8 files changed, 29 insertions, 17 deletions
diff --git a/benchmark/benchmark_read_buffer.cpp b/benchmark/benchmark_read_buffer.cpp
index 31a1f599..431f42aa 100644
--- a/benchmark/benchmark_read_buffer.cpp
+++ b/benchmark/benchmark_read_buffer.cpp
@@ -1,7 +1,7 @@
 #include "utests/utest_helper.hpp"
 #include <sys/time.h>
 
-int benchmark_read_buffer(void)
+double benchmark_read_buffer(void)
 {
   struct timeval start,stop;
 
@@ -43,7 +43,9 @@ int benchmark_read_buffer(void)
   free(buf_data[0]);
   buf_data[0] = NULL;
 
-  return time_subtract(&stop, &start, 0);
+  double elapsed = time_subtract(&stop, &start, 0);
+
+  return BANDWIDTH(sz * sizeof(float) * 2 * 100, elapsed);
 }
 
 MAKE_BENCHMARK_FROM_FUNCTION(benchmark_read_buffer);
diff --git a/benchmark/benchmark_read_image.cpp b/benchmark/benchmark_read_image.cpp
index 48aa9871..e3aa5bd0 100644
--- a/benchmark/benchmark_read_image.cpp
+++ b/benchmark/benchmark_read_image.cpp
@@ -2,7 +2,7 @@
 #include "utests/utest_helper.hpp"
 #include <sys/time.h>
 
-int benchmark_read_image(void)
+double benchmark_read_image(void)
 {
   struct timeval start,stop;
 
@@ -61,7 +61,9 @@ int benchmark_read_image(void)
   free(buf_data[0]);
   buf_data[0] = NULL;
 
-  return time_subtract(&stop, &start, 0);
+  double elapsed = time_subtract(&stop, &start, 0);
+
+  return BANDWIDTH(sz * sizeof(float) * 2 * 100, elapsed);
 }
 
 MAKE_BENCHMARK_FROM_FUNCTION(benchmark_read_image);
diff --git a/benchmark/benchmark_use_host_ptr_buffer.cpp b/benchmark/benchmark_use_host_ptr_buffer.cpp
index 80b6c5c3..9e3d1551 100644
--- a/benchmark/benchmark_use_host_ptr_buffer.cpp
+++ b/benchmark/benchmark_use_host_ptr_buffer.cpp
@@ -1,7 +1,7 @@
 #include "utests/utest_helper.hpp"
 #include <sys/time.h>
 
-int benchmark_use_host_ptr_buffer(void)
+double benchmark_use_host_ptr_buffer(void)
 {
   struct timeval start,stop;
 
@@ -32,7 +32,9 @@ int benchmark_use_host_ptr_buffer(void)
   free(buf_data[0]);
   buf_data[0] = NULL;
 
-  return time_subtract(&stop, &start, 0);
+  double elapsed = time_subtract(&stop, &start, 0);
+
+  return BANDWIDTH(n*sizeof(uint32_t)*100*2, elapsed);
 }
 
 MAKE_BENCHMARK_FROM_FUNCTION(benchmark_use_host_ptr_buffer);
diff --git a/benchmark/enqueue_copy_buf.cpp b/benchmark/enqueue_copy_buf.cpp
index f012cf71..549c8b16 100644
--- a/benchmark/enqueue_copy_buf.cpp
+++ b/benchmark/enqueue_copy_buf.cpp
@@ -28,7 +28,7 @@ void test_copy_buf(size_t sz, size_t src_off, size_t dst_off, size_t cb)
     src_off, dst_off, cb*sizeof(char), 0, NULL, NULL));
 }
 
-int enqueue_copy_buf(void)
+double enqueue_copy_buf(void)
 {
   size_t i;
   const size_t sz = 127 *1023 * 1023;
@@ -41,7 +41,9 @@ int enqueue_copy_buf(void)
   }
 
   gettimeofday(&stop,0);
-  return time_subtract(&stop, &start, 0);
+  double elapsed = time_subtract(&stop, &start, 0);
+
+  return BANDWIDTH(sz * sizeof(char) * 10, elapsed);
 }
 
 MAKE_BENCHMARK_FROM_FUNCTION(enqueue_copy_buf);
diff --git a/utests/utest.hpp b/utests/utest.hpp
index b028b64f..7ae8b876 100644
--- a/utests/utest.hpp
+++ b/utests/utest.hpp
@@ -30,6 +30,7 @@
 #include "utest_exception.hpp"
 #include <vector>
 #include <iostream>
+#include <iomanip>
 
 /*! struct for statistics */
 struct RStatistics
@@ -135,10 +136,10 @@ struct UTest
 
 #define BENCHMARK(EXPR) \
  do { \
-    int ret = 0;\
+    double ret = 0;\
     try { \
       ret = EXPR; \
-      std::cout << "    [Result: " << ret << "]    [SUCCESS]" << std::endl; \
+      std::cout << "    [Result: " << std::fixed<< std::setprecision(3) << ret << " GB/S]    [SUCCESS]" << std::endl; \
       UTest::retStatistics.passCount += 1; \
     } \
     catch (Exception e) { \
@@ -147,5 +148,8 @@ struct UTest
       UTest::retStatistics.failCount++; \
     } \
   } while (0)
+
+#define BANDWIDTH(BYTES, MSEC) \
+  ((double)(BYTES)) / ((MSEC) * 1e6);
 #endif /* __UTEST_UTEST_HPP__ */
 
diff --git a/utests/utest_helper.cpp b/utests/utest_helper.cpp
index 591054e9..d3c378e9 100644
--- a/utests/utest_helper.cpp
+++ b/utests/utest_helper.cpp
@@ -681,7 +681,7 @@ int cl_INT_ULP(int int_number)
   return 0;
 }
 
-int time_subtract(struct timeval *y, struct timeval *x, struct timeval *result)
+double time_subtract(struct timeval *y, struct timeval *x, struct timeval *result)
 {
   if ( x->tv_sec > y->tv_sec )
     return   -1;
@@ -699,6 +699,6 @@ int time_subtract(struct timeval *y, struct timeval *x, struct timeval *result)
     }
   }
 
-  int msec = 1000.0*(y->tv_sec - x->tv_sec) + (y->tv_usec - x->tv_usec)/1000.0;
+  double msec = 1000.0*(y->tv_sec - x->tv_sec) + (y->tv_usec - x->tv_usec)/1000.0;
   return msec;
-}
-\ No newline at end of file
+}
diff --git a/utests/utest_helper.hpp b/utests/utest_helper.hpp
index 5d8e835c..6d097662 100644
--- a/utests/utest_helper.hpp
+++ b/utests/utest_helper.hpp
@@ -231,7 +231,7 @@ extern float cl_FLT_ULP(float float_number);
 extern int cl_INT_ULP(int int_number);
 
 /* subtract the time */
-int time_subtract(struct timeval *y, struct timeval *x, struct timeval *result);
+double time_subtract(struct timeval *y, struct timeval *x, struct timeval *result);
 
 #endif /* __UTEST_HELPER_HPP__ */
 
diff --git a/utests/vload_bench.cpp b/utests/vload_bench.cpp
index a7703fcd..ddfaaee7 100644
--- a/utests/vload_bench.cpp
+++ b/utests/vload_bench.cpp
@@ -34,8 +34,8 @@ static double vload_bench(const char *kernelFunc, uint32_t N, uint32_t offset, b
     OCL_FINISH();
     gettimeofday(&end, NULL);
     double elapsed = (end.tv_sec - start.tv_sec) * 1e6 + (end.tv_usec - start.tv_usec);
-    double bandwidth = (globals[0] * (N_ITERATIONS) * sizeof(T) * N) / elapsed;
-    printf("\t%2.1fGB/S\n", bandwidth/1000.);
+    double bandwidth = (globals[0] * (N_ITERATIONS) * sizeof(T) * N) / (elapsed * 1000.);
+    printf("\t%2.1fGB/S\n", bandwidth);
     return bandwidth;
   } else {
     // Check result
@@ -71,7 +71,7 @@ VLOAD_TEST(float, float)
 #endif
 
 #define VLOAD_BENCH(T, kT) \
-static int vload_bench_ ##kT(void) \
+static double vload_bench_ ##kT(void) \
 { \
   uint8_t vectorSize[] = {2, 3, 4, 8, 16}; \
   double totBandwidth = 0; \