Move helper methods.

1 files changed, 50 insertions, 0 deletions

diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp b/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp
index 76072aa6906..a58de582547 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/private_helpers.hpp
@@ -8,6 +8,56 @@
 namespace vespalib::hwaccelrated::helper {
 namespace {
 
+template <typename ACCUM, typename T, size_t UNROLL>
+ACCUM
+multiplyAdd(const T * a, const T * b, size_t sz)
+{
+    ACCUM partial[UNROLL];
+    for (size_t i(0); i < UNROLL; i++) {
+        partial[i] = 0;
+    }
+    size_t i(0);
+    for (; i + UNROLL <= sz; i+= UNROLL) {
+        for (size_t j(0); j < UNROLL; j++) {
+            partial[j] += a[i+j] * b[i+j];
+        }
+    }
+    for (;i < sz; i++) {
+        partial[i%UNROLL] += a[i] * b[i];
+    }
+    ACCUM sum(0);
+    for (size_t j(0); j < UNROLL; j++) {
+        sum += partial[j];
+    }
+    return sum;
+}
+
+template<size_t UNROLL, typename Operation>
+void
+bitOperation(Operation operation, void * aOrg, const void * bOrg, size_t bytes) {
+
+    const size_t sz(bytes/sizeof(uint64_t));
+    {
+        uint64_t *a(static_cast<uint64_t *>(aOrg));
+        const uint64_t *b(static_cast<const uint64_t *>(bOrg));
+        size_t i(0);
+        for (; i + UNROLL <= sz; i += UNROLL) {
+            for (size_t j(0); j < UNROLL; j++) {
+                a[i + j] = operation(a[i + j], b[i + j]);
+            }
+        }
+        for (; i < sz; i++) {
+            a[i] = operation(a[i], b[i]);
+        }
+    }
+
+    uint8_t *a(static_cast<uint8_t *>(aOrg));
+    const uint8_t *b(static_cast<const uint8_t *>(bOrg));
+    for (size_t i(sz*sizeof(uint64_t)); i < bytes; i++) {
+        a[i] = operation(a[i], b[i]);
+    }
+}
+
 inline size_t
 populationCount(const uint64_t *a, size_t sz) {
     size_t count(0);