PaddlePaddle
diff --git a/‎paddle/fluid/pir/dialect/operator/ir/manual_op.cc
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion b/‎paddle/fluid/pir/dialect/operator/ir/manual_op.cc
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion
diff --git a/‎paddle/fluid/pir/dialect/operator/ir/manual_op_vjp.cc
Copy file name to clipboard
+20-20Lines changed: 20 additions & 20 deletions b/‎paddle/fluid/pir/dialect/operator/ir/manual_op_vjp.cc
Copy file name to clipboard
+20-20Lines changed: 20 additions & 20 deletions
diff --git a/‎python/paddle/autograd/ir_backward.py
Copy file name to clipboard
+134-43Lines changed: 134 additions & 43 deletions b/‎python/paddle/autograd/ir_backward.py
Copy file name to clipboard
+134-43Lines changed: 134 additions & 43 deletions
diff --git a/‎python/paddle/tensor/array.py
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion b/‎python/paddle/tensor/array.py
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion
diff --git a/‎test/ir/pir/test_while_api.py
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎test/ir/pir/test_while_api.py
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
@@ -1501,7 +1501,7 @@ OpInfoTuple ArrayReadOp::GetOpInfo() {
                   false,
                   false,
                   false,
-                  true),
+                  false),
       OpInputInfo(
           "i", "paddle::dialect::ScalarAttribute", false, false, true, false)};
 
 
@@ -196,7 +196,7 @@ std::vector<std::vector<pir::OpResult>> ArrayWrite_Op::Vjp(
     pir::Operation* op,
     const std::vector<std::vector<pir::Value>>& inputs_,
     const std::vector<std::vector<pir::Value>>& outputs,
-    const std::vector<std::vector<pir::Value>>& out_grads,
+    const std::vector<std::vector<pir::Value>>& in_grads,
     const std::vector<std::vector<bool>>& stop_gradients) {
   PADDLE_ENFORCE_EQ(
       inputs_.size(),
@@ -212,19 +212,21 @@ std::vector<std::vector<pir::OpResult>> ArrayWrite_Op::Vjp(
           outputs.size()));
 
   PADDLE_ENFORCE_EQ(
-      out_grads.size(),
+      in_grads.size(),
       1,
       platform::errors::InvalidArgument(
           "ArrayWrite_ op's outputs size should be 1, but now is %d.",
           outputs.size()));
 
   VLOG(6) << "Vjp prepare call  ArrayWrite_'s vjp inteface";
-  pir::OpResult tensor_res =
-      paddle::dialect::array_read(out_grads[0][0], inputs_[2][0]);
-
-  std::vector<std::vector<pir::OpResult>> res{{tensor_res}};
-  if (stop_gradients[0][0]) {
-    res = {{}};
+  pir::OpResult x_grad =
+      paddle::dialect::array_read(in_grads[0][0], inputs_[2][0]);
+  pir::OpResult zero = paddle::dialect::zeros_like(inputs_[1][0]);
+  paddle::dialect::array_write_(in_grads[0][0], zero, inputs_[2][0]);
+  std::vector<std::vector<pir::OpResult>> res(1);
+  res[0].resize(1);
+  if (!stop_gradients[0][0]) {
+    res[0][0] = x_grad;
   }
   return res;
 }
@@ -247,27 +249,25 @@ std::vector<std::vector<pir::OpResult>> ArrayReadOp::Vjp(
       platform::errors::InvalidArgument(
           "Array_read op's outputs size should be 1, but now is %d.",
           outputs.size()));
-
+  // x = array_read(input, i)
+  // out_grads[0][0] is x_grad
+  // out_grads[1][0] is input_array_grad
   PADDLE_ENFORCE_EQ(
       out_grads.size(),
-      1,
+      2,
       platform::errors::InvalidArgument(
           "Array_read op's outputs size should be 1, but now is %d.",
           outputs.size()));
 
   VLOG(6) << "Vjp prepare call  Array_read's vjp inteface";
 
-  paddle::dialect::DenseTensorType outgrad_type =
-      out_grads[0][0].type().dyn_cast<paddle::dialect::DenseTensorType>();
-  pir::Value new_array = paddle::dialect::create_array(
-      paddle::dialect::TransToPhiDataType(outgrad_type.dtype()));
-  pir::OpResult tensor_res =
-      paddle::dialect::array_write_(new_array, out_grads[0][0], inputs_[1][0]);
+  pir::Value array_grad_i_origin =
+      paddle::dialect::array_read(out_grads[1][0], inputs_[1][0]);
+  pir::Value array_grad_i =
+      paddle::dialect::add(array_grad_i_origin, out_grads[0][0]);
+  paddle::dialect::array_write_(out_grads[1][0], array_grad_i, inputs_[1][0]);
 
-  std::vector<std::vector<pir::OpResult>> res{{tensor_res}};
-  if (stop_gradients[0][0]) {
-    res = {{}};
-  }
+  std::vector<std::vector<pir::OpResult>> res;
   return res;
 }
 }  // namespace dialect
 
@@ -63,17 +63,26 @@ def check_all_puts(block, inputs, outputs):
 
 
 def append_full_like(float_value, copy_value, value, state, backward_ops):
-    value_grad = paddle.full_like(
-        copy_value,
-        float_value,
-        dtype=copy_value.dtype,
-    )
-    full_like_op = value_grad.get_defining_op()
-    full_op = full_like_op.operand_source(1).get_defining_op()
+    if copy_value.is_tensorarray():
+        value_grad = paddle._pir_ops.create_array_like(
+            copy_value,
+            float_value,
+        )
+        full_like_op = value_grad.get_defining_op()
+        backward_ops_ = [full_like_op]
+    else:
+        value_grad = paddle.full_like(
+            copy_value,
+            float_value,
+            dtype=copy_value.dtype,
+        )
+        full_like_op = value_grad.get_defining_op()
+        full_op = full_like_op.operand_source(1).get_defining_op()
+        backward_ops_ = [full_like_op, full_op]
     update_bwdop_structure(
         backward_ops,
         state.op_to_opgrad[value.get_defining_op()],
-        [full_like_op, full_op],
+        backward_ops_,
     )
     state.value_to_valuegrad[value] = [[value_grad]]
     return value_grad
@@ -367,6 +376,16 @@ def inverse_sort_op(ops):
     return sorted_list
 
 
+def inplace_net(op_list):
+    op_name_list = [op.name() for op in op_list]
+    if (
+        "pd_op.array_write_" in op_name_list
+        or "pd_op.array_read" in op_name_list
+    ):
+        return True
+    return False
+
+
 def append_backward_ops(
     base_op,
     base_inputs,
@@ -425,13 +444,28 @@ def return_map_value(value, map):
             output = map[output]
         return output
 
+    def return_map_value_list(grad_value, map):
+        output = []
+        for i in range(len(grad_value)):
+            if grad_value[i] in map:
+                output.append(map[grad_value[i]])
+            else:
+                output.append(grad_value[i])
+        return output
+
     def append_add_n(value):
         # value is input of more than one fwd_op,
         # so more than one bwd_op create input_grad,
         # need add sum op to accumulate gradient
-        add_n_value = paddle.add_n(
-            [item[0] for item in state.value_to_valuegrad[value]]
-        )
+        if value.is_tensorarray():
+            add_n_value = paddle._pir_ops.add_n_array(
+                [item[0] for item in state.value_to_valuegrad[value]]
+            )
+        else:
+            add_n_value = paddle.add_n(
+                [item[0] for item in state.value_to_valuegrad[value]]
+            )
+
         add_n_op = add_n_value.get_defining_op()
         combine_op = add_n_op.operand_source(0).get_defining_op()
         update_bwdop_structure(
@@ -446,7 +480,12 @@ def make_output_with_output_grad(op):
         zero_flag = [False] * op.num_results()
         outputs = []
         output_grads = []
-        for i, value in enumerate(op.results()):
+        if op.name() == "pd_op.array_write_":
+            output_list = [op.operand_source(0)]
+        else:
+            output_list = op.results()
+
+        for i, value in enumerate(output_list):
             new_value = [
                 return_map_value(value, control_flow_value_to_copyvalue_map)
             ]
@@ -496,9 +535,39 @@ def make_output_with_output_grad(op):
             outputs.append(new_value)
             grad_value = state.value_to_valuegrad[value][0]
             output_grads.append(
-                [bwd_value_to_block_argument_map[grad_value[0]]]
-                if grad_value[0] in bwd_value_to_block_argument_map
-                else grad_value
+                return_map_value_list(
+                    grad_value, bwd_value_to_block_argument_map
+                )
+            )
+
+        if op.name() == "pd_op.array_read":
+            value = op.operand_source(0)
+            while value in state.inside_value_to_outside_value_map:
+                value = state.inside_value_to_outside_value_map[value]
+
+            if value in state.value_to_valuegrad:
+                if len(state.value_to_valuegrad[value]) > 1:
+                    append_add_n(value)
+
+            if (
+                value not in state.value_to_valuegrad
+                or state.value_to_valuegrad[value] == []
+            ):
+                append_full_like(
+                    0.0,
+                    return_map_value(
+                        value, control_flow_value_to_copyvalue_map
+                    ),
+                    value,
+                    state,
+                    backward_ops,
+                )
+
+            grad_value = state.value_to_valuegrad[value][0]
+            output_grads.append(
+                return_map_value_list(
+                    grad_value, bwd_value_to_block_argument_map
+                )
             )
 
         return zero_flag, outputs, output_grads
@@ -692,7 +761,11 @@ def argument_to_value(while_op):
     else:
         forward_ops = effective_forward_ops
 
-    inverse_effective_forward_ops = inverse_sort_op(forward_ops)
+    if inplace_net(forward_ops):
+        inverse_effective_forward_ops = reversed(forward_ops)
+    else:
+        inverse_effective_forward_ops = inverse_sort_op(forward_ops)
+
     clear_effective_forward_ops = []
     for op in inverse_effective_forward_ops:
         if op.name() != "builtin.combine" and op.name() != "builtin.split":
@@ -743,7 +816,9 @@ def argument_to_value(while_op):
                 else:
                     # all(zero_flag) support this op has no contribution for grad
                     # should be delete (prune sub_graph)
-                    if len(output_grads) == 0 or all(zero_flag):
+                    if (
+                        len(output_grads) == 0 or all(zero_flag)
+                    ) and op.name() != "pd_op.while":
                         continue
 
                     if op.name() == "pd_op.if":
@@ -787,17 +862,28 @@ def argument_to_value(while_op):
                         for i, input in enumerate(
                             get_used_external_value(while_block)
                         ):
-                            append_full_like(
-                                0.0, input, input, sub_state, backward_ops
-                            )
+                            if input in sub_state.value_to_valuegrad:
+                                if len(sub_state.value_to_valuegrad[input]) > 1:
+                                    append_add_n(input)
+
+                            if (
+                                input not in sub_state.value_to_valuegrad
+                                or sub_state.value_to_valuegrad[input] == []
+                            ):
+                                append_full_like(
+                                    0.0, input, input, sub_state, backward_ops
+                                )
+
                             grad_value = sub_state.value_to_valuegrad[input][0]
+                            for tmp in state.value_to_valuegrad[input]:
+                                state.value_to_sumvaluegrad[input].append(tmp)
+                            state.value_to_valuegrad[input] = []
                             output_grads.append(
-                                [bwd_value_to_block_argument_map[grad_value[0]]]
-                                if grad_value[0]
-                                in bwd_value_to_block_argument_map
-                                else grad_value
+                                return_map_value_list(
+                                    grad_value,
+                                    bwd_value_to_block_argument_map,
+                                )
                             )
-
                         build_pipe_for_block(while_block)
                         with dynamic_shape_prim_vjp_guard(op, inputs):
                             input_grads = paddle.framework.core.call_vjp(
@@ -953,6 +1039,11 @@ def calc_gradient_helper(outputs, inputs, grad_outputs, no_grad_set):
     block = outputs[0].get_defining_op().get_parent_block()
     state = State(block)
 
+    total_ops = []
+    if block.parent_block is not None:
+        total_ops += block.parent_block.ops
+    total_ops += block.ops
+
     # update no_grad_set if some value stop_gradient=True
     update_no_grad_set_by_stopgradient(block, no_grad_set)
     complete_outputs, _, backward_ops = prepare_grad_outputs(
@@ -961,14 +1052,14 @@ def calc_gradient_helper(outputs, inputs, grad_outputs, no_grad_set):
 
     inputs_set = ValueSet(inputs)
     outputs_set = ValueSet(complete_outputs)
-    total_ops = []
-    if block.parent_block is not None:
-        total_ops += block.parent_block.ops
-    total_ops += block.ops
 
-    effective_forward_ops, _ = prune_ops(
-        total_ops, inputs_set, outputs_set, no_grad_set
-    )
+    if inplace_net(total_ops):
+        effective_forward_ops = total_ops
+    else:
+        effective_forward_ops, _ = prune_ops(
+            total_ops, inputs_set, outputs_set, no_grad_set
+        )
+
     update_no_grad_set_after_prune(
         total_ops, effective_forward_ops, no_grad_set, inputs, complete_outputs
     )
@@ -993,18 +1084,18 @@ def calc_gradient_helper(outputs, inputs, grad_outputs, no_grad_set):
     outputs_set, inputs_set, no_gradvar_set = create_backward_prune_set(
         outputs_fwd_set, inputs_fwd_set, no_grad_set, state
     )
+    if not inplace_net(backward_ops):
+        _, remove_ops = prune_ops(
+            backward_ops, inputs_set, outputs_set, no_gradvar_set
+        )
 
-    _, remove_ops = prune_ops(
-        backward_ops, inputs_set, outputs_set, no_gradvar_set
-    )
-
-    state.turn_map()
-    for bwd_op in inverse_sort_op(remove_ops):
-        if bwd_op.result(0) in ValueSet(grad_outputs):
-            continue
-        if bwd_op.result(0).use_empty():
-            remove_op(block, bwd_op, state)
-    state.turn_map()
+        state.turn_map()
+        for bwd_op in inverse_sort_op(remove_ops):
+            if bwd_op.result(0) in ValueSet(grad_outputs):
+                continue
+            if bwd_op.result(0).use_empty():
+                remove_op(block, bwd_op, state)
+        state.turn_map()
 
     input_grad_map = state.value_to_valuegrad
 
 
@@ -230,7 +230,7 @@ def array_write(x, i, array=None):
         if array is None:
             array = paddle._pir_ops.create_array(x.dtype)
 
-        array = paddle._pir_ops.array_write_(array, x, i)
+        paddle._pir_ops.array_write_(array, x, i)
         return array
     else:
         check_variable_and_dtype(i, 'i', ['int64'], 'array_write')
 
@@ -177,7 +177,7 @@ def test_backward(self):
             )
             self.assertEqual(
                 main_program.global_block()
-                .ops[-3]
+                .ops[-1]
                 .as_while_op()
                 .body()
                 .ops[-4]
@@ -187,7 +187,7 @@ def test_backward(self):
 
             self.assertEqual(
                 main_program.global_block()
-                .ops[-3]
+                .ops[-1]
                 .as_while_op()
                 .body()
                 .ops[-5]