def __init__(self, gpuf, direction):
common.check_type('gpuf', gpuf, gpu.Fields)
common.check_value('direction', direction, ('+', '-', '+-'))
qtask = cpu.QueueTask()
if '+' in direction:
self.cpuf_p = cpuf_p = cpu.Fields(qtask,
3,
gpuf.ny,
gpuf.nz,
gpuf.coeff_use,
gpuf.precision_float,
use_cpu_core=0)
self.gf_p_h = gpu.GetFields(gpuf, ['hy', 'hz'], (-2, 0, 0),
(-2, -1, -1))
self.sf_p_h = cpu.SetFields(cpuf_p, ['hy', 'hz'], (0, 0, 0),
(0, -1, -1), True)
self.gf_p_e = cpu.GetFields(cpuf_p, ['ey', 'ez'], (1, 0, 0),
(1, -1, -1))
self.sf_p_e = gpu.SetFields(gpuf, ['ey', 'ez'], (-1, 0, 0),
(-1, -1, -1), True)
self.gf_h = cpu.GetFields(cpuf_p, ['hy', 'hz'], (1, 0, 0),
(1, -1, -1))
self.sf_e = cpu.SetFields(cpuf_p, ['ey', 'ez'], (-1, 0, 0),
(-1, -1, -1), True)
self.tmp_recv_e = np.zeros(self.gf_h.host_array.shape, gpuf.dtype)
if '-' in direction:
self.cpuf_m = cpuf_m = cpu.Fields(qtask,
3,
gpuf.ny,
gpuf.nz,
gpuf.coeff_use,
gpuf.precision_float,
use_cpu_core=0)
self.gf_m_e = gpu.GetFields(gpuf, ['ey', 'ez'], (1, 0, 0),
(1, -1, -1))
self.sf_m_e = cpu.SetFields(cpuf_m, ['ey', 'ez'], (-1, 0, 0),
(-1, -1, -1), True)
self.gf_m_h = cpu.GetFields(cpuf_m, ['hy', 'hz'], (1, 0, 0),
(1, -1, -1))
self.sf_m_h = gpu.SetFields(gpuf, ['hy', 'hz'], (0, 0, 0),
(0, -1, -1), True)
self.gf_e = cpu.GetFields(cpuf_m, ['ey', 'ez'], (1, 0, 0),
(1, -1, -1))
self.sf_h = cpu.SetFields(cpuf_m, ['hy', 'hz'], (0, 0, 0),
(0, -1, -1), True)
self.tmp_recv_h = np.zeros(self.gf_e.host_array.shape, gpuf.dtype)
# global variables
self.direction = direction
self.qtask = qtask
def runTest(self):
ufunc, nx, ny, nz, coeff_use, precision_float, use_cpu_core, tmax = self.args
qtask = cpu.QueueTask()
fields = cpu.Fields(qtask, nx, ny, nz, coeff_use, precision_float,
use_cpu_core)
cpu.Core(fields)
fields_ref = naive.Fields(nx,
ny,
nz,
precision_float,
segment_nbytes=16)
naive.Core(fields_ref)
# allocations
ns = fields.ns
dtype = fields.dtype
ehs = common_random.generate_ehs(nx, ny, nz, dtype, ufunc)
fields.set_ehs(*ehs)
fields_ref.set_ehs(*ehs)
ces, chs = common_random.generate_cs(nx, ny, nz, dtype, coeff_use)
if 'e' in coeff_use:
fields.set_ces(*ces)
fields_ref.set_ces(*ces)
if 'h' in coeff_use:
fields.set_chs(*chs)
fields_ref.set_chs(*chs)
# verify
strf_list = ['ex', 'ey', 'ez', 'hx', 'hy', 'hz']
if ufunc == 'e':
for tstep in xrange(0, tmax):
fields.update_e()
fields_ref.update_e()
fields.enqueue_barrier()
for strf, eh in zip(strf_list, ehs)[:3]:
norm = np.linalg.norm(fields.get(strf) - fields_ref.get(strf))
max_diff = np.abs(fields.get(strf) -
fields_ref.get(strf)).max()
self.assertEqual(
norm, 0,
'%s, %s, %g, %g' % (self.args, strf, norm, max_diff))
elif ufunc == 'h':
for tstep in xrange(0, tmax):
fields.update_h()
fields_ref.update_h()
fields.enqueue_barrier()
for strf, eh in zip(strf_list, ehs)[3:]:
norm = np.linalg.norm(fields.get(strf) - fields_ref.get(strf))
max_diff = np.abs(fields.get(strf) -
fields_ref.get(strf)).max()
self.assertEqual(norm, 0, '%s, %s, %g, %g' % \
(self.args, strf, norm, max_diff) )
def __init__(self, gpuf, direction, tmax):
common.check_type('gpuf', gpuf, gpu.Fields)
common.check_value('direction', direction, ('+', '-', '+-'))
qtask = cpu.QueueTask()
if '+' in direction:
self.cpuf_p = cpuf_p = cpu.Fields(qtask,
3,
gpuf.ny,
gpuf.nz,
gpuf.coeff_use,
gpuf.precision_float,
use_cpu_core=1)
self.gf_p_h = gpu.GetFields(gpuf, ['hy', 'hz'], (-2, 0, 0),
(-2, -1, -1))
self.sf_p_h = cpu.SetFields(cpuf_p, ['hy', 'hz'], (0, 0, 0),
(0, -1, -1), True)
self.gf_p_e = cpu.GetFields(cpuf_p, ['ey', 'ez'], (1, 0, 0),
(1, -1, -1))
self.sf_p_e = gpu.SetFields(gpuf, ['ey', 'ez'], (-1, 0, 0),
(-1, -1, -1), True)
self.gf_h = gf_h = cpu.GetFields(cpuf_p, ['hy', 'hz'], (1, 0, 0),
(1, -1, -1))
self.sf_e = cpu.SetFields(cpuf_p, ['ey', 'ez'], (-1, 0, 0),
(-1, -1, -1), True)
self.req_send_h = comm.Send_init(gf_h.host_array, rank + 1, tag=0)
self.tmp_recv_e_list = [
np.zeros(gf_h.host_array.shape, gpuf.dtype) for i in range(2)
]
self.req_recv_e_list = [
comm.Recv_init(tmp_recv_e, rank + 1, tag=1)
for tmp_recv_e in self.tmp_recv_e_list
]
self.switch_e = 0
if '-' in direction:
self.cpuf_m = cpuf_m = cpu.Fields(qtask,
3,
gpuf.ny,
gpuf.nz,
gpuf.coeff_use,
gpuf.precision_float,
use_cpu_core=1)
self.gf_m_e = gpu.GetFields(gpuf, ['ey', 'ez'], (1, 0, 0),
(1, -1, -1))
self.sf_m_e = cpu.SetFields(cpuf_m, ['ey', 'ez'], (-1, 0, 0),
(-1, -1, -1), True)
self.gf_m_h = cpu.GetFields(cpuf_m, ['hy', 'hz'], (1, 0, 0),
(1, -1, -1))
self.sf_m_h = gpu.SetFields(gpuf, ['hy', 'hz'], (0, 0, 0),
(0, -1, -1), True)
self.gf_e = gf_e = cpu.GetFields(cpuf_m, ['ey', 'ez'], (1, 0, 0),
(1, -1, -1))
self.sf_h = cpu.SetFields(cpuf_m, ['hy', 'hz'], (0, 0, 0),
(0, -1, -1), True)
self.req_send_e = comm.Send_init(gf_e.host_array, rank - 1, tag=1)
self.tmp_recv_h_list = [
np.zeros(gf_e.host_array.shape, gpuf.dtype) for i in range(2)
]
self.req_recv_h_list = [
comm.Recv_init(tmp_recv_h, rank - 1, tag=0)
for tmp_recv_h in self.tmp_recv_h_list
]
self.switch_h = 0
# global variables
self.direction = direction
self.qtask = qtask
self.tmax = tmax
self.tstep = 1
#nx, ny, nz = 240, 256, 256 # 540 MB #nx, ny, nz = 544, 544, 544 # 5527 MB #nx, ny, nz = 512, 512, 512 # 4608 MB #nx, ny, nz = 480, 480, 480 # 3796 MB nx, ny, nz = 800, 256, 256 # 576 MB #nx, ny, nz = 128, 128, 128 # 72 MB coeff_use = 'e' precision_float = 'single' # instances gpu_devices = common_gpu.gpu_device_list(print_info=False) context = cl.Context(gpu_devices) device = gpu_devices[0] qtask = cpu.QueueTask() fields = Fields(context, device, qtask, nx, ny, nz, coeff_use, precision_float) Core(fields) tmax = 250 if is_plot else 1000 if rank == 0: direction = '+' elif rank == size - 1: direction = '-' else: direction = '+-' #exch = node.ExchangeMpiNonBlock(fields, direction) #exch = node.ExchangeMpiBufferBlock(fields, direction) #exch = node.ExchangeMpiBufferBlockSplit(fields, direction) exch = node.ExchangeMpiBufferNonBlockSplitEnqueue(fields, direction, tmax) if '+' in direction: cpu.Core(exch.cpuf_p) if '-' in direction: cpu.Core(exch.cpuf_m)
use_cpu_core = 1
# instances
gpu_devices = common_gpu.gpu_device_list(print_info=False)
context = cl.Context(gpu_devices)
device = gpu_devices[0]
gpuf = gpu.Fields(context, device, nx, ny, nz, coeff_use, precision_float)
tmax = 250 if is_plot else 1000
if rank == 0: direction = '+'
elif rank == size - 1: direction = '-'
else: direction = '+-'
buffer_dict = {}
if '+' in direction:
buffer_dict['x+'] = cpu.Fields(cpu.QueueTask(), 2, ny, nz, coeff_use, precision_float, use_cpu_core)
if '-' in direction:
buffer_dict['x-'] = cpu.Fields(cpu.QueueTask(), 2, ny, nz, coeff_use, precision_float, use_cpu_core)
nodef = node.Fields([gpuf], buffer_dict)
node.Core(nodef)
#network.ExchangeMpiBlock(nodef, direction)
#network.ExchangeMpiNonBlock(nodef, direction)
network.ExchangeMpiBuffer(nodef)
if is_plot:
is_master = True if rank == 0 else False
getf = gpu.GetFields(gpuf, 'ez', (0, 0, 0.5), (-1, -1, 0.5))
# instances
gpu_devices = common_gpu.gpu_device_list(print_info=False)
context = cl.Context(gpu_devices)
device = gpu_devices[0]
gpuf = gpu.Fields(context, device, nx, ny, nz, coeff_use, precision_float)
tmax = 250 if is_plot else 10000
#if rank == 0: direction = '+'
#elif rank == size - 1: direction = '-'
#else: direction = '+-'
direction = '+-'
buffer_dict = {}
if '+' in direction:
buffer_dict['x+'] = cpu.Fields(cpu.QueueTask(), 2, ny, nz, coeff_use,
precision_float, use_cpu_core)
if '-' in direction:
buffer_dict['x-'] = cpu.Fields(cpu.QueueTask(), 2, ny, nz, coeff_use,
precision_float, use_cpu_core)
nodef = node.Fields([gpuf], buffer_dict)
node.Core(nodef)
#network.ExchangeMpiBlock(nodef, direction)
#network.ExchangeMpiNonBlock(nodef, direction)
network.ExchangeMpiBuffer(nodef)
if is_plot:
is_master = True if rank == 1 else False
getf = gpu.GetFields(gpuf, 'ez', (0, 0, 0.5), (-1, -1, 0.5))
def runTest(self):
ufunc, nx, ny, nz, coeff_use, precision_float, tmax = self.args
gpu_devices = common_gpu.gpu_device_list(print_info=False)
context = cl.Context(gpu_devices)
device = gpu_devices[0]
qtask = cpu.QueueTask()
fields = gpu.Fields(context, device, qtask, nx, ny, nz, coeff_use,
precision_float)
gpu.Core(fields)
fields_ref = naive.Fields(nx,
ny,
nz,
precision_float,
segment_nbytes=64)
naive.Core(fields_ref)
# allocations
ns = fields.ns
dtype = fields.dtype
strf_list = ['ex', 'ey', 'ez', 'hx', 'hy', 'hz']
ehs = common_random.generate_ehs(nx, ny, nz, dtype, ufunc)
fields.set_eh_bufs(*ehs)
fields_ref.set_ehs(*ehs)
ces, chs = common_random.generate_cs(nx, ny, nz, dtype, coeff_use)
if 'e' in coeff_use:
fields.set_ce_bufs(*ces)
fields_ref.set_ces(*ces)
if 'h' in coeff_use:
fields.set_ch_bufs(*chs)
fields_ref.set_chs(*chs)
tmpf = np.zeros(fields.ns_pitch, dtype=dtype)
# update
if ufunc == 'e':
for tstep in xrange(0, tmax):
fields.update_e()
fields_ref.update_e()
qtask.enqueue_barrier()
for strf, eh in zip(strf_list, ehs)[:3]:
cl.enqueue_copy(fields.queue, tmpf, fields.get_buf(strf))
norm = np.linalg.norm(fields_ref.get(strf) - tmpf)
max_diff = np.abs(fields_ref.get(strf) - tmpf).max()
self.assertEqual(
norm, 0,
'%s, %s, %g, %g' % (self.args, strf, norm, max_diff))
elif ufunc == 'h':
for tstep in xrange(0, tmax):
fields.update_h()
fields_ref.update_h()
for strf, eh in zip(strf_list, ehs)[3:]:
cl.enqueue_copy(fields.queue, tmpf, fields.get_buf(strf))
norm = np.linalg.norm(fields_ref.get(strf) - tmpf)
max_diff = np.abs(fields_ref.get(strf) - tmpf).max()
self.assertEqual(
norm, 0,
'%s, %s, %g, %g' % (self.args, strf, norm, max_diff))