diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 027f82ba..df20cbe5 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -39,10 +39,10 @@ cuda_add_executable(
   common/decoder_main.cpp
   gpu/decoder/ape_penalty.cu
   gpu/decoder/encoder_decoder.cu
-  gpu/decoder/nth_element.cu
   gpu/dl4mt/encoder.cu
   gpu/dl4mt/gru.cu
   gpu/mblas/matrix.cu
+  gpu/mblas/nth_element.cu
   gpu/npz_converter.cu
   common/loader_factory.cpp
   $<TARGET_OBJECTS:libcommon>
@@ -55,8 +55,8 @@ cuda_add_library(amunmt SHARED
   python/amunmt.cpp
   gpu/decoder/ape_penalty.cu
   gpu/decoder/encoder_decoder.cu
-  gpu/decoder/nth_element.cu
   gpu/mblas/matrix.cu
+  gpu/mblas/nth_element.cu
   gpu/dl4mt/encoder.cu
   gpu/dl4mt/gru.cu
   gpu/npz_converter.cu
diff --git a/src/gpu/decoder/best_hyps.h b/src/gpu/decoder/best_hyps.h
index a91b3e1c..9fc42f82 100644
--- a/src/gpu/decoder/best_hyps.h
+++ b/src/gpu/decoder/best_hyps.h
@@ -2,7 +2,7 @@
 
 #include "common/scorer.h"
 #include "gpu/mblas/matrix.h"
-#include "gpu/decoder/nth_element.h"
+#include "gpu/mblas/nth_element.h"
 
 #include <thrust/device_vector.h>
 #include <thrust/functional.h>
diff --git a/src/gpu/mblas/matrix.h b/src/gpu/mblas/matrix.h
index ffd21e99..ad6bf0f0 100644
--- a/src/gpu/mblas/matrix.h
+++ b/src/gpu/mblas/matrix.h
@@ -27,8 +27,6 @@
 
 #include "gpu/decoder/encoder_decoder.h"
 #include "gpu/types-gpu.h"
-#include "gpu/nth_element.h"
-
 
 namespace GPU {
 namespace mblas {
diff --git a/src/gpu/decoder/nth_element.cu b/src/gpu/mblas/nth_element.cu
similarity index 99%
rename from src/gpu/decoder/nth_element.cu
rename to src/gpu/mblas/nth_element.cu
index 94befd6e..e8da3ec8 100644
--- a/src/gpu/decoder/nth_element.cu
+++ b/src/gpu/mblas/nth_element.cu
@@ -1,4 +1,4 @@
-#include "gpu/decoder/nth_element.h"
+#include "gpu/mblas/nth_element.h"
 #include <iostream>
 
 
diff --git a/src/gpu/decoder/nth_element.h b/src/gpu/mblas/nth_element.h
similarity index 100%
rename from src/gpu/decoder/nth_element.h
rename to src/gpu/mblas/nth_element.h