如果我有一个 std::string
:
std::string s{"hello"};
以及一个就地修改它的循环,如下所示:
for (auto &c: s)
c = std::toupper(c);
我可以将其替换为等效的 transform
:
std::transform(s.begin(), s.end(), s.begin(),
[](unsigned char c) -> unsigned char
{ return std::toupper(c); });
这些片段生成相同的 assembly 。他们也有类似的performance .
但是,如果我有 std::vector<std::string>
:
std::vector<std::string> v {"hello", "how", "are", "you"};
并就地修改它,如下所示:
for (auto & s : v)
for (auto &c: s)
c = std::toupper(c);
等效的变换应该是:
std::transform(std::begin(v), std::end(v), std::begin(v),
[](auto s) {
std::transform(std::begin(s), std::end(s), std::begin(s),
[](unsigned char c) -> unsigned char { return std::toupper(c); });
return s;
});
但是,transform
版本生成的数量超过一半 assembly ,和performs相应地很差,这让我感到惊讶。
是std::transform
在这种情况下不是零成本抽象,或者我只是错误地使用它?
最佳答案
通过引用传递和返回所有内容。否则,您将复制该字符串的多个拷贝。请注意更改:[](auto& s) -> std::string& {
std::transform(std::begin(v), std::end(v), std::begin(v),
[](auto& s) -> std::string& {
std::transform(std::begin(s), std::end(s), std::begin(s),
[](unsigned char c) -> unsigned char { return std::toupper(c); });
return s;
});
我在您的链接中添加了两个新的快速工作台函数。一种将输入字符串作为引用传递的字符串。另一个也通过引用返回。即:
static void Transform2(benchmark::State& state) {
// Code before the loop is not measured
std::vector<std::string> v {"hello", "how", "are", "you"};
for (auto _ : state) {
std::transform(std::begin(v), std::end(v), std::begin(v),
[](auto& s) {
std::transform(std::begin(s), std::end(s), std::begin(s),
[](unsigned char c) -> unsigned char { return std::toupper(c); });
return s;
});
}
}
BENCHMARK(Transform2);
static void Transform3(benchmark::State& state) {
// Code before the loop is not measured
std::vector<std::string> v {"hello", "how", "are", "you"};
for (auto _ : state) {
std::transform(std::begin(v), std::end(v), std::begin(v),
[](auto& s) -> std::string& {
std::transform(std::begin(s), std::end(s), std::begin(s),
[](unsigned char c) -> unsigned char { return std::toupper(c); });
return s;
});
}
}
BENCHMARK(Transform3);
根据我运行基准测试时的幸运程度,Transform3 的性能几乎(有时等于)InPlace 测试实现。
关于c++ - 嵌套的 std::transform 效率低吗?,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/61263702/