Fix significant GOMP barrier overhead in exhaustive_L2sqr_blas.

Signed-off-by: Mulugeta Mammo <[email protected]>

Author: mulugetam

faiss/utils/distances.cpp

@@ -321,7 +321,9 @@ void exhaustive_L2sqr_blas_default_impl(
                        ip_block.get(),
                        &nyi);
             }
-#pragma omp parallel for
+// Commented out - causes significant GOMP barrier overhead.
+// #pragma omp parallel for
+//
             for (int64_t i = i0; i < i1; i++) {
                 float* ip_line = ip_block.get() + (i - i0) * (j1 - j0);
 
@@ -423,7 +425,9 @@ void exhaustive_L2sqr_blas_cmax_avx2(
                        ip_block.get(),
                        &nyi);
             }
-#pragma omp parallel for
+// Commented out - causes significant GOMP barrier overhead.
+// #pragma omp parallel for
+//
             for (int64_t i = i0; i < i1; i++) {
                 float* ip_line = ip_block.get() + (i - i0) * (j1 - j0);
 
@@ -633,7 +637,9 @@ void exhaustive_L2sqr_blas_cmax_sve(
                        ip_block.get(),
                        &nyi);
             }
-#pragma omp parallel for
+// Commented out - causes significant GOMP barrier overhead.
+// #pragma omp parallel for
+//
             for (int64_t i = i0; i < i1; i++) {
                 const size_t count = j1 - j0;
                 float* ip_line = ip_block.get() + (i - i0) * count;