mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 00:21:07 +01:00
3b2b6ae1a8
Summary: This speeds up the inlining pass of FairSeq model from 180s -> 13s. Pull Request resolved: https://github.com/pytorch/pytorch/pull/35424 Differential Revision: D20657271 Pulled By: eellison fbshipit-source-id: 7a9006858c2f1b157f5a3f36ed2b3774cc186de8
77 lines
2.2 KiB
C++
77 lines
2.2 KiB
C++
#include <torch/csrc/jit/api/function_impl.h>
|
|
#include <torch/csrc/jit/frontend/error_report.h>
|
|
#include <torch/csrc/jit/passes/inliner.h>
|
|
#include <torch/csrc/jit/passes/peephole.h>
|
|
#include "torch/csrc/jit/passes/constant_propagation.h"
|
|
|
|
namespace torch {
|
|
namespace jit {
|
|
namespace {
|
|
c10::FunctionSchema defaultSchemaFor(const Function& function) {
|
|
std::vector<c10::Argument> args;
|
|
std::vector<c10::Argument> returns;
|
|
Graph& g = *function.graph();
|
|
size_t num_inputs = function.num_inputs();
|
|
for (size_t i = 0; i < num_inputs; ++i) {
|
|
const Value* v = g.inputs().at(i);
|
|
std::string name = v->hasDebugName() ? v->debugNameBase()
|
|
: ("argument_" + c10::to_string(i));
|
|
args.emplace_back(std::move(name), unshapedType(g.inputs()[i]->type()));
|
|
}
|
|
for (size_t i = 0; i < g.outputs().size(); ++i) {
|
|
returns.emplace_back("", unshapedType(g.outputs()[i]->type()));
|
|
}
|
|
return {function.name(), "", std::move(args), std::move(returns)};
|
|
}
|
|
} // namespace
|
|
|
|
void placeholderCreator(GraphFunction&) {
|
|
throw RecursiveMethodCallError();
|
|
}
|
|
|
|
void GraphFunction::run(Stack& stack) {
|
|
get_executor().run(stack);
|
|
}
|
|
|
|
void GraphFunction::run(Stack&& stack) {
|
|
run(stack);
|
|
}
|
|
|
|
IValue GraphFunction::operator()(
|
|
std::vector<IValue> stack,
|
|
const Kwargs& kwargs) {
|
|
getSchema().checkAndNormalizeInputs(stack, kwargs);
|
|
run(stack);
|
|
return stack.front();
|
|
}
|
|
|
|
void GraphFunction::ensure_defined() {
|
|
if (function_creator_) {
|
|
auto creator = function_creator_;
|
|
function_creator_ = placeholderCreator;
|
|
creator(*this);
|
|
function_creator_ = nullptr;
|
|
}
|
|
check_single_output();
|
|
}
|
|
|
|
const c10::FunctionSchema& GraphFunction::getSchema() const {
|
|
if (schema_ == nullptr) {
|
|
schema_ = std::make_unique<c10::FunctionSchema>(defaultSchemaFor(*this));
|
|
}
|
|
return *schema_;
|
|
}
|
|
|
|
void preoptimizeGraph(std::shared_ptr<Graph>& graph) {
|
|
// Peephole Optimize cleans up many "is None" checks and creates constant prop
|
|
// opportunities
|
|
PeepholeOptimize(graph);
|
|
// AliasDb construction can be slow, so run it just on immutable types
|
|
// to clean up constant Ifs & other easy wins
|
|
ConstantPropagationImmutableTypes(graph);
|
|
Inline(*graph);
|
|
}
|
|
|
|
} // namespace jit
|
|
} // namespace torch
|