TinyFPGA BX USB Serial

Origins

Luke Valenty's USB module, as adapted by Lawrie Griffiths and others. Luke's code was created with the purpose of providing a "bit banged" USB port to SPI bridge for his (awesome) TinyFPGA boards.

The original is here - https://github.com/tinyfpga/TinyFPGA-Bootloader

Lawrie Griffiths dug into the dark innards of this code and did the bulk of the work to change it to being a USB - SERIAL bridge, looking to the user like a serial port, rather than an SPI master. He also created some great examples. His work is here - https://github.com/lawrie/tiny_usb_examples

However there were a couple of problems:

• arachne-pnr would give spotty results, sometimes not working at all. This is easy to understand when considering that arachne-pnr does not place in such a way as to minimize timings! NextNPR is a must.
• the original code has a few places where it violates Verilog rules (mostly fixed by smunaut and submitted as a pull request tinyfpga/TinyFPGA-Bootloader#21)

This Project

This project makes the above changes, and fixes one or two other small issues. Importantly, it also replaces the front end to the code with a pipeline interface, which is what makes this project more fork-y and less pull-request-y with respect to Lawrie's repo.

The key to the streaming pipeline interface is that there are three ports: data, valid, and ready. A sender will place new data on the data port and raise the valid flag. If a receiver is ready to receive, it will raise the ready flag. So in any connection there will be two signals generated by the sender (data and valid) and one by the receiver ready. And here's the key part if valid and ready are both raised at the same time the data is considered to be transfered. This interface can clunk along - the sender raising valid now and then, and provided the ready signal is raised, data will flow, but pretty amazingly, this interface can really rock - if there is data to be had at every clock cycle, just leave valid and ready on and one word of data will flow at each clock.

This works pretty well and has been tested at full speed using some Python code.

The USB code alone (with virtually no other functionality at all) uses 14% of the device's LUT's. Here's the NextPNR output:

Info: Device utilisation:
Info: 	         ICESTORM_LC:  1093/ 7680    14%
Info: 	        ICESTORM_RAM:     2/   32     6%
Info: 	               SB_IO:     9/  256     3%
Info: 	               SB_GB:     8/    8   100%
Info: 	        ICESTORM_PLL:     1/    2    50%
Info: 	         SB_WARMBOOT:     0/    1     0%

This code is passing timing.

Info: Max frequency for clock 'clk_48mhz_$glb_clk': 51.60 MHz (PASS at 48.00 MHz)

This is quite notable, since other adaptations of the Valenty code are forced to run most of the code in a much slower clock domain to pass timing.

Here's how the code is placed in the iCE40

An Aside - Xilinx 7

Since this is fully generic verilog code, there is no special work to do to make the code work in other architectures beyond instantiating the correct IO primitives. Delightfully, this codebase also works on Xilinx parts.

The file usb_uart_x7.v is the wrapper that does the IO primitive work for Xilinx 7 code, so the primitives don't clutter the higher code.

With a little ad-hoc hardware (an Adafruit USB mini breakout board, 68 ohm resistors connecting D+ and D- to FPGA IO, and a 1.5k pull up on the D+) a second port was added to a Digilent Cmod A7 board.

And connecting the soft usb to the onboard serial port results in a simple USB-USB bridge.

In contrast to the appearance of a 14% occupied iCE40, here's a 2% occupied Xilinx Artix 7 (20k LUTSs) running the simple SERIAL - Valenty USB bridge.

The wrapper code is included here and can easily be included in Xilinx 7 code that need a soft USB port, but the files for the above example are part of another to be soon released project. The present experiment was performed in Xilinx's Titanic dev environment, Vivado, but it might work in the open source Project Xray.

Another Aside - ECP 5

Some experimentation with the ECP5 Evaluation Board reveals that it too will run this code. In the shot below, we see the new working USB port on the standard Eval board's Pmod connector. A second, hand wired Pmod connector is visible behind although not in use.

The file usb_uart_ecp5.v is the wrapper that connects the ECP5 bidirectional IO primitive into the code.

The Eval board has one of those ubiquitous FTDI FT2232H chips for programming and IO, but the second channel is wired for I2C out of the box. To activate the UART side (which requires a UART on the FPGA!) a couple of modifications need to be made to the board.

First The FTDI has to be told that the protocol it should run on the B channel is UART. This means getting into FT Prog (Windows only, annoyingly) and setting the hardware setting to UART. This has the effect that when the chip powers up, it knows it should present as a UART (not a FIFO which is the default, and unsupported by the hardware).

Then connect the pads at each of R34 and R35 to connect the FTDI with the ECP5. They are shown as Do Not Insert 0 ohm resistors on the schematic. These lines are connected as I2C to several places by default, but only one connection is populated, FPGA lines (E6 D7). If they are left as inputs (high impedance) they don't disturb anything.

With R34 and R35 nicely bridged the FTDI.TX -> FPGA.RX line works, which is easily verified by piping the whole incoming char to the 8-leds. Also easy to see if you have baud rate issues here.

Finally, the FTDI when in UART mode wants to make a third line (FTDI pin 40) an RTS signal. That signal which is connected to the FTDI receiver (pin39) via a 0 ohm resistor kills all (FPGA TX -> FTDI RX) traffic. Removing R24 solves this.

It would be fun to see how many USB ports could be added to this board.

Interface

The interface to the code looks like the following:

usb_uart u_u (
  .clk_48mhz  (clk_48mhz),
  .reset      (reset),

  // pins
  .pin_usb_p( pin_usb_p ),
  .pin_usb_n( pin_usb_n ),

  // uart pipeline in
  .uart_in_data( uart_in_data ),
  .uart_in_valid( uart_in_valid ),
  .uart_in_ready( uart_in_ready ),

  // uart pipeline out
  .uart_out_data( uart_out_data ),
  .uart_out_valid( uart_out_valid ),
  .uart_out_ready( uart_out_ready ),
);

Note that there are no baud rates, start, stop or parity specifications as this is only a pretend UART.

Clock

You will need a 48Mhz clock. This can be generated by pll from the TinyFPGA BX's 16Mhz oscillator.

SB_PLL40_CORE #(
		.FEEDBACK_PATH("SIMPLE"),
		.DIVR(4'b0000),		// DIVR =  0
		.DIVF(7'b0101111),	// DIVF = 47
		.DIVQ(3'b100),		// DIVQ =  4
		.FILTER_RANGE(3'b001)	// FILTER_RANGE = 1
	) uut (
		.LOCK(locked),
		.RESETB(1'b1),
		.BYPASS(1'b0),
		.REFERENCECLK(clock_in),
		.PLLOUTCORE(clock_out)
		);

In this project, the pll is contained in its own module (pll.v) that is created by the IceStorm project tool icepll.

Amazingly, the 48MHz signal can also be generated by dividing down a faster clock. Using an icepll configured pll at 192MHz, a 48MHz signal can be derived the simple way, by divider, and (possibly unnecessarily) put through the global buffer network for distribution. This worked for a few experiments, but probably requires more development and testing to confirm. Perhaps it can be done better.

wire clk_192mhz;
wire clk_locked;

// Use an icepll generated pll to give us 192MHz
pll pll192( .clock_in(pin_clk), .clock_out(clk_192mhz), .locked(clk_locked) );

reg [4:0] reset_counter = 0;

// Generate the slower clock
reg       clk_48mhz_logic;
reg [1:0] clk_divider;
always @(posedge clk_192mhz ) begin
    if ( ~clk_locked ) begin
        clk_divider <= 0;
    end else begin
        case (clk_divider)
            0: begin
                clk_48mhz_logic <= 1;
                clk_divider <= 1;
            end
            1: begin
                clk_divider <= 2;
            end
            2: begin
                clk_48mhz_logic <= 0;
                clk_divider <= 3;
            end
            3: begin
                clk_divider <= 0;
            end
      endcase
  end
end

// This clock has to go places.  Put it through a global buffer
wire clk_48mhz;
SB_GB gbc(
    .USER_SIGNAL_TO_GLOBAL_BUFFER (clk_48mhz_logic),
    .GLOBAL_BUFFER_OUTPUT ( clk_48mhz ) );

Device Pins

The pins in the interface to the module (pin_usb_p and pin_usb_n) are the raw device pins, the direction control logic is internal to usb_uart_i40. The original usb_uart module is retained below to facilitate reuse in other architectures.

Somewhere the USB pull up pin has to be asserted. This is done in the top level code, since usb_uart doesn't manipulate it.

assign pin_pu = 1'b1;

Pipeline Interface

There are two pipeline interfaces to the module, uart_in and uart_out. The former being a pipeline receiver and the latter being a pipeline sender. The naming can be a bit confusing. uart_in is into the module, out of the device, into the host, uart_out is out of the host, into the device, out of the module.

These can just be tied together (as they are in this project) to create a UART loopback.

Use

Development has been done on Ubuntu.

Clone the repo

git clone [email protected]:davidthings/tinyfpga_bx_usbserial.git

and enter the directory

cd tinyfpga_bx_usbserial

And make it!

make

The make process has got to do a few things so it may take a minute.

If it completes successfully, press the "program" button on the TinyFPGA BX and program it

make prog

If all went well, you will then see a new port which you can connect to with a serial terminal emulator (on Ubuntu the port is /dev/ttyACM0 and a good terminal program is gtkterminal since it doesn't seem to get upset if you forget to disconnect before reprogramming).

For this demo, characters typed are rather unspectacularly just echo'ed back.

make gui

Let's you perform the place and route manually and generate pretty pictures like the ones above.

Dependencies

Nothing is required beyond the usual tools needed for TinyFPGA development. This is a command-line makefile project.

Icestorm

http://www.clifford.at/icestorm/

Make sure you get NextPNR.

TinyFPGA BX

https://tinyfpga.com/bx/guide.html

Debug

There is some debug instrumentation left in the code. This will just disappear if you don't use it. For the Device Host bug, there were 16 lines connected to the hardware to present as much state as possible from the USB stack for debugging. Here's one resulting screenshot where the interface is working to near capacity.

• D0 - trigger (to send a 40-ish byte block if the last one was done)
• D1 Pipeline Data[ 0 ] (toggles when presented with a sequence of odd-even data)
• D2 Pipeline Valid - there is data available
• D3 Pipeline Ready - the data can be accepted
• D4 USB SERIAL IN State Machine[0] - controlling the pipeline from FPGA to Host (usb_uart_bridge_ep.v)
• D5 USB SERIAL IN State Machine[1]
• D6 in_ep_data_done (asserted when a partial outgoing buffer should be sent)
• D7 in_ep_data_free (asserted while the usb serial sysyem can accept data)
• D8 - EP State Machine[0] - current_ep_state (usb_fs_in_pe.v)
• D9 - EP State Machine[1]
• D10 - IN TX State Machine[0] - in_xfr_state (usb_fs_in_pe.v)
• D11 - IN TX State Machine[1]
• D12 - Bytes available to Tx - this indicates that there are bytes to send the USB transmitter
• D13 - Byte to Tx - this is the strobe of the next byte into the USB transmitter
• D14 - Receiver Message Start - the USB receiver has received a message start (an IN token or Ack here)
• D15 - Receiver Message End - the USB receiver has received a message start (an IN token or Ack here)

Overall, you can see the data source pipelining data (D1) into the buffer whenever the USB port is ready (D3). When a buffer is full, the interface waits (EP State Machine D9,D8 = 1,0) until it gets an IN Token (D14,D15). When this happens, the waiting buffer is sent as fast as the USB lines can send send them. The interface waits for an ACK (D14,D15) and when all is well, signals up the stack that it's ready to fill another buffer. And the cycle restarts.

Issues

Pipeline Interface

Run slowly, the pipeline is pretty simple. Data is available (valid), receiver (signaling ready) receives it. Wait. Repeat. You can certainly use it in a simple way by de-asserting ready every cycle forcing the interface into a two step (valid ready, valid ~ready). The logic to do this is pretty simple.

Where the pipeline shines is when data is streaming - one in, one out every clock cycle, (as happens in multibyte transfers), but for a simple seeming interface, and one that appears in near all FPGA code, a pipeline like this can be fiendishly complex to implement, especially for modules both receiving and sending data. Implementers must have a firm grasp on their ='s vs <='s, know their always @(*)'s from their always @(posedge clk)'s, and really understand combinatorial vs registered signals.

Here's a taste. Consider a pipeline module, m with both upstream and downstream communicating partners. The main area of complexity happens when data is streaming (valid and ready are asserted by all). Every clock, data is flowing from the upstream module into m, and from m into the downstream module below it. Suddenly the downstream module gets busy or is itself held up for some reason and it lowers ready. When this happens module m has to latch the data that it is currently sending (since data is transferred only when both parties agree), hold valid high and wait. Meanwhile the upstream module's data was considered transfered. So the poor module m in the middle has to both latch its data word and store the new overflow one and just hang out. Then when the downstream module finally signals that it's ready (raising the ready line), module m first needs to give it the original data, then the stored overflow data, and only then can it start accepting data again (raising its ready to the upstream module). Getting all this in your head and implemented correctly can take days (or weeks)

Mr ZipCPU talks about pipelining a lot here - https://zipcpu.com/blog/2017/08/14/strategies-for-pipelining.html . He uses "STB" and "Busy" as his signals, and his development of the subject is pretty nice. I haven't used his code. There are a few places where I have skeptical squint-eye, and so to not end up balling and throwing things every time I implemented something pipeline-y, I ended up writing my own pipeline code from scratch.

So, feel free to use this interface, or put another one on it! Maybe someone will create a whole suite of pipelined modules to do stuff...

Command Line

There are rumors that some people don't like command-line tools. If you (or a friend) think you might be one of these people, please feel free to connect this project to your other fancy environment (Apio, PlatformIO) and submit the appropriate files so others may experience these delights.

Accuracy / Mistakes / Etc.

Please feel free to suggest fixes / improvements / changes.

Fixed Issues

Removed Extra Endpoint

There was an extra unused endpoint burried in the code. Removing it results 2% less LUT use, and improves the chances of a design passing timing. Thanks very much to Olof Kindren for taking the time to understand this code and post the PR.

Design Passing Timing

Originally some long constants were creating very long carry chains. The code could never be made to pass at 48MHz. This was dangerous since some parts under some conditions might not work. Now it does pass!

Info: Max frequency for clock 'clk_48mhz_$glb_clk': 51.60 MHz (PASS at 48.00 MHz)

For reference, there were a few places like the following:

usb/usb_fs_out_pe.v:185

              if (ep_get_addr[ep_num][5:0] >= (ep_put_addr[ep_num][5:0] - 2)) begin

usb/usb_fs_out_pe.v:212

          ep_get_addr_next[ep_num][5:0] <= ep_get_addr[ep_num][5:0] + 1;

In these cases, the constant is interpreted as 32b and the whole operation is conducted at that width. 32b long carry chains can not be done at 48MHz in the iCE40. The fix was for these needed to become explicitly shorter.

usb/usb_fs_out_pe.v:185

              if (ep_get_addr[ep_num][5:0] >= (ep_put_addr[ep_num][5:0] - 6'H2)) begin

usb/usb_fs_out_pe.v:212

          ep_get_addr_next[ep_num][5:0] <= ep_get_addr[ep_num][5:0] + 6H'1;

So that the full 32b addition and subtraction was not conducted. This is possibly an issue with yosys (and may even be already fixed) since the Xilinx Vivado tools just silently ignore all the extra bits. This kind of problem occurs in a few other places. See the relevant commit for more.

Device-Host Bug

If the device sent data when the host was not ready, bad things happened. If the device sent more than 32 bytes in a single burst, bad things happened. Only good things happen now from single bytes up to streams almost at link capacity.

32 Character Bug

Turns out that was all me. Fixed now.

Back Pressure Bugs

The OUT EP interface (characters from the host) would stutter a little sometimes when the pipeline signaled ~ready. This is fixed now.

Send Timeout

The IN EP (characters to the host) employed a very aggressive policy for buffer termination. Whenever the pipeline went dry - ~valid it would send the bytes already received, initiating a whole packet turnaround. This is not ideal when a pipeline may stop for a cycle or two periodically. So now when there's a pause, the IN EP waits for up to 8 cycles before dispatching the packet.