You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

trace.cpp 2.1 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061
  1. /**
  2. * \file imperative/python/src/trace.cpp
  3. * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
  4. *
  5. * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
  6. *
  7. * Unless required by applicable law or agreed to in writing,
  8. * software distributed under the License is distributed on an
  9. * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
  10. * implied.
  11. */
  12. #include "./trace.h"
  13. #include "./helper.h"
  14. #include "megbrain/imperative/ops/autogen.h"
  15. namespace py = pybind11;
  16. namespace mgb::imperative::python {
  17. apply_result_t apply_trace(ApplyContext& ctx) {
  18. apply_result_t outputs;
  19. if (ctx.backward) {
  20. // reach here when compiled=True
  21. auto args = py::tuple(ctx.nargs + 1);
  22. args[0] = py::cast(ctx.op);
  23. for (size_t i = 0; i < ctx.nargs; i++) {
  24. args[i + 1] = py::cast(ctx.args[i]->m_var);
  25. }
  26. py::object pyout = py::reinterpret_steal<py::object>(
  27. PyObject_Call(cpp_apply_backward_varnode, args.ptr(), nullptr));
  28. if (!pyout) throw py::error_already_set();
  29. // assumption: python function always returns PyList
  30. auto tup = py::reinterpret_borrow<py::list>(pyout);
  31. for (size_t i = 0; i < tup.size(); i++) {
  32. auto pitem = tup[i].cast<cg::VarNode*>();
  33. outputs.emplace_back(std::make_shared<Tensor>(pitem));
  34. }
  35. return outputs;
  36. }
  37. auto args = py::tuple(ctx.nargs + 1);
  38. args[0] = py::cast(ctx.op);
  39. for (size_t i = 0; i < ctx.nargs; i++) {
  40. args[i + 1] = TensorWrapper::make(ctx.args[i]->shared_from_this());
  41. }
  42. auto pyout = PyObject_Call(cpp_apply_with_tracing, args.ptr(), nullptr);
  43. if (!pyout) throw py::error_already_set();
  44. // assumption: python function always returns PyList
  45. auto tup = py::reinterpret_steal<py::list>(pyout);
  46. for (size_t i = 0; i < tup.size(); i++) {
  47. auto tw = TensorWrapper::try_cast(tup[i].ptr());
  48. outputs.emplace_back(tw->m_tensor);
  49. }
  50. return outputs;
  51. }
  52. } // namespace mgb::imperative::python

MegEngine 安装包中集成了使用 GPU 运行代码所需的 CUDA 环境,不用区分 CPU 和 GPU 版。 如果想要运行 GPU 程序,请确保机器本身配有 GPU 硬件设备并安装好驱动。 如果你想体验在云端 GPU 算力平台进行深度学习开发的感觉,欢迎访问 MegStudio 平台