Saturday, August 13, 2016

google interview: efficiently find a number in a partial sorted list

// http://www.impactinterview.com/2009/10/140-google-interview-questions/#software_engineer
// You are given a list of numbers.  When you reach the end of the list you
// will come back to the beginning of the list (a circular list).  Write the
// most efficient algorithm to find the minimum # in this list.  Find any given
// # in the list.  The numbers in the list are always increasing but you
// don’t know where the circular list begins, ie: 38, 40, 55, 89, 6, 13, 20,
// 23, 36.

// First of all note very carefully this is for a *unidirectional circular linked
// list* if we alter the design of the data structure there are much better ways
// to do this.. but i believe that is out of bounds as a potential answer.

// hmm..  essentially the find min number algorithm is a component of the find
// any number algorithm..  Since the question isn't written as find the min,
// then fix the list start and then find any # ill assume that it was meant as
// a helper step to get to the find algorithm..  so in the interest of getting
// it done lets just cut to the find any # algorithm and code a few evolutions
// showing efficiency improvements..

#include <iostream>
#include <vector>
#include <chrono>
#include <functional>

// ****************************************
// *************** CYCLIC LIST ************
// ****************************************

class List
{
std::vector<int> data_;
int offset_;

public:
class Iterator
{
List* list_;
int index_;

public:
Iterator() :
list_(NULL),
index_(-1)
{}

Iterator(List* list) :
list_(list),
index_(list->offset_)
{}

// WARNING DO NOT USE in find.. that was not spec'ed
// this is for testing only
Iterator(List* list, int offset) :
list_(list),
index_(offset)
{}

Iterator& operator++()
{
index_ = (index_ + 1) % list_->data_.size();
return *this;
}

Iterator operator+(int delta)
{
int offset = (index_ + delta) % list_->data_.size();
while (offset < 0) offset += list_->data_.size();

return Iterator(list_, offset);
}

// WARNING DO NOT UE in find.. that was not spec'ed
// this is for testing only
Iterator& operator--()
{
index_ = (list_->data_.size() + index_ - 1) % list_->data_.size();
return *this;
}

int operator*()
{
return list_->data_[index_];
}

bool operator==(const Iterator& other) const
{
return other.list_ == list_ and
other.index_ == index_;
}

bool operator!=(const Iterator& other) const
{
return not (other == *this);
}

bool valid() const
{
return list_ != NULL;
}
};

List(std::initializer_list<int> data,
int offset = 0) :
data_(data),
offset_(offset)
{}

List(int size) :
data_(size),
offset_(rand() % size)
{
// random fill
int value = 0;
for (int i = 0; i < size; ++i)
{
value += 2 + rand() % 5;
data_[i] = value;
}
}

bool empty() const
{
return data_.empty();
}

Iterator begin()
{
return Iterator(this);
}

Iterator random()
{
return Iterator(this, rand() % data_.size());
}
};

// ****************************************
// *************** FIND NAIVE *************
// ****************************************

List::Iterator find_naive(List& data, int target)
{
if (data.empty()) return List::Iterator();

List::Iterator start = data.begin();
List::Iterator i = start;

do
{
if (*i == target) return i;
++i;
}
while (i != start);

return List::Iterator();
}

// ****************************************
// ************* FIND NAIVE V2 ************
// ****************************************

// Attempt to simply the code a bit so that we handle the cyclic iteration and
// partial sorted issue cleaner for latter scaling

List::Iterator find_naive2(List& data, int target)
{
if (data.empty()) return  List::Iterator();

//termination is an issue because start *is* the end ..
// simplify this by checking the start..
List::Iterator end = data.begin();
if (target == *end) return end;

// then start at the first item after and now "start" is really the end
List::Iterator i = end + 1;

while (i != end and *i != target) ++i;
if (*i == target) return i;

return  List::Iterator();
}

// Of course you could try a divide and conquer idea..  but your dealing with a
// list that you want to divide in half.  Dividing it it half means you need to
// step the iterator the mid point..  So assuming you know the size of the list
// the total search on a fully sorted list will *always* cost u n/2 + n/4 + n/8
// ...  moves. Which totals to n-1 (ie if n=8 is 4+2+1=7)..  So the performance
// is actually worst on average for the naive version ... of course this might
// not be true because the cost of iterators might be very low compared to the
// compare.. but i dont think it is... so im skipping it..

// ****************************************
// ****************************************

// What we can really do is try to take advantage of the sorted order of data
// and terminate early..  In order to do that what you have to do is skip to
// the correct sorted part for your target and then search... (hence the find
// min number stuff before)

// Keep in mind this gets much less robust to problems with the data..  ie
// if the list isnt sort (and parted) before hand..

List::Iterator find_skip_start(List& data, int target)
{
// handle the empty dataset case..
if (data.empty()) return  List::Iterator();

//termination is an issue because start *is* the end ..
// simplify this by checking the start..
List::Iterator end = data.begin();
if (target == *end) return end;

// then start at the first item after and now "start" is really the end
List::Iterator i = end + 1;

// if the target is in the second part
if (target < *end)
{
int split_edge = *end;
while (*i >= split_edge and i != end) ++i;
}

// then search for the target
while (*i < target and i != end) ++i;

if (*i == target) return i;
return  List::Iterator();
}

// ****************************************
// ******* DOUBLE SKIP IN SECTION *********
// ****************************************

// Now because its sorted you dont have to check items every time you move
// forward.  You can step and then roll back if you went to far..  this makes
// some big assumptions about the cost of iterator movement vs integer
// compares.. my "List" class isnt great but it works out.. and in
// theory should cut out every second compare..

List::Iterator find_double_skip(List& data, int target)
{
// handle the empty dataset case..
if (data.empty()) return  List::Iterator();

//termination is an issue because start *is* the end ..
// simplify this by checking the start..
List::Iterator end = data.begin();
if (target == *end) return end;

// then start at the first item after and now "start" is really the end
List::Iterator i  = end + 1;

if (target < *end)
{
int split_edge = *end;

List::Iterator iprev;
while (*i >= split_edge and i != end)
{
++i;
iprev = i;
if (i != end) ++i;
}

// check if we need to back up the double step.
if (iprev.valid() and *iprev < split_edge) i = iprev;
}

{
// double skip till target found or exceeded
List::Iterator iprev;
while (*i < target and i != end)
{
++i;
iprev = i;
if (i != end) ++i;
}

// check if found the target at either of the pointers..
if (iprev.valid() and *iprev == target) return iprev;
if (*i == target) return i;
}

return  List::Iterator();
}

// ****************************************
// ********** N SKIP IN SECTION ***********
// ****************************************

// Optimization..  of course if u can do better with steps of 2 then u can
// generalize it to any number of steps..  for example using steps of 20..  now
// we are really playing balance games..  trading the speeds of various
// comparisons, assignments etc..  Then u have to tune the size of skip for

List::Iterator find_n_skip(List& data, int target)
{
const int skip = 20;

// handle the empty dataset case..
if (data.empty()) return  List::Iterator();

//termination is an issue because start *is* the end ..
// simplify this by checking the start..
List::Iterator end = data.begin();
if (target == *end) return end;

// then start at the first item after and now "start" is really the end
List::Iterator i  = end + 1;

if (target < *end)
{
List::Iterator iprev = i;
int split_edge = *end;
while (*i >= split_edge and i != end)
{
++i;
iprev = i;
int steps = skip;
while (--steps > 0 and i != end) ++i;
}

// when found return to the last known ok point and search forward 1 by 1
i = iprev;
while (*i >= split_edge and i != end) ++i;
}

{
List::Iterator iprev = i;
while (*i < target and i != end)
{
++i;
iprev = i;
int steps = skip;
while (--steps > 0 and i != end) ++i;
}

// when found return to the last known ok point and search forward 1 by 1
i = iprev;
while (*i < target and i != end) ++i;

if (*i == target) return i;
}

return  List::Iterator();
}

void test(std::string dataTag,
std::string algoTag,
std::function<List::Iterator (List& , int)> func,
List& data,
int target,
bool hasTarget)
{
auto start = std::chrono::system_clock::now();
List::Iterator result = func(data, target);
auto end   = std::chrono::system_clock::now();

std::cout << algoTag << " for " << dataTag << ": ";
if (not result.valid())
std::cout << (hasTarget ? "FAIL" : "pass");
else
std::cout << ((hasTarget and (target == *result)) ? "pass" : "FAIL");

std::cout << " -- time:"
<< std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count()
<< " nSec" << std::endl;
}

void testSet(std::string tag,
List& data,
int target,
bool hasTarget)
{
test(tag, "naive ",  find_naive,       data, target, hasTarget);
test(tag, "naive2",  find_naive2,      data, target, hasTarget);
test(tag, "skip  ",  find_skip_start,  data, target, hasTarget);
test(tag, "2xskip",  find_double_skip, data, target, hasTarget);
test(tag, "nskip ",  find_n_skip,      data, target, hasTarget);
std::cout << "\n";
}

int main()
{
srand(std::chrono::system_clock::now().time_since_epoch().count());

{
List list({10,11,12,1,2,3,4,5,6,7,8});
testSet("small     ", list, 4, true);
}

{
List list(1000000);
testSet("big rand  ", list, *list.random(), true);
testSet("big nrand ", list, *list.random()-1, false);  // note that  random always is +2 ~ +7 so -1 means i will not be in the data..

testSet("big early ", list, *(list.begin() + 10), true);    // i hope.. its random so it might not be..
testSet("big nearly", list, *(list.begin() + 10)-1, false); // check early exit in first half on fail

testSet("big last  ", list, *(--list.begin()), true);
testSet("big nlast ", list, -1, false);
}
}



naive  for small     : pass -- time:3207 nSec
naive2 for small     : pass -- time:2175 nSec
skip   for small     : pass -- time:1996 nSec
2xskip for small     : pass -- time:2151 nSec
nskip  for small     : pass -- time:4232 nSec

naive  for big rand  : pass -- time:14229998 nSec
naive2 for big rand  : pass -- time:13228995 nSec
skip   for big rand  : pass -- time:12786059 nSec
2xskip for big rand  : pass -- time:10874261 nSec
nskip  for big rand  : pass -- time:9004453 nSec

naive  for big nrand : pass -- time:43750054 nSec
naive2 for big nrand : pass -- time:37382249 nSec
skip   for big nrand : pass -- time:17939681 nSec
2xskip for big nrand : pass -- time:15509062 nSec
nskip  for big nrand : pass -- time:12927864 nSec

naive  for big early : pass -- time:738 nSec
naive2 for big early : pass -- time:611 nSec
skip   for big early : pass -- time:632 nSec
2xskip for big early : pass -- time:708 nSec
nskip  for big early : pass -- time:1064 nSec

naive  for big nearly: pass -- time:28980196 nSec
naive2 for big nearly: pass -- time:27551947 nSec
skip   for big nearly: pass -- time:1034 nSec
2xskip for big nearly: pass -- time:719 nSec
nskip  for big nearly: pass -- time:1217 nSec

naive  for big last  : pass -- time:27341796 nSec
naive2 for big last  : pass -- time:27780499 nSec
skip   for big last  : pass -- time:28165415 nSec
2xskip for big last  : pass -- time:24593599 nSec
nskip  for big last  : pass -- time:22455148 nSec

naive  for big nlast : pass -- time:27195820 nSec
naive2 for big nlast : pass -- time:26584976 nSec
skip   for big nlast : pass -- time:8640245 nSec
2xskip for big nlast : pass -- time:7304119 nSec
nskip  for big nlast : pass -- time:6601818 nSec


Tuesday, August 9, 2016

google interview: measure the speed of a context switch

// http://www.impactinterview.com/2009/10/140-google-interview-questions/#software_engineer
// Write a C program which measures the the speed of a context switch on a
// UNIX/Linux system.

// wow.. just wow... the questioner hates you.. so moving on...

// A context switch occurs when the kernel or hardware thinks its time to do
// something else...  you have no control of it..  *period*...  the only true
// way to measure it is to get into the kernel remove *everything* and check it
// directly..  without running any other programs and device drivers etc, i
// mean hell painting the screen, that blinking cursor your looking at while
// typing thats *all* a context switch.

// so write code to measure this.. well.. i dont know the kernel so thats out..
// so i dont have the skills to do the *real* job.. the best i can do is guess..

// basically i can measure the whole process: program executing, ctx switch, do
// something else, ctx switch, complete execution

// so we are going to try to measure something that should have very little
// variation and should scale well..  ie multiply and accumulate a register.
// we will do the ops few times and profile them. Then change the total number
// of times we do it and reprofile..  of course smaller context switches may
// be getting buried in the testing method etc..  not to mention the timers
// effects etc etc

// what we are looking for is a statistical blip...  the mult accum should just
// shift right when we scale it..  anything that stays still isnt part of our
// test..  its likely to be some part of the kernels context switching or device
// drivers

// Anyway results looked as expected..  the test execution time moved to the
// right when the number of cycles was increased..  there is a set of a large
// reoccurring spikes in the execution times..  but i doubt these are the cost
// of a context switch, more likely some device driver, service, interrupt
// handlers bumping into my execution etc

#include <iostream>
#include <chrono>
#include <map>

int main()
{
std::map<int,int> counts;
for (int k = 0; k < 100; ++k)
{
std::cout << ".";
for (int j = 0; j < 1000000; ++j)
{
auto start = std::chrono::system_clock::now();
for (int i = 0; i < 100000; ++i)
{
i += i*i;
}
auto end = std::chrono::system_clock::now();

int delta = std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count();

// sigh.. crud but will work for now
if (counts.find(delta) != counts.end())
++(counts[delta]);
else
counts[delta] = 1;
}
}
std::cout << "\n";

for (std::map<int,int>::iterator cit = counts.begin();
cit != counts.end();
++cit)
{
std::cout << " " << cit->first << ":" << cit->second << "\n";
}
}


The results copied out into open office and graphed etc. Note that the "fuzz" in the red section maybe significant also.. but im not really a maths guy so I cant be certain, and my use of stats here is really flimsy so grain of salt people... Update: i just noticed that i didnt align the graphs very well(this wasn't intentional) anyway the mistake is a bit disadvantageous to the results so ill leave it be.

Saturday, August 6, 2016

google interview: detect a cycle in a graph

// detect a cycle in a directed graph

#include <iostream>
#include <vector>
#include <unordered_set>
#include <unordered_map>
#include <memory>
#include <algorithm>
#include <functional>
#include <chrono>

struct Node
{
static int gID;
int id_;

std::vector<Node*> kids_;

Node() : id_(++gID) {}
void link(Node& kid) { kids_.push_back(&kid); }
};

int Node::gID = 0;

// *************************************************
// *************** PRE BOARD WORK  *****************
// *************************************************
//
//  Basic loop with Node "R" going into unknown "Graph"
//                  +---+    << check for self link node in path
//    ~~~~~~~~~     |   |
//   (         ) <--R<--+
//    ) Graph (     ^
//   (         ) -?-+
//    ~~~~~~~~~
//
// Recursive case for "A" a child Node of "R"
//                  +---+---+    << check for prior node in path
//    ~~~~~~~~~     |   v   |
//   (         ) <--A<--R<--+
//    ) Graph (     ^   ^
//   (         ) -?-+-?-+
//    ~~~~~~~~~
//
// Recursive case for "B" with child Node "A"
//                  +---+---+--+    << check for *ANY* prior node in path
//    ~~~~~~~~~     |   v   v  |
//   (         ) <--B<--A<--R<-+
//    ) Graph (     ^   ^   ^
//   (         ) -?-+-?-+-?-+
//    ~~~~~~~~~
//
// etc
//
// Therefore first algo is to DFS visit all nodes and check if current nodes
// kids loop back into the path of nodes before the current one.

// *************************************************
// ***************** BASIC RECURSE *****************
// *************************************************

bool hasCycle_recurseImpl(Node* v,
std::vector<Node*>& path)
{
// Mark the current node as part of pathway
path.push_back(v);

// then for all kids
for(Node* i : v->kids_)
{
// if the kid is path of the path you have a loop done
std::vector<Node*>::iterator pit =
std::find(path.begin(), path.end(), i);

if (pit != path.end())
return true;

// recurse into kid
if (hasCycle_recurseImpl(i, path) )
return true;
}

// remove the node from path
path.pop_back();
return false;
}

bool hasCycle_recurse(Node* root)
{
std::vector<Node*> path;

if (root == NULL) return false;

return hasCycle_recurseImpl(root, path);
}

// *************************************************
// ***************** UNORDERED PATH ****************
// *************************************************

// optimization of path - ordering is of the path is *not* important.  As a
// result we can choose a data structure with O(1) search, insert and delete ie
// a hash

bool hasCycle_unordPathImpl(Node* v,
std::unordered_set<Node*>& path)
{
// Mark the current node as part of pathway
path.insert(v);

// then for all kids
for(Node* i : v->kids_)
{
// if the kid is path of the path you have a loop done
if (path.find(i) != path.end())
return true;

// recurse into kid
if (hasCycle_unordPathImpl(i, path) )
return true;
}

// remove the node from path
path.erase(v);
return false;
}

bool hasCycle_unordPath(Node* root)
{
std::unordered_set<Node*> path;

if (root == NULL) return false;

return hasCycle_unordPathImpl(root, path);
}

// *************************************************
// ****************** NO REVISIT  ******************
// *************************************************

// optimization of visiting - never revisit a node.  The logic behind this is
// easy but it took me a bit to actually see it.
//
// If you visit a node then the DFS will terminate on 2 possible conditions.
// 1) you found a cycle,
// 2) you exhausted all possible nodes below this one and didn't end up in a
// cycle.
//
// Now here is what took me a bit to get..  If you ever enter a path that forms
// a cycle you have no possible way to exit unless you found end of the cycle
// that means that you can never see a visited node unless it is about to start
// a cycle or has been exhausted and will *never* form a cycle even if we add to
// the current path. IE if it could have formed a cycle with the current node
// then it should have traversed into the current node already before it was
// exhausted.

bool hasCycle_1visitImpl(Node* v,
std::unordered_set<Node*>& visited,
std::unordered_set<Node*>& path)
{
if(visited.find(v) == visited.end())
{
// Mark the current node as visited and part of pathway
visited.insert(v);
path.insert(v);

// then for all kids
for(Node* i : v->kids_)
{
// if the kid is path of the path you have a loop done
if (path.find(i) != path.end())
return true;

// recurse into kid
if (hasCycle_1visitImpl(i, visited, path) )
return true;
}

// remove the node from path
path.erase(v);
}
return false;
}

bool hasCycle_1visit(Node* root)
{
std::unordered_set<Node*> visited;
std::unordered_set<Node*> path;

if (root == NULL) return false;

return hasCycle_1visitImpl(root, visited, path);
}

// *************************************************
// *** CALL STACK, PATH and VISIT DUPLICATION  *****
// *************************************************

// call stack, path hash and visit hash optimization -
// the recursive call stack is basically the same as the path stack and the path
// stack is the same as the visit stack all 3 data structures can be made into
// one..  using an iterative approach that stores the parent node to remove call
// stack, and a stateful hash entry to double as visit and child idx marking which
// kid is currently in use when in the path

bool hasCycle_iterative(Node* root)
{
typedef std::pair<Node*,int>  ParentKidIdx;
std::unordered_map<Node*, ParentKidIdx> path;

if (root == NULL) return false;

// note the oddity with the new nodes current kids we start with the end kid
// this way it naturally ends in a negative idx when kids are exhusted
path[root] = ParentKidIdx(NULL, root->kids_.size());
Node* current  = root;

while (current != NULL)
{
std::unordered_map<Node*, ParentKidIdx>::iterator pit
= path.find(current);

if (pit == path.end())
std::cout << "IMPOSSIBLE! we have a node without a state\n";

// just for easy of undertanding std::pair's member names suck..
// and as u can see this is a bit confusing
Node*  parent = pit->second.first;
int&   kidIdx = pit->second.second;  // note its *writable*

// find the next kid that that current node should work on ..
Node* newKid = NULL;
while (kidIdx >= 0 and newKid == NULL)
{
// try to get the next kid
--kidIdx;
if (kidIdx >= 0)
{
// extract the kid
newKid = current->kids_[kidIdx];

// check if the kid has a state already
std::unordered_map<Node*, ParentKidIdx>::iterator kit
= path.find(newKid);

if (kit != path.end())
{
// child is has aleady been visited but whats its current
// state at??
if (kit->second.second >= 0)
// it is activite and in the current path..so this is a
// cycle
return true;

// it was visited and exhusted..  waste of time move to
// next one
newKid = NULL;
}
else
{
// kid has never been visied before make its init state
// entry and then visited it
path[newKid] = ParentKidIdx(current,
newKid->kids_.size());
}
}
}

if (newKid != NULL)
{
// we found a new kid to explore.. jump in to it
current = newKid;
}
else
{
// we have exhusted all the current ones kids..  move up to the
// parent
current = parent;
}
}
return false;
}

// *************************************************
// ******************* TEST ************************
// *************************************************

void testCase(std::string tag,
std::function<bool(Node*)> hasCycle,
Node* root,
bool expected)
{
auto start = std::chrono::system_clock::now();
bool result = hasCycle(root);
auto end   = std::chrono::system_clock::now();

std::cout << tag << ": " << (expected == result ? "pass" : "FAIL")
<< " -- time:"
<< std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count()
<< " nSec" << std::endl;
}

void testCaseSet(std::string set,
Node* root,
bool expected)
{
std::cout << "\n *** " << set << " ***\n";
testCase("hasCycle_recurse", hasCycle_recurse,       root, expected);
testCase("hasCycle_unord  ", hasCycle_unordPath,     root, expected);
testCase("hasCycle_1visit ", hasCycle_1visit,        root, expected);
testCase("hasCycle_iter   ", hasCycle_iterative,     root, expected);
}

void test()
{
{
testCaseSet("null",
NULL, false);
}

{
Node a0;
testCaseSet("single node",
&a0, false);
}

{
Node a0;
testCaseSet("single node cycle",
&a0, true);
}

{
Node a0;
Node a1;
testCaseSet("two node",
&a0, false);
}

{
Node a0;
Node a1;

&a0, false);
}

{
Node a0;
Node a1;

testCaseSet("two node cycle",
&a0, true);
}

{
// visit breaker
Node a0;
Node a1;
Node a2;

testCaseSet("triple node double end link - reversed",
&a0, false);
}

{
// visit breaker
Node a0;
Node a1;
Node a2;

&a0, false);
}

{
Node a0;
Node a1;
Node a2;
Node a3;

testCaseSet("some random tree",
&a0, true);
}

{
// graph speed breaker wide first node with repeated long tail..
Node chest;
std::vector<Node> shoulders(100);
for (int i = 0; i < shoulders.size(); ++i)
{
}

std::vector<Node> tail(100);
for (int i = 1; i < tail.size(); ++i)
{
}

Node looper;

}
std::cout << std::endl;
}

int main()
{
test();
}


*** null ***
hasCycle_recurse: pass -- time:2574 nSec
hasCycle_unord  : pass -- time:13314 nSec
hasCycle_1visit : pass -- time:4531 nSec
hasCycle_iter   : pass -- time:4097 nSec

*** single node ***
hasCycle_recurse: pass -- time:4975 nSec
hasCycle_unord  : pass -- time:12330 nSec
hasCycle_1visit : pass -- time:9873 nSec
hasCycle_iter   : pass -- time:9279 nSec

*** single node cycle ***
hasCycle_recurse: pass -- time:4614 nSec
hasCycle_unord  : pass -- time:8321 nSec
hasCycle_1visit : pass -- time:9863 nSec
hasCycle_iter   : pass -- time:8335 nSec

*** two node ***
hasCycle_recurse: pass -- time:19714 nSec
hasCycle_unord  : pass -- time:9582 nSec
hasCycle_1visit : pass -- time:13318 nSec
hasCycle_iter   : pass -- time:11426 nSec

*** two node - double link ***
hasCycle_recurse: pass -- time:5507 nSec
hasCycle_unord  : pass -- time:11714 nSec
hasCycle_1visit : pass -- time:13890 nSec
hasCycle_iter   : pass -- time:11678 nSec

*** two node cycle ***
hasCycle_recurse: pass -- time:5299 nSec
hasCycle_unord  : pass -- time:7683 nSec
hasCycle_1visit : pass -- time:12348 nSec
hasCycle_iter   : pass -- time:9779 nSec

*** triple node double end link - reversed ***
hasCycle_recurse: pass -- time:7694 nSec
hasCycle_unord  : pass -- time:13821 nSec
hasCycle_1visit : pass -- time:18194 nSec
hasCycle_iter   : pass -- time:12800 nSec

*** triple node double end link ***
hasCycle_recurse: pass -- time:7322 nSec
hasCycle_unord  : pass -- time:13999 nSec
hasCycle_1visit : pass -- time:16998 nSec
hasCycle_iter   : pass -- time:12793 nSec

*** some random tree ***
hasCycle_recurse: pass -- time:4840 nSec
hasCycle_unord  : pass -- time:7792 nSec
hasCycle_1visit : pass -- time:11853 nSec
hasCycle_iter   : pass -- time:15752 nSec

hasCycle_recurse: pass -- time:18295450 nSec
hasCycle_unord  : pass -- time:14848133 nSec
hasCycle_1visit : pass -- time:615861 nSec
hasCycle_iter   : pass -- time:512595 nSec

*** head shoulder tail - looped ***
hasCycle_recurse: pass -- time:131233 nSec
hasCycle_unord  : pass -- time:136192 nSec
hasCycle_1visit : pass -- time:242850 nSec
hasCycle_iter   : pass -- time:176559 nSec



Wednesday, August 3, 2016

Google interview: Implement an AVL tree ...

Seems to me that google is most likely to ask about Tree inserts over erase. Also they might ask for traversal, infix printing etc etc... basically the cheapest way to do this is just code the whole AVL tree eco system period...

AVL trees look complex but they basically are a BST tree with an additional pair of rebalancing rotations that you apply to them after an insert or erase.. this makes them somewhat close to the theoretical log(n) depth of a well balanced BST tree...

There are 2 rotate operations that are depicted as follows. they are reversed depending of the side that are viewing them from (some people call them 4 with i think just over complicates them), personally i know them as inside and outside rotates, you can tell by looking at the position of the nodes grand child.. if its under the grandparent its inside, otherwise its outside.

a) grandchild-outside-child (left side)
T1, T2, T3 and T4 are subtrees.
z                            y
/ \                         /   \
y   T4  Right Rotate (z)    x      z
/ \      --------------->  /  \    /  \
x   T3                     T1  T2  T3  T4
/ \
T1   T2

b) grandchild-inside-child (left-side)
z                            z                            x
/ \                         /   \                        /   \
y   T4  Left Rotate (y)     x    T4   Right Rotate(z)   y      z
/ \      -------------->    /  \       -------------->  / \    / \
T1   x                       y    T3                     T1  T2 T3  T4
/ \                     / \
T2   T3                 T1   T2


The trick to knowing when to perform the rebalance ops is to check the height balance between the nodes left and right side of the current one. If the height balance is off by more than +/- 1 you need to rotate, the sign tells you the direction. Then to determine if its the inner or outer case you check the balance of the child node.. as you can see from the diagram above the outer case should already have an inbalance in the *same* direction(sign) as the parents inbalance if it does not then you have the inner case and need to adjust it to make the other case before doing the rotate to correct..

#include <stdint.h>
#include <vector>
#include <algorithm>
#include <iostream>
#include <functional>

template <typename T>
class AvlTree
{
public:
// late binding via constructor - dont pollute type
typedef std::function<bool(const T&, const T&)> Comparer;

private:
struct Node
{
uint32_t height_;
Node*    left_;
Node*    right_;
T        item_;

Node(const T& item) :
height_(1),
left_(NULL),
right_(NULL),
item_(item)
{}

~Node()
{
if (left_  != NULL) delete left_;
if (right_ != NULL) delete right_;
}

int height() const
{
int leftHeight  = (left_  != NULL) ? left_->height_  : 0;
int rightHeight = (right_ != NULL) ? right_->height_ : 0;

return std::max(leftHeight, rightHeight) + 1;
}

int balance() const
{
int leftHeight  = (left_  != NULL) ? left_->height_  : 0;
int rightHeight = (right_ != NULL) ? right_->height_ : 0;

return leftHeight - rightHeight;
}

};
public:
class Iterator
{
// infix DFS reshaped into stack for an iterator.. fun but crazy...
//
// Ok..  So i got creative here..  rather than use parent pointers in the
// tree I implemented the iterator as a DFS parent stack..  now you do
// *not* want to do this in production because its actually a lot more
// fragile than the parent pointer solution..
//
// consider the following trade offs
// - the find operation needs to create the stack which is costly and
//  iterator returned from the find will often *not* move making it
//  needless
// - the iterator is less robust to tree updates.  with a parent
//  pointer the iterator just points at one node so changing anything that
//  is not the current node is generally ok..  However with the stack
//  implementation any update that effects the parent chain breaks the
//  iterators stack and makes a large issue

// TODO can a bi-directional version be made??...

typedef std::vector<Node*> Stack;

Stack stack_;

void enstackLeft(Node* node);
void next();

public:
Iterator(Node* root);   // not publicily usable because Node is not public!
Iterator();
Iterator& operator++();
void      operator++(int);
const T&  operator*() const;
const T*  operator->() const;

T&  operator*() ;
T*  operator->();

int height() const; // yep mostly for testing

bool  operator==(const Iterator& other) const;
bool  operator!=(const Iterator& other) const;
};

class Handle
{
// a light weight *immovable* iterator .. so *not* an iterator
// TODO it *is* possible to lazy covert to a real iterator if ++'ed
//  but whatever the stacked based iterator is crazy
Node* node_;
public:
Handle() :
node_(NULL)
{}

Handle(Node* node) :
node_(node)
{}

const T& operator*()  const { return node_->item_; }
T& operator*()        { return node_->item_; }

const T* operator->() const { return &(node_->item_); }
T* operator->()       { return &(node_->item_); }

operator bool () const      { return node_ != NULL; }
};

private:
Node*   root_;
Comparer cmp_;

// TODO compilation firewalling issues...
//  ie can this be converted to a base Node without T and a factory to
//  create and handle Nodes with T.

void rotateLeft (Node*& node);
void rotateRight(Node*& node);
void rebalance  (Node*& node);

void insertImpl (const T& item,
Node*& node);

void replacementEraseRight(Node*& node,
Node*& other);

void eraseImpl(const T& item,
Node*& node);

public:
AvlTree() :
root_(NULL),
cmp_(std::less<T>())
{}

AvlTree(Comparer cmp) :
root_(NULL),
cmp_(cmp)
{}

~AvlTree()
{
if (root_ != NULL) delete root_;
}

void insert(const T& item);
void erase(const T& item);
Handle find(const T& item);
void print(std::ostream& os);

Iterator begin();
Iterator end();
};

// *************************************************
// ******************** ITERATOR *******************
// *************************************************

template <typename T>
AvlTree<T>::Iterator::Iterator() :
stack_()
{}

template <typename T>
AvlTree<T>::Iterator::Iterator(AvlTree<T>::Node* root) :
stack_()
{
//when we add new nodes we go down the left side and enstack then
enstackLeft(root);
}

template <typename T>
void AvlTree<T>::Iterator::enstackLeft(typename AvlTree<T>::Node* node)
{
while (node != NULL)
{
stack_.push_back(node);
node = node->left_;
}
}

template <typename T>
void AvlTree<T>::Iterator::next()
{
// node has just completed left side and self visit.. remove me from the stack
Node* current = stack_.back();
stack_.pop_back();

// and enstack add my right branch
enstackLeft(current->right_);
}

template <typename T>
const T& AvlTree<T>::Iterator::operator*() const
{
return stack_.back()->item_;
}

template <typename T>
T& AvlTree<T>::Iterator::operator*()
{
return stack_.back()->item_;
}

template <typename T>
const T* AvlTree<T>::Iterator::operator->() const
{
return &(stack_.back()->item_);
}

template <typename T>
T* AvlTree<T>::Iterator::operator->()
{
return &(stack_.back()->item_);
}

template <typename T>
int AvlTree<T>::Iterator::height() const
{
return stack_.back()->height_;
}

template <typename T>
typename AvlTree<T>::Iterator& AvlTree<T>::Iterator::operator++()
{
next();
return *this;
}

// NOTE void on x++ to termiate the post iterator junk
template <typename T>
void AvlTree<T>::Iterator::operator++(int)
{
// Node* prev = stack_.back();
next();
// return ...;
}

template <typename T>
bool AvlTree<T>::Iterator::operator==(const AvlTree<T>::Iterator& other) const
{
// TODO short cut the last item?? would be quicker
return other.stack_ == stack_;
}

template <typename T>
bool AvlTree<T>::Iterator::operator!=(const AvlTree<T>::Iterator& other) const
{
// TODO short cut the last item?? would be quicker
return not (other == *this);
}

template <typename T>
typename AvlTree<T>::Iterator AvlTree<T>::begin()
{
return Iterator(root_);
}

template <typename T>
typename AvlTree<T>::Iterator AvlTree<T>::end()
{
return Iterator();
}

// *************************************************
// ********************** FIND *********************
// *************************************************

// this is a bit of a problem using stack based iterators makes find
// more weighty using nodes exposes the internals
// lightwieght_iterator?? one that cant move but gives you access.. a "handle"
template <typename T>
typename AvlTree<T>::Handle AvlTree<T>::find(const T& item)
{
Node* node = root_;
while (node != NULL)
{
if      (cmp_(item, node->item_))   // less *than*
{
node = node->left_;
}
else if (cmp_(node->item_, item))   // greater *than*
{
node = node->right_;
}
else
{
return typename AvlTree<T>::Handle(node);
}
}
return typename AvlTree<T>::Handle();
}

// *************************************************
// ********************** PRINT ********************
// *************************************************

template <typename T>
void AvlTree<T>::print(std::ostream& os)
{
for (AvlTree<uint32_t>::Iterator it = begin();
it != end();
++it)
{
os << *it << "(" << it.height()  << ")" << " ";
}
}

template <typename T>
std::ostream& operator<<(std::ostream& os, AvlTree<T>& tree)
{
tree.print(os);
return os;
}

// *************************************************
// ******************** REBALANCE ******************
// *************************************************

template <typename T>
void AvlTree<T>::rotateLeft(typename AvlTree<T>::Node*& node)
{
// given the current node make its left the top
Node* right  = node->right_;
if (right == NULL) return;  // this will explode.. stop!

node->right_ = right->left_;
right->left_  = node;

node->height_  = node->height();
right->height_ = right->height();

// keep in mind that this is using Node*& so the following updates the
// nodes parent left/right or the root!
node = right;
}

template <typename T>
void AvlTree<T>::rotateRight(typename AvlTree<T>::Node*& node)
{
// given the current node make its left the top
Node* left  = node->left_;
if (left == NULL) return; // this will explode.. stop!

node->left_ = left->right_;
left->right_ = node;

node->height_ = node->height();
left->height_ = left->height();

// keep in mind that this is using Node*& so the following updates the
// nodes parent left/right or the root!
node = left;
}

// attempt to rebalance in a general way for both insert delete..
template <typename T>
void AvlTree<T>::rebalance(typename AvlTree<T>::Node*& node)
{
if (node == NULL) return;

// correct node height..
node->height_ = node->height();

// and compute AVL rotations
int balance = node->balance();

// done if balance is good...
if (balance > -2 and balance < 2)  return;

Node* child = NULL;
if (balance < -1)
{
// right side is heavy
child = node->right_;
}
else if (balance > 1)
{
// left side is heavy
child = node->left_;
}

int childBalance = child->balance();

// std::cout << "   ++ childBalance: " << childBalance << "\n";

// now rotate to fix inbalance
if (balance > 0)
{
// planning to rotate right..
if (childBalance <= 0)
//but if the inner child left heavy or balanced correct it
rotateLeft(node->left_);

// left side so rotate right
rotateRight(node);
}
else
{
// planning to rotate left..
if (childBalance >= 0)
//but if the inner child right heavy or balanced correct it
rotateRight(node->right_);

rotateLeft(node);
}
}

// *************************************************
// ********************* INSERT ********************
// *************************************************

template <typename T>
void AvlTree<T>::insertImpl(const T& item,
typename AvlTree<T>::Node*& node)
{
// BST tree insert
if (node == NULL)
{
node = new Node(item);
return;
}

if (cmp_(item, node->item_))
{
insertImpl(item, node->left_);
}
else
{
insertImpl(item, node->right_);
}

// now AVL tree rebalance..
rebalance(node);
}

template <typename T>
void AvlTree<T>::insert(const T& item)
{
insertImpl(item, root_);
}

// *************************************************
// ********************* ERASE *********************
// *************************************************

template <typename T>
void AvlTree<T>::replacementEraseRight(typename AvlTree<T>::Node*& node,
typename AvlTree<T>::Node*& other)
{
// we are no longer searching for the erase target..
// we are now searching for the next object after it
// the right side min

if (other->left_ != NULL)
{
// go right to the end of the left branching and do replacement erase
replacementEraseRight(node,other->left_);

// then rebalance the tree from here up
rebalance(other);
}
else
{
// swap replacement with target node so nothing odd happens
std::swap(node->item_, other->item_);

// now unlink and erase other from tree
Node* right = other->right_;
other->right_ = NULL;
delete other;

// relink others child in its place
other = right;
}
}

template <typename T>
void AvlTree<T>::eraseImpl(const T& item,
typename AvlTree<T>::Node*& node)
{
// BST tree erase
// didnt exist in the tree??
if (node == NULL) return;

//
if (cmp_(item, node->item_))          // a < b
{
eraseImpl(item, node->left_);
}
else if (cmp_(node->item_, item))     // b < a
{
eraseImpl(item, node->right_);
}
else                                  // b == a
{
// match!
if ((node->right_ == NULL) or (node->right_ == NULL))
{
Node* child = (node->right_ == NULL) ? node->left_ : node->right_;

// unlink node.. and kill it
node->right_ = NULL;
node->left_ = NULL;
delete node;

// replace with child
node = child;
}
else
{
// special case both right and left have stuff.. but we have 1 parent
// means we have to take a one of the closet (in value) nodes and replace this one
// we have 2 choices for the replacement the max of left or min of right...
replacementEraseRight(node, node->right_);

// TODO.. shouldn't this do a balance check and then go left or right based on the heavy branch??
}
}

// and then AVL tree rebalance..
rebalance(node);
}

template <typename T>
void AvlTree<T>::erase(const T& item)
{
eraseImpl(item, root_);
}

// *************************************************
// ********************* AVL MAP *******************
// *************************************************

// TODO cleanup AvlTree to AvlMap coverison class
template <typename Key,
typename Value>
class AvlMap
{
typedef std::pair<Key,Value> Pair;
typedef std::function<bool(const Pair&, const Pair&)> Comparer;
typedef AvlTree<Pair> Tree;

AvlTree<Pair> tree_;
public:
typedef typename Tree::Iterator Iterator;

AvlMap() :
tree_([] (const Pair& v1, const Pair& v2) { return v1.first < v2.first; })
{}

Value& operator[](const Key& key)
{
// hmm not such a great implementation...
Pair target(key,Value());
typename AvlTree<Pair>::Handle it = tree_.find(target);
if (it) return it->second;

tree_.insert(target);
return tree_.find(target)->second;
}

Iterator begin() { return tree_.begin(); }
Iterator end()   { return tree_.end();   }
};

// *************************************************
// *********************** TEST ********************
// *************************************************

void test()
{
// TODO honestly this inst real testing.. need to make it check the results..
{
AvlTree<uint32_t> tree;

tree.begin();
tree.end();

tree.find(1);
}

{
AvlTree<uint32_t> tree;

tree.insert(50); std::cout << tree << "\n";
tree.insert(25); std::cout << tree << "\n";
tree.insert(10); std::cout << tree << "\n";
tree.insert(40); std::cout << tree << "\n";
tree.insert(30); std::cout << tree << "\n";
tree.insert(45); std::cout << tree << "\n";
tree.insert(75); std::cout << tree << "\n";

AvlTree<uint32_t>::Handle handle = tree.find(75);
std::cout << "find:" << *handle << "\n";

tree.erase(50); std::cout << tree << "\n";
tree.erase(25); std::cout << tree << "\n";
tree.erase(10); std::cout << tree << "\n";
tree.erase(40); std::cout << tree << "\n";
tree.erase(30); std::cout << tree << "\n";
tree.erase(45); std::cout << tree << "\n";
tree.erase(75); std::cout << tree << "\n";
}

// outer rotate check
{
AvlTree<uint32_t> tree;

tree.insert(15); std::cout << tree << "\n";
tree.insert(77); std::cout << tree << "\n";
tree.insert(35); std::cout << tree << "\n";

// pair erase check
tree.erase( 1); std::cout << tree << "\n";
tree.erase(35); std::cout << tree << "\n";
tree.erase(15); std::cout << tree << "\n";
tree.erase(77); std::cout << tree << "\n";
tree.erase( 1); std::cout << tree << "\n";
}

{
AvlTree<uint32_t> tree;

tree.insert(93); std::cout << tree << "\n";
tree.insert(86); std::cout << tree << "\n";
tree.insert(87); std::cout << tree << "\n";

// pair erase check
tree.erase(99); std::cout << tree << "\n";
tree.erase(86); std::cout << tree << "\n";
tree.erase(87); std::cout << tree << "\n";
tree.erase(93); std::cout << tree << "\n";
tree.erase(99); std::cout << tree << "\n";
}
}

void randomTreeTest()
{
AvlTree<uint32_t> tree;

for (int i = 0; i < (4096-1); ++i)
{
uint32_t value = std::rand() % 100;
// std::cout << value << " ";
tree.insert(value);
}
std::cout << "\n";

// frequency count the nodes at each layer..
// thought it was ironic to use AvlTree as the map also.. just to prove that its flexible
AvlMap<int,int> freq;
for (AvlTree<uint32_t>::Iterator it = tree.begin();
it != tree.end();
++it)
{
++(freq[it.height()]);
}
std::cout << "\n";

// dump it
for (AvlMap<int,int>::Iterator it = freq.begin();
it != freq.end();
++it)
{
std::cout << it->first << ":" << it->second  << " ";
}
std::cout << "\n";
}

int main()
{
test();

randomTreeTest();
}



And the output looks like:
50(1)
25(1) 50(2)
10(1) 25(2) 50(1)
10(1) 25(3) 40(1) 50(2)
10(1) 25(3) 30(1) 40(2) 50(1)
10(1) 25(2) 30(1) 40(3) 45(1) 50(2)
10(1) 25(2) 30(1) 40(3) 45(1) 50(2) 75(1)
find:75
10(1) 25(2) 30(1) 40(3) 45(1) 75(2)
10(1) 30(2) 40(3) 45(1) 75(2)
30(1) 40(3) 45(1) 75(2)
30(1) 45(2) 75(1)
45(2) 75(1)
75(1)

15(1)
15(2) 77(1)
15(1) 35(2) 77(1)
15(1) 35(2) 77(1)
15(1) 77(2)
77(1)

93(1)
86(1) 93(2)
86(1) 87(2) 93(1)
86(1) 87(2) 93(1)
87(2) 93(1)
93(1)

1:1992 2:995 3:518 4:266 5:129 6:90 7:45 8:26 9:16 10:8 11:5 12:2 13:2 14:1