Coverage for src/flag_gems/ops/all.py: 68%

1import logging 

2import math 

3 

4import torch 

5import triton 

6import triton.language as tl 

7 

8from flag_gems import runtime 

9from flag_gems.runtime import torch_device_fn 

10from flag_gems.utils import dim_compress, libentry, libtuner 

11from flag_gems.utils import triton_lang_extension as tle 

12 

13logger = logging.getLogger(__name__) 

14 

15# torch.all: Tests if all elements in input evaluate to True. If the dtype of input 

16# is not BOOL, then test if all elements in input evaluate to non-zero value 

17# In triton function, test if all elements in input evaluate to non-zero value is ok. 

18 

19 

20@triton.jit 

21def reduce_all(a, b): 

22 return a and b 

23 

24 

25@libentry() 

26@libtuner( 

27 configs=runtime.get_tuned_config("naive_reduction"), 

28 key=["M", "N"], 

29) 

30@triton.jit 

31def all_kernel_dim( 

32 inp, 

33 out, 

34 M, 

35 N, 

36 BLOCK_M: tl.constexpr, 

37 BLOCK_N: tl.constexpr, 

38): 

39 # Map the program id to the row of inp it should compute. 

40 pid = tle.program_id(0) 

41 rows = pid * BLOCK_M + tl.arange(0, BLOCK_M)[:, None] 

42 inp = inp + rows * N 

43 out = out + rows 

44 row_mask = rows < M 

45 

46 _all = tl.full([BLOCK_M, BLOCK_N], value=1, dtype=tl.int1) 

47 for off in range(0, N, BLOCK_N): 

48 cols = off + tl.arange(0, BLOCK_N)[None, :] 

49 col_mask = cols < N 

50 mask = row_mask and col_mask 

51 

52 a = tl.load(inp + cols, mask, other=1.0) 

53 _all = _all and (a != 0) 

54 all = tl.reduce(_all, axis=1, combine_fn=reduce_all) 

55 tl.store(out, all[:, None], row_mask) 

56 

57 

58@libentry() 

59@triton.jit 

60def all_kernel_1( 

61 inp, 

62 mid, 

63 n_elements, 

64 mid_size, 

65 BLOCK_SIZE: tl.constexpr, 

66): 

67 pid = tle.program_id(0) 

68 offset = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) 

69 inp_ptrs = inp + offset 

70 mask = offset < n_elements 

71 inp_val = tl.load(inp_ptrs, mask=mask, other=1.0) 

72 all_val = tl.reduce(inp_val != 0, axis=0, combine_fn=reduce_all) 

73 mid_ptr = mid + pid 

74 tl.store(mid_ptr, all_val) 

75 

76 

77@libentry() 

78@triton.jit 

79def all_kernel_2(mid, out, MID_SIZE, BLOCK_MID: tl.constexpr): 

80 offset = tl.arange(0, BLOCK_MID) 

81 mid_ptrs = mid + offset 

82 mask = offset < MID_SIZE 

83 mid_val = tl.load(mid_ptrs, mask=mask, other=1).to(tl.int1) 

84 all_val = tl.reduce(mid_val, axis=0, combine_fn=reduce_all) 

85 tl.store(out, all_val) 

86 

87 

88def all(inp): 

89 logger.debug("GEMS ALL") 

90 n_elements = inp.numel() 

91 block_size = triton.next_power_of_2(math.ceil(math.sqrt(n_elements))) 

92 mid_size = triton.cdiv(n_elements, block_size) 

93 block_mid = triton.next_power_of_2(mid_size) 

94 

95 mid = torch.empty((mid_size,), dtype=torch.bool, device=inp.device) 

96 out = torch.empty([], dtype=torch.bool, device=inp.device) 

97 

98 with torch_device_fn.device(inp.device): 

99 all_kernel_1[(mid_size, 1)](inp, mid, n_elements, mid_size, block_size) 

100 all_kernel_2[(1, 1)](mid, out, mid_size, block_mid) 

101 

102 return out 

103 

104 

105def all_dim(inp, dim=None, keepdim=False): 

106 logger.debug("GEMS ALL DIM") 

107 shape = list(inp.shape) 

108 if dim is None: 

109 out = all(inp) 

110 if keepdim: 

111 out = torch.reshape(out, [1] * inp.ndim) 

112 else: 

113 assert dim >= -inp.ndim and dim < inp.ndim, "Invalid dim" 

114 dim = dim % inp.ndim 

115 inp = dim_compress(inp, dim) 

116 N = shape[dim] 

117 shape[dim] = 1 

118 M = inp.numel() // N 

119 

120 out = torch.empty(shape, dtype=torch.bool, device=inp.device) 

121 

122 grid = lambda meta: (triton.cdiv(M, meta["BLOCK_M"]),) 

123 with torch_device_fn.device(inp.device): 

124 all_kernel_dim[grid](inp, out, M, N) 

125 if not keepdim: 

126 out = out.squeeze(dim=dim) 

127 return out 

128 

129 

130def all_dims(inp, dim=None, keepdim=False): 

131 logger.debug("GEMS ALL DIMS") 

132 

133 if dim is None or isinstance(dim, int): 

134 return all_dim(inp, dim=dim, keepdim=keepdim) 

135 assert ((i >= -inp.ndim and i < inp.ndim) for i in dim), "Invalid dim" 

136 

137 shape = list(inp.shape) 

138 dim = [d % inp.ndim for d in dim] 

139 inp = dim_compress(inp, dim) 

140 N = 1 

141 for i in dim: 

142 N *= shape[i] 

143 shape[i] = 1 

144 M = inp.numel() // N 

145 

146 out = torch.empty(shape, dtype=torch.bool, device=inp.device) 

147 

148 grid = lambda meta: (triton.cdiv(M, meta["BLOCK_M"]),) 

149 with torch_device_fn.device(inp.device): 

150 all_kernel_dim[grid](inp, out, M, N) 

151 if not keepdim: 

152 out = out.squeeze(dim=dim) 

153 return out