void InitializeAll(unsigned seed, double dev) {
SetSeed(seed);
SetDev(dev);
}
int
clemuKernelJob :: initJob(void *_device,
const char * _program_location,
const char * _program_nm,
const char*_kernel_entry_name,
int _domainDim,
int _domain[],
int _group[],
int _nmr_arg,
ClKrnlArg _args[])
{
int ret = CL_EMU_SUCCESS;
int actual_nbr_instances = 1;
clemuGPU *dev = (clemuGPU*)_device;
int nbr_wavefronts;
// TO AVOID MEM FAULT
SetDev(dev);
for( int i = 0; i < _domainDim; i++)
{
m_nbr_grpDim[i] = (_domain[i] + _group[i] - 1) / _group[i];
actual_nbr_instances *= _group[i];
}
m_nbr_groups = m_nbr_grpDim[0] * m_nbr_grpDim[1] * m_nbr_grpDim[2];
nbr_wavefronts = (actual_nbr_instances + dev->GetWFSz() - 1) / dev->GetWFSz();
m_nmr_dim = _domainDim;
m_grpDim[0] = m_grpDim[1] = m_grpDim[2] = 1;
m_domainDim[0] = m_domainDim[1] = m_domainDim[2] = 1;
for(int k = 0; k < _domainDim; k++)
{
m_grpDim[k] = _group[k];
m_domainDim[k] = m_nbr_grpDim[k] * _group[k];
}
m_actual_nbr_instances = actual_nbr_instances;
m_nbr_wavefronts = nbr_wavefronts;
strcpy_s(m_kernel_name,_MAX_KERNEL_NAME,_kernel_entry_name);
// BIND KERNEL
m_kernel = clemuFindCompiledKernel(_program_location,
_program_nm,
_kernel_entry_name);
m_args = 0;
m_nmb_arg = _nmr_arg;
m_localmem_sz = 0;
if ( m_nmb_arg > 0 )
{
m_args = new ClKrnlArg[m_nmb_arg];
assert(m_args);
for( int l = 0; l < m_nmb_arg; l++)
{
m_args[l] = _args[l];
m_args[l].arg_ptr = &m_args[l].m_arg.iarg;
if ( (m_args[l].m_flags & CL_ARG_LCL_MEM_SZ) == CL_ARG_LCL_MEM_SZ )
{
m_localmem_sz = m_args[l].m_len;
}
}
}
m_last_wavesz = actual_nbr_instances % dev->GetWFSz();
m_last_wavesz = (m_last_wavesz == 0 ) ? dev->GetWFSz() : m_last_wavesz;
assert(m_kernel);
m_nbr_active_group = dev->GetNbrSIMD() * dev->GetNbrGroupsPerSIMD();
if ( !m_groups )
{
m_groups = new clemuKernelGroup[m_nbr_active_group];
assert(m_groups);
memset(m_groups, 0, sizeof(clemuKernelGroup)*m_nbr_active_group);
}
for(int j = 0; j < m_nbr_active_group; j++)
{
m_groups[j].m_parent = this;
m_groups[j].m_kernel = m_kernel;
#if 0
if ( !m_groups[j].m_fiber.m_FIBER_id )
{
m_groups[j].m_fiber.m_FIBER_id =
CreateFiberEx( CLEMU_STACKCOMMITSIZE,
CLEMU_STACKRESERVESIZE,
FIBER_FLAG_FLOAT_SWITCH,
clemu_groupthread_proc,
&m_groups[j]);
assert(m_groups[j].m_fiber.m_FIBER_id);
m_groups[j].m_fiber.m_FLS_index = 0;
}
if ( !m_groups[j].m_td_fibers )
{
m_groups[j].m_td_fibers = new CL_EMU_FIBER[m_actual_nbr_instances];
assert(m_groups[j].m_td_fibers);
memset(m_groups[j].m_td_fibers, 0, sizeof(CL_EMU_FIBER) * m_actual_nbr_instances);
}
for(int l = 0, tid = 0; l < m_nbr_wavefronts; l++)
{
int wf_sz = GetRealWFSz(l);
for(int m = 0; m < wf_sz; m++, tid++)
{
if ( !m_groups[j].m_td_fibers[l*wf_sz + m].m_FIBER_id )
{
m_groups[j].m_td_fibers[l*wf_sz + m].m_FIBER_id =
CreateFiberEx( CLEMU_STACKCOMMITSIZE,
CLEMU_STACKRESERVESIZE,
FIBER_FLAG_FLOAT_SWITCH,
clemu_wavefrontthread_proc,
&m_groups[j].m_td_fibers[l*wf_sz + m]);
assert(m_groups[j].m_td_fibers[l*wf_sz + m].m_FIBER_id);
m_groups[j].m_td_fibers[l*wf_sz + m].m_FLS_index = 0; //FlsAlloc(0);
m_groups[j].m_td_fibers[l*wf_sz + m].m_group_ptr = &m_groups[j];
}
}
}
#endif
// replicate arguments
if ( !m_groups[j].m_args && m_nmb_arg > 0 )
{
m_groups[j].m_nmb_arg = m_nmb_arg;
m_groups[j].m_args = new ClKrnlArg[m_nmb_arg];
memset(m_groups[j].m_args, 0, sizeof(ClKrnlArg) * m_nmb_arg);
}
for( int l = 0; l < m_nmb_arg; l++)
{
memcpy(&m_groups[j].m_args[l], &m_args[l], sizeof(ClKrnlArg));
}
}
return(ret);
}