multi-threaded global filter creation

3 files changed, 65 insertions, 3 deletions

diff --git a/searchlib/src/vespa/searchlib/queryeval/global_filter.cpp b/searchlib/src/vespa/searchlib/queryeval/global_filter.cpp
index 2aff91974a7..04e4d154821 100644
--- a/searchlib/src/vespa/searchlib/queryeval/global_filter.cpp
+++ b/searchlib/src/vespa/searchlib/queryeval/global_filter.cpp
@@ -1,7 +1,14 @@
 // Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
 
 #include "global_filter.h"
+#include "blueprint.h"
 #include <vespa/vespalib/util/require.h>
+#include <vespa/vespalib/util/thread_bundle.h>
+#include <vespa/searchlib/common/bitvector.h>
+#include <cassert>
+
+using vespalib::Runnable;
+using vespalib::ThreadBundle;
 
 namespace search::queryeval {
 
@@ -49,6 +56,27 @@ struct MultiBitVectorFilter : public GlobalFilter {
     }
 };
 
+std::unique_ptr<BitVector> make_part(Blueprint &blueprint, uint32_t begin, uint32_t end) {
+    bool strict = true;
+    auto constraint = Blueprint::FilterConstraint::UPPER_BOUND;
+    auto filter_iterator = blueprint.createFilterSearch(strict, constraint);
+    filter_iterator->initRange(begin, end);
+    auto result = filter_iterator->get_hits(begin);
+    // count bits in parallel and cache the results for later
+    result->countTrueBits();
+    return result;
+}
+
+struct MakePart : Runnable {
+    Blueprint &blueprint;
+    uint32_t begin;
+    uint32_t end;
+    std::unique_ptr<BitVector> result;
+    MakePart(Blueprint &blueprint_in, uint32_t begin_in, uint32_t end_in) noexcept
+      : blueprint(blueprint_in), begin(begin_in), end(end_in), result() {}
+    void run() override { result = make_part(blueprint, begin, end); }
+};
+
 }
 
 GlobalFilter::GlobalFilter() = default;
@@ -83,9 +111,10 @@ GlobalFilter::create(std::unique_ptr<BitVector> vector)
 std::shared_ptr<GlobalFilter>
 GlobalFilter::create(std::vector<std::unique_ptr<BitVector>> vectors)
 {
-    uint32_t total_size = 0;
+    uint32_t total_size = 1;
     uint32_t total_count = 0;
     std::vector<uint32_t> splits;
+    splits.reserve(vectors.size());
     for (size_t i = 0; i < vectors.size(); ++i) {
         bool last = ((i + 1) == vectors.size());
         total_count += vectors[i]->countTrueBits();
@@ -100,4 +129,32 @@ GlobalFilter::create(std::vector<std::unique_ptr<BitVector>> vectors)
                                                   total_size, total_count);
 }
 
+std::shared_ptr<GlobalFilter>
+GlobalFilter::create(Blueprint &blueprint, uint32_t docid_limit, ThreadBundle &thread_bundle)
+{
+    uint32_t num_threads = thread_bundle.size();
+    std::vector<MakePart> parts;
+    parts.reserve(num_threads);
+    uint32_t docid = 1;
+    uint32_t per_thread = (docid_limit - docid) / num_threads;
+    uint32_t rest_docs = (docid_limit - docid) % num_threads;
+    while (docid < docid_limit) {
+        uint32_t part_size = per_thread + (parts.size() < rest_docs);
+        parts.emplace_back(blueprint, docid, docid + part_size);
+        docid += part_size;
+    }
+    assert(parts.size() <= num_threads);
+    assert((docid == docid_limit) || parts.empty());
+    thread_bundle.run(parts);
+    if (parts.size() == 1) {
+        return create(std::move(parts[0].result));
+    }
+    std::vector<std::unique_ptr<BitVector>> vectors;
+    vectors.reserve(parts.size());
+    for (MakePart &part: parts) {
+        vectors.push_back(std::move(part.result));
+    }
+    return create(std::move(vectors));
+}
+
 }
diff --git a/searchlib/src/vespa/searchlib/queryeval/global_filter.h b/searchlib/src/vespa/searchlib/queryeval/global_filter.h
index e93864db2c8..e162bccdbc5 100644
--- a/searchlib/src/vespa/searchlib/queryeval/global_filter.h
+++ b/searchlib/src/vespa/searchlib/queryeval/global_filter.h
@@ -3,11 +3,15 @@
 #pragma once
 
 #include <memory>
-#include <vespa/searchlib/common/bitvector.h>
 #include <vector>
 
+namespace vespalib { class ThreadBundle; }
+namespace search { class BitVector; }
+
 namespace search::queryeval {
 
+class Blueprint;
+
 /**
  * Hold ownership of a global filter that can be taken into account by
  * adaptive query operators. The owned 'bitvector' should be a
@@ -35,6 +39,7 @@ public:
     static std::shared_ptr<GlobalFilter> create(std::vector<uint32_t> docids, uint32_t size);
     static std::shared_ptr<GlobalFilter> create(std::unique_ptr<BitVector> vector);
     static std::shared_ptr<GlobalFilter> create(std::vector<std::unique_ptr<BitVector>> vectors);
+    static std::shared_ptr<GlobalFilter> create(Blueprint &blueprint, uint32_t docid_limit, vespalib::ThreadBundle &thread_bundle);
 };
 
 } // namespace
diff --git a/searchlib/src/vespa/searchlib/queryeval/simpleresult.h b/searchlib/src/vespa/searchlib/queryeval/simpleresult.h
index a7bb30ec487..4a8b7a3429f 100644
--- a/searchlib/src/vespa/searchlib/queryeval/simpleresult.h
+++ b/searchlib/src/vespa/searchlib/queryeval/simpleresult.h
@@ -20,7 +20,7 @@ public:
     /**
      * Create an empty result
      **/
-    SimpleResult() : _hits() {}
+    SimpleResult() noexcept : _hits() {}
 
     /**
      * Create a result with the given hits.