Optimisation hint for array of random numbers - c++

Question

To provide context, I'm working through Programming Praxis Bingo Challenge and wanted to see how fast I could make this code run.
static void fisher_yates(T& source) {
const size_t len = source.size();
for(size_t i = 1; i < len;++i) {
std::swap(source[i],source[rand() % (i+1)]);
}
}
std::array<int,25> generate_table() {
std::array<int,25> bingo_grid;
for(int i = 0 ; i < 25;++i) {
switch(i) {
case 0: case 1: case 2: case 3: case 4:
bingo_grid[i] = rand() % 15 + 1;
break;
case 5: case 6: case 7: case 8: case 9:
bingo_grid[i] = rand() % 15 + 16;
break;
case 10: case 11: case 12: case 13: case 14:
bingo_grid[i] = rand() % 15 + 31;
break;
case 15: case 16: case 17: case 18: case 19:
bingo_grid[i] = rand() % 15 + 46;
break;
case 20: case 21: case 22: case 23: case 24:
bingo_grid[i] = rand() % 15 + 61;
break;
}
}
bingo_grid[12] = 0;
return bingo_grid;
}
bool is_bingoed(const std::array<int,25>& grid) {
// Check columns
if(grid[0] == 0) {
if(grid[1] == 0 && grid[2] == 0 && grid[3] == 0 && grid[4] == 0)
return true;
if(grid[0] == 0 && grid[6] == 0 && grid[18] == 0 && grid[24] == 0)
return true;
if(grid[5] == 0 && grid[10] == 0 && grid[15] == 0 && grid[20] == 0)
return true;
}
if(grid[1] == 0) {
if(grid[6] == 0 && grid[11] == 0 && grid[16] == 0 && grid[21] == 0)
return true;
}
if(grid[2] == 0) {
if(grid[7] == 0 && grid[17] == 0 && grid[22] == 0)
return true;
}
if(grid[3] == 0) {
if(grid[8] == 0 && grid[13] == 0 && grid[18] == 0 && grid[23] == 0)
return true;
}
if(grid[4] == 0) {
if(grid[9] == 0 && grid[14] == 0 && grid[19] == 0 && grid[24] == 0)
return true;
if(grid[8] == 0 && grid[16] == 0 && grid[21] == 0)
return true;
}
if(grid[6] == 0) {
if(grid[6] == 0 && grid[7] == 0 && grid[8] == 0 && grid[9] == 0)
return true;
}
if(grid[12] == 0) {
if(grid[10] == 0 && grid[11] == 0 && grid[13] == 0 && grid[14] == 0)
return true;
}
if(grid[18] == 0) {
if(grid[15] == 0 && grid[16] == 0 && grid[17] == 0 && grid[19] == 0)
return true;
}
return false;
}
static bool mark_card(const int card,std::array<int,25>& bingo_grid) {
for(auto &i : bingo_grid)
if(card == i) {
i = 0;
return true;
}
return false;
}
int play_game() {
// Bingo is 5 columns, each column(n) is random permutation of 1-15*n
// Fisher-Yates to generate random permutations
// Create 500 playing cards
const int max = 500;
std::vector<std::array<int,25>> bingo_cards;
bingo_cards.reserve(max);
for(int i = 0; i<max;++i) {
bingo_cards.push_back(generate_table());
//display_bingo(bingo_cards[i]);
}
// Random shuffle 75 cards
auto iter = boost::counting_range(1,76);
std::vector<int> cards(std::begin(iter),std::end(iter));
fisher_yates(cards);
bool is_finished = false;
int counter = 0;
for(auto card : cards) {
for(auto& playing_card : bingo_cards) {
if(mark_card(card,playing_card)) {
//display_bingo(playing_card);
if(is_bingoed(playing_card))
return counter;
}
}
counter++;
}
return counter;
}
int bingo() {
srand(time(NULL));
int total = 0;
for(int i = 0 ; i < 10000;i++) {
total+=play_game();
}
boost::singleton_pool<boost::pool_allocator_tag, sizeof(int)>::release_memory();
return total / 10000;
}
The original version used a boost::multi_array to represent the grid. After profiling, I changed it to a std::array which got me a significant speed up. I then moved from using fisher_yates shuffle to generate bingo cards to using a random number generator.
Then finally I changed the is_bingoed test function to reduce the number of checks per call to speed up the game-over check.
All this has helped. Right now if I profile this code, the generate_table function takes up 72% of the time, mark_card() is 18%, and is_bingoed() about 6%. I'm looking for hints to see what can be done to improve the speed of either.
My first thought with is_bingoed() is to use the SSE intrinsics to do a compare with 0 (maybe use XOR?) but I don't have any ideas on the generate_table() or mark_car(). This is more of a self challenge for fun but wondered what others thought?
Current timing is it takes 4.6s on a 2Ghz Q6660 (down from 35s originally)

Answer 1

Just focussing on your most expensive function, generate_table, you can simplify this part of the code and make it less branchy, which may help:
for(int i = 0 ; i < 25;++i) {
switch(i) {
case 0: case 1: case 2: case 3: case 4:
bingo_grid[i] = rand() % 15 + 1;
break;
case 5: case 6: case 7: case 8: case 9:
bingo_grid[i] = rand() % 15 + 16;
break;
case 10: case 11: case 12: case 13: case 14:
bingo_grid[i] = rand() % 15 + 31;
break;
case 15: case 16: case 17: case 18: case 19:
bingo_grid[i] = rand() % 15 + 46;
break;
case 20: case 21: case 22: case 23: case 24:
bingo_grid[i] = rand() % 15 + 61;
break;
}
}
e.g.
for(int i = 0 ; i < 25;++i) {
int r = rand() % 15 + 1;
bingo_grid[i] = r + (i / 5) * 15;
}
Beyond that I'd look at a faster rand() and also see if you can get rid of the divide and and modulo.
On a separate note, your algorithm may be flawed in that there is nothing to prevent duplicate numbers in bingo_grid.

Answer 2

Changing the is_bingoed() method to use SSE instructions (using Agner Fog's library) and Paul R's generate_table() reduced the time to just 1.05s. And using Intel's fast_rand() function got it down to 0.38s.
So I thought I'd paste the code changes for others who might be interested.
static unsigned int g_seed;
//Used to seed the generator.
inline void fast_srand( int seed )
{
g_seed = seed;
}
//fastrand routine returns one integer, similar output value range as C lib.
inline int fastrand()
{
g_seed = (214013*g_seed+2531011);
return (g_seed>>16)&0x7FFF;
}
bool is_bingoed(const std::array<int,25>& grid) {
// Check columns
Vec8i vec(grid[0],grid[1],grid[2],grid[3],grid[4],0,0,0);
Vec4i vec2(grid[6],grid[18],grid[24],0);
Vec4i vec3(grid[5],grid[10],grid[15],20);
Vec8i vec4(grid[1],grid[6],grid[11],grid[16],grid[21],0,0,0);
Vec4i vec5(grid[2],grid[7],grid[17],grid[22]);
Vec8i vec6(grid[3],grid[8],grid[13],grid[18],grid[23],0,0,0);
Vec8i vec7(grid[4],grid[9],grid[14],grid[19],grid[24],0,0,0);
Vec4i vec8(grid[8],grid[16],grid[21],grid[4]);
Vec4i vec9(grid[6],grid[7],grid[8],grid[9]);
Vec8i vec10(grid[12],grid[10],grid[11],grid[13],grid[14],0,0,0);
Vec8i vec11(grid[18],grid[15],grid[16],grid[17],grid[19],0,0,0);
if(horizontal_and(vec) && horizontal_and(vec2) && horizontal_and(vec3) && horizontal_and(vec4) &&
horizontal_and(vec5) && horizontal_and(vec6) && horizontal_and(vec7) && horizontal_and(vec8)) {
return false;
}
if(horizontal_and(vec9) && horizontal_and(vec10) && horizontal_and(vec11)) {
return false;
}
return true;
}