Move the right address context on the outside.

3 files changed, 68 insertions, 75 deletions

diff --git a/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.cpp b/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.cpp
index 8c704c61415..d6d0a1abf4a 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.cpp
+++ b/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.cpp
@@ -9,9 +9,8 @@ namespace vespalib::tensor {
 DenseTensorAddressCombiner::~DenseTensorAddressCombiner() = default;
 
 DenseTensorAddressCombiner::DenseTensorAddressCombiner(const eval::ValueType &combined, const eval::ValueType &lhs,
-                                                       const eval::ValueType &rhs, CellsRef rhsCells)
-    : _rightAddress(rhs, rhsCells),
-      _combinedAddress(combined),
+                                                       const eval::ValueType &rhs)
+    : _combinedAddress(combined),
       _left(),
       _commonRight(),
       _right()
@@ -39,7 +38,7 @@ DenseTensorAddressCombiner::DenseTensorAddressCombiner(const eval::ValueType &co
     }
 }
 
-DenseTensorAddressCombiner::AddressContext::AddressContext(const eval::ValueType &type)
+AddressContext::AddressContext(const eval::ValueType &type)
     : _type(type),
       _accumulatedSize(_type.dimensions().size()),
       _address(type.dimensions().size(), 0)
@@ -52,7 +51,7 @@ DenseTensorAddressCombiner::AddressContext::AddressContext(const eval::ValueType
     }
 }
 
-DenseTensorAddressCombiner::AddressContext::~AddressContext() = default;
+AddressContext::~AddressContext() = default;
 
 eval::ValueType
 DenseTensorAddressCombiner::combineDimensions(const eval::ValueType &lhs, const eval::ValueType &rhs)
diff --git a/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.h b/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.h
index 0d4fa794f38..eea523045e9 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.h
+++ b/eval/src/vespa/eval/tensor/dense/dense_tensor_address_combiner.h
@@ -9,6 +9,42 @@
 
 namespace vespalib::tensor {
 
+class AddressContext {
+public:
+    using Address = DenseTensorCellsIterator::Address;
+    using size_type = eval::ValueType::Dimension::size_type;
+    using Mapping = std::vector<std::pair<uint32_t, uint32_t>>;
+    AddressContext(const eval::ValueType &type);
+    ~AddressContext();
+    size_type dimSize(uint32_t dim) const { return _type.dimensions()[dim].size; }
+    size_type wholeDimStep(uint32_t dim) const { return _accumulatedSize[dim] * dimSize(dim); }
+    size_t index() const {
+        size_t cellIdx(0);
+        for (uint32_t i(0); i < _address.size(); i++) {
+            cellIdx += _address[i]*_accumulatedSize[i];
+        }
+        return cellIdx;
+    }
+    void update(const Address & addr, const Mapping & mapping) {
+        for (const auto & m : mapping) {
+            _address[m.first] = addr[m.second];
+        }
+    }
+    bool updateCommon(const Address & addr, const Mapping & mapping) {
+        for (const auto & m : mapping) {
+            if (addr[m.first] >= dimSize(m.second)) {
+                return false;
+            }
+            _address[m.second] = addr[m.first];
+        }
+        return true;
+    }
+
+    const eval::ValueType &_type;
+    std::vector<size_t>    _accumulatedSize;
+    Address                _address;
+
+};
 
 /**
  * Combines two dense tensor addresses to a new tensor address.
@@ -18,48 +54,13 @@ namespace vespalib::tensor {
 class DenseTensorAddressCombiner
 {
 public:
-    using Mapping = std::vector<std::pair<uint32_t, uint32_t>>;
+    using Mapping = AddressContext::Mapping;
 
 private:
     using Address = DenseTensorCellsIterator::Address;
     using CellsRef = vespalib::ConstArrayRef<double>;
     using size_type = eval::ValueType::Dimension::size_type;
 
-    class AddressContext {
-    public:
-        AddressContext(const eval::ValueType &type);
-        ~AddressContext();
-        size_type dimSize(uint32_t dim) const { return _type.dimensions()[dim].size; }
-        size_type wholeDimStep(uint32_t dim) const { return _accumulatedSize[dim] * dimSize(dim); }
-        size_t index() const {
-            size_t cellIdx(0);
-            for (uint32_t i(0); i < _address.size(); i++) {
-                cellIdx += _address[i]*_accumulatedSize[i];
-            }
-            return cellIdx;
-        }
-        void update(const Address & addr, const Mapping & mapping) {
-            for (const auto & m : mapping) {
-                _address[m.first] = addr[m.second];
-            }
-        }
-
-        const eval::ValueType &_type;
-        std::vector<size_t>    _accumulatedSize;
-        Address                _address;
-
-    };
-    class CellAddressContext : public AddressContext {
-    public:
-        CellAddressContext(const eval::ValueType &type, CellsRef cells) : AddressContext(type), _cells(cells) { }
-        double cell() const { return cell(index()); }
-        double cell(size_t cellIdx) const { return _cells[cellIdx]; }
-    private:
-        CellsRef               _cells;
-    };
-
-
-    CellAddressContext     _rightAddress;
     AddressContext         _combinedAddress;
 
     Mapping                _left;
@@ -67,54 +68,43 @@ private:
     Mapping                _right;
 
 public:
-    DenseTensorAddressCombiner(const eval::ValueType &combined, const eval::ValueType &lhs,
-                               const eval::ValueType &rhs, CellsRef rhsCells);
+    DenseTensorAddressCombiner(const eval::ValueType &combined, const eval::ValueType &lhs, const eval::ValueType &rhs);
     ~DenseTensorAddressCombiner();
     void updateLeftAndCommon(const Address & addr) { _combinedAddress.update(addr, _left); }
 
     bool hasAnyRightOnlyDimensions() const { return ! _right.empty(); }
 
-    const Address &address() const { return _combinedAddress._address; }
-
-    bool updateCommonRight() {
-        for (const auto & m : _commonRight) {
-            if (_combinedAddress._address[m.first] >= _rightAddress.dimSize(m.second)) {
-                return false;
-            }
-            _rightAddress._address[m.second] = _combinedAddress._address[m.first];
-        }
-        return true;
-    }
-    double rightCell() { return _rightAddress.cell(); }
+    const Address & address() const { return _combinedAddress._address; }
+    const Mapping & commonRight() const { return _commonRight; }
 
     template <typename Func>
-    void for_each(Func && func) {
+    void for_each(const AddressContext & rightAddress, const CellsRef & rhsCells, Func && func) {
         const int32_t lastDimension = _right.size() - 1;
         int32_t curDimension = lastDimension;
-        size_t rightCellIdx = _rightAddress.index();
+        size_t rightCellIdx = rightAddress.index();
         size_t combinedCellIdx = _combinedAddress.index();
         while (curDimension >= 0) {
             const uint32_t rdim = _right[curDimension].second;
             const uint32_t cdim = _right[curDimension].first;
             size_type & cindex = _combinedAddress._address[cdim];
             if (curDimension == lastDimension) {
-                for (cindex = 0; cindex < _rightAddress.dimSize(rdim); cindex++) {
-                    func(combinedCellIdx, _rightAddress.cell(rightCellIdx));
-                    rightCellIdx += _rightAddress._accumulatedSize[rdim];
+                for (cindex = 0; cindex < rightAddress.dimSize(rdim); cindex++) {
+                    func(combinedCellIdx, rhsCells[rightCellIdx]);
+                    rightCellIdx += rightAddress._accumulatedSize[rdim];
                     combinedCellIdx += _combinedAddress._accumulatedSize[cdim];
                 }
                 cindex = 0;
-                rightCellIdx -= _rightAddress.wholeDimStep(rdim);
+                rightCellIdx -= rightAddress.wholeDimStep(rdim);
                 combinedCellIdx -= _combinedAddress.wholeDimStep(cdim);
                 curDimension--;
             } else {
-                if (cindex < _rightAddress.dimSize(rdim)) {
+                if (cindex < rightAddress.dimSize(rdim)) {
                     cindex++;
-                    rightCellIdx += _rightAddress._accumulatedSize[rdim];
+                    rightCellIdx += rightAddress._accumulatedSize[rdim];
                     combinedCellIdx += _combinedAddress._accumulatedSize[cdim];
                     curDimension++;
                 } else {
-                    rightCellIdx -= _rightAddress.wholeDimStep(rdim);
+                    rightCellIdx -= rightAddress.wholeDimStep(rdim);
                     combinedCellIdx -= _combinedAddress.wholeDimStep(cdim);
                     cindex = 0;
                     curDimension--;
diff --git a/eval/src/vespa/eval/tensor/dense/dense_tensor_apply.hpp b/eval/src/vespa/eval/tensor/dense/dense_tensor_apply.hpp
index 868a49627c8..f61dbd05def 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_tensor_apply.hpp
+++ b/eval/src/vespa/eval/tensor/dense/dense_tensor_apply.hpp
@@ -11,17 +11,18 @@ namespace vespalib::tensor::dense {
 template <typename Function>
 std::unique_ptr<Tensor>
 apply(DenseTensorAddressCombiner & combiner, DirectDenseTensorBuilder & builder,
-      const DenseTensorView &lhs, Function &&func) __attribute__((noinline));
+      const DenseTensorView &lhs, AddressContext & rhsAddr, const DenseTensorView::CellsRef & rhsCells,
+      Function &&func) __attribute__((noinline));
 
 template <typename Function>
 std::unique_ptr<Tensor>
 apply(DenseTensorAddressCombiner & combiner, DirectDenseTensorBuilder & builder,
-      const DenseTensorView &lhs, Function &&func)
+      const DenseTensorView &lhs, AddressContext & rhsAddr, const DenseTensorView::CellsRef & rhsCells, Function &&func)
 {
     for (DenseTensorCellsIterator lhsItr = lhs.cellsIterator(); lhsItr.valid(); lhsItr.next()) {
         combiner.updateLeftAndCommon(lhsItr.address());
-        if (combiner.updateCommonRight()) {
-            combiner.for_each([&func, &builder, &lhsItr](size_t combined, double rhsCell) {
+        if (rhsAddr.updateCommon(combiner.address(), combiner.commonRight())) {
+            combiner.for_each(rhsAddr, rhsCells, [&func, &builder, &lhsItr](size_t combined, double rhsCell) {
                 builder.insertCell(combined, func(lhsItr.cell(), rhsCell));
             });
         }
@@ -33,17 +34,19 @@ apply(DenseTensorAddressCombiner & combiner, DirectDenseTensorBuilder & builder,
 template <typename Function>
 std::unique_ptr<Tensor>
 apply_no_rightonly_dimensions(DenseTensorAddressCombiner & combiner, DirectDenseTensorBuilder & builder,
-                              const DenseTensorView &lhs, Function &&func)  __attribute__((noinline));
+                              const DenseTensorView &lhs, AddressContext & rhsAddr,
+                              const DenseTensorView::CellsRef & rhsCells, Function &&func)  __attribute__((noinline));
 
 template <typename Function>
 std::unique_ptr<Tensor>
 apply_no_rightonly_dimensions(DenseTensorAddressCombiner & combiner, DirectDenseTensorBuilder & builder,
-                              const DenseTensorView &lhs, Function &&func)
+                              const DenseTensorView &lhs, AddressContext & rhsAddr,
+                              const DenseTensorView::CellsRef & rhsCells, Function &&func)
 {
     for (DenseTensorCellsIterator lhsItr = lhs.cellsIterator(); lhsItr.valid(); lhsItr.next()) {
         combiner.updateLeftAndCommon(lhsItr.address());
-        if (combiner.updateCommonRight()) {
-            builder.insertCell(combiner.address(), func(lhsItr.cell(), combiner.rightCell()));
+        if (rhsAddr.updateCommon(combiner.address(), combiner.commonRight())) {
+            builder.insertCell(combiner.address(), func(lhsItr.cell(), rhsCells[rhsAddr.index()]));
         }
     }
     return builder.build();
@@ -54,12 +57,13 @@ std::unique_ptr<Tensor>
 apply(const DenseTensorView &lhs, const DenseTensorView &rhs, Function &&func)
 {
     eval::ValueType resultType = DenseTensorAddressCombiner::combineDimensions(lhs.fast_type(), rhs.fast_type());
-    DenseTensorAddressCombiner combiner(resultType, lhs.fast_type(), rhs.fast_type(), rhs.cellsRef());
+    DenseTensorAddressCombiner combiner(resultType, lhs.fast_type(), rhs.fast_type());
     DirectDenseTensorBuilder builder(resultType);
+    AddressContext rhsAddress(rhs.fast_type());
     if (combiner.hasAnyRightOnlyDimensions()) {
-        return apply(combiner, builder, lhs, std::move(func));
+        return apply(combiner, builder, lhs, rhsAddress, rhs.cellsRef(), std::move(func));
     } else {
-        return apply_no_rightonly_dimensions(combiner, builder, lhs, std::move(func));
+        return apply_no_rightonly_dimensions(combiner, builder, lhs, rhsAddress, rhs.cellsRef(), std::move(func));
     }
 }