diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0450e19..3695218 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -105,6 +105,7 @@ if(NOT MSVC)
 endif()
 
 find_package(Threads REQUIRED)
+find_package(OpenCL REQUIRED)
 
 
 ################################################################################
@@ -128,6 +129,11 @@ add_executable(cuda_memtest
         cuda_memtest.cpp
         )
 
+add_executable(ocl_memtest
+        ocl_tests.cpp
+        ocl_memtest.cpp
+        )
+
 if(CUDA_MEMTEST_BACKEND STREQUAL "cuda")
     target_link_libraries(cuda_memtest INTERFACE CUDA::cudart)
     target_link_libraries(cuda_memtest INTERFACE CUDA::cuda_driver)
@@ -152,6 +158,10 @@ endif()
 
 if(NOT MSVC)
     target_link_libraries(cuda_memtest PRIVATE Threads::Threads)
+    target_link_libraries(ocl_memtest
+        PRIVATE Threads::Threads
+        PRIVATE OpenCL::OpenCL
+        )
 endif()
 
 ## annotate with RPATH's
@@ -175,11 +185,14 @@ endif()
 option(CUDA_MEMTEST_RELEASE "disable all runtime asserts" ON)
 if(CUDA_MEMTEST_RELEASE)
     target_compile_definitions(cuda_memtest PRIVATE NDEBUG)
+    target_compile_definitions(ocl_memtest PRIVATE NDEBUG)
 endif(CUDA_MEMTEST_RELEASE)
 
 ################################################################################
 # Install cuda_memtest
 ################################################################################
 
-install(TARGETS cuda_memtest
+install(TARGETS
+        cuda_memtest
+        ocl_memtest
         RUNTIME DESTINATION bin)
diff --git a/ocl_memtest.cpp b/ocl_memtest.cpp
index 69deb5b..22b0e69 100644
--- a/ocl_memtest.cpp
+++ b/ocl_memtest.cpp
@@ -18,7 +18,7 @@ unsigned int exit_on_error = 0;
 
 
 
-#define KERNEL_FILE "ocl_memtest_kernels.cpp"
+#define KERNEL_FILE "ocl_memtest_kernels.cl"
 
 #define MAX_KERNEL_FILE_SIZE (1024*1024)
 
diff --git a/ocl_memtest_kernels.cpp b/ocl_memtest_kernels.cl
similarity index 97%
rename from ocl_memtest_kernels.cpp
rename to ocl_memtest_kernels.cl
index c7ee6f8..0755d10 100644
--- a/ocl_memtest_kernels.cpp
+++ b/ocl_memtest_kernels.cl
@@ -595,9 +595,9 @@ kernel0_local_write(__global char* ptr, unsigned long memsize)
   
 
   for(i=idx; i < n; i+= total_num_threads){
-    __global unsigned long * start_p= (__global unsigned long)(ptr + i*BLOCKSIZE);
-    __global unsigned long* end_p = (__global unsigned long*)(ptr + (i+1)*BLOCKSIZE);
-    __global unsigned long * p =start_p;
+    __global unsigned long * start_p = (__global unsigned long*)(ptr + i*BLOCKSIZE);
+    __global unsigned long * end_p = (__global unsigned long*)(ptr + (i+1)*BLOCKSIZE);
+    __global unsigned long * p = start_p;
     unsigned int pattern = 1;
     unsigned int mask = 8;
       
@@ -645,9 +645,9 @@ kernel0_local_read(__global char* ptr, unsigned long memsize,
   
 
   for(i=idx; i < n; i+= total_num_threads){
-    __global unsigned long * start_p= (__global unsigned long)(ptr + i*BLOCKSIZE);
-    __global unsigned long* end_p = (__global unsigned long*)(ptr + (i+1)*BLOCKSIZE);
-    __global unsigned long * p =start_p;
+    __global unsigned long * start_p = (__global unsigned long*)(ptr + i*BLOCKSIZE);
+    __global unsigned long * end_p = (__global unsigned long*)(ptr + (i+1)*BLOCKSIZE);
+    __global unsigned long * p = start_p;
     unsigned int pattern = 1;
     unsigned int mask = 8;