Coverage for src/flag_gems/runtime/backend/_mthreads/ops/randn.py: 0%

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 

15try: 

16 pair_uniform_to_normal = tl.pair_uniform_to_normal 

17except AttributeError: 

18 

19 @triton.jit 

20 def pair_uniform_to_normal(u1, u2): 

21 """Box-Muller transform""" 

22 u1 = tl.maximum(1.0e-7, u1) 

23 th = 6.283185307179586 * u2 

24 r = tl.sqrt(-2.0 * tl.log(u1)) 

25 return r * tl.cos(th), r * tl.sin(th) 

26 

27 

28device_ = device 

29logger = logging.getLogger( 

30 f'flag_gems.runtime.backend._mthreads.ops.{__name__.split(".")[-1]}' 

31) 

32 

33 

34@triton.heuristics(runtime.get_heuristic_config("randn")) 

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

36def randn_kernel( 

37 out_ptr, 

38 N, 

39 philox_seed, 

40 philox_offset, 

41 BLOCK: tl.constexpr, 

42): 

43 philox_seed = philox_seed.to(tl.int64) 

44 philox_offset = philox_offset.to(tl.int64) 

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

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

47 i4 = (tl.program_id(0) * BLOCK + tl.arange(0, BLOCK)).to(tl.uint32) 

48 c0 += i4 

49 _O = c0 * 0 

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

51 r0 = uint_to_uniform_float(r0) 

52 r1 = uint_to_uniform_float(r1) 

53 r2 = uint_to_uniform_float(r2) 

54 r3 = uint_to_uniform_float(r3) 

55 n0, n1 = pair_uniform_to_normal(r0, r1) 

56 n2, n3 = pair_uniform_to_normal(r2, r3) 

57 off_0 = ((tl.program_id(0) * BLOCK * 4).to(tl.int64) + tl.arange(0, BLOCK)).to( 

58 tl.int64 

59 ) 

60 off_1 = off_0 + BLOCK 

61 off_2 = off_1 + BLOCK 

62 off_3 = off_2 + BLOCK 

63 tl.store(out_ptr + off_0, n0, mask=off_0 < N, eviction_policy="evict_first") 

64 tl.store(out_ptr + off_1, n1, mask=off_1 < N, eviction_policy="evict_first") 

65 tl.store(out_ptr + off_2, n2, mask=off_2 < N, eviction_policy="evict_first") 

66 tl.store(out_ptr + off_3, n3, mask=off_3 < N, eviction_policy="evict_first") 

67 

68 

69UNROLL = 4 

70 

71 

72def randn(size, *, dtype=None, layout=None, device=None, pin_memory=None): 

73 logger.debug("GEMS_MTHREADS RANDN") 

74 if dtype is None: 

75 dtype = torch.get_default_dtype() 

76 if device is None: 

77 device = torch.device(device_.name) 

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

79 N = volume(size) 

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

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

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

83 increment = triton.cdiv(N, UNROLL) 

84 philox_seed, philox_offset = philox_backend_seed_offset(increment) 

85 with torch_device_fn.device(device): 

86 randn_kernel[grid_fn](out, N, philox_seed, philox_offset) 

87 return out