It had been quite a while since I had seen John 00 Fleming. He definitely has his own style of hypnotic trance, and that's exactly what I needed for that day. It also had been quite a while since I last went to AudioSF, so it was nice to be back, and the place was pretty busy :)
thanks Vic for clueing me in about the event
John is known for always having a great time when playing with a big smile :)
my LED googles were pretty popular
wee :)
People were having fun, we've been caged up for too long ;)
Thanks for John for taking picturs with us at the end
I drove to the top of Montebello for nice views and a sunset hike to Black Mountain:
Sure, you can also bike up (and it's a tough ride), but it's a fairly long bike ride back in the dark, so driving was easier.
Beautiful hike with sunset views, we'll have to do it again.
Another TFSF party, it had been a while since I last went to Halcyon. So glad to be back for a fun party with Blurr and Jonny Atom opening.
Unfortunately we only got 50% of Kyau and Albert as Steven got covid and wasn't able to travel, but Ralph did a great job on his own :)
We were early, and had the club to ourselves for a short while :)
It was good to be able to go back to Pure for a nice trance night. It's so close to home that it's great news to have more local events. Andrew Rayel started at 0:30, and while it was only a 90mn set, it was packed with energy and lots of tunes as Andrew did some very tight mixing with more than one track per minute at times. That was fun :)
The crowd had lots of energy and the nigh was truly enjoyable. Thanks to Pure for the night:
First party of the year, despite the increasing covid issues, a fair amount of people came, we had the show in two rooms and it was pretty packed. Everyone came to have a good time, and we sure did. Thanks or the set, Markus!
The v4 outfit was 64x96 resolution (3 P4 panels of 64x32), while the new v4 outfit changes to 128x192 (3 P2 panels of 128x64), or 4 times more pixels running in the same exact footprint (my body size, which ideally remains constant :) ).
Because I now have 4x more pixels, and that would have been too much for an ESP32 or teensy (before teensy 4, which could have done it, but it lacks wifi), I had to switch to something with more memory and horsepower, I went with rPi and Henner Zeller's rpi-rgb-panel library, which also supported more panels like mine that had an FM6126 or FM6127 chip that required a special init string.
The other reason for rPi is that to get a high refresh rate and avoid visible refresh bars when taking pictures with cameras, it was better to run the 3 panels on 3 different channels to give them maximum refresh rate (300-400Hz is possible that way), which is supported by that library, using the active-3 board.
So, easy, I just have to port all my code from arduino/ESP32 to rPi/linux, right? Damn, that's actually a lot of work, and I didn't want to do this, so I was able to do something better, I found ArduinoOnPC, and was able to fork and modify it to add Framebuffer::GFX support and added 3 display outputs:
This combined work allows running my arduino code on linux, mostly unmodified, which means it runs on rPi. Then, the FastLED_RPIRGBPanel_GFX glue driver I wrote to make rpi-rgb-panel compatible with FrameBuffer::GFX, allows running all my 2D code, unmodified and have it sent to RGBPanels through the rpi-rgb-panel driver, I'm glad I didn't have to write :)
So, this is what the prototype looked like now 2 years ago:
In the process, I fixed a lot of bugs in all the 2D code I had when I went to a 384x256 panel which obviously overflowed all code that relied on x and y being smaller than 256, and also blew up FastLED code that assumed that there could only ever be 64K pixels :)
While this was a big milestone and proved that my crazy idea running arduino code on rPi using my same 2D library, was possible, there was a lot of work left to do for my outfit, including changing all the code to deal with a much higher resolution, and days (yes days) of work finding close to 200 animated GIFs in the higher resolution, and rescaling them for my non standard 128x192 resolution. Also, all the fonts had to be changed, and a bunch of other stuff, which took months and months of work (all in all over a year):
So, this is what it looks like, ESP32 SmartMatrix 64x96, compared to rPi rpi-rgb-panel 128x192, triple channel. Some gifs, I found the exact one in higher resolution:
this gif was so cool, it's only with the higher resolution that I found out it was Pulp Fiction
Other gifs, I found a much better one:
Here is a good video showing the resolution difference between the 2 chips:
Hardware
The hardware got a bit more complicated, especially as my wiring wasn't giving reliable enough 5V power to the rPi causing random failures. Eventually I had to feed the battery voltage (16V) via a different cable and step it down to 5V/USB behind the panels to get full power to the rPi (otherwise when the panels drew too much from the main 5V source, it dipped it a bit too much and caused issues).
It does not look great, it's meant to be serviceable and easy to debug, and that part is against my body, so people don't see it :)
For wiring reasons, I had a nice trick with the front panels would shift out bits, and I would send them to the back panels with ribbon cables. If you know how RGBPanels work, it does mean that a BCM plane meant to be displayed for an interval t1 in the front, ends up being displayed for an interval t2, when shifted to the back. Because of random luck of the order of BCM planes, it happened to work well enough with SmartMatrix, so it saved wiring for me (no need to splice the output to go to front and rear panels). Unfortunately with rpi-rgb-panel, the BCM planes are displayed in the opposite order it seems, so the output shifted on the rear panels, is visibly not good:
Unfortunately, there is no good fix to this outside of splicing cables, which I didn't want to do, so instead I slightly hacked the rpi-rgb-panel library to shift output bits twice. This is a bit wasteful for refresh rate, but things were fast enough with 3 channels, that I could afford the software hack and losing half of my refresh speed.
Another issue with the ribbon cables is that the active 3 board is wired to have the cable stick up (i.e. towards my belly). I worked with the nice folks at electrodragon to get bare boards without the connectors soldered and looked at how to best make it work for minimal height footprint. Bottom in blue is the original which sticks up the wrong way, upper right is what I would love to have but isn't possible because the traces on the board would have to be rewired (pin order is wrong when the plug is put underneath), so I had to settle with upper right, some angled connectors and I had to move the key hole from one side to the other for the cable to go in the right way.
I really wish I could have done this, but the wire order would be wrong
Then, I also had to protect the panels with my kitchen cutting board sheets that I've been using. It's not perfect, but they act as diffusers and protect the pixels a bit, because the P2 pixels are so small that they fall off if you look at them wrong:
Battery Use
The new setup uses a bit more battery, first because of the rPi, and also because the new panels use slightly more power, although not 4 times more because the pixels are 4 times slower, so the amount of light is somewhat similar. I had to upgrade my setup to allow adding a 3rd 80Wh battery for longer festivals (12 to 14H with 240Wh). With 2 batteries in the black box, I get about 8-9H.
ESP32 - rPi Integration
Because all my code was written for ESP32, including Wifi code that generated pages on the fly from code, thanks to https://github.com/distrakt/OmEspHelpers , and the ESP32 still runs the neopixels on arms and legs (plus IR input, although it's become a bit obsolete now), I took a very unusual approach of running my code on both CPUs a tthe same time.
The ESP32 runs the demo, and blind outputs it to an RGBPanel that isn't there anymore. At the same time, its debug serial output is connected to rPi which reads it as a text input over a ttyUSB serial port. The rPi code can run in independent mode (where I control it via ssh from my phone, haha), or it detect a hearbeat from the ESP32 over serial, and read commands from the ESP32, including what demo to run. So, the ESP32 controls what demo is run, tells the rPi to switch to that demo and display to the RGBPanels. That makes the rPi a bit more than just a glorified graphics card/GPU since it generates what pixels need to be displayed instead of just being given a pre-computed framebuffer to display.
I had to make the code smart over time so that the rPi can connect and disconnect from the ESP32 and run independently if the connection dies (which it used to when I had power issues that cause the FTDI ttyUSB to fail routinely when running on batteries).
The rPi can also back-control the ESP32, so when I test at home, I ssh into the rPi, and the rPi uses the serial connection to the ESP32 to tell it what to do, or I can use the web server on the ESP32 and tell it what to do directly.
This means the rPi can work on its own without the ESP32 being needed, except for:
- IR input (it's not really necessary, and linux IR code is very different, so it would be a full rewrite)
- Wifi commands (none of Wifi code works on linux and would have to be entirely rewritten)
- FastLED output would not work well on rPi since it's timing dependent, and also there are no IO pins left on mine with 3 channel output
If I were to re-write a lot of code to make wifi work on linux, that would make the rPi independent and not need the ESP32 anymore (except for the neopixel strips), so I just didn't bother. Also I can brag about having a dual CPU system with synchronization between the chips, which was fun to write and debug.
This bit of video shows how the 2 communicate:
Using Linux integration for development
There isn't much to say about this, thanks to my ArduinoOnPC work mentioned earlier in this page, the exact same linux code works on my rPi and my linux laptop, so I can write and debug my code on linux, which is so much faster.
This shows an example of what it looks like:
Wifi and OmEspHelpers
Ah yes, Wifi, that was a fair amount of code, especially on ESP32 where Wifi is more basic and can cause crashes if you get weird conflicts between interrupts, Flash (SPIFFS or FATFS), PSRAM, IRAM, and Wifi. After looking for an easy to use solution, I settled on https://github.com/distrakt/OmEspHelpers because I could generate the HTML pages from code (saving lost time to update the Flash FS each time, which is slow with 16MB, and not having to worry about syncing HTML tags between static HTML pages code).
This is the end result, the main screen allows selecting which demos run (neopixel strips + main screen), at which speed, how bright:
there are over 200 2D demos to cycle through, some are machine generated, some are animated gifs
The diagnosis screens give more info on the device, and allow editing the config file that selects which demos run by default depending on the screen size, and whether 'bestof' is selected, or not. The config file also allows choosing the order demos run, in:
Demo of wifi:
Glasses
I got tired of the El Wire glasses, they were unreliable, got dimmer over time, required high voltage (I literally got shocked by that current when wires got frayed), so I got rid of them.
I did try laser cut glass glasses, they look kind of cool, but they are big and impossible to fold.
I ended up getting neopixel glasses which had good battery life, but after I dropped them once, a pixel fell out, and that stopped the rest of the string from working. Thankfully I was able to take a spare neopixel from a strip and replace the missing one. I didn't have the right tools or skill, so I was not able to solder it, but I used glue and that worked too :)
LED Shoes
I've been using LED shoes since 2016. They're fun, but not that reliable: the LED strips keep breaking, if it's not the control module itself on occasion, and of course the battery is too small, which is why I wrote this page on how tohack/improve them, but that also makes them less reliable. I tried to find other options, but have not yet. I had a look at LED laces, and those are not very reliable either (or very bright). I tried to hack these to power from the power supply I added and upgraded in the shoes, but it was probably not the best idea:
I got everything somewhat working in April 2021 for a first show (about 1.5 years from when I started), and then worked through multiple electrical and reliably problems (including serial ttyUSB stability issues, and power issues I had to fix by adding a second higher voltage feed to the rPi). Also fixed sync issues between the chips and other improvements in graphics and menus. I considered V5 mostly done and reliable just in time for ADE 2021, EDC 2021, and Dreamstate 2021
I had the outfit mostly working (with a few occasional hardware issues) in time for Creamfields and untold in Aug-Sept 2021:
This even got me on Romanian national TV :)
I had fun at ADE also, and add time to make custom displays for specific events:
Solarstone looks better than his picture, haha
I have some early code that allows me to send text from my phone for special occasions, need to improve the interface:
Future
Higher resolution still? Probably not, I'm happy with this resolution, I can get pictures and text with proper fonts. Besides, that would just get me to the level of strapping a TFT screen to me with HDMI input, which would kind of be cheating and render all my code worthless since basically I'd just be carrying a TV.
I'll add a better interface to send text messasges
Uploading images from the field would be cool
Adding a USB webcam and doing video
Removing the ESP32 and running everything from the rPI would make sense, but I don't feel like rewriting all my web code.
Redoing the power system to work off USB battery packs that could be strapped under the panels (tricky because of amps needed, and would make the whole outfit heavy, because batteries are heavy)
I happened to walk by the former French Mint (was in Paris, now moved somewhere bigger and cheaper for higher production volume). It's a very interesting museum that shows how the coins were designed, made, and even goes into each metal's pros and cons, as well as the history of coins over French history. Quite fascinating.
original old equipment, not used anymore
a cool artifact is they had an original KG prototype #17, used to verify that the main one, wasn't losing weight
movie that showed how coins are made today
Really cool thing is they showed how coins and medals were made:
an artist was working on this big model
big models
the big model goes on the right, and is used to make the target size small coin model on the left
I had already written FastLED_SPITFT::GFX to display Framebuffer::GFX code on TFTs, but that was using the Adafruit drivers that were of varying quality (one didn't support HWSPI on ESP32, so it was very slow).
Later, I found out about Arduino_GFX from Leung CHAN, which is a unified driver for a lot of TFTs of much better quality than the adafruit drivers. There is support for: GC9A01, GC9106, HX8347C, HX8347D, HX8352C, HX8357A, HX8357B, ILI9225, ILI9341, ILI9341, ILI9342, ILI9481, ILI9486, ILI9488, ILI9488, ILI9806, JBT6K71, NT35310, NT35510, NT39125, R61529, SEPS525, SSD1283A, SSD1331, SSD1351, SSD1351, ST7735, ST7735, ST7735, ST7789, ST7789, ST7789, ST7789, ST7789, ST7796
all with a single driver, a single interface, and better speed than Adafruit drivers. Good job Leung, thanks.
Now, why would you use my FastLED_ArduinoGFX::TFT layer, especially when Arduino::GFX has some support for Canvas (equivalent to FrameBuffer?)
First, using a Framebuffer (or canvas) allows for running clear and doing a flush all at once when drawing the next frame is finished
Framebuffer allows you to read back pixels from the framebuffer to shift them or mirror them, or even to dim them (FastLED allows diming the entire framebuffer to give effects that fade old pixels with time)
FastLED and SmartMatrix (or rpi-rgb-panel) all have code that is based on a framebuffer with RGB88 pixels (24bits) which is why Framebuffer::GFX comes in and why I wrote it. All this code works against Framebuffer::GFX and can be displayed on any supported backend.
Support for FastLED and LEDMatrix 2D APIs (which in turn require FastLED CRGB (RGB88) pixel storage, again supported by Framebuffer::GFX but not the Adafruit or Arduino_GFX TFT drivers.
See the Framebuffer::GFX that explains support
So, this is why FastLED_ArduinoGFX::TFT is here. If you want more APIs than just Adafruit::GFX, and you want your code to also work on all other supported backends (FastLED Matrix, SmartMatrix, RGBPanel on rPi, or even running directly on linux to write/debug your arduino code on linux with gdb or ASAN memory sanitizer)
This same LEDMatrix demo now works a lot faster thanks to Arduino_GFX copying data to the TFT faster:
This is 24bpp FastLED/LEDMatrix code running on a 16bpp TFT via Framebuffer::GFX
Basic code example
This basic example is the simplest and skips neomatrix_config.h by defining things "in line", in the code. It's easier to understand, but defeats the main advantage of neomatrix_config.h, which is to have all your definitions outsdie of your code, and allows updating your hardware info in a single file while having all your demo code still work on new hardware by just modifying one common file.
Have a look at this simple file:
https://github.com/marcmerlin/FastLED_ArduinoGFX_TFT/blob/master/examples/MatrixGFXDemo/MatrixGFXDemo.ino
Running Framebuffer::GFX code on a TFT, FastLED_NeoMatrix_SmartMatrix_LEDMatrix_GFX_Demos and neomatrix_config.h
If you have bigger displays like an ILI9341, that's 320*240*24bpp or 230KB for a full 24bpp framebuffer. That fits on a teensy 3.6 or better, but not on an ESP32 where the memory is not contiguous (unless you use PSRAM which neomatrix_config will automatically use for you).
This means that without enough memory, you can define a smaller framebuffer that only covers portion of the TFT, and then render the framebuffer at the desired offset. Check out this example to see how it works: https://github.com/marcmerlin/FastLED_ArduinoGFX_TFT/tree/master/examples/SplitILI934Display
This is the end result, you can see that the ILI9341 is spilt in two, the top half is mapped to a framebuffer, the 2nd part uses direct adafruit::GFX rendering through Arduino_GFX:
tftw/tftw are size of the physical TFT, mw/mh are size of the framebuffer, framebuffer can be smaller than the TFT if there isn't enough RAM to have a framebuffer as big as the TFT, like ILI9341 on ESP32 without PSRAM
gfx_scale is used to keep track of a framebuffer smaller than the TFT (for instance on ESP32 without PSRAM, you can have a half height framebuffer that is then displayed twice, or can be reset with new data and applied to the top of the screen and later the bottom of the screen)
tft_name is simply used to keep track of what TFT name that index is, used for debugging
