Coverage for src/flag_gems/ops/rand.py: 54%

1import logging 

2 

3import torch 

4import triton 

5import triton.language as tl 

6 

7from flag_gems import runtime 

8from flag_gems.runtime import device, torch_device_fn 

9from flag_gems.utils.random_utils import ( 

10 philox_backend_seed_offset, 

11 uint_to_uniform_float, 

12) 

13from flag_gems.utils.shape_utils import volume 

14 

15logger = logging.getLogger(__name__) 

16device_ = device 

17 

18 

19@triton.heuristics(runtime.get_heuristic_config("rand")) 

20@triton.jit(do_not_specialize=["philox_seed", "philox_offset"]) 

21def rand_kernel( 

22 out_ptr, 

23 N, 

24 philox_seed, 

25 philox_offset, 

26 BLOCK: tl.constexpr, 

27): 

28 philox_seed = philox_seed.to(tl.int64) 

29 philox_offset = philox_offset.to(tl.int64) 

30 c0 = (philox_offset & 0xFFFFFFFF).to(tl.uint32) 

31 c1 = ((philox_offset >> 32) & 0xFFFFFFFF).to(tl.uint32) 

32 i4 = tl.program_id(0) * BLOCK + tl.arange(0, BLOCK) 

33 c0 += i4 

34 _O = c0 * 0 

35 r0, r1, r2, r3 = tl.philox(philox_seed, c0, c1, _O, _O) 

36 r0 = uint_to_uniform_float(r0) 

37 r1 = uint_to_uniform_float(r1) 

38 r2 = uint_to_uniform_float(r2) 

39 r3 = uint_to_uniform_float(r3) 

40 off_0 = tl.program_id(0) * BLOCK * 4 + tl.arange(0, BLOCK) 

41 off_1 = off_0 + BLOCK 

42 off_2 = off_1 + BLOCK 

43 off_3 = off_2 + BLOCK 

44 tl.store(out_ptr + off_0, r0, mask=off_0 < N, eviction_policy="evict_first") 

45 tl.store(out_ptr + off_1, r1, mask=off_1 < N, eviction_policy="evict_first") 

46 tl.store(out_ptr + off_2, r2, mask=off_2 < N, eviction_policy="evict_first") 

47 tl.store(out_ptr + off_3, r3, mask=off_3 < N, eviction_policy="evict_first") 

48 

49 

50UNROLL = 4 

51 

52 

53def rand(size, *, dtype=None, layout=None, device=None, pin_memory=None): 

54 logger.debug("GEMS RAND") 

55 if dtype is None: 

56 dtype = torch.get_default_dtype() 

57 if device is None: 

58 device = torch.device(device_.name) 

59 

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

61 N = volume(size) 

62 grid_fn = lambda meta: (triton.cdiv(N, meta["BLOCK"] * UNROLL),) 

63 # (TODO) Using Triton autotuner makes kernel parameters opaque to the caller, 

64 # hence we cannot obtain the per thread offset as in Pytorch. 

65 increment = triton.cdiv(N, UNROLL) 

66 philox_seed, philox_offset = philox_backend_seed_offset(increment) 

67 with torch_device_fn.device(device): 

68 rand_kernel[grid_fn](out, N, philox_seed, philox_offset) 

69 return out