7.5. Benchmarking#

Tip

This notebook benchmarks JAX on a single CPU core. Compare with Julia results as reported in ComPWA/polarimetry#27. See also the Extended benchmark #68 discussion.

Note

This notebook uses only one run and one loop for %timeit, because JAX seems to cache its return values.

Hide code cell source 
from __future__ import annotations

import logging
from collections import defaultdict

import numpy as np
import pandas as pd
import sympy as sp
from IPython.display import Markdown
from psutil import cpu_count

from polarimetry import formulate_polarimetry
from polarimetry.data import (
    create_data_transformer,
    generate_meshgrid_sample,
    generate_phasespace_sample,
)
from polarimetry.io import (
    mute_jax_warnings,
    perform_cached_doit,
    perform_cached_lambdify,
)
from polarimetry.lhcb import (
    load_model_builder,
    load_model_parameters,
)
from polarimetry.lhcb.particle import load_particles

LOGGER = logging.getLogger()
LOGGER.setLevel(logging.ERROR)
mute_jax_warnings()

model_choice = 0
model_file = "../../data/model-definitions.yaml"
particles = load_particles("../../data/particle-definitions.yaml")
amplitude_builder = load_model_builder(model_file, particles, model_choice)
imported_parameter_values = load_model_parameters(
    model_file, amplitude_builder.decay, model_choice, particles
)
reference_subsystem = 1
model = amplitude_builder.formulate(reference_subsystem)
model.parameter_defaults.update(imported_parameter_values)

timing_parametrized = defaultdict(dict)
timing_substituted = defaultdict(dict)

print("Physical cores:", cpu_count(logical=False))
print("Total cores:", cpu_count(logical=True))

Physical cores: 16
Total cores: 16

%%time
polarimetry_exprs = formulate_polarimetry(amplitude_builder, reference_subsystem)
unfolded_polarimetry_exprs = [
    perform_cached_doit(expr.doit().xreplace(model.amplitudes))
    for expr in polarimetry_exprs
]
unfolded_intensity_expr = perform_cached_doit(model.full_expression)

CPU times: user 1min 30s, sys: 100 ms, total: 1min 30s
Wall time: 1min 31s

7.5.1. DataTransformer performance#

n_events = 100_000
phsp_sample = generate_phasespace_sample(model.decay, n_events, seed=0)
transformer = create_data_transformer(model)
%timeit -n1 -r1 transformer(phsp_sample)  # first run, so no cache and JIT-compilation
%timeit -n1 -r1 transformer(phsp_sample)  # second run with cache
%timeit -n1 -r1 transformer(phsp_sample)  # third run with cache
phsp_sample = transformer(phsp_sample)
random_point = {k: v[0] if len(v.shape) > 0 else v for k, v in phsp_sample.items()}

303 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
8.19 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
7.93 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

res = 54
grid_sample = generate_meshgrid_sample(model.decay, res)
%timeit -n1 -r1 transformer(grid_sample)  # first run, without cache, but already compiled
%timeit -n1 -r1 transformer(grid_sample)  # second run with cache
%timeit -n1 -r1 transformer(grid_sample)  # third run with cache
grid_sample = transformer(grid_sample)

274 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
2.34 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
1.92 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

7.5.2. Parametrized function#

Total number of mathematical operations:

  • \(\alpha_x\): 133,630

  • \(\alpha_y\): 133,634

  • \(\alpha_z\): 133,630

  • \(I_\mathrm{tot}\): 43,198

%%time
parametrized_polarimetry_funcs = [
    perform_cached_lambdify(
        expr,
        parameters=model.parameter_defaults,
        backend="jax",
    )
    for expr in unfolded_polarimetry_exprs
]
parametrized_intensity_func = perform_cached_lambdify(
    unfolded_intensity_expr,
    parameters=model.parameter_defaults,
    backend="jax",
)

CPU times: user 14.2 s, sys: 6.9 ms, total: 14.2 s
Wall time: 14.3 s

rng = np.random.default_rng(seed=0)
original_parameters = dict(parametrized_intensity_func.parameters)
modified_parameters = {
    k: rng.uniform(0.9, 1.1) * v
    for k, v in parametrized_intensity_func.parameters.items()
}

7.5.2.1. One data point#

7.5.2.1.1. JIT-compilation#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(random_point)

<TimeitResult : 1.08 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](random_point)
array = parametrized_polarimetry_funcs[1](random_point)
array = parametrized_polarimetry_funcs[2](random_point)

<TimeitResult : 5.62 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.1.2. Compiled performance#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(random_point)

<TimeitResult : 420 µs ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](random_point)
array = parametrized_polarimetry_funcs[1](random_point)
array = parametrized_polarimetry_funcs[2](random_point)

<TimeitResult : 841 µs ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.2. 54x54 grid sample#

7.5.2.2.1. Compiled but uncached#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(grid_sample)

<TimeitResult : 1.19 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](grid_sample)
array = parametrized_polarimetry_funcs[1](grid_sample)
array = parametrized_polarimetry_funcs[2](grid_sample)

<TimeitResult : 7.08 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.2.2. Second run with cache#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(grid_sample)

<TimeitResult : 3.84 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](grid_sample)
array = parametrized_polarimetry_funcs[1](grid_sample)
array = parametrized_polarimetry_funcs[2](grid_sample)

<TimeitResult : 5.69 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.3. 100.000 event phase space sample#

7.5.2.3.1. Compiled but uncached#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(phsp_sample)

<TimeitResult : 1.22 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](phsp_sample)
array = parametrized_polarimetry_funcs[1](phsp_sample)
array = parametrized_polarimetry_funcs[2](phsp_sample)

<TimeitResult : 7.14 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.3.2. Second run with cache#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(phsp_sample)

<TimeitResult : 15.6 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](phsp_sample)
array = parametrized_polarimetry_funcs[1](phsp_sample)
array = parametrized_polarimetry_funcs[2](phsp_sample)

<TimeitResult : 66.3 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.4. Recompilation after parameter modification#

parametrized_intensity_func.update_parameters(modified_parameters)
for func in parametrized_polarimetry_funcs:
    func.update_parameters(modified_parameters)

7.5.2.4.1. Compiled but uncached#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(phsp_sample)

<TimeitResult : 1.17 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](phsp_sample)
array = parametrized_polarimetry_funcs[1](phsp_sample)
array = parametrized_polarimetry_funcs[2](phsp_sample)

<TimeitResult : 7.17 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.2.4.2. Second run with cache#

%%timeit -n1 -r1 -q -o
array = parametrized_intensity_func(phsp_sample)

<TimeitResult : 14.7 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = parametrized_polarimetry_funcs[0](phsp_sample)
array = parametrized_polarimetry_funcs[1](phsp_sample)
array = parametrized_polarimetry_funcs[2](phsp_sample)

<TimeitResult : 78.5 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

parametrized_intensity_func.update_parameters(original_parameters)
for func in parametrized_polarimetry_funcs:
    func.update_parameters(original_parameters)

7.5.3. All parameters substituted#

subs_polarimetry_exprs = [
    expr.xreplace(model.parameter_defaults) for expr in unfolded_polarimetry_exprs
]
subs_intensity_expr = unfolded_intensity_expr.xreplace(model.parameter_defaults)

Number of mathematical operations after substituting all parameters:

  • \(\alpha_x\): 29,552

  • \(\alpha_y\): 29,556

  • \(\alpha_z\): 29,552

  • \(I_\mathrm{tot}\): 9,624

%%time
polarimetry_funcs = [
    perform_cached_lambdify(expr, backend="jax") for expr in subs_polarimetry_exprs
]
intensity_func = perform_cached_lambdify(subs_intensity_expr, backend="jax")

CPU times: user 5.19 s, sys: 6.93 ms, total: 5.2 s
Wall time: 5.23 s

7.5.3.1. One data point#

7.5.3.1.1. JIT-compilation#

%%timeit -n1 -r1 -q -o
array = intensity_func(random_point)

<TimeitResult : 689 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = polarimetry_funcs[0](random_point)
array = polarimetry_funcs[1](random_point)
array = polarimetry_funcs[2](random_point)

<TimeitResult : 3.58 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.3.1.2. Compiled performance#

%%timeit -n1 -r1 -q -o
array = intensity_func(random_point)

<TimeitResult : 181 µs ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = polarimetry_funcs[0](random_point)
array = polarimetry_funcs[1](random_point)
array = polarimetry_funcs[2](random_point)

<TimeitResult : 446 µs ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.3.2. 54x54 grid sample#

7.5.3.2.1. Compiled but uncached#

%%timeit -n1 -r1 -q -o
array = intensity_func(grid_sample)

<TimeitResult : 842 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = polarimetry_funcs[0](grid_sample)
array = polarimetry_funcs[1](grid_sample)
array = polarimetry_funcs[2](grid_sample)

<TimeitResult : 4.53 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.3.2.2. Second run with cache#

%%timeit -n1 -r1 -q -o
array = intensity_func(grid_sample)

<TimeitResult : 2.69 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = polarimetry_funcs[0](grid_sample)
array = polarimetry_funcs[1](grid_sample)
array = polarimetry_funcs[2](grid_sample)

<TimeitResult : 8.47 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.3.3. 100.000 event phase space sample#

7.5.3.3.1. Compiled but uncached#

%%timeit -n1 -r1 -q -o
array = intensity_func(phsp_sample)

<TimeitResult : 891 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = polarimetry_funcs[0](phsp_sample)
array = polarimetry_funcs[1](phsp_sample)
array = polarimetry_funcs[2](phsp_sample)

<TimeitResult : 4.75 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.3.3.2. Second run with cache#

%%timeit -n1 -r1 -q -o
array = intensity_func(phsp_sample)

<TimeitResult : 13.9 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

%%timeit -n1 -r1 -q -o
array = polarimetry_funcs[0](phsp_sample)
array = polarimetry_funcs[1](phsp_sample)
array = polarimetry_funcs[2](phsp_sample)

<TimeitResult : 67 ms ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)>

7.5.4. Summary#

Hide code cell source 
def collect_sorted_row_title() -> list[str]:
    row_titles = {}
    row_titles.update(timing_parametrized["intensity"])
    row_titles.update(timing_parametrized["polarimetry"])
    row_titles.update(timing_substituted["intensity"])
    row_titles.update(timing_substituted["polarimetry"])
    return list(row_titles)


def remove_loop_info(timing) -> str:
    if timing is None:
        return ""
    pattern = " ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)"
    return str(timing).replace(pattern, "")


row_titles = collect_sorted_row_title()
values = [
    (
        remove_loop_info(timing_parametrized["intensity"].get(row)),
        remove_loop_info(timing_parametrized["polarimetry"].get(row)),
        remove_loop_info(timing_substituted["intensity"].get(row)),
        remove_loop_info(timing_substituted["polarimetry"].get(row)),
    )
    for row in row_titles
]
columns = pd.MultiIndex.from_tuples(
    [
        ("parametrized", "I"),
        ("parametrized", "ɑ"),
        ("substituted", "I"),
        ("substituted", "ɑ"),
    ],
)
df = pd.DataFrame(values, index=row_titles, columns=columns)
df.style.set_table_styles(
    [
        dict(selector="th", props=[("text-align", "left")]),
        dict(selector="td", props=[("text-align", "left")]),
    ]
)
  parametrized substituted
  I ɑ I ɑ
random point (compilation) 1.08 s 5.62 s 689 ms 3.58 s
random point (cached) 420 µs 841 µs 181 µs 446 µs
54x54 grid 1.19 s 7.08 s 842 ms 4.53 s
54x54 grid (cached) 3.84 ms 5.69 ms 2.69 ms 8.47 ms
100,000 phsp 1.22 s 7.14 s 891 ms 4.75 s
100,000 phsp (cached) 15.6 ms 66.3 ms 13.9 ms 67 ms
modified 100,000 phsp 1.17 s 7.17 s
modified 100,000 phsp (cached) 14.7 ms 78.5 ms