~prepare_main_thread() noexcept
{
BOOL result = ConvertFiberToThread();
HPX_ASSERT(FALSE != result);
HPX_UNUSED(result);
}
void CoCleanup()
{
ConvertFiberToThread();
}
int
clemuGPU :: scheduleNextJob(void)
{
int ret = CL_EMU_SUCCESS;
// TO DO : SELECT JOB
// SELECT SE, SIMD
clemuKernelJob *nextJob = m_job[0];
int cur_gbl_tid;
clemuGPU* dev = (clemuGPU*)(nextJob->GetDev());
nextJob->m_next_active_group = 0;
nextJob->m_fiber.m_FIBER_id = ConvertThreadToFiberEx(nextJob,FIBER_FLAG_FLOAT_SWITCH);
nextJob->m_fiber.m_FLS_index = 0;
// TO DO: NBR GROUPS > SIMDS
// run
for(int j = 0, grp_id = 0, wf_id = 0; j < nextJob->m_nbr_grpDim[1]; j++)
{
for(int i = 0; i < nextJob->m_nbr_grpDim[0]; i++, grp_id++, wf_id += nextJob->m_nbr_wavefronts)
{
int active_grp = nextJob->m_next_active_group;
// activate
if ( !nextJob->m_groups[active_grp].m_fiber.m_FIBER_id )
{
nextJob->m_groups[active_grp].m_fiber.m_FIBER_id =
CreateFiberEx( CLEMU_STACKCOMMITSIZE,
CLEMU_STACKRESERVESIZE,
FIBER_FLAG_FLOAT_SWITCH,
clemu_groupthread_proc,
&nextJob->m_groups[active_grp]);
assert(nextJob->m_groups[active_grp].m_fiber.m_FIBER_id);
nextJob->m_groups[active_grp].m_fiber.m_FLS_index = 0;
}
if ( !nextJob->m_groups[active_grp].m_td_fibers )
{
nextJob->m_groups[active_grp].m_td_fibers = new CL_EMU_FIBER[nextJob->GetNbrInstances()];
assert(nextJob->m_groups[active_grp].m_td_fibers);
memset(nextJob->m_groups[active_grp].m_td_fibers, 0, sizeof(CL_EMU_FIBER) * nextJob->GetNbrInstances());
}
for(int l = 0, tid = 0; l < nextJob->GetNbrWavefronts(); l++)
{
int wf_sz = nextJob->GetRealWFSz(l);
for(int m = 0; m < wf_sz; m++, tid++)
{
if ( !nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FIBER_id )
{
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FIBER_id =
CreateFiberEx( CLEMU_STACKCOMMITSIZE,
CLEMU_STACKRESERVESIZE,
FIBER_FLAG_FLOAT_SWITCH,
clemu_wavefrontthread_proc,
&nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m]);
assert(nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FIBER_id);
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FLS_index = 0; //FlsAlloc(0);
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_group_ptr = &nextJob->m_groups[active_grp];
}
}
}
nextJob->m_groups[active_grp].m_parent = nextJob;
nextJob->m_groups[active_grp].m_grpCoord[0] = i;
nextJob->m_groups[active_grp].m_grpCoord[1] = j;
nextJob->m_groups[active_grp].m_grpCoord[2] = 0;
nextJob->m_groups[active_grp].m_group_id = grp_id;
nextJob->m_groups[active_grp].m_se_id = 0;
nextJob->m_groups[active_grp].m_simd_id = active_grp % dev->GetNbrSIMD();
nextJob->m_groups[active_grp].m_first_hwwavefront = wf_id;
cur_gbl_tid = grp_id * nextJob->m_actual_nbr_instances;
nextJob->m_groups[active_grp].m_kernel = nextJob->GetKernel();
nextJob->m_groups[active_grp].m_endof_group = 0;
// local memory setup
// EXPLANATION:
// The size comes either from invokation (argument) or from teh kernel declaration.
// if it comes from declaration for teh first group there will be 0 sz which is max size.
// second group will det the correect size and will be separated from teh first group
int local_mem_sz = (nextJob->m_localmem_sz > clemuGetCompiledKernelMemSz(nextJob->m_groups[active_grp].m_kernel)) ? nextJob->m_localmem_sz : clemuGetCompiledKernelMemSz(nextJob->m_groups[active_grp].m_kernel);
nextJob->m_groups[active_grp].m_localmem_sz = local_mem_sz;
nextJob->m_groups[active_grp].m_localmem_ptr = (unsigned char*)dev->AssignLclMem(nextJob->m_groups[active_grp].m_simd_id,active_grp,local_mem_sz);
if ( nextJob->m_groups[active_grp].m_nmb_arg > 0 )
{
for( int l = 0; l < nextJob->m_groups[active_grp].m_nmb_arg; l++)
{
if ( (nextJob->m_groups[active_grp].m_args[l].m_flags & CL_ARG_LCL_MEM_SZ) == CL_ARG_LCL_MEM_SZ )
{
nextJob->m_groups[active_grp].m_args[l].m_flags |= CL_ARG_LDS_PTR;
nextJob->m_groups[active_grp].m_args[l].m_arg.parg = nextJob->m_groups[active_grp].m_localmem_ptr;
break;
}
}
}
// nextJob->m_groups[grp_id].m_wf_fibers = new CL_EMU_FIBER[nextJob->m_nbr_wavefronts];
// assert(nextJob->m_groups[grp_id].m_wf_fibers);
if ( !nextJob->m_groups[active_grp].m_td_fibers )
{
nextJob->m_groups[active_grp].m_td_fibers = new CL_EMU_FIBER[nextJob->m_actual_nbr_instances];
assert(nextJob->m_groups[active_grp].m_td_fibers);
}
for(int l = 0, tid = 0; l < nextJob->m_nbr_wavefronts; l++)
{
int wf_sz = nextJob->GetRealWFSz(l);
for(int m = 0; m < wf_sz; m++, tid++)
{
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_grp_tid = l * wf_sz + m;
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_gbl_tid = cur_gbl_tid + nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_grp_tid;
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_wfid = l;
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_wf_tid = m;
}
}
// schedule group
nextJob->m_curgroup = active_grp;
SwitchToFiber(nextJob->m_groups[active_grp].m_fiber.m_FIBER_id);
// free
if ( nextJob->m_groups[active_grp].m_td_fibers )
{
for(int l = 0, tid = 0; l < nextJob->GetNbrWavefronts(); l++)
{
int wf_sz = nextJob->GetRealWFSz(l);
for(int m = 0; m < wf_sz; m++, tid++)
{
if ( nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FIBER_id )
{
DeleteFiber(nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FIBER_id);
nextJob->m_groups[active_grp].m_td_fibers[l*wf_sz + m].m_FIBER_id = 0;
}
}
}
delete [] nextJob->m_groups[active_grp].m_td_fibers;
nextJob->m_groups[active_grp].m_td_fibers = 0;
}
if ( nextJob->m_groups[active_grp].m_fiber.m_FIBER_id )
{
DeleteFiber(nextJob->m_groups[active_grp].m_fiber.m_FIBER_id);
nextJob->m_groups[active_grp].m_fiber.m_FIBER_id = 0;
}
nextJob->m_next_active_group = (active_grp + 1) % nextJob->m_nbr_active_group;
}
}
nextJob->FreeJob();
ConvertFiberToThread();
nextJob->m_fiber.m_FIBER_id = 0;
return(ret);
}
