import torch
import torch_mlu
import torch_mlu_ops as tmo
from common import benchmark_forward, save_to_csv
import argparse
from tabulate import tabulate
import os

e2e_time_param_dict_list = [{"batch": 1, "seq_len": 1024, "head_num": 32, "head_size": 128, "rotary_dim": 128,
                             "interleaved": False, "discrete": True, "dynamic_ntk": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "head_num": 40, "head_size": 128, "rotary_dim": 64,
                             "interleaved": True, "discrete": False, "dynamic_ntk": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "head_num": 52, "head_size": 128, "rotary_dim": 128,
                             "interleaved": False, "discrete": True, "dynamic_ntk": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "head_num": 64, "head_size": 128, "rotary_dim": 128,
                             "interleaved": False, "discrete": True, "dynamic_ntk": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "head_num": 25, "head_size": 64, "rotary_dim": 64,
                             "interleaved": False, "discrete": True, "dynamic_ntk": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "head_num": 64, "head_size": 96, "rotary_dim": 96,
                             "interleaved": False, "discrete": True, "dynamic_ntk": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 4, "seq_len": 1, "head_num": 80, "head_size": 128, "rotary_dim": 128,
                             "interleaved": False, "discrete": True, "dynamic_ntk": False, "input_dtype": [torch.float16, 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 = ["batch", "seq_len", "head_num", "head_size", "rotary_dim", "interleaved", "discrete", "dynamic_ntk", "input_dtype", "hardware_time(us)", "e2e_latency(us)"]
  contents = []
  for params_dict in e2e_time_param_dict_list:
    batch = params_dict["batch"]
    seq_len = params_dict["seq_len"]
    head_num = params_dict["head_num"]
    head_size = params_dict["head_size"]
    # full/partial
    rotary_dim = params_dict["rotary_dim"]
    # cross/fold
    interleaved = params_dict["interleaved"]
    # discrete
    discrete = params_dict["discrete"]
    # dynamic_ntk
    dynamic_ntk = params_dict["dynamic_ntk"]
    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():
        continue
      input = torch.randn(batch, seq_len, head_num, head_size).to(device).to(dtype) # [batch, seqlen, head_num, head_size]
      if dynamic_ntk:
        sin_cache = torch.randn(batch, seq_len, rotary_dim).to(device).to(dtype)
        cos_cache = torch.randn(batch, seq_len, rotary_dim).to(device).to(dtype)
      else:
        sin_cache = torch.randn(seq_len, rotary_dim).to(device).to(dtype)
        cos_cache = torch.randn(seq_len, rotary_dim).to(device).to(dtype)
      if discrete:
        pos_ids = torch.randint(0, seq_len, (batch * seq_len,)).to(device).to(torch.int32)
      else:
        pos_ids = None
      hardware_time, e2e_time = benchmark_forward(tmo.apply_rotary,
                                                  input,
                                                  sin_cache,
                                                  cos_cache,
                                                  pos_ids,
                                                  None,
                                                  interleaved,
                                                  discrete,
                                                  dynamic_ntk,
                                                  seq_len,
                                                  repeats=args.repeat_times)
      content = [f"{batch}", f"{seq_len}", f"{head_num}", f"{head_size}", f"{rotary_dim}", f"{interleaved}", f"{discrete}", f"{dynamic_ntk}", f"{dtype}", f"{hardware_time}", f"{e2e_time}"]
      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()