Compares filtering implementations (numpy, cython)

The benchmark looks into different ways to implement thresholding: every value of a vector superior to mx is replaced by mx (numpy.clip()). It compares several implementation to numpy.

• cfilter_dmax cfilter_dmax

• cfilter_dmax2 cfilter_dmax2

• cfilter_dmax4 cfilter_dmax4

• cfilter_dmax16 cfilter_dmax16

• cyfilter_dmax cyfilter_dmax

• filter_dmax_cython filter_dmax_cython

• filter_dmax_cython_optim filter_dmax_cython_optim

• pyfilter_dmax pyfilter_dmax

import pprint
import numpy
import matplotlib.pyplot as plt
from pandas import DataFrame
from teachcompute.validation.cython.experiment_cython import (
    pyfilter_dmax,
    filter_dmax_cython,
    filter_dmax_cython_optim,
    cyfilter_dmax,
    cfilter_dmax,
    cfilter_dmax2,
    cfilter_dmax16,
    cfilter_dmax4,
)
from teachcompute.ext_test_case import measure_time_dim


def get_vectors(fct, n, h=200, dtype=numpy.float64):
    ctxs = [
        dict(
            va=numpy.random.randn(n).astype(dtype),
            fil=fct,
            mx=numpy.float64(0),
            x_name=n,
        )
        for n in range(10, n, h)
    ]
    return ctxs


def numpy_filter(va, mx):
    va[va > mx] = mx


all_res = []
for fct in [
    numpy_filter,
    pyfilter_dmax,
    filter_dmax_cython,
    filter_dmax_cython_optim,
    cyfilter_dmax,
    cfilter_dmax,
    cfilter_dmax2,
    cfilter_dmax16,
    cfilter_dmax4,
]:
    print(fct)
    ctxs = get_vectors(fct, 1000 if fct == pyfilter_dmax else 40000)
    res = list(measure_time_dim("fil(va, mx)", ctxs, verbose=1))
    for r in res:
        r["fct"] = fct.__name__
    all_res.extend(res)

pprint.pprint(all_res[:2])

<function numpy_filter at 0x7ffa185cbe20>

  0%|          | 0/200 [00:00<?, ?it/s]
 37%|███▋      | 74/200 [00:00<00:00, 734.22it/s]
 74%|███████▍  | 148/200 [00:00<00:00, 335.10it/s]
 96%|█████████▋| 193/200 [00:00<00:00, 278.68it/s]
100%|██████████| 200/200 [00:00<00:00, 282.56it/s]
<cyfunction pyfilter_dmax at 0x7ff9972718a0>

  0%|          | 0/5 [00:00<?, ?it/s]
100%|██████████| 5/5 [00:00<00:00, 68.45it/s]
<cyfunction filter_dmax_cython at 0x7ff9972717d0>

  0%|          | 0/200 [00:00<?, ?it/s]
 54%|█████▍    | 108/200 [00:00<00:00, 1063.95it/s]
100%|██████████| 200/200 [00:00<00:00, 557.82it/s]
<cyfunction filter_dmax_cython_optim at 0x7ff997271700>

  0%|          | 0/200 [00:00<?, ?it/s]
 53%|█████▎    | 106/200 [00:00<00:00, 1050.56it/s]
100%|██████████| 200/200 [00:00<00:00, 590.42it/s]
<cyfunction cyfilter_dmax at 0x7ff997271220>

  0%|          | 0/200 [00:00<?, ?it/s]
 48%|████▊     | 95/200 [00:00<00:00, 945.32it/s]
 95%|█████████▌| 190/200 [00:00<00:00, 565.05it/s]
100%|██████████| 200/200 [00:00<00:00, 575.55it/s]
<cyfunction cfilter_dmax at 0x7ff997271970>

  0%|          | 0/200 [00:00<?, ?it/s]
 50%|█████     | 100/200 [00:00<00:00, 996.86it/s]
100%|██████████| 200/200 [00:00<00:00, 489.97it/s]
100%|██████████| 200/200 [00:00<00:00, 530.03it/s]
<cyfunction cfilter_dmax2 at 0x7ff997271be0>

  0%|          | 0/200 [00:00<?, ?it/s]
 48%|████▊     | 96/200 [00:00<00:00, 949.86it/s]
 96%|█████████▌| 191/200 [00:00<00:00, 433.61it/s]
100%|██████████| 200/200 [00:00<00:00, 448.65it/s]
<cyfunction cfilter_dmax16 at 0x7ff997271b10>

  0%|          | 0/200 [00:00<?, ?it/s]
 36%|███▌      | 71/200 [00:00<00:00, 707.81it/s]
 71%|███████   | 142/200 [00:00<00:00, 365.45it/s]
 94%|█████████▍| 188/200 [00:00<00:00, 267.68it/s]
100%|██████████| 200/200 [00:00<00:00, 290.62it/s]
<cyfunction cfilter_dmax4 at 0x7ff997271a40>

  0%|          | 0/200 [00:00<?, ?it/s]
 26%|██▋       | 53/200 [00:00<00:00, 528.62it/s]
 53%|█████▎    | 106/200 [00:00<00:00, 238.76it/s]
 69%|██████▉   | 138/200 [00:00<00:00, 169.93it/s]
 80%|████████  | 161/200 [00:00<00:00, 147.09it/s]
 90%|████████▉ | 179/200 [00:01<00:00, 127.18it/s]
 97%|█████████▋| 194/200 [00:01<00:00, 113.76it/s]
100%|██████████| 200/200 [00:01<00:00, 143.99it/s]
[{'average': np.float64(8.415899997089582e-07),
  'context_size': 184,
  'deviation': np.float64(1.3342516226394618e-07),
  'fct': 'numpy_filter',
  'max_exec': np.float64(1.2366200007818407e-06),
  'min_exec': np.float64(7.848800032661529e-07),
  'number': 50,
  'repeat': 10,
  'ttime': np.float64(8.415899997089581e-06),
  'warmup_time': 3.251600014664291e-05,
  'x_name': 10},
 {'average': np.float64(8.49112000196328e-07),
  'context_size': 184,
  'deviation': np.float64(8.249650193622526e-08),
  'fct': 'numpy_filter',
  'max_exec': np.float64(1.0923399986495497e-06),
  'min_exec': np.float64(8.093399992503691e-07),
  'number': 50,
  'repeat': 10,
  'ttime': np.float64(8.49112000196328e-06),
  'warmup_time': 8.533000027455273e-06,
  'x_name': 210}]

Let’s display the results

cc = DataFrame(all_res)
cc["N"] = cc["x_name"]

fig, ax = plt.subplots(2, 2, figsize=(10, 10))
cc[cc.N <= 1100].pivot(index="N", columns="fct", values="average").plot(
    logy=True, ax=ax[0, 0]
)
cc[cc.fct != "pyfilter_dmax"].pivot(index="N", columns="fct", values="average").plot(
    logy=True, ax=ax[0, 1]
)
cc[cc.fct != "pyfilter_dmax"].pivot(index="N", columns="fct", values="average").plot(
    logy=True, logx=True, ax=ax[1, 1]
)
cc[(cc.fct.str.contains("cfilter") | cc.fct.str.contains("numpy"))].pivot(
    index="N", columns="fct", values="average"
).plot(logy=True, ax=ax[1, 0])
ax[0, 0].set_title("Comparison of filter implementations")
ax[0, 1].set_title("Comparison of filter implementations\nwithout pyfilter_dmax")

Text(0.5, 1.0, 'Comparison of filter implementations\nwithout pyfilter_dmax')

The results depends on the machine, its number of cores, the compilation settings of numpy or this module.