nuclear@0: #include <stdio.h>
nuclear@8: #include <string.h>
nuclear@2: #include <math.h>
nuclear@0: #include <assert.h>
John@14: #include "ogl.h"
nuclear@0: #include "ocl.h"
nuclear@22: #include "scene.h"
nuclear@32: #include "timer.h"
nuclear@0: 
nuclear@12: // kernel arguments
nuclear@12: enum {
nuclear@12: 	KARG_FRAMEBUFFER,
nuclear@12: 	KARG_RENDER_INFO,
nuclear@12: 	KARG_FACES,
nuclear@12: 	KARG_MATLIB,
nuclear@12: 	KARG_LIGHTS,
nuclear@12: 	KARG_PRIM_RAYS,
nuclear@12: 	KARG_XFORM,
John@14: 	KARG_INVTRANS_XFORM,
nuclear@28: 	KARG_KDTREE,
John@14: 
John@14: 	NUM_KERNEL_ARGS
nuclear@12: };
John@11: 
nuclear@2: struct RendInfo {
nuclear@22: 	float ambient[4];
nuclear@2: 	int xsz, ysz;
nuclear@9: 	int num_faces, num_lights;
nuclear@2: 	int max_iter;
nuclear@28: 	int kd_depth;
nuclear@12: };
nuclear@2: 
nuclear@1: struct Ray {
nuclear@8: 	float origin[4], dir[4];
nuclear@12: };
nuclear@1: 
nuclear@3: struct Light {
nuclear@8: 	float pos[4], color[4];
nuclear@12: };
nuclear@1: 
nuclear@3: static Ray get_primary_ray(int x, int y, int w, int h, float vfov_deg);
nuclear@3: 
nuclear@13: static Face *faces;
nuclear@3: static Ray *prim_rays;
nuclear@3: static CLProgram *prog;
nuclear@3: static int global_size;
nuclear@3: 
nuclear@4: static Light lightlist[] = {
nuclear@22: 	{{-8, 15, 18, 0}, {1, 1, 1, 1}}
nuclear@4: };
nuclear@4: 
nuclear@7: 
nuclear@4: static RendInfo rinf;
nuclear@4: 
nuclear@4: 
nuclear@13: bool init_renderer(int xsz, int ysz, Scene *scn)
nuclear@0: {
nuclear@4: 	// render info
nuclear@22: 	rinf.ambient[0] = rinf.ambient[1] = rinf.ambient[2] = 0.0;
nuclear@16: 	rinf.ambient[3] = 0.0;
nuclear@16: 
nuclear@4: 	rinf.xsz = xsz;
nuclear@4: 	rinf.ysz = ysz;
nuclear@13: 	rinf.num_faces = scn->get_num_faces();
nuclear@4: 	rinf.num_lights = sizeof lightlist / sizeof *lightlist;
nuclear@4: 	rinf.max_iter = 6;
nuclear@28: 	rinf.kd_depth = kdtree_depth(scn->kdtree);
nuclear@4: 
nuclear@3: 	/* calculate primary rays */
nuclear@3: 	prim_rays = new Ray[xsz * ysz];
nuclear@2: 
nuclear@2: 	for(int i=0; i<ysz; i++) {
nuclear@2: 		for(int j=0; j<xsz; j++) {
nuclear@2: 			prim_rays[i * xsz + j] = get_primary_ray(j, i, xsz, ysz, 45.0);
nuclear@2: 		}
nuclear@0: 	}
nuclear@0: 
nuclear@2: 	/* setup opencl */
nuclear@3: 	prog = new CLProgram("render");
nuclear@3: 	if(!prog->load("rt.cl")) {
nuclear@8: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@24: 	if(!(faces = (Face*)scn->get_face_buffer())) {
nuclear@13: 		fprintf(stderr, "failed to create face buffer\n");
nuclear@13: 		return false;
nuclear@13: 	}
nuclear@13: 
nuclear@28: 	const KDNodeGPU *kdbuf = scn->get_kdtree_buffer();
nuclear@28: 	if(!kdbuf) {
nuclear@28: 		fprintf(stderr, "failed to create kdtree buffer\n");
nuclear@28: 		return false;
nuclear@28: 	}
nuclear@32: 	// XXX now we can actually destroy the original kdtree and keep only the GPU version
nuclear@28: 
nuclear@3: 	/* setup argument buffers */
nuclear@12: 	prog->set_arg_buffer(KARG_FRAMEBUFFER, ARG_WR, xsz * ysz * 4 * sizeof(float));
nuclear@12: 	prog->set_arg_buffer(KARG_RENDER_INFO, ARG_RD, sizeof rinf, &rinf);
John@14: 	prog->set_arg_buffer(KARG_FACES, ARG_RD, rinf.num_faces * sizeof(Face), faces);
John@14: 	prog->set_arg_buffer(KARG_MATLIB, ARG_RD, scn->get_num_materials() * sizeof(Material), scn->get_materials());
nuclear@12: 	prog->set_arg_buffer(KARG_LIGHTS, ARG_RD, sizeof lightlist, lightlist);
nuclear@12: 	prog->set_arg_buffer(KARG_PRIM_RAYS, ARG_RD, xsz * ysz * sizeof *prim_rays, prim_rays);
nuclear@12: 	prog->set_arg_buffer(KARG_XFORM, ARG_RD, 16 * sizeof(float));
nuclear@12: 	prog->set_arg_buffer(KARG_INVTRANS_XFORM, ARG_RD, 16 * sizeof(float));
nuclear@29: 	prog->set_arg_buffer(KARG_KDTREE, ARG_RD, scn->get_kdtree_buffer_size(), kdbuf);
nuclear@12: 
John@14: 	if(prog->get_num_args() < NUM_KERNEL_ARGS) {
John@14: 		return false;
John@14: 	}
John@14: 
nuclear@16: 	if(!prog->build()) {
nuclear@16: 		return false;
nuclear@16: 	}
nuclear@16: 
nuclear@12: 	delete [] prim_rays;
nuclear@2: 
nuclear@3: 	global_size = xsz * ysz;
nuclear@3: 	return true;
nuclear@3: }
nuclear@3: 
nuclear@3: void destroy_renderer()
nuclear@3: {
nuclear@3: 	delete prog;
nuclear@3: }
nuclear@3: 
nuclear@3: bool render()
nuclear@3: {
nuclear@32: 	long tm0 = get_msec();
nuclear@32: 
nuclear@3: 	if(!prog->run(1, global_size)) {
nuclear@3: 		return false;
nuclear@0: 	}
John@15: 
nuclear@32: 	long tm_run = get_msec() - tm0;
nuclear@32: 
nuclear@13: 	CLMemBuffer *mbuf = prog->get_arg_buffer(KARG_FRAMEBUFFER);
nuclear@12: 	void *fb = map_mem_buffer(mbuf, MAP_RD);
nuclear@13: 	if(!fb) {
nuclear@13: 		fprintf(stderr, "FAILED\n");
nuclear@13: 		return false;
nuclear@13: 	}
nuclear@13: 
nuclear@12: 	glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, rinf.xsz, rinf.ysz, GL_RGBA, GL_FLOAT, fb);
nuclear@2: 	unmap_mem_buffer(mbuf);
nuclear@32: 
nuclear@32: 	long tm_upd = get_msec() - tm0 - tm_run;
nuclear@32: 
nuclear@32: 	printf("render %ld msec (%ld run, %ld upd)\n", tm_run + tm_upd, tm_run, tm_upd);
nuclear@3: 	return true;
nuclear@0: }
nuclear@2: 
nuclear@27: #define MIN(a, b)	((a) < (b) ? (a) : (b))
nuclear@21: static void dbg_set_gl_material(Material *mat)
nuclear@21: {
nuclear@21: 	static Material def_mat = {{0.7, 0.7, 0.7, 1}, {0, 0, 0, 0}, 0, 0, 0};
nuclear@21: 
nuclear@21: 	if(!mat) mat = &def_mat;
nuclear@21: 
nuclear@21: 	glMaterialfv(GL_FRONT_AND_BACK, GL_AMBIENT_AND_DIFFUSE, mat->kd);
nuclear@21: 	glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, mat->ks);
nuclear@27: 	glMaterialf(GL_FRONT_AND_BACK, GL_SHININESS, MIN(mat->spow, 128.0f));
nuclear@21: }
nuclear@21: 
nuclear@27: void dbg_render_gl(Scene *scn, bool show_tree, bool show_obj)
nuclear@8: {
nuclear@22: 	glPushAttrib(GL_ENABLE_BIT | GL_TRANSFORM_BIT | GL_LIGHTING_BIT);
nuclear@8: 
nuclear@21: 	for(int i=0; i<rinf.num_lights; i++) {
nuclear@21: 		float lpos[4];
nuclear@21: 
nuclear@21: 		memcpy(lpos, lightlist[i].pos, sizeof lpos);
nuclear@21: 		lpos[3] = 1.0;
nuclear@21: 
nuclear@21: 		glLightfv(GL_LIGHT0 + i, GL_POSITION, lpos);
nuclear@21: 		glLightfv(GL_LIGHT0 + i, GL_DIFFUSE, lightlist[i].color);
nuclear@22: 		glEnable(GL_LIGHT0 + i);
nuclear@21: 	}
nuclear@21: 
nuclear@12: 	glDisable(GL_TEXTURE_2D);
nuclear@12: 	glEnable(GL_DEPTH_TEST);
John@15: 	glEnable(GL_LIGHTING);
nuclear@12: 
nuclear@12: 	glMatrixMode(GL_PROJECTION);
nuclear@12: 	glPushMatrix();
nuclear@12: 	glLoadIdentity();
nuclear@12: 	gluPerspective(45.0, (float)rinf.xsz / (float)rinf.ysz, 0.5, 1000.0);
nuclear@12: 
nuclear@27: 	if(show_obj) {
nuclear@27: 		Material *materials = scn->get_materials();
John@14: 
nuclear@27: 		int num_faces = scn->get_num_faces();
nuclear@27: 		int cur_mat = -1;
nuclear@21: 
nuclear@27: 		for(int i=0; i<num_faces; i++) {
nuclear@27: 			if(faces[i].matid != cur_mat) {
nuclear@27: 				if(cur_mat != -1) {
nuclear@27: 					glEnd();
nuclear@27: 				}
nuclear@27: 				dbg_set_gl_material(materials ? materials + faces[i].matid : 0);
nuclear@27: 				cur_mat = faces[i].matid;
nuclear@27: 				glBegin(GL_TRIANGLES);
nuclear@21: 			}
nuclear@27: 
nuclear@27: 			for(int j=0; j<3; j++) {
nuclear@27: 				glNormal3fv(faces[i].v[j].normal);
nuclear@27: 				glVertex3fv(faces[i].v[j].pos);
nuclear@27: 			}
John@14: 		}
nuclear@27: 		glEnd();
nuclear@27: 	}
nuclear@12: 
nuclear@27: 	if(show_tree) {
nuclear@27: 		scn->draw_kdtree();
nuclear@12: 	}
nuclear@12: 
nuclear@12: 	glPopMatrix();
nuclear@12: 	glPopAttrib();
nuclear@22: 
nuclear@22: 	assert(glGetError() == GL_NO_ERROR);
nuclear@12: }
nuclear@12: 
nuclear@12: void set_xform(float *matrix, float *invtrans)
nuclear@12: {
nuclear@12: 	CLMemBuffer *mbuf_xform = prog->get_arg_buffer(KARG_XFORM);
nuclear@12: 	CLMemBuffer *mbuf_invtrans = prog->get_arg_buffer(KARG_INVTRANS_XFORM);
nuclear@12: 	assert(mbuf_xform && mbuf_invtrans);
nuclear@12: 
nuclear@12: 	float *mem = (float*)map_mem_buffer(mbuf_xform, MAP_WR);
nuclear@12: 	memcpy(mem, matrix, 16 * sizeof *mem);
nuclear@12: 	unmap_mem_buffer(mbuf_xform);
nuclear@12: 
nuclear@12: 	mem = (float*)map_mem_buffer(mbuf_invtrans, MAP_WR);
nuclear@12: 	memcpy(mem, invtrans, 16 * sizeof *mem);
nuclear@12: 	unmap_mem_buffer(mbuf_invtrans);
nuclear@8: }
nuclear@8: 
nuclear@3: static Ray get_primary_ray(int x, int y, int w, int h, float vfov_deg)
nuclear@2: {
nuclear@2: 	float vfov = M_PI * vfov_deg / 180.0;
nuclear@2: 	float aspect = (float)w / (float)h;
nuclear@2: 
nuclear@2: 	float ysz = 2.0;
nuclear@2: 	float xsz = aspect * ysz;
nuclear@2: 
nuclear@2: 	float px = ((float)x / (float)w) * xsz - xsz / 2.0;
nuclear@2: 	float py = 1.0 - ((float)y / (float)h) * ysz;
nuclear@2: 	float pz = 1.0 / tan(0.5 * vfov);
nuclear@2: 
nuclear@4: 	px *= 100.0;
nuclear@4: 	py *= 100.0;
nuclear@4: 	pz *= 100.0;
nuclear@2: 
nuclear@18: 	Ray ray = {{0, 0, 0, 1}, {px, py, -pz, 1}};
nuclear@2: 	return ray;
nuclear@2: }