Use native implementation of grey-to-RGB conversion for display update

[WasabiFan]

Continuation of #497.

This change should do it for the EV3; IIRC, other platforms will need their own solution to handle the extra bit used for the green channel.

The things that blocked me from opening a PR for that change originally were A) needing to test on the EV3 and B) needing to try with python3. (or, I suppose B is a prereq to testing with our library on A.) Hopefully I can get some time to build in Docker and figure out the Python targets.

[WasabiFan]

Do we have interest in pursuing this? I never actually got to testing it on the EV3 but I can if the perf boost is beneficial.

[dwalton76]

I'll give it a try on my EV3

[dwalton76]

Testing with this:

#!/usr/bin/env python3

from ev3dev2.display import Display
from time import sleep
import datetime as dt
import logging

logging.basicConfig(level=logging.INFO,
                    format='%(asctime)s %(filename)22s %(levelname)8s: %(message)s')
log = logging.getLogger(__name__)


display = Display()
display_font = "luBS24"
x_grid = 4
y_grid = 4
total_ms = 0

log.info("start")
for x in range(20):
    start_time = dt.datetime.now()
    display.text_grid("%s%s%s" % (x, x, x), clear_screen=True, x=x_grid, y=y_grid, font=display_font)
    display.update()
    end_time = dt.datetime.now()
    delta = end_time - start_time
    total_ms += (delta.seconds * 1000) + int(delta.microseconds / 1000)

log.info("end")
log.info("total_ms: %s" % total_ms)
display.reset_screen()
display.update()

I ran it three times to get a baseline:

• 12401ms
• 12144ms
• 12586ms

[WasabiFan]

That's without the Pillow patch, right?

[dwalton76]

right....Pillow is still building :( Am going to test on master on your fork then greyscale-to-rgb-packer

[dwalton76]

With master on your fork I got

• 20500ms
• 22233ms
• 21487ms

But this does not include the time that it takes to import our Display library...that has gone up significantly. I did not time it before but it now takes 47s just to from ev3dev2.display import Display :(

[dwalton76]

With the patch I get

• 21942ms
• 21294ms
• 21157ms

[dwalton76]

What version of Pillow ships with ev3dev? (not in front of my ev3 at the moment). Something to try would be to apply @WasabiFan 's patch to that version and see what we get.

[WasabiFan]

I believe this also requires changing the Python library to use the native packer. Should be...

im.tobytes("raw", "RGB")

[dwalton76]

Today when we do

self.mmap[:] = self._img.convert("RGB").tobytes("raw", "XRGB")

we get back a list of size 91136 but with

self.mmap[:] = self._img.convert("RGB").tobytes("raw", "RGB")

the list is only of length 68352, this causes an IndexError: mmap slice assignment is wrong size exception. The original size of self.mmap is determined by FbMem._get_fix_info(fid)'s smem_len...the fix_info.smem_len is used in _map_fb_memory.

For grins I hacked _map_fb_memory to use a size of 68352, re-ran my test with "RGB" and got 22050ms, 20909ms, and 21051ms.

[WasabiFan]

Which platform are you running this on? It is only applicable on EV3.

[dwalton76]

Am running on an EV3.

System info (from ev3dev-sysinfo)

Image file:         ev3dev-stretch-ev3-generic-2019-03-03
Kernel version:     4.14.96-ev3dev-2.3.2-ev3
Brickman:           0.10.0
BogoMIPS:           148.88
Bluetooth:          
Board:              board0
BOARD_INFO_HW_REV=7
BOARD_INFO_MODEL=LEGO MINDSTORMS EV3
BOARD_INFO_ROM_REV=6
BOARD_INFO_SERIAL_NUM=00165346201F
BOARD_INFO_TYPE=main

[dwalton76]

What version of Pillow ships with ev3dev? (not in front of my ev3 at the moment). Something to try would be to apply @WasabiFan 's patch to that version and see what we get.

We are running 4.0.0-4

robot@ev3dev:~$ dpkg -l | grep "Python Imaging"
ii  python3-pil:armel                           4.0.0-4                           armel        Python Imaging Library (Python3)
robot@ev3dev:~$ 

[WasabiFan]

Right now, your measurements are showing that the change doesn't improve performance, right?