Coverage for src/flag_gems/ops/ones.py: 81%

1import logging 

2 

3import torch 

4import triton 

5import triton.language as tl 

6 

7from flag_gems.runtime import device, torch_device_fn 

8from flag_gems.utils import libentry 

9from flag_gems.utils import triton_lang_extension as tle 

10from flag_gems.utils.shape_utils import volume 

11 

12device_ = device 

13logger = logging.getLogger(__name__) 

14 

15 

16@libentry() 

17@triton.jit 

18def ones_kernel( 

19 output_ptr, 

20 n_elements, 

21 BLOCK_SIZE: tl.constexpr, 

22): 

23 pid = tle.program_id(axis=0) 

24 block_start = pid * BLOCK_SIZE 

25 offsets = block_start + tl.arange(0, BLOCK_SIZE) 

26 mask = offsets < n_elements 

27 tl.store(output_ptr + offsets, 1.0, mask=mask) 

28 

29 

30def ones(size, *, dtype=None, layout=None, device=None, pin_memory=None): 

31 logger.debug("GEMS ONES") 

32 if dtype is None: 

33 dtype = torch.get_default_dtype() 

34 if device is None: 

35 device = torch.device(device_.name) 

36 

37 out = torch.empty(size, device=device, dtype=dtype) 

38 N = volume(size) 

39 BLOCK_SIZE = 1024 

40 grid = (triton.cdiv(N, BLOCK_SIZE),) 

41 with torch_device_fn.device(device): 

42 ones_kernel[grid](out, N, BLOCK_SIZE) 

43 return out