From 482cacd420ada7c0bd0fe46a500b62cb76439955 Mon Sep 17 00:00:00 2001
From: Pierre-Emmanuel Viel <p.emmanuel.viel@gmail.com>
Date: Wed, 1 Jul 2020 18:27:07 +0200
Subject: [PATCH] Mix of 32 and 64bits vector types prevents vectorisation for
 distance computation. Argument "a" is of type ElementType* that is either
 int* or float*, while b was double*. Mixing types prevents the possibility to
 use SSE or AVX instructions. On implementation without SIMD instructions,
 this doesn't show any impact on performance.

---
 .../include/opencv2/flann/kmeans_index.h      | 22 +++++++++----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/modules/flann/include/opencv2/flann/kmeans_index.h b/modules/flann/include/opencv2/flann/kmeans_index.h
index f743d75224..b556b4ad8a 100644
--- a/modules/flann/include/opencv2/flann/kmeans_index.h
+++ b/modules/flann/include/opencv2/flann/kmeans_index.h
@@ -726,15 +726,6 @@ private:
         }
 
 
-        cv::AutoBuffer<double> dcenters_buf(branching*veclen_);
-        Matrix<double> dcenters(dcenters_buf.data(), branching, veclen_);
-        for (int i=0; i<centers_length; ++i) {
-            ElementType* vec = dataset_[centers_idx[i]];
-            for (size_t k=0; k<veclen_; ++k) {
-                dcenters[i][k] = double(vec[k]);
-            }
-        }
-
         std::vector<DistanceType> radiuses(branching);
         cv::AutoBuffer<int> count_buf(branching);
         int* count = count_buf.data();
@@ -748,10 +739,10 @@ private:
         int* belongs_to = belongs_to_buf.data();
         for (int i=0; i<indices_length; ++i) {
 
-            DistanceType sq_dist = distance_(dataset_[indices[i]], dcenters[0], veclen_);
+            DistanceType sq_dist = distance_(dataset_[indices[i]], dataset_[centers_idx[0]], veclen_);
             belongs_to[i] = 0;
             for (int j=1; j<branching; ++j) {
-                DistanceType new_sq_dist = distance_(dataset_[indices[i]], dcenters[j], veclen_);
+                DistanceType new_sq_dist = distance_(dataset_[indices[i]], dataset_[centers_idx[j]], veclen_);
                 if (sq_dist>new_sq_dist) {
                     belongs_to[i] = j;
                     sq_dist = new_sq_dist;
@@ -763,6 +754,15 @@ private:
             count[belongs_to[i]]++;
         }
 
+        cv::AutoBuffer<double> dcenters_buf(branching*veclen_);
+        Matrix<double> dcenters(dcenters_buf.data(), branching, veclen_);
+        for (int i=0; i<centers_length; ++i) {
+            ElementType* vec = dataset_[centers_idx[i]];
+            for (size_t k=0; k<veclen_; ++k) {
+                dcenters[i][k] = double(vec[k]);
+            }
+        }
+
         bool converged = false;
         int iteration = 0;
         while (!converged && iteration<iterations_) {