Skip to content

[Dy2St][PIR] Replace output with inplace source #63040

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
SigureMo merged 5 commits into from
Apr 1, 2024
Merged

[Dy2St][PIR] Replace output with inplace source #63040

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8c700e8
[Dy2St] Replace output with inplace source
SigureMo Mar 26, 2024
490a3af
handle forward outputs and backward grads
SigureMo Mar 27, 2024
07151b5
target should be key, fix multiple target map to same source
SigureMo Mar 27, 2024
618e7ad
only process parameter op
SigureMo Mar 29, 2024
4468b8b
cleanup debug info
SigureMo Mar 31, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
54 changes: 30 additions & 24 deletions paddle/fluid/eager/to_static/run_program_op_node.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,39 +91,45 @@ static bool IsVariableRefArray(const Tensor &tensor) {

static auto GetNameFromValue(const ::pir::Block *block,
const std::vector<::pir::Value> &values,
bool is_input) {
bool allow_input,
bool allow_output) {
PADDLE_ENFORCE_EQ(
allow_input || allow_output,
true,
paddle::platform::errors::InvalidArgument(
"GetNameFromValue should allow input or output at least one."));
// we use name here, later value is used directly.
std::unordered_map<::pir::Value, std::string> value2name;
if (is_input) {
if (allow_input) {
for (auto &kwarg : block->kwargs()) {
value2name[kwarg.second] = kwarg.first;
}
}
for (auto &op : *block) {
std::string name;
if (is_input && op.name() == "pd_op.data") {
if (allow_input && op.name() == "pd_op.data") {
name =
op.attributes().at("name").dyn_cast<pir::StrAttribute>().AsString();
value2name[op.results()[0].Value::impl()] = name;
} else if (!is_input && op.name() == "builtin.set_parameter") {
} else if (allow_output && op.name() == "builtin.set_parameter") {
name = op.attributes()
.at("parameter_name")
.dyn_cast<pir::StrAttribute>()
.AsString();
value2name[op.operand(0).source()] = name;
} else if (!is_input && op.name() == "builtin.shadow_output") {
} else if (allow_output && op.name() == "builtin.shadow_output") {
name = op.attributes()
.at("output_name")
.dyn_cast<pir::StrAttribute>()
.AsString();
value2name[op.operand(0).source()] = name;
} else if (is_input && op.name() == "builtin.parameter") {
} else if (allow_input && op.name() == "builtin.parameter") {
name = op.attributes()
.at("parameter_name")
.dyn_cast<pir::StrAttribute>()
.AsString();
value2name[op.result(0).Value::impl()] = name;
} else if (is_input && op.name() == "builtin.constant") {
} else if (allow_input && op.name() == "builtin.constant") {
if (op.isa<pir::ConstantTensorOp>()) {
name = op.dyn_cast<pir::ConstantTensorOp>().tensor_name();
value2name[op.result(0).Value::impl()] = name;
Expand Down Expand Up @@ -248,12 +254,7 @@ static void ShareTensorsIntoScopeByValue(
const std::vector<Tensor> &tensors,
const std::vector<::pir::Value> &values,
paddle::framework::Scope *scope) {
auto names = GetNameFromValue(block, values, true);
if (VLOG_IS_ON(4)) {
for (auto &s : names) {
VLOG(4) << "ShareTensorIntoScopeByValue name: " << s;
}
}
auto names = GetNameFromValue(block, values, true, false);
ShareTensorsIntoScopeWithName(tensors, names, scope);
}

Expand All @@ -262,11 +263,16 @@ static void ShareTensorsFromScopeByValue(
const std::vector<Tensor *> &tensors,
const std::vector<::pir::Value> &values,
paddle::framework::Scope *scope) {
auto names = GetNameFromValue(block, values, false);
// NOTE(SigureMo): If the program has an inplace chain connecting
// an input value to an output value, the output value will be
// replaced with the input value, so we set the `allow_input` to
// `true` in `GetNameFromValue`
auto names = GetNameFromValue(block, values, true, true);
for (size_t i = 0; i < tensors.size(); ++i) {
auto &name = names[i];
auto &value = values[i];
VLOG(2) << "share " << name << " from scope";
VLOG(4) << "Share Tensor From Scope: " << name;

if (value.impl() == nullptr) {
// skip stop_gradient.
continue;
Expand Down Expand Up @@ -524,20 +530,20 @@ inline void PirRunProgramAPI(
// *backward_program);

// update interpretercore skip_gc_var
auto skip_names =
details::GetNameFromValue(forward_global_block, middle_values, false);
auto skip_names = details::GetNameFromValue(
forward_global_block, middle_values, false, true);
auto skip_names_set =
std::set<std::string>(skip_names.begin(), skip_names.end());
auto no_need_buffer_values = PADDLE_GET_CONST(std::vector<::pir::Value>,
attrs.at("no_need_buffers"));
auto no_need_buffer_names = details::GetNameFromValue(
forward_global_block, no_need_buffer_values, false);
forward_global_block, no_need_buffer_values, false, true);
for (auto &name : no_need_buffer_names) {
VLOG(4) << "Find no need buffer vars with name:" << name;
skip_names_set.erase(name);
}
skip_names =
details::GetNameFromValue(forward_global_block, output_values, false);
skip_names = details::GetNameFromValue(
forward_global_block, output_values, false, true);
skip_names_set.insert(skip_names.begin(), skip_names.end());
details::print_collection(skip_names_set);
interpreter_core->SetSkipGcVars(skip_names_set);
Expand Down Expand Up @@ -1127,11 +1133,11 @@ inline void PirRunProgramGradAPI(

// get all eager gc vars
std::set<std::string> skip_eager_delete_vars;
auto skip_names =
details::GetNameFromValue(backward_global_block, x_grad_values, false);
auto skip_names = details::GetNameFromValue(
backward_global_block, x_grad_values, false, true);
skip_eager_delete_vars.insert(skip_names.begin(), skip_names.end());
skip_names =
details::GetNameFromValue(backward_global_block, p_grad_values, false);
skip_names = details::GetNameFromValue(
backward_global_block, p_grad_values, false, true);
skip_eager_delete_vars.insert(skip_names.begin(), skip_names.end());
interpreter_core->SetSkipGcVars(skip_eager_delete_vars);
cache.UpdateSkipEagerDeleteVars(program_id,
Expand Down
Loading