/* 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 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);
}
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);
}
int main(int argc, const char* argv[]) {
icl_args* args = icl_init_args();
icl_parse_args(argc, argv, args);
chdir(PATH);
int size = args->size;
icl_print_args(args);
cl_float4* output = (cl_float4*)malloc(sizeof(cl_float4) * size);
icl_init_devices(args->device_type);
icl_start_energy_measurement();
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(args->device_id);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "sinewave.cl", "sinewave", "", ICL_SOURCE);
size_t szLocalWorkSize = args->local_size;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
for (int i = 0; i < args->loop_iteration; ++i) {
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(cl_float4) * size);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 2,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&size);
icl_read_buffer(buf_output, CL_TRUE, sizeof(cl_float4) * size, &output[0], NULL, NULL);
icl_release_buffer(buf_output);
}
icl_release_kernel(kernel);
}
icl_stop_energy_measurement();
// for the test check
printf("Result check: OK\n");
icl_release_args(args);
icl_release_devices();
free(output);
}
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");
}
int main(int argc, const char* argv[]) {
icl_args* args = icl_init_args();
icl_parse_args(argc, argv, args);
icl_print_args(args);
chdir(PATH);
int size = args->size;
int* input = (int*)malloc(sizeof(int) * size);
int* output = (int *)malloc(sizeof(int) * size);
for(int i=0; i < size; ++i) {
input[i] = i;
}
icl_init_devices(ICL_ALL);
icl_start_energy_measurement();
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(0);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "simple.cl", "simple", "", ICL_SOURCE);
size_t szLocalWorkSize = args->local_size;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
for (int i = 0; i < args->loop_iteration; ++i) {
icl_buffer* buf_input = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(int) * size);
icl_write_buffer(buf_input, CL_TRUE, sizeof(int) * size, &input[0], NULL, NULL);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 3,
(size_t)0, (void *)buf_input,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&size);
icl_read_buffer(buf_output, CL_TRUE, sizeof(int) * size, &output[0], NULL, NULL);
icl_release_buffers(2, buf_input, buf_output);
}
icl_release_kernel(kernel);
}
icl_stop_energy_measurement();
if (args->check_result) {
printf("======================\n= Simple program working\n");
unsigned int check = 1;
for(unsigned int i = 0; i < size; ++i) {
if(output[i] != input[i]) {
check = 0;
printf("= fail at %d, expected %d / actual %d", i, i, output[i]);
break;
}
}
printf("======================\n");
printf("Result check: %s\n", check ? "OK" : "FAIL");
} else {
printf("Result check: OK\n");
}
icl_release_args(args);
icl_release_devices();
free(input);
free(output);
}
int main(int argc, const char* argv[]) {
icl_args* args = icl_init_args();
icl_parse_args(argc, argv, args);
icl_print_args(args);
chdir(PATH);
int size = args->size;
float* input1 = (float*) malloc(sizeof(float) * size);
float* input2 = (float*) malloc(sizeof(float) * size);
float* alpha = (float*) malloc(sizeof(float) * size);
float* beta = (float*) malloc(sizeof(float) * size);
float* output = (float*) malloc(sizeof(float) * size);
fill_random_float(input2, size, 1, -1.0f, 1.0f);
qsort(input2, size, sizeof(float), float_compare);
float step = 2.0f / size;
for(int i=0; i < size; i++)
input1[i] = -1.0f + i * step;
fill_random_float(alpha, size, 1, -1.0f, 1.0f);
fill_random_float(beta, size, 1, -1.0f, 1.0f);
icl_init_devices(args->device_type);
icl_start_energy_measurement();
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(args->device_id);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "lin_reg.cl", "lin_reg", "", ICL_SOURCE);
size_t szLocalWorkSize = args->local_size;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
for (int i = 0; i < args->loop_iteration; ++i) {
icl_buffer* buf_input1 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(float) * size);
icl_buffer* buf_input2 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(float) * size);
icl_buffer* buf_alpha = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(float) * size);
icl_buffer* buf_beta = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(float) * size);
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(float) * size);
icl_write_buffer(buf_input1, CL_TRUE, sizeof(float) * size, &input1[0], NULL, NULL);
icl_write_buffer(buf_input2, CL_TRUE, sizeof(float) * size, &input2[0], NULL, NULL);
icl_write_buffer(buf_alpha, CL_TRUE, sizeof(float) * size, &alpha[0], NULL, NULL);
icl_write_buffer(buf_beta, CL_TRUE, sizeof(float) * size, &beta[0], NULL, NULL);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 6,
(size_t)0, (void *)buf_input1,
(size_t)0, (void *)buf_input2,
(size_t)0, (void *)buf_alpha,
(size_t)0, (void *)buf_beta,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&size);
icl_read_buffer(buf_output, CL_TRUE, sizeof(float) * size, &output[0], NULL, NULL);
icl_release_buffers(5, buf_input1, buf_input2, buf_alpha, buf_beta, buf_output);
}
icl_release_kernel(kernel);
}
icl_stop_energy_measurement();
if (args->check_result) {
printf("======================\n= Linear Regression Done\n");
float* output2 = (float *)malloc(sizeof(float) * size);
for(unsigned int j = 0; j < size; ++j) {
const int gid = j;
float a = alpha[gid];
float b = beta[gid];
float error = 0;
for(int i=0; i<size; i++) {
float e = (a * input1[i] + b) - input2[i];
error += e * e;
}
output2[gid] = error;
}
bool check = compare_float(output, output2, size, 0.000001);
printf("======================\n");
printf("Result check: %s\n", check ? "OK" : "FAIL");
free(output2);
} else {
printf("Result check: OK\n");
}
icl_release_args(args);
icl_release_devices();
free(input1);
free(input2);
free(alpha);
free(beta);
free(output);
}
int main(int argc, const char* argv[]) {
chdir(PATH);
icl_args* args = icl_init_args();
icl_parse_args(argc, argv, args);
icl_print_args(args);
int size = args->size;
cl_uint* ma = (cl_uint*) malloc(sizeof(cl_uint) * size);
cl_uint* b = (cl_uint*) malloc(sizeof(cl_uint) * size);
cl_uint* c = (cl_uint*) malloc(sizeof(cl_uint) * size);
cl_uint* seed = (cl_uint*) malloc(sizeof(cl_uint) * size);
cl_float4* result = (cl_float4*)malloc(sizeof(cl_float4) * size);
for (cl_uint i = 0; i < size; ++i) {
ma[i] = i; b[i] = i; c[i] = i; seed[i] = i;
}
icl_init_devices(args->device_type);
icl_start_energy_measurement();
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(args->device_id);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "mers_twister.cl", "mersenne_twister", "", ICL_SOURCE);
size_t szLocalWorkSize = args->local_size;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
for (int i = 0; i < args->loop_iteration; ++i) {
icl_buffer* buf_ma = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(cl_uint) * size);
icl_buffer* buf_b = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(cl_uint) * size);
icl_buffer* buf_c = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(cl_uint) * size);
icl_buffer* buf_seed = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(cl_uint) * size);
icl_buffer* buf_result = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(cl_float4) * size);
icl_write_buffer(buf_ma, CL_FALSE, sizeof(cl_uint) * size, &ma[0], NULL, NULL);
icl_write_buffer(buf_b, CL_FALSE, sizeof(cl_uint) * size, &b[0], NULL, NULL);
icl_write_buffer(buf_c, CL_FALSE, sizeof(cl_uint) * size, &c[0], NULL, NULL);
icl_write_buffer(buf_seed, CL_TRUE, sizeof(cl_uint) * size, &seed[0], NULL, NULL);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 6,
(size_t)0, (void *)buf_ma,
(size_t)0, (void *)buf_b,
(size_t)0, (void *)buf_c,
(size_t)0, (void *)buf_seed,
(size_t)0, (void *)buf_result,
sizeof(cl_int), (void *)&size
);
icl_read_buffer(buf_result, CL_TRUE, sizeof(cl_float4) * size, &result[0], NULL, NULL);
icl_release_buffers(5, buf_ma, buf_b, buf_c, buf_seed, buf_result);
}
icl_release_kernel(kernel);
}
icl_stop_energy_measurement();
if (args->check_result) {
printf("======================\n= mersenne twister test\n");
printf("Check not Implemented\n");
printf("Result check: OK\n");
} else {
printf("Result check: OK\n");
}
icl_release_devices();
free(ma);
free(b);
free(c);
free(seed);
free(result);
return 0;
}
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;
}
void scan(icl_buffer *data, icl_buffer *flag, UINT n) {
// use size for actual n, not overapproximation as in allocation
UINT numWorkGroups = ((n + wx - 1) / wx);
UINT sizeScanBuff = ((numWorkGroups + wx -1) / wx) * wx;
size_t gx = numWorkGroups * wx;
INT init = -1;
#if TIMING
clFinish(dev->queue);
icl_start_timer(timer);
#endif
icl_run_kernel(perBlockScanByKey, 1, &gx, &wx, NULL, perBlockScanEvent, 8,
(size_t)0, (void *)flag,
(size_t)0, (void *)data,
sizeof(INT), &init,
sizeof(UINT), &n,
sizeof(UINT) * wx, NULL,
sizeof(INT) * wx, NULL,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)preSumArray);
/*
V_OPENCL( kernels[0].setArg( 0, firstKey.getBuffer()), "Error setArg kernels[ 0 ]" ); // Input keys
V_OPENCL( kernels[0].setArg( 1, firstValue.getBuffer()),"Error setArg kernels[ 0 ]" ); // Input buffer
V_OPENCL( kernels[0].setArg( 2, result.getBuffer( ) ), "Error setArg kernels[ 0 ]" ); // Output buffer
V_OPENCL( kernels[0].setArg( 3, init ), "Error setArg kernels[ 0 ]" ); // Initial value exclusive
V_OPENCL( kernels[0].setArg( 4, numElements ), "Error setArg kernels[ 0 ]" ); // Size of scratch buffer
V_OPENCL( kernels[0].setArg( 5, ldsKeySize, NULL ), "Error setArg kernels[ 0 ]" ); // Scratch buffer
V_OPENCL( kernels[0].setArg( 6, ldsValueSize, NULL ), "Error setArg kernels[ 0 ]" ); // Scratch buffer
V_OPENCL( kernels[0].setArg( 7, *binaryPredicateBuffer),"Error setArg kernels[ 0 ]" ); // User provided functor
V_OPENCL( kernels[0].setArg( 8, *binaryFunctionBuffer ),"Error setArg kernels[ 0 ]" ); // User provided functor
V_OPENCL( kernels[0].setArg( 9, *keySumArray ), "Error setArg kernels[ 0 ]" ); // Output per block sum
V_OPENCL( kernels[0].setArg(10, *preSumArray ), "Error setArg kernels[ 0 ]" ); // Output per block sum
V_OPENCL( kernels[0].setArg(11, doExclusiveScan ), "Error setArg kernels[ 0 ]" ); // Exclusive scan?
*/
UINT workPerThread = sizeScanBuff / wx;
icl_run_kernel(intraBlockInclusiveScanByKey, 1, &wx, &wx, NULL, intraBlockEvent, 7,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)preSumArray,
(size_t)0, (void *)postSumArray,
sizeof(UINT), &numWorkGroups,
sizeof(UINT) * wx, NULL,
sizeof(INT) * wx, NULL,
sizeof(UINT), &workPerThread);
/*
INT* output = (UINT*)malloc(n * sizeof(INT));
icl_read_buffer(keySumArray, CL_TRUE, n * sizeof(INT), output, NULL, NULL);
for(int i = 0; i < n; ++i)
printf("%d ", output[i]);
printf("\n"); */
/*
V_OPENCL( kernels[1].setArg( 0, *keySumArray ), "Error setArg kernels[ 1 ]" ); // Input keys
V_OPENCL( kernels[1].setArg( 1, *preSumArray ), "Error setArg kernels[ 1 ]" ); // Input buffer
V_OPENCL( kernels[1].setArg( 2, *postSumArray ), "Error setArg kernels[ 1 ]" ); // Output buffer
V_OPENCL( kernels[1].setArg( 3, numWorkGroupsK0 ), "Error setArg kernels[ 1 ]" ); // Size of scratch buffer
V_OPENCL( kernels[1].setArg( 4, ldsKeySize, NULL ), "Error setArg kernels[ 1 ]" ); // Scratch buffer
V_OPENCL( kernels[1].setArg( 5, ldsValueSize, NULL ), "Error setArg kernels[ 1 ]" ); // Scratch buffer
V_OPENCL( kernels[1].setArg( 6, workPerThread ), "Error setArg kernels[ 1 ]" ); // User provided functor
V_OPENCL( kernels[1].setArg( 7, *binaryPredicateBuffer ),"Error setArg kernels[ 1 ]" ); // User provided functor
V_OPENCL( kernels[1].setArg( 8, *binaryFunctionBuffer ),"Error setArg kernels[ 1 ]" ); // User provided functor
*/
icl_run_kernel(perBlockAdditionByKey, 1, &gx, &wx, NULL, perBlockAdditionEvent, 5,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)postSumArray,
(size_t)0, (void *)flag,
(size_t)0, (void *)data,
sizeof(UINT), &n);
/*
V_OPENCL( kernels[2].setArg( 0, *keySumArray ), "Error setArg kernels[ 2 ]" ); // Input buffer
V_OPENCL( kernels[2].setArg( 1, *postSumArray ), "Error setArg kernels[ 2 ]" ); // Input buffer
V_OPENCL( kernels[2].setArg( 2, firstKey.getBuffer()), "Error setArg kernels[ 2 ]" ); // Output buffer
V_OPENCL( kernels[2].setArg( 3, result.getBuffer()), "Error setArg kernels[ 2 ]" ); // Output buffer
V_OPENCL( kernels[2].setArg( 4, numElements ), "Error setArg kernels[ 2 ]" ); // Size of scratch buffer
V_OPENCL( kernels[2].setArg( 5, *binaryPredicateBuffer ),"Error setArg kernels[ 2 ]" ); // User provided functor
V_OPENCL( kernels[2].setArg( 6, *binaryFunctionBuffer ),"Error setArg kernels[ 2 ]" ); // User provided functor
*/
#if TIMING
clFinish(dev->queue);
icl_stop_timer(timer);
perBlockScanTime += icl_profile_event(perBlockScanEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
intraBlockTime += icl_profile_event(intraBlockEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
perBlockAdditionTime += icl_profile_event(perBlockAdditionEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
#endif
}
int main(int argc, const char* argv[]) {
icl_args* args = icl_init_args();
icl_parse_args(argc, argv, args);
icl_print_args(args);
chdir(PATH);
// int dim = 128;
int size = args->size;
int nRef = 100000;
float* ref = (float*)malloc(sizeof(float) * nRef /* dim*/);
float* query = (float*)malloc(sizeof(float) * size /* dim*/);
float* dists = (float *)malloc(sizeof(float) * size);
int* neighbors = (int*)malloc(sizeof(int) * size);
srand(42);
for(int i=0; i < nRef/*dim*/; ++i) {
ref[i] = rand();
}
for(int i=0; i < size/**dim*/; ++i) {
query[i] = rand();
}
icl_init_devices(ICL_ALL);
icl_start_energy_measurement();
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(0);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "knn.cl", "knn", "", ICL_SOURCE);
size_t szLocalWorkSize = args->local_size;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
for (int i = 0; i < args->loop_iteration; ++i) {
icl_buffer* buf_ref = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(float) * nRef /* dim*/);
icl_buffer* buf_query = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(float) * size /* dim*/);
icl_buffer* buf_dists = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(float) * size);
icl_buffer* buf_neighbors = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(int) * size);
icl_write_buffer(buf_ref, CL_TRUE, sizeof(float) * nRef /*dim*/, &ref[0], NULL, NULL);
icl_write_buffer(buf_query, CL_TRUE, sizeof(float) * size /*dim*/, &query[0], NULL, NULL);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 6,
(size_t)0, (void *)buf_ref,
(size_t)0, (void *)buf_query,
(size_t)0, (void *)buf_dists,
(size_t)0, (void *)buf_neighbors,
sizeof(cl_int), (void *)&nRef,
sizeof(cl_int), (void *)&size);
// sizeof(cl_int), (void *)&dim);
icl_read_buffer(buf_dists, CL_TRUE, sizeof(float) * size, &dists[0], NULL, NULL);
icl_read_buffer(buf_neighbors, CL_TRUE, sizeof(int) * size, &neighbors[0], NULL, NULL);
icl_release_buffers(4, buf_ref, buf_query, buf_dists, buf_neighbors);
}
icl_release_kernel(kernel);
}
icl_stop_energy_measurement();
if (args->check_result) {
printf("======================\n= KNN program working\n");
unsigned int check = 1;
unsigned int sum = 0;
for(int i = 0; i < size; ++i) {
if(dists[i] < 0)
check = 0;
if(neighbors[i] < 0 || neighbors[i] >= nRef)
check = 0;
}
printf("======================\n");
printf("Result check: %s\n", check ? "OK" : "FAIL");
} else {
printf("Result check: OK\n");
}
icl_release_args(args);
icl_release_devices();
free(ref);
free(query);
free(dists);
free(neighbors);
}
int main(int argc, const char* argv[])
{
chdir(PATH);
icl_args* args = icl_init_args();
icl_parse_args(argc, argv, args);
icl_print_args(args);
int size = args->size;
// this is the size of chunking - so far as big as the local size
int chunkSize = 16;
cl_float16* input = (cl_float16*)malloc(sizeof(cl_float16) * size);
float* output = (float*)malloc(sizeof(float) * size);
fillrandom_float((float*)input,size, chunkSize, 0.001f ,100000.f);
icl_init_devices(args->device_type);
icl_start_energy_measurement();
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(args->device_id);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "reduction_chunking.cl", "reduce", "", ICL_SOURCE);
size_t szLocalWorkSize = args->local_size;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
for (int i = 0; i < args->loop_iteration; ++i) {
icl_buffer* buf_input = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(cl_float16) * size);
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(float) * size);
icl_write_buffer(buf_input, CL_FALSE, sizeof(cl_float16) * size, &input[0], NULL, NULL);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, NULL, NULL, 4,
(size_t)0, (void *)buf_input,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&chunkSize,
sizeof(cl_int), (void *)&size
);
icl_read_buffer(buf_output, CL_TRUE, sizeof(float) * size, &output[0], NULL, NULL);
icl_release_buffers(2, buf_input, buf_output);
}
icl_release_kernel(kernel);
}
icl_stop_energy_measurement();
// printf("Chunks' minimum \n");
// out_float_hbuffer(output, size);
if (args->check_result) {
printf("======================\n= Reduction test\n");
unsigned int check = 1;
float host_min = 100000.f;
float* testInput = (float*)input;
for(unsigned int i = 0; i < size*chunkSize; ++i)
if(testInput[i] < host_min) host_min = testInput[i];
printf("Host minimum is %f\n", host_min);
float device_min = 100000.f;
for(unsigned int i = 0; i < size; ++i)
if(output[i] < device_min) device_min = output[i];
printf("Device minimum is %f\n", device_min);
printf("Result check: %s\n", (device_min == host_min) ? "OK" : "FAIL");
} else {
printf("Result check: OK\n");
}
icl_release_devices();
free(input);
free(output);
#ifdef _MSC_VER
icl_prompt();
#endif
return 0;
}
/**************************************************************************
Function: ocl_jacobi
This routine contains the main iteration loop for the Jacobi iteration
using OpenCL kernel.
params:
a two arrays to compute solution into
max_iter maximum number of iterations
size size of array for this MPI rank
tolerance all differences should be les than this tolerance value
mpi_ranks number of MPI ranks in each dimension
rank_pos cartesian position of this rank
origin origin for this rank
d discretion size
mpi_comm MPI communications structure
local_workblock_size size of local workblock for OpenCL kernel
device_type OpenCL device type
full_copy boolean if full buffer copy is to be done
**************************************************************************/
static void ocl_jacobi(value_type *a[2],
unsigned int max_iter,
size_t size[DIMENSIONS],
value_type tolerance,
value_type d[DIMENSIONS],
size_t local_workblock_size[DIMENSIONS],
cl_device_type device_type,
unsigned int full_copy) {
size_t array_size;
unsigned int i, j, rc, iter = 0;
size_t delta_buffer_size, delta_size[DIMENSIONS];
size_t tile_delta_size, tile_cache_size;
value_type max_diff, timer;
icl_device* device_id;
icl_kernel* kernel;
cl_int err;
icl_buffer *a_buf[2], *delta_buf;
value_type *delta;
/* convenience for y stride in array */
cl_uint ystride = size[Y]+2*GHOST_CELL_WIDTH;
/* init devices */
icl_init_devices(device_type);
/* find OpenCL device */
device_id = icl_get_device(0);
/* build the kernel and verify the kernel */
kernel = icl_create_kernel(device_id, "jacsolver_kernel.cl", "ocl_jacobi_local_copy", "", ICL_SOURCE);
/* calculate size of kernel local memory - also used later for kernel params */
tile_delta_size = local_workblock_size[X] * local_workblock_size[Y];
tile_cache_size = (local_workblock_size[X]+2*GHOST_CELL_WIDTH) * (local_workblock_size[Y]+2*GHOST_CELL_WIDTH);
/* verify the device has enough resources for this device */
/* I'm an optimist, we just hope for the best
if ((cluGetAvailableLocalMem(device_id, kernel) < tile_delta_size + tile_cache_size) ||
(! cluCheckLocalWorkgroupSize(device_id, kernel, DIMENSIONS, local_workblock_size))) {
local_workblock_size[X] = 1;
local_workblock_size[Y] = 1;
}
*/
printf("Estimating solution using OpenCL Jacobi iteration with %d x %d workblock.\n", (int)local_workblock_size[X], (int)local_workblock_size[Y]);
fflush(stdout);
/* init arrays by setting the initial value and the boundary conditions */
set_initial_solution(a[OLD], size, INITIAL_GUESS);
set_initial_solution(a[NEW], size, INITIAL_GUESS);
set_boundary_conditions(a[OLD], size, d);
set_boundary_conditions(a[NEW], size, d);
/* print the initial solution guess */
print_array("Init ", a[NEW], size, d);
/* allocate memory for differences */
delta_size[X] = size[X] / local_workblock_size[X];
delta_size[Y] = size[Y] / local_workblock_size[Y];
delta_buffer_size = delta_size[X] * delta_size[Y];
delta = (value_type *)malloc(sizeof(value_type) * delta_buffer_size);
/* initialize deltas so that first execution of kernel with overlapping
* reduction on the host will work correctly and not prematurely exit
*/
for (i=0; i<delta_size[X]; ++i) {
for (j=0; j<delta_size[Y]; ++j) {
delta[i * delta_size[Y] + j] = 1.0;
}
}
/* create buffers for OpenCL device using host memory */
array_size = (size[X]+2*GHOST_CELL_WIDTH) * ystride;
a_buf[OLD] = icl_create_buffer(device_id, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sizeof(value_type) * array_size);
a_buf[NEW] = icl_create_buffer(device_id, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sizeof(value_type) * array_size);
delta_buf = icl_create_buffer(device_id, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sizeof(value_type) * delta_buffer_size);
/* copy over buffers to device */
icl_write_buffer(a_buf[OLD], CL_TRUE, sizeof(value_type) * array_size, a[OLD], NULL, NULL);
icl_write_buffer(a_buf[NEW], CL_TRUE, sizeof(value_type) * array_size, a[NEW], NULL, NULL);
/* set the kernel execution type - data parallel */
// cluSetKernelNDRange(clu, kernel, DIMENSIONS, NULL, size, local_workblock_size);
/* iterate until maximum difference is less than the given tolerance
or number of iterations is too high */
do {
/* swap array pointers for next iteration */
SWAP_PTR(a[OLD], a[NEW]);
SWAP_BUF(a_buf[OLD], a_buf[NEW]);
icl_run_kernel(kernel, DIMENSIONS, size, local_workblock_size, NULL, NULL, 6,
(size_t)0,(void *) a_buf[OLD],
(size_t)0, (void *) a_buf[NEW],
sizeof(value_type) * tile_delta_size, NULL,
sizeof(value_type) * tile_cache_size, NULL,
(size_t)0, (void *) delta_buf,
sizeof(cl_uint), (void *) &ystride);
/* while the kernel is running, calculate the reduction for the previous iteration */
max_diff = ocl_jacobi_reduce(delta, delta_size);
/* enqueue a synchronous copy of the delta. This will not occur until the kernel
* has finished. The deltas for each workgroup is a much smaller array to process
*/
icl_read_buffer(delta_buf, CL_TRUE, sizeof(value_type) * delta_buffer_size, delta, NULL, NULL);
// clEnqueueReadBuffer(queue, a_buf[NEW], CL_TRUE, 0, sizeof(value_type) * array_size, a[NEW], 0, NULL, NULL));
/* output status for user, overwrite the same line */
if ((0 == iter % 100)) {
printf("Iteration=%5d, max difference=%0.7f, target=%0.7f\r",
iter, max_diff, tolerance);
fflush(stdout);
}
/* increment the iteration counter */
iter++;
} while (max_diff > tolerance && max_iter >= iter); /* do loop */
/* read back the final result */
icl_read_buffer(a_buf[NEW], CL_TRUE, sizeof(value_type) * array_size, a[NEW], NULL, NULL);
/* output final iteration count and maximum difference value */
printf("Iteration=%5d, max difference=%0.7f, execution time=%.3f seconds\n", iter-1, max_diff, timer);
fflush(stdout);
/* finish usage of OpenCL device */
icl_release_buffers(3, a_buf[OLD], a_buf[NEW], delta_buf);
icl_release_kernel(kernel);
free(delta);
}
int main(int argc, char* argv[]) {
int size = 1000;
int* input1 = (int*)malloc(sizeof(int) * size);
int* input2 = (int*) malloc(sizeof(int) * size);
int* output = (int *)malloc(sizeof(int) * size);
for(int i=0; i < size; ++i) {
input1[i] = i;
input2[i] = 1;
}
#ifndef INSIEME
icl_timer* time1 = icl_init_timer(ICL_SEC);
icl_start_timer(time1);
#endif
icl_init_devices(ICL_CPU);
#ifndef INSIEME
printf("TIME for initialization: %f\n", icl_stop_timer(time1));
#endif
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(0);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "vec_mul.cl", "vec_mul", "", ICL_SOURCE);
icl_buffer* buf_input1 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_input2 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(int) * size);
icl_event* wb1 = icl_create_event();
icl_event* wb2 = icl_create_event();
icl_event* rb = icl_create_event();
icl_write_buffer(buf_input1, CL_FALSE, sizeof(int) * size, &input1[0], NULL, wb1);
icl_write_buffer(buf_input2, CL_FALSE, sizeof(int) * size, &input2[0], NULL, wb2);
size_t szLocalWorkSize = 256;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
icl_event* rk = icl_create_event();
icl_event* wb_all = icl_create_event_list(2, wb1, wb2);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, wb_all, rk, 4,
(size_t)0, (void *)buf_input1,
(size_t)0, (void *)buf_input2,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&size);
icl_read_buffer(buf_output, CL_TRUE, sizeof(int) * size, &output[0], rk, rb);
printf("Time wb1 %f\n", icl_profile_event(wb1, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time wb2 %f\n", icl_profile_event(wb2, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time rk %f\n", icl_profile_event(rk, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time rb %f\n", icl_profile_event(rb, ICL_STARTED, ICL_FINISHED, ICL_SEC));
icl_release_events(5, wb1, wb2, wb_all, rk, rb);
icl_release_buffers(3, buf_input1, buf_input2, buf_output);
icl_release_kernel(kernel);
}
#ifndef INSIEME
icl_restart_timer(time1);
#endif
icl_release_devices();
#ifndef INSIEME
printf("TIME for releasing the devices: %f\n", icl_stop_timer(time1));
icl_release_timer(time1);
#endif
// CHECK for output
printf("======================\n= Vector Mul Done\n");
unsigned int check = 1;
for(unsigned int i = 0; i < size; ++i) {
if(output[i] != i*size) {
check = 0;
printf("= fail at %d, expected %d / actual %d", i, i*3/2, output[i]);
break;
}
}
printf("= result check: %s\n======================\n", check ? "OK" : "FAIL");
free(input1);
free(input2);
free(output);
}