Coverage for src/flag_gems/runtime/backend/_cambricon/ops/relu.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 

11from ..utils.pointwise_dynamic import pointwise_dynamic 

12 

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

14 

15 

16@libentry() 

17@libtuner( 

18 configs=[ 

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

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

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

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

23 ], 

24 key=["n_elements"], 

25) 

26@triton.jit 

27def relu_kernel( 

28 X_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 tl.store(OUT_ptr + offsets, tl.where(x > 0, x, 0), mask=mask) 

43 

44 

45# keep backward using pointwise_dynamic 

46@pointwise_dynamic(promotion_methods=[(0, "DEFAULT")]) 

47@triton.jit 

48def relu_backward(x, dy): 

49 return tl.where(x > 0, dy, 0) 

50 

51 

52def relu(self): 

53 logger.debug("GEMS_CAMBRICON RELU FORWARD") 

54 A = self.contiguous() 

55 out = torch.empty_like(A) 

56 N = A.numel() 

57 if N == 0: 

58 return out 

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

60 with torch_device_fn.device(A.device): 

61 relu_kernel[grid_fn](A, out, N) 

62 return out 

63 

64 

65def relu_(A): 

66 logger.debug("GEMS_CAMBRICON RELU_ FORWARD") 

67 A_contig = A.contiguous() 

68 N = A_contig.numel() 

69 if N == 0: 

70 return A 

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

72 with torch_device_fn.device(A.device): 

73 relu_kernel[grid_fn](A_contig, A_contig, N) 

74 if not A.is_contiguous(): 

75 A.copy_(A_contig) 

76 return A