This time I want to give a broad overview of how the FPGA part of the design works.
The OpenVizsla FPGA design (available in the GitHub OV repository was written using Migen, “a Python toolbox for building complex digital hardware”. Migen allows to write logic as the result of a Python script, and compile them to Verilog (and ultimately an FPGA bitstream). This has interesting implications, as it allows for the full power of the Python language when building language features.
Let me bring up an FPGA-centric block diagram of the OpenVizsla design. Don’t panic - I will explain the individual blocks and provide some examples. You can even click parts in the block diagram to see some detailed description. Now if that isn’t magic… (And please excuse my limited website usability skills.)
(Did I tell you that you can click in the block diagram to see more details of the blocks?)
The main SDRAM chip
A regular MT48LC16M16A2 or compatible. It has a 16-bit data bus, and runs (by default) at 100 MHz, which gives a theoretical bandwidth limit of 200 MByte/s. The effective data rate will be less, because there are dead cycles when switching between pages and when doing refresh.
Wikipedia has an article that describes the basics.
SDRAM Host Controller
This is the internal SDRAM logic. Internally, a
HostIf-Interface is used to interact with the SDRAM controller.
HostIf, you first issue a request using the
i_*-signals. This can
be either a read or write (
i_wr=1 is a write issue,
i_wr=0 is a read
issue), to a specific start address (
i_addr). The controller
acknowledges the issue, after seeing your
i_stb=1, by asserting
After that, there is the data phase. Whenever
d_stb (which is controlled
by the SDRAM controller) is asserted, data is transferred, unless
For writes this means that the data on
written to the next address, and for reads, data is available on
There is no other flow control, so the client has to take the data (or
deliver the data) as fast as the controller wants. If you can’t keep the
rate, you have to end the issue (
d_term=1), and then start a new issue.
Here is an example transfer on this bus:
- Write AD50, AD51, …, ADB3 to address 000050, 000051, …, 0000B3
- Read back from 000055, …, 000058
d_stb de-assertions are arbitrary, meaning that they depend on how busy
the system is. A reason for
d_stb could be that we hit a column limit (and
the next page has to be selected), or that another master took over.
HostIf-clients always need to support
d_stb de-assertions at any time.
- ov_types.py defines the HostIF interface.
- SDRAMCTL interfaces to the physical SDRAM chip, and provides a single
HostIf-interface on the other side.
- SDRAMMux time-slices the
HostIf-interface to multiple masters.
- SDRAMBIST is one of such masters, and first writes a test-pattern, and then reads it back.
- SDRAMBISTCfg provides a CSR interface to the
BankArray scans modules that inherit from
AutoCSR and builds a list of
Bank then contains a number of a CSRs. A
addresses on the CSR bus into write enables and selects the right register
to return data from.
FTDI data mux (
FTDI_sync245 modules connects to the FTDI chip on the one side, and to
CmdProc module on the other side, and provides bidirectional data
transfer. It handles the FTDI SyncFIFO protocol in a state machine and
provides the clock domain crossing from the internal clock to the FTDI 60MHz
CmdProc connects to the FIFOs provided by
FTDI_sync245, and connects
the HOST-to-FPGA FIFO to a
BusDecode instance, and the FPGA-to-HOST FIFO
BusDecode processing data bytes sent by the host, and generates CSR
master transactions. Here is an example CSR master transaction. The data
from the host was
55 80 03 00 D8 (with the last byte being a checksum),
and this decodes to a CSR write at address
0003 with data being
It also generates the CSR response packets; for each read and write transfer on the bus, you get a copy of the transaction back via the FTDI.
In the opposite direction the
BusEncode takes multiple sources - such as
the CSR response, or the USB packet data, or the LFSR test data, and
switches them into a single stream. It relies on each source to assert the
last signal whenever a packet is complete. Switching then happens at those
Here we see two pending packets (as indicated by
stb being asserted in
both clients), and the round-robin muxer first selects client1 (which is a
CSR response packet) and then, once client1 asserts
last, it switches to
USB capture logic
This large block is in charge of:
- Configuration of the external ULPI chip to sniff mode.
- Packetizing the ULPI data, and insertion of control (
The ULPI and data processing is decidedly out of scope of this blog entry.
But to understand the big picture: The output of the
is a packetized byte stream again, similar to the CSR slave results. It gets
muxed into the stream sent to the analysis host in the
The LFSR Generator
The LFSR stream generator has a
Source that outputs a packetized LFSR stream. It is configured using two CSRs:
RANDTEST_SIZE: The number of LFSR bytes in each packet.
RANDTEST_CFG: A single ‘go’ bit. If it is set, packets will be generated, otherwise the generation will stop.
- FTDI_randtest implements it.
External IO FPGA Logic
BTN_status connects to the external physical button, and provides the
ability to read the button status via a CSR register
LED_outputs allows to select from a number of LED sources via a per-LED
Source 0 is the
LEDS_OUT register, so by default, you can display an
arbitrary LED pattern by writing it to
In the top-level module
OV3, a number of LED sources are selected:
# from ovhw/top.py: # GPIOs (leds/buttons) self.submodules.leds = LED_outputs(plat.request('leds'), [ [self.bist.busy, self.ftdi_bus.tx_ind], [0, self.ftdi_bus.rx_ind],  ], active=0)
This means that LED0 can be switched between
OUT, the BIST
busy signal, and the FTDI bus
LED 1 can be switched between
OUT, a static 0, and the FTDI bus
active=0 causes the
LED_outputs module to invert the leds
leds_raw.eq(leds if active else ~leds)) since they are active-low.
LEDs are connected with a current-limit resistor to pins P57, P58, P59 on the FPGA against VCC. To drive a LED, the pin must be driven to GND.
- LD1,LD2,LD3,LD4 in the Schematics
The external push-button connects P67 of the FPGA with GND. A 10K pull-up resistor provides a positive input when the button is not pressed.
- SW1 in the Schematics
The ULPI and data processing is decidedly out of scope of this blog entry.
An FT2232H is used for communication with the host.
The FTDI website has a document that describes the SyncFIFO mode that we’re using.
SyncFIFO mode allows transferring data at a very high speed using an 8-bit bidirectional BUS clocked at 60MHz, and some control lines. The actual speed is depending mostly on how fast the host can process the data, but is of course inherently limited to the USB 2.0 High-Speed efficiency of about 90%.
FTDI outputs, FPGA inputs:
RXF#: Data is transferred from FTDI to FPGA when both
RXF#is driven high when there is no data to be read.
TXE#: Data is transferred from FPGA to FTDI when both
TXE#is driven high when there is no space for data to be stored.
CLKOUT: 60 MHz clock generated by FTDI chip. All signals are synchronous to this clock.
FPGA outputs, FTDI inputs:
RD#: Acknowledges the current byte, and causes the FTDI to load the next byte onto the bus if
RXF#is low during that cycle.
WR#: Strobe for the current byte if
TXE#is low that cycle.
OE#: Controls bus direction. Must be asserted at least one cycle before driving
SIWU: “Send Immediate” pin. Can be used to flush the FTDI buffer to lower latency, but we’re not doing this right now.
On the left we see the FT2232H interface chip to the analysis host. It
provides a bi-direction fast data link to a host PC running
incoming command stream from the PC is parsed in
CmdProc, turned into CSR
(Configuration registers) master transactions.
BankArray is a collection
of CSR slaves. One such slave is the
BTN_status component which allows
reading the state of the external hardware button, another one is
controlling the LEDs, other allow access to the ULPI controller, others can
invoke BIST sequences in SDRAM or configure SDRAM-to-host streaming.
Migen automatically wires up modules to a CSR infrastructure, so it’s easy to add registers that can be accessed from the host.
Again, please click on the various elements in the block diagram to see a description and waveforms!