torchFFI vs regular lua code: HUGE improvement

testffi.lua

-- test of FFI vs regular lua code
-- EC march 21st 2014

torch.setdefaulttensortype('torch.FloatTensor')

require 'sys'
require 'image'
require 'torchffi'
local bit = require('bit')
local ffi = require('ffi')
local C = ffi.C


-- blob parser LUA:
function parse(tin, threshold, blobs, scale)
  --loop over pixels
  for y=1, tin:size(1) do
     for x=1, tin:size(2) do
        local val = tin[y][x]
        if (val > threshold) then
          entry = {}
          entry[1] = x
          entry[2] = y
          entry[3] = scale
          table.insert(blobs,entry)
      end
    end
  end
end


-- blob parser FFI:
function parseFFI(tin, threshold, blobs, scale)
  --loop over pixels
  for y=0, 511 do
     for x=0, 511 do
        local val = tin[512*y+x]
        if (val > threshold) then
          entry = {}
          entry[1] = x
          entry[2] = y
          entry[3] = scale
          table.insert(blobs,entry)
      end
    end
  end
end

-- TEST:
tima = image.lena()[1]
timap = torch.data(tima)

sys.tic()
parse(tima, 10, 0, 0)
t = sys.toc()
print('Time in lua: ', t)

sys.tic()
parseFFI(timap, 10, 0, 0)
t = sys.toc()
print('Time with FFI: ', t)

th -i testffi.lua
Time in lua: 0.22888493537903
Time with FFI: 0.0010042190551758

Basically, you are sending a 2D Tensor in the first case and and an array pointer in the second. Is it correct? So, there is a 230× improvement??? :o
And, ffi is the luajit thing, correct?