Add symiirorder1 to cupyx.scipy.signal

[andfoy]

See #7403

[mergify]

This pull request is now in conflicts. Could you fix it @andfoy? 🙏

[takagi]

/test mini

[mergify]

This pull request is now in conflicts. Could you fix it @andfoy? 🙏

[takagi]

/test mini

[takagi]

I merge this as the CI failures are not related. Thanks, @andfoy!

[andfoy]

These are the benchmark results for symiirorder1 after PR #7522.

For larger input sizes, the GPU implementation achieves a ~3X improvement over SciPy.

Benchmark script

from itertools import product

import cupy as cp
import numpy as np
from cupy import testing
from cupyx.profiler import benchmark

from cupyx.scipy.signal import symiirorder1 as symiir1_gpu
from scipy.signal import symiirorder1 as symiir1_cpu

from tqdm import tqdm


input_sizes = [5, 10, 100, 1000, 5000, 10000, 50000, 100000, 200000, 500000,
               700000, 1_000_000, 2_000_000, 10_000_000]
modules = [(cp, symiir1_gpu, 'CuPy'), (np, symiir1_cpu, 'SciPy')]
# kwargs = [('order', [1, 2, 4, 6, 10])]
kwargs = [('precision', [1])]
# kwargs = [('extrapolate', [True])]
kwargs = list(product(*[list(product([x[0]], x[1])) for x in kwargs]))
measurement = {}


def gather_time(prof):
    cpu_time = prof.cpu_times.mean() * 1000
    gpu_time = prof.gpu_times.mean() * 1000
    return max(cpu_time, gpu_time)


def input_creation(xp, fn, size, precision=-1, **kwargs):
    size = size[0]
    x = testing.shaped_random((size,), xp, dtype=xp.float32)
    c0 = xp.asarray([2.0], dtype=xp.float32)
    z1 = xp.asarray([0.5], dtype=xp.float32)
    sig = (x, c0, z1, precision)
    return fn, sig


def closure(fn, sig):
    # try:
    return fn(*sig)
    # except ValueError:
    #     pass


for kwarg_comb in kwargs:
    kwarg_measurement = measurement.get(kwarg_comb, {})
    measurement[kwarg_comb] = kwarg_measurement
    kwarg_comb = dict(kwarg_comb)
    for size in tqdm(input_sizes):
        size_measurement = kwarg_measurement.get(size, {})
        kwarg_measurement[size] = size_measurement
        for xp, cls, _id in modules:
            b, x = input_creation(xp, cls, (size,), **kwarg_comb)
            prof = benchmark(closure, (b, x), n_repeat=100)
            size_measurement[_id] = gather_time(prof)


lines = []
for kwarg_comb in kwargs:
    kwarg_line = ', '.join([f'{x}={y}' for x, y in kwarg_comb])
    lines.append(f'### `{kwarg_line}`\n')
    lines.append('| Size | CuPy (ms) | SciPy (ms) | Speedup |')
    lines.append('|:----:|:---------:|:----------:|:-------:|')
    kwarg_measurement = measurement[kwarg_comb]
    for size in input_sizes:
        comp = f'{size}'
        size_measurement = kwarg_measurement[size]
        times = []
        for _, _, _id in modules:
            time = size_measurement[_id]
            times.append(time)
            time_info = f'{time:3f}'
            comp = f'{comp} | {time_info}'
        speedup = times[1] / times[0]
        comp = f'{comp} | {speedup:3f}'
        lines.append(f'| {comp} |')
    lines.append('\n')

print('\n'.join(lines))

precision=1

Size	CuPy (ms)	SciPy (ms)	Speedup
5	2.344661	0.002708	0.001155
10	2.356515	0.002852	0.001210
100	2.360504	0.004143	0.001755
1000	2.421581	0.012302	0.005080
5000	2.387976	0.042065	0.017615
10000	2.418097	0.078274	0.032370
50000	2.394409	0.369082	0.154143
100000	2.417459	0.731418	0.302557
200000	2.501704	1.468732	0.587093
500000	2.781974	3.646715	1.310837
700000	2.806415	5.084350	1.811688
1000000	3.172863	7.267981	2.290670
2000000	5.308269	14.556299	2.742193
10000000	24.033003	85.798085	3.570011

[takagi]

Thanks, it's fine!