import torch
import torch_mlu
import torch_mlu_ops as tmo
from common import *
import argparse
from tabulate import tabulate
import os

e2e_time_param_dict_list = [{"batch": 1, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 16, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 72, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 1024, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 4096, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 8192, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 32768, "seq_len": 5, "hidden_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "input_dtype": [torch.bfloat16]}]

def main():
  parser = argparse.ArgumentParser()
  parser.add_argument('--repeat_times', type=int, default=10, help='repeat times for testing')
  parser.add_argument('--csv', action='store_true', help='write the report data to csv')
  parser.add_argument('-o', type=str, help='specify the output folder name under --csv mode')

  args = parser.parse_args()
  device = 'mlu'
  titles = ["input_shape", "act_mode", "is_gated", "input_dtype", "hardware_time(us)", "e2e_latency(us)", "IO efficiency"]
  contents = []
  bd = get_band_width()
  for params_dict in e2e_time_param_dict_list:
    batch = params_dict["batch"]
    seq_len = params_dict["seq_len"]
    hidden_size = params_dict["hidden_size"]
    act_mode = params_dict["act_mode"]
    is_gated = params_dict["is_gated"]
    input_dtype_list = params_dict["input_dtype"]
    for dtype in input_dtype_list:
      if dtype == torch.bfloat16 and not torch_mlu.mlu.is_bf16_supported():
        dtype = torch.half
      input = torch.randn(batch, seq_len, hidden_size).to(device).to(dtype)
      hardware_time, e2e_time = benchmark_forward(tmo.active,
                                                  input,
                                                  act_mode,
                                                  is_gated,
                                                  repeats=args.repeat_times)
      io_bytes = input.element_size() * input.nelement() * (2 - 0.5 * is_gated)
      io_eff = io_bytes / hardware_time / bd
      content = [f"{batch,seq_len,hidden_size}", f"{act_mode}", f"{is_gated}", f"{dtype}", f"{hardware_time}", f"{e2e_time}", f"{io_eff}"]
      contents.append(content)
  table = [titles] + contents
  print(tabulate(table, headers="firstrow", tablefmt="grid"))

  if args.csv:
    current_file_path = __file__
    _, file_name = os.path.split(current_file_path)
    save_to_csv(table, args.o, file_name)

if __name__=="__main__":
  main()