Coverage for src/flag_gems/runtime/backend/_cambricon/ops/neg.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=1, num_warps=1), 

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

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

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

22 ], 

23 key=["n_elements"], 

24) 

25@triton.jit 

26def neg_func( 

27 X_ptr, 

28 OUT_ptr, 

29 n_elements, 

30 BLOCK_SIZE: tl.constexpr, 

31): 

32 pid = tl.program_id(0) 

33 num_jobs = tl.num_programs(0) 

34 block_start = pid * BLOCK_SIZE 

35 step = num_jobs * BLOCK_SIZE 

36 block_start = block_start.to(tl.int64) 

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

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

39 mask = offsets < n_elements 

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

41 tl.store(OUT_ptr + offsets, -x, mask=mask) 

42 

43 

44def neg(A): 

45 logger.debug("GEMS_CAMBRICON NEG") 

46 A = A.contiguous() 

47 out = torch.empty_like(A) 

48 N = A.numel() 

49 if N == 0: 

50 return out 

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

52 with torch_device_fn.device(A.device): 

53 neg_func[grid_fn](A, out, N) 

54 return out 

55 

56 

57def neg_(A): 

58 logger.debug("GEMS_CAMBRICON NEG_") 

59 A_contig = A.contiguous() 

60 N = A_contig.numel() 

61 if N == 0: 

62 return A 

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

64 with torch_device_fn.device(A.device): 

65 neg_func[grid_fn](A_contig, A_contig, N) 

66 if not A.is_contiguous(): 

67 A.copy_(A_contig) 

68 return A