void updateTree(icl_buffer *nodelist, icl_buffer *particlesD, icl_buffer *treeD, UINT nParticles, UINT treeHeight, icl_device* dev) {
// compile OpenCL kernels
icl_kernel* update = icl_create_kernel(dev, "kernel/updateTree.cl", "updateTree", KERNEL_BUILD_MACRO, ICL_SOURCE);
UINT nNodes = nParticles * 2 - 1;
#if timing == 1
icl_timer* timer_update = icl_init_timer(ICL_MILLI);
#endif
size_t localSizeB = 256;
size_t globalSizeB = ((nNodes + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_update);
#endif
for(int l = (int)treeHeight; l >= 0; --l)
{
icl_run_kernel(update, 1, &globalSizeB, &localSizeB, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)treeD,
(size_t)0, (void *)particlesD,
sizeof(int), &l,
sizeof(UINT), &nNodes);
}
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_update);
printf("tree update %f\n\n", timer_update->current_time);
#endif
}
void ocl_make_fieldbu_irregular(int nr_lines, int field_nr_points, int linelength, const cl_double4* vecs, const double* y, cl_double4* RF, double resolutionfactor)
{
icl_write_buffer(vecs_buf, CL_FALSE, sizeof(cl_double4) * nr_lines, &vecs[0], NULL, NULL);
icl_write_buffer(y_buf, CL_FALSE, sizeof(cl_double) * linelength*nr_lines+1, &y[0], NULL, NULL);
icl_write_buffer(RF_buf, CL_FALSE, sizeof(cl_double4) * field_nr_points, &RF[0], NULL, NULL);
size_t size = field_nr_points;
size_t szLocalWorkSize = LOCAL_GROUP_SIZE;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier) multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
printf("\n\nocl_make_field_irregular with blocking and unrolling (%d), global size %d, local size %d\n", LOOP_UNROLL, szGlobalWorkSize, szLocalWorkSize);
icl_run_kernel(irr_fieldbu_kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 6,
sizeof(cl_int), (void *)&nr_lines,
sizeof(cl_int), (void *)&linelength,
(size_t)0, (void *)vecs_buf,
(size_t)0, (void *)y_buf,
(size_t)0, (void *)RF_buf,
sizeof(cl_double), (void *)&resolutionfactor
);
icl_read_buffer(RF_buf, CL_TRUE, sizeof(cl_double4) * field_nr_points, &RF[0], NULL, NULL);
icl_finish(device);
}
/* unrool of all loops */
void ocl_make_field3(int nr_lines, int fielddim_x, int fielddim_y, int fielddim_z, int linelength,
const cl_double4* vecs, const double* y, double* RF,
double resolutionfactor)
{
icl_write_buffer(vecs_buf, CL_FALSE, sizeof(cl_double4) * nr_lines, &vecs[0], NULL, NULL);
icl_write_buffer(y_buf, CL_FALSE, sizeof(cl_double) * linelength*nr_lines+1, &y[0], NULL, NULL);
size_t szLocalWorkSize = LOCAL_GROUP_SIZE;
size_t size = nr_lines * fielddim_x * fielddim_y * fielddim_z;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier) multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
printf("\n\nocl_make_field3, global size (%d), local size (%d)\n", szGlobalWorkSize, szLocalWorkSize);
icl_run_kernel(reg_field3_kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 9,
sizeof(cl_int), (void *)&nr_lines,
sizeof(cl_int), (void *)&fielddim_x,
sizeof(cl_int), (void *)&fielddim_y,
sizeof(cl_int), (void *)&fielddim_z,
sizeof(cl_int), (void *)&linelength,
(size_t)0, (void *)vecs_buf,
(size_t)0, (void *)y_buf,
(size_t)0, (void *)RF_buf,
sizeof(cl_double), (void *)&resolutionfactor
);
icl_read_buffer(RF_buf, CL_TRUE, sizeof(cl_double) * fielddim_x * fielddim_y * fielddim_z, &RF[0], NULL, NULL);
icl_finish(device);
}
void ocl_make_field1bu(int nr_lines, int fielddim_x, int fielddim_y, int fielddim_z, int linelength,
const cl_double4* vecs, const double* y, double* RF,
double resolutionfactor)
{
icl_write_buffer(vecs_buf, CL_FALSE, sizeof(cl_double4) * nr_lines, &vecs[0], NULL, NULL);
icl_write_buffer(y_buf, CL_FALSE, sizeof(cl_double) * linelength*nr_lines+1, &y[0], NULL, NULL);
size_t szGlobalWorkSize[1] = { fielddim_x*fielddim_y*fielddim_z };
size_t szLocalWorkSize[1] = { LOCAL_GROUP_SIZE };
printf("\n\nocl_make_field1 with blocking and unrolling factor %d, global size (%d), local size (%d)\n", LOOP_UNROLL, szGlobalWorkSize[0], szLocalWorkSize[0]);
icl_run_kernel(reg_field1bu_kernel, 1, szGlobalWorkSize, szLocalWorkSize, NULL, NULL, 9,
sizeof(cl_int), (void *)&nr_lines,
sizeof(cl_int), (void *)&fielddim_x,
sizeof(cl_int), (void *)&fielddim_y,
sizeof(cl_int), (void *)&fielddim_z,
sizeof(cl_int), (void *)&linelength,
(size_t)0, (void *)vecs_buf,
(size_t)0, (void *)y_buf,
(size_t)0, (void *)RF_buf,
sizeof(cl_double), (void *)&resolutionfactor
);
icl_read_buffer(RF_buf, CL_TRUE, sizeof(cl_double) * fielddim_x * fielddim_y * fielddim_z, &RF[0], NULL, NULL);
icl_finish(device);
}
void drift(const FLOAT dt, icl_buffer* particles, const UINT nParticles, icl_device* dev)
{
icl_kernel* tw = icl_create_kernel(dev, "kernel/timestep.cl", "drift_particles", "-Iinclude", ICL_SOURCE);
size_t localSize = 256;
size_t globalSize = ((nParticles + 255) / 256) * 256;
icl_run_kernel(tw, 1, &globalSize, &localSize, NULL, NULL, 3,
(size_t)0, (void *)particles,
sizeof(UINT), &nParticles,
sizeof(FLOAT), &dt);
icl_finish(dev);
icl_release_kernel(tw);
}
void ocl_make_field_outcore(int nr_lines, int fielddim_x, int fielddim_y, int fielddim_z, int linelength,
const cl_double4* vecs, const double* y, double* RF,
double resolutionfactor)
{
size_t xy = fielddim_x*fielddim_y;
size_t z = fielddim_z;
size_t overall_size = xy * z;
size_t chunk_size = OUTOFCORE_SIZE;
icl_write_buffer(vecs_buf, CL_FALSE, sizeof(cl_double4) * nr_lines, &vecs[0], NULL, NULL);
icl_write_buffer(y_buf, CL_FALSE, sizeof(cl_double) * linelength*nr_lines+1, &y[0], NULL, NULL);
printf("ocl_make_field out-of-core with overall size of %d, chunked by %d\n", overall_size, chunk_size);
// for each chunk, we send and receive back a new part of the field
for(size_t offset = 0, bid = 0; offset < overall_size; offset += chunk_size, bid++)
{
printf("block %d offset %d,", bid, offset);
size_t current_chunk_size = MIN(overall_size-offset, chunk_size);
size_t szGlobalWorkSize[1] = { current_chunk_size };
size_t szLocalWorkSize[1] = { LOCAL_GROUP_SIZE };
unsigned long long in_offset = offset;
// printf("\nocl_make_field out-of-core, global size (%d), local size (%d)\n", szGlobalWorkSize[0], szLocalWorkSize[0]);
icl_run_kernel(reg_field_outofcore_kernel, 1, szGlobalWorkSize, szLocalWorkSize, NULL, NULL, 10,
sizeof(cl_int), (void*) &nr_lines,
sizeof(cl_int), (void*) &fielddim_x,
sizeof(cl_int), (void*) &fielddim_y,
sizeof(cl_int), (void*) &fielddim_z,
sizeof(cl_int), (void*) &linelength,
(size_t)0, (void*) vecs_buf,
(size_t)0, (void*) y_buf,
(size_t)0, (void*) RF_buf,
sizeof(cl_double), (void*) &resolutionfactor,
sizeof(cl_ulong), (void*) &in_offset
);
icl_read_buffer(&RF_buf[0], CL_FALSE, sizeof(cl_double) * current_chunk_size, &RF[offset], NULL, NULL);
}
icl_finish(device);
printf("\n");
}
UINT buildTree(icl_buffer *nodelist, icl_buffer *particlesD, icl_buffer *treeD, UINT nParticles, icl_device* dev)
{
UINT level = 1;
UINT nNodes = nParticles * 2 - 1;
icl_timer* timer = icl_init_timer(ICL_MILLI);
// void icl_start_timer(icl_timer* timer);
double time = 0;
// overapproximate size of temporal lists
/* struct Node** activelist = (struct Node**)malloc(nParticles * sizeof(struct Node*)); UINT activeN = 0;
struct Node** smalllist = (struct Node**)malloc(nParticles * sizeof(struct Node*)); UINT smallN = 0;
struct Node** nextlist = (struct Node**)malloc(nParticles * sizeof(struct Node*)); UINT nextN = 0;*/
icl_buffer* activelist = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(NodeId) * nParticles);
icl_buffer* smalllist = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(NodeId) * nParticles);
icl_buffer* nextlist = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(NodeId) * nParticles);
icl_buffer* sizes = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(UINT) * 5); // holds the current size of each of 3 buffers:
// 0 activelist
// 1 nodelist
// 2 smalllist
// 3 old max level
// 4 new max level
UINT maxNchunks = ((nParticles / fmin(T, chunk_size)) * 2) -1;
// assert(maxNchunks <= 256 && "adapt implementation"); // TODO allow more than 256 chunks per node
icl_buffer* chunks = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(struct Chunk) * maxNchunks);
icl_buffer* bboxes = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(struct BBox) * maxNchunks);
size_t localSize = 1;
size_t globalSize = 1;
/*
struct Particle* particles = (struct Particle*)malloc(3000 * sizeof(struct Particle));
icl_read_buffer(particlesD, CL_TRUE, sizeof(struct Particle) * 3000, &particles[0], NULL, NULL);
printf("%f %f %f\n", particles[0].pos.x, particles[0].pos.y, particles[0].pos.z);
*/
// compile OpenCL kernels
icl_kernel* init = icl_create_kernel(dev, "kernel/init.cl", "init", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* resetChunks = icl_create_kernel(dev, "kernel/init.cl", "memset_chunks", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* gp2c = icl_create_kernel(dev, "kernel/groupToChunks.cl", "groupParticlesIntoChunks", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* cBBox = icl_create_kernel(dev, "kernel/chunkedBBox.cl", "chunkedBBox", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* bBox = icl_create_kernel(dev, "kernel/bBox.cl", "bBox", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* sln = icl_create_kernel(dev, "kernel/splitLargeNodes.cl", "splitLargeNodes", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* sortP = icl_create_kernel(dev, "kernel/sortParticlesToChilds.cl", "sortParticlesToChilds", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* snf = icl_create_kernel(dev, "kernel/smallNodeFiltering.cl", "smallNodeFiltering", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* pnl = icl_create_kernel(dev, "kernel/packNextlist.cl", "packNextlist", KERNEL_BUILD_MACRO, ICL_SOURCE);
//////////////////////////////////////////////////////////////////////////
icl_kernel* preScan = icl_create_kernel(dev, "kernel/sortP_prescan.cl", "sortP_prescan_chunked", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* postScan = icl_create_kernel(dev, "kernel/sortP_postscan.cl", "sortP_postscan_chunked", KERNEL_BUILD_MACRO, ICL_SOURCE);
segmented_scan_init(nParticles, dev, KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* memset_int_s = icl_create_kernel(dev, "kernel/init.cl", "memset_int_s", KERNEL_BUILD_MACRO, ICL_SOURCE);
// approach with segmented scan
icl_buffer *scan_data = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(int) * nParticles);
icl_buffer *scan_flag = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(int) * nParticles);
icl_buffer* buffered_particles = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(struct Particle) * nParticles);
#if timing == 1
icl_timer* timer_gp2c = icl_init_timer(ICL_MILLI);
icl_timer* timer_cBBox = icl_init_timer(ICL_MILLI);
icl_timer* timer_bBox = icl_init_timer(ICL_MILLI);
icl_timer* timer_sln = icl_init_timer(ICL_MILLI);
icl_timer* timer_sortP = icl_init_timer(ICL_MILLI);
icl_timer* timer_snf = icl_init_timer(ICL_MILLI);
icl_timer* timer_pnl = icl_init_timer(ICL_MILLI);
icl_timer* timer_ran = icl_init_timer(ICL_MILLI);
icl_timer* timer_prescan = icl_init_timer(ICL_MILLI);
icl_timer* timer_scan = icl_init_timer(ICL_MILLI);
icl_timer* timer_postscan = icl_init_timer(ICL_MILLI);
#endif
//add root node to the activelist and initialize size lists
icl_run_kernel(init, 1, &globalSize, &localSize, NULL, NULL, 3,
(size_t)0, (void *)activelist,
(size_t)0, (void *)sizes,
(size_t)0, (void *)particlesD);
UINT activeN = 1;
icl_finish(dev);
// smallest power of 2 bigger or equal to maxxNchnunks
UINT pow2maxNchunks = pow2roundup(maxNchunks);
// processLargeNode
while(activeN != 0)
{
icl_start_timer(timer);
// group triangles into chunks
size_t localSize1 = min(pow2maxNchunks, 256);
#if timing == 1
icl_start_timer(timer_gp2c);
#endif
size_t globalSize1 = ((maxNchunks + localSize1 -1) / localSize1) * localSize1;
// reset chunks
icl_run_kernel(resetChunks, 1, &globalSize1, &localSize1, NULL, NULL, 2,
(size_t)0, (void *)chunks,
sizeof(UINT), &maxNchunks);
globalSize1 = localSize1 * activeN;
// split every node in chunk of chunk_size
icl_run_kernel(gp2c, 1, &globalSize1, &localSize1, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)chunks,
sizeof(UINT), &activeN);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_gp2c);
#endif
// compute per chunk bounding box
size_t localSize2 = chunk_size;
size_t globalSize2 = maxNchunks * chunk_size;
#if timing == 1
icl_start_timer(timer_cBBox);
#endif
icl_run_kernel(cBBox, 1, &globalSize2, &localSize2, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)chunks,
(size_t)0, (void *)bboxes);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_cBBox);
#endif
// compute per node bounding box
size_t localSize3 = min(pow2maxNchunks, 256);
size_t globalSize3 = localSize3 * activeN;
#if timing == 1
icl_start_timer(timer_bBox);
#endif
icl_run_kernel(bBox, 1, &globalSize3, &localSize3, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)bboxes,
sizeof(UINT), &activeN);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_bBox);
#endif
// split large nodes
size_t localSize4 = 256;
size_t globalSize4 = ((activeN + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_sln);
#endif
icl_run_kernel(sln, 1, &globalSize4, &localSize4, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)sizes,
sizeof(UINT), &activeN);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_sln);
#endif
///////////////////////////////////////////////////////////////////////////////
// XXx replaced with segmented scan
//globalSize = (activeN+1) * 256;
#if timing == 1
icl_start_timer(timer_sortP);
#endif
#if DEVICE == ICL_CPU
// sort particles to child nodes
size_t localSize5 = 256;
size_t globalSize5 = ((activeN + 255) / 256) * 256;
icl_run_kernel(sortP, 1, &globalSize5, &localSize5, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)activelist,
(size_t)0, (void *)nextlist,
sizeof(UINT), &activeN
);
#else
// init scan_flag to 1
cl_int initFlag = 1;
size_t np = (size_t)((nParticles + localSize4 -1 ) / localSize4) * localSize4;
icl_run_kernel(memset_int_s, 1, &np, &localSize4, NULL, NULL, 3,
(size_t)0, (void *)scan_flag,
sizeof(cl_int), &initFlag,
sizeof(UINT), &nParticles
);
#if timing == 1
icl_start_timer(timer_prescan);
#endif
// pre-scan fills data0 and data1 with 1 and 0 whenever value < pivot
localSize = chunk_size;
// globalSize = activeN * 256;
globalSize = maxNchunks * chunk_size;
icl_run_kernel(preScan, 1, &globalSize, &localSize, NULL, NULL, 7,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)chunks,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)activelist,
sizeof(UINT), &activeN,
(size_t)0, (void *)scan_data,
(size_t)0, (void *)scan_flag
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_prescan);
#endif
#if timing == 1
icl_start_timer(timer_scan);
#endif
// scan for
scan(scan_data, scan_flag, nParticles);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_scan);
#endif
// copy partially sorted data to the final
icl_copy_buffer(particlesD, buffered_particles, sizeof(struct Particle) * nParticles, NULL, NULL);
// swap(particlesD, buffered_particles);
#if timing == 1
icl_start_timer(timer_postscan);
#endif
localSize = chunk_size;
globalSize = maxNchunks * chunk_size;
icl_run_kernel(postScan, 1, &globalSize, &localSize, NULL, NULL, 7,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)chunks,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)activelist,
sizeof(UINT), &activeN,
(size_t)0, (void *)buffered_particles,
(size_t)0, (void *)scan_data
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_postscan);
#endif
icl_finish(dev);
#endif
///////////////////////////////////////////////////////////////////////////////
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_sortP);
#endif
// small node filtering
size_t localSize6 = 256;
size_t globalSize6 = ((activeN*2 + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_snf);
#endif
icl_run_kernel(snf, 1, &globalSize6, &localSize6, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)smalllist,
(size_t)0, (void *)sizes
//, sizeof(UINT), &nParticles
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_snf);
#endif
// packing of nextlist
size_t localSize7 = 1;
size_t globalSize7 = 1;
#if timing == 1
icl_start_timer(timer_pnl);
#endif
icl_run_kernel(pnl, 1, &globalSize7, &localSize7, NULL, NULL, 2,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)sizes
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_pnl);
#endif
// swap nextlist and activelist
swap(&nextlist, &activelist);
#if timing == 1
icl_start_timer(timer_ran);
#endif
// read size of next activelist set in kernel
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT), &activeN, NULL, NULL);
icl_finish(dev);
#if timing == 1
icl_stop_timer(timer_ran);
#endif
++level;
//printf("%d: ActiveN %d\n", level, activeN);
time = icl_stop_timer(timer);
}
icl_release_kernel(init);
icl_release_kernel(gp2c);
icl_release_kernel(cBBox);
icl_release_kernel(bBox);
icl_release_kernel(sln);
icl_release_kernel(sortP);
icl_release_kernel(snf);
icl_release_kernel(pnl);
//////////////////////////////////////////////////////////////////////////
icl_release_kernel(preScan);
icl_release_kernel(postScan);
segmented_scan_release();
icl_release_buffers(3, scan_data, scan_flag, buffered_particles);
#if timing == 1
printf("gp2c %f\ncBBox %f\nbBox %f\nsln %f\nsortP %f\nsnf %f\npnl %f\nran %f\n\n",
timer_gp2c->current_time,
timer_cBBox->current_time,
timer_bBox->current_time,
timer_sln->current_time,
timer_sortP->current_time,
timer_snf->current_time,
timer_pnl->current_time,
timer_ran->current_time);
icl_release_timer(timer_gp2c);
icl_release_timer(timer_cBBox);
icl_release_timer(timer_bBox);
icl_release_timer(timer_sln);
icl_release_timer(timer_sortP);
icl_release_timer(timer_snf);
icl_release_timer(timer_pnl);
printf("prescan %f\nscan %f\npostscan %f\n\n", timer_prescan->current_time, timer_scan->current_time, timer_postscan->current_time);
icl_release_timer(timer_prescan);
icl_release_timer(timer_scan);
icl_release_timer(timer_postscan);
#endif
/*
icl_read_buffer(nodelist, CL_TRUE, sizeof(struct Node) * 6000, tree->nodelist, NULL, NULL);
printf("node: %d, left %d, right %d", tree->nodelist[0].particlesHigh - tree->nodelist[0].particlesLow,
tree->nodelist[1].particlesHigh - tree->nodelist[1].particlesLow, tree->nodelist[2].particlesHigh - tree->nodelist[2].particlesLow);
printBox(tree->nodelist[49].bounding_box);
printBox(tree->nodelist[53].bounding_box);
printBox(tree->nodelist[54].bounding_box);
for(int i = 0; i < 6000; ++i)
if(tree->nodelist[i].bounding_box.box[0].x != 0.0)
printBox(tree->nodelist[i].bounding_box);
*/
//small nodes stage
// preprocessSmallNodes(smalllist);
icl_release_buffers(3, activelist, chunks, bboxes);
icl_kernel* sasl = icl_create_kernel(dev, "kernel/swapActiveAndSmalllist.cl", "swapActiveAndSmalllist", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* ssn = icl_create_kernel(dev, "kernel/splitSmallNodes.cl", "splitSmallNodes", KERNEL_BUILD_MACRO, ICL_SOURCE);
#if timing == 1
icl_timer* timer_ssn = icl_init_timer(ICL_MILLI);
icl_timer* timer_sasl = icl_init_timer(ICL_MILLI);
icl_timer* timer_rsn = icl_init_timer(ICL_MILLI);
#endif
size_t localSize8 = 1;
size_t globalSize8 = 1;
UINT setMaxLevel = 0;
icl_run_kernel(sasl, 1, &globalSize8, &localSize8, NULL, NULL, 2,
(size_t)0, (void *)sizes,
sizeof(UINT), &level);
// get number of small nodes
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT), &activeN, NULL, NULL);
while(activeN != 0)
{
icl_start_timer(timer);
// compute SVH and determine the split plane
size_t localSize9 = 256;
size_t globalSize9 = ((activeN + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_ssn);
#endif
icl_run_kernel(ssn, 1, &globalSize9, &localSize9, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)smalllist,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)sizes);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_ssn);
#endif
size_t localSizeA = 1;
size_t globalSizeA = 1;
#if timing == 1
icl_start_timer(timer_sasl);
#endif
icl_run_kernel(sasl, 1, &globalSizeA, &localSizeA, NULL, NULL, 2,
(size_t)0, (void *)sizes,
sizeof(UINT), &setMaxLevel);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_sasl);
#endif
swap(&nextlist, &smalllist);
// read size of next activelist set in kernel
#if timing == 1
icl_start_timer(timer_rsn);
#endif
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT), &activeN, NULL, NULL);
//printf("small size %d\n", activeN);
icl_finish(dev);
#if timing == 1
icl_stop_timer(timer_rsn);
#endif
time = icl_stop_timer(timer);
}
icl_release_buffer(smalllist);
icl_release_buffer(nextlist);
icl_release_kernel(sasl);
icl_release_kernel(ssn);
#if timing == 1
printf("ssn %f\nsasl %f\nrsn %f\n\n", timer_ssn->current_time, timer_sasl->current_time, timer_rsn->current_time);
icl_release_timer(timer_ssn);
icl_release_timer(timer_sasl);
icl_release_timer(timer_rsn);
#endif
UINT s[5];
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT) * 5, &s, NULL, NULL);
icl_release_buffer(sizes);
icl_kernel* upPass = icl_create_kernel(dev, "kernel/upPass.cl", "upPass", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* downPass = icl_create_kernel(dev, "kernel/kdDownPass.cl", "downPass", KERNEL_BUILD_MACRO, ICL_SOURCE);
#if timing == 1
icl_timer* timer_upPass = icl_init_timer(ICL_MILLI);
icl_timer* timer_downPass = icl_init_timer(ICL_MILLI);
icl_timer* timer_rt = icl_init_timer(ICL_MILLI);
#endif
UINT treeHeight = s[4];
printf("Tree height: %d\n", treeHeight);
size_t localSizeB = 256;
size_t globalSizeB = ((nNodes + 255) / 256) * 256;
icl_start_timer(timer);
#if timing == 1
icl_start_timer(timer_upPass);
#endif
for(int l = (int)treeHeight; l >= 0; --l)
{
icl_run_kernel(upPass, 1, &globalSizeB, &localSizeB, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)particlesD,
sizeof(int), &l,
sizeof(UINT), &nNodes);
}
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_upPass);
icl_start_timer(timer_downPass);
#endif
for(UINT l = 0; l <= treeHeight; ++l)
{
icl_run_kernel(downPass, 1, &globalSizeB, &localSizeB, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)treeD,
sizeof(UINT), &l,
sizeof(UINT), &nNodes);
}
icl_finish(dev);
#if timing == 1
icl_stop_timer(timer_downPass);
#endif
time = icl_stop_timer(timer);
icl_release_kernel(upPass);
icl_release_kernel(downPass);
#if timing == 1
icl_start_timer(timer_rt);
#endif
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_rt);
printf("upPass %f\ndownPass %f\nread Tree %f\n\n", timer_upPass->current_time, timer_downPass->current_time, timer_rt->current_time);
icl_release_timer(timer_upPass);
icl_release_timer(timer_downPass);
icl_release_timer(timer_rt);
#endif
// struct Node* kdTree = (struct Node*)malloc(sizeof(struct Node) * nNodes);
// icl_read_buffer(treeD, CL_TRUE, sizeof(struct Node) * nNodes, kdTree, NULL, NULL);
// printf("%d", tree->nodelist[0].left_child);
printf("\nTime: %f\n", time);
icl_release_timer(timer);
return treeHeight;
}
