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 = [{"num_batch": 1, "seq_len": 1, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 32, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 72, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 1024, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 2048, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 4096, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 8192, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 32768, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 1, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 2, "seq_len": 16, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 2, "seq_len": 36, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 8, "seq_len": 128, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 16, "seq_len": 128, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 4, "seq_len": 1024, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 2, "seq_len": 4096, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 16, "seq_len": 2048, "num_expert": 32, "topk": 5, "num_expert_group": -1, "topk_group": -1, "normalize": True, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 1, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 16, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 64, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 1024, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 2048, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 8192, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": False, "input_dtype": torch.bfloat16},
                            {"num_batch": 1, "seq_len": 32768, "num_expert": 160, "topk": 6, "num_expert_group": 8, "topk_group": 3, "normalize": 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()
  titles = ["num_batch", "seq_len", "num_expert", "topk", "num_expert_group", "topk_group", "normalize", "input_dtype", "hardware_time(us)", "e2e_latency(us)", "IO efficiency"]
  contents = []
  bd = get_band_width()

  for params_dict in e2e_time_param_dict_list:
    num_batch = params_dict["num_batch"]
    seq_len = params_dict["seq_len"]
    num_expert = params_dict["num_expert"]
    topk = params_dict["topk"]
    num_expert_group = params_dict["num_expert_group"]
    topk_group = params_dict["topk_group"]
    normalize = params_dict["normalize"]
    dtype = params_dict["input_dtype"]
    if dtype == torch.bfloat16 and not torch_mlu.mlu.is_bf16_supported():
      dtype = torch.half
    input = torch.randn(num_batch, seq_len, num_expert, dtype=dtype, device='mlu')
    mask = torch.randint(0, 2, (1, seq_len, num_expert), dtype = dtype, device='mlu')
    if num_expert_group > 1:
        mask = None
    normed_by = "softmax_logit"
    reduce_weight = torch.empty(num_batch, topk, dtype=torch.float, device='mlu')
    expert_id = torch.empty(num_batch, topk, dtype=torch.int32, device='mlu')
    hardware_time, e2e_time = benchmark_forward(tmo.moe_softmax_topk,
                                                input,
                                                topk,
                                                normalize,
                                                num_expert_group,
                                                topk_group,
                                                mask,
                                                normed_by,
                                                repeats=args.repeat_times)
    io_bytes = input.element_size() * input.nelement() + \
               reduce_weight.element_size() * reduce_weight.nelement() + \
               expert_id.element_size() * expert_id.nelement()
    io_eff = io_bytes / hardware_time / bd
    content = [f"{num_batch}", f"{seq_len}", f"{num_expert}", f"{topk}", f"{num_expert_group}", f"{topk_group}", f"{normalize}", 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()