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

e2e_time_param_dict_list = [{"batch": 1, "seq_len": 1024, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": True, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 1, "seq_len": 4096, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": False, "input_dtype": [torch.bfloat16]},
                            {"batch": 1, "seq_len": 8192, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 1, "seq_len": 32768, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 1, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 16, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 32, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 64, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 128, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 256, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},
                            {"batch": 512, "seq_len": 1, "inner_size": 1024,
                             "act_mode": "gelu", "is_gated": False, "has_bias": True,
                             "is_ep": True, "input_dtype": [torch.bfloat16]},]

def gen_data(num_expert,
             total_tokens,
             inner_size,
             output_stride,
             dtype,
             is_gated,
             has_bias,
             is_ep):
    ci = inner_size * (1 + is_gated)
    input = torch.randn(total_tokens, ci, dtype=dtype, device='mlu')
    cusum_token_count, token_count = generate_token_count(num_expert, total_tokens)
    output = torch.empty((total_tokens, inner_size), dtype=dtype, device='mlu')
    output.as_strided(output.size(), (output_stride, 1))
    start_expert_id = random.randint(0, num_expert - 1) if is_ep else 0
    expert_size = random.randint(1, num_expert - start_expert_id) if is_ep else num_expert
    bias = torch.randn(num_expert, ci, dtype=dtype, device='mlu') if has_bias else None
    return input, bias, token_count, cusum_token_count, output, start_expert_id, expert_size

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 = ["input_shape", "act_mode", "is_gated", "has_bias", "expert_num", "start_expert_id",
            "expert_size", "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"]
    inner_size = params_dict["inner_size"]
    act_mode = params_dict["act_mode"]
    is_gated = params_dict["is_gated"]
    input_dtype_list = params_dict["input_dtype"]
    has_bias = params_dict["has_bias"]
    is_ep = params_dict["is_ep"]
    for dtype in input_dtype_list:
      if dtype == torch.bfloat16 and not torch_mlu.mlu.is_bf16_supported():
        dtype = torch.half
      expert_num = expert_num = random.randint(1, 256)
      input, bias, token_count, cusum_token_count, output, start_expert_id, expert_size = \
            gen_data(expert_num, batch * seq_len, inner_size, inner_size, dtype, is_gated, has_bias, is_ep)

      real_bias = bias[start_expert_id:start_expert_id + expert_size] if has_bias else None
      hardware_time, e2e_time = benchmark_forward(tmo.moe_active,
                                                  input,
                                                  act_mode,
                                                  is_gated,
                                                  output,
                                                  real_bias,
                                                  cusum_token_count.mlu() if has_bias or is_ep else None,
                                                  start_expert_id,
                                                  expert_size,
                                                  repeats=args.repeat_times)

      io_bytes = input.element_size() * input.nelement() * (2 - 0.5 * is_gated) + \
                 real_bias.element_size() * real_bias.nelement() + \
                 (cusum_token_count.element_size() * cusum_token_count.nelement()) if has_bias or is_ep else 0
      io_eff = io_bytes / hardware_time / bd
      content = [f"{batch,seq_len,inner_size}", f"{act_mode}", f"{is_gated}", f"{has_bias}", f"{expert_num}",
                 f"{start_expert_id}", f"{expert_size}", 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()