def gpu_dag(dag):
""" The GPU version of a CPU dag - including gpu communication """
nc_dag, sent, recvd = non_comm_dag(dag)
recvs = {
cpu_to_gpu_var(var)[0].clone(): {
'fn': GpuFromHost(),
'args': (var, )
}
for var, it in dag.items() if isrecv(it['fn'])
}
sends = {
it['args'][0]: {
'fn': HostFromGpu(),
'args': (cpu_to_gpu_var(it['args'][0])[0].clone(), )
}
for _, it in dag.items() if issend(it['fn'])
}
def gpu_item((k, v)):
i, op, o = v['args'], v['fn'], (k, )
gi, gop, go = gpu_job(i, op, o)
return (go[0], {'fn': gop, 'args': gi})
gdag = dict(map(gpu_item, nc_dag.items()))
return merge(gdag, recvs, sends)
def gpu_dag_transfers(dag):
""" Edges/jobs for moving data from gpu version dag to cpu version
>>> dag = {c: {'fn': dot 'args': (a, b)}}
>>> gpu_dag_transfers(dag)
{c: {'fn': HostFromGpu(), 'args': (gpu_c,)},
gpu_a: {'fn': GpuFromHost(), 'args': (a, )},
gpu_b: {'fn': GpuFromHost(), 'args': (b, )}}
"""
recv_inputs = {
cpu_to_gpu_var(var)[0].clone(): {
'fn': GpuFromHost(),
'args': (var, )
}
for var in inputs_of(dag) if isinstance(var, theano.Variable)
}
send_outputs = {
var: {
'fn': HostFromGpu(),
'args': (cpu_to_gpu_var(var)[0].clone(), )
}
for var in outputs_of(dag) if isinstance(var, theano.Variable)
}
return merge(recv_inputs, send_outputs)
def gpu_dag_transfers(dag):
""" Edges/jobs for moving data from gpu version dag to cpu version
>>> dag = {c: {'fn': dot 'args': (a, b)}}
>>> gpu_dag_transfers(dag)
{c: {'fn': HostFromGpu(), 'args': (gpu_c,)},
gpu_a: {'fn': GpuFromHost(), 'args': (a, )},
gpu_b: {'fn': GpuFromHost(), 'args': (b, )}}
"""
recv_inputs = {cpu_to_gpu_var(var)[0].clone(): {'fn': GpuFromHost(),
'args': (var,)}
for var in inputs_of(dag)
if isinstance(var, theano.Variable)}
send_outputs = {var: {'fn': HostFromGpu(),
'args': (cpu_to_gpu_var(var)[0].clone(),)}
for var in outputs_of(dag)
if isinstance(var, theano.Variable)}
return merge(recv_inputs, send_outputs)
def gpu_dag(dag):
""" The GPU version of a CPU dag - including gpu communication """
nc_dag, sent, recvd = non_comm_dag(dag)
recvs = {cpu_to_gpu_var(var)[0].clone(): {'fn': GpuFromHost(),
'args': (var,)}
for var, it in dag.items()
if isrecv(it['fn'])}
sends = {it['args'][0]: {'fn': HostFromGpu(),
'args':(cpu_to_gpu_var(it['args'][0])[0].clone(),)}
for _, it in dag.items()
if issend(it['fn'])}
def gpu_item((k, v)):
i, op, o = v['args'], v['fn'], (k,)
gi, gop, go = gpu_job(i, op, o)
return (go[0], {'fn': gop, 'args': gi})
gdag = dict(map(gpu_item, nc_dag.items()))
return merge(gdag, recvs, sends)
if __name__ == '__main__':
import time
from sys import argv, stdout
from ape.theano_gpu_util import cpu_to_gpu_var, togpu_data
import theano
import numpy as np
import ast
ns = argv[1]
ns = ast.literal_eval(ns)
results = []
# Make function to send variables from gpu to cpu
x = theano.tensor.matrix('x')
gx, cx = cpu_to_gpu_var(x)
send = theano.function((gx,), cx)
for n in ns:
xx = np.ones((n,1), dtype=np.float32)
gxx = togpu_data(xx)
starttime = time.time()
cxx = send(gxx)
endtime = time.time()
duration = endtime - starttime
results.append((xx.nbytes, duration))
stdout.write("[%s]\n"%(',\n'.join(map(str, results))))
if __name__ == '__main__':
import time
from sys import argv, stdout
from ape.theano_gpu_util import cpu_to_gpu_var, togpu_data
import theano
import numpy as np
import ast
ns = argv[1]
ns = ast.literal_eval(ns)
results = []
# Make function to send variables from gpu to cpu
x = theano.tensor.matrix('x')
gx, cx = cpu_to_gpu_var(x)
send = theano.function((gx, ), cx)
for n in ns:
xx = np.ones((n, 1), dtype=np.float32)
gxx = togpu_data(xx)
starttime = time.time()
cxx = send(gxx)
endtime = time.time()
duration = endtime - starttime
results.append((xx.nbytes, duration))
stdout.write("[%s]\n" % (',\n'.join(map(str, results))))