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

1import logging 

2 

3import torch 

4import triton 

5import triton.language as tl 

6 

7from flag_gems.runtime import torch_device_fn 

8from flag_gems.utils import libentry, libtuner 

9 

10from ..utils import TOTAL_CORE_NUM 

11 

12logger = logging.getLogger("flag_gems").getChild(__name__.lstrip(".")) 

13 

14 

15@libentry() 

16@libtuner( 

17 configs=[ 

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

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

20 triton.Config(kwargs={"BLOCK_SIZE": 65536}, num_stages=3, num_warps=1), 

21 triton.Config(kwargs={"BLOCK_SIZE": 131072}, num_stages=3, num_warps=1), 

22 ], 

23 key=["n_elements"], 

24) 

25@triton.jit 

26def logical_or_kernel( 

27 X_ptr, 

28 Y_ptr, 

29 OUT_ptr, 

30 n_elements, 

31 BLOCK_SIZE: tl.constexpr, 

32): 

33 pid = tl.program_id(0) 

34 num_jobs = tl.num_programs(0) 

35 block_start = pid * BLOCK_SIZE 

36 step = num_jobs * BLOCK_SIZE 

37 block_start = block_start.to(tl.int64) 

38 for off in range(block_start, n_elements, step): 

39 offsets = off + tl.arange(0, BLOCK_SIZE) 

40 mask = offsets < n_elements 

41 x = tl.load(X_ptr + offsets, mask=mask) 

42 y = tl.load(Y_ptr + offsets, mask=mask) 

43 result = (x != 0) | (y != 0) 

44 tl.store(OUT_ptr + offsets, result, mask=mask) 

45 

46 

47def logical_or(A, B): 

48 logger.debug("GEMS_CAMBRICON LOGICAL_OR") 

49 A = A.contiguous() 

50 B = B.contiguous() 

51 out = torch.empty(A.shape, dtype=torch.bool, device=A.device) 

52 N = A.numel() 

53 if N == 0: 

54 return out 

55 grid_fn = lambda meta: (min(triton.cdiv(N, meta["BLOCK_SIZE"]), TOTAL_CORE_NUM),) 

56 with torch_device_fn.device(A.device): 

57 logical_or_kernel[grid_fn](A, B, out, N) 

58 return out 

59 

60 

61def logical_or_(A, B): 

62 logger.debug("GEMS_CAMBRICON LOGICAL_OR_") 

63 A_contig = A.contiguous() 

64 B = B.contiguous() 

65 N = A_contig.numel() 

66 if N == 0: 

67 return A 

68 grid_fn = lambda meta: (min(triton.cdiv(N, meta["BLOCK_SIZE"]), TOTAL_CORE_NUM),) 

69 with torch_device_fn.device(A.device): 

70 logical_or_kernel[grid_fn](A_contig, B, A_contig, N) 

71 if not A.is_contiguous(): 

72 A.copy_(A_contig) 

73 return A