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.

op_mapping_info.proto 1.4 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273
  1. syntax = "proto3";
  2. package aicpu.dump;
  3. message Shape {
  4. repeated uint64 dim = 1;
  5. }
  6. message Output {
  7. int32 data_type = 1;
  8. int32 format = 2;
  9. Shape shape = 3;
  10. uint64 address = 4;
  11. string original_name = 5;
  12. int32 original_output_index = 6;
  13. int32 original_output_data_type = 7;
  14. int32 original_output_format = 8;
  15. uint64 size = 9;
  16. }
  17. message Input {
  18. int32 data_type =1;
  19. int32 format = 2;
  20. Shape shape = 3;
  21. uint64 address = 4;
  22. uint64 size = 5;
  23. }
  24. enum BufferType {
  25. L1 = 0;
  26. }
  27. message OpBuffer {
  28. BufferType buffer_type = 1;
  29. uint64 address = 2;
  30. uint64 size = 3;
  31. }
  32. message Op {
  33. string op_name = 1;
  34. string op_type = 2;
  35. }
  36. message Task {
  37. uint32 task_id = 1;
  38. uint32 stream_id = 2;
  39. Op op = 3;
  40. repeated Output output = 4;
  41. bool end_graph = 5;
  42. repeated Input input = 6;
  43. repeated OpBuffer buffer = 7;
  44. }
  45. message OpMappingInfo {
  46. string dump_path = 1;
  47. oneof model_name_param {
  48. string model_name = 2;
  49. }
  50. oneof model_id_param {
  51. uint32 model_id = 3;
  52. }
  53. oneof step_id {
  54. uint64 step_id_addr = 4;
  55. }
  56. oneof iterations_per_loop {
  57. uint64 iterations_per_loop_addr = 5;
  58. }
  59. oneof loop_cond {
  60. uint64 loop_cond_addr = 6;
  61. }
  62. uint32 flag = 7; // 0x01 load, 0x00 unload
  63. repeated Task task = 8;
  64. string dump_step = 9;
  65. }

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示