Use openblas for dotproduct on everything up to and including avx2.

9 files changed, 16 insertions, 116 deletions

diff --git a/vespalib/src/tests/exception_classes/mmap.cpp b/vespalib/src/tests/exception_classes/mmap.cpp
index 81b5e0de30e..2a6896bb0e0 100644
--- a/vespalib/src/tests/exception_classes/mmap.cpp
+++ b/vespalib/src/tests/exception_classes/mmap.cpp
@@ -2,7 +2,7 @@
 #include <vespa/vespalib/util/alloc.h>
 #include <vector>
 #include <cassert>
-#include <string.h>
+#include <cstring>
 #include <cstdlib>
 #include <sys/resource.h>
 
diff --git a/vespalib/src/vespa/vespalib/CMakeLists.txt b/vespalib/src/vespa/vespalib/CMakeLists.txt
index 4a753a66394..92149d3f0ea 100644
--- a/vespalib/src/vespa/vespalib/CMakeLists.txt
+++ b/vespalib/src/vespa/vespalib/CMakeLists.txt
@@ -30,6 +30,8 @@ vespa_add_library(vespalib
     ${VESPA_GCC_LIB}
 )
 
+set(BLA_VENDOR OpenBLAS)
+vespa_add_target_package_dependency(vespalib BLAS)
 vespa_add_target_package_dependency(vespalib OpenSSL)
 vespa_add_target_package_dependency(vespalib RE2)
 
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/avx.cpp
index 39ea0d2d73b..ec6dc164323 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx.cpp
@@ -5,18 +5,6 @@
 
 namespace vespalib::hwaccelrated {
 
-float
-AvxAccelrator::dotProduct(const float * af, const float * bf, size_t sz) const
-{
-    return avx::dotProductSelectAlignment<float, 32>(af, bf, sz);
-}
-
-double
-AvxAccelrator::dotProduct(const double * af, const double * bf, size_t sz) const
-{
-    return avx::dotProductSelectAlignment<double, 32>(af, bf, sz);
-}
-
 size_t
 AvxAccelrator::populationCount(const uint64_t *a, size_t sz) const {
     return helper::populationCount(a, sz);
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx.h b/vespalib/src/vespa/vespalib/hwaccelrated/avx.h
index 624531a9ca5..e7f090b4695 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx.h
@@ -12,8 +12,6 @@ namespace vespalib::hwaccelrated {
 class AvxAccelrator : public Sse2Accelrator
 {
 public:
-    float dotProduct(const float * a, const float * b, size_t sz) const override;
-    double dotProduct(const double * a, const double * b, size_t sz) const override;
     size_t populationCount(const uint64_t *a, size_t sz) const override;
 };
 
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp
index ea8a3ead538..f0d03a995e4 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.cpp
@@ -5,18 +5,6 @@
 
 namespace vespalib::hwaccelrated {
 
-float
-Avx2Accelrator::dotProduct(const float * af, const float * bf, size_t sz) const
-{
-    return avx::dotProductSelectAlignment<float, 32>(af, bf, sz);
-}
-
-double
-Avx2Accelrator::dotProduct(const double * af, const double * bf, size_t sz) const
-{
-    return avx::dotProductSelectAlignment<double, 32>(af, bf, sz);
-}
-
 size_t
 Avx2Accelrator::populationCount(const uint64_t *a, size_t sz) const {
     return helper::populationCount(a, sz);
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h
index cf91bc81cfd..7e1784698f1 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/avx2.h
@@ -12,8 +12,6 @@ namespace vespalib::hwaccelrated {
 class Avx2Accelrator : public AvxAccelrator
 {
 public:
-    float dotProduct(const float * a, const float * b, size_t sz) const override;
-    double dotProduct(const double * a, const double * b, size_t sz) const override;
     size_t populationCount(const uint64_t *a, size_t sz) const override;
 };
 
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp
index b70ebb4051a..c0a9258cd74 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/generic.cpp
@@ -2,6 +2,7 @@
 
 #include "generic.h"
 #include "private_helpers.hpp"
+#include <cblas.h>
 
 namespace vespalib::hwaccelrated {
 
@@ -37,8 +38,8 @@ 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));
+        auto a(static_cast<uint64_t *>(aOrg));
+        auto 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++) {
@@ -50,8 +51,8 @@ bitOperation(Operation operation, void * aOrg, const void * bOrg, size_t bytes)
         }
     }
 
-    uint8_t *a(static_cast<uint8_t *>(aOrg));
-    const uint8_t *b(static_cast<const uint8_t *>(bOrg));
+    auto a(static_cast<uint8_t *>(aOrg));
+    auto *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]);
     }
@@ -62,36 +63,36 @@ bitOperation(Operation operation, void * aOrg, const void * bOrg, size_t bytes)
 float
 GenericAccelrator::dotProduct(const float * a, const float * b, size_t sz) const
 {
-    return multiplyAdd<float, float, 4>(a, b, sz);
+    return cblas_sdot(sz, a, 1, b, 1);
 }
 
 double
 GenericAccelrator::dotProduct(const double * a, const double * b, size_t sz) const
 {
-    return multiplyAdd<double, double, 4>(a, b, sz);
+    return cblas_ddot(sz, a, 1, b, 1);
 }
 
 int64_t
 GenericAccelrator::dotProduct(const int8_t * a, const int8_t * b, size_t sz) const
 {
-    return multiplyAdd<int64_t, int8_t, 4>(a, b, sz);
+    return multiplyAdd<int64_t, int8_t, 8>(a, b, sz);
 }
 
 int64_t
 GenericAccelrator::dotProduct(const int16_t * a, const int16_t * b, size_t sz) const
 {
-    return multiplyAdd<int64_t, int16_t, 4>(a, b, sz);
+    return multiplyAdd<int64_t, int16_t, 8>(a, b, sz);
 }
 int64_t
 GenericAccelrator::dotProduct(const int32_t * a, const int32_t * b, size_t sz) const
 {
-    return multiplyAdd<int64_t, int32_t, 4>(a, b, sz);
+    return multiplyAdd<int64_t, int32_t, 8>(a, b, sz);
 }
 
 long long
 GenericAccelrator::dotProduct(const int64_t * a, const int64_t * b, size_t sz) const
 {
-    return multiplyAdd<long long, int64_t, 4>(a, b, sz);
+    return multiplyAdd<long long, int64_t, 8>(a, b, sz);
 }
 
 void
@@ -114,12 +115,12 @@ GenericAccelrator::andNotBit(void * aOrg, const void * bOrg, size_t bytes) const
 void
 GenericAccelrator::notBit(void * aOrg, size_t bytes) const
 {
-    uint64_t *a(static_cast<uint64_t *>(aOrg));
+    auto a(static_cast<uint64_t *>(aOrg));
     const size_t sz(bytes/sizeof(uint64_t));
     for (size_t i(0); i < sz; i++) {
         a[i] = ~a[i];
     }
-    uint8_t *ac(static_cast<uint8_t *>(aOrg));
+    auto ac(static_cast<uint8_t *>(aOrg));
     for (size_t i(sz*sizeof(uint64_t)); i < bytes; i++) {
         ac[i] = ~ac[i];
     }
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp b/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp
index a0f584f8a9f..64a26d49f2b 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/sse2.cpp
@@ -5,78 +5,6 @@
 
 namespace vespalib::hwaccelrated {
 
-namespace {
-
-bool validAlignment16(const void * p) {
-    return (reinterpret_cast<uint64_t>(p) & 0xful) == 0;
-}
-
-bool validAlignment16(const void * a, const void * b) {
-    return validAlignment16(a) && validAlignment16(b);
-}
-
-}
-
-float
-Sse2Accelrator::dotProduct(const float * af, const float * bf, size_t sz) const
-{
-    if ( ! validAlignment16(af, bf)) {
-        return GenericAccelrator::dotProduct(af, bf, sz);
-    }
-    typedef float v4sf __attribute__ ((vector_size (16)));
-    const size_t ChunkSize(16);
-    const size_t VectorsPerChunk(ChunkSize/4);
-    v4sf partial[VectorsPerChunk] = { {0.0, 0.0, 0.0, 0.0}, {0.0, 0.0, 0.0, 0.0}, {0.0, 0.0, 0.0, 0.0}, {0.0, 0.0, 0.0, 0.0} };
-    const v4sf * a = reinterpret_cast<const v4sf *>(af);
-    const v4sf * b = reinterpret_cast<const v4sf *>(bf);
-
-    const size_t numChunks(sz/ChunkSize);
-    for (size_t i(0); i < numChunks; i++) {
-        for (size_t j(0); j < VectorsPerChunk; j++) {
-            partial[j] += a[VectorsPerChunk*i+j] * b[VectorsPerChunk*i+j];
-        }
-    }
-    float sum(0);
-    for (size_t i(numChunks*ChunkSize); i < sz; i++) {
-        sum += af[i] * bf[i];
-    }
-    for (size_t i(1); i < VectorsPerChunk; i++) {
-        partial[0] += partial[i];
-    }
-    sum += partial[0][0] + partial[0][1] + partial[0][2] + partial[0][3];
-    return sum; 
-}
-
-double
-Sse2Accelrator::dotProduct(const double * af, const double * bf, size_t sz) const
-{
-    if ( ! validAlignment16(af, bf)) {
-        return GenericAccelrator::dotProduct(af, bf, sz);
-    }
-    typedef double v2sd __attribute__ ((vector_size (16)));
-    const size_t ChunkSize(8);
-    const size_t VectorsPerChunk(ChunkSize/2);
-    v2sd partial[VectorsPerChunk] = { {0.0, 0.0}, {0.0, 0.0}, {0.0, 0.0}, {0.0, 0.0} };
-    const v2sd * a = reinterpret_cast<const v2sd *>(af);
-    const v2sd * b = reinterpret_cast<const v2sd *>(bf);
-
-    const size_t numChunks(sz/ChunkSize);
-    for (size_t i(0); i < numChunks; i++) {
-        for (size_t j(0); j < VectorsPerChunk; j++) {
-            partial[j] += a[VectorsPerChunk*i+j] * b[VectorsPerChunk*i+j];
-        }
-    }
-    double sum(0);
-    for (size_t i(numChunks*ChunkSize); i < sz; i++) {
-        sum += af[i] * bf[i];
-    }
-    for (size_t i(1); i < VectorsPerChunk; i++) {
-        partial[0] += partial[i];
-    }
-    sum += partial[0][0] + partial[0][1];
-    return sum; 
-}
-
 size_t
 Sse2Accelrator::populationCount(const uint64_t *a, size_t sz) const {
     return helper::populationCount(a, sz);
diff --git a/vespalib/src/vespa/vespalib/hwaccelrated/sse2.h b/vespalib/src/vespa/vespalib/hwaccelrated/sse2.h
index d0fbefe5f03..0b2462423b1 100644
--- a/vespalib/src/vespa/vespalib/hwaccelrated/sse2.h
+++ b/vespalib/src/vespa/vespalib/hwaccelrated/sse2.h
@@ -12,9 +12,6 @@ namespace vespalib::hwaccelrated {
 class Sse2Accelrator : public GenericAccelrator
 {
 public:
-    float dotProduct(const float * a, const float * b, size_t sz) const override;
-    double dotProduct(const double * a, const double * b, size_t sz) const override;
-
     size_t populationCount(const uint64_t *a, size_t sz) const override;
 };