m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge remote-tracking branch 'upstream/develop' into dev

Browse Source 
test=develop
revert-16839-cmakelist_change
heqiaozhi 7 years ago
parent 85a0746e9d 06809ebbb1
commit 759940786e
119 changed files with 2594 additions and 1071 deletions
Show Stats Download Patch File Download Diff File

3
CMakeLists.txt
Unescape View File

@ -71,7 +71,8 @@ option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plan
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(ON_INFER "Turn on inference optimization." OFF)
option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF)
option(WITH_INFERENCE_API_TEST "Test fluid inference C++ high-level api interface" OFF)
option(WITH_HIGH_LEVEL_API_TEST "Test fluid python high-level api interface" OFF)
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION})
option(WITH_FAST_MATH "Make use of fast math library, might affect the precision to some extent" ON)

1
cmake/external/protobuf.cmake vendored
Unescape View File

@ -221,6 +221,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
-DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
-DCMAKE_INSTALL_LIBDIR=lib
-DBUILD_SHARED_LIBS=OFF
CMAKE_CACHE_ARGS
-DCMAKE_INSTALL_PREFIX:PATH=${PROTOBUF_INSTALL_DIR}
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}

18
paddle/fluid/API.spec
Unescape View File

@ -13,6 +13,7 @@ paddle.fluid.name_scope (ArgSpec(args=['prefix'], varargs=None, keywords=None, d
paddle.fluid.cuda_places (ArgSpec(args=['device_ids'], varargs=None, keywords=None, defaults=(None,)), ('document', '7d9a51fc9cf3c5245b5227080a8064c3'))
paddle.fluid.cpu_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', '4c0cd83f0b401fc2ff84c70974e5d210'))
paddle.fluid.cuda_pinned_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd0c3ebd813c39958c92b78e3eef7e912'))
paddle.fluid.in_dygraph_mode (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', 'f06314a1cb30c96b5808dde2219c2dae'))
paddle.fluid.Executor.__init__ (ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.Executor.close (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'f5369953dd0c443961cf79f7a00e1a03'))
paddle.fluid.Executor.infer_from_dataset (ArgSpec(args=['self', 'program', 'dataset', 'scope', 'thread', 'debug', 'fetch_list', 'fetch_info', 'print_period'], varargs=None, keywords=None, defaults=(None, None, None, 0, False, None, None, 100)), ('document', '9c7decb955b9c4f718114179c8985581'))
@ -117,6 +118,8 @@ paddle.fluid.layers.reduce_mean (ArgSpec(args=['input', 'dim', 'keep_dim', 'name
paddle.fluid.layers.reduce_max (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', '66a622db727551761ce4eb73eaa7f6a4'))
paddle.fluid.layers.reduce_min (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'd50ac552b5d131468ed466d08bb2d38c'))
paddle.fluid.layers.reduce_prod (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'fcd8301a0ce15f219c7a4bcd0c1e8eca'))
paddle.fluid.layers.reduce_all (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', '646ca4d4a2cc16084f59de44b6927eca'))
paddle.fluid.layers.reduce_any (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'f36661060aeeaf6c6b1331e41b3726fa'))
paddle.fluid.layers.sequence_first_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '2b290d3d77882bfe9bb8d331cac8cdd3'))
paddle.fluid.layers.sequence_last_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'c16a892f44f7fe71bfa5afc32d3f34ce'))
paddle.fluid.layers.sequence_slice (ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'fdcea0e8b5bc7d8d4b1b072c521014e6'))
@ -124,7 +127,7 @@ paddle.fluid.layers.dropout (ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed
paddle.fluid.layers.split (ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)), ('document', '652625345c2acb900029c78cc75f8aa6'))
paddle.fluid.layers.ctc_greedy_decoder (ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'ebbf2adbd79683dc93db03454dfa18c2'))
paddle.fluid.layers.edit_distance (ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None)), ('document', '97f0262f97602644c83142789d784571'))
paddle.fluid.layers.l2_normalize (ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)), ('document', '6e428384ce6a77207fa2c70d9f011990'))
paddle.fluid.layers.l2_normalize (ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)), ('document', '35c6a241bcc1a1fc89508860d82ad62b'))
paddle.fluid.layers.matmul (ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'alpha', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, None)), ('document', 'b4cbe1ac451005df6dad12e9ffdccca9'))
paddle.fluid.layers.topk (ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd3570c02f71bcd78e60b3f31dc8f5b32'))
paddle.fluid.layers.warpctc (ArgSpec(args=['input', 'label', 'blank', 'norm_by_times', 'use_cudnn'], varargs=None, keywords=None, defaults=(0, False, False)), ('document', 'aaba49c038ba927f0a8e45c0c9a686ab'))
@ -155,10 +158,10 @@ paddle.fluid.layers.label_smooth (ArgSpec(args=['label', 'prior_dist', 'epsilon'
paddle.fluid.layers.roi_pool (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)), ('document', 'c317aa595deb31649083c8faa91cdb97'))
paddle.fluid.layers.roi_align (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)), ('document', '12c5bbb8b38c42e623fbc47611d766e1'))
paddle.fluid.layers.dice_loss (ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)), ('document', '1ba0508d573f65feecf3564dce22aa1d'))
paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', '7a1966d7c3a48f1fc0881cdaf5d83b0b'))
paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', 'd1b08c11bb9277386fcf6ae70b6622d1'))
paddle.fluid.layers.image_resize_short (ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)), ('document', '06211aefc50c5a3e940d7204d859cdf7'))
paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'e4fb4ed511b2293b8f04f7e872afbfd7'))
paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', '735fa9758a6d7ff3b47d7b827f961c1d'))
paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'c45591fbc4f64a178fbca219e1546a58'))
paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', 'ae6d73cdc7f3a138d8a338ecdb33c1ae'))
paddle.fluid.layers.gather (ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None), ('document', '98f1c86716b9b7f4dda83f20e2adeee2'))
paddle.fluid.layers.scatter (ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '65f8e9d8ddfd0b412f940579c4faa342'))
paddle.fluid.layers.sequence_scatter (ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '15b522457dfef103f0c20ca9d397678b'))
@ -203,6 +206,7 @@ paddle.fluid.layers.gaussian_random_batch_size_like (ArgSpec(args=['input', 'sha
paddle.fluid.layers.sum (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', 'a418e3ccb5e2ac21bd60f5cc221d5860'))
paddle.fluid.layers.slice (ArgSpec(args=['input', 'axes', 'starts', 'ends'], varargs=None, keywords=None, defaults=None), ('document', '01dbb91e7c74cb11336cd531013de51a'))
paddle.fluid.layers.shape (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '17db0f814eb7bb5a3fac1ca6e60e16d8'))
paddle.fluid.layers.rank (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'ee1386c42ecc8f424fe3fb21862fefc2'))
paddle.fluid.layers.logical_and (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', 'cdcf20c494c92060d10feb9374532f42'))
paddle.fluid.layers.logical_or (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '0eae3f726a4afe590757552fa3ced012'))
paddle.fluid.layers.logical_xor (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', 'b0daaa3fa4a0aa62f9b58c43d959eb25'))
@ -235,7 +239,7 @@ paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], vararg
paddle.fluid.layers.kldiv_loss (ArgSpec(args=['x', 'target', 'reduction', 'name'], varargs=None, keywords=None, defaults=('mean', None)), ('document', '776d536cac47c89073abc7ee524d5aec'))
paddle.fluid.layers.tree_conv (ArgSpec(args=['nodes_vector', 'edge_set', 'output_size', 'num_filters', 'max_depth', 'act', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(1, 2, 'tanh', None, None, None)), ('document', '34ea12ac9f10a65dccbc50100d12e607'))
paddle.fluid.layers.npair_loss (ArgSpec(args=['anchor', 'positive', 'labels', 'l2_reg'], varargs=None, keywords=None, defaults=(0.002,)), ('document', '46994d10276dd4cb803b4062b5d14329'))
paddle.fluid.layers.pixel_shuffle (ArgSpec(args=['x', 'upscale_factor'], varargs=None, keywords=None, defaults=None), ('document', 'ad669cdf83e72a69ebc5ed79e36486de'))
paddle.fluid.layers.pixel_shuffle (ArgSpec(args=['x', 'upscale_factor'], varargs=None, keywords=None, defaults=None), ('document', '731b21c62a4add60a33bd76d802ffc5c'))
paddle.fluid.layers.fsp_matrix (ArgSpec(args=['x', 'y'], varargs=None, keywords=None, defaults=None), ('document', 'b76ccca3735bea4a58a0dbf0d77c5393'))
paddle.fluid.layers.continuous_value_model (ArgSpec(args=['input', 'cvm', 'use_cvm'], varargs=None, keywords=None, defaults=(True,)), ('document', '88046160ef4bbd28f18fa6484d95b75c'))
paddle.fluid.layers.data (ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)), ('document', '33bbd42027d872b3818b3d64ec52e139'))
@ -272,6 +276,7 @@ paddle.fluid.layers.has_inf (ArgSpec(args=['x'], varargs=None, keywords=None, de
paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '2e53e83127dbfd86e7098bdfe9a549e8'))
paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
paddle.fluid.layers.linspace (ArgSpec(args=['start', 'stop', 'num', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '495e21e9a848c2d075a102802fc67756'))
paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@ -361,8 +366,7 @@ paddle.fluid.layers.inverse_time_decay (ArgSpec(args=['learning_rate', 'decay_st
paddle.fluid.layers.polynomial_decay (ArgSpec(args=['learning_rate', 'decay_steps', 'end_learning_rate', 'power', 'cycle'], varargs=None, keywords=None, defaults=(0.0001, 1.0, False)), ('document', '882634f420f626642f0874481263da40'))
paddle.fluid.layers.piecewise_decay (ArgSpec(args=['boundaries', 'values'], varargs=None, keywords=None, defaults=None), ('document', 'c717d9d1d78a53c809d01b8bc56f3cae'))
paddle.fluid.layers.noam_decay (ArgSpec(args=['d_model', 'warmup_steps'], varargs=None, keywords=None, defaults=None), ('document', 'd9a95746353fd574be36dc28d8726c28'))
paddle.fluid.layers.append_LARS (ArgSpec(args=['params_grads', 'learning_rate', 'weight_decay'], varargs=None, keywords=None, defaults=None), ('document', 'd24fa1e7d62ac8a534fc6a86002f84f8'))
paddle.fluid.layers.cosine_decay (ArgSpec(args=['learning_rate', 'step_each_epoch', 'epochs'], varargs=None, keywords=None, defaults=None), ('document', '9588c64c26ffaef3c466e404a6af9d9b'))
paddle.fluid.layers.cosine_decay (ArgSpec(args=['learning_rate', 'step_each_epoch', 'epochs'], varargs=None, keywords=None, defaults=None), ('document', 'f8b2727bccf0f368c997d7cf05847e49'))
paddle.fluid.layers.linear_lr_warmup (ArgSpec(args=['learning_rate', 'warmup_steps', 'start_lr', 'end_lr'], varargs=None, keywords=None, defaults=None), ('document', '2ef3f5ca5cd71ea4217c418e5a7a0565'))
paddle.fluid.contrib.InitState.__init__ (ArgSpec(args=['self', 'init', 'shape', 'value', 'init_boot', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, None, False, 'float32')), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.StateCell.__init__ (ArgSpec(args=['self', 'inputs', 'states', 'out_state', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))

30
paddle/fluid/framework/data_feed.cc
Unescape View File

@ -72,7 +72,6 @@ bool DataFeed::PickOneFile(std::string* filename) {
}
VLOG(3) << "file_idx_=" << *file_idx_;
*filename = filelist_[(*file_idx_)++];
// LOG(ERROR) << "pick file:" << *filename;
return true;
}
@ -466,6 +465,17 @@ void MultiSlotDataFeed::Init(
if (slot.is_used()) {
use_slots_.push_back(all_slots_[i]);
use_slots_is_dense_.push_back(slot.is_dense());
std::vector<int> local_shape;
if (slot.is_dense()) {
// for batch size holder if is_dense
if (slot.shape(0) > 0) {
local_shape.push_back(0);
}
}
for (size_t i = 0; i < slot.shape_size(); ++i) {
local_shape.push_back(slot.shape(i));
}
use_slots_shape_.push_back(local_shape);
}
}
feed_vec_.resize(use_slots_.size());
@ -752,8 +762,8 @@ void MultiSlotDataFeed::PutToFeedVec(
LoD data_lod{offset};
feed_vec_[i]->set_lod(data_lod);
if (use_slots_is_dense_[i]) {
int dim = total_instance / batch_size_;
feed_vec_[i]->Resize({batch_size_, dim});
use_slots_shape_[i][0] = batch_size_;
feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
}
}
#endif
@ -785,6 +795,16 @@ void MultiSlotInMemoryDataFeed::Init(
if (slot.is_used()) {
use_slots_.push_back(all_slots_[i]);
use_slots_is_dense_.push_back(slot.is_dense());
std::vector<int> local_shape;
if (slot.is_dense()) {
if (slot.shape(0) > 0) {
local_shape.push_back(0);
}
}
for (size_t i = 0; i < slot.shape_size(); ++i) {
local_shape.push_back(slot.shape(i));
}
use_slots_shape_.push_back(local_shape);
}
}
feed_vec_.resize(use_slots_.size());
@ -940,8 +960,8 @@ void MultiSlotInMemoryDataFeed::PutToFeedVec(
LoD data_lod{offset};
feed_vec_[i]->set_lod(data_lod);
if (use_slots_is_dense_[i]) {
int dim = total_instance / batch_size_;
feed_vec_[i]->Resize({batch_size_, dim});
use_slots_shape_[i][0] = batch_size_;
feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
}
}
#endif

1
paddle/fluid/framework/data_feed.h
Unescape View File

@ -142,6 +142,7 @@ class DataFeed {
// object)
std::vector<std::string> all_slots_;
std::vector<std::string> all_slots_type_;
std::vector<std::vector<int>> use_slots_shape_;
std::vector<int>
use_slots_index_; // -1: not used; >=0: the index of use_slots_

1
paddle/fluid/framework/data_feed.proto
Unescape View File

@ -19,6 +19,7 @@ message Slot {
required string type = 2;
optional bool is_dense = 3 [ default = false ];
optional bool is_used = 4 [ default = false ];
repeated int32 shape = 5; // we can define N-D Tensor
}
message MultiSlotDesc { repeated Slot slots = 1; }

6
paddle/fluid/framework/details/build_strategy.cc
Unescape View File

@ -150,6 +150,11 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
AppendPass("runtime_context_cache_pass");
}
if (strategy_.cache_expected_kernel_) {
VLOG(10) << "Add expected_kernel_cache_pass";
AppendPass("expected_kernel_cache_pass");
}
AppendMultiDevPass(strategy_);
if (strategy_.fuse_all_reduce_ops_) {
@ -337,3 +342,4 @@ USE_PASS(fuse_adam_op_pass);
USE_PASS(fuse_sgd_op_pass);
USE_PASS(fuse_all_reduce_op_pass);
USE_PASS(runtime_context_cache_pass);
USE_PASS(expected_kernel_cache_pass);

11
paddle/fluid/framework/details/build_strategy.h
Unescape View File

@ -83,11 +83,11 @@ struct BuildStrategy {
bool sync_batch_norm_{false};
bool memory_optimize_{true};
// TODO(dzhwinter):
// make enable_inplace, memory_optimize_
// memory_early_delete_ true by default
bool enable_inplace_{true};
// FIXME(liuwei1031) disable memory_optimzie and enable_inplace in 1.4
// to open them by default, we need to solve the fetch variable issue
bool memory_optimize_{false};
bool enable_inplace_{false};
bool enable_sequential_execution_{false};
@ -108,6 +108,7 @@ struct BuildStrategy {
bool remove_unnecessary_lock_{true};
bool cache_runtime_context_{false};
bool cache_expected_kernel_{true};
// NOTE:
// Before you add new options, think if it's a general strategy that works

9
paddle/fluid/framework/details/inplace_op_pass.cc
Unescape View File

@ -305,6 +305,12 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
if (var_nodes_[in_var_name].back() != in_node) {
VLOG(4) << "SKIP since " << in_var_name
<< " is also used as output by other ops";
continue;
}
bool can_replace = true;
if (in_var_name == out_var_name) {
can_replace = false;
@ -527,6 +533,9 @@ void GraphView::Build(ir::Graph* g) {
};
for (auto& node : g->Nodes()) {
if (!node->IsOp()) continue;
// avoid optimize the variable used in sub-blocks
if (OpHasSubBlock(node->Op())) update_skip_set(node);
if (node->Name() == "send") update_skip_set(node);
if (node->Name() == "recv") update_skip_set(node);
if (node->Name() == "prefetch") update_skip_set(node);

6
paddle/fluid/framework/details/op_registry.h
Unescape View File

@ -233,6 +233,12 @@ struct OpInfoFiller<T, kNoNeedBufferVarsInference> {
}
};
// A fake OpInfoFiller of void
template <>
struct OpInfoFiller<void, kUnknown> {
void operator()(const char* op_type, OpInfo* info) const {}
};
} // namespace details
} // namespace framework

50
paddle/fluid/framework/downpour_worker.cc
Unescape View File

@ -21,40 +21,40 @@ namespace framework {
void DownpourWorker::Initialize(const TrainerDesc& desc) {
param_ = desc.downpour_param();
for (size_t i = 0; i < param_.sparse_table_size(); ++i) {
for (int i = 0; i < param_.sparse_table_size(); ++i) {
uint64_t table_id =
static_cast<uint64_t>(param_.sparse_table(i).table_id());
TableParameter table = param_.sparse_table(i);
sparse_key_names_[table_id].resize(table.sparse_key_name_size());
for (size_t j = 0; j < table.sparse_key_name_size(); ++j) {
for (int j = 0; j < table.sparse_key_name_size(); ++j) {
sparse_key_names_[table_id][j] = table.sparse_key_name(j);
}
sparse_value_names_[table_id].resize(table.sparse_value_name_size());
for (size_t j = 0; j < table.sparse_value_name_size(); ++j) {
for (int j = 0; j < table.sparse_value_name_size(); ++j) {
sparse_value_names_[table_id][j] = table.sparse_value_name(j);
}
sparse_grad_names_[table_id].resize(table.sparse_grad_name_size());
for (size_t j = 0; j < table.sparse_grad_name_size(); ++j) {
for (int j = 0; j < table.sparse_grad_name_size(); ++j) {
sparse_grad_names_[table_id][j] = table.sparse_grad_name(j);
}
label_var_name_[table_id] = table.label_var_name();
}
for (size_t i = 0; i < param_.dense_table_size(); ++i) {
for (int i = 0; i < param_.dense_table_size(); ++i) {
uint64_t table_id = static_cast<uint64_t>(param_.dense_table(i).table_id());
auto table = param_.dense_table(i);
dense_value_names_[table_id].resize(table.dense_value_name_size());
for (size_t j = 0; j < table.dense_value_name_size(); ++j) {
for (int j = 0; j < table.dense_value_name_size(); ++j) {
dense_value_names_[table_id][j] = table.dense_value_name(j);
}
dense_grad_names_[table_id].resize(table.dense_grad_name_size());
for (size_t j = 0; j < table.dense_grad_name_size(); ++j) {
for (int j = 0; j < table.dense_grad_name_size(); ++j) {
dense_grad_names_[table_id][j] = table.dense_grad_name(j);
}
}
skip_ops_.resize(param_.skip_ops_size());
for (size_t i = 0; i < param_.skip_ops_size(); ++i) {
for (int i = 0; i < param_.skip_ops_size(); ++i) {
skip_ops_[i] = param_.skip_ops(i);
}
@ -83,14 +83,14 @@ void DownpourWorker::CollectLabelInfo(size_t table_idx) {
LoDTensor* tensor = var->GetMutable<LoDTensor>();
int64_t* label_ptr = tensor->data<int64_t>();
int global_index = 0;
size_t global_index = 0;
for (size_t i = 0; i < sparse_key_names_[table_id].size(); ++i) {
VLOG(3) << "sparse_key_names_[" << i
<< "]: " << sparse_key_names_[table_id][i];
Variable* fea_var = thread_scope_->FindVar(sparse_key_names_[table_id][i]);
LoDTensor* tensor = fea_var->GetMutable<LoDTensor>();
int64_t* ids = tensor->data<int64_t>();
int fea_idx = 0;
size_t fea_idx = 0;
// tensor->lod()[0].size() == batch_size + 1
for (auto lod_idx = 1u; lod_idx < tensor->lod()[0].size(); ++lod_idx) {
for (; fea_idx < tensor->lod()[0][lod_idx]; ++fea_idx) {
@ -138,7 +138,7 @@ void DownpourWorker::FillSparseValue(size_t table_idx) {
auto& tensor_lod = tensor->lod()[0];
LoD data_lod{tensor_lod};
tensor_emb->set_lod(data_lod);
for (auto index = 0u; index < len; ++index) {
for (int index = 0; index < len; ++index) {
if (ids[index] == 0u) {
memcpy(ptr + table.emb_dim() * index, init_value.data() + 2,
sizeof(float) * table.emb_dim());
@ -192,7 +192,7 @@ void DownpourWorker::TrainFilesWithProfiler() {
read_time += timeline.ElapsedSec();
total_time += timeline.ElapsedSec();
VLOG(3) << "program config size: " << param_.program_config_size();
for (size_t i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).pull_sparse_table_id(i));
@ -244,8 +244,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
}
if (need_to_push_sparse_) {
for (size_t i = 0;
i < param_.program_config(0).push_sparse_table_id_size(); ++i) {
for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).push_sparse_table_id(i));
TableParameter table;
@ -268,8 +268,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
if (need_to_push_dense_) {
timeline.Start();
for (size_t i = 0;
i < param_.program_config(0).push_dense_table_id_size(); ++i) {
for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).push_dense_table_id(i));
fleet_ptr_->PushDenseVarsAsync(
@ -315,8 +315,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
}
if (need_to_push_dense_) {
for (size_t i = 0;
i < param_.program_config(0).push_dense_table_id_size(); ++i) {
for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).push_dense_table_id(i));
pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);
@ -362,7 +362,7 @@ void DownpourWorker::TrainFiles() {
int cur_batch;
while ((cur_batch = device_reader_->Next()) > 0) {
// pull sparse here
for (size_t i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).pull_sparse_table_id(i));
@ -397,8 +397,8 @@ void DownpourWorker::TrainFiles() {
if (need_to_push_sparse_) {
// push gradients here
for (size_t i = 0;
i < param_.program_config(0).push_sparse_table_id_size(); ++i) {
for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).push_sparse_table_id(i));
TableParameter table;
@ -416,8 +416,8 @@ void DownpourWorker::TrainFiles() {
}
if (need_to_push_dense_) {
for (size_t i = 0;
i < param_.program_config(0).push_dense_table_id_size(); ++i) {
for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).push_dense_table_id(i));
fleet_ptr_->PushDenseVarsAsync(
@ -461,8 +461,8 @@ void DownpourWorker::TrainFiles() {
}
if (need_to_push_dense_) {
for (size_t i = 0;
i < param_.program_config(0).push_dense_table_id_size(); ++i) {
for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
++i) {
uint64_t tid = static_cast<uint64_t>(
param_.program_config(0).push_dense_table_id(i));
pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);

2
paddle/fluid/framework/ir/expected_kernel_cache_pass.cc
Unescape View File

@ -23,7 +23,7 @@ namespace ir {
void ExpectedKernelCachePass::ApplyImpl(ir::Graph* graph) const {
VLOG(3) << "Applies Expected Kernel Cache strategy.";
for (const Node* n : graph->Nodes()) {
if (n->IsOp()) {
if (n->IsOp() && n->Op()) {
n->Op()->SetAttr(kEnableCacheExpectedKernel, true);
}
}

36
paddle/fluid/framework/ir/graph_helper.cc
Unescape View File

@ -31,10 +31,10 @@ namespace paddle {
namespace framework {
namespace ir {
namespace {
void SortHelper(
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
ir::Node *node, std::unordered_set<ir::Node *> *visited,
std::vector<ir::Node *> *ret) {
void SortHelper(const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>,
ir::NodeComp> &adj_list,
ir::Node *node, std::unordered_set<ir::Node *> *visited,
std::vector<ir::Node *> *ret) {
visited->insert(node);
for (auto adj : adj_list.at(node)) {
@ -50,7 +50,8 @@ void SortHelper(
bool HasCircleHelper(
ir::Node *node,
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
&adj_list,
std::unordered_set<ir::Node *> *visited,
std::unordered_set<ir::Node *> *in_trace,
std::vector<std::vector<ir::Node *>> *circles) {
@ -84,7 +85,8 @@ bool HasCircleHelper(
}
bool HasCircleInternal(
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
&adj_list,
std::vector<std::vector<ir::Node *>> *circles) {
std::unordered_set<ir::Node *> visited;
std::unordered_set<ir::Node *> in_trace;
@ -107,8 +109,8 @@ bool FindCircleSubGraph(const Graph &graph,
}
std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
BuildOperationAdjList(graph);
std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
adj_list = BuildOperationAdjList(graph);
PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr));
std::unordered_set<ir::Node *> visited;
std::vector<ir::Node *> ret;
@ -117,34 +119,30 @@ std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
SortHelper(adj_list, adj.first, &visited, &ret);
}
}
return ret;
}
// Build operator inlink edge table.
std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
const Graph &graph) {
std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list;
std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
BuildOperationAdjList(const Graph &graph) {
std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
adj_list;
for (auto &n : graph.Nodes()) {
if (!n->IsOp()) continue;
if (adj_list.find(n) == adj_list.end()) {
adj_list[n] = std::unordered_set<ir::Node *>();
adj_list[n] = std::set<ir::Node *, ir::NodeComp>();
}
std::vector<ir::Node *> nodes;
for (auto &var : n->inputs) {
for (auto &adj_n : var->inputs) {
PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
<< " -> " << n->Name() << reinterpret_cast<void *>(n)
<< " via " << var->Name() << reinterpret_cast<void *>(var);
nodes.push_back(adj_n);
adj_list[n].insert(adj_n);
}
}
std::sort(nodes.begin(), nodes.end(), [](ir::Node *node1, ir::Node *node2) {
return node1->id() > node2->id();
});
adj_list[n].insert(std::make_move_iterator(nodes.begin()),
std::make_move_iterator(nodes.end()));
}
return adj_list;
}

12
paddle/fluid/framework/ir/graph_helper.h
Unescape View File

@ -16,6 +16,7 @@ limitations under the License. */
#include <map>
#include <memory>
#include <set>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
@ -25,6 +26,13 @@ namespace paddle {
namespace framework {
namespace ir {
// Compare nodes via node id.
struct NodeComp {
bool operator()(ir::Node *const &node1, ir::Node *const &node2) const {
return node1->id() < node2->id();
}
};
// Test if the graph contains circle.
bool HasCircle(const Graph &graph);
@ -57,8 +65,8 @@ std::vector<Node *> TopologyVarientSort(const Graph &graph, SortKind sort_kind);
void CleanIndividualNodes(Graph *graph);
// Build an adjacency list of operations for the `graph`.
std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
const Graph &graph);
std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
BuildOperationAdjList(const Graph &graph);
template <typename T>
std::vector<T *> FilterByNodeWrapper(const Graph &graph) {

1
paddle/fluid/framework/op_desc.cc
Unescape View File

@ -241,6 +241,7 @@ OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
outputs_ = outputs;
attrs_ = attrs;
need_update_ = true;
block_ = nullptr;
}
OpDesc::OpDesc(const OpDesc &other, BlockDesc *block) {

15
paddle/fluid/framework/operator.cc
Unescape View File

@ -880,7 +880,16 @@ std::vector<KernelConfig>* OperatorWithKernel::GetKernelConfig(
void OperatorWithKernel::RunImpl(const Scope& scope,
const platform::Place& place) const {
if (!HasAttr(kEnableCacheRuntimeContext)) {
// To reduce the elapsed time of HasAttr, we use bool variable to record the
// result of HasAttr.
if (!enable_cache_runtime_context && HasAttr(kEnableCacheRuntimeContext))
enable_cache_runtime_context = true;
if (!enable_cache_expected_kernel && HasAttr(kEnableCacheExpectedKernel))
enable_cache_expected_kernel = true;
if (!all_kernels_must_compute_runtime_shape &&
HasAttr(kAllKernelsMustComputeRuntimeShape))
all_kernels_must_compute_runtime_shape = true;
if (!enable_cache_runtime_context) {
RuntimeContext ctx(Inputs(), Outputs(), scope);
RunImpl(scope, place, &ctx);
} else {
@ -899,7 +908,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx = pool.Get(place);
if (!HasAttr(kEnableCacheExpectedKernel) || !kernel_type_) {
if (!enable_cache_expected_kernel || !kernel_type_) {
ChooseKernel(*runtime_ctx, scope, place);
}
@ -918,7 +927,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
dev_ctx = pool.Get(kernel_type_->place_);
}
if (!HasAttr(kAllKernelsMustComputeRuntimeShape)) {
if (!all_kernels_must_compute_runtime_shape) {
RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, *runtime_ctx);
this->InferShape(&infer_shape_ctx);
}

3
paddle/fluid/framework/operator.h
Unescape View File

@ -506,6 +506,9 @@ class OperatorWithKernel : public OperatorBase {
mutable std::unique_ptr<OpKernelFunc> kernel_func_;
mutable std::unique_ptr<RuntimeContext> runtime_ctx_;
mutable const Scope* pre_scope_ = nullptr;
mutable bool enable_cache_runtime_context = false;
mutable bool enable_cache_expected_kernel = false;
mutable bool all_kernels_must_compute_runtime_shape = false;
};
extern bool OpSupportGPU(const std::string& op_type);

4
paddle/fluid/framework/parallel_executor.cc
Unescape View File

@ -221,7 +221,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
PADDLE_ENFORCE(!member_->use_cuda_,
"gpu mode does not support async_mode_ now!");
graphs.push_back(graph);
for (int i = 1; i < places.size(); ++i) {
for (size_t i = 1; i < places.size(); ++i) {
auto *tmp_graph = new ir::Graph(graph->OriginProgram());
async_graphs_.emplace_back(tmp_graph);
graphs.push_back(tmp_graph);
@ -315,7 +315,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
graph = build_strategy.Apply(graph, {member_->places_[0]}, loss_var_name,
{member_->local_scopes_[0]}, 1,
member_->use_cuda_, member_->nccl_ctxs_.get());
for (int i = 1; i < member_->places_.size(); ++i) {
for (size_t i = 1; i < member_->places_.size(); ++i) {
graphs[i] =
build_strategy.Apply(graphs[i], {member_->places_[i]}, loss_var_name,
{member_->local_scopes_[i]}, 1,

2
paddle/fluid/framework/trainer_desc.proto
Unescape View File

@ -76,7 +76,7 @@ message PullDenseWorkerParameter {
message TableParameter {
// dense table only
optional int64 table_id = 1;
optional uint64 table_id = 1;
repeated string dense_value_name = 2;
repeated string dense_grad_name = 3;
repeated int32 push_dense_wait_times = 5;

3
paddle/fluid/inference/api/analysis_predictor.cc
Unescape View File

@ -259,6 +259,9 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
return false;
}
PADDLE_ENFORCE_NOT_NULL(input_ptr);
PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
if (platform::is_cpu_place(place_)) {
// TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),

1
paddle/fluid/inference/api/api.cc
Unescape View File

@ -54,6 +54,7 @@ PaddleBuf &PaddleBuf::operator=(const PaddleBuf &other) {
memory_owned_ = other.memory_owned_;
} else {
Resize(other.length());
PADDLE_ENFORCE(!(other.length() > 0 && other.data() == nullptr));
memcpy(data_, other.data(), other.length());
length_ = other.length();
memory_owned_ = true;

5
paddle/fluid/inference/api/api_impl.cc
Unescape View File

@ -169,6 +169,7 @@ std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
// Hot fix the bug that result diff in multi-thread.
// TODO(Superjomn) re-implement a real clone here.
PADDLE_ENFORCE_NOT_NULL(dynamic_cast<NativePaddlePredictor *>(cls.get()));
if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init(nullptr)) {
LOG(ERROR) << "fail to call Init";
return nullptr;
@ -210,6 +211,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
return false;
}
PADDLE_ENFORCE_NOT_NULL(input_ptr);
PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
if (platform::is_cpu_place(place_)) {
// TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
@ -316,6 +319,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
}
std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
PADDLE_ENFORCE_NOT_NULL(
dynamic_cast<NativePaddlePredictor *>(predictor.get()));
if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init(nullptr)) {
return nullptr;
}

2
paddle/fluid/inference/api/paddle_pass_builder.cc
Unescape View File

@ -123,8 +123,8 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
// will enhance this pass later.
"runtime_context_cache_pass", //
"attention_lstm_fuse_pass", //
"seqpool_concat_fuse_pass", //
"seqconv_eltadd_relu_fuse_pass", //
// "seqpool_concat_fuse_pass", //
// "embedding_fc_lstm_fuse_pass", //
"fc_lstm_fuse_pass", //
"mul_lstm_fuse_pass", //

1
paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc
Unescape View File

@ -47,6 +47,7 @@ struct DataRecord {
num_lines++;
std::vector<std::string> data;
split(line, '\t', &data);
PADDLE_ENFORCE(data.size() >= 4);
// load title1 data
std::vector<int64_t> title1_data;
split_to_int64(data[0], ' ', &title1_data);

3
paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
Unescape View File

@ -150,6 +150,9 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
if (use_mkldnn) {
cfg->EnableMKLDNN();
}
// Enable seqpool_concat_fuse_pass, disabled by default since it takes much
// time
cfg->pass_builder()->InsertPass(2, "seqpool_concat_fuse_pass");
}
void profile(bool use_mkldnn = false) {

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

Write Preview
Loading…
Cancel
Save