Does
b = tf.slice(a, [...], [...])
allocate a new memory buffer and then copy from a's buffer? or do a and b share the same buffer?
And, what about
... = tf.nn.convolution(tf.slice(a, [...], [...]), ...)
In this case, the slice is not being named. Does this incur an allocation and copy?
In general, is there some resource for learning these memory management details?
Thanks!
I was curious of this as well and went into the kernel implementation of tf.slice.
The only time memory is not allocated is when the slice is the identity slice (so there are no changes), and when the slice is dimension 0 aligned.
if (is_identity) {
VLOG(1) << "Slice identity";
context->set_output(0, input);
*done = true;
return;
}
if (slice_dim0 &&
IsDim0SliceAligned<T>(input.shape(), (*begin)[0], (*size)[0])) {
VLOG(1) << "Slice dim 0: " << input.shape().DebugString();
CHECK_GE(input.dims(), 1); // Otherwise, is_identity should be true.
context->set_output(0, input.Slice((*begin)[0], (*begin)[0] + (*size)[0]));
*done = true;
return;
}
Otherwise the following will be called
OP_REQUIRES_OK(context, context->allocate_output(0, *output_shape, result));
