Merge branch 'develop' into yolo_box

revert-16190-refine_parallel_executor
Kaipeng Deng 6 years ago committed by GitHub
commit 74037cc1c8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

@ -179,7 +179,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
else:
build_strategy.reduce_strategy = fluid.BuildStrategy(
).ReduceStrategy.AllReduce
build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
avg_loss = train_args[0]

@ -110,7 +110,7 @@ function(op_library TARGET)
# Define operators that don't need pybind here.
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op")
"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1)
endif()

@ -91,7 +91,7 @@ paddle.fluid.layers.pool2d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'po
paddle.fluid.layers.pool3d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)), ('document', '043de7333b79ee0ac55053c14ed81625'))
paddle.fluid.layers.adaptive_pool2d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)), ('document', '859b887174d06f361658f69cb7c06d95'))
paddle.fluid.layers.adaptive_pool3d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)), ('document', '120f4323a3d7ed9c0916f15a59f0e497'))
paddle.fluid.layers.batch_norm (ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False)), ('document', 'c527b71b8a4c60dca8df8a745c2b598d'))
paddle.fluid.layers.batch_norm (ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False)), ('document', '320c6973b02ea179fa89fecc80796464'))
paddle.fluid.layers.data_norm (ArgSpec(args=['input', 'act', 'epsilon', 'param_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var'], varargs=None, keywords=None, defaults=(None, 1e-05, None, 'NCHW', False, None, None, None, False)), ('document', 'e45e09e65a2658e07cad987222f0d9ab'))
paddle.fluid.layers.beam_search_decode (ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b0b8d53821716cd50c42e09b593f3feb'))
paddle.fluid.layers.conv2d_transpose (ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)), ('document', '03993955ab1e6d3044c44e6f17fc85e9'))
@ -293,6 +293,7 @@ paddle.fluid.layers.sigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=
paddle.fluid.layers.logsigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '81ccb7acafd06c7728e11581f5d342e3'))
paddle.fluid.layers.exp (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e6b3e769413d96aab4176f96db25984b'))
paddle.fluid.layers.tanh (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e9d586a0b5bd05f67ee78048f9d503b6'))
paddle.fluid.layers.atan (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '3a46e0b5f9ce82348406478e610f14c9'))
paddle.fluid.layers.tanh_shrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '1e521554b9fdda9061ec6d306f0709b7'))
paddle.fluid.layers.softshrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '9eef31597bbafa2bd49691e072296e13'))
paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '072a8541e0f632366bba10f67cb0db27'))
@ -300,6 +301,8 @@ paddle.fluid.layers.abs (ArgSpec(args=['x', 'name'], varargs=None, keywords=None
paddle.fluid.layers.ceil (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c75d67dc5fe28f68e4cfffead4f698ad'))
paddle.fluid.layers.floor (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '647b16c5da5ef909649ae02abb434973'))
paddle.fluid.layers.cos (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '485f2686bcc2fe37a4bd893769c8a3e2'))
paddle.fluid.layers.acos (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '920a47734482276c069ba24c61c26b25'))
paddle.fluid.layers.asin (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'cf4ee2c9b9d7293556f8c5173dfb5d2c'))
paddle.fluid.layers.sin (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '01f1766aa76eff1df30147505b59f7c4'))
paddle.fluid.layers.round (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b47f5da13913d3e56bdb1e612a73f3f2'))
paddle.fluid.layers.reciprocal (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'cc6ac2f14f03c52aaa83a59bf83b8d26'))
@ -327,7 +330,7 @@ paddle.fluid.layers.generate_mask_labels (ArgSpec(args=['im_info', 'gt_classes',
paddle.fluid.layers.iou_similarity (ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '587845f60c5d97ffdf2dfd21da52eca1'))
paddle.fluid.layers.box_coder (ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0)), ('document', '032d0f4b7d8f6235ee5d91e473344f0e'))
paddle.fluid.layers.polygon_box_transform (ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '0e5ac2507723a0b5adec473f9556799b'))
paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '991e934c3e09abf0edec7c9c978b4691'))
paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'gtscore', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(None, True, None)), ('document', '57fa96922e42db8f064c3fb77f2255e8'))
paddle.fluid.layers.yolo_box (ArgSpec(args=['x', 'img_size', 'anchors', 'class_num', 'conf_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5566169a5ab993d177792c023c7fb340'))
paddle.fluid.layers.box_clip (ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '397e9e02b451d99c56e20f268fa03f2e'))
paddle.fluid.layers.multiclass_nms (ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)), ('document', 'ca7d1107b6c5d2d6d8221039a220fde0'))

@ -16,6 +16,7 @@ limitations under the License. */
#include <glog/logging.h>
#include <memory>
#include <utility>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
@ -49,6 +50,11 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
AppendPass("sequential_execution_pass");
}
// Add op fusion.
if (strategy.sync_batch_norm_) {
AppendPass("sync_batch_norm_pass");
}
// Add op fusion.
if (strategy.fuse_relu_depthwise_conv_) {
AppendPass("fuse_relu_depthwise_conv_pass");
@ -227,6 +233,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
} // namespace framework
} // namespace paddle
USE_PASS(sync_batch_norm_pass);
USE_PASS(fuse_relu_depthwise_conv_pass);
USE_PASS(fuse_elewise_add_act_pass);
USE_PASS(graph_viz_pass);

@ -77,6 +77,8 @@ struct BuildStrategy {
bool fuse_relu_depthwise_conv_{false};
bool sync_batch_norm_{false};
bool memory_optimize_{true};
// TODO(dzhwinter):
// make enable_inplace, memory_optimize_

@ -16,6 +16,7 @@
#include <algorithm>
#include <deque>
#include <iterator>
#include <memory>
#include <stack>
#include <string>
#include <unordered_map>
@ -263,6 +264,10 @@ void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
ir::Graph* graph) const {
VLOG(4) << "Try to inplace op " << op->Name();
// FIXME(liuwei1031): Graph is not aware of the existence of BlockDescs and
// ProgramDescs.
// The operations related to BlockDesc or ProgramDesc should perform on Graph
// or Node directly!
PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
"op_desc is nullptr");
// some pre-requirments need to meet if the op want to inplaced.

@ -337,7 +337,6 @@ bool NodeCanReused(const VarDesc& node) {
auto type = node.GetType();
// only these types holds bulk of gpu memory
if (!(type == proto::VarType::LOD_TENSOR ||
type == proto::VarType::SELECTED_ROWS ||
type == proto::VarType::LOD_TENSOR_ARRAY)) {
return false;
}

@ -24,6 +24,7 @@
#include <sstream>
#include <string>
#include <type_traits>
#include <unordered_set>
#include <vector>
#include "gflags/gflags.h"
#include "paddle/fluid/framework/data_type.h"
@ -191,6 +192,10 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
// immediately to make the subblock variable reuse strategy take
// effect. Because it is a single op in graph. No need to
// update the ir nodes.
// FIXME(liuwei1031): Graph is not aware of the existence of
// BlockDescs and ProgramDescs.
// The operations related to BlockDesc or ProgramDesc should perform
// on Graph or Node directly!
sub_op_desc->Rename(var->Name(), cache->Name());
if (sub_op_desc->Block() != nullptr &&
sub_op_desc->Block()->HasVar(var->Name())) {

@ -34,11 +34,11 @@ limitations under the License. */
#ifdef PADDLE_WITH_NGRAPH
#include "paddle/fluid/operators/ngraph/ngraph_engine.h"
DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
#endif
DECLARE_bool(benchmark);
DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
namespace paddle {
namespace framework {
@ -194,9 +194,6 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
bool force_disable_gc) {
platform::RecordBlock b(block_id);
if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
#ifdef PADDLE_WITH_NGRAPH
if (FLAGS_use_ngraph) operators::NgraphEngine::EnableNgraph(pdesc);
#endif
auto ctx = Prepare(pdesc, block_id, skip_ref_cnt_vars, force_disable_gc);
RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
}
@ -372,6 +369,12 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
#ifdef PADDLE_WITH_NGRAPH
if (FLAGS_use_ngraph) {
paddle::operators::NgraphEngine::FuseNgraphOps(
ctx->prog_.Block(ctx->block_id_), &ctx->ops_);
}
#endif
return ctx;
}

@ -46,6 +46,7 @@ cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
pass_library(graph_to_program_pass base)
pass_library(graph_viz_pass base)
pass_library(lock_free_optimize_pass base)
pass_library(cpu_quantize_squash_pass inference)
pass_library(fc_fuse_pass inference)
pass_library(attention_lstm_fuse_pass inference)
pass_library(infer_clean_graph_pass inference)
@ -66,6 +67,7 @@ pass_library(conv_elementwise_add_fuse_pass inference)
pass_library(conv_affine_channel_fuse_pass inference)
pass_library(transpose_flatten_concat_fuse_pass inference)
pass_library(identity_scale_op_clean_pass base)
pass_library(sync_batch_norm_pass base)
# There may be many transpose-flatten structures in a model, and the output of
# these structures will be used as inputs to the concat Op. This pattern will
@ -100,6 +102,8 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
cc_test(test_cpu_quantize_squash_pass SRCS cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
if (WITH_MKLDNN)
cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
cc_test(test_conv_bias_mkldnn_fuse_pass SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc DEPS conv_bias_mkldnn_fuse_pass naive_executor)

@ -0,0 +1,146 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file eint8_outcept 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 eint8_outpress or
// implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/cpu_quantize_squash_pass.h"
#include <string>
#include <vector>
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/string/pretty_log.h"
namespace paddle {
namespace framework {
namespace ir {
using string::PrettyLogDetail;
void CPUQuantizeSquashPass::FindNodesToKeep(
Graph* graph,
std::unordered_map<const Node*, int>* nodes_keep_counter) const {
GraphPatternDetector gpd;
patterns::DequantAny deq_any_pattern{gpd.mutable_pattern(), "deqant_any"};
deq_any_pattern();
int found_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_IR_NODE_FROM_SUBGRAPH(dequant_out, dequant_out, deq_any_pattern);
if (nodes_keep_counter->find(dequant_out) == nodes_keep_counter->end())
(*nodes_keep_counter)[dequant_out] = 1;
else
(*nodes_keep_counter)[dequant_out] += 1;
found_count++;
};
gpd(graph, handler);
AddStatis(found_count);
}
void CPUQuantizeSquashPass::Squash(
Graph* graph,
std::unordered_map<const Node*, int>* nodes_keep_counter) const {
GraphPatternDetector gpd;
patterns::DequantQuantAny squash_pattern{gpd.mutable_pattern(), "squash"};
squash_pattern();
int found_squash_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "squash requantize-quantize ops pair";
GET_IR_NODE_FROM_SUBGRAPH(dequant_in, dequant_in, squash_pattern);
GET_IR_NODE_FROM_SUBGRAPH(dequant_op, dequant_op, squash_pattern);
GET_IR_NODE_FROM_SUBGRAPH(dequant_out, dequant_out, squash_pattern);
GET_IR_NODE_FROM_SUBGRAPH(quant_op, quant_op, squash_pattern);
GET_IR_NODE_FROM_SUBGRAPH(quant_out, quant_out, squash_pattern);
GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, squash_pattern);
auto* next_op_desc = next_op->Op();
float dequant_scale = boost::get<float>(dequant_op->Op()->GetAttr("Scale"));
float quant_scale = boost::get<float>(quant_op->Op()->GetAttr("Scale"));
PADDLE_ENFORCE(nodes_keep_counter->find(dequant_out) !=
nodes_keep_counter->end());
// check if dequantize op should be kept or removed, decrease the counter
bool keep_dequant = (*nodes_keep_counter)[dequant_out]-- > 1;
if (dequant_scale == quant_scale) {
// squash dequantize-quantize to nothing
auto quant_out_var_name = quant_out->Name();
auto next_op_inputs = next_op_desc->InputNames();
for (const auto& name : next_op_inputs) {
auto var_name = next_op_desc->Input(name)[0];
if (var_name.compare(quant_out_var_name) == 0) {
next_op_desc->SetInput(
name, std::vector<std::string>({dequant_in->Name()}));
break;
}
}
if (keep_dequant)
GraphSafeRemoveNodes(graph, {quant_op, quant_out});
else
GraphSafeRemoveNodes(graph,
{dequant_op, quant_op, dequant_out, quant_out});
IR_NODE_LINK_TO(dequant_in, next_op);
found_squash_count++;
} else {
// squash dequantize-quantize to requantize op
OpDesc desc;
desc.SetType("requantize");
desc.SetInput("Input", std::vector<std::string>({dequant_in->Name()}));
desc.SetOutput("Output", std::vector<std::string>({quant_out->Name()}));
desc.SetAttr("Scale_in", dequant_scale);
desc.SetAttr("Scale_out", quant_scale);
auto requant_op = g->CreateOpNode(&desc);
if (keep_dequant)
GraphSafeRemoveNodes(graph, {quant_op});
else
GraphSafeRemoveNodes(graph, {dequant_op, quant_op, dequant_out});
IR_NODE_LINK_TO(dequant_in, requant_op);
IR_NODE_LINK_TO(requant_op, quant_out);
found_squash_count++;
}
};
gpd(graph, handler);
AddStatis(found_squash_count);
PrettyLogDetail("--- squashed %d dequantize-quantize pairs",
found_squash_count);
}
std::unique_ptr<ir::Graph> CPUQuantizeSquashPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("cpu_quantize_squash_pass", graph.get());
std::unordered_map<const Node*, int> nodes_keep_counter;
FindNodesToKeep(graph.get(), &nodes_keep_counter);
Squash(graph.get(), &nodes_keep_counter);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(cpu_quantize_squash_pass,
paddle::framework::ir::CPUQuantizeSquashPass);

@ -0,0 +1,58 @@
// Copyright (c) 2019 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.
#pragma once
#include <memory>
#include <string>
#include <unordered_map>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
/*
* Squash dequantize->quantize pair pattern into requantize op
*/
class CPUQuantizeSquashPass : public FusePassBase {
public:
virtual ~CPUQuantizeSquashPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
/*
* For each dequantize's output find the number of operators it is an input to
*/
void FindNodesToKeep(
Graph* graph,
std::unordered_map<const Node*, int>* nodes_keep_counter) const;
/*
* Squash dequantize-quantize ops pairs into requantize or nothing
*/
void Squash(Graph* graph,
std::unordered_map<const Node*, int>* nodes_keep_counter) const;
const std::string name_scope_{"squash"};
};
} // namespace ir
} // namespace framework
} // namespace paddle

@ -0,0 +1,179 @@
// Copyright (c) 2019 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 "paddle/fluid/framework/ir/cpu_quantize_squash_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs, bool use_mkldnn,
float scale = 0) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetAttr("name", name);
if (type == "conv2d") {
op->SetInput("Input", {inputs[0]});
if (inputs.size() > 1) op->SetInput("Filter", {inputs[1]});
if (inputs.size() > 2) op->SetInput("Bias", {inputs[2]});
op->SetOutput("Output", {outputs[0]});
} else if (type == "quantize") {
op->SetInput("Input", {inputs[0]});
op->SetOutput("Output", {outputs[0]});
op->SetAttr("Scale", scale);
} else if (type == "dequantize") {
op->SetInput("Input", {inputs[0]});
op->SetOutput("Output", {outputs[0]});
op->SetAttr("Scale", scale);
}
}
// (a,w1,b1)->Conv1->d
// d->Dequant->e
// e->Quant->f
// (f,w2,b2)->Conv2->i
ProgramDesc BuildProgramDesc(bool use_mkldnn, float scale1, float scale2) {
ProgramDesc prog;
for (auto& v : std::initializer_list<std::string>(
{"a", "w1", "b1", "d", "e", "f", "w2", "b2", "i"})) {
auto* var = prog.MutableBlock(0)->Var(v);
if (v.find("w") == 0 || v.find("b") == 0) {
var->SetPersistable(true);
}
}
SetOp(&prog, "conv2d", "Conv1", {"a", "w1", "b1"}, {"d"}, use_mkldnn);
SetOp(&prog, "dequantize", "Dequant", {"d"}, {"e"}, use_mkldnn, scale1);
SetOp(&prog, "quantize", "Quant", {"e"}, {"f"}, use_mkldnn, scale2);
SetOp(&prog, "conv2d", "Conv2", {"f", "w2", "b2"}, {"i"}, use_mkldnn);
return prog;
}
static const std::initializer_list<std::string> variable_names{
"a", "b", "c", "d", "e", "f", "g", "h"};
// a->Conv1->b
// b->Dequant->c
//
// c->Quant1->d and d->Conv2->e
//
// c->Conv3->f
//
// c->Quant2->g and g->Conv4->h
//
ProgramDesc BuildProgramDesc2(bool use_mkldnn, float scale1, float scale2,
float scale3) {
ProgramDesc prog;
for (auto& v : variable_names) {
prog.MutableBlock(0)->Var(v);
}
SetOp(&prog, "conv2d", "Conv1", {"a"}, {"b"}, use_mkldnn);
SetOp(&prog, "dequantize", "Dequant", {"b"}, {"c"}, use_mkldnn, scale1);
SetOp(&prog, "quantize", "Quant1", {"c"}, {"d"}, use_mkldnn, scale2);
SetOp(&prog, "conv2d", "Conv2", {"d"}, {"e"}, use_mkldnn);
SetOp(&prog, "conv2d", "Conv3", {"c"}, {"f"}, use_mkldnn);
SetOp(&prog, "quantize", "Quant2", {"c"}, {"g"}, use_mkldnn, scale3);
SetOp(&prog, "conv2d", "Conv4", {"g"}, {"h"}, use_mkldnn);
return prog;
}
void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
const char* var_name) {
auto x = scope->Var(var_name);
auto tensor = x->GetMutable<LoDTensor>();
tensor->mutable_data(place, proto::VarType::FP32,
::paddle::memory::Allocator::kDefault, 1);
}
void MainTest(const ProgramDesc& prog, int removed_nodes_num) {
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
// Init scope, as it is used in pass
auto place = paddle::platform::CPUPlace();
NaiveExecutor exe{place};
Scope scope;
exe.CreateVariables(prog, 0, true, &scope);
for (auto& v : variable_names) {
InitTensorHolder(&scope, place, v.c_str());
}
graph->Set(kParamScopeAttr, new framework::Scope*(&scope));
auto pass = PassRegistry::Instance().Get("cpu_quantize_squash_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_EQ(original_nodes_num - removed_nodes_num, current_nodes_num);
}
TEST(CpuQuantizeSquashPass, equal_scales) {
auto scale = 1.2345f;
auto use_mkldnn = true;
// Remove 4 nodes: Dequant, Quant, e, f
auto remove_nodes = 4;
MainTest(BuildProgramDesc(use_mkldnn, scale, scale), remove_nodes);
use_mkldnn = !use_mkldnn;
MainTest(BuildProgramDesc(use_mkldnn, scale, scale), remove_nodes);
}
TEST(CpuQuantizeSquashPass, inequal_scales) {
auto scale1 = 1.2345f;
auto scale2 = 21.0f;
auto use_mkldnn = true;
// Remove 3 nodes: Dequant, Quant, e
// Insert 1 node: requantize
auto remove_nodes = 2;
MainTest(BuildProgramDesc(use_mkldnn, scale1, scale2), remove_nodes);
use_mkldnn = !use_mkldnn;
MainTest(BuildProgramDesc(use_mkldnn, scale1, scale2), remove_nodes);
}
TEST(CpuQuantizeSquashPass, branch_to_equal_inequal_and_fp32) {
// Delete both quantize ops,
// bypass dequantize in both branches,
// insert requantize on one branch
auto scale = 1.2345f;
auto scale2 = 21.0f;
auto use_mkldnn = true;
// Remove 3 nodes: Quant1, Quant2, g
// Insert 1 node: requantize
auto remove_nodes = 2;
MainTest(BuildProgramDesc2(use_mkldnn, scale, scale, scale2), remove_nodes);
use_mkldnn = !use_mkldnn;
MainTest(BuildProgramDesc2(use_mkldnn, scale, scale, scale2), remove_nodes);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(cpu_quantize_squash_pass);

@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include <algorithm>
#include <unordered_set>
#include <unordered_map>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/op_proto_maker.h"
@ -152,6 +152,39 @@ void Graph::ResolveHazard(
}
}
std::shared_ptr<Graph> Graph::Clone() {
auto cloned_graph = std::make_shared<Graph>(this->program_);
cloned_graph->ReleaseNodes();
cloned_graph->num_node_created_ = 0;
std::unordered_map<ir::Node *, ir::Node *> origin_to_cloned;
for (auto *n : this->node_set_) {
ir::Node *cloned_node = nullptr;
if (n->IsCtrlVar()) {
cloned_node = cloned_graph->CreateControlDepVar();
} else if (!n->var_desc_ && !n->op_desc_) { // empty node
cloned_node = cloned_graph->CreateEmptyNode(n->Name(), n->NodeType());
} else if (n->IsVar()) {
cloned_node = cloned_graph->CreateVarNode(n->Var());
} else if (n->IsOp()) {
cloned_node = cloned_graph->CreateOpNode(n->Op());
}
if (cloned_node) {
origin_to_cloned[n] = cloned_node;
} else {
PADDLE_THROW("The cloned node's type is not supported!");
}
}
for (auto *n : this->node_set_) {
for (auto it = n->inputs.begin(); it != n->inputs.end(); it++) {
origin_to_cloned[n]->inputs.push_back(origin_to_cloned[*it]);
}
for (auto it = n->outputs.begin(); it != n->outputs.end(); it++) {
origin_to_cloned[n]->outputs.push_back(origin_to_cloned[*it]);
}
}
return cloned_graph;
}
bool IsControlDepVar(const ir::Node &var) {
return var.Name().find(ir::Node::kControlDepVarName) != std::string::npos;
}

@ -17,6 +17,7 @@ limitations under the License. */
#include <map>
#include <memory>
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/node.h"
@ -199,7 +200,12 @@ class Graph {
// WARN: After a series of passes, the current graph can be quite
// different from OriginProgram. Caller shouldn't assume much from
// the returned OriginProgram.
const ProgramDesc &OriginProgram() const { return program_; }
const ProgramDesc &OriginProgram() const {
LOG(WARNING) << "WARN: After a series of passes, the current graph can be "
"quite different from OriginProgram. So, please avoid "
"using the `OriginProgram()` method!";
return program_;
}
// This method takes ownership of `node`.
ir::Node *AddNode(ir::Node *node) {
@ -212,6 +218,10 @@ class Graph {
void ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
// Create a new and duplicated graph.
// WARN: The method only clones the graph structure, not its attributes.
std::shared_ptr<Graph> Clone();
private:
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);

@ -1301,6 +1301,51 @@ PDNode *patterns::ConvAffineChannel::operator()(
return ac_out_var;
}
PDNode *patterns::DequantQuantAny::operator()() {
auto *dequant_in = pattern->NewNode(dequant_in_repr())
->AsInput()
->assert_is_op_input("dequantize", "Input");
auto *dequant_op =
pattern->NewNode(dequant_op_repr())->assert_is_op("dequantize");
auto *dequant_out = pattern->NewNode(dequant_out_repr())
->AsOutput()
->assert_is_op_output("dequantize", "Output");
auto *quant_op = pattern->NewNode(quant_op_repr())
->assert_is_op("quantize")
->AsIntermediate();
auto *quant_out = pattern->NewNode(quant_out_repr())
->AsOutput()
->assert_is_op_output("quantize");
auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();
dequant_op->LinksFrom({dequant_in}).LinksTo({dequant_out});
quant_op->LinksFrom({dequant_out}).LinksTo({quant_out});
next_op->LinksFrom({quant_out});
return quant_out;
}
PDNode *patterns::DequantAny::operator()() {
auto *dequant_op =
pattern->NewNode(dequant_op_repr())->assert_is_op("dequantize");
auto *dequant_out = pattern->NewNode(dequant_out_repr())
->AsOutput()
->assert_is_op_output("dequantize", "Output");
auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();
dequant_op->LinksTo({dequant_out});
next_op->LinksFrom({dequant_out});
return dequant_out;
}
// a -> transpose_op(1) -> transpose_out_a -> flatten_op(1) -> flatten_out_a
// b -> transpose_op(2) -> transpose_out_b -> flatten_op(2) -> flatten_out_b
// ...

@ -18,8 +18,11 @@
#include <gtest/gtest_prod.h>
#endif
#include <memory>
#include <numeric>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
@ -766,6 +769,34 @@ struct ConvAffineChannel : public PatternBase {
PATTERN_DECL_NODE(ac_out); // Out
};
// Dequantize + Quantize + anyOP
// This pattern is used for squashing the dequantize-quantize pairs.
struct DequantQuantAny : public PatternBase {
DequantQuantAny(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "dequant_quant_any") {}
PDNode* operator()();
PATTERN_DECL_NODE(dequant_in);
PATTERN_DECL_NODE(dequant_op);
PATTERN_DECL_NODE(dequant_out);
PATTERN_DECL_NODE(quant_op);
PATTERN_DECL_NODE(quant_out);
PATTERN_DECL_NODE(next_op);
};
// Dequantize + anyOP
// This quantize is used for getting number of ops the Dequantize's
// output is an input to.
struct DequantAny : public PatternBase {
DequantAny(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "dequant_any") {}
PDNode* operator()();
PATTERN_DECL_NODE(dequant_op);
PATTERN_DECL_NODE(dequant_out);
PATTERN_DECL_NODE(next_op);
};
struct TransposeFlattenConcat : public PatternBase {
TransposeFlattenConcat(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "transpose_flatten_concat") {}

@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include <typeindex>
#include <typeinfo>

@ -0,0 +1,45 @@
/* Copyright (c) 2019 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 "paddle/fluid/framework/ir/sync_batch_norm_pass.h"
#include <memory>
#include <string>
#include <utility>
namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> SyncBatchNormPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
VLOG(3) << "Use synchronous batch norm";
for (const Node* n : graph->Nodes()) {
if (n->IsOp()) {
auto* op = n->Op();
if (op->Type() == "batch_norm") {
op->SetType("sync_batch_norm");
}
if (op->Type() == "batch_norm_grad") {
op->SetType("sync_batch_norm_grad");
}
}
}
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(sync_batch_norm_pass, paddle::framework::ir::SyncBatchNormPass);

@ -0,0 +1,32 @@
/* Copyright (c) 2019 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. */
#pragma once
#include <memory>
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
class SyncBatchNormPass : public Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle

@ -0,0 +1,80 @@
// Copyright (c) 2019 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 "paddle/fluid/framework/ir/sync_batch_norm_pass.h"
#include <gtest/gtest.h>
namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
op->SetAttr("name", name);
op->SetInput("X", inputs);
op->SetOutput("Out", outputs);
}
// (a, conv_w)->conv2d->b
// (b, bn_scale, bn_bias, mean, var)->batch_norm
// ->(c, mean, var, save_mean, save_inv_var)
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v : std::vector<std::string>({"a", "conv_w", "b", "bn_scale",
"bn_bias", "mean", "var", "c",
"save_mean", "save_inv_var"})) {
auto* var = prog.MutableBlock(0)->Var(v);
if (v == "conv_w" || v == "bn_scale" || v == "bn_bias" || v == "mean" ||
v == "var") {
var->SetPersistable(true);
}
}
SetOp(&prog, "conv2d", "conv", std::vector<std::string>({"a", "conv_w"}),
std::vector<std::string>({"b"}));
SetOp(&prog, "batch_norm", "bn",
std::vector<std::string>({"b", "bn_scale", "bn_bias", "mean", "var"}),
std::vector<std::string>(
{"c", "mean", "var", "save_mean", "save_inv_var"}));
return prog;
}
TEST(IsTestPass, basic) {
auto prog = BuildProgramDesc();
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
auto pass = PassRegistry::Instance().Get("sync_batch_norm_pass");
graph = pass->Apply(std::move(graph));
for (auto* node : graph->Nodes()) {
if (node->IsOp()) {
auto* op = node->Op();
auto op_name = boost::get<std::string>(op->GetAttr("name"));
if (op_name == "bn") {
ASSERT_EQ(op->Type(), "sync_batch_norm");
}
}
}
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(sync_batch_norm_pass);

@ -186,14 +186,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
VLOG(3) << place << " " << DebugStringEx(&scope);
} catch (platform::EnforceNotMet exception) {
if (Attrs().count("sub_block") != 0) {
throw;
throw std::move(exception);
}
auto& callstack = Attr<std::vector<std::string>>(
OpProtoAndCheckerMaker::OpCreationCallstackAttrName());
if (callstack.empty()) {
throw;
throw std::move(exception);
}
std::ostringstream sout;
sout << "Invoke operator " << Type() << " error.\n";
@ -204,7 +204,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
sout << "C++ Callstacks: \n";
sout << exception.err_str_;
exception.err_str_ = sout.str();
throw;
throw std::move(exception);
} catch (...) {
std::rethrow_exception(std::current_exception());
}
@ -926,8 +926,10 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
dev_ctx = pool.Get(expected_kernel_key.place_);
}
RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx);
this->InferShape(&infer_shape_ctx);
if (!HasAttr(kAllKernelsMustComputeRuntimeShape)) {
RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx);
this->InferShape(&infer_shape_ctx);
}
// TODO(panyx0718): ExecutionContext should only depend on RuntimeContext
// not Scope. Imperative mode only pass inputs and get outputs.
kernel_iter->second(

@ -62,6 +62,15 @@ constexpr char kZeroVarSuffix[] = "@ZERO";
/// Variables with this suffix are the new Gradient.
constexpr char kNewGradSuffix[] = "@NEWGRAD@";
/// If an Op has this attribute, all its kernels should calculate output
/// variable's shape in the corresponding Compute() function. And
/// OperatorWithKernel::RunImpl() would skip call this Op's InferShape()
/// function in its runtime for speedup.
/// TODO(luotao): Note that this temporal attribute would be deleted after all
/// ops contain it.
constexpr char kAllKernelsMustComputeRuntimeShape[] =
"@ALL_KERNELS_MUST_COMPUTE_RUNTIME_SHAPE@";
// define some kernel priority
/* Define multiple kernel type fallback order*/
extern std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;

@ -14,8 +14,10 @@ limitations under the License. */
#include "paddle/fluid/framework/parallel_executor.h"
#include <algorithm>
#include <memory>
#include <string>
#include <tuple>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/graph_helper.h"
@ -181,13 +183,14 @@ std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
return member_->local_scopes_;
}
ParallelExecutor::ParallelExecutor(
const std::vector<platform::Place> &places,
const std::unordered_set<std::string> &bcast_vars,
const std::string &loss_var_name, Scope *scope,
const std::vector<Scope *> &local_scopes,
const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
ir::Graph *graph)
ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
const std::vector<std::string> &bcast_vars,
const std::string &loss_var_name,
Scope *scope,
const std::vector<Scope *> &local_scopes,
const ExecutionStrategy &exec_strategy,
const BuildStrategy &build_strategy,
ir::Graph *graph)
: member_(new ParallelExecutorPrivate(places)) {
member_->global_scope_ = scope;
member_->use_cuda_ = exec_strategy.use_cuda_;
@ -250,13 +253,41 @@ ParallelExecutor::ParallelExecutor(
member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
member_->places_, nccl_id, build_strategy.num_trainers_,
build_strategy.trainer_id_));
std::unique_ptr<platform::NCCLContextMap> dev_nccl_ctxs;
dev_nccl_ctxs.reset(new platform::NCCLContextMap(member_->places_));
// Initialize device context's nccl comm
// Note, more than one ParallelExecutor with same place, the nccl comm will
// be rewrite and there will be some problem.
for (size_t dev_id = 0; dev_id < member_->places_.size(); ++dev_id) {
auto &nccl_ctx = dev_nccl_ctxs->at(dev_id);
platform::DeviceContextPool &pool =
platform::DeviceContextPool::Instance();
auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
pool.Get(member_->places_[dev_id]));
dev_ctx->set_nccl_comm(nccl_ctx.comm());
}
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
}
if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
BCastParamsToDevices(bcast_vars);
// broadcast parameters from the 0th device to others:
auto need_broadcast = [&]() -> bool {
if (build_strategy.num_trainers_ > 1) {
// 1. num_tariners would be grater than 1 for nccl distributed training.
return true;
} else if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
// 2. Only one trainer process, but ParallelExecutor hold multiple
// devices.
return true;
}
return false;
};
if (need_broadcast()) {
BCastParamsToDevices(bcast_vars, build_strategy.trainer_id_);
}
// Startup Program has been run. All local scopes has correct parameters.
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
@ -338,7 +369,7 @@ ParallelExecutor::ParallelExecutor(
}
void ParallelExecutor::BCastParamsToDevices(
const std::unordered_set<std::string> &vars) const {
const std::vector<std::string> &vars, int trainer_id) const {
// the initializing bcast, all vars would be bcast from device(0).
for (auto &var : vars) {
framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
@ -362,7 +393,7 @@ void ParallelExecutor::BCastParamsToDevices(
auto place = member_->places_[i];
void *buffer;
if (i == 0) {
if (i == 0 && trainer_id == 0) {
buffer = const_cast<void *>(main_tensor.data<void>());
} else {
auto local_scope = member_->local_scopes_[i];

@ -14,9 +14,11 @@ limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/details/build_strategy.h"
@ -45,7 +47,7 @@ class ParallelExecutor {
public:
explicit ParallelExecutor(const std::vector<platform::Place> &places,
const std::unordered_set<std::string> &bcast_vars,
const std::vector<std::string> &bcast_vars,
const std::string &loss_var_name, Scope *scope,
const std::vector<Scope *> &local_scopes,
const ExecutionStrategy &exec_strategy,
@ -70,7 +72,10 @@ class ParallelExecutor {
const std::string &fetched_var_name);
private:
void BCastParamsToDevices(const std::unordered_set<std::string> &vars) const;
// broadcast the parameters from the 0th device.
// trainer_id the trainer index in nccl distributed training.
void BCastParamsToDevices(const std::vector<std::string> &vars,
int trainer_id = 0) const;
bool EnableParallelGraphExecution(const ir::Graph &graph,
const ExecutionStrategy &exec_strategy,
const BuildStrategy &build_strategy) const;

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save