int main( int argc, char** argv )
{
/* loop variables */
int x,y;
/* Create matrixes */
unsigned int * input_array;
unsigned int * serial_array;
unsigned int * output_array;
unsigned int * filter_list;
/* Initialize the data. Values don't matter much. */
posix_memalign ( (void**)&input_array, 4096, DATA_LEN * sizeof(unsigned int));
// input_array = (unsigned int*) posix_memalign ( DATA_LEN * sizeof(unsigned int), 4096);
for (x=0; x<DATA_LEN; x++)
{
input_array[x] = x % 2048;
}
/* start with an empty *all zeros* output array */
posix_memalign ( (void**)&output_array, 4096, DATA_LEN * sizeof(unsigned int));
memset ( output_array, 0, DATA_LEN );
posix_memalign ( (void**)&serial_array, 4096, DATA_LEN * sizeof(unsigned int));
memset ( serial_array, 0, DATA_LEN );
/* Initialize the filter. Values don't matter much. */
filter_list = (unsigned int*) malloc ( FILTER_LEN * sizeof(unsigned int));
for (y=0; y<FILTER_LEN; y++)
{
filter_list[y] = y;
}
/* Execute at a variety of filter lengths */
for ( int filter_len =16; filter_len<=FILTER_LEN; filter_len*=2)
{
serialDataFirst ( DATA_LEN, input_array, serial_array, filter_len, filter_list );
memset ( output_array, 0, DATA_LEN );
serialFilterFirst ( DATA_LEN, input_array, output_array, filter_len, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
//omp_set_num_threads(16);
parallelFilterFirst ( DATA_LEN, input_array, output_array, filter_len, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
parallelDataFirst ( DATA_LEN, input_array, output_array, filter_len, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
}
}
int main( int argc, char** argv )
{
/* loop variables */
int x,y;
/* Create matrixes */
unsigned int * input_array;
unsigned int * serial_array;
unsigned int * output_array;
unsigned int * filter_list;
/* Initialize the data. Values don't matter much. */
posix_memalign ( (void**)&input_array, 4096, DATA_LEN * sizeof(unsigned int));
// input_array = (unsigned int*) posix_memalign ( DATA_LEN * sizeof(unsigned int), 4096);
for (x=0; x<DATA_LEN; x++)
{
input_array[x] = x % 2048;
}
/* start with an empty *all zeros* output array */
posix_memalign ( (void**)&output_array, 4096, DATA_LEN * sizeof(unsigned int));
memset ( output_array, 0, DATA_LEN );
posix_memalign ( (void**)&serial_array, 4096, DATA_LEN * sizeof(unsigned int));
memset ( serial_array, 0, DATA_LEN );
/* Initialize the filter. Values don't matter much. */
filter_list = (unsigned int*) malloc ( FILTER_LEN * sizeof(unsigned int));
for (y=0; y<FILTER_LEN; y++)
{
filter_list[y] = y;
}
int n;
printf("Please input a DIGIT(1: Loop efficiency; 2: Parallel speedup; 3: Parallel scaleup; 4: Optimized verion speedup; 5: Optimized version scaleup): \n");
scanf("%d", &n);
// serialDataFirst ( DATA_LEN, input_array, serial_array, 512, filter_list );
// memset ( output_array, 0, DATA_LEN );
if (n == 4) {
printf("Optimized verion speedup:\n");
serialDataFirst ( DATA_LEN, input_array, serial_array, 512, filter_list );
memset ( output_array, 0, DATA_LEN );
/*************************** optimized speedup ***************************/
for ( int i = 1; i <= 20; i++)
{
for (int j = 1; j <= 16; j = j * 2) {
omp_set_num_threads(j);
optimizedParallelDataFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
omp_set_num_threads(j);
optimizedParallelFilterFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
}
printf("\n");
}
}
if (n == 5) {
printf("Optimized verion scaleup:\n");
serialDataFirst ( DATA_LEN, input_array, serial_array, 512, filter_list );
memset ( output_array, 0, DATA_LEN );
/*************************** optimized scaleup (need to alter DATA_LEN as well) ***************************/
for ( int i = 1; i <= 20; i++)
{
omp_set_num_threads(4);
optimizedParallelDataFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
omp_set_num_threads(4);
optimizedParallelFilterFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
}
printf("\n");
}
if (n == 1) {
printf("Loop efficiency:\n");
/*************************** loop efficiency part ***************************/
for ( int i = 1; i <= 20; i++) {
/* Execute at a variety of filter lengths */
for ( int filter_len =1; filter_len<=FILTER_LEN; filter_len*=2) //filter_len =16
{
serialDataFirst ( DATA_LEN, input_array, serial_array, filter_len, filter_list );
memset ( output_array, 0, DATA_LEN );
serialFilterFirst ( DATA_LEN, input_array, output_array, filter_len, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
}
}
}
if (n == 2) {
printf("Parallel speedup:\n");
serialDataFirst ( DATA_LEN, input_array, serial_array, 512, filter_list );
memset ( output_array, 0, DATA_LEN );
/*************************** speedup ***************************/
for ( int i = 1; i <= 20; i++) {
for (int i = 1; i <= 16; i = i * 2) {
omp_set_num_threads(i);
parallelFilterFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
omp_set_num_threads(i);
parallelDataFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
}
}
}
if (n == 3) {
printf("Parallel scaleup:\n");
serialDataFirst ( DATA_LEN, input_array, serial_array, 512, filter_list );
memset ( output_array, 0, DATA_LEN );
/*************************** scaleup (need to alter DATA_LEN as well)***************************/
for (int i = 1; i <= 20; i++) {
omp_set_num_threads(4);
parallelFilterFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
omp_set_num_threads(4);
parallelDataFirst ( DATA_LEN, input_array, output_array, 512, filter_list );
checkData ( serial_array, output_array );
memset ( output_array, 0, DATA_LEN );
}
}
}