Coverage for src/flag_gems/runtime/backend/_cambricon/ops/arange.py: 0%

1import logging 

2import math 

3 

4import torch 

5import triton 

6import triton.language as tl 

7 

8from flag_gems import runtime 

9from flag_gems.utils import libentry, libtuner 

10 

11from ..utils import TOTAL_CORE_NUM 

12 

13logger = logging.getLogger("flag_gems").getChild(__name__.lstrip(".")) 

14 

15 

16@libentry() 

17@libtuner( 

18 configs=[ 

19 triton.Config(kwargs={"BLOCK_SIZE": 1024}, num_stages=3, num_warps=1), 

20 triton.Config(kwargs={"BLOCK_SIZE": 4096}, num_stages=3, num_warps=1), 

21 triton.Config(kwargs={"BLOCK_SIZE": 8192}, num_stages=3, num_warps=1), 

22 triton.Config(kwargs={"BLOCK_SIZE": 16384}, num_stages=3, num_warps=1), 

23 ], 

24 key=["size"], 

25 strategy=["log"], 

26) 

27@triton.jit 

28def arange_func(y_ptr, start, end, step, size, BLOCK_SIZE: tl.constexpr): 

29 pid = tl.program_id(axis=0) 

30 num_jobs = tl.num_programs(axis=0) 

31 block_start = pid * BLOCK_SIZE 

32 block_step = num_jobs * BLOCK_SIZE 

33 block_start = block_start 

34 for block_start_offset in range(block_start, size, block_step): 

35 offset = tl.arange(0, BLOCK_SIZE) + block_start_offset 

36 arange_val = offset * step + start 

37 tl.store(y_ptr + offset, arange_val, mask=offset < size) 

38 

39 

40def arange_start( 

41 start, end, step=1, *, dtype=None, layout=None, device=None, pin_memory=None 

42): 

43 logger.debug("GEMS_CAMBRICON ARANGE") 

44 if dtype is torch.int64: 

45 sgn = (step > 0) - (step < 0) 

46 size = (end - start + step - sgn) // step 

47 else: 

48 size = math.ceil((end - start) / step) 

49 

50 assert ( 

51 size < torch.iinfo(torch.int32).max 

52 ), f"Size {size} is not less than the maximum int32 value max_int32" 

53 

54 grid = lambda META: (min(triton.cdiv(size, META["BLOCK_SIZE"]), TOTAL_CORE_NUM),) 

55 

56 if dtype is None: 

57 dtype = torch.int64 

58 

59 if pin_memory is None: 

60 pin_memory = False 

61 

62 if device is None: 

63 device = ( 

64 runtime.device.name 

65 ) # Note(Zhengzekang): Torch default value is CPU, but triton is target to GPU. 

66 

67 result = torch.empty((size,), device=device, dtype=dtype, pin_memory=pin_memory) 

68 arange_func[grid](result, start, end, step, size) 

69 return result 

70 

71 

72def arange(end, *, dtype=None, layout=None, device=None, pin_memory=None): 

73 return arange_start( 

74 0, end, 1, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory 

75 )