Changeset: 8fe4dd86cfe2 for MonetDB
URL: http://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=8fe4dd86cfe2
Modified Files:
monetdb5/extras/crackers/crackers_core_unordered.mx
Branch: holindex
Log Message:
Handle elements that do not "fit" in vectors after the merge phase.
diffs (170 lines):
diff --git a/monetdb5/extras/crackers/crackers_core_unordered.mx
b/monetdb5/extras/crackers/crackers_core_unordered.mx
--- a/monetdb5/extras/crackers/crackers_core_unordered.mx
+++ b/monetdb5/extras/crackers/crackers_core_unordered.mx
@@ -923,9 +923,8 @@ static str CRKvectorized_x_@2_@1 (
BUN last_left = first_left + ml - 1, first_right = last_right + 1 - mr;
BUN valueCount = ( (ml && mr && last_left + 1 < first_right) ? ml + mr
: last_right - first_left + 1 );
const BUN vectorCount = valueCount/vector_elements;
- const BUN remaining_elements = valueCount % vector_elements;
- BUN lowerReadCursor = first_left, upperReadCursor = last_right -
remaining_elements + 1;
- BUN lowerWriteCursor = first_left, upperWriteCursor = last_right -
remaining_elements;
+ BUN lowerReadCursor = first_left, upperReadCursor = last_right + 1;
+ BUN lowerWriteCursor = first_left, upperWriteCursor = last_right;
BUN vectorI = 0, vectorR = 3, skip, skip_left, skip_right;
cursorDeltas deltas;
BAT *localBuffer = BATnew(TYPE_oid, TYPE_@1, vector_elements*3);
@@ -933,12 +932,11 @@ static str CRKvectorized_x_@2_@1 (
@1 *src_t = (@1 *) Tloc(buffer, BUNfirst(buffer));
oid *src_h_local = (oid*) Hloc(localBuffer, BUNfirst(localBuffer));
@1 *src_t_local = (@1 *) Tloc(localBuffer, BUNfirst(localBuffer));
- first_right = last_right - remaining_elements + 1 - mr;
assert(buffer || localBuffer);
assert(pos_r);
- //assert(valueCount%(2*vector_elements) == 0);
- //assert(!(ml && mr && last_left + 1 < first_right) ||
(ml%(2*vector_elements) == 0 && mr%(2*vector_elements) == 0));
+ assert(valueCount%(2*vector_elements) == 0);
+ assert(!(ml && mr && last_left + 1 < first_right) ||
(ml%(2*vector_elements) == 0 && mr%(2*vector_elements) == 0));
memcpy(src_h_local, src_h + first_left, sizeof(oid)*2*vector_elements);
memcpy(src_h_local + 2*vector_elements, src_h +
upperReadCursor-vector_elements, sizeof(oid)*2*vector_elements);
@@ -1046,52 +1044,6 @@ static str CRKvectorized_x_@2_@1 (
assert (vectorR == vectorCount);
//assert(lowerReadCursor == upperReadCursor || (lowerReadCursor ==
first_right && upperReadCursor == last_left + 1));
- if(remaining_elements > 0)
- {
- BUN qualifying_elements = 0;
- BUN lowerCursor = last_right - remaining_elements + 1,
upperCursor = last_right;
- oid *tmp_h = GDKmalloc(remaining_elements * sizeof(oid));
- @1 *tmp_t = GDKmalloc(remaining_elements * sizeof(@1 ));
-
- while (src_t[lowerCursor] @7 pivot)
- {
- lowerCursor++;
- qualifying_elements++;
- }
- while (src_t[upperCursor] @8 pivot)
- upperCursor--;
-
- while (lowerCursor < upperCursor) {
- if (src_t[lowerCursor] @8 pivot && src_t[upperCursor]
@7 pivot) {
- @1 tmp = src_t[lowerCursor];
- oid tmp_oid = src_h[lowerCursor];
- src_t[lowerCursor] = src_t[upperCursor];
- src_t[upperCursor] = tmp;
- src_h[lowerCursor] = src_h[upperCursor];
- src_h[upperCursor] = tmp_oid;
- lowerCursor++;
- upperCursor--;
- qualifying_elements++;
- } else if (src_t[lowerCursor] @7 pivot) {
- lowerCursor++;
- qualifying_elements++;
- } else if (src_t[upperCursor] @8 pivot) {
- upperCursor--;
- }
- }
- if (lowerCursor == upperCursor && src_t[lowerCursor] @7 pivot)
- qualifying_elements += 1;
-
- memcpy(tmp_h, src_h + lowerWriteCursor,
sizeof(oid)*qualifying_elements);
- memcpy(src_h + lowerWriteCursor, src_h + (last_right -
remaining_elements + 1), sizeof(oid)*qualifying_elements);
- memcpy(src_h + (last_right - remaining_elements + 1), tmp_h,
sizeof(oid)*qualifying_elements);
- memcpy(tmp_t, src_t + lowerWriteCursor,
sizeof(@1)*qualifying_elements);
- memcpy(src_t + lowerWriteCursor, src_t + (last_right -
remaining_elements + 1), sizeof(@1)*qualifying_elements);
- memcpy(src_t + (last_right - remaining_elements + 1), tmp_t,
sizeof(@1)*qualifying_elements);
- lowerWriteCursor = lowerWriteCursor + qualifying_elements;
- }
-
-
*pos_r = lowerWriteCursor;
return MAL_SUCCEED;
@@ -1114,7 +1066,7 @@ cracking_MT_vectorized_crackThread_@2_@1
return NULL;
}
-static str CRKvectorized_MT_@2_@1 (const BAT *b, const BUN first, const BUN
last, const @1 pivot, oid *pos, int nthreads, int alt, const @1 pivot_P, int
vector_elements)
+static str CRKvectorized_MT_@2_@1 (const BAT *b, BUN first, BUN last, const @1
pivot, oid *pos, int nthreads, int alt, const @1 pivot_P, int vector_elements)
{
BUN n = last - first + 1; /* total # tuples / values */
BUN ml; /* # tuples / values in left slice */
@@ -1126,6 +1078,7 @@ static str CRKvectorized_MT_@2_@1 (const
int i, j;
oid *src_h = (oid*) Hloc(b, BUNfirst(b)), *tmp_h;
@1 *src_t = (@1 *) Tloc(b, BUNfirst(b)), *tmp_t;
+ BUN remaining_elements = 0;
/* adjust nthreads */
if (nthreads == 0) {
@@ -1136,7 +1089,13 @@ static str CRKvectorized_MT_@2_@1 (const
/* more threads / smaller slices does not make sense */
nthreads = (int) (n / 1000) + 1;
}
+
mm = (n / nthreads);
+ remaining_elements = (mm % (2 * vector_elements)) * nthreads;
+ mm = (n - remaining_elements) / nthreads;
+ remaining_elements += ((n - remaining_elements) % nthreads);
+ last -= remaining_elements;
+ mm -= ((n - remaining_elements) % nthreads);
if (alt == 1) {
ml = 0;
mr = 0;
@@ -1314,6 +1273,55 @@ static str CRKvectorized_MT_@2_@1 (const
assert(f != BUN_NONE);
assert((f == last + 1 && src_t[f-1] @7 pivot) || src_t[f] @8 pivot);
assert((f == first && src_t[f] @8 pivot) || src_t[f-1] @7 pivot);
+
+ if(remaining_elements > 0)
+ {
+ BUN qualifying_elements = 0;
+ BUN lowerCursor = last + 1, upperCursor = last +
remaining_elements;
+ oid *tmp_head = GDKmalloc(remaining_elements * sizeof(oid));
+ @1 *tmp_tail = GDKmalloc(remaining_elements * sizeof(@1 ));
+
+ while (src_t[lowerCursor] @7 pivot)
+ {
+ lowerCursor++;
+ qualifying_elements++;
+ }
+ while (src_t[upperCursor] @8 pivot)
+ upperCursor--;
+
+ while (lowerCursor < upperCursor) {
+ if (src_t[lowerCursor] @8 pivot && src_t[upperCursor]
@7 pivot) {
+ @1 tmp = src_t[lowerCursor];
+ oid tmp_oid = src_h[lowerCursor];
+ src_t[lowerCursor] = src_t[upperCursor];
+ src_t[upperCursor] = tmp;
+ src_h[lowerCursor] = src_h[upperCursor];
+ src_h[upperCursor] = tmp_oid;
+ lowerCursor++;
+ upperCursor--;
+ qualifying_elements++;
+ } else if (src_t[lowerCursor] @7 pivot) {
+ lowerCursor++;
+ qualifying_elements++;
+ } else if (src_t[upperCursor] @8 pivot) {
+ upperCursor--;
+ }
+ }
+ if (lowerCursor == upperCursor && src_t[lowerCursor] @7 pivot)
+ qualifying_elements += 1;
+
+ memcpy(tmp_head, src_h + f, sizeof(oid)*qualifying_elements);
+ memcpy(src_h + f, src_h + last + 1,
sizeof(oid)*qualifying_elements);
+ memcpy(src_h + last + 1, tmp_head,
sizeof(oid)*qualifying_elements);
+ memcpy(tmp_tail, src_t + f, sizeof(@1)*qualifying_elements);
+ memcpy(src_t + f, src_t + last + 1,
sizeof(@1)*qualifying_elements);
+ memcpy(src_t + last + 1, tmp_tail,
sizeof(@1)*qualifying_elements);
+
+ f += qualifying_elements;
+ GDKfree(tmp_head);
+ GDKfree(tmp_tail);
+ }
+
*pos = (BUN) (f == 0 ? 0 : f - 1);
GDKfree(tmp_h);
_______________________________________________
checkin-list mailing list
[email protected]
https://www.monetdb.org/mailman/listinfo/checkin-list
