# Implement bit-wise matrix multiplication

The MMIX architecture is a fairly simple big-endian RISC design. It has a couple of interesting features, one of them is the MXOR instruction, which is what you should implement.

The MXOR instruction interprets its arguments (two 64-bit registers) as two 8×8 matrices of bits and performs a matrix multiplication where exclusive-or is used for addition and logical and is used for multiplication.

An argument in binary representation where the leftmost digit is most significant

a0 a1 a2 … a63

is interpreted as a matrix like this:

a0 a1 … a7
a8 a9 … a15

a56 a57 … a63

Your task is to write a function, subroutine or similar in a language of choice which can be evoked from a single-letter name (i.e. f(x,y)) that implements this function. The two arguments of the function shall be of equal implementation-defined unsigned 64 bit integer type, the function shall return a value of the same type.

he winner of this competition is the shortest submission (measured in bytes or characters if the language you use uses a meaningful character encoding). You are encouraged to post a readable commented version of your code in addition to the golfed solution so other's can understand how it works.

• Some test cases would be nice. – Peter Taylor Aug 20 '14 at 19:27
• Do the input and outputs need to be numbers that correspond to the bitstrings, or can we use lists of bits? – xnor Aug 20 '14 at 19:44
• Am I allowed to use the language's matrix multiplication (if present in the standard lib)? – Howard Aug 20 '14 at 20:19
• Put another way, matrix multiplication where the elements of the matrix are in GF(2). – hobbs Aug 20 '14 at 21:53
• @Howard: why not? – FUZxxl Aug 20 '14 at 22:13

### APL, 28 characters

f←{2⊥,↑≠.×/{8 8⍴(64⍴2)⊤⍵}¨⍵}


Usage:

> f 12345678 87654321
3977539


You can read it almost literally if you go right-to-left. (64⍴2)⊤ transforms each argument into 64 bit representation and 8 8⍴ makes an 8x8 matrix out of the results. ≠.×/ then applies the inner product to both arguments, where multiplication × and xor ≠ is used. The matrix is then transformed back to vector form ,↑, before 2⊥ undoes the basis transformation.

• Save 6 bytes by converting to tradfn, swapping ⊤'s arguments using ⍨, and flattening with ∊. Also, you don't need to count the naming of your function in your code length. Try it online! – Adám Aug 9 '17 at 12:48

## x86 machine code, 4544 40 bytes

  40072c:       b1 08                   mov    $0x8,%cl 40072e: 48 c1 c7 08 rol$0x8,%rdi
400732:       b5 08                   mov    $0x8,%ch 400734: 48 0f 6e c6 movq %rsi,%mm0 400738: 0f d7 d0 pmovmskb %mm0,%edx 40073b: 40 20 fa and %dil,%dl 40073e: 7a 01 jp 400741 <amul+0x15> 400740: f9 stc 400741: 48 11 c0 adc %rax,%rax 400744: 48 d1 c6 rol %rsi 400747: fe cd dec %ch 400749: 75 e9 jne 400734 <amul+0x8> 40074b: 48 c1 ce 08 ror$0x8,%rsi
40074f:       fe c9                   dec    %cl
400751:       75 db                   jne    40072e <amul+0x2>
400753:       c3                      retq


The code uses two tricky moves. The first is the pmovmskb intsruction which takes every 8th bit of an mmx register and packs them together into an integer register. Very useful for extracting columns. The second is the parity flag which is used to accumulate the dot product (the and/jp/stc combo).

Here's the expanded code with comments:

    # rowloop
mov $8, %cl 1: rol$8, %rdi    # get next row of A into low 8 bits

# columnloop
mov $8, %ch 2: # extract first column movq %rsi, %mm0 pmovmskb %mm0, %edx # dot with current row testb %dil, %dl # shift result and add parity bit jp 3f stc 3: adc %rax, %rax # shift over columns of B rol %rsi dec %ch jne 2b # restore B ror$8, %rsi

dec %cl
jne 1b
ret

• Looks cool! Very neat solution. – FUZxxl Aug 21 '14 at 21:29
• The use of the parity bit is really ingenious though. Makes the solution so much easier. – FUZxxl Aug 22 '14 at 4:15
• BTW, you can replace xor %rax,%rax by xor %eax,%eax for one byte. The high byte is zeroed out automatically. sub %eax,%eax might be even shorter. – FUZxxl Aug 22 '14 at 4:16
• You might also want to replace setnp %dl; lea (%rdx,%rax,2),%rax by jnp 1f; stc; 1:adc %rax,%rax if that improves the size (should bring one byte). – FUZxxl Aug 22 '14 at 4:21
• Oh yes, you can also replace dec %cl;jne 1b by loop 1b since %ch will always be zero when encountering this instruction. – FUZxxl Aug 22 '14 at 4:33

## GolfScript (49 48 45 chars)

This takes a completely different approach to my earlier GolfScript entry, so I think it wants a separate answer.

{0{1$M&16.?M/:K*3$K&M*&^@2/@M)/@}255:M*++}:f;


It's pure bit-bashing.

If we consider a as a series of matrix elements we get

a_00 a_01 a_02 a_03 a_04 a_05 a_06 a_07 a_10 a_11 ...


Similarly b. Now, c_ij = a_i0 & b_0j ^ a_i1 & b1j ^ ... ^ a_i7 & b7j. So by grouping over the common index we get

c =   (a_00 a_00 ... a_00 a_00 a_10 a_10 ...) & (b_00 b_01 b_02 ... b_07 b_00 b_01 ...)
^ (a_01 a_01 ... a_01 a_01 a_11 a_11 ...) & (b_10 b_11 b_12 ... b_17 b_10 b_11 ...)
^ ...
=   (((a >> 7) & 0x0101010101010101) * 0xff) & (((b >> 56) & 0xff) * 0x0101010101010101)
^ (((a >> 6) & 0x0101010101010101) * 0xff) & (((b >> 48) & 0xff) * 0x0101010101010101)
^ ...


That's enough background to be able to explain the code:

{                 # Function declaration boilerplate; stack is a b
0               # Accumulator for the result
{               # Loop: stack holds a>>k b>>8k partial-sum
1$M& # M holds 255 (see below); get (b>>8k) & 0xff 16.?M/:K # Derive 0x0101010101010101 and store in K * # Multiply: stack is a>>k b>>8k partial-sum rhs-of-and 3$K&M*        #   Similarly for a: add lhs-of-and to the stack
&^            #   Do the pointwise multiplication and addition
@2/@M)/@      #   Restore loop invariant with a>>=1 and b>>=8
}255:M*         # Store 255 in M and execute the loop that many times
# 8 times suffices to get the correct result; after that b>>8k == 0 and the
# partial sum doesn't change. After 64 iterations, a>>k == 0
++              # So we can ditch their remnants by adding to the partial sum
}:f;              # Function declaration boilerplate


Thanks to Dennis for the observation that 16.? is equal to 2 64? (or 2.6??); since it's one char shorter, I've stolen it. Thanks to Howard for suggesting a 2-char improvement which became 48->45.

• Now that's an interesting approach. Nice bit magic! – Dennis Aug 21 '14 at 15:35
• You can replace ;; with * since b>>64 is always zero. – Howard Aug 21 '14 at 15:38
• If you directly calculate the result on the stack (instead of collecting terms in an array) you may even save one more char: {0{1$255:M&16.?M/:K*3$K&M*&^@2/@M)/@}8*+\;}:f;. – Howard Aug 21 '14 at 15:57
• @Howard, good points both. It's even possible to save one more by shifting a far enough right to become zero. – Peter Taylor Aug 22 '14 at 7:29

# CJam, 35 bytes

{2{GG#+2b1>8/\}*zm*{z{:&}%:^}%2b}:F


Try it online by pasting the following Code:

{2{GG#+2b1>8/\}*zm*{z{:&}%:^}%2b}:F;  " Define function and clear stack.              ";
12345678 87654321 F                   " Call function.                                ";


### How it works

2{      " Do the following twice:                                                     ";
GG#+  " Add 16 ** 16 = 2 ** 64 to the topmost integer on the stack.                 ";
2b    " Convert it into an array of its digits in base two.                         ";
1>    " Discard the first digit.                                                    ";
8/    " Split into arrays of 8 digits.                                              ";
\     " Swap the resulting array with the second topmost stack element.             ";
}*      " This converts two 64-bit integers into 8×8 matrices of its base 2 digits.   ";
zm*     " Transpose the second matrix and compute the cartesian product of both.      ";
{       " For each pair (one row of the first and one column of the second matrix):   ";
z     " Transpose the array.                                                        ";
{:&}% " For each array of the array, compute the logical AND of its elements.       ";
:^    " Compute the logical XOR of the results.                                     ";
}%      " Note that ^ and & are addition and multiplication in GF(2).                 ";
2b      " Convert the array of base 2 digits into and integer.                        ";

• 8 ** 8 = 2 ** 24, surely? – Peter Taylor Aug 21 '14 at 15:14
• Of course... G is 16, so it's a wrong comment for correct code. Thanks! – Dennis Aug 21 '14 at 15:15
• Nice trick, though. If I steal it, I can save 1 char. – Peter Taylor Aug 21 '14 at 15:17

## GolfScript (65 chars)

{[{2.6??+2base(;8/\}2*zip 8*\{{.}7*}%]zip{zip{~&}%{^}*}%2base}:f;


This turns out to be similar to BeetDemGuise's Python solution, although it uses more zips and less eval.

Dissection:

{
# Stack: a b (as integers)
[                  # Start gathering into array for subsequent zip
{                # Repeat twice:
2.6??+2base(;  # Convert to 64 bits
8/             # Break into an 8x8 matrix
\              # Flip top of stack
}2*
# Stack: a b (as matrices)
# Each cell of the matrix product needs a row of a and a column of b
zip 8*           # Get 8 copies of columns of b, ordered
#   [col_0 col_1 ... col_7 col_0 col_1 ...]
\{{.}7*}%        # Get 8 copies of rows of a, ordered
#   [row_0 row_0 row_0 ... row_7 row_7]
]                  # Gather into array
zip                # Zip: each array entry is now an array [col_j row_i]
{                  # Map each array item
zip              # Pair up each bit from col_j with corresponding bit of row_i
{~&}%            # Multiply the bits in each pair
}%

2base              # Convert back from an array of 64 bits to an integer
}:f;                 # Assign to f and don't leave a copy lying about on the stack


## J, 50 chars

Here is a version that does not use J's built-in general matrix product conjunction ..

   f=:#.@,@:(|."1)@(~:/@:*"2 1"1 2|.)&(8 8$]#:~64#2:)  ### Testing (with 4×4 matrices)  g=:#.@,@:(|."1)@(~:/@:*"2 1"1 2|.)&(4 4$]#:~16#2:)
44584 g 28371
46550


And the matrices visualized, for manual verification.

   (4 4$]#:~16#2:)&.> 44584;28371;46550 ┌───────┬───────┬───────┐ │1 0 1 0│0 1 1 0│1 0 1 1│ │1 1 1 0│1 1 1 0│0 1 0 1│ │0 0 1 0│1 1 0 1│1 1 0 1│ │1 0 0 0│0 0 1 1│0 1 1 0│ └───────┴───────┴───────┘  ### Exploded view  &( ) to each of the arguments, 64#2: generate 64 2's (radix for below) ]#:~ decode from radix 8 8$             reshape as 8×8 array
(              |.)                 reverse the right argument
"2 1"1 2                    for each pair of rows,
*                              multiply
@:                               composed with
~:/                                 xor folded
(|."1)@                                   reverse each row
,@:                                          ravel (flatten array)
#.@                                             back to binary again
f=:                                                assign function to name 'f'

• Instead of 64#2:, isn't there a shorter construct with 64$2? – FUZxxl Aug 21 '14 at 11:08 • @FUZxxl no, we can substitute # with $ (they do the same in this case), but we need the 2 to be a verb for the whole paren to be a train. 2: is a constant verb returning 2. – FireFly Aug 21 '14 at 11:40
• A, right. That makes sense. – FUZxxl Aug 21 '14 at 14:01

# Python 2 - 183

def m(a,b):a,b=map(lambda x:'{:064b}'.format(x),[a,b]);r=range(8);return int(''.join(eval('^'.join(o+'&'+t for o,t in zip(a[8*i:8+8*i],v)))for i in r for v in[b[j::8]for j in r]),2)


This is the first thing that came to my mind and so it can mostly likely be golfed down quite a bit more. In any case, its pretty straight-forward string manipulation:

def m(a,b):
# Convert the integer inputs to their 64bit binary versions.
a,b=map(lambda x:'{:064b}'.format(x),[a,b])

n=''
# For all rows in matrix a...
for row in [a[8*i:8+8*i] for i in range(8)]:
# For all columns in matrix b...
for col in [b[j::8] for j in range(8)]:
# Get each pair.
pairs = zip(row, col)

# Multiply each pair.
mul_pairs = ['&'.join(z) for z in pairs]

# XOR them all together.
result = '^'.join(mul_pairs)

# Append the result to n.
n+=eval(result)

return int(n,2)


# J (34)

This should work I think. It would be nice if you gave some test cases...

f=:~:/ .*&.(8 8&$@(_64&{.) :.,@#:)  This works with the test cases supplied in the comments (when changing 8 8 to 4 4 and changing 64 to 16) ### Explanation The whole function can be simplified to the following: F&.G  Visual representation: What this means: G is applied to both arguments. Then F is applied with G x as its left argument, and G y as its right argument. Then, the inverse of G is applied to the result of F. "Inverse?", you might ask. With inverse I mean a function applied _1 times, so for example the inverse of #: (dec -> bin vector) is #:^:_1 (or #.) (bin vector -> dec). So, applying this theory, this would be G: • #: takes the binary form of the argument • _64&{. takes the last 64 items of the list. When there aren't 64 or more items, it will be padded with 0's • 8 8&$ shapes it to an 8*8 matrix
• :., sets the reverse of (4 4&$@(_16&{.) to be just ,: ravel and this would be F: • ~:/ .* is matrix multiplication • Nope, I have a longer version using . that seems to work though. You need to return the corresponding number of the matrix, and #: will have issues with leading zero bits. Also, the space before the . is significant (to disambiguate from /.). Oh, and you want ~: (xor) in place of + (or take the result modulo 2 manually). – FireFly Aug 20 '14 at 21:22 • @FireFly Made some changes, would that work? – ɐɔıʇǝɥʇuʎs Aug 20 '14 at 21:26 • Try these for testcases (4×4 arrays, so change the 8 8 correspondingly): 44584 × 33825 = 44584 (33825 is identity); 44584 × 28371 = 46550. – FireFly Aug 20 '14 at 21:27 • What I have is currently f=:#.@,@(~:/ .*)&(8 8$]#:~64#2:), but it's so similar to yours that it'd be weird to post it as its own answer. Maybe you can work off of that. – FireFly Aug 20 '14 at 21:53
• It would be nice if you could add some text that explains how it works. – FUZxxl Aug 20 '14 at 22:43