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

# for e2e time test
e2e_time_param_dict_list = [{"batch": 1, "seq_q": 32768, "seq_kv": 32768, "head_num": 8,
                             "head_num_kv": 1, "head_size": 128, "use_causal": True,
                             "softmax_scale": 1e-6, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_q": 16384, "seq_kv": 16384, "head_num": 8,
                             "head_num_kv": 1, "head_size": 128, "use_causal": True,
                             "softmax_scale": 1e-6, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_q": 8192,  "seq_kv": 24576, "head_num": 8,
                             "head_num_kv": 1, "head_size": 128, "use_causal": True,
                             "softmax_scale": 1e-6, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_q": 4096,  "seq_kv": 28672, "head_num": 8,
                             "head_num_kv": 1, "head_size": 128, "use_causal": True,
                             "softmax_scale": 1e-6, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"batch": 1, "seq_q": 4096,  "seq_kv": 32768, "head_num": 8,
                             "head_num_kv": 1, "head_size": 128, "use_causal": True,
                             "softmax_scale": 1e-6, "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_q", "seq_kv", "head_num", "head_num_kv", "head_size", "use_causal", "input_dtype", "hardware_time(us)", "e2e_latency(us)"]
  contents = []
  for params_dict in e2e_time_param_dict_list:
    batch = params_dict["batch"]
    seq_q = params_dict["seq_q"]
    seq_kv = params_dict["seq_kv"]
    head_num = params_dict["head_num"]
    head_num_kv = params_dict["head_num_kv"]
    head_size = params_dict["head_size"]
    use_causal = params_dict["use_causal"]
    softmax_scale = params_dict["softmax_scale"]
    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
      if seq_q == seq_kv:
        qkv = torch.randn(batch, seq_q, head_num + 2 * head_num_kv, head_size).to(dtype).to(device)
        q = qkv[:, :, : head_num, :]
        k = qkv[:, :, head_num : head_num + head_num_kv, :]
        v = qkv[:, :, head_num + head_num_kv : head_num + head_num * 2, :]
      elif seq_q < seq_kv:
        q = torch.randn(batch, seq_q, head_num, head_size).to(device).to(dtype)
        kv = torch.randn(batch, seq_kv, head_num_kv * 2, head_size).to(device).to(dtype)
        k = kv[:, :, : head_num_kv, :]
        v = kv[:, :, head_num_kv :, :]
      hardware_time, e2e_time = benchmark_forward(tmo.flash_attention,
                                                  q = q,
                                                  k = k,
                                                  v = v,
                                                  out = None,
                                                  cu_seq_lens_q = None,
                                                  cu_seq_lens_kv = None,
                                                  alibi_slope = None,
                                                  attn_bias = None,
                                                  max_seq_len_q = seq_q,
                                                  max_seq_len_kv = seq_kv,
                                                  softmax_scale = softmax_scale,
                                                  is_causal = use_causal,
                                                  window_size_left = -1,
                                                  window_size_right = -1,
                                                  compute_dtype = dtype,
                                                  return_lse = False,
                                                  block_tables = None,
                                                  k_cache_quant_scale = None,
                                                  v_cache_quant_scale = None,
                                                  repeats=args.repeat_times)
      content = [f"{batch}", f"{seq_q}", f"{seq_kv}", f"{head_num}", f"{head_num_kv}", f"{head_size}", f"{use_causal}", 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()