int main2mod(unsigned long long spe_id, unsigned long long program_data_ea, unsigned long long env)
{
unsigned tagid = spe_id&31;
uint32 i,j;
// get program data
mfc_get(&pd, program_data_ea, sizeof(spu_program_data), tagid, 0, 0);
mfc_write_tag_mask(1<<tagid);
mfc_read_tag_status_all();
// precompute partial working states based on ihv & partial msg block
pre_compute(pd.ihv1, pd.ihv2, pd.m1, pd.m2);
if (pd.collisiondata > 0)
{
j = pd.collisiondata*8;
vec_uint32* bufferptr = &buffer[j];
// get the trail buffer
for (i = 0; i < j; i += 128)
mfc_get(&buffer[i], &pd.buffer[i], sizeof(vec_uint32)*128, tagid, 0, 0);
mfc_write_tag_mask(1<<tagid);
mfc_read_tag_status_all();
// process collision trails
reduce_trailsmod(pd.ihv1, pd.ihv2mod, pd.m1, pd.m2, &pd.astart, &buffer[0], bufferptr);
reduce_trails2mod(pd.ihv1, pd.ihv2mod, pd.m1, pd.m2, &pd.astart, &buffer[0], bufferptr);
find_collmod(pd.ihv1, pd.ihv2mod, pd.m1, pd.m2, &pd.astart, &buffer[0], bufferptr);
// store the trail buffer
for (i = 0; i < j; i += 128)
mfc_put(&buffer[i], &pd.buffer[i], sizeof(vec_uint32)*128, tagid, 0, 0);
mfc_write_tag_mask(1<<tagid);
mfc_read_tag_status_all();
} else {
// fill the trail buffer in steps and do intermediate DMA transfers
vec_uint32* bufferptr = &buffer[0];
for (i = 0; i < BUFFERSIZE; i += 256)
{
bufferptr = generate_trailsmod(pd.ihv1, pd.ihv2mod, pd.m1, pd.m2, &pd.astart, bufferptr, &buffer[i+256]);
mfc_put(&buffer[i], &pd.buffer[i], sizeof(vec_uint32)*128, tagid, 0, 0);
mfc_put(&buffer[i+128], &pd.buffer[i+128], sizeof(vec_uint32)*128, tagid, 0, 0);
}
}
// transfer the current program data back
mfc_put(&pd, program_data_ea, sizeof(spu_program_data), tagid, 0, 0);
// wait for dma transfers to complete
mfc_write_tag_mask(1<<tagid);
mfc_read_tag_status_all();
return 0;
}
void read_from_stdin()
{
int n, i, value;
int pos = 0;
while((n = read(0, buf + pos, BUFSIZE - pos)) > 0){
pos += n;
}
gettimeofday(&t[1], NULL);
buf[pos] = '\n';
buf[pos + 1] = '\0';
char *space = index(buf, ' ');
N = atoi(buf);
D = atoi(space + 1);
gettimeofday(&t[2], NULL);
/// read item prices.
char *src = index(space + 1, '\n') + 1;
value = 0;
i = 0;
while(i < N){
if(*src != '\n'){
value = 10 * value + digit[(int)*src];
}else{
count_and_offset[value]++;
value = 0;
i++;
}
src ++;
}
gettimeofday(&t[3], NULL);
/// read campaign prices.
int *dst = cprices;
i = 0;
value = 0;
while(i < D){
if(*src != '\n'){
value = 10 * value + digit[(int)*src];
}else{
*(dst++) = value;
value = 0;
i++;
}
src ++;
}
gettimeofday(&t[4], NULL);
pre_compute();
}
