int main(int argc, char **argv){
unsigned long hits = 0, tries = 0;
int errs = 0;
threefry2x64_ctr_t c = {{0}}, r;
threefry2x64_key_t k = {{R123_64BIT(0xdeadbeef12345678)}};
threefry2x32_ctr_t ch = {{0}}, rh;
threefry2x32_key_t kh = {{0xdecafbad}};
(void)argc; (void)argv; /* unused */
printf("%lu uniform doubles from threefry2x64\n", NTRIES);
while (tries < NTRIES) {
double x, y;
c.v[0]++; /* increment the counter */
r = threefry2x64(c, k);
x = 2.*u01_open_open_64_53(r.v[0]) - 1.;
y = 2.*u01_open_open_64_53(r.v[1]) - 1.;
if( x*x + y*y < 1.0 )
hits++;
tries++;
}
errs += pi_check(hits, tries);
printf("%lu uniform doubles from threefry2x32\n", NTRIES);
hits = tries = 0;
while (tries < NTRIES) {
double x, y;
ch.v[0]++; /* increment the counter */
rh = threefry2x32(ch, kh);
x = 2.*u01_open_open_32_53(rh.v[0]) - 1.;
y = 2.*u01_open_open_32_53(rh.v[1]) - 1.;
if( x*x + y*y < 1.0 )
hits++;
tries++;
}
errs += pi_check(hits, tries);
printf("%lu uniform floats from threefry2x32\n", NTRIES);
hits = tries = 0;
while (tries < NTRIES) {
float x, y;
ch.v[0]++; /* increment the counter */
rh = threefry2x32(ch, kh);
x = 2.f*u01_open_open_32_24(rh.v[0]) - 1.f;
y = 2.f*u01_open_open_32_24(rh.v[1]) - 1.f;
if( x*x + y*y < 1.0 )
hits++;
tries++;
}
errs += pi_check(hits, tries);
return errs;
}
int main(int argc, char **argv){
unsigned long hits = 0, tries = 0;
const int64_t two_to_the_62 = ((int64_t)1)<<62;
threefry2x64_key_t key = {{0, 0}};
threefry2x64_ctr_t ctr = {{0, 0}};
enum { int32s_per_counter = sizeof(ctr)/sizeof(int32_t) };
(void)argc;(void)argv; /* unused */
printf("Throwing %lu darts at a square board using threefry2x64\n", NTRIES);
/* make the most of each bijection by looping over as many
int32_t's as we can find in the ctr_type. */
assert( int32s_per_counter%2 == 0 );
while(tries < NTRIES){
/* Use a union to avoid strict aliasing issues. */
union{
threefry2x64_ctr_t ct;
int32_t i32[int32s_per_counter];
}u;
size_t j;
/* Don't worry about the 'carry'. We're not going to loop
more than 2^64 times. */
ctr.v[0]++;
u.ct = threefry2x64(ctr, key);
for(j=0; j<int32s_per_counter; j+=2){
int64_t x = u.i32[j];
int64_t y = u.i32[j+1];
if( (x*x + y*y) < two_to_the_62 )
hits++;
tries++;
}
}
return pi_check(hits, tries);
}
int main(int argc, char **argv)
{
const char *kernelname = "counthits";
unsigned count =10000;
cl_int err;
cl_context cl_context;
cl_program program;
cl_kernel cl_kernel;
cl_mem cl_out;
cl_command_queue cl_queue;
size_t i, nthreads, hits_sz;
size_t cores, work_group_size;
cl_uint2 * hits_host;
double d = 0.; // timer
d = timer(&d);
progname = argv[0];
CHECK(cl::Platform::get(&platformList));
CHECKERR( cl_context = createCLContext(CL_DEVICE_TYPE_GPU,cl_vendor::VENDOR_AMD, &err) );
std::vector<cl::Device> devices;
CHECKERR( devices = cl_context.getInfo<CL_CONTEXT_DEVICES>(&err) );
size_t length = 0;
const char * sourceStr = loadFileToString("pi_opencl_kernel.ocl","",&length);
cl::Program::Sources sources(1, std::make_pair(sourceStr, length));
program = cl::Program(cl_context, sources);
CHECK( program.build(devices,"-I ..\\include") );
CHECKERR(work_group_size = devices[0].getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>(&err) );
CHECKERR(cores = devices[0].getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>(&err) );
cores *= 16*4; //Tahiti.
if (work_group_size > 64) work_group_size /= 2;
nthreads = cores * work_group_size*32; //2048*128 = 262144
if (count == 0)
count = NTRIES/nthreads; //38
printf("Count: %lu\n",count);
hits_sz = nthreads * sizeof(hits_host[0]);//2097152
CHECKNOTZERO(hits_host = (cl_uint2 *)malloc(hits_sz));
CHECKERR ( cl_out = cl::Buffer( cl_context, CL_MEM_WRITE_ONLY | CL_MEM_USE_HOST_PTR, hits_sz, hits_host, &err));
CHECKERR ( cl_kernel = cl::Kernel(program,kernelname,&err) );
CHECK ( cl_kernel.setArg( 0, count) );
CHECK ( cl_kernel.setArg( 1, cl_out) );
CHECKERR (cl_queue = cl::CommandQueue(cl_context, devices[0], 0, &err) );
cl::Event event;
CHECK( cl_queue.enqueueNDRangeKernel(cl_kernel,cl::NullRange,cl::NDRange(nthreads), cl::NDRange(work_group_size), NULL, &event) );
event.wait();
CHECK( cl_queue.enqueueReadBuffer(cl_out, CL_TRUE, 0,hits_sz, hits_host) );
unsigned long hits = 0, tries = 0;
for (i = 0; i < nthreads; i++) {
#ifdef _DEBUG
printf("%lu %u %u\n", (unsigned long)i, hits_host[i].s[0], hits_host[i].s[1]);
#endif
hits += hits_host[i].s[0];
tries += hits_host[i].s[1];
}
return pi_check(hits, tries);
}
