Coverage for src/flag_gems/runtime/backend/_ascend/ops/var_mean.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 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(f'flag_gems.runtime._ascend.ops.{__name__.split(".")[-1]}') 

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_mean_welford_kernel( 

30 X, 

31 Var, 

32 Mean, 

33 M, 

34 N, 

35 correction, 

36 BLOCK_M: tl.constexpr, 

37 BLOCK_N: tl.constexpr, 

38): 

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

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

41 X = X + pid * N 

42 Var = Var + pid 

43 Mean = Mean + pid 

44 row_mask = pid < M 

45 

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

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

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

49 

50 for off in range(0, N, BLOCK_N): 

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

52 col_mask = cols < N 

53 mask = row_mask and col_mask 

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

55 

56 count = _count + mask 

57 cnt = tl.maximum(count, 1) 

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

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

60 _mean = cur_mean 

61 _count = count 

62 

63 # 手动实现 tl.reduce 的功能，沿着 axis=1 进行归约 

64 # 使用 tl.sum 来进行归约，这等价于 welford 算法在这种情况下的行为 

65 

66 # 计算每行的总计数 

67 total_count = tl.sum(_count, axis=1) # shape: (BLOCK_M,) 

68 

69 # 计算加权平均值 

70 weighted_sum = tl.sum(_mean * _count, axis=1) # shape: (BLOCK_M,) 

71 mean = weighted_sum / tl.maximum(total_count, 1) # shape: (BLOCK_M,) 

72 

73 # 计算方差累积值 

74 # 对于每个元素，计算其对总体方差的贡献 

75 mean_expanded = mean[:, None] # shape: (BLOCK_M, 1) 

76 

77 # 计算每个局部统计量对总体方差的贡献 

78 # 这是 Welford 算法的并行化版本 

79 local_var_contrib = _acc + _count * (_mean - mean_expanded) * ( 

80 _mean - mean_expanded 

81 ) 

82 acc = tl.sum(local_var_contrib, axis=1) # shape: (BLOCK_M,) 

83 

84 var = acc / (N - correction) 

85 mean = mean[:, None] 

86 var = var[:, None] 

87 

88 # Write mean / var 

89 tl.store(Mean, mean, row_mask) 

90 tl.store(Var, var, row_mask) 

91 

92 

93@libentry() 

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

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

96def var_mean_welford_kernel_simple( 

97 X, 

98 Var, 

99 Mean, 

100 M, 

101 N, 

102 correction, 

103 BLOCK_M: tl.constexpr, 

104 BLOCK_N: tl.constexpr, 

105): 

106 # 程序ID映射 

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

108 X = X + pid * N 

109 Var = Var + pid 

110 Mean = Mean + pid 

111 row_mask = pid < M 

112 

113 # 每行单独处理 

114 for row in range(BLOCK_M): 

115 if row < BLOCK_M: 

116 current_row_mask = (tl.arange(0, BLOCK_M) == row)[:, None] & row_mask 

117 

118 if tl.sum(current_row_mask.to(tl.int32)) > 0: 

119 # 初始化当前行的统计量 

120 running_mean = 0.0 

121 running_M = 0.0 

122 count = 0 

123 

124 # 按块处理当前行 

125 for off in range(0, N, BLOCK_N): 

126 cols = off + tl.arange(0, BLOCK_N) 

127 col_mask = cols < N 

128 

129 # 加载数据 

130 x_vals = tl.load(X + row * N + cols, col_mask, other=0.0).to( 

131 tl.float32 

132 ) 

133 

134 # 对块内每个有效元素进行在线更新 

135 for i in range(BLOCK_N): 

136 if i < BLOCK_N and (off + i) < N: 

137 count += 1 

138 x = x_vals[i] 

139 

140 delta = x - running_mean 

141 running_mean += delta / count 

142 delta2 = x - running_mean 

143 running_M += delta * delta2 

144 

145 # 计算方差 

146 variance = running_M / (N - correction) if N > correction else 0.0 

147 

148 # 存储结果 

149 tl.store(Mean + row, running_mean, current_row_mask[:, 0]) 

150 tl.store(Var + row, variance, current_row_mask[:, 0]) 

151 

152 

153@libentry() 

154@triton.jit 

155def var_mean_kernel_1( 

156 X, 

157 Acc, 

158 Average, 

159 Count, 

160 N, 

161 BLOCK_N: tl.constexpr, 

162): 

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

164 pid = tle.program_id(0) 

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

166 

167 X = X + offset 

168 Acc = Acc + pid 

169 Average = Average + pid 

170 Count = Count + pid 

171 mask = offset < N 

172 

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

174 

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

176 average = tl.sum(x) / count 

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

178 

179 tl.store(Average, average) 

180 tl.store(Acc, acc) 

181 tl.store(Count, count) 

182 

183 

184@libentry() 

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

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

187def var_mean_kernel_2( 

188 Acc, 

189 Average, 

190 Count, 

191 Var, 

192 Mean, 

193 N, 

194 correction, 

195 BLOCK_NUM, 

196 BLOCK_N: tl.constexpr, 

197): 

198 offset = tl.arange(0, BLOCK_N) 

199 mask = offset < BLOCK_NUM 

200 Acc = Acc + offset 

201 Average = Average + offset 

202 Count = Count + offset 

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

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

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

206 

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

208 # 手动实现 tl.reduce 的功能，沿着 axis=0 进行归约 

209 # 计算总计数 

210 total_count = tl.sum(count) 

211 

212 # 计算加权平均值 

213 weighted_sum = tl.sum(average * count) 

214 mean = weighted_sum / tl.maximum(total_count, 1) 

215 

216 # 计算方差累积值 

217 # 对于每个块，计算其对总体方差的贡献 

218 # 这是 Welford 算法的并行化版本 

219 local_var_contrib = acc + count * (average - mean) * (average - mean) 

220 nvar = tl.sum(local_var_contrib) 

221 

222 var = nvar / (N - correction) 

223 tl.store(Mean, mean) 

224 tl.store(Var, var) 

225 

226 

227def var_mean(x, dim=None, *, correction=None, keepdim=False): 

228 logger.debug("GEMS_ASCEND VAR MEAN") 

229 if correction is None: 

230 correction = 1.0 

231 

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

233 dim = list(range(x.ndim)) 

234 shape = [1] * x.ndim 

235 N = x.numel() 

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

237 mean = torch.empty(shape, dtype=x.dtype, device=x.device) 

238 BLOCK_N = 1024 

239 BLOCK_NUM = triton.cdiv(N, BLOCK_N) 

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

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

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

243 

244 with torch_device_fn.device(x.device): 

245 var_mean_kernel_1[(BLOCK_NUM,)](x, acc, average, count, N, BLOCK_N=BLOCK_N) 

246 var_mean_kernel_2[(1,)]( 

247 acc, average, count, var, mean, N, correction, BLOCK_NUM 

248 ) 

249 else: 

250 shape = list(x.shape) 

251 dim = [d % x.ndim for d in dim] 

252 x = dim_compress(x, dim) 

253 N = 1 

254 for i in dim: 

255 N *= shape[i] 

256 shape[i] = 1 

257 M = x.numel() // N 

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

259 mean = torch.empty(shape, dtype=x.dtype, device=x.device) 

260 

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

262 with torch_device_fn.device(x.device): 

263 var_mean_welford_kernel[grid](x, var, mean, M, N, correction) 

264 

265 if not keepdim: 

266 var = var.squeeze(dim=dim) 

267 mean = mean.squeeze(dim=dim) 

268 return var, mean