optimzie reshape related fusion (#53066)

paddle/fluid/framework/paddle2cinn/build_cinn_pass.cc

@@ -510,8 +510,16 @@ void AnalyseClusterVariables(
     bool is_inference_stage,
     const std::unordered_set<std::string>& skip_gc_var_names) {
   // collecting all input and output of op
+  std::unordered_set<std::string> unused_outputs;
+  std::unordered_set<std::string> legacy_ops{"reshape2", "transpose2"};
   for (auto* op_node : cluster) {
     const auto& op_name = op_node->Name();
+    if (legacy_ops.count(op_name) && op_node->Op()->HasOutput("XShape")) {
+      for (const auto& var_name :
+           (*(op_node->Op()->MutableOutputs()))["XShape"]) {
+        unused_outputs.insert(var_name);
+      }
+    }
     for (auto* input_var_node : op_node->inputs) {
       if (!deny_var_set.count(input_var_node->Name())) {
         // ignore deny var node
@@ -527,9 +535,12 @@ void AnalyseClusterVariables(
   // remove output node from cluster_inputs,
   // and add cluster_internals node
   for (auto* var_node : *cluster_outputs) {
-    if (cluster_inputs->count(var_node) > 0) {
+    if ((cluster_inputs->count(var_node) > 0) ||
+        (unused_outputs.count(var_node->Name()))) {
       // if a input node also exists in output list, remove
-      cluster_inputs->erase(var_node);
+      if (cluster_inputs->count(var_node) > 0) {
+        cluster_inputs->erase(var_node);
+      }
 
       // the internal node is must an output node of sub-graph,
       // but not any input node of out-graph.
@@ -538,8 +549,12 @@ void AnalyseClusterVariables(
       for (size_t i = 0; i < var_node->outputs.size() && is_only_used_internal;
            ++i) {
         is_only_used_internal &= (cluster.count(var_node->outputs[i]) > 0);
+        VLOG(3) << "var_node->outputs[" << i << "]: " << var_node->Name()
+                << ", is_only_used_internal: " << is_only_used_internal;
       }
+
       if (is_only_used_internal) {
+        VLOG(3) << "insert internal var: " << var_node->Name();
         cluster_internals->insert(var_node);
       }
     }

python/paddle/incubate/autograd/composite_rules.py

@@ -164,10 +164,12 @@ def layernorm_composite(x, scale, bias, epsilon, begin_norm_axis):
     out = difference * rsqrt_var
 
     if scale is not None:
-        scale = reshape(scale, x.shape[begin_norm_axis:])
+        if x.shape[begin_norm_axis:] is not scale.shape:
+            scale = reshape(scale, x.shape[begin_norm_axis:])
         out = out * scale
     if bias is not None:
-        bias = reshape(bias, x.shape[begin_norm_axis:])
+        if x.shape[begin_norm_axis:] is not bias.shape:
+            bias = reshape(bias, x.shape[begin_norm_axis:])
         out = out + bias
 
     mean_ = reshape(mean_, [-1])