From dc7015c5f24e7849bc48b99dec07252cc3ec6715 Mon Sep 17 00:00:00 2001
From: evghenii <egaburov@dds.nl>
Date: Sat, 9 Nov 2013 19:08:08 +0100
Subject: [PATCH 1/2] added wc-timer for host code

---
 examples_cuda/rt/rt.cpp | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/examples_cuda/rt/rt.cpp b/examples_cuda/rt/rt.cpp
index 9ec340da..79221a07 100644
--- a/examples_cuda/rt/rt.cpp
+++ b/examples_cuda/rt/rt.cpp
@@ -133,10 +133,12 @@ int main(int argc, char *argv[]) {
   if (filename == NULL)
     usage();
 
+  fprintf(stderr, " --- 1 --- \n");
+
 #define READ(var, n)                                            \
   if (fread(&(var), sizeof(var), n, f) != (unsigned int)n) {  \
     fprintf(stderr, "Unexpected EOF reading scene file\n"); \
-    return 1;                                               \
+    assert(0); \
   } else /* eat ; */                                                     
 
   //
@@ -149,6 +151,7 @@ int main(int argc, char *argv[]) {
     perror(fnbuf);
     return 1;
   }
+  fprintf(stderr, " --- 2 --- \n");
 
   //
   // Nothing fancy, and trouble if we run on a big-endian system, just
@@ -170,6 +173,7 @@ int main(int argc, char *argv[]) {
     perror(fnbuf);
     return 1;
   }
+  fprintf(stderr, " --- 3 --- \n");
 
   // The BVH file starts with an int that gives the total number of BVH
   // nodes
@@ -177,6 +181,7 @@ int main(int argc, char *argv[]) {
   READ(nNodes, 1);
 
   LinearBVHNode *nodes = new LinearBVHNode[nNodes];
+#pragma omp parallel for
   for (unsigned int i = 0; i < nNodes; ++i) {
     // Each node is 6x floats for a boox, then an integer for an offset
     // to the second child node, then an integer that encodes the type
@@ -195,10 +200,13 @@ int main(int argc, char *argv[]) {
     READ(nodes[i].pad, 1);
   }
 
+  fprintf(stderr, " --- 4 --- \n");
+
   // And then read the triangles 
   uint nTris;
   READ(nTris, 1);
   Triangle *triangles = new Triangle[nTris];
+#pragma omp parallel for
   for (uint i = 0; i < nTris; ++i) {
     // 9x floats for the 3 vertices
     float v[9];
@@ -214,6 +222,7 @@ int main(int argc, char *argv[]) {
   }
   fclose(f);
 
+  fprintf(stderr, " --- 5 --- \n");
   int height = int(baseHeight * scale);
   int width = int(baseWidth * scale);
 

From 3a549e5c2fa1749f30dc4c436462adddb57fbeba Mon Sep 17 00:00:00 2001
From: evghenii <egaburov@dds.nl>
Date: Sat, 9 Nov 2013 19:26:19 +0100
Subject: [PATCH 2/2] xeonphi tests added for rt

---
 examples_cuda/rt/rt.cpp   | 15 ++++++++++-----
 examples_cuda/tasksys.cpp |  2 ++
 2 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/examples_cuda/rt/rt.cpp b/examples_cuda/rt/rt.cpp
index 79221a07..d4a9ba51 100644
--- a/examples_cuda/rt/rt.cpp
+++ b/examples_cuda/rt/rt.cpp
@@ -181,7 +181,6 @@ int main(int argc, char *argv[]) {
   READ(nNodes, 1);
 
   LinearBVHNode *nodes = new LinearBVHNode[nNodes];
-#pragma omp parallel for
   for (unsigned int i = 0; i < nNodes; ++i) {
     // Each node is 6x floats for a boox, then an integer for an offset
     // to the second child node, then an integer that encodes the type
@@ -206,7 +205,6 @@ int main(int argc, char *argv[]) {
   uint nTris;
   READ(nTris, 1);
   Triangle *triangles = new Triangle[nTris];
-#pragma omp parallel for
   for (uint i = 0; i < nTris; ++i) {
     // 9x floats for the 3 vertices
     float v[9];
@@ -255,19 +253,26 @@ int main(int argc, char *argv[]) {
   //
   // Run 3 iterations with ispc + 1 core, record the minimum time
   //
+  fprintf(stderr, " --- 6 --- \n");
+  fflush(stderr);
   double minTimeISPCtasks = 1e30;
-  for (int i = 0; i < 3; ++i) {
-    reset_and_start_timer();
+  for (int i = 0; i < 3; ++i) 
+  {
+  //  reset_and_start_timer();
     const double t0 = rtc();
     raytrace_ispc_tasks(width, height, baseWidth, baseHeight, raster2camera,
         camera2world, image, id, nodes, triangles);
     double dt = rtc() - t0; //get_elapsed_mcycles();
     minTimeISPCtasks = std::min(dt, minTimeISPCtasks);
   }
-  printf("[rt ispc + tasks]:\t\t[%.3f] million cycles for %d x %d image\n", 
+  fprintf(stderr, "[rt ispc + tasks]:\t\t[%.3f] million cycles for %d x %d image\n", 
       minTimeISPCtasks, width, height);
+  fflush(stderr);
 
+  fprintf(stderr, " --- 7 --- \n");
+  fflush(stderr);
   writeImage(id, image, width, height, "rt-ispc-tasks.ppm");
+  return 0;
 
   memset(id, 0, width*height*sizeof(int));
   memset(image, 0, width*height*sizeof(float));
diff --git a/examples_cuda/tasksys.cpp b/examples_cuda/tasksys.cpp
index cccff55b..55cbccd5 100644
--- a/examples_cuda/tasksys.cpp
+++ b/examples_cuda/tasksys.cpp
@@ -59,7 +59,9 @@
 #define ISPC_USE_PTHREADS
 #define ISPC_USE_PTHREADS_FULLY_SUBSCRIBED
 #define ISPC_USE_CILK
+*/
 #define ISPC_USE_OMP
+/*
 #define ISPC_USE_TBB_TASK_GROUP
 #define ISPC_USE_TBB_PARALLEL_FOR