增加warp +=情况下的显式sync

faiss/gpu/impl/PQScanMultiPassNoPrecomputed-inl.cuh

@@ -248,13 +248,13 @@ __global__ void pqScanNoPrecomputedMultiPass(
 #pragma unroll
                 for (int byte = 0; byte < kBytesPerCode32; ++byte) {
                     auto code = getByte(code32[word], byte * 8, 8);
-                    // HC Debug
-                    __syncthreads();
 
                     auto offset = codesPerSubQuantizer *
                             (word * kBytesPerCode32 + byte);
 
                     dist += ConvertTo<float>::to(codeDist[offset + code]);
+                    // HC Debug explicit sync
+                    __syncthreads();
                 }
             }
         }