From 86652738c0b085f4ed06635e5d0a008ced520e4b Mon Sep 17 00:00:00 2001
From: Evghenii <egaburov@dds.nl>
Date: Thu, 14 Nov 2013 22:54:37 +0100
Subject: [PATCH] working on rt

---
 examples_cuda/rt/Makefile_gpu |  7 ++++---
 examples_cuda/rt/rt.ispc      | 20 +++++++-------------
 2 files changed, 11 insertions(+), 16 deletions(-)

diff --git a/examples_cuda/rt/Makefile_gpu b/examples_cuda/rt/Makefile_gpu
index 5c1aac0b..82d1d5d2 100644
--- a/examples_cuda/rt/Makefile_gpu
+++ b/examples_cuda/rt/Makefile_gpu
@@ -1,6 +1,6 @@
 PROG=rt_cu
 ISPC_SRC=rt1.ispc
-CXX_SRC=rt_cu.cpp
+CXX_SRC=rt_cu.cpp 
 
 CXX=g++
 CXXFLAGS=-O3 -I$(CUDATK)/include
@@ -8,7 +8,7 @@ LD=g++
 LDFLAGS=-lcuda
 
 ISPC=ispc
-ISPCFLAGS=-O3 --math-lib=default --target=nvptx64,avx
+ISPCFLAGS=-O3 --math-lib=default --target=nvptx64 --opt=fast-math
 
 LLVM32 = $(HOME)/usr/local/llvm/bin-3.2
 LLVM   = $(HOME)/usr/local/llvm/bin-3.3
@@ -27,6 +27,7 @@ CXX_OBJ=$(CXX_SRC:%.cpp=%.o)
 all: $(PROG)
 
 
+$(CXX_OBJ) : kernel.ptx
 $(PROG): $(CXX_OBJ) kernel.ptx
 	/bin/cp kernel.ptx __kernels.ptx
 	$(LD) -o $@ $(CXX_OBJ) $(LDFLAGS)
@@ -36,7 +37,7 @@ $(PROG): $(CXX_OBJ) kernel.ptx
 
 
 %_ispc_nvptx64.bc: %.ispc
-	$(ISPC) $(ISPCFLAGS) --emit-llvm -o `basename $< .ispc`_ispc.bc -h `basename $< .ispc`_ispc.h $< --emit-llvm
+	$(ISPC) $(ISPCFLAGS) --emit-llvm -o `basename $< .ispc`_ispc_nvptx64.bc -h `basename $< .ispc`_ispc.h $< --emit-llvm
 
 %.ptx: %.bc
 	$(LLVM32DIS) $<
diff --git a/examples_cuda/rt/rt.ispc b/examples_cuda/rt/rt.ispc
index 230a5288..fbd611fd 100644
--- a/examples_cuda/rt/rt.ispc
+++ b/examples_cuda/rt/rt.ispc
@@ -295,21 +295,15 @@ static inline void raytrace_tile(uniform int x0, uniform int x1,
     uniform float widthScale = (float)(baseWidth) / (float)(width);
     uniform float heightScale = (float)(baseHeight) / (float)(height);
 
-//    foreach_tiled (y = y0 ... y1, x = x0 ... x1) 
-    for (uniform int y = y0; y < y1; y++)
-      for (uniform int xb = x0; xb < x1; xb += programCount)
+    foreach_tiled (y = y0 ... y1, x = x0 ... x1) 
     {
-      const int x = xb + programIndex;
-        Ray ray;
-        generateRay(raster2camera, camera2world, x*widthScale, y*heightScale, ray);
-        BVHIntersect(nodes, triangles, ray);
+      Ray ray;
+      generateRay(raster2camera, camera2world, x*widthScale, y*heightScale, ray);
+      BVHIntersect(nodes, triangles, ray);
 
-        int offset = y * width + x;
-        if (x < x1)
-        {
-          image[offset] = ray.maxt;
-          id[offset] = ray.hitId;
-        }
+      int offset = y * width + x;
+      image[offset] = ray.maxt;
+      id[offset] = ray.hitId;
     }
 }