def get_launch_params(self, slvr, D):
polchans_per_block = D['BLOCKDIMX']
ants_per_block = D['BLOCKDIMY']
times_per_block = D['BLOCKDIMZ']
ntime, na, npolchan = slvr.dim_local_size('ntime', 'na', 'npolchan')
polchan_blocks = mbu.blocks_required(npolchan, polchans_per_block)
ant_blocks = mbu.blocks_required(na, ants_per_block)
time_blocks = mbu.blocks_required(ntime, times_per_block)
return {
'block' : (polchans_per_block, ants_per_block, times_per_block),
'grid' : (polchan_blocks, ant_blocks, time_blocks),
}
def get_kernel_params(self, solver):
ntime, na, nchan = solver.dim_local_size('ntime', 'na', 'nchan')
D = FLOAT_PARAMS if solver.is_float() else DOUBLE_PARAMS
chans_per_block = D['BLOCKDIMX'] if nchan > D['BLOCKDIMX'] else nchan
ants_per_block = D['BLOCKDIMY'] if na > D['BLOCKDIMY'] else na
times_per_block = D['BLOCKDIMZ'] if ntime > D['BLOCKDIMZ'] else ntime
chan_blocks = mbu.blocks_required(nchan, chans_per_block)
ant_blocks = mbu.blocks_required(na, ants_per_block)
time_blocks = mbu.blocks_required(ntime, times_per_block)
return {
'block' : (chans_per_block, ants_per_block, times_per_block),
'grid' : (chan_blocks, ant_blocks, time_blocks),
}
def get_kernel_params(self, solver, nx, ny):
slvr = solver
D = FLOAT_PARAMS if slvr.is_float() else DOUBLE_PARAMS
#x_per_block = D['BLOCKDIMX'] if nx > D['BLOCKDIMX'] else nx
#y_per_block = D['BLOCKDIMY'] if ny > D['BLOCKDIMX'] else ny
TILE_DIM = D['TILE_DIM']
x_per_block = D['BLOCKDIMX']
y_per_block = D['BLOCKDIMY']
x_blocks = mbu.blocks_required(nx, TILE_DIM)
y_blocks = mbu.blocks_required(ny, TILE_DIM)
return {
'block' : (x_per_block, y_per_block,1),
'grid' : (x_blocks, y_blocks, 1),
}
def get_kernel_params(self, solver):
slvr = solver
nsrc, ntime, nbl, nchan = slvr.dim_local_size('nsrc', 'ntime', 'nbl',
'nchan')
D = FLOAT_PARAMS if slvr.is_float() else DOUBLE_PARAMS
njones = nbl * nchan * ntime * nsrc
jones_per_block = D['BLOCKDIMX'] if njones > D['BLOCKDIMX'] else njones
jones_blocks = mbu.blocks_required(njones, jones_per_block)
return {
'block' : (jones_per_block,1,1), \
'grid' : (jones_blocks,1,1)
}
def get_kernel_params(self, solver):
slvr = solver
nsrc, ntime, nbl, nchan = slvr.dim_local_size(
'nsrc', 'ntime', 'nbl', 'nchan')
D = FLOAT_PARAMS if slvr.is_float() else DOUBLE_PARAMS
njones = nbl*nchan*ntime*nsrc
jones_per_block = D['BLOCKDIMX'] if njones > D['BLOCKDIMX'] else njones
jones_blocks = mbu.blocks_required(njones,jones_per_block)
return {
'block' : (jones_per_block,1,1), \
'grid' : (jones_blocks,1,1)
}
