缠论引擎提速 2.6x，瓶颈是逐行 Series 查找而非指标计算

原以为浪费在 add_indicators 算了太多用不到的指标，实测它只占全量构建的 1.3%——talib 是向量化 C 代码，便宜。真正的两处： cal_kl_data 占 96%：每根 K 线 df.iloc[i] 新建一个 40 列 Series，再在其上做几十次逐键查找。改为预取 ndarray 后 2 万根 1946ms → 824ms。 ChanKLC.cal_all_ema_status 占 25%：每次合并 KLU 都立即重算，而它产出的 ema_status / ema52_pos / ema52_status 全仓无任何读取方（含前端）。改为惰性求值，保留属性形式以防将来有人读。顺带删掉 get_klc_list 里累加一整轮后直接丢弃的 ema_up_list / ema_down_list。另加 TF_DF(lean=True)：只构建到中枢，跳过线段/走势中枢/MACD 状态机——这些只服务 bsp_list 与 web 展示，笔和中枢不依赖。研究与实盘走这条快 3.6x。结果 2 万根 5m：full 1946 → 754ms，lean → 543ms。 step46_engine_parity.py 是配套的安全网，改引擎前先跑一次 --save。它对 KLC 端点与分型、笔起止价与 is_sure、中枢 zg/zd/available_ts/阶梯、信号全部输出列，以及 26 个被下游消费的 dataframe 列取哈希。本次三处改动逐步验证，另用 git stash 切回改动前代码在 20 万根 × 5 用例上做了跨版本逐位对拍，全部一致；增量路径与 web API 也各验一遍。 Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-28 04:04:44 +08:00
parent 3b14bba247
commit 0b4d7693b8
7 changed files with 1300 additions and 80 deletions
@@ -147,41 +147,53 @@ class KlineBuilderMixin:
 		return df['volume_ratio']

 	def cal_kl_data(self, dataframe:DataFrame):
-		fields = "time,open,high,low,close,volume"
+		"""按行构造 KLU 链。
+
+		这里刻意不用 `dataframe.iloc[i]`：那会为每一根新建一个几十列的 Series，
+		随后 set_indicators 再在其上做几十次逐键查找。实测这两件事合计占 TF_DF
+		构建耗时的 96%。改为先把用到的列取成 ndarray，循环里只做整数下标访问。
+		"""
+		n = len(dataframe)
+		if n == 0:
+			return []
+
+		times = self._format_times(dataframe['date'])
+		o_a = dataframe['open'].to_numpy(dtype=float)
+		h_a = dataframe['high'].to_numpy(dtype=float)
+		l_a = dataframe['low'].to_numpy(dtype=float)
+		c_a = dataframe['close'].to_numpy(dtype=float)
+		v_a = dataframe['volume'].to_numpy(dtype=float)
+
+		has_ind = 'macd' in dataframe.columns
+		ind_cols = {}
+		if has_ind:
+			for _attr, col in ChanKLU.INDICATOR_FIELDS:
+				if col in dataframe.columns:
+					ind_cols[col] = dataframe[col].to_numpy(dtype=float)
+
 		klu_list = []
 		last_klu = None
-		for i in range(0, len(dataframe)):
-			item = dataframe.iloc[i]
-			date = item['date']
-			o = item['open']
-			h = item['high']
-			l = item['low']
-			c = item['close']
-			v = item['volume']
-			# time_obj = date.fromtimestamp(date)
-			# date = date + timedelta(hours=8)
-			time_str = date.strftime('%Y-%m-%d %H:%M:%S')
-			item_data = [
-				time_str,
-				o,
-				h,
-				l,
-				c,
-				v
-			]
-			# klu =  KLU(self.create_item_dict(item_data, GetColumnNameFromFieldList(fields)))
-			klu = ChanKLU(time_str, o, h, l, c, v)
-			# print(klu.time, klu.open, klu.high, klu.low, klu.close, klu.volume)	
+		for i in range(n):
+			klu = ChanKLU(times[i], o_a[i], h_a[i], l_a[i], c_a[i], v_a[i])
 			klu.set_idx(i)
 			klu_list.append(klu)
 			if last_klu:
 				last_klu.set_next(klu)
 				klu.set_pre(last_klu)
 			last_klu = klu
-			if 'macd' in item:
-				klu.set_indicators(item)
+			if has_ind:
+				klu.set_indicators_from(ind_cols, i)
 		return klu_list

+	@staticmethod
+	def _format_times(col):
+		"""向量化 strftime。非 datetime 列（少见）退回逐个格式化。"""
+		fmt = '%Y-%m-%d %H:%M:%S'
+		try:
+			return col.dt.strftime(fmt).to_numpy()
+		except AttributeError:
+			return np.array([d.strftime(fmt) for d in col], dtype=object)
+
 	def get_kl_data(self, dataframe:DataFrame):
 		return self.cal_kl_data(dataframe)

@@ -224,35 +236,20 @@ class KlineBuilderMixin:

 	def get_klc_list(self, klu_list):
 		klc_list = []
-		last_klu = None
-		# ChanMACD.__init__ 已调用 cal_macd_state，切勿再调一次（会重复堆积 seg/unittf）
-		macd = ChanMACD(klu_list)
-		klu_list = macd.klu_list
-		self._last_chan_macd = macd
-		ema_up_list = []
-		ema_down_list = []
-		ema_up_count = 0
-		ema_down_count = 0
+		# ChanMACD.__init__ 已调用 cal_macd_state，切勿再调一次（会重复堆积 seg/unittf）。
+		# lean 模式跳过整套 MACD 状态机：它只服务于 bsp/背驰/web 展示，笔与中枢不依赖它。
+		if getattr(self, 'lean', False):
+			self._last_chan_macd = None
+		else:
+			macd = ChanMACD(klu_list)
+			klu_list = macd.klu_list
+			self._last_chan_macd = macd
+
 		last_klu = None
 		for klu in klu_list:
-			ema = klu.ema52
-			last_ema = last_klu.ema52 if last_klu else 0
-			if klu.close >= ema:
-				ema_up_count += 1
-			elif klu.close < ema:
-				ema_down_count += 1
-			if last_klu and last_klu.close >= last_ema and klu.close < ema:
-				ema_up_list.append(ema_up_count)
-				#print(last_klu.time, ema_up_count, "UP END")
-				ema_up_count = 0
-			elif last_klu and last_klu.close < last_ema and klu.close >= ema:
-				ema_down_list.append(ema_down_count)
-				#print(last_klu.time, ema_down_count, "DOWN END")
-				ema_down_count = 0
 			self._push_klu_into_klc_list(klc_list, klu, last_klu)
 			last_klu = klu
 		klc_list = self.cal_trend(klc_list)
-		#print(ema52_up_list, ema52_down_list)
 		return klc_list


@@ -38,10 +38,18 @@ from chanlun.pipeline.builders.seg import SegBuilderMixin
 from chanlun.pipeline.builders.zs import ZsBuilderMixin

 class TF_DF(IndicatorsBuilderMixin, KlineBuilderMixin, BiBuilderMixin, SegBuilderMixin, ZsBuilderMixin, BspBuilderMixin, FastBspBuilderMixin, IncrementalBuilderMixin):
-	def __init__(self, df=None, interval=0, timeframe=None):
+	def __init__(self, df=None, interval=0, timeframe=None, lean=False):
+		"""lean=True 只构建到中枢，跳过线段/走势中枢/MACD 状态机。
+
+		研究与实盘只吃 bi_list → 中枢 → fast_bsp 这条链；线段、zs、big_zs 和整套
+		MACD 背驰状态机是 web 展示与 bsp_list 才用的。实测这些占全量构建的约四成。
+		注意 lean 下 bsp_list/seg_list/chanmacd 均为空，**不要给 web 用**。
+		"""
+		self.lean = lean
 		if df is not None:
-			self.init_TF_DF(df, interval, timeframe)
-	def init_TF_DF(self, df, interval, timeframe):
+			self.init_TF_DF(df, interval, timeframe, lean=lean)
+	def init_TF_DF(self, df, interval, timeframe, lean=False):
+		self.lean = lean
 		self.timeframe = timeframe
 		self.interval = interval
 		# 检查 DataFrame 是否为空或没有 date 列
@@ -69,6 +77,10 @@ class TF_DF(IndicatorsBuilderMixin, KlineBuilderMixin, BiBuilderMixin, SegBuilde
 		self.klc_list = self.get_klc_list(self.klu_list)
 		self.bi_list = self.cal_bi_list(self.klc_list)
 		self.bi_zs_list = self.cal_bi_zs_list_pure(self.bi_list)
+		if self.lean:
+			self.big_zs_list = []
+			self.chanmacd = None
+			return
 		self.seg_list = self.get_seg_list(self.bi_list)
 		self.zs_list = self.get_zs_list(self.bi_list, self.seg_list)
 		self.big_zs_list = self.get_big_zs_list(self.zs_list)