static void mincross_clust(Agraph_t *ug)
{
Agraph_t *g;
g = GD_model(ug);
if (run(g)) {
presort(ug); /* move the external nodes */
subclustports(ug);
do {
mincross_sweep(g,GD_pass(g)%2,GD_pass(g)%4<2);
} while (run(g));
transpose_sweep(g,TRUE);
restorebest(g);
}
}
void
tiled_mergesort(unsigned int a[], int N)
{
/* track keeps an eye on where i started last */
/* outer track keeps and eye on which cache_sized/2 segemnt i started in last */
int i,j; /* indices for the first array */
int level1_count;
int level2_count;
int extra_level1_count = 0;
int extra_level2;
int final_extra;
unsigned int* aux_data;
unsigned int* aux;
unsigned int* level2_start;
unsigned int* level2_aux_start;
unsigned int minusA;
unsigned int** level1_finish;
unsigned int** level1_other;
int odd = 0;
describe_predictor(&global_predictor[0], "forwards middle");
describe_predictor(&global_predictor[1], "forwards next");
describe_predictor(&global_predictor[2], "forwards end");
describe_predictor(&global_predictor[3], "forwards setup");
describe_predictor(&global_predictor[4], "forwards equal");
describe_predictor(&global_predictor[5], "reverse middle");
describe_predictor(&global_predictor[6], "reverse next");
describe_predictor(&global_predictor[7], "reverse end");
describe_predictor(&global_predictor[8], "reverse setup");
describe_predictor(&global_predictor[9], "reverse equal");
describe_predictor(&global_predictor[10], "insertion outer");
describe_predictor(&global_predictor[11], "insertion inner");
describe_predictor(&global_predictor[12], "insertion reverse outer");
describe_predictor(&global_predictor[13], "insertion reverse inner");
#ifdef _USE_ROLLED_LOOPS
describe_predictor(&global_predictor[14], "forwards left");
describe_predictor(&global_predictor[15], "forwards right");
describe_predictor(&global_predictor[16], "reverse left");
describe_predictor(&global_predictor[17], "reverse right");
#else
describe_predictor(&global_predictor[14], "forwards left 0");
describe_predictor(&global_predictor[15], "forwards left 1");
describe_predictor(&global_predictor[16], "forwards left 2");
describe_predictor(&global_predictor[17], "forwards left 3");
describe_predictor(&global_predictor[18], "forwards left 4");
describe_predictor(&global_predictor[19], "forwards left 5");
describe_predictor(&global_predictor[20], "forwards left 6");
describe_predictor(&global_predictor[21], "forwards left 7");
describe_predictor(&global_predictor[22], "forwards right 0");
describe_predictor(&global_predictor[23], "forwards right 1");
describe_predictor(&global_predictor[24], "forwards right 2");
describe_predictor(&global_predictor[25], "forwards right 3");
describe_predictor(&global_predictor[26], "forwards right 4");
describe_predictor(&global_predictor[27], "forwards right 5");
describe_predictor(&global_predictor[28], "forwards right 6");
describe_predictor(&global_predictor[29], "forwards right 7");
describe_predictor(&global_predictor[30], "reverse left 0");
describe_predictor(&global_predictor[31], "reverse left 1");
describe_predictor(&global_predictor[32], "reverse left 2");
describe_predictor(&global_predictor[33], "reverse left 3");
describe_predictor(&global_predictor[34], "reverse left 4");
describe_predictor(&global_predictor[35], "reverse left 5");
describe_predictor(&global_predictor[36], "reverse left 6");
describe_predictor(&global_predictor[37], "reverse left 7");
describe_predictor(&global_predictor[38], "reverse right 0");
describe_predictor(&global_predictor[39], "reverse right 1");
describe_predictor(&global_predictor[40], "reverse right 2");
describe_predictor(&global_predictor[41], "reverse right 3");
describe_predictor(&global_predictor[42], "reverse right 4");
describe_predictor(&global_predictor[43], "reverse right 5");
describe_predictor(&global_predictor[44], "reverse right 6");
describe_predictor(&global_predictor[45], "reverse right 7");
#endif
/* a quick explanation, cause I keep needing to be reminded how I did this:
* an address is split into 3 parts: the tag, the index and the offset.
* Suppose theres a 32 bit address, a 32 byte cache line and 65536 cache
* blocks, as in our tests. In this case, the 32 bit address is split into
* a 5 bit offset (2^5 = 32 byte cache line), a 16 bit index (2^16 = 65536
* cache blocks) and the rest is the tag. Therefore, a and aux need to
* have exactly the opposite index. minusA is the index aux needs to have,
* which we mask in. If this results in an address lower than the one we
* started with (in aux), then thats out of bounds, and increase it by
* 65536.
*/
/* get the index we need*/
minusA = get_index(a) ^ (1 << (BLOCK_BITS - 1));
aux_data = memalign(ALIGNMENT, (N + 2*LIMIT) * sizeof(unsigned int));
/* clear the index bits, and mask in the desired index */
aux = (unsigned int*)(((unsigned int)aux_data & (~BLOCK_AND_LINE_MASK)) | (minusA << LINE_BITS));
if (aux < aux_data) /* then the new index is less than the old one */
{
aux = (unsigned int*)((unsigned int)aux + (1 << (BLOCK_AND_LINE_BITS)));
}
if (N <= 2048) /* fits in the level 1 cache */
{
if (get_count(N) & 1) set_presort_count(ODD_COUNT);
else set_presort_count(EVEN_COUNT);
presort_flexible(a, N);
merge(a, N, presort_count, aux);
goto end;
}
/* OUT(get_count(N)); */
/* OUT(N); */
level2_count = N / LIMIT; /* the number of standard LIMIT sized passes */
level1_count = LIMIT / 1024; /* the number of standard 4k sized passes, per level 2 iteration */
extra_level2 = N % LIMIT; /* the number of extra items left, after the level 2 passes*/
extra_level1_count = extra_level2 / 1024; /* number of extra level 1 passes */
final_extra = extra_level2 % 1024; /* number of items left over */
level2_start = a;
level2_aux_start = aux;
/*make sure it ends up in a, not aux */
/* odd means it should end up in aux. and the final merge will get it into a */
/* even means it should end up in a, and the final mergre will do an even number of steps */
/* obviously, we should take this out. but the final bit is quite complex,
* it doesnt come up in the tests (due to using powers of 2) and the extra
* code doesnt cause too much of a hit. More importantly, it doesnt need to
* be optimal, as we're only interested in data cache and brnahc
* predictors, so the fact that theres a few extra instructions to get
* loaded doesnt make a difference. */
if (get_count(N) & 1)
{
level1_finish = &level2_aux_start;
level1_other = &level2_start;
set_presort_count(ODD_COUNT);
odd = 1;
}
else
{
level1_finish = &level2_start;
level1_other = &level2_aux_start;
set_presort_count(EVEN_COUNT);
odd = 0;
}
/* printf("a = %p\n", a);
printf("aux = %p\n", aux); */
for(i = 0; i < level2_count-1; i+=2) /* sort it level 2 */
{
/* printf("going into level2: i=%d\n", i); */
/* merge them all into LIMIT sized bits */
presort(level2_start, LIMIT);
merge(level2_start, LIMIT, presort_count, level2_aux_start);
level2_start += LIMIT;
level2_aux_start += LIMIT;
/* now do it in reverse */
presort(level2_start, LIMIT);
merge_reverse(level2_start, LIMIT, presort_count, level2_aux_start);
level2_start += LIMIT;
level2_aux_start += LIMIT;
}
if (i < level2_count)
{
/* merge them all into LIMIT sized bits */
presort(level2_start, LIMIT);
merge(level2_start, LIMIT, presort_count, level2_aux_start);
level2_start += LIMIT;
level2_aux_start += LIMIT;
}
/* this bit is too complicated to remove the 2 level tiling from. It doesnt
* alter the results either, since we use powers of two */
/* sort the remaining bits */
/* level2 start is in the right place */
/* OUT(extra_level1_count); */
if (extra_level2) /* there is a maximum of 1 extra tevel 2 sort */
{
int extra_level1_single = extra_level1_count & 0x1;
unsigned int* level1_start = level2_start;
unsigned int* level1_aux_start = level2_aux_start;
extra_level1_count &= ~0x1; /*clear the last bit */
/* OUT(extra_level2); */
/* OUT(extra_level1_count); */
for(j = 0; j < extra_level1_count; j+=2) /* merge the level 1 cache first */
{
presort(level1_start, 1024);
merge(level1_start, 1024, presort_count, level1_aux_start); /* after this they end up in aux */
level1_start += 1024;
level1_aux_start += 1024;
/* now reverse it */
presort(level1_start, 1024);
merge_reverse(level1_start, 1024, presort_count, level1_aux_start);
level1_start += 1024;
level1_aux_start += 1024;
/* these end up in aux */
}
/* OUT(extra_level1_single); */
if (extra_level1_single)/* if there a full one left, its forward */
{
presort(level1_start, 1024);
merge(level1_start, 1024, presort_count, level1_aux_start); /* after this they end up in aux */
level1_start += 1024;
level1_aux_start += 1024;
}
/* OUT(final_extra); */
if (final_extra) /* theres less than a full level1 sized chunk */
{
/* these will be sorted in one go */
/* if it turns out the number should be 2049, I may need to change this */
if (final_extra <= double_presort_count)
{
if (!odd) set_presort_count(ODD_COUNT);
}
else
{
if (get_count(final_extra) & 1)
{
if (odd) set_presort_count(EVEN_COUNT);
else set_presort_count(ODD_COUNT);
}
}
/* should this be reversed or not */
if (!extra_level1_single)
{
presort_flexible(level1_start, final_extra);
merge(level1_start, final_extra, presort_count, level1_aux_start);
}
else
{
presort_flexible_reverse(level1_start, final_extra);
merge_reverse(level1_start, final_extra, presort_count, level1_aux_start);
}
}
/* merge the whole extra into 1 */
merge(*level1_finish, extra_level2, 1024, *level1_other);
}
/* now merge everything together */
if (N > LIMIT)
{
if (odd) /* its in aux - we took steps to ensure it*/
{
merge(aux, N, LIMIT, a);
}
else
{
merge(a, N, LIMIT, aux);
}
}
end:
free(aux_data);
init_predictor(&global_predictor[0]);
init_predictor(&global_predictor[1]);
init_predictor(&global_predictor[2]);
init_predictor(&global_predictor[3]);
init_predictor(&global_predictor[4]);
init_predictor(&global_predictor[5]);
init_predictor(&global_predictor[6]);
init_predictor(&global_predictor[7]);
init_predictor(&global_predictor[8]);
init_predictor(&global_predictor[9]);
init_predictor(&global_predictor[10]);
init_predictor(&global_predictor[11]);
init_predictor(&global_predictor[12]);
init_predictor(&global_predictor[13]);
}
