diff --git a/test_conformance/select/test_select.cpp b/test_conformance/select/test_select.cpp
index 7fa3bc08..b0cda09f 100644
--- a/test_conformance/select/test_select.cpp
+++ b/test_conformance/select/test_select.cpp
@@ -303,6 +303,10 @@ static int doTest(cl_command_queue queue, cl_context context, Type stype, Type c
     cl_mem dest = NULL;
     void *ref = NULL;
     void *sref = NULL;
+    void *src1_host = NULL;
+    void *src2_host = NULL;
+    void *cmp_host = NULL;
+    void *dest_host = NULL;
 
     cl_ulong blocks = type_size[stype] * 0x100000000ULL / BUFFER_SIZE;
     size_t block_elements = BUFFER_SIZE / type_size[stype];
@@ -359,6 +363,30 @@ static int doTest(cl_command_queue queue, cl_context context, Type stype, Type c
     dest = clCreateBuffer( context, CL_MEM_WRITE_ONLY, BUFFER_SIZE, NULL, &err );
     if( err ) { log_error( "Error: could not allocate dest buffer\n" );  ++s_test_fail; goto exit; }
 
+    src1_host = malloc(BUFFER_SIZE);
+    if (NULL == src1_host)
+    {
+        log_error("Error: could not allocate src1_host buffer\n");
+        goto exit;
+    }
+    src2_host = malloc(BUFFER_SIZE);
+    if (NULL == src2_host)
+    {
+        log_error("Error: could not allocate src2_host buffer\n");
+        goto exit;
+    }
+    cmp_host = malloc(BUFFER_SIZE);
+    if (NULL == cmp_host)
+    {
+        log_error("Error: could not allocate cmp_host buffer\n");
+        goto exit;
+    }
+    dest_host = malloc(BUFFER_SIZE);
+    if (NULL == dest_host)
+    {
+        log_error("Error: could not allocate dest_host buffer\n");
+        goto exit;
+    }
 
     // We block the test as we are running over the range of compare values
     // "block the test" means "break the test into blocks"
@@ -387,13 +415,6 @@ static int doTest(cl_command_queue queue, cl_context context, Type stype, Type c
         // Setup the input data to change for each block
         initCmpBuffer(s3, cmptype, i * cmp_stride, block_elements);
 
-        // Create the reference result
-        Select sfunc = (cmptype == ctype[stype][0]) ? vrefSelects[stype][0] : vrefSelects[stype][1];
-        (*sfunc)(ref, s1, s2, s3, block_elements);
-
-        sfunc = (cmptype == ctype[stype][0]) ? refSelects[stype][0] : refSelects[stype][1];
-        (*sfunc)(sref, s1, s2, s3, block_elements);
-
         if( (err = clEnqueueUnmapMemObject( queue, src1, s1, 0, NULL, NULL )))
         { log_error( "Error: coult not unmap src1\n" );  ++s_test_fail; goto exit; }
         if( (err = clEnqueueUnmapMemObject( queue, src2, s2, 0, NULL, NULL )))
@@ -401,6 +422,40 @@ static int doTest(cl_command_queue queue, cl_context context, Type stype, Type c
         if( (err = clEnqueueUnmapMemObject( queue, cmp, s3, 0, NULL, NULL )))
         { log_error( "Error: coult not unmap cmp\n" );  ++s_test_fail; goto exit; }
 
+        // Create the reference result
+        err = clEnqueueReadBuffer(queue, src1, CL_TRUE, 0, BUFFER_SIZE,
+                                  src1_host, 0, NULL, NULL);
+        if (err)
+        {
+            log_error("Error: Reading buffer from src1 to src1_host failed\n");
+            ++s_test_fail;
+            goto exit;
+        }
+        err = clEnqueueReadBuffer(queue, src2, CL_TRUE, 0, BUFFER_SIZE,
+                                  src2_host, 0, NULL, NULL);
+        if (err)
+        {
+            log_error("Error: Reading buffer from src2 to src2_host failed\n");
+            ++s_test_fail;
+            goto exit;
+        }
+        err = clEnqueueReadBuffer(queue, cmp, CL_TRUE, 0, BUFFER_SIZE, cmp_host,
+                                  0, NULL, NULL);
+        if (err)
+        {
+            log_error("Error: Reading buffer from cmp to cmp_host failed\n");
+            ++s_test_fail;
+            goto exit;
+        }
+
+        Select sfunc = (cmptype == ctype[stype][0]) ? vrefSelects[stype][0]
+                                                    : vrefSelects[stype][1];
+        (*sfunc)(ref, src1_host, src2_host, cmp_host, block_elements);
+
+        sfunc = (cmptype == ctype[stype][0]) ? refSelects[stype][0]
+                                             : refSelects[stype][1];
+        (*sfunc)(sref, src1_host, src2_host, cmp_host, block_elements);
+
         for (vecsize = 0; vecsize < VECTOR_SIZE_COUNT; ++vecsize)
         {
             size_t vector_size = element_count[vecsize] * type_size[stype];
@@ -415,7 +470,6 @@ static int doTest(cl_command_queue queue, cl_context context, Type stype, Type c
             if((err = clSetKernelArg(kernels[vecsize], 3,  sizeof cmp, &cmp) ))
             { log_error( "Error: Cannot set kernel arg dest! %d\n", err ); ++s_test_fail; goto exit; }
 
-
             // Wipe destination
             void *d = clEnqueueMapBuffer( queue, dest, CL_TRUE, CL_MAP_WRITE, 0, BUFFER_SIZE, 0, NULL, NULL, &err );
             if( err ){ log_error( "Error: Could not map dest" );  ++s_test_fail; goto exit; }
@@ -429,18 +483,22 @@ static int doTest(cl_command_queue queue, cl_context context, Type stype, Type c
                 goto exit;
             }
 
-            d = clEnqueueMapBuffer( queue, dest, CL_TRUE, CL_MAP_READ, 0, BUFFER_SIZE, 0, NULL, NULL, &err );
-            if( err ){ log_error( "Error: Could not map dest # 2" );  ++s_test_fail; goto exit; }
-
-            if ((*checkResults[stype])(d, vecsize == 0 ? sref : ref, block_elements, element_count[vecsize])!=0){
-                log_error("vec_size:%d indx: 0x%16.16llx\n", (int)element_count[vecsize], i);
+            err = clEnqueueReadBuffer(queue, dest, CL_TRUE, 0, BUFFER_SIZE,
+                                      dest_host, 0, NULL, NULL);
+            if (err)
+            {
+                log_error(
+                    "Error: Reading buffer from dest to dest_host failed\n");
                 ++s_test_fail;
                 goto exit;
             }
 
-            if( (err = clEnqueueUnmapMemObject( queue, dest, d, 0, NULL, NULL ) ) )
+            if ((*checkResults[stype])(dest_host, vecsize == 0 ? sref : ref,
+                                       block_elements, element_count[vecsize])
+                != 0)
             {
-                log_error( "Error: Could not unmap dest" );
+                log_error("vec_size:%d indx: 0x%16.16llx\n",
+                          (int)element_count[vecsize], i);
                 ++s_test_fail;
                 goto exit;
             }
@@ -459,6 +517,10 @@ exit:
     if( dest)   clReleaseMemObject( dest );
     if( ref )   free(ref );
     if( sref )  free(sref );
+    if (src1_host) free(src1_host);
+    if (src2_host) free(src2_host);
+    if (cmp_host) free(cmp_host);
+    if (dest_host) free(dest_host);
 
     for (vecsize = 0; vecsize < VECTOR_SIZE_COUNT; vecsize++) {
         clReleaseKernel(kernels[vecsize]);