SSE整数除法

SSE integer division?

本文关键字:除法 整数 SSE      更新时间:2023-10-16

有_mm_div_ps用于浮点数除法,有_mm_mullo_epi16用于整数乘法。但是有整数除法(16位值)吗?我怎样才能进行这样的分割?

数学告诉我们确实有可能跑得更快

Agner Fog的(http://www.agner.org/optimize/#vectorclass)方法在使用单个除法时效果很好。此外,如果除数在编译时已知,或者在运行时不经常更改,则该方法还有更多的好处。

然而,当对__m128i元素执行SIMD除法时,在编译时除数和被除数都没有可用的信息,我们别无选择,只能转换为浮点数并执行计算。另一方面,使用_mm_div_ps不会给我们带来惊人的速度提升,因为这个指令在大多数微架构上有11到14个周期的可变延迟,如果我们考虑骑士登陆,有时可以达到38个周期。最重要的是,该指令不是完全流水线化的,并且根据微架构具有3-6个周期的倒数吞吐量。

但是我们可以避免使用_mm_div_ps而使用_mm_rcp_ss

不幸的是,__m128 _mm_rcp_ss (__m128 a)之所以快,只是因为它提供了近似。即(取自英特尔内部指南):

计算a中低单精度(32位)浮点元素的近似倒数,将结果存储在dst的低元素中,并将a中高3个打包的元素复制到dst的高元素中。此近似的最大相对误差小于1.5*2^-12。

因此,为了从_mm_rcp_ss中获益,我们需要补偿由于近似而造成的损失。在这个方向上有一个伟大的工作,可以在Niels Möller和Torbjörn Granlund的改进不变整数除法中找到:

由于当前处理器缺乏有效的除法指令,因此除法是使用预先计算的除数倒数的单字近似值作为乘法来执行的,然后是几个调整步骤。

为了计算16位有符号整数除法,我们只需要一个调整步骤,并据此建立我们的解。

<标题>SSE2 h1> 解决方案只能使用SSE2,如果可用,将使用FMA

。然而,使用普通除法可能和使用近似法一样快(甚至更快)。在AVX存在的情况下,该解决方案可以得到改进,因为使用一个AVX寄存器可以同时处理高电平和低电平。

由于我们只处理16位,我们可以使用暴力测试在几秒钟内轻松验证解决方案的正确性:

void print_epi16(__m128i a)
{
int i; int16_t tmp[8];
_mm_storeu_si128( (__m128i*) tmp, a);
for (i = 0; i < 8; i += 1) {
printf("%8d ", (int) tmp[i]);
}
printf("n");
}
bool run_mm_div_epi16(const int16_t *a, const int16_t *b)
{
const size_t n = 8;
int16_t result_expected[n];
int16_t result_obtained[n];
//
// Derive the expected result
//
for (size_t i = 0; i < n; i += 1) {
result_expected[i] = a[i] / b[i];
}
//
// Now perform the computation
//
const __m128i va = _mm_loadu_si128((__m128i *) a);
const __m128i vb = _mm_loadu_si128((__m128i *) b);
const __m128i vr = _mm_div_epi16(va, vb);
_mm_storeu_si128((__m128i *) result_obtained, vr);
//
// Check for array equality
//
bool eq = std::equal(result_obtained, result_obtained + n, result_expected);
if (!eq) {
cout << "Testing of _mm_div_epi16 failed" << endl << endl;
cout << "a: ";
print_epi16(va);
cout << "b: ";
print_epi16(vb);
cout << endl;
cout << "results_obtained: ";
print_epi16(vr);
cout << "results_expected: ";
print_epi16(_mm_loadu_si128((__m128i *) result_expected));
cout << endl;
}
return eq;
}
void test_mm_div_epi16()
{
const int n = 8;
bool correct = true;
//
// Brute-force testing
//
int16_t a[n];
int16_t b[n];
for (int32_t i = INT16_MIN; correct && i <= INT16_MAX; i += n) {
for (int32_t j = 0; j < n; j += 1) {
a[j] = (int16_t) (i + j);
}
for (int32_t j = INT16_MIN; correct && j < 0; j += 1) {
const __m128i jv = _mm_set1_epi16((int16_t) j);
_mm_storeu_si128((__m128i *) b, jv);
correct = correct && run_mm_div_epi16(a, b);
}
for (int32_t j = 1; correct && j <= INT16_MAX; j += 1) {
const __m128i jv = _mm_set1_epi16((int16_t) j);
_mm_storeu_si128((__m128i *) b, jv);
correct = correct && run_mm_div_epi16(a, b);
}
}
if (correct) {
cout << "Done!" << endl;
} else {
cout << "_mm_div_epi16 can not be validated" << endl;
}
}
<标题>AVX2 h1>
static inline __m256i _mm256_div_epi16(const __m256i &a_epi16, const __m256i &b_epi16) {
//
// Setup the constants.
//
const __m256 two = _mm256_set1_ps(2.00000051757f);
//
// Convert to two 32-bit integers
//
const __m256i a_hi_epi32       = _mm256_srai_epi32(a_epi16, 16);
const __m256i a_lo_epi32_shift = _mm256_slli_epi32(a_epi16, 16);
const __m256i a_lo_epi32       = _mm256_srai_epi32(a_lo_epi32_shift, 16);
const __m256i b_hi_epi32       = _mm256_srai_epi32(b_epi16, 16);
const __m256i b_lo_epi32_shift = _mm256_slli_epi32(b_epi16, 16);
const __m256i b_lo_epi32       = _mm256_srai_epi32(b_lo_epi32_shift, 16);
//
// Convert to 32-bit floats
//
const __m256 a_hi = _mm256_cvtepi32_ps(a_hi_epi32);
const __m256 a_lo = _mm256_cvtepi32_ps(a_lo_epi32);
const __m256 b_hi = _mm256_cvtepi32_ps(b_hi_epi32);
const __m256 b_lo = _mm256_cvtepi32_ps(b_lo_epi32);
//
// Calculate the reciprocal
//
const __m256 b_hi_rcp = _mm256_rcp_ps(b_hi);
const __m256 b_lo_rcp = _mm256_rcp_ps(b_lo);
//
// Calculate the inverse
//
const __m256 b_hi_inv_1 = _mm256_fnmadd_ps(b_hi_rcp, b_hi, two);
const __m256 b_lo_inv_1 = _mm256_fnmadd_ps(b_lo_rcp, b_lo, two);
//
// Compensate for the loss
//
const __m256 b_hi_rcp_1 = _mm256_mul_ps(b_hi_rcp, b_hi_inv_1);
const __m256 b_lo_rcp_1 = _mm256_mul_ps(b_lo_rcp, b_lo_inv_1);
//
// Perform the division by multiplication
//
const __m256 hi = _mm256_mul_ps(a_hi, b_hi_rcp_1);
const __m256 lo = _mm256_mul_ps(a_lo, b_lo_rcp_1);
//
// Convert back to integers
//
const __m256i hi_epi32 = _mm256_cvttps_epi32(hi);
const __m256i lo_epi32 = _mm256_cvttps_epi32(lo);
//
// Blend the low and the high-parts
//
const __m256i hi_epi32_shift = _mm256_slli_epi32(hi_epi32, 16);
return _mm256_blend_epi16(lo_epi32, hi_epi32_shift, 0xAA);
}

可以使用上面描述的相同方法来执行代码验证。

<标题>

性能我们可以使用度量每周期(F/C)来评估性能。在这种情况下,我们希望看到每个循环可以执行多少除法。为此,我们定义了两个向量ab,并执行逐点除法。ab都使用xorshift32填充随机数据,初始化为uint32_t state = 3853970173;

我使用RDTSC来测量周期,使用热缓存执行15次重复,并使用中位数作为结果。为了避免频率缩放和资源共享对测量的影响,禁用了Turbo Boost和Hyper-Threading。为了运行代码,我使用Intel Xeon CPU E3-1285L v33.10GHz Haswell, 32GB RAM和25.6 GB/s带宽作为主存,运行Debian GNU/Linux 8 (jessie),kernel 3.16.43-2+deb8u3gcc使用的是4.9.2-10。结果如下:

SSE2实现

我们将普通除法与上述算法进行比较:

===============================================================
= Compiler & System info
===============================================================
Current CPU          : Intel(R) Xeon(R) CPU E3-1285L v3 @ 3.10GHz
CXX Compiler ID      : GNU
CXX Compiler Path    : /usr/bin/c++
CXX Compiler Version : 4.9.2
CXX Compiler Flags   : -O3 -std=c++11 -msse2 -mno-fma
--------------------------------------------------------------------------------
|   Size    | Division F/C |  Division B/W   | Approx. F/C | Approximation B/W |
--------------------------------------------------------------------------------
|       128 |  0.5714286   |   26911.45 MB/s |  0.5019608  |   23634.21 MB/s   |
|       256 |  0.5714286   |   26909.17 MB/s |  0.5039370  |   23745.44 MB/s   |
|       512 |  0.5707915   |   26928.14 MB/s |  0.5039370  |   23763.79 MB/s   |
|      1024 |  0.5707915   |   26936.33 MB/s |  0.5039370  |   23776.85 MB/s   |
|      2048 |  0.5709507   |   26938.51 MB/s |  0.5039370  |   23780.25 MB/s   |
|      4096 |  0.5708711   |   26940.56 MB/s |  0.5039990  |   23782.65 MB/s   |
|      8192 |  0.5708711   |   26940.16 MB/s |  0.5039370  |   23781.85 MB/s   |
|     16384 |  0.5704735   |   26921.76 MB/s |  0.4954040  |   23379.24 MB/s   |
|     32768 |  0.5704537   |   26921.26 MB/s |  0.4954639  |   23382.13 MB/s   |
|     65536 |  0.5703147   |   26914.53 MB/s |  0.4943539  |   23330.13 MB/s   |
|    131072 |  0.5691680   |   26860.21 MB/s |  0.4929539  |   23264.40 MB/s   |
|    262144 |  0.5690618   |   26855.60 MB/s |  0.4929187  |   23262.22 MB/s   |
|    524288 |  0.5691378   |   26858.75 MB/s |  0.4929488  |   23263.56 MB/s   |
|   1048576 |  0.5677474   |   26794.14 MB/s |  0.4918968  |   23214.34 MB/s   |
|   2097152 |  0.5371243   |   25348.39 MB/s |  0.4700511  |   22183.07 MB/s   |
|   4194304 |  0.5128146   |   24200.82 MB/s |  0.4529809  |   21377.28 MB/s   |
|   8388608 |  0.5036971   |   23770.36 MB/s |  0.4438345  |   20945.84 MB/s   |
|  16777216 |  0.5005390   |   23621.14 MB/s |  0.4409909  |   20811.32 MB/s   |
|  33554432 |  0.4992792   |   23561.90 MB/s |  0.4399777  |   20763.49 MB/s   |
--------------------------------------------------------------------------------

我们可以观察到,普通除法将比建议的近似步骤略快。在这种情况下,我们可以得出结论,在Haswell微架构上使用SSE2近似将是次优的。

然而,如果我们在旧的Sandy Bridge机器上运行相同的结果,例如Intel Xeon(R) CPU X5680 @ 3.33GHz,我们已经可以看到近似的好处:

===============================================================
= Compiler & System info
===============================================================
Current CPU          : Intel(R) Xeon(R) CPU X5680  @ 3.33GHz
CXX Compiler ID      : GNU
CXX Compiler Path    : /usr/bin/c++
CXX Compiler Version : 4.8.5
CXX Compiler Flags   : -O3 -std=c++11 -msse2 -mno-fma
--------------------------------------------------------------------------------
|   Size    | Division F/C |  Division B/W   | Approx. F/C | Approximation B/W |
--------------------------------------------------------------------------------
|       128 |  0.2857143   |   14511.41 MB/s |  0.3720930  |   18899.89 MB/s   |
|       256 |  0.2853958   |   14512.51 MB/s |  0.3715530  |   18898.91 MB/s   |
|       512 |  0.2853958   |   14510.53 MB/s |  0.3715530  |   18896.44 MB/s   |
|      1024 |  0.2853162   |   14511.81 MB/s |  0.3700759  |   18824.00 MB/s   |
|      2048 |  0.2853162   |   14511.04 MB/s |  0.3708130  |   18860.31 MB/s   |
|      4096 |  0.2852964   |   14511.16 MB/s |  0.3711826  |   18879.27 MB/s   |
|      8192 |  0.2852666   |   14510.23 MB/s |  0.3713172  |   18886.39 MB/s   |
|     16384 |  0.2852616   |   14509.86 MB/s |  0.3712920  |   18885.60 MB/s   |
|     32768 |  0.2852244   |   14507.93 MB/s |  0.3712709  |   18884.86 MB/s   |
|     65536 |  0.2851003   |   14501.41 MB/s |  0.3701114  |   18826.14 MB/s   |
|    131072 |  0.2850711   |   14499.95 MB/s |  0.3685017  |   18743.58 MB/s   |
|    262144 |  0.2850745   |   14500.47 MB/s |  0.3684799  |   18742.78 MB/s   |
|    524288 |  0.2848062   |   14486.66 MB/s |  0.3681040  |   18723.63 MB/s   |
|   1048576 |  0.2846679   |   14479.64 MB/s |  0.3671284  |   18674.02 MB/s   |
|   2097152 |  0.2840133   |   14446.52 MB/s |  0.3664623  |   18640.01 MB/s   |
|   4194304 |  0.2745241   |   13963.13 MB/s |  0.3488823  |   17745.24 MB/s   |
|   8388608 |  0.2741900   |   13946.39 MB/s |  0.3476036  |   17680.37 MB/s   |
|  16777216 |  0.2740689   |   13940.32 MB/s |  0.3477076  |   17685.97 MB/s   |
|  33554432 |  0.2746752   |   13970.75 MB/s |  0.3482017  |   17711.36 MB/s   |
--------------------------------------------------------------------------------

如果能看到它在更老的机器上的表现就更好了,比如Nehalem(假设它支持RCP)。

SSE41+FMAImplementation

我们将普通除法与上面提出的算法进行比较,启用FMASSE41

===============================================================
= Compiler & System info
===============================================================
Current CPU          : Intel(R) Xeon(R) CPU E3-1285L v3 @ 3.10GHz
CXX Compiler ID      : GNU
CXX Compiler Path    : /usr/bin/c++
CXX Compiler Version : 4.9.2
CXX Compiler Flags   : -O3 -std=c++11 -msse4.1 -mfma
--------------------------------------------------------------------------------
|   Size    | Division F/C |  Division B/W   | Approx. F/C | Approximation B/W |
--------------------------------------------------------------------------------
|       128 |  0.5714286   |   26884.20 MB/s |  0.5423729  |   25506.41 MB/s   |
|       256 |  0.5701559   |   26879.92 MB/s |  0.5412262  |   25503.95 MB/s   |
|       512 |  0.5701559   |   26904.68 MB/s |  0.5423729  |   25584.65 MB/s   |
|      1024 |  0.5704735   |   26911.46 MB/s |  0.5429480  |   25622.57 MB/s   |
|      2048 |  0.5704735   |   26915.03 MB/s |  0.5433802  |   25640.09 MB/s   |
|      4096 |  0.5703941   |   26917.72 MB/s |  0.5435965  |   25651.63 MB/s   |
|      8192 |  0.5703544   |   26915.85 MB/s |  0.5436687  |   25656.76 MB/s   |
|     16384 |  0.5699972   |   26898.44 MB/s |  0.5262583  |   24834.54 MB/s   |
|     32768 |  0.5699873   |   26898.93 MB/s |  0.5262076  |   24833.21 MB/s   |
|     65536 |  0.5698882   |   26894.48 MB/s |  0.5250567  |   24778.35 MB/s   |
|    131072 |  0.5697024   |   26885.50 MB/s |  0.5224302  |   24654.59 MB/s   |
|    262144 |  0.5696950   |   26884.72 MB/s |  0.5223095  |   24649.49 MB/s   |
|    524288 |  0.5696937   |   26885.37 MB/s |  0.5223308  |   24650.21 MB/s   |
|   1048576 |  0.5690340   |   26854.14 MB/s |  0.5220133  |   24634.71 MB/s   |
|   2097152 |  0.5455717   |   25746.56 MB/s |  0.5041949  |   23794.65 MB/s   |
|   4194304 |  0.5125461   |   24188.11 MB/s |  0.4756604  |   22447.05 MB/s   |
|   8388608 |  0.5043430   |   23800.67 MB/s |  0.4659974  |   21991.51 MB/s   |
|  16777216 |  0.5017375   |   23677.94 MB/s |  0.4614457  |   21776.58 MB/s   |
|  33554432 |  0.5005865   |   23623.50 MB/s |  0.4596277  |   21690.63 MB/s   |
--------------------------------------------------------------------------------

FMA+SSE4.1确实给了我们一定程度的改进,但这还不够好。

AVX2+FMA实现

最后,我们可以看到比较AVX2普通除法和近似方法的真正好处:

===============================================================
= Compiler & System info
===============================================================
Current CPU          : Intel(R) Xeon(R) CPU E3-1285L v3 @ 3.10GHz
CXX Compiler ID      : GNU
CXX Compiler Path    : /usr/bin/c++
CXX Compiler Version : 4.9.2
CXX Compiler Flags   : -O3 -std=c++11 -march=haswell
--------------------------------------------------------------------------------
|   Size    | Division F/C |  Division B/W   | Approx. F/C | Approximation B/W |
--------------------------------------------------------------------------------
|       128 |  0.5663717   |   26672.73 MB/s |  0.9481481  |   44627.89 MB/s   |
|       256 |  0.5651214   |   26653.72 MB/s |  0.9481481  |   44651.56 MB/s   |
|       512 |  0.5644983   |   26640.36 MB/s |  0.9463956  |   44660.99 MB/s   |
|      1024 |  0.5657459   |   26689.41 MB/s |  0.9552239  |   45044.21 MB/s   |
|      2048 |  0.5662151   |   26715.40 MB/s |  0.9624060  |   45405.33 MB/s   |
|      4096 |  0.5663717   |   26726.27 MB/s |  0.9671783  |   45633.64 MB/s   |
|      8192 |  0.5664500   |   26732.42 MB/s |  0.9688941  |   45724.83 MB/s   |
|     16384 |  0.5699377   |   26896.04 MB/s |  0.9092624  |   42909.11 MB/s   |
|     32768 |  0.5699675   |   26897.85 MB/s |  0.9087077  |   42883.21 MB/s   |
|     65536 |  0.5699625   |   26898.59 MB/s |  0.9001456  |   42480.91 MB/s   |
|    131072 |  0.5699253   |   26896.38 MB/s |  0.8926057  |   42124.09 MB/s   |
|    262144 |  0.5699117   |   26895.58 MB/s |  0.8928610  |   42137.13 MB/s   |
|    524288 |  0.5698622   |   26892.87 MB/s |  0.8928002  |   42133.63 MB/s   |
|   1048576 |  0.5685829   |   26833.13 MB/s |  0.8894302  |   41974.25 MB/s   |
|   2097152 |  0.5558453   |   26231.90 MB/s |  0.8371921  |   39508.55 MB/s   |
|   4194304 |  0.5224387   |   24654.67 MB/s |  0.7436747  |   35094.81 MB/s   |
|   8388608 |  0.5143588   |   24273.46 MB/s |  0.7185252  |   33909.08 MB/s   |
|  16777216 |  0.5107452   |   24103.19 MB/s |  0.7133449  |   33664.28 MB/s   |
|  33554432 |  0.5101245   |   24074.10 MB/s |  0.7125114  |   33625.03 MB/s   |
--------------------------------------------------------------------------------
结论

此方法绝对可以提供对普通除法的加速。实际上可以获得多少速度,这实际上取决于底层架构,以及除法如何与应用程序逻辑的其余部分交互。

请参阅Agner Fog的vectorclass他已经实现了一个快速算法,使用SSE/AVX对8位,16位和32位字(但不是64位)进行整数除法http://www.agner.org/optimize/#vectorclass

查看文件vector128 .h的代码和算法的描述,作为他写得很好的手册VectorClass.pdf

这是他的手册中描述算法的片段。

"整数的除法在x86指令集和它的扩展中没有指令对于整向量除法很有用,如果它们存在。因此,vector类库使用了一种快速整数算法部门。该算法的基本原理可以用公式表示:A/b≈A * (2n/b)>> n计算步骤如下:1. 为n找一个合适的值2. 计算2n/b3.计算舍入误差的必要修正4. 做乘法和右移,并应用修正四舍五入错误

如果多个数被同一个除数除,这个公式是有利的b.步骤1、2和3只需要做一次,而步骤4要重复做一次a.文件中描述了数学细节vectori128.h。(参见T. Granlund和P. L. Montgomery:按不变式除法使用乘法的整数,SIGPLAN论文集"…

编辑:靠近文件末尾的vectori128.h展示了如何使用标量变量进行短除法"计算用于快速除法的参数要比做除法。因此,使用相同的除数对象是有利的很多次了。例如,将80个无符号短整数除以10:

short x = 10;
uint16_t dividends[80], quotients[80];         // numbers to work with
Divisor_us div10(x);                          // make divisor object for dividing by 10
Vec8us temp;                                   // temporary vector
for (int i = 0; i < 80; i += 8) {              // loop for 4 elements per iteration
temp.load(dividends+i);                    // load 4 elements
temp /= div10;                             // divide each element by 10
temp.store(quotients+i);                   // store 4 elements
}

"

编辑:用vector of shorts进行整型除法

#include <stdio.h>
#include "vectorclass.h"
int main() {    
short numa[] = {10, 20, 30, 40, 50, 60, 70, 80};
short dena[] = {10, 20, 30, 40, 50, 60, 70, 80};
Vec8s num = Vec8s().load(numa);
Vec8s den = Vec8s().load(dena);
Vec4f num_low = to_float(extend_low(num));
Vec4f num_high = to_float(extend_high(num));
Vec4f den_low = to_float(extend_low(den));
Vec4f den_high = to_float(extend_high(den));
Vec4f qf_low = num_low/den_low;
Vec4f qf_high = num_high/den_high;
Vec4i q_low = truncate_to_int(qf_low);
Vec4i q_high = truncate_to_int(qf_high);
Vec8s q = compress(q_low, q_high);
for(int i=0; i<8; i++) {
printf("%d ", q[i]);
} printf("n");
}

对于8位除法,可以通过创建幻数表来实现。

参见"Hacker’s Delight",第238页

签署:

static inline __m128i _mm_div_epi16(const __m128i &a_epi16, const __m128i &b_epi16) {
//
// Setup the constants.
//
const __m128  two     = _mm_set1_ps(2.00000051757f);
const __m128i lo_mask = _mm_set1_epi32(0xFFFF);
//
// Convert to two 32-bit integers
//
const __m128i a_hi_epi32       = _mm_srai_epi32(a_epi16, 16);
const __m128i a_lo_epi32_shift = _mm_slli_epi32(a_epi16, 16);
const __m128i a_lo_epi32       = _mm_srai_epi32(a_lo_epi32_shift, 16);
const __m128i b_hi_epi32       = _mm_srai_epi32(b_epi16, 16);
const __m128i b_lo_epi32_shift = _mm_slli_epi32(b_epi16, 16);
const __m128i b_lo_epi32       = _mm_srai_epi32(b_lo_epi32_shift, 16);
//
// Convert to 32-bit floats
//
const __m128 a_hi = _mm_cvtepi32_ps(a_hi_epi32);
const __m128 a_lo = _mm_cvtepi32_ps(a_lo_epi32);
const __m128 b_hi = _mm_cvtepi32_ps(b_hi_epi32);
const __m128 b_lo = _mm_cvtepi32_ps(b_lo_epi32);
//
// Calculate the reciprocal
//
const __m128 b_hi_rcp = _mm_rcp_ps(b_hi);
const __m128 b_lo_rcp = _mm_rcp_ps(b_lo);
//
// Calculate the inverse
//
#ifdef __FMA__
const __m128 b_hi_inv_1 = _mm_fnmadd_ps(b_hi_rcp, b_hi, two);
const __m128 b_lo_inv_1 = _mm_fnmadd_ps(b_lo_rcp, b_lo, two);
#else
const __m128 b_mul_hi   = _mm_mul_ps(b_hi_rcp, b_hi);
const __m128 b_mul_lo   = _mm_mul_ps(b_lo_rcp, b_lo);
const __m128 b_hi_inv_1 = _mm_sub_ps(two, b_mul_hi);
const __m128 b_lo_inv_1 = _mm_sub_ps(two, b_mul_lo);
#endif
//
// Compensate for the loss
//
const __m128 b_hi_rcp_1 = _mm_mul_ps(b_hi_rcp, b_hi_inv_1);
const __m128 b_lo_rcp_1 = _mm_mul_ps(b_lo_rcp, b_lo_inv_1);
//
// Perform the division by multiplication
//
const __m128 hi = _mm_mul_ps(a_hi, b_hi_rcp_1);
const __m128 lo = _mm_mul_ps(a_lo, b_lo_rcp_1);
//
// Convert back to integers
//
const __m128i hi_epi32 = _mm_cvttps_epi32(hi);
const __m128i lo_epi32 = _mm_cvttps_epi32(lo);
//
// Zero-out the unnecessary parts
//
const __m128i hi_epi32_shift = _mm_slli_epi32(hi_epi32, 16);
#ifdef __SSE4_1__
//
// Blend the bits, and return
//
return _mm_blend_epi16(lo_epi32, hi_epi32_shift, 0xAA);
#else
//
// Blend the bits, and return
//
const __m128i lo_epi32_mask = _mm_and_si128(lo_epi32, const_mm_div_epi16_lo_mask);
return _mm_or_si128(hi_epi32_shift, lo_epi32_mask);
#endif
}

0无符号:

__m128i _mm_div_epu8(__m128i n, __m128i den)
{
static const uint16_t magic_number_table[256] =
{
0x0001, 0x0000, 0x8000, 0x5580, 0x4000, 0x3340, 0x2ac0, 0x04a0, 0x2000, 0x1c80, 0x19a0, 0x0750, 0x1560, 0x13c0, 0x0250, 0x1120,
0x1000, 0x0f10, 0x0e40, 0x0d80, 0x0cd0, 0x0438, 0x03a8, 0x0328, 0x0ab0, 0x0a40, 0x09e0, 0x0980, 0x0128, 0x00d8, 0x0890, 0x0048,
0x0800, 0x07c8, 0x0788, 0x0758, 0x0720, 0x06f0, 0x06c0, 0x0294, 0x0668, 0x0640, 0x021c, 0x05f8, 0x05d8, 0x01b4, 0x0194, 0x0578,
0x0558, 0x013c, 0x0520, 0x0508, 0x04f0, 0x04d8, 0x04c0, 0x04a8, 0x0094, 0x0480, 0x006c, 0x0458, 0x0448, 0x0034, 0x0024, 0x0014,
0x0400, 0x03f4, 0x03e4, 0x03d4, 0x03c8, 0x03b8, 0x03ac, 0x039c, 0x0390, 0x0384, 0x0378, 0x036c, 0x0360, 0x0354, 0x014a, 0x0340,
0x0334, 0x032c, 0x0320, 0x0318, 0x010e, 0x0304, 0x02fc, 0x02f4, 0x02ec, 0x02e4, 0x02dc, 0x02d4, 0x02cc, 0x02c4, 0x02bc, 0x02b4,
0x02ac, 0x02a4, 0x02a0, 0x0298, 0x0290, 0x028c, 0x0284, 0x007e, 0x0278, 0x0072, 0x026c, 0x0066, 0x0260, 0x025c, 0x0254, 0x0250,
0x004a, 0x0244, 0x0240, 0x023c, 0x0036, 0x0032, 0x022c, 0x0228, 0x0224, 0x001e, 0x001a, 0x0016, 0x0012, 0x000e, 0x000a, 0x0006,
0x0200, 0x00fd, 0x01fc, 0x01f8, 0x01f4, 0x01f0, 0x01ec, 0x01e8, 0x01e4, 0x01e0, 0x01dc, 0x01d8, 0x01d6, 0x01d4, 0x01d0, 0x01cc,
0x01c8, 0x01c4, 0x01c2, 0x01c0, 0x01bc, 0x01b8, 0x01b6, 0x01b4, 0x01b0, 0x01ae, 0x01ac, 0x01a8, 0x01a6, 0x01a4, 0x01a0, 0x019e,
0x019c, 0x0198, 0x0196, 0x0194, 0x0190, 0x018e, 0x018c, 0x018a, 0x0188, 0x0184, 0x0182, 0x0180, 0x017e, 0x017c, 0x017a, 0x0178,
0x0176, 0x0174, 0x0172, 0x0170, 0x016e, 0x016c, 0x016a, 0x0168, 0x0166, 0x0164, 0x0162, 0x0160, 0x015e, 0x015c, 0x015a, 0x0158,
0x0156, 0x0154, 0x0152, 0x0051, 0x0150, 0x014e, 0x014c, 0x014a, 0x0148, 0x0047, 0x0146, 0x0144, 0x0142, 0x0140, 0x003f, 0x013e,
0x013c, 0x013a, 0x0039, 0x0138, 0x0136, 0x0134, 0x0033, 0x0132, 0x0130, 0x002f, 0x012e, 0x012c, 0x012a, 0x0029, 0x0128, 0x0126,
0x0025, 0x0124, 0x0122, 0x0021, 0x0120, 0x001f, 0x011e, 0x011c, 0x001b, 0x011a, 0x0019, 0x0118, 0x0116, 0x0015, 0x0114, 0x0013,
0x0112, 0x0110, 0x000f, 0x010e, 0x000d, 0x010c, 0x000b, 0x010a, 0x0009, 0x0108, 0x0007, 0x0106, 0x0005, 0x0104, 0x0003, 0x0102
};
static const uint16_t shift_table[256] =
{
0x0001, 0x0100, 0x0100, 0x0080, 0x0100, 0x0040, 0x0040, 0x0020, 0x0100, 0x0080, 0x0020, 0x0010, 0x0020, 0x0040, 0x0010, 0x0020,
0x0100, 0x0010, 0x0040, 0x0080, 0x0010, 0x0008, 0x0008, 0x0008, 0x0010, 0x0040, 0x0020, 0x0080, 0x0008, 0x0008, 0x0010, 0x0008,
0x0100, 0x0008, 0x0008, 0x0008, 0x0020, 0x0010, 0x0040, 0x0004, 0x0008, 0x0040, 0x0004, 0x0008, 0x0008, 0x0004, 0x0004, 0x0008,
0x0008, 0x0004, 0x0020, 0x0008, 0x0010, 0x0008, 0x0040, 0x0008, 0x0004, 0x0080, 0x0004, 0x0008, 0x0008, 0x0004, 0x0004, 0x0004,
0x0100, 0x0004, 0x0004, 0x0004, 0x0008, 0x0008, 0x0004, 0x0004, 0x0010, 0x0004, 0x0008, 0x0004, 0x0020, 0x0004, 0x0002, 0x0040,
0x0004, 0x0004, 0x0020, 0x0008, 0x0002, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004, 0x0004,
0x0004, 0x0004, 0x0020, 0x0008, 0x0010, 0x0004, 0x0004, 0x0002, 0x0008, 0x0002, 0x0004, 0x0002, 0x0020, 0x0004, 0x0004, 0x0010,
0x0002, 0x0004, 0x0040, 0x0004, 0x0002, 0x0002, 0x0004, 0x0008, 0x0004, 0x0002, 0x0002, 0x0002, 0x0002, 0x0002, 0x0002, 0x0002,
0x0100, 0x0001, 0x0004, 0x0008, 0x0004, 0x0010, 0x0004, 0x0008, 0x0004, 0x0020, 0x0004, 0x0008, 0x0002, 0x0004, 0x0010, 0x0004,
0x0008, 0x0004, 0x0002, 0x0040, 0x0004, 0x0008, 0x0002, 0x0004, 0x0010, 0x0002, 0x0004, 0x0008, 0x0002, 0x0004, 0x0020, 0x0002,
0x0004, 0x0008, 0x0002, 0x0004, 0x0010, 0x0002, 0x0004, 0x0002, 0x0008, 0x0004, 0x0002, 0x0080, 0x0002, 0x0004, 0x0002, 0x0008,
0x0002, 0x0004, 0x0002, 0x0010, 0x0002, 0x0004, 0x0002, 0x0008, 0x0002, 0x0004, 0x0002, 0x0020, 0x0002, 0x0004, 0x0002, 0x0008,
0x0002, 0x0004, 0x0002, 0x0001, 0x0010, 0x0002, 0x0004, 0x0002, 0x0008, 0x0001, 0x0002, 0x0004, 0x0002, 0x0040, 0x0001, 0x0002,
0x0004, 0x0002, 0x0001, 0x0008, 0x0002, 0x0004, 0x0001, 0x0002, 0x0010, 0x0001, 0x0002, 0x0004, 0x0002, 0x0001, 0x0008, 0x0002,
0x0001, 0x0004, 0x0002, 0x0001, 0x0020, 0x0001, 0x0002, 0x0004, 0x0001, 0x0002, 0x0001, 0x0008, 0x0002, 0x0001, 0x0004, 0x0001,
0x0002, 0x0010, 0x0001, 0x0002, 0x0001, 0x0004, 0x0001, 0x0002, 0x0001, 0x0008, 0x0001, 0x0002, 0x0001, 0x0004, 0x0001, 0x0002
};
static const uint16_t mask_table[256] =
{
0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0xffff, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0xffff, 0x0000, 0xffff,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0xffff, 0x0000,
0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0xffff, 0xffff,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000,
0xffff, 0x0000, 0x0000, 0x0000, 0xffff, 0xffff, 0x0000, 0x0000, 0x0000, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000,
0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000, 0x0000, 0xffff, 0x0000, 0x0000,
0xffff, 0x0000, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0x0000, 0xffff, 0x0000, 0xffff,
0x0000, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000, 0xffff, 0x0000
};
uint8_t load_den[16];
_mm_storeu_si128((__m128i*)load_den, den);
uint16_t mul[16];
uint16_t mask[16];
uint16_t shift[16];
for (size_t i = 0; i < 16; i++)
{
const uint16_t cur_den = load_den[i];
mul[i] = magic_number_table[cur_den];
mask[i] = mask_table[cur_den];
shift[i] = shift_table[cur_den];
}
#if 0
__m128i a = _mm_unpacklo_epi8(n, _mm_setzero_si128());
__m128i b = _mm_unpackhi_epi8(n, _mm_setzero_si128());
__m256i c = _mm256_castsi128_si256(a);
c = _mm256_insertf128_si256(c, b, 1);
#else
// Thanks to Peter Cordes
__m256i c = _mm256_cvtepu8_epi16(n);
#endif
__m256i magic = _mm256_loadu_si256((const __m256i*)mul);
__m256i high = _mm256_mulhi_epu16(magic, c);
__m256i low = _mm256_mullo_epi16(magic, c);
__m256i low_down = _mm256_srli_epi16(low, 8);
__m256i high_up = _mm256_slli_epi16(high, 8);
__m256i low_high = _mm256_or_si256(low_down, high_up);
__m256i target_up = _mm256_mullo_epi16(c, _mm256_loadu_si256((const __m256i*)shift));
__m256i cal1 = _mm256_sub_epi16(target_up, low_high);
__m256i cal2 = _mm256_srli_epi16(cal1, 1);
__m256i cal3 = _mm256_add_epi16(cal2, low_high);
__m256i cal4 = _mm256_srli_epi16(cal3, 7);
__m256i res = _mm256_blendv_epi8(high, cal4, _mm256_loadu_si256((const __m256i*)mask));
__m128i v0h = _mm256_extractf128_si256(res, 0);
__m128i v0l = _mm256_extractf128_si256(res, 1);
return _mm_packus_epi16(v0h, v0l);
}

这是由于原始解决方案的新来者:这个Agner Fog的子程序库在优化方面为我创造了奇迹

下面是对同一个变量值进行多次分割的情况(比如在大循环中)

#include <asmlib.h>
unsigned int a, b, d;
unsigned int divisor = any_random_value;
div_u32 OptimumDivision(divisor);
a/OptimumDivision;
b/OptimumDivision;

这是unsigned int -如果你需要负值,使用div_i32代替,这在我的测试中更快,即使手册说相反

我得到了3倍甚至更多的性能