void ocl_alloc_buffers (sotl_device_t *dev)
{
device_create_buffers(dev);
if (sotl_verbose)
sotl_log(INFO, "%.2f MB of memory allocated for %d atoms on device [%s]\n",
dev->mem_allocated / (1024.0 * 1024.0),
dev->atom_set.natoms, dev->name);
}
void omp_init (sotl_device_t *dev)
{
#ifdef _SPHERE_MODE_
sotl_log(ERROR, "Sequential implementation does currently not support SPHERE_MODE\n");
exit (1);
#endif
borders_enabled = 1;
dev->compute = SOTL_COMPUTE_OMP; // dummy op to avoid warning
}
void ocl_init(sotl_device_t *dev)
{
cl_int err = 0;
// Create context
//
#ifdef HAVE_LIBGL
if(dev->display) {
#ifdef __APPLE__
CGLContextObj cgl_context = CGLGetCurrentContext ();
CGLShareGroupObj sharegroup = CGLGetShareGroup (cgl_context);
cl_context_properties properties[] = {
CL_CONTEXT_PROPERTY_USE_CGL_SHAREGROUP_APPLE,
(cl_context_properties) sharegroup,
0
};
#else
cl_context_properties properties[] = {
CL_GL_CONTEXT_KHR,
(cl_context_properties) glXGetCurrentContext (),
CL_GLX_DISPLAY_KHR,
(cl_context_properties) glXGetCurrentDisplay (),
CL_CONTEXT_PLATFORM, (cl_context_properties) dev->platform->id,
0
};
#endif
dev->context = clCreateContext (properties, 1, &dev->id, NULL, NULL, &err);
} else
#endif
{
dev->context = clCreateContext (0, 1, &dev->id, NULL, NULL, &err);
}
check (err, "Failed to create compute context \n");
// Load program source
//
const char *opencl_prog;
opencl_prog = file_get_contents(PROGRAM_NAME);
if (!opencl_prog) {
sotl_log(CRITICAL, "Failed to read contents of the OpenCL program '%s'.\n", PROGRAM_NAME);
}
// Build program
//
dev->program = clCreateProgramWithSource (dev->context, 1, &opencl_prog, NULL, &err);
check (err, "Failed to create program");
char options[OPENCL_PROG_MAX_STRING_SIZE_OPTIONS];
// Tile size of 32 works better on Xeon/Xeon Phi
//
if(dev->type != CL_DEVICE_TYPE_GPU)
dev->tile_size = 32;
else
dev->tile_size = TILE_SIZE;
#ifdef SLIDE
dev->slide_steps = SLIDE;
#else
dev->slide_steps = 1;
#endif
sprintf (options,
OPENCL_BUILD_OPTIONS
"-DTILE_SIZE=%d "
"-DSUBCELL=%d "
"-DDELTA_T=%.10f "
#ifdef SLIDE
"-DSLIDE=%d "
#endif
OPENCL_PROG_STRING_OPTIONS " -I "OCL_INCLUDE " "
LENNARD_STRING,
dev->tile_size,
SUBCELL,
1.0f,
#ifdef SLIDE
dev->slide_steps,
#endif
OPENCL_PROG_PARAM_OPTIONS,
LENNARD_PARAM);
// If not a GPU, do not use Tiling
//
if (dev->type != CL_DEVICE_TYPE_GPU)
strcat (options, " -DNO_LOCAL_MEM");
#ifdef TILE_CACHE
// On GPU, use a cache of TILES
//
if (dev->type == CL_DEVICE_TYPE_GPU)
strcat (options, " -DTILE_CACHE");
#endif
// Multi-accelerators configuration
//
if (sotl_have_multi())
strcat (options, " -DHAVE_MULTI");
#ifdef FORCE_N_UPDATE
if (!sotl_have_multi())
strcat (options, " -DFORCE_N_UPDATE");
#endif
#ifdef TORUS
strcat (options, " -DXY_TORUS -DZ_TORUS");
#endif
#ifdef XEON_VECTORIZATION
strcat (options, " -DXEON_VECTORIZATION");
#endif
#if defined(__APPLE__)
strcat(options, " -DHAVE_PRINTF");
#endif
if (sotl_verbose)
sotl_log(INFO, "--- Compiler flags ---\n%s\n----------------------\n",
options);
err = clBuildProgram (dev->program, 0, NULL, options, NULL, NULL);
if(sotl_verbose || err != CL_SUCCESS) {
size_t len;
// Display compiler error log
//
clGetProgramBuildInfo (dev->program, dev->id,
CL_PROGRAM_BUILD_LOG, 0, NULL, &len);
{
char buffer[len + 1];
clGetProgramBuildInfo (dev->program, dev->id,
CL_PROGRAM_BUILD_LOG,
sizeof (buffer), buffer, NULL);
sotl_log(INFO, "---- Compiler log ----\n%s\n----------------------\n",
buffer);
}
check (err, "Failed to build program");
}
// Create an OpenCL command queue
//
dev->queue = clCreateCommandQueue (dev->context, dev->id,
CL_QUEUE_PROFILING_ENABLE, &err);
check (err, "Failed to create a command queue");
cl_create_kernels(dev);
}