[hybrid performance] Optimize pipeline send wait (#34086)

Author: wangxicoding

paddle/fluid/operators/nop_op.cc

@@ -0,0 +1,66 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#include <string>
+
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+class NopOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {}
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(framework::proto::VarType::FP32,
+                                   ctx.GetPlace());
+  }
+};
+
+class NopOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "(Tensor) The input tensor of nop op.").AsDuplicable();
+    AddOutput("Out", "(Tensor) The output tensor of nop op.").AsDuplicable();
+    AddComment(R"DOC(
+Nop Operator
+
+Do nothing, except let the input and output tensors occupy the memory and
+establish the dependency between input and output tensors.
+)DOC");
+  }
+};
+
+template <typename T>
+class NopKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {}
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_WITHOUT_GRADIENT(nop, ops::NopOp, ops::NopOpMaker);
+
+REGISTER_OP_CPU_KERNEL(nop, ops::NopKernel<float>);
+
+REGISTER_OP_CUDA_KERNEL(nop, ops::NopKernel<float>);
+
+REGISTER_OP_NPU_KERNEL(nop, ops::NopKernel<float>);

python/paddle/fluid/optimizer.py

@@ -4221,6 +4221,8 @@ def __init__(self, optimizer, num_microbatches=1, start_cpu_core_id=0):
         self._param_device_map = None
         self._pipeline_pair = []
         self._pp_ring_map = dict()
+        self.output_var_to_op = None
+        self.input_var_to_op = None
 
     # insert allreduce op to sync global information for global
     # gradient clip and amp
@@ -4657,6 +4659,9 @@ def _check_validation(self, block):
             int(self._op_role.Optimize),
             int(self._op_role.Backward) | int(self._op_role.Loss),
         ]
+        pre_stage_id = None
+        decrease_flag = False
+        in_optimize = False
         for op in block.ops:
             if not op._has_kernel(op.type):
                 assert op.type == "conditional_block" and (
@@ -4666,25 +4671,49 @@ def _check_validation(self, block):
             assert op.has_attr(self._op_role_key), (
                 "op ({}) has no {} attribute.".format(op.type,
                                                       self._op_role_key))
-            assert int(op.attr(self._op_role_key)) in valid_op_role_value, \
+            op_role = op.attr(self._op_role_key)
+            assert int(op_role) in valid_op_role_value, \
                 "op_role {} for op {} must be one of {}".format(
-                    op.attr(self._op_role_key),
+                    op_role,
                     op.type,
                     valid_op_role_value)
+            if int(op_role) == int(self._op_role.Optimize):
+                in_optimize = True
+
             assert op.has_attr(self._op_device_key), (
                 "op ({}) has no {} attribute.".format(op.type,
                                                       self._op_device_key))
 
             device = op.attr(self._op_device_key)
             assert device, ("op_device attribute for op "
                             "{} has not been set.".format(op.type))
-            if device == "gpu:all": continue
+            if device == "gpu:all" or device == "npu:all": continue
+
             dev_type = device.split(':')[0]
+            stage_id = int(device.split(':')[1])
             assert dev_type == "gpu" or dev_type == 'npu', (
                 "Now only gpu and npu devices are supported "
                 "for pipeline parallelism.")
-            if not device in device_list:
+
+            if device not in device_list:
                 device_list.append(device)
+
+            if not in_optimize:
+                if pre_stage_id is not None:
+                    interval = stage_id - pre_stage_id
+                    assert abs(interval) <= 1, \
+                        "The stage interval of two consecutive ops in the pipeline must be < = 1," \
+                        "but the interval of op={} and prev op is {}".format(op, interval)
+                    # stage must be in order, such as Forward(0 1 2 3 4), Backward(4 3 2 1 0)
+                    # if stage is unordered, such as Forward(0 1 2 3 4 3 4), will report error
+                    if interval == -1:
+                        decrease_flag = True
+                    if interval == 1:
+                        assert decrease_flag is False, \
+                            "Pipeline stage must be in order, " \
+                            "please check the stage of op={}".format(op)
+                pre_stage_id = stage_id
+
         return device_list
 
     def _insert_sendrecv_ops_for_boundaries(self, block):
@@ -4826,14 +4855,16 @@ def _insert_send_recv(cur_id, prev_id):
                             })
                         extra_index_info['index'] += 1
                         insert_index = None
+
                         if int(op_role) == int(self._op_role.Backward):
                             insert_index = extra_index_info[
                                 'first_optimize_index']
                             new_op_role = self._op_role.Optimize
                         else:
                             insert_index = index
                             new_op_role = self._op_role.Backward
-                        block._insert_op_without_sync(
+
+                        sync_comm_op = block._insert_op_without_sync(
                             index=insert_index + extra_index_info['index'],
                             type='c_sync_comm_stream',
                             inputs={'X': [var]},
@@ -4843,8 +4874,11 @@ def _insert_send_recv(cur_id, prev_id):
                                 self._op_role_key: new_op_role,
                                 'ring_id': ring_id,
                             })
+
                         if int(op_role) == int(self._op_role.Forward):
+                            sync_comm_op._set_attr('pipeline_flag', '')
                             extra_index_info['index'] += 1
+
                         var_shape = list(var.shape)
                         var_shape[0] = self.micro_batch_size if var_shape[
                             0] < 0 else var_shape[0]
@@ -5153,17 +5187,55 @@ def _get_input_output_info(self, block):
         Get info of op input and output.
         '''
         # A map from output var to op which generate it.
-        self.output_var_to_op = dict()
+        output_var_to_op = defaultdict(list)
         # A map from var to op which takes it as input.
-        self.input_var_to_op = dict()
+        input_var_to_op = defaultdict(list)
 
-        for index, op in enumerate(list(block.ops)):
+        for index, op in enumerate(block.ops):
             for var_name in op.input_arg_names:
-                ops = self.input_var_to_op.setdefault(var_name, [])
-                ops.append([op, index])
+                input_var_to_op[var_name].append([op, index])
             for var_name in op.output_arg_names:
-                ops = self.output_var_to_op.setdefault(var_name, [])
-                ops.append([op, index])
+                output_var_to_op[var_name].append([op, index])
+
+        return output_var_to_op, input_var_to_op
+
+    def _optimize_forward_send_sync(self, program):
+        """
+        optimize forward send's sync_comm_stream schedule
+        """
+        if self.schedule_mode != '1F1B': return
+
+        block = program.block(0)
+
+        backward_recv_index = None
+        for index, op in enumerate(block.ops):
+            if op.type == 'recv_v2' and self._is_backward_op(op):
+                backward_recv_index = index
+                break
+
+        if backward_recv_index is None: return
+
+        offset = 0
+        for index, op in enumerate(list(block.ops)):
+            if index >= backward_recv_index: break
+            if op.type == 'c_sync_comm_stream' and op.has_attr('pipeline_flag'):
+                var_name = op.input_arg_names[0]
+                var = block.var(var_name)
+                block._remove_op(index + offset, sync=False)
+                offset -= 1
+                # NOTE:
+                # 1. When the backward recv is completed, it indicates
+                # that the forward send is completed too. So we only need
+                # to use the NOP op to prevent memory release.
+                # 2. Because we removed sync_comm_op,
+                # we will insert NOP after recv_op.
+                block._insert_op_without_sync(
+                    index=backward_recv_index,
+                    type='nop',
+                    inputs={'X': [var]},
+                    outputs={'Out': [var]},
+                    attrs={self._op_role_key: self._op_role.Backward})
+        block._sync_with_cpp()
 
     def minimize(self,
                  loss,
@@ -5200,7 +5272,8 @@ def minimize(self,
             loss, startup_program, parameter_list, no_grad_set)
         self._param_device_map = self._origin_optimizer._param_device_map
 
-        self._get_input_output_info(main_block)
+        self.output_var_to_op, self.input_var_to_op = \
+            self._get_input_output_info(main_block)
         # Step1: add default op_device attribute for ops.
         self._add_op_device_attr(main_block)
         device_list = self._check_validation(main_block)
@@ -5229,6 +5302,10 @@ def device_cmp(device1, device2):
         for p in program_list:
             self._create_vars(p.global_block(), main_block)
 
+        self.local_rank %= len(device_list)
+        # Step3.5: optimize forward send sync_comm to overlap send and recv
+        self._optimize_forward_send_sync(program_list[self.local_rank])
+
         # Step4: Special Case: process persistable vars that exist in
         # multiple sections
         # FIXME 
@@ -5238,7 +5315,6 @@ def device_cmp(device1, device2):
         # Step5: Add sub blocks for section programs
         self._add_sub_blocks(main_block, program_list)
 
-        self.local_rank %= len(device_list)
         place_list = []
         for dev in device_list:
             dev_index = int(dev.split(":")[1])