qspec_map = {}
+        input_qspec_map[first_input_node] = quantization_config.input_activation
+        share_qparams_with_input_act0_qspec = SharedQuantizationSpec(
+            (first_input_node, node)
+        )
+
+        for input_node in input_nodes[1:]:
+            if input_node not in input_qspec_map:
+                input_qspec_map[input_node] = share_qparams_with_input_act0_qspec
+
+        node.meta[QUANT_ANNOTATION_KEY] = QuantizationAnnotation(
+            input_qspec_map=input_qspec_map,
+            output_qspec=share_qparams_with_input_act0_qspec,
+            _annotated=True,
+        )
+
     def annotate_matmul_input1(node: Node):
         quantization_config_8a8w = get_8a8w_qnn_ptq_config(
             act_symmetric=True, act_observer=MinMaxObserver
@@ -247,6 +281,12 @@ def annotate_matmul_input1(node: Node):
             ]:
                 annotate_single_in_single_out(node, quantization_config_8a8w)
                 node = node.args[0]
+            elif node.target == torch.ops.aten.stack.default:
+                annotate_stack(node, quantization_config_8a8w)
+                node = node.args[0]
+            elif node.target == torch.ops.aten.flatten.using_ints:
+                annotate_single_in_share_out(node, quantization_config_8a8w)
+                node = node.args[0]
             elif node.target == torch.ops.aten.cat.default:
                 annotate_cat(node, quantization_config_8a8w)
                 # For v, we tag 8a until conv op.