def gpu_job(i, op, o):
""" Convert a cpu job to a gpu job
inputs:
i - iterable of input variables
op - a cpu op
o - iterable of output variables
outputs:
gi - iterable of input variables
gop - a gpu op
go - iterable of output variables
"""
node = op.make_node(*i)
for v,nv in zip(o, node.outputs):
nv.name = v.name
gii, goo = cpu_to_gpu_graph(node.inputs, node.outputs)
if len(list_of_nodes(gii, goo)) != 1:
raise ValueError("We have assumed that translations of single cpu-ops "
"would result in single gpu-ops. This computation "
"has invalidated that assumption")
op = goo[0].owner.op
go = map(lambda x: x.clone(), goo)
gi = map(lambda x: x.clone(), gii)
for v, gv in zip(i+o, gi+go):
gv.name = gpu_name(v.name)
return gi, op, go
def gpu_job(i, op, o):
""" Convert a cpu job to a gpu job
inputs:
i - iterable of input variables
op - a cpu op
o - iterable of output variables
outputs:
gi - iterable of input variables
gop - a gpu op
go - iterable of output variables
"""
node = op.make_node(*i)
for v, nv in zip(o, node.outputs):
nv.name = v.name
gii, goo = cpu_to_gpu_graph(node.inputs, node.outputs)
if len(list_of_nodes(gii, goo)) != 1:
raise ValueError("We have assumed that translations of single cpu-ops "
"would result in single gpu-ops. This computation "
"has invalidated that assumption")
op = goo[0].owner.op
go = map(lambda x: x.clone(), goo)
gi = map(lambda x: x.clone(), gii)
for v, gv in zip(i + o, gi + go):
gv.name = gpu_name(v.name)
return gi, op, go
def internal_gpu_theano_graph(dag):
"""
inputs - a dicdag with send/recvs
outputs -
gins/gots - a theano i/o graph with gpu variables and ops
# sents - tuple of sent gpu variables
# recvs - a tuple of recved gpu variables
"""
non_comm, sent, recved = non_comm_dag(dag)
inputs = inputs_of( non_comm)
outputs = outputs_of(non_comm)
ins, outs = dicdag.theano.dag_to_theano_graph(non_comm, inputs, outputs)
gins, gouts = cpu_to_gpu_graph(ins, outs)
return gins, gouts
def internal_gpu_theano_graph(dag):
"""
inputs - a dicdag with send/recvs
outputs -
gins/gots - a theano i/o graph with gpu variables and ops
# sents - tuple of sent gpu variables
# recvs - a tuple of recved gpu variables
"""
non_comm, sent, recved = non_comm_dag(dag)
inputs = inputs_of(non_comm)
outputs = outputs_of(non_comm)
ins, outs = dicdag.theano.dag_to_theano_graph(non_comm, inputs, outputs)
gins, gouts = cpu_to_gpu_graph(ins, outs)
return gins, gouts
def time_computation(inputs, outputs, numeric_inputs, niter):
gpu_inputs, gpu_outputs = cpu_to_gpu_graph(inputs, outputs)
# TODO: replace this with a can_run_on(job, machine) function
# use this function in tompkins
if not all(isinstance(n.op, theano.sandbox.cuda.GpuOp)
for n in theano.gof.graph.list_of_nodes(gpu_inputs, gpu_outputs)):
return 99999.9
gpu_numeric_inputs = map(togpu_data, numeric_inputs)
f = theano.function(gpu_inputs, gpu_outputs)
starttime = time.time()
debugprint("Computing")
for n in xrange(niter):
outputs = f(*numeric_inputs)
endtime = time.time()
duration = endtime - starttime
return duration/niter
def time_computation(inputs, outputs, numeric_inputs, niter):
gpu_inputs, gpu_outputs = cpu_to_gpu_graph(inputs, outputs)
# TODO: replace this with a can_run_on(job, machine) function
# use this function in tompkins
if not all(
isinstance(n.op, theano.sandbox.cuda.GpuOp)
for n in theano.gof.graph.list_of_nodes(gpu_inputs, gpu_outputs)):
return 99999.9
gpu_numeric_inputs = map(togpu_data, numeric_inputs)
f = theano.function(gpu_inputs, gpu_outputs)
starttime = time.time()
debugprint("Computing")
for n in xrange(niter):
outputs = f(*numeric_inputs)
endtime = time.time()
duration = endtime - starttime
return duration / niter
