//! This function is documented in the header file
void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
{
if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
{
init_timings(nbv->gpu_nbv->timings);
}
}
static void stmp3xxxfb_enable_controller(struct stmp3xxx_fb_data *data)
{
struct stmp3xxx_platform_fb_entry *pentry = data->pdata->cur;
if (!data || !data->pdata || !data->pdata->cur)
return;
stmp3xxx_init_lcdif();
init_timings(data);
pentry->init_panel(data->dev, data->phys_start,
data->info.fix.smem_len, data->pdata->cur);
pentry->run_panel();
if (pentry->blank_panel)
pentry->blank_panel(FB_BLANK_UNBLANK);
}
//! This function is documented in the header file
void nbnxn_gpu_init(gmx_nbnxn_ocl_t **p_nb,
const gmx_device_info_t *deviceInfo,
const interaction_const_t *ic,
const NbnxnListParameters *listParams,
const nbnxn_atomdata_t *nbat,
int rank,
gmx_bool bLocalAndNonlocal)
{
gmx_nbnxn_ocl_t *nb;
cl_int cl_error;
cl_command_queue_properties queue_properties;
assert(ic);
if (p_nb == nullptr)
{
return;
}
snew(nb, 1);
snew(nb->atdat, 1);
snew(nb->nbparam, 1);
snew(nb->plist[eintLocal], 1);
if (bLocalAndNonlocal)
{
snew(nb->plist[eintNonlocal], 1);
}
nb->bUseTwoStreams = static_cast<cl_bool>(bLocalAndNonlocal);
nb->timers = new cl_timers_t();
snew(nb->timings, 1);
/* set device info, just point it to the right GPU among the detected ones */
nb->dev_info = deviceInfo;
snew(nb->dev_rundata, 1);
/* init nbst */
pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
init_plist(nb->plist[eintLocal]);
/* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
nb->bDoTime = static_cast<cl_bool>(getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
/* Create queues only after bDoTime has been initialized */
if (nb->bDoTime)
{
queue_properties = CL_QUEUE_PROFILING_ENABLE;
}
else
{
queue_properties = 0;
}
nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
/* local/non-local GPU streams */
nb->stream[eintLocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
if (CL_SUCCESS != cl_error)
{
gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
rank,
nb->dev_info->device_name,
cl_error);
}
if (nb->bUseTwoStreams)
{
init_plist(nb->plist[eintNonlocal]);
nb->stream[eintNonlocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
if (CL_SUCCESS != cl_error)
{
gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
rank,
nb->dev_info->device_name,
cl_error);
}
}
if (nb->bDoTime)
{
init_timers(nb->timers, nb->bUseTwoStreams == CL_TRUE);
init_timings(nb->timings);
}
nbnxn_ocl_init_const(nb, ic, listParams, nbat);
/* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
* TODO: decide about NVIDIA
*/
nb->bPrefetchLjParam =
(getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr) &&
((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
|| (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
/* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
* but sadly this is not supported in OpenCL (yet?). Consider adding it if
* it becomes supported.
*/
nbnxn_gpu_compile_kernels(nb);
nbnxn_gpu_init_kernels(nb);
/* clear energy and shift force outputs */
nbnxn_ocl_clear_e_fshift(nb);
*p_nb = nb;
if (debug)
{
fprintf(debug, "Initialized OpenCL data structures.\n");
}
}
