Numbaの高速ぶりには感謝しております。

下記サイトでもNumba使われてますが、Numbaが断然速い。本PCでもJavaScriptコードでは約3秒。Numbaは掲載記事通りのパラなら3.5秒程度。fastmathやparallelパラ付けで0.7秒位。断然速い。
https://itnext.io/compare-c-js-python-python-numba-php7-php8-and-golang-in-prime-number-calculation-55e82b6f82a9

PythonのコードをよりNumbaへ置き換えたいのですがDictがあると高速化はできていません。
わかる方がいらしたら教えて頂けると幸いです。

キーは数字/Numba有り

NoPythonモードでビルドできます

import numpy as np
from numba import njit
import time
start = time.time()


@njit(cache=True, parallel=True, fastmath=True)
def dict_try4nb():
    d = dict()
    k = {1: np.arange(1), 2: np.arange(2)}
    # The following tells the compiler what the key type and the
    # value
    # type are for `d`.
    d[3] = np.arange(3)
    d[5] = np.arange(5)
    return d, k


d, k = dict_try4nb()
print(d)    # {3: [0 1 2], 5: [0 1 2 3 4]}
print(k)    # {1: [0], 2: [0 1]}
print(f"Cost: {time.time()-start:.1f}sec")

結果

Cost: 14.6sec # Cache前なので遅い!?
Cost: 4.3sec
Cost: 4.1sec

キーは数字/Numba無し

import numpy as np
from numba import njit
import time
start = time.time()


def dict_try4nb():
    d = dict()
    k = {1: np.arange(1), 2: np.arange(2)}
    # The following tells the compiler what the key type and the
    # value
    # type are for `d`.
    d[3] = np.arange(3)
    d[5] = np.arange(5)
    return d, k


d, k = dict_try4nb()
print(d)    # {3: [0 1 2], 5: [0 1 2 3 4]}
print(k)    # {1: [0], 2: [0 1]}
print(f"Cost: {time.time()-start:.1f}sec")

結果

Cost: 0.0sec
Cost: 0.0sec
Cost: 0.0sec

キーは文字列/Numba有り

NoPythonモードはエラー生じます。解決策ご存知でしたら教えて頂けるとたすかります。


from random import randint
import numpy as np
from numba import njit, prange, jit
import time


@jit(cache=True, parallel=True, forceobj=True)
def dict_str_try4nb(n):
    start = time.time()
    d = dict()
    k = {1: np.arange(1), 2: np.arange(2)}
    # The following tells the compiler what the key type and the
    # value
    # type are for `d`.
    d[3] = np.arange(3)
    d[5] = np.arange(5)
    for i in prange(n):
        # d[i]=np.arange(10)
        d[f"テスト{i}"] = randint(1, 1000)
    return start, d, k


start, d, k = dict_str_try4nb(100_000)
# print(d)    # {3: [0 1 2], 5: [0 1 2 3 4]}
# print(k)    # {1: [0], 2: [0 1]}
print(f"Cost: {time.time()-start:.1f}sec")

結果
徐々に早くなっている？

Cost: 2.1sec
Cost: 1.7sec
Cost: 1.5sec

キーは文字列/Numba無し

from random import randint
import numpy as np
from numba import njit, prange, jit
import time


def dict_str_try4nb(n):
    start = time.time()
    d = dict()
    k = {1: np.arange(1), 2: np.arange(2)}
    # The following tells the compiler what the key type and the
    # value
    # type are for `d`.
    d[3] = np.arange(3)
    d[5] = np.arange(5)
    for i in prange(n):
        # d[i]=np.arange(10)
        d[f"テスト{i}"] = randint(1, 1000)
    return start, d, k


start, d, k = dict_str_try4nb(100_000)
# print(d)    # {3: [0 1 2], 5: [0 1 2 3 4]}
# print(k)    # {1: [0], 2: [0 1]}
print(f"Cost: {time.time()-start:.1f}sec")

結果

Cost: 1.0sec
Cost: 1.1sec
Cost: 1.0sec

np.arrangeをprimitive型値へ/キーは数値

Normal(Numba無し)

from random import randint
import numpy as np
from numba import njit, prange, jit
import time
start = time.time()

def dict_str_try4nb(n):

    d = dict()
    k = {1: 1, 2: 2}
    # The following tells the compiler what the key type and the
    # value
    # type are for `d`.
    d[3] = 3
    d[5] = 5
    for i in prange(n):
        # d[i]=np.arange(10)
        d[1] = randint(1, 1000)
    return start, d, k


start, d, k = dict_str_try4nb(100_000)
# print(d)    # {3: [0 1 2], 5: [0 1 2 3 4]}
# print(k)    # {1: [0], 2: [0 1]}
print(f"Cost: {time.time()-start:.1f}sec")

結果

Cost: 0.1sec
Cost: 0.1sec
Cost: 0.1sec

Numba有り

Cacheやfastmathパラを印加して最適化

from random import randint
import numpy as np
from numba import njit, prange, jit
import time

start = time.time()


@njit(cache=True, parallel=True, fastmath=True)
def dict_str_try4nb(n):
    d = dict()
    k = {1: 1, 2: 2}
    # The following tells the compiler what the key type and the
    # value
    # type are for `d`.
    d[3] = 3
    d[5] = 5
    for i in prange(n):
        # d[i]=np.arange(10)
        d[1] = randint(1, 1000)
    return d, k


d, k = dict_str_try4nb(100_000)
# print(d)    # {3: [0 1 2], 5: [0 1 2 3 4]}
# print(k)    # {1: [0], 2: [0 1]}
print(f"Cost: {time.time()-start:.1f}sec")

結果
キーを数値でもNumba無しより遅い。

Cost: 0.4sec
Cost: 0.2sec
Cost: 0.2sec