Nsly0204
    • Create new note
    • Create a note from template
      • Sharing URL Link copied
      • /edit
      • View mode
        • Edit mode
        • View mode
        • Book mode
        • Slide mode
        Edit mode View mode Book mode Slide mode
      • Customize slides
      • Note Permission
      • Read
        • Only me
        • Signed-in users
        • Everyone
        Only me Signed-in users Everyone
      • Write
        • Only me
        • Signed-in users
        • Everyone
        Only me Signed-in users Everyone
      • Engagement control Commenting, Suggest edit, Emoji Reply
    • Invite by email
      Invitee

      This note has no invitees

    • Publish Note

      Share your work with the world Congratulations! 🎉 Your note is out in the world Publish Note

      Your note will be visible on your profile and discoverable by anyone.
      Your note is now live.
      This note is visible on your profile and discoverable online.
      Everyone on the web can find and read all notes of this public team.
      See published notes
      Unpublish note
      Please check the box to agree to the Community Guidelines.
      View profile
    • Commenting
      Permission
      Disabled Forbidden Owners Signed-in users Everyone
    • Enable
    • Permission
      • Forbidden
      • Owners
      • Signed-in users
      • Everyone
    • Suggest edit
      Permission
      Disabled Forbidden Owners Signed-in users Everyone
    • Enable
    • Permission
      • Forbidden
      • Owners
      • Signed-in users
    • Emoji Reply
    • Enable
    • Versions and GitHub Sync
    • Note settings
    • Note Insights New
    • Engagement control
    • Transfer ownership
    • Delete this note
    • Save as template
    • Insert from template
    • Import from
      • Dropbox
      • Google Drive
      • Gist
      • Clipboard
    • Export to
      • Dropbox
      • Google Drive
      • Gist
    • Download
      • Markdown
      • HTML
      • Raw HTML
Menu Note settings Note Insights Versions and GitHub Sync Sharing URL Create Help
Create Create new note Create a note from template
Menu
Options
Engagement control Transfer ownership Delete this note
Import from
Dropbox Google Drive Gist Clipboard
Export to
Dropbox Google Drive Gist
Download
Markdown HTML Raw HTML
Back
Sharing URL Link copied
/edit
View mode
  • Edit mode
  • View mode
  • Book mode
  • Slide mode
Edit mode View mode Book mode Slide mode
Customize slides
Note Permission
Read
Only me
  • Only me
  • Signed-in users
  • Everyone
Only me Signed-in users Everyone
Write
Only me
  • Only me
  • Signed-in users
  • Everyone
Only me Signed-in users Everyone
Engagement control Commenting, Suggest edit, Emoji Reply
  • Invite by email
    Invitee

    This note has no invitees

  • Publish Note

    Share your work with the world Congratulations! 🎉 Your note is out in the world Publish Note

    Your note will be visible on your profile and discoverable by anyone.
    Your note is now live.
    This note is visible on your profile and discoverable online.
    Everyone on the web can find and read all notes of this public team.
    See published notes
    Unpublish note
    Please check the box to agree to the Community Guidelines.
    View profile
    Engagement control
    Commenting
    Permission
    Disabled Forbidden Owners Signed-in users Everyone
    Enable
    Permission
    • Forbidden
    • Owners
    • Signed-in users
    • Everyone
    Suggest edit
    Permission
    Disabled Forbidden Owners Signed-in users Everyone
    Enable
    Permission
    • Forbidden
    • Owners
    • Signed-in users
    Emoji Reply
    Enable
    Import from Dropbox Google Drive Gist Clipboard
       Owned this note    Owned this note      
    Published Linked with GitHub
    • Any changes
      Be notified of any changes
    • Mention me
      Be notified of mention me
    • Unsubscribe
    # 一對一問答紀錄:從整數平方根到浮點數開平方根。 ## 整數開平方根 在接下來的諸多演算法中,可以歸納為以下大綱: 1. 迭代起始值: 二元搜索法 `cls()` 、 查表法。 2. 演算法: Newton-Paphson method / Heron's method (以下簡稱牛頓法)、 digit-by-digit method (以下簡稱逐位逼近法)。 3. 透過追蹤迭代變數減少位移運算。 > Qs: why is branch free important? ### 牛頓法 以下將跟隨 [Python's integer square root algorithm](https://github.com/mdickinson/snippets/blob/main/papers/isqrt/isqrt.pdf) 逐步探討牛頓法整數開平方根的演進。 參閱 [從 √2 的存在談開平方根的快速運算](https://hackmd.io/@sysprog/sqrt?stext=18180%3A3%3A0%3A1747215575%3A9b9x2Q),Heron's method 保證: $$a^{(k+1)}= \frac{ a^{(k)} + \lfloor \frac{n}{a^{(k)}} \rfloor}{2} $$ 最終 $a^{(k+1)}$ 會收斂到 $\sqrt{n}$,於是有了以下實作。 ```c int mySqrt(int n) { if (n == 0) return 0; if (n < 4) return 1; unsigned int ans = n / 2; if (ans > 65535) // 65535 = 2^16 - 1 ans = 65535; while (ans * ans > n || (ans + 1) * (ans + 1) <= n) ans = (ans + n / ans) / 2; return ans; } ``` 加入用二元搜索法 `clz()` 取得初始值,和追蹤迭代變數 `d` (和 [2025 Quiz2-2](https://hackmd.io/@sysprog/linux2025-quiz2) 中的 `m` 有異曲同工之妙) ,有更快的實作如下: ```c uint64_t simple_newton_sqrt(uint64_t x) { int total_bits = 64; int shift = (total_bits - 1 - clz64(x)); uint64_t a = 1ULL << shift; uint64_t d; while (1) { d = x / a; if (a < d) return a; a = (a + d) << 1; } } ``` #### Listing 4: Adaptive-precision Heron’s method, iterative 文章中基於現有實作,發現每次迭代過程中準確的有效位數增加緩慢,提出以下方法予以優化: 與其直接計算 $\sqrt{n}$ ,改為計算 $\sqrt{\lfloor \frac{n}{j^2} \rfloor}$ ,且隨著迭代,整數 $j\rightarrow 0$,期待 $\sqrt{\lfloor \frac{n}{j^2} \rfloor}\rightarrow \sqrt{n}$。 進行證明之前: 1. 定義 **near square root** :若一個正整數 $a$ 是正整數 $n$ 的 near square root 則 $(a-1)^2 \lt n \lt (a+1)^2$。 2. 定義 $b=\lfloor \frac{n}{j^2} \rfloor$ 的 near square root,移項開根號可得 $jb\rightarrow \sqrt{n}$。 3. (Theorem 10) $a=kb+\lfloor \frac{n}{4kb} \rfloor$ 是 n 的 near square root ,若$4k^2 \lt n$ 綜合以上,由 2,不妨令 $j=2k$,$b=\lfloor \frac{n}{4k^2} \rfloor$ 的 near square root ,如果 $4k^2 \lt n$ , (Theorem 10) 保證 $a=kb+\lfloor \frac{n}{4kb} \rfloor$ 是 n 的 near square root。 由 Theorem 10, $a$ 同時是 n 的 near square root,,我們可以有更精密的估計: $$\lfloor \frac{jb+\lfloor \frac{n}{jb} \rfloor}{2} \rfloor \sim \sqrt{n}$$ 為求使用位元運算,令 $k=2^e$ ,根據 3. 的條件,我們可以得到最大的縮小倍率 $k$ : $$\begin{align} 4k^2 \le n &\Longleftrightarrow 2^{2+4e} \lt n \\ &\Longleftrightarrow 2+4e+1 \le length(n) \\ &\Longleftrightarrow e \le \lfloor \frac{length(n)-3}{4} \rfloor \end{align}$$ 定義 `recur_sqrt(x)` 演算法步驟: 1. 根據 $\lfloor \frac{n}{4k^2} \rfloor$ 計算最大 $k=2^e$ 當作估計的縮小倍率。 2. 因為$a=kb+\lfloor \frac{n}{4kb} \rfloor$ 是 n 的 near square root,計算有效位數較少的 `recur_sqrt(b)` (near square root of $\lfloor \frac{n}{4k^2} \rfloor$),取代 `recur_sqrt(a)` ,引入遞迴寫成 `a` = recur_sqrt($\lfloor \frac{n}{4k^2} \rfloor$)。 3. 回傳正確的近似 $a=kb+\lfloor \frac{n}{4kb} \rfloor$ (near square root of $n$) ```c uint64_t recur_sqrt(uint64_t n) { if (n < 4) return 1; uint64_t e = (64 - clz64(n) - 3) >> 2; // 4k^2 < n uint64_t a = recur_sqrt(n >> (e+2)); return (a << e) + (n >> (e+2)) / a; } uint64_t raw_adaptive_sqrt(uint64_t x) { uint64_t a = recur_sqrt(x); return (a*a <= x) ? a : a-1; } ``` 文中提供不使用遞迴的版本,具體來說做了以下代數代換: $$a=kb+\lfloor \frac{n}{4kb} \rfloor \rightarrow a=kb+\lfloor \frac{n}{2^{2(c-d)}kb} \rfloor$$ $$a = \text{near square root of} \,\,\lfloor \frac{n}{2^{2(c-d)}} \rfloor = \lfloor \frac{n}{4^{(c-d)}} \rfloor$$ $$k=2^e \rightarrow 2^{d-e-1}$$ 注意到會隨迭代改變的變數是 $a$、$d$、$s$ ,其中 $d=\lfloor \frac{c}{2^s} \rfloor$ ,$s$ 隨迭代次數遞減,代表最終 $d \rightarrow c$ , $k \rightarrow 2^{c-e-1}$。 又 $b$ = near square root of $a$,令 $m= \lfloor \frac{n}{4^{(c-d)}} \rfloor$,$a$ = near square root of $n$,於是我們有: $$a=kb+\lfloor \frac{n}{2^{2(c-d)}kb} \rfloor=2^{d-e-1}b+\lfloor \frac{m}{4(2^{d-e-1})b} \rfloor$$ $$a=2^{d-e-1}b+\lfloor \frac{m}{4(2^{d-e-1})b} \rfloor \rightarrow 2^{c-e-1}b+\lfloor \frac{m}{4(2^{c-e-1})b} \rfloor$$ 最後一式已然滿足上述 Theorem 10 的形式,最後證明縮倍比率 $k$ 的最大值為: $$\begin{align} 4k^2=4(4^{d-e-1})^2&=4\times 4^{d-e-1}\times 4^{d-e-1} \\ &\le 4\times 4^{d-e-1}\times 4^{e} \\ &= 4^d \end{align}$$ 至此我們成功證明 $a$ 同時是 $m$ 和 $n$ 的 near square root **注意到這裡沒有推倒過程,文章中直接嘗試證明以下實作的合理性**,不想看推導的可以跳過,根據程式碼定義: * $c=\lfloor\frac{log_2(n)-1}{2}\rfloor\le\lfloor \log_4(n)\rfloor$ (換底公式: $\log_{a^n}b=\log_{a}b^{\frac{1}{n}}$) * $s=\log_2c$ * $d=\lfloor \frac{c}{2^s} \rfloor$ * $e=\lfloor \frac{d}{2} \rfloor$ 利用 $Lemma\,\,12$ 得以證明 $4k^2 \le n$ , Theorem 10 成立,即所求 $a$ 是 $m$ 的 near square root。 $$ 4k^2 \le n \Longleftrightarrow 4(4^{d-e-1})^2 \le 4^d \lt m $$ ```c uint64_t adaptive_newton_sqrt(uint64_t x) { int total_bits = 64; uint64_t c = (total_bits - clz64(x) - 1) >> 1; uint64_t s = total_bits - clz64(c); uint64_t d = 0, a = 1, e; while (s > 0) { e = d; s--; d = c >> s; a = (a << (d-e-1)) + (x >> ((c<<1) - d - e + 1)) / a; } return (a*a <= x) ? a : (a-1); } ``` 值得一提的是 `d` 基本上沒有數學意義,一樣是我們為了減少運算所紀錄的迭代中間產物。 #### Listing 7: Fixed-precision variant, valid for $0 ≤ n < 2^{64}$ 從上一個實作可以發現迭代次數上限 $s=\lfloor\log_2\lfloor\log_2n\rfloor\rfloor$,以 64 位元無號整數來說,$c=31$,迭代次數必然是 $s=5$,由此我們如果固定美次迭代的縮小倍率 $k=2^{d-e-1}$ ,則我們可以展開迭代成以下形式,讀者可以自行帶入數值證明: ```c uint64_t fix_newton_sqrt(uint64_t x) { int total_bits = 64; /* notice that we want x to shift with interval 2^2 to make sure the */ uint64_t e = clz64(x) >> 1; // e = (32 − n.bit_length ()) / 2 uint64_t m = x << (e << 1); // make x > 2^62 /* for 2^62 < integer < 2^64 , fixed-percision Newton's method is * guarantee to converge with 5 iteration. */ uint64_t a = 1 + (m >> 62); // s=4 a = (a << 1) + (m >> 59) / a; // s=3 a = (a << 3) + (m >> 53) / a; // s=2 a = (a << 7) + (m >> 41) / a; // s=1 a = (a << 15) + (m >> 17) / a; // s=0 a = a >> e; // shift back return (a*a <= x) ? a : (a-1); } ``` **Listing 8: LUT + Fixed percision + Newton's Method** ```c /* isqrt32_tab[k] = isqrt(256*(k+64) - 1) for 0 <= k < 192 */ static const uint8_t isqrt32_tab[192] = { 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 143, 144, 145, 146, 147, 148, 149, 150, 150, 151, 152, 153, 154, 155, 155, 156, 157, 158, 159, 159, 160, 161, 162, 163, 163, 164, 165, 166, 167, 167, 168, 169, 170, 170, 171, 172, 173, 173, 174, 175, 175, 176, 177, 178, 178, 179, 180, 181, 181, 182, 183, 183, 184, 185, 185, 186, 187, 187, 188, 189, 189, 190, 191, 191, 192, 193, 193, 194, 195, 195, 196, 197, 197, 198, 199, 199, 200, 201, 201, 202, 203, 203, 204, 204, 205, 206, 206, 207, 207, 208, 209, 209, 210, 211, 211, 212, 212, 213, 214, 214, 215, 215, 216, 217, 217, 218, 218, 219, 220, 221, 221, 222, 222, 223, 223, 224, 225, 225, 226, 226, 227, 227, 228, 229, 229, 230, 230, 231, 231, 232, 232, 233, 234, 234, 235, 236, 236, 237, 237, 238, 238, 239, 239, 240, 241, 241, 242, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247, 247, 248, 248, 249, 249, 250, 250, 251, 251, 252, 252, 253, 253, 254, 254, 255, }; uint16_t table_sqrti(uint32_t x) { if (x == 0) return x; /* make sure x > 2^30 */ int shift = clz32(x) & 30; x <<= shift; /* get 2^14 < a < 2^16 from table */ printf("loop: %d\n", (int)((x >> 24) - 64)); uint16_t a = 1 + isqrt32_tab[(x >> 24) - 64]; /* Newton's method */ a = (a << 7) + (x >> 9) / a; /* near square correction */ if (x < a * a) a -= 1; return a >> (shift >> 1); } ``` 從 listing 7 可以直覺的想到,因為迭代的精度 (a 準確的位數) 是以固定的位數增長,換句話說,我們可以利用查表取代前幾次的牛頓法迭代運算,具體來說,根據 $Lemma\,13$ 對於一個數 $x,\,\,x > 2^{30}$ ,如果我們先透過查表找到 $t=\lfloor\frac{x}{2^{15}}\rfloor,\,\,2^{14}<t<2^{16}$ , lemma 保證迭代後的 $t$ 收斂到 $x$ 的 near sqaure root,由此可以發展出 listing 8 實作。 Look Up Table(LUT)製作方法單純是把 $t$ 預先做好 near square root ,注意到因為這邊是使用 8 位元大小存放LUT,所以在最後一項 $t=\lfloor2^{16}-1\rfloor$ 同時有 255 和 256 兩個 near sqaure root 時,避免使用需要 9 位元的 256。 ### Digit-by-digit Method (逐位逼近法) 詳見[測驗 2](https://hackmd.io/a_okd5kAS0CdzP22ZmHlVg?both=&stext=7584%3A4%3A0%3A1747209583%3ATugJuK)。 ```c uint64_t sqrti(uint64_t x) { uint64_t m, y = 0; if (x <= 1) return x; int total_bits = 64; /* clz64(x) returns the count of leading zeros in x. * (total_bits - 1 - clz64(x)) gives the index of the highest set bit. * Rounding that index down to an even number ensures our starting m is a * power of 4. */ int shift = (total_bits - 1 - clz64(x)) & ~1; m = 1ULL << shift; while (m) { uint64_t b = y + m; y >>= 1; if (x >= b) { x -= b; y += m; } m >>= 2; } return y; } ``` 至此我們完整了解牛頓法開平方根,而[整數開平方根 Quiz2-2](https://hackmd.io/@DboOgKS6RtOmMLCltFsBig/SkS4ZDUeeg?fbclid=IwY2xjawKQ_D5leHRuA2FlbQIxMQBicmlkETFHNlRUTW9jQ3c1MXBRYVVvAR5REqHnVCBKAoORuVxYHyyf4ECOZrZiMQcmHoZtCDuGht2VY_eC5UEqU5WerQ_aem_ZvO-9XJyk4KNS977HM_StQ&stext=2%3A15%3A0%3A1747230693%3AWpynbZ)中提到,牛頓法相比逐位逼近法迭代次數更少,運算更快,然而 linux 核心卻仍採用逐位逼近法。 ## 浮點數開平方根 ### IEEE754 ![雙精度浮點數](https://hackmd.io/_uploads/BkJ143Wbge.png) 引述自維基百科,[IEEE754](https://en.wikipedia.org/wiki/IEEE_754#Rounding_rules) 中提供四種近似到最接近整數的 `round()` 方法: * 捨入到最接近:修約到最接近,在一樣接近的情況下偶數優先(Ties To Even,這是預設的修約方式):會將結果修約為最接近且可以表示的值,但是當存在兩個數一樣接近的時候,則取其中的偶數(在二進位中是以0結尾的)。 * 朝+∞方向捨入:會將結果朝正無限大的方向捨入。 * 朝-∞方向捨入:會將結果朝負無限大的方向捨入。 * 朝0方向捨入:會將結果朝0的方向捨入。 相較於無條件進位和無條件捨去,通常默認的修約方法是最接近四捨五入的 **RoundTiesToEven**,以下提供精度到小數點後 8 位元的範例,我們假設換算成整數只需要小數點後三位,即最尾端有四位元需要被近似後捨棄。 ``` LSB 的一半: 1000 # 正常情況 大於 1000 進位,小於捨棄 對於 1.001_1001,round 後為 1.010。 對於 1.001_0111,round 後為 1.001。 # 正好一半: round to nearest even 對於 1.001_1000,經舍入處理後為 1.010。 對於 -1.010_1000,經舍入處理後為 -1.010。 ``` 單精度浮點數之非正規數例外情況: ![image](https://hackmd.io/_uploads/SJC9rIo-gx.png) 結論來說,本次單精度浮點數開根號實作需要處理以下例外情況: * 對正負 0 開根號 $\rightarrow$ 回傳正負 0。 * 對負數開根號,根據 IEEE 754 ,屬於數學尚未定義數,回傳 qNaN 。 * 對太小的數根號使發生 Underflow ,回傳小於等於原數的值即可(不需處理)。 * 對正無窮大開根號 $\rightarrow$ 回傳正無窮大。 * 對 NaN 開根號 $\rightarrow$ 回傳 NaN 。 ### 為何需要避免使用 FPU 在作業三 [Linux 核心的浮點數運算](https://hackmd.io/@sysprog/linux2025-kxo/%2F%40sysprog%2Flinux2025-kxo-a%3Fstext%3D3060%253A14%253A0%253A1748433549%253A9UogU2)一文中提到: linux kernel 若要使用 FPU 通常需要觸發 trap(exception) 並切換至 floating point mode ,而這個 trap 的成本並不小,核心通常沒有餘裕負擔,所以幾乎所有的核心程式碼都使用定點數運算。 這也代表大部分的時間 FPU context 的內容與 user context 正在進行的 process 不吻合,若有需要要使用 FPU 的情況,需要自行操作對 FPU state 的存取,並且避免 context switch。建議盡量使用定點數運算。 舉例來說,有以下可能的問題 * Lazy FP state leak: * Lazy FP state 因亂序執行失效 * 並非所有硬體都具備 FPU ### 牛頓法 ```c typedef union { float value; uint32_t v_int; } ieee_float_shape_type; /* Get a 32 bit int from a float. */ #define EXTRACT_WORDS(ix0, d) \ do { \ ieee_float_shape_type ew_u; \ ew_u.value = (d); \ (ix0) = ew_u.v_int; \ } while (0) float mySqrt(float n) { uint32_t sign_mask = 0x80000000; uint32_t exp_mask = 0x7F800000; uint32_t frac_mask = 0x007fffff; int32_t ix0, exponent, i; uint32_t mentissa; float result = 0.0, x = n, d; EXTRACT_WORDS(ix0, x); /* take care of zero */ if (ix0 <= 0) { if ((ix0 & (~sign_mask)) == 0) return x; /* sqrt(+-0) = +-0 */ if (ix0 < 0) return (x - x) / (x - x); /* sqrt(-ve) = sNaN */ } /* take care of positive infinity */ exponent = ix0 >> 23; mentissa = ix0 & frac_mask; if (exponent == 0) { if ((ix0 & frac_mask) == 0) return x; } /* take care of NaNs */ else if (exponent == 0xFF) { if (mentissa != 0) return x; } printf("1 exponent: %d, ix0: %X \n", exponent,ix0); printf("x: %f, n: %f \n", x,n); for (i = 0; i < 10; i++) { d = n / x; x = (x + d) / 2; printf("x: %f, n: %f \n", x,n); } return x; } ``` 參考以上程式碼,若我們直接用牛頓法實作,相比於 `sqrti` 我們需要考慮收斂的問題,且在輸入小於 1 時特別明顯,以輸入 `1e-19` 為例,可以發現 10 次迭代依然無法收斂。 ``` 1 exponent: 63, ix0: 1FEC1E4A x: 0.000000, n: 0.000000 x: 0.500000, n: 0.000000 x: 0.250000, n: 0.000000 x: 0.125000, n: 0.000000 x: 0.062500, n: 0.000000 x: 0.031250, n: 0.000000 x: 0.015625, n: 0.000000 x: 0.007812, n: 0.000000 x: 0.003906, n: 0.000000 x: 0.001953, n: 0.000000 x: 0.000977, n: 0.000000 correct: 0.000000, my: 0.000977 ``` 顯然,現有程式仍有很大缺陷,我們當然可以透過尋找更好的初始值解決迭代次數的問題,可參考本文最後用倒數平方根做猜想的牛頓法實現。 ### Digit-by-digit Method ```c p = s = 0; /* q = sqrt(x) */ r = 0x01000000; /* r = moving bit from right to left */ while(r!=0) { if(s+r <= ix0) { ix0 -= (s+r); s += 2*r; q += r; } ix0 += ix0; r >>= 1; } ``` 參考[直式開方的步驟原理說明](https://www.youtube.com/watch?v=iQYGKId1c3U)。 ![image](https://hackmd.io/_uploads/S17RGnnblx.png) * $x_{i}$ 是待開平方數,對應圖中 6241。 * $q_{i}^2$ 是已決定的平方根,對應圖中 70。 * $q_{i+1}=q$`y` 代表加入下一位後的商,對應圖中直式除法的商 `7y`。 要求下一位平方根,可以總結成以下步驟: 1. 增加兩位,計算當前誤差值 $13(41)=x_{i}^2-q_{i}^2$ 2. 決定下一位元 `y`,若 $(2\times70+y)\times y > 當前差值$ 則商 $q_{i+1}$ 填入對應的 `y`。 **改寫至二進制** 觀察以下二進制小數平方: $$\begin{align} (1.1)^2 &= (1+\frac{1}{2})^2=\frac{9}{4}= 2.25 \\ (1.101)^2 &= (1+\frac{1}{2}+\frac{1}{8})^2=\frac{169}{64}\approx 2.6406 \\ (1.11)^2 &= (1+\frac{1}{2}+\frac{1}{4})^2=\frac{49}{16}\approx 3.0625 \end{align}$$ 可以發現多增加一位小數點的平方 $(1.101_2)^2$ 必定被包在 $(1.10_2)^2$ 和 $(1.11_2)^2$ 之間,故和十進位不同,我們算誤差值只需要一次多觀察一位,也就是說第 i+1 次的迭代誤差值 `x` 可以寫成: $$\begin{align} x^{(i+1)} &= x^{(i)}\times2 - (2q+b_{i+1}\times2 )b_{i+1}\times2 \\ &= x^{(i)}\times2 - (2q+2^{-(i)}\times2 )2^{-(i)}\times2 \end{align}$$ 其中 $x^{(i+1)}$ 代表 $q$ 目前計算到第 i+1 位,也可以說是小數點後第 i 位,對應 $b_{i+1}=2^{-(i)}$。 注意到多觀察一位隱含同時把剩下的誤差和原本的商右移,可以參照十進制時,誤差 $13.41\,\, ⭢\,\, 13.41\times10^2$ 和商 $7\,\, ⭢\,\, 70$ ,對應 $x^{(i)}\times2$ 和 $b_{i+1} \times2$。 由此,我們可以得到==不需要 FPU== 的實作。 ```c float DBD_sqrtf(float n) { float x = n, z; uint32_t sign_mask = 0x80000000; uint32_t frac_mask = 0x007fffff; int32_t ix0,s,q,m,t,i; uint32_t r; EXTRACT_WORDS(ix0,x); /* take care of zero */ if (ix0 <= 0) { if ((ix0 & (~sign_mask)) == 0) return x; /* sqrt(+-0) = +-0 */ if (ix0 < 0) return (x - x) / (x - x); /* sqrt(-ve) = sNaN */ } /* normalize x */ m = (ix0 >> 23); if (m == 0) { /* subnormal x */ for (i = 0; (ix0 & 0x00800000) == 0; i++) ix0 <<= 1; m -= i - 1; } m -= 127; /* unbias exponent */ /* retrieve correct mentissa 1.XXX */ ix0 = (ix0 & frac_mask) | 0x00800000; if (m & 1) /* odd m, double x to make it even */ ix0 += ix0; m >>= 1; /* (exponent)m = [m/2] */ /* generate sqrt(x) digit by digit */ ix0 += ix0; q = s = 0; /* q = sqrt(x) */ r = 0x01000000; /* r = moving bit from right to left */ while(r!=0) { t = s+r; if(t <= ix0) { // if reamin > add + r s = t+r; // add + r ix0 -= t; // remain - add q += r; // add one bit } ix0 += ix0; r >>= 1; } /* use floating add to find out rounding direction */ if(ix0!=0) { z = one-tiny; /* trigger inexact flag */ if (z>=one) { z = one+tiny; if (z>one) q += 2; else q += (q & 1); } } ix0 = (q >> 1) + 0x3f000000; ix0 += (m << 23); INSERT_WORDS(z,ix0); return z; } ``` ### 倒數平方根 + 牛頓法 這部份 [從 √2 的存在談開平方根的快速運算](https://hackmd.io/@sysprog/sqrt?stext=18180%3A3%3A0%3A1747215575%3A9b9x2Q) 已有介紹 `Q_rsqrt(x)`,實作上也很直接,即先用倒數平方根做初始值的猜測 $$\sqrt{N} = \text{Q_rsqrt(x)} \times N$$在加入牛頓法提高精度 >TODO: 比較有無牛頓法的差異 ### TODO: 比較 Digit-by-digit 和 Inverse Square Root 實現 `sqrtf()` 之效能差異 ### 參考資料 * [glibc/math_private.h](https://codebrowser.dev/glibc/glibc/sysdeps/generic/math_private.h.html) * [e_sqrtf.c](https://github.com/openbsd/src/blob/dd79ebd3e2ae7cad5e9ae8d576005e34146143b2/lib/libm/src/e_sqrtf.c#L23) * [/ARM/sqrtf.c](https://github.com/simonbyrne/apple-libm/blob/master/Source/ARM/sqrtf.c)

    Import from clipboard

    Paste your markdown or webpage here...

    Advanced permission required

    Your current role can only read. Ask the system administrator to acquire write and comment permission.

    This team is disabled

    Sorry, this team is disabled. You can't edit this note.

    This note is locked

    Sorry, only owner can edit this note.

    Reach the limit

    Sorry, you've reached the max length this note can be.
    Please reduce the content or divide it to more notes, thank you!

    Import from Gist

    Import from Snippet

    or

    Export to Snippet

    Are you sure?

    Do you really want to delete this note?
    All users will lose their connection.

    Create a note from template

    Create a note from template

    Oops...
    This template has been removed or transferred.
    Upgrade
    All
    • All
    • Team
    No template.

    Create a template

    Upgrade

    Delete template

    Do you really want to delete this template?
    Turn this template into a regular note and keep its content, versions, and comments.

    This page need refresh

    You have an incompatible client version.
    Refresh to update.
    New version available!
    See releases notes here
    Refresh to enjoy new features.
    Your user state has changed.
    Refresh to load new user state.

    Sign in

    Forgot password

    or

    By clicking below, you agree to our terms of service.

    Sign in via Facebook Sign in via Twitter Sign in via GitHub Sign in via Dropbox Sign in with Wallet
    Wallet ( )
    Connect another wallet

    New to HackMD? Sign up

    Help

    • English
    • 中文
    • Français
    • Deutsch
    • 日本語
    • Español
    • Català
    • Ελληνικά
    • Português
    • italiano
    • Türkçe
    • Русский
    • Nederlands
    • hrvatski jezik
    • język polski
    • Українська
    • हिन्दी
    • svenska
    • Esperanto
    • dansk

    Documents

    Help & Tutorial

    How to use Book mode

    Slide Example

    API Docs

    Edit in VSCode

    Install browser extension

    Contacts

    Feedback

    Discord

    Send us email

    Resources

    Releases

    Pricing

    Blog

    Policy

    Terms

    Privacy

    Cheatsheet

    Syntax Example Reference
    # Header Header 基本排版
    - Unordered List
    • Unordered List
    1. Ordered List
    1. Ordered List
    - [ ] Todo List
    • Todo List
    > Blockquote
    Blockquote
    **Bold font** Bold font
    *Italics font* Italics font
    ~~Strikethrough~~ Strikethrough
    19^th^ 19th
    H~2~O H2O
    ++Inserted text++ Inserted text
    ==Marked text== Marked text
    [link text](https:// "title") Link
    ![image alt](https:// "title") Image
    `Code` Code 在筆記中貼入程式碼
    ```javascript
    var i = 0;
    ```
    var i = 0;
    :smile: :smile: Emoji list
    {%youtube youtube_id %} Externals
    $L^aT_eX$ LaTeX
    :::info
    This is a alert area.
    :::

    This is a alert area.

    Versions and GitHub Sync
    Get Full History Access

    • Edit version name
    • Delete

    revision author avatar     named on  

    More Less

    Note content is identical to the latest version.
    Compare
      Choose a version
      No search result
      Version not found
    Sign in to link this note to GitHub
    Learn more
    This note is not linked with GitHub
     

    Feedback

    Submission failed, please try again

    Thanks for your support.

    On a scale of 0-10, how likely is it that you would recommend HackMD to your friends, family or business associates?

    Please give us some advice and help us improve HackMD.

     

    Thanks for your feedback

    Remove version name

    Do you want to remove this version name and description?

    Transfer ownership

    Transfer to
      Warning: is a public team. If you transfer note to this team, everyone on the web can find and read this note.

        Link with GitHub

        Please authorize HackMD on GitHub
        • Please sign in to GitHub and install the HackMD app on your GitHub repo.
        • HackMD links with GitHub through a GitHub App. You can choose which repo to install our App.
        Learn more  Sign in to GitHub

        Push the note to GitHub Push to GitHub Pull a file from GitHub

          Authorize again
         

        Choose which file to push to

        Select repo
        Refresh Authorize more repos
        Select branch
        Select file
        Select branch
        Choose version(s) to push
        • Save a new version and push
        • Choose from existing versions
        Include title and tags
        Available push count

        Pull from GitHub

         
        File from GitHub
        File from HackMD

        GitHub Link Settings

        File linked

        Linked by
        File path
        Last synced branch
        Available push count

        Danger Zone

        Unlink
        You will no longer receive notification when GitHub file changes after unlink.

        Syncing

        Push failed

        Push successfully