diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py index 10b633a4fc1063aab5c0d34b994f9c233e228f17..df159a334e86d62e175bce3b363b74ec78c1fd64 100644 --- a/benchmark/fluid/fluid_benchmark.py +++ b/benchmark/fluid/fluid_benchmark.py @@ -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] diff --git a/cmake/operators.cmake b/cmake/operators.cmake index 11a5b1b4554e7899c3ee7092a9295234743750d7..34c6cbd73ddd67860ef4e74ad7ce98b9b954d9ad 100644 --- a/cmake/operators.cmake +++ b/cmake/operators.cmake @@ -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() diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index df3497de209e3b6ede6986e1ac5f92c4427ca9bd..057ce9341b7bd4296806d02ec4349b09b23d2f1f 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -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')) diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 7ddf1ab44fe096739f4d241994e5cb686970a7c5..ad19d729ebde4a9c81c283518f3cb2ac28152443 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -38,10 +38,10 @@ if(WITH_GPU) nv_library(tensor SRCS tensor.cc .tensor_util.cu DEPS place memory data_type device_context) add_dependencies(tensor tensor_util) else() - nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context ) + nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context profiler) endif(WIN32) else() - cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context ) + cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context profiler) endif() cc_test(tensor_test SRCS tensor_test.cc DEPS tensor) @@ -174,7 +174,7 @@ else() cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op) endif() -target_link_libraries(executor garbage_collector) +target_link_libraries(executor garbage_collector while_op_helper) cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index dc308fd2592bb158f46f6eac9dd0df25787559fe..9f06455ea5410bcab081ed212a34960f8fe6f0bf 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -61,7 +61,8 @@ cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_ cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle) cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper) -cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass) +cc_library(while_op_eager_deletion_pass SRCS while_op_eager_deletion_pass.cc DEPS while_op_helper graph_helper pass computation_op_handle) +cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass while_op_eager_deletion_pass) cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper) cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass) diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index 2cfc76e47f41862731fb2de5d1d03287acd4d9d7..932d0b4538eb2ec5df97d0bde806c33f825b6f68 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -16,6 +16,7 @@ limitations under the License. */ #include #include +#include #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 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); diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index d755a2505aead37538bef2b01a193dba87dc1567..122411641dacde57ef3851f05bc92d86c1f83866 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -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_ diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index 1e3dbb1e44ecb16872e3bf4dee31e31cc69c9818..e98b16e6b3a07bfa0994295306e3bfa9e4174834 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include @@ -31,6 +32,8 @@ class ComputationOpHandle : public OpHandleBase { ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place, size_t scope_idx); + OperatorBase *GetOp() { return op_.get(); } + std::string Name() const override; const Scope *GetScope() const { return scope_; } diff --git a/paddle/fluid/framework/details/eager_deletion_op_handle.cc b/paddle/fluid/framework/details/eager_deletion_op_handle.cc index 03fbfd7f24a8a987db72f45be777acc7ece577a6..dbc90737f2286db6e74d3271f39d004c25e4a949 100644 --- a/paddle/fluid/framework/details/eager_deletion_op_handle.cc +++ b/paddle/fluid/framework/details/eager_deletion_op_handle.cc @@ -12,6 +12,10 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include +#include + #include "paddle/fluid/framework/details/eager_deletion_op_handle.h" #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/scope.h" @@ -45,6 +49,7 @@ EagerDeletionOpHandle::EagerDeletionOpHandle( } } #endif + PADDLE_ENFORCE(!var_names_.empty(), "Var names cannot be empty"); } EagerDeletionOpHandle::~EagerDeletionOpHandle() { @@ -60,15 +65,20 @@ EagerDeletionOpHandle::~EagerDeletionOpHandle() { std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; } void EagerDeletionOpHandle::RunImpl() { - auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get(); + Scope *exec_scope = nullptr; std::deque> garbages; for (auto &name : var_names_) { auto it = ref_cnts_->find(name); - // Var not found, not reference count has not decreased to 0 + // Reference count has not decreased to 0 if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) { continue; } + if (!exec_scope) { + exec_scope = scope_->FindVar(kLocalExecScopeName)->Get(); + } + + // Var not found auto *var = exec_scope->FindVar(name); if (var == nullptr) { continue; diff --git a/paddle/fluid/framework/details/eager_deletion_pass.cc b/paddle/fluid/framework/details/eager_deletion_pass.cc index 4e42d0b4972d567dd769cad6ff8b9d45380ab77a..377bb915e0ce175d4e3fb74cb1ace21e5f46d9d8 100644 --- a/paddle/fluid/framework/details/eager_deletion_pass.cc +++ b/paddle/fluid/framework/details/eager_deletion_pass.cc @@ -12,20 +12,173 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include #include #include +#include #include #include "paddle/fluid/framework/details/computation_op_handle.h" #include "paddle/fluid/framework/details/eager_deletion_op_handle.h" -#include "paddle/fluid/framework/details/eager_deletion_pass.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/ir/graph_helper.h" +DEFINE_double(memory_fraction_of_eager_deletion, 1.0, + "Fraction of eager deletion. If less than 1.0, all variables in " + "the program would be sorted according to its memory size, and " + "only the FLAGS_memory_fraction_of_eager_deletion of the largest " + "variables would be deleted."); + namespace paddle { namespace framework { namespace details { +// op -> variables which can be deleted after op runs +using OpToVarNameSetMap = + std::unordered_map>; + +// Check whether the variable is LoDTensor based on static VarDesc info +static bool IsLoDTensor(VarDesc *var) { + return var->Proto()->type().type() == proto::VarType::LOD_TENSOR; +} + +// Get memory size of LoDTensor +static int64_t GetMemorySize( + const std::unordered_map> &vars, + const std::string &var_name) { + auto *var_desc = TryGetLatestVarDesc(vars.at(var_name)); + PADDLE_ENFORCE_NOT_NULL(var_desc); + PADDLE_ENFORCE(IsLoDTensor(var_desc)); + auto dims = var_desc->GetShape(); + return SizeOfType(var_desc->GetDataType()) * + std::accumulate(dims.begin(), dims.end(), static_cast(1), + std::multiplies()); +} + +// Split all variables in the graph into LoDTensor and Non-LoDTensor (e.g. +// SelectedRows, LoDTensorArray) +// Since partial GC is based on static analysis of memory size of each variable +// So we should skip SelectedRows and LoDTensorArray here +static void SplitIntoLoDTensorAndNonLoDTensorVars( + const OpToVarNameSetMap &m, const GraphVars &vars, + OpToVarNameSetMap *lod_tensors, OpToVarNameSetMap *other_vars) { + lod_tensors->clear(); + other_vars->clear(); + + for (auto &op_vars_pair : m) { + for (auto &var_name : op_vars_pair.second) { + auto *var_desc = TryGetLatestVarDesc( + vars[op_vars_pair.first->GetScopeIdx()].at(var_name)); + if (IsLoDTensor(var_desc)) { + (*lod_tensors)[op_vars_pair.first].insert(var_name); + } else { + (*other_vars)[op_vars_pair.first].insert(var_name); + } + } + } +} + +struct GCVarInfo { + GCVarInfo(const std::string &name, int64_t memory_size, + ComputationOpHandle *op, size_t scope_idx) + : name_(name), + memory_size_(memory_size), + op_(op), + scope_idx_(scope_idx) {} + + std::string name_; // variable name + int64_t memory_size_; // memory size + ComputationOpHandle *op_; // op after which the variable could be deleted + size_t scope_idx_; // scope index where the variable locates + + int64_t AbsMemorySize() const { return std::abs(memory_size_); } +}; + +// Delete delete_lod_tensor_only is not used currently +static OpToVarNameSetMap ShrinkGCVars( + const OpToVarNameSetMap &m, const GraphVars &vars, + const std::vector &places, double fraction_of_memory_size, + bool delete_lod_tensor_only = false) { + // Do not perform gc when fraction_of_memory_size = 0 + if (fraction_of_memory_size <= 0.0) return {}; + + /** + * Step 1: Split all variables into LoDTensor and Non-LoDTensor. + * We can only calculate memory size of LoDTensors + */ + OpToVarNameSetMap lod_tensors, other_vars; + SplitIntoLoDTensorAndNonLoDTensorVars(m, vars, &lod_tensors, &other_vars); + + // Perform complete gc when fraction_of_memory_size >= 1 + if (fraction_of_memory_size >= 1.0) { + return delete_lod_tensor_only ? lod_tensors : m; + } + + /** + * Step 2: build GCVarInfos, and calculate total memory sizes of each device + */ + + // place -> variable info (name, memory size, place, scope_idx) + std::map> place_to_vars; + + // place -> total memory sizes + std::map place_to_size; + for (auto &op_vars_pair : lod_tensors) { + auto *op = op_vars_pair.first; + auto &var_names = op_vars_pair.second; + auto scope_idx = op->GetScopeIdx(); + auto &place = places[scope_idx]; + + for (auto &var_name : var_names) { + auto var_size = GetMemorySize(vars[scope_idx], var_name); + GCVarInfo var_info(var_name, var_size, op, scope_idx); + place_to_size[place] += var_info.AbsMemorySize(); + place_to_vars[place].emplace_back(std::move(var_info)); + } + } + + /** + * Step 3: sort GCVarInfos, and only delete the largest variables. + */ + OpToVarNameSetMap partial_vars; + for (auto &place_to_var_pair : place_to_vars) { + auto &place = place_to_var_pair.first; + auto &gc_vars = place_to_var_pair.second; + std::sort(gc_vars.begin(), gc_vars.end(), + [](const GCVarInfo &var1, const GCVarInfo &var2) { + return var1.AbsMemorySize() > var2.AbsMemorySize(); + }); + + int64_t accumulated_size = 0; + int64_t size_threshold = + static_cast(fraction_of_memory_size * place_to_size[place]); + for (size_t i = 0; i < gc_vars.size() && accumulated_size < size_threshold; + ++i) { + partial_vars[gc_vars[i].op_].insert(gc_vars[i].name_); + accumulated_size += gc_vars[i].AbsMemorySize(); + } + } + + /** + * Step 4: Combine other vars (SelectedRows, LoDTensorArray) + */ + if (!delete_lod_tensor_only) { + for (auto &op_vars_pair : other_vars) { + partial_vars[op_vars_pair.first].insert(op_vars_pair.second.begin(), + op_vars_pair.second.end()); + } + } + + return partial_vars; +} + +class EagerDeletionPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + std::unique_ptr EagerDeletionPass::ApplyImpl( std::unique_ptr graph) const { auto &ref_cnts = @@ -43,9 +196,7 @@ std::unique_ptr EagerDeletionPass::ApplyImpl( // a reverse map of last_live_ops // i.e., last op --> variable names which can be deleted. - std::unordered_map> - op_vars_map; - + OpToVarNameSetMap op_vars_map; for (auto &var_ops_map : last_live_ops) { for (auto &var_ops_pair : var_ops_map) { const std::string &var_name = var_ops_pair.first; @@ -55,6 +206,9 @@ std::unique_ptr EagerDeletionPass::ApplyImpl( } } + op_vars_map = ShrinkGCVars(op_vars_map, vars, places, + FLAGS_memory_fraction_of_eager_deletion); + for (auto &pair : op_vars_map) { auto *op = pair.first; auto &var_names = pair.second; @@ -85,8 +239,13 @@ std::unique_ptr EagerDeletionPass::ApplyImpl( eager_deletion_op->AddOutput(dummy_leaf); } + VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " + << FLAGS_memory_fraction_of_eager_deletion; VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)"; - return graph; + + auto while_op_eager_deletion_pass = + ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass"); + return while_op_eager_deletion_pass->Apply(std::move(graph)); } } // namespace details @@ -99,3 +258,5 @@ REGISTER_PASS(eager_deletion_pass, .RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars) .RequirePassAttr(paddle::framework::details::kAllPlaces) .RequirePassAttr(paddle::framework::details::kGarbageCollector); + +USE_PASS(while_op_eager_deletion_pass); diff --git a/paddle/fluid/framework/details/eager_deletion_pass.h b/paddle/fluid/framework/details/eager_deletion_pass.h deleted file mode 100644 index d7a7a9709d970841060778806451bc21cb2c7571..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/eager_deletion_pass.h +++ /dev/null @@ -1,32 +0,0 @@ -// Copyright (c) 2018 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 "paddle/fluid/framework/ir/graph.h" -#include "paddle/fluid/framework/ir/pass.h" - -namespace paddle { -namespace framework { -namespace details { - -class EagerDeletionPass : public ir::Pass { - protected: - std::unique_ptr ApplyImpl( - std::unique_ptr graph) const override; -}; - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/inplace_op_pass.cc b/paddle/fluid/framework/details/inplace_op_pass.cc index c91fc81b2defc9fe6b5720ce652a9aa94b27735e..8d4717ad19d4ca0525eac4d1a0dfe6d0076a8c09 100644 --- a/paddle/fluid/framework/details/inplace_op_pass.cc +++ b/paddle/fluid/framework/details/inplace_op_pass.cc @@ -16,6 +16,7 @@ #include #include #include +#include #include #include #include @@ -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. diff --git a/paddle/fluid/framework/details/memory_optimize_helper.cc b/paddle/fluid/framework/details/memory_optimize_helper.cc index c89a33fc959247afb74dab49056fc3fca8b9bd89..533d3269be350de35954e575965fe7a089941058 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper.cc @@ -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; } diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc index e7284ea64438557161a0c97a6a7f45fb9bb245ca..80720af32d5670928a6ad2b9efbeadf6452b0273 100644 --- a/paddle/fluid/framework/details/memory_optimize_pass.cc +++ b/paddle/fluid/framework/details/memory_optimize_pass.cc @@ -24,6 +24,7 @@ #include #include #include +#include #include #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())) { diff --git a/paddle/fluid/framework/details/reference_count_pass.cc b/paddle/fluid/framework/details/reference_count_pass.cc index 13a042d8e6ed7f18c76387b666d681df0eabd0b5..6092143449bc8e20117e7021bd44553cf64ae5b5 100644 --- a/paddle/fluid/framework/details/reference_count_pass.cc +++ b/paddle/fluid/framework/details/reference_count_pass.cc @@ -12,9 +12,13 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include #include #include #include +#include +#include +#include #include #include "paddle/fluid/framework/details/computation_op_handle.h" @@ -189,15 +193,6 @@ ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx, return shrink_func(computation_op); } -static VarDesc *TryGetLatestVarDesc(const std::vector &vars) { - VarDesc *var_desc = nullptr; - std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool { - var_desc = var_handle->Node()->Var(); - return var_desc != nullptr; - }); - return var_desc; -} - std::unique_ptr ReferenceCountPass::ApplyImpl( std::unique_ptr graph) const { auto &ref_cnts = Get>(kGlobalReferenceCount); diff --git a/paddle/fluid/framework/details/reference_count_pass_helper.cc b/paddle/fluid/framework/details/reference_count_pass_helper.cc index 89bd08c2d041d795205b29bb29aba311d1dbd932..94de0e6ab0a91d90a7f2c4c4fc14eb78663c95fe 100644 --- a/paddle/fluid/framework/details/reference_count_pass_helper.cc +++ b/paddle/fluid/framework/details/reference_count_pass_helper.cc @@ -13,9 +13,22 @@ // limitations under the License. #include "paddle/fluid/framework/details/reference_count_pass_helper.h" +#include "paddle/fluid/framework/details/var_handle.h" +#include "paddle/fluid/framework/var_desc.h" namespace paddle { namespace framework { -namespace details {} // namespace details +namespace details { + +VarDesc *TryGetLatestVarDesc(const std::vector &vars) { + VarDesc *var_desc = nullptr; + std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool { + var_desc = var_handle->Node()->Var(); + return var_desc != nullptr; + }); + return var_desc; +} + +} // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/reference_count_pass_helper.h b/paddle/fluid/framework/details/reference_count_pass_helper.h index 1c083dbf001b08e40a54cc89b21c3dea1f18f16a..ce700119c54ddd711315dfa45d61b9241cfda651 100644 --- a/paddle/fluid/framework/details/reference_count_pass_helper.h +++ b/paddle/fluid/framework/details/reference_count_pass_helper.h @@ -16,6 +16,7 @@ #include #include +#include #include #include #include @@ -25,6 +26,10 @@ namespace paddle { namespace framework { + +class VarDesc; +class VarHandle; + namespace details { class ComputationOpHandle; @@ -43,9 +48,11 @@ const char kGarbageCollector[] = "garbage_collector"; const char kAllPlaces[] = "all_places"; using LastLiveOpsOfVars = - std::unordered_map>; + std::unordered_map>; const char kLastLiveOpsOfVars[] = "last_live_ops_of_var"; +VarDesc *TryGetLatestVarDesc(const std::vector &vars); + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/while_op_eager_deletion_pass.cc b/paddle/fluid/framework/details/while_op_eager_deletion_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..fd6b6dd2274d9721b8754e16cd7b4f1ab596380d --- /dev/null +++ b/paddle/fluid/framework/details/while_op_eager_deletion_pass.cc @@ -0,0 +1,62 @@ +// 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/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/operators/controlflow/while_op_helper.h" + +namespace paddle { +namespace framework { +namespace details { + +class WhileOpEagerDeletionPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override { + auto all_ops = ir::FilterByNodeWrapper(*graph); + + // Find all while_op and while_grad_op + std::unordered_map, + std::vector>> + target_ops; + for (auto *op : all_ops) { + auto compute_op = dynamic_cast(op); + if (compute_op == nullptr) continue; + + if (compute_op->Name() == "while") { + target_ops[compute_op->GetScopeIdx()].first.emplace_back( + compute_op->GetOp()); + } else if (compute_op->Name() == "while_grad") { + target_ops[compute_op->GetScopeIdx()].second.emplace_back( + compute_op->GetOp()); + } + } + + for (auto &ops_pair : target_ops) { + auto &while_ops = ops_pair.second.first; + auto &while_grad_ops = ops_pair.second.second; + operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp( + while_ops, while_grad_ops); + } + return graph; + } +}; + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(while_op_eager_deletion_pass, + paddle::framework::details::WhileOpEagerDeletionPass); diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index c31d0beec306fe165164837cd15c95b4efd76af0..99192292b0be992d5ff0ecebba6294b9ba27e958 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -14,6 +14,10 @@ limitations under the License. */ #include "paddle/fluid/framework/executor.h" #include +#include +#include +#include +#include #include "paddle/fluid/framework/feed_fetch_method.h" #include "paddle/fluid/framework/lod_rank_table.h" @@ -23,17 +27,18 @@ limitations under the License. */ #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/framework/transfer_scope_cache.h" #include "paddle/fluid/framework/variable_helper.h" +#include "paddle/fluid/operators/controlflow/while_op_helper.h" #include "paddle/fluid/operators/distributed/distributed.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" #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 { @@ -75,11 +80,11 @@ static std::unordered_map GetNonPersistableReferenceCounts( ExecutorPrepareContext::ExecutorPrepareContext( const framework::ProgramDesc& prog, size_t block_id, - const std::vector& skip_ref_cnt_vars) - : prog_(prog), block_id_(block_id) { - if (GetEagerDeletionThreshold() >= 0) { - global_ref_cnts_ = GetNonPersistableReferenceCounts(prog.Block(block_id), - skip_ref_cnt_vars); + const std::vector& keep_vars, bool force_disable_gc) + : prog_(prog), block_id_(block_id), force_disable_gc_(force_disable_gc) { + if (GetEagerDeletionThreshold() >= 0 && !force_disable_gc_) { + global_ref_cnts_ = + GetNonPersistableReferenceCounts(prog.Block(block_id), keep_vars); } } @@ -184,13 +189,12 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, } void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, - bool create_local_scope, bool create_vars) { + bool create_local_scope, bool create_vars, + const std::vector& skip_ref_cnt_vars, + 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); + auto ctx = Prepare(pdesc, block_id, skip_ref_cnt_vars, force_disable_gc); RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars); } @@ -357,20 +361,27 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, std::unique_ptr Executor::Prepare( const ProgramDesc& program, int block_id, - const std::vector& skip_ref_cnt_vars) { - std::unique_ptr ctx( - new ExecutorPrepareContext(program, block_id, skip_ref_cnt_vars)); + const std::vector& skip_ref_cnt_vars, bool force_disable_gc) { + std::unique_ptr ctx(new ExecutorPrepareContext( + program, block_id, skip_ref_cnt_vars, force_disable_gc)); PADDLE_ENFORCE_LT(static_cast(block_id), program.Size()); auto& block = program.Block(block_id); 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; } std::vector> Executor::Prepare( const ProgramDesc& program, const std::vector& block_ids, - const std::vector>& skip_ref_cnt_vars) { + const std::vector>& skip_ref_cnt_vars, + bool force_disable_gc) { PADDLE_ENFORCE( skip_ref_cnt_vars.empty() || skip_ref_cnt_vars.size() == block_ids.size(), "skip_ref_cnt_vars should be either empty or equals to block number %d", @@ -380,9 +391,11 @@ std::vector> Executor::Prepare( for (auto& bid : block_ids) { ExecutorPrepareContext* ctx; if (skip_ref_cnt_vars.empty()) { - ctx = new ExecutorPrepareContext(program, bid); + ctx = new ExecutorPrepareContext(program, bid, std::vector(), + force_disable_gc); } else { - ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx]); + ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx], + force_disable_gc); } PADDLE_ENFORCE_LT(static_cast(bid), program.Size()); auto& block = program.Block(bid); @@ -409,8 +422,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, int64_t max_memory_size = GetEagerDeletionThreshold(); std::unique_ptr gc; - // skip while_op and while_grad_op temporarily - if (max_memory_size >= 0 && !keep_kids) { + // FIXME(zjl): recurrent_op is rather complex, we would + // disable gc forcely in recurrent_op + if (!ctx->force_disable_gc_ && max_memory_size >= 0) { ctx->ResetReferenceCount(); #ifdef PADDLE_WITH_CUDA if (platform::is_gpu_place(place_)) { @@ -428,6 +442,11 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, #ifdef PADDLE_WITH_CUDA } #endif + // If gc is enabled and block size > 1 + if (gc && ctx->prog_.Size() > 1) { + operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(ctx->block_id_, + ctx->ops_); + } } for (auto& op : ctx->ops_) { diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 5a040ac641588ad4d89d1f6e4c0d6c296eff38eb..65cb9e51ab2c9208b6bfbbed54f4136ffbd627ff 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -15,7 +15,9 @@ limitations under the License. */ #pragma once #include +#include #include +#include #include #include "paddle/fluid/framework/garbage_collector.h" #include "paddle/fluid/framework/op_info.h" @@ -30,7 +32,8 @@ namespace framework { struct ExecutorPrepareContext { ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id, const std::vector& skip_ref_cnt_vars = - std::vector()); + std::vector(), + bool force_disable_gc = false); ~ExecutorPrepareContext(); @@ -38,6 +41,7 @@ struct ExecutorPrepareContext { const framework::ProgramDesc& prog_; size_t block_id_; + bool force_disable_gc_; std::vector> ops_; std::unordered_map global_ref_cnts_; @@ -66,7 +70,10 @@ class Executor { * Scope */ void Run(const ProgramDesc& prog, Scope* scope, int block_id, - bool create_local_scope = true, bool create_vars = true); + bool create_local_scope = true, bool create_vars = true, + const std::vector& skip_ref_cnt_vars = + std::vector(), + bool force_disable_gc = false); // This API is very slow. void Run(const ProgramDesc& program, Scope* scope, @@ -79,12 +86,14 @@ class Executor { static std::unique_ptr Prepare( const ProgramDesc& program, int block_id, const std::vector& skip_ref_cnt_vars = - std::vector()); + std::vector(), + bool force_disable_gc = false); static std::vector> Prepare( const ProgramDesc& program, const std::vector& block_ids, const std::vector>& skip_ref_cnt_vars = - std::vector>()); + std::vector>(), + bool force_disable_gc = false); void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id); diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index ca6b0229e906c0f8bfbf9ee6781013cb4ef7bbce..faf7768a7bd198ad148a8fd50b1fa6c2e0d2886a 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -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) diff --git a/paddle/fluid/framework/ir/cpu_quantize_squash_pass.cc b/paddle/fluid/framework/ir/cpu_quantize_squash_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..de62a69de4f25912c5f56973de0dca5343bbe906 --- /dev/null +++ b/paddle/fluid/framework/ir/cpu_quantize_squash_pass.cc @@ -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 +#include +#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* 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* 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(dequant_op->Op()->GetAttr("Scale")); + float quant_scale = boost::get(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({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({dequant_in->Name()})); + desc.SetOutput("Output", std::vector({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 CPUQuantizeSquashPass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init("cpu_quantize_squash_pass", graph.get()); + + std::unordered_map 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); diff --git a/paddle/fluid/framework/ir/cpu_quantize_squash_pass.h b/paddle/fluid/framework/ir/cpu_quantize_squash_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..b823a2cef35b2f9994df9c9473246db3d69843e7 --- /dev/null +++ b/paddle/fluid/framework/ir/cpu_quantize_squash_pass.h @@ -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 +#include +#include +#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 ApplyImpl( + std::unique_ptr 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* nodes_keep_counter) const; + + /* + * Squash dequantize-quantize ops pairs into requantize or nothing + */ + void Squash(Graph* graph, + std::unordered_map* nodes_keep_counter) const; + + const std::string name_scope_{"squash"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/cpu_quantize_squash_pass_tester.cc b/paddle/fluid/framework/ir/cpu_quantize_squash_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..3a3eb53f79955b37f5f9af6a09b2f9c8e934aa3e --- /dev/null +++ b/paddle/fluid/framework/ir/cpu_quantize_squash_pass_tester.cc @@ -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 +#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& inputs, + const std::vector& 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( + {"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 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(); + tensor->mutable_data(place, proto::VarType::FP32, + ::paddle::memory::Allocator::kDefault, 1); +} + +void MainTest(const ProgramDesc& prog, int removed_nodes_num) { + std::unique_ptr 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); diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index 5e954fa9c419b249bb8a4be5a78c01da85b017b2..6a9340b870df324f7dea03181bdb2b097e13e705 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include +#include #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::Clone() { + auto cloned_graph = std::make_shared(this->program_); + cloned_graph->ReleaseNodes(); + cloned_graph->num_node_created_ = 0; + std::unordered_map 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; } diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h index cfd974e4bd679fdd06739f4c943bb197865020fb..fff015d4a6f0c631017458ceb039ae3f1deb0e2c 100644 --- a/paddle/fluid/framework/ir/graph.h +++ b/paddle/fluid/framework/ir/graph.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include #include +#include #include #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> &var_nodes); + // Create a new and duplicated graph. + // WARN: The method only clones the graph structure, not its attributes. + std::shared_ptr Clone(); + private: std::map> InitFromProgram( const ProgramDesc &program); diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index c0c34d186b00814fe6c6fd42beb78133233a1357..08354b526a04626152e8e4df9354d34222347cae 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -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 // ... diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h index c8be586f546dc604375401b13a801841efbf08d2..3db4bba10d6c722db6bace3ace407c94d5826387 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.h +++ b/paddle/fluid/framework/ir/graph_pattern_detector.h @@ -18,8 +18,11 @@ #include #endif +#include #include #include +#include +#include #include #include #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") {} diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h index 9eade9eaa8f00fe6e76063344f47968f4e97cf7f..72fb876d98dc84164398583baf22c49014af483a 100644 --- a/paddle/fluid/framework/ir/node.h +++ b/paddle/fluid/framework/ir/node.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include #include diff --git a/paddle/fluid/framework/ir/sync_batch_norm_pass.cc b/paddle/fluid/framework/ir/sync_batch_norm_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..b37003991505140b0d531a4ea2b481c6d4b09d75 --- /dev/null +++ b/paddle/fluid/framework/ir/sync_batch_norm_pass.cc @@ -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 +#include +#include + +namespace paddle { +namespace framework { +namespace ir { + +std::unique_ptr SyncBatchNormPass::ApplyImpl( + std::unique_ptr 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); diff --git a/paddle/fluid/framework/ir/sync_batch_norm_pass.h b/paddle/fluid/framework/ir/sync_batch_norm_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..51cce3dca69330071f7d12efef08e2006e8bd7ac --- /dev/null +++ b/paddle/fluid/framework/ir/sync_batch_norm_pass.h @@ -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 +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +class SyncBatchNormPass : public Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/sync_batch_norm_pass_tester.cc b/paddle/fluid/framework/ir/sync_batch_norm_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..9c94c1746a6590df5a43c099b9c4c3678ca6e393 --- /dev/null +++ b/paddle/fluid/framework/ir/sync_batch_norm_pass_tester.cc @@ -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 + +namespace paddle { +namespace framework { +namespace ir { + +void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, + const std::vector& inputs, + const std::vector& 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({"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({"a", "conv_w"}), + std::vector({"b"})); + SetOp(&prog, "batch_norm", "bn", + std::vector({"b", "bn_scale", "bn_bias", "mean", "var"}), + std::vector( + {"c", "mean", "var", "save_mean", "save_inv_var"})); + return prog; +} + +TEST(IsTestPass, basic) { + auto prog = BuildProgramDesc(); + + std::unique_ptr 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(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); diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index df1689764d21fcbb054a0bf32ef725541bdaefe3..44821aadf6d1f6777484dab5d25d01ff3b42596b 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -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>( 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( diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 55629636a816982c4debe4b5b7138558ac309eb5..822bf5c9ceaa31e1283fa3cf1dbe42a43894a5dd 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -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> kKernelPriority; diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 3e1d61813ca83ebdf9435036117e79abe501b24b..56f108cea2e5d7dadbea2e2cbec39dbe7f4ba094 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -14,8 +14,10 @@ limitations under the License. */ #include "paddle/fluid/framework/parallel_executor.h" #include +#include #include #include +#include #include #include "paddle/fluid/framework/ir/graph_helper.h" @@ -181,13 +183,14 @@ std::vector &ParallelExecutor::GetLocalScopes() { return member_->local_scopes_; } -ParallelExecutor::ParallelExecutor( - const std::vector &places, - const std::unordered_set &bcast_vars, - const std::string &loss_var_name, Scope *scope, - const std::vector &local_scopes, - const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy, - ir::Graph *graph) +ParallelExecutor::ParallelExecutor(const std::vector &places, + const std::vector &bcast_vars, + const std::string &loss_var_name, + Scope *scope, + const std::vector &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 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( + 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 &vars) const { + const std::vector &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(main_tensor.data()); } else { auto local_scope = member_->local_scopes_[i]; diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index ddf60b39466e72822142e1dad2cfe9a97b6cf6f2..d4658b9623fe8c23b6a8b2903e3c48d794ba1652 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -14,9 +14,11 @@ limitations under the License. */ #pragma once +#include #include #include #include +#include #include #include "paddle/fluid/framework/details/build_strategy.h" @@ -45,7 +47,7 @@ class ParallelExecutor { public: explicit ParallelExecutor(const std::vector &places, - const std::unordered_set &bcast_vars, + const std::vector &bcast_vars, const std::string &loss_var_name, Scope *scope, const std::vector &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 &vars) const; + // broadcast the parameters from the 0th device. + // trainer_id the trainer index in nccl distributed training. + void BCastParamsToDevices(const std::vector &vars, + int trainer_id = 0) const; bool EnableParallelGraphExecution(const ir::Graph &graph, const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy) const; diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index 89166bfd15f26e066d32a7191217a9b9a8977bda..a7f09df4917532e7261cee471c711897c8eb3447 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -18,6 +18,7 @@ #include #include #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace framework { @@ -137,16 +138,19 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, #ifdef PADDLE_WITH_CUDA else if (platform::is_gpu_place(src_place) && // NOLINT platform::is_cpu_place(dst_place)) { + platform::RecordEvent record_event("TensorCopy:GPU->CPU"); auto src_gpu_place = boost::get(src_place); auto dst_cpu_place = boost::get(dst_place); memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); } else if (platform::is_cpu_place(src_place) && platform::is_gpu_place(dst_place)) { + platform::RecordEvent record_event("TensorCopy:CPU->GPU"); auto src_cpu_place = boost::get(src_place); auto dst_gpu_place = boost::get(dst_place); memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr); } else if (platform::is_gpu_place(src_place) && platform::is_gpu_place(dst_place)) { + platform::RecordEvent record_event("TensorCopy:GPU->GPU"); if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) { VLOG(3) << "Skip copy the same data from " << src_place << " to " << dst_place; @@ -157,6 +161,7 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); } else if (platform::is_cuda_pinned_place(src_place) && platform::is_gpu_place(dst_place)) { + platform::RecordEvent record_event("TensorCopy:CUDAPinned->GPU"); auto src_pinned_place = boost::get(src_place); auto dst_gpu_place = boost::get(dst_place); memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size, diff --git a/paddle/fluid/imperative/layer.cc b/paddle/fluid/imperative/layer.cc index 012dfc1c7f66027bc5375794e0d70ed78e70e781..5530823b90f6580692456253b0eb9d0af4e3240b 100644 --- a/paddle/fluid/imperative/layer.cc +++ b/paddle/fluid/imperative/layer.cc @@ -159,10 +159,9 @@ class Autograd { for (auto it : candidate->pre_ops_) { for (OpBase* pre_op : it.second) { if (!pre_op) continue; - VLOG(5) << "op dep " << candidate->op_desc_->Type() << " trace id " + VLOG(5) << "op dep " << candidate->Type() << " trace id " << candidate->trace_id_ << " <---- " << it.first << " <---- " - << pre_op->op_desc_->Type() << " trace id " - << pre_op->trace_id_; + << pre_op->Type() << " trace id " << pre_op->trace_id_; if (visited.find(pre_op) == visited.end()) { visited.insert(pre_op); queue.push_back(pre_op); @@ -180,10 +179,12 @@ std::unique_ptr VarBase::NewVarBase(const platform::Place& dst_place, PADDLE_ENFORCE(var_->IsInitialized(), "Variable must be initialized when getting numpy tensor"); - std::unique_ptr new_var(new VarBase()); + // TODO(minqiyang): change this after move unique_name generator to CXX + const framework::LoDTensor& self_tensor = var_->Get(); + std::unique_ptr new_var(new VarBase( + "Itmp", self_tensor.type(), self_tensor.dims(), dst_place, true, false)); framework::LoDTensor* tensor = new_var->var_->GetMutable(); - tensor->Resize(var_->Get().dims()); tensor->set_lod(var_->Get().lod()); if (blocking) { @@ -199,52 +200,62 @@ std::unique_ptr VarBase::NewVarBase(const platform::Place& dst_place, } if (platform::is_gpu_place(dst_place)) { - VLOG(3) << "copy tensor " << var_desc_->Name() << " from gpu"; + VLOG(3) << "copy tensor " << Name() << " from gpu"; } return new_var; } framework::LoDTensor& VarBase::GradValue() { - VLOG(3) << "get var grad " << var_desc_->Name(); + VLOG(3) << "get var grad " << Name(); + PADDLE_ENFORCE_NOT_NULL(grads_, + "Could not get grad value from no grad variable"); return *(grads_->var_->GetMutable()); } std::map> OpBase::ApplyGrad() { if (grad_op_descs_.empty() && backward_id_ <= 0) { - VLOG(3) << "op with no grad: " << op_desc_->Type(); + VLOG(3) << "op with no grad: " << Type(); return {}; } - VLOG(3) << "apply op grad: " << op_desc_->Type(); - std::vector grad_outputs; + VLOG(3) << "apply op grad: " << Type(); + std::vector tmp_grad_outputs; if (backward_id_ > 0) { VLOG(3) << "py_layer_grad"; - grad_outputs.resize(1); - grad_outputs[0][framework::GradVarName(PyLayer::kFwdOut)] = + tmp_grad_outputs.resize(1); + tmp_grad_outputs[0][framework::GradVarName(PyLayer::kFwdOut)] = PyLayer::ApplyGrad( backward_id_, grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)]); } else { - grad_outputs.resize(grad_op_descs_.size()); - for (size_t k = 0; k < grad_op_descs_.size(); ++k) { + const size_t grad_op_count = grad_op_descs_.size(); + + tmp_grad_outputs.resize(grad_op_count); + for (size_t k = 0; k < grad_op_count; ++k) { framework::OpDesc* grad_op_desc = grad_op_descs_[k]; - VLOG(3) << "op grad " << grad_op_desc->Type(); - for (auto it : grad_output_vars_[k]) { - auto& outputs = grad_outputs[k][it.first]; + auto& grad_output_variable_map = grad_output_vars_[k]; + + VLOG(3) << "apply grad op " << grad_op_desc->Type(); + + // Allocate tmp grad output variable + for (auto it : grad_output_variable_map) { + auto& outputs = tmp_grad_outputs[k][it.first]; + outputs.reserve(it.second.size()); for (size_t i = 0; i < it.second.size(); ++i) { // Allocate a new variable Variable* tmp_var = new framework::Variable(); tmp_var->GetMutable(); - outputs.push_back(tmp_var); + outputs.emplace_back(tmp_var); } } - framework::RuntimeContext ctx(grad_input_vars_[k], grad_outputs[k]); + // Run grad op + framework::RuntimeContext ctx(grad_input_vars_[k], tmp_grad_outputs[k]); // No need to do compile time infer shape here. // grad_op_desc_->InferShape(*block_); - grad_op_desc->InferVarType(block_); + // grad_op_desc->InferVarType(block_); std::unique_ptr opbase = framework::OpRegistry::CreateOp(*grad_op_desc); @@ -260,9 +271,10 @@ std::map> OpBase::ApplyGrad() { } } + // Add tmp grad outputs to original grad vars for (size_t k = 0; k < grad_output_vars_.size(); ++k) { for (auto it : grad_output_vars_[k]) { - auto& outputs = grad_outputs[k][it.first]; + auto& outputs = tmp_grad_outputs[k][it.first]; auto& origin_outputs = it.second; PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size()); @@ -316,19 +328,14 @@ void PyLayer::RegisterFunc(int func_id, const py::object& py_func) { int PyLayer::NumFuncs() { return py_funcs_.size(); } -std::vector PyLayer::Apply(int func_id, - const std::vector& inputs) { +std::vector PyLayer::Apply(int func_id, + const std::vector& inputs) { std::vector invars; for (const VarBase* in : inputs) { invars.push_back(in->var_); } PADDLE_ENFORCE(py_funcs_.find(func_id) != py_funcs_.end()); - std::vector outvars = CallPythonFunc(py_funcs_[func_id], invars); - std::vector ret; - for (Variable* v : outvars) { - ret.push_back(new VarBase(v, new VarBase(true))); - } - return ret; + return CallPythonFunc(py_funcs_[func_id], invars); } std::vector PyLayer::ApplyGrad( diff --git a/paddle/fluid/imperative/layer.h b/paddle/fluid/imperative/layer.h index 7a9f33dc1e6cbc0c3ec1e649906fb0a8de047189..618a5b7a03295ce679dc6a88e0eac57069e78b8b 100644 --- a/paddle/fluid/imperative/layer.h +++ b/paddle/fluid/imperative/layer.h @@ -112,31 +112,53 @@ class OpBase; */ class VarBase { public: - VarBase() : VarBase(new framework::Variable(), new VarBase(true)) {} - - explicit VarBase(bool stop_gradient) - : VarBase(new framework::Variable(), - stop_gradient ? nullptr : new VarBase(true), stop_gradient) {} - - VarBase(framework::Variable* var, VarBase* grad) - : VarBase(var, grad, false) {} + // Internal interface, create VarBase from exist variable + VarBase(const std::string& name, framework::Variable* var, VarBase* grad, + bool stop_gradient) + : VarBase(name, var->Get().type(), + var->Get().dims(), + var->Get().place(), var, grad, + stop_gradient, false) {} + + // Python interface + VarBase(const std::string& name, const framework::proto::VarType::Type dtype, + const std::vector& shape, const platform::Place& place, + bool stop_gradient, bool persistable) + : VarBase(name, dtype, framework::make_ddim(shape), place, stop_gradient, + persistable) {} + + // Internal interface, create VarBase from with ddim + VarBase(const std::string& name, const framework::proto::VarType::Type dtype, + const framework::DDim& shape, const platform::Place& place, + bool stop_gradient, bool persistable) + : VarBase(name, dtype, shape, place, nullptr, nullptr, stop_gradient, + persistable) {} private: - VarBase(framework::Variable* var, VarBase* grad, bool stop_gradient) - : name_(), - var_desc_(nullptr), + VarBase(const std::string& name, framework::proto::VarType::Type dtype, + const framework::DDim& shape, const platform::Place& place, + framework::Variable* var, VarBase* grad, bool stop_gradient, + bool persistable) + : name_(name), + dtype_(dtype), + place_(place), var_(var), grads_(grad), - block_(nullptr), - persistable_(false), stop_gradient_(stop_gradient), + persistable_(persistable), pre_op_(nullptr), pre_op_out_name_(), - pre_op_out_idx_(-1) {} + pre_op_out_idx_(-1) { + if (!var_) { + var_ = new framework::Variable(); + auto tensor = var_->GetMutable(); + tensor->Resize(shape); + tensor->mutable_data(place_, dtype_); + } + } public: virtual ~VarBase() { - // TODO(minqiyang): remove var desc from block desc if (var_) { delete var_; var_ = nullptr; @@ -151,14 +173,30 @@ class VarBase { pre_op_out_idx_ = -1; } - inline OpBase* PreOp() const { return pre_op_; } - inline int PreOpOutIdx() const { return pre_op_out_idx_; } + inline void SetName(const std::string& name) { name_ = name; } + inline std::string Name() const { return name_; } + + inline std::vector Shape() const { + if (var_->IsInitialized()) { + return framework::vectorize(var_->Get().dims()); + } else { + return {}; + } + } + + inline framework::proto::VarType::Type DType() const { return dtype_; } inline void SetStopGradient(bool stop_gradient) { stop_gradient_ = stop_gradient; } inline bool IsStopGradient() const { return stop_gradient_; } + inline void SetPersistable(bool persistable) { persistable_ = persistable; } + inline bool IsPersistable() const { return persistable_; } + + inline OpBase* PreOp() const { return pre_op_; } + inline int PreOpOutIdx() const { return pre_op_out_idx_; } + void RunBackward(); inline void ResetPreOp(OpBase* op) { @@ -180,7 +218,7 @@ class VarBase { } void ClearGradient() { - VLOG(1) << "clear gradient of " << var_desc_->Name(); + VLOG(1) << "clear gradient of " << Name(); if (grads_ && grads_->var_ && grads_->var_->IsInitialized()) { auto grads_t = grads_->var_->GetMutable(); operators::math::set_constant( @@ -196,23 +234,20 @@ class VarBase { const bool blocking) const; inline std::string GradName() const { - PADDLE_ENFORCE( - var_desc_, - "Couldn't get gradient variable's name, please call backward() first"); - return string::Sprintf("%s@IGrad", var_desc_->Name()); + return string::Sprintf("%s@IGrad", Name()); } std::string name_; - framework::VarDesc* var_desc_; + framework::proto::VarType::Type dtype_; + platform::Place place_; framework::Variable* var_; VarBase* grads_; - framework::BlockDesc* block_; - bool persistable_; - private: bool stop_gradient_; + bool persistable_; + OpBase* pre_op_; std::string pre_op_out_name_; int pre_op_out_idx_; @@ -223,11 +258,11 @@ class VarBase { */ class PYBIND11_HIDDEN OpBase { public: - OpBase() - : op_desc_(nullptr), + OpBase(const std::string& type) + : type_(type), + trace_id_(-1), forward_id_(-1), backward_id_(-1), - trace_id_(-1), place_(platform::CPUPlace()), backward_hooks_() {} @@ -249,13 +284,34 @@ class PYBIND11_HIDDEN OpBase { std::map> ApplyGrad(); + inline std::string Type() const { return type_; } + inline std::string GradOpType(size_t index) const { + PADDLE_ENFORCE_NOT_NULL(grad_op_descs_[index]); + return grad_op_descs_[index]->Type(); + } + void RegisterBackwardHooks(const py::object& callable); void InvokeBackwardHooks(); - // One of `op_desc_` or `forward_id_` is set, not both. - // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_. - framework::OpDesc* op_desc_; + void TrackPreOp(const VarBase* inp_var, const std::string& inp_name) { + if (inp_var->PreOp() && !inp_var->IsStopGradient()) { + VLOG(3) << "add pre op " << inp_var->PreOp()->Type() << " in slot " + << inp_name; + pre_ops_[inp_name].push_back(inp_var->PreOp()); + pre_ops_out_idx_[inp_name].push_back(inp_var->PreOpOutIdx()); + } else { + VLOG(3) << "no pre op in slot " << inp_name + << " input var stop_gradient: " << inp_var->IsStopGradient(); + pre_ops_[inp_name].push_back(nullptr); + // pre_ops_out_idx_[inp_name].push_back(-1); + } + } + + std::string type_; + // One of `trace_id_` or `forward_id_` is set, not both. + // For pure python PyLayer, use `forward_id_`, otherwise, use trace_id_. + int trace_id_; int forward_id_; // When has backward, one of `grad_op_descs_` or `backward_id_` is set, @@ -263,7 +319,6 @@ class PYBIND11_HIDDEN OpBase { // Note: each fwd op corresponds to a vector of bwd ops. std::vector grad_op_descs_; int backward_id_; - int trace_id_; platform::Place place_; @@ -277,8 +332,6 @@ class PYBIND11_HIDDEN OpBase { // Outputs to a vector of bwd ops. std::vector grad_output_vars_; - framework::BlockDesc* block_; - std::vector backward_hooks_; }; @@ -303,8 +356,8 @@ class PyLayer { static int NumFuncs(); - static std::vector Apply(int func_id, - const std::vector& inputs); + static std::vector Apply( + int func_id, const std::vector& inputs); static std::vector ApplyGrad( int func_id, const std::vector& inputs); diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc index 0cb1676372fdd35a762e897d269550f2d1e1ac36..7ee92b4d8c46d8814400dbc02847d701005f3d5b 100644 --- a/paddle/fluid/imperative/tracer.cc +++ b/paddle/fluid/imperative/tracer.cc @@ -56,15 +56,19 @@ void CreateGradOp(const framework::OpDesc& op_desc, } } -void InitVar(framework::Variable* var, framework::Variable* grad_var, - platform::DeviceContext* dev_ctx) { +void InitGrad(VarBase* var, platform::DeviceContext* dev_ctx) { + PADDLE_ENFORCE_NOT_NULL(var, "Could not get valid var base"); PADDLE_ENFORCE_NOT_NULL(dev_ctx, "Could not get valid device from forward op"); - auto& var_t = var->Get(); - grad_var->GetMutable()->mutable_data( - var_t.dims(), dev_ctx->GetPlace()); - operators::math::set_constant( - *dev_ctx, grad_var->GetMutable(), 0.0); + + if (var->grads_ == nullptr) { + auto& var_t = var->var_->Get(); + var->grads_ = new VarBase(var->GradName(), framework::proto::VarType::FP32, + framework::vectorize(var_t.dims()), + dev_ctx->GetPlace(), true, false); + auto grad_t = var->grads_->var_->GetMutable(); + operators::math::set_constant(*dev_ctx, grad_t, 0.0); + } } platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs) { @@ -85,6 +89,62 @@ platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs) { return result; } +framework::VariableNameMap CreateInputVarNameMap( + const OpBase* op, const VarBasePtrMap& varbase_map) { + framework::VariableNameMap result; + + auto& info_map = framework::OpInfoMap::Instance(); + auto* op_info = info_map.GetNullable(op->Type()); + if (op_info == nullptr || op_info->proto_ == nullptr) { + return result; + } + + for (auto& in : op_info->Proto().inputs()) { + auto it = varbase_map.find(in.name()); + if (it == varbase_map.end()) { + PADDLE_ENFORCE(in.dispensable()); + result[in.name()] = {}; + } else { + auto var_vector = it->second; + std::vector args; + args.reserve(var_vector.size()); + for (VarBase* var_base : var_vector) { + args.emplace_back(var_base->Name()); + } + result[in.name()] = args; + } + } + return result; +} + +framework::VariableNameMap CreateOutputVarNameMap( + const OpBase* op, const VarBasePtrMap& varbase_map) { + framework::VariableNameMap result; + + auto& info_map = framework::OpInfoMap::Instance(); + auto* op_info = info_map.GetNullable(op->Type()); + if (op_info == nullptr || op_info->proto_ == nullptr) { + return result; + } + + for (auto& out : op_info->Proto().outputs()) { + auto it = varbase_map.find(out.name()); + if (it == varbase_map.end()) { + PADDLE_ENFORCE(out.dispensable()); + result[out.name()] = {}; + } else { + auto var_vector = it->second; + std::vector args; + args.reserve(var_vector.size()); + for (VarBase* var_base : var_vector) { + args.emplace_back(var_base->Name()); + } + result[out.name()] = args; + } + } + return result; +} + Tracer::Tracer(framework::BlockDesc* root_block) : root_block_(root_block) { if (!FLAGS_tracer_profile_fname.empty()) { std::call_once(gTracerProfileOnce, [] { @@ -101,7 +161,7 @@ Tracer::Tracer(framework::BlockDesc* root_block) : root_block_(root_block) { std::set Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, const VarBasePtrMap& outputs, - framework::BlockDesc* block, + framework::AttributeMap attrs_map, const platform::Place expected_place, const bool stop_gradient) { #ifdef WITH_GPERFTOOLS @@ -110,40 +170,27 @@ std::set Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, } #endif - std::map vars; - - framework::OpDesc* op_desc = op->op_desc_; - VLOG(3) << "tracer tracing " << op_desc->Type() << " trace id " - << op->trace_id_; - op_desc->InferShape(*block); - op_desc->InferVarType(block); - - std::unique_ptr op_base = - framework::OpRegistry::CreateOp(*op_desc); - framework::VariableValueMap invars_map; framework::VariableValueMap outvars_map; + // Construct input_vars_map and output_vars_map + std::map current_vars_map; op->input_vars_ = inputs; for (auto it : op->input_vars_) { auto& invars = invars_map[it.first]; invars.reserve(it.second.size()); for (VarBase* inp : it.second) { - PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr", - op->op_desc_->Type(), inp->var_desc_->Name()); + PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr", op->Type(), + inp->Name()); invars.emplace_back(inp->var_); - vars[inp->var_desc_->Name()] = inp; - if (inp->PreOp() && !inp->IsStopGradient()) { - op->pre_ops_[it.first].push_back(inp->PreOp()); - op->pre_ops_out_idx_[it.first].push_back(inp->PreOpOutIdx()); - VLOG(3) << "add pre op " << inp->PreOp()->op_desc_->Type(); - } else { - op->pre_ops_[it.first].push_back(nullptr); + op->TrackPreOp(inp, it.first); + if (!stop_gradient) { + current_vars_map[inp->Name()] = inp; } - VLOG(3) << "input vname " << inp->var_desc_->Name() << " " - << inp->var_->IsInitialized() << " stop_gradient " - << inp->IsStopGradient(); + VLOG(3) << "input var name: " << inp->Name() + << " inited: " << inp->var_->IsInitialized() + << " stop_grad: " << inp->IsStopGradient(); } } @@ -152,25 +199,38 @@ std::set Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, auto& outvars = outvars_map[it.first]; const std::vector& outputs = it.second; outvars.reserve(outputs.size()); - for (size_t i = 0; i < outputs.size(); ++i) { + for (size_t i = 0U; i < outputs.size(); ++i) { VarBase* out = outputs[i]; outvars.emplace_back(out->var_); - vars[out->var_desc_->Name()] = out; - - framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name()); - if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) { - out->var_->GetMutable(); - } else { - LOG(ERROR) << "tracer doesn't support yet"; - } out->TrackPreOp(op, it.first, i, stop_gradient); + if (!stop_gradient) { + current_vars_map[out->Name()] = out; + } - VLOG(3) << "output vname " << out->var_desc_->Name() << " " - << out->var_->IsInitialized(); + VLOG(3) << "input var name: " << out->Name() + << " inited: " << out->var_->IsInitialized() + << " stop_grad: " << out->IsStopGradient(); } } - VLOG(3) << "tracer running " << op_desc->Type(); + // Check attrs and create op + framework::VariableNameMap invars_name_map = + CreateInputVarNameMap(op, inputs); + framework::VariableNameMap outvars_name_map = + CreateOutputVarNameMap(op, outputs); + + auto& info = framework::OpInfoMap::Instance().Get(op->Type()); + if (info.Checker() != nullptr) { + info.Checker()->Check(&attrs_map); + } + + std::unique_ptr op_base = + framework::OpRegistry::CreateOp(op->Type(), invars_name_map, + outvars_name_map, attrs_map); + + // TODO(minqiyang): Support infer var type in imperative mode + // Run forward op + VLOG(3) << "tracer running " << op->Type(); framework::RuntimeContext ctx(invars_map, outvars_map); // TODO(panyx0718): Cache p. @@ -186,36 +246,44 @@ std::set Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, framework::ExecutionContext(prepared_op.op, scope, *prepared_op.dev_ctx, prepared_op.ctx, prepared_op.kernel_configs)); + // construct backward op std::set vars_saved_for_backward; - if (!stop_gradient) { + VLOG(5) << "start construct backward op"; + + // construct grad op descs + std::unique_ptr fwd_op_desc(new framework::OpDesc( + op->Type(), invars_name_map, outvars_name_map, attrs_map)); std::unique_ptr> grad_to_var( new std::unordered_map()); - CreateGradOp(*op_desc, {}, {block}, &op->grad_op_descs_, grad_to_var.get()); + // NOTE(minqiyang): We don't support control flow op in imperative now + // Add grad_block_ when we want to support it + CreateGradOp(*fwd_op_desc, {}, {}, &op->grad_op_descs_, grad_to_var.get()); - op->grad_input_vars_.resize(op->grad_op_descs_.size()); - op->grad_output_vars_.resize(op->grad_op_descs_.size()); + VLOG(5) << "create grad op desc: " << op->grad_op_descs_[0]->Type(); - for (size_t i = 0; i < op->grad_op_descs_.size(); ++i) { + const size_t grad_op_count = op->grad_op_descs_.size(); + + op->grad_input_vars_.resize(grad_op_count); + op->grad_output_vars_.resize(grad_op_count); + + for (size_t i = 0; i < grad_op_count; ++i) { framework::OpDesc* grad_op_desc = op->grad_op_descs_[i]; for (auto it : grad_op_desc->Inputs()) { auto& grad_in_vars = op->grad_input_vars_[i][it.first]; + grad_in_vars.reserve(it.second.size()); for (const std::string& grad_invar : it.second) { - block->FindRecursiveOrCreateVar(grad_invar); auto var_it = grad_to_var->find(grad_invar); if (var_it == grad_to_var->end()) { - auto fwd_var_it = vars.find(grad_invar); - PADDLE_ENFORCE(fwd_var_it != vars.end()); + auto fwd_var_it = current_vars_map.find(grad_invar); + PADDLE_ENFORCE(fwd_var_it != current_vars_map.end()); // Forward inputs or outputs. - grad_in_vars.push_back(fwd_var_it->second->var_); + grad_in_vars.emplace_back(fwd_var_it->second->var_); } else { - VarBase* var = vars[var_it->second]; - if (!var->grads_->var_->IsInitialized()) { - InitVar(var->var_, var->grads_->var_, - prepared_op.GetDeviceContext()); - } + VarBase* var = current_vars_map[var_it->second]; + InitGrad(var, prepared_op.GetDeviceContext()); // Douts. - grad_in_vars.push_back(var->grads_->var_); + grad_in_vars.emplace_back(var->grads_->var_); } vars_saved_for_backward.insert(it.first); @@ -225,48 +293,48 @@ std::set Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, for (auto it : grad_op_desc->Outputs()) { auto& grad_out_vars = op->grad_output_vars_[i][it.first]; for (const std::string& grad_outvar : it.second) { - block->FindRecursiveOrCreateVar(grad_outvar); auto var_it = grad_to_var->find(grad_outvar); PADDLE_ENFORCE(var_it != grad_to_var->end(), "Could not found the grad op output var, should this " "operator %s's stop gradient be True", - op_desc->Type()); - VarBase* var = vars[var_it->second]; - if (!var->grads_->var_->IsInitialized()) { - InitVar(var->var_, var->grads_->var_, - prepared_op.GetDeviceContext()); - } + op->Type()); + VarBase* var = current_vars_map[var_it->second]; + InitGrad(var, prepared_op.GetDeviceContext()); grad_out_vars.push_back(var->grads_->var_); } } } } - op->block_ = block; return vars_saved_for_backward; } std::vector Tracer::PyTrace(OpBase* op, const std::vector& inputs, bool stop_gradient) { - VLOG(3) << "py_trace"; + VLOG(3) << "py_trace " << op->Type(); + op->input_vars_[PyLayer::kFwdInp] = inputs; - op->output_vars_[PyLayer::kFwdOut] = PyLayer::Apply(op->forward_id_, inputs); + + std::vector ret_vars = + PyLayer::Apply(op->forward_id_, inputs); + for (VarBase* inp : inputs) { - if (inp->PreOp() && !inp->IsStopGradient()) { - op->pre_ops_[PyLayer::kFwdInp].push_back(inp->PreOp()); - op->pre_ops_out_idx_[PyLayer::kFwdInp].push_back(inp->PreOpOutIdx()); - } else { - op->pre_ops_[PyLayer::kFwdInp].push_back(nullptr); - } + op->TrackPreOp(inp, PyLayer::kFwdInp); } - auto& outputs = op->output_vars_[PyLayer::kFwdOut]; - for (size_t i = 0; i < outputs.size(); ++i) { - VarBase* out = outputs[i]; + std::vector& outputs = op->output_vars_[PyLayer::kFwdOut]; + outputs.reserve(ret_vars.size()); + for (size_t i = 0U; i != ret_vars.size(); ++i) { + framework::Variable* v = ret_vars[i]; + VarBase* out = new VarBase(string::Sprintf("%s_out_%d", op->Type(), i), v, + nullptr, stop_gradient); + outputs.emplace_back(out); out->TrackPreOp(op, PyLayer::kFwdOut, i, stop_gradient); } + if (!stop_gradient) { + VLOG(5) << "start construct backward op"; op->grad_input_vars_.resize(1); op->grad_output_vars_.resize(1); auto& grad_input_vars = @@ -281,23 +349,16 @@ std::vector Tracer::PyTrace(OpBase* op, grad_input_vars.push_back(out->var_); } + // TODO(minqiyang): Add GPU support for PyLayer, only support CPU now platform::CPUPlace place; for (VarBase* out : outputs) { + InitGrad(out, platform::DeviceContextPool::Instance().Get(place)); grad_input_vars.push_back(out->grads_->var_); - if (!grad_input_vars.back()->IsInitialized()) { - // TODO(minqiyang): Add GPU support for PyLayer, only support CPU now - InitVar(out->var_, grad_input_vars.back(), - platform::DeviceContextPool::Instance().Get(place)); - } } - for (const VarBase* inp : inputs) { + for (VarBase* inp : inputs) { + InitGrad(inp, platform::DeviceContextPool::Instance().Get(place)); grad_output_vars.push_back(inp->grads_->var_); - if (!grad_output_vars.back()->IsInitialized()) { - // TODO(minqiyang): Add GPU support for PyLayer, only support CPU now - InitVar(inp->var_, grad_output_vars.back(), - platform::DeviceContextPool::Instance().Get(place)); - } } } return outputs; diff --git a/paddle/fluid/imperative/tracer.h b/paddle/fluid/imperative/tracer.h index 8a0267c37f7c98a172fe0fa573955dc420952c0a..7b65d55e9eff1444d84a3fba284ecbb8b47d1733 100644 --- a/paddle/fluid/imperative/tracer.h +++ b/paddle/fluid/imperative/tracer.h @@ -17,6 +17,8 @@ #include #include #include +#include +#include #include #include "paddle/fluid/framework/op_desc.h" @@ -34,7 +36,8 @@ void CreateGradOp(const framework::OpDesc& op_desc, framework::OpDesc** grad_op_desc, std::unordered_map* grad_to_var); -void InitVar(framework::Variable* var, framework::Variable* grad_var); +void InitVar(const VarBase* var, framework::Variable* grad_var, + platform::DeviceContext* dev_ctx); platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs); @@ -46,7 +49,7 @@ class Tracer { std::set Trace(OpBase* op, const VarBasePtrMap& inputs, const VarBasePtrMap& outputs, - framework::BlockDesc* block, + framework::AttributeMap attrs_map, const platform::Place expected_place, const bool stop_gradient = false); diff --git a/paddle/fluid/inference/api/details/zero_copy_tensor.cc b/paddle/fluid/inference/api/details/zero_copy_tensor.cc index cf02901d963858d2a44b7c588a5c6a49358b0d3f..9a40cf4b60a64c3d0452a4367ccb7ac36de6b3b8 100644 --- a/paddle/fluid/inference/api/details/zero_copy_tensor.cc +++ b/paddle/fluid/inference/api/details/zero_copy_tensor.cc @@ -126,15 +126,20 @@ void ZeroCopyTensor::copy_to_cpu(T *data) { } template void ZeroCopyTensor::copy_from_cpu(const float *data); template void ZeroCopyTensor::copy_from_cpu(const int64_t *data); +template void ZeroCopyTensor::copy_from_cpu(const int32_t *data); template void ZeroCopyTensor::copy_to_cpu(float *data); template void ZeroCopyTensor::copy_to_cpu(int64_t *data); +template void ZeroCopyTensor::copy_to_cpu(int32_t *data); template float *ZeroCopyTensor::data(PaddlePlace *place, int *size) const; template int64_t *ZeroCopyTensor::data(PaddlePlace *place, int *size) const; +template int32_t *ZeroCopyTensor::data(PaddlePlace *place, + int *size) const; template float *ZeroCopyTensor::mutable_data(PaddlePlace place); template int64_t *ZeroCopyTensor::mutable_data(PaddlePlace place); +template int32_t *ZeroCopyTensor::mutable_data(PaddlePlace place); void *ZeroCopyTensor::FindTensor() const { PADDLE_ENFORCE(!name_.empty(), diff --git a/paddle/fluid/inference/api/helper.h b/paddle/fluid/inference/api/helper.h index 1ce3fe5af74424cd2d66940c739dd2c2eebef047..258a79fa4e884177490fab79778151ae52537aa0 100644 --- a/paddle/fluid/inference/api/helper.h +++ b/paddle/fluid/inference/api/helper.h @@ -139,9 +139,8 @@ static void TensorAssignData(PaddleTensor *tensor, } template -static int ZeroCopyTensorAssignData(ZeroCopyTensor *tensor, - const std::vector> &data) { - int size{0}; +static void ZeroCopyTensorAssignData(ZeroCopyTensor *tensor, + const std::vector> &data) { auto *ptr = tensor->mutable_data(PaddlePlace::kCPU); int c = 0; for (const auto &f : data) { @@ -149,7 +148,15 @@ static int ZeroCopyTensorAssignData(ZeroCopyTensor *tensor, ptr[c++] = v; } } - return size; +} + +template +static void ZeroCopyTensorAssignData(ZeroCopyTensor *tensor, + const PaddleBuf &data) { + auto *ptr = tensor->mutable_data(PaddlePlace::kCPU); + for (size_t i = 0; i < data.length() / sizeof(T); i++) { + ptr[i] = *(reinterpret_cast(data.data()) + i); + } } static bool CompareTensor(const PaddleTensor &a, const PaddleTensor &b) { diff --git a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc index 3f6c933f2bcc6ed5410cb95a48f5ee6869280fe4..5157bd280d0f3ee327d5cee7799477b5e6fd3f71 100644 --- a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc @@ -107,6 +107,9 @@ void SetConfig(AnalysisConfig *cfg) { cfg->DisableGpu(); cfg->SwitchSpecifyInputNames(); cfg->SwitchIrOptim(); + if (FLAGS_zero_copy) { + cfg->SwitchUseFeedFetchOps(false); + } } void SetInput(std::vector> *inputs) { @@ -131,7 +134,7 @@ TEST(Analyzer_Pyramid_DNN, profile) { TestPrediction(reinterpret_cast(&cfg), input_slots_all, &outputs, FLAGS_num_threads); - if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { + if (FLAGS_num_threads == 1 && !FLAGS_test_all_data && !FLAGS_zero_copy) { PADDLE_ENFORCE_EQ(outputs.size(), 1UL); size_t size = GetSize(outputs[0]); PADDLE_ENFORCE_GT(size, 0); @@ -166,6 +169,19 @@ TEST(Analyzer_Pyramid_DNN, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare result of AnalysisConfig and AnalysisConfig + ZeroCopy +TEST(Analyzer_Pyramid_DNN, compare_zero_copy) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + std::vector outputs_name; + outputs_name.emplace_back("cos_sim_2.tmp_0"); + CompareAnalysisAndZeroCopy(reinterpret_cast(&cfg), + input_slots_all, outputs_name); +} + // Compare Deterministic result TEST(Analyzer_Pyramid_DNN, compare_determine) { AnalysisConfig cfg; diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc index 36282b3efe5756da55b056c09e94aa352e3dcf8a..dcf4b38ce8a9230148738cfd0840ca96b0c7cf8c 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc @@ -207,6 +207,9 @@ void SetConfig(AnalysisConfig *cfg) { cfg->DisableGpu(); cfg->SwitchSpecifyInputNames(); cfg->SwitchIrOptim(); + if (FLAGS_zero_copy) { + cfg->SwitchUseFeedFetchOps(false); + } } void SetInput(std::vector> *inputs) { @@ -285,133 +288,17 @@ TEST(Analyzer_rnn1, multi_thread) { input_slots_all, &outputs, 2 /* multi_thread */); } -// Validate that the AnalysisPredictor + ZeroCopyTensor really works by testing -// on the complex RNN1 model. -TEST(Analyzer_rnn1, ZeroCopy) { - AnalysisConfig config; - SetConfig(&config); - config.SwitchUseFeedFetchOps(false); - - PaddlePlace place; - - auto predictor = CreatePaddlePredictor(config); - - config.SwitchUseFeedFetchOps(true); - auto native_predictor = - CreatePaddlePredictor(config.ToNativeConfig()); - - config.SwitchUseFeedFetchOps( - true); // the analysis predictor needs feed/fetch. - auto analysis_predictor = CreatePaddlePredictor(config); - -#define NEW_TENSOR(name__) \ - auto name__##_tensor = predictor->GetInputTensor(#name__); - NEW_TENSOR(data_lod_attention); - NEW_TENSOR(cell_init); - NEW_TENSOR(data); - NEW_TENSOR(week); - NEW_TENSOR(minute); - NEW_TENSOR(hidden_init); - - // Prepare data for AnalysisPredictor - DataRecord data(FLAGS_infer_data, FLAGS_batch_size); - PrepareZeroCopyInputs(data_lod_attention_tensor.get(), cell_init_tensor.get(), - data_tensor.get(), hidden_init_tensor.get(), - week_tensor.get(), minute_tensor.get(), &data, - FLAGS_batch_size); - - // Prepare data for NativePredictor - std::vector> native_inputs; - SetInput(&native_inputs); - std::vector native_outputs; - std::vector analysis_outputs; - - auto output_tensor = predictor->GetOutputTensor("final_output.tmp_1"); - // Run analysis predictor - - int num_ops; - auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); - ASSERT_TRUE(fuse_statis.count("fc_fuse")); - ASSERT_EQ(fuse_statis.at("fc_fuse"), 1); - ASSERT_EQ(fuse_statis.at("fc_nobias_lstm_fuse"), 2); // bi-directional LSTM - ASSERT_EQ(fuse_statis.at("seq_concat_fc_fuse"), 1); - ASSERT_EQ(num_ops, - 13); // After graph optimization, only 13 operators exists. - - Timer timer; - double total_time{0}; - for (int i = 0; i < FLAGS_repeat; i++) { - timer.tic(); - predictor->ZeroCopyRun(); - total_time += timer.toc(); - } - LOG(INFO) << "ZeroCopy output: " << DescribeZeroCopyTensor(*output_tensor); - - ASSERT_TRUE(native_predictor->Run(native_inputs.front(), &native_outputs)); - LOG(INFO) << "native output " << DescribeTensor(native_outputs.front()); - - int output_size{0}; // this is the number of elements not memory size - auto *zero_copy_data = output_tensor->data(&place, &output_size); - auto *native_data = static_cast(native_outputs.front().data.data()); - for (int i = 0; i < output_size; i++) { - EXPECT_NEAR(zero_copy_data[i], native_data[i], 1e-3); - } -} - -TEST(Analyzer_rnn1, ZeroCopyMultiThread) { - AnalysisConfig config; - SetConfig(&config); - config.SwitchUseFeedFetchOps(false); - -#define NEW_TENSOR(name__) \ - auto name__##_tensor = predictor->GetInputTensor(#name__); - - std::vector> predictors; - predictors.emplace_back(CreatePaddlePredictor(config)); - for (int tid = 1; tid < FLAGS_num_threads; tid++) { - predictors.emplace_back(predictors.front()->Clone()); - } - double total_time_of_threads{0}; - std::vector threads; - - for (int tid = 0; tid < FLAGS_num_threads; tid++) { - threads.emplace_back([&, tid] { - auto &predictor = predictors[tid]; - NEW_TENSOR(data_lod_attention); - NEW_TENSOR(cell_init); - NEW_TENSOR(data); - NEW_TENSOR(week); - NEW_TENSOR(minute); - NEW_TENSOR(hidden_init); - - // Prepare data for AnalysisPredictor - DataRecord data(FLAGS_infer_data, FLAGS_batch_size); - Timer timer; - double total_time{0}; - - for (int i = 0; i < FLAGS_repeat; i++) { - PrepareZeroCopyInputs(data_lod_attention_tensor.get(), - cell_init_tensor.get(), data_tensor.get(), - hidden_init_tensor.get(), week_tensor.get(), - minute_tensor.get(), &data, FLAGS_batch_size); - - timer.tic(); - predictor->ZeroCopyRun(); - total_time += timer.toc(); - } - - total_time_of_threads += total_time; - - LOG(INFO) << "thread time: " << total_time / FLAGS_repeat; - }); - } - - for (auto &t : threads) { - t.join(); - } +// Compare result of AnalysisConfig and AnalysisConfig + ZeroCopy +TEST(Analyzer_rnn1, compare_zero_copy) { + AnalysisConfig cfg; + SetConfig(&cfg); - LOG(INFO) << "average time: " - << total_time_of_threads / FLAGS_num_threads / FLAGS_repeat; + std::vector> input_slots_all; + SetInput(&input_slots_all); + std::vector outputs_name; + outputs_name.emplace_back("final_output.tmp_1"); + CompareAnalysisAndZeroCopy(reinterpret_cast(&cfg), + input_slots_all, outputs_name); } } // namespace inference diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc index cca2ab1ee148b568e714c24dded7cd72403f0e5f..19fa5528da4d11d2eb1a2f932f60a84c3f5468e7 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc @@ -144,6 +144,9 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) { cfg->SwitchSpecifyInputNames(); cfg->SwitchIrDebug(); cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads); + if (FLAGS_zero_copy) { + cfg->SwitchUseFeedFetchOps(false); + } if (use_mkldnn) { cfg->EnableMKLDNN(); } @@ -184,10 +187,10 @@ TEST(Analyzer_seq_pool1, compare_determine) { input_slots_all); } -void analysis_fuse_statis(bool use_zerocopy) { +// Check the fuse status +TEST(Analyzer_seq_pool1, fuse_statis) { AnalysisConfig cfg; SetConfig(&cfg); - cfg.SwitchUseFeedFetchOps(!use_zerocopy); int num_ops; auto predictor = CreatePaddlePredictor(cfg); auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); @@ -203,137 +206,17 @@ void analysis_fuse_statis(bool use_zerocopy) { EXPECT_EQ(num_ops, 171); } -// Check the fuse status -TEST(Analyzer_seq_pool1, fuse_statis) { analysis_fuse_statis(false); } - -void PrepareZeroCopyInputs( - const std::unique_ptr &predictor, - std::vector> *inputs) { - DataRecord data(FLAGS_infer_data, FLAGS_batch_size); - // only feed one batch - const auto &one_batch = data.NextBatch(); - inputs->clear(); - for (size_t i = 0; i < one_batch.size(); ++i) { - auto &slot = one_batch[i]; - auto tensor = predictor->GetInputTensor(slot.name + "_embed"); - tensor->Reshape(slot.shape); - tensor->SetLoD({slot.lod}); - ZeroCopyTensorAssignData(tensor.get(), slot.data); - inputs->emplace_back(std::move(tensor)); - } -} - -// return the output values -std::vector zerocopy_profile(int repeat_times) { - AnalysisConfig config; - SetConfig(&config); - config.SwitchUseFeedFetchOps(false); - auto predictor = CreatePaddlePredictor(config); - std::vector> inputs; - PrepareZeroCopyInputs(predictor, &inputs); - auto output_tensor = predictor->GetOutputTensor(out_var_name); - Timer timer; - LOG(INFO) << "Warm up run..."; - timer.tic(); - predictor->ZeroCopyRun(); - PrintTime(FLAGS_batch_size, 1, 1, 0, timer.toc(), 1); - if (FLAGS_profile) { - paddle::platform::ResetProfiler(); - } - LOG(INFO) << "Run " << repeat_times << " times..."; - timer.tic(); - for (int i = 0; i < repeat_times; i++) { - predictor->ZeroCopyRun(); - } - PrintTime(FLAGS_batch_size, repeat_times, 1, 0, timer.toc() / repeat_times, - 1); - - LOG(INFO) << "ZeroCopy output: " << DescribeZeroCopyTensor(*output_tensor); - PaddlePlace place; - int output_size{0}; - auto *pdata = output_tensor->data(&place, &output_size); - std::vector res(output_size); - for (int i = 0; i < output_size; ++i) { - res[i] = pdata[i]; - } - return res; -} - -TEST(Analyzer_seq_pool1, zerocopy_profile) { zerocopy_profile(FLAGS_repeat); } - -TEST(Analyzer_seq_pool1, zerocopy_profile_threads) { - AnalysisConfig config; - SetConfig(&config); - config.SwitchUseFeedFetchOps(false); - - std::vector> predictors; - predictors.emplace_back(CreatePaddlePredictor(config)); - for (int tid = 1; tid < FLAGS_num_threads; tid++) { - predictors.emplace_back(predictors.front()->Clone()); - } - double total_time_of_threads{0}; - std::vector threads; - - for (int tid = 0; tid < FLAGS_num_threads; tid++) { - threads.emplace_back([&, tid] { - auto &predictor = predictors[tid]; - std::vector> inputs; - PrepareZeroCopyInputs(predictor, &inputs); - auto output_tensor = predictor->GetOutputTensor(out_var_name); - Timer timer; - double total_time{0}; - - LOG(INFO) << "Warm up run..."; - timer.tic(); - predictor->ZeroCopyRun(); - PrintTime(FLAGS_batch_size, 1, FLAGS_num_threads, tid, timer.toc(), 1); - if (FLAGS_profile) { - paddle::platform::ResetProfiler(); - } - int repeat_times = FLAGS_repeat; - LOG(INFO) << "Run " << repeat_times << " times..."; - timer.tic(); - - for (int i = 0; i < repeat_times; i++) { - predictor->ZeroCopyRun(); - } - total_time += timer.toc(); - total_time_of_threads += total_time; - - LOG(INFO) << "thread time: " << total_time / repeat_times; - }); - } - - for (auto &t : threads) { - t.join(); - } - - LOG(INFO) << "average time: " - << total_time_of_threads / FLAGS_num_threads / FLAGS_repeat; -} - -TEST(Analyzer_seq_pool1, zerocopy_fuse_statis) { analysis_fuse_statis(true); } +// Compare result of AnalysisConfig and AnalysisConfig + ZeroCopy +TEST(Analyzer_seq_pool1, compare_zero_copy) { + AnalysisConfig cfg; + SetConfig(&cfg); -TEST(Analyzer_seq_pool1, zerocopy_compare_native) { - AnalysisConfig config; - SetConfig(&config); - config.SwitchUseFeedFetchOps(true); - auto predictor = CreatePaddlePredictor(config.ToNativeConfig()); - std::vector native_outputs; std::vector> input_slots_all; SetInput(&input_slots_all); - ASSERT_TRUE(predictor->Run(input_slots_all[0], &native_outputs)); - EXPECT_EQ(native_outputs.size(), 1UL); - - auto zerocopy_output = zerocopy_profile(1); - EXPECT_EQ(zerocopy_output.size() * sizeof(float), - native_outputs.front().data.length()); - auto *native_data = static_cast(native_outputs.front().data.data()); - for (size_t i = 0; i < zerocopy_output.size(); ++i) { - EXPECT_LT( - std::fabs((zerocopy_output[i] - native_data[i]) / zerocopy_output[i]), - 1e-3); - } + std::vector outputs_name; + outputs_name.emplace_back(out_var_name); + CompareAnalysisAndZeroCopy(reinterpret_cast(&cfg), + input_slots_all, outputs_name); } } // namespace analysis diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h index 41daff83c482c5f95d02afee9637d19d469ca507..a4881afe58a03902556ddb8a057c5f0579e4d1d2 100644 --- a/paddle/fluid/inference/tests/api/tester_helper.h +++ b/paddle/fluid/inference/tests/api/tester_helper.h @@ -50,6 +50,7 @@ DEFINE_bool(use_analysis, true, DEFINE_bool(record_benchmark, false, "Record benchmark after profiling the model"); DEFINE_double(accuracy, 1e-3, "Result Accuracy."); +DEFINE_bool(zero_copy, false, "Use ZeroCopy to speedup Feed/Fetch."); DECLARE_bool(profile); DECLARE_int32(paddle_num_threads); @@ -67,6 +68,7 @@ void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) { LOG(INFO) << analysis_config->ToNativeConfig(); } +// Compare result between two PaddleTensor void CompareResult(const std::vector &outputs, const std::vector &ref_outputs) { EXPECT_GT(outputs.size(), 0UL); @@ -108,6 +110,50 @@ void CompareResult(const std::vector &outputs, } } +// Compare result between a PaddleTensor and a ZeroCopyTensor +void CompareResult(const std::vector &outputs, + const std::vector &ref_outputs) { + EXPECT_GT(outputs.size(), 0UL); + EXPECT_EQ(outputs.size(), ref_outputs.size()); + for (size_t i = 0; i < outputs.size(); i++) { + auto &out = outputs[i]; + auto &ref_out = ref_outputs[i]; + size_t size = VecReduceToInt(out.shape); + EXPECT_GT(size, 0UL); + int ref_size = 0; // this is the number of elements not memory size + PaddlePlace place; + switch (out.dtype) { + case PaddleDType::INT64: { + int64_t *pdata = static_cast(out.data.data()); + int64_t *pdata_ref = ref_out.data(&place, &ref_size); + EXPECT_EQ(size, ref_size); + for (size_t j = 0; j < size; ++j) { + EXPECT_EQ(pdata_ref[j], pdata[j]); + } + break; + } + case PaddleDType::FLOAT32: { + float *pdata = static_cast(out.data.data()); + float *pdata_ref = ref_out.data(&place, &ref_size); + EXPECT_EQ(size, ref_size); + for (size_t j = 0; j < size; ++j) { + CHECK_LE(std::abs(pdata_ref[j] - pdata[j]), FLAGS_accuracy); + } + break; + } + case PaddleDType::INT32: { + int32_t *pdata = static_cast(out.data.data()); + int32_t *pdata_ref = ref_out.data(&place, &ref_size); + EXPECT_EQ(size, ref_size); + for (size_t j = 0; j < size; ++j) { + EXPECT_EQ(pdata_ref[j], pdata[j]); + } + break; + } + } + } +} + std::unique_ptr CreateTestPredictor( const PaddlePredictor::Config *config, bool use_analysis = true) { const auto *analysis_config = @@ -205,61 +251,106 @@ void GetInputPerBatch(const std::vector> &in, } } -void TestOneThreadPrediction( - const PaddlePredictor::Config *config, - const std::vector> &inputs, - std::vector *outputs, bool use_analysis = true) { - int batch_size = FLAGS_batch_size; - int num_times = FLAGS_repeat; - auto predictor = CreateTestPredictor(config, use_analysis); +void ConvertPaddleTensorToZeroCopyTensor( + PaddlePredictor *predictor, const std::vector &inputs) { + for (size_t i = 0; i < inputs.size(); i++) { + auto input = inputs[i]; + auto tensor = predictor->GetInputTensor(input.name); + tensor->Reshape(input.shape); + tensor->SetLoD({input.lod}); + if (input.dtype == PaddleDType::INT64) { + ZeroCopyTensorAssignData(tensor.get(), input.data); + } else if (input.dtype == PaddleDType::FLOAT32) { + ZeroCopyTensorAssignData(tensor.get(), input.data); + } else if (input.dtype == PaddleDType::INT32) { + ZeroCopyTensorAssignData(tensor.get(), input.data); + } else { + LOG(ERROR) << "unsupported feed type " << input.dtype; + } + } +} - // warmup run - LOG(INFO) << "Warm up run..."; - { - Timer warmup_timer; - warmup_timer.tic(); +void PredictionWarmUp(PaddlePredictor *predictor, + const std::vector> &inputs, + std::vector *outputs, int num_threads, + int tid) { + int batch_size = FLAGS_batch_size; + LOG(INFO) << "Running thread " << tid << ", warm up run..."; + if (FLAGS_zero_copy) { + ConvertPaddleTensorToZeroCopyTensor(predictor, inputs[0]); + } + Timer warmup_timer; + warmup_timer.tic(); + if (!FLAGS_zero_copy) { predictor->Run(inputs[0], outputs, batch_size); - PrintTime(batch_size, 1, 1, 0, warmup_timer.toc(), 1); - if (FLAGS_profile) { - paddle::platform::ResetProfiler(); - } + } else { + predictor->ZeroCopyRun(); + } + PrintTime(batch_size, 1, num_threads, tid, warmup_timer.toc(), 1); + if (FLAGS_profile) { + paddle::platform::ResetProfiler(); } +} - LOG(INFO) << "Run " << num_times << " times..."; - { - Timer run_timer; - run_timer.tic(); +void PredictionRun(PaddlePredictor *predictor, + const std::vector> &inputs, + std::vector *outputs, int num_threads, + int tid) { + int batch_size = FLAGS_batch_size; + int num_times = FLAGS_repeat; + LOG(INFO) << "Thread " << tid << " run " << num_times << " times..."; + Timer run_timer; + double elapsed_time = 0; #ifdef WITH_GPERFTOOLS - ProfilerStart("paddle_inference.prof"); + ProfilerStart("paddle_inference.prof"); #endif - for (int i = 0; i < num_times; i++) { - for (size_t j = 0; j < inputs.size(); j++) { - predictor->Run(inputs[j], outputs, batch_size); + if (!FLAGS_zero_copy) { + run_timer.tic(); + for (size_t i = 0; i < inputs.size(); i++) { + for (int j = 0; j < num_times; j++) { + predictor->Run(inputs[i], outputs, batch_size); + } + } + elapsed_time = run_timer.toc(); + } else { + for (size_t i = 0; i < inputs.size(); i++) { + ConvertPaddleTensorToZeroCopyTensor(predictor, inputs[i]); + run_timer.tic(); + for (int j = 0; j < num_times; j++) { + predictor->ZeroCopyRun(); } + elapsed_time += run_timer.toc(); } + } #ifdef WITH_GPERFTOOLS - ProfilerStop(); + ProfilerStop(); #endif - double latency = run_timer.toc() / (num_times > 1 ? num_times : 1); - PrintTime(batch_size, num_times, 1, 0, latency, inputs.size()); - if (FLAGS_record_benchmark) { - Benchmark benchmark; - benchmark.SetName(FLAGS_model_name); - benchmark.SetBatchSize(batch_size); - benchmark.SetLatency(latency); - benchmark.PersistToFile("benchmark_record.txt"); - } + PrintTime(batch_size, num_times, num_threads, tid, elapsed_time / num_times, + inputs.size()); + if (FLAGS_record_benchmark) { + Benchmark benchmark; + benchmark.SetName(FLAGS_model_name); + benchmark.SetBatchSize(batch_size); + benchmark.SetLatency(elapsed_time / num_times); + benchmark.PersistToFile("benchmark_record.txt"); } } +void TestOneThreadPrediction( + const PaddlePredictor::Config *config, + const std::vector> &inputs, + std::vector *outputs, bool use_analysis = true) { + auto predictor = CreateTestPredictor(config, use_analysis); + PredictionWarmUp(predictor.get(), inputs, outputs, 1, 0); + PredictionRun(predictor.get(), inputs, outputs, 1, 0); +} + void TestMultiThreadPrediction( const PaddlePredictor::Config *config, const std::vector> &inputs, std::vector *outputs, int num_threads, bool use_analysis = true) { - int batch_size = FLAGS_batch_size; - int num_times = FLAGS_repeat; std::vector threads; std::vector> predictors; predictors.emplace_back(CreateTestPredictor(config, use_analysis)); @@ -267,7 +358,6 @@ void TestMultiThreadPrediction( predictors.emplace_back(predictors.front()->Clone()); } - size_t total_time{0}; for (int tid = 0; tid < num_threads; ++tid) { threads.emplace_back([&, tid]() { // Each thread should have local inputs and outputs. @@ -280,34 +370,8 @@ void TestMultiThreadPrediction( ->SetMkldnnThreadID(static_cast(tid) + 1); } #endif - - // warmup run - LOG(INFO) << "Running thread " << tid << ", warm up run..."; - { - Timer warmup_timer; - warmup_timer.tic(); - predictor->Run(inputs[0], outputs, batch_size); - PrintTime(batch_size, 1, num_threads, tid, warmup_timer.toc(), 1); - if (FLAGS_profile) { - paddle::platform::ResetProfiler(); - } - } - - LOG(INFO) << "Thread " << tid << " run " << num_times << " times..."; - { - Timer timer; - timer.tic(); - for (int i = 0; i < num_times; i++) { - for (const auto &input : inputs) { - ASSERT_TRUE(predictor->Run(input, &outputs_tid)); - } - } - - auto time = timer.toc(); - total_time += time; - PrintTime(batch_size, num_times, num_threads, tid, time / num_times, - inputs.size()); - } + PredictionWarmUp(predictor.get(), inputs, outputs, num_threads, tid); + PredictionRun(predictor.get(), inputs, outputs, num_threads, tid); }); } for (int i = 0; i < num_threads; ++i) { @@ -367,6 +431,31 @@ void CompareNativeAndAnalysis( CompareResult(analysis_outputs, native_outputs); } +void CompareAnalysisAndZeroCopy( + PaddlePredictor::Config *config, + const std::vector> &inputs, + const std::vector &outputs_name) { + int batch_size = FLAGS_batch_size; + // analysis + std::vector analysis_outputs; + auto predictor = CreateTestPredictor(config, true); + predictor->Run(inputs[0], &analysis_outputs, batch_size); + // analysis + zero_copy + std::vector zerocopy_outputs; + reinterpret_cast(config)->SwitchUseFeedFetchOps(false); + predictor = CreateTestPredictor(config, true); + ConvertPaddleTensorToZeroCopyTensor(predictor.get(), inputs[0]); + predictor->ZeroCopyRun(); + for (size_t i = 0; i < outputs_name.size(); i++) { + ZeroCopyTensor zerocopy_output = + *predictor->GetOutputTensor(outputs_name[i]).get(); + zerocopy_outputs.emplace_back(zerocopy_output); + LOG(INFO) << "ZeroCopy output: " << DescribeZeroCopyTensor(zerocopy_output); + } + // compare + CompareResult(analysis_outputs, zerocopy_outputs); +} + template std::string LoDTensorSummary(const framework::LoDTensor &tensor) { std::stringstream ss; diff --git a/paddle/fluid/inference/tests/test.cmake b/paddle/fluid/inference/tests/test.cmake index 6c5fe043ffa3f3dcafe2dbbebd6244467f859abf..f551b322fe00892be79dd966235504bb4f54c718 100644 --- a/paddle/fluid/inference/tests/test.cmake +++ b/paddle/fluid/inference/tests/test.cmake @@ -30,19 +30,20 @@ function(inference_download_and_uncompress INSTALL_DIR URL FILENAME) ${EXTERNAL_PROJECT_NAME} ${EXTERNAL_PROJECT_LOG_ARGS} PREFIX ${INSTALL_DIR} - URL ${URL}/${FILENAME} + DOWNLOAD_COMMAND wget -q -O ${INSTALL_DIR}/${FILENAME} ${URL}/${FILENAME} && + ${CMAKE_COMMAND} -E tar xzf ${INSTALL_DIR}/${FILENAME} DOWNLOAD_DIR ${INSTALL_DIR} DOWNLOAD_NO_PROGRESS 1 CONFIGURE_COMMAND "" BUILD_COMMAND "" UPDATE_COMMAND "" - INSTALL_COMMAND ${CMAKE_COMMAND} -E copy_directory ${UNPACK_DIR} ${INSTALL_DIR} + INSTALL_COMMAND "" ) endfunction() set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec") -if (NOT EXISTS ${WORD2VEC_INSTALL_DIR}) - inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz") +if(NOT EXISTS ${WORD2VEC_INSTALL_DIR} AND NOT WIN32) + inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz") endif() set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model") diff --git a/paddle/fluid/memory/CMakeLists.txt b/paddle/fluid/memory/CMakeLists.txt index e7268077643c3988c59a52bf54873f1e8db4619b..7eb663ea280e65f3c10304aa47c9970df099b901 100644 --- a/paddle/fluid/memory/CMakeLists.txt +++ b/paddle/fluid/memory/CMakeLists.txt @@ -1,6 +1,6 @@ add_subdirectory(detail) add_subdirectory(allocation) -cc_library(malloc SRCS malloc.cc DEPS place enforce allocator_facade) +cc_library(malloc SRCS malloc.cc DEPS place enforce allocator_facade profiler) cc_library(memcpy SRCS memcpy.cc DEPS place) cc_library(memory diff --git a/paddle/fluid/memory/allocation/CMakeLists.txt b/paddle/fluid/memory/allocation/CMakeLists.txt index 4b7b9064dcde9b5209264257d51bbd976ba8eb85..7c44e18f8f39cfcdf749441ba7530e5227c44b5f 100644 --- a/paddle/fluid/memory/allocation/CMakeLists.txt +++ b/paddle/fluid/memory/allocation/CMakeLists.txt @@ -3,7 +3,7 @@ cc_library(cpu_allocator SRCS cpu_allocator.cc DEPS allocator) cc_library(best_fit_allocator SRCS best_fit_allocator.cc DEPS allocator) cc_library(locked_allocator SRCS locked_allocator.cc DEPS allocator) cc_library(buffered_allocator SRCS buffered_allocator.cc DEPS allocator) -cc_library(legacy_allocator SRCS legacy_allocator.cc DEPS allocator buddy_allocator) +cc_library(legacy_allocator SRCS legacy_allocator.cc DEPS allocator buddy_allocator profiler) cc_test(buffered_allocator_test SRCS buffered_allocator_test.cc DEPS best_fit_allocator locked_allocator buffered_allocator cpu_allocator) if (WITH_GPU) diff --git a/paddle/fluid/memory/allocation/legacy_allocator.cc b/paddle/fluid/memory/allocation/legacy_allocator.cc index 1936f9d4cd83c53cf7b322ab29a3e0d92e042abc..c233bf4edf5462dc48f6c3f4f22a517a03585b45 100644 --- a/paddle/fluid/memory/allocation/legacy_allocator.cc +++ b/paddle/fluid/memory/allocation/legacy_allocator.cc @@ -12,8 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/memory/allocation/legacy_allocator.h" - +#include #include #include #include @@ -23,9 +22,11 @@ #endif #include "glog/logging.h" +#include "paddle/fluid/memory/allocation/legacy_allocator.h" #include "paddle/fluid/memory/detail/buddy_allocator.h" #include "paddle/fluid/memory/detail/system_allocator.h" #include "paddle/fluid/platform/gpu_info.h" +#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/string/printf.h" #include "paddle/fluid/string/split.h" @@ -328,18 +329,22 @@ size_t Usage::operator()(const platform::CUDAPinnedPlace &cuda_pinned) const { } // namespace legacy namespace allocation { - LegacyMemMonitor GPUMemMonitor; Allocation *LegacyAllocator::AllocateImpl(size_t size, Allocator::Attr attr) { void *ptr = boost::apply_visitor(legacy::AllocVisitor(size), place_); - return new Allocation(ptr, size, place_); + auto *tmp_alloc = new Allocation(ptr, size, place_); + platform::MemEvenRecorder::Instance().PushMemRecord( + static_cast(tmp_alloc), place_, size); + return tmp_alloc; } void LegacyAllocator::Free(Allocation *allocation) { boost::apply_visitor( legacy::FreeVisitor(allocation->ptr(), allocation->size()), allocation->place()); + platform::MemEvenRecorder::Instance().PopMemRecord( + static_cast(allocation), place_); delete allocation; } diff --git a/paddle/fluid/memory/memcpy.cc b/paddle/fluid/memory/memcpy.cc index 2a6f70a01e303aa1b608248cbeb8dcfa24837a0c..1408163e4b5278ddcd65eb4f2900109d772a589a 100644 --- a/paddle/fluid/memory/memcpy.cc +++ b/paddle/fluid/memory/memcpy.cc @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/memory/memcpy.h" #include // for memcpy +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace memory { @@ -29,14 +30,23 @@ void Copy(platform::CPUPlace, void* dst, #ifdef PADDLE_WITH_CUDA static constexpr size_t kMaxGpuAsyncCopyBytes = 64 * 1024; // 64K +// NOTE(zcd): Do not use GpuMemcpySync as much as possible. +// because GpuMemcpySync issues the copying command to the default stream, +// which will make two commands from different streams cannot run concurrently. +// Reference: +// https://devblogs.nvidia.com/gpu-pro-tip-cuda-7-streams-simplify-concurrency/ + template <> void Copy( platform::CPUPlace dst_place, void* dst, platform::CUDAPlace src_place, const void* src, size_t num, cudaStream_t stream) { platform::SetDeviceId(src_place.device); + if (stream) { + platform::RecordEvent record_event("GpuMemcpyAsync:GPU->CPU"); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream); } else { + platform::RecordEvent record_event("GpuMemcpySync:GPU->CPU"); platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost); // FIXME(zjl): do we really need it? if (num <= kMaxGpuAsyncCopyBytes) { @@ -51,8 +61,10 @@ void Copy( const void* src, size_t num, cudaStream_t stream) { platform::SetDeviceId(dst_place.device); if (stream) { + platform::RecordEvent record_event("GpuMemcpyAsync:CPU->GPU"); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream); } else { + platform::RecordEvent record_event("GpuMemcpySync:CPU->GPU"); platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice); // FIXME(zjl): do we really need it? if (num <= kMaxGpuAsyncCopyBytes) { @@ -68,15 +80,19 @@ void Copy( if (dst_place == src_place) { platform::SetDeviceId(src_place.device); if (stream) { + platform::RecordEvent record_event("GpuMemcpyAsync(same_gpu):GPU->GPU"); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToDevice, stream); } else { + platform::RecordEvent record_event("GpuMemcpySync(same_gpu):GPU->GPU"); platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice); } } else { if (stream) { + platform::RecordEvent record_event("GpuMemcpyPeerAsync:GPU->GPU"); platform::GpuMemcpyPeerAsync(dst, dst_place.device, src, src_place.device, num, stream); } else { + platform::RecordEvent record_event("GpuMemcpyPeerSync:GPU->GPU"); platform::GpuMemcpyPeerSync(dst, dst_place.device, src, src_place.device, num); } @@ -111,8 +127,10 @@ void Copy( cudaStream_t stream) { platform::SetDeviceId(src_place.device); if (stream) { + platform::RecordEvent record_event("GpuMemcpyAsync:GPU->CUDAPinned"); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream); } else { + platform::RecordEvent record_event("GpuMemcpySync:GPU->CUDAPinned"); platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost); } } @@ -124,8 +142,10 @@ void Copy( cudaStream_t stream) { platform::SetDeviceId(dst_place.device); if (stream) { + platform::RecordEvent record_event("GpuMemcpyAsync:CUDAPinned->GPU"); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream); } else { + platform::RecordEvent record_event("GpuMemcpySync:CUDAPinned->GPU"); platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice); } } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index a3f2a69aef52b6f55aa09e6dee2c22c048626c0d..2f8e0b3a3038b40bd46b818d39588a91aaf2f0b4 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -44,10 +44,10 @@ if (WITH_DISTRIBUTE) SET(OP_PREFETCH_DEPS ${OP_PREFETCH_DEPS} parameter_prefetch) endif() -register_operators(EXCLUDES py_func_op warpctc_op conv_fusion_op DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS}) +register_operators(EXCLUDES py_func_op warpctc_op conv_fusion_op sync_batch_norm_op DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS}) -# warpctc_op needs cudnn 7 above if (WITH_GPU) + # warpctc_op needs cudnn 7 above if (${CUDNN_MAJOR_VERSION} VERSION_LESS 7) op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu.cc) else() @@ -58,6 +58,8 @@ if (WITH_GPU) op_library(conv_fusion_op) file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_fusion);\n") endif() + op_library(sync_batch_norm_op) + file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sync_batch_norm);\n") else() op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale) endif() diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 2feb8e4c4787440fd086c597fa2a7f97204e34ac..f79960317aa1bac7ae9f8d80e4886dde8fe8ebcb 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -13,7 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/activation_op.h" +#include #include +#include #include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h" #include "paddle/fluid/platform/port.h" #ifdef PADDLE_WITH_CUDA @@ -269,6 +271,48 @@ $$out = \\frac{x}{1 + \|x\|}$$ )DOC"; +class AcosOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "Input of acos operator"); + AddOutput("Out", "Output of acos operator"); + AddComment(R"DOC( +Arccosine Activation Operator. + +$$out = \cos^{-1}(x)$$ + +)DOC"); + } +}; + +class AsinOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "Input of asin operator"); + AddOutput("Out", "Output of asin operator"); + AddComment(R"DOC( +Arcsine Activation Operator. + +$$out = \sin^{-1}(x)$$ + +)DOC"); + } +}; + +class AtanOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "Input of atan operator"); + AddOutput("Out", "Output of atan operator"); + AddComment(R"DOC( +Arctanh Activation Operator. + +$$out = \tanh^{-1}(x)$$ + +)DOC"); + } +}; + class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -543,7 +587,10 @@ namespace ops = paddle::operators; __macro(SoftShrink, softshrink); \ __macro(Abs, abs); \ __macro(Cos, cos); \ + __macro(Acos, acos); \ __macro(Sin, sin); \ + __macro(Asin, asin); \ + __macro(Atan, atan); \ __macro(Round, round); \ __macro(Log, log); \ __macro(Square, square); \ diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h index 1f5ae7fb5cd2e1c14190602d2c35e6c3755cfd70..ff7e623f6f383ed2a8b8a40b3186d9c439ff1d86 100644 --- a/paddle/fluid/operators/activation_op.h +++ b/paddle/fluid/operators/activation_op.h @@ -39,9 +39,8 @@ namespace operators { Please refer to the layer_helper.py and get the details. */ static std::unordered_set InplaceOpSet = { - "sigmoid", "exp", "relu", "tanh", "sqrt", "ceil", - "floor", "reciprocal", "relu6", "soft_relu", "hard_sigmoid", -}; + "sigmoid", "exp", "relu", "tanh", "sqrt", "ceil", + "floor", "reciprocal", "relu6", "soft_relu", "hard_sigmoid"}; static bool IsInplace(const std::string& op) { bool inplace = InplaceOpSet.count(op); @@ -553,6 +552,101 @@ struct SinFunctor : public BaseActivationFunctor { } }; +template +struct Acos { + HOSTDEVICE T operator()(const T& val) const { return acos(val); } +}; + +template <> +struct Acos { + HOSTDEVICE platform::float16 operator()(const platform::float16& val) const { + return platform::float16(acos(static_cast(val))); + } +}; + +// Acos(x) = acos(x) +template +struct AcosFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.unaryExpr(Acos()); + } +}; + +// acos'(x) = -1/sqrt(1-x^2) +template +struct AcosGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = + -dout * static_cast(1) / (static_cast(1) - x.square()).sqrt(); + } +}; + +template +struct Asin { + HOSTDEVICE T operator()(const T& val) const { return asin(val); } +}; + +template <> +struct Asin { + HOSTDEVICE platform::float16 operator()(const platform::float16& val) const { + return platform::float16(asin(static_cast(val))); + } +}; + +// Asin(x) = asin(x) +template +struct AsinFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.unaryExpr(Asin()); + } +}; + +// asin'(x) = 1/sqrt(1-x^2) +template +struct AsinGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = + dout * static_cast(1) / (static_cast(1) - x.square()).sqrt(); + } +}; + +template +struct Atan { + HOSTDEVICE T operator()(const T& val) const { return atan(val); } +}; + +template <> +struct Atan { + HOSTDEVICE platform::float16 operator()(const platform::float16& val) const { + return platform::float16(atan(static_cast(val))); + } +}; + +// Atan(x) = atan(x) +template +struct AtanFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.unaryExpr(Atan()); + } +}; + +// atan'(x) = 1 / (1 + x^2) +template +struct AtanGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * static_cast(1) / (static_cast(1) + x.square()); + } +}; + // round(x) = [x] template struct RoundFunctor : public BaseActivationFunctor { @@ -1001,13 +1095,16 @@ struct SwishGradFunctor : public BaseActivationFunctor { __macro(relu, ReluFunctor, ReluGradFunctor); \ __macro(gelu, GeluFunctor, GeluGradFunctor); \ __macro(tanh, TanhFunctor, TanhGradFunctor); \ + __macro(atan, AtanFunctor, AtanGradFunctor); \ __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \ __macro(sqrt, SqrtFunctor, SqrtGradFunctor); \ __macro(abs, AbsFunctor, AbsGradFunctor); \ __macro(ceil, CeilFunctor, ZeroGradFunctor); \ __macro(floor, FloorFunctor, ZeroGradFunctor); \ __macro(cos, CosFunctor, CosGradFunctor); \ + __macro(acos, AcosFunctor, AcosGradFunctor); \ __macro(sin, SinFunctor, SinGradFunctor); \ + __macro(asin, AsinFunctor, AsinGradFunctor); \ __macro(round, RoundFunctor, ZeroGradFunctor); \ __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \ __macro(log, LogFunctor, LogGradFunctor); \ diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index feac4125381bd897dac89943af44850012e4761d..c0ad959309a7036639c4bc15621a2bd0296526f5 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -13,7 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/batch_norm_op.h" +#include #include +#include #include "paddle/fluid/framework/data_layout.h" #ifdef PADDLE_WITH_MKLDNN #include "paddle/fluid/platform/mkldnn_helper.h" @@ -22,147 +24,150 @@ limitations under the License. */ namespace paddle { namespace operators { -class BatchNormOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), ""); - PADDLE_ENFORCE(ctx->HasInput("Scale"), ""); - PADDLE_ENFORCE(ctx->HasInput("Bias"), ""); - PADDLE_ENFORCE(ctx->HasInput("Mean"), ""); - PADDLE_ENFORCE(ctx->HasInput("Variance"), ""); - PADDLE_ENFORCE(ctx->HasOutput("Y"), ""); - PADDLE_ENFORCE(ctx->HasOutput("MeanOut"), ""); - PADDLE_ENFORCE(ctx->HasOutput("VarianceOut"), ""); - PADDLE_ENFORCE(ctx->HasOutput("SavedMean"), ""); - PADDLE_ENFORCE(ctx->HasOutput("SavedVariance"), ""); - - // make sure Mean/MeanOut and Variance/VarianceOut share memory in Python - PADDLE_ENFORCE_EQ(ctx->Inputs("Mean")[0], ctx->Outputs("MeanOut")[0], - "Mean and MeanOut should share the same memory"); - PADDLE_ENFORCE_EQ(ctx->Inputs("Variance")[0], - ctx->Outputs("VarianceOut")[0], - "Variance and VarianceOut should share the same memory"); - - const auto x_dims = ctx->GetInputDim("X"); - const DataLayout data_layout = framework::StringToDataLayout( - ctx->Attrs().Get("data_layout")); - - PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, - "Input X must have 2 to 5 dimensions."); - - const int64_t C = - (data_layout == DataLayout::kNCHW ? x_dims[1] - : x_dims[x_dims.size() - 1]); - - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], C); - - ctx->SetOutputDim("Y", x_dims); - ctx->SetOutputDim("MeanOut", {C}); - ctx->SetOutputDim("VarianceOut", {C}); - ctx->SetOutputDim("SavedMean", {C}); - ctx->SetOutputDim("SavedVariance", {C}); - ctx->ShareLoD("X", "Y"); +void BatchNormOp::InferShape(framework::InferShapeContext *ctx) const { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Scale"), + "Input(Scale) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Bias"), + "Input(Bias) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Mean"), + "Input(Mean) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Variance"), + "Input(Variance) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), + "Output(Y) of ConvOp should not be null."); + bool is_test = ctx->Attrs().Get("is_test"); + if (!is_test) { + PADDLE_ENFORCE(ctx->HasOutput("MeanOut"), + "Output(MeanOut) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("VarianceOut"), + "Output(VarianceOut) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("SavedMean"), + "Output(SavedMean) of ConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("SavedVariance"), + "Output(SavedVariance) of ConvOp should not be null."); } - protected: - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext &ctx) const override { - auto input_data_type = ctx.Input("X")->type(); - // By default, the type of the scale, bias, mean, - // and var tensors should both be float. (For float or float16 input tensor) - // or double (For double input tensor). - auto bn_param_type = framework::proto::VarType::FP32; - if (input_data_type == framework::proto::VarType::FP64) { - bn_param_type = framework::proto::VarType::FP64; - } - PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Scale")->type(), - "Scale input should be of float type"); - PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Bias")->type(), - "Bias input should be of float type"); - PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Mean")->type(), - "Mean input should be of float type"); - PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Variance")->type(), - "Variance input should be of float type"); - - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::LibraryType library = framework::LibraryType::kPlain; - framework::DataLayout layout = framework::DataLayout::kAnyLayout; + // make sure Mean/MeanOut and Variance/VarianceOut share memory in Python + PADDLE_ENFORCE_EQ(ctx->Inputs("Mean")[0], ctx->Outputs("MeanOut")[0], + "Mean and MeanOut should share the same memory"); + PADDLE_ENFORCE_EQ(ctx->Inputs("Variance")[0], ctx->Outputs("VarianceOut")[0], + "Variance and VarianceOut should share the same memory"); + + const auto x_dims = ctx->GetInputDim("X"); + const DataLayout data_layout = framework::StringToDataLayout( + ctx->Attrs().Get("data_layout")); + + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "Input X must have 2 to 5 dimensions."); + + const int64_t C = + (data_layout == DataLayout::kNCHW ? x_dims[1] + : x_dims[x_dims.size() - 1]); + + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], C); + + ctx->SetOutputDim("Y", x_dims); + ctx->SetOutputDim("MeanOut", {C}); + ctx->SetOutputDim("VarianceOut", {C}); + ctx->SetOutputDim("SavedMean", {C}); + ctx->SetOutputDim("SavedVariance", {C}); + ctx->ShareLoD("X", "Y"); +} + +framework::OpKernelType BatchNormOp::GetExpectedKernelType( + const framework::ExecutionContext &ctx) const { + auto input_data_type = ctx.Input("X")->type(); + // By default, the type of the scale, bias, mean, + // and var tensors should both be float. (For float or float16 input tensor) + // or double (For double input tensor). + auto bn_param_type = framework::proto::VarType::FP32; + if (input_data_type == framework::proto::VarType::FP64) { + bn_param_type = framework::proto::VarType::FP64; + } + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Scale")->type(), + "Scale input should be of float type"); + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Bias")->type(), + "Bias input should be of float type"); + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Mean")->type(), + "Mean input should be of float type"); + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Variance")->type(), + "Variance input should be of float type"); + + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN - if (library == framework::LibraryType::kPlain && - platform::CanMKLDNNBeUsed(ctx)) { - library = framework::LibraryType::kMKLDNN; - layout = framework::DataLayout::kMKLDNN; - } -#endif - - return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, - library); + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } -}; +#endif -class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() override { - AddAttr("is_test", - "(bool, default false) Set to true for inference only, false " - "for training. Some layers may run faster when this is true.") - .SetDefault(false); - AddAttr("momentum", "").SetDefault(0.9); - AddAttr("epsilon", "") - .SetDefault(1e-5) - .AddCustomChecker([](const float &epsilon) { - PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f, - "'epsilon' should be between 0.0 and 0.001."); - }); - AddAttr("data_layout", "").SetDefault("NCHW"); - AddInput("X", "The input tensor"); - AddInput("Scale", - "Scale is a 1-dimensional tensor of size C " - "that is applied to the output"); - AddInput("Bias", - "Bias is a 1-dimensional tensor of size C " - "that is applied to the output"); - AddInput("Mean", - "The global mean (for training) or " - "estimated mean (for testing)"); - AddInput("Variance", - "The global variance (for training) " - "or estimated Variance (for testing)"); - AddOutput("Y", "result after normalization"); - AddOutput("MeanOut", - "Share memory with Mean. " - "Store the global mean when training"); - AddOutput("VarianceOut", - "Share memory with Variance. " - "Store the global Variance when training"); - AddOutput("SavedMean", - "Mean of the current mini batch, " - "will apply to output when training") - .AsIntermediate(); - AddOutput("SavedVariance", - "Variance of the current mini batch, " - "will apply to output when training") - .AsIntermediate(); - AddAttr("use_mkldnn", - "(bool, default false) Only used in mkldnn kernel") - .SetDefault(false); - AddAttr("fuse_with_relu", - "(bool, default false) Only used in mkldnn kernel") - .SetDefault(false); - AddAttr("use_global_stats", - "(bool, default false) Whether to use global mean and " - "variance. In inference or test mode, set use_global_stats " - "to true or is_test true. the behavior is equivalent. " - "In train mode, when setting use_global_stats True, the " - "global mean and variance are also used during train time, " - "the BN acts as scaling and shiffting.") - .SetDefault(false); - AddComment(R"DOC( + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, + library); +} + +void BatchNormOpMaker::Make() { + AddAttr("is_test", + "(bool, default false) Set to true for inference only, false " + "for training. Some layers may run faster when this is true.") + .SetDefault(false); + AddAttr("momentum", "").SetDefault(0.9); + AddAttr("epsilon", "") + .SetDefault(1e-5) + .AddCustomChecker([](const float &epsilon) { + PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f, + "'epsilon' should be between 0.0 and 0.001."); + }); + AddAttr("data_layout", "").SetDefault("NCHW"); + AddInput("X", "The input tensor"); + AddInput("Scale", + "Scale is a 1-dimensional tensor of size C " + "that is applied to the output"); + AddInput("Bias", + "Bias is a 1-dimensional tensor of size C " + "that is applied to the output"); + AddInput("Mean", + "The global mean (for training) or " + "estimated mean (for testing)"); + AddInput("Variance", + "The global variance (for training) " + "or estimated Variance (for testing)"); + AddOutput("Y", "result after normalization"); + AddOutput("MeanOut", + "Share memory with Mean. " + "Store the global mean when training"); + AddOutput("VarianceOut", + "Share memory with Variance. " + "Store the global Variance when training"); + AddOutput("SavedMean", + "Mean of the current mini batch, " + "will apply to output when training") + .AsIntermediate(); + AddOutput("SavedVariance", + "Variance of the current mini batch, " + "will apply to output when training") + .AsIntermediate(); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); + AddAttr("fuse_with_relu", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); + AddAttr("use_global_stats", + "(bool, default false) Whether to use global mean and " + "variance. In inference or test mode, set use_global_stats " + "to true or is_test true. the behavior is equivalent. " + "In train mode, when setting use_global_stats True, the " + "global mean and variance are also used during train time, " + "the BN acts as scaling and shiffting.") + .SetDefault(false); + AddComment(R"DOC( Batch Normalization. Batch Norm has been implemented as discussed in the paper: @@ -173,17 +178,7 @@ The required data format for this layer is one of the following: 2. NCHW `[batch, in_channels, in_height, in_width]` )DOC"); - } -}; - -class BatchNormOpInferVarType - : public framework::PassInDtypeAndVarTypeToOutput { - protected: - std::unordered_map GetInputOutputWithSameType() - const override { - return std::unordered_map{{"X", /*->*/ "Y"}}; - } -}; +} template class BatchNormKernel @@ -336,82 +331,75 @@ class BatchNormKernel } }; -class BatchNormGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext *ctx) const override { - // check input - PADDLE_ENFORCE(ctx->HasInput("X")); - PADDLE_ENFORCE(ctx->HasInput("Scale"), "Input(scale) should not be null."); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), - "Input(Y@GRAD) should not be null."); - PADDLE_ENFORCE(ctx->HasInput("SavedMean"), - "Input(SavedMean) should not be null."); - PADDLE_ENFORCE(ctx->HasInput("SavedVariance"), - "Input(SavedVariance) should not be null"); - - // check output - PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), ""); - if (ctx->HasOutput(framework::GradVarName("Scale"))) { - PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), - "Output(Scale@GRAD) and Output(Bias@GRAD) should not be " - "null at same time"); - } - const bool use_global_stats = ctx->Attrs().Get("use_global_stats"); - if (use_global_stats) { - PADDLE_ENFORCE(!ctx->Attrs().Get("use_mkldnn"), - "Using global stats during training is not supported " - "in gradient op kernel of batch_norm_mkldnn_op now."); - } +void BatchNormGradOp::InferShape(framework::InferShapeContext *ctx) const { + // check input + PADDLE_ENFORCE(ctx->HasInput("X")); + PADDLE_ENFORCE(ctx->HasInput("Scale"), "Input(scale) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("SavedMean"), + "Input(SavedMean) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("SavedVariance"), + "Input(SavedVariance) should not be null"); + + // check output + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), ""); + if (ctx->HasOutput(framework::GradVarName("Scale"))) { + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), + "Output(Scale@GRAD) and Output(Bias@GRAD) should not be " + "null at same time"); + } + const bool use_global_stats = ctx->Attrs().Get("use_global_stats"); + if (use_global_stats) { + PADDLE_ENFORCE(!ctx->Attrs().Get("use_mkldnn"), + "Using global stats during training is not supported " + "in gradient op kernel of batch_norm_mkldnn_op now."); + } - const auto x_dims = ctx->GetInputDim("X"); - const DataLayout data_layout = framework::StringToDataLayout( - ctx->Attrs().Get("data_layout")); - const int C = - (data_layout == DataLayout::kNCHW ? x_dims[1] - : x_dims[x_dims.size() - 1]); + const auto x_dims = ctx->GetInputDim("X"); + const DataLayout data_layout = framework::StringToDataLayout( + ctx->Attrs().Get("data_layout")); + const int C = (data_layout == DataLayout::kNCHW ? x_dims[1] + : x_dims[x_dims.size() - 1]); - ctx->SetOutputDim(framework::GradVarName("X"), x_dims); - if (ctx->HasOutput(framework::GradVarName("Scale"))) { - ctx->SetOutputDim(framework::GradVarName("Scale"), {C}); - ctx->SetOutputDim(framework::GradVarName("Bias"), {C}); - } + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + if (ctx->HasOutput(framework::GradVarName("Scale"))) { + ctx->SetOutputDim(framework::GradVarName("Scale"), {C}); + ctx->SetOutputDim(framework::GradVarName("Bias"), {C}); } +} - protected: - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext &ctx) const override { - const auto *var = ctx.InputVar(framework::GradVarName("Y")); - if (var == nullptr) { - PADDLE_THROW("can't find Y@GRAD"); - } - const Tensor *t = nullptr; - if (var->IsType()) { - t = &var->Get(); - } else if (var->IsType()) { - t = &var->Get(); - } - if (t == nullptr) { - PADDLE_THROW("can't find Y@GRAD"); - } +framework::OpKernelType BatchNormGradOp::GetExpectedKernelType( + const framework::ExecutionContext &ctx) const { + const auto *var = ctx.InputVar(framework::GradVarName("Y")); + if (var == nullptr) { + PADDLE_THROW("can't find Y@GRAD"); + } + const Tensor *t = nullptr; + if (var->IsType()) { + t = &var->Get(); + } else if (var->IsType()) { + t = &var->Get(); + } + if (t == nullptr) { + PADDLE_THROW("can't find Y@GRAD"); + } - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::LibraryType library = framework::LibraryType::kPlain; - framework::DataLayout layout = framework::DataLayout::kAnyLayout; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN - if (library == framework::LibraryType::kPlain && - platform::CanMKLDNNBeUsed(ctx)) { - library = framework::LibraryType::kMKLDNN; - layout = framework::DataLayout::kMKLDNN; - } + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } #endif - return framework::OpKernelType(ctx.Input("X")->type(), - ctx.GetPlace(), layout, library); - } -}; + return framework::OpKernelType(ctx.Input("X")->type(), ctx.GetPlace(), + layout, library); +} template class BatchNormGradKernel @@ -572,37 +560,31 @@ class BatchNormGradKernel } }; -class BatchNormGradMaker : public framework::SingleGradOpDescMaker { - public: - using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; - - protected: - std::unique_ptr Apply() const override { - auto *op = new framework::OpDesc(); - op->SetType("batch_norm_grad"); - op->SetInput("X", Input("X")); - op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); - - op->SetInput("Scale", Input("Scale")); - op->SetInput("Bias", Input("Bias")); - op->SetInput("SavedMean", Output("SavedMean")); - op->SetInput("SavedVariance", Output("SavedVariance")); - - // used when setting use_global_stats True during training - if (boost::get(GetAttr("use_global_stats"))) { - op->SetInput("Mean", Output("MeanOut")); - op->SetInput("Variance", Output("VarianceOut")); - } +std::unique_ptr BatchNormGradMaker::Apply() const { + auto *op = new framework::OpDesc(); + op->SetType(GradOpType()); + op->SetInput("X", Input("X")); + op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); + + op->SetInput("Scale", Input("Scale")); + op->SetInput("Bias", Input("Bias")); + op->SetInput("SavedMean", Output("SavedMean")); + op->SetInput("SavedVariance", Output("SavedVariance")); + + // used when setting use_global_stats True during training + if (boost::get(GetAttr("use_global_stats"))) { + op->SetInput("Mean", Output("MeanOut")); + op->SetInput("Variance", Output("VarianceOut")); + } - op->SetAttrMap(Attrs()); + op->SetAttrMap(Attrs()); - op->SetOutput(framework::GradVarName("X"), InputGrad("X")); - op->SetOutput(framework::GradVarName("Scale"), InputGrad("Scale")); - op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias")); + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetOutput(framework::GradVarName("Scale"), InputGrad("Scale")); + op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias")); - return std::unique_ptr(op); - } -}; + return std::unique_ptr(op); +} class BatchNormInplaceInToOut : public framework::InplaceInToOut { public: @@ -642,10 +624,10 @@ class BatchNormGradInplaceInToOut : public framework::InplaceInToOut { namespace ops = paddle::operators; REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, - ops::BatchNormOpInferVarType, ops::BatchNormGradMaker, - ops::BatchNormInplaceInToOut); -REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp, - ops::BatchNormGradInplaceInToOut); + ops::BatchNormOpInferVarType, ops::BatchNormGradMaker) +// ops::BatchNormInplaceInToOut); +REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp) +// ops::BatchNormGradInplaceInToOut); REGISTER_OP_CPU_KERNEL( batch_norm, ops::BatchNormKernel, diff --git a/paddle/fluid/operators/batch_norm_op.cu b/paddle/fluid/operators/batch_norm_op.cu index 1c45746a92ad057a97d9f65aa256df616fc37f3d..36d297ec5523b9e8a136c536165bdb4d3a380c25 100644 --- a/paddle/fluid/operators/batch_norm_op.cu +++ b/paddle/fluid/operators/batch_norm_op.cu @@ -33,26 +33,6 @@ using CudnnDataType = platform::CudnnDataType; template using BatchNormParamType = typename CudnnDataType::BatchNormParamType; -void ExtractNCWHD(const framework::DDim &dims, const DataLayout &data_layout, - int *N, int *C, int *H, int *W, int *D) { - *N = dims[0]; - if (dims.size() == 2) { - *C = dims[1]; - *H = 1; - *W = 1; - *D = 1; - } else { - *C = data_layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1]; - *H = data_layout == DataLayout::kNCHW ? dims[2] : dims[1]; - *W = dims.size() > 3 - ? (data_layout == DataLayout::kNCHW ? dims[3] : dims[2]) - : 1; - *D = dims.size() > 4 - ? (data_layout == DataLayout::kNCHW ? dims[4] : dims[3]) - : 1; - } -} - template class BatchNormKernel : public framework::OpKernel { @@ -196,22 +176,6 @@ class BatchNormKernel } }; -template -static __global__ void KeBNBackwardData(const T *dy, - const BatchNormParamType *scale, - const BatchNormParamType *variance, - const double epsilon, const int C, - const int HxW, const int num, T *dx) { - int gid = blockIdx.x * blockDim.x + threadIdx.x; - int stride = blockDim.x * gridDim.x; - for (int i = gid; i < num; i += stride) { - const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C; - BatchNormParamType inv_var = 1.0 / sqrt(variance[c] + epsilon); - dx[i] = static_cast(static_cast>(dy[i]) * - scale[c] * inv_var); - } -} - template static __global__ void KeBNBackwardScaleBias( const T *dy, const T *x, const BatchNormParamType *mean, @@ -248,6 +212,22 @@ static __global__ void KeBNBackwardScaleBias( } } +template +static __global__ void KeBNBackwardData(const T *dy, + const BatchNormParamType *scale, + const BatchNormParamType *variance, + const double epsilon, const int C, + const int HxW, const int num, T *dx) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = gid; i < num; i += stride) { + const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C; + BatchNormParamType inv_var = 1.0 / sqrt(variance[c] + epsilon); + dx[i] = static_cast(static_cast>(dy[i]) * + scale[c] * inv_var); + } +} + template class BatchNormGradKernel : public framework::OpKernel { @@ -383,7 +363,7 @@ class BatchNormGradKernel KeBNBackwardScaleBias<<< grid2, block, 0, dev_ctx.stream()>>>( d_y->data(), x->data(), running_mean_data, running_var_data, - epsilon, C, H * W, num, d_scale->data>(), + epsilon, N, C, H * W * D, d_scale->data>(), d_bias->data>()); } } else { @@ -394,10 +374,10 @@ class BatchNormGradKernel running_var_data, epsilon, C, H * W, num, d_x->data()); } if (d_scale && d_bias) { - KeBNBackwardScaleBias<<< + KeBNBackwardScaleBias<<< grid2, block, 0, dev_ctx.stream()>>>( d_y->data(), x->data(), running_mean_data, running_var_data, - epsilon, C, H * W, num, d_scale->data>(), + epsilon, N, C, H * W * D, d_scale->data>(), d_bias->data>()); } } diff --git a/paddle/fluid/operators/batch_norm_op.h b/paddle/fluid/operators/batch_norm_op.h index 5e3d630d6889e445c5e84fa836d2d81bb7266779..6e89d73eb236ee7844c7de3c273e0b0f275a3e33 100644 --- a/paddle/fluid/operators/batch_norm_op.h +++ b/paddle/fluid/operators/batch_norm_op.h @@ -13,6 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -35,17 +38,84 @@ template using ConstEigenVectorArrayMap = Eigen::Map>; +class BatchNormOp : 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; +}; + +class BatchNormGradOp : 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; +}; + +class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override; +}; + +class BatchNormGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override; + + virtual std::string GradOpType() const { + return this->ForwardOpType() + "_grad"; + } +}; + +class BatchNormOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Y"}}; + } +}; + template class BatchNormKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override; + void Compute(const framework::ExecutionContext &ctx) const override; }; template class BatchNormGradKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override; + void Compute(const framework::ExecutionContext &ctx) const override; }; +inline void ExtractNCWHD(const framework::DDim &dims, + const DataLayout &data_layout, int *N, int *C, int *H, + int *W, int *D) { + *N = dims[0]; + if (dims.size() == 2) { + *C = dims[1]; + *H = 1; + *W = 1; + *D = 1; + } else { + *C = data_layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1]; + *H = data_layout == DataLayout::kNCHW ? dims[2] : dims[1]; + *W = dims.size() > 3 + ? (data_layout == DataLayout::kNCHW ? dims[3] : dims[2]) + : 1; + *D = dims.size() > 4 + ? (data_layout == DataLayout::kNCHW ? dims[4] : dims[3]) + : 1; + } +} + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/controlflow/CMakeLists.txt b/paddle/fluid/operators/controlflow/CMakeLists.txt index b614e9b03502634a29333f331e25201a0f77ba38..7aa1c44eaafe53034b19ee52c59cc94d3a1269da 100644 --- a/paddle/fluid/operators/controlflow/CMakeLists.txt +++ b/paddle/fluid/operators/controlflow/CMakeLists.txt @@ -1,4 +1,5 @@ include(operators) register_operators(DEPS naive_executor) +cc_library(while_op_helper SRCS while_op_helper.cc DEPS operator) file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n") diff --git a/paddle/fluid/operators/controlflow/while_op.cc b/paddle/fluid/operators/controlflow/while_op.cc index 0360cf5273591946570cac47e2578e43f498b550..8352ba4f2b846af58d2d041ebf5201ee15f8481c 100644 --- a/paddle/fluid/operators/controlflow/while_op.cc +++ b/paddle/fluid/operators/controlflow/while_op.cc @@ -18,6 +18,7 @@ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/controlflow/while_op_helper.h" #include "paddle/fluid/operators/detail/safe_ref.h" namespace paddle { @@ -26,14 +27,6 @@ namespace operators { using StepScopeVar = std::vector; using LoDTensor = framework::LoDTensor; -static constexpr char kStepBlock[] = "sub_block"; -static constexpr char kCondition[] = "Condition"; -static constexpr char kStepScopes[] = "StepScopes"; -static constexpr char kX[] = "X"; -static constexpr char kXGRAD[] = "X@GRAD"; -static constexpr char kOutputs[] = "Out"; -static constexpr char kSkipEagerDeletionVars[] = "skip_eager_deletion_vars"; - namespace { // NOLINT static std::string GetSkipEagerDeletionVarsDebugString( const std::vector &vars) { diff --git a/paddle/fluid/operators/controlflow/while_op_helper.cc b/paddle/fluid/operators/controlflow/while_op_helper.cc new file mode 100644 index 0000000000000000000000000000000000000000..2cbd94a061b5b369d67b6e0995d6b8fd45801828 --- /dev/null +++ b/paddle/fluid/operators/controlflow/while_op_helper.cc @@ -0,0 +1,291 @@ +// 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/operators/controlflow/while_op_helper.h" +#include +#include +#include +#include "paddle/fluid/framework/program_desc.h" + +namespace paddle { +namespace operators { + +// OpVariant is a wrapper class of OpDesc and OperatorBase +// So that API would be the same. +class OpVariant { + struct InputsVisitor + : public boost::static_visitor { + template + const framework::VariableNameMap *operator()(const OpType *op) const { + return &(op->Inputs()); + } + }; + + struct OutputsVisitor + : public boost::static_visitor { + template + const framework::VariableNameMap *operator()(const OpType *op) const { + return &(op->Outputs()); + } + }; + + struct AttributeMapVisitor + : public boost::static_visitor { + const framework::AttributeMap *operator()( + const framework::OpDesc *op) const { + return &(op->GetAttrMap()); + } + + const framework::AttributeMap *operator()( + const framework::OperatorBase *op) const { + return &(op->Attrs()); + } + }; + + struct RawPointerVisitor : public boost::static_visitor { + template + const void *operator()(const OpType *op) const { + return op; + } + }; + + public: + OpVariant(const framework::OperatorBase *op) : op_(op) {} // NOLINT + + OpVariant(const framework::OpDesc *op) : op_(op) {} // NOLINT + + const framework::VariableNameMap &Inputs() const { + return *boost::apply_visitor(InputsVisitor(), op_); + } + + const framework::VariableNameMap &Outputs() const { + return *boost::apply_visitor(OutputsVisitor(), op_); + } + + const framework::AttributeMap &Attrs() const { + return *boost::apply_visitor(AttributeMapVisitor(), op_); + } + + template + const AttrType &Attr(const std::string &name) const { + auto &attrs = Attrs(); + auto it = attrs.find(name); + PADDLE_ENFORCE(it != attrs.end(), "Cannot find attribute %s", name); + return boost::get(it->second); + } + + bool operator==(const OpVariant &other) const { + return RawPointer() == other.RawPointer(); + } + + const void *RawPointer() const { + return boost::apply_visitor(RawPointerVisitor(), op_); + } + + int which() const { return static_cast(op_.which()); } + + struct Hasher { + size_t operator()(const OpVariant &op) const { + return reinterpret_cast(op.RawPointer()); + } + }; + + private: + const boost::variant + op_; +}; + +static std::string GetDebugString(const std::vector &names) { + if (names.empty()) return ""; + std::string ret = names[0]; + for (size_t i = 1; i < names.size(); ++i) { + ret += (" " + names[i]); + } + return ret; +} + +// Set skip variables of while_op and while_grad_op +// These variables should be skipped when eager deletion enables. +// It is because: +// 1. while_grad_op needs some variables defined in while_op. +// 2. while_grad_op needs variables from the previous time step. +static void SetSkipVars(const OpVariant &op, std::vector attr) { + auto &attrs = const_cast(op.Attrs()); + VLOG(2) << "Prepare to skip " << attr.size() + << " var(s): " << GetDebugString(attr); + attrs[kSkipEagerDeletionVars] = std::move(attr); +} + +// Check whether the forward while_op and while_grad_op match +// The program may have many while_ops. +static bool IsMatchedWhileOpAndWhileGradOp(const OpVariant &fwd_op, + const OpVariant &grad_op) { + return fwd_op.Inputs().at(kX) == grad_op.Inputs().at(kX) && + fwd_op.Outputs().at(kOutputs) == grad_op.Inputs().at(kOutputs); +} + +// Test whether the variable is skippable in forward while_op +// The variable is skippable in while_op when the variable used in while_grad +// is not from grad_block. +static bool IsSkippableVar(const std::string &name, + framework::BlockDesc *grad_block) { + return name != framework::kEmptyVarName && !grad_block->HasVar(name); +} + +static void ModifyWhileOpAndWhileGradOpAttr(const OpVariant &fwd_op, + const OpVariant &bwd_op) { + auto *grad_block = bwd_op.Attr(kStepBlock); + + // Find all skippable variables in forward while_op + std::unordered_set forward_skip_vars; + for (auto *op_desc : grad_block->AllOps()) { + for (auto &in_arg_name : op_desc->InputArgumentNames()) { + if (IsSkippableVar(in_arg_name, grad_block)) { + forward_skip_vars.insert(in_arg_name); + } + } + + for (auto &out_arg_name : op_desc->OutputArgumentNames()) { + if (IsSkippableVar(out_arg_name, grad_block)) { + forward_skip_vars.insert(out_arg_name); + } + } + } + + SetSkipVars(fwd_op, std::vector(forward_skip_vars.begin(), + forward_skip_vars.end())); + + // Find all skippable variables in while_grad_op + // The skipped variables are those which would be used across time steps. + auto &fwd_input = fwd_op.Inputs().at(kX); + auto &in_grads = bwd_op.Outputs().at(framework::GradVarName(kX)); + PADDLE_ENFORCE_EQ( + fwd_input.size(), in_grads.size(), + "Backward input gradient number does not match forward input number."); + + std::unordered_set backward_skip_vars; + for (size_t i = 0; i < in_grads.size(); ++i) { + if (in_grads[i] == framework::kEmptyVarName) { + continue; + } + backward_skip_vars.insert(in_grads[i]); + backward_skip_vars.insert(framework::GradVarName(fwd_input[i])); + } + + SetSkipVars(bwd_op, std::vector(backward_skip_vars.begin(), + backward_skip_vars.end())); +} + +// Find all while_ops and while_grad_ops in the graph or program +// The while_grad_op and while_op may located in different blocks +// So we should traverse all blocks in the program and find them out. +static void FindAllWhileAndWhileGradOp(std::vector *while_ops, + std::vector *while_grad_ops) { + PADDLE_ENFORCE_GE(while_ops->size(), while_grad_ops->size()); + + if (while_ops->empty()) return; + + const auto *program = + while_ops->front().Attr(kStepBlock)->Program(); + for (size_t i = 1; i < program->Size(); ++i) { + auto &block = program->Block(i); + for (size_t j = 0; j < block.OpSize(); ++j) { + auto *op = block.Op(j); + if (op->Type() == "while") { + while_ops->emplace_back(op); + } else if (op->Type() == "while_grad") { + while_grad_ops->emplace_back(op); + } + } + } + + PADDLE_ENFORCE_GE(while_ops->size(), while_grad_ops->size(), + "There are extra while_grad ops in the graph or program"); +} + +static void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOpImpl( + std::vector *while_ops, std::vector *while_grad_ops) { + FindAllWhileAndWhileGradOp(while_ops, while_grad_ops); + + VLOG(2) << "Found while op num: " << while_ops->size() + << ", while grad op num: " << while_grad_ops->size(); + + if (while_grad_ops->empty()) { + return; + } + + std::unordered_set while_op_set( + while_ops->begin(), while_ops->end()); + + for (auto &bwd_op : *while_grad_ops) { + const OpVariant *matched_fwd_op = nullptr; + for (auto &fwd_op : while_op_set) { + if (IsMatchedWhileOpAndWhileGradOp(fwd_op, bwd_op)) { + PADDLE_ENFORCE(matched_fwd_op == nullptr, + "Found multiple matched while ops"); + matched_fwd_op = &fwd_op; + } + } + PADDLE_ENFORCE_NOT_NULL(matched_fwd_op, + "Cannot find matched forward while op."); + ModifyWhileOpAndWhileGradOpAttr(*matched_fwd_op, bwd_op); + while_op_set.erase(*matched_fwd_op); + } +} + +void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp( + int block_id, + const std::vector> &all_ops) { + // If block_id is not 0, returns + // This is because all while_ops and while_grad_ops in the whole program + // would be processed when block_id is 0 (i.e. when Executor::Run() or + // ParallelExecutor constructs). + + // What's more, all while_ops and while_grad_ops must be processed when + // block_id is zero. If not, while_op may run first and erase variables + // used in while_grad_op, and in this moment, while_grad_ops may be not + // constructed yet. + if (block_id != 0) return; + + std::vector fwd_ops, bwd_ops; + for (auto &op : all_ops) { + if (op->Type() == "while") { + fwd_ops.emplace_back(op.get()); + } else if (op->Type() == "while_grad") { + bwd_ops.emplace_back(op.get()); + } + } + PrepareSafeEagerDeletionOnWhileOpAndWhileGradOpImpl(&fwd_ops, &bwd_ops); +} + +void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp( + const std::vector &while_ops, + const std::vector &while_grad_ops) { + std::vector fwd_ops, bwd_ops; + fwd_ops.reserve(while_ops.size()); + for (auto *op : while_ops) { + fwd_ops.emplace_back(op); + } + + bwd_ops.reserve(while_grad_ops.size()); + for (auto *op : while_grad_ops) { + bwd_ops.emplace_back(op); + } + + PrepareSafeEagerDeletionOnWhileOpAndWhileGradOpImpl(&fwd_ops, &bwd_ops); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/controlflow/while_op_helper.h b/paddle/fluid/operators/controlflow/while_op_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..456ba8642b9bd32a1236d112cc8b387ae6a279d3 --- /dev/null +++ b/paddle/fluid/operators/controlflow/while_op_helper.h @@ -0,0 +1,43 @@ +// 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 +#include +#include +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/variant.h" + +namespace paddle { +namespace operators { + +static constexpr char kStepBlock[] = "sub_block"; +static constexpr char kCondition[] = "Condition"; +static constexpr char kStepScopes[] = "StepScopes"; +static constexpr char kX[] = "X"; +static constexpr char kXGRAD[] = "X@GRAD"; +static constexpr char kOutputs[] = "Out"; +static constexpr char kSkipEagerDeletionVars[] = "skip_eager_deletion_vars"; + +void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp( + int block_id, + const std::vector> &all_ops); + +void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp( + const std::vector &while_ops, + const std::vector &while_grad_ops); + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/crf_decoding_op.h b/paddle/fluid/operators/crf_decoding_op.h index 72774a878d98b431da05cf870139752421b2df8d..d6b54038ec5648c72d606a6c7b9c8356cb74521b 100644 --- a/paddle/fluid/operators/crf_decoding_op.h +++ b/paddle/fluid/operators/crf_decoding_op.h @@ -82,8 +82,9 @@ class CRFDecodingOpKernel : public framework::OpKernel { Tensor track; int* track_value = track.mutable_data(emission_dims, platform::CPUPlace()); - auto ker = jit::Get, - platform::CPUPlace>(tag_num); + auto ker = + jit::KernelFuncs, platform::CPUPlace>::Cache() + .At(tag_num); ker(static_cast(seq_len), x, w, alpha_value, track_value, tag_num); T max_score = -std::numeric_limits::max(); int max_i = 0; diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index 3adc7baebddd06ced74afea1e77017beb57582e8..a617b9fb1d948340d25853252be79fdd08fe0438 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -13,18 +13,21 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" +#include #include +#include namespace paddle { namespace operators { -class CrossEntropyOp : public framework::OperatorWithKernel { +class CrossEntropyOpBase : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null."); auto x_dims = ctx->GetInputDim("X"); @@ -43,7 +46,8 @@ class CrossEntropyOp : public framework::OperatorWithKernel { "Input(X) and Input(Label) shall have the same shape " "except the last dimension."); } - if (ctx->Attrs().Get("soft_label")) { + + if (IsSoftLabel(ctx)) { if (check) { PADDLE_ENFORCE_EQ(x_dims[rank - 1], label_dims[rank - 1], "If Attr(soft_label) == true, the last dimension of " @@ -69,21 +73,24 @@ class CrossEntropyOp : public framework::OperatorWithKernel { return framework::OpKernelType(ctx.Input("X")->type(), ctx.device_context()); } + + virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const { + return ctx->Attrs().Get("soft_label"); + } }; -class CrossEntropyGradientOp : public framework::OperatorWithKernel { +class CrossEntropyGradientOpBase : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + void InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), "Input(Y@GRAD) shoudl be not null."); PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), "Output(X@GRAD) should be not null."); - auto x_dims = ctx->GetInputDim("X"); + auto x_dims = GetXDim(ctx); auto label_dims = ctx->GetInputDim("Label"); auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y")); int rank = x_dims.size(); @@ -108,9 +115,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { "The Input(X) and Input(Y@Grad) should have the same " "shape except the last dimension."); } - PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1, - "The last dimension of Input(Y@Grad) should be 1."); - if (ctx->Attrs().Get("soft_label")) { + if (IsSoftLabel(ctx)) { if (check) { PADDLE_ENFORCE_EQ( x_dims[rank - 1], label_dims[rank - 1], @@ -123,7 +128,10 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { "Input(Label) should be 1."); } ctx->SetOutputDim(framework::GradVarName("X"), x_dims); - ctx->ShareLoD("X", framework::GradVarName("X")); + PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1, + "The last dimension of Input(Y@Grad) should be 1."); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + ctx->ShareLoD(VarNameWithXLoD(), framework::GradVarName("X")); } protected: @@ -131,8 +139,28 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { // is determined by its input "X". framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType(ctx.Input("X")->type(), - ctx.device_context()); + return framework::OpKernelType( + ctx.Input(framework::GradVarName("Y"))->type(), + ctx.device_context()); + } + + virtual framework::DDim GetXDim(framework::InferShapeContext* ctx) const { + return ctx->GetInputDim("X"); + } + + virtual const char* VarNameWithXLoD() const { return "X"; } + + virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const { + return ctx->Attrs().Get("soft_label"); + } +}; + +class CrossEntropyOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Y"}}; } }; @@ -200,22 +228,132 @@ or not. But the output only shares the LoD information with input X. } }; -class CrossEntropyOpInferVarType - : public framework::PassInDtypeAndVarTypeToOutput { +class CrossEntropyGradientOp : public CrossEntropyGradientOpBase { + public: + using CrossEntropyGradientOpBase::CrossEntropyGradientOpBase; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + CrossEntropyGradientOpBase::InferShape(ctx); + } +}; + +class CrossEntropyOp2 : public CrossEntropyOpBase { + public: + using CrossEntropyOpBase::CrossEntropyOpBase; + + void InferShape(framework::InferShapeContext* ctx) const override { + CrossEntropyOpBase::InferShape(ctx); + + PADDLE_ENFORCE(ctx->HasOutput("XShape"), + "Output(XShape) should be not null."); + + PADDLE_ENFORCE(ctx->HasOutput("MatchX"), + "Output(MatchX) should be not null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_dims_vec = framework::vectorize(x_dims); + x_dims_vec.push_back(0); + ctx->SetOutputDim("XShape", framework::make_ddim(x_dims_vec)); + x_dims[x_dims.size() - 1] = 1; + ctx->SetOutputDim("MatchX", x_dims); + ctx->ShareLoD("X", /*->*/ "XShape"); + } + protected: - std::unordered_map GetInputOutputWithSameType() - const override { - return std::unordered_map{{"X", /*->*/ "Y"}}; + bool IsSoftLabel(framework::InferShapeContext* ctx) const override { + return false; + } +}; + +class CrossEntropyGradientOp2 : public CrossEntropyGradientOpBase { + public: + using CrossEntropyGradientOpBase::CrossEntropyGradientOpBase; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("MatchX"), "Input(MatchX) must exist"); + CrossEntropyGradientOpBase::InferShape(ctx); + } + + protected: + virtual framework::DDim GetXDim(framework::InferShapeContext* ctx) const { + auto x_shape = ctx->GetInputDim("XShape"); + return framework::DDim(x_shape.Get(), x_shape.size() - 1); + } + + virtual const char* VarNameWithXLoD() const { return "XShape"; } + + virtual bool IsSoftLabel(framework::InferShapeContext* ctx) const { + return false; } }; + +class CrossEntropyOpMaker2 : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor, default Tensor), a tensor whose last dimension " + "size is equal to the number of classes. This input is a " + "probability computed by the previous operator, which is almost " + "always the result of a softmax operator."); + AddInput( + "Label", + "(Tensor), the tensor which represents the ground truth. It has the " + "same shape with 'X' except the last dimension. One hot Tensor."); + AddOutput("Y", + "(Tensor, default Tensor), a tensor whose shape is same " + "with 'X' except that the last dimension size is 1. It " + "represents the cross entropy loss."); + AddOutput("XShape", "Temporaily variable to save shape and LoD of X."); + AddOutput("MatchX", + "X value that matches label, used for gradient computation."); + AddAttr("ignore_index", + "(int, default -100), Specifies a target value that is" + "ignored and does not contribute to the input gradient." + "Only valid if soft_label is set to False") + .SetDefault(-100); + AddComment(R"DOC( +Hard-label CrossEntropy Operator. + +The input 'X' and 'Label' will first be logically flattened to 2-D matrixs. +The matrix's second dimension(row length) is as same as the original last +dimension, and the first dimension(column length) is the product of all other +original dimensions. Then the softmax computation will take palce on each raw +of flattened matrixs. + +Only support hard label. + +Both the input X and Label can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD information with input X. + +)DOC"); + } +}; + +class CrossEntropyGradOpDescMaker2 : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + std::unique_ptr op(new framework::OpDesc()); + op->SetType("cross_entropy_grad2"); + op->SetInput("Label", Input("Label")); + op->SetInput("MatchX", Output("MatchX")); + op->SetInput("XShape", Output("XShape")); + op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetAttrMap(Attrs()); + return op; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; using CPUCtx = paddle::platform::CPUDeviceContext; -REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker, - ops::CrossEntropyOpInferVarType, +REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOpBase, + ops::CrossEntropyOpMaker, ops::CrossEntropyOpInferVarType, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp); REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, @@ -223,3 +361,14 @@ REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, REGISTER_OP_CPU_KERNEL(cross_entropy_grad, ops::CrossEntropyGradientOpKernel, ops::CrossEntropyGradientOpKernel); + +REGISTER_OPERATOR(cross_entropy2, ops::CrossEntropyOp2, + ops::CrossEntropyOpMaker2, ops::CrossEntropyOpInferVarType, + ops::CrossEntropyGradOpDescMaker2); +REGISTER_OPERATOR(cross_entropy_grad2, ops::CrossEntropyGradientOp2); +REGISTER_OP_CPU_KERNEL(cross_entropy2, + ops::CrossEntropyOpKernel2, + ops::CrossEntropyOpKernel2); +REGISTER_OP_CPU_KERNEL(cross_entropy_grad2, + ops::CrossEntropyGradientOpKernel2, + ops::CrossEntropyGradientOpKernel2); diff --git a/paddle/fluid/operators/cross_entropy_op.cu b/paddle/fluid/operators/cross_entropy_op.cu index fcd34383a85f6984a8f27ce0625364f8fd5e31d6..243e7f52c1e3c4c210e91f708ae5d6de97e4afbc 100644 --- a/paddle/fluid/operators/cross_entropy_op.cu +++ b/paddle/fluid/operators/cross_entropy_op.cu @@ -27,3 +27,13 @@ REGISTER_OP_CUDA_KERNEL( cross_entropy_grad, ops::CrossEntropyGradientOpKernel, ops::CrossEntropyGradientOpKernel, ops::CrossEntropyGradientOpKernel); + +REGISTER_OP_CUDA_KERNEL(cross_entropy2, + ops::CrossEntropyOpKernel2, + ops::CrossEntropyOpKernel2, + ops::CrossEntropyOpKernel2); + +REGISTER_OP_CUDA_KERNEL( + cross_entropy_grad2, ops::CrossEntropyGradientOpKernel2, + ops::CrossEntropyGradientOpKernel2, + ops::CrossEntropyGradientOpKernel2); diff --git a/paddle/fluid/operators/cross_entropy_op.h b/paddle/fluid/operators/cross_entropy_op.h index f123e11542d85c904a81fe2a87f59ab52511cc15..7eb663773ed072760c47a2914377b5306ceeb7af 100644 --- a/paddle/fluid/operators/cross_entropy_op.h +++ b/paddle/fluid/operators/cross_entropy_op.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math.h" #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/for_range.h" @@ -137,5 +138,124 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel { } }; +template +struct HardLabelCrossEntropyForwardFunctor { + HardLabelCrossEntropyForwardFunctor(const T* x, T* y, T* match_x, + const int64_t* label, + int64_t ignore_index, + int64_t feature_size) + : x_(x), + y_(y), + match_x_(match_x), + label_(label), + ignore_index_(ignore_index), + feature_size_(feature_size) {} + + HOSTDEVICE void operator()(int64_t idx) const { + auto label = label_[idx]; + if (label != ignore_index_) { + auto match_x = x_[idx * feature_size_ + label]; + y_[idx] = -math::TolerableValue()(real_log(match_x)); + match_x_[idx] = match_x; + } else { + y_[idx] = 0; + match_x_[idx] = 0; // any value is ok + } + } + + const T* x_; + T* y_; + T* match_x_; + const int64_t* label_; + int64_t ignore_index_; + int64_t feature_size_; +}; + +template +struct HardLabelCrossEntropyBackwardFunctor { + HardLabelCrossEntropyBackwardFunctor(T* dx, const T* dy, const T* match_x, + const int64_t* label, + int64_t ignore_index, + int64_t feature_size) + : dx_(dx), + dy_(dy), + match_x_(match_x), + label_(label), + ignore_index_(ignore_index), + feature_size_(feature_size) {} + + HOSTDEVICE void operator()(int64_t idx) const { + auto row_idx = idx / feature_size_; + auto col_idx = idx % feature_size_; + auto label = label_[row_idx]; + if (label == col_idx && label != ignore_index_) { + dx_[idx] = -dy_[row_idx] / match_x_[row_idx]; + } else { + dx_[idx] = 0; + } + } + + T* dx_; + const T* dy_; + const T* match_x_; + const int64_t* label_; + int64_t ignore_index_; + int64_t feature_size_; +}; + +template +class CrossEntropyOpKernel2 : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* label = ctx.Input("Label"); + auto* y = ctx.Output("Y"); + auto* match_x = ctx.Output("MatchX"); + + auto& x_dims = x->dims(); + auto feature_size = x_dims[x_dims.size() - 1]; + auto batch_size = framework::product(x->dims()) / feature_size; + + auto* p_x = x->data(); + auto* p_label = label->data(); + auto* p_y = y->mutable_data(ctx.GetPlace()); + auto* p_match_x = match_x->mutable_data(ctx.GetPlace()); + + auto ignore_index = ctx.Attr("ignore_index"); + + platform::ForRange for_range( + ctx.template device_context(), batch_size); + for_range(HardLabelCrossEntropyForwardFunctor( + p_x, p_y, p_match_x, p_label, ignore_index, feature_size)); + } +}; + +template +class CrossEntropyGradientOpKernel2 : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Input(framework::GradVarName("Y")); + auto* match_x = ctx.Input("MatchX"); + auto* label = ctx.Input("Label"); + + auto* p_dx = dx->mutable_data(ctx.GetPlace()); + auto* p_dy = dy->data(); + auto* p_match_x = match_x->data(); + auto* p_label = label->data(); + + int64_t ignore_index = ctx.Attr("ignore_index"); + int rank = dx->dims().size(); + int64_t feature_size = dx->dims()[rank - 1]; + int64_t batch_size = framework::product(dx->dims()) / feature_size; + + platform::ForRange for_range( + ctx.template device_context(), + batch_size * feature_size); + for_range(HardLabelCrossEntropyBackwardFunctor( + p_dx, p_dy, p_match_x, p_label, ignore_index, feature_size)); + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detection/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h index 6d406f8196f9964c85bb94541fa7a7a23857539b..d4c7e8cf7723bf83d3cd8bf36b9ae6c5f1c35b10 100644 --- a/paddle/fluid/operators/detection/box_coder_op.h +++ b/paddle/fluid/operators/detection/box_coder_op.h @@ -20,7 +20,7 @@ namespace operators { enum class BoxCodeType { kEncodeCenterSize = 0, kDecodeCenterSize = 1 }; -inline BoxCodeType GetBoxCodeType(const std::string& type) { +inline BoxCodeType GetBoxCodeType(const std::string &type) { if (type == "encode_center_size") { return BoxCodeType::kEncodeCenterSize; } else if (type == "decode_center_size") { @@ -32,24 +32,23 @@ inline BoxCodeType GetBoxCodeType(const std::string& type) { template class BoxCoderKernel : public framework::OpKernel { public: - void EncodeCenterSize(const framework::Tensor* target_box, - const framework::Tensor* prior_box, - const framework::Tensor* prior_box_var, + void EncodeCenterSize(const framework::Tensor *target_box, + const framework::Tensor *prior_box, + const framework::Tensor *prior_box_var, const bool normalized, - const std::vector variance, T* output) const { + const std::vector variance, T *output) const { int64_t row = target_box->dims()[0]; int64_t col = prior_box->dims()[0]; int64_t len = prior_box->dims()[1]; - auto* target_box_data = target_box->data(); - auto* prior_box_data = prior_box->data(); - const T* prior_box_var_data = nullptr; - if (prior_box_var) prior_box_var_data = prior_box_var->data(); #ifdef PADDLE_WITH_MKLML #pragma omp parallel for collapse(2) #endif for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { + auto *target_box_data = target_box->data(); + auto *prior_box_data = prior_box->data(); + size_t offset = i * col * len + j * len; T prior_box_width = prior_box_data[j * len + 2] - prior_box_data[j * len] + (normalized == false); T prior_box_height = prior_box_data[j * len + 3] - @@ -69,7 +68,6 @@ class BoxCoderKernel : public framework::OpKernel { target_box_data[i * len + 1] + (normalized == false); - size_t offset = i * col * len + j * len; output[offset] = (target_box_center_x - prior_box_center_x) / prior_box_width; output[offset + 1] = @@ -78,44 +76,61 @@ class BoxCoderKernel : public framework::OpKernel { std::log(std::fabs(target_box_width / prior_box_width)); output[offset + 3] = std::log(std::fabs(target_box_height / prior_box_height)); - if (prior_box_var) { - int prior_var_offset = j * len; - output[offset] /= prior_box_var_data[prior_var_offset]; - output[offset + 1] /= prior_box_var_data[prior_var_offset + 1]; - output[offset + 2] /= prior_box_var_data[prior_var_offset + 2]; - output[offset + 3] /= prior_box_var_data[prior_var_offset + 3]; - } else if (!(variance.empty())) { + } + } + + if (prior_box_var) { + const T *prior_box_var_data = prior_box_var->data(); +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for collapse(3) +#endif + for (int64_t i = 0; i < row; ++i) { + for (int64_t j = 0; j < col; ++j) { for (int k = 0; k < 4; ++k) { + size_t offset = i * col * len + j * len; + int prior_var_offset = j * len; + output[offset + k] /= prior_box_var_data[prior_var_offset + k]; + } + } + } + } else if (!(variance.empty())) { +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for collapse(3) +#endif + for (int64_t i = 0; i < row; ++i) { + for (int64_t j = 0; j < col; ++j) { + for (int k = 0; k < 4; ++k) { + size_t offset = i * col * len + j * len; output[offset + k] /= static_cast(variance[k]); } } } } } + template - void DecodeCenterSize(const framework::Tensor* target_box, - const framework::Tensor* prior_box, - const framework::Tensor* prior_box_var, + void DecodeCenterSize(const framework::Tensor *target_box, + const framework::Tensor *prior_box, + const framework::Tensor *prior_box_var, const bool normalized, std::vector variance, - T* output) const { + T *output) const { int64_t row = target_box->dims()[0]; int64_t col = target_box->dims()[1]; int64_t len = target_box->dims()[2]; - auto* target_box_data = target_box->data(); - auto* prior_box_data = prior_box->data(); - const T* prior_box_var_data = nullptr; - if (var_size == 2) prior_box_var_data = prior_box_var->data(); - int prior_box_offset = 0; - T var_data[4] = {1., 1., 1., 1.}; - T* var_ptr = var_data; #ifdef PADDLE_WITH_MKLML #pragma omp parallel for collapse(2) #endif for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { + auto *target_box_data = target_box->data(); + auto *prior_box_data = prior_box->data(); + + T var_data[4] = {1., 1., 1., 1.}; + T *var_ptr = var_data; size_t offset = i * col * len + j * len; - prior_box_offset = axis == 0 ? j * len : i * len; + int prior_box_offset = axis == 0 ? j * len : i * len; + T prior_box_width = prior_box_data[prior_box_offset + 2] - prior_box_data[prior_box_offset] + (normalized == false); @@ -131,10 +146,10 @@ class BoxCoderKernel : public framework::OpKernel { T target_box_width = 0, target_box_height = 0; int prior_var_offset = axis == 0 ? j * len : i * len; if (var_size == 2) { - std::memcpy(var_ptr, prior_box_var_data + prior_var_offset, + std::memcpy(var_ptr, prior_box_var->data() + prior_var_offset, 4 * sizeof(T)); } else if (var_size == 1) { - var_ptr = reinterpret_cast(variance.data()); + var_ptr = reinterpret_cast(variance.data()); } T box_var_x = *var_ptr; T box_var_y = *(var_ptr + 1); @@ -162,11 +177,11 @@ class BoxCoderKernel : public framework::OpKernel { } } - void Compute(const framework::ExecutionContext& context) const override { - auto* prior_box = context.Input("PriorBox"); - auto* prior_box_var = context.Input("PriorBoxVar"); - auto* target_box = context.Input("TargetBox"); - auto* output_box = context.Output("OutputBox"); + void Compute(const framework::ExecutionContext &context) const override { + auto *prior_box = context.Input("PriorBox"); + auto *prior_box_var = context.Input("PriorBoxVar"); + auto *target_box = context.Input("TargetBox"); + auto *output_box = context.Output("OutputBox"); std::vector variance = context.Attr>("variance"); const int axis = context.Attr("axis"); if (target_box->lod().size()) { @@ -194,7 +209,7 @@ class BoxCoderKernel : public framework::OpKernel { output_box->mutable_data({row, col, len}, context.GetPlace()); - T* output = output_box->data(); + T *output = output_box->data(); if (code_type == BoxCodeType::kEncodeCenterSize) { EncodeCenterSize(target_box, prior_box, prior_box_var, normalized, variance, output); diff --git a/paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc b/paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc index 04e8800bbc888540c4df21360c767688eb19c423..f2f4d3fee053a1e5bacd3c2165dba960f3befea4 100644 --- a/paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc +++ b/paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc @@ -110,8 +110,9 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel { constexpr int simd_width = 16; int C = c / simd_width; - auto multiply = jit::Get, - platform::CPUPlace>(0); + auto multiply = jit::KernelFuncs, + platform::CPUPlace>::Cache() + .At(0); #pragma omp parallel for collapse(2) for (int ni = 0; ni < n; ni++) { for (int ci = 0; ci < C; ci++) { diff --git a/paddle/fluid/operators/expand_op.cc b/paddle/fluid/operators/expand_op.cc index 44a2f37b66772425a835c26e94c37b500e8a5d19..fcb2be93635eeaeaae25c3a845fd06aa1a73e2e7 100644 --- a/paddle/fluid/operators/expand_op.cc +++ b/paddle/fluid/operators/expand_op.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/expand_op.h" +#include #include namespace paddle { @@ -138,12 +139,28 @@ class ExpandGradOp : public framework::OperatorWithKernel { } }; +class ExpandGradOpDescMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + std::unique_ptr op(new framework::OpDesc()); + op->SetType("expand_grad"); + op->SetInput("X", Input("X")); + op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetAttrMap(Attrs()); + return op; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(expand, ops::ExpandOp, ops::ExpandOpMaker, - paddle::framework::DefaultGradOpDescMaker); + ops::ExpandGradOpDescMaker); REGISTER_OPERATOR(expand_grad, ops::ExpandGradOp); REGISTER_OP_CPU_KERNEL( expand, ops::ExpandKernel, diff --git a/paddle/fluid/operators/fake_dequantize_op.cc b/paddle/fluid/operators/fake_dequantize_op.cc index 5d6488c67e0db440c8d4609736523643dd666dcc..68c7227e5a7123e1e751dd55e243ee481bf36540 100644 --- a/paddle/fluid/operators/fake_dequantize_op.cc +++ b/paddle/fluid/operators/fake_dequantize_op.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/fake_dequantize_op.h" #include +#include namespace paddle { namespace operators { @@ -76,6 +77,63 @@ $$Out = \frac{scale*X}{ max_range }$$ } }; +class FakeChannelWiseDequantizeMaxAbsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE( + ctx->HasInput("X"), + "Input(X) of FakeChannelWiseDequantizeMaxAbsOp should not be null."); + PADDLE_ENFORCE(ctx->HasInputs("Scales"), + "Input(Scales) of FakeChannelWiseDequantizeMaxAbsOp " + "should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Output(Out) of FakeChannelWiseDequantizeMaxAbsOp should not be null."); + + ctx->ShareDim("X", /*->*/ "Out"); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class FakeChannelWiseDequantizeMaxAbsOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor) The input with float-32/64 type is the " + "low precision tensor."); + AddInput("Scales", + "(Tensors) The scales in quantization stage. " + "Now, `Scales` is a vector with at most two tensors. " + "If Scales has two elements, the second tensor should only have " + "one value.") + .AsDuplicable(); + AddOutput("Out", + "(Tensor) The output is the dequantized high " + "precision tensor."); + AddAttr>( + "quant_bits", + "Quantization bit numbers in quantization stage. " + "The size of `quant_bits` should be equal to the size of `Scales`.") + .SetDefault({8}); + + AddComment(R"DOC( +FakeChannelWiseDequantizeMaxAbsOp operator. + +This calculation is an opposite operation of FakeChannelWiseQuantizeMaxAbsOp: + +$$Out_c = \frac{X_c\prod_{i=1}^{n}Scales_{ic}}{\prod_{i=1}^{n}(2^{quant\_bits_i-1}-1)}$$ + +In the above formula, the range value of $c$ can be represented as $0 \leq c \lt \ the\ channel\ number\ of\ X$. +Besides, the size of $quant\_bits$ should be equal to the size of $Scales$, and it is called $n$ in the formula. + +Notes: In general, the per-channel quantization is only applied to weights and the activations use per-layer quantization. +)DOC"); + } +}; + } // namespace operators } // namespace paddle @@ -88,3 +146,11 @@ REGISTER_OPERATOR(fake_dequantize_max_abs, ops::FakeDequantizeMaxAbsOp, REGISTER_OP_CPU_KERNEL(fake_dequantize_max_abs, ops::FakeDequantizeMaxAbsKernel, ops::FakeDequantizeMaxAbsKernel); + +REGISTER_OPERATOR(fake_channel_wise_dequantize_max_abs, + ops::FakeChannelWiseDequantizeMaxAbsOp, + ops::FakeChannelWiseDequantizeMaxAbsOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(fake_channel_wise_dequantize_max_abs, + ops::FakeChannelWiseDequantizeMaxAbsKernel, + ops::FakeChannelWiseDequantizeMaxAbsKernel); diff --git a/paddle/fluid/operators/fake_dequantize_op.cu b/paddle/fluid/operators/fake_dequantize_op.cu index 225bcc45bc65bc9268d1e866a4358731eaf0c3ef..35dcc69279d0119e75c4c5072e7817c839b9e819 100644 --- a/paddle/fluid/operators/fake_dequantize_op.cu +++ b/paddle/fluid/operators/fake_dequantize_op.cu @@ -55,3 +55,7 @@ using CUDA = paddle::platform::CUDADeviceContext; REGISTER_OP_CUDA_KERNEL(fake_dequantize_max_abs, ops::FakeDequantizeMaxAbsKernel, ops::FakeDequantizeMaxAbsKernel); +REGISTER_OP_CUDA_KERNEL( + fake_channel_wise_dequantize_max_abs, + ops::FakeChannelWiseDequantizeMaxAbsKernel, + ops::FakeChannelWiseDequantizeMaxAbsKernel); diff --git a/paddle/fluid/operators/fake_dequantize_op.h b/paddle/fluid/operators/fake_dequantize_op.h index d9923a10daa01ca06ebabb27cf9285b0628634bc..d05f2038531bbe9c35da54c94d2ef4d659acca70 100644 --- a/paddle/fluid/operators/fake_dequantize_op.h +++ b/paddle/fluid/operators/fake_dequantize_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -45,5 +46,42 @@ class FakeDequantizeMaxAbsKernel : public framework::OpKernel { } }; +template +class FakeChannelWiseDequantizeMaxAbsKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& ctx) const { + auto* in = ctx.Input("X"); + auto scales = ctx.MultiInput("Scales"); + auto* out = ctx.Output("Out"); + + PADDLE_ENFORCE_EQ(scales[0]->numel(), in->dims()[0], + "The number of first scale values must be the same with " + "first dimension value of Input(X)."); + + auto quant_bits = ctx.Attr>("quant_bits"); + int max_range = std::pow(2, quant_bits[0] - 1) - 1; + + auto& dev_ctx = ctx.template device_context(); + out->mutable_data(dev_ctx.GetPlace()); + + auto dequant = DequantizeFunctor(); + for (int64_t i = 0; i < in->dims()[0]; i++) { + framework::Tensor one_channel_in = in->Slice(i, i + 1); + framework::Tensor one_channel_out = out->Slice(i, i + 1); + framework::Tensor one_channel_scale = scales[0]->Slice(i, i + 1); + dequant(dev_ctx, &one_channel_in, &one_channel_scale, + static_cast(max_range), &one_channel_out); + } + + if (scales.size() == 2) { + PADDLE_ENFORCE_EQ( + scales[1]->numel(), 1, + "The second scale tensor should only have one value at now."); + max_range = std::pow(2, quant_bits[1] - 1) - 1; + dequant(dev_ctx, out, scales[1], static_cast(max_range), out); + } + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/fake_quantize_op.cc b/paddle/fluid/operators/fake_quantize_op.cc index 3bb07d383548e6f4be810c96d2a916c0fe5e45f5..70186e5efa29b1324ff7f3954720276156fddaf1 100644 --- a/paddle/fluid/operators/fake_quantize_op.cc +++ b/paddle/fluid/operators/fake_quantize_op.cc @@ -134,6 +134,60 @@ $$Out = round(X/scale * range)$$ } }; +class FakeChannelWiseQuantizeAbsMaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FakeChannelWiseQuantizeOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Output(Out) of FakeChannelWiseQuantizeOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("OutScales"), + "Output(Scales) of FakeChannelWiseQuantizeOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->SetOutputDim("OutScales", {ctx->GetInputDim("X")[0]}); + ctx->ShareLoD("X", /*->*/ "Out"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.GetPlace()); + } +}; + +class FakeChannelWiseQuantizeAbsMaxOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) Input is float data type."); + AddOutput("Out", + "(Tensor) Output of quantized low level tensor, " + "but also saved as float data type."); + AddOutput("OutScales", "(Tensor) Current channel wise scale"); + AddAttr("bit_length", "(int, default 8)") + .SetDefault(8) + .AddCustomChecker([](const int& bit_length) { + PADDLE_ENFORCE(bit_length >= 1 && bit_length <= 16, + "'bit_length' should be between 1 and 16."); + }); + AddComment(R"DOC( +The scale of FakeChannelWiseQuantize operator is a vector. +In detail, each channel of the input X has a scale value. + +$$scale_c = max(abs(X_c))$$ +$$range = 2^{bit\_length - 1} - 1$$ +$$Out_c = round(\frac{X_c * range} {scale_c})$$ +In above three formulas, the range value of c is as follow: +$$0 \leq c \lt \ the\ channel\ number\ of\ X$$ +)DOC"); + } +}; + class FakeQuantizeRangeAbsMaxOp : public framework::OperatorWithKernel { public: FakeQuantizeRangeAbsMaxOp(const std::string& type, @@ -218,3 +272,10 @@ REGISTER_OPERATOR(fake_quantize_range_abs_max, ops::FakeQuantizeRangeAbsMaxOp, paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL(fake_quantize_range_abs_max, ops::FakeQuantizeRangeAbsMaxKernel); + +REGISTER_OPERATOR(fake_channel_wise_quantize_abs_max, + ops::FakeChannelWiseQuantizeAbsMaxOp, + ops::FakeChannelWiseQuantizeAbsMaxOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(fake_channel_wise_quantize_abs_max, + ops::FakeChannelWiseQuantizeAbsMaxKernel); diff --git a/paddle/fluid/operators/fake_quantize_op.cu b/paddle/fluid/operators/fake_quantize_op.cu index a0ff6396210c2b3a7f8bd6b9f274b875d7fd4933..5da16a7c7314c62034bff67bcc8d099e2799c3de 100644 --- a/paddle/fluid/operators/fake_quantize_op.cu +++ b/paddle/fluid/operators/fake_quantize_op.cu @@ -174,5 +174,7 @@ namespace ops = paddle::operators; using CUDA = paddle::platform::CUDADeviceContext; REGISTER_OP_CUDA_KERNEL(fake_quantize_abs_max, ops::FakeQuantizeAbsMaxKernel); +REGISTER_OP_CUDA_KERNEL(fake_channel_wise_quantize_abs_max, + ops::FakeChannelWiseQuantizeAbsMaxKernel); REGISTER_OP_CUDA_KERNEL(fake_quantize_range_abs_max, ops::FakeQuantizeRangeAbsMaxKernel); diff --git a/paddle/fluid/operators/fake_quantize_op.h b/paddle/fluid/operators/fake_quantize_op.h index 7ace7573ec5c03ab8788cfc0aab614b7f80ea073..8b47600e7d99ad9e4e40ae162582d4c8461224ad 100644 --- a/paddle/fluid/operators/fake_quantize_op.h +++ b/paddle/fluid/operators/fake_quantize_op.h @@ -63,6 +63,39 @@ class FakeQuantizeAbsMaxKernel : public framework::OpKernel { } }; +template +class FakeChannelWiseQuantizeAbsMaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in = context.Input("X"); + + auto* out = context.Output("Out"); + auto* out_scales = context.Output("OutScales"); + T* out_scales_data = out_scales->mutable_data(context.GetPlace()); + out->mutable_data(context.GetPlace()); + + int bit_length = context.Attr("bit_length"); + int bin_cnt = std::pow(2, bit_length - 1) - 1; + + auto& dev_ctx = context.template device_context(); + auto find_abs_max = FindAbsMaxFunctor(); + for (int64_t i = 0; i < in->dims()[0]; i++) { + framework::Tensor one_channel = in->Slice(i, i + 1); + const T* one_channel_data = one_channel.data(); + find_abs_max(dev_ctx, one_channel_data, one_channel.numel(), + &out_scales_data[i]); + } + auto clip_quant = ClipAndFakeQuantFunctor(); + for (int64_t i = 0; i < in->dims()[0]; i++) { + framework::Tensor one_channel_in = in->Slice(i, i + 1); + framework::Tensor one_channel_out = out->Slice(i, i + 1); + framework::Tensor one_channel_scale = out_scales->Slice(i, i + 1); + clip_quant(dev_ctx, one_channel_in, one_channel_scale, bin_cnt, + &one_channel_out); + } + } +}; + template class FakeQuantizeRangeAbsMaxKernel : public framework::OpKernel { public: diff --git a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc index 80caf70b08e65932d6ccb90a5293d072b2b2bc72..a0026427e2514735711f7eba26fcf861cb498d5e 100644 --- a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc +++ b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.cc @@ -23,9 +23,6 @@ class FusedEmbeddingSeqPoolOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - if (ctx->IsRuntime()) { - return; - } PADDLE_ENFORCE(ctx->HasInput("W"), "Input W of FusedEmbeddingSeqPoolOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Ids"), @@ -91,6 +88,8 @@ class FusedEmbeddingSeqPoolOpMaker : public framework::OpProtoAndCheckerMaker { "(boolean, default false) " "Sparse update.") .SetDefault(false); + AddAttr(framework::kAllKernelsMustComputeRuntimeShape, "") + .SetDefault(true); AddComment(R"DOC( FusedEmbeddingSeqPool Operator. diff --git a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h index f13c02038606e52337b7ef85545e37054e54b631..4651c2b2ba81a404b64818fec81cef79634ff036 100644 --- a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h +++ b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h @@ -52,8 +52,9 @@ struct EmbeddingVSumFunctor { out_width, jit::SeqPoolType::kSum); for (size_t i = 0; i != ids_lod.size() - 1; ++i) { attr.index_height = ids_lod[i + 1] - ids_lod[i]; - auto emb_seqpool = jit::Get, - platform::CPUPlace>(attr); + auto emb_seqpool = + jit::KernelFuncs, platform::CPUPlace>::Cache() + .At(attr); emb_seqpool(table, ids + ids_lod[i] * idx_width, output + i * out_width, &attr); } @@ -120,6 +121,8 @@ class FusedEmbeddingSeqPoolGradKernel : public framework::OpKernel { auto *ids = context.Input("Ids"); auto *d_output = context.Input(framework::GradVarName("Out")); auto *d_table = context.Output(framework::GradVarName("W")); + // runtime shape + d_table->set_height(table_dim[0]); auto *ids_data = ids->data(); int64_t ids_num = ids->numel(); @@ -135,8 +138,9 @@ class FusedEmbeddingSeqPoolGradKernel : public framework::OpKernel { T *d_table_data = d_table_value->mutable_data(context.GetPlace()); const T *d_output_data = d_output->data(); - auto vbroadcast = jit::Get, - platform::CPUPlace>(out_width); + auto vbroadcast = + jit::KernelFuncs, platform::CPUPlace>::Cache() + .At(out_width); for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { int64_t h = static_cast(lod[i + 1] - lod[i]); const T *src = d_output_data + i * out_width; diff --git a/paddle/fluid/operators/fused/fusion_gru_op.cc b/paddle/fluid/operators/fused/fusion_gru_op.cc index 66acba49e5ac25c5097042225ccfe30b258040fa..ba5f0747c4d04bbb41f34dc7f895b22d38392ea6 100644 --- a/paddle/fluid/operators/fused/fusion_gru_op.cc +++ b/paddle/fluid/operators/fused/fusion_gru_op.cc @@ -182,29 +182,32 @@ class FusionGRUKernel : public framework::OpKernel { const int total_T = x_dims[0]; \ const int D3 = wh_dims[1] -#define INIT_OTHER_DEFINES \ - auto* h0 = ctx.Input("H0"); \ - auto* wx = ctx.Input("WeightX"); \ - auto* bias = ctx.Input("Bias"); \ - auto* hidden_out = ctx.Output("Hidden"); \ - bool is_reverse = ctx.Attr("is_reverse"); \ - const int M = x_dims[1]; \ - const int D = wh_dims[0]; \ - const int D2 = D * 2; \ - const jit::gru_attr_t attr( \ - D, jit::to_kerneltype(ctx.Attr("gate_activation")), \ - jit::to_kerneltype(ctx.Attr("activation"))); \ - jit::gru_t one_step; \ - auto ComputeH1 = \ - jit::Get, platform::CPUPlace>(attr); \ - auto ComputeHtPart1 = \ - jit::Get, platform::CPUPlace>(attr); \ - auto ComputeHtPart2 = \ - jit::Get, platform::CPUPlace>(attr); \ - const T* x_data = x->data(); \ - const T* wx_data = wx->data(); \ - const T* wh_data = wh->data(); \ - auto place = ctx.GetPlace(); \ +#define INIT_OTHER_DEFINES \ + auto* h0 = ctx.Input("H0"); \ + auto* wx = ctx.Input("WeightX"); \ + auto* bias = ctx.Input("Bias"); \ + auto* hidden_out = ctx.Output("Hidden"); \ + bool is_reverse = ctx.Attr("is_reverse"); \ + const int M = x_dims[1]; \ + const int D = wh_dims[0]; \ + const int D2 = D * 2; \ + const jit::gru_attr_t attr( \ + D, jit::to_kerneltype(ctx.Attr("gate_activation")), \ + jit::to_kerneltype(ctx.Attr("activation"))); \ + jit::gru_t one_step; \ + auto ComputeH1 = \ + jit::KernelFuncs, platform::CPUPlace>::Cache().At( \ + attr); \ + auto ComputeHtPart1 = \ + jit::KernelFuncs, platform::CPUPlace>::Cache() \ + .At(attr); \ + auto ComputeHtPart2 = \ + jit::KernelFuncs, platform::CPUPlace>::Cache() \ + .At(attr); \ + const T* x_data = x->data(); \ + const T* wx_data = wx->data(); \ + const T* wh_data = wh->data(); \ + auto place = ctx.GetPlace(); \ T* xx_data = xx->mutable_data(place) void SeqCompute(const framework::ExecutionContext& ctx) const { diff --git a/paddle/fluid/operators/fused/fusion_lstm_op.cc b/paddle/fluid/operators/fused/fusion_lstm_op.cc index b11b7c11bfe0ae4c79d5bb39844bce618649c44d..c8c07bd126d5b4eac688d43fd794856f8222525a 100644 --- a/paddle/fluid/operators/fused/fusion_lstm_op.cc +++ b/paddle/fluid/operators/fused/fusion_lstm_op.cc @@ -235,32 +235,34 @@ class FuisonLSTMKernel : public framework::OpKernel { const int D = wh_dims[0]; \ const int D4 = wh_dims[1] -#define INIT_OTHER_DEFINES \ - const T* x_data = x->data(); \ - const T* wx_data = wx->data(); \ - const T* wh_data = wh->data(); \ - /* diagonal weight*/ \ - const T* wp_data = bias->data() + D4; \ - /* for peephole only*/ \ - T* checked_cell_data = nullptr; \ - auto place = ctx.GetPlace(); \ - if (use_peepholes) { \ - /* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \ - auto* checked_cell = ctx.Output("CheckedCell"); \ - checked_cell_data = checked_cell->mutable_data(place); \ - } \ - const jit::lstm_attr_t attr( \ - D, jit::to_kerneltype(ctx.Attr("gate_activation")), \ - jit::to_kerneltype(ctx.Attr("candidate_activation")), \ - jit::to_kerneltype(ctx.Attr("cell_activation")), \ - use_peepholes); \ - jit::lstm_t one_step; \ - one_step.wp = wp_data; \ - one_step.checked = checked_cell_data; \ - auto ComputeC1H1 = \ - jit::Get, platform::CPUPlace>(attr); \ - auto ComputeCtHt = \ - jit::Get, platform::CPUPlace>(attr) +#define INIT_OTHER_DEFINES \ + const T* x_data = x->data(); \ + const T* wx_data = wx->data(); \ + const T* wh_data = wh->data(); \ + /* diagonal weight*/ \ + const T* wp_data = bias->data() + D4; \ + /* for peephole only*/ \ + T* checked_cell_data = nullptr; \ + auto place = ctx.GetPlace(); \ + if (use_peepholes) { \ + /* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \ + auto* checked_cell = ctx.Output("CheckedCell"); \ + checked_cell_data = checked_cell->mutable_data(place); \ + } \ + const jit::lstm_attr_t attr( \ + D, jit::to_kerneltype(ctx.Attr("gate_activation")), \ + jit::to_kerneltype(ctx.Attr("candidate_activation")), \ + jit::to_kerneltype(ctx.Attr("cell_activation")), \ + use_peepholes); \ + jit::lstm_t one_step; \ + one_step.wp = wp_data; \ + one_step.checked = checked_cell_data; \ + auto ComputeC1H1 = \ + jit::KernelFuncs, platform::CPUPlace>::Cache().At( \ + attr); \ + auto ComputeCtHt = \ + jit::KernelFuncs, platform::CPUPlace>::Cache().At( \ + attr) // Wh GEMM #define GEMM_WH_ADDON(bs, prev, out) \ diff --git a/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc b/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc index 8ecdf2ed9d40e7f5dc9226c635a8c8e6406a76ba..6be35de65f48525b2da7d5c9ef260b2d0798b67b 100644 --- a/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc +++ b/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc @@ -82,9 +82,11 @@ template static void fc_relu(const T* x, const T* w, const T* b, T* y, const jit::matmul_attr_t& attr) { auto matmul = - jit::Get, platform::CPUPlace>(attr); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr); auto addbias_relu = - jit::Get, platform::CPUPlace>(attr.n); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr.n); matmul(x, w, y, &attr); T* dst = y; for (int i = 0; i < attr.m; ++i) { diff --git a/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc b/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc index d48bdafe0aa38cb860b54b2e41ebad3421b93bce..25916768c08e7222ba95bd6e1999400a923b21a3 100644 --- a/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc +++ b/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc @@ -98,7 +98,7 @@ class FusionSeqPoolConcatKernel : public framework::OpKernel { attr.type = jit::SeqPoolType::kSqrt; } auto seqpool = - jit::Get, platform::CPUPlace>( + jit::KernelFuncs, platform::CPUPlace>::Cache().At( attr); size_t n = ins.size(); size_t dst_step_size = n * w; diff --git a/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc b/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc index 8493f4468fc994964116d99dc85dd34fb19a44cc..53679ebddee1ceec102b5861c54b398aa4da4cde 100644 --- a/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc +++ b/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc @@ -94,19 +94,23 @@ class FusionSquaredMatSubKernel : public framework::OpKernel { int o_numel = attr.m * attr.n; auto vsquare_x = - jit::Get, platform::CPUPlace>(attr.m * - attr.k); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr.m * attr.k); auto vsquare_y = - jit::Get, platform::CPUPlace>(attr.k * - attr.n); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr.k * attr.n); auto vsquare_xy = - jit::Get, platform::CPUPlace>(o_numel); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + o_numel); auto vsub = - jit::Get, platform::CPUPlace>(o_numel); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + o_numel); auto vscal = - jit::Get, platform::CPUPlace>(o_numel); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + o_numel); auto matmul = - jit::Get, platform::CPUPlace>(attr); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr); const T* x_data = x->data(); const T* y_data = y->data(); diff --git a/paddle/fluid/operators/hash_op.cc b/paddle/fluid/operators/hash_op.cc index 7a29f80ff1ce413519ea9cea6a35747bdced5885..f6395fb32feac175976cb96e1c0bee7347cb3ea8 100644 --- a/paddle/fluid/operators/hash_op.cc +++ b/paddle/fluid/operators/hash_op.cc @@ -26,9 +26,6 @@ class HashOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} void InferShape(framework::InferShapeContext *ctx) const override { - if (ctx->IsRuntime()) { - return; - } PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of HashOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), @@ -57,6 +54,8 @@ $$Out = scale * X$$ )DOC"); AddAttr("num_hash", "").SetDefault(1); AddAttr("mod_by", "").SetDefault(100000); + AddAttr(framework::kAllKernelsMustComputeRuntimeShape, "") + .SetDefault(true); } }; diff --git a/paddle/fluid/operators/jit/CMakeLists.txt b/paddle/fluid/operators/jit/CMakeLists.txt index 35775d7ec9efcdbad69e4491792f7d4e513832ad..47d6c83f2adf8c4b7476410ce7c1d435633a8bfe 100644 --- a/paddle/fluid/operators/jit/CMakeLists.txt +++ b/paddle/fluid/operators/jit/CMakeLists.txt @@ -5,7 +5,7 @@ file(APPEND ${jit_file} "\#pragma once\n") file(APPEND ${jit_file} "\#include \"paddle/fluid/operators/jit/helper.h\"\n") file(APPEND ${jit_file} "\#include \"paddle/fluid/operators/jit/registry.h\"\n\n") -set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce place) +set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce place xxhash) file(GLOB jit_kernel_cc_srcs RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cc") list(REMOVE_ITEM jit_kernel_cc_srcs test.cc benchmark.cc) diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc index 3088280bb90174e6195a349c07a3435e131e2b33..fbb04a166ef52efd9bd05f27ca656d928d97fb96 100644 --- a/paddle/fluid/operators/jit/benchmark.cc +++ b/paddle/fluid/operators/jit/benchmark.cc @@ -59,8 +59,6 @@ BenchJITKernel* InsertBenchmark(BenchJITKernel* b) { InsertBenchmark(new BenchJITKernel_##name##_##dtype##_##place##_()); \ void BenchJITKernel_##name##_##dtype##_##place##_::Run() -#define BENCH_FP32_CPU(name) BENCH_JITKERNEL(name, FP32, CPU) - void RUN_ALL_BENCHMARK() { for (auto p : g_all_benchmarks) { if (!FLAGS_filter.empty() && FLAGS_filter != p->Name()) { @@ -90,11 +88,11 @@ std::vector TestSizes() { return s; } -template +template struct BenchFunc { // return this function avg time // TODO(TJ): clear cache every time - double operator()(const typename KernelTuples::func_type tgt, Args... args) { + double operator()(const typename KernelTuple::func_type tgt, Args... args) { for (int i = 0; i < FLAGS_burning; ++i) { tgt(args...); } @@ -109,40 +107,17 @@ struct BenchFunc { namespace jit = paddle::operators::jit; -template -void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { - BenchFunc benchmark; +template +void BenchAllImpls(const typename KernelTuple::attr_type& attr, Args... args) { + BenchFunc benchmark; std::vector> infos; - // test refer - auto refer = jit::GetRefer(); - if (!refer) { - LOG(FATAL) << "Refer can not be empty!"; + auto funcs = jit::GetAllCandidateFuncsWithTypes(attr); + for (auto f : funcs) { + infos.push_back(std::make_pair(f.first, benchmark(f.second, args...))); } - infos.push_back(std::make_pair("Refer", benchmark(refer, args...))); - // test jitcode - auto jitcode = jit::GetJitCode(attr); - if (jitcode) { - infos.push_back(std::make_pair("JitCode", benchmark(jitcode, args...))); - } - // test all impls in more - jit::KernelKey kkey(KT, PlaceType()); - auto& pool = jit::KernelPool().Instance().AllKernels(); - auto iter = pool.find(kkey); - if (iter != pool.end()) { - auto& impls = iter->second; - for (auto& impl : impls) { - auto i = dynamic_cast*>(impl.get()); - if (i && i->UseMe(attr)) { - auto more = i->GetFunc(); - infos.push_back( - std::make_pair(i->ImplType(), benchmark(more, args...))); - } - } - } // Test result from Get function - auto tgt = jit::Get(attr); + auto tgt = jit::KernelFuncs::Cache().At(attr); if (!tgt) { LOG(FATAL) << "Target can not be empty!"; } @@ -150,7 +125,8 @@ void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { // print std::ostringstream loginfos; - loginfos << "Kernel Type " << jit::to_string(KT) << ": " << attr << ": "; + loginfos << "Kernel Type " << jit::to_string(KernelTuple::kernel_type) << ": " + << attr << ": "; for (auto pair : infos) { loginfos << pair.first << " takes " << pair.second << " us; "; } @@ -159,8 +135,9 @@ void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { using Tensor = paddle::framework::Tensor; -template -void BenchXYZNKernel() { +template +void BenchKernelXYZN() { + using T = typename KernelTuple::data_type; for (int d : TestSizes()) { Tensor x, y, z; x.Resize({d}); @@ -171,16 +148,16 @@ void BenchXYZNKernel() { T* z_data = z.mutable_data(PlaceType()); RandomVec(d, x_data); RandomVec(d, y_data); - BenchAllImpls, PlaceType>(d, x.data(), - y.data(), z_data, d); + BenchAllImpls(d, x.data(), y.data(), z_data, + d); // test inplace - BenchAllImpls, PlaceType>(d, x.data(), z_data, - z_data, d); + BenchAllImpls(d, x.data(), z_data, z_data, d); } } -template -void BenchAXYNKernel() { +template +void BenchKernelAXYN() { + using T = typename KernelTuple::data_type; for (int d : TestSizes()) { const T a = static_cast(3); Tensor x, y; @@ -189,26 +166,26 @@ void BenchAXYNKernel() { T* x_data = x.mutable_data(PlaceType()); T* y_data = y.mutable_data(PlaceType()); RandomVec(d, x_data); - BenchAllImpls, PlaceType>(d, &a, x.data(), y_data, - d); + BenchAllImpls(d, &a, x.data(), y_data, d); // test inplace - BenchAllImpls, PlaceType>(d, &a, x.data(), x_data, - d); + BenchAllImpls(d, &a, x.data(), x_data, d); } } -template -void BenchXRNKernel() { +template +void BenchKernelXRN() { + using T = typename KernelTuple::data_type; for (int d : TestSizes()) { Tensor x; RandomVec(d, x.mutable_data({d}, PlaceType())); T res; - BenchAllImpls, PlaceType>(d, x.data(), &res, d); + BenchAllImpls(d, x.data(), &res, d); } } -template -void BenchXYNKernel() { +template +void BenchKernelXYN() { + using T = typename KernelTuple::data_type; for (int d : TestSizes()) { Tensor x, y; x.Resize({d}); @@ -216,12 +193,13 @@ void BenchXYNKernel() { T* x_data = x.mutable_data(PlaceType()); T* y_data = y.mutable_data(PlaceType()); RandomVec(d, x_data); - BenchAllImpls, PlaceType>(d, x.data(), y_data, d); + BenchAllImpls(d, x.data(), y_data, d); } } -template -void BenchLSTMKernel() { +template +void BenchKernelLSTM() { + using T = typename KernelTuple::data_type; for (bool use_peephole : {true, false}) { for (int d : TestSizes()) { const jit::lstm_attr_t attr(d, jit::kVSigmoid, jit::kVTanh, jit::kVTanh, @@ -252,13 +230,14 @@ void BenchLSTMKernel() { step.wp = wp_data; step.checked = checked_data; } - BenchAllImpls, PlaceType>(attr, &step, &attr); + BenchAllImpls(attr, &step, &attr); } } } -template -void BenchGRUKernel() { +template +void BenchKernelGRU() { + using T = typename KernelTuple::data_type; for (int d : TestSizes()) { const jit::gru_attr_t attr(d, jit::kVSigmoid, jit::kVTanh); auto place = PlaceType(); @@ -275,12 +254,13 @@ void BenchGRUKernel() { step.gates = x_data; step.ht_1 = ht_1_data; step.ht = ht_data; - BenchAllImpls, PlaceType>(attr, &step, &attr); + BenchAllImpls(attr, &step, &attr); } } -template -void BenchSeqPoolKernel() { +template +void BenchKernelSeqPool() { + using T = typename KernelTuple::data_type; std::vector pool_types = { jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt}; for (auto type : pool_types) { @@ -294,15 +274,15 @@ void BenchSeqPoolKernel() { RandomVec(h * w, x.mutable_data(PlaceType()), -2.f, 2.f); const T* x_data = x.data(); T* y_data = y.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>(attr, x_data, - y_data, &attr); + BenchAllImpls(attr, x_data, y_data, &attr); } } } } -template -void BenchEmbSeqPoolKernel() { +template +void BenchKernelEmbSeqPool() { + using T = typename KernelTuple::data_type; std::vector pool_types = {jit::SeqPoolType::kSum}; int64_t tbl_h = 1e4; for (int tbl_w : {10, 16, 256}) { @@ -324,16 +304,17 @@ void BenchEmbSeqPoolKernel() { tbl_h - 1); const int64_t* idx_data = idx.data(); T* o_data = out.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>( - attr, table_data, idx_data, o_data, &attr); + BenchAllImpls(attr, table_data, idx_data, + o_data, &attr); } } } } } -template -void BenchSgdKernel() { +template +void BenchKernelSgd() { + using T = typename KernelTuple::data_type; const T lr = 0.1; auto UnDuplicatedRandomVec = [](int n, const int64_t lower, const int64_t upper) -> std::vector { @@ -364,15 +345,16 @@ void BenchSgdKernel() { const T* grad_data = grad.data(); const int64_t* rows_data = rows.data(); jit::sgd_attr_t attr(param_h, grad_w, rows_size, grad_w, rows_size); - BenchAllImpls, PlaceType>( - attr, &lr, param_data, grad_data, rows_data, param_data, &attr); + BenchAllImpls(attr, &lr, param_data, grad_data, + rows_data, param_data, &attr); } } } } -template -void BenchMatMulKernel() { +template +void BenchKernelMatMul() { + using T = typename KernelTuple::data_type; for (int m : {1, 2, 3, 4}) { for (int n : TestSizes()) { for (int k : TestSizes()) { @@ -386,15 +368,16 @@ void BenchMatMulKernel() { const T* b_data = b.data(); T* c_data = c.mutable_data(PlaceType()); const jit::matmul_attr_t attr{m, n, k}; - BenchAllImpls, PlaceType>(attr, a_data, b_data, - c_data, &attr); + BenchAllImpls(attr, a_data, b_data, c_data, + &attr); } } } } -template -void BenchSoftmaxKernel() { +template +void BenchKernelSoftmax() { + using T = typename KernelTuple::data_type; for (int bs : {1, 2, 10}) { for (int n : TestSizes()) { Tensor x, y; @@ -403,14 +386,14 @@ void BenchSoftmaxKernel() { RandomVec(bs * n, x.mutable_data(PlaceType()), -2.f, 2.f); const T* x_data = x.data(); T* y_data = y.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>(n, x_data, y_data, n, - bs); + BenchAllImpls(n, x_data, y_data, n, bs); } } } -template -void BenchLayerNormKernel() { +template +void BenchKernelLayerNorm() { + using T = typename KernelTuple::data_type; const T epsilon = 9.99999975e-06; for (int n : {1, 2, 10}) { for (int x_dim_0 : {1, 9, 17, 50}) { @@ -439,16 +422,17 @@ void BenchLayerNormKernel() { T* var_data = var.data(); T* out_data = out.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>( - right, x_data, out_data, mean_data, var_data, scale_data, bias_data, - left, epsilon, right); + BenchAllImpls(right, x_data, out_data, + mean_data, var_data, scale_data, + bias_data, left, epsilon, right); } } } } -template -void BenchCRFDecodingKernel() { +template +void BenchKernelCRFDecoding() { + using T = typename KernelTuple::data_type; constexpr int state_trans_base_idx = 2; for (int seq_len : {1, 11, 17, 50}) { for (int tag_num : TestSizes()) { @@ -468,14 +452,15 @@ void BenchCRFDecodingKernel() { T* alpha_data = alpha.mutable_data(PlaceType()); int* track_data = track.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>( - tag_num, seq_len, x_data, w_data, alpha_data, track_data, tag_num); + BenchAllImpls(tag_num, seq_len, x_data, w_data, + alpha_data, track_data, tag_num); } } } -template -void BenchVBroadcastKernel() { +template +void BenchKernelVBroadcast() { + using T = typename KernelTuple::data_type; for (int64_t w : {1, 16, 64, 100, 256}) { Tensor x; x.Resize({w}); @@ -485,78 +470,86 @@ void BenchVBroadcastKernel() { Tensor y; y.Resize({h * w}); T* y_data = y.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>( - w, x_data, y_data, static_cast(h), w); + BenchAllImpls(w, x_data, y_data, + static_cast(h), w); } } } -using T = float; -using CPUPlace = paddle::platform::CPUPlace; +#define BenchKernelVMul BenchKernelXYZN +#define BenchKernelVAdd BenchKernelXYZN +#define BenchKernelVAddRelu BenchKernelXYZN +#define BenchKernelVSub BenchKernelXYZN -// xyzn -BENCH_FP32_CPU(kVMul) { BenchXYZNKernel(); } -BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel(); } -BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel(); } -BENCH_FP32_CPU(kVSub) { BenchXYZNKernel(); } +#define BenchKernelVScal BenchKernelAXYN +#define BenchKernelVAddBias BenchKernelAXYN -// axyn -BENCH_FP32_CPU(kVScal) { BenchAXYNKernel(); } -BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel(); } +#define BenchKernelVRelu BenchKernelXYN +#define BenchKernelVIdentity BenchKernelXYN +#define BenchKernelVSquare BenchKernelXYN +#define BenchKernelVExp BenchKernelXYN +#define BenchKernelVSigmoid BenchKernelXYN +#define BenchKernelVTanh BenchKernelXYN +#define BenchKernelVCopy BenchKernelXYN -// xrn -BENCH_FP32_CPU(kHSum) { BenchXRNKernel(); } -BENCH_FP32_CPU(kHMax) { BenchXRNKernel(); } +#define BenchKernelHMax BenchKernelXRN +#define BenchKernelHSum BenchKernelXRN -// xyn -BENCH_FP32_CPU(kVRelu) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVSquare) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVExp) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVTanh) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVCopy) { BenchXYNKernel(); } - -// lstm and peephole -BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel(); } -BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel(); } - -// gru functions -BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel(); } -BENCH_FP32_CPU(kGRUHtPart1) { BenchGRUKernel(); } -BENCH_FP32_CPU(kGRUHtPart2) { BenchGRUKernel(); } - -// seq pool function -BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel(); } - -// embedding seq pool function -BENCH_FP32_CPU(kEmbSeqPool) { - BenchEmbSeqPoolKernel(); -} +#define BenchKernelLSTMCtHt BenchKernelLSTM +#define BenchKernelLSTMC1H1 BenchKernelLSTM -// sgd function -BENCH_FP32_CPU(kSgd) { BenchSgdKernel(); } +#define BenchKernelGRUH1 BenchKernelGRU +#define BenchKernelGRUHtPart1 BenchKernelGRU +#define BenchKernelGRUHtPart2 BenchKernelGRU -// matmul -BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel(); } +using CPUPlace = paddle::platform::CPUPlace; -// softmax -BENCH_FP32_CPU(kSoftmax) { BenchSoftmaxKernel(); } +#define BENCH_FP32_CPU(name) \ + BENCH_JITKERNEL(name, FP32, CPU) { \ + BenchKernel##name, CPUPlace>(); \ + } -// layernorm -BENCH_FP32_CPU(kLayerNorm) { - BenchLayerNormKernel(); -} +// xyzn +BENCH_FP32_CPU(VMul); +BENCH_FP32_CPU(VAdd); +BENCH_FP32_CPU(VAddRelu); +BENCH_FP32_CPU(VSub); -// crfdecoding -BENCH_FP32_CPU(kCRFDecoding) { - BenchCRFDecodingKernel(); -} +// axyn +BENCH_FP32_CPU(VScal); +BENCH_FP32_CPU(VAddBias); -// vbroadcast function -BENCH_FP32_CPU(kVBroadcast) { - BenchVBroadcastKernel(); -} +// xyn +BENCH_FP32_CPU(VRelu); +BENCH_FP32_CPU(VIdentity); +BENCH_FP32_CPU(VSquare); +BENCH_FP32_CPU(VExp); +BENCH_FP32_CPU(VSigmoid); +BENCH_FP32_CPU(VTanh); +BENCH_FP32_CPU(VCopy); + +// xrn +BENCH_FP32_CPU(HMax); +BENCH_FP32_CPU(HSum); + +// LSTM +BENCH_FP32_CPU(LSTMCtHt); +BENCH_FP32_CPU(LSTMC1H1); + +// GRU +BENCH_FP32_CPU(GRUH1); +BENCH_FP32_CPU(GRUHtPart1); +BENCH_FP32_CPU(GRUHtPart2); + +BENCH_FP32_CPU(LayerNorm); +BENCH_FP32_CPU(CRFDecoding); + +BENCH_FP32_CPU(SeqPool); +BENCH_FP32_CPU(EmbSeqPool); +BENCH_FP32_CPU(MatMul); +BENCH_FP32_CPU(Softmax); +BENCH_FP32_CPU(Sgd); +BENCH_FP32_CPU(VBroadcast); // Benchmark all jit kernels including jitcode, mkl and refer. // To use this tool, run command: ./benchmark [options...] diff --git a/paddle/fluid/operators/jit/gen/act.cc b/paddle/fluid/operators/jit/gen/act.cc index e7a7375879064eb27c94315fe7b93eece7866b92..ad68e792c7a8ec4fb600a5b04153ad45895d761a 100644 --- a/paddle/fluid/operators/jit/gen/act.cc +++ b/paddle/fluid/operators/jit/gen/act.cc @@ -13,6 +13,7 @@ * limitations under the License. */ #include "paddle/fluid/operators/jit/gen/act.h" +#include #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -81,7 +82,7 @@ void VActJitCode::genCode() { #define DECLARE_ACT_CREATOR(name) \ class name##Creator : public JitCodeCreator { \ public: \ - bool UseMe(const int& attr) const override; \ + bool CanBeUsed(const int& attr) const override; \ size_t CodeSize(const int& d) const override; \ std::unique_ptr CreateJitCode(const int& attr) const override { \ return make_unique(attr, CodeSize(attr)); \ @@ -96,27 +97,27 @@ DECLARE_ACT_CREATOR(VSigmoid); DECLARE_ACT_CREATOR(VTanh); // TODO(TJ): tuning use me -bool VReluCreator::UseMe(const int& d) const { +bool VReluCreator::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx); } -bool VSquareCreator::UseMe(const int& d) const { +bool VSquareCreator::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx); } -bool VIdentityCreator::UseMe(const int& d) const { +bool VIdentityCreator::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx); } -bool VExpCreator::UseMe(const int& d) const { +bool VExpCreator::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx) && d < 32; } -bool VSigmoidCreator::UseMe(const int& d) const { +bool VSigmoidCreator::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx); } -bool VTanhCreator::UseMe(const int& d) const { +bool VTanhCreator::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx); } diff --git a/paddle/fluid/operators/jit/gen/blas.cc b/paddle/fluid/operators/jit/gen/blas.cc index 5da24c359edd2df93333fe0ca8a18cdc7385aadb..c126b9077ae50f528074210bae39227a9fcd3277 100644 --- a/paddle/fluid/operators/jit/gen/blas.cc +++ b/paddle/fluid/operators/jit/gen/blas.cc @@ -13,6 +13,7 @@ * limitations under the License. */ #include "paddle/fluid/operators/jit/gen/blas.h" +#include #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -142,7 +143,7 @@ void NCHW16CMulNCJitCode::genCode() { class NCHW16CMulNCCreator : public JitCodeCreator { public: - bool UseMe(const int& attr) const override { + bool CanBeUsed(const int& attr) const override { return platform::MayIUse(platform::avx512f); } size_t CodeSize(const int& d) const override { return 256 * 1024; } @@ -154,7 +155,7 @@ class NCHW16CMulNCCreator : public JitCodeCreator { #define DECLARE_BLAS_CREATOR(name) \ class name##Creator : public JitCodeCreator { \ public: \ - bool UseMe(const int& attr) const override { \ + bool CanBeUsed(const int& attr) const override { \ return platform::MayIUse(platform::avx) && attr <= 1024; \ } \ size_t CodeSize(const int& d) const override { \ diff --git a/paddle/fluid/operators/jit/gen/embseqpool.cc b/paddle/fluid/operators/jit/gen/embseqpool.cc index 23837a3fb9886ae8a839d4b31bd57916168ea53c..331a4b0d0753b37843c3d112256abfbabe9a4913 100644 --- a/paddle/fluid/operators/jit/gen/embseqpool.cc +++ b/paddle/fluid/operators/jit/gen/embseqpool.cc @@ -14,6 +14,7 @@ #include "paddle/fluid/operators/jit/gen/embseqpool.h" #include // offsetof +#include #include #include "paddle/fluid/operators/jit/gen/act.h" // for exp_float_consts ones #include "paddle/fluid/operators/jit/registry.h" @@ -121,7 +122,7 @@ void EmbSeqPoolJitCode::genCode() { class EmbSeqPoolCreator : public JitCodeCreator { public: - bool UseMe(const emb_seq_pool_attr_t& attr) const override { + bool CanBeUsed(const emb_seq_pool_attr_t& attr) const override { return platform::MayIUse(platform::avx) && attr.table_width % YMM_FLOAT_BLOCK == 0; } diff --git a/paddle/fluid/operators/jit/gen/gru.cc b/paddle/fluid/operators/jit/gen/gru.cc index 13f7a14111a80632a06c7fc632da47c0802828f7..b5b0cffa80612c61829766027013f172962b5069 100644 --- a/paddle/fluid/operators/jit/gen/gru.cc +++ b/paddle/fluid/operators/jit/gen/gru.cc @@ -14,6 +14,7 @@ #include "paddle/fluid/operators/jit/gen/gru.h" #include // offsetof +#include #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -86,7 +87,7 @@ void GRUJitCode::genCode() { class name##Creator : public JitCodeCreator { \ public: \ /* TODO(TJ): enable more */ \ - bool UseMe(const gru_attr_t& attr) const override { \ + bool CanBeUsed(const gru_attr_t& attr) const override { \ return platform::MayIUse(platform::avx) && attr.d % 8 == 0; \ } \ size_t CodeSize(const gru_attr_t& attr) const override { \ diff --git a/paddle/fluid/operators/jit/gen/hopv.cc b/paddle/fluid/operators/jit/gen/hopv.cc index e7884017198623d996fe98a55691da6e342d656a..462ac68a932e14b1200d503a937a35454c0e0618 100644 --- a/paddle/fluid/operators/jit/gen/hopv.cc +++ b/paddle/fluid/operators/jit/gen/hopv.cc @@ -13,6 +13,7 @@ * limitations under the License. */ #include "paddle/fluid/operators/jit/gen/hopv.h" +#include #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -76,7 +77,7 @@ void HOPVJitCode::genCode() { #define DECLARE_HOP_CREATOR(name) \ class name##Creator : public JitCodeCreator { \ public: \ - bool UseMe(const int& attr) const override { \ + bool CanBeUsed(const int& attr) const override { \ return platform::MayIUse(platform::avx); \ } \ size_t CodeSize(const int& d) const override { \ diff --git a/paddle/fluid/operators/jit/gen/jitcode.h b/paddle/fluid/operators/jit/gen/jitcode.h index 39847d1b65f771976c4dde5a3e34cc40e33851e6..228db7cc721099750da30adeaa828ae31f521422 100644 --- a/paddle/fluid/operators/jit/gen/jitcode.h +++ b/paddle/fluid/operators/jit/gen/jitcode.h @@ -73,7 +73,7 @@ class JitCode : public GenBase, public Xbyak::CodeGenerator { virtual void genCode() = 0; size_t getSize() const override { return CodeGenerator::getSize(); } - const unsigned char* getCodeInternal() override { + const unsigned char* getCodeInternal() const override { const Xbyak::uint8* code = CodeGenerator::getCode(); return code; } diff --git a/paddle/fluid/operators/jit/gen/lstm.cc b/paddle/fluid/operators/jit/gen/lstm.cc index 08bafb5a81882072129a4bfa86d5aff2d33a79a1..2c3bc985e9a8b224835d848d30e0a3ef641ed2f9 100644 --- a/paddle/fluid/operators/jit/gen/lstm.cc +++ b/paddle/fluid/operators/jit/gen/lstm.cc @@ -14,6 +14,7 @@ #include "paddle/fluid/operators/jit/gen/lstm.h" #include // offsetof +#include #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -114,7 +115,7 @@ void LSTMJitCode::genCode() { class name##Creator : public JitCodeCreator { \ public: \ /* TODO(TJ): enable more */ \ - bool UseMe(const lstm_attr_t& attr) const override { \ + bool CanBeUsed(const lstm_attr_t& attr) const override { \ return platform::MayIUse(platform::avx) && attr.d % 8 == 0; \ } \ size_t CodeSize(const lstm_attr_t& attr) const override { \ diff --git a/paddle/fluid/operators/jit/gen/matmul.cc b/paddle/fluid/operators/jit/gen/matmul.cc index ae3858eab20aeb80553d8fcec4088a6632c9c17d..d9955c8cc655f86bbc6c8135bdfa6c83493727f2 100644 --- a/paddle/fluid/operators/jit/gen/matmul.cc +++ b/paddle/fluid/operators/jit/gen/matmul.cc @@ -14,8 +14,8 @@ #include "paddle/fluid/operators/jit/gen/matmul.h" #include // offsetof +#include #include - #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -98,7 +98,7 @@ void MatMulJitCode::genCode() { class MatMulCreator : public JitCodeCreator { public: - bool UseMe(const matmul_attr_t& attr) const override { + bool CanBeUsed(const matmul_attr_t& attr) const override { return attr.m == 1 && platform::MayIUse(platform::avx512f) && attr.n % ZMM_FLOAT_BLOCK == 0 && attr.k < 512; } diff --git a/paddle/fluid/operators/jit/gen/seqpool.cc b/paddle/fluid/operators/jit/gen/seqpool.cc index 530d24ee1fb7d9da84102641e1d4d2ab08ab1860..d9e5904add4486ddf126093865f7e0571c1909e4 100644 --- a/paddle/fluid/operators/jit/gen/seqpool.cc +++ b/paddle/fluid/operators/jit/gen/seqpool.cc @@ -13,6 +13,7 @@ * limitations under the License. */ #include "paddle/fluid/operators/jit/gen/seqpool.h" +#include #include "paddle/fluid/operators/jit/gen/act.h" // for exp_float_consts ones #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -57,7 +58,7 @@ void SeqPoolJitCode::genCode() { class SeqPoolCreator : public JitCodeCreator { public: - bool UseMe(const seq_pool_attr_t& attr) const override { + bool CanBeUsed(const seq_pool_attr_t& attr) const override { return platform::MayIUse(platform::avx); } size_t CodeSize(const seq_pool_attr_t& attr) const override { diff --git a/paddle/fluid/operators/jit/gen/sgd.cc b/paddle/fluid/operators/jit/gen/sgd.cc index a745a27f9543a75f6915c9316aad62fa41305bb1..e65d3500b496c811b2da39752417ce5ef3ab3914 100644 --- a/paddle/fluid/operators/jit/gen/sgd.cc +++ b/paddle/fluid/operators/jit/gen/sgd.cc @@ -14,6 +14,7 @@ #include "paddle/fluid/operators/jit/gen/sgd.h" #include // offsetof +#include #include #include "paddle/fluid/operators/jit/registry.h" #include "paddle/fluid/platform/cpu_info.h" @@ -104,7 +105,7 @@ void SgdJitCode::genCode() { class SgdCreator : public JitCodeCreator { public: - bool UseMe(const sgd_attr_t& attr) const override { + bool CanBeUsed(const sgd_attr_t& attr) const override { return platform::MayIUse(platform::avx) && attr.grad_width % YMM_FLOAT_BLOCK == 0; } diff --git a/paddle/fluid/operators/jit/gen/vbroadcast.cc b/paddle/fluid/operators/jit/gen/vbroadcast.cc index 3f9fbdbd821acae0940c5a7b8d9a5eb2432712ff..66a8d75fd4de5bae3ba37cf7fe7b1645938aa855 100644 --- a/paddle/fluid/operators/jit/gen/vbroadcast.cc +++ b/paddle/fluid/operators/jit/gen/vbroadcast.cc @@ -69,7 +69,7 @@ void VBroadcastJitCode::genCode() { class VBroadcastCreator : public JitCodeCreator { public: - bool UseMe(const int64_t& w) const override { + bool CanBeUsed(const int64_t& w) const override { return platform::MayIUse(platform::avx) && w % YMM_FLOAT_BLOCK == 0; } size_t CodeSize(const int64_t& w) const override { diff --git a/paddle/fluid/operators/jit/gen_base.cc b/paddle/fluid/operators/jit/gen_base.cc index f3603875ad7bda1fc688f9c053e0d37f7bb31f02..4c49eff49e3efc0664a084f9fa2bb897db0c6f1d 100644 --- a/paddle/fluid/operators/jit/gen_base.cc +++ b/paddle/fluid/operators/jit/gen_base.cc @@ -31,7 +31,7 @@ namespace paddle { namespace operators { namespace jit { -// refer do not need useme, it would be the last one. +// refer do not need CanBeUsed, it would be the last one. void GenBase::dumpCode(const unsigned char* code) const { if (code) { static int counter = 0; diff --git a/paddle/fluid/operators/jit/gen_base.h b/paddle/fluid/operators/jit/gen_base.h index a7c7a35a7ea35bd80333b04f001d4ab5b5d1e06b..033c603c07c288ba621ceaa912ea0c476fe86cd6 100644 --- a/paddle/fluid/operators/jit/gen_base.h +++ b/paddle/fluid/operators/jit/gen_base.h @@ -31,9 +31,10 @@ class GenBase : public Kernel { virtual ~GenBase() = default; virtual std::string name() const = 0; virtual size_t getSize() const = 0; - virtual const unsigned char* getCodeInternal() = 0; + virtual const unsigned char* getCodeInternal() const = 0; + const char* ImplType() const override { return "JitCode"; } template - Func getCode() { + Func getCode() const { const unsigned char* code = this->getCodeInternal(); if (FLAGS_dump_jitcode) { this->dumpCode(code); @@ -65,7 +66,7 @@ class JitCodeCreator : public GenCreator { virtual ~JitCodeCreator() = default; // condition when this jit code can be used. - virtual bool UseMe(const Attr& attr) const = 0; + virtual bool CanBeUsed(const Attr& attr) const = 0; // estimate this code size virtual size_t CodeSize(const Attr& attr) const = 0; diff --git a/paddle/fluid/operators/jit/helper.h b/paddle/fluid/operators/jit/helper.h index d85c719c1c58c88ec244f1f6ad8343d66391241d..1ac5318d461c2e8bc4f43569602a88f95a76befb 100644 --- a/paddle/fluid/operators/jit/helper.h +++ b/paddle/fluid/operators/jit/helper.h @@ -16,6 +16,8 @@ #include #include +#include +#include // for std::move #include #include "paddle/fluid/operators/jit/gen_base.h" #include "paddle/fluid/operators/jit/kernel_base.h" @@ -27,35 +29,34 @@ namespace paddle { namespace operators { namespace jit { -template +template inline typename std::enable_if< - std::is_same::value && + std::is_same::value && std::is_same::value, - typename KernelTuples::func_type>::type -GetJitCode(const typename KernelTuples::attr_type& attr) { - using Func = typename KernelTuples::func_type; - using Attr = typename KernelTuples::attr_type; - size_t key = JitCodeKey(attr); - auto& codes = JitCodePool().Instance(); + const Kernel*>::type +GetJitCode(const typename KernelTuple::attr_type& attr) { + using Attr = typename KernelTuple::attr_type; + int64_t key = JitCodeKey(attr); + auto& codes = JitCodePool::Instance(); if (codes.Has(key)) { - return codes.AllKernels().at(key)->template getCode(); + return codes.AllKernels().at(key).get(); } // creator is not related with attr, so can use KernelKey as key - KernelKey kkey(KT, PlaceType()); + KernelKey kkey(KernelTuple::kernel_type, PlaceType()); // pool: (KernelKey(type, place), vector) - auto& creator_map = JitCodeCreatorPool().Instance().AllCreators(); + auto& creator_map = JitCodeCreatorPool::Instance().AllCreators(); auto iter = creator_map.find(kkey); if (iter != creator_map.end()) { auto& creators = iter->second; for (auto& cur : creators) { auto i = dynamic_cast*>(cur.get()); - if (i && i->UseMe(attr)) { + if (i && i->CanBeUsed(attr)) { auto p = i->CreateJitCode(attr); if (p) { - auto f = p->template getCode(); + auto res = p.get(); codes.Insert(key, std::move(p)); - return f; + return res; } } } @@ -63,87 +64,153 @@ GetJitCode(const typename KernelTuples::attr_type& attr) { return nullptr; } -template +template inline typename std::enable_if< - !std::is_same::value || + !std::is_same::value || !std::is_same::value, - typename KernelTuples::func_type>::type -GetJitCode(const typename KernelTuples::attr_type& attr) { + const Kernel*>::type +GetJitCode(const typename KernelTuple::attr_type& attr) { return nullptr; } // Refer code do not related with attr, which is just for cast // Refer is always on CPUPlace -template -inline typename KernelTuples::func_type GetRefer() { - auto& ref_pool = ReferKernelPool().Instance().AllKernels(); - KernelKey kkey(KT, platform::CPUPlace()); +template +inline const Kernel* GetReferKernel() { + auto& ref_pool = ReferKernelPool::Instance().AllKernels(); + KernelKey kkey(KernelTuple::kernel_type, platform::CPUPlace()); auto ref_iter = ref_pool.find(kkey); PADDLE_ENFORCE(ref_iter != ref_pool.end(), "Every Kernel should have reference function."); auto& ref_impls = ref_iter->second; for (auto& impl : ref_impls) { - auto i = dynamic_cast*>(impl.get()); + auto i = dynamic_cast*>(impl.get()); if (i) { - return i->GetFunc(); + return i; } } return nullptr; } -template -typename KernelTuples::func_type Get( - const typename KernelTuples::attr_type& attr) { - auto jitfunc = GetJitCode(attr); - if (jitfunc) { - return jitfunc; +template +inline typename KernelTuple::func_type GetReferFunc() { + auto ker = GetReferKernel(); + auto p = dynamic_cast*>(ker); + PADDLE_ENFORCE(p, "The Refer kernel should exsit"); + return p->GetFunc(); +} + +// Return all Kernels that can be used +template +std::vector GetAllCandidateKernels( + const typename KernelTuple::attr_type& attr) { + // the search order shoudl be jitcode > more > refer + std::vector res; + auto jitker = GetJitCode(attr); + if (jitker) { + res.emplace_back(jitker); } - // pool: (KernelKey(type, place), vector) - KernelKey kkey(KT, PlaceType()); - auto& pool = KernelPool().Instance().AllKernels(); + // more kernelpool: (KernelKey(type, place), vector) + KernelKey kkey(KernelTuple::kernel_type, PlaceType()); + auto& pool = KernelPool::Instance().AllKernels(); auto iter = pool.find(kkey); if (iter != pool.end()) { auto& impls = iter->second; for (auto& impl : impls) { - auto i = dynamic_cast*>(impl.get()); - if (i && i->UseMe(attr)) { - return i->GetFunc(); + auto i = dynamic_cast*>(impl.get()); + if (i && i->CanBeUsed(attr)) { + res.emplace_back(i); } } } // The last implementation should be reference function on CPUPlace. - return GetRefer(); + auto ref = GetReferKernel(); + PADDLE_ENFORCE(ref != nullptr, "Refer Kernel can not be empty."); + res.emplace_back(ref); + return res; +} + +template +std::vector> +GetAllCandidateFuncsWithTypes(const typename KernelTuple::attr_type& attr) { + using Func = typename KernelTuple::func_type; + auto kers = GetAllCandidateKernels(attr); + std::vector> res; + for (auto k : kers) { + std::string name = k->ImplType(); + if (name == "JitCode") { + auto i = dynamic_cast(k); + PADDLE_ENFORCE(i, "jitcode kernel cast can not fail."); + res.emplace_back(std::make_pair(name, i->template getCode())); + } else { + auto i = dynamic_cast*>(k); + PADDLE_ENFORCE(i, "kernel cast can not fail."); + res.emplace_back(std::make_pair(name, i->GetFunc())); + } + } + return res; +} + +template +std::vector GetAllCandidateFuncs( + const typename KernelTuple::attr_type& attr) { + auto funcs = GetAllCandidateFuncsWithTypes(attr); + std::vector res; + for (auto& i : funcs) { + res.emplace_back(i.second); + } + return res; +} + +template +typename KernelTuple::func_type GetDefaultBestFunc( + const typename KernelTuple::attr_type& attr) { + auto funcs = GetAllCandidateFuncs(attr); + PADDLE_ENFORCE_GE(funcs.size(), 1UL); + // Here could do some runtime benchmark of this attr and return the best one. + // But yet just get the first one as the default best one, + // which is searched in order and tuned by offline. + return funcs[0]; } -template +template class KernelFuncs { public: KernelFuncs() = default; static KernelFuncs& Cache() { - static thread_local KernelFuncs g_func_cache; + static thread_local KernelFuncs g_func_cache; return g_func_cache; } - bool Has(int key) const { return funcs_.find(key) != funcs_.end(); } - - void Insert(int key, typename KernelTuples::func_type func) { - funcs_.emplace(key, func); - } - - typename KernelTuples::func_type At(int key) { + // the exposed interface to use + typename KernelTuple::func_type At( + const typename KernelTuple::attr_type& attr) { + // Maybe here is not good enough, not all kernels should have jitcode + int64_t key = JitCodeKey(attr); if (Has(key)) { return funcs_.at(key); } - auto func = Get(key); + // If do not have this attr in cache then get the default best + auto func = GetDefaultBestFunc(attr); Insert(key, func); return func; } + typename KernelTuple::func_type operator[]( + const typename KernelTuple::attr_type& attr) { + return At(attr); + } + + protected: + bool Has(int64_t key) const { return funcs_.find(key) != funcs_.end(); } + void Insert(int64_t key, typename KernelTuple::func_type func) { + funcs_.emplace(key, func); + } + private: - std::unordered_map funcs_; + std::unordered_map funcs_; DISABLE_COPY_AND_ASSIGN(KernelFuncs); }; diff --git a/paddle/fluid/operators/jit/kernel_base.h b/paddle/fluid/operators/jit/kernel_base.h index 96e162a21bff2a5624f35ada615c9a9a17ad3c75..bd34d7dfc72a139e70983c56c3220bd01d572bcd 100644 --- a/paddle/fluid/operators/jit/kernel_base.h +++ b/paddle/fluid/operators/jit/kernel_base.h @@ -62,26 +62,55 @@ typedef enum { kSqrt, } SeqPoolType; +// x, y, z, n template -struct XYZNTuples { +struct XYZNTuple { typedef T data_type; typedef int attr_type; typedef void (*func_type)(const T*, const T*, T*, int); }; +// a, x, y, n template -struct AXYNTuples : public XYZNTuples {}; +struct AXYNTuple : public XYZNTuple {}; +// x, y, n template -struct XYNTuples { +struct XYNTuple { typedef T data_type; typedef int attr_type; typedef void (*func_type)(const T*, T*, int); }; -// x, return and int +// x, returned value, n template -struct XRNTuples : public XYNTuples {}; +struct XRNTuple : public XYNTuple {}; + +#define DECLARE_KERNELTUPLE(kernel_tuple, type) \ + template \ + struct type##Tuple : public kernel_tuple { \ + static constexpr KernelType kernel_type = k##type; \ + } + +// Tuple should be corresponding to the KernelType +DECLARE_KERNELTUPLE(XYZNTuple, VMul); +DECLARE_KERNELTUPLE(XYZNTuple, VAdd); +DECLARE_KERNELTUPLE(XYZNTuple, VAddRelu); +DECLARE_KERNELTUPLE(XYZNTuple, VSub); + +DECLARE_KERNELTUPLE(AXYNTuple, VScal); +DECLARE_KERNELTUPLE(AXYNTuple, VAddBias); + +DECLARE_KERNELTUPLE(XYNTuple, VRelu); +DECLARE_KERNELTUPLE(XYNTuple, VIdentity); +DECLARE_KERNELTUPLE(XYNTuple, VSquare); +DECLARE_KERNELTUPLE(XYNTuple, VExp); +DECLARE_KERNELTUPLE(XYNTuple, VSigmoid); +DECLARE_KERNELTUPLE(XYNTuple, VTanh); +DECLARE_KERNELTUPLE(XYNTuple, VCopy); + +DECLARE_KERNELTUPLE(XRNTuple, HMax); +DECLARE_KERNELTUPLE(XRNTuple, HSum); typedef struct { void* gates; // gates: x_ch, x_ih, x_fh, x_oh @@ -122,21 +151,31 @@ typedef struct rnn_attr_s gru_attr_t; typedef struct lstm_attr_s lstm_attr_t; template -struct LSTMTuples { +struct LSTMTuple { typedef T data_type; typedef lstm_attr_t attr_type; typedef void (*func_type)(lstm_t*, const lstm_attr_t*); }; template -struct GRUTuples { +struct GRUTuple { typedef T data_type; typedef gru_attr_t attr_type; typedef void (*func_type)(gru_t*, const gru_attr_t*); }; +DECLARE_KERNELTUPLE(LSTMTuple, LSTMCtHt); +DECLARE_KERNELTUPLE(LSTMTuple, LSTMC1H1); + +DECLARE_KERNELTUPLE(GRUTuple, GRUH1); +DECLARE_KERNELTUPLE(GRUTuple, GRUHtPart1); +DECLARE_KERNELTUPLE(GRUTuple, GRUHtPart2); + +#undef DECLARE_KERNELTUPLE + template -struct VBroadcastTuples { +struct VBroadcastTuple { + static constexpr KernelType kernel_type = kVBroadcast; typedef T data_type; typedef int64_t attr_type; typedef void (*func_type)(const T*, T*, int64_t, int64_t); @@ -151,7 +190,8 @@ typedef struct seq_pool_attr_s { } seq_pool_attr_t; template -struct SeqPoolTuples { +struct SeqPoolTuple { + static constexpr KernelType kernel_type = kSeqPool; typedef T data_type; typedef seq_pool_attr_t attr_type; typedef void (*func_type)(const T*, T*, const seq_pool_attr_t*); @@ -176,7 +216,8 @@ typedef struct emb_seq_pool_attr_s { } emb_seq_pool_attr_t; template -struct EmbSeqPoolTuples { +struct EmbSeqPoolTuple { + static constexpr KernelType kernel_type = kEmbSeqPool; typedef T data_type; typedef emb_seq_pool_attr_t attr_type; typedef void (*func_type)(const T*, const int64_t*, T*, @@ -198,7 +239,8 @@ typedef struct sgd_attr_s { } sgd_attr_t; template -struct SgdTuples { +struct SgdTuple { + static constexpr KernelType kernel_type = kSgd; typedef T data_type; typedef sgd_attr_t attr_type; typedef void (*func_type)(const T*, const T*, const T*, const int64_t*, T*, @@ -214,21 +256,24 @@ typedef struct matmul_attr_s { } matmul_attr_t; template -struct MatMulTuples { +struct MatMulTuple { + static constexpr KernelType kernel_type = kMatMul; typedef T data_type; typedef matmul_attr_t attr_type; typedef void (*func_type)(const T*, const T*, T*, const matmul_attr_t*); }; template -struct CRFDecodingTuples { +struct CRFDecodingTuple { + static constexpr KernelType kernel_type = kCRFDecoding; typedef T data_type; typedef int attr_type; typedef void (*func_type)(const int, const T*, const T*, T*, int*, int); }; template -struct LayerNormTuples { +struct LayerNormTuple { + static constexpr KernelType kernel_type = kLayerNorm; typedef T data_type; typedef int attr_type; typedef void (*func_type)(T*, T*, T*, T*, const T*, const T*, int, @@ -236,7 +281,8 @@ struct LayerNormTuples { }; template -struct SoftmaxTuples { +struct SoftmaxTuple { + static constexpr KernelType kernel_type = kSoftmax; typedef T data_type; typedef int attr_type; typedef void (*func_type)(const T*, T*, int, int); @@ -244,7 +290,8 @@ struct SoftmaxTuples { // nChw16c = nChw16c .* NC template -struct NCHW16CMulNCTuples { +struct NCHW16CMulNCTuple { + static constexpr KernelType kernel_type = kNCHW16CMulNC; typedef T data_type; typedef int attr_type; typedef void (*func_type)(const T*, const T*, T*, int, int); @@ -255,28 +302,29 @@ class Kernel { public: Kernel() = default; virtual ~Kernel() = default; + virtual const char* ImplType() const = 0; DISABLE_COPY_AND_ASSIGN(Kernel); }; -template +template class KernelMore : public Kernel { public: - using T = typename KernelTuples::data_type; - using Func = typename KernelTuples::func_type; - using Attr = typename KernelTuples::attr_type; + using T = typename KernelTuple::data_type; + using Func = typename KernelTuple::func_type; + using Attr = typename KernelTuple::attr_type; virtual Func GetFunc() const { return func; } - virtual bool UseMe(const Attr& attr) const = 0; - virtual const char* ImplType() const = 0; + // specify this kernel can be used, means it should not fail if use it. + virtual bool CanBeUsed(const Attr& attr) const = 0; protected: Func func{nullptr}; }; -template -class ReferKernel : public KernelMore { +template +class ReferKernel : public KernelMore { public: // Refer code can always be used - bool UseMe(const typename KernelTuples::attr_type& attr) const override { + bool CanBeUsed(const typename KernelTuple::attr_type& attr) const override { return true; } const char* ImplType() const override { return "Refer"; } diff --git a/paddle/fluid/operators/jit/kernel_key.cc b/paddle/fluid/operators/jit/kernel_key.cc index 1c2fddcae79d8b89e1169d5bcb364b3ff2e42dd3..1ad220b3972a3d3920610ab8f7ea416892a80d22 100644 --- a/paddle/fluid/operators/jit/kernel_key.cc +++ b/paddle/fluid/operators/jit/kernel_key.cc @@ -13,6 +13,7 @@ * limitations under the License. */ #include "paddle/fluid/operators/jit/kernel_key.h" +#include // XXH64: 13.8 GB/s #include "paddle/fluid/platform/enforce.h" namespace paddle { @@ -20,71 +21,46 @@ namespace operators { namespace jit { template <> -size_t JitCodeKey(const int& d) { +int64_t JitCodeKey(const int& d) { return d; } template <> -size_t JitCodeKey(const int64_t& d) { +int64_t JitCodeKey(const int64_t& d) { return d; } -// TODO(TJ): refine and benchmark JitCodeKey generatation -constexpr int act_type_shift = 3; // suppot 2^3 act types -static inline int act_type_convert(KernelType type) { - if (type == kVIdentity) { - return 0; - } else if (type == kVExp) { - return 1; - } else if (type == kVRelu) { - return 2; - } else if (type == kVSigmoid) { - return 3; - } else if (type == kVTanh) { - return 4; - } - PADDLE_THROW("Unsupported act type %d", type); - return 0; -} - template <> -size_t JitCodeKey(const lstm_attr_t& attr) { - size_t key = attr.d; - int gate_key = act_type_convert(attr.act_gate) << 1; - int cand_key = act_type_convert(attr.act_cand) << (1 + act_type_shift); - int cell_key = act_type_convert(attr.act_cell) << (1 + act_type_shift * 2); - return (key << (1 + act_type_shift * 3)) + gate_key + cand_key + cell_key + - attr.use_peephole; +int64_t JitCodeKey(const gru_attr_t& attr) { + return XXH64(&attr, sizeof(gru_attr_t), 0); } template <> -size_t JitCodeKey(const gru_attr_t& attr) { - size_t key = attr.d; - return (key << (act_type_shift * 2)) + act_type_convert(attr.act_gate) + - (act_type_convert(attr.act_cand) << act_type_shift); +int64_t JitCodeKey(const lstm_attr_t& attr) { + int keys[5] = { + attr.d, static_cast(attr.act_gate), static_cast(attr.act_cand), + static_cast(attr.act_cell), static_cast(attr.use_peephole)}; + return XXH64(keys, sizeof(int) * 5, 0); } template <> -size_t JitCodeKey(const seq_pool_attr_t& attr) { - size_t key = attr.w; - constexpr int pool_type_shift = 3; - return (key << pool_type_shift) + static_cast(attr.type); +int64_t JitCodeKey(const seq_pool_attr_t& attr) { + int keys[2] = {attr.w, static_cast(attr.type)}; + return XXH64(keys, sizeof(int) * 2, 0); } template <> -size_t JitCodeKey(const matmul_attr_t& attr) { - size_t key = attr.m; - constexpr int shift = 21; - return (key << shift * 2) + ((static_cast(attr.n)) << shift) + attr.k; +int64_t JitCodeKey(const matmul_attr_t& attr) { + return XXH64(&attr, sizeof(int) * 3, 0); // m, n, k } template <> -size_t JitCodeKey(const emb_seq_pool_attr_t& attr) { +int64_t JitCodeKey(const emb_seq_pool_attr_t& attr) { return attr.table_width; } template <> -size_t JitCodeKey(const sgd_attr_t& attr) { +int64_t JitCodeKey(const sgd_attr_t& attr) { return attr.grad_width; } diff --git a/paddle/fluid/operators/jit/kernel_key.h b/paddle/fluid/operators/jit/kernel_key.h index 611a0210d614196ad0b05d583303688c1d964e04..b2cf92f23e8ccff5fff7c6e193f7118fbb4765f0 100644 --- a/paddle/fluid/operators/jit/kernel_key.h +++ b/paddle/fluid/operators/jit/kernel_key.h @@ -46,7 +46,7 @@ struct KernelKey { // Every JitCode should have a method to get the key from attribution template -size_t JitCodeKey(const Attr& attr); +int64_t JitCodeKey(const Attr& attr); } // namespace jit } // namespace operators diff --git a/paddle/fluid/operators/jit/kernel_pool.h b/paddle/fluid/operators/jit/kernel_pool.h index 3e15242af28839ee0759e1a5b3930d6d6bfaa0ff..04710a54ac9ddf2ecb8f6a1f2ca33ef158d2d73f 100644 --- a/paddle/fluid/operators/jit/kernel_pool.h +++ b/paddle/fluid/operators/jit/kernel_pool.h @@ -17,6 +17,7 @@ #include // for unique_ptr #include #include +#include // for move #include #include "paddle/fluid/operators/jit/gen_base.h" #include "paddle/fluid/operators/jit/kernel_base.h" @@ -30,7 +31,7 @@ namespace jit { template class JitCodePool { typedef std::unique_ptr GenBasePtr; - typedef std::unordered_map JitCodeMap; + typedef std::unordered_map JitCodeMap; public: JitCodePool() = default; @@ -41,9 +42,9 @@ class JitCodePool { const JitCodeMap& AllKernels() { return codes_; } - bool Has(size_t key) const { return codes_.find(key) != codes_.end(); } + bool Has(int64_t key) const { return codes_.find(key) != codes_.end(); } - void Insert(size_t key, GenBasePtr value) { + void Insert(int64_t key, GenBasePtr value) { codes_.emplace(key, std::move(value)); } diff --git a/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc b/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc index 16c91f8246dda34b1436fd4edd507e9ff603de6b..1254d00189a315b4f743f48e56b3eb53c92ec694 100644 --- a/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc +++ b/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc @@ -161,7 +161,7 @@ void CRFDecoding(const int seq_len, const float* x, const float* w, } } -bool CRFDecodingKernel::UseMe(const int& d) const { +bool CRFDecodingKernel::CanBeUsed(const int& d) const { #ifdef __AVX512F__ constexpr int block = ZMM_FLOAT_BLOCK; #else diff --git a/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.h b/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.h index 24179d90ddcc6e7f44ffa4b2ca0886fbca5c81bf..49b1a1fea4b16f435120bb37c7d9c8c07a4cc4f5 100644 --- a/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.h +++ b/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.h @@ -26,11 +26,11 @@ namespace intrinsic { void CRFDecoding(const int seq_len, const float* x, const float* w, float* alpha, int* track, int tag_num); -class CRFDecodingKernel : public KernelMore> { +class CRFDecodingKernel : public KernelMore> { public: CRFDecodingKernel() { this->func = CRFDecoding; } - bool UseMe( - const typename CRFDecodingTuples::attr_type&) const override; + bool CanBeUsed( + const typename CRFDecodingTuple::attr_type&) const override; const char* ImplType() const override { return "Intrinsic"; } }; diff --git a/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc b/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc index e9b6e401c6825b21191881d4e57fe09b48d2f4ee..a4e3246f10495b67871c08fd8cb7ccd1cf085c9e 100644 --- a/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc +++ b/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc @@ -153,7 +153,7 @@ void LayerNorm(float* x, float* out, float* mean, float* var, } } -bool LayerNormKernel::UseMe(const int& d) const { +bool LayerNormKernel::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx) && d >= YMM_FLOAT_BLOCK; } diff --git a/paddle/fluid/operators/jit/more/intrinsic/layer_norm.h b/paddle/fluid/operators/jit/more/intrinsic/layer_norm.h index 89da2940f4420c418f9bd5260c4b74606cc9168f..7b9f676050d806314edd1e46611416a8b7170add 100644 --- a/paddle/fluid/operators/jit/more/intrinsic/layer_norm.h +++ b/paddle/fluid/operators/jit/more/intrinsic/layer_norm.h @@ -27,10 +27,11 @@ void LayerNorm(float* x, float* out, float* mean, float* var, const float* scale, const float* bias, int height, const float epsilon, int right); -class LayerNormKernel : public KernelMore> { +class LayerNormKernel : public KernelMore> { public: LayerNormKernel() { this->func = LayerNorm; } - bool UseMe(const typename LayerNormTuples::attr_type&) const override; + bool CanBeUsed( + const typename LayerNormTuple::attr_type&) const override; const char* ImplType() const override { return "Intrinsic"; } }; diff --git a/paddle/fluid/operators/jit/more/mix/mix.cc b/paddle/fluid/operators/jit/more/mix/mix.cc index 0036d1c238b17768c4df61af22a85588990e1815..6e709a16d232e2fa1a77e74e228b763fed4dd75b 100644 --- a/paddle/fluid/operators/jit/more/mix/mix.cc +++ b/paddle/fluid/operators/jit/more/mix/mix.cc @@ -23,6 +23,8 @@ namespace jit { namespace more { namespace mix { +using CPUPlace = platform::CPUPlace; + void VSigmoid(const T* x, T* y, int n) { const float min = SIGMOID_THRESHOLD_MIN; const float max = SIGMOID_THRESHOLD_MAX; @@ -30,7 +32,7 @@ void VSigmoid(const T* x, T* y, int n) { y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]); y[i] = static_cast(0) - y[i]; } - auto compute = Get, platform::CPUPlace>(n); + auto compute = KernelFuncs, CPUPlace>::Cache().At(n); compute(y, y, n); for (int i = 0; i < n; ++i) { y[i] = static_cast(1) / (static_cast(1) + y[i]); @@ -39,9 +41,9 @@ void VSigmoid(const T* x, T* y, int n) { void VTanh(const T* x, T* y, int n) { const T a = 2, b = -1; - auto compute_scal = Get, platform::CPUPlace>(n); - auto compute_addbias = Get, platform::CPUPlace>(n); - auto compute_sigmoid = Get, platform::CPUPlace>(n); + auto compute_scal = KernelFuncs, CPUPlace>::Cache().At(n); + auto compute_addbias = KernelFuncs, CPUPlace>::Cache().At(n); + auto compute_sigmoid = KernelFuncs, CPUPlace>::Cache().At(n); compute_scal(&a, x, y, n); compute_sigmoid(y, y, n); compute_scal(&a, y, y, n); @@ -49,16 +51,12 @@ void VTanh(const T* x, T* y, int n) { } void Softmax(const T* x, T* y, int n, int bs) { - auto compute_hmax = - KernelFuncs, platform::CPUPlace>::Cache().At(n); - auto compute_hsum = - KernelFuncs, platform::CPUPlace>::Cache().At(n); - auto compute_vscal = - KernelFuncs, platform::CPUPlace>::Cache().At(n); + auto compute_hmax = KernelFuncs, CPUPlace>::Cache().At(n); + auto compute_hsum = KernelFuncs, CPUPlace>::Cache().At(n); + auto compute_vscal = KernelFuncs, CPUPlace>::Cache().At(n); auto compute_vaddbias = - KernelFuncs, platform::CPUPlace>::Cache().At(n); - auto compute_vexp = - KernelFuncs, platform::CPUPlace>::Cache().At(n); + KernelFuncs, CPUPlace>::Cache().At(n); + auto compute_vexp = KernelFuncs, CPUPlace>::Cache().At(n); for (int i = 0; i < bs; ++i) { T scalar; @@ -76,13 +74,13 @@ void Softmax(const T* x, T* y, int n, int bs) { void (*getActFunc(KernelType type, int d))(const T*, T*, int) { // NOLINT if (type == kVSigmoid) { - return Get, platform::CPUPlace>(d); + return KernelFuncs, CPUPlace>::Cache().At(d); } else if (type == kVRelu) { - return Get, platform::CPUPlace>(d); + return KernelFuncs, CPUPlace>::Cache().At(d); } else if (type == kVTanh) { - return Get, platform::CPUPlace>(d); + return KernelFuncs, CPUPlace>::Cache().At(d); } else if (type == kVIdentity) { - return Get, platform::CPUPlace>(d); + return KernelFuncs, CPUPlace>::Cache().At(d); } PADDLE_THROW("Not support type: %s", type); return nullptr; @@ -98,9 +96,9 @@ void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr) { const int d = attr->d; const int d2 = d * 2; const int d3 = d * 3; - auto vmul_d = Get, platform::CPUPlace>(d); - auto vadd_d = Get, platform::CPUPlace>(d); - auto vadd_d2 = Get, platform::CPUPlace>(d2); + auto vmul_d = KernelFuncs, CPUPlace>::Cache().At(d); + auto vadd_d = KernelFuncs, CPUPlace>::Cache().At(d); + auto vadd_d2 = KernelFuncs, CPUPlace>::Cache().At(d2); auto act_gate_d = getActFunc(attr->act_gate, d); auto act_gate_d2 = getActFunc(attr->act_gate, d2); auto act_gate_d3 = getActFunc(attr->act_gate, d3); @@ -140,8 +138,8 @@ void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr) { int d = attr->d; int d2 = d * 2; int d3 = d * 3; - auto vmul_d = Get, platform::CPUPlace>(d); - auto vadd_d = Get, platform::CPUPlace>(d); + auto vmul_d = KernelFuncs, CPUPlace>::Cache().At(d); + auto vadd_d = KernelFuncs, CPUPlace>::Cache().At(d); auto act_gate_d = getActFunc(attr->act_gate, d); auto act_cand_d = getActFunc(attr->act_cand, d); auto act_cell_d = getActFunc(attr->act_cell, d); @@ -169,7 +167,7 @@ void GRUH1(gru_t* step, const gru_attr_t* attr) { int d2 = d * 2; auto act_gate = getActFunc(attr->act_gate, d); auto act_cand = getActFunc(attr->act_cand, d); - auto vmul_d = Get, platform::CPUPlace>(d); + auto vmul_d = KernelFuncs, CPUPlace>::Cache().At(d); act_gate(gates, gates, d); act_cand(gates + d2, gates + d2, d); vmul_d(gates, gates + d2, ht, d); @@ -182,7 +180,7 @@ void GRUHtPart1(gru_t* step, const gru_attr_t* attr) { T* ht = reinterpret_cast(step->ht); const T* ht_1 = reinterpret_cast(step->ht_1); auto act_gate = getActFunc(attr->act_gate, attr->d); - auto vmul_d = Get, platform::CPUPlace>(attr->d); + auto vmul_d = KernelFuncs, CPUPlace>::Cache().At(attr->d); act_gate(gates + attr->d, gates + attr->d, attr->d); vmul_d(ht_1, gates + attr->d, ht, attr->d); } @@ -206,21 +204,21 @@ void GRUHtPart2(gru_t* step, const gru_attr_t* attr) { } // TODO(TJ): tuning me -bool VSigmoidKernel::UseMe(const int& d) const { return true; } +bool VSigmoidKernel::CanBeUsed(const int& d) const { return true; } -bool VTanhKernel::UseMe(const int& d) const { return true; } +bool VTanhKernel::CanBeUsed(const int& d) const { return true; } -bool SoftmaxKernel::UseMe(const int& d) const { return true; } +bool SoftmaxKernel::CanBeUsed(const int& d) const { return true; } -bool LSTMCtHtKernel::UseMe(const lstm_attr_t& attr) const { return true; } +bool LSTMCtHtKernel::CanBeUsed(const lstm_attr_t& attr) const { return true; } -bool LSTMC1H1Kernel::UseMe(const lstm_attr_t& attr) const { return true; } +bool LSTMC1H1Kernel::CanBeUsed(const lstm_attr_t& attr) const { return true; } -bool GRUH1Kernel::UseMe(const gru_attr_t& attr) const { return true; } +bool GRUH1Kernel::CanBeUsed(const gru_attr_t& attr) const { return true; } -bool GRUHtPart1Kernel::UseMe(const gru_attr_t& attr) const { return true; } +bool GRUHtPart1Kernel::CanBeUsed(const gru_attr_t& attr) const { return true; } -bool GRUHtPart2Kernel::UseMe(const gru_attr_t& attr) const { return true; } +bool GRUHtPart2Kernel::CanBeUsed(const gru_attr_t& attr) const { return true; } } // namespace mix } // namespace more @@ -230,16 +228,16 @@ bool GRUHtPart2Kernel::UseMe(const gru_attr_t& attr) const { return true; } namespace mix = paddle::operators::jit::more::mix; -#define REGISTER_MORE_KERNEL(key, func) \ - REGISTER_JITKERNEL_MORE(key, mix, mix::func##Kernel) - -REGISTER_MORE_KERNEL(kVSigmoid, VSigmoid); -REGISTER_MORE_KERNEL(kVTanh, VTanh); -REGISTER_MORE_KERNEL(kSoftmax, Softmax); -REGISTER_MORE_KERNEL(kLSTMCtHt, LSTMCtHt); -REGISTER_MORE_KERNEL(kLSTMC1H1, LSTMC1H1); -REGISTER_MORE_KERNEL(kGRUH1, GRUH1); -REGISTER_MORE_KERNEL(kGRUHtPart1, GRUHtPart1); -REGISTER_MORE_KERNEL(kGRUHtPart2, GRUHtPart2); +#define REGISTER_MORE_KERNEL(func) \ + REGISTER_JITKERNEL_MORE(k##func, mix, mix::func##Kernel) + +REGISTER_MORE_KERNEL(VSigmoid); +REGISTER_MORE_KERNEL(VTanh); +REGISTER_MORE_KERNEL(Softmax); +REGISTER_MORE_KERNEL(LSTMCtHt); +REGISTER_MORE_KERNEL(LSTMC1H1); +REGISTER_MORE_KERNEL(GRUH1); +REGISTER_MORE_KERNEL(GRUHtPart1); +REGISTER_MORE_KERNEL(GRUHtPart2); #undef REGISTER_MORE_KERNEL diff --git a/paddle/fluid/operators/jit/more/mix/mix.h b/paddle/fluid/operators/jit/more/mix/mix.h index d64af192197a0b339a39a1862c028875da2f3900..994d485909c874a8a15418ad946c79a10265c748 100644 --- a/paddle/fluid/operators/jit/more/mix/mix.h +++ b/paddle/fluid/operators/jit/more/mix/mix.h @@ -34,27 +34,27 @@ void GRUH1(gru_t* step, const gru_attr_t* attr); void GRUHtPart1(gru_t* step, const gru_attr_t* attr); void GRUHtPart2(gru_t* step, const gru_attr_t* attr); -#define DECLARE_MORE_KERNEL(name, tuples) \ - class name##Kernel : public KernelMore> { \ - public: \ - name##Kernel() { this->func = name; } \ - bool UseMe(const typename tuples::attr_type&) const override; \ - const char* ImplType() const override { return "Mixed"; } \ +#define DECLARE_MORE_KERNEL(name) \ + class name##Kernel : public KernelMore> { \ + public: \ + name##Kernel() { this->func = name; } \ + bool CanBeUsed(const typename name##Tuple::attr_type&) const override; \ + const char* ImplType() const override { return "Mixed"; } \ } // XYN -DECLARE_MORE_KERNEL(VSigmoid, XYNTuples); -DECLARE_MORE_KERNEL(VTanh, XYNTuples); +DECLARE_MORE_KERNEL(VSigmoid); +DECLARE_MORE_KERNEL(VTanh); // XRN -DECLARE_MORE_KERNEL(Softmax, SoftmaxTuples); +DECLARE_MORE_KERNEL(Softmax); -DECLARE_MORE_KERNEL(LSTMCtHt, LSTMTuples); -DECLARE_MORE_KERNEL(LSTMC1H1, LSTMTuples); +DECLARE_MORE_KERNEL(LSTMCtHt); +DECLARE_MORE_KERNEL(LSTMC1H1); -DECLARE_MORE_KERNEL(GRUH1, GRUTuples); -DECLARE_MORE_KERNEL(GRUHtPart1, GRUTuples); -DECLARE_MORE_KERNEL(GRUHtPart2, GRUTuples); +DECLARE_MORE_KERNEL(GRUH1); +DECLARE_MORE_KERNEL(GRUHtPart1); +DECLARE_MORE_KERNEL(GRUHtPart2); #undef DECLARE_MORE_KERNEL diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.cc b/paddle/fluid/operators/jit/more/mkl/mkl.cc index 4f51353bce834325e6c659399a374e4fbc40d4b7..4f600b38144f53798e3d4c66264fc5bfa671a4f7 100644 --- a/paddle/fluid/operators/jit/more/mkl/mkl.cc +++ b/paddle/fluid/operators/jit/more/mkl/mkl.cc @@ -130,105 +130,106 @@ void ASum(const double* x, double* res, int n) { // TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512 template <> -bool VMulKernel::UseMe(const int& d) const { +bool VMulKernel::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx512f) && d > 512; } template <> -bool VAddKernel::UseMe(const int& d) const { +bool VAddKernel::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx) && d > 512; } template <> -bool VScalKernel::UseMe(const int& d) const { +bool VScalKernel::CanBeUsed(const int& d) const { return platform::MayIUse(platform::avx512f) && d > 512; } template <> -bool VExpKernel::UseMe(const int& d) const { +bool VExpKernel::CanBeUsed(const int& d) const { return d > 7; } template <> -bool VSquareKernel::UseMe(const int& d) const { +bool VSquareKernel::CanBeUsed(const int& d) const { return d > 7; } template <> -bool VCopyKernel::UseMe(const int& d) const { +bool VCopyKernel::CanBeUsed(const int& d) const { return d > 15; } template <> -bool VBroadcastKernel::UseMe(const int64_t& d) const { +bool VBroadcastKernel::CanBeUsed(const int64_t& d) const { return d > 127; } template <> -bool VBroadcastKernel::UseMe(const int64_t& attr) const { +bool VBroadcastKernel::CanBeUsed(const int64_t& attr) const { return true; } template <> -bool VSigmoidKernel::UseMe(const int& d) const { +bool VSigmoidKernel::CanBeUsed(const int& d) const { return d > 7; } template <> -bool VTanhKernel::UseMe(const int& d) const { +bool VTanhKernel::CanBeUsed(const int& d) const { return d > 7; } template <> -bool SeqPoolKernel::UseMe(const seq_pool_attr_t& attr) const { +bool SeqPoolKernel::CanBeUsed(const seq_pool_attr_t& attr) const { return true; } template <> -bool SeqPoolKernel::UseMe(const seq_pool_attr_t& attr) const { +bool SeqPoolKernel::CanBeUsed(const seq_pool_attr_t& attr) const { return true; } template <> -bool EmbSeqPoolKernel::UseMe(const emb_seq_pool_attr_t& attr) const { +bool EmbSeqPoolKernel::CanBeUsed(const emb_seq_pool_attr_t& attr) const { return true; } template <> -bool EmbSeqPoolKernel::UseMe(const emb_seq_pool_attr_t& attr) const { +bool EmbSeqPoolKernel::CanBeUsed( + const emb_seq_pool_attr_t& attr) const { return true; } template <> -bool SgdKernel::UseMe(const sgd_attr_t& attr) const { +bool SgdKernel::CanBeUsed(const sgd_attr_t& attr) const { return true; } template <> -bool SgdKernel::UseMe(const sgd_attr_t& attr) const { +bool SgdKernel::CanBeUsed(const sgd_attr_t& attr) const { return true; } template <> -bool MatMulKernel::UseMe(const matmul_attr_t& attr) const { +bool MatMulKernel::CanBeUsed(const matmul_attr_t& attr) const { return platform::MayIUse(platform::avx); } template <> -bool MatMulKernel::UseMe(const matmul_attr_t& attr) const { +bool MatMulKernel::CanBeUsed(const matmul_attr_t& attr) const { return true; } template <> -bool SoftmaxKernel::UseMe(const int& d) const { +bool SoftmaxKernel::CanBeUsed(const int& d) const { // tuned on avx2 return platform::MayIUse(platform::avx) && d < 60; } -#define AWALYS_USE_ME_WITH_DOUBLE(func) \ - template <> \ - bool func##Kernel::UseMe(const int& d) const { \ - return true; \ +#define AWALYS_USE_ME_WITH_DOUBLE(func) \ + template <> \ + bool func##Kernel::CanBeUsed(const int& d) const { \ + return true; \ } AWALYS_USE_ME_WITH_DOUBLE(VMul); @@ -250,23 +251,23 @@ AWALYS_USE_ME_WITH_DOUBLE(Softmax); namespace mkl = paddle::operators::jit::more::mkl; -#define REGISTER_MKL_KERNEL(key, func) \ - REGISTER_JITKERNEL_MORE(key, mkl, mkl::func##Kernel, \ +#define REGISTER_MKL_KERNEL(func) \ + REGISTER_JITKERNEL_MORE(k##func, mkl, mkl::func##Kernel, \ mkl::func##Kernel) -REGISTER_MKL_KERNEL(kMatMul, MatMul); -REGISTER_MKL_KERNEL(kVMul, VMul); -REGISTER_MKL_KERNEL(kVAdd, VAdd); -REGISTER_MKL_KERNEL(kVScal, VScal); -REGISTER_MKL_KERNEL(kVExp, VExp); -REGISTER_MKL_KERNEL(kVSquare, VSquare); -REGISTER_MKL_KERNEL(kVCopy, VCopy); -REGISTER_MKL_KERNEL(kVBroadcast, VBroadcast); -REGISTER_MKL_KERNEL(kVSigmoid, VSigmoid); -REGISTER_MKL_KERNEL(kVTanh, VTanh); -REGISTER_MKL_KERNEL(kSeqPool, SeqPool); -REGISTER_MKL_KERNEL(kEmbSeqPool, EmbSeqPool); -REGISTER_MKL_KERNEL(kSoftmax, Softmax); -REGISTER_MKL_KERNEL(kSgd, Sgd); +REGISTER_MKL_KERNEL(MatMul); +REGISTER_MKL_KERNEL(VMul); +REGISTER_MKL_KERNEL(VAdd); +REGISTER_MKL_KERNEL(VScal); +REGISTER_MKL_KERNEL(VExp); +REGISTER_MKL_KERNEL(VSquare); +REGISTER_MKL_KERNEL(VCopy); +REGISTER_MKL_KERNEL(VBroadcast); +REGISTER_MKL_KERNEL(VSigmoid); +REGISTER_MKL_KERNEL(VTanh); +REGISTER_MKL_KERNEL(SeqPool); +REGISTER_MKL_KERNEL(EmbSeqPool); +REGISTER_MKL_KERNEL(Softmax); +REGISTER_MKL_KERNEL(Sgd); #undef REGISTER_MKL_KERNEL diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.h b/paddle/fluid/operators/jit/more/mkl/mkl.h index db2d6faed4fdcfebedb9d9eb752831259af30186..f51dca654cd3d93dcd396af7895aebf5ee915c22 100644 --- a/paddle/fluid/operators/jit/more/mkl/mkl.h +++ b/paddle/fluid/operators/jit/more/mkl/mkl.h @@ -175,41 +175,38 @@ void Sgd(const T* lr, const T* param, const T* grad, const int64_t* rows, } } -#define DECLARE_MKL_KERNEL(name, tuples) \ - template \ - class name##Kernel : public KernelMore> { \ - public: \ - name##Kernel() { this->func = name; } \ - bool UseMe(const typename tuples::attr_type&) const override; \ - const char* ImplType() const override { return "MKL"; } \ +#define DECLARE_MKL_KERNEL(name) \ + template \ + class name##Kernel : public KernelMore> { \ + public: \ + name##Kernel() { this->func = name; } \ + bool CanBeUsed(const typename name##Tuple::attr_type&) const override; \ + const char* ImplType() const override { return "MKL"; } \ } // ABCMNK -DECLARE_MKL_KERNEL(MatMul, MatMulTuples); +DECLARE_MKL_KERNEL(MatMul); // XYZN -DECLARE_MKL_KERNEL(VMul, XYZNTuples); -DECLARE_MKL_KERNEL(VAdd, XYZNTuples); +DECLARE_MKL_KERNEL(VMul); +DECLARE_MKL_KERNEL(VAdd); // AXYN -DECLARE_MKL_KERNEL(VScal, AXYNTuples); +DECLARE_MKL_KERNEL(VScal); // XYN -DECLARE_MKL_KERNEL(VExp, XYNTuples); -DECLARE_MKL_KERNEL(VSigmoid, XYNTuples); -DECLARE_MKL_KERNEL(VTanh, XYNTuples); -DECLARE_MKL_KERNEL(VSquare, XYNTuples); -DECLARE_MKL_KERNEL(VCopy, XYNTuples); - -DECLARE_MKL_KERNEL(SeqPool, SeqPoolTuples); - -DECLARE_MKL_KERNEL(EmbSeqPool, EmbSeqPoolTuples); - -DECLARE_MKL_KERNEL(Softmax, SoftmaxTuples); - -DECLARE_MKL_KERNEL(Sgd, SgdTuples); - -DECLARE_MKL_KERNEL(VBroadcast, VBroadcastTuples); +DECLARE_MKL_KERNEL(VExp); +DECLARE_MKL_KERNEL(VSigmoid); +DECLARE_MKL_KERNEL(VTanh); +DECLARE_MKL_KERNEL(VSquare); +DECLARE_MKL_KERNEL(VCopy); + +// others +DECLARE_MKL_KERNEL(SeqPool); +DECLARE_MKL_KERNEL(EmbSeqPool); +DECLARE_MKL_KERNEL(Softmax); +DECLARE_MKL_KERNEL(Sgd); +DECLARE_MKL_KERNEL(VBroadcast); #undef DECLARE_MKL_KERNEL diff --git a/paddle/fluid/operators/jit/refer/refer.cc b/paddle/fluid/operators/jit/refer/refer.cc index c279d1b2ca4f53bb6bc5da0cab41e9086ed475bd..0d1c4770903fc59160e308b958270e5826928d61 100644 --- a/paddle/fluid/operators/jit/refer/refer.cc +++ b/paddle/fluid/operators/jit/refer/refer.cc @@ -17,51 +17,43 @@ namespace refer = paddle::operators::jit::refer; -#define REGISTER_REFER_KERNEL(key, func) \ - REGISTER_JITKERNEL_REFER(key, refer::func##Kernel, \ +#define REGISTER_REFER_KERNEL(func) \ + REGISTER_JITKERNEL_REFER(k##func, refer::func##Kernel, \ refer::func##Kernel) -REGISTER_REFER_KERNEL(kVMul, VMul); -REGISTER_REFER_KERNEL(kVAdd, VAdd); -REGISTER_REFER_KERNEL(kVAddRelu, VAddRelu); -REGISTER_REFER_KERNEL(kVSub, VSub); - -REGISTER_REFER_KERNEL(kVScal, VScal); -REGISTER_REFER_KERNEL(kVAddBias, VAddBias); - -REGISTER_REFER_KERNEL(kVRelu, VRelu); -REGISTER_REFER_KERNEL(kVCopy, VCopy); -REGISTER_REFER_KERNEL(kVIdentity, VIdentity); -REGISTER_REFER_KERNEL(kVSquare, VSquare); -REGISTER_REFER_KERNEL(kVExp, VExp); -REGISTER_REFER_KERNEL(kVSigmoid, VSigmoid); -REGISTER_REFER_KERNEL(kVTanh, VTanh); - -REGISTER_REFER_KERNEL(kLSTMCtHt, LSTMCtHt); -REGISTER_REFER_KERNEL(kLSTMC1H1, LSTMC1H1); - -REGISTER_REFER_KERNEL(kGRUH1, GRUH1); -REGISTER_REFER_KERNEL(kGRUHtPart1, GRUHtPart1); -REGISTER_REFER_KERNEL(kGRUHtPart2, GRUHtPart2); - -REGISTER_REFER_KERNEL(kCRFDecoding, CRFDecoding); -REGISTER_REFER_KERNEL(kLayerNorm, LayerNorm); - -REGISTER_REFER_KERNEL(kNCHW16CMulNC, NCHW16CMulNC); - -REGISTER_REFER_KERNEL(kSeqPool, SeqPool); - -REGISTER_REFER_KERNEL(kMatMul, MatMul); - -REGISTER_REFER_KERNEL(kHMax, HMax); -REGISTER_REFER_KERNEL(kHSum, HSum); - -REGISTER_REFER_KERNEL(kSoftmax, Softmax); - -REGISTER_REFER_KERNEL(kEmbSeqPool, EmbSeqPool); - -REGISTER_REFER_KERNEL(kSgd, Sgd); - -REGISTER_REFER_KERNEL(kVBroadcast, VBroadcast); +REGISTER_REFER_KERNEL(VMul); +REGISTER_REFER_KERNEL(VAdd); +REGISTER_REFER_KERNEL(VAddRelu); +REGISTER_REFER_KERNEL(VSub); + +REGISTER_REFER_KERNEL(VScal); +REGISTER_REFER_KERNEL(VAddBias); + +REGISTER_REFER_KERNEL(VRelu); +REGISTER_REFER_KERNEL(VCopy); +REGISTER_REFER_KERNEL(VIdentity); +REGISTER_REFER_KERNEL(VSquare); +REGISTER_REFER_KERNEL(VExp); +REGISTER_REFER_KERNEL(VSigmoid); +REGISTER_REFER_KERNEL(VTanh); + +REGISTER_REFER_KERNEL(LSTMCtHt); +REGISTER_REFER_KERNEL(LSTMC1H1); + +REGISTER_REFER_KERNEL(GRUH1); +REGISTER_REFER_KERNEL(GRUHtPart1); +REGISTER_REFER_KERNEL(GRUHtPart2); + +REGISTER_REFER_KERNEL(CRFDecoding); +REGISTER_REFER_KERNEL(LayerNorm); +REGISTER_REFER_KERNEL(NCHW16CMulNC); +REGISTER_REFER_KERNEL(SeqPool); +REGISTER_REFER_KERNEL(MatMul); +REGISTER_REFER_KERNEL(HMax); +REGISTER_REFER_KERNEL(HSum); +REGISTER_REFER_KERNEL(Softmax); +REGISTER_REFER_KERNEL(EmbSeqPool); +REGISTER_REFER_KERNEL(Sgd); +REGISTER_REFER_KERNEL(VBroadcast); #undef REGISTER_REFER_KERNEL diff --git a/paddle/fluid/operators/jit/refer/refer.h b/paddle/fluid/operators/jit/refer/refer.h index b3b2097828c5b6d647fd6bfe14a6e8bff04409e0..cac705a484127b4813ef2d0996bf5aaee2b9f1b3 100644 --- a/paddle/fluid/operators/jit/refer/refer.h +++ b/paddle/fluid/operators/jit/refer/refer.h @@ -490,60 +490,54 @@ void Sgd(const T* lr, const T* param, const T* grad, const int64_t* rows, } } -#define DECLARE_REFER_KERNEL(name, tuples) \ - template \ - class name##Kernel : public ReferKernel> { \ - public: \ - name##Kernel() { this->func = name; } \ +#define DECLARE_REFER_KERNEL(name) \ + template \ + class name##Kernel : public ReferKernel> { \ + public: \ + name##Kernel() { this->func = name; } \ } // const T* x, const T* y, T* z, int n -DECLARE_REFER_KERNEL(VMul, XYZNTuples); -DECLARE_REFER_KERNEL(VAdd, XYZNTuples); -DECLARE_REFER_KERNEL(VAddRelu, XYZNTuples); -DECLARE_REFER_KERNEL(VSub, XYZNTuples); +DECLARE_REFER_KERNEL(VMul); +DECLARE_REFER_KERNEL(VAdd); +DECLARE_REFER_KERNEL(VAddRelu); +DECLARE_REFER_KERNEL(VSub); // const T* a, const T* x, T* y, int n -DECLARE_REFER_KERNEL(VScal, AXYNTuples); -DECLARE_REFER_KERNEL(VAddBias, AXYNTuples); +DECLARE_REFER_KERNEL(VScal); +DECLARE_REFER_KERNEL(VAddBias); // const T* x, T* y, int n -DECLARE_REFER_KERNEL(VRelu, XYNTuples); -DECLARE_REFER_KERNEL(VIdentity, XYNTuples); -DECLARE_REFER_KERNEL(VExp, XYNTuples); -DECLARE_REFER_KERNEL(VSigmoid, XYNTuples); -DECLARE_REFER_KERNEL(VTanh, XYNTuples); -DECLARE_REFER_KERNEL(VSquare, XYNTuples); -DECLARE_REFER_KERNEL(VCopy, XYNTuples); +DECLARE_REFER_KERNEL(VRelu); +DECLARE_REFER_KERNEL(VIdentity); +DECLARE_REFER_KERNEL(VExp); +DECLARE_REFER_KERNEL(VSigmoid); +DECLARE_REFER_KERNEL(VTanh); +DECLARE_REFER_KERNEL(VSquare); +DECLARE_REFER_KERNEL(VCopy); // lstm_t*, const lstm_attr_t* -DECLARE_REFER_KERNEL(LSTMCtHt, LSTMTuples); -DECLARE_REFER_KERNEL(LSTMC1H1, LSTMTuples); +DECLARE_REFER_KERNEL(LSTMCtHt); +DECLARE_REFER_KERNEL(LSTMC1H1); // gru_t*, const gru_attr_t* -DECLARE_REFER_KERNEL(GRUH1, GRUTuples); -DECLARE_REFER_KERNEL(GRUHtPart1, GRUTuples); -DECLARE_REFER_KERNEL(GRUHtPart2, GRUTuples); - -DECLARE_REFER_KERNEL(CRFDecoding, CRFDecodingTuples); -DECLARE_REFER_KERNEL(LayerNorm, LayerNormTuples); - -DECLARE_REFER_KERNEL(NCHW16CMulNC, NCHW16CMulNCTuples); - -DECLARE_REFER_KERNEL(SeqPool, SeqPoolTuples); - -DECLARE_REFER_KERNEL(MatMul, MatMulTuples); - -DECLARE_REFER_KERNEL(HMax, XRNTuples); -DECLARE_REFER_KERNEL(HSum, XRNTuples); - -DECLARE_REFER_KERNEL(Softmax, SoftmaxTuples); - -DECLARE_REFER_KERNEL(EmbSeqPool, EmbSeqPoolTuples); - -DECLARE_REFER_KERNEL(Sgd, SgdTuples); - -DECLARE_REFER_KERNEL(VBroadcast, VBroadcastTuples); +DECLARE_REFER_KERNEL(GRUH1); +DECLARE_REFER_KERNEL(GRUHtPart1); +DECLARE_REFER_KERNEL(GRUHtPart2); + +DECLARE_REFER_KERNEL(HMax); +DECLARE_REFER_KERNEL(HSum); + +// others +DECLARE_REFER_KERNEL(CRFDecoding); +DECLARE_REFER_KERNEL(LayerNorm); +DECLARE_REFER_KERNEL(NCHW16CMulNC); +DECLARE_REFER_KERNEL(SeqPool); +DECLARE_REFER_KERNEL(MatMul); +DECLARE_REFER_KERNEL(Softmax); +DECLARE_REFER_KERNEL(EmbSeqPool); +DECLARE_REFER_KERNEL(Sgd); +DECLARE_REFER_KERNEL(VBroadcast); #undef DECLARE_REFER_KERNEL diff --git a/paddle/fluid/operators/jit/registry.h b/paddle/fluid/operators/jit/registry.h index cb32c487208fe8fe9e72c069db8833c736316aec..567a903236979ff4ac6095033f53d2a473f4eb2c 100644 --- a/paddle/fluid/operators/jit/registry.h +++ b/paddle/fluid/operators/jit/registry.h @@ -17,6 +17,7 @@ #include #include #include +#include // for std::move #include "paddle/fluid/operators/jit/kernel_base.h" #include "paddle/fluid/operators/jit/kernel_pool.h" #include "paddle/fluid/platform/place.h" @@ -49,8 +50,8 @@ struct JitKernelRegistrarFunctor { void operator()(KernelType kt) const { KernelKey kkey(kt, PlaceType()); - Pool().Instance().Insert(kkey, - std::move(make_unique())); + Pool::Instance().Insert(kkey, + std::move(make_unique())); constexpr auto size = std::tuple_size>::value; JitKernelRegistrarFunctor diff --git a/paddle/fluid/operators/jit/test.cc b/paddle/fluid/operators/jit/test.cc index cdec14dc4383897f4ae24fc89b99fe00c713cf42..6c099a7a062472e2701401ddc58bb9051074f810 100644 --- a/paddle/fluid/operators/jit/test.cc +++ b/paddle/fluid/operators/jit/test.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include #include #include #include @@ -64,413 +65,23 @@ std::vector TestSizes() { namespace jit = paddle::operators::jit; using CPUPlace = paddle::platform::CPUPlace; -template -struct TestFuncWithRefer { - void operator()(const typename KernelTuples::func_type tgt, Args... args) { - LOG(FATAL) << "Should specify this function."; - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector, - std::vector> { - void operator()(const typename jit::XYZNTuples::func_type tgt, - const std::vector& x, const std::vector& y, - const std::vector& zref) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(zref.size(), x.size()); - EXPECT_EQ(zref.size(), y.size()); - const T* x_data = x.data(); - const T* y_data = y.data(); - const T* zref_data = zref.data(); - const int d = zref.size(); - - std::vector ztgt(d); - T* ztgt_data = ztgt.data(); - // test normal - tgt(x_data, y_data, ztgt_data, d); - ExpectEQ(ztgt_data, zref_data, d); - // test inplace x - std::copy(x.begin(), x.end(), ztgt.begin()); - tgt(ztgt_data, y_data, ztgt_data, d); - ExpectEQ(ztgt_data, zref_data, d); - // test inplace y - std::copy(y.begin(), y.end(), ztgt.begin()); - tgt(x_data, ztgt_data, ztgt_data, d); - ExpectEQ(ztgt_data, zref_data, d); - } -}; - -template -struct TestFuncWithRefer, T, std::vector, - std::vector> { - void operator()(const typename jit::AXYNTuples::func_type tgt, const T a, - const std::vector& x, const std::vector& yref) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(yref.size(), x.size()); - const T* x_data = x.data(); - const T* yref_data = yref.data(); - const int d = yref.size(); - std::vector ytgt(d); - T* ytgt_data = ytgt.data(); - // test normal - tgt(&a, x_data, ytgt_data, d); - ExpectEQ(ytgt_data, yref_data, d); - // test inplace x - std::copy(x.begin(), x.end(), ytgt.begin()); - tgt(&a, ytgt_data, ytgt_data, d); - ExpectEQ(ytgt_data, yref_data, d); - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector, - int, int> { - void operator()(const typename jit::SoftmaxTuples::func_type tgt, - const std::vector& x, const std::vector& yref, int n, - int bs) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(yref.size(), x.size()); - EXPECT_EQ(x.size(), static_cast(n * bs)); - const T* x_data = x.data(); - const T* yref_data = yref.data(); - std::vector ytgt(n * bs); - T* ytgt_data = ytgt.data(); - // test normal - tgt(x_data, ytgt_data, n, bs); - ExpectEQ(ytgt_data, yref_data, n * bs); - // test inplace x - std::copy(x.begin(), x.end(), ytgt.begin()); - tgt(ytgt_data, ytgt_data, n, bs); - ExpectEQ(ytgt_data, yref_data, n * bs); - } -}; - -template -struct TestFuncWithRefer, std::vector, T> { - void operator()(const typename jit::XRNTuples::func_type tgt, - const std::vector& x, const T ref_res) { - EXPECT_TRUE(tgt != nullptr); - T tgt_res; - tgt(x.data(), &tgt_res, x.size()); - ExpectEQ(&tgt_res, &ref_res, 1); - } -}; - -template -struct TestFuncWithRefer, std::vector, - std::vector, int64_t, - typename jit::VBroadcastTuples::attr_type> { - void operator()(const typename jit::VBroadcastTuples::func_type tgt, - const std::vector& x, const std::vector& yref, - int64_t h, - const typename jit::VBroadcastTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(x.size(), static_cast(attr)); - EXPECT_EQ(yref.size(), x.size() * h); - std::vector y(yref.size()); - const T* x_data = x.data(); - const T* yref_data = yref.data(); - T* y_data = y.data(); - tgt(x_data, y_data, h, attr); - ExpectEQ(y_data, yref_data, yref.size()); - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector> { - void operator()(const typename jit::XYNTuples::func_type tgt, - const std::vector& x, const std::vector& yref) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(yref.size(), x.size()); - const T* x_data = x.data(); - const T* yref_data = yref.data(); - const int d = yref.size(); - std::vector ytgt(d); - T* ytgt_data = ytgt.data(); - // test normal - tgt(x_data, ytgt_data, d); - ExpectEQ(ytgt_data, yref_data, d); - // test inplace x - std::copy(x.begin(), x.end(), ytgt.begin()); - tgt(ytgt_data, ytgt_data, d); - ExpectEQ(ytgt_data, yref_data, d); - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector, - std::vector, std::vector, std::vector, - typename jit::LSTMTuples::attr_type> { - void operator()(const typename jit::LSTMTuples::func_type tgt, - const std::vector& xsrc, const std::vector& wp, - const std::vector& ct_1, const std::vector& ct_ref, - const std::vector& ht_ref, - const typename jit::LSTMTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(ct_ref.size(), ht_ref.size()); - EXPECT_EQ(ct_1.size(), ht_ref.size()); - EXPECT_EQ(xsrc.size(), 4 * ht_ref.size()); - EXPECT_EQ(wp.size(), 3 * ht_ref.size()); - - // x could be changed after compute, so copy to save src - int d = ht_ref.size(); - std::vector x(xsrc.size()), ct(ct_ref.size()), ht(ht_ref.size()); - std::vector checked(2 * d); - std::copy(xsrc.begin(), xsrc.end(), x.begin()); - - const T* ct_1_data = ct_1.data(); - const T* wp_data = wp.data(); - const T* ct_ref_data = ct_ref.data(); - const T* ht_ref_data = ht_ref.data(); - T* x_data = x.data(); - T* ct_data = ct.data(); - T* ht_data = ht.data(); - T* checked_data = checked.data(); - - jit::lstm_t step; - step.gates = x_data; - step.ct_1 = ct_1_data; - step.ct = ct_data; - step.ht = ht_data; - if (attr.use_peephole) { - step.wp = wp_data; - step.checked = checked_data; - } - - tgt(&step, &attr); - ExpectEQ(ct_data, ct_ref_data, d); - ExpectEQ(ht_data, ht_ref_data, d); - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector, - std::vector, - typename jit::GRUTuples::attr_type> { - void operator()(const typename jit::GRUTuples::func_type tgt, - const std::vector& xsrc, const std::vector& ht_1, - const std::vector& ht_ref, - const typename jit::GRUTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(ht_1.size(), ht_ref.size()); - EXPECT_EQ(xsrc.size(), 3 * ht_ref.size()); - - // x could be changed after compute, so copy to save src - int d = ht_ref.size(); - std::vector x(xsrc.size()), ht(ht_ref.size()); - std::copy(xsrc.begin(), xsrc.end(), x.begin()); - const T* ht_1_data = ht_1.data(); - const T* ht_ref_data = ht_ref.data(); - T* x_data = x.data(); - T* ht_data = ht.data(); - jit::gru_t step; - step.gates = x_data; - step.ht_1 = ht_1_data; - step.ht = ht_data; - tgt(&step, &attr); - ExpectEQ(ht_data, ht_ref_data, d); - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector, - typename jit::SeqPoolTuples::attr_type> { - void operator()(const typename jit::SeqPoolTuples::func_type tgt, - const std::vector& x, const std::vector& yref, - const typename jit::SeqPoolTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(x.size() % yref.size(), static_cast(0)); - int w = yref.size(); - std::vector y(w); - const T* x_data = x.data(); - const T* yref_data = yref.data(); - T* y_data = y.data(); - tgt(x_data, y_data, &attr); - ExpectEQ(y_data, yref_data, w); - } -}; - -template -struct TestFuncWithRefer, std::vector, - std::vector, std::vector, - typename jit::EmbSeqPoolTuples::attr_type> { - void operator()(const typename jit::EmbSeqPoolTuples::func_type tgt, - const std::vector& table, const std::vector& idx, - const std::vector& oref, - const typename jit::EmbSeqPoolTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(table.size(), - static_cast(attr.table_height * attr.table_width)); - EXPECT_EQ(idx.size(), - static_cast(attr.index_height * attr.index_width)); - EXPECT_EQ(oref.size(), - static_cast(attr.table_width * attr.index_width)); - const T* table_data = table.data(); - const int64_t* idx_data = idx.data(); - const T* oref_data = oref.data(); - int o_w = oref.size(); - std::vector out(o_w); - T* o_data = out.data(); - tgt(table_data, idx_data, o_data, &attr); - ExpectEQ(o_data, oref_data, o_w); - } -}; - -template -struct TestFuncWithRefer, T, std::vector, std::vector, - std::vector, std::vector, - typename jit::SgdTuples::attr_type> { - void operator()(const typename jit::SgdTuples::func_type tgt, const T lr, - const std::vector& param, const std::vector& grad, - const std::vector& rows, const std::vector& oref, - const typename jit::SgdTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(param.size(), - static_cast(attr.param_height * attr.param_width)); - EXPECT_EQ(grad.size(), - static_cast(attr.grad_height * attr.grad_width)); - EXPECT_EQ(rows.size(), static_cast(attr.selected_rows_size)); - EXPECT_EQ(param.size(), oref.size()); - const T* param_data = param.data(); - const T* grad_data = grad.data(); - const int64_t* rows_data = rows.data(); - const T* oref_data = oref.data(); - - std::vector out(oref.size()); - T* o_data = out.data(); - tgt(&lr, param_data, grad_data, rows_data, o_data, &attr); - // only the selected rows should be equal - for (size_t i = 0; i < rows.size(); ++i) { - ExpectEQ(o_data + rows[i] * attr.grad_width, - oref_data + rows[i] * attr.grad_width, attr.grad_width); - } - - // inplace - std::copy(param.begin(), param.end(), out.begin()); - tgt(&lr, o_data, grad_data, rows_data, o_data, &attr); - for (size_t i = 0; i < rows.size(); ++i) { - ExpectEQ(o_data + rows[i] * attr.grad_width, - oref_data + rows[i] * attr.grad_width, attr.grad_width); - } - } -}; - -template -struct TestFuncWithRefer, std::vector, std::vector, - std::vector, - typename jit::MatMulTuples::attr_type> { - void operator()(const typename jit::MatMulTuples::func_type tgt, - const std::vector& a, const std::vector& b, - const std::vector& cref, - const typename jit::MatMulTuples::attr_type& attr) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(a.size(), static_cast(attr.m * attr.k)); - EXPECT_EQ(b.size(), static_cast(attr.k * attr.n)); - EXPECT_EQ(cref.size(), static_cast(attr.m * attr.n)); - std::vector c(cref.size()); - const T* a_data = a.data(); - const T* b_data = b.data(); - const T* cref_data = cref.data(); - T* c_data = c.data(); - tgt(a_data, b_data, c_data, &attr); - ExpectEQ(c_data, cref_data, attr.m * attr.n); - } -}; - -template -struct TestFuncWithRefer, std::vector, - std::vector, std::vector, std::vector, - std::vector, std::vector, int, float, int> { - void operator()(const typename jit::LayerNormTuples::func_type tgt, - std::vector& x, std::vector& outref, // NOLINT - std::vector& mean, std::vector& var, // NOLINT - const std::vector& scale, const std::vector& bias, - int left, const float epsilon, int right) { - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(x.size(), static_cast(left * right)); - EXPECT_EQ(outref.size(), static_cast(left * right)); - EXPECT_EQ(mean.size(), static_cast(left)); - EXPECT_EQ(var.size(), static_cast(left)); - EXPECT_EQ(scale.size(), static_cast(right)); - EXPECT_EQ(bias.size(), static_cast(right)); - std::vector outtgt(outref.size()); - const T* scale_data = scale.data(); - const T* bias_data = bias.data(); - T* x_data = x.data(); - T* mean_data = mean.data(); - T* var_data = var.data(); - T* outref_data = outref.data(); - T* outtgt_data = outtgt.data(); - - tgt(x_data, outtgt_data, mean_data, var_data, scale_data, bias_data, left, - epsilon, right); - ExpectEQ(outtgt_data, outref_data, left * right); - } -}; - -template -struct TestFuncWithRefer, int, std::vector, - std::vector, std::vector, std::vector, - int> { - void operator()(const typename jit::CRFDecodingTuples::func_type tgt, - const int seq_len, const std::vector& x, - const std::vector& w, std::vector& alpharef, // NOLINT - std::vector& trackref, int tag_num) { // NOLINT - constexpr int state_trans_base_idx = 2; - EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(x.size(), static_cast(seq_len * tag_num)); - EXPECT_EQ(w.size(), - static_cast((tag_num + state_trans_base_idx) * tag_num)); - EXPECT_EQ(alpharef.size(), static_cast(seq_len * tag_num)); - EXPECT_EQ(trackref.size(), static_cast(seq_len * tag_num)); - std::vector alphatgt(alpharef.size()); - std::vector tracktgt(trackref.size()); - - memcpy(trackref.data(), tracktgt.data(), tag_num * sizeof(int)); - tgt(seq_len, (const T*)x.data(), (const T*)w.data(), alphatgt.data(), - tracktgt.data(), tag_num); - ExpectEQ(alpharef.data(), alphatgt.data(), seq_len * tag_num); - ExpectEQ(trackref.data(), tracktgt.data(), seq_len * tag_num); - } -}; - -template -void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { - TestFuncWithRefer test; - // test jitcode - auto jitcode = jit::GetJitCode(attr); - if (jitcode) { - VLOG(10) << "Test Jitcode Kernel "; - test(jitcode, args...); +void TestAllImpls(const typename KernelTuple::attr_type& attr, + const Tester& verifier, const Args&... args) { + auto funcs = jit::GetAllCandidateFuncsWithTypes(attr); + for (auto f : funcs) { + VLOG(10) << "Test Kernel " << f.first; + verifier(f.second, args...); } - // test all impls in more - jit::KernelKey kkey(KT, PlaceType()); - auto& pool = jit::KernelPool().Instance().AllKernels(); - auto iter = pool.find(kkey); - if (iter != pool.end()) { - auto& impls = iter->second; - for (auto& impl : impls) { - auto i = dynamic_cast*>(impl.get()); - if (i && i->UseMe(attr)) { - auto more = i->GetFunc(); - VLOG(10) << "Test More Kernel : " << i->ImplType(); - test(more, args...); - } - } - } - // test result from Get function - // VLOG(10) << "Test Get function "; - auto tgt = jit::Get(attr); - test(tgt, args...); } -template -void TestKernelXYZNTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelXYZN() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); for (int d : TestSizes()) { - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector x(d), y(d), zref(d); @@ -494,16 +105,42 @@ void TestKernelXYZNTuples() { ExpectEQ(xinp_data, zref_data, d); ExpectEQ(yinp_data, zref_data, d); - TestAllImpls, PlaceType, std::vector, - std::vector, std::vector>(d, x, y, zref); + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& x, const std::vector& y, + const std::vector& zref) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(zref.size(), x.size()); + EXPECT_EQ(zref.size(), y.size()); + const T* x_data = x.data(); + const T* y_data = y.data(); + const T* zref_data = zref.data(); + const int d = zref.size(); + + std::vector ztgt(d); + T* ztgt_data = ztgt.data(); + // test normal + tgt(x_data, y_data, ztgt_data, d); + ExpectEQ(ztgt_data, zref_data, d); + // test inplace x + std::copy(x.begin(), x.end(), ztgt.begin()); + tgt(ztgt_data, y_data, ztgt_data, d); + ExpectEQ(ztgt_data, zref_data, d); + // test inplace y + std::copy(y.begin(), y.end(), ztgt.begin()); + tgt(x_data, ztgt_data, ztgt_data, d); + ExpectEQ(ztgt_data, zref_data, d); + }; + + TestAllImpls(d, verifier, x, y, zref); } } -template -void TestKernelAXYNTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelAXYN() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); for (int d : TestSizes()) { - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); const T a = static_cast(3); @@ -520,34 +157,33 @@ void TestKernelAXYNTuples() { ref(&a, xinp_data, xinp_data, d); ExpectEQ(xinp_data, yref_data, d); - TestAllImpls, PlaceType, T, std::vector, - std::vector>(d, a, x, yref); - } -} - -template -void TestKernelXRNTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - auto last_acc = FLAGS_acc; - FLAGS_acc = 1e-4; - for (int d : TestSizes()) { - auto ref = jit::GetRefer>(); - EXPECT_TRUE(ref != nullptr); - std::vector x(d); - RandomVec(d, x.data()); - T ref_res; - ref(x.data(), &ref_res, d); - TestAllImpls, PlaceType, std::vector, T>(d, x, - ref_res); + auto verifier = [](const typename KernelTuple::func_type tgt, const T a, + const std::vector& x, const std::vector& yref) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(yref.size(), x.size()); + const T* x_data = x.data(); + const T* yref_data = yref.data(); + const int d = yref.size(); + std::vector ytgt(d); + T* ytgt_data = ytgt.data(); + // test normal + tgt(&a, x_data, ytgt_data, d); + ExpectEQ(ytgt_data, yref_data, d); + // test inplace x + std::copy(x.begin(), x.end(), ytgt.begin()); + tgt(&a, ytgt_data, ytgt_data, d); + ExpectEQ(ytgt_data, yref_data, d); + }; + TestAllImpls(d, verifier, a, x, yref); } - FLAGS_acc = last_acc; } -template -void TestKernelXYNTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelXYN() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); for (int d : TestSizes()) { - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector x(d), yref(d); @@ -562,15 +198,57 @@ void TestKernelXYNTuples() { ref(x_data, yref_data, d); ref(xinp_data, xinp_data, d); ExpectEQ(xinp_data, yref_data, d); + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& x, const std::vector& yref) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(yref.size(), x.size()); + const T* x_data = x.data(); + const T* yref_data = yref.data(); + const int d = yref.size(); + std::vector ytgt(d); + T* ytgt_data = ytgt.data(); + // test normal + tgt(x_data, ytgt_data, d); + ExpectEQ(ytgt_data, yref_data, d); + // test inplace x + std::copy(x.begin(), x.end(), ytgt.begin()); + tgt(ytgt_data, ytgt_data, d); + ExpectEQ(ytgt_data, yref_data, d); + }; + TestAllImpls(d, verifier, x, yref); + } +} - TestAllImpls, PlaceType, std::vector, - std::vector>(d, x, yref); +template +void TestKernelXRN() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); + auto last_acc = FLAGS_acc; + FLAGS_acc = 1e-4; + for (int d : TestSizes()) { + auto ref = jit::GetReferFunc(); + EXPECT_TRUE(ref != nullptr); + std::vector x(d); + RandomVec(d, x.data()); + T ref_res; + ref(x.data(), &ref_res, d); + + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& x, const T ref_res) { + EXPECT_TRUE(tgt != nullptr); + T tgt_res; + tgt(x.data(), &tgt_res, x.size()); + ExpectEQ(&tgt_res, &ref_res, 1); + }; + TestAllImpls(d, verifier, x, ref_res); } + FLAGS_acc = last_acc; } -template -void TestKernelLSTMTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelLSTM() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); std::vector all_acts = {"sigmoid", "tanh", "relu", "identity"}; auto test_sizes = TestSizes(); test_sizes.erase(std::remove(test_sizes.begin(), test_sizes.end(), 1000)); @@ -582,7 +260,7 @@ void TestKernelLSTMTuples() { const jit::lstm_attr_t attr( d, jit::to_kerneltype(act_gate), jit::to_kerneltype(act_cand), jit::to_kerneltype(act_cell), use_peephole); - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector xsrc(4 * d), wp(3 * d), ct_1(d); std::vector ct_ref(d), ht_ref(d), checked(2 * d); @@ -609,10 +287,51 @@ void TestKernelLSTMTuples() { } ref(&step, &attr); VLOG(10) << attr; - TestAllImpls, PlaceType, std::vector, - std::vector, std::vector, std::vector, - std::vector>(attr, xsrc, wp, ct_1, ct_ref, ht_ref, - attr); + + auto verifier = []( + const typename KernelTuple::func_type tgt, + const std::vector& xsrc, const std::vector& wp, + const std::vector& ct_1, const std::vector& ct_ref, + const std::vector& ht_ref, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(ct_ref.size(), ht_ref.size()); + EXPECT_EQ(ct_1.size(), ht_ref.size()); + EXPECT_EQ(xsrc.size(), 4 * ht_ref.size()); + EXPECT_EQ(wp.size(), 3 * ht_ref.size()); + + // x could be changed after compute, so copy to save src + int d = ht_ref.size(); + std::vector x(xsrc.size()), ct(ct_ref.size()), + ht(ht_ref.size()); + std::vector checked(2 * d); + std::copy(xsrc.begin(), xsrc.end(), x.begin()); + + const T* ct_1_data = ct_1.data(); + const T* wp_data = wp.data(); + const T* ct_ref_data = ct_ref.data(); + const T* ht_ref_data = ht_ref.data(); + T* x_data = x.data(); + T* ct_data = ct.data(); + T* ht_data = ht.data(); + T* checked_data = checked.data(); + + jit::lstm_t step; + step.gates = x_data; + step.ct_1 = ct_1_data; + step.ct = ct_data; + step.ht = ht_data; + if (attr.use_peephole) { + step.wp = wp_data; + step.checked = checked_data; + } + + tgt(&step, &attr); + ExpectEQ(ct_data, ct_ref_data, d); + ExpectEQ(ht_data, ht_ref_data, d); + }; + TestAllImpls(attr, verifier, xsrc, wp, ct_1, + ct_ref, ht_ref, attr); } } } @@ -620,9 +339,10 @@ void TestKernelLSTMTuples() { } } -template -void TestKernelGRUTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelGRU() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); std::vector all_acts = {"sigmoid", "tanh", "relu", "identity"}; auto test_sizes = TestSizes(); test_sizes.erase(std::remove(test_sizes.begin(), test_sizes.end(), 1000)); @@ -631,7 +351,7 @@ void TestKernelGRUTuples() { for (auto& act_cand : all_acts) { const jit::gru_attr_t attr(d, jit::to_kerneltype(act_gate), jit::to_kerneltype(act_cand)); - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector xsrc(3 * d), ht_1(d), ht_ref(d); RandomVec(3 * d, xsrc.data()); @@ -648,17 +368,218 @@ void TestKernelGRUTuples() { step.ht = ht_ref_data; ref(&step, &attr); VLOG(10) << attr; - TestAllImpls, PlaceType, std::vector, - std::vector, std::vector>(attr, xsrc, ht_1, ht_ref, - attr); + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& xsrc, + const std::vector& ht_1, + const std::vector& ht_ref, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(ht_1.size(), ht_ref.size()); + EXPECT_EQ(xsrc.size(), 3 * ht_ref.size()); + + // x could be changed after compute, so copy to save src + int d = ht_ref.size(); + std::vector x(xsrc.size()), ht(ht_ref.size()); + std::copy(xsrc.begin(), xsrc.end(), x.begin()); + const T* ht_1_data = ht_1.data(); + const T* ht_ref_data = ht_ref.data(); + T* x_data = x.data(); + T* ht_data = ht.data(); + jit::gru_t step; + step.gates = x_data; + step.ht_1 = ht_1_data; + step.ht = ht_data; + tgt(&step, &attr); + ExpectEQ(ht_data, ht_ref_data, d); + }; + TestAllImpls(attr, verifier, xsrc, ht_1, ht_ref, + attr); } } } } -template -void TestKernelSeqPoolTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelNCHW16CMulNC() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); + const int n = 3, c = 16 * 4, h = 10, w = 10; + auto ref = jit::GetReferFunc(); + EXPECT_TRUE(ref != nullptr); + int sz = n * c * h * w; + std::vector x(sz), y(n * c), zref(sz); + std::vector ztgt(sz), zjit(sz); + RandomVec(sz, x.data()); + RandomVec(n * c, y.data()); + + const T* x_data = x.data(); + const T* y_data = y.data(); + T* zref_data = zref.data(); + T* ztgt_data = ztgt.data(); + T* zjit_data = zjit.data(); + constexpr int simd_width = ZMM_FLOAT_BLOCK; + int C = c / simd_width; + auto tgt = jit::KernelFuncs::Cache().At(0); + auto funcs = jit::GetAllCandidateFuncs(0); + EXPECT_GT(funcs.size(), 0UL); + auto jitcode = funcs[0]; + EXPECT_TRUE(tgt != nullptr); + + if (std::is_same::value && + paddle::platform::MayIUse(paddle::platform::avx512f)) { + EXPECT_TRUE(jitcode != nullptr); + } + for (int ni = 0; ni < n; ni++) { + for (int ci = 0; ci < C; ci++) { + auto ptr_x = + x_data + ni * C * h * w * simd_width + ci * h * w * simd_width; + auto ptr_y = y_data + ni * C * simd_width + ci * simd_width; + auto ptr_zref = + zref_data + ni * C * h * w * simd_width + ci * h * w * simd_width; + auto ptr_ztgt = + ztgt_data + ni * C * h * w * simd_width + ci * h * w * simd_width; + + ref(ptr_x, ptr_y, ptr_zref, h, w); + tgt(ptr_x, ptr_y, ptr_ztgt, h, w); + + if (jitcode) { + auto ptr_zjit = + zjit_data + ni * C * h * w * simd_width + ci * h * w * simd_width; + jitcode(ptr_x, ptr_y, ptr_zjit, h, w); + } + } + } + ExpectEQ(ztgt_data, zref_data, sz); + if (jitcode) { + ExpectEQ(zjit_data, zref_data, sz); + } +} + +template +void TestKernelLayerNorm() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); + const T epsilon = 9.99999975e-06; + for (int n : {1, 2, 10}) { + for (int x_dim_0 : {1, 9, 17, 50}) { + int left = n * x_dim_0; + for (int x_dim_1 : TestSizes()) { + int right = x_dim_1; + auto ref = jit::GetReferFunc(); + EXPECT_TRUE(ref != nullptr); + int sz = left * right; + std::vector x(sz), mean(left), var(left), scale(right), bias(right), + outref(sz); + RandomVec(sz, x.data()); + RandomVec(left, mean.data()); + RandomVec(left, var.data()); + RandomVec(right, scale.data()); + RandomVec(right, bias.data()); + + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* outref_data = outref.data(); + + ref(x_data, outref_data, mean_data, var_data, scale_data, bias_data, + left, epsilon, right); + + auto verifier = []( + const typename KernelTuple::func_type tgt, const std::vector& x_, + const std::vector& outref_, const std::vector& mean_, + const std::vector& var_, const std::vector& scale, + const std::vector& bias, const int& left, const float& epsilon, + const typename KernelTuple::attr_type& right) { + EXPECT_TRUE(tgt != nullptr); + std::vector outtgt(outref_.size()); + std::vector x(x_.size()); + std::vector mean(mean_.size()); + std::vector var(var_.size()); + std::vector outref(outref_.size()); + std::copy(x_.begin(), x_.end(), x.begin()); + std::copy(mean_.begin(), mean_.end(), mean.begin()); + std::copy(var_.begin(), var_.end(), var.begin()); + std::copy(outref_.begin(), outref_.end(), outref.begin()); + + EXPECT_EQ(x.size(), static_cast(left * right)); + EXPECT_EQ(outref.size(), static_cast(left * right)); + EXPECT_EQ(mean.size(), static_cast(left)); + EXPECT_EQ(var.size(), static_cast(left)); + EXPECT_EQ(scale.size(), static_cast(right)); + EXPECT_EQ(bias.size(), static_cast(right)); + + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* outref_data = outref.data(); + T* outtgt_data = outtgt.data(); + tgt(x_data, outtgt_data, mean_data, var_data, scale_data, bias_data, + left, epsilon, right); + ExpectEQ(outtgt_data, outref_data, left * right); + }; + TestAllImpls(right, verifier, x, outref, mean, + var, scale, bias, left, epsilon, + right); + } + } + } +} + +template +void TestKernelCRFDecoding() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); + constexpr int state_trans_base_idx = 2; + auto test_sizes = TestSizes(); + test_sizes.erase(std::remove(test_sizes.begin(), test_sizes.end(), 2000)); + for (int seq_len : {1, 11, 17, 50}) { + for (int tag_num : test_sizes) { + auto ref = jit::GetReferFunc(); + EXPECT_TRUE(ref != nullptr); + int x_sz = seq_len * tag_num; + int w_sz = (tag_num + state_trans_base_idx) * tag_num; + std::vector x(x_sz), w(w_sz), alpharef(x_sz); + std::vector trackref(x_sz); + RandomVec(x_sz, x.data()); + RandomVec(w_sz, w.data()); + + ref(seq_len, (const T*)x.data(), (const T*)w.data(), alpharef.data(), + trackref.data(), tag_num); + + auto verifier = []( + const typename KernelTuple::func_type tgt, const int& seq_len, + const std::vector& x, const std::vector& w, + const std::vector& alpharef, const std::vector& trackref, + const typename KernelTuple::attr_type& tag_num) { + constexpr int state_trans_base_idx = 2; + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size(), static_cast(seq_len * tag_num)); + EXPECT_EQ(w.size(), static_cast( + (tag_num + state_trans_base_idx) * tag_num)); + EXPECT_EQ(alpharef.size(), static_cast(seq_len * tag_num)); + EXPECT_EQ(trackref.size(), static_cast(seq_len * tag_num)); + std::vector alphatgt(alpharef.size()); + std::vector tracktgt(trackref.size()); + memcpy(tracktgt.data(), trackref.data(), tag_num * sizeof(int)); + tgt(seq_len, (const T*)x.data(), (const T*)w.data(), alphatgt.data(), + tracktgt.data(), tag_num); + ExpectEQ(alpharef.data(), alphatgt.data(), seq_len * tag_num); + ExpectEQ(trackref.data(), tracktgt.data(), seq_len * tag_num); + }; + TestAllImpls(tag_num, verifier, seq_len, x, w, + alpharef, trackref, tag_num); + } + } +} + +template +void TestKernelSeqPool() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); std::vector pool_types = { jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt}; auto test_sizes = TestSizes(); @@ -668,7 +589,7 @@ void TestKernelSeqPoolTuples() { jit::seq_pool_attr_t attr(w, type); for (int h : test_sizes) { attr.h = h; - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector x(h * w), yref(w); RandomVec(h * w, x.data()); @@ -676,16 +597,86 @@ void TestKernelSeqPoolTuples() { T* yref_data = yref.data(); ref(x_data, yref_data, &attr); VLOG(10) << attr; - TestAllImpls, PlaceType, std::vector, - std::vector>(attr, x, yref, attr); + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& x, const std::vector& yref, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size() % yref.size(), static_cast(0)); + int w = yref.size(); + std::vector y(w); + const T* x_data = x.data(); + const T* yref_data = yref.data(); + T* y_data = y.data(); + tgt(x_data, y_data, &attr); + ExpectEQ(y_data, yref_data, w); + }; + TestAllImpls(attr, verifier, x, yref, attr); } } } } -template -void TestKernelMatMulTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelEmbSeqPool() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); + int64_t tbl_h = 1e4; + std::vector pool_types = { + jit::SeqPoolType::kSum}; // only support sum yet + auto test_sizes = TestSizes(); + test_sizes.erase(std::remove(test_sizes.begin(), test_sizes.end(), 1000)); + for (int tbl_w : test_sizes) { + std::vector table(tbl_h * tbl_w); + RandomVec(tbl_h * tbl_w, table.data()); + const T* table_data = table.data(); + for (auto type : pool_types) { + for (int idx_w : {1, 2, 10, 16}) { + for (int idx_h : {1, 2, 9, 13, 16}) { + auto ref = jit::GetReferFunc(); + EXPECT_TRUE(ref != nullptr); + std::vector idx(idx_h * idx_w); + RandomVec(idx_h * idx_w, idx.data(), 0, tbl_h - 1); + int64_t out_w = tbl_w * idx_w; + std::vector oref(out_w); + const int64_t* idx_data = idx.data(); + T* o_data = oref.data(); + jit::emb_seq_pool_attr_t attr(tbl_h, tbl_w, idx_h, idx_w, out_w, + type); + ref(table_data, idx_data, o_data, &attr); + + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& table, + const std::vector& idx, + const std::vector& oref, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(table.size(), static_cast(attr.table_height * + attr.table_width)); + EXPECT_EQ(idx.size(), static_cast(attr.index_height * + attr.index_width)); + EXPECT_EQ(oref.size(), + static_cast(attr.table_width * attr.index_width)); + const T* table_data = table.data(); + const int64_t* idx_data = idx.data(); + const T* oref_data = oref.data(); + int o_w = oref.size(); + std::vector out(o_w); + T* o_data = out.data(); + tgt(table_data, idx_data, o_data, &attr); + ExpectEQ(o_data, oref_data, o_w); + }; + TestAllImpls(attr, verifier, table, idx, oref, + attr); + } + } + } + } +} + +template +void TestKernelMatMul() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); auto last_acc = FLAGS_acc; // export MKL_CBWR=AVX would make MKL force to use AVX // export KMP_DETERMINISTIC_REDUCTION=yes would make the result deterministic @@ -693,7 +684,7 @@ void TestKernelMatMulTuples() { for (int m : {1, 2, 3, 4}) { for (int n : {1, 2, 3, 4}) { for (int k : TestSizes()) { - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector a(m * k), b(k * n), c(m * n); RandomVec(m * k, a.data()); @@ -703,20 +694,36 @@ void TestKernelMatMulTuples() { T* c_data = c.data(); const jit::matmul_attr_t attr{m, n, k}; ref(a_data, b_data, c_data, &attr); - TestAllImpls, PlaceType, std::vector, - std::vector, std::vector>(attr, a, b, c, attr); + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& a, const std::vector& b, + const std::vector& cref, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(a.size(), static_cast(attr.m * attr.k)); + EXPECT_EQ(b.size(), static_cast(attr.k * attr.n)); + EXPECT_EQ(cref.size(), static_cast(attr.m * attr.n)); + std::vector c(cref.size()); + const T* a_data = a.data(); + const T* b_data = b.data(); + const T* cref_data = cref.data(); + T* c_data = c.data(); + tgt(a_data, b_data, c_data, &attr); + ExpectEQ(c_data, cref_data, attr.m * attr.n); + }; + TestAllImpls(attr, verifier, a, b, c, attr); } } } FLAGS_acc = last_acc; } -template -void TestKernelSoftmaxTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelSoftmax() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); for (int bs : {1, 2, 10}) { for (int n : TestSizes()) { - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); std::vector x(bs * n), y(bs * n); RandomVec(bs * n, x.data()); @@ -730,51 +737,33 @@ void TestKernelSoftmaxTuples() { ref(xinp_data, xinp_data, n, bs); ExpectEQ(xinp_data, y_data, n * bs); - TestAllImpls, PlaceType, std::vector, - std::vector>(n, x, y, n, bs); - } - } -} - -template -void TestKernelEmbSeqPoolTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - int64_t tbl_h = 1e4; - std::vector pool_types = { - jit::SeqPoolType::kSum}; // only support sum yet - auto test_sizes = TestSizes(); - test_sizes.erase(std::remove(test_sizes.begin(), test_sizes.end(), 1000)); - for (int tbl_w : test_sizes) { - std::vector table(tbl_h * tbl_w); - RandomVec(tbl_h * tbl_w, table.data()); - const T* table_data = table.data(); - for (auto type : pool_types) { - for (int idx_w : {1, 2, 10, 16}) { - for (int idx_h : {1, 2, 9, 13, 16}) { - auto ref = jit::GetRefer>(); - EXPECT_TRUE(ref != nullptr); - std::vector idx(idx_h * idx_w); - RandomVec(idx_h * idx_w, idx.data(), 0, tbl_h - 1); - int64_t out_w = tbl_w * idx_w; - std::vector oref(out_w); - const int64_t* idx_data = idx.data(); - T* o_data = oref.data(); - jit::emb_seq_pool_attr_t attr(tbl_h, tbl_w, idx_h, idx_w, out_w, - type); - ref(table_data, idx_data, o_data, &attr); - - TestAllImpls, PlaceType, std::vector, - std::vector, std::vector>(attr, table, idx, - oref, attr); - } - } + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& x, const std::vector& yref, + int n, int bs) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(yref.size(), x.size()); + EXPECT_EQ(x.size(), static_cast(n * bs)); + const T* x_data = x.data(); + const T* yref_data = yref.data(); + std::vector ytgt(n * bs); + T* ytgt_data = ytgt.data(); + // test normal + tgt(x_data, ytgt_data, n, bs); + ExpectEQ(ytgt_data, yref_data, n * bs); + // test inplace x + std::copy(x.begin(), x.end(), ytgt.begin()); + tgt(ytgt_data, ytgt_data, n, bs); + ExpectEQ(ytgt_data, yref_data, n * bs); + }; + TestAllImpls(n, verifier, x, y, n, bs); } } } -template -void TestKernelSgdTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); +template +void TestKernelSgd() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); const T lr = 0.1; auto UnDuplicatedRandomVec = [](int n, const int64_t lower, const int64_t upper) -> std::vector { @@ -802,7 +791,7 @@ void TestKernelSgdTuples() { RandomVec(rows_size * grad_w, grad.data()); const int64_t* rows_data = rows.data(); const T* grad_data = grad.data(); - auto ref = jit::GetRefer>(); + auto ref = jit::GetReferFunc(); EXPECT_TRUE(ref != nullptr); jit::sgd_attr_t attr(param_h, grad_w, rows_size, grad_w, rows_size); ref(&lr, param_data, grad_data, rows_data, out_data, &attr); @@ -818,227 +807,488 @@ void TestKernelSgdTuples() { grad_w); } - TestAllImpls, PlaceType, T, std::vector, - std::vector, std::vector, std::vector>( - attr, lr, param, grad, rows, param_out, attr); + auto verifier = []( + const typename KernelTuple::func_type tgt, const T lr, + const std::vector& param, const std::vector& grad, + const std::vector& rows, const std::vector& oref, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(param.size(), + static_cast(attr.param_height * attr.param_width)); + EXPECT_EQ(grad.size(), + static_cast(attr.grad_height * attr.grad_width)); + EXPECT_EQ(rows.size(), static_cast(attr.selected_rows_size)); + EXPECT_EQ(param.size(), oref.size()); + const T* param_data = param.data(); + const T* grad_data = grad.data(); + const int64_t* rows_data = rows.data(); + const T* oref_data = oref.data(); + + std::vector out(oref.size()); + T* o_data = out.data(); + tgt(&lr, param_data, grad_data, rows_data, o_data, &attr); + // only the selected rows should be equal + for (size_t i = 0; i < rows.size(); ++i) { + ExpectEQ(o_data + rows[i] * attr.grad_width, + oref_data + rows[i] * attr.grad_width, attr.grad_width); + } + + // inplace + std::copy(param.begin(), param.end(), out.begin()); + tgt(&lr, o_data, grad_data, rows_data, o_data, &attr); + for (size_t i = 0; i < rows.size(); ++i) { + ExpectEQ(o_data + rows[i] * attr.grad_width, + oref_data + rows[i] * attr.grad_width, attr.grad_width); + } + }; + TestAllImpls(attr, verifier, lr, param, grad, + rows, param_out, attr); } } } } -template -void TestKernelNCHW16CMulNCTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - const int n = 3, c = 16 * 4, h = 10, w = 10; - auto ref = jit::GetRefer>(); - EXPECT_TRUE(ref != nullptr); - int sz = n * c * h * w; - std::vector x(sz), y(n * c), zref(sz); - std::vector ztgt(sz), zjit(sz); - RandomVec(sz, x.data()); - RandomVec(n * c, y.data()); - - const T* x_data = x.data(); - const T* y_data = y.data(); - T* zref_data = zref.data(); - T* ztgt_data = ztgt.data(); - T* zjit_data = zjit.data(); - constexpr int simd_width = ZMM_FLOAT_BLOCK; - int C = c / simd_width; - auto tgt = jit::Get, PlaceType>(0); - auto jitcode = jit::GetJitCode, PlaceType>(0); - EXPECT_TRUE(tgt != nullptr); - - if (std::is_same::value && - paddle::platform::MayIUse(paddle::platform::avx512f)) { - EXPECT_TRUE(jitcode != nullptr); - } - for (int ni = 0; ni < n; ni++) { - for (int ci = 0; ci < C; ci++) { - auto ptr_x = - x_data + ni * C * h * w * simd_width + ci * h * w * simd_width; - auto ptr_y = y_data + ni * C * simd_width + ci * simd_width; - auto ptr_zref = - zref_data + ni * C * h * w * simd_width + ci * h * w * simd_width; - auto ptr_ztgt = - ztgt_data + ni * C * h * w * simd_width + ci * h * w * simd_width; - - ref(ptr_x, ptr_y, ptr_zref, h, w); - tgt(ptr_x, ptr_y, ptr_ztgt, h, w); +template +void TestKernelVBroadcast() { + using T = typename KernelTuple::data_type; + VLOG(10) << "Test JITKernel: " << jit::to_string(KernelTuple::kernel_type); + for (int w : TestSizes()) { + std::vector x(w); + RandomVec(w, x.data()); + const T* x_data = x.data(); + for (int64_t h : {1, 2, 6}) { + auto ref = jit::GetReferFunc(); + EXPECT_TRUE(ref != nullptr); + std::vector y(w * h); + T* y_data = y.data(); + ref(x_data, y_data, h, w); - if (jitcode) { - auto ptr_zjit = - zjit_data + ni * C * h * w * simd_width + ci * h * w * simd_width; - jitcode(ptr_x, ptr_y, ptr_zjit, h, w); - } + auto verifier = [](const typename KernelTuple::func_type tgt, + const std::vector& x, const std::vector& yref, + const int64_t& h, + const typename KernelTuple::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size(), static_cast(attr)); + EXPECT_EQ(yref.size(), x.size() * h); + std::vector y(yref.size()); + const T* x_data = x.data(); + const T* yref_data = yref.data(); + T* y_data = y.data(); + tgt(x_data, y_data, h, attr); + ExpectEQ(y_data, yref_data, yref.size()); + }; + TestAllImpls(static_cast(w), verifier, x, + y, h, static_cast(w)); } } - ExpectEQ(ztgt_data, zref_data, sz); - if (jitcode) { - ExpectEQ(zjit_data, zref_data, sz); - } } -template -void TestKernelLayerNormTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - const T epsilon = 9.99999975e-06; - for (int n : {1, 2, 10}) { - for (int x_dim_0 : {1, 9, 17, 50}) { - int left = n * x_dim_0; - for (int x_dim_1 : TestSizes()) { - int right = x_dim_1; - auto ref = jit::GetRefer>(); - EXPECT_TRUE(ref != nullptr); - int sz = left * right; - std::vector x(sz), mean(left), var(left), scale(right), bias(right), - outref(sz); - RandomVec(sz, x.data()); - RandomVec(left, mean.data()); - RandomVec(left, var.data()); - RandomVec(right, scale.data()); - RandomVec(right, bias.data()); - - const T* scale_data = scale.data(); - const T* bias_data = bias.data(); - T* x_data = x.data(); - T* mean_data = mean.data(); - T* var_data = var.data(); - T* outref_data = outref.data(); +// test pool +TEST(JITKernel_pool, jitcreator) { + const auto& jitcreators = jit::JitCodeCreatorPool::Instance().AllCreators(); +#if defined(_WIN32) || defined(__APPLE__) || defined(__OSX__) + EXPECT_EQ(jitcreators.size(), 0UL); +#else + EXPECT_EQ(jitcreators.size(), 25UL); +#endif +} - ref(x_data, outref_data, mean_data, var_data, scale_data, bias_data, - left, epsilon, right); +TEST(JITKernel_pool, jitpool) { + // jitpool is related with attr + const auto& kers = jit::JitCodePool().Instance().AllKernels(); + EXPECT_EQ(kers.size(), 0UL); + jit::GetAllCandidateKernels, CPUPlace>(3); +// after call GetAllCandidateKernels, it will create jitcode Automatically +#if defined(_WIN32) || defined(__APPLE__) || defined(__OSX__) + EXPECT_EQ(kers.size(), 0UL); +#else + EXPECT_EQ(kers.size(), 1UL); +#endif +} - TestAllImpls, PlaceType, std::vector, - std::vector, std::vector, std::vector, - std::vector, std::vector, int, float>( - right, x, outref, mean, var, scale, bias, left, epsilon, right); - } - } - } +TEST(JITKernel_pool, more) { + const auto& kers = jit::KernelPool::Instance().AllKernels(); +#if defined(__APPLE__) || defined(__OSX__) + EXPECT_EQ(kers.size(), 10UL); +#else +#ifdef PADDLE_WITH_MKLML + EXPECT_EQ(kers.size(), 21UL); +#else + EXPECT_EQ(kers.size(), 8UL); +#endif +#endif } -template -void TestKernelCRFDecodingTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - constexpr int state_trans_base_idx = 2; - auto test_sizes = TestSizes(); - test_sizes.erase(std::remove(test_sizes.begin(), test_sizes.end(), 2000)); - for (int seq_len : {1, 11, 17, 50}) { - for (int tag_num : test_sizes) { - auto ref = jit::GetRefer>(); - EXPECT_TRUE(ref != nullptr); - int x_sz = seq_len * tag_num; - int w_sz = (tag_num + state_trans_base_idx) * tag_num; - std::vector x(x_sz), w(w_sz), alpharef(x_sz); - std::vector trackref(x_sz); - RandomVec(x_sz, x.data()); - RandomVec(w_sz, w.data()); +TEST(JITKernel_pool, refer) { + const auto& kers = jit::ReferKernelPool::Instance().AllKernels(); + EXPECT_EQ(kers.size(), 29UL); +} - ref(seq_len, (const T*)x.data(), (const T*)w.data(), alpharef.data(), - trackref.data(), tag_num); +// test helper +TEST(JITKernel_helper, GetAllCandidateKernels) { + auto fp_kers = + jit::GetAllCandidateKernels, CPUPlace>(10); +#if defined(_WIN32) || defined(__APPLE__) || defined(__OSX__) + EXPECT_GE(fp_kers.size(), 1UL); // refer +#else +#ifdef PADDLE_WITH_MKLML + EXPECT_GE(fp_kers.size(), 3UL); // jitcode, mkl, refer +#else + EXPECT_GE(fp_kers.size(), 2UL); // jitcode, refer +#endif +#endif + + auto db_kers = + jit::GetAllCandidateKernels, CPUPlace>(10); +#if defined(_WIN32) || defined(__APPLE__) || defined(__OSX__) + EXPECT_GE(db_kers.size(), 1UL); // refer +#else +#ifdef PADDLE_WITH_MKLML + EXPECT_GE(db_kers.size(), 2UL); // mkl, refer +#else + EXPECT_GE(db_kers.size(), 1UL); // refer +#endif +#endif +} - TestAllImpls, PlaceType, int, - std::vector, std::vector, std::vector, - std::vector, int>(tag_num, seq_len, x, w, alpharef, - trackref, tag_num); - } - } +TEST(JITKernel_helper, GetAllCandidateFuncsWithTypes) { + auto fp_kers = + jit::GetAllCandidateFuncsWithTypes, CPUPlace>(10); +#if defined(__APPLE__) || defined(__OSX__) + EXPECT_GE(fp_kers.size(), 1UL); // refer +#else +#if !defined(PADDLE_WITH_MKLML) || defined(_WIN32) + EXPECT_GE(fp_kers.size(), 2UL); // jitcode/mkl, refer +#else + EXPECT_GE(fp_kers.size(), 3UL); // jitcode, mkl, refer +#endif +#endif + + auto db_kers = + jit::GetAllCandidateFuncsWithTypes, CPUPlace>(10); +#if defined(__APPLE__) || defined(__OSX__) || !defined(PADDLE_WITH_MKLML) + EXPECT_GE(db_kers.size(), 1UL); // refer +#else + EXPECT_GE(db_kers.size(), 2UL); // mkl, refer +#endif } -template -void TestKernelVBroadcastTuples() { - VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - for (int w : TestSizes()) { - std::vector x(w); - RandomVec(w, x.data()); - const T* x_data = x.data(); - for (int64_t h : {1, 2, 6}) { - auto ref = jit::GetRefer>(); - EXPECT_TRUE(ref != nullptr); - std::vector y(w * h); - T* y_data = y.data(); - ref(x_data, y_data, h, w); +TEST(JITKernel_helper, KernelFuncs) { + auto f1 = jit::KernelFuncs, CPUPlace>::Cache().At(3); + auto f2 = jit::KernelFuncs, CPUPlace>::Cache()[3]; + EXPECT_TRUE(f1 != nullptr); + EXPECT_TRUE(f1 == f2); + + auto f3 = jit::KernelFuncs, CPUPlace>::Cache()[5]; +#if defined(_WIN32) || defined(__APPLE__) || defined(__OSX__) + EXPECT_TRUE(f2 == f3); +#else + EXPECT_TRUE(f2 != f3); +#endif +} - TestAllImpls, PlaceType, std::vector, - std::vector, int64_t>(static_cast(w), x, y, h, - static_cast(w)); - } +TEST(JITKernel_helper, GetAllCandidateFuncs) { + auto funcs = jit::GetAllCandidateFuncs, CPUPlace>(10); + auto kers = jit::GetAllCandidateKernels, CPUPlace>(10); + EXPECT_EQ(funcs.size(), kers.size()); + + std::vector x(10), tgt(10); + RandomVec(10, x.data()); + auto best = jit::GetDefaultBestFunc, CPUPlace>(10); + best(x.data(), tgt.data(), 10); + for (auto f : funcs) { + std::vector y(10); + f(x.data(), y.data(), 10); + ExpectEQ(y.data(), tgt.data(), 10); } } -#define TEST_CPU_KERNEL(test_tuple, kernel_type) \ - TEST(JITKernel, kernel_type) { \ - TestKernel##test_tuple(); \ - TestKernel##test_tuple(); \ +TEST(JITKernel_helper, pack_weights) { + const int N = 8 * 60, K = 2; + float src[K][N], yref[K][N], y[K * N]; + float* x = &(src[0][0]); + float* ref = &(yref[0][0]); + for (int i = 0; i < N * K; ++i) { + *(x + i) = static_cast(i); + } + int block = 0; + std::vector groups; + if (paddle::platform::MayIUse(paddle::platform::avx512f)) { + block = ZMM_FLOAT_BLOCK; + groups.push_back(30); + } else { + block = YMM_FLOAT_BLOCK; + groups.insert(groups.end(), {14, 14, 14, 14, 4}); } -TEST_CPU_KERNEL(XYZNTuples, kVMul); -TEST_CPU_KERNEL(XYZNTuples, kVAdd); -TEST_CPU_KERNEL(XYZNTuples, kVAddRelu); -TEST_CPU_KERNEL(XYZNTuples, kVSub); - -TEST_CPU_KERNEL(AXYNTuples, kVScal); -TEST_CPU_KERNEL(AXYNTuples, kVAddBias); + int offset = 0; + int acc = 0; + for (int g : groups) { + g = g * block; + for (int k = 0; k < K; ++k) { + for (int i = 0; i < g; ++i) { + *(ref + offset) = src[k][i + acc]; + offset++; + } + } + acc += g; + } -TEST_CPU_KERNEL(XRNTuples, kHMax); -TEST_CPU_KERNEL(XRNTuples, kHSum); + jit::pack_weights(x, y, N, K); + ExpectEQ(y, ref, N * K); +} -TEST_CPU_KERNEL(XYNTuples, kVRelu); -TEST_CPU_KERNEL(XYNTuples, kVIdentity); -TEST_CPU_KERNEL(XYNTuples, kVSquare); -TEST_CPU_KERNEL(XYNTuples, kVExp); -TEST_CPU_KERNEL(XYNTuples, kVSigmoid); -TEST_CPU_KERNEL(XYNTuples, kVTanh); -TEST_CPU_KERNEL(XYNTuples, kVCopy); +TEST(JITKernel_helper, attr) { + std::ostringstream out; + // KernelTypes + out << jit::to_string(jit::kNone) << jit::to_string(jit::kCRFDecoding) + << jit::to_string(jit::kEmbSeqPool) << jit::to_string(jit::kGRUH1) + << jit::to_string(jit::kGRUHtPart1) << jit::to_string(jit::kGRUHtPart2) + << jit::to_string(jit::kHSum) << jit::to_string(jit::kHMax) + << jit::to_string(jit::kLSTMCtHt) << jit::to_string(jit::kLSTMC1H1) + << jit::to_string(jit::kLayerNorm) << jit::to_string(jit::kMatMul) + << jit::to_string(jit::kNCHW16CMulNC) << jit::to_string(jit::kSeqPool) + << jit::to_string(jit::kSoftmax) << jit::to_string(jit::kVAdd) + << jit::to_string(jit::kVAddBias) << jit::to_string(jit::kVAddRelu) + << jit::to_string(jit::kVBroadcast) << jit::to_string(jit::kVCopy) + << jit::to_string(jit::kVExp) << jit::to_string(jit::kVIdentity) + << jit::to_string(jit::kVMul) << jit::to_string(jit::kVRelu) + << jit::to_string(jit::kVScal) << jit::to_string(jit::kSgd) + << jit::to_string(jit::kVSigmoid) << jit::to_string(jit::kVSquare) + << jit::to_string(jit::kVSub) << jit::to_string(jit::kVTanh); + EXPECT_EQ(out.str().size(), 234); + + // SeqPoolTypes + out.str(""); + out << jit::to_string(jit::kSum) << jit::to_string(jit::kAvg) + << jit::to_string(jit::kSqrt); + EXPECT_EQ(out.str().size(), 13); + + EXPECT_EQ(jit::to_kerneltype("relu"), jit::kVRelu); + EXPECT_EQ(jit::to_kerneltype("Identity"), jit::kVIdentity); + EXPECT_EQ(jit::to_kerneltype("VEXP"), jit::kVExp); + EXPECT_EQ(jit::to_kerneltype("SigmoiD"), jit::kVSigmoid); + EXPECT_EQ(jit::to_kerneltype("VTanh"), jit::kVTanh); + + out.str(""); + out << jit::lstm_attr_t(8, jit::kVIdentity, jit::kVSigmoid, jit::kVTanh); + EXPECT_EQ(out.str().size(), 89); + + out.str(""); + out << jit::gru_attr_t(8, jit::kVIdentity, jit::kVSigmoid); + EXPECT_EQ(out.str().size(), 52); + + out.str(""); + out << jit::seq_pool_attr_t(8, jit::SeqPoolType::kSum); + EXPECT_EQ(out.str().size(), 44); + + out.str(""); + out << jit::emb_seq_pool_attr_t(1, 2, 3, 4, 5, jit::SeqPoolType::kAvg); + EXPECT_EQ(out.str().size(), 93); + + out.str(""); + out << jit::sgd_attr_t(1, 2, 3, 4, 5); + EXPECT_EQ(out.str().size(), 81); + + out.str(""); + out << jit::matmul_attr_t(1, 2, 3); + EXPECT_EQ(out.str().size(), 14); +} -TEST_CPU_KERNEL(LSTMTuples, kLSTMCtHt); -TEST_CPU_KERNEL(LSTMTuples, kLSTMC1H1); +// test keys +TEST(JITKernel_key, int) { + EXPECT_TRUE(jit::JitCodeKey(2) == jit::JitCodeKey(2)); + EXPECT_TRUE(jit::JitCodeKey(2) == jit::JitCodeKey(2)); + EXPECT_TRUE(jit::JitCodeKey(2) != jit::JitCodeKey(3)); +} -TEST_CPU_KERNEL(GRUTuples, kGRUH1); -TEST_CPU_KERNEL(GRUTuples, kGRUHtPart1); -TEST_CPU_KERNEL(GRUTuples, kGRUHtPart2); +TEST(JITKernel_key, gru) { + jit::gru_attr_t attr1(8, jit::kVSigmoid, jit::kVTanh); + jit::gru_attr_t attr2(8, jit::kVSigmoid, jit::kVTanh); + jit::gru_attr_t attr3(9, jit::kVSigmoid, jit::kVTanh); + jit::gru_attr_t attr4(9, jit::kVSigmoid, jit::kVIdentity); + jit::gru_attr_t attr5(9, jit::kVTanh, jit::kVIdentity); -TEST_CPU_KERNEL(NCHW16CMulNCTuples, kNCHW16CMulNC); + auto key1 = jit::JitCodeKey(attr1); + auto key2 = jit::JitCodeKey(attr2); + auto key3 = jit::JitCodeKey(attr3); + auto key4 = jit::JitCodeKey(attr4); + auto key5 = jit::JitCodeKey(attr5); -TEST_CPU_KERNEL(SeqPoolTuples, kSeqPool); -TEST_CPU_KERNEL(MatMulTuples, kMatMul); -TEST_CPU_KERNEL(SoftmaxTuples, kSoftmax); -TEST_CPU_KERNEL(EmbSeqPoolTuples, kEmbSeqPool); -TEST_CPU_KERNEL(SgdTuples, kSgd); -TEST_CPU_KERNEL(LayerNormTuples, kLayerNorm); -TEST_CPU_KERNEL(CRFDecodingTuples, kCRFDecoding); -TEST_CPU_KERNEL(VBroadcastTuples, kVBroadcast); + EXPECT_TRUE(key1 == key2); + EXPECT_TRUE(key2 != key3); + EXPECT_TRUE(key2 != key4); + EXPECT_TRUE(key2 != key5); + EXPECT_TRUE(key3 != key4); + EXPECT_TRUE(key3 != key5); + EXPECT_TRUE(key4 != key5); +} TEST(JITKernel_key, lstm) { jit::lstm_attr_t attr1(8, jit::kVIdentity, jit::kVSigmoid, jit::kVTanh); - jit::lstm_attr_t attr2(9, jit::kVIdentity, jit::kVSigmoid, jit::kVTanh); + jit::lstm_attr_t attr2(8, jit::kVIdentity, jit::kVSigmoid, jit::kVTanh); jit::lstm_attr_t attr3(9, jit::kVIdentity, jit::kVSigmoid, jit::kVTanh); jit::lstm_attr_t attr4(9, jit::kVRelu, jit::kVSigmoid, jit::kVTanh); + jit::lstm_attr_t attr5(9, jit::kVRelu, jit::kVSigmoid, jit::kVTanh, true); + jit::lstm_attr_t attr6(9, jit::kVRelu, jit::kVSigmoid, jit::kVTanh, true); auto key1 = jit::JitCodeKey(attr1); auto key2 = jit::JitCodeKey(attr2); auto key3 = jit::JitCodeKey(attr3); auto key4 = jit::JitCodeKey(attr4); + auto key5 = jit::JitCodeKey(attr5); + auto key6 = jit::JitCodeKey(attr6); - EXPECT_TRUE(key1 != key2); - EXPECT_TRUE(key2 == key3); + EXPECT_TRUE(key1 == key2); + EXPECT_TRUE(key2 != key3); + EXPECT_TRUE(key2 != key4); + EXPECT_TRUE(key2 != key5); EXPECT_TRUE(key3 != key4); + EXPECT_TRUE(key3 != key5); + EXPECT_TRUE(key4 != key5); + EXPECT_TRUE(key5 == key6); } -TEST(JITKernel_key, gru) { - jit::gru_attr_t attr1(8, jit::kVSigmoid, jit::kVTanh); - jit::gru_attr_t attr2(9, jit::kVSigmoid, jit::kVTanh); - jit::gru_attr_t attr3(9, jit::kVSigmoid, jit::kVTanh); - jit::gru_attr_t attr4(9, jit::kVSigmoid, jit::kVIdentity); +TEST(JITKernel_key, seq_pool) { + jit::seq_pool_attr_t attr1(2, jit::SeqPoolType::kSum, 1); + jit::seq_pool_attr_t attr2(2, jit::SeqPoolType::kSum, 3); + jit::seq_pool_attr_t attr3(3, jit::SeqPoolType::kSum, 3); + jit::seq_pool_attr_t attr4(3, jit::SeqPoolType::kAvg, 3); - auto key1 = jit::JitCodeKey(attr1); - auto key2 = jit::JitCodeKey(attr2); - auto key3 = jit::JitCodeKey(attr3); - auto key4 = jit::JitCodeKey(attr4); + auto key1 = jit::JitCodeKey(attr1); + auto key2 = jit::JitCodeKey(attr2); + auto key3 = jit::JitCodeKey(attr3); + auto key4 = jit::JitCodeKey(attr4); - EXPECT_TRUE(key1 != key2); + EXPECT_TRUE(key1 == key2); + EXPECT_TRUE(key2 != key3); + EXPECT_TRUE(key2 != key4); + EXPECT_TRUE(key3 != key4); +} + +TEST(JITKernel_key, matmul) { + jit::matmul_attr_t attr1(1, 2, 3); + jit::matmul_attr_t attr2(1, 2, 3); + jit::matmul_attr_t attr3(1, 3, 3); + jit::matmul_attr_t attr4(2, 3, 4); + + auto key1 = jit::JitCodeKey(attr1); + auto key2 = jit::JitCodeKey(attr2); + auto key3 = jit::JitCodeKey(attr3); + auto key4 = jit::JitCodeKey(attr4); + + EXPECT_TRUE(key1 == key2); + EXPECT_TRUE(key2 != key3); + EXPECT_TRUE(key2 != key4); + EXPECT_TRUE(key3 != key4); +} + +TEST(JITKernel_key, emb_seq_pool) { + jit::emb_seq_pool_attr_t attr1(1, 2, 3, 4, 5, jit::SeqPoolType::kSum); + jit::emb_seq_pool_attr_t attr2(1, 2, 3, 4, 5, jit::SeqPoolType::kSum); + jit::emb_seq_pool_attr_t attr3(10, 2, 9, 8, 7, jit::SeqPoolType::kAvg); + jit::emb_seq_pool_attr_t attr4(10, 3, 9, 8, 7, jit::SeqPoolType::kSum); + jit::emb_seq_pool_attr_t attr5(1, 6, 3, 4, 5, jit::SeqPoolType::kSum); + + auto key1 = jit::JitCodeKey(attr1); + auto key2 = jit::JitCodeKey(attr2); + auto key3 = jit::JitCodeKey(attr3); + auto key4 = jit::JitCodeKey(attr4); + auto key5 = jit::JitCodeKey(attr5); + + EXPECT_TRUE(key1 == key2); + EXPECT_TRUE(key2 == key3); + EXPECT_TRUE(key2 != key4); + EXPECT_TRUE(key2 != key5); + EXPECT_TRUE(key4 != key5); +} + +TEST(JITKernel_key, sgd) { + jit::sgd_attr_t attr1(1, 2, 3, 4, 5); + jit::sgd_attr_t attr2(1, 2, 3, 4, 5); + jit::sgd_attr_t attr3(9, 8, 7, 4, 6); + jit::sgd_attr_t attr4(1, 2, 3, 6, 5); + jit::sgd_attr_t attr5(10, 9, 8, 7, 6); + + auto key1 = jit::JitCodeKey(attr1); + auto key2 = jit::JitCodeKey(attr2); + auto key3 = jit::JitCodeKey(attr3); + auto key4 = jit::JitCodeKey(attr4); + auto key5 = jit::JitCodeKey(attr5); + + EXPECT_TRUE(key1 == key2); EXPECT_TRUE(key2 == key3); EXPECT_TRUE(key3 != key4); + EXPECT_TRUE(key3 != key5); + EXPECT_TRUE(key4 != key5); } -// TODO(TJ): add more test about key and pool + +// test kernerls +#define TestKernelVMul TestKernelXYZN +#define TestKernelVAdd TestKernelXYZN +#define TestKernelVAddRelu TestKernelXYZN +#define TestKernelVSub TestKernelXYZN + +#define TestKernelVScal TestKernelAXYN +#define TestKernelVAddBias TestKernelAXYN + +#define TestKernelVRelu TestKernelXYN +#define TestKernelVIdentity TestKernelXYN +#define TestKernelVSquare TestKernelXYN +#define TestKernelVExp TestKernelXYN +#define TestKernelVSigmoid TestKernelXYN +#define TestKernelVTanh TestKernelXYN +#define TestKernelVCopy TestKernelXYN + +#define TestKernelHMax TestKernelXRN +#define TestKernelHSum TestKernelXRN + +#define TestKernelLSTMCtHt TestKernelLSTM +#define TestKernelLSTMC1H1 TestKernelLSTM + +#define TestKernelGRUH1 TestKernelGRU +#define TestKernelGRUHtPart1 TestKernelGRU +#define TestKernelGRUHtPart2 TestKernelGRU + +#define TEST_CPU_KERNEL(kernel_type) \ + TEST(JITKernel, kernel_type) { \ + TestKernel##kernel_type, CPUPlace>(); \ + TestKernel##kernel_type, CPUPlace>(); \ + } + +TEST_CPU_KERNEL(VMul); +TEST_CPU_KERNEL(VAdd); +TEST_CPU_KERNEL(VAddRelu); +TEST_CPU_KERNEL(VSub); + +TEST_CPU_KERNEL(VScal); +TEST_CPU_KERNEL(VAddBias); + +TEST_CPU_KERNEL(VRelu); +TEST_CPU_KERNEL(VIdentity); +TEST_CPU_KERNEL(VSquare); +TEST_CPU_KERNEL(VExp); +TEST_CPU_KERNEL(VSigmoid); +TEST_CPU_KERNEL(VTanh); +TEST_CPU_KERNEL(VCopy); + +TEST_CPU_KERNEL(HMax); +TEST_CPU_KERNEL(HSum); + +TEST_CPU_KERNEL(LSTMCtHt); +TEST_CPU_KERNEL(LSTMC1H1); + +TEST_CPU_KERNEL(GRUH1); +TEST_CPU_KERNEL(GRUHtPart1); +TEST_CPU_KERNEL(GRUHtPart2); + +TEST_CPU_KERNEL(NCHW16CMulNC); +TEST_CPU_KERNEL(LayerNorm); +TEST_CPU_KERNEL(CRFDecoding); + +TEST_CPU_KERNEL(SeqPool); +TEST_CPU_KERNEL(EmbSeqPool); +TEST_CPU_KERNEL(MatMul); +TEST_CPU_KERNEL(Softmax); +TEST_CPU_KERNEL(Sgd); +TEST_CPU_KERNEL(VBroadcast); diff --git a/paddle/fluid/operators/layer_norm_op.h b/paddle/fluid/operators/layer_norm_op.h index f564a103963bd93732165596712230b0f37f7f26..8627c83b43cc0ff0f56417c0f7f67effa494cd37 100644 --- a/paddle/fluid/operators/layer_norm_op.h +++ b/paddle/fluid/operators/layer_norm_op.h @@ -230,8 +230,8 @@ class LayerNormKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(bias->numel(), right); auto ker = - jit::Get, platform::CPUPlace>( - right); + jit::KernelFuncs, platform::CPUPlace>::Cache() + .At(right); ker(x.data(), out.data(), mean->data(), var->data(), scale->data(), bias->data(), static_cast(left), static_cast(epsilon), right); diff --git a/paddle/fluid/operators/math.h b/paddle/fluid/operators/math.h new file mode 100644 index 0000000000000000000000000000000000000000..8cc24200d37dffaff1deda2f0181e5875141add0 --- /dev/null +++ b/paddle/fluid/operators/math.h @@ -0,0 +1,42 @@ +// 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 "paddle/fluid/platform/float16.h" +#include "paddle/fluid/platform/hostdevice.h" + +#include "math.h" // NOLINT + +namespace paddle { +namespace operators { + +inline HOSTDEVICE platform::float16 real_exp(platform::float16 x) { + return static_cast(::expf(static_cast(x))); +} + +inline HOSTDEVICE float real_exp(float x) { return ::expf(x); } + +inline HOSTDEVICE double real_exp(double x) { return ::exp(x); } + +inline HOSTDEVICE platform::float16 real_log(platform::float16 x) { + return static_cast(::logf(static_cast(x))); +} + +inline HOSTDEVICE float real_log(float x) { return ::logf(x); } + +inline HOSTDEVICE double real_log(double x) { return ::log(x); } + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index cb200ec8d6ea533d546f3e01a16a48c88b14f677..44cbdf2e9882195819bc3ca047dbac6e2fa4e631 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -12,6 +12,7 @@ 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/operators/math.h" #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/platform/cuda_device_function.h" #include "paddle/fluid/platform/cuda_primitives.h" @@ -20,17 +21,6 @@ namespace paddle { namespace operators { namespace math { -namespace { - -__device__ __forceinline__ float real_log(float x) { return logf(x); } - -__device__ __forceinline__ double real_log(double x) { return log(x); } - -__device__ __forceinline__ platform::float16 real_log( - const platform::float16& val) { - return static_cast(logf(static_cast(val))); -} - template __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, const int N, const int D, @@ -61,7 +51,6 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, Y[blockIdx.x] = -val; } } -} // namespace template class CrossEntropyFunctor { diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h index 0ad57c51be79cd3577b43c9af777bff710308fac..66ce57594a14d8c94737b5dbe83af413628ef1cf 100644 --- a/paddle/fluid/operators/math/fc_compute.h +++ b/paddle/fluid/operators/math/fc_compute.h @@ -30,17 +30,16 @@ inline void FCCompute(const BlasT& blas, const int M, return; } if (relu) { - auto compute = jit::KernelFuncs, - platform::CPUPlace>::Cache() - .At(N); + auto compute = + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + N); for (int i = 0; i < M; i++) { T* dst = Y + i * N; compute(B, dst, dst, N); } } else { - auto compute = jit::KernelFuncs, - platform::CPUPlace>::Cache() - .At(N); + auto compute = + jit::KernelFuncs, platform::CPUPlace>::Cache().At(N); #ifdef PADDLE_WITH_MKLML #pragma omp parallel for #endif diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc index 2a47502614b9cd3df4583992669ab4bf78228181..7af44f2b2ca56f615ca0c8ad4590958af2abe9eb 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cc +++ b/paddle/fluid/operators/math/sequence_pooling.cc @@ -256,8 +256,8 @@ class SequencePoolFunctor { static_cast(input.numel() / input.dims()[0]), jit::SeqPoolType::kSum); auto seqpool = - jit::Get, platform::CPUPlace>( - attr); + jit::KernelFuncs, platform::CPUPlace>::Cache() + .At(attr); for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { attr.h = static_cast(lod[i + 1] - lod[i]); seqpool(src, dst, &attr); diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h index a1cb3f972826a67721b00ce6df0ec48cc34d6e03..d77b6712c548370a99e350b73ab86b170c0e17dc 100644 --- a/paddle/fluid/operators/math/softmax_impl.h +++ b/paddle/fluid/operators/math/softmax_impl.h @@ -82,8 +82,7 @@ class SoftmaxFunctor> { const int kClassDim = 1; // 2D data. Batch x C auto compute_softmax = - jit::KernelFuncs, - platform::CPUPlace>::Cache() + jit::KernelFuncs, platform::CPUPlace>::Cache() .At(in_dims[kClassDim]); compute_softmax(in_data, out_data, in_dims[kClassDim], in_dims[kBatchDim]); } diff --git a/paddle/fluid/operators/ngraph/ngraph_engine.cc b/paddle/fluid/operators/ngraph/ngraph_engine.cc index 41037d9039bb53038af80eafa269ee9246dc9980..cd32200e925193b393f4531b87ed6b1e4291109d 100644 --- a/paddle/fluid/operators/ngraph/ngraph_engine.cc +++ b/paddle/fluid/operators/ngraph/ngraph_engine.cc @@ -29,7 +29,6 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_desc.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/operators/ngraph/ngraph_bridge.h" #include "paddle/fluid/operators/ngraph/ngraph_engine.h" @@ -42,44 +41,75 @@ static ngraph::Shape Ddim2Shape(const framework::DDim& dims) { for (int i = 0; i < dims.size(); ++i) { int k = dims[i]; k = k == 0 ? 1 : k; - sp.push_back(k); + sp.emplace_back(k); } return sp; } +static framework::DDim Shape2Ddim(const ngraph::Shape& shape) { + std::vector dims; + for (size_t i = 0; i < shape.size(); ++i) { + int64_t k = shape[i]; + dims.emplace_back(k); + } + return framework::make_ddim(dims); +} + static std::map pd2ng_type_map = { {framework::proto::VarType::FP32, ngraph::element::f32}, {framework::proto::VarType::FP64, ngraph::element::f64}, {framework::proto::VarType::INT32, ngraph::element::i32}, {framework::proto::VarType::INT64, ngraph::element::i64}, - {framework::proto::VarType::BOOL, ngraph::element::boolean}, -}; - -std::unordered_map> - NgraphEngine::func_cache_ = {}; + {framework::proto::VarType::BOOL, ngraph::element::boolean}}; + +static std::map + ng2pd_type_map = { + {ngraph::element::f32, framework::proto::VarType::FP32}, + {ngraph::element::f64, framework::proto::VarType::FP64}, + {ngraph::element::i32, framework::proto::VarType::INT32}, + {ngraph::element::i64, framework::proto::VarType::INT64}, + {ngraph::element::boolean, framework::proto::VarType::BOOL}}; + +std::vector NgraphEngine::feed_vars = {}; +std::vector NgraphEngine::fetch_vars = {}; +framework::Variable* NgraphEngine::pre_var_ptr = nullptr; +const framework::BlockDesc* NgraphEngine::p_bdesc = nullptr; + +std::unordered_map NgraphEngine::engine_cache = {}; +std::unordered_map>> + NgraphEngine::t_in_cache_ = {}; std::shared_ptr NgraphEngine::backend_ = ngraph::runtime::Backend::create("CPU"); static std::vector> NgraphOpIntervals( - framework::BlockDesc* block) { + std::vector>* ops) { + NgraphEngine::feed_vars.clear(); + NgraphEngine::fetch_vars.clear(); std::vector> intervals; - auto ops = block->AllOps(); - int size = ops.size(); + + int size = ops->size(); int left = 0; - while (left < size && ops.at(left)->Type() != framework::kFeedOpType) { + while (left < size && ops->at(left)->Type() != framework::kFeedOpType) { ++left; } if (left == size) { return intervals; } - while (left < size && ops.at(left)->Type() == framework::kFeedOpType) { + + while (left < size && ops->at(left)->Type() == framework::kFeedOpType) { + for (auto& var_name_item : ops->at(left)->Outputs()) { + for (auto& var_name : var_name_item.second) { + NgraphEngine::feed_vars.emplace_back(var_name); + } + } ++left; } int right = left; - while (right < size && ops.at(right)->Type() != framework::kFetchOpType) { + while (right < size && ops->at(right)->Type() != framework::kFetchOpType) { ++right; } if (right == size) { @@ -87,85 +117,124 @@ static std::vector> NgraphOpIntervals( } if (left >= right) return intervals; + int index = right; + while (index < size && ops->at(index)->Type() == framework::kFetchOpType) { + for (auto& var_name_item : ops->at(index)->Inputs()) { + for (auto& var_name : var_name_item.second) { + NgraphEngine::fetch_vars.emplace_back(var_name); + } + } + ++index; + } + // (left, right - 1) represents indices between feed and fetch int pivot = left; while (pivot < right) { - auto op_type = ops.at(pivot)->Type(); + auto op_type = ops->at(pivot)->Type(); if (NgraphBridge::isRegister(op_type)) { ++pivot; } else { int start = pivot, end = start; while (pivot < right && - (!NgraphBridge::isRegister(ops.at(pivot)->Type()))) { + (!NgraphBridge::isRegister(ops->at(pivot)->Type()))) { ++pivot; ++end; } std::vector interval = {start, end}; - intervals.push_back(interval); + intervals.emplace_back(interval); } } // end while return intervals; } -static void SubstituteNgraphOp(framework::BlockDesc* block, - std::string block_str, - std::vector interval) { - framework::ProgramDesc program; - block->RemoveOp(interval.at(0), interval.at(1)); - auto* ng_op = block->InsertOp(interval.at(0)); - ng_op->SetType("ngraph_engine"); - ng_op->SetAttr("interval", interval); - ng_op->SetAttr("graph", block_str); +static void SubstituteNgraphOp( + std::vector>* ops, + std::string engine_key, std::string block_str, std::vector interval) { + framework::OpDesc ng_op_desc(nullptr); + ng_op_desc.SetType("ngraph_engine"); + ng_op_desc.SetAttr("interval", interval); + ng_op_desc.SetAttr("engine_key", engine_key); + ng_op_desc.SetAttr("graph", block_str); + + ops->erase(ops->begin() + interval[0], ops->begin() + interval[1]); + ops->insert(ops->begin() + interval[0], + framework::OpRegistry::CreateOp(ng_op_desc)); } -// TODO(baojun-nervana): Move EnableNgraph to compile time per PR #15089 -void NgraphEngine::EnableNgraph(const framework::ProgramDesc& program) { -#ifdef PADDLE_WITH_NGRAPH - VLOG(4) << "use_ngraph=True"; - for (size_t bid = 0; bid < program.Size(); ++bid) { - // TODO(baojun-nervana): Remove the const_cast - auto* block = - const_cast(program).MutableBlock(bid); - std::string block_str = block->Proto()->SerializeAsString(); - auto intervals = NgraphOpIntervals(block); - for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) { - SubstituteNgraphOp(block, block_str, *it); - } +std::string SerializedBlock(const std::vector& op_descs) { + framework::proto::BlockDesc block_proto; + framework::BlockDesc block_desc(nullptr, &block_proto); + block_desc.Proto()->set_parent_idx(-1); + block_desc.Proto()->set_idx(0); + + for (auto* op_desc : op_descs) { + auto* op = block_desc.AppendOp(); + *op->Proto() = *op_desc->Proto(); + } + return block_desc.Proto()->SerializeAsString(); +} + +std::string GenerateEngineKey(const framework::BlockDesc& bdesc) { + framework::proto::BlockDesc block_proto; + framework::BlockDesc block_desc(nullptr, &block_proto); + block_desc.Proto()->set_parent_idx(-1); + block_desc.Proto()->set_idx(0); + + for (auto& op_desc : bdesc.AllOps()) { + auto* op = block_desc.AppendOp(); + *op->Proto() = *op_desc->Proto(); + } + auto engine_key = std::to_string( + std::hash()(block_desc.Proto()->SerializeAsString())); + return engine_key; +} + +std::string GenerateEngineKey(const std::vector& engine_inputs, + const std::vector& engine_outputs, + int size) { + std::string engine_hash_key = ""; + for (auto name : engine_inputs) { + engine_hash_key += name; + } + for (auto name : engine_outputs) { + engine_hash_key += name; + } + engine_hash_key += std::to_string(size); + auto engine_key = std::to_string(std::hash()(engine_hash_key)); + return engine_key; +} + +void NgraphEngine::FuseNgraphOps( + const framework::BlockDesc& block_desc, + std::vector>* ops) { + NgraphEngine::p_bdesc = &block_desc; + auto intervals = NgraphOpIntervals(ops); + std::string engine_key = + GenerateEngineKey(feed_vars, fetch_vars, ops->size()); + for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) { + SubstituteNgraphOp(ops, engine_key, "", *it); } -#else - LOG(WARNING) - << "'NGRAPH' is not supported, Please re-compile with WITH_NGRAPH option"; -#endif } NgraphEngine::NgraphEngine(const framework::Scope& scope, const platform::Place& place, - const std::string& serialized_graph, - const std::vector& interval) + const framework::ExecutionContext& ctx) : scope_(scope), place_(place) { + std::string serialized_graph = ctx.Attr("graph"); + auto interval = ctx.Attr>("interval"); + std::string engine_key = ctx.Attr("engine_key"); + var_in_node_map_ = std::make_shared< std::unordered_map>>(); var_node_map_ = std::make_shared< std::unordered_map>>(); - func_cache_key_ = std::to_string(interval[0]) + std::to_string(interval[1]) + - serialized_graph; - - framework::proto::BlockDesc bdesc; - bdesc.ParseFromString(serialized_graph); - framework::BlockDesc block(nullptr, &bdesc); - - Prepare(block, interval); - - BuildNgIO(); - - GetNgFunction(); + GetNgFunction(engine_key, interval); } -void NgraphEngine::Prepare(const framework::BlockDesc& block, - const std::vector& interval) { - for (auto& var : block.AllVars()) { +void NgraphEngine::Prepare(const std::vector& interval) { + for (auto& var : p_bdesc->AllVars()) { if (!(var->GetType() == framework::proto::VarType::SELECTED_ROWS || var->GetType() == framework::proto::VarType::LOD_TENSOR || var->GetType() == framework::proto::VarType::LOD_TENSOR_ARRAY)) { @@ -192,108 +261,57 @@ void NgraphEngine::Prepare(const framework::BlockDesc& block, } } - auto ops_desc = block.AllOps(); - int idx = interval[0]; - while (idx < interval[1]) { - auto op_desc = ops_desc.at(idx); - auto op = framework::OpRegistry::CreateOp(*op_desc); - fused_ops_.push_back(std::move(op)); - ++idx; - } - - while (ops_desc.at(idx)->Type() != framework::kFetchOpType) { - auto op_desc = ops_desc.at(idx); - for (auto& var_name_item : op_desc->Inputs()) { - for (auto& var_name : var_name_item.second) { - post_op_inputs_.insert(var_name); - } - } - ++idx; - } - - while (idx < static_cast(ops_desc.size()) && - ops_desc.at(idx)->Type() == framework::kFetchOpType) { - std::string fetch_target_name = ops_desc.at(idx)->Input("X")[0]; - fetches_.insert(fetch_target_name); - ++idx; - } - - if (ops_desc.at(interval.at(0) - 1)->Type() == framework::kFeedOpType && - ops_desc.at(interval.at(1))->Type() == framework::kFetchOpType) { - ng_op_state_ = OpState::FULL; + std::vector ops_desc; + for (auto op_desc : p_bdesc->AllOps()) { + ops_desc.emplace_back(op_desc); } - for (auto* op_desc : ops_desc) { + for (auto op_desc : ops_desc) { if (op_desc->Type().find("_grad") != std::string::npos) { - ng_op_state_ = ng_op_state_ == OpState::FULL ? OpState::FULL_TRAIN - : OpState::PARTIAL_TRAIN; + this->is_test_ = false; break; } } - if (ng_op_state_ != OpState::FULL_TRAIN && - ng_op_state_ != OpState::PARTIAL_TRAIN) { - ng_op_state_ = ng_op_state_ == OpState::FULL ? OpState::FULL_TEST - : OpState::PARTIAL_TEST; + if (interval[0] > 0 && + ops_desc.at(interval[0] - 1)->Type() == framework::kFeedOpType && + interval[1] < static_cast(ops_desc.size()) && + ops_desc.at(interval.at(1))->Type() == framework::kFetchOpType) { + this->op_state_ = OpState::FULL; } -} -void NgraphEngine::GetNgInputShape( - std::shared_ptr op) { - framework::RuntimeContext ctx(op->Inputs(), op->Outputs(), scope_); - op->RuntimeInferShape(scope_, place_, ctx); - for (auto& var_name_item : op->Inputs()) { - for (auto& var_name : var_name_item.second) { - auto* var = scope_.FindVar(var_name); - if (var && var->IsType()) { - auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); - auto sp = Ddim2Shape(tensor_pd->dims()); - if (std::find(var_in_.begin(), var_in_.end(), var_name) != - var_in_.end()) { - if (var_node_map_->find(var_name) == var_node_map_->end()) { - // auto ng_type = pd2ng_type_map.at(GetDataTypeOfVar(var)); - auto ng_type = var_type_map_.at(var_name); - auto prm = - std::make_shared(ng_type, sp, true); - (*var_node_map_)[var_name] = prm; - (*var_in_node_map_)[var_name] = prm; - } - } - } - } + if (this->op_state_ == OpState::FULL) { + this->op_state_ = this->is_test_ ? OpState::FULL_TEST : OpState::FULL_TRAIN; + } else { + this->op_state_ = + this->is_test_ ? OpState::PARTIAL_TEST : OpState::PARTIAL_TRAIN; } -} -void NgraphEngine::BuildNgNodes() { - for (auto& op : fused_ops_) { - for (auto& var_name_item : op->Outputs()) { + int idx = interval[0]; + while (idx < interval[1]) { + this->fused_ops_.emplace_back( + framework::OpRegistry::CreateOp(*(ops_desc[idx]))); + ++idx; + } + while (ops_desc.at(idx)->Type() != framework::kFetchOpType) { + auto op_desc = ops_desc.at(idx); + for (auto& var_name_item : op_desc->Inputs()) { for (auto& var_name : var_name_item.second) { - if (var_node_map_->find(var_name) == var_node_map_->end()) { - auto* var = scope_.FindVar(var_name); - if (var && var->IsType()) { - auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); - auto& ddim = tensor_pd->dims(); - auto ng_shape = Ddim2Shape(ddim); - auto ng_type = var_type_map_.at(var_name); - auto prm = std::make_shared(ng_type, - ng_shape, true); - (*var_node_map_)[var_name] = prm; - } - } + this->post_op_inputs_.insert(var_name); } } + ++idx; } - NgraphBridge ngb(var_node_map_); - for (auto& op : fused_ops_) { - ngb.BuildNgNode(op); - } + + BuildNgIO(ops_desc, interval); } -void NgraphEngine::BuildNgIO() { +void NgraphEngine::BuildNgIO(const std::vector& ops_desc, + const std::vector& interval) { std::unordered_set inputs; std::unordered_set outputs; - - for (auto& op : fused_ops_) { + for (int i = interval[0]; i < interval[1]; ++i) { + auto op = ops_desc[i]; for (auto& var_name_item : op->Inputs()) { for (auto& var_name : var_name_item.second) { inputs.insert(var_name); @@ -302,15 +320,11 @@ void NgraphEngine::BuildNgIO() { std::find(var_in_.begin(), var_in_.end(), var_name) == var_in_.end()) { // fill var_in here to keep lhs and rhs order - var_in_.push_back(var_name); + this->var_in_.emplace_back(var_name); } } } - if (op->Type() != "fill_constant") { - GetNgInputShape(op); - } - for (auto& var_name_item : op->Outputs()) { PADDLE_ENFORCE_LE(var_name_item.second.size(), 1, "op %s has more than 1 output - Not handling yet", @@ -322,172 +336,278 @@ void NgraphEngine::BuildNgIO() { } // var_out.clear(); - for (auto& op : fused_ops_) { + for (int i = interval[0]; i < interval[1]; ++i) { + auto op = ops_desc[i]; for (auto& var_name_item : op->Outputs()) { PADDLE_ENFORCE_LE(var_name_item.second.size(), 1, "op %s has more than 1 output - Not handling yet", op->Type()); for (auto& var_name : var_name_item.second) { - switch (ng_op_state_) { + switch (this->op_state_) { case OpState::PARTIAL_TEST: if (post_op_inputs_.find(var_name) != post_op_inputs_.end() || - fetches_.find(var_name) != fetches_.end()) { - var_out_.push_back(var_name); + find(fetch_vars.begin(), fetch_vars.end(), var_name) != + fetch_vars.end()) { + this->var_out_.emplace_back(var_name); } break; case OpState::FULL_TEST: - if (fetches_.find(var_name) != fetches_.end()) { - var_out_.push_back(var_name); + if (find(fetch_vars.begin(), fetch_vars.end(), var_name) != + fetch_vars.end()) { + this->var_out_.emplace_back(var_name); } break; case OpState::PARTIAL_TRAIN: - if (fetches_.find(var_name) != fetches_.end() || + if (find(fetch_vars.begin(), fetch_vars.end(), var_name) != + fetch_vars.end() || post_op_inputs_.find(var_name) != post_op_inputs_.end() || persistables_.find(var_name) != persistables_.end()) { - var_out_.push_back(var_name); + this->var_out_.emplace_back(var_name); } break; case OpState::FULL_TRAIN: - if (fetches_.find(var_name) != fetches_.end() || + if (find(fetch_vars.begin(), fetch_vars.end(), var_name) != + fetch_vars.end() || persistables_.find(var_name) != persistables_.end()) { - var_out_.push_back(var_name); + this->var_out_.emplace_back(var_name); } break; default: - var_out_.push_back(var_name); + this->var_out_.emplace_back(var_name); } } } } + for (size_t i = 0; i < var_in_.size(); ++i) { + auto var_name = var_in_[i]; + if (persistables_.find(var_name) == persistables_.end()) { + var_in_updates_.emplace_back(i); + } + } } -void NgraphEngine::BuildNgFunction() { +void NgraphEngine::GetNgInputShape() { + for (auto& var_name : var_in_) { + auto* var = scope_.FindVar(var_name); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + auto sp = Ddim2Shape(tensor_pd->dims()); + auto ng_type = var_type_map_[var_name]; + auto prm = std::make_shared(ng_type, sp, true); + (*var_node_map_)[var_name] = prm; + (*var_in_node_map_)[var_name] = prm; + } + } +} + +void NgraphEngine::BuildNgNodes() { + for (auto& op : fused_ops_) { + for (auto& var_name_item : op->Outputs()) { + for (auto& var_name : var_name_item.second) { + if (var_node_map_->find(var_name) == var_node_map_->end()) { + auto* var = scope_.FindVar(var_name); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + auto& ddim = tensor_pd->dims(); + auto ng_shape = Ddim2Shape(ddim); + auto ng_type = var_type_map_[var_name]; + auto prm = std::make_shared(ng_type, + ng_shape, true); + (*var_node_map_)[var_name] = prm; + } + } + } + } + } + + NgraphBridge ngb(var_node_map_); + for (auto& op : fused_ops_) { + ngb.BuildNgNode(op); + } +} + +void NgraphEngine::RunInferShape() { + for (auto& op : fused_ops_) { + framework::RuntimeContext ctx(op->Inputs(), op->Outputs(), scope_); + op->RuntimeInferShape(scope_, place_, ctx); + } +} + +void NgraphEngine::BuildNgFunction(const std::vector& interval) { + Prepare(interval); + RunInferShape(); + GetNgInputShape(); BuildNgNodes(); ngraph_function_ = nullptr; ngraph::NodeVector func_outputs; ngraph::ParameterVector func_inputs; for (auto& vo : var_out_) { - func_outputs.push_back(var_node_map_->at(vo)); + func_outputs.emplace_back(var_node_map_->at(vo)); } for (auto& vi : var_in_) { std::shared_ptr prm = std::dynamic_pointer_cast( var_in_node_map_->at(vi)); - func_inputs.push_back(prm); + func_inputs.emplace_back(prm); } ngraph_function_ = std::make_shared(func_outputs, func_inputs); } -void NgraphEngine::GetNgFunction() { - bool cache_on = true; - if (cache_on) { - std::string input_shape_str; - for (auto& var_name : var_in_) { - auto shape = var_node_map_->at(var_name)->get_shape(); - for (size_t i = 0; i < shape.size(); ++i) { - input_shape_str += std::to_string(shape.at(i)); +void NgraphEngine::GetNgFunction(std::string engine_key, + const std::vector& interval) { + bool use_cache = true; + if (use_cache) { + this->func_cache_key_ = ""; + for (int i = 0; i < std::min(static_cast(feed_vars.size()), 10); ++i) { + auto* var = scope_.FindVar(feed_vars[i]); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + auto dims = tensor_pd->dims(); + for (int j = 0; j < dims.size(); ++j) { + func_cache_key_ += std::to_string(dims[j]); + } } } - func_cache_key_ = input_shape_str + func_cache_key_; - if (func_cache_.find(func_cache_key_) != func_cache_.end()) { - ngraph_function_ = func_cache_.at(func_cache_key_); - } else { - BuildNgFunction(); - func_cache_[func_cache_key_] = ngraph_function_; + func_cache_key_ += std::to_string(interval[0]) + "_" + + std::to_string(interval[1]) + engine_key; + func_cache_key_ = std::to_string(std::hash()(func_cache_key_)); + + if (engine_cache.find(func_cache_key_) != engine_cache.end()) { + if (engine_cache[func_cache_key_].persistables.size() == 0) { + engine_cache.clear(); + t_in_cache_.clear(); + } else { + auto var_name = engine_cache[func_cache_key_].persistables.begin(); + framework::Variable* var = scope_.FindVar(*var_name); + if (var != pre_var_ptr) { + engine_cache.clear(); + t_in_cache_.clear(); + } + pre_var_ptr = var; + } + } + + if (engine_cache.find(func_cache_key_) == engine_cache.end()) { + BuildNgFunction(interval); + engine_cache[func_cache_key_].ngraph_function = this->ngraph_function_; + engine_cache[func_cache_key_].persistables = this->persistables_; + engine_cache[func_cache_key_].var_in_updates = this->var_in_updates_; + engine_cache[func_cache_key_].var_in = this->var_in_; + engine_cache[func_cache_key_].var_out = this->var_out_; + engine_cache[func_cache_key_].is_test = this->is_test_; } } else { - BuildNgFunction(); + BuildNgFunction(interval); } } void NgraphEngine::Run(const framework::Scope& scope, const platform::Place& place) const { - std::vector> t_in; - std::vector> t_out; + std::shared_ptr ng_func; + const std::set* p_persistables; + const std::vector* p_var_in_updates; + const std::vector* p_var_in; + const std::vector* p_var_out; + bool is_test; + + bool use_cache = true; + if (use_cache) { + PADDLE_ENFORCE(engine_cache.find(func_cache_key_) != engine_cache.end(), + "Cannot find cached data to run ngraph function"); + ng_func = engine_cache[func_cache_key_].ngraph_function; + p_persistables = &(engine_cache[func_cache_key_].persistables); + p_var_in_updates = &(engine_cache[func_cache_key_].var_in_updates); + p_var_in = &(engine_cache[func_cache_key_].var_in); + p_var_out = &(engine_cache[func_cache_key_].var_out); + is_test = engine_cache[func_cache_key_].is_test; + } else { + ng_func = ngraph_function_; + p_persistables = &this->persistables_; + p_var_in_updates = &this->var_in_updates_; + p_var_in = &this->var_in_; + p_var_out = &this->var_out_; + is_test = this->is_test_; + } - for (size_t i = 0; i < var_in_.size(); ++i) { - auto vi = var_in_.at(i); - auto sp = var_node_map_->at(vi)->get_shape(); - std::shared_ptr ti; - auto* var = scope.FindVar(vi); - if (var && var->IsType()) { - auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var); - PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()), - "Ensure ngraph tensor layout align with paddle tensor"); - auto ng_type = var_type_map_.at(vi); - if (ng_type == ngraph::element::f32) { - auto pd_arr = tensor_pd->mutable_data(place); - ti = backend_->create_tensor(ngraph::element::f32, sp, pd_arr); - } else if (ng_type == ngraph::element::i32) { - const int* arr = tensor_pd->data(); - ti = backend_->create_tensor(ngraph::element::i32, sp, - const_cast(arr)); - } else if (ng_type == ngraph::element::i64) { - auto pd_arr = tensor_pd->mutable_data(place); - ti = backend_->create_tensor(ngraph::element::i64, sp, pd_arr); - } else if (ng_type == ngraph::element::f64) { - auto pd_arr = tensor_pd->mutable_data(place); - ti = backend_->create_tensor(ngraph::element::f64, sp, pd_arr); - } else if (ng_type == ngraph::element::boolean) { - auto pd_arr = tensor_pd->mutable_data(place); - ti = backend_->create_tensor(ngraph::element::boolean, sp, pd_arr); + std::vector>* p_t_in; + std::vector> t_in = {}; + + auto m_parameters = ng_func->get_parameters(); + auto m_results = ng_func->get_results(); + if (is_test && use_cache && + t_in_cache_.find(func_cache_key_) != t_in_cache_.end()) { + p_t_in = &(t_in_cache_[func_cache_key_]); + for (size_t i = 0; i < p_var_in_updates->size(); ++i) { + int index = p_var_in_updates->at(i); + auto vi = p_var_in->at(index); + auto sp = m_parameters[index]->get_shape(); + auto ng_type = m_parameters[index]->get_element_type(); + std::shared_ptr ti; + auto* var = scope.FindVar(vi); + if (var && var->IsType()) { + auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var); + void* pd_arr = tensor_pd->mutable_data(place, ng2pd_type_map[ng_type]); + ti = backend_->create_tensor(ng_type, sp, pd_arr); + (*p_t_in)[index] = ti; } else { - PADDLE_THROW("Data type not handling for var %s", vi); + PADDLE_THROW("Cannot find var or tensor with var name %s", vi); } + } + } else { + if (is_test && use_cache) { + p_t_in = &(t_in_cache_[func_cache_key_]); } else { - PADDLE_THROW("Cannot find var or tensor with var name %s", vi); + p_t_in = &t_in; } - bool is_test = (ng_op_state_ == OpState::PARTIAL_TEST || - ng_op_state_ == OpState::FULL_TEST) - ? true - : false; - bool is_persistable = - (persistables_.find(vi) != persistables_.end()) ? true : false; - if (is_test && is_persistable) { - ti->set_stale(false); + + for (size_t i = 0; i < p_var_in->size(); ++i) { + auto vi = p_var_in->at(i); + auto sp = m_parameters[i]->get_shape(); + auto ng_type = m_parameters[i]->get_element_type(); + std::shared_ptr ti; + auto* var = scope.FindVar(vi); + if (var && var->IsType()) { + auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var); + void* pd_arr = tensor_pd->mutable_data(place, ng2pd_type_map[ng_type]); + PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()), + "Ensure ngraph tensor layout align with paddle tensor"); + ti = backend_->create_tensor(ng_type, sp, pd_arr); + } else { + PADDLE_THROW("Cannot find var or tensor with var name %s", vi); + } + bool is_persistable = + (p_persistables->find(vi) != p_persistables->end()) ? true : false; + if (is_test && is_persistable) { + ti->set_stale(false); + } + (*p_t_in).emplace_back(ti); } - t_in.push_back(ti); } - for (size_t i = 0; i < var_out_.size(); ++i) { - auto vo = var_out_[i]; + std::vector> t_out = {}; + for (size_t i = 0; i < p_var_out->size(); ++i) { + auto vo = p_var_out->at(i); auto* var = scope.FindVar(vo); - std::shared_ptr to; if (var && var->IsType()) { + auto sp = m_results[i]->get_shape(); + var->GetMutable()->Resize(Shape2Ddim(sp)); auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var); - auto dd = tensor_pd->dims(); - ngraph::Shape sp = Ddim2Shape(dd); - auto ng_type = var_type_map_.at(vo); - if (ng_type == ngraph::element::f32) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ng_type, sp, pd_arr); - } else if (ng_type == ngraph::element::i64) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ng_type, sp, pd_arr); - } else if (ng_type == ngraph::element::i32) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ng_type, sp, pd_arr); - } else if (ng_type == ngraph::element::f64) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ng_type, sp, pd_arr); - } else if (ng_type == ngraph::element::boolean) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ng_type, sp, pd_arr); - } else { - PADDLE_THROW("Data type not handled in for var %s", vo); - } - t_out.push_back(to); + auto ng_type = m_results[i]->get_element_type(); + void* pd_arr = tensor_pd->mutable_data(place, ng2pd_type_map[ng_type]); + std::shared_ptr to = + backend_->create_tensor(ng_type, sp, pd_arr); + t_out.emplace_back(to); } else { PADDLE_THROW("Cannot find var or tensor with var name %s", vo); } } - auto handle = backend_->compile(ngraph_function_); - handle->call_with_validate(t_out, t_in); + auto handle = backend_->compile(ng_func); + handle->call_with_validate(t_out, *p_t_in); } // NgraphEngine::Run } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/ngraph/ngraph_engine.h b/paddle/fluid/operators/ngraph/ngraph_engine.h index bf5ff2a743b0edb69163e674d36c56a02c0b4153..fef51464b5702e61d052f28050f6aefaecf0f615 100644 --- a/paddle/fluid/operators/ngraph/ngraph_engine.h +++ b/paddle/fluid/operators/ngraph/ngraph_engine.h @@ -12,12 +12,18 @@ 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. */ +#ifndef PADDLE_FLUID_OPERATORS_NGRAPH_NGRAPH_ENGINE_H_ +#define PADDLE_FLUID_OPERATORS_NGRAPH_NGRAPH_ENGINE_H_ +#include +#include #include #include +#include #include #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/var_desc.h" #include "ngraph/ngraph.hpp" @@ -33,29 +39,47 @@ enum class OpState { /* nGraph support state on ops */ UNKNOWN /* Output all for debug purpose */ }; +// cache engine repetitives +struct EngineCache { + std::shared_ptr ngraph_function; + std::set persistables; + std::vector var_in; + std::vector var_out; + std::vector var_in_updates; + bool is_test = true; +}; + // perform graph build through bridge and execute computation class NgraphEngine { public: explicit NgraphEngine(const framework::Scope& scope, const platform::Place& place, - const std::string& serialized_graph, - const std::vector& interval); + const framework::ExecutionContext& ctx); void Run(const framework::Scope& scope, const platform::Place& place) const; - static void EnableNgraph(const framework::ProgramDesc& program); + static const framework::BlockDesc* p_bdesc; + static std::vector feed_vars, fetch_vars; + + static void FuseNgraphOps( + const framework::BlockDesc& prog, + std::vector>* ops); private: - static std::unordered_map> - func_cache_; + static std::unordered_map engine_cache; + static std::unordered_map< + std::string, std::vector>> + t_in_cache_; + static framework::Variable* pre_var_ptr; + const framework::Scope& scope_; const platform::Place& place_; std::vector> fused_ops_; std::unordered_map var_type_map_; - std::unordered_set persistables_; - std::unordered_set fetches_; + std::set persistables_; std::unordered_set post_op_inputs_; - OpState ng_op_state_ = OpState::UNKNOWN; + OpState op_state_ = OpState::UNKNOWN; + bool is_test_{true}; std::string func_cache_key_; // ngraph backend eg. CPU @@ -66,6 +90,8 @@ class NgraphEngine { std::vector var_in_; // var_name of outputs from fetch in order std::vector var_out_; + // non-persitable var_in + std::vector var_in_updates_; // map input vars to nodes std::shared_ptr< std::unordered_map>> @@ -74,20 +100,23 @@ class NgraphEngine { std::shared_ptr< std::unordered_map>> var_node_map_; - // prepare info for nraph engine - void Prepare(const framework::BlockDesc& block, - const std::vector& interval); + // prepare info for ngraph engine need + void Prepare(const std::vector& interval); + // get ngraph engine input and output list + void BuildNgIO(const std::vector& op_descs, + const std::vector& interval); // get ngraph input and define ngraph input parameters - void GetNgInputShape(std::shared_ptr op); + void GetNgInputShape(); // Call ngraph bridge to map ops void BuildNgNodes(); - // get the ngraph input and output var list - void BuildNgIO(); + // run paddle RuntimeInferShape to get the tensor shape + void RunInferShape(); // build ngraph function call - void BuildNgFunction(); + void BuildNgFunction(const std::vector& interval); // Check cache for ngraph function or otherwise build the function - void GetNgFunction(); + void GetNgFunction(std::string engine_key, const std::vector& interval); }; } // namespace operators } // namespace paddle +#endif // PADDLE_FLUID_OPERATORS_NGRAPH_NGRAPH_ENGINE_H_ diff --git a/paddle/fluid/operators/ngraph/ngraph_engine_op.cc b/paddle/fluid/operators/ngraph/ngraph_engine_op.cc index 3051ca123b29658d3e9a35239ad00f621a297cb5..f941f917c82b3b74a35739c08112233fd0a3477c 100644 --- a/paddle/fluid/operators/ngraph/ngraph_engine_op.cc +++ b/paddle/fluid/operators/ngraph/ngraph_engine_op.cc @@ -29,6 +29,7 @@ class NgraphEngineOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Xs", "A list of inputs.").AsDispensable(); AddOutput("Ys", "A list of outputs").AsDispensable(); AddAttr("graph", "the graph."); + AddAttr("engine_key", "the engine hash key."); AddAttr>("interval", "op interval supported by ngraph"); AddComment("ngraph engine operator."); } diff --git a/paddle/fluid/operators/ngraph/ngraph_engine_op.h b/paddle/fluid/operators/ngraph/ngraph_engine_op.h index 2f194a9b8766316fc645f7e22e21fff048fb7d63..c9b2a3970e17c1a06fa0cc67aa15df304a30656e 100644 --- a/paddle/fluid/operators/ngraph/ngraph_engine_op.h +++ b/paddle/fluid/operators/ngraph/ngraph_engine_op.h @@ -46,10 +46,8 @@ class NgraphEngineKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto& scope = ctx.scope(); auto place = ctx.GetPlace(); - std::string serialized_graph = ctx.Attr("graph"); - auto interval = ctx.Attr>("interval"); - NgraphEngine ngraph_engine(scope, place, serialized_graph, interval); + NgraphEngine ngraph_engine(scope, place, ctx); ngraph_engine.Run(scope, place); } }; diff --git a/paddle/fluid/operators/optimizers/sgd_op.h b/paddle/fluid/operators/optimizers/sgd_op.h index c9c9f530fe846c1713ad176e05a377996d04470b..5dd5f67e004c63e294152239ab7bd3db26542eed 100644 --- a/paddle/fluid/operators/optimizers/sgd_op.h +++ b/paddle/fluid/operators/optimizers/sgd_op.h @@ -48,7 +48,8 @@ class SGDOpKernel : public framework::OpKernel { T *out_data = param_out->mutable_data(ctx.GetPlace()); auto sgd = - jit::Get, platform::CPUPlace>(attr); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr); sgd(lr, param_data, grad_data, &rows_idx, out_data, &attr); } else if (grad_var->IsType()) { // TODO(qijun): In Sparse SGD operator, in-place update is enforced. @@ -82,7 +83,8 @@ class SGDOpKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(attr.grad_width, attr.param_width); auto sgd = - jit::Get, platform::CPUPlace>(attr); + jit::KernelFuncs, platform::CPUPlace>::Cache().At( + attr); sgd(lr, param_data, grad_data, rows_data, out_data, &attr); } else { PADDLE_THROW("Unsupported Variable Type of Grad"); diff --git a/paddle/fluid/operators/reader/buffered_reader.cc b/paddle/fluid/operators/reader/buffered_reader.cc index 52e96c4fb3a058057f5acd5e30b7a0e869aefacc..134807092d59329ce93381da67a98b8230db5767 100644 --- a/paddle/fluid/operators/reader/buffered_reader.cc +++ b/paddle/fluid/operators/reader/buffered_reader.cc @@ -17,6 +17,7 @@ #include #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace reader { @@ -50,8 +51,9 @@ BufferedReader::BufferedReader( .Get(place_))) ->stream(); events.resize(buffer_size); - for (auto &event : events) + for (auto &event : events) { PADDLE_ENFORCE(cudaEventCreateWithFlags(&event, cudaEventDisableTiming)); + } PADDLE_ENFORCE(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking)); } #endif @@ -84,12 +86,15 @@ void BufferedReader::ReadAsync(size_t i) { #ifdef PADDLE_WITH_CUDA // NOTE(liangdun): using async copy instead of TensorCopySync - // TensorCopySync would block other stream + // TensorCopySync would block other stream, because TensorCopySync + // issues the copying command to the default stream, it will make two + // commands from different streams cannot run concurrently. if (platform::is_gpu_place(place_)) { platform::SetDeviceId(boost::get(place_).device); PADDLE_ENFORCE(cudaStreamWaitEvent(stream, events[i], 0)); TensorVec &gpu = gpu_buffer_[i]; gpu.resize(cpu.size()); + platform::RecordEvent record_event("BufferedReader:MemoryCopy"); for (size_t i = 0; i < cpu.size(); ++i) { gpu[i].Resize(cpu[i].dims()); gpu[i].set_layout(cpu[i].layout()); @@ -98,20 +103,20 @@ void BufferedReader::ReadAsync(size_t i) { auto gpu_ptr = gpu[i].mutable_data(place_, cpu[i].type()); auto size = cpu[i].numel() * paddle::framework::SizeOfType(cpu[i].type()); - if (platform::is_cuda_pinned_place(cpu_place)) + if (platform::is_cuda_pinned_place(cpu_place)) { memory::Copy(boost::get(place_), gpu_ptr, boost::get(cpu_place), cpu_ptr, size, stream); - else if ((platform::is_gpu_place(cpu_place))) + } else if ((platform::is_gpu_place(cpu_place))) { memory::Copy(boost::get(place_), gpu_ptr, boost::get(cpu_place), cpu_ptr, size, stream); - else - // if cpu place is not pinned, async copy is slower than sync copy, - // so we use sync copy instead. + } else { + // TODO(zcd): The default stream should not be used here. memory::Copy(boost::get(place_), gpu_ptr, boost::get(cpu_place), cpu_ptr, size, 0); + } gpu[i].set_lod(cpu[i].lod()); } PADDLE_ENFORCE(cudaStreamSynchronize(stream)); diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 88c968a0eaae8a2ac6f14ede9348c837bcd92d76..2898a62ddbac524ceb212cac5f34aeda3b1e01cb 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -282,7 +282,9 @@ class RecurrentOp : public RecurrentBase { // Every inputs are linked now, execute! executor.Run(*program, &cur_scope, block->ID(), - false /*create_local_scope*/); + false /*create_local_scope*/, true /*create_vars*/, + std::vector() /*skip_ref_cnt_vars*/, + true /*force_disable_gc*/); // get device context from pool platform::DeviceContextPool &pool = @@ -398,7 +400,9 @@ class RecurrentGradOp : public RecurrentBase { VLOG(5) << "Recurrent memory linking finished "; // Run step block with cur_scope executor.Run(*program, &cur_scope, block->ID(), - false /*create_local_scope*/); + false /*create_local_scope*/, true /*create_vars*/, + std::vector() /*skip_ref_cnt_vars*/, + true /*force_disable_gc*/); VLOG(5) << "executor.Run finished "; diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index 37f69426b62fedf8cbeca68105fb86fb4ea72eab..2b429380fbfc007f5936bff96e0924d93abc81f5 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -219,14 +219,6 @@ class ReshapeKernel { std::vector(shape_data, shape_data + shape_tensor->numel()); out_dims = ReshapeOp::ValidateShape(shape, in->dims()); } - if (!in->lod().empty()) { - PADDLE_ENFORCE_EQ( - out_dims[0], in->dims()[0], - "Reshape operator cannot reshape an input sequence batch " - "into an output sequence batch that has a different " - "number of time steps. Please consider using " - "sequence_reshape op."); - } out->mutable_data(ctx.GetPlace(), in->type()); framework::TensorCopy( diff --git a/paddle/fluid/operators/selu_op.h b/paddle/fluid/operators/selu_op.h index bdb506885c932708803fe8d84ee705aee0fe02b4..b2fc834c42f65ff3521b6267ed2f32fabbab4e4d 100644 --- a/paddle/fluid/operators/selu_op.h +++ b/paddle/fluid/operators/selu_op.h @@ -15,13 +15,12 @@ limitations under the License. */ #pragma once #include #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math.h" #include "paddle/fluid/platform/for_range.h" + namespace paddle { namespace operators { -static HOSTDEVICE float real_exp(float x) { return expf(x); } -static HOSTDEVICE float real_exp(double x) { return exp(x); } - template struct SeluFunctor { SeluFunctor(const T* x_data_ptr, float alpha, float scale, T* y_data_ptr) diff --git a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc index d3dcd1f96a986d2450c8af780a12183f7dfc66d5..f357c9c08d042b69259f229955922f2f11b52c63 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc +++ b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc @@ -22,9 +22,6 @@ class SequenceEnumerateOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - if (ctx->IsRuntime()) { - return; - } PADDLE_ENFORCE( ctx->HasInput("X"), "Input(X) of SequecceEnumerate operator should not be null."); @@ -62,6 +59,8 @@ class SequenceEnumerateOpMaker : public framework::OpProtoAndCheckerMaker { }); AddAttr("pad_value", "(int) The enumerate sequence padding value.") .SetDefault(0); + AddAttr(framework::kAllKernelsMustComputeRuntimeShape, "") + .SetDefault(true); AddComment(R"DOC( Sequence Enumerate Operator. diff --git a/paddle/fluid/operators/sequence_ops/sequence_softmax_op.cu b/paddle/fluid/operators/sequence_ops/sequence_softmax_op.cu index cc5e9821903fb7a726f52177df1d17757f697411..a9dc0a4fda253db9bb0d33c4a25fbba36492f35b 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_softmax_op.cu +++ b/paddle/fluid/operators/sequence_ops/sequence_softmax_op.cu @@ -14,6 +14,7 @@ limitations under the License. */ #include #include // NOLINT +#include "paddle/fluid/operators/math.h" #include "paddle/fluid/operators/sequence_ops/sequence_softmax_op.h" namespace paddle { @@ -21,9 +22,6 @@ namespace operators { using LoDTensor = framework::LoDTensor; -__device__ __forceinline__ float real_exp(float x) { return expf(x); } -__device__ __forceinline__ double real_exp(double x) { return exp(x); } - template using BlockReduce = cub::BlockReduce; diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cu b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cu index 2a4570ef5cec0bee07efd69a2efd1a079ff33df5..aea69de6434a38aa834ff14f6d3d15ad5bbfc3e6 100644 --- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cu +++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cu @@ -12,6 +12,7 @@ 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 "cub/cub.cuh" +#include "paddle/fluid/operators/math.h" #include "paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.h" #include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/hostdevice.h" @@ -21,11 +22,6 @@ namespace operators { using Tensor = framework::Tensor; -static HOSTDEVICE float real_exp(float x) { return expf(x); } -static HOSTDEVICE float real_exp(double x) { return exp(x); } -static HOSTDEVICE float real_log(float x) { return logf(x); } -static HOSTDEVICE float real_log(double x) { return log(x); } - static constexpr int kNumCUDAThreads = 512; static constexpr int kNumMaxinumNumBlocks = 4096; diff --git a/paddle/fluid/operators/slice_op.cu b/paddle/fluid/operators/slice_op.cu index 5efecb78d1a4eaffc3a9c62e1e82a9bcb5922748..1af57b89a3506a7211cca9233ef624b0a83e77a0 100644 --- a/paddle/fluid/operators/slice_op.cu +++ b/paddle/fluid/operators/slice_op.cu @@ -12,18 +12,138 @@ 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 +#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/slice_op.h" +#include "paddle/fluid/platform/cuda_device_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/float16.h" + +namespace paddle { +namespace operators { + +using platform::PADDLE_CUDA_NUM_THREADS; + +template +__global__ void Padding(const paddle::platform::float16* d_out, + const int* out_dims, const int* in_dims, + const int* offsets, int64_t n, + paddle::platform::float16* d_in) { + int64_t out_idx = threadIdx.x + blockDim.x * blockIdx.x; + if (out_idx < n) { + int coords[D] = {0}; + for (int i = D - 1; i >= 0; --i) { + coords[i] = out_idx % out_dims[i]; + out_idx /= out_dims[i]; + coords[i] += offsets[i]; + } + + int64_t in_idx = 0; + for (int i = 0; i < D - 1; ++i) { + in_idx += coords[i] * in_dims[i + 1]; + } + in_idx += coords[D - 1]; + + d_in[in_idx] = d_out[out_idx]; + } +} + +template <> +class SliceGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_out = ctx.Input(framework::GradVarName("Out")); + auto* d_in = ctx.Output(framework::GradVarName("Input")); + d_in->mutable_data(ctx.GetPlace()); + + auto out_dims = d_out->dims(); + auto in_dims = d_in->dims(); + int rank = out_dims.size(); + std::vector offsets(rank, 0); + auto axes = ctx.Attr>("axes"); + auto starts = ctx.Attr>("starts"); + + for (size_t i = 0; i < starts.size(); ++i) { + if (starts[i] < 0) { + starts[i] += in_dims[axes[i]]; + } + offsets[axes[i]] = std::max(starts[i], 0); + } + + math::SetConstant + set_zero; + auto& dev_ctx = + ctx.template device_context(); + set_zero(dev_ctx, d_in, static_cast(0)); + + int64_t numel = d_out->numel(); + dim3 blocks((numel - 1) / PADDLE_CUDA_NUM_THREADS + 1, 1, 1); + dim3 threads(PADDLE_CUDA_NUM_THREADS, 1, 1); + auto stream = ctx.cuda_device_context().stream(); + + auto out_shape = framework::vectorize2int(out_dims); + thrust::device_vector out_dims_vec(out_shape.begin(), out_shape.end()); + auto in_shape = framework::vectorize2int(in_dims); + thrust::device_vector in_dims_vec(in_shape.begin(), in_shape.end()); + thrust::device_vector offsets_vec(offsets.begin(), offsets.end()); + const int* out_dims_ptr = thrust::raw_pointer_cast(out_dims_vec.data()); + const int* in_dims_ptr = thrust::raw_pointer_cast(in_dims_vec.data()); + const int* offsets_ptr = thrust::raw_pointer_cast(offsets_vec.data()); + + switch (rank) { + case 1: + Padding<1><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 2: + Padding<2><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 3: + Padding<3><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 4: + Padding<4><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 5: + Padding<5><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 6: + Padding<6><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + } + } +}; + +} // namespace operators +} // namespace paddle namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( slice, ops::SliceKernel, ops::SliceKernel, ops::SliceKernel, - ops::SliceKernel); + ops::SliceKernel, + ops::SliceKernel); REGISTER_OP_CUDA_KERNEL( slice_grad, ops::SliceGradKernel, ops::SliceGradKernel, ops::SliceGradKernel, - ops::SliceGradKernel); + ops::SliceGradKernel, + ops::SliceGradKernel); diff --git a/paddle/fluid/operators/sync_batch_norm_op.cc b/paddle/fluid/operators/sync_batch_norm_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d6cf27fd779eeddc94c1839e46892a99f61bd1bf --- /dev/null +++ b/paddle/fluid/operators/sync_batch_norm_op.cc @@ -0,0 +1,20 @@ +/* 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/operators/batch_norm_op.h" + +namespace ops = paddle::operators; +REGISTER_OPERATOR(sync_batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, + ops::BatchNormOpInferVarType, ops::BatchNormGradMaker); +REGISTER_OPERATOR(sync_batch_norm_grad, ops::BatchNormGradOp); diff --git a/paddle/fluid/operators/sync_batch_norm_op.cu b/paddle/fluid/operators/sync_batch_norm_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..a5984bfaaaf96f7a412176bb9868dc44488acf3f --- /dev/null +++ b/paddle/fluid/operators/sync_batch_norm_op.cu @@ -0,0 +1,452 @@ +/* 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 +#include +#include +#include +#include "cub/cub.cuh" +#include "paddle/fluid/framework/data_layout.h" +#include "paddle/fluid/operators/batch_norm_op.h" +#include "paddle/fluid/platform/cudnn_helper.h" +#include "paddle/fluid/platform/float16.h" +#include "paddle/fluid/platform/nccl_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using DataLayout = framework::DataLayout; +template +using CudnnDataType = platform::CudnnDataType; + +template +__global__ void KeLocalStats(const T *x, int N, int M, int C, T *mean_var) { + typedef cub::BlockReduce BlockReduce; + __shared__ typename BlockReduce::TempStorage temp_storage; + for (int k = blockIdx.x; k < C; k += gridDim.x) { + T x_sum = 0; + T x2_sum = 0; + for (int i = threadIdx.x; i < N * M; i += BlockDim) { + int id = layout == framework::DataLayout::kNCHW + ? (i / M) * C * M + k * M + i % M + : i * C + k; + T x_in = x[id]; + x_sum += x_in; + x2_sum += x_in * x_in; + } + __syncthreads(); + T out = BlockReduce(temp_storage).Reduce(x_sum, cub::Sum()); + __syncthreads(); + if (threadIdx.x == 0) { + mean_var[k] = out / (N * M); + } + out = BlockReduce(temp_storage).Reduce(x2_sum, cub::Sum()); + __syncthreads(); + if (threadIdx.x == 0) { + mean_var[k + C] = out / (N * M); + } + } + if (blockIdx.x == 0 && threadIdx.x == 0) { + mean_var[2 * C] = static_cast(1.0); + } +} + +template +__global__ void KeSyncAndMovingStats(T *means, T *variances, T *num_dev, + const int C, const T momentum, + const double epsilon, T *sv_mean_data, + T *sv_inv_var_data, T *moving_means, + T *moving_variances) { + // sync stats across multi-devices + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = gid; i < C; i += stride) { + T mean = means[i] / (*num_dev); + T var = variances[i] / (*num_dev); + var = var - mean * mean; + + // sync stats + sv_mean_data[i] = mean; + sv_inv_var_data[i] = 1.0 / sqrt(var + epsilon); + variances[i] = var; + + // moving stats + moving_means[i] = moving_means[i] * momentum + mean * (1. - momentum); + moving_variances[i] = + moving_variances[i] * momentum + var * (1. - momentum); + } +} + +template +static __global__ void KeNormAffine(const T *x, const T *scale, const T *bias, + const T *mean, const T *variance, + const double epsilon, const int C, + const int M, const int num, T *y) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = gid; i < num; i += stride) { + const int c = layout == framework::DataLayout::kNCHW ? (i / M) % C : i % C; + y[i] = (x[i] - mean[c]) / sqrt(variance[c] + epsilon) * scale[c] + bias[c]; + } +} + +template +class SyncBatchNormKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + double epsilon = static_cast(ctx.Attr("epsilon")); + const float momentum = ctx.Attr("momentum"); + const bool is_test = ctx.Attr("is_test"); + const std::string layout_str = ctx.Attr("data_layout"); + const DataLayout layout = framework::StringToDataLayout(layout_str); + const bool use_global_stats = ctx.Attr("use_global_stats"); + PADDLE_ENFORCE( + !use_global_stats, + "sync_batch_norm doesn't support to set use_global_stats True. ", + "Please use batch_norm in this case."); + + const auto *x = ctx.Input("X"); + const auto &x_dims = x->dims(); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); + int N, C, H, W, D; + ExtractNCWHD(x_dims, layout, &N, &C, &H, &W, &D); + int x_numel = x->numel(); + + const T *x_d = x->data(); + const T *s_d = ctx.Input("Scale")->data(); + const T *b_d = ctx.Input("Bias")->data(); + + auto *y = ctx.Output("Y"); + T *y_d = y->mutable_data(ctx.GetPlace()); + + const T *mean_data = nullptr; + const T *var_data = nullptr; + + auto &dev_ctx = ctx.cuda_device_context(); + auto stream = dev_ctx.stream(); + auto *comm = dev_ctx.nccl_comm(); + const int block = 512; + int max_threads = dev_ctx.GetMaxPhysicalThreadCount(); + + paddle::memory::AllocationPtr alloc_ptr{nullptr}; + + if (is_test) { + const auto *est_mean = ctx.Input("Mean"); + const auto *est_var = ctx.Input("Variance"); + mean_data = est_mean->data(); + var_data = est_var->data(); + } else { + auto &allocator = + platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx); + // x, x^2, 1, here 1 is used to calc device num + // device num also can be got from platform::DeviceContextPool + const int bytes = (C * 2 + 1) * sizeof(T); + alloc_ptr = allocator.Allocate(bytes); + + T *stats = reinterpret_cast(alloc_ptr->ptr()); + const int threads = 256; + int grid = std::min(C, (max_threads + threads - 1) / threads); + if (layout == framework::DataLayout::kNCHW) { + KeLocalStats< + T, threads, + framework::DataLayout::kNCHW><<>>( + x_d, N, H * W * D, C, stats); + } else { + KeLocalStats< + T, threads, + framework::DataLayout::kNHWC><<>>( + x_d, N, H * W * D, C, stats); + } + + Tensor c_g_st; + T *c_g_st_d = c_g_st.mutable_data({2 * C + 1}, platform::CPUPlace()); + auto gplace = boost::get(ctx.GetPlace()); + memory::Copy(platform::CPUPlace(), c_g_st_d, gplace, stats, bytes, 0); + + int dtype = platform::ToNCCLDataType(x->type()); + // In-place operation + PADDLE_ENFORCE(platform::dynload::ncclAllReduce( + stats, stats, 2 * C + 1, static_cast(dtype), ncclSum, + comm, stream)); + + // moving mean/variance + auto *mean_out = ctx.Output("MeanOut"); + auto *variance_out = ctx.Output("VarianceOut"); + T *est_mean_data = mean_out->mutable_data(ctx.GetPlace()); + T *est_var_data = variance_out->mutable_data(ctx.GetPlace()); + + auto *saved_mean = ctx.Output("SavedMean"); + auto *saved_inv_variance = ctx.Output("SavedVariance"); + T *sv_mean_data = saved_mean->mutable_data(ctx.GetPlace()); + T *sv_inv_var_data = saved_inv_variance->mutable_data(ctx.GetPlace()); + + // Note, Input('Mean')/Input('Variance') share variable with + // Output('MeanOut')/Output('VarianceOut') + KeSyncAndMovingStats<<<(C + block - 1) / block, block, 0, stream>>>( + stats, stats + C, stats + 2 * C, C, momentum, epsilon, sv_mean_data, + sv_inv_var_data, est_mean_data, est_var_data); + + mean_data = sv_mean_data; + var_data = stats + C; + } + + int grid2 = (std::min(x_numel, max_threads) + block - 1) / block; + if (layout == framework::DataLayout::kNCHW) { + KeNormAffine<<>>( + x_d, s_d, b_d, mean_data, var_data, epsilon, C, H * W * D, x_numel, + y_d); + } else { + KeNormAffine<<>>( + x_d, s_d, b_d, mean_data, var_data, epsilon, C, H * W * D, x_numel, + y_d); + } + } +}; + +template +__global__ void KeBackwardLocalStats(const T *dy, const T *x, const T *means, + int N, int M, int C, T *sum_dy_prod) { + typedef cub::BlockReduce BlockReduce; + __shared__ typename BlockReduce::TempStorage temp_storage; + for (int k = blockIdx.x; k < C; k += gridDim.x) { + T sum1 = 0; + T sum2 = 0; + T mean = means[k]; + for (int i = threadIdx.x; i < N * M; i += blockDim.x) { + int id = layout == framework::DataLayout::kNCHW + ? (i / M) * C * M + k * M + i % M + : i * C + k; + T g = dy[id]; + sum1 += g; + sum2 += g * (x[id] - mean); + } + + __syncthreads(); + T out = BlockReduce(temp_storage).Reduce(sum1, cub::Sum()); + __syncthreads(); + if (threadIdx.x == 0) { + sum_dy_prod[k] = out; + } + out = BlockReduce(temp_storage).Reduce(sum2, cub::Sum()); + __syncthreads(); + if (threadIdx.x == 0) { + sum_dy_prod[k + C] = out; + } + } + if (blockIdx.x == 0 && threadIdx.x == 0) { + sum_dy_prod[2 * C] = static_cast(1.0); + } +} + +template +static __global__ void KeBNBackwardScaleBias(const T *dy, const T *x, + const T *mean, + const T *inv_variance, + const double epsilon, const int N, + const int C, const int HxW, + T *dscale, T *dbias) { + const int outer_size = C; + const int inner_size = N * HxW; + typedef cub::BlockReduce BlockReduce; + __shared__ typename BlockReduce::TempStorage temp_storage; + + for (int i = blockIdx.x; i < outer_size; i += gridDim.x) { + T ds_sum = static_cast(0); + T db_sum = static_cast(0); + + T inv_var_i = inv_variance[i]; + T mean_i = mean[i]; + for (int j = threadIdx.x; j < inner_size; j += blockDim.x) { + const int id = layout == framework::DataLayout::kNCHW + ? ((j / HxW) * C + i) * HxW + (j % HxW) + : j * outer_size + i; + ds_sum += dy[id] * (x[id] - mean_i); + db_sum += dy[id]; + } + __syncthreads(); + double os = BlockReduce(temp_storage) + .Reduce(static_cast(ds_sum), cub::Sum()); + __syncthreads(); + double ob = BlockReduce(temp_storage) + .Reduce(static_cast(db_sum), cub::Sum()); + __syncthreads(); + if (threadIdx.x == 0) { + dscale[i] = static_cast(os * inv_var_i); + dbias[i] = static_cast(ob); + } + __syncthreads(); + } +} + +template +static __global__ void KeBNBackwardData(const T *dy, const T *x, const T *beta, + const T *mean, const T *inv_variance, + const T *g_sum_dy, + const T *g_sum_dy_prod, + const T *num_dev, const double epsilon, + const int C, const int HxW, + const int num, T *dx) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + T scale = static_cast(C) / num; + T dev_num = num_dev[0]; + for (int i = gid; i < num; i += stride) { + const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C; + T inv_var = inv_variance[c]; + T s_d = beta[c]; + T gvar = -1.0 * (g_sum_dy_prod[c] / dev_num) * s_d * inv_var * + (inv_var * inv_var); + T gmean = -1.0 * (g_sum_dy[c] / dev_num) * s_d * inv_var; + + dx[i] = + dy[i] * s_d * inv_var + gmean * scale + gvar * scale * (x[i] - mean[c]); + } +} + +// Deriving the Gradient for the Backward Pass of Batch Normalization +// https://kevinzakka.github.io/2016/09/14/batch_normalization/ +template +class SyncBatchNormGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + double epsilon = static_cast(ctx.Attr("epsilon")); + const std::string layout_str = ctx.Attr("data_layout"); + + const DataLayout layout = framework::StringToDataLayout(layout_str); + const auto *x = ctx.Input("X"); + const auto *d_y = ctx.Input(framework::GradVarName("Y")); + const auto *scale = ctx.Input("Scale"); + + const auto &x_dims = x->dims(); + + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); + int N, C, H, W, D; + ExtractNCWHD(x_dims, layout, &N, &C, &H, &W, &D); + + // init output + auto *d_x = ctx.Output(framework::GradVarName("X")); + auto *d_scale = ctx.Output(framework::GradVarName("Scale")); + auto *d_bias = ctx.Output(framework::GradVarName("Bias")); + + d_x->mutable_data(ctx.GetPlace()); + if (d_scale && d_bias) { + d_scale->mutable_data(ctx.GetPlace()); + d_bias->mutable_data(ctx.GetPlace()); + } + PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL); + PADDLE_ENFORCE_EQ(scale->dims()[0], C); + + std::vector dims; + std::vector strides; + if (layout == DataLayout::kNCHW) { + dims = {N, C, H, W, D}; + strides = {C * H * W * D, H * W * D, W * D, D, 1}; + } else { + dims = {N, C, H, W, D}; + strides = {H * W * C * D, 1, W * D * C, D * C, C}; + } + + const T *x_d = x->data(); + const T *dy_d = d_y->data(); + + auto &dev_ctx = ctx.cuda_device_context(); + auto stream = dev_ctx.stream(); + auto *comm = dev_ctx.nccl_comm(); + + const T *saved_mean = ctx.Input("SavedMean")->data(); + const T *saved_inv_var = ctx.Input("SavedVariance")->data(); + auto &allocator = + platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx); + const int bytes = (C * 2 + 1) * sizeof(T); + auto alloc_ptr = allocator.Allocate(bytes); + T *stats = reinterpret_cast(alloc_ptr->ptr()); + + const int threads = 256; + int max_threads = dev_ctx.GetMaxPhysicalThreadCount(); + int grid = std::min(C, (max_threads + threads - 1) / threads); + int x_numel = x->numel(); + int fsize = H * W * D; + + if (layout == framework::DataLayout::kNCHW) { + KeBackwardLocalStats< + T, threads, + framework::DataLayout::kNCHW><<>>( + dy_d, x_d, saved_mean, N, fsize, C, stats); + } else { + KeBackwardLocalStats< + T, threads, + framework::DataLayout::kNHWC><<>>( + dy_d, x_d, saved_mean, N, fsize, C, stats); + } + int dtype = platform::ToNCCLDataType(x->type()); + // In-place operation + PADDLE_ENFORCE(platform::dynload::ncclAllReduce( + stats, stats, 2 * C + 1, static_cast(dtype), ncclSum, + comm, stream)); + + const int block = 512; + int grid2 = (std::min(x_numel, max_threads) + block - 1) / block; + if (layout == framework::DataLayout::kNCHW) { + if (d_scale && d_bias) { + KeBNBackwardScaleBias< + T, threads, + framework::DataLayout::kNCHW><<>>( + dy_d, x_d, saved_mean, saved_inv_var, epsilon, N, C, fsize, + d_scale->data(), d_bias->data()); + } + if (d_x) { + KeBNBackwardData< + T, framework::DataLayout::kNCHW><<>>( + dy_d, x_d, scale->data(), saved_mean, saved_inv_var, stats, + stats + C, stats + 2 * C, epsilon, C, fsize, x->numel(), + d_x->data()); + } + } else { + if (d_scale && d_bias) { + KeBNBackwardScaleBias< + T, threads, + framework::DataLayout::kNHWC><<>>( + dy_d, x_d, saved_mean, saved_inv_var, epsilon, N, C, fsize, + d_scale->data(), d_bias->data()); + } + if (d_x) { + KeBNBackwardData< + T, framework::DataLayout::kNHWC><<>>( + dy_d, x_d, scale->data(), saved_mean, saved_inv_var, stats, + stats + C, stats + 2 * C, epsilon, C, fsize, x->numel(), + d_x->data()); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace plat = paddle::platform; +REGISTER_OP_CUDA_KERNEL( + sync_batch_norm, ops::SyncBatchNormKernel, + ops::SyncBatchNormKernel); +REGISTER_OP_CUDA_KERNEL( + sync_batch_norm_grad, + ops::SyncBatchNormGradKernel, + ops::SyncBatchNormGradKernel); diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index 920b43b2b1990af58b73888bf7a652d57c20563c..ada9a197361c75fb0128b49b7465154d58f94304 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -57,7 +57,6 @@ DeviceContextPool::DeviceContextPool( for (auto& p : places) { set.insert(p); } - for (auto& p : set) { if (platform::is_cpu_place(p)) { #ifdef PADDLE_WITH_MKLDNN @@ -317,6 +316,7 @@ CUDADeviceContext::~CUDADeviceContext() { eigen_stream_.reset(); eigen_device_.reset(); PADDLE_ENFORCE(cudaStreamDestroy(stream_)); + PADDLE_ENFORCE(dynload::ncclCommDestroy(nccl_comm_)); } Place CUDADeviceContext::GetPlace() const { return place_; } diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index d376f90ad5754d70f3b9f30957eb2e2f584f8da9..3f7ce3d9446925ef5ec0da6651aaa789aaaa1420 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -265,6 +265,12 @@ class CUDADeviceContext : public DeviceContext { /*! \brief Return cuda stream in the device context. */ cudaStream_t stream() const; + /*! \brief Return nccl communicators. */ + ncclComm_t nccl_comm() const { return nccl_comm_; } + + /*! \brief Set nccl communicators. */ + void set_nccl_comm(ncclComm_t comm) { nccl_comm_ = comm; } + template void RecordEvent(cudaEvent_t ev, Callback callback) { callback(); @@ -289,6 +295,13 @@ class CUDADeviceContext : public DeviceContext { std::unique_ptr cublas_handle_; std::unique_ptr cublas_tensor_core_handle_; + // NCCL communicator (single process version) for NCCL collective operations. + // NCCL collective operations provides fast collectives over multiple GPUs + // both within and across nodes. + // But, this collectives is used for collectives over multiple GPUs within + // nodes. + ncclComm_t nccl_comm_{nullptr}; + int compute_capability_; int runtime_version_; int driver_version_; diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index 0179daa55715be9787bc7cc8a693319024d404b7..8458b17f82a976bad37df58dddc2c0d80c8eb13e 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -11,7 +11,6 @@ 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/platform/device_tracer.h" #include #include @@ -30,6 +29,7 @@ limitations under the License. */ #include "glog/logging.h" #include "google/protobuf/text_format.h" #include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/platform/device_tracer.h" #include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/string/printf.h" @@ -222,19 +222,24 @@ void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, } case CUPTI_ACTIVITY_KIND_DRIVER: { auto *api = reinterpret_cast(record); - if (api->start != 0 && api->end != 0) - // -1 device id represents CUDA api call - tracer->AddCPURecords( + if (api->start != 0 && api->end != 0) { + // -1 device id represents ActiveKind api call + tracer->AddActiveKindRecords( DriverKind(api->cbid), api->start, api->end, -1, - GetThreadIdFromSystemThreadId(api->threadId)); + GetThreadIdFromSystemThreadId(api->threadId), + api->correlationId); + } break; } case CUPTI_ACTIVITY_KIND_RUNTIME: { auto *api = reinterpret_cast(record); - if (api->start != 0 && api->end != 0) - tracer->AddCPURecords( + if (api->start != 0 && api->end != 0) { + // -1 device id represents ActiveKind api call + tracer->AddActiveKindRecords( RuntimeKind(api->cbid), api->start, api->end, -1, - GetThreadIdFromSystemThreadId(api->threadId)); + GetThreadIdFromSystemThreadId(api->threadId), + api->correlationId); + } break; } default: { break; } @@ -313,6 +318,43 @@ class DeviceTracerImpl : public DeviceTracer { stream_id, correlation_id, bytes}); } + void AddMemInfoRecord(uint64_t start_ns, uint64_t end_ns, size_t bytes, + const Place &place, const std::string &alloc_in, + const std::string &free_in, int64_t thread_id) { + if (0 == start_ns || 0 == end_ns) { + VLOG(3) << alloc_in << ", " << free_in << " Cannot be traced."; + return; + } + thread_local std::forward_list *local_mem_info_record = + nullptr; + if (local_mem_info_record == nullptr) { + std::lock_guard l(trace_mu_); + mem_info_record_.emplace_front(); + local_mem_info_record = &mem_info_record_.front(); + } + local_mem_info_record->emplace_front(MemInfoRecord{ + start_ns, end_ns, bytes, place, thread_id, alloc_in, free_in}); + } + + void AddActiveKindRecords(const std::string &anno, uint64_t start_ns, + uint64_t end_ns, int64_t device_id, + int64_t thread_id, uint32_t correlation_id) { + if (anno.empty()) { + VLOG(1) << "Empty timeline annotation."; + return; + } + thread_local std::forward_list + *local_active_kind_records = nullptr; + if (local_active_kind_records == nullptr) { + std::lock_guard l(trace_mu_); + active_kind_records_.emplace_front(); + local_active_kind_records = &active_kind_records_.front(); + } + // lock is not needed, only one thread call this function. + local_active_kind_records->push_front(ActiveKindRecord{ + anno, start_ns, end_ns, device_id, thread_id, correlation_id}); + } + void AddKernelRecords(std::string name, uint64_t start, uint64_t end, int64_t device_id, int64_t stream_id, uint32_t correlation_id) { @@ -355,6 +397,7 @@ class DeviceTracerImpl : public DeviceTracer { } const std::vector cbids { CUPTI_RUNTIME_TRACE_CBID_cudaMemcpy_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaSetupArgument_v3020, CUPTI_RUNTIME_TRACE_CBID_cudaMemcpyAsync_v3020, CUPTI_RUNTIME_TRACE_CBID_cudaMemset_v3020, CUPTI_RUNTIME_TRACE_CBID_cudaMemsetAsync_v3020, @@ -385,6 +428,8 @@ class DeviceTracerImpl : public DeviceTracer { correlations_.clear(); for (auto &tmp : correlations_pairs) tmp.clear(); for (auto &tmp : cpu_records_) tmp.clear(); + for (auto &tmp : mem_info_record_) tmp.clear(); + for (auto &tmp : active_kind_records_) tmp.clear(); } void GenEventKernelCudaElapsedTime() { @@ -415,9 +460,12 @@ class DeviceTracerImpl : public DeviceTracer { proto::Profile profile_pb; profile_pb.set_start_ns(start_ns_); profile_pb.set_end_ns(end_ns_); - if (correlations_.empty()) - for (auto &tmp : correlations_pairs) + if (correlations_.empty()) { + for (auto &tmp : correlations_pairs) { for (auto &pair : tmp) correlations_[pair.first] = pair.second; + } + } + for (const KernelRecord &r : kernel_records_) { auto *event = profile_pb.add_events(); event->set_type(proto::Event::GPUKernel); @@ -437,7 +485,8 @@ class DeviceTracerImpl : public DeviceTracer { event->set_device_id(r.device_id); } VLOG(1) << "KernelRecord event miss: " << miss << " find: " << find; - for (auto &tmp : cpu_records_) + + for (auto &tmp : cpu_records_) { for (const CPURecord &r : tmp) { auto *event = profile_pb.add_events(); event->set_type(proto::Event::CPU); @@ -447,6 +496,25 @@ class DeviceTracerImpl : public DeviceTracer { event->set_sub_device_id(r.thread_id); event->set_device_id(r.device_id); } + } + + for (auto &tmp : active_kind_records_) { + for (const ActiveKindRecord &r : tmp) { + auto *event = profile_pb.add_events(); + event->set_type(proto::Event::CPU); + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + event->set_name(c->second->name()); + event->set_detail_info(r.name); + } else { + event->set_name(r.name); + } + event->set_start_ns(r.start_ns); + event->set_end_ns(r.end_ns); + event->set_sub_device_id(r.thread_id); + event->set_device_id(r.device_id); + } + } miss = find = 0; for (const MemRecord &r : mem_records_) { auto *event = profile_pb.add_events(); @@ -467,6 +535,31 @@ class DeviceTracerImpl : public DeviceTracer { event->mutable_memcopy()->set_bytes(r.bytes); } VLOG(1) << "MemRecord event miss: " << miss << " find: " << find; + + for (auto &tmp : mem_info_record_) { + for (const auto &r : tmp) { + auto *event = profile_pb.add_mem_events(); + event->set_device_id(0); + if (platform::is_cpu_place(r.place)) { + event->set_place(proto::MemEvent::CPUPlace); + } else if (platform::is_gpu_place(r.place)) { + event->set_place(proto::MemEvent::CUDAPlace); + event->set_device_id( + boost::get(r.place).GetDeviceId()); + } else if (platform::is_cuda_pinned_place(r.place)) { + event->set_place(proto::MemEvent::CUDAPinnedPlace); + } else { + PADDLE_THROW("The current place is not supported."); + } + event->set_alloc_in(r.alloc_in); + event->set_free_in(r.free_in); + event->set_start_ns(r.start_ns); + event->set_end_ns(r.end_ns); + event->set_bytes(r.bytes); + event->set_thread_id(r.thread_id); + } + } + std::ofstream profile_f; profile_f.open(profile_path, std::ios::out | std::ios::trunc | std::ios::binary); @@ -510,6 +603,8 @@ class DeviceTracerImpl : public DeviceTracer { std::forward_list kernel_records_; std::forward_list mem_records_; std::forward_list> cpu_records_; + std::forward_list> mem_info_record_; + std::forward_list> active_kind_records_; std::forward_list>> correlations_pairs; std::unordered_map correlations_; @@ -531,7 +626,7 @@ Event *CurAnnotation() { return annotation_stack.back(); } std::string CurAnnotationName() { - if (annotation_stack.empty()) return ""; + if (annotation_stack.empty()) return "Unknown"; return annotation_stack.back()->name(); } @@ -613,6 +708,7 @@ void initCuptiCbidStr() { REGISTER_RUNTIME_CBID_STR(cudaUnbindTexture_v3020); REGISTER_RUNTIME_CBID_STR(cudaSetupArgument_v3020); REGISTER_RUNTIME_CBID_STR(cudaLaunch_v3020); + REGISTER_RUNTIME_CBID_STR(cudaDeviceGetPCIBusId_v4010); #if CUDA_VERSION >= 9000 REGISTER_RUNTIME_CBID_STR(cudaLaunchCooperativeKernel_v9000); REGISTER_RUNTIME_CBID_STR(cudaLaunchCooperativeKernelMultiDevice_v9000); diff --git a/paddle/fluid/platform/device_tracer.h b/paddle/fluid/platform/device_tracer.h index d4418d836d66e329af8ed3f5ec05f49d47146b3e..85168a046fb3fa4317956737871cde56e15bedfb 100644 --- a/paddle/fluid/platform/device_tracer.h +++ b/paddle/fluid/platform/device_tracer.h @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/platform/dynload/cupti.h" #include "paddle/fluid/platform/event.h" +#include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/profiler.pb.h" @@ -47,6 +48,7 @@ class DeviceTracer { int64_t stream_id; uint32_t correlation_id; }; + struct CPURecord { std::string name; uint64_t start_ns; @@ -54,6 +56,7 @@ class DeviceTracer { int64_t device_id; int64_t thread_id; }; + struct MemRecord { std::string name; uint64_t start_ns; @@ -64,6 +67,25 @@ class DeviceTracer { uint64_t bytes; }; + struct MemInfoRecord { + uint64_t start_ns; + uint64_t end_ns; + size_t bytes; + Place place; + int64_t thread_id; + std::string alloc_in; + std::string free_in; + }; + + struct ActiveKindRecord { + std::string name; + uint64_t start_ns; + uint64_t end_ns; + int64_t device_id; + int64_t thread_id; + uint32_t correlation_id; + }; + virtual ~DeviceTracer() {} // Needs to be called once before use. virtual void Enable() = 0; @@ -85,6 +107,16 @@ class DeviceTracer { virtual void AddCPURecords(const std::string& anno, uint64_t start_ns, uint64_t end_ns, int64_t device_id, int64_t thread_id) = 0; + virtual void AddActiveKindRecords(const std::string& anno, uint64_t start_ns, + uint64_t end_ns, int64_t device_id, + int64_t thread_id, + uint32_t correlation_id) = 0; + + virtual void AddMemInfoRecord(uint64_t start_ns, uint64_t end_ns, + size_t bytes, const Place& place, + const std::string& alloc_in, + const std::string& free_in, + int64_t thread_id) = 0; // Add a cuda kernel stats. `correlation_id` will be mapped to annotation // added before for human readability. diff --git a/paddle/fluid/platform/event.h b/paddle/fluid/platform/event.h index 2dcf966754cbed2670acb9c3548c23355be5503c..e9bdb82a50fa4166cecdaea1de01d2f458f3da9a 100644 --- a/paddle/fluid/platform/event.h +++ b/paddle/fluid/platform/event.h @@ -13,10 +13,12 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once + #include #ifdef PADDLE_WITH_CUDA #include #endif +#include "paddle/fluid/platform/place.h" namespace paddle { namespace platform { @@ -64,5 +66,36 @@ class Event { #endif #endif }; + +class MemEvent { + public: + MemEvent(EventType type, uint64_t start_ns, uint64_t end_ns, size_t bytes, + Place place, int64_t thread_id, const std::string& annotation) + : type_(type), + start_ns_(start_ns), + end_ns_(end_ns), + bytes_(bytes), + place_(place), + thread_id_(thread_id), + annotation_(annotation) {} + + const EventType& type() const { return type_; } + uint64_t start_ns() const { return start_ns_; } + uint64_t end_ns() const { return end_ns_; } + size_t bytes() const { return bytes_; } + Place place() const { return place_; } + int64_t thread_id() const { return thread_id_; } + const std::string& annotation() const { return annotation_; } + + private: + EventType type_; + uint64_t start_ns_ = 0; + uint64_t end_ns_ = 0; + size_t bytes_; + Place place_; + int64_t thread_id_; + std::string annotation_; +}; + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc index 4dcf7e79043af008cb2067d90d12d629c5c2d0d9..d53a4029e1bad9eded693d2d9bd8e01e13bb73e7 100644 --- a/paddle/fluid/platform/init.cc +++ b/paddle/fluid/platform/init.cc @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #include // for strdup #include +#include +#include #include #include @@ -140,6 +142,7 @@ void InitDevices(bool init_p2p, const std::vector devices) { places.emplace_back(platform::CPUPlace()); platform::DeviceContextPool::Init(places); platform::DeviceTemporaryAllocator::Init(); + #ifndef PADDLE_WITH_MKLDNN platform::SetNumThreads(FLAGS_paddle_num_threads); #endif diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 6ae21ee8294bedc388f837aad3e20a2b9aca98a2..0428c40f985d78f0262eb0a73984bc59ab43aac2 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -16,9 +16,11 @@ #pragma once #include +#include #include #include // NOLINT #include +#include #include #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/platform/dynload/nccl.h" @@ -78,6 +80,8 @@ struct NCCLContext { cudaStream_t stream() const { return ctx_->stream(); } + ncclComm_t comm() const { return comm_; } + int device_id() const { return boost::get(ctx_->GetPlace()).device; } diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 9a285a6b533dcb48013e3b3e4d34dc27186173ac..6d055a442106d88af39c771f3ddf156ba616c99f 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -13,7 +13,6 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/platform/profiler.h" - #include #include #include @@ -21,6 +20,8 @@ limitations under the License. */ #include // NOLINT #include #include +#include + #ifdef PADDLE_WITH_CUDA #include #endif // PADDLE_WITH_CUDA @@ -36,8 +37,6 @@ DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not."); namespace paddle { namespace platform { -struct EventList; - static int64_t profiler_lister_id = 0; static bool should_send_profile_state = false; std::mutex profiler_mu; @@ -53,43 +52,15 @@ static uint32_t g_next_thread_id = 0; // The global mutex static std::mutex g_all_event_lists_mutex; // The total event lists of all threads -static std::list> g_all_event_lists; +static std::list>> g_all_event_lists; // The thread local event list only can be accessed by the specific thread -static thread_local std::shared_ptr g_event_list; - -struct EventList { - constexpr static size_t kMB = 1024 * 1024; - constexpr static size_t kEventBlockSize = 16 * kMB; - constexpr static size_t kEventSize = sizeof(Event); - constexpr static size_t kEventAlign = alignof(Event); - constexpr static size_t kNumBlock = - kEventBlockSize / - ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign); - - template - Event* Record(Args&&... args) { - if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) { - event_blocks.emplace_front(); - event_blocks.front().reserve(kNumBlock); - } - event_blocks.front().emplace_back(std::forward(args)...); - return &event_blocks.front().back(); - } - - std::vector Reduce() { - std::vector result; - for (auto& block : event_blocks) { - result.insert(result.begin(), std::make_move_iterator(block.begin()), - std::make_move_iterator(block.end())); - } - event_blocks.clear(); - return result; - } +static thread_local std::shared_ptr> g_event_list; - void Clear() { event_blocks.clear(); } - - std::forward_list> event_blocks; -}; +static std::list>> g_all_mem_event_lists; +static thread_local std::shared_ptr> g_mem_event_list; +static std::mutex g_all_mem_event_lists_mutex; +static thread_local int32_t g_mem_thread_id; +static uint32_t g_mem_next_thread_id = 0; inline uint64_t GetTimeInNsec() { using clock = std::conditional &GetMemEventList() { + if (!g_mem_event_list) { + g_mem_event_list = std::make_shared>(); + std::lock_guard guard(g_all_mem_event_lists_mutex); + g_mem_thread_id = g_mem_next_thread_id++; + g_all_mem_event_lists.emplace_front(g_mem_event_list); + } + return *g_mem_event_list; +} + +void PushMemEvent(uint64_t start_ns, uint64_t end_ns, size_t bytes, + const Place &place, const std::string &annotation) { + GetMemEventList().Record(EventType::kPushRange, start_ns, end_ns, bytes, + place, g_mem_thread_id, annotation); +} + +void PopMemEvent(uint64_t start_ns, uint64_t end_ns, size_t bytes, + const Place &place, const std::string &annotation) { + GetMemEventList().Record(EventType::kPopRange, start_ns, end_ns, bytes, place, + g_mem_thread_id, annotation); +} + +inline EventList &GetEventList() { if (!g_event_list) { std::lock_guard guard(g_all_event_lists_mutex); - g_event_list = std::make_shared(); + g_event_list = std::make_shared>(); g_thread_id = g_next_thread_id++; g_all_event_lists.emplace_front(g_event_list); RecoreCurThreadId(g_thread_id); @@ -131,26 +124,26 @@ inline EventList& GetEventList() { return *g_event_list; } -void Mark(const std::string& name) { +void Mark(const std::string &name) { GetEventList().Record(EventType::kMark, name, g_thread_id); } -Event* PushEvent(const std::string& name) { +Event *PushEvent(const std::string &name) { return GetEventList().Record(EventType::kPushRange, name, g_thread_id); } -void PopEvent(const std::string& name) { +void PopEvent(const std::string &name) { GetEventList().Record(EventType::kPopRange, name, g_thread_id); } -RecordEvent::RecordEvent(const std::string& name) +RecordEvent::RecordEvent(const std::string &name) : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; // lock is not needed, the code below is thread-safe is_enabled_ = true; name_ = name; - Event* e = PushEvent(name_); + Event *e = PushEvent(name_); // Maybe need the same push/pop behavior. SetCurAnnotation(e); } @@ -158,7 +151,7 @@ RecordEvent::RecordEvent(const std::string& name) RecordEvent::~RecordEvent() { if (g_state == ProfilerState::kDisabled || !is_enabled_) return; // lock is not needed, the code below is thread-safe - DeviceTracer* tracer = GetDeviceTracer(); + DeviceTracer *tracer = GetDeviceTracer(); if (tracer) { tracer->AddCPURecords(CurAnnotationName(), start_ns_, PosixInNsec(), BlockDepth(), g_thread_id); @@ -167,7 +160,56 @@ RecordEvent::~RecordEvent() { PopEvent(name_); } -RecordRPCEvent::RecordRPCEvent(const std::string& name) { +MemEvenRecorder MemEvenRecorder::recorder; + +void MemEvenRecorder::PushMemRecord(const void *ptr, const Place &place, + size_t size) { + if (g_state == ProfilerState::kDisabled) return; + std::lock_guard guard(mtx_); + auto &events = address_memevent_[place]; + PADDLE_ENFORCE(events.count(ptr) == 0, ""); + events.emplace(ptr, std::unique_ptr( + new MemEvenRecorder::RecordMemEvent(place, size))); +} + +void MemEvenRecorder::PopMemRecord(const void *ptr, const Place &place) { + if (g_state == ProfilerState::kDisabled) return; + std::lock_guard guard(mtx_); + auto &events = address_memevent_[place]; + auto iter = events.find(ptr); + // The ptr maybe not in address_memevent + if (iter != events.end()) { + events.erase(iter); + } +} + +void MemEvenRecorder::Flush() { + std::lock_guard guard(mtx_); + address_memevent_.clear(); +} + +MemEvenRecorder::RecordMemEvent::RecordMemEvent(const Place &place, + size_t bytes) + : place_(place), + bytes_(bytes), + start_ns_(PosixInNsec()), + alloc_in_(CurAnnotationName()) { + PushMemEvent(start_ns_, end_ns_, bytes_, place_, alloc_in_); +} + +MemEvenRecorder::RecordMemEvent::~RecordMemEvent() { + DeviceTracer *tracer = GetDeviceTracer(); + end_ns_ = PosixInNsec(); + + auto annotation_free = CurAnnotationName(); + if (tracer) { + tracer->AddMemInfoRecord(start_ns_, end_ns_, bytes_, place_, alloc_in_, + annotation_free, g_mem_thread_id); + } + PopMemEvent(start_ns_, end_ns_, bytes_, place_, annotation_free); +} + +RecordRPCEvent::RecordRPCEvent(const std::string &name) { if (FLAGS_enable_rpc_profiler) { event_.reset(new platform::RecordEvent(name)); } @@ -185,7 +227,7 @@ RecordBlock::RecordBlock(int block_id) RecordBlock::~RecordBlock() { // lock is not needed, the code below is thread-safe if (g_state == ProfilerState::kDisabled || !is_enabled_) return; - DeviceTracer* tracer = GetDeviceTracer(); + DeviceTracer *tracer = GetDeviceTracer(); if (tracer) { // We try to put all blocks at the same nested depth in the // same timeline lane. and distinguish the using thread_id. @@ -232,11 +274,16 @@ void EnableProfiler(ProfilerState state) { void ResetProfiler() { SynchronizeAllDevice(); GetDeviceTracer()->Reset(); + MemEvenRecorder::Instance().Flush(); std::lock_guard guard(g_all_event_lists_mutex); for (auto it = g_all_event_lists.begin(); it != g_all_event_lists.end(); ++it) { (*it)->Clear(); } + for (auto it = g_all_mem_event_lists.begin(); + it != g_all_mem_event_lists.end(); ++it) { + (*it)->Clear(); + } } std::vector> GetAllEvents() { @@ -249,6 +296,15 @@ std::vector> GetAllEvents() { return result; } +std::vector> GetMemEvents() { + std::lock_guard guard(g_all_mem_event_lists_mutex); + std::vector> result; + for (auto &it : g_all_mem_event_lists) { + result.emplace_back((*it).Reduce()); + } + return result; +} + // The information of each event given in the profiling report struct EventItem { std::string name; @@ -263,8 +319,8 @@ struct EventItem { }; // Print results -void PrintProfiler(const std::vector>& events_table, - const std::string& sorted_domain, const size_t name_width, +void PrintProfiler(const std::vector> &events_table, + const std::string &sorted_domain, const size_t name_width, const size_t data_width, bool merge_thread) { // Output header information std::cout << "\n------------------------->" @@ -302,7 +358,7 @@ void PrintProfiler(const std::vector>& events_table, << std::setw(data_width) << "Ratio." << std::endl; for (size_t i = 0; i < events_table.size(); ++i) { for (size_t j = 0; j < events_table[i].size(); ++j) { - const EventItem& event_item = events_table[i][j]; + const EventItem &event_item = events_table[i][j]; std::cout << std::setw(name_width) << event_item.name << std::setw(data_width) << event_item.calls << std::setw(data_width) << event_item.total_time; @@ -326,54 +382,54 @@ void PrintProfiler(const std::vector>& events_table, } // Parse the event list and output the profiling report -void ParseEvents(const std::vector>& events, +void ParseEvents(const std::vector> &events, bool merge_thread, EventSortingKey sorted_by = EventSortingKey::kDefault) { if (g_state == ProfilerState::kDisabled) return; if (merge_thread && events.size() < 2) return; std::string sorted_domain; - std::function sorted_func; + std::function sorted_func; switch (sorted_by) { case EventSortingKey::kCalls: sorted_domain = "number of calls"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.calls > b.calls; }; break; case EventSortingKey::kTotal: sorted_domain = "total time"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.total_time > b.total_time; }; break; case EventSortingKey::kMin: sorted_domain = "minimum time"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.min_time > b.min_time; }; break; case EventSortingKey::kMax: sorted_domain = "maximum time"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.max_time > b.max_time; }; break; case EventSortingKey::kAve: sorted_domain = "average time"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.ave_time > b.ave_time; }; break; case EventSortingKey::kGPUTime: sorted_domain = "average time"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.gpu_time > b.gpu_time; }; break; case EventSortingKey::kCPUTime: sorted_domain = "average time"; - sorted_func = [](const EventItem& a, const EventItem& b) { + sorted_func = [](const EventItem &a, const EventItem &b) { return a.cpu_time > b.cpu_time; }; break; @@ -381,7 +437,7 @@ void ParseEvents(const std::vector>& events, sorted_domain = "event first end time"; } - const std::vector>* analyze_events; + const std::vector> *analyze_events; std::vector> merged_events_list; if (merge_thread) { std::vector merged_events; @@ -469,7 +525,7 @@ void ParseEvents(const std::vector>& events, } } // average time - for (auto& item : event_items) { + for (auto &item : event_items) { item.ave_time = item.total_time / item.calls; item.ratio = item.total_time / total; } @@ -493,15 +549,77 @@ void ParseEvents(const std::vector>& events, merge_thread); } +struct MemoryProfierReport { + size_t alloc_times{0}; + size_t alloc_size{0}; + size_t free_times{0}; + size_t free_size{0}; +}; + +// Print results +void PrintMemProfiler( + const std::map> + &annotation_report, + const size_t name_width, const size_t data_width) { + // Output header information + std::cout << "\n------------------------->" + << " Memory Profiling Report " + << "<-------------------------\n\n"; + + // Output events table + std::cout.setf(std::ios::left); + std::cout << std::setw(name_width) << "Event" << std::setw(data_width) + << "Alloc Calls" << std::setw(data_width) << "Size(MB)" + << std::setw(data_width) << "Free Calls" << std::setw(data_width) + << "Size(MB)" << std::endl; + + for (auto &tmp : annotation_report) { + for (auto &e : tmp.second) { + auto event_name = string::Sprintf("%s:%s", tmp.first, e.first); + std::cout << std::setw(name_width) << event_name; + std::cout << std::setw(data_width) << e.second.alloc_times; + std::cout << std::setw(data_width) + << e.second.alloc_size / (1024.0 * 1024.0); + std::cout << std::setw(data_width) << e.second.free_times; + std::cout << std::setw(data_width) + << e.second.free_size / (1024.0 * 1024.0) << std::endl; + } + } + std::cout << std::endl; +} + +// parse memory events +void ParseMemEvents(const std::vector> &events) { + if (g_state == ProfilerState::kDisabled) return; + // place, annotation, alloc times, alloc size + std::map> + annotation_report; + + for (auto &tmp : events) { + for (auto &e : tmp) { + if (e.type() == EventType::kPushRange) { + annotation_report[e.place()][e.annotation()].alloc_times += 1; + annotation_report[e.place()][e.annotation()].alloc_size += e.bytes(); + } else if (e.type() == EventType::kPopRange) { + annotation_report[e.place()][e.annotation()].free_times += 1; + annotation_report[e.place()][e.annotation()].free_size += e.bytes(); + } + } + } + PrintMemProfiler(annotation_report, 55, 18); +} + void DisableProfiler(EventSortingKey sorted_key, - const std::string& profile_path) { + const std::string &profile_path) { SynchronizeAllDevice(); + MemEvenRecorder::Instance().Flush(); + std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled) return; // Mark the profiling stop. Mark("_stop_profiler_"); - DeviceTracer* tracer = GetDeviceTracer(); + DeviceTracer *tracer = GetDeviceTracer(); if (tracer->IsEnabled()) { tracer->Disable(); tracer->GenProfile(profile_path); @@ -511,6 +629,11 @@ void DisableProfiler(EventSortingKey sorted_key, std::vector> all_events = GetAllEvents(); ParseEvents(all_events, true, sorted_key); ParseEvents(all_events, false, sorted_key); + if (VLOG_IS_ON(5)) { + std::vector> all_mem_events = GetMemEvents(); + ParseMemEvents(all_mem_events); + } + ResetProfiler(); g_state = ProfilerState::kDisabled; should_send_profile_state = true; diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index aec0ae34292d62905de0e1f459b2b6db4554ebb7..8d11855b70de824159f19f2997b876564e7719b1 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -15,10 +15,17 @@ limitations under the License. */ #pragma once #include #include +#include +#include +#include // NOLINT #include +#include +#include +#include #include #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/event.h" +#include "paddle/fluid/platform/place.h" #ifdef PADDLE_WITH_CUDA #include "paddle/fluid/platform/gpu_info.h" #endif @@ -34,8 +41,41 @@ enum ProfilerState { void Mark(const std::string& name); -Event* PushEvent(const std::string& name); +void PushMemEvent(uint64_t start_ns, uint64_t end_ns, size_t bytes, + const Place& place); +void PopMemEvent(uint64_t start_ns, uint64_t end_ns, size_t bytes, + const Place& place); + +struct MemEvenRecorder { + public: + void PushMemRecord(const void* ptr, const Place& place, size_t size); + void PopMemRecord(const void* ptr, const Place& place); + void Flush(); + static MemEvenRecorder& Instance() { return recorder; } + private: + struct RecordMemEvent { + RecordMemEvent(const Place& place, size_t bytes); + ~RecordMemEvent(); + + Place place_; + size_t bytes_; + uint64_t start_ns_; + uint64_t end_ns_; + std::string alloc_in_; + std::string free_in_; + }; + + static MemEvenRecorder recorder; + std::map>> + address_memevent_; + std::mutex mtx_; + MemEvenRecorder() {} + DISABLE_COPY_AND_ASSIGN(MemEvenRecorder); +}; + +Event* PushEvent(const std::string& name); void PopEvent(const std::string& name); struct RecordEvent { @@ -87,6 +127,41 @@ enum EventSortingKey { kGPUTime }; +template +struct EventList { + constexpr static size_t kMB = 1024 * 1024; + constexpr static size_t kEventBlockSize = 16 * kMB; + constexpr static size_t kEventSize = sizeof(T); + constexpr static size_t kEventAlign = alignof(T); + constexpr static size_t kNumBlock = + kEventBlockSize / + ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign); + + template + T* Record(Args&&... args) { + if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) { + event_blocks.emplace_front(); + event_blocks.front().reserve(kNumBlock); + } + event_blocks.front().emplace_back(std::forward(args)...); + return &event_blocks.front().back(); + } + + std::vector Reduce() { + std::vector result; + for (auto& block : event_blocks) { + result.insert(result.begin(), std::make_move_iterator(block.begin()), + std::make_move_iterator(block.end())); + } + event_blocks.clear(); + return result; + } + + void Clear() { event_blocks.clear(); } + + std::forward_list> event_blocks; +}; + // Enable the profiling function. void EnableProfiler(ProfilerState state); diff --git a/paddle/fluid/platform/profiler.proto b/paddle/fluid/platform/profiler.proto index e761d7b266e92fd5d47b5b6073ffc8bea1dc877d..cfa3c6906f83f750c8d6dc654f29b8fe95ec17ac 100644 --- a/paddle/fluid/platform/profiler.proto +++ b/paddle/fluid/platform/profiler.proto @@ -34,8 +34,25 @@ message Event { optional string detail_info = 9; } +message MemEvent { + enum Place { + CUDAPlace = 0; + CPUPlace = 1; + CUDAPinnedPlace = 2; + } + optional uint64 start_ns = 1; + optional uint64 end_ns = 2; + optional uint64 bytes = 3; + optional Place place = 4; + optional uint64 thread_id = 5; + optional uint32 device_id = 6; + optional string alloc_in = 7; + optional string free_in = 8; +} + message Profile { repeated Event events = 1; optional uint64 start_ns = 2; optional uint64 end_ns = 3; + repeated MemEvent mem_events = 4; } \ No newline at end of file diff --git a/paddle/fluid/pybind/imperative.cc b/paddle/fluid/pybind/imperative.cc index aeabed19abfda3c857f54e5ada54d52bf95e2602..6bbda69297a48ce27ce23282c4e08d49ee3cce6c 100644 --- a/paddle/fluid/pybind/imperative.cc +++ b/paddle/fluid/pybind/imperative.cc @@ -13,10 +13,18 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/pybind/imperative.h" + +#include +#include +#include +#include + #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/imperative/tracer.h" #include "paddle/fluid/imperative/type_defs.h" +#include "paddle/fluid/pybind/pybind_boost_headers.h" + namespace paddle { namespace pybind { @@ -31,20 +39,20 @@ void BindTracer(pybind11::module* m) { [](imperative::Tracer& self, imperative::OpBase* op, const imperative::VarBasePtrMap& inputs, const imperative::VarBasePtrMap& outputs, - framework::BlockDesc* block, + framework::AttributeMap attrs_map, const platform::CPUPlace expected_place, const bool stop_gradient = false) { - return self.Trace(op, inputs, outputs, block, expected_place, + return self.Trace(op, inputs, outputs, attrs_map, expected_place, stop_gradient); }) .def("trace", [](imperative::Tracer& self, imperative::OpBase* op, const imperative::VarBasePtrMap& inputs, const imperative::VarBasePtrMap& outputs, - framework::BlockDesc* block, + framework::AttributeMap attrs_map, const platform::CUDAPlace expected_place, const bool stop_gradient = false) { - return self.Trace(op, inputs, outputs, block, expected_place, + return self.Trace(op, inputs, outputs, attrs_map, expected_place, stop_gradient); }) .def("py_trace", &imperative::Tracer::PyTrace, diff --git a/paddle/fluid/pybind/imperative.h b/paddle/fluid/pybind/imperative.h index 8c48b2a7153c566930a074bd0bab1f054c13c2d5..8496cbfcb18798ee8ce1714431b7877bb2b7d377 100644 --- a/paddle/fluid/pybind/imperative.h +++ b/paddle/fluid/pybind/imperative.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include +#include #include #include "paddle/fluid/imperative/layer.h" #include "pybind11/pybind11.h" @@ -36,6 +37,8 @@ class Layer : public imperative::Layer { class PYBIND11_HIDDEN PyOpBase : public imperative::OpBase { public: using imperative::OpBase::OpBase; // Inherit constructors + + PyOpBase(const std::string& name) : OpBase(name) {} }; class PyVarBase : public imperative::VarBase { diff --git a/paddle/fluid/pybind/ir.cc b/paddle/fluid/pybind/ir.cc index 68f74a8531fff0c49c8a62d12f5cde7af77faf8a..c69ccd507210f976c1cb8ad072928b96693a948d 100644 --- a/paddle/fluid/pybind/ir.cc +++ b/paddle/fluid/pybind/ir.cc @@ -18,6 +18,7 @@ #include #include #include +#include #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" @@ -54,12 +55,14 @@ void BindGraph(py::module *m) { "The graph is a Directed Acyclic Single Static Assignment Graph, see " "`paddle::ir::Graph` for details.") .def(py::init()) + .def("clone", &Graph::Clone) .def("has", &Graph::Has) .def("get_int", &Graph::Get) .def("get_float", &Graph::Get) .def("get_double", &Graph::Get) .def("get_string", &Graph::Get) - .def("get_marked_nodes", &Graph::Get>) + .def("get_marked_nodes", &Graph::Get>, + return_value_policy::reference) .def("set", [](Graph &self, const std::string &attr_name, int attr) { return self.Set(attr_name, new int(attr)); }) .def("set", @@ -103,7 +106,8 @@ void BindGraph(py::module *m) { .def("retrieve_node", &Graph::RetrieveNode, return_value_policy::reference) .def("resolve_hazard", &Graph::ResolveHazard) - .def("origin_program_desc", &Graph::OriginProgram); + .def("origin_program_desc", &Graph::OriginProgram, + return_value_policy::reference); } void BindNode(py::module *m) { diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index e729be4a95a58510f1e0162af4216feaa400d971..7b5e417504fa16426279c8ed3c24d6d62e6be404 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -23,97 +23,7 @@ limitations under the License. */ #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/var_desc.h" -// Cast boost::variant for PyBind. -// Copy from -// https://github.com/pybind/pybind11/issues/576#issuecomment-269563199 -namespace pybind11 { -namespace detail { - -#if !defined(PYBIND11_HIDDEN) -#ifdef _WIN32 -#define PYBIND11_HIDDEN __declspec(dllexport) -#else -#define PYBIND11_HIDDEN __attribute__((visibility("hidden"))) -#endif -#endif - -// Can be replaced by a generic lambda in C++14 -struct PYBIND11_HIDDEN paddle_variant_caster_visitor - : public boost::static_visitor { - return_value_policy policy; - handle parent; - - paddle_variant_caster_visitor(return_value_policy policy, handle parent) - : policy(policy), parent(parent) {} - - template - handle operator()(T const &src) const { - return make_caster::cast(src, policy, parent); - } -}; - -template -struct paddle_variant_caster; - -template