def __init__(s, nxs, ny, nz, target_device='all', print_device_info=True):
s.gpu_devices = utils.get_gpu_devices()
if print_device_info:
utils.print_gpu_info(s.gpu_devices)
utils.print_cpu_info()
ngpu_dev = len(s.gpu_devices)
s.context, s.queues = utils.create_context_queues(s.gpu_devices)
s.ngpu = len(s.gpu_devices)
s.Ls = 256
if target_device == int:
s.Gs = utils.get_optimal_global_work_size(
s.gpu_devices[target_device])
else:
s.Gs = utils.get_optimal_global_work_size(s.gpu_devices[0])
if type(nxs) == list:
if len(nxs) == s.ngpu:
s.nxs = nxs
s.nx_gpu = np.array(nxs).sum()
else:
print(
'Error: len(nxs) %d is not matched with the number of target devices %d.'
% (len(nxs), s.ngpu))
sys.exit()
elif type(nxs) == int:
if nxs % s.ngpu == 0:
s.nxs = [nxs / s.ngpu for i in xrange(s.ngpu)]
s.nx_gpu = nxs
else:
print(
'Error: nxs %d is not multiple of the number of target devices %d.'
% (nxs, s.ngpu))
sys.exit()
else:
print('Error: nxs type %s is invalid.' % type(nxs))
sys.exit()
s.ny, s.nz = ny, nz
s.check_grid_size()
s.allocations()
s.get_program(print_source=False)
def __init__(s, nxs, ny, nz, target_device='all', print_device_info=True):
s.gpu_devices = utils.get_gpu_devices()
if print_device_info:
utils.print_gpu_info(s.gpu_devices)
utils.print_cpu_info()
ngpu_dev = len(s.gpu_devices)
s.context, s.queues = utils.create_context_queues(s.gpu_devices)
s.ngpu = len(s.gpu_devices)
s.Ls = 256
if target_device == int:
s.Gs = utils.get_optimal_global_work_size(s.gpu_devices[target_device])
else:
s.Gs = utils.get_optimal_global_work_size(s.gpu_devices[0])
if type(nxs) == list:
if len(nxs) == s.ngpu:
s.nxs = nxs
s.nx_gpu = np.array(nxs).sum()
else:
print('Error: len(nxs) %d is not matched with the number of target devices %d.' %(len(nxs), s.ngpu))
sys.exit()
elif type(nxs) == int:
if nxs % s.ngpu == 0:
s.nxs = [nxs/s.ngpu for i in xrange(s.ngpu)]
s.nx_gpu = nxs
else:
print('Error: nxs %d is not multiple of the number of target devices %d.' %(nxs, s.ngpu))
sys.exit()
else:
print('Error: nxs type %s is invalid.' %type(nxs))
sys.exit()
s.ny, s.nz = ny, nz
s.check_grid_size()
s.allocations()
s.get_program(print_source=False)
def __init__(s, nxs, ny, nz, target_device='all', print_verbose=True):
s.print_verbose = print_verbose
s.gpu_devices = utils.get_gpu_devices(s.print_verbose)
if s.print_verbose:
utils.print_gpu_info(s.gpu_devices)
utils.print_cpu_info()
ngpu_dev = len(s.gpu_devices)
s.lsize = 256
s.gsizes = []
s.nnx = 1
s.ngpu = ngpu_dev
s.context, s.queues = utils.create_context_queues(s.gpu_devices)
td = target_device
if ngpu_dev > 0:
for device in s.gpu_devices:
s.gsizes.append( utils.get_optimal_global_work_size(device) )
if td == 'cpu':
s.ngpu = 0
target_str = 'CPU'
elif td in ['gpu%d' % i for i in range(ngpu_dev)]:
s.ngpu = 1
gpu_num = int(td.strip('gpu'))
s.gsizes = [ s.gsizes[gpu_num] ]
s.gpu_devices = [ s.gpu_devices[gpu_num] ]
s.context, s.queues = utils.create_context_queues(s.gpu_devices)
target_str = 'Single GPU #%d' % gpu_num
elif td in ['gpu']:
s.nnx = ngpu_dev
target_str = '%d GPUs' % s.ngpu
elif td in ['all']:
s.nnx = ngpu_dev + 1
target_str = 'CPU + %d GPUs' % s.ngpu
else:
print('Error: Invalid target_device option.')
print(' Possible options: %s' %(['all', 'cpu', 'gpu'] + ['gpu%d' % i for i in range(ngpu_dev)]))
sys.exit()
else:
if td in ['all', 'cpu']:
s.nnx = 1
s.ngpu = 0
target_str = 'CPU'
else:
print('Error: Invalid target_device option.')
print(' There are no GPU devices.')
print(' Possible options: %s' %(['all', 'cpu']))
sys.exit()
if type(nxs) == list:
if len(nxs) == s.nnx:
s.nxs = nxs
s.nx_total = np.array(nxs).sum()
else:
print('Error: len(nxs) %d is not matched with the number of target devices %d.' %(len(nxs), s.nnx))
sys.exit()
elif type(nxs) == int:
s.nx_total = nxs
if s.nnx == 1:
s.nxs = [nxs]
else:
#s.nxs = utils.get_optimal_nxs()
s.nxs = [nxs/s.ngpu for i in xrange(s.ngpu)]
else:
print('Error: nxs type %s is invalid.' % type(nxs))
print(' Possible types: %s' %(['list', 'int']))
sys.exit()
if s.print_verbose:
print('Target Device : %s' % target_str)
print('s.nnx = %d' % s.nnx)
print('s.ngpu = %d' % s.ngpu)
print('s.nxs = %s' % s.nxs)
print('')
s.ny, s.nz = ny, nz
s.check_grid_size()
s.allocations()
s.get_program(print_ksource=False)
s.prepare_updates()
def get_dt(func, args):
tmax = 1000
t0 = datetime.now()
for tstep in xrange(1, tmax+1): func(*args)
dt0 = datetime.now() - t0
dt = (dt0.seconds + dt0.microseconds * 1e-6) / tmax
return dt
nxs = range(96, 480+1, 32) # nx**2, 72.0 KiB ~ 1.76 MiB
gpu_devices = utils.get_gpu_devices()
utils.print_gpu_info(gpu_devices)
context, queues = utils.create_context_queues(gpu_devices)
queue = queues[0]
mf = cl.mem_flags
print('nx\tdtoh\t\thtod\t\tdtoh(pinned)\thtod(pinned)')
for nx in nxs:
#for nx in [480]:
print(nx),
f = np.random.rand(2*nx*nx).astype(np.float32)
dbuf = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=f)
hbuf = np.zeros(f.shape, dtype=f.dtype)
# dtoh (pageable)
#cl.enqueue_read_buffer(queue, dbuf, hbuf)
#print(np.linalg.norm(f - hbuf))
#assert np.linalg.norm(f - hbuf) == 0
def get_dt(func, args):
tmax = 1000
t0 = datetime.now()
for tstep in xrange(1, tmax + 1):
func(*args)
dt0 = datetime.now() - t0
dt = (dt0.seconds + dt0.microseconds * 1e-6) / tmax
return dt
nxs = range(96, 480 + 1, 32) # nx**2, 72.0 KiB ~ 1.76 MiB
gpu_devices = utils.get_gpu_devices()
utils.print_gpu_info(gpu_devices)
context, queues = utils.create_context_queues(gpu_devices)
queue = queues[0]
mf = cl.mem_flags
print('nx\tdtoh\t\thtod\t\tdtoh(pinned)\thtod(pinned)')
for nx in nxs:
#for nx in [480]:
print(nx),
f = np.random.rand(2 * nx * nx).astype(np.float32)
dbuf = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=f)
hbuf = np.zeros(f.shape, dtype=f.dtype)
# dtoh (pageable)
#cl.enqueue_read_buffer(queue, dbuf, hbuf)
#print(np.linalg.norm(f - hbuf))
#assert np.linalg.norm(f - hbuf) == 0
def __init__(s, nxs, ny, nz, target_device='all', print_verbose=True):
s.print_verbose = print_verbose
s.gpu_devices = utils.get_gpu_devices(s.print_verbose)
if s.print_verbose:
utils.print_gpu_info(s.gpu_devices)
utils.print_cpu_info()
ngpu_dev = len(s.gpu_devices)
s.lsize = 256
s.gsizes = []
s.nnx = 1
s.ngpu = ngpu_dev
s.context, s.queues = utils.create_context_queues(s.gpu_devices)
td = target_device
if ngpu_dev > 0:
for device in s.gpu_devices:
s.gsizes.append(utils.get_optimal_global_work_size(device))
if td == 'cpu':
s.ngpu = 0
target_str = 'CPU'
elif td in ['gpu%d' % i for i in range(ngpu_dev)]:
s.ngpu = 1
gpu_num = int(td.strip('gpu'))
s.gsizes = [s.gsizes[gpu_num]]
s.gpu_devices = [s.gpu_devices[gpu_num]]
s.context, s.queues = utils.create_context_queues(
s.gpu_devices)
target_str = 'Single GPU #%d' % gpu_num
elif td in ['gpu']:
s.nnx = ngpu_dev
target_str = '%d GPUs' % s.ngpu
elif td in ['all']:
s.nnx = ngpu_dev + 1
target_str = 'CPU + %d GPUs' % s.ngpu
else:
print('Error: Invalid target_device option.')
print(' Possible options: %s' %
(['all', 'cpu', 'gpu'] +
['gpu%d' % i for i in range(ngpu_dev)]))
sys.exit()
else:
if td in ['all', 'cpu']:
s.nnx = 1
s.ngpu = 0
target_str = 'CPU'
else:
print('Error: Invalid target_device option.')
print(' There are no GPU devices.')
print(' Possible options: %s' % (['all', 'cpu']))
sys.exit()
if type(nxs) == list:
if len(nxs) == s.nnx:
s.nxs = nxs
s.nx_total = np.array(nxs).sum()
else:
print(
'Error: len(nxs) %d is not matched with the number of target devices %d.'
% (len(nxs), s.nnx))
sys.exit()
elif type(nxs) == int:
s.nx_total = nxs
if s.nnx == 1:
s.nxs = [nxs]
else:
#s.nxs = utils.get_optimal_nxs()
s.nxs = [nxs / s.ngpu for i in xrange(s.ngpu)]
else:
print('Error: nxs type %s is invalid.' % type(nxs))
print(' Possible types: %s' % (['list', 'int']))
sys.exit()
if s.print_verbose:
print('Target Device : %s' % target_str)
print('s.nnx = %d' % s.nnx)
print('s.ngpu = %d' % s.ngpu)
print('s.nxs = %s' % s.nxs)
print('')
s.ny, s.nz = ny, nz
s.check_grid_size()
s.allocations()
s.get_program(print_ksource=False)
s.prepare_updates()