1.1 --- a/src/rt.cc	Sat Aug 28 21:50:17 2010 +0100
     1.2 +++ b/src/rt.cc	Sun Aug 29 04:20:42 2010 +0100
     1.3 @@ -6,6 +6,7 @@
     1.4  #include "ocl.h"
     1.5  #include "scene.h"
     1.6  #include "timer.h"
     1.7 +#include "common.h"
     1.8  
     1.9  // kernel arguments
    1.10  enum {
    1.11 @@ -122,7 +123,8 @@
    1.12  		return false;
    1.13  	}
    1.14  
    1.15 -	if(!prog->build()) {
    1.16 +	const char *opt = "-Isrc -cl-mad-enable -cl-single-precision-constant -cl-fast-relaxed-math";
    1.17 +	if(!prog->build(opt)) {
    1.18  		return false;
    1.19  	}
    1.20  
    1.21 @@ -291,29 +293,19 @@
    1.22  
    1.23  	float mag = sqrt(px * px + py * py + pz * pz);
    1.24  
    1.25 -	px = px * 500.0 / mag;
    1.26 -	py = py * 500.0 / mag;
    1.27 -	pz = pz * 500.0 / mag;
    1.28 +	px = px * RAY_MAG / mag;
    1.29 +	py = py * RAY_MAG / mag;
    1.30 +	pz = pz * RAY_MAG / mag;
    1.31  
    1.32  	Ray ray = {{0, 0, 0, 1}, {px, py, -pz, 1}};
    1.33  	return ray;
    1.34  }
    1.35  
    1.36 -static int next_pow2(int x)
    1.37 -{
    1.38 -	x--;
    1.39 -	x = (x >> 1) | x;
    1.40 -	x = (x >> 2) | x;
    1.41 -	x = (x >> 4) | x;
    1.42 -	x = (x >> 8) | x;
    1.43 -	x = (x >> 16) | x;
    1.44 -	return x + 1;
    1.45 -}
    1.46 -
    1.47  static float *create_kdimage(const KDNodeGPU *kdtree, int num_nodes, int *xsz_ret, int *ysz_ret)
    1.48  {
    1.49 -	int xsz = 16;
    1.50 -	int ysz = next_pow2(num_nodes);
    1.51 +	int ysz = MIN(num_nodes, KDIMG_MAX_HEIGHT);
    1.52 +	int columns = (num_nodes - 1) / KDIMG_MAX_HEIGHT + 1;
    1.53 +	int xsz = KDIMG_NODE_WIDTH * columns;
    1.54  
    1.55  	printf("creating kdtree image %dx%d (%d nodes)\n", xsz, ysz, num_nodes);
    1.56  
    1.57 @@ -321,7 +313,10 @@
    1.58  	memset(img, 0, 4 * xsz * ysz * sizeof *img);
    1.59  
    1.60  	for(int i=0; i<num_nodes; i++) {
    1.61 -		float *ptr = img + i * 4 * xsz;
    1.62 +		int x = KDIMG_NODE_WIDTH * (i / KDIMG_MAX_HEIGHT);
    1.63 +		int y = i % KDIMG_MAX_HEIGHT;
    1.64 +
    1.65 +		float *ptr = img + (y * xsz + x) * 4;
    1.66  
    1.67  		*ptr++ = kdtree[i].aabb.min[0];
    1.68  		*ptr++ = kdtree[i].aabb.min[1];