C++ – Searching for a fast/efficient histogram algorithm (with pre-specified bins)

algorithmchistogramMATLAB

I don't do much coding outside of Matlab, but I have a need to export my Matlab code to another language, most likely C. My Matlab code includes a histogram function, histc(), that places my input data (which is double-precision, not integer) into a specified array of bins, to form a histogram.

I'm sure I can piece together a couple nested loops to generate a histogram function, but I need this function to be fast and memory-light, as it will be accessed repeatedly and often.

To avoid re-inventing the wheel, anyone know if C language has any existing histogram function(s) available for use, or whether people needing such a thing generally create it themselves?

Anyone know an efficient algorithm for creating a histogram? Pseudo-code is fine.

Thanks in advance.

Best Answer

The "ideal" histogram algorithm will depend upon the range you expect to capture. Generally any histogram algorithm will look like this:

const int NSAMPLES = whatever;
double samples[NSAMPLES] = { 1.0, 3.93, 1e30, ... }; // your data set
const int NBUCKETS = 10; // or whatever
int counts[NBUCKETS] = { 0 };
for (int i = 0; i != NSAMPLES; ++i) {
    counts[TRANSFER(samples[i])]++;
}

where TRANSFER() is some function that maps your inputs to a bin (0th or Nth bin mapping to "out of range" of applicable).

The exact implementation of TRANSFER() depends a lot on the expected distribution of your sample and where you are interested in detail. Some common approaches I have seen:

uniform distribution in range [a,b] (requires linear transform)
logarithmic distribution of unsigned integer values (best when combined with some bit twiddling hacks to quickly determine the nearest power-of-two or similar).

If you don't know the distribution up-front, then you really can't have an efficient mechanism to bin them effectively: you'll either have to guess (biased or uninformative results) or store everything and sort it at the end, binning into equal-sized buckets (poor performance).

Options

Low Memory (32-bit int, 32-bit machine)(from here):

unsigned int
reverse(register unsigned int x)
{
    x = (((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1));
    x = (((x & 0xcccccccc) >> 2) | ((x & 0x33333333) << 2));
    x = (((x & 0xf0f0f0f0) >> 4) | ((x & 0x0f0f0f0f) << 4));
    x = (((x & 0xff00ff00) >> 8) | ((x & 0x00ff00ff) << 8));
    return((x >> 16) | (x << 16));

}

From the famous Bit Twiddling Hacks page:

Fastest (lookup table):

static const unsigned char BitReverseTable256[] = 
{
  0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0, 0x10, 0x90, 0x50, 0xD0, 0x30, 0xB0, 0x70, 0xF0, 
  0x08, 0x88, 0x48, 0xC8, 0x28, 0xA8, 0x68, 0xE8, 0x18, 0x98, 0x58, 0xD8, 0x38, 0xB8, 0x78, 0xF8, 
  0x04, 0x84, 0x44, 0xC4, 0x24, 0xA4, 0x64, 0xE4, 0x14, 0x94, 0x54, 0xD4, 0x34, 0xB4, 0x74, 0xF4, 
  0x0C, 0x8C, 0x4C, 0xCC, 0x2C, 0xAC, 0x6C, 0xEC, 0x1C, 0x9C, 0x5C, 0xDC, 0x3C, 0xBC, 0x7C, 0xFC, 
  0x02, 0x82, 0x42, 0xC2, 0x22, 0xA2, 0x62, 0xE2, 0x12, 0x92, 0x52, 0xD2, 0x32, 0xB2, 0x72, 0xF2, 
  0x0A, 0x8A, 0x4A, 0xCA, 0x2A, 0xAA, 0x6A, 0xEA, 0x1A, 0x9A, 0x5A, 0xDA, 0x3A, 0xBA, 0x7A, 0xFA,
  0x06, 0x86, 0x46, 0xC6, 0x26, 0xA6, 0x66, 0xE6, 0x16, 0x96, 0x56, 0xD6, 0x36, 0xB6, 0x76, 0xF6, 
  0x0E, 0x8E, 0x4E, 0xCE, 0x2E, 0xAE, 0x6E, 0xEE, 0x1E, 0x9E, 0x5E, 0xDE, 0x3E, 0xBE, 0x7E, 0xFE,
  0x01, 0x81, 0x41, 0xC1, 0x21, 0xA1, 0x61, 0xE1, 0x11, 0x91, 0x51, 0xD1, 0x31, 0xB1, 0x71, 0xF1,
  0x09, 0x89, 0x49, 0xC9, 0x29, 0xA9, 0x69, 0xE9, 0x19, 0x99, 0x59, 0xD9, 0x39, 0xB9, 0x79, 0xF9, 
  0x05, 0x85, 0x45, 0xC5, 0x25, 0xA5, 0x65, 0xE5, 0x15, 0x95, 0x55, 0xD5, 0x35, 0xB5, 0x75, 0xF5,
  0x0D, 0x8D, 0x4D, 0xCD, 0x2D, 0xAD, 0x6D, 0xED, 0x1D, 0x9D, 0x5D, 0xDD, 0x3D, 0xBD, 0x7D, 0xFD,
  0x03, 0x83, 0x43, 0xC3, 0x23, 0xA3, 0x63, 0xE3, 0x13, 0x93, 0x53, 0xD3, 0x33, 0xB3, 0x73, 0xF3, 
  0x0B, 0x8B, 0x4B, 0xCB, 0x2B, 0xAB, 0x6B, 0xEB, 0x1B, 0x9B, 0x5B, 0xDB, 0x3B, 0xBB, 0x7B, 0xFB,
  0x07, 0x87, 0x47, 0xC7, 0x27, 0xA7, 0x67, 0xE7, 0x17, 0x97, 0x57, 0xD7, 0x37, 0xB7, 0x77, 0xF7, 
  0x0F, 0x8F, 0x4F, 0xCF, 0x2F, 0xAF, 0x6F, 0xEF, 0x1F, 0x9F, 0x5F, 0xDF, 0x3F, 0xBF, 0x7F, 0xFF
};

unsigned int v; // reverse 32-bit value, 8 bits at time
unsigned int c; // c will get v reversed

// Option 1:
c = (BitReverseTable256[v & 0xff] << 24) | 
    (BitReverseTable256[(v >> 8) & 0xff] << 16) | 
    (BitReverseTable256[(v >> 16) & 0xff] << 8) |
    (BitReverseTable256[(v >> 24) & 0xff]);

// Option 2:
unsigned char * p = (unsigned char *) &v;
unsigned char * q = (unsigned char *) &c;
q[3] = BitReverseTable256[p[0]]; 
q[2] = BitReverseTable256[p[1]]; 
q[1] = BitReverseTable256[p[2]]; 
q[0] = BitReverseTable256[p[3]];

You can extend this idea to 64-bit ints, or trade off memory for speed (assuming your L1 Data Cache is large enough), and reverse 16 bits at a time with a 64K-entry lookup table.

Others

Simple

unsigned int v;     // input bits to be reversed
unsigned int r = v & 1; // r will be reversed bits of v; first get LSB of v
int s = sizeof(v) * CHAR_BIT - 1; // extra shift needed at end

for (v >>= 1; v; v >>= 1)
{   
  r <<= 1;
  r |= v & 1;
  s--;
}
r <<= s; // shift when v's highest bits are zero

Faster (32-bit processor)

unsigned char b = x;
b = ((b * 0x0802LU & 0x22110LU) | (b * 0x8020LU & 0x88440LU)) * 0x10101LU >> 16;

Faster (64-bit processor)

unsigned char b; // reverse this (8-bit) byte
b = (b * 0x0202020202ULL & 0x010884422010ULL) % 1023;

If you want to do this on a 32-bit int, just reverse the bits in each byte, and reverse the order of the bytes. That is:

unsigned int toReverse;
unsigned int reversed;
unsigned char inByte0 = (toReverse & 0xFF);
unsigned char inByte1 = (toReverse & 0xFF00) >> 8;
unsigned char inByte2 = (toReverse & 0xFF0000) >> 16;
unsigned char inByte3 = (toReverse & 0xFF000000) >> 24;
reversed = (reverseBits(inByte0) << 24) | (reverseBits(inByte1) << 16) | (reverseBits(inByte2) << 8) | (reverseBits(inByte3);

Results

I benchmarked the two most promising solutions, the lookup table, and bitwise-AND (the first one). The test machine is a laptop w/ 4GB of DDR2-800 and a Core 2 Duo T7500 @ 2.4GHz, 4MB L2 Cache; YMMV. I used gcc 4.3.2 on 64-bit Linux. OpenMP (and the GCC bindings) were used for high-resolution timers.

reverse.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>

unsigned int
reverse(register unsigned int x)
{
    x = (((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1));
    x = (((x & 0xcccccccc) >> 2) | ((x & 0x33333333) << 2));
    x = (((x & 0xf0f0f0f0) >> 4) | ((x & 0x0f0f0f0f) << 4));
    x = (((x & 0xff00ff00) >> 8) | ((x & 0x00ff00ff) << 8));
    return((x >> 16) | (x << 16));

}

int main()
{
    unsigned int *ints = malloc(100000000*sizeof(unsigned int));
    unsigned int *ints2 = malloc(100000000*sizeof(unsigned int));
    for(unsigned int i = 0; i < 100000000; i++)
      ints[i] = rand();

    unsigned int *inptr = ints;
    unsigned int *outptr = ints2;
    unsigned int *endptr = ints + 100000000;
    // Starting the time measurement
    double start = omp_get_wtime();
    // Computations to be measured
    while(inptr != endptr)
    {
      (*outptr) = reverse(*inptr);
      inptr++;
      outptr++;
    }
    // Measuring the elapsed time
    double end = omp_get_wtime();
    // Time calculation (in seconds)
    printf("Time: %f seconds\n", end-start);

    free(ints);
    free(ints2);

    return 0;
}

reverse_lookup.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>

static const unsigned char BitReverseTable256[] = 
{
  0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0, 0x10, 0x90, 0x50, 0xD0, 0x30, 0xB0, 0x70, 0xF0, 
  0x08, 0x88, 0x48, 0xC8, 0x28, 0xA8, 0x68, 0xE8, 0x18, 0x98, 0x58, 0xD8, 0x38, 0xB8, 0x78, 0xF8, 
  0x04, 0x84, 0x44, 0xC4, 0x24, 0xA4, 0x64, 0xE4, 0x14, 0x94, 0x54, 0xD4, 0x34, 0xB4, 0x74, 0xF4, 
  0x0C, 0x8C, 0x4C, 0xCC, 0x2C, 0xAC, 0x6C, 0xEC, 0x1C, 0x9C, 0x5C, 0xDC, 0x3C, 0xBC, 0x7C, 0xFC, 
  0x02, 0x82, 0x42, 0xC2, 0x22, 0xA2, 0x62, 0xE2, 0x12, 0x92, 0x52, 0xD2, 0x32, 0xB2, 0x72, 0xF2, 
  0x0A, 0x8A, 0x4A, 0xCA, 0x2A, 0xAA, 0x6A, 0xEA, 0x1A, 0x9A, 0x5A, 0xDA, 0x3A, 0xBA, 0x7A, 0xFA,
  0x06, 0x86, 0x46, 0xC6, 0x26, 0xA6, 0x66, 0xE6, 0x16, 0x96, 0x56, 0xD6, 0x36, 0xB6, 0x76, 0xF6, 
  0x0E, 0x8E, 0x4E, 0xCE, 0x2E, 0xAE, 0x6E, 0xEE, 0x1E, 0x9E, 0x5E, 0xDE, 0x3E, 0xBE, 0x7E, 0xFE,
  0x01, 0x81, 0x41, 0xC1, 0x21, 0xA1, 0x61, 0xE1, 0x11, 0x91, 0x51, 0xD1, 0x31, 0xB1, 0x71, 0xF1,
  0x09, 0x89, 0x49, 0xC9, 0x29, 0xA9, 0x69, 0xE9, 0x19, 0x99, 0x59, 0xD9, 0x39, 0xB9, 0x79, 0xF9, 
  0x05, 0x85, 0x45, 0xC5, 0x25, 0xA5, 0x65, 0xE5, 0x15, 0x95, 0x55, 0xD5, 0x35, 0xB5, 0x75, 0xF5,
  0x0D, 0x8D, 0x4D, 0xCD, 0x2D, 0xAD, 0x6D, 0xED, 0x1D, 0x9D, 0x5D, 0xDD, 0x3D, 0xBD, 0x7D, 0xFD,
  0x03, 0x83, 0x43, 0xC3, 0x23, 0xA3, 0x63, 0xE3, 0x13, 0x93, 0x53, 0xD3, 0x33, 0xB3, 0x73, 0xF3, 
  0x0B, 0x8B, 0x4B, 0xCB, 0x2B, 0xAB, 0x6B, 0xEB, 0x1B, 0x9B, 0x5B, 0xDB, 0x3B, 0xBB, 0x7B, 0xFB,
  0x07, 0x87, 0x47, 0xC7, 0x27, 0xA7, 0x67, 0xE7, 0x17, 0x97, 0x57, 0xD7, 0x37, 0xB7, 0x77, 0xF7, 
  0x0F, 0x8F, 0x4F, 0xCF, 0x2F, 0xAF, 0x6F, 0xEF, 0x1F, 0x9F, 0x5F, 0xDF, 0x3F, 0xBF, 0x7F, 0xFF
};

int main()
{
    unsigned int *ints = malloc(100000000*sizeof(unsigned int));
    unsigned int *ints2 = malloc(100000000*sizeof(unsigned int));
    for(unsigned int i = 0; i < 100000000; i++)
      ints[i] = rand();

    unsigned int *inptr = ints;
    unsigned int *outptr = ints2;
    unsigned int *endptr = ints + 100000000;
    // Starting the time measurement
    double start = omp_get_wtime();
    // Computations to be measured
    while(inptr != endptr)
    {
    unsigned int in = *inptr;  

    // Option 1:
    //*outptr = (BitReverseTable256[in & 0xff] << 24) | 
    //    (BitReverseTable256[(in >> 8) & 0xff] << 16) | 
    //    (BitReverseTable256[(in >> 16) & 0xff] << 8) |
    //    (BitReverseTable256[(in >> 24) & 0xff]);

    // Option 2:
    unsigned char * p = (unsigned char *) &(*inptr);
    unsigned char * q = (unsigned char *) &(*outptr);
    q[3] = BitReverseTable256[p[0]]; 
    q[2] = BitReverseTable256[p[1]]; 
    q[1] = BitReverseTable256[p[2]]; 
    q[0] = BitReverseTable256[p[3]];

      inptr++;
      outptr++;
    }
    // Measuring the elapsed time
    double end = omp_get_wtime();
    // Time calculation (in seconds)
    printf("Time: %f seconds\n", end-start);

    free(ints);
    free(ints2);

    return 0;
}

I tried both approaches at several different optimizations, ran 3 trials at each level, and each trial reversed 100 million random unsigned ints. For the lookup table option, I tried both schemes (options 1 and 2) given on the bitwise hacks page. Results are shown below.

Bitwise AND

mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -o reverse reverse.c
mrj10@mjlap:~/code$ ./reverse
Time: 2.000593 seconds
mrj10@mjlap:~/code$ ./reverse
Time: 1.938893 seconds
mrj10@mjlap:~/code$ ./reverse
Time: 1.936365 seconds
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O2 -o reverse reverse.c
mrj10@mjlap:~/code$ ./reverse
Time: 0.942709 seconds
mrj10@mjlap:~/code$ ./reverse
Time: 0.991104 seconds
mrj10@mjlap:~/code$ ./reverse
Time: 0.947203 seconds
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O3 -o reverse reverse.c
mrj10@mjlap:~/code$ ./reverse
Time: 0.922639 seconds
mrj10@mjlap:~/code$ ./reverse
Time: 0.892372 seconds
mrj10@mjlap:~/code$ ./reverse
Time: 0.891688 seconds

Lookup Table (option 1)

mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -o reverse_lookup reverse_lookup.c
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.201127 seconds              
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.196129 seconds              
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.235972 seconds              
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O2 -o reverse_lookup reverse_lookup.c
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 0.633042 seconds              
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 0.655880 seconds              
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 0.633390 seconds              
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O3 -o reverse_lookup reverse_lookup.c
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 0.652322 seconds              
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 0.631739 seconds              
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 0.652431 seconds

Lookup Table (option 2)

mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -o reverse_lookup reverse_lookup.c
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.671537 seconds
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.688173 seconds
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.664662 seconds
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O2 -o reverse_lookup reverse_lookup.c
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.049851 seconds
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.048403 seconds
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.085086 seconds
mrj10@mjlap:~/code$ gcc -fopenmp -std=c99 -O3 -o reverse_lookup reverse_lookup.c
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.082223 seconds
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.053431 seconds
mrj10@mjlap:~/code$ ./reverse_lookup
Time: 1.081224 seconds

Conclusion

Use the lookup table, with option 1 (byte addressing is unsurprisingly slow) if you're concerned about performance. If you need to squeeze every last byte of memory out of your system (and you might, if you care about the performance of bit reversal), the optimized versions of the bitwise-AND approach aren't too shabby either.

Caveat

Yes, I know the benchmark code is a complete hack. Suggestions on how to improve it are more than welcome. Things I know about:

I don't have access to ICC. This may be faster (please respond in a comment if you can test this out).
A 64K lookup table may do well on some modern microarchitectures with large L1D.
-mtune=native didn't work for -O2/-O3 (ld blew up with some crazy symbol redefinition error), so I don't believe the generated code is tuned for my microarchitecture.
There may be a way to do this slightly faster with SSE. I have no idea how, but with fast replication, packed bitwise AND, and swizzling instructions, there's got to be something there.
I know only enough x86 assembly to be dangerous; here's the code GCC generated on -O3 for option 1, so somebody more knowledgable than myself can check it out:

32-bit

.L3:
movl    (%r12,%rsi), %ecx
movzbl  %cl, %eax
movzbl  BitReverseTable256(%rax), %edx
movl    %ecx, %eax
shrl    $24, %eax
mov     %eax, %eax
movzbl  BitReverseTable256(%rax), %eax
sall    $24, %edx
orl     %eax, %edx
movzbl  %ch, %eax
shrl    $16, %ecx
movzbl  BitReverseTable256(%rax), %eax
movzbl  %cl, %ecx
sall    $16, %eax
orl     %eax, %edx
movzbl  BitReverseTable256(%rcx), %eax
sall    $8, %eax
orl     %eax, %edx
movl    %edx, (%r13,%rsi)
addq    $4, %rsi
cmpq    $400000000, %rsi
jne     .L3

EDIT: I also tried using uint64_t types on my machine to see if there was any performance boost. Performance was about 10% faster than 32-bit, and was nearly identical whether you were just using 64-bit types to reverse bits on two 32-bit int types at a time, or whether you were actually reversing bits in half as many 64-bit values. The assembly code is shown below (for the former case, reversing bits for two 32-bit int types at a time):

.L3:
movq    (%r12,%rsi), %rdx
movq    %rdx, %rax
shrq    $24, %rax
andl    $255, %eax
movzbl  BitReverseTable256(%rax), %ecx
movzbq  %dl,%rax
movzbl  BitReverseTable256(%rax), %eax
salq    $24, %rax
orq     %rax, %rcx
movq    %rdx, %rax
shrq    $56, %rax
movzbl  BitReverseTable256(%rax), %eax
salq    $32, %rax
orq     %rax, %rcx
movzbl  %dh, %eax
shrq    $16, %rdx
movzbl  BitReverseTable256(%rax), %eax
salq    $16, %rax
orq     %rax, %rcx
movzbq  %dl,%rax
shrq    $16, %rdx
movzbl  BitReverseTable256(%rax), %eax
salq    $8, %rax
orq     %rax, %rcx
movzbq  %dl,%rax
shrq    $8, %rdx
movzbl  BitReverseTable256(%rax), %eax
salq    $56, %rax
orq     %rax, %rcx
movzbq  %dl,%rax
shrq    $8, %rdx
movzbl  BitReverseTable256(%rax), %eax
andl    $255, %edx
salq    $48, %rax
orq     %rax, %rcx
movzbl  BitReverseTable256(%rdx), %eax
salq    $40, %rax
orq     %rax, %rcx
movq    %rcx, (%r13,%rsi)
addq    $8, %rsi
cmpq    $400000000, %rsi
jne     .L3

Matlab – Finding entropy in opencv

The big slow down in your code is this: log(gray_frequency)/log(2.0f)).

You should not call cvNormalizeHist(). You know the bins are going to sum to 81, so just subtract 81 * log(81)/log(2) from the calculated entropy (but of course that is a constant not calcualted every time in your loop). If you don't normalize the hisgram, its entries will be integers and you can use them to access a lookup table.

Since you have a 9x9 kernel the maximum value of gray_frequency is 81 (as long as you don't normalize the histogram) and you can easily replace those two calls to log() by a single lookup of a precalculated table. This will make a huge difference. You can initialize a table like this:

    double entropy_table[82]; // 0 .. 81
    const double log2 = log(2.0);
    entropy_table[0] = 0.0;
    for(int i = 1; i < 82; i ++)
    {
        entropy_table[i] = i * log(double(i)) / log2;
    }

Then later it is just:

entroy -= entropy_table[gray_frequency];

Also you may find implementing your own histgram code is a win. E.g. if you have a small kernel you can keep track of which bins you are going to use and only clear those. But since you are using 81/256 bins this mightn't be worth it.

Another place you can get a speed up is in borrder pixel handling. You are checking this for every pixel. But oif you had separate loops for the boarder pixels and the inner pixels a re lot of calls to max and min could be avoided.

If that still isn't fast enough, you may consider using parallel_for on stripes. As a good example on how to do that, have a look at the source code for OpenCV's morphological filter.