Merge pull request #150 from githubnext/autoloop/perf-comparison-8724e9f9

[Autoloop] [Autoloop: perf-comparison]

Author: mrjf

benchmarks/pandas/bench_any_all.py

@@ -0,0 +1,36 @@
+"""Benchmark: any_all — Series.any / all and DataFrame.any / all on 100k rows."""
+import json, time
+import numpy as np
+import pandas as pd
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 50
+
+s = pd.Series(np.arange(SIZE) % 2 == 0)
+df = pd.DataFrame({
+    "a": np.arange(SIZE) % 3 != 0,
+    "b": np.arange(SIZE) > 0,
+    "c": np.ones(SIZE, dtype=bool),
+})
+
+for _ in range(WARMUP):
+    s.any()
+    s.all()
+    df.any()
+    df.all()
+
+start = time.perf_counter()
+for _ in range(ITERATIONS):
+    s.any()
+    s.all()
+    df.any()
+    df.all()
+total = (time.perf_counter() - start) * 1000
+
+print(json.dumps({
+    "function": "any_all",
+    "mean_ms": total / ITERATIONS,
+    "iterations": ITERATIONS,
+    "total_ms": total,
+}))

benchmarks/pandas/bench_astype_df_fn.py

@@ -0,0 +1,31 @@
+"""Benchmark: astype standalone — DataFrame.astype with per-column and uniform dtype on 100k-row DataFrame."""
+import json, time
+import pandas as pd
+import numpy as np
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 50
+
+df = pd.DataFrame({
+    "a": np.arange(SIZE, dtype=np.float64),
+    "b": np.arange(SIZE, dtype=np.int64),
+    "c": np.where(np.arange(SIZE) % 2 == 0, 1, 0).astype(np.int64),
+})
+
+for _ in range(WARMUP):
+    df.astype({"a": "float32", "b": "int32"})
+    df.astype("float64")
+
+start = time.perf_counter()
+for _ in range(ITERATIONS):
+    df.astype({"a": "float32", "b": "int32"})
+    df.astype("float64")
+total = (time.perf_counter() - start) * 1000
+
+print(json.dumps({
+    "function": "astype_df_fn",
+    "mean_ms": round(total / ITERATIONS, 3),
+    "iterations": ITERATIONS,
+    "total_ms": round(total, 3),
+}))

benchmarks/pandas/bench_cat_freq_crosstab.py

@@ -0,0 +1,39 @@
+"""
+Benchmark: pd.Series.value_counts (freq table) and pd.crosstab for categorical data on 100k elements.
+Outputs JSON: {"function": "cat_freq_crosstab", "mean_ms": ..., "iterations": ..., "total_ms": ...}
+"""
+import json
+import time
+import pandas as pd
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 20
+
+cats_a = ["alpha", "beta", "gamma", "delta", "epsilon"]
+cats_b = ["north", "south", "east", "west"]
+data_a = pd.Categorical([cats_a[i % len(cats_a)] for i in range(SIZE)], categories=cats_a)
+data_b = pd.Categorical([cats_b[i % len(cats_b)] for i in range(SIZE)], categories=cats_b)
+s_a = pd.Series(data_a)
+s_b = pd.Series(data_b)
+
+for _ in range(WARMUP):
+    s_a.value_counts(sort=False)
+    pd.crosstab(s_a, s_b)
+    pd.crosstab(s_a, s_b, normalize=True)
+
+times = []
+for _ in range(ITERATIONS):
+    t0 = time.perf_counter()
+    s_a.value_counts(sort=False)
+    pd.crosstab(s_a, s_b)
+    pd.crosstab(s_a, s_b, normalize=True)
+    times.append((time.perf_counter() - t0) * 1000)
+
+total_ms = sum(times)
+print(json.dumps({
+    "function": "cat_freq_crosstab",
+    "mean_ms": total_ms / ITERATIONS,
+    "iterations": ITERATIONS,
+    "total_ms": total_ms,
+}))

benchmarks/pandas/bench_cat_intersect_diff.py

@@ -0,0 +1,53 @@
+"""
+Benchmark: pandas category set operations — intersection and difference of
+categorical Series categories (100k-element, 20 categories each).
+Mirrors tsb's catIntersectCategories / catDiffCategories.
+Outputs JSON: {"function": "cat_intersect_diff", "mean_ms": ..., "iterations": ..., "total_ms": ...}
+"""
+import json
+import time
+import pandas as pd
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 30
+
+cats_a = [f"cat_a_{i}" for i in range(20)]
+cats_b = [f"cat_{'a' if i < 10 else 'b'}_{i}" for i in range(20)]
+
+data_a = [cats_a[i % len(cats_a)] for i in range(SIZE)]
+data_b = [cats_b[i % len(cats_b)] for i in range(SIZE)]
+
+s_a = pd.Categorical(data_a, categories=cats_a)
+s_b = pd.Categorical(data_b, categories=cats_b)
+
+def cat_intersect(a, b):
+    """Return new Categorical with categories = intersection of a.categories and b.categories."""
+    b_set = set(b.categories)
+    intersected = [c for c in a.categories if c in b_set]
+    return pd.Categorical(a, categories=intersected)
+
+def cat_diff(a, b):
+    """Return new Categorical with categories = a.categories - b.categories."""
+    b_set = set(b.categories)
+    remaining = [c for c in a.categories if c not in b_set]
+    return pd.Categorical(a, categories=remaining)
+
+for _ in range(WARMUP):
+    cat_intersect(s_a, s_b)
+    cat_diff(s_a, s_b)
+
+times = []
+for _ in range(ITERATIONS):
+    t0 = time.perf_counter()
+    cat_intersect(s_a, s_b)
+    cat_diff(s_a, s_b)
+    times.append((time.perf_counter() - t0) * 1000)
+
+total_ms = sum(times)
+print(json.dumps({
+    "function": "cat_intersect_diff",
+    "mean_ms": total_ms / ITERATIONS,
+    "iterations": ITERATIONS,
+    "total_ms": total_ms,
+}))

benchmarks/pandas/bench_cat_ops_from_codes.py

@@ -0,0 +1,49 @@
+"""
+Benchmark: pd.Categorical.from_codes, reorder_categories by freq, ordered categorical on 100k elements.
+Outputs JSON: {"function": "cat_ops_from_codes", "mean_ms": ..., "iterations": ..., "total_ms": ...}
+"""
+import json
+import time
+import pandas as pd
+import numpy as np
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 20
+
+categories = ["alpha", "beta", "gamma", "delta", "epsilon"]
+codes = [i % len(categories) for i in range(SIZE)]
+order = ["epsilon", "delta", "gamma", "beta", "alpha"]
+
+def cat_from_codes():
+    return pd.Categorical.from_codes(codes, categories=categories)
+
+def cat_sort_by_freq(c):
+    s = pd.Series(c)
+    freq_order = s.value_counts().index.tolist()
+    return s.astype(pd.CategoricalDtype(categories=freq_order, ordered=False))
+
+def cat_to_ordinal(c):
+    s = pd.Series(c)
+    return s.astype(pd.CategoricalDtype(categories=order, ordered=True))
+
+for _ in range(WARMUP):
+    c = cat_from_codes()
+    cat_sort_by_freq(c)
+    cat_to_ordinal(c)
+
+times = []
+for _ in range(ITERATIONS):
+    t0 = time.perf_counter()
+    c = cat_from_codes()
+    cat_sort_by_freq(c)
+    cat_to_ordinal(c)
+    times.append((time.perf_counter() - t0) * 1000)
+
+total_ms = sum(times)
+print(json.dumps({
+    "function": "cat_ops_from_codes",
+    "mean_ms": total_ms / ITERATIONS,
+    "iterations": ITERATIONS,
+    "total_ms": total_ms,
+}))

benchmarks/pandas/bench_cat_ops_setops.py

@@ -0,0 +1,51 @@
+"""
+Benchmark: categorical union/intersect/diff categories on 100k element Series.
+Outputs JSON: {"function": "cat_ops_setops", "mean_ms": ..., "iterations": ..., "total_ms": ...}
+"""
+import json
+import time
+import pandas as pd
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 20
+
+cats_a = ["alpha", "beta", "gamma", "delta"]
+cats_b = ["gamma", "delta", "epsilon", "zeta"]
+data_a = [cats_a[i % len(cats_a)] for i in range(SIZE)]
+data_b = [cats_b[i % len(cats_b)] for i in range(SIZE)]
+s_a = pd.Series(data_a, dtype="category")
+s_b = pd.Series(data_b, dtype="category")
+
+def cat_union(a, b):
+    cats = list(dict.fromkeys(list(a.cat.categories) + [c for c in b.cat.categories if c not in a.cat.categories]))
+    return a.astype(pd.CategoricalDtype(categories=cats))
+
+def cat_intersect(a, b):
+    cats = [c for c in a.cat.categories if c in set(b.cat.categories)]
+    return a.astype(pd.CategoricalDtype(categories=cats))
+
+def cat_diff(a, b):
+    cats = [c for c in a.cat.categories if c not in set(b.cat.categories)]
+    return a.astype(pd.CategoricalDtype(categories=cats))
+
+for _ in range(WARMUP):
+    cat_union(s_a, s_b)
+    cat_intersect(s_a, s_b)
+    cat_diff(s_a, s_b)
+
+times = []
+for _ in range(ITERATIONS):
+    t0 = time.perf_counter()
+    cat_union(s_a, s_b)
+    cat_intersect(s_a, s_b)
+    cat_diff(s_a, s_b)
+    times.append((time.perf_counter() - t0) * 1000)
+
+total_ms = sum(times)
+print(json.dumps({
+    "function": "cat_ops_setops",
+    "mean_ms": total_ms / ITERATIONS,
+    "iterations": ITERATIONS,
+    "total_ms": total_ms,
+}))

benchmarks/pandas/bench_combine_first_fn.py

@@ -0,0 +1,29 @@
+"""Benchmark: combineFirstSeries standalone — pd.Series.combine_first() on 50k-element Series with 30% NaN."""
+import json, time
+import pandas as pd
+import numpy as np
+
+SIZE = 50_000
+WARMUP = 5
+ITERATIONS = 30
+
+rng = np.random.default_rng(42)
+data1 = rng.standard_normal(SIZE)
+data1[::3] = float("nan")  # ~30% nulls
+s1 = pd.Series(data1)
+s2 = pd.Series(np.arange(SIZE, dtype=np.float64) * 2.0)
+
+for _ in range(WARMUP):
+    s1.combine_first(s2)
+
+start = time.perf_counter()
+for _ in range(ITERATIONS):
+    s1.combine_first(s2)
+total = (time.perf_counter() - start) * 1000
+
+print(json.dumps({
+    "function": "combine_first_fn",
+    "mean_ms": round(total / ITERATIONS, 3),
+    "iterations": ITERATIONS,
+    "total_ms": round(total, 3),
+}))

benchmarks/pandas/bench_combine_first_series.py

@@ -0,0 +1,33 @@
+"""Benchmark: Series.combine_first (standalone equivalent) — fill missing values from another Series.
+Mirrors tsb bench_combine_first_series.ts for pandas.
+"""
+import json, time
+import pandas as pd
+import numpy as np
+
+SIZE = 10_000
+WARMUP = 5
+ITERATIONS = 50
+
+data1 = [None if i % 3 == 0 else i * 0.5 for i in range(SIZE)]
+data2 = [i * 0.1 for i in range(SIZE)]
+s1 = pd.Series(data1)
+s2 = pd.Series(data2)
+
+for _ in range(WARMUP):
+    s1.combine_first(s2)
+
+times = []
+for _ in range(ITERATIONS):
+    t0 = time.perf_counter()
+    s1.combine_first(s2)
+    times.append((time.perf_counter() - t0) * 1000)
+
+total = sum(times)
+mean = total / ITERATIONS
+print(json.dumps({
+    "function": "combine_first_series",
+    "mean_ms": round(mean, 3),
+    "iterations": ITERATIONS,
+    "total_ms": round(total, 3),
+}))

benchmarks/pandas/bench_combine_first_series_fn.py

@@ -0,0 +1,41 @@
+"""
+Benchmark: Series.combine_first() — fill NaN values from another Series (union of indexes).
+Mirrors tsb bench_combine_first_series_fn.ts (standalone combineFirstSeries fn).
+Outputs JSON: {"function": "combine_first_series_fn", "mean_ms": ..., "iterations": ..., "total_ms": ...}
+"""
+import json
+import time
+import numpy as np
+import pandas as pd
+
+SIZE = 100_000
+WARMUP = 5
+ITERATIONS = 30
+
+rng = np.random.default_rng(42)
+raw = rng.uniform(0, 10, SIZE)
+mask = rng.integers(0, 4, SIZE) == 0  # ~25% nulls
+d1 = pd.array(raw, dtype="Float64")
+for idx in range(SIZE):
+    if mask[idx]:
+        d1[idx] = pd.NA
+
+s1 = pd.Series(d1, dtype="Float64")
+s2 = pd.Series(rng.uniform(0, 10, SIZE))
+
+for _ in range(WARMUP):
+    s1.combine_first(s2)
+
+times = []
+for _ in range(ITERATIONS):
+    t0 = time.perf_counter()
+    s1.combine_first(s2)
+    times.append((time.perf_counter() - t0) * 1000)
+
+total_ms = sum(times)
+print(json.dumps({
+    "function": "combine_first_series_fn",
+    "mean_ms": round(total_ms / ITERATIONS, 3),
+    "iterations": ITERATIONS,
+    "total_ms": round(total_ms, 3),
+}))