Switch from texture to float3 array for upper and lower bounds. 10% speed boost!

1 files changed, 7 insertions, 5 deletions

diff --git a/src/mesh.h b/src/mesh.h
index b4714c4..bb30bef 100644
--- a/src/mesh.h
+++ b/src/mesh.h
@@ -13,8 +13,8 @@ __device__ unsigned int g_start_node;
 __device__ unsigned int g_first_node;
 
 /* lower/upper bounds for the bounding box associated with each node/leaf */
-texture<float4, 1, cudaReadModeElementType> upper_bounds;
-texture<float4, 1, cudaReadModeElementType> lower_bounds;
+__device__ float3 *g_lower_bounds;
+__device__ float3 *g_upper_bounds;
 
 /* map to child node/triangle indices */
 texture<unsigned int, 1, cudaReadModeElementType> node_map;
@@ -38,8 +38,8 @@ __device__ int convert(int c)
    intersects the bounding box return true, else return false. */
 __device__ bool intersect_node(const float3 &origin, const float3 &direction, const int &i)
 {
-	float3 lower_bound = make_float3(tex1Dfetch(lower_bounds, i));
-	float3 upper_bound = make_float3(tex1Dfetch(upper_bounds, i));
+	float3 lower_bound = g_lower_bounds[i];
+	float3 upper_bound = g_upper_bounds[i];
 
 	return intersect_box(origin, direction, lower_bound, upper_bound);
 }
@@ -134,13 +134,15 @@ __device__ int intersect_mesh(const float3 &origin, const float3& direction, flo
 extern "C"
 {
 
-__global__ void set_global_mesh_variables(uint4 *triangles, float3 *vertices, unsigned int *colors, unsigned int start_node, unsigned int first_node)
+  __global__ void set_global_mesh_variables(uint4 *triangles, float3 *vertices, unsigned int *colors, unsigned int start_node, unsigned int first_node, float3 *lower_bounds, float3 *upper_bounds)
 {
 	g_triangles = triangles;
 	g_vertices = vertices;
 	g_colors = colors;
 	g_start_node = start_node;
 	g_first_node = first_node;
+	g_lower_bounds = lower_bounds;
+	g_upper_bounds = upper_bounds;
 }
 
 __global__ void set_colors(unsigned int *colors)