Coverage for src/flag_gems/ops/var.py: 50%

1import logging 

2 

3import torch 

4import triton 

5import triton.language as tl 

6 

7from flag_gems import runtime 

8from flag_gems.runtime import torch_device_fn 

9from flag_gems.utils import dim_compress, libentry 

10from flag_gems.utils import triton_lang_extension as tle 

11 

12logger = logging.getLogger(__name__) 

13 

14 

15@triton.jit 

16def welford_func(mean_x, count_x, M_x, mean_y, count_y, M_y): 

17 count = count_x + count_y 

18 _count = tl.maximum(count, 1) 

19 mc_x = mean_x * count_x 

20 mc_y = mean_y * count_y 

21 mean = (mc_x + mc_y) / _count 

22 M = M_x + mc_x * mean_x + M_y + mc_y * mean_y - count * mean * mean 

23 return mean, count, M 

24 

25 

26@libentry() 

27@triton.autotune(configs=runtime.get_tuned_config("var_mean"), key=["M", "N"]) 

28@triton.jit(do_not_specialize=["correction"]) 

29def var_welford_kernel( 

30 X, 

31 Var, 

32 M, 

33 N, 

34 correction, 

35 BLOCK_M: tl.constexpr, 

36 BLOCK_N: tl.constexpr, 

37): 

38 # Map the program id to the row of X it should compute. 

39 pid = tle.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)[:, None] 

40 X = X + pid * N 

41 Var = Var + pid 

42 row_mask = pid < M 

43 

44 _mean = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) 

45 _acc = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) 

46 _count = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) 

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 x = tl.load(X + cols, mask, other=0.0).to(tl.float32) 

53 

54 count = _count + mask 

55 cnt = tl.maximum(count, 1) 

56 cur_mean = (_mean * _count + x) / cnt 

57 _acc += (x - cur_mean) * (x - _mean) * mask 

58 _mean = cur_mean 

59 _count = count 

60 

61 mean, _, acc = tl.reduce((_mean, _count, _acc), axis=1, combine_fn=welford_func) 

62 var = acc / (N - correction) 

63 var = var[:, None] 

64 # Write var 

65 tl.store(Var, var, row_mask) 

66 

67 

68@libentry() 

69@triton.jit 

70def var_kernel_1( 

71 X, 

72 Acc, 

73 Average, 

74 Count, 

75 N, 

76 BLOCK_N: tl.constexpr, 

77): 

78 # Map the program id to the row of X it should compute. 

79 pid = tle.program_id(0) 

80 offset = pid * BLOCK_N + tl.arange(0, BLOCK_N) 

81 

82 X = X + offset 

83 Acc = Acc + pid 

84 Average = Average + pid 

85 Count = Count + pid 

86 mask = offset < N 

87 

88 x = tl.load(X, mask, other=0.0).to(tl.float32) 

89 

90 count = tl.sum(mask.to(tl.float32)) 

91 average = tl.sum(x) / count 

92 acc = tl.sum(x * x) - count * average * average 

93 

94 tl.store(Average, average) 

95 tl.store(Acc, acc) 

96 tl.store(Count, count) 

97 

98 

99@libentry() 

100@triton.heuristics(runtime.get_heuristic_config("var_mean")) 

101@triton.jit(do_not_specialize=["correction"]) 

102def var_kernel_2( 

103 Acc, 

104 Average, 

105 Count, 

106 Var, 

107 N, 

108 correction, 

109 BLOCK_NUM, 

110 BLOCK_N: tl.constexpr, 

111): 

112 offset = tl.arange(0, BLOCK_N) 

113 mask = offset < BLOCK_NUM 

114 Acc = Acc + offset 

115 Average = Average + offset 

116 Count = Count + offset 

117 acc = tl.load(Acc, mask, other=0.0).to(tl.float32) 

118 average = tl.load(Average, mask, other=0.0).to(tl.float32) 

119 count = tl.load(Count, mask, other=0.0).to(tl.float32) 

120 

121 mean, _, nvar = tl.reduce((average, count, acc), axis=0, combine_fn=welford_func) 

122 

123 var = nvar / (N - correction) 

124 tl.store(Var, var) 

125 

126 

127def var(x, dim=None, *, correction=None, keepdim=False): 

128 logger.debug("GEMS VAR") 

129 if correction is None: 

130 correction = 1.0 

131 

132 if dim is None or len(dim) == x.ndim: 

133 dim = list(range(x.ndim)) 

134 shape = [1] * x.ndim 

135 N = x.numel() 

136 var = torch.empty(shape, dtype=x.dtype, device=x.device) 

137 BLOCK_N = 1024 

138 BLOCK_NUM = triton.cdiv(N, BLOCK_N) 

139 acc = torch.empty([BLOCK_NUM], dtype=x.dtype, device=x.device) 

140 average = torch.empty([BLOCK_NUM], dtype=x.dtype, device=x.device) 

141 count = torch.empty([BLOCK_NUM], dtype=x.dtype, device=x.device) 

142 

143 with torch_device_fn.device(x.device): 

144 var_kernel_1[(BLOCK_NUM,)](x, acc, average, count, N, BLOCK_N=BLOCK_N) 

145 var_kernel_2[(1,)](acc, average, count, var, N, correction, BLOCK_NUM) 

146 else: 

147 shape = list(x.shape) 

148 dim = [d % x.ndim for d in dim] 

149 x = dim_compress(x, dim) 

150 N = 1 

151 for i in dim: 

152 N *= shape[i] 

153 shape[i] = 1 

154 M = x.numel() // N 

155 var = torch.empty(shape, dtype=x.dtype, device=x.device) 

156 

157 grid = lambda META: (triton.cdiv(M, META["BLOCK_M"]),) 

158 with torch_device_fn.device(x.device): 

159 var_welford_kernel[grid](x, var, M, N, correction) 

160 

161 if not keepdim: 

162 var = var.squeeze(dim=dim) 

163 return var 

164 

165 

166def var_dim(x, dim=None, *, correction=None, keepdim=False): 

167 logger.debug("GEMS VAR_DIM") 

168 return var(x, dim=dim, correction=correction, keepdim=keepdim) 

169 

170 

171def var_correction(x, dim=None, *, correction=None, keepdim=False): 

172 logger.debug("GEMS VAR_CORRECTION") 

173 return var(x, dim=dim, correction=correction, keepdim=keepdim)