diff --git a/test_conformance/math_brute_force/binary_double.cpp b/test_conformance/math_brute_force/binary_double.cpp
index e51327df..cc46ebd3 100644
--- a/test_conformance/math_brute_force/binary_double.cpp
+++ b/test_conformance/math_brute_force/binary_double.cpp
@@ -636,12 +636,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     for (j = 0; j < buffer_elements; j++)
         r[j] = (cl_double)func.f_ff(s[j], s2[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_ulong *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -651,16 +652,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_ulong *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                            CL_TRUE, CL_MAP_READ, 0,
-                                            buffer_size, 0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_ulong *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/binary_float.cpp b/test_conformance/math_brute_force/binary_float.cpp
index 5a37a407..6bea0c9a 100644
--- a/test_conformance/math_brute_force/binary_float.cpp
+++ b/test_conformance/math_brute_force/binary_float.cpp
@@ -684,12 +684,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
 
     if (isFDim && ftz) RestoreFPState(&oldMode);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_uint *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -699,16 +700,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_uint *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                           CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                           0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     if (!skipVerification)
     {
         // Verify data
diff --git a/test_conformance/math_brute_force/binary_i_double.cpp b/test_conformance/math_brute_force/binary_i_double.cpp
index c3c3eb13..7e19b243 100644
--- a/test_conformance/math_brute_force/binary_i_double.cpp
+++ b/test_conformance/math_brute_force/binary_i_double.cpp
@@ -635,12 +635,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     for (j = 0; j < buffer_elements; j++)
         r[j] = (cl_double)func.f_fi(s[j], s2[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_ulong *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -650,16 +651,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_ulong *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                            CL_TRUE, CL_MAP_READ, 0,
-                                            buffer_size, 0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_ulong *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/binary_i_float.cpp b/test_conformance/math_brute_force/binary_i_float.cpp
index 93cb910b..08790c9d 100644
--- a/test_conformance/math_brute_force/binary_i_float.cpp
+++ b/test_conformance/math_brute_force/binary_i_float.cpp
@@ -627,12 +627,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     s2 = (cl_int *)gIn2 + thread_id * buffer_elements;
     for (j = 0; j < buffer_elements; j++) r[j] = (float)func.f_fi(s[j], s2[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_uint *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -642,16 +643,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_uint *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                           CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                           0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_uint *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/binary_operator_double.cpp b/test_conformance/math_brute_force/binary_operator_double.cpp
index 09540b03..dc1edf96 100644
--- a/test_conformance/math_brute_force/binary_operator_double.cpp
+++ b/test_conformance/math_brute_force/binary_operator_double.cpp
@@ -629,12 +629,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     for (j = 0; j < buffer_elements; j++)
         r[j] = (cl_double)func.f_ff(s[j], s2[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_ulong *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -644,16 +645,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_ulong *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                            CL_TRUE, CL_MAP_READ, 0,
-                                            buffer_size, 0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_ulong *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/binary_operator_float.cpp b/test_conformance/math_brute_force/binary_operator_float.cpp
index a59be163..bc037654 100644
--- a/test_conformance/math_brute_force/binary_operator_float.cpp
+++ b/test_conformance/math_brute_force/binary_operator_float.cpp
@@ -674,12 +674,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
 
     if (ftz) RestoreFPState(&oldMode);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_uint *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -689,16 +690,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_uint *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                           CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                           0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_uint *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/macro_binary_double.cpp b/test_conformance/math_brute_force/macro_binary_double.cpp
index 7c40a18b..4085bbd5 100644
--- a/test_conformance/math_brute_force/macro_binary_double.cpp
+++ b/test_conformance/math_brute_force/macro_binary_double.cpp
@@ -596,13 +596,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     s2 = (cl_double *)gIn2 + thread_id * buffer_elements;
     for (j = 0; j < buffer_elements; j++) r[j] = dfunc.i_ff(s[j], s2[j]);
 
-
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_long *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -612,16 +612,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_long *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                           CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                           0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_long *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/macro_binary_float.cpp b/test_conformance/math_brute_force/macro_binary_float.cpp
index 12c14dc5..7c51895a 100644
--- a/test_conformance/math_brute_force/macro_binary_float.cpp
+++ b/test_conformance/math_brute_force/macro_binary_float.cpp
@@ -589,13 +589,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     s2 = (float *)gIn2 + thread_id * buffer_elements;
     for (j = 0; j < buffer_elements; j++) r[j] = func.i_ff(s[j], s2[j]);
 
-
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_int *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -605,16 +605,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_int *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                          CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                          0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        goto exit;
-    }
-
     // Verify data
     t = (cl_int *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/macro_unary_double.cpp b/test_conformance/math_brute_force/macro_unary_double.cpp
index 4fadbfc3..d2501e22 100644
--- a/test_conformance/math_brute_force/macro_unary_double.cpp
+++ b/test_conformance/math_brute_force/macro_unary_double.cpp
@@ -413,12 +413,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     cl_double *s = (cl_double *)p;
     for (j = 0; j < buffer_elements; j++) r[j] = dfunc.i_f(s[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_long *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -428,16 +429,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_long *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                           CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                           0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        return error;
-    }
-
     // Verify data
     cl_long *t = (cl_long *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/macro_unary_float.cpp b/test_conformance/math_brute_force/macro_unary_float.cpp
index 4e72748a..94a315d0 100644
--- a/test_conformance/math_brute_force/macro_unary_float.cpp
+++ b/test_conformance/math_brute_force/macro_unary_float.cpp
@@ -416,12 +416,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     float *s = (float *)p;
     for (j = 0; j < buffer_elements; j++) r[j] = ref_func(s[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_int *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -431,16 +432,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (cl_int *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                          CL_TRUE, CL_MAP_READ, 0, buffer_size,
-                                          0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        return error;
-    }
-
     // Verify data
     cl_int *t = (cl_int *)r;
     for (j = 0; j < buffer_elements; j++)
diff --git a/test_conformance/math_brute_force/unary_double.cpp b/test_conformance/math_brute_force/unary_double.cpp
index 3dc1548c..3b2fe473 100644
--- a/test_conformance/math_brute_force/unary_double.cpp
+++ b/test_conformance/math_brute_force/unary_double.cpp
@@ -436,12 +436,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     cl_double *s = (cl_double *)p;
     for (j = 0; j < buffer_elements; j++) r[j] = (cl_double)func.f_f(s[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_ulong *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -450,16 +451,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
             return error;
         }
     }
-    // Wait for the last buffer
-    out[j] = (cl_ulong *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                            CL_TRUE, CL_MAP_READ, 0,
-                                            buffer_size, 0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        return error;
-    }
-
 
     // Verify data
     cl_ulong *t = (cl_ulong *)r;
diff --git a/test_conformance/math_brute_force/unary_float.cpp b/test_conformance/math_brute_force/unary_float.cpp
index 73365d19..83d5e807 100644
--- a/test_conformance/math_brute_force/unary_float.cpp
+++ b/test_conformance/math_brute_force/unary_float.cpp
@@ -491,12 +491,13 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
     float *s = (float *)p;
     for (j = 0; j < buffer_elements; j++) r[j] = (float)func.f_f(s[j]);
 
-    // Read the data back -- no need to wait for the first N-1 buffers. This is
-    // an in order queue.
-    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
+    // Read the data back -- no need to wait for the first N-1 buffers but wait
+    // for the last buffer. This is an in order queue.
+    for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
     {
+        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
         out[j] = (cl_uint *)clEnqueueMapBuffer(
-            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0,
+            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
             buffer_size, 0, NULL, NULL, &error);
         if (error || NULL == out[j])
         {
@@ -506,16 +507,6 @@ static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
         }
     }
 
-    // Wait for the last buffer
-    out[j] = (uint32_t *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
-                                            CL_TRUE, CL_MAP_READ, 0,
-                                            buffer_size, 0, NULL, NULL, &error);
-    if (error || NULL == out[j])
-    {
-        vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error);
-        return error;
-    }
-
     // Verify data
     uint32_t *t = (uint32_t *)r;
     for (j = 0; j < buffer_elements; j++)