flex_cblas_sdot and flex_cblas_snrm2

apps/utils/compute_groundtruth.cpp

@@ -26,8 +26,10 @@
 #else
 #include <stdlib.h>
 #endif
+
 #include "filter_utils.h"
 #include "utils.h"
+#include "../../src/math_utils.cpp"
 
 // WORKS FOR UPTO 2 BILLION POINTS (as we use INT INSTEAD OF UNSIGNED)
 
@@ -75,8 +77,8 @@ void compute_l2sq(float *const points_l2sq, const float *const matrix, const int
     assert(points_l2sq != NULL);
 #pragma omp parallel for schedule(static, 65536)
     for (int64_t d = 0; d < num_points; ++d)
-        points_l2sq[d] = cblas_sdot((int64_t)dim, matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1,
-                                    matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1);
+        points_l2sq[d] = math_utils::flex_cblas_sdot((int64_t)dim, matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1,
+                                                     matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1);
 }
 
 void distsq_to_points(const size_t dim,

apps/utils/compute_groundtruth_for_filters.cpp

@@ -29,6 +29,7 @@
 
 #include "filter_utils.h"
 #include "utils.h"
+#include "../../src/math_utils.cpp"
 
 // WORKS FOR UPTO 2 BILLION POINTS (as we use INT INSTEAD OF UNSIGNED)
 
@@ -76,8 +77,8 @@ void compute_l2sq(float *const points_l2sq, const float *const matrix, const int
     assert(points_l2sq != NULL);
 #pragma omp parallel for schedule(static, 65536)
     for (int64_t d = 0; d < num_points; ++d)
-        points_l2sq[d] = cblas_sdot((int64_t)dim, matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1,
-                                    matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1);
+        points_l2sq[d] = math_utils::flex_cblas_sdot((int64_t)dim, matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1,
+                                                     matrix + (ptrdiff_t)d * (ptrdiff_t)dim, 1);
 }
 
 void distsq_to_points(const size_t dim,

src/math_utils.cpp

@@ -4,13 +4,42 @@
 #include <limits>
 #include <malloc.h>
 #include <math_utils.h>
-#include <mkl.h>
 #include "logger.h"
 #include "utils.h"
 
+#ifdef USE_OPENBLAS
+#include <cblas.h>
+#else
+#include <mkl.h>
+#endif
+
 namespace math_utils
 {
 
+#ifdef USE_OPENBLAS
+using Flex_INT = int;
+#else
+using Flex_INT = MKL_INT;
+#endif
+
+float flex_cblas_sdot(const int64_t N, const float *X, const int incX, const float *Y, const int incY) noexcept
+{
+#ifdef USE_OPENBLAS
+    return sdot(N, X, incX, Y, incY);
+#else
+    return cblas_sdot(N, X, incX, Y, incY);
+#endif
+}
+
+float flex_cblas_snrm2(const Flex_INT N, const float *X, const Flex_INT incX) noexcept
+{
+#ifdef USE_OPENBLAS
+    return snrm2(N, X, incX);
+#else
+    return cblas_snrm2(N, X, incX);
+#endif
+}
+
 float calc_distance(float *vec_1, float *vec_2, size_t dim)
 {
     float dist = 0;
@@ -29,7 +58,7 @@ void compute_vecs_l2sq(float *vecs_l2sq, float *data, const size_t num_points, c
 #pragma omp parallel for schedule(static, 8192)
     for (int64_t n_iter = 0; n_iter < (int64_t)num_points; n_iter++)
     {
-        vecs_l2sq[n_iter] = cblas_snrm2((MKL_INT)dim, (data + (n_iter * dim)), 1);
+        vecs_l2sq[n_iter] = flex_cblas_snrm2((MKL_INT)dim, (data + (n_iter * dim)), 1);
         vecs_l2sq[n_iter] *= vecs_l2sq[n_iter];
     }
 }