!12314 [auto-monad] Change backend execution order sorting policy

From: @hwhewei
Reviewed-by: @ginfung,@zh_qh
Signed-off-by: @zh_qh

mindspore/ccsrc/backend/session/kernel_graph.cc

@@ -201,21 +201,17 @@ void KernelGraph::VisitNodeDescendants(const AnfNodePtr &node, std::queue<AnfNod
 }
 
 void KernelGraph::SetExecOrderByDefault() {
-  std::queue<AnfNodePtr> seed_nodes;
-  UpdateNodeEdgeList(&seed_nodes);
+  std::queue<AnfNodePtr> zero_input_nodes;
+  UpdateNodeEdgeList(&zero_input_nodes);
   execution_order_.clear();
   std::unordered_set<AnfNodePtr> visited_nodes;
-  std::queue<AnfNodePtr> zero_input_nodes;
   AnfNodePtr last_communication_node = nullptr;
   std::queue<AnfNodePtr> communication_descendants;
-  while (!seed_nodes.empty() || last_communication_node != nullptr) {
+  while (!zero_input_nodes.empty() || last_communication_node != nullptr) {
     // seed nodes first, then visit last all reduce node descendant
-    if (seed_nodes.empty()) {
+    if (last_communication_node != nullptr) {
       VisitNodeDescendants(last_communication_node, &communication_descendants, &visited_nodes);
       last_communication_node = nullptr;
-    } else {
-      zero_input_nodes.push(seed_nodes.front());
-      seed_nodes.pop();
     }
     // all reduce node descendant first, then common queue
     while (!zero_input_nodes.empty() || !communication_descendants.empty()) {
@@ -901,14 +897,11 @@ void KernelGraph::UpdateNodeEdgeList(std::queue<AnfNodePtr> *seed_nodes) {
       seed_nodes->push(node);
       continue;
     }
-    auto cnode = dyn_cast<CNode>(node);
+    auto cnode = node->cast<CNodePtr>();
     if (cnode == nullptr) {
       continue;
     }
-    auto &inputs = cnode->inputs();
-    // We push inputs from right to left, so that them can be evaluated from left to right.
-    for (auto iter = inputs.rbegin(); iter != inputs.rend(); ++iter) {
-      auto &input = *iter;
+    for (auto &input : cnode->inputs()) {
       PushNoVisitedNode(input, &que, &visited_nodes);
       AddDependEdge(node, input, 1);
     }

tests/st/auto_monad/test_auto_monad.py

@@ -1427,3 +1427,43 @@ def test_if_cast():
     r1 = net(beta1, beta2)
     expect = Tensor(np.array([3]).astype(np.float32))
     np.testing.assert_array_equal(r1.asnumpy(), expect.asnumpy())
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_multi_add_assign():
+    class Net(Cell):
+        def __init__(self, i1):
+            super(Net, self).__init__()
+            self.add = P.Add()
+            self.sub = P.Sub()
+            self.mul = P.Mul()
+            self.assign = P.Assign()
+            self.p = Parameter(i1, name='para')
+
+        def construct(self, a, d, e):
+            res1 = self.add(self.add(self.add(self.p, a), a), a)
+            mul = self.mul(d, e)
+            self.assign(self.p, mul)
+            res2 = self.sub(self.p, e)
+            return res2, res1
+
+    def numpy_out(p, a, d, e):
+        res1 = p + a + a + a
+        res_as = d * e
+        res2 = d * e - e
+        return res2, res1, res_as
+
+    p = (np.abs(np.random.normal(0, 1, [3])) + 1).astype(np.float32)
+    i0 = (np.abs(np.random.normal(0, 1, [3])) + 1).astype(np.float32)
+    i1 = (np.abs(np.random.normal(0, 1, [3])) + 1).astype(np.float32)
+    i2 = (np.abs(np.random.normal(0, 1, [3])) + 1).astype(np.float32)
+
+    net = Net(Tensor(p))
+    r2, r1 = net(Tensor(i0), Tensor(i1), Tensor(i2))
+
+    outputs = [r2.asnumpy(), r1.asnumpy(), net.p.data.asnumpy()]
+    expects = numpy_out(p, i0, i1, i2)
+    np.testing.assert_array_equal(outputs, expects)

tests/st/networks/models/bert/bert_performance/test_bert_tdt_lossscale.py

@@ -229,7 +229,7 @@ def test_bert_performance():
 
     # assertion occurs while the loss value, overflow state or loss_scale value is wrong
     loss_value = np.array(callback.loss_list)
-    expect_loss_value = [11.3660, 11.3265, 11.3264]
+    expect_loss_value = [11.3246, 11.2834, 11.2833]
     print("loss value: {}".format(loss_value))
     assert np.allclose(loss_value, expect_loss_value, 0, 0.0005)
 

tests/st/networks/models/bert/bert_precision/test_bert_tdt_lossscale.py

@@ -229,8 +229,8 @@ def test_bert_precision(enable_graph_kernel=False):
         expect_loss_value = [12.206627, 11.840489, 11.798470, 11.796345, 11.790964, 12.366766, 11.971539, 12.576565,
                              12.185522, 12.386192]
     else:
-        expect_loss_value = [12.206587, 11.966410, 11.965916, 11.975922, 11.970262, 12.608881, 12.174048, 12.840656,
-                             12.407923, 12.631133]
+        expect_loss_value = [12.206587, 11.940709, 11.930911, 11.937369, 11.932178, 12.556069, 12.130172, 12.783402,
+                             12.359581, 12.578078]
     print("loss value: {}".format(loss_value))
     assert np.allclose(loss_value, expect_loss_value, 0, 0.0005)