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, "hidden_size": 1600, "inner_size": 6400,
                             "gated_ffn": False, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 2048, "inner_size": 8192,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 4096, "inner_size": 11008,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 4096, "inner_size": 14336,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 4096, "inner_size": 16384,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 5120, "inner_size": 13824,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 5120, "inner_size": 27392,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 6656, "inner_size": 17920,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 8192, "inner_size": 22016,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 8192, "inner_size": 24576,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 8192, "inner_size": 28672,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 8192, "inner_size": 49152,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 12288, "inner_size": 32768,
                             "gated_ffn": True, "act_mode": "gelu", "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_len": 1024, "hidden_size": 14336, "inner_size": 57344,
                             "gated_ffn": True, "act_mode": "gelu", "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", "hidden_size", "inner_size", "gated_ffn", "act_mode", "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"]
    hidden_size = params_dict["hidden_size"]
    inner_size = params_dict["inner_size"]
    gated_ffn = params_dict["gated_ffn"]
    act_mode = params_dict["act_mode"]
    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, hidden_size).to(device).to(dtype)
      up_proj_weight = torch.randn(inner_size, hidden_size).to(device).to(dtype)
      up_proj_bias = torch.randn(inner_size).to(device).to(dtype)
      down_proj_weight = torch.randn(hidden_size, inner_size).to(device).to(dtype)
      down_proj_bias = torch.randn(hidden_size).to(device).to(dtype)
      gate_up_proj_weight, gate_up_proj_bias = None, None
      if gated_ffn:
        gate_up_proj_weight = torch.randn(inner_size, hidden_size).to(device).to(dtype)
        gate_up_proj_bias = torch.randn(inner_size).to(device).to(dtype)
      hardware_time, e2e_time = benchmark_forward(tmo.ffn,
                                                  input,
                                                  up_proj_weight,
                                                  up_proj_bias,
                                                  down_proj_weight,
                                                  down_proj_bias,
                                                  gate_up_proj_weight,
                                                  gate_up_proj_bias,
                                                  act_mode,
                                                  repeats=args.repeat_times)
      content = [f"{batch}", f"{seq_len}", f"{hidden_size}", f"{inner_size}", f"{gated_ffn}", f"{act_mode}", 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()