Sunday, January 8, 2017

Higher order functions as an enabler for lazy evaluation

Yesterdays post about Generating lambdas for clarity and performance showed how to make use of higher order functions to improve clarity while giving the optimiser a chance to improve performance, but the example used retained the original inflexible design.

Here's a short recap. Given a struct Employee, there is a function that filters out developers based on their salary. The original code looked like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
struct Employee
{
  using Number = unsigned;
  Number id;
  unsigned salary;
  std::string title;
  std::string name;
};

using staff = std::vector<Employee>;

auto devs_who_make_at_least(const staff& s, unsigned floor)
{
  std::vector<staff::value_type const*> rv;
  for (auto& e : s)
  {
    if (e.salary >= floor && e.title == "Developer")
    {
      rv.push_back(&e);
    }
  }
  return rv;
}

The code is by no means difficult to understand, but it lacks in flexibility, and loops like this spread over many places in the code base becomes a maintenance liability.

Enter a kind of filter function template, that like the devs_who_make_at_least() function, returns a vector of pointers to the elements given, but using a provided predicate instead of a hard coded condition:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
template <typename Container, typename Pred>
auto elements_matching(Container const& c, Pred&& p)
{
  std::vector<typename Container::value_type const*> rv;
  for (auto& e : c)
  {
    if (p(e))
    {
      rv.push_back(&e);
    }
  }
  return rv;
}

This is more generic. It works on any STL-like container and accepts any predicate that works on the value type. A few sample higher order functions were introduced to select elements just as in the original example:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
auto earns_at_least(unsigned floor)
{
  return [floor](const Employee& e) { return e.salary >= floor; };
}

auto has_title(const std::string& s)
{
  return [s](const Employee& e) { return e.title == s; };
}

template <typename T, typename U>
auto pred_and(T&& t, U&& u)
{
  return [t,u](auto&& x) { return t(x) && u(x); };
}

Here pred_and() is used to create a new predicate from two others, returning true only if both contained predicates return true. Creating a vector of pointers to all developers that earn 1M or more simply becomes:

1
2
3
auto megadevs = elements_matching(employees,
                                  pred_and(earns_at_least(1000000), 
                                           has_title("Developer")));

This was also shown to give performance gains over the original code.

So far so good, but what if the desired result is not a vector of pointers? What if we just want to calculate the accumulated salary of those selected? It doesn't make sense to pay the cost of populating an array for that. What if not all developers are needed, but you just need to browse through the selected ones until some other criteria is met? Then it is unnecessary to have populated the vector with all of them.

Enter lazy evaluation. The idea is to create yet another level of indirection, and defer calling the predicate until the actual iteration is done.

What is wanted is something like this:
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
auto megadevs = filter(employees,
                       pred_and(earns_at_least(1000000), 
                                has_title("Developer")));

// No iteration over employees yet
// No vector created.

uint64_t sum = 0;
for (auto i = megadevs.begin();// calls pred and increments until true or end
     i != megadevs.end();      // simple comparison
     ++i)                      // calls pred on incremented until true or end
{
  sum += i->salary;            // refers to element in employees
}

From this a design emerges. The type of megadevs, returned by the call to filter(), must have member functions begin() and end(). The iterators returned must have access to the iterators into employees, and must have access to the predicate.

Here's an outline for the type returned by filter():

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
template <typename Container, typename Pred>
class filter_t
{
  using CI = typename Container::const_iterator;
 public:
  class iterator;

  template <typename P>
  filter_t(Container& c_, P&& p_)
   : c(c_), p(std::forward<P>(p_))  { }

  iterator begin() {
    auto i = c.begin();
    auto e = c.end();
    while (i != e && !p(*i))
      ++i;
    return { i, e, p };
  }

  iterator end() { return { c.end(), c.end(), p }; }
private:
  const Container& c;
  Pred             p;
};

Each instance must refer to the original container, and must access the predicate, so these types are template parameters. The alias CI is just there for typographical purposes, making blog code shorter. The class iterator is deferred for a while. I'll get back to it soon. The interesting part is the member function begin(). We want it to create an iterator that refers to the first element in the container that fulfils the predicate, or end if there is no such element. It does this by checking the predicate and incrementing until true. The returned filter iterator must have the found iterator, the end, and the predicate, otherwise it cannot implement its increment operators.
The natural implementation would be to use std::find_if() instead of incrementing and checking manually, but this was unexpectedly slow.
Now we can look at the iterator class:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
template <typename Container, typename Pred>
class filter_t<Container, Pred>::iterator
{
public:
  using value_type = typename std::iterator_traits<CI>::value_type;
  using reference = typename std::iterator_traits<CI>::reference;;
  using pointer = typename std::iterator_traits<CI>::pointer;
  using difference_type = typename std::iterator_traits<CI>::difference_type;
  using iterator_category = std::forward_iterator_tag;

  iterator(CI b, CI e, Pred& p_) : iter(b), iend(e), p(p_) { }

  reference operator*() const { return *iter; }

  iterator& operator++()
  {
    ++iter;
    while (iter != iend && !p(*iter))
      ++iter;
    return *this;
  }

  iterator operator++(int)
  {
    iterator rv(*this);
    ++*this;
    return rv;
  }

  bool operator==(const iterator& rh) const
  {
    return iter == rh.iter;
  }

  bool operator!=(const iterator& rh) const
  {
    return iter != rh.iter;
  }
private:
  CI       iter;
  CI const iend;
  Pred&    p;
};

It's quite a code block, but most is obvious. The aliases on lines 5-9 are there to make the iterator cooperate with algorithms in the standard library. See std::iterator_traits<> for details.

The other interesting part is operator++() on lines 15-21. As when created from the begin() member function of filter_t<>, it searches for the next iterator in the referenced container until the predicate matches or the end is reached. Also here, a hand written loop is used, after performance measurements showed an unexpected need.

With this in place, the filter() function template becomes nearly trivial:

1
2
3
4
5
6
template <typename Container, typename Predicate>
inline filter_t<Container, std::remove_reference_t<Predicate>>
filter(Container& c, Predicate&& p)
{
  return { c, std::forward<Predicate>(p) };
}

Here's a good time to warn that this is a proof of concept code, intended to show the idea. For real world use, you would for example need variants with mutable access and variants that take ownership of the container if it is an rvalue, otherwise you cannot safely chain operations. These variants change nothing in performance, but they do make the implementation rather more complex.

Now is a good time to see if this did any good. A slight modification is made from the sample program of yesterday. The program populates a vector of 5000 employees. 1/4 of them with title = "Developer", and an unrealistically uniform random distribution salary in the range 50000 - 250000. With those the program did 500000 loops, each filtering out the developers that make 150000 or more and calculates the accumulated salary of the first 500.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
auto devs = filter(employees,
                   pred_and(earns_at_least(150000), 
                            has_title("Developer")));

int count = 500;
uint64_t sum = 0;
for (auto & d : devs)
{
  sum += d.salary;
  if (--count == 0) break;
}

Using g++ 6.2 and -O3, perf yields this result:

 Performance counter stats for './a.out':

   10 814 982 186      branch-instructions                                         
      752 594 101      branch-misses             #    6,96% of all branches         
   30 224 548 004      cpu-cycles                                                  
   29 445 937 781      instructions              #    0,97  insn per cycle         

     10,177892229 seconds time elapsed

Slightly rewriting the test program to add using the vector version from yesterday as:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
auto devs = elements_matching(employees,
                              pred_and(earns_at_least(150000), 
                                       has_title("Developer")));

int count = 500;
uint64_t sum = 0;
for (auto & d : devs)
{
  sum += d->salary;
  if (--count == 0) break;
}

gives the result:

 Performance counter stats for './a.out':

   14 881 795 600      branch-instructions                                         
    1 061 800 168      branch-misses             #    7,13% of all branches         
   45 275 693 140      cpu-cycles                                                  
   47 380 071 935      instructions              #    1,05  insn per cycle         

     15,215398962 seconds time elapsed

Here the advantage of lazy evaluation becomes obvious. We're only interested in the first 500 matches, so not populating a vector with the remaining is a gain. However, in all fairness, it should be said that there is a cost to this generality. The filter iterator shows much worse branch prediction results, and for large data sets, this can be noticeable.

There it is. Higher order functions are by no means a necessity for lazy evaluation, but when you have them, it's not a huge job to implement, and using it becomes natural. The lazy evaluation can give large performance gains, but there is an added complexity which may back fire.

Expanding further on this sooner or later leads to a range library, but that is another story.

Saturday, January 7, 2017

Generate lambdas for clarity and performance

Higher order functions, functions that operate on other functions or returns functions, are familiar to those who have had some experience with functional programming, but they often seems magical to those who have not. Some of those with experience of using higher order functions have a gut feeling that they are expensive to use and prefer to avoid them.

[ edited 2017-01-08: added performance numbers for gcc at the bottom of the post ]
[ edited 2017-01-14: Perfect forwarding of functions in higher order function ]

Take this somewhat silly traditional C++ code:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
struct Employee
{
  using Number = unsigned;
  Number id;
  unsigned salary;
  std::string title;
  std::string name;
};

using staff = std::vector<Employee>;

auto devs_who_make_at_least(const staff& s, unsigned floor)
{
  std::vector<staff::value_type const*> rv;
  for (auto& e : s)
  {
    if (e.salary >= floor && e.title == "Developer")
    {
      rv.push_back(&e);
    }
  }
  return rv;
}

It's not very complex, and most C++ developers understand what it does in a few seconds. If this is the only place in the entire code base that elements are selected from a vector of Employees, then all is well. But what if there are several places? The code becomes cluttered with almost identical copies of the same loop. This screams algorithm. None in the standard library is a perfect fit, though. Let's hand roll a simple one:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
template <typename Container, typename Pred>
auto elements_matching(Container const& c, Pred&& p)
{
  std::vector<typename Container::value_type const*> rv;
  for (auto& e : c)
  {
    if (p(e))
    {
      rv.push_back(&e);
    }
  }
  return rv;
}


It's not ideal, but it works with any container type and returns a vector with pointers to the elements in the container that matches the predicate.

With this, the devs_who_make_at_least() function becomes simple:


1
2
3
4
5
6
7
auto devs_who_make_at_least(const staff& s, unsigned floor)
{
  auto pred = [floor](const Employee& e) {
    return e.salary >= floor && e.title == "Developer";
  };
  return elements_matching(s,pred);
}

This is cool, but what if there's more than one place in the code where you want to select employees based on salary or title?

C++14 introduced the auto return type, which is great for writing functions that generates lambdas.

Take for example:

1
2
3
4
5
6
7
8
9
auto earns_at_least(unsigned floor)
{
  return [floor](const Employee& e) { return e.salary >= floor; };
}

auto has_title(const std::string& s)
{
  return [s](const Employee& e) { return e.title == s; };
}

Each of the above functions returns a lambda that has captured the function parameter. These can be quite expressively used with the elements_matching algorithm introduced earlier:

1
2
3
4
staff employees;

devs = elements_matching(employees, has_title("Developer"));
megaearners = elements_matching(employees, earns_at_least(1000000));

This is powerful on its own right. The code is easy to read. Most, even very inexperienced C++ developers, grasp the intent directly, even if the mechanism may be unclear.

But what if we want more than one criteria, like selecting developers with a certain salary? Let's introduce an and higher order function, that returns true if two contained predicates both return true.

1
2
3
4
5
6
template <typename T, typename U>
auto pred_and(T&& t, U&& u)
{
  return [t = std::forward<T>(t),u = std::forward<U>(u)](auto&& x)
         { return t(x) && u(x); };
}

This function template pred_and(), creates a new predicate using two predicates. It will be true if and only if both t(x) and u(x) evaluates to true. It naturally short circuits the logic, so that if t(x) is false, u(x) is never evaluated.

Now finding the super well paid developers who make more that 1M becomes so easy it doesn't even need a separate function anymore.

1
2
3
auto megadevs = elements_matching(employees,
                                  pred_and(earns_at_least(1000000), 
                                           has_title("Developer")));

So what about performance, then? Surely this magic has a cost in confusing the optimiser?

I created a small test program that populates a vector of 5000 employees. 1/4 of them with title = "Developer", and an unrealistically uniform random distribution salary in the range 50000 - 250000. With those the program did 500000 loops, each filtering out the developers that make 150000 or more. This was build with clang++ 3.9.

The output from the linux tool perf are,

first from the hand made loop at the top of this post:

 Performance counter stats for './a.out':

   31 078 321 937      branch-instructions                                  
    2 055 192 607      branch-misses             #    6,61% of all branches         
   97 906 147 917      cpu-cycles                                           
  111 202 137 490      instructions              #    1,14  insn per cycle  

     32,974801101 seconds time elapsed

Then from using the algorithm and higher order function:

 Performance counter stats for './a.out':

   14 258 216 181      branch-instructions                                  
    1 513 180 390      branch-misses             #   10,61% of all branches         
   54 898 012 739      cpu-cycles                                           
   42 665 314 382      instructions              #    0,78  insn per cycle  

     18,747075752 seconds time elapsed

I must admit I was puzzled by this huge performance advantage of using higher order functions, especially when there has been no attempt to optimise it at all. It turns out there's a small mistake in the hand written loop. It compares the member .title, which is a std::string with the string literal "Developer". This means it must make character by character comparison. The higher order function has_title() captures a std::string, and equal comparison of std::string begins with checking their lengths. If the lengths are different there's no reason to look at the characters. That is the only reason I've seen for the huge gain.

Changing the hand written loop to compare title with a std::string gives this result:

 Performance counter stats for './a.out':

   13 971 299 475      branch-instructions                                  
    1 449 744 816      branch-misses             #   10,38% of all branches         
   51 964 400 535      cpu-cycles                                           
   39 992 503 929      instructions              #    0,77  insn per cycle  

     17,448599597 seconds time elapsed

So, it is better performing. Not much. But it is. However, the hand written loop gets copied all over the code base, the algorithm and higher order functions do not. The performance bug could've been made in the has_title() function as well, but it would be one place to fix, not many. Likewise, the elements_matching() algorithm could be optimised, for example with partial loop unrolling to do several comparisons per revolution. That too would be an effort spent once in one place, and not all over the code.

That was the results with clang++-3.9. Let's see how g++ 6.2 fares in comparison. First the higher order function version:

 Performance counter stats for './a.out':

   14 006 122 557      branch-instructions                                      
    1 027 464 337      branch-misses             #    7,34% of all branches         
   43 066 722 628      cpu-cycles                                               
   44 394 114 000      instructions              #    1,03  insn per cycle      

     14,505986613 seconds time elapsed

And then the hand written loop, with the fix to compare with std::string:

 Performance counter stats for './a.out':

   14 039 057 224      branch-instructions                                       
    1 852 376 974      branch-misses             #   13,19% of all branches         
   58 797 998 124      cpu-cycles                                                
   40 858 411 471      instructions              #    0,69  insn per cycle       

     19,743206041 seconds time elapsed

Here it's clear that gcc does a slightly worse job than clang with optimising the hand written loop, but does a substantially better job at optimising the higher order function version.

As can be seen, the exact result does vary between compilers, but the chances of losing any performance to speak of when using higher order functions are small, and there is the potential for very noticeable performance gains.

So, use algorithms and use higher order functions, for code clarity and for performance.