#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#define NUM_WORKERS 7
#define NUM_ELEMENTS 1000000
typedef struct
{
float *input;
float *output;
int start;
int count;
} Job;
void *spe_worker(void *arg)
{
Job *job = (Job *)arg;
float local_buffer[4096];
int processed = 0;
while (processed < job->count)
{
int chunk = job->count - processed;
if (chunk > 4096)
chunk = 4096;
for (int i = 0; i < chunk; i++)
local_buffer[i] = job->input[job->start + processed + i];
for (int i = 0; i < chunk; i++)
{
float x = local_buffer[i];
local_buffer[i] = x * x + 2.0f * x + 1.0f;
}
for (int i = 0; i < chunk; i++)
job->output[job->start + processed + i] = local_buffer[i];
processed += chunk;
}
return NULL;
}
int main(void)
{
float *input = malloc(sizeof(float) * NUM_ELEMENTS);
float *output = malloc(sizeof(float) * NUM_ELEMENTS);
if (!input || !output)
return 1;
for (int i = 0; i < NUM_ELEMENTS; i++)
input[i] = (float)i;
pthread_t spe[NUM_WORKERS];
Job jobs[NUM_WORKERS];
int elements_per_spe = NUM_ELEMENTS / NUM_WORKERS;
for (int i = 0; i < NUM_WORKERS; i++)
{
jobs[i].input = input;
jobs[i].output = output;
jobs[i].start = i * elements_per_spe;
jobs[i].count =
(i == NUM_WORKERS - 1)
? NUM_ELEMENTS - jobs[i].start
: elements_per_spe;
pthread_create(
&spe[i],
NULL,
spe_worker,
&jobs[i]
);
}
for (int i = 0; i < NUM_WORKERS; i++)
pthread_join(spe[i], NULL);
printf("Result[0] = %f\n", output[0]);
printf("Result[999999] = %f\n", output[999999]);
free(input);
free(output);
return 0;
}