def BuildAssignInstruction(builder):
new_parallel_desc_symbol = boxing_util.TryReplaceDeviceTag(
builder, var_blob_object.parallel_desc_symbol, "cpu")
consumer_op_arg_parallel_attr = op_arg_util.OpArgParallelAttribute(
new_parallel_desc_symbol,
var_blob_object.op_arg_parallel_attr.sbp_parallel,
var_blob_object.op_arg_parallel_attr.opt_mirrored_parallel,
)
tmp_blob_object = boxing_util.BoxingTo(builder, value_blob_object,
consumer_op_arg_parallel_attr)
boxing_util.Assign(builder, var_blob_object, tmp_blob_object)
def NoBoxingCudaD2HStatelessCall(
self, op_attribute, in_parallel_conf, bn_in_op2blob_object={}
):
op_parallel_desc_sym = self.GetParallelDescSymbol(in_parallel_conf)
blob_parallel_desc_sym = boxing_util.TryReplaceDeviceTag(
self, op_parallel_desc_sym, "cpu"
)
self._CheckRefInBlobObjectParallelDesc(
op_attribute, blob_parallel_desc_sym, bn_in_op2blob_object=bn_in_op2blob_object,
)
def GetDirectBlobObject(blob_object, op_arg_parallel_attr):
return blob_object
self._StatelessCall(
"copy_d2h",
op_attribute,
op_parallel_desc_sym=op_parallel_desc_sym,
blob_parallel_desc_sym=blob_parallel_desc_sym,
bn_in_op2blob_object=bn_in_op2blob_object,
get_delegate_blob_object=GetDirectBlobObject,
)
