我的 C++11 程序正在执行序列化数据的在线处理,循环需要运行数百万个内存位置。计算效率是必须的,我担心的是,在这样的循环中调用函数或类会产生不必要的操作,从而影响效率,例如在不同的变量范围之间传递操作所需的几个指针值。
为了举例说明,让我们考虑以下虚拟示例,其中“某事”是重复的操作。请注意,“something”中的代码使用循环范围内的变量。
do {
something(&span,&foo);
spam++
foo++
if ( spam == spam_spam ) {
something(&span,&foo);
other_things(&span,&foo);
something(&span,&foo);
}
else {
something(&span,&foo);
still_other_things(&span,&foo);
something(&span,&foo);
}
}
while (foo<bar);
有没有办法重复代码块并避免使用不必要的操作移动和复制变量?在此类循环中使用函数和类实际上是否意味着额外的操作以及如何避免它?
更新
按照建议,我使用下面提供的代码运行了一些测试。我测试了几个关于如何调用简单增量 1 亿次的选项。我在 Hyper-V 下的 x86_64 虚拟机上通过 RHEL 7 Server 7.6 使用 GCC。
最初,使用“g++ -std=c++17 -o test.o test.cpp”编译
简单循环计算(基线):211.046ms
内联函数:468.768 毫秒
Lambda 函数:253.466 毫秒
定义宏:211.995ms
函数传递值:466.986ms
函数传递指针:344.646ms
带 void 的函数:190.557 毫秒
对象方法与成员操作:231.458ms
对象方法传递值:227.615ms
从这些结果中,我意识到编译器没有采用内联建议,即使在尝试按照 g++ doesn't inline functions 中的建议将其膨胀后也是如此。
后来,按照 Mat 在同一篇文章的回答中的建议,我使用“g++ -std=c++17 -O2 -o test.o test.cpp”打开了编译器优化,并得到了以下结果与未优化的测试相比,迭代次数相同。
简单循环计算(基线):62.9254ms
内联函数:65.0564 毫秒
Lambda 函数:32.8637 毫秒
定义宏:63.0299ms
函数传递值:64.2876ms
函数传递指针:63.3416ms
带 void 的函数:32.1073ms
对象方法与成员操作:63.3847ms
对象方法传递值:62.5151ms
到此为止的结论:
内联函数不是好的选择,因为无法确定编译器将如何真正接受它,结果可能与使用标准函数一样糟糕。
“定义宏”和“lambda 函数”是内联的更好替代方法。每个都有其优点和特点,#define 更灵活。
使用对象成员和方法可以很好地平衡解决任何情况下的问题,同时以更易于维护和优化的形式维护代码。
调整编译器是值得的;
遵循用于测试的代码:
// Libraries
#include <iostream>
#include <cmath>
#include <chrono>
// Namespaces
using namespace std;
using namespace std::chrono;
// constants that control program behaviour
const long END_RESULT = 100000000;
const double AVERAGING_LENGTH = 40.0;
const int NUMBER_OF_ALGORITHM = 9;
const long INITIAL_VALUE = 0;
const long INCREMENT = 1;
// Global variables used for test with void function and to general control of the program;
long global_variable;
long global_increment;
// Function that returns the execution time for a simple loop
int64_t simple_loop_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Perform the computation for baseline
do {
local_variable += local_increment;
} while ( local_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return(duration_cast<microseconds>( timer_stop - timer_start ).count());
}
// Functions that computes the execution time when using inline code within the loop
inline long increment_variable() __attribute__((always_inline));
inline long increment_variable(long local_variable, long local_increment) {
return local_variable += local_increment;
}
int64_t inline_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Perform the computation for baseline
do {
local_variable = increment_variable(local_variable,local_increment);
} while ( local_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// Functions that computes the execution time when using lambda code within the loop
int64_t labda_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// define lambda function
auto lambda_increment = [&] {
local_variable += local_increment;
};
// Perform the computation for baseline
do {
lambda_increment();
} while ( local_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// define lambda function
#define define_increment() local_variable += local_increment;
// Functions that computes the execution time when using lambda code within the loop
int64_t define_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Perform the computation for baseline
do {
define_increment();
} while ( local_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// Functions that compute the execution time when calling a function within the loop passing variable values
long increment_with_values_function(long local_variable, long local_increment) {
return local_variable += local_increment;
}
int64_t function_values_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Perform the computation for baseline
do {
local_variable = increment_with_values_function(local_variable,local_increment);
} while ( local_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// Functions that compute the execution time when calling a function within the loop passing variable pointers
long increment_with_pointers_function(long *local_variable, long *local_increment) {
return *local_variable += *local_increment;
}
int64_t function_pointers_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Perform the computation for baseline
do {
local_variable = increment_with_pointers_function(&local_variable,&local_increment);
} while ( local_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// Functions that compute the execution time when calling a function within the loop without passing variables
void increment_with_void_function(void) {
global_variable += global_increment;
}
int64_t function_void_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// set global variables
global_variable = local_variable;
global_increment = local_increment;
// Perform the computation for baseline
do {
increment_with_void_function();
} while ( global_variable != END_RESULT);
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// Object and Function that compute the duration when using a method of the object where data is stored without passing variables
struct object {
long object_variable = 0;
long object_increment = 1;
object(long local_variable, long local_increment) {
object_variable = local_variable;
object_increment = local_increment;
}
void increment_object(void){
object_variable+=object_increment;
}
void increment_object_with_value(long local_increment){
object_variable+=local_increment;
}
};
int64_t object_members_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Create object
object object_instance = {local_variable,local_increment};
// Perform the computation for baseline
do {
object_instance.increment_object();
} while ( object_instance.object_variable != END_RESULT);
// Get the results out of the object
local_variable = object_instance.object_variable;
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
// Function that compute the duration when using a method of the object where data is stored passing variables
int64_t object_values_computation(long local_variable, long local_increment) {
// Starts the clock to measure the execution time for the baseline
high_resolution_clock::time_point timer_start = high_resolution_clock::now();
// Create object
object object_instance = {local_variable,local_increment};
// Perform the computation for baseline
do {
object_instance.increment_object_with_value(local_increment);
} while ( object_instance.object_variable != END_RESULT);
// Get the results out of the object
local_variable = object_instance.object_variable;
// Stop the clock to measure performance of the silly version
high_resolution_clock::time_point timer_stop = high_resolution_clock::now();
return duration_cast<microseconds>( timer_stop - timer_start ).count();
}
int main() {
// Create array to store execution time results for all tests
pair<string,int64_t> duration_sum[NUMBER_OF_ALGORITHM]={
make_pair("Simple loop computation (baseline): ",0.0),
make_pair("Inline Function: ",0.0),
make_pair("Lambda Function: ",0.0),
make_pair("Define Macro: ",0.0)
make_pair("Function passing values: ",0.0),
make_pair("Function passing pointers: ",0.0),
make_pair("Function with void: ",0.0),
make_pair("Object method operating with members: ",0.0),
make_pair("Object method passing values: ",0.0),
};
// loop to compute average of several execution times
for ( int i = 0; i < AVERAGING_LENGTH; i++) {
// Compute the execution time for a simple loop as the baseline
duration_sum[0].second = duration_sum[0].second + simple_loop_computation(INITIAL_VALUE, INCREMENT);
// Compute the execution time when using inline code within the loop (expected same as baseline)
duration_sum[1].second = duration_sum[1].second + inline_computation(INITIAL_VALUE, INCREMENT);
// Compute the execution time when using lambda code within the loop (expected same as baseline)
duration_sum[2].second = duration_sum[2].second + labda_computation(INITIAL_VALUE, INCREMENT);
// Compute the duration when using a define macro
duration_sum[3].second = duration_sum[3].second + define_computation(INITIAL_VALUE, INCREMENT);
// Compute the execution time when calling a function within the loop passing variables values
duration_sum[4].second = duration_sum[4].second + function_values_computation(INITIAL_VALUE, INCREMENT);
// Compute the execution time when calling a function within the loop passing variables pointers
duration_sum[5].second = duration_sum[5].second + function_pointers_computation(INITIAL_VALUE, INCREMENT);
// Compute the execution time when calling a function within the loop without passing variables
duration_sum[6].second = duration_sum[6].second + function_void_computation(INITIAL_VALUE, INCREMENT);
// Compute the duration when using a method of the object where data is stored without passing variables
duration_sum[7].second = duration_sum[7].second + object_members_computation(INITIAL_VALUE, INCREMENT);
// Compute the duration when using a method of the object where data is stored passing variables
duration_sum[8].second = duration_sum[8].second + object_values_computation(INITIAL_VALUE, INCREMENT);
}
double average_baseline_duration = 0.0;
// Print out results
for ( int i = 0; i < NUMBER_OF_ALGORITHM; i++) {
// compute averave from sum
average_baseline_duration = ((double)duration_sum[i].second/AVERAGING_LENGTH)/1000.0;
// Print the result
cout << duration_sum[i].first << average_baseline_duration << "ms \n";
}
return 0;
}
最佳答案
如果代码足够短,可以声明为内联,编译器会把它放入内联。如果不是,那么重复它可能无济于事。
但是,老实说,这是最不有效的优化形式。关注高效的算法和缓存高效的数据结构。
关于c++ - 如何在不使用函数或类的情况下重复代码段以实现 C++ 中的高性能循环,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/55383866/