- Stick to void * and byte offsets.

- Correct spelling error.

10 files changed, 37 insertions, 31 deletions

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 233609d505b..8588a5510f7 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp
@@ -21,12 +21,12 @@ Avx2Accelrator::squaredEuclideanDistance(const double * a, const double * b, siz
 }
 
 void
-Avx2Accelrator::and64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) const {
+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 uint64_t *, bool>> &src, uint64_t *dest) const {
+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 292961a6f4d..b6f3d299748 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h
@@ -15,8 +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 uint64_t *, bool>> &src, uint64_t *dest) const override;
-    void or64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) 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 9cfae4757b9..4dade08e77a 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx512.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx512.cpp
@@ -33,12 +33,12 @@ Avx512Accelrator::squaredEuclideanDistance(const double * a, const double * b, s
 }
 
 void
-Avx512Accelrator::and64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) const {
+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 uint64_t *, bool>> &src, uint64_t *dest) const {
+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 ee422b57171..a54d57407b2 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx512.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx512.h
@@ -17,8 +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 uint64_t *, bool>> &src, uint64_t *dest) const override;
-    void or64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) 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 460ae7e7388..f9dfaacf626 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp
@@ -166,12 +166,12 @@ GenericAccelrator::squaredEuclideanDistance(const double * a, const double * b,
 }
 
 void
-GenericAccelrator::and64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) const {
+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 uint64_t *, bool>> &src, uint64_t *dest) const {
+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 8ce320cd4c4..2335b40fe85 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/generic.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/generic.h
@@ -25,8 +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 uint64_t *, bool>> &src, uint64_t *dest) const override;
-    void or64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) 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 068833357c3..de917c5f065 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.cpp
@@ -149,7 +149,7 @@ 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 uint64_t *, bool>> vRefs;
+    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));
     }
@@ -160,7 +160,7 @@ verifyOr64(const IAccelrated & accel, const std::vector<std::vector<uint64_t>> &
     }
 
     uint64_t dest[8] __attribute((aligned(64)));
-    accel.or64(offset, vRefs, dest);
+    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");
@@ -171,7 +171,7 @@ 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 uint64_t *, bool>> vRefs;
+    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));
     }
@@ -181,7 +181,7 @@ verifyAnd64(const IAccelrated & accel, const std::vector<std::vector<uint64_t>>
     }
 
     uint64_t dest[8] __attribute((aligned(64)));
-    accel.and64(offset, vRefs, dest);
+    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");
@@ -239,7 +239,7 @@ RuntimeVerificator::RuntimeVerificator()
     GenericAccelrator generic;
     verify(generic);
 
-    const IAccelrated & thisCpu(IAccelrated::getAccelrator());
+    const IAccelrated & thisCpu(IAccelrated::getAccelerator());
     verify(thisCpu);
 }
 
@@ -272,7 +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 f352fb292ce..2594a48dd33 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/iaccelrated.h
@@ -30,11 +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 sources
-    virtual void and64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) const = 0;
-    virtual void or64(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &src, uint64_t *dest) 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 2759cc35ba9..6fc49f969f2 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp
@@ -31,17 +31,22 @@ T get(const void * base, bool invert) {
     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 uint64_t *, bool>> & src, uint64_t * dest) {
+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 = reinterpret_cast<Chunk *>(dest);
-    const Chunk * tmp = reinterpret_cast<const Chunk *>(src[0].first+offset);
+    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 = reinterpret_cast<const Chunk *>(src[i].first+offset);
+        tmp = cast<Chunk>(src[i].first, offset);
         for (size_t n=0; n < Chunks; n++) {
             chunk[n] &= get<Chunk>(tmp+n, src[i].second);
         }
@@ -50,15 +55,15 @@ andChunks(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> &
 
 template<unsigned ChunkSize, unsigned Chunks>
 void
-orChunks(size_t offset, const std::vector<std::pair<const uint64_t *, bool>> & src, uint64_t * dest) {
+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 = reinterpret_cast<Chunk *>(dest);
-    const Chunk * tmp = reinterpret_cast<const Chunk *>(src[0].first+offset);
+    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 = reinterpret_cast<const Chunk *>(src[i].first+offset);
+        tmp = cast<Chunk>(src[i].first, offset);
         for (size_t n=0; n < Chunks; n++) {
             chunk[n] |= get<Chunk>(tmp+n, src[i].second);
         }