Lantian Driver

Posted on
  • 2019-07-08
  • 13:10
  • Computers and Clients

Electric propulsion system test equipment for drone designers. Improve your drone propulsion system by testing your brushless motors and propellers up to 75 kgf of thrust and 48 Nm of torque. Measure your propeller and motor efficiency and find the best combination for your UAV. Data acquisition software included. Since a WM8731 driver module written by former students is provided on the course website, we thought adding audio was pretty easy. Audio Chip Outputs Noise¶ In the operating systems course I took in the previous semester (previous to this FPGA course), I added Sound Blaster 16 sound card support to our group's operating system. Power consumption. The operating voltage of the module is from 1.9 to 3.6V, but the good news is that the logic pins are 5-volt tolerant, so we can easily connect it to an Arduino or any 5V logic microcontroller without using any logic level converter.

  1. Networking
  2. Audio
  3. Video
  4. Keyboard

Last semester, the school offered a digital systems course, which involves development on FPGAs. At the end of the course, we need to gather in groups and achieve some complicated functionality with the flexible architecture of FPGA, such as creating a game or running a convoluted neural network. We are free to add extra functionalities as we wish.

Our group has done a game similar to Raiden, or controlling a fighter jet and attacking enemies with bullets. In addition to course requirements, we implemented these extra functionalities:

  • A 640x480 VGA framebuffer with 16 bit color depth, stored on SRAM chip
    • Naturally, Simplified Chinese fonts are included (full UTF-8 Chinese range, but without punctuation marks, since they are out of that range and we're kinda lazy)
    • Fast screen scrolling by adjusting Y offset (for flying effects)
    • SRAM controller and chip are working at 2x bus frequency, so both the CPU and the VGA controller can access data simultaneously without competition / lockups
  • Up to 8 jets (friend + enemy), fully customizable image (no palette or indexing involved), free movement across the screen
  • Up to 56 bullets (friend + enemy), customizable size and color, free movement across the screen
  • Loop playback of background music (~5min) with WM8731 sound chip, 2 channels, 16 bit depth, 8000Hz sample rate
    • Used a module provided by course (written by former students) but had pitfalls; will be discussed later
  • Internet connection with Marvell 88E1111 chip and RJ45 connectors, uploading gameplay records and downloading rankings
    • Used open source code on GitHub for data transmission (
    • Custom MDIO communication code for modifying network chip registers
    • LwIP as TCP/IP protocol stack
  • User input via USB keyboard, connected to onboard USB chip
    • Since our keyboards are unstable (common in our lab), we used a separate CPU for controlling the keyboard, so they can be reset and debugged separately, and the main CPU can be reset less often

Here is the demonstration video (in Chinese):

We used the DE2-115 development board with FPGA chip of Cyclone IV EP4CE115F29C7, and Quartus Prime 18.1 Free Edition as development software.

Thie article is a record of problems I encountered and resolved when implementing the extra features.


WARNING: This article do not contain any materials covered in the course, and is NOT HELPFUL to other projects except the final project. If you referenced this article or our code, you MUST list it in your report, or you may be punished for plagiarism. Microsensys driver download for windows 10.

We open sourced the code at

We implemented networking on the development board to upload gameplay records and generate score rankings. This is the extra feature we consider the most important. This part involves pitfalls we had with the following components:

  • Intel Triple Speed Ethernet IP (The 10/100/100M Adaptive Ethernet module that comes with Quartus)
  • Ethernet IP by GitHub User Alex Forencich (
  • LwIP Embedded TCP/IP Protocol Stack (

Intel Triple Speed Ethernet IP Won't Work¶

DE2-115 development board has 2 Marvell 88E1111 network chips builtin, corresponding to the 2 RJ45 ethernet jacks. While interfacing with the chip, we found that Quartus has provided an Ethernet IP with rich options / parameter selections, and went for it.

We had to admit that such official Intel IPs have comprehensive functionalities, such as support for various interfaces such as GMII and RGMII, builtin FIFO buffers, easy MDIO communication with Ethernet chip, and a lot of stuff we didn't need. But when we added the IP and tried to use it, no data is sent from the Ethernet chip, TX LED isn't blinking, and no data is received despite RX LED is blinking. It still wouldn't work even if we followed the official Intel guide step by step.

After debugging for as long as 3 days, we found that this Ethernet IP by Intel needs to be paid for. Since we apparently didn't do that, the IP works in trial mode, and will only work when the development board is connected to PC and Licensing window is open on Quartus Programming UI. In other circumstances, it will disable itself by not sending any data and not processing any received data.

Out of rage and frustration, we decided to replace it with an open source IP on GitHub, the second one listed above. While the IP is open source and free, it does have its pitfalls; read on.

Lacking MDIO Support with Open Source Module¶

Marvell 88E1111 Ethernet chip has a MDIO interface, providing read/write access to 32 registers. They are used to set link speed (10/100/1000M), obtain connection status (plugged in or not), etc. But we dug into the open source IP and found no trace of MDIO support.

Intel does provide a independent MDIO module free of charge, but it works with a newer protocol version of Clause 45, and cannot communicate normally with the chip, which speaks older Clause 22.

So we managed to create our own.

This module directly exposes MDIO registers on Avalon-MM memory bus, and direct memory operations can be issued to them.

Open Source IP Receives but Doesn't Send, TX LED Not Blinking¶

Now we have received packets on the FPGA, which means we are close to success. But we did a packet capture with Wireshark on the computer connected to the other end and got nothing.

If the TX LED doesn't blink, it means that data transferred from FPGA to Ethernet chip is corrupted, and the Ethernet chip will not send anything it doesn't recognize. Investigation shows that the IFG Delay parameter of the Open Source IP is set incorrectly. This parameter controls the number of gap cycles between two Ethernet frames over the cable, and is usually set to 12. We had removed the parameter previously and ignored it..

Open Source IP Receives but Doesn't Send, TX LED Blinking¶

.. And we're here, adding that parameter.

After a 3-hour reading of the code of the Open Source IP, we found that it requires data from the same packet to be sent continuously, not accepting delay in any of the cycles. The 1000Mb Ethernet module runs at 125 MHz and sends 8 bits of data to the Ethernet chip every time, but our CPU and DMA runs at 50 MHz and also sends 8 bits each cycle. Therefore the Open Source IP needs to wait for data, and when this happens it simply tells the Ethernet chip that 'data is corrupted'. Ethernet chip on the FPGA board proceeds to send the 'corrupted' signal, and the network card on the computer other end silently drops the packet, before the operating system is aware of anything.

The solution is also simple: the DMA module supports 32-bit width on the sender side, so 32-bit width is used with DMA and input side of FIFO (output side connected to Open Source IP is still 8 bits). We also increased the DMA FIFO buffer size a bit out of stability considerations and simple convenience. This time, the bandwidth of DMA is increased from 400 Mbps to 1600 Mbps, and the Open Source IP no longer needs to wait for data.

Since a WM8731 driver module written by former students is provided on the course website, we thought adding audio was pretty easy.

We were wrong.

Audio Chip Outputs Noise¶

In the operating systems course I took in the previous semester (previous to this FPGA course), I added Sound Blaster 16 sound card support to our group's operating system. At the beginning I designed the interface of the sound module to be the same as SB16, providing a memory region where audio data is stored, and triggering an interrupt whenever half of the data is read.

And the audio chip output only noise.

We thought that this interface is too complicated, and did it again with DMA and FIFO, similar to Ethernet.

Still only noise.

We began to read through the driver module, but the module is written in VHDL, a totally different language from SystemVerilog taught in the course. The design is also quite weird, as we traced an input signal and found that it ended up unused!?

We finally discovered that a signal, which we expect to be a 8000 Hz pulse signal (equal to sample rate) that indicates ready for next sample point, is not what we thought it to be.

We had to find our own 8000 Hz signal source, and we saw Intel's Interval Timer IP.

For the sake of simplicity (and laziness), we added a 8000 Hz interrupt to the main CPU, and output the audio signal via PIO in the interrupt handler. Simple but effective.

Our finished game doesn't consume a lot of CPU cycles, and can handle such high frequency interrupts.

The CPU handled one interrupt and stopped.

We read through the datasheet of Interval Timer, and found we need to write to a register after each interrupt.

After one rich day of debugging, we finally heard beautiful music coming from our board.

The video part is the difficult part of the whole project. We began with a pure framebuffer without sprites, yet our mere 50 MHz FREE CPU core, without cache, pipelining or instruction prediction, couldn't handle the workload of refreshing each pixel. Here the problem came:

FPS Less Than 1¶

Yes, that's how bad the performance went. Due to the limitation of SRAM (10 ns per operation, or 100 MHz; bus at 50 MHz max since SRAM is at double frequency with the multiplexer module), we cannot overclock it anymore.

We were not going to try the paid version of CPU IP, since we were scared of what happened with our Ethernet module.

We though compiler optimizations might do something, and added -O2. FPS instantly increased to around 7-8.

FPS Less Than 30¶

Lantian Drivers License

But this is definitely not enough for a shooter game. The minimum acceptable FPS is 30, and reaching 60 will be the best. Because of this we had to build a sprite system, where each object could be moved quickly by modifying their coordinates. Thanks to this, we reached a framerate of 60, and the game could be played normally.

In addition we added a DMA controller to the architecture. The DMA controller is a dedicated module for memory copies, and is much more efficient than the FREE CPU.

But we're not done yet:

64 Sprites Cause Logical Timeouts¶

Lantian Driverpack

Fighter jets and bullets have their transparency. When processing each pixel, a long combinational logic processing each layer is required. This means that the logic cannot properly finish in 40 ns (or 25 MHz, the pixel frequency of 640x480 VGA), and the VGA output may have artifacts.

We ended up with a tree shaped design. Since the Cyclone IV FPGA has 4 input LUTs, we designed such a structure:

  1. All 64 objects are split to 32 groups. Transparency is handled in each group, and 32 outputs are generated;
  2. Split 32 outputs to 16 groups and repeat;
  3. Split 16 outputs to 8 groups and repeat;
  4. Split 8 outputs to 4 groups and repeat;
  5. Split 4 outputs to 2 groups and repeat;
  6. Process the last 2 outputs and generate the final output.

The combinational logic path is reduced from 64 comparisons to 6.

Storing Data to SDRAM¶

Lastly, we need a way to store all the data, including fighter jets, backgrounds and audio data, to SDRAM.

In the course we are suggested to use the DE2-115ControlPanel, which loads a custom program to the FPGA and writes to SDRAM. But in previous homeworks, we found that the data on SDRAM may be randomly corrupted after switching back to our own programs.

What we did is to write to the SDRAM when we upload the program to CPU. Create a new ELF section in the BSP Editor in Eclipse, name it anything starting with a dot, such as .resources, but not the same name as anything in Platform Designer. Then, assign it to be stored in SDRAM.

Arrays can be created similar to this:

The array will be uploaded to and stored in SDRAM in this case. Since no extra programming on the FPGA itself is required, there is no risk of data damage.

Although this part is covered by course materials, the USB keyboard will extremely unstable when simply following the guides. We often had to reset the keyboard dozens of times so it can be recognized, but when it works, it works stably until the next reset.

Avoid Reset's Interference of Debugging¶

The first thing we thought of is to split the reset functionality of the main program and the USB keyboard part, so they can be debugged individually.

To be precise, we created an individual CPU that is solely used to communicate with the keyboard. The keyboard CPU and the USB chip are reset with an individual button, so it will not interfere with the main program. The two CPUs communicate with a dual-port on-chip memory, without locking mechanisms (unnecessary since it's write-only on one end and read-only on the other end).

Avoid Reset's Interference of Muscle Health¶

We added a timeout to the USB CPU's program. The CPU will reset itself if it cannot connect to the keyboard after a time period.

Lantian Drivers Ed

Actual testing shows that it is helpful, but not much. But since we are unable to solve the USB keyboard issue (even the professor cannot fix it), we had to live with such a workaround.