Skip to content

Commit

Permalink
fix loading cache (PaddlePaddle#457)
Browse files Browse the repository at this point in the history
  • Loading branch information
@gglin001
gglin001 committed Feb 21, 2022
1 parent cfe4f6c commit 978b6f6
Show file tree
Hide file tree
Showing 5 changed files with 140 additions and 138 deletions.
4 changes: 2 additions & 2 deletions paddle/fluid/framework/ir/ipu/optimizer_extract_pass.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,7 @@ const bool is_regularization_op(const std::string& op_namescope) {
}

void IpuOptimizerExtractPass::ApplyImpl(ir::Graph* graph) const {
// 这里构建的 op 符合 popart 的定义, 涉及到的一些值需要在 LowerOptimier 时获得
// optimizer values will be extracted when lowering optimizer in ipu_backend
OpDesc new_op("popart_optimizer", {}, {}, {});
new_op.SetAttr("op_role", 0);
new_op.SetAttr("with_lr_sched", false);
Expand Down Expand Up @@ -86,7 +86,7 @@ void IpuOptimizerExtractPass::ApplyImpl(ir::Graph* graph) const {
bool is_regularization = is_regularization_op(op_namescope);

VLOG(10) << "found optimizer releated op: " << op_type;
// initial larning_rate will be set in LowerOptimier
// initial larning_rate will be set in ipu_backend
set_ops.insert(op_type);
if (op_type == "sgd") {
auto type = std::string{"sgd"};
Expand Down
8 changes: 3 additions & 5 deletions paddle/fluid/framework/ir/ipu/popart_canonicalization_pass.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@

#include "paddle/fluid/framework/ir/ipu/popart_canonicalization_pass.h"

#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/ir/pass_tester_helper.h"
#include "paddle/fluid/platform/device/ipu/popart_canonicalization/canonicalization_utils.h"

Expand All @@ -28,11 +29,8 @@ void PopartCanonicalizationPass::ApplyImpl(ir::Graph* graph) const {

auto custom_ops = Get<std::unordered_set<std::string>>("custom_ops");
std::vector<std::string> missing_ops;
auto nodes = graph->Nodes();
for (auto* node : nodes) {
if (!node->IsOp()) {
continue;
}
auto sorted_ops = TopologySortOperations(*graph);
for (auto* node : sorted_ops) {
auto* op = node->Op();
auto op_type = op->Type();

Expand Down
16 changes: 8 additions & 8 deletions paddle/fluid/platform/device/ipu/ipu_backend.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -43,17 +43,17 @@ void IpuBackend::Compile(Graph* graph,
const std::vector<std::string>& feed_list,
const std::vector<std::string>& fetch_list) {
VLOG(10) << "enter IpuBackend::Compile";
compiler_->Prepare();
executor_->SetCompilerResources(compiler_->GetResources());

compiler_->InitInputs(graph, feed_list);
compiler_->LowerConstants(graph, scope_);
compiler_->LowerWeights(graph, scope_);
compiler_->LowerBody(graph);
compiler_->Prepare(graph);
compiler_->InitInputs(feed_list);
compiler_->LowerConstants(scope_);
compiler_->LowerWeights(scope_);
compiler_->LowerBody();
compiler_->InitOutputs(fetch_list);
if (ipu_strategy_->is_training) {
compiler_->LowerOptimier(graph, scope_);
compiler_->LowerOptimizer(scope_);
}
executor_->SetCompilerResources(compiler_->GetResources());

is_compiled_ = true;
// when call compile, means a new graph
is_prepared_ = false;
Expand Down
190 changes: 93 additions & 97 deletions paddle/fluid/platform/device/ipu/ipu_compiler.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -98,16 +98,30 @@ TO GetCastSigAttrAllowNull(std::string attr, OpDesc* op_desc) {
}
}

GraphHelper::GraphHelper(const Graph* g) {
graph = g;
sorted_ops = framework::ir::TopologySortOperations(*g);
for (auto* node : g->Nodes()) {
nodes_id_map[node->id()] = node;
if (node->IsVar()) {
vars_name_map[node->Name()] = node;
sorted_vars_id.push_back(node->id());
}
}
std::sort(sorted_vars_id.begin(), sorted_vars_id.end());
}

Compiler::Compiler() { RegisterOpFunc(); }

Compiler::~Compiler() {
builder_.reset();
resources_.reset();
}

void Compiler::Prepare() {
void Compiler::Prepare(const Graph* graph) {
builder_ = popart::Builder::create();
resources_ = std::make_unique<CompilerResources>();
graph_helper_ = std::make_unique<GraphHelper>(graph);
}

void Compiler::RegisterOpFunc() {
Expand Down Expand Up @@ -171,93 +185,24 @@ void Compiler::RegisterOpFunc() {
#undef INT_VEC
}

void Compiler::LowerBody(const Graph* graph) {
VLOG(10) << "enter Compiler::LowerBody";
auto nodes = framework::ir::TopologySortOperations(*graph);
for (auto* node : nodes) {
auto* op_desc = node->Op();
auto op_type = op_desc->Type();
VLOG(10) << "lowering op: " << op_type;

if (op_type == "popart_constant") {
// pass
} else if (op_type == "popart_optimizer") {
// pass
} else if (op_type == "popart_checkpointoutput") {
auto inputs = GetOpInputs(op_desc);
auto outputs = GetOpOutputs(op_desc);
auto output_ids = builder_->checkpointOutput(inputs);
InsertTensors(outputs, output_ids);
} else if (op_type == "popart_custom_op") {
auto inputs = GetOpInputs(op_desc);
auto outputs = GetOpOutputs(op_desc);
auto debug_context = BuildDebugContext(op_desc);
auto attributes = std::map<std::string, popart::any>{};
for (auto& attr : op_desc->GetAttrMap()) {
CustomOpAttrVisitor visitor(&attributes, attr.first);
boost::apply_visitor(visitor, attr.second);
}
auto __op_type =
BOOST_GET_CONST(std::string, op_desc->GetAttr("__op_type"));
VLOG(10) << "Build graph from custom op: " << __op_type;
auto it = custom_ops_.find(__op_type);
auto output_ids =
builder_->customOp(it->second.popart_op, it->second.popart_op.version,
inputs, outputs.size(), attributes, debug_context);
SetIpuIndexStage(output_ids, op_desc);
InsertTensors(outputs, output_ids);
} else if (op_type == "popart_printtensor") {
auto inputs = GetOpInputs(op_desc);
auto outputs = GetOpOutputs(op_desc);
auto debug_context = BuildDebugContext(op_desc);
auto print_gradient =
BOOST_GET_CONST(int64_t, op_desc->GetAttr("print_gradient"));
auto title = BOOST_GET_CONST(std::string, op_desc->GetAttr("title"));
auto output_ids = builder_->aiGraphcoreOpset1().printtensor(
inputs, print_gradient, debug_context, title);
SetIpuIndexStage(output_ids, op_desc);
InsertTensors(outputs, output_ids);
} else {
auto itr = name_function_.find(op_type);
if (itr != name_function_.end()) {
itr->second(node->Op());
} else {
PADDLE_THROW(platform::errors::NotFound(
"%s is not registered, please check for unsupported operators for "
"running on IPU",
op_type));
}
}
}
VLOG(10) << "leave Compiler::LowerBody";
}

void Compiler::InitInputs(Graph* graph,
const std::vector<std::string>& feed_list) {
void Compiler::InitInputs(const std::vector<std::string>& feed_list) {
for (const auto& feed_name : feed_list) {
feed_list_.push_back(feed_name);
for (const Node* n : graph->Nodes()) {
if (n->IsVar()) {
auto* var_desc = n->Var();
if (feed_name == var_desc->Name()) {
VLOG(10) << "feed_name= " << var_desc->Name();
auto data_type = VarType2PopartType(var_desc->GetDataType());
popart::TensorInfo input_info{data_type, var_desc->GetShape()};
VLOG(10) << "popart input_info = " << input_info;
popart::TensorId tensor_id =
builder_->addInputTensor(input_info, feed_name);
VLOG(10) << "popart input tensor id = " << tensor_id;
resources_->inputs.push_back(tensor_id);
resources_->tensors.emplace(var_desc->Name(), tensor_id);
}
}
}
auto* node = graph_helper_->vars_name_map[feed_name];
auto* var_desc = node->Var();
VLOG(10) << "feed_name= " << var_desc->Name();
auto data_type = VarType2PopartType(var_desc->GetDataType());
popart::TensorInfo input_info{data_type, var_desc->GetShape()};
VLOG(10) << "popart input_info = " << input_info;
popart::TensorId tensor_id =
builder_->addInputTensor(input_info, feed_name);
VLOG(10) << "popart input tensor id = " << tensor_id;
resources_->inputs.push_back(tensor_id);
resources_->tensors.emplace(var_desc->Name(), tensor_id);
}
}

void Compiler::InitOutputs(const std::vector<std::string>& fetch_list) {
for (const auto& fetch_name : fetch_list) {
fetch_list_.push_back(fetch_name);
auto tensor = resources_->tensors.find(fetch_name);
PADDLE_ENFORCE_NE(
tensor, resources_->tensors.end(),
Expand All @@ -271,14 +216,10 @@ void Compiler::InitOutputs(const std::vector<std::string>& fetch_list) {
}
}

void Compiler::LowerConstants(const Graph* graph, const Scope* scope) {
void Compiler::LowerConstants(const Scope* scope) {
auto& kid_scope = scope->NewScope();
VLOG(10) << "enter Compiler::LowerConstants";
for (auto* node : graph->Nodes()) {
if (!node->IsOp()) {
continue;
}

for (auto* node : graph_helper_->sorted_ops) {
auto* op_desc = node->Op();
auto op_type = op_desc->Type();
if (op_type == "popart_constant") {
Expand Down Expand Up @@ -308,17 +249,16 @@ void Compiler::LowerConstants(const Graph* graph, const Scope* scope) {
VLOG(10) << "leave Compiler::LowerConstants";
}

void Compiler::LowerWeights(const Graph* graph, const Scope* scope) {
void Compiler::LowerWeights(const Scope* scope) {
VLOG(10) << "enter Compiler::LowerWeights";
PADDLE_ENFORCE_NOT_NULL(scope,
platform::errors::PreconditionNotMet(
"You should call set_scope before LowerWeights"));
// at this step, the graph doesn't contains optimizer related states
for (const auto* node : graph->Nodes()) {
for (auto id : graph_helper_->sorted_vars_id) {
auto* node = graph_helper_->nodes_id_map[id];
if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
if (node->Var()->Persistable() && node->inputs.empty()) {
auto var_name = node->Var()->Name();
if (resources_->tensors.count(var_name) != 0) {
VLOG(10) << "found existed one, skip lowering Weight: " << var_name;
continue;
}
VLOG(10) << "lowering weight: " << var_name;
Expand All @@ -344,12 +284,68 @@ void Compiler::LowerWeights(const Graph* graph, const Scope* scope) {
VLOG(10) << "leave Compiler::LowerWeights";
}

void Compiler::LowerOptimier(const Graph* graph, const Scope* scope) {
for (auto* node : graph->Nodes()) {
if (!node->IsOp()) {
continue;
void Compiler::LowerBody() {
VLOG(10) << "enter Compiler::LowerBody";
for (auto* node : graph_helper_->sorted_ops) {
auto* op_desc = node->Op();
auto op_type = op_desc->Type();
VLOG(10) << "lowering op: " << op_type;

if (op_type == "popart_constant") {
// pass
} else if (op_type == "popart_optimizer") {
// pass
} else if (op_type == "popart_checkpointoutput") {
auto inputs = GetOpInputs(op_desc);
auto outputs = GetOpOutputs(op_desc);
auto output_ids = builder_->checkpointOutput(inputs);
InsertTensors(outputs, output_ids);
} else if (op_type == "popart_custom_op") {
auto inputs = GetOpInputs(op_desc);
auto outputs = GetOpOutputs(op_desc);
auto debug_context = BuildDebugContext(op_desc);
auto attributes = std::map<std::string, popart::any>{};
for (auto& attr : op_desc->GetAttrMap()) {
CustomOpAttrVisitor visitor(&attributes, attr.first);
boost::apply_visitor(visitor, attr.second);
}
auto __op_type =
BOOST_GET_CONST(std::string, op_desc->GetAttr("__op_type"));
VLOG(10) << "Build graph from custom op: " << __op_type;
auto it = custom_ops_.find(__op_type);
auto output_ids =
builder_->customOp(it->second.popart_op, it->second.popart_op.version,
inputs, outputs.size(), attributes, debug_context);
SetIpuIndexStage(output_ids, op_desc);
InsertTensors(outputs, output_ids);
} else if (op_type == "popart_printtensor") {
auto inputs = GetOpInputs(op_desc);
auto outputs = GetOpOutputs(op_desc);
auto debug_context = BuildDebugContext(op_desc);
auto print_gradient =
BOOST_GET_CONST(int64_t, op_desc->GetAttr("print_gradient"));
auto title = BOOST_GET_CONST(std::string, op_desc->GetAttr("title"));
auto output_ids = builder_->aiGraphcoreOpset1().printtensor(
inputs, print_gradient, debug_context, title);
SetIpuIndexStage(output_ids, op_desc);
InsertTensors(outputs, output_ids);
} else {
auto itr = name_function_.find(op_type);
if (itr != name_function_.end()) {
itr->second(node->Op());
} else {
PADDLE_THROW(platform::errors::NotFound(
"%s is not registered, please check for unsupported operators for "
"running on IPU",
op_type));
}
}
}
VLOG(10) << "leave Compiler::LowerBody";
}

void Compiler::LowerOptimizer(const Scope* scope) {
for (auto* node : graph_helper_->sorted_ops) {
auto* op_desc = node->Op();
auto op_type = op_desc->Type();
if (op_type == "popart_optimizer") {
Expand Down
Loading

0 comments on commit 978b6f6

Please sign in to comment.