Improve large dense aggregate reads with tile metadata only. (#4657)

The dense reader was creating a bitmap to compute all aggregate results
at the top level of the read. For large aggregate reads where we don't
need to load any tiles, this could be quite large. It also would turn
out to be completely unnecessary. This fix moves the bitmap to the lower
level of the read, where a smaller bitmap can be created only if
necessary.

---
TYPE: IMPROVEMENT
DESC: Improve large dense aggregate reads with tile metadata only.

(cherry picked from commit 598e84e)

tiledb/sm/query/readers/dense_reader.cc

@@ -1319,13 +1319,6 @@ Status DenseReader::process_aggregates(
   const auto& tile_coords = subarray.tile_coords();
   const auto global_order = layout_ == Layout::GLOBAL_ORDER;
 
-  std::vector<uint8_t> aggregate_bitmap;
-  if (condition_.has_value()) {
-    aggregate_bitmap = qc_result;
-  } else {
-    aggregate_bitmap.resize(subarray.cell_num(), 1);
-  }
-
   // Process values in parallel.
   auto status = parallel_for_2d(
       storage_manager_->compute_tp(),
@@ -1358,7 +1351,7 @@ Status DenseReader::process_aggregates(
               tile_subarrays[t],
               global_order ? tile_offsets[t] : 0,
               range_info,
-              aggregate_bitmap,
+              qc_result,
               range_thread_idx,
               num_range_threads));
         }
@@ -1863,7 +1856,7 @@ Status DenseReader::aggregate_tiles(
     const Subarray& tile_subarray,
     const uint64_t global_cell_offset,
     const std::vector<RangeInfo<DimType>>& range_info,
-    std::vector<uint8_t>& aggregate_bitmap,
+    const std::vector<uint8_t>& qc_result,
     const uint64_t range_thread_idx,
     const uint64_t num_range_threads) {
   // For easy reference
@@ -1909,6 +1902,14 @@ Status DenseReader::aggregate_tiles(
       cell_offset = iter.dest_offset_row_col();
     }
 
+    std::vector<uint8_t> aggregate_bitmap(iter.cell_slab_length(), 1);
+    if (condition_.has_value()) {
+      memcpy(
+          aggregate_bitmap.data(),
+          qc_result.data() + cell_offset,
+          iter.cell_slab_length());
+    }
+
     // Iterate through all fragment domains and copy data.
     for (uint64_t fd = 0; fd < frag_domains.size(); fd++) {
       // If the cell slab overlaps this fragment domain range, copy data.
@@ -1936,7 +1937,7 @@ Status DenseReader::aggregate_tiles(
               iter.pos_in_tile() + start,
               iter.pos_in_tile() + end + 1,
               tile_tuples[fd],
-              &aggregate_bitmap[cell_offset + start])};
+              aggregate_bitmap.data() + start)};
           for (auto& aggregate : aggregates) {
             aggregate->aggregate_data(aggregate_buffer);
           }
@@ -1952,7 +1953,7 @@ Status DenseReader::aggregate_tiles(
                 start_cell,
                 start_cell + 1,
                 tile_tuples[fd],
-                &aggregate_bitmap[cell_offset + start + i])};
+                aggregate_bitmap.data() + start + i)};
             for (auto& aggregate : aggregates) {
               aggregate->aggregate_data(aggregate_buffer);
             }
@@ -1964,7 +1965,7 @@ Status DenseReader::aggregate_tiles(
         // fragments.
         if (fd != frag_domains.size() - 1) {
           for (uint64_t c = start; c <= end; c++) {
-            aggregate_bitmap[cell_offset + c] = 0;
+            aggregate_bitmap[c] = 0;
           }
         }
 

tiledb/sm/query/readers/dense_reader.h

@@ -400,7 +400,7 @@ class DenseReader : public ReaderBase, public IQueryStrategy {
       const Subarray& tile_subarray,
       const uint64_t global_cell_offset,
       const std::vector<RangeInfo<DimType>>& range_info,
-      std::vector<uint8_t>& aggregate_bitmap,
+      const std::vector<uint8_t>& qc_result,
       const uint64_t range_thread_idx,
       const uint64_t num_range_threads);