This is an interview question from google. I am not able to solve it by myself. Can somebody shed some light?
Write a program to print the sequence of keystrokes such that it generates the maximum number of character 'A's. You are allowed to use only 4 keys: A, Ctrl+A, Ctrl+C and Ctrl+V. Only N keystrokes are allowed. All Ctrl+ characters are considered as one keystroke, so Ctrl+A is one keystroke.
For example, the sequence A, Ctrl+A, Ctrl+C, Ctrl+V generates two A's in 4 keystrokes.
Ctrl+A is Select All
Ctrl+C is Copy
Ctrl+V is Paste
I did some mathematics. For any N, using x numbers of A's , one Ctrl+A, one Ctrl+C and y Ctrl+V, we can generate max ((N-1)/2)2 number of A's. For some N > M, it is better to use as many Ctrl+A's, Ctrl+C and Ctrl+V sequences as it doubles the number of A's.
The sequence Ctrl+A, Ctrl+V, Ctrl+C will not overwrite the existing selection. It will append the copied selection to selected one.
There's a dynamic programming solution. We start off knowing 0 keys can make us 0 A's. Then we iterate through for i up to n, doing two things: pressing A once and pressing select all + copy followed by paste j times (actually j-i-1 below; note the trick here: the contents are still in the clipboard, so we can paste it multiple times without copying each time). We only have to consider up to 4 consecutive pastes, since select, copy, paste x 5 is equivalent to select, copy, paste, select, copy, paste and the latter is better since it leaves us with more in the clipboard. Once we've reached n, we have the desired result.
The complexity might appear to be O(N), but since the numbers grow at an exponential rate it is actually O(N2) due to the complexity of multiplying the large numbers. Below is a Python implementation. It takes about 0.5 seconds to calculate for N=50,000.
def max_chars(n):
dp = [0] * (n+1)
for i in xrange(n):
dp[i+1] = max(dp[i+1], dp[i]+1) # press a
for j in xrange(i+3, min(i+7, n+1)):
dp[j] = max(dp[j], dp[i]*(j-i-1)) # press select all, copy, paste x (j-i-1)
return dp[n]
In the code, j represents the total number of keys pressed after our new sequence of keypresses. We already have i keypresses at this stage, and 2 new keypresses go to select-all and copy. Therefore we're hitting paste j-i-2 times. Since pasting adds to the existing sequence of dp[i] A's, we need to add 1 making it j-i-1. This explains the j-i-1 in the 2nd-last line.
Here are some results (n => number of A's):
7 => 9
8 => 12
9 => 16
10 => 20
100 => 1391569403904
1,000 => 3268160001953743683783272702066311903448533894049486008426303248121757146615064636953144900245
174442911064952028008546304
50,000 => a very large number!
I agree with #SB that you should always state your assumptions: Mine is that you don't need to paste twice to double the number of characters. This gets the answer for 7, so unless my solution is wrong the assumption must be right.
In case someone wonders why I'm not checking sequences of the form Ctrl+A, Ctrl+C, A, Ctrl+V: The end result will always be the same as A, Ctrl+A, Ctrl+C, Ctrl+V which I do consider.
By using marcog's solution I found a pattern that starts at n=16. To illustrate this here are the keystrokes for n=24 up to n=29, I replaced ^A with S (select), ^C with C (copy), and ^V with P (paste) for readability:
24: A,A,A,A,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P
4 * 4 * 4 * 4 * 4 = 1024
25: A,A,A,A,S,C,P,P,P,S,C,P,P,S,C,P,P,S,C,P,P,S,C,P,P
4 * 4 * 3 * 3 * 3 * 3 = 1296
26: A,A,A,A,S,C,P,P,P,S,C,P,P,P,S,C,P,P,S,C,P,P,S,C,P,P
4 * 4 * 4 * 3 * 3 * 3 = 1728
27: A,A,A,A,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P,S,C,P,P,S,C,P,P
4 * 4 * 4 * 4 * 3 * 3 = 2304
28: A,A,A,A,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P,S,C,P,P
4 * 4 * 4 * 4 * 4 * 3 = 3072
29: A,A,A,A,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P,S,C,P,P,P
4 * 4 * 4 * 4 * 4 * 4 = 4096
After an initial 4 As, the ideal pattern is to select, copy, paste, paste, paste and repeat. This will multiply the number of As by 4 every 5 keystrokes. If this 5 keystroke pattern cannot consume the remaining keystrokes on its own some number of 4 keystroke patterns (SCPP) consume the final keystrokes, replacing SCPPP (or removing one of the pastes) as necessary. The 4 keystroke patterns multiply the total by 3 every 4 keystrokes.
Using this pattern here is some Python code that gets the same results as marcog's solution, but is O(1) edit: This is actually O(log n) due to exponentiation, thanks to IVlad for pointing that out.
def max_chars(n):
if n <= 15:
return (0, 1, 2, 3, 4, 5, 6, 9, 12, 16, 20, 27, 36, 48, 64, 81)[n]
e3 = (4 - n) % 5
e4 = n // 5 - e3
return 4 * (4 ** e4) * (3 ** e3)
Calculating e3:
There are always between 0 and 4 SCPP patterns at the end of the keystroke list, for n % 5 == 4 there are 4, n % 5 == 1 there are 3, n % 5 == 2 there are 2, n % 5 == 3 there are 1, and n % 5 == 4 there are 0. This can be simplified to (4 - n) % 5.
Calculating e4:
The total number of patterns increases by 1 whenever n % 5 == 0, as it turns out this number increases to exactly n / 5. Using floor division we can get the total number of patterns, the total number for e4 is the total number of patterns minus e3. For those unfamiliar with Python, // is the future-proof notation for floor division.
Here's how I would approach it:
assume CtrlA = select all
assume CtrlC = copy selection
assume CtrlV = paste copied selection
given some text, it takes 4 keystrokes to duplicate it:
CtrlA to select it all
CtrlC to copy it
CtrlV to paste (this will paste over the selection - STATE YOUR ASSUMPTIONS)
CtrlV to paste again which doubles it.
From there, you can consider doing 4 or 5 A's, then looping through the above. Note that doing ctrl + a, c, v, v will grow your text exponentially as you loop through. If remaining strokes < 4, just keep doing a CtrlV
The key to interviews # places like Google is to state your assumptions, and communicate your thinking. they want to know how you solve problems.
It's solveable in O(1): Like with the Fibonacci numbers, there is a formula to calculate the number of printed As (and the sequence of keystrokes):
1) We can simplify the problem description:
Having only [A],[C-a]+[C-c],[C-v] and an empty copy-paste-buffer
equals
having only [C-a]+[C-c],[C-v] and "A" in the copy-paste-buffer.
2) We can describe the sequence of keystrokes as a string of N chars out of {'*','V','v'}, where 'v' means [C-v] and '*' means [C-a] and 'V' means [C-c]. Example: "vvvv*Vvvvv*Vvvv"
The length of that string still equals N.
The product of the lengths of the Vv-words in that string equals the number of produced As.
3) Given a fixed length N for that string and a fixed number K of words, the outcome will be maximal iff all words have nearly equal lengths. Their pair-wise difference is not more than ±1.
Now, what is the optimal number K, if N is given?
4) Suppose, we want to increase the number of words by appending one single word of length L, then we have to reduce L+1 times any previous word by one 'v'. Example: "…*Vvvv*Vvvv*Vvvv*Vvvv" -> "…*Vvv*Vvv*Vvv*Vvv*Vvv"
Now, what is the optimal word length L?
(5*5*5*5*5) < (4*4*4*4*4)*4 , (4*4*4*4) > (3*3*3*3)*3
=> Optimal is L=4.
5) Suppose, we have a sufficient large N to generate a string with many words of length 4, but a few keystrokes are left; how should we use them?
If there are 5 or more left: Append another word with length 4.
If there are 0 left: Done.
If there are 4 left: We could either
a) append one word with length 3: 4*4*4*4*3=768.
b) or increase 4 words to lenght 5: 5*5*5*5=625. => Appending one word is better.
If there are 3 left: We could either
a) or append one word with length 3 by adjusting the previus word from length 4 to 3: 4*4*4*2=128 < 4*4*3*3=144.
b) increase 3 words to lenght 5: 5*5*5=125. => Appending one word is better.
If there are 2 left: We could either
a) or append one word with length 3 by adjusting the previus two words from length 4 to 3: 4*4*1=16 < 3*3*3=27.
b) increase 2 words to lenght 5: 5*5=25. => Appending one word is better.
If there is 1 left: We could either
a) or append one word with length 3 by adjusting the previus three words from length 4 to 3: 4*4*4*0=0 < 3*3*3*3=81.
b) increase one word to lenght 5: 4*4*5=80. => Appending one word is better.
6) Now, what if we don't have a "sufficient large N" to use the rules in 5)? We have to stick with plan b), if possible!
The strings for small N are:
1:"v", 2:"vv", 3:"vvv", 4:"vvvv"
5:"vvvvv" → 5 (plan b)
6:"vvvvvv" → 6 (plan b)
7:"vvv*Vvv" → 9 (plan a)
8:"vvvv*Vvv" → 12 (plan a)
9:"vvvv*Vvvv" → 16
10:"vvvv*Vvvvv" → 20 (plan b)
11:"vvv*Vvv*Vvv" → 29 (plan a)
12:"vvvv*Vvv*Vvv" → 36 (plan a)
13:"vvvv*Vvvv*Vvv" → 48 (plan a)
14:"vvvv*Vvvv*Vvvv" → 64
15:"vvv*Vvv*Vvv*Vvv" → 81 (plan a)
…
7) Now, what is the optimal number K of words in a string of length N?
If N < 7 then K=1 else if 6 < N < 11 then K=2 ; otherwise: K=ceil((N+1)/5)
Written in C/C++/Java: int K = (N<7)?(1) : (N<11)?(2) : ((N+5)/5);
And if N > 10, then the number of words with length 3 will be: K*5-1-N. With this, we can calculate the number of printed As:
If N > 10, the number of As will be: 4^{N+1-4K}·3^{5K-N-1}
Using CtrlA + CtrlC + CtrlV is an advantage only after 4 'A's.
So I would do something like this (in pseudo-BASIC-code, since you haven't specified any proper language):
// We should not use the clipboard for the first four A's:
FOR I IN 1 TO MIN(N, 4)
PRINT 'CLICK A'
NEXT
LET N1 = N - 4
// Generates the maximum number of pastes allowed:
FOR I IN 1 TO (N1 DIV 3) DO
PRINT 'CTRL-A'
PRINT 'CTRL-C'
PRINT 'CTRL-V'
LET N1 = N1 - 3
NEXT
// If we still have same keystrokes left, let's use them with simple CTRL-Vs
FOR I IN N1 TO N
PRINT 'CTRL-V'
NEXT
Edit
Back to using a single CtrlV in the main loop.
Added some comments to explain what I'm trying to do here.
Fixed an issue with the "first four A's" block.
It takes 3 keystrokes to double your number of As. It only makes sense to start doubling when you have 3 or more As already printed. You want your last allowed keystroke to be a CtrlV to make sure you are doubling the biggest number you can, so in order to align it we will fill in any extra keystrokes after the first three As at the beginning with more As.
for (i = 3 + n%3; i>0 && n>0; n--, i--) {
print("a");
}
for (; n>0; n = n-3) {
print("ctrl-a");
print("ctrl-c");
print("ctrl-v");
}
Edit:
This is terrible, I completely got ahead of myself and didn't consider multiple pastes for each copy.
Edit 2:
I believe pasting 3 times is optimal, when you have enough keystrokes to do it. In 5 keystrokes you multiply your number of As by 4. This is better than multiplying by 3 using 4 keystrokes and better than multiplying by 5 using 6 keystrokes. I compared this by giving each method the same number of keystrokes, enough so they each would finish a cycle at the same time (60), letting the 3-multiplier do 15 cycles, the 4-multiplier do 12 cycles, and the 5-multiplier do 10 cycles. 3^15 = 14,348,907, 4^12=16,777,216, and 5^10=9,765,625. If there are only 4 keystrokes left, doing a 3-multiplier is better than pasting 4 more times, essentially making the previous 4 multiplier become an 8-multiplier. If there are only 3 keystrokes left, a 2-multiplier is best.
Assume you have x characters in the clipboard and x characters in the text area; let's call it "state x".
Let's press "Paste" a few times (i denote it by m-1 for convenience), then "Select-all" and "Copy"; after this sequence, we get to "state m*x".
Here, we wasted a total of m+1 keystrokes.
So the asymptotic growth is (at least) something like f^n, where f = m^(1/(m+1)).
I believe it's the maximum possible asymptotic growth, though i cannot prove it (yet).
Trying various values of m shows that the maximum for f is obtained for m=4.
Let's use the following algorithm:
Press A a few times
Press Select-all
Press Copy
Repeat a few times:
Press Paste
Press Paste
Press Paste
Press Select-all
Press Copy
While any keystrokes left:
Press Paste
(not sure it's the optimal one).
The number of times to press A at the beginning is 3: if you press it 4 times, you miss the opportunity to double the number of A's in 3 more keystrokes.
The number of times to press Paste at the end is no more than 5: if you have 6 or more keystrokes left, you can use Paste, Paste, Paste, Select-all, Copy, Paste instead.
So, we get the following algorithm:
If (less than 6 keystrokes - special case)
While (any keystrokes left)
A
Else
First 5 keystrokes: A, A, A, Select-all, Copy
While (more than 5 keystrokes left)
Paste, Paste, Paste, Select-all, Copy
While (any keystrokes left)
Paste
(not sure it's the optimal one). The number of characters after executing this is something like
3 * pow(4, floor((n - 6) / 5)) * (2 + (n - 1) % 5).
Sample values: 1,2,3,4,5,6,9,12,15,18,24,36,48,60,72,96,144,192,240,288,...
What follows uses the OP's second edit that pasting does not replace existing text.
Notice a few things:
^A and ^C can be considered a single action that takes two keystrokes, since it never makes sense to do them individually. In fact, we can replace all instances of ^A^C with ^K^V, where ^K is a one-key "cut" operation (let's abbreviate it X). We shall see that dealing with ^K is much nicer than the two-cost ^A^C.
Let's assume that an 'A' starts in the clipboard. Then ^V (let's abbreviate it Y) is strictly superior to A and we can drop the latter from all consideration. (In the actual problem, if the clipboard starts empty, in what follows we'll just replace Y with A instead of ^V up until the first X.)
Every reasonable keystroke sequence can thus be interpreted as a group of Ys separated by Xs, for example YYYXYXYYXY. Denote by V(s) the number of 'A's produced by the sequence s. Then V(nXm) = V(n)*V(m), because X essentially replaces every Y in m with V(n) 'A's.
The copy-paste problem is thus isomorphic to the following problem: "using m+1 numbers which sum to N-m, maximimze their product." For example, when N=6, the answer is m=1 and the numbers (2,3). 6 = 2*3 = V(YYXYYY) = V(AA^A^C^V^V) (or V(YYYXYY) = V(AAA^A^C^V). )
We can make a few observations:
For a fixed value of m, the numbers to choose are ceil( (N-m)/(m+1) ) and floor( (N-m)/(m+1) ) (in whatever combination makes the sum work out; more specifically you will need (N-m) % (m+1) ceils and the rest floors). This is because, for a < b, (a+1)*(b-1) >= a*b.
Unfortunately I don't see an easy way to find the value of m. If this were my interview I would propose two solutions at this point:
Option 1. Loop over all possible m. An O(n log n) solution.
C++ code:
long long ipow(int a, int b)
{
long long val=1;
long long mul=a;
while(b>0)
{
if(b%2)
val *= mul;
mul *= mul;
b/=2;
}
return val;
}
long long trym(int N, int m)
{
int floor = (N-m)/(m+1);
int ceil = 1+floor;
int numceils = (N-m)%(m+1);
return ipow(floor, m+1-numceils) * ipow(ceil, numceils);
}
long long maxAs(int N)
{
long long maxval=0;
for(int m=0; m<N; m++)
{
maxval = std::max(maxval, trym(N,m));
}
return maxval;
}
Option 2. Allow m to attain non-integer values and find its optimal value by taking the derivative of [(N-m)/(m+1)]^m with respect to m and solving for its root. There is no analytic solution, but the root can be found using e.g. Newton's method. Then use the floor and ceiling of that root for the value of m, and choose whichever is best.
public int dp(int n)
{
int arr[] = new int[n];
for (int i = 0; i < n; i++)
arr[i] = i + 1;
for (int i = 2; i < n - 3; i++)
{
int numchars = arr[i] * 2;
int j = i + 3;
arr[j] = Math.max(arr[j], numchars);
while (j < n - 1)
{
numchars = numchars + arr[i];
arr[++j] = Math.max(arr[j], numchars);
}
}
return arr[n - 1];
}
Here is my approach and solution with code below.
Approach:
There are three distinct operations that can be performed.
Keystroke A - Outputs one character 'A'
Keystroke (Ctrl-A) + (Ctrl-C) - Outputs nothing essentially. These two keystrokes can be combined into one operation because each of these keystrokes individually make no sense. Also, this keystroke sets up the output for the next paste operation.
Keystroke (Ctrl-V) - Output for this keystroke really depends on the previous (second) operation and hence we would need to account for that in our code.
Now given the three distinct operations and their respective outputs, we have to run through all the permutations of these operations.
Assumption:
Now, some version of this problem states that the sequence of keystrokes, Ctrl+A -> Ctrl+C -> Ctrl+V, overwrite the highlighted selection. To factor in this assumption, only one line of code needs to be added to the solution below where the printed variable in case 2 is set to 0
case 2:
//Ctrl-A and then Ctrl-C
if((count+2) < maxKeys)
{
pOutput = printed;
//comment the below statement to NOT factor
//in the assumption described above
printed = 0;
}
For this solution
The code below will print a couple of sequences and the last sequence is the correct answer for any given N. e.g. for N=11 this will be the correct sequence
With the assumption
A, A, A, A, A, C, S, V, V, V, V, :20:
Without the assumption
A, A, A, C, S, V, V, C, S, V, V, :27:
I have decided to retain the assumption for this solution.
Keystroke Legend:
'A' - A
'C' - Ctrl+A
'S' - Ctrl+C
'V' - Ctrl+V
Code:
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
void maxAprinted(int count, int maxKeys, int op, int printed, int pOutput, int *maxPrinted, char *seqArray)
{
if(count > maxKeys)
return;
if(count == maxKeys)
{
if((*maxPrinted) < printed)
{
//new sequence found which is an improvement over last sequence
(*maxPrinted) = printed;
printf("\n");
int i;
for(i=0; i<maxKeys; i++)
printf(" %c,",seqArray[i]);
}
return;
}
switch(op)
{
case 1:
//A keystroke
printed++;
seqArray[count] = 'A';
count++;
break;
case 2:
//Ctrl-A and then Ctrl-C
if((count+2) < maxKeys)
{
pOutput = printed;
//comment the below statement to NOT factor
//in the assumption described above
printed = 0;
}
seqArray[count] = 'C';
count++;
seqArray[count] = 'S';
count++;
break;
case 3:
//Ctrl-V
printed = printed + pOutput;
seqArray[count] = 'V';
count++;
break;
}
maxAprinted(count, maxKeys, 1, printed, pOutput, maxPrinted, seqArray);
maxAprinted(count, maxKeys, 2, printed, pOutput, maxPrinted, seqArray);
maxAprinted(count, maxKeys, 3, printed, pOutput, maxPrinted, seqArray);
}
int main()
{
const int keyStrokes = 11;
//this array stores the sequence of keystrokes
char *sequence;
sequence = (char*)malloc(sizeof(char)*(keyStrokes + 1));
//stores the max count for As printed for a sqeuence
//updated in the recursive call.
int printedAs = 0;
maxAprinted(0, keyStrokes, 1, 0, 0, &printedAs, sequence);
printf(" :%d:", printedAs);
return 0;
}
Using the tricks mentioned in answers above, Mathematically, Solution can be explained in one equation as,
4 + 4^[(N-4)/5] + ((N-4)%5)*4^[(N-4)/5].
where [] is greatest integer factor
There is a trade-off between printing m-A's manually, then using Ctrl+A, Ctrl+C, and N-m-2 Ctrl+V. The best solution is in the middle. If max key strokes = 10, the best solution is typing 5 A's or 4 A's.
try using this Look at this http://www.geeksforgeeks.org/how-to-print-maximum-number-of-a-using-given-four-keys/ and maybe optimize a bit looking for the results around the mid point.
Here is my solution with dynamic programming, without a nested loop, and which also prints the actual characters that you'd need to type:
N = 52
count = [0] * N
res = [[]] * N
clipboard = [0] * N
def maybe_update(i, new_count, new_res, new_clipboard):
if new_count > count[i] or (
new_count == count[i] and new_clipboard > clipboard[i]):
count[i] = new_count
res[i] = new_res
clipboard[i] = new_clipboard
for i in range(1, N):
# First option: type 'A'.
# Using list concatenation for 'res' to avoid O(n^2) string concatenation.
maybe_update(i, count[i - 1] + 1, res[i - 1] + ['A'], clipboard[i - 1])
# Second option: type 'CTRL+V'.
maybe_update(i, count[i - 1] + clipboard[i - 1], res[i - 1] + ['v'],
clipboard[i - 1])
# Third option: type 'CTRL+A, CTRL+C, CTRL+V'.
# Assumption: CTRL+V always appends.
if i >= 3:
maybe_update(i, 2 * count[i - 3], res[i - 3] + ['acv'], count[i - 3])
for i in range(N):
print '%2d %7d %6d %-52s' % (i, count[i], clipboard[i], ''.join(res[i]))
This is the output ('a' means 'CTRL+A', etc.)
0 0 0
1 1 0 A
2 2 0 AA
3 3 0 AAA
4 4 0 AAAA
5 5 0 AAAAA
6 6 3 AAAacv
7 9 3 AAAacvv
8 12 3 AAAacvvv
9 15 3 AAAacvvvv
10 18 9 AAAacvvacv
11 27 9 AAAacvvacvv
12 36 9 AAAacvvacvvv
13 45 9 AAAacvvacvvvv
14 54 27 AAAacvvacvvacv
15 81 27 AAAacvvacvvacvv
16 108 27 AAAacvvacvvacvvv
17 135 27 AAAacvvacvvacvvvv
18 162 81 AAAacvvacvvacvvacv
19 243 81 AAAacvvacvvacvvacvv
20 324 81 AAAacvvacvvacvvacvvv
21 405 81 AAAacvvacvvacvvacvvvv
22 486 243 AAAacvvacvvacvvacvvacv
23 729 243 AAAacvvacvvacvvacvvacvv
24 972 243 AAAacvvacvvacvvacvvacvvv
25 1215 243 AAAacvvacvvacvvacvvacvvvv
26 1458 729 AAAacvvacvvacvvacvvacvvacv
27 2187 729 AAAacvvacvvacvvacvvacvvacvv
28 2916 729 AAAacvvacvvacvvacvvacvvacvvv
29 3645 729 AAAacvvacvvacvvacvvacvvacvvvv
30 4374 2187 AAAacvvacvvacvvacvvacvvacvvacv
31 6561 2187 AAAacvvacvvacvvacvvacvvacvvacvv
32 8748 2187 AAAacvvacvvacvvacvvacvvacvvacvvv
33 10935 2187 AAAacvvacvvacvvacvvacvvacvvacvvvv
34 13122 6561 AAAacvvacvvacvvacvvacvvacvvacvvacv
35 19683 6561 AAAacvvacvvacvvacvvacvvacvvacvvacvv
36 26244 6561 AAAacvvacvvacvvacvvacvvacvvacvvacvvv
37 32805 6561 AAAacvvacvvacvvacvvacvvacvvacvvacvvvv
38 39366 19683 AAAacvvacvvacvvacvvacvvacvvacvvacvvacv
39 59049 19683 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvv
40 78732 19683 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvv
41 98415 19683 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvvv
42 118098 59049 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacv
43 177147 59049 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvv
44 236196 59049 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvv
45 295245 59049 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvvv
46 354294 177147 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacv
47 531441 177147 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacvv
48 708588 177147 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvv
49 885735 177147 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvvv
50 1062882 531441 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacv
51 1594323 531441 AAAacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacvvacvv
If N key Strokes are allowed, then the result is N-3.
A's -> N-3
CTRL+A -> Selecting those N Characters :+1
CTRL+C -> Copying those N Characters :+1
Ctrl+V -> Pasting the N Characters. :+1
i.e., (Since we have selected the whole characters using CTRL+A) Replacing these existing N-3 characters with the copied N-3 Characters(which is overriding the same characters) and the result is N-3.
After reading the base64 wiki ...
I'm trying to figure out how's the formula working :
Given a string with length of n , the base64 length will be
Which is : 4*Math.Ceiling(((double)s.Length/3)))
I already know that base64 length must be %4==0 to allow the decoder know what was the original text length.
The max number of padding for a sequence can be = or ==.
wiki :The number of output bytes per input byte is approximately 4 / 3 (33%
overhead)
Question:
How does the information above settle with the output length ?
Each character is used to represent 6 bits (log2(64) = 6).
Therefore 4 chars are used to represent 4 * 6 = 24 bits = 3 bytes.
So you need 4*(n/3) chars to represent n bytes, and this needs to be rounded up to a multiple of 4.
The number of unused padding chars resulting from the rounding up to a multiple of 4 will obviously be 0, 1, 2 or 3.
4 * n / 3 gives unpadded length.
And round up to the nearest multiple of 4 for padding, and as 4 is a power of 2 can use bitwise logical operations.
((4 * n / 3) + 3) & ~3
For reference, the Base64 encoder's length formula is as follows:
As you said, a Base64 encoder given n bytes of data will produce a string of 4n/3 Base64 characters. Put another way, every 3 bytes of data will result in 4 Base64 characters. EDIT: A comment correctly points out that my previous graphic did not account for padding; the correct formula for padding is 4(Ceiling(n/3)).
The Wikipedia article shows exactly how the ASCII string Man encoded into the Base64 string TWFu in its example. The input string is 3 bytes, or 24 bits, in size, so the formula correctly predicts the output will be 4 bytes (or 32 bits) long: TWFu. The process encodes every 6 bits of data into one of the 64 Base64 characters, so the 24-bit input divided by 6 results in 4 Base64 characters.
You ask in a comment what the size of encoding 123456 would be. Keeping in mind that every every character of that string is 1 byte, or 8 bits, in size (assuming ASCII/UTF8 encoding), we are encoding 6 bytes, or 48 bits, of data. According to the equation, we expect the output length to be (6 bytes / 3 bytes) * 4 characters = 8 characters.
Putting 123456 into a Base64 encoder creates MTIzNDU2, which is 8 characters long, just as we expected.
Integers
Generally we don't want to use doubles because we don't want to use the floating point ops, rounding errors etc. They are just not necessary.
For this it is a good idea to remember how to perform the ceiling division: ceil(x / y) in doubles can be written as (x + y - 1) / y (while avoiding negative numbers, but beware of overflow).
Readable
If you go for readability you can of course also program it like this (example in Java, for C you could use macro's, of course):
public static int ceilDiv(int x, int y) {
return (x + y - 1) / y;
}
public static int paddedBase64(int n) {
int blocks = ceilDiv(n, 3);
return blocks * 4;
}
public static int unpaddedBase64(int n) {
int bits = 8 * n;
return ceilDiv(bits, 6);
}
// test only
public static void main(String[] args) {
for (int n = 0; n < 21; n++) {
System.out.println("Base 64 padded: " + paddedBase64(n));
System.out.println("Base 64 unpadded: " + unpaddedBase64(n));
}
}
Inlined
Padded
We know that we need 4 characters blocks at the time for each 3 bytes (or less). So then the formula becomes (for x = n and y = 3):
blocks = (bytes + 3 - 1) / 3
chars = blocks * 4
or combined:
chars = ((bytes + 3 - 1) / 3) * 4
your compiler will optimize out the 3 - 1, so just leave it like this to maintain readability.
Unpadded
Less common is the unpadded variant, for this we remember that each we need a character for each 6 bits, rounded up:
bits = bytes * 8
chars = (bits + 6 - 1) / 6
or combined:
chars = (bytes * 8 + 6 - 1) / 6
we can however still divide by two (if we want to):
chars = (bytes * 4 + 3 - 1) / 3
Unreadable
In case you don't trust your compiler to do the final optimizations for you (or if you want to confuse your colleagues):
Padded
((n + 2) / 3) << 2
Unpadded
((n << 2) | 2) / 3
So there we are, two logical ways of calculation, and we don't need any branches, bit-ops or modulo ops - unless we really want to.
Notes:
Obviously you may need to add 1 to the calculations to include a null termination byte.
For Mime you may need to take care of possible line termination characters and such (look for other answers for that).
(In an attempt to give a succinct yet complete derivation.)
Every input byte has 8 bits, so for n input bytes we get:
n × 8 input bits
Every 6 bits is an output byte, so:
ceil(n × 8 / 6) = ceil(n × 4 / 3) output bytes
This is without padding.
With padding, we round that up to multiple-of-four output bytes:
ceil(ceil(n × 4 / 3) / 4) × 4 = ceil(n × 4 / 3 / 4) × 4 = ceil(n / 3) × 4 output bytes
See Nested Divisions (Wikipedia) for the first equivalence.
Using integer arithmetics, ceil(n / m) can be calculated as (n + m – 1) div m,
hence we get:
(n * 4 + 2) div 3 without padding
(n + 2) div 3 * 4 with padding
For illustration:
n with padding (n + 2) div 3 * 4 without padding (n * 4 + 2) div 3
------------------------------------------------------------------------------
0 0 0
1 AA== 4 AA 2
2 AAA= 4 AAA 3
3 AAAA 4 AAAA 4
4 AAAAAA== 8 AAAAAA 6
5 AAAAAAA= 8 AAAAAAA 7
6 AAAAAAAA 8 AAAAAAAA 8
7 AAAAAAAAAA== 12 AAAAAAAAAA 10
8 AAAAAAAAAAA= 12 AAAAAAAAAAA 11
9 AAAAAAAAAAAA 12 AAAAAAAAAAAA 12
10 AAAAAAAAAAAAAA== 16 AAAAAAAAAAAAAA 14
11 AAAAAAAAAAAAAAA= 16 AAAAAAAAAAAAAAA 15
12 AAAAAAAAAAAAAAAA 16 AAAAAAAAAAAAAAAA 16
Finally, in the case of MIME Base64 encoding, two additional bytes (CR LF) are needed per every 76 output bytes, rounded up or down depending on whether a terminating newline is required.
Here is a function to calculate the original size of an encoded Base 64 file as a String in KB:
private Double calcBase64SizeInKBytes(String base64String) {
Double result = -1.0;
if(StringUtils.isNotEmpty(base64String)) {
Integer padding = 0;
if(base64String.endsWith("==")) {
padding = 2;
}
else {
if (base64String.endsWith("=")) padding = 1;
}
result = (Math.ceil(base64String.length() / 4) * 3 ) - padding;
}
return result / 1000;
}
I think the given answers miss the point of the original question, which is how much space needs to be allocated to fit the base64 encoding for a given binary string of length n bytes.
The answer is (floor(n / 3) + 1) * 4 + 1
This includes padding and a terminating null character. You may not need the floor call if you are doing integer arithmetic.
Including padding, a base64 string requires four bytes for every three-byte chunk of the original string, including any partial chunks. One or two bytes extra at the end of the string will still get converted to four bytes in the base64 string when padding is added. Unless you have a very specific use, it is best to add the padding, usually an equals character. I added an extra byte for a null character in C, because ASCII strings without this are a little dangerous and you'd need to carry the string length separately.
For all people who speak C, take a look at these two macros:
// calculate the size of 'output' buffer required for a 'input' buffer of length x during Base64 encoding operation
#define B64ENCODE_OUT_SAFESIZE(x) ((((x) + 3 - 1)/3) * 4 + 1)
// calculate the size of 'output' buffer required for a 'input' buffer of length x during Base64 decoding operation
#define B64DECODE_OUT_SAFESIZE(x) (((x)*3)/4)
Taken from here.
While everyone else is debating algebraic formulas, I'd rather just use BASE64 itself to tell me:
$ echo "Including padding, a base64 string requires four bytes for every three-byte chunk of the original string, including any partial chunks. One or two bytes extra at the end of the string will still get converted to four bytes in the base64 string when padding is added. Unless you have a very specific use, it is best to add the padding, usually an equals character. I added an extra byte for a null character in C, because ASCII strings without this are a little dangerous and you'd need to carry the string length separately."| wc -c
525
$ echo "Including padding, a base64 string requires four bytes for every three-byte chunk of the original string, including any partial chunks. One or two bytes extra at the end of the string will still get converted to four bytes in the base64 string when padding is added. Unless you have a very specific use, it is best to add the padding, usually an equals character. I added an extra byte for a null character in C, because ASCII strings without this are a little dangerous and you'd need to carry the string length separately." | base64 | wc -c
710
So it seems the formula of 3 bytes being represented by 4 base64 characters seems correct.
I don't see the simplified formula in other responses. The logic is covered but I wanted a most basic form for my embedded use:
Unpadded = ((4 * n) + 2) / 3
Padded = 4 * ((n + 2) / 3)
NOTE: When calculating the unpadded count we round up the integer division i.e. add Divisor-1 which is +2 in this case
Seems to me that the right formula should be:
n64 = 4 * (n / 3) + (n % 3 != 0 ? 4 : 0)
I believe that this one is an exact answer if n%3 not zero, no ?
(n + 3-n%3)
4 * ---------
3
Mathematica version :
SizeB64[n_] := If[Mod[n, 3] == 0, 4 n/3, 4 (n + 3 - Mod[n, 3])/3]
Have fun
GI
Simple implementantion in javascript
function sizeOfBase64String(base64String) {
if (!base64String) return 0;
const padding = (base64String.match(/(=*)$/) || [])[1].length;
return 4 * Math.ceil((base64String.length / 3)) - padding;
}
If there is someone interested in achieve the #Pedro Silva solution in JS, I just ported this same solution for it:
const getBase64Size = (base64) => {
let padding = base64.length
? getBase64Padding(base64)
: 0
return ((Math.ceil(base64.length / 4) * 3 ) - padding) / 1000
}
const getBase64Padding = (base64) => {
return endsWith(base64, '==')
? 2
: 1
}
const endsWith = (str, end) => {
let charsFromEnd = end.length
let extractedEnd = str.slice(-charsFromEnd)
return extractedEnd === end
}
In windows - I wanted to estimate size of mime64 sized buffer, but all precise calculation formula's did not work for me - finally I've ended up with approximate formula like this:
Mine64 string allocation size (approximate)
= (((4 * ((binary buffer size) + 1)) / 3) + 1)
So last +1 - it's used for ascii-zero - last character needs to allocated to store zero ending - but why "binary buffer size" is + 1 - I suspect that there is some mime64 termination character ? Or may be this is some alignment issue.