optimize the implemention of k-selection on faiss gpu version (#2467)

* optimize the implemention of k-selection on faiss gpu version
Signed-off-by: Ncmli <chengming.li@zilliz.com>

* update change log, re-classified the pr from feature to improvement
Signed-off-by: Ncmli <chengming.li@zilliz.com>

* update change log
Signed-off-by: Ncmli <chengming.li@zilliz.com>
Co-authored-by: Ncmli <chengming.li@zilliz.com>

CHANGELOG.md

@@ -12,6 +12,7 @@ Please mark all change in change log and use the issue from GitHub
 ## Improvement
 -   \#2381 Upgrade FAISS to 1.6.3
 -   \#2441 Improve Knowhere code coverage
+-   \#2466 optimize k-selection implemention of faiss gpu version
 
 ## Task
 

core/src/index/thirdparty/faiss/gpu/utils/Select.cuh

@@ -124,16 +124,8 @@ struct BlockSelect {
 
   __device__ inline void addThreadQ(K k, V v) {
     if (Dir ? Comp::gt(k, warpKTop) : Comp::lt(k, warpKTop)) {
-      // Rotate right
-#pragma unroll
-      for (int i = NumThreadQ - 1; i > 0; --i) {
-        threadK[i] = threadK[i - 1];
-        threadV[i] = threadV[i - 1];
-      }
-
-      threadK[0] = k;
-      threadV[0] = v;
-      ++numVals;
+        threadK[numVals] = k;
+        threadV[numVals ++] = v;
     }
   }