Is this good?

int array[COLS][ROWS];
long sum = 0;

// initialize array.

for (int r=0; r<ROWS; ++r)
    for (int c=0; c<COLS; ++c)
        sum += array[c][r];

Much better!

int array[COLS][ROWS];
long sum = 0;

// initialize array.

for (int c=0; c<COLS; ++c)
    for (int r=0; r<ROWS; ++r)
        sum += array[c][r];

Same amount of instructions,

same amount of computation,

but

DIFFERENT SPEEDS!!!

You forgot about the magician inside the machine

Modern software gave you valuable abstractions, but you might have forgotten that you are talking with a machine. 

It is getting worse and worse!!

Core

L1 cache

L2 cache

L3 cache

...

• L1 cache: 32 KB, 4 cycles
• L2 cache: 256 kB, 10 cycles
• L3 cache: 8MB, ~30 cycles
• DRAM: X GB, ~90 cycles

Core

L1 cache

L2 cache

Core

L1 cache

L2 cache

Cache is copied using contiguous segments

• Typically 64 bytes...
• The same cache line is copied on multiple CPUs.
• If any CPU modifies it, the entire line is marked "dirty".

int doWork(const vector<int>& data)
{
    unsigned long sum = 0;

    for (int d : data)
        sum += d * d;

    return (int)(sum / data.size());
}

Case A: iterate through a vector

auto data = vector<int>(SIZE);
initData( data );
doWork( data ); //benchmark this

int doWork(const vector<int*>& data)
{
    unsigned long sum = 0;

    for (int* d : data)
        sum += (*d) * (*d);

    return (int)(sum / data.size());
}

Case B: simple indirection

auto data     = vector<int>(SIZE);
auto pointers = vector<int*>(data.size());
for (size_t i = 0; i < data.size(); ++i)
    pointers[i] = &data[i]; // pointing to contiguous memory
}

initData( data );

doWork( pointers );

int doWork(const vector<int*>& data)
{
    unsigned long sum = 0;

    for (int* d : data)
        sum += (*d) * (*d);

    return (int)(sum / data.size());
}

Case C: shuffled memory

auto pointers = vector<int*>( dataSize );
for (size_t i = 0; i < dataSize; ++i)
    pointers[i] = new int; // pointing to "sparse memory"
}

initData( pointers );

doWork( pointers );

Example of antipattern

struct Object{
    bool toBeUpdated;
    // ... 
    //stuff
    //....
};


std::vector< Object > objectList;


//main loop
for(int i=0; i objectList.size(); i++){

    if( objectList[i].toBeUpdated )
        updateObject( objectList[i] );
}

Top get the value of the field "toBeUpdated", you are obliged to read a chunk of the Object

Fix

std::vector< bool > toBeUpdated;
std::vector< Object > objectList;


//main loop
for(int i=0; i objectList.size(); i++){

    if( toBeUpdated[i] )
        updateObject( objectList[i] );
}

You don't need to read the Object from memory unless "toBeUpdated[i]" is true

There is more...

false sharing!

void computeOdds(int *matrix_chunk, int chunk_size, int &partial_count)
{
    for( int i = 0; i < chunk_size; ++i )
        if( matrix_chunk[i] %2 == 0 )
            partial_count++;
}

int computeOddsInParallel( int *matrix, int DIM )
{
    int result[POOL_SIZE];
    
    // Each of P parallel workers processes 1/P-th  of the data
    // let's suppose for the sake of simplicity that DIM is a multiple of POOL_SIZE...
    int chunkSize = DIM/POOL_SIZE; 
    
    for( int p = 0; p < POOL_SIZE; ++p )
    {
        // push partial work inside a worker
        pool.run( std::bind(computeOdds, &matrix[ p*chunkSize ],  chunkSize, result[p] ));
    } 
    
    // Wait for the parallel work to complete…
    pool.join();
    
    // Finally, do the sequential "reduction" step to combine the results
    int odds = 0;
    for( int p = 0; p < POOL_SIZE; ++p )
      odds += result[p];

    return odds;
}

Making it parallel (some pseudo code used)

void computeOdds(int *matrix_chunk,
                 int chunk_size, 
                 int &partial_count)
{
    int count = 0;
    for( int i = 0; i < chunk_size; ++i )
        if( matrix_chunk[i] %2 == 0 )
            count++;
    partial_count = count;
}

void computeOdds(int *matrix_chunk,
                 int chunk_size, 
                 int &partial_count)
{
    
    for( int i = 0; i < chunk_size; ++i )
        if( matrix_chunk[i] %2 == 0 )
            partial_count ++;

}

int result[POOLS_SIZE] was on a single cache line. 

Resources

• http://www.drdobbs.com/parallel/eliminate-false-sharing/217500206

• http://gameprogrammingpatterns.com/data-locality.html

• http://bitbashing.io/memory-performance.html

• https://www.youtube.com/watch?v=WDIkqP4JbkE&index=31&list=PLXIxC-YkNzOKdmwQvTJvfZa-E69AXNY70

• https://www.youtube.com/watch?v=rX0ItVEVjHc
• https://www.youtube.com/watch?v=KN8MFFvRl50