diff options
16 files changed, 285 insertions, 68 deletions
diff --git a/eval/src/tests/ann/nns-l2.h b/eval/src/tests/ann/nns-l2.h index 82a95741200..de24df50b6c 100644 --- a/eval/src/tests/ann/nns-l2.h +++ b/eval/src/tests/ann/nns-l2.h @@ -36,7 +36,7 @@ template <typename FltType = float> struct L2DistCalc { const vespalib::hwaccelrated::IAccelrated & _hw; - L2DistCalc() : _hw(vespalib::hwaccelrated::IAccelrated::getAccelrator()) {} + L2DistCalc() : _hw(vespalib::hwaccelrated::IAccelrated::getAccelerator()) {} using Arr = vespalib::ArrayRef<FltType>; using ConstArr = vespalib::ConstArrayRef<FltType>; diff --git a/searchlib/src/vespa/searchlib/common/bitvector.cpp b/searchlib/src/vespa/searchlib/common/bitvector.cpp index 96234e373dc..0a33e23de72 100644 --- a/searchlib/src/vespa/searchlib/common/bitvector.cpp +++ b/searchlib/src/vespa/searchlib/common/bitvector.cpp @@ -167,7 +167,7 @@ BitVector::countInterval(Range range_in) const ++endw; } if (startw < endw) { - res += IAccelrated::getAccelrator().populationCount(bitValues + startw, endw - startw); + res += IAccelrated::getAccelerator().populationCount(bitValues + startw, endw - startw); } if (partialEnd) { res += Optimized::popCount(bitValues[endw] & ~endBits(last)); @@ -185,13 +185,13 @@ BitVector::orWith(const BitVector & right) if (right.size() > 0) { ssize_t commonBytes = numActiveBytes(getStartIndex(), right.size()) - sizeof(Word); if (commonBytes > 0) { - IAccelrated::getAccelrator().orBit(getActiveStart(), right.getWordIndex(getStartIndex()), commonBytes); + IAccelrated::getAccelerator().orBit(getActiveStart(), right.getWordIndex(getStartIndex()), commonBytes); } Index last(right.size() - 1); getWordIndex(last)[0] |= (right.getWordIndex(last)[0] & ~endBits(last)); } } else { - IAccelrated::getAccelrator().orBit(getActiveStart(), right.getWordIndex(getStartIndex()), getActiveBytes()); + IAccelrated::getAccelerator().orBit(getActiveStart(), right.getWordIndex(getStartIndex()), getActiveBytes()); } repairEnds(); invalidateCachedCount(); @@ -216,7 +216,7 @@ BitVector::andWith(const BitVector & right) verifyInclusiveStart(*this, right); uint32_t commonBytes = std::min(getActiveBytes(), numActiveBytes(getStartIndex(), right.size())); - IAccelrated::getAccelrator().andBit(getActiveStart(), right.getWordIndex(getStartIndex()), commonBytes); + IAccelrated::getAccelerator().andBit(getActiveStart(), right.getWordIndex(getStartIndex()), commonBytes); if (right.size() < size()) { clearInterval(right.size(), size()); } @@ -235,13 +235,13 @@ BitVector::andNotWith(const BitVector& right) if (right.size() > 0) { ssize_t commonBytes = numActiveBytes(getStartIndex(), right.size()) - sizeof(Word); if (commonBytes > 0) { - IAccelrated::getAccelrator().andNotBit(getActiveStart(), right.getWordIndex(getStartIndex()), commonBytes); + IAccelrated::getAccelerator().andNotBit(getActiveStart(), right.getWordIndex(getStartIndex()), commonBytes); } Index last(right.size() - 1); getWordIndex(last)[0] &= ~(right.getWordIndex(last)[0] & ~endBits(last)); } } else { - IAccelrated::getAccelrator().andNotBit(getActiveStart(), right.getWordIndex(getStartIndex()), getActiveBytes()); + IAccelrated::getAccelerator().andNotBit(getActiveStart(), right.getWordIndex(getStartIndex()), getActiveBytes()); } repairEnds(); @@ -250,7 +250,7 @@ BitVector::andNotWith(const BitVector& right) void BitVector::notSelf() { - IAccelrated::getAccelrator().notBit(getActiveStart(), getActiveBytes()); + IAccelrated::getAccelerator().notBit(getActiveStart(), getActiveBytes()); setGuardBit(); invalidateCachedCount(); } diff --git a/searchlib/src/vespa/searchlib/features/dotproductfeature.cpp b/searchlib/src/vespa/searchlib/features/dotproductfeature.cpp index a8737a19eec..37fd98c9f20 100644 --- a/searchlib/src/vespa/searchlib/features/dotproductfeature.cpp +++ b/searchlib/src/vespa/searchlib/features/dotproductfeature.cpp @@ -256,7 +256,7 @@ namespace dotproduct::array { template <typename BaseType> DotProductExecutorBase<BaseType>::DotProductExecutorBase(const V & queryVector) : FeatureExecutor(), - _multiplier(IAccelrated::getAccelrator()), + _multiplier(IAccelrated::getAccelerator()), _queryVector(queryVector) { } diff --git a/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.cpp b/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.cpp index 105d57b22b1..d36d16a679a 100644 --- a/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.cpp +++ b/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.cpp @@ -1,19 +1,19 @@ // Copyright 2017 Yahoo Holdings. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root. -#include <vespa/searchlib/queryeval/multibitvectoriterator.h> -#include <vespa/searchlib/queryeval/andsearch.h> -#include <vespa/searchlib/queryeval/andnotsearch.h> -#include <vespa/searchlib/queryeval/sourceblendersearch.h> -#include <vespa/searchlib/queryeval/orsearch.h> +#include "multibitvectoriterator.h" +#include "andsearch.h" +#include "andnotsearch.h" +#include "sourceblendersearch.h" #include <vespa/searchlib/common/bitvectoriterator.h> -#include <vespa/searchlib/attribute/attributeiterators.h> #include <vespa/searchlib/fef/termfieldmatchdata.h> #include <vespa/searchlib/fef/termfieldmatchdataarray.h> #include <vespa/vespalib/util/optimized.h> +#include <vespa/vespalib/hwaccelrated/iaccelrated.h> namespace search::queryeval { using vespalib::Trinary; +using vespalib::hwaccelrated::IAccelrated; namespace { @@ -21,7 +21,15 @@ template<typename Update> class MultiBitVectorIterator : public MultiBitVectorIteratorBase { public: - MultiBitVectorIterator(Children children) : MultiBitVectorIteratorBase(std::move(children)) { } + explicit MultiBitVectorIterator(Children children) + : MultiBitVectorIteratorBase(std::move(children)), + _update(), + _accel(IAccelrated::getAccelerator()), + _lastWords() + { + static_assert(sizeof(_lastWords) == 64, "Latswords should have 64 byte size"); + memset(&_lastWords, 0, sizeof(_lastWords)); + } protected: void updateLastValue(uint32_t docId); void strictSeek(uint32_t docId); @@ -29,33 +37,55 @@ private: void doSeek(uint32_t docId) override; Trinary is_strict() const override { return Trinary::False; } bool acceptExtraFilter() const override { return Update::isAnd(); } - Update _update; + Update _update; + const IAccelrated & _accel; + alignas(64) Word _lastWords[8]; }; template<typename Update> class MultiBitVectorIteratorStrict : public MultiBitVectorIterator<Update> { public: - MultiBitVectorIteratorStrict(MultiSearch::Children children) : MultiBitVectorIterator<Update>(std::move(children)) { } + explicit MultiBitVectorIteratorStrict(MultiSearch::Children children) + : MultiBitVectorIterator<Update>(std::move(children)) + { } private: void doSeek(uint32_t docId) override { this->strictSeek(docId); } Trinary is_strict() const override { return Trinary::True; } }; +struct And { + using Word = BitWord::Word; + void operator () (const IAccelrated & accel, size_t offset, const std::vector<std::pair<const void *, bool>> & src, Word *dest) { + accel.and64(offset*sizeof(uint64_t), src, dest); + } + static bool isAnd() { return true; } +}; + +struct Or { + using Word = BitWord::Word; + void operator () (const IAccelrated & accel, size_t offset, const std::vector<std::pair<const void *, bool>> & src, Word *dest) { + accel.or64(offset*sizeof(uint64_t), src, dest); + } + static bool isAnd() { return false; } +}; + template<typename Update> void MultiBitVectorIterator<Update>::updateLastValue(uint32_t docId) { if (docId >= _lastMaxDocIdLimit) { - if (__builtin_expect(docId < _numDocs, true)) { - const uint32_t index(wordNum(docId)); - _lastValue = _bvs[0][index]; - for(uint32_t i(1); i < _bvs.size(); i++) { - _lastValue = _update(_lastValue, _bvs[i][index]); - } - _lastMaxDocIdLimit = (index + 1) * WordLen; - } else { + if (__builtin_expect(docId >= _numDocs, false)) { setAtEnd(); + return; + } + const uint32_t index(wordNum(docId)); + if (docId >= _lastMaxDocIdLimitRequireFetch) { + uint32_t baseIndex = index & ~(sizeof(_lastWords)/sizeof(Word) - 1); + _update(_accel, baseIndex, _bvs, _lastWords); + _lastMaxDocIdLimitRequireFetch = (baseIndex + (sizeof(_lastWords)/sizeof(Word))) * WordLen; } + _lastValue = _lastWords[index % (sizeof(_lastWords)/sizeof(Word))]; + _lastMaxDocIdLimit = (index + 1) * WordLen; } } @@ -75,7 +105,7 @@ template<typename Update> void MultiBitVectorIterator<Update>::strictSeek(uint32_t docId) { - for (updateLastValue(docId), _lastValue=_lastValue & checkTab(docId); + for (updateLastValue(docId), _lastValue = _lastValue & checkTab(docId); (_lastValue == 0) && __builtin_expect(! isAtEnd(), true); updateLastValue(_lastMaxDocIdLimit)); if (__builtin_expect(!isAtEnd(), true)) { @@ -88,21 +118,6 @@ MultiBitVectorIterator<Update>::strictSeek(uint32_t docId) } } -struct And { - typedef BitWord::Word Word; - Word operator () (const Word a, const Word b) { - return a & b; - } - static bool isAnd() { return true; } -}; - -struct Or { - typedef BitWord::Word Word; - Word operator () (const Word a, const Word b) { - return a | b; - } - static bool isAnd() { return false; } -}; typedef MultiBitVectorIterator<And> AndBVIterator; typedef MultiBitVectorIteratorStrict<And> AndBVIteratorStrict; @@ -136,14 +151,15 @@ bool canOptimize(const MultiSearch & s) { MultiBitVectorIteratorBase::MultiBitVectorIteratorBase(Children children) : MultiSearch(std::move(children)), _numDocs(std::numeric_limits<unsigned int>::max()), - _lastValue(0), _lastMaxDocIdLimit(0), + _lastMaxDocIdLimitRequireFetch(0), + _lastValue(0), _bvs() { _bvs.reserve(getChildren().size()); - for (size_t i(0); i < getChildren().size(); i++) { - const auto * bv = static_cast<const BitVectorIterator *>(getChildren()[i].get()); - _bvs.emplace_back(reinterpret_cast<const Word *>(bv->getBitValues()), bv->isInverted()); + for (const auto & child : getChildren()) { + const auto * bv = static_cast<const BitVectorIterator *>(child.get()); + _bvs.emplace_back(bv->getBitValues(), bv->isInverted()); _numDocs = std::min(_numDocs, bv->getDocIdLimit()); } } @@ -155,6 +171,7 @@ MultiBitVectorIteratorBase::initRange(uint32_t beginId, uint32_t endId) { MultiSearch::initRange(beginId, endId); _lastMaxDocIdLimit = 0; + _lastMaxDocIdLimitRequireFetch = 0; } SearchIterator::UP @@ -163,9 +180,10 @@ MultiBitVectorIteratorBase::andWith(UP filter, uint32_t estimate) (void) estimate; if (filter->isBitVector() && acceptExtraFilter()) { const auto & bv = static_cast<const BitVectorIterator &>(*filter); - _bvs.emplace_back(reinterpret_cast<const Word *>(bv.getBitValues()), bv.isInverted()); + _bvs.emplace_back(bv.getBitValues(), bv.isInverted()); insert(getChildren().size(), std::move(filter)); _lastMaxDocIdLimit = 0; // force reload + _lastMaxDocIdLimitRequireFetch = 0; } return filter; } diff --git a/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.h b/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.h index cde9ffcbfe5..29e92584ffe 100644 --- a/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.h +++ b/searchlib/src/vespa/searchlib/queryeval/multibitvectoriterator.h @@ -11,7 +11,7 @@ namespace search::queryeval { class MultiBitVectorIteratorBase : public MultiSearch, protected BitWord { public: - ~MultiBitVectorIteratorBase(); + ~MultiBitVectorIteratorBase() override; void initRange(uint32_t beginId, uint32_t endId) override; void addUnpackIndex(size_t index) { _unpackInfo.add(index); } /** @@ -20,26 +20,21 @@ public: */ static SearchIterator::UP optimize(SearchIterator::UP parent); protected: - MultiBitVectorIteratorBase(Children children); - class MetaWord { - public: - MetaWord(const Word * words, bool inverted) : _words(words), _inverted(inverted) { } - Word operator [] (uint32_t index) const { return _inverted ? ~_words[index] : _words[index]; } - private: - const Word * _words; - bool _inverted; - }; + MultiBitVectorIteratorBase(Children hildren); + using MetaWord = std::pair<const void *, bool>; uint32_t _numDocs; - Word _lastValue; // Last value computed uint32_t _lastMaxDocIdLimit; // next documentid requiring recomputation. + uint32_t _lastMaxDocIdLimitRequireFetch; + Word _lastValue; // Last value computed std::vector<MetaWord> _bvs; private: virtual bool acceptExtraFilter() const = 0; UP andWith(UP filter, uint32_t estimate) override; void doUnpack(uint32_t docid) override; - UnpackInfo _unpackInfo; static SearchIterator::UP optimizeMultiSearch(SearchIterator::UP parent); + + UnpackInfo _unpackInfo; }; } diff --git a/searchlib/src/vespa/searchlib/tensor/distance_functions.h b/searchlib/src/vespa/searchlib/tensor/distance_functions.h index 79f987c740c..d37495e85da 100644 --- a/searchlib/src/vespa/searchlib/tensor/distance_functions.h +++ b/searchlib/src/vespa/searchlib/tensor/distance_functions.h @@ -17,7 +17,7 @@ template <typename FloatType> class SquaredEuclideanDistance : public DistanceFunction { public: SquaredEuclideanDistance() - : _computer(vespalib::hwaccelrated::IAccelrated::getAccelrator()) + : _computer(vespalib::hwaccelrated::IAccelrated::getAccelerator()) {} double calc(const vespalib::tensor::TypedCells& lhs, const vespalib::tensor::TypedCells& rhs) const override { auto lhs_vector = lhs.typify<FloatType>(); @@ -60,7 +60,7 @@ template <typename FloatType> class AngularDistance : public DistanceFunction { public: AngularDistance() - : _computer(vespalib::hwaccelrated::IAccelrated::getAccelrator()) + : _computer(vespalib::hwaccelrated::IAccelrated::getAccelerator()) {} double calc(const vespalib::tensor::TypedCells& lhs, const vespalib::tensor::TypedCells& rhs) const override { auto lhs_vector = lhs.typify<FloatType>(); diff --git a/vespalib/src/tests/dotproduct/dotproductbenchmark.cpp b/vespalib/src/tests/dotproduct/dotproductbenchmark.cpp index d6e1aef9394..e95e8a5c58b 100644 --- a/vespalib/src/tests/dotproduct/dotproductbenchmark.cpp +++ b/vespalib/src/tests/dotproduct/dotproductbenchmark.cpp @@ -60,7 +60,7 @@ template <typename T> FullBenchmark<T>::FullBenchmark(size_t numDocs, size_t numValues) : _values(numDocs*numValues), _query(numValues), - _dp(IAccelrated::getAccelrator()) + _dp(IAccelrated::getAccelerator()) { for (size_t i(0); i < numDocs; i++) { for (size_t j(0); j < numValues; j++) { diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp index 7ff393c87f8..8588a5510f7 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp @@ -20,4 +20,14 @@ Avx2Accelrator::squaredEuclideanDistance(const double * a, const double * b, siz return avx::euclideanDistanceSelectAlignment<double, 32>(a, b, sz); } +void +Avx2Accelrator::and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const { + helper::andChunks<32u, 2u>(offset, src, dest); +} + +void +Avx2Accelrator::or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const { + helper::orChunks<32u, 2u>(offset, src, dest); +} + } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h index 3e0dbb28110..b6f3d299748 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h +++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h @@ -15,6 +15,8 @@ public: size_t populationCount(const uint64_t *a, size_t sz) const override; double squaredEuclideanDistance(const float * a, const float * b, size_t sz) const override; double squaredEuclideanDistance(const double * a, const double * b, size_t sz) const override; + void and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const override; + void or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const override; }; } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx512.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/avx512.cpp index 0941e6d6ad8..4dade08e77a 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/avx512.cpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx512.cpp @@ -32,4 +32,14 @@ Avx512Accelrator::squaredEuclideanDistance(const double * a, const double * b, s return avx::euclideanDistanceSelectAlignment<double, 64>(a, b, sz); } +void +Avx512Accelrator::and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const { + helper::andChunks<64, 1>(offset, src, dest); +} + +void +Avx512Accelrator::or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const { + helper::orChunks<64, 1>(offset, src, dest); +} + } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx512.h b/vespalib/src/vespa/vespalib/hwaccelrated/avx512.h index 209ec06c857..a54d57407b2 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/avx512.h +++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx512.h @@ -17,6 +17,8 @@ public: size_t populationCount(const uint64_t *a, size_t sz) const override; double squaredEuclideanDistance(const float * a, const float * b, size_t sz) const override; double squaredEuclideanDistance(const double * a, const double * b, size_t sz) const override; + void and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const override; + void or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const override; }; } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp index f9684e88c63..f9dfaacf626 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp @@ -165,4 +165,14 @@ GenericAccelrator::squaredEuclideanDistance(const double * a, const double * b, return euclideanDistanceT<double, 4>(a, b, sz); } +void +GenericAccelrator::and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const { + helper::andChunks<16, 4>(offset, src, dest); +} + +void +GenericAccelrator::or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const { + helper::orChunks<16,4>(offset, src, dest); +} + } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/generic.h b/vespalib/src/vespa/vespalib/hwaccelrated/generic.h index 50a3d59d49d..2335b40fe85 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/generic.h +++ b/vespalib/src/vespa/vespalib/hwaccelrated/generic.h @@ -25,6 +25,8 @@ public: size_t populationCount(const uint64_t *a, size_t sz) const override; double squaredEuclideanDistance(const float * a, const float * b, size_t sz) const override; double squaredEuclideanDistance(const double * a, const double * b, size_t sz) const override; + void and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const override; + void or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const override; }; } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.cpp index bb132165e53..de917c5f065 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.cpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.cpp @@ -46,7 +46,8 @@ std::vector<T> createAndFill(size_t sz) { } template<typename T> -void verifyDotproduct(const IAccelrated & accel) +void +verifyDotproduct(const IAccelrated & accel) { const size_t testLength(255); srand(1); @@ -66,7 +67,8 @@ void verifyDotproduct(const IAccelrated & accel) } template<typename T> -void verifyEuclideanDistance(const IAccelrated & accel) { +void +verifyEuclideanDistance(const IAccelrated & accel) { const size_t testLength(255); srand(1); std::vector<T> a = createAndFill<T>(testLength); @@ -84,7 +86,8 @@ void verifyEuclideanDistance(const IAccelrated & accel) { } } -void verifyPopulationCount(const IAccelrated & accel) +void +verifyPopulationCount(const IAccelrated & accel) { const uint64_t words[7] = {0x123456789abcdef0L, // 32 0x0000000000000000L, // 0 @@ -101,6 +104,118 @@ void verifyPopulationCount(const IAccelrated & accel) } } +void +fill(std::vector<uint64_t> & v, size_t n) { + v.reserve(n); + for (size_t i(0); i < n; i++) { + v.emplace_back(random()); + } +} + +void +simpleAndWith(std::vector<uint64_t> & dest, const std::vector<uint64_t> & src) { + for (size_t i(0); i < dest.size(); i++) { + dest[i] &= src[i]; + } +} + +void +simpleOrWith(std::vector<uint64_t> & dest, const std::vector<uint64_t> & src) { + for (size_t i(0); i < dest.size(); i++) { + dest[i] |= src[i]; + } +} + +std::vector<uint64_t> +simpleInvert(const std::vector<uint64_t> & src) { + std::vector<uint64_t> inverted; + inverted.reserve(src.size()); + for (size_t i(0); i < src.size(); i++) { + inverted.push_back(~src[i]); + } + return inverted; +} + +std::vector<uint64_t> +optionallyInvert(bool invert, std::vector<uint64_t> v) { + return invert ? simpleInvert(std::move(v)) : std::move(v); +} + +bool shouldInvert(bool invertSome) { + return invertSome ? (random() & 1) : false; +} + +void +verifyOr64(const IAccelrated & accel, const std::vector<std::vector<uint64_t>> & vectors, + size_t offset, size_t num_vectors, bool invertSome) +{ + std::vector<std::pair<const void *, bool>> vRefs; + for (size_t j(0); j < num_vectors; j++) { + vRefs.emplace_back(&vectors[j][0], shouldInvert(invertSome)); + } + + std::vector<uint64_t> expected = optionallyInvert(vRefs[0].second, vectors[0]); + for (size_t j = 1; j < num_vectors; j++) { + simpleOrWith(expected, optionallyInvert(vRefs[j].second, vectors[j])); + } + + uint64_t dest[8] __attribute((aligned(64))); + accel.or64(offset*sizeof(uint64_t), vRefs, dest); + int diff = memcmp(&expected[offset], dest, sizeof(dest)); + if (diff != 0) { + LOG_ABORT("Accelerator fails to compute correct 64 bytes OR"); + } +} + +void +verifyAnd64(const IAccelrated & accel, const std::vector<std::vector<uint64_t>> & vectors, + size_t offset, size_t num_vectors, bool invertSome) +{ + std::vector<std::pair<const void *, bool>> vRefs; + for (size_t j(0); j < num_vectors; j++) { + vRefs.emplace_back(&vectors[j][0], shouldInvert(invertSome)); + } + std::vector<uint64_t> expected = optionallyInvert(vRefs[0].second, vectors[0]); + for (size_t j = 1; j < num_vectors; j++) { + simpleAndWith(expected, optionallyInvert(vRefs[j].second, vectors[j])); + } + + uint64_t dest[8] __attribute((aligned(64))); + accel.and64(offset*sizeof(uint64_t), vRefs, dest); + int diff = memcmp(&expected[offset], dest, sizeof(dest)); + if (diff != 0) { + LOG_ABORT("Accelerator fails to compute correct 64 bytes AND"); + } +} + +void +verifyOr64(const IAccelrated & accel) { + std::vector<std::vector<uint64_t>> vectors(3) ; + for (auto & v : vectors) { + fill(v, 16); + } + for (size_t offset = 0; offset < 8; offset++) { + for (size_t i = 1; i < vectors.size(); i++) { + verifyOr64(accel, vectors, offset, i, false); + verifyOr64(accel, vectors, offset, i, true); + } + } +} + +void +verifyAnd64(const IAccelrated & accel) { + std::vector<std::vector<uint64_t>> vectors(3); + for (auto & v : vectors) { + fill(v, 16); + } + for (size_t offset = 0; offset < 8; offset++) { + for (size_t i = 1; i < vectors.size(); i++) { + verifyAnd64(accel, vectors, offset, i, false); + verifyAnd64(accel, vectors, offset, i, true); + } + } +} + class RuntimeVerificator { public: @@ -114,6 +229,8 @@ private: verifyEuclideanDistance<float>(accelrated); verifyEuclideanDistance<double>(accelrated); verifyPopulationCount(accelrated); + verifyAnd64(accelrated); + verifyOr64(accelrated); } }; @@ -122,7 +239,7 @@ RuntimeVerificator::RuntimeVerificator() GenericAccelrator generic; verify(generic); - const IAccelrated & thisCpu(IAccelrated::getAccelrator()); + const IAccelrated & thisCpu(IAccelrated::getAccelerator()); verify(thisCpu); } @@ -155,7 +272,7 @@ static Selector _G_selector; RuntimeVerificator _G_verifyAccelrator; const IAccelrated & -IAccelrated::getAccelrator() +IAccelrated::getAccelerator() { static IAccelrated::UP accelrator = _G_selector.create(); return *accelrator; diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.h b/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.h index 0292ad14643..2594a48dd33 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.h +++ b/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.h @@ -4,6 +4,7 @@ #include <memory> #include <cstdint> +#include <vector> namespace vespalib::hwaccelrated { @@ -29,8 +30,12 @@ public: virtual size_t populationCount(const uint64_t *a, size_t sz) const = 0; virtual double squaredEuclideanDistance(const float * a, const float * b, size_t sz) const = 0; virtual double squaredEuclideanDistance(const double * a, const double * b, size_t sz) const = 0; + // AND 64 bytes from multiple, optionally inverted sources + virtual void and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const = 0; + // OR 64 bytes from multiple, optionally inverted sources + virtual void or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const = 0; - static const IAccelrated & getAccelrator() __attribute__((noinline)); + static const IAccelrated & getAccelerator() __attribute__((noinline)); }; } diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp b/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp index f5daf2b9081..6fc49f969f2 100644 --- a/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp +++ b/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp @@ -24,5 +24,51 @@ populationCount(const uint64_t *a, size_t sz) { return count; } +template<typename T> +T get(const void * base, bool invert) { + T v; + memcpy(&v, base, sizeof(T)); + return __builtin_expect(invert, false) ? ~v : v; +} + +template <typename T> +const T * cast(const void * ptr, size_t offsetBytes) { + return static_cast<const T *>(static_cast<const void *>(static_cast<const char *>(ptr) + offsetBytes)); +} + +template<unsigned ChunkSize, unsigned Chunks> +void +andChunks(size_t offset, const std::vector<std::pair<const void *, bool>> & src, void * dest) { + typedef uint64_t Chunk __attribute__ ((vector_size (ChunkSize))); + Chunk * chunk = static_cast<Chunk *>(dest); + const Chunk * tmp = cast<Chunk>(src[0].first, offset); + for (size_t n=0; n < Chunks; n++) { + chunk[n] = get<Chunk>(tmp+n, src[0].second); + } + for (size_t i(1); i < src.size(); i++) { + tmp = cast<Chunk>(src[i].first, offset); + for (size_t n=0; n < Chunks; n++) { + chunk[n] &= get<Chunk>(tmp+n, src[i].second); + } + } +} + +template<unsigned ChunkSize, unsigned Chunks> +void +orChunks(size_t offset, const std::vector<std::pair<const void *, bool>> & src, void * dest) { + typedef uint64_t Chunk __attribute__ ((vector_size (ChunkSize))); + Chunk * chunk = static_cast<Chunk *>(dest); + const Chunk * tmp = cast<Chunk>(src[0].first, offset); + for (size_t n=0; n < Chunks; n++) { + chunk[n] = get<Chunk>(tmp+n, src[0].second); + } + for (size_t i(1); i < src.size(); i++) { + tmp = cast<Chunk>(src[i].first, offset); + for (size_t n=0; n < Chunks; n++) { + chunk[n] |= get<Chunk>(tmp+n, src[i].second); + } + } +} + } } |