Add support for copy on write with store_binary and frozen string

[kou]

It reduces memory usage when NArray just uses data as read only data because store_binary doesn't copy the passed binary data until NArray writes the passed binary data.

Benchamrk:

require "numo/narray"

GC.disable

shape = [100, 100]
array = Numo::Int32.new(*shape).seq
data = array.to_binary
frozen_data = array.to_binary.freeze
marshaled_data = Marshal.dump(array)

def report_memory_usage(label, previous)
  usage = File.readlines("/proc/self/status").grep(/^VmRSS/)[0][/^VmRSS:\s*(\d+)/, 1].to_i
  puts "#{label}: #{usage - previous} kB"
  usage
end

usage = report_memory_usage("initial", 0)

10000.times do
  Numo::Int32.new(*shape).seq
end
usage = report_memory_usage("seq", usage)

10000.times do
  Numo::Int32.from_binary(data, shape)
end
usage = report_memory_usage("from not frozen data", usage)

10000.times do
  Numo::Int32.from_binary(frozen_data, shape)
end
usage = report_memory_usage("from frozen data", usage)

10000.times do
  Numo::Int32.from_binary(frozen_data, shape)[0, 0] = 1
end
usage = report_memory_usage("from frozen data with write", usage)

10000.times do
  Marshal.load(marshaled_data)
end
usage = report_memory_usage("marshal", usage)

With this change:

initial: 16488 kB
seq: 393360 kB
from not frozen data: 392040 kB
from frozen data: 1848 kB
from frozen data with write: 392660 kB
marshal: 396528 kB

Without this change:

initial: 16276 kB
seq: 393360 kB
from not frozen data: 391776 kB
from frozen data: 392040 kB
from frozen data with write: 392040 kB
marshal: 786456 kB

See "fron frozen data" entry. It reduces memory usage from 392040 kB to 1848 kB (0.4%).

See also "marshal" entry. It reduces memory usage from 786456 kB to 396528 kB (50%).

[kou]

Are you interested in this feature? Or do you dislike this feature?

If Numo::NArray has this feature, we can convert Apache Arrow data to Numo::NArray without data copy.