Cache bf.ndarray.as_BFarray() to reduce overhead

- The cost of the full call is quite significant on things like small memcopies

python/bifrost/ndarray.py

@@ -39,7 +39,7 @@
 import ctypes
 import numpy as np
 from memory import raw_malloc, raw_free, raw_get_space, space_accessible
-from bifrost.libbifrost import _bf, _check
+from bifrost.libbifrost import _bf, _check, _fast_call
 import device
 from DataType import DataType
 from Space import Space
@@ -78,15 +78,16 @@ def copy_array(dst, src):
 	    space_accessible(src_bf.bf.space, ['system'])):
 		np.copyto(dst_bf, src_bf)
 	else:
-		_check(_bf.ArrayCopy(dst_bf.as_BFarray(), src_bf.as_BFarray()))
+		_fast_call(_bf.ArrayCopy, dst_bf.as_BFarray(), src_bf.as_BFarray())
 		if dst_bf.bf.space != src_bf.bf.space:
 			# TODO: Decide where/when these need to be called
 			device.stream_synchronize()
 	return dst
 
 def memset_array(dst, value):
 	dst_bf = asarray(dst)
-	_check(_bf.ArrayMemset(dst_bf.as_BFarray(), value))
+	#_check(_bf.ArrayMemset(dst_bf.as_BFarray(), value))
+	_fast_call(_bf.ArrayMemset, dst_bf.as_BFarray(), value)
 	return dst
 
 # Stores Bifrost-specific metadata that augments Numpy's metadata
@@ -138,6 +139,7 @@ def __new__(cls, base=None, space=None, shape=None, dtype=None,
 				# Allow conjugated to be redefined
 				if conjugated is not None:
 					obj.bf.conjugated = conjugated
+					obj._update_BFarray()
 			else:
 				if not isinstance(base, np.ndarray):
 					# Convert base to np.ndarray
@@ -222,6 +224,7 @@ def __new__(cls, base=None, space=None, shape=None, dtype=None,
 			obj = np.ndarray.__new__(cls, shape, dtype_np,
 			                         data_buffer, offset, strides)
 			obj.bf = BFArrayInfo(space, dtype, native, conjugated, ownbuffer)
+			obj._update_BFarray()
 		return obj
 	def __array_finalize__(self, obj):
 		if obj is None:
@@ -241,10 +244,18 @@ def __array_finalize__(self, obj):
 			native     = obj.dtype.isnative
 			conjugated = False
 			self.bf = BFArrayInfo(space, dtype, native, conjugated)
+		self._update_BFarray()
 	def __del__(self):
 		if hasattr(self, 'bf') and self.bf.ownbuffer:
 			raw_free(self.bf.ownbuffer, self.bf.space)
+	def _update_BFarray(self):
+		# (Re-)cache the BFarray structure
+		# Note: This must be called after any updates to self.bf.*
+		self._BFarray = None
+		self._BFarray = self.as_BFarray()
 	def as_BFarray(self):
+		if self._BFarray is not None:
+			return self._BFarray
 		a = _bf.BFarray()
 		a.data      = self.ctypes.data
 		a.space     = Space(self.bf.space).as_BFspace()
@@ -282,6 +293,7 @@ def view(self, dtype=None, type_=None):
 			dtype_np = np.dtype(dtype_bf.as_numpy_dtype())
 			v = super(ndarray, self).view(dtype_np)
 			v.bf.dtype = dtype_bf
+			v._update_BFarray()
 			return v
 	#def astype(self, dtype):
 	#	dtype_bf = DataType(dtype)
@@ -304,6 +316,7 @@ def tofile(self, fid, sep="", format="%s"):
 	def byteswap(self, inplace=False):
 		if inplace:
 			self.bf.native = not self.bf.native
+			self._update_BFarray()
 			return super(ndarray, self).byteswap(True)
 		else:
 			return ndarray(self).byteswap(True)