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
import random

e2e_time_param_dict_list = [{"max_batch": 20, "batch": 10, "cache_mem_len": 1024, "max_context_len": 512,
                             "max_seq_offset": 20, "head_num_q": 32, "head_num_kv": 32, "head_size": 128,
                             "packed": True, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"max_batch": 20, "batch": 10, "cache_mem_len": 1024, "max_context_len": 512,
                             "max_seq_offset": 20, "head_num_q": 32, "head_num_kv": 32, "head_size": 128,
                             "packed": False, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"max_batch": 20, "batch": 10, "cache_mem_len": 1024, "max_context_len": 512,
                             "max_seq_offset": 20, "head_num_q": 32, "head_num_kv": 32, "head_size": 128,
                             "packed": True, "input_dtype": [torch.float16, torch.bfloat16]},
                            {"max_batch": 20, "batch": 10, "cache_mem_len": 1024, "max_context_len": 512,
                             "max_seq_offset": 20, "head_num_q": 32, "head_num_kv": 32, "head_size": 128,
                             "packed": 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()
  titles = ["batch", "max_context_len", "head_num_q", "head_num_kv", "head_size", "packed", "input_dytpe", "hardware_time(us)", "e2e_latency(us)"]
  contents = []
  for params_dict in e2e_time_param_dict_list:
    max_batch = params_dict["max_batch"]
    batch = params_dict["batch"]
    cache_mem_len = params_dict["cache_mem_len"]
    max_context_len = params_dict["max_context_len"]
    max_seq_offset = params_dict["max_seq_offset"]
    head_num_q = params_dict["head_num_q"]
    head_num_kv = params_dict["head_num_kv"]
    head_size = params_dict["head_size"]
    packed = params_dict["packed"]
    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
      context_lens = torch.randint(size=(batch, ), low=max_context_len,
                                        high=max_context_len+1,
                                        dtype=torch.int32, device='mlu')
      # max_seq_offset = max_context_len // 3 + 1
      context_seq_offsets = torch.randint(size=(batch, ), low=max_seq_offset, high=max_seq_offset+1,
                                          dtype=torch.int32, device='mlu')
      cache_seq_offsets = torch.randint(size=(batch, ), low=-1,
                                        high=(cache_mem_len - max_context_len) // 3 + 1,
                                        dtype=torch.int32, device='mlu')

      cu_context_lens = torch.cumsum(context_lens, dim=-1)
      cu_context_lens = torch.nn.functional.pad(cu_context_lens, (1,0), "constant", 0).to(torch.int32)
      total_seqlen = cu_context_lens[-1]
      total_heads = head_num_q + 2 * head_num_kv
      if packed > 0:
          context = torch.randn((total_seqlen, total_heads, head_size),
                                dtype=torch.float, device='mlu')
      else:
          context = torch.randn((batch, max_context_len + max_seq_offset, total_heads, head_size),
                                dtype=torch.float, device='mlu')
      cache = torch.randn((2, max_batch, head_num_kv, cache_mem_len, head_size), dtype=torch.float, device='mlu')
      context = context.to(dtype)
      cache = cache.to(dtype)
      key = context[..., head_num_q : head_num_q + head_num_kv, :]
      value = context[..., head_num_q + head_num_kv : head_num_q + 2 * head_num_kv, :]
      key_cache = cache[0]
      value_cache = cache[1]

      cache_bs_id = None
      cache_bs_id = random.sample([*range(0, max_batch)], batch)
      cache_bs_id = torch.IntTensor(cache_bs_id).mlu()

      if packed > 0:
        hardware_time, e2e_time = benchmark_forward(tmo.reshape_linear_cache,
                                                    key, value,
                                                    key_cache, value_cache,
                                                    cu_context_lens, max_context_len,
                                                    packed > 0, None,
                                                    cache_bs_id, cache_seq_offsets,
                                                    repeats=args.repeat_times)

      else:
        hardware_time, e2e_time = benchmark_forward(tmo.reshape_linear_cache,
                                                    key, value,
                                                    key_cache, value_cache,
                                                    context_lens, max_context_len,
                                                    packed > 0, context_seq_offsets,
                                                    cache_bs_id, cache_seq_offsets,
                                                    repeats=args.repeat_times)

      content = [f"{batch}", f"{max_context_len}", f"{head_num_q}", f"{head_num_kv}", f"{head_size}", f"{packed}", 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()