Use vespalib::hash_set instead of std::set to reduce number of allocation and epeed it up. Also use faster 2^N AND based hash tables.

5 files changed, 74 insertions, 48 deletions

diff --git a/eval/src/vespa/eval/tensor/sparse/CMakeLists.txt b/eval/src/vespa/eval/tensor/sparse/CMakeLists.txt
index 8e1d18a87b7..a25d2abb477 100644
--- a/eval/src/vespa/eval/tensor/sparse/CMakeLists.txt
+++ b/eval/src/vespa/eval/tensor/sparse/CMakeLists.txt
@@ -1,6 +1,7 @@
 # Copyright 2017 Yahoo Holdings. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
 vespa_add_library(eval_tensor_sparse OBJECT
     SOURCES
+    direct_sparse_tensor_builder.cpp
     sparse_tensor.cpp
     sparse_tensor_add.cpp
     sparse_tensor_address_builder.cpp
diff --git a/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.cpp b/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.cpp
new file mode 100644
index 00000000000..4a28b54d201
--- /dev/null
+++ b/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.cpp
@@ -0,0 +1,62 @@
+// Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+#include "direct_sparse_tensor_builder.h"
+
+namespace vespalib::tensor {
+
+void
+DirectSparseTensorBuilder::copyCells(const Cells &cells_in)
+{
+    for (const auto &cell : cells_in) {
+        SparseTensorAddressRef oldRef = cell.first;
+        SparseTensorAddressRef newRef(oldRef, _stash);
+        _cells[newRef] = cell.second;
+    }
+}
+
+DirectSparseTensorBuilder::DirectSparseTensorBuilder()
+    : _stash(SparseTensor::STASH_CHUNK_SIZE),
+      _type(eval::ValueType::double_type()),
+      _cells()
+{
+}
+
+DirectSparseTensorBuilder::DirectSparseTensorBuilder(const eval::ValueType &type_in)
+    : _stash(SparseTensor::STASH_CHUNK_SIZE),
+      _type(type_in),
+      _cells()
+{
+}
+
+DirectSparseTensorBuilder::DirectSparseTensorBuilder(const eval::ValueType &type_in, const Cells &cells_in)
+    : _stash(SparseTensor::STASH_CHUNK_SIZE),
+      _type(type_in),
+      _cells()
+{
+    copyCells(cells_in);
+}
+
+DirectSparseTensorBuilder::~DirectSparseTensorBuilder() = default;
+
+Tensor::UP
+DirectSparseTensorBuilder::build() {
+    return std::make_unique<SparseTensor>(std::move(_type), std::move(_cells), std::move(_stash));
+}
+
+void
+DirectSparseTensorBuilder::insertCell(SparseTensorAddressRef address, double value) {
+    // This address should not already exist and a new cell should be inserted.
+    insertCell(address, value, [](double, double) -> double { HDR_ABORT("should not be reached"); });
+}
+
+void
+DirectSparseTensorBuilder::insertCell(SparseTensorAddressBuilder &address, double value) {
+    // This address should not already exist and a new cell should be inserted.
+    insertCell(address.getAddressRef(), value, [](double, double) -> double { HDR_ABORT("should not be reached"); });
+}
+
+void DirectSparseTensorBuilder::reserve(uint32_t estimatedCells) {
+    _cells.resize(estimatedCells*2);
+}
+
+}
\ No newline at end of file
diff --git a/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.h b/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.h
index d54c8810a81..f9182a199be 100644
--- a/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.h
+++ b/eval/src/vespa/eval/tensor/sparse/direct_sparse_tensor_builder.h
@@ -25,43 +25,13 @@ private:
     Cells _cells;
 
 public:
-    void
-    copyCells(const Cells &cells_in)
-    {
-        for (const auto &cell : cells_in) {
-            SparseTensorAddressRef oldRef = cell.first;
-            SparseTensorAddressRef newRef(oldRef, _stash);
-            _cells[newRef] = cell.second;
-        }
-    }
-
-    DirectSparseTensorBuilder()
-        : _stash(SparseTensor::STASH_CHUNK_SIZE),
-          _type(eval::ValueType::double_type()),
-          _cells()
-    {
-    }
-
-    DirectSparseTensorBuilder(const eval::ValueType &type_in)
-        : _stash(SparseTensor::STASH_CHUNK_SIZE),
-          _type(type_in),
-          _cells()
-    {
-    }
-
-    DirectSparseTensorBuilder(const eval::ValueType &type_in, const Cells &cells_in)
-        : _stash(SparseTensor::STASH_CHUNK_SIZE),
-          _type(type_in),
-          _cells()
-    {
-        copyCells(cells_in);
-    }
+    void copyCells(const Cells &cells_in);
+    DirectSparseTensorBuilder();
+    DirectSparseTensorBuilder(const eval::ValueType &type_in);
+    DirectSparseTensorBuilder(const eval::ValueType &type_in, const Cells &cells_in);
+    ~DirectSparseTensorBuilder();
 
-    ~DirectSparseTensorBuilder() {};
-
-    Tensor::UP build() {
-        return std::make_unique<SparseTensor>(std::move(_type), std::move(_cells), std::move(_stash));
-    }
+    Tensor::UP build();
 
     template <class Function>
     void insertCell(SparseTensorAddressRef address, double value, Function &&func)
@@ -75,10 +45,7 @@ public:
         }
     }
 
-    void insertCell(SparseTensorAddressRef address, double value) {
-        // This address should not already exist and a new cell should be inserted.
-        insertCell(address, value, [](double, double) -> double { HDR_ABORT("should not be reached"); });
-    }
+    void insertCell(SparseTensorAddressRef address, double value);
 
     template <class Function>
     void insertCell(SparseTensorAddressBuilder &address, double value, Function &&func)
@@ -86,14 +53,11 @@ public:
         insertCell(address.getAddressRef(), value, func);
     }
 
-    void insertCell(SparseTensorAddressBuilder &address, double value) {
-        // This address should not already exist and a new cell should be inserted.
-        insertCell(address.getAddressRef(), value, [](double, double) -> double { HDR_ABORT("should not be reached"); });
-    }
+    void insertCell(SparseTensorAddressBuilder &address, double value);
 
     eval::ValueType &fast_type() { return _type; }
     Cells &cells() { return _cells; }
-    void reserve(uint32_t estimatedCells) { _cells.resize(estimatedCells*2); }
+    void reserve(uint32_t estimatedCells);
 };
 
 }
diff --git a/eval/src/vespa/eval/tensor/sparse/sparse_tensor.cpp b/eval/src/vespa/eval/tensor/sparse/sparse_tensor.cpp
index 1fc93e8234f..d183c33f5cd 100644
--- a/eval/src/vespa/eval/tensor/sparse/sparse_tensor.cpp
+++ b/eval/src/vespa/eval/tensor/sparse/sparse_tensor.cpp
@@ -245,5 +245,4 @@ SparseTensor::remove(const CellValues &cellAddresses) const
 
 }
 
-VESPALIB_HASH_MAP_INSTANTIATE_H_E_M(vespalib::tensor::SparseTensorAddressRef, double, vespalib::hash<vespalib::tensor::SparseTensorAddressRef>,
-        std::equal_to<vespalib::tensor::SparseTensorAddressRef>, vespalib::hashtable_base::and_modulator);
+VESPALIB_HASH_MAP_INSTANTIATE(vespalib::tensor::SparseTensorAddressRef, double);
diff --git a/eval/src/vespa/eval/tensor/sparse/sparse_tensor.h b/eval/src/vespa/eval/tensor/sparse/sparse_tensor.h
index 880cd32c605..e5ea639b460 100644
--- a/eval/src/vespa/eval/tensor/sparse/sparse_tensor.h
+++ b/eval/src/vespa/eval/tensor/sparse/sparse_tensor.h
@@ -22,7 +22,7 @@ class SparseTensor : public Tensor
 {
 public:
     using Cells = hash_map<SparseTensorAddressRef, double, hash<SparseTensorAddressRef>,
-            std::equal_to<SparseTensorAddressRef>, hashtable_base::and_modulator>;
+                           std::equal_to<>, hashtable_base::and_modulator>;
 
     static constexpr size_t STASH_CHUNK_SIZE = 16384u;