# HG changeset patch
# User John Tsiombikas <nuclear@member.fsf.org>
# Date 1283052042 -3600
# Node ID 8047637961a251e02cc1a41f374506409ca92d6a
# Parent  e7f79c6ad246256bfe3745cd40374715a12c00ec
fixed the issue of hitting maximum vertical image sizes for large kdtrees

diff -r e7f79c6ad246 -r 8047637961a2 Makefile
--- a/Makefile	Sat Aug 28 21:50:17 2010 +0100
+++ b/Makefile	Sun Aug 29 04:20:42 2010 +0100
@@ -3,15 +3,18 @@
 dep = $(obj:.o=.d)
 bin = test
 
+opt = -O3 -ffast-math
+dbg = -g
+
 CXX = g++
-CXXFLAGS = -pedantic -Wall -g $(def)
+CXXFLAGS = -pedantic -Wall $(dbg) $(opt) $(def)
 LDFLAGS = $(libgl) $(libcl) -lpthread
 
 ifeq ($(shell uname -s), Darwin)
 	libgl = -framework OpenGL -framework GLUT
 	libcl = -framework OpenCL
 else
-	libgl = -lGL -lglut
+	libgl = -lGL -lGLU -lglut
 	libcl = -lOpenCL
 	def = -DCLGL_INTEROP
 endif
diff -r e7f79c6ad246 -r 8047637961a2 rt.cl
--- a/rt.cl	Sat Aug 28 21:50:17 2010 +0100
+++ b/rt.cl	Sun Aug 29 04:20:42 2010 +0100
@@ -1,4 +1,5 @@
 /* vim: set ft=opencl:ts=4:sw=4 */
+#include "common.h"
 
 struct RendInfo {
 	float4 ambient;
@@ -58,7 +59,6 @@
 	float4 min, max;
 };
 
-#define MAX_NODE_FACES	32
 struct KDNode {
 	struct AABBox aabb;
 	int face_idx[MAX_NODE_FACES];
@@ -67,7 +67,6 @@
 	int padding;
 };
 
-#define RAY_MAG		500.0
 #define MIN_ENERGY	0.001
 #define EPSILON		1e-5
 
@@ -176,7 +175,7 @@
 	return dcol + scol;
 }
 
-#define STACK_SIZE	64
+#define STACK_SIZE	MAX_TREE_DEPTH
 bool find_intersection(struct Ray ray, const struct Scene *scn, struct SurfPoint *spres, read_only image2d_t kdimg)
 {
 	struct SurfPoint sp0;
@@ -369,20 +368,22 @@
 // read a KD-tree node from a texture scanline
 void read_kdnode(int idx, struct KDNode *node, read_only image2d_t kdimg)
 {
+	int startx = KDIMG_NODE_WIDTH * (idx / KDIMG_MAX_HEIGHT);
+
 	int2 tc;
-	tc.x = 0;
-	tc.y = idx;
+	tc.x = startx;
+	tc.y = idx % KDIMG_MAX_HEIGHT;
 
 	node->aabb.min = read_imagef(kdimg, kdsampler, tc); tc.x++;
 	node->aabb.max = read_imagef(kdimg, kdsampler, tc);
 
-	tc.x = 2 + MAX_NODE_FACES / 4;
+	tc.x = startx + 2 + MAX_NODE_FACES / 4;
 	float4 pix = read_imagef(kdimg, kdsampler, tc);
 	node->num_faces = (int)pix.x;
 	node->left = (int)pix.y;
 	node->right = (int)pix.z;
 
-	tc.x = 2;
+	tc.x = startx + 2;
 	for(int i=0; i<node->num_faces; i+=4) {
 		float4 pix = read_imagef(kdimg, kdsampler, tc); tc.x++;
 		node->face_idx[i] = (int)pix.x;
diff -r e7f79c6ad246 -r 8047637961a2 src/ocl.cc
--- a/src/ocl.cc	Sat Aug 28 21:50:17 2010 +0100
+++ b/src/ocl.cc	Sun Aug 29 04:20:42 2010 +0100
@@ -61,8 +61,6 @@
 		return false;
 	}
 
-
-
 #ifndef CLGL_INTEROP
 	cl_context_properties *prop = 0;
 #else
@@ -499,12 +497,9 @@
 	return num_args;
 }
 
-bool CLProgram::build()
+bool CLProgram::build(const char *opt)
 {
 	int err;
-
-	const char *opt = "-cl-mad-enable -cl-single-precision-constant -cl-fast-relaxed-math";
-
 	if((err = clBuildProgram(prog, 0, 0, opt, 0, 0)) != 0) {
 		size_t sz;
 		clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, 0, 0, &sz);
diff -r e7f79c6ad246 -r 8047637961a2 src/ocl.h
--- a/src/ocl.h	Sat Aug 28 21:50:17 2010 +0100
+++ b/src/ocl.h	Sun Aug 29 04:20:42 2010 +0100
@@ -103,7 +103,7 @@
 	CLMemBuffer *get_arg_buffer(int arg);
 	int get_num_args() const;
 
-	bool build();
+	bool build(const char *opt = 0);
 
 	bool run() const;
 	bool run(int dim, ...) const;
diff -r e7f79c6ad246 -r 8047637961a2 src/rt.cc
--- a/src/rt.cc	Sat Aug 28 21:50:17 2010 +0100
+++ b/src/rt.cc	Sun Aug 29 04:20:42 2010 +0100
@@ -6,6 +6,7 @@
 #include "ocl.h"
 #include "scene.h"
 #include "timer.h"
+#include "common.h"
 
 // kernel arguments
 enum {
@@ -122,7 +123,8 @@
 		return false;
 	}
 
-	if(!prog->build()) {
+	const char *opt = "-Isrc -cl-mad-enable -cl-single-precision-constant -cl-fast-relaxed-math";
+	if(!prog->build(opt)) {
 		return false;
 	}
 
@@ -291,29 +293,19 @@
 
 	float mag = sqrt(px * px + py * py + pz * pz);
 
-	px = px * 500.0 / mag;
-	py = py * 500.0 / mag;
-	pz = pz * 500.0 / mag;
+	px = px * RAY_MAG / mag;
+	py = py * RAY_MAG / mag;
+	pz = pz * RAY_MAG / mag;
 
 	Ray ray = {{0, 0, 0, 1}, {px, py, -pz, 1}};
 	return ray;
 }
 
-static int next_pow2(int x)
-{
-	x--;
-	x = (x >> 1) | x;
-	x = (x >> 2) | x;
-	x = (x >> 4) | x;
-	x = (x >> 8) | x;
-	x = (x >> 16) | x;
-	return x + 1;
-}
-
 static float *create_kdimage(const KDNodeGPU *kdtree, int num_nodes, int *xsz_ret, int *ysz_ret)
 {
-	int xsz = 16;
-	int ysz = next_pow2(num_nodes);
+	int ysz = MIN(num_nodes, KDIMG_MAX_HEIGHT);
+	int columns = (num_nodes - 1) / KDIMG_MAX_HEIGHT + 1;
+	int xsz = KDIMG_NODE_WIDTH * columns;
 
 	printf("creating kdtree image %dx%d (%d nodes)\n", xsz, ysz, num_nodes);
 
@@ -321,7 +313,10 @@
 	memset(img, 0, 4 * xsz * ysz * sizeof *img);
 
 	for(int i=0; i<num_nodes; i++) {
-		float *ptr = img + i * 4 * xsz;
+		int x = KDIMG_NODE_WIDTH * (i / KDIMG_MAX_HEIGHT);
+		int y = i % KDIMG_MAX_HEIGHT;
+
+		float *ptr = img + (y * xsz + x) * 4;
 
 		*ptr++ = kdtree[i].aabb.min[0];
 		*ptr++ = kdtree[i].aabb.min[1];
diff -r e7f79c6ad246 -r 8047637961a2 src/scene.h
--- a/src/scene.h	Sat Aug 28 21:50:17 2010 +0100
+++ b/src/scene.h	Sun Aug 29 04:20:42 2010 +0100
@@ -4,6 +4,7 @@
 #include <stdio.h>
 #include <vector>
 #include <list>
+#include "common.h"
 
 struct Vertex {
 	float pos[4];
@@ -51,7 +52,6 @@
 	KDNode();
 };
 
-#define MAX_NODE_FACES	32
 struct KDNodeGPU {
 	AABBox aabb;
 	int face_idx[MAX_NODE_FACES];