Files
pytorch/torch/csrc/jit/mobile/code.h
Zhengxu Chen d459e79500 [jit][edge] Remove usage of shared_ptr<mobile::Code>. (#68037)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/68037

Right now mobile::Code doesn't outlive its enclosing Function, and all accesses to Code happens inside interpreter loop which doesn't outlive the module, so we don't need to use std::shared_ptr here. This also should saves us 1-2 KB for binary size, because shared_ptr seems to bloat on arm64 android.
ghstack-source-id: 145818696

Test Plan: eyes.

Reviewed By: qihqi, tugsbayasgalan

Differential Revision: D32264616

fbshipit-source-id: d83f538d6604cf75fd7728a25127b4849ce7ab2a
2021-12-16 13:11:46 -08:00

38 lines
1.1 KiB
C++

#pragma once
#include <vector>
#include <ATen/core/ivalue.h>
#include <ATen/core/operator_name.h>
#include <torch/csrc/jit/runtime/instruction.h>
namespace torch {
namespace jit {
namespace mobile {
using Stack = std::vector<c10::IValue>;
using DebugHandle = int64_t;
class Function;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
struct Code {
std::vector<Instruction> instructions_;
std::vector<DebugHandle> debug_handles_;
std::vector<c10::OperatorName> op_names_;
std::vector<int> operator_input_sizes_;
std::vector<std::function<void(Stack&)>> operators_;
std::vector<c10::IValue> constants_;
std::vector<c10::TypePtr> types_;
// TODO After we actually export CALL instructions we can remove this.
// We may need a two-stage importing scheme, where we firstly construct all
// function objects, and then append referenced function pointers. This could
// be done in parseMethods().
std::vector<mobile::Function*> functions_;
size_t register_size_ = 0; // Aggregated output size.
};
} // namespace mobile
} // namespace jit
} // namespace torch