If Tx source synchronous clock is not routed through clock capable pins
the interface and driving logic must run on the Rx interface clock.
This introduces a dependency, Rx interface must be bring up before the
Tx. In this mode a Tx only operation is not possible.
This is done through a synthesis parameter.
Expose this parameter to the software so it can query if the limitations
exists in the implementation.
If the REMOVE_NULL_BEAT_EN is set, in FIFO mode, all the beats with a
NULL TKEEP will be removed from the AXI stream.
This feature is used initially in data_offload, to create a continues and
cyclic TX data stream for DACs, when the IPs in the path have different data
widths.
Dual clock mode is introduced in link layer to support different
datapath widths on the transport layer than on physical layer.
- Link clock : lane rate / 40 for input datapath width of 4 octets 8b10b
- Device clock : Link clock * input data path width / output datapath width
Supports four clock configurations, single or dual clock mode with or
without external device clock.
The configuration interface reflects the dual clock domain.
Support multiple clock monitors in a block.
Before this change the clock monitor had to be named with a fix name
preventing multiple instances of the clock monitor.
Add parameter that describes interface to link layer, this must be
integer multiple of octets per frame.
Add parameter that describes interface to user/DMA, this must be
multiple of bytes so software can process the samples easier.
Make synthesis parameters accessible for the drivers.
Rework implementation to reflect the parameters of the actual core and
not of the AXI interfacing core.
To support deterministic latency with non-power of two octets per frame
(F=3,6) the interface width towards the transport layer must be resized
to match integer multiple of frames.
e.g Input datapath width = 4; Output datpath width = 6;
for F=3 one beat contains 2 frames
for F=6 one beat contains 1 frame
The width change is realized with a gearbox.
Due the interface width change the single clock domain core is split
in two clock domains.
- Link clock : lane rate / 40 for input datapath width of 4 octets 8b10b
- lane rate / 20 for input datapath width of 8 octets 8b10b
- lane rate / 66 for input datapath width of 8 octets 64b66b
- Device clock : Link clock * input data path width / output datapath width
Interface to transport layer and SYSREF handling is moved to device clock domain.
The configuration interface reflects the dual clock domain.
If Input and Output datapath width matches, the gearbox is no longer
required, a single clock can be connected to both clocks.
In order to keep resource utilization low and for better timing closure
allow disabling of the character replacement logic.
If the parameter is set the frame alignment monitoring is limited to links
where scrambling is on.
Add support to JESD204 RX and TX core for 8-byte 8b/10b link mode,
and frame alignment character replace/insert with or without scrambling.
Add support for xcelium simulator to jesd204/tb
Increased cores minor version.
There are boards (e.g. AD4630-24) which take the SCLK and echo back to
the FPGA through a level shifter - doing this removes the effect of
round-trip timing delays from the level shifter. This is commonly done
whenever isolators are used since they are very slow.
By setting the ECHO_SCLK parameter to 1, the IP will use the incoming
echoed SCLK clock to latch the SDI line(s). The sdi_data_valid is still
synchronous to the SPI clock, and it's generated after the last valid
SDI latch.
The designer's responsibility is to time the SDI shift registers in order to
respect the design requirements.
The new REG_PRBS_CNTRL and REG_PRBS_STATUS registers expose controls of internal
PRBS generators and checkers allowing the testing the multi-gigabit serial link
at the physical layer without the need of the link layer bringup.
In phase aligned mode the fPLL uses an external feedback path to better
align the phase of the PLL output to the phase of the external reference
clock.
This mode is required for deterministic latency to be able to sample SYSREF
which is source synchronous to the external reference clock signal.
So far phase aligned mode had been disabled since manual PLL calibration
would fail in this mode under certain (unknown) circumstances and dynamic
reconfiguration of the PLL would not work.
The latest Intel Arria 10 transceiver datasheet contains instructions for
the proper calibration sequence to make it work when the PLL is configured
for phase aligned mode. Software has been updated accordingly, so enable
phase aligned mode.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Refactor the AXI4 stream FIFO implementation.
- Define a single address generator which supports both single and double
clock mode. (synchronous and asynchronous)
- Fix FIFO status bits (empty/full). NOTE: In asynchronous mode the
flags can have a several clock cycle delay in function of the upstream/downstream
clock ratio.
- In synchronous none FIFO mode (ADDRESS_WIDTH==0), the module acts as
an AXI4 stream pipeline.
Allow channels received from dma to re-map to other channels, e.g. allowing
broadcasting the same channel to all channels.
The feature is selectable with synthesis parameter and disabled by default.
Build a large mux from smaller ones defined by the REQ_MUX_SZ parameter
Use EN_REG to add a register at the output of the small muxes to help
timing closure.
This commit adds two fields:
1. source channel selection - Sets the channel number the for the source data.
2. DMA enable mask - When this bit is set do not drive the enable line
towards the DMA interface.
This feature will allow the user to hold(indefinitely) the last sample, from an
ongoing DMA transfer, simple or cyclic(stooped by user or trigger).
This commit also adds as functionality option:
-synchronized stop between the two channels(DMAs)
-stop by trigger
This reverts commit 0402ce85e4
and reverts commit 164aa97ec3.
The trigger pulse generation must be handled outside of the
SPI Engine framework.
It is recommanded to be done in system level using a PWM
generator or an external signal.
fixed critical warnings generated when the NUM_OF_CLK_MONS parameter
is set to 0 and the constraints written in up_clock_mon_constr.xdc
cannot be applied; removed up_clock_mon_constr.xdc from ip core.
When the link is disabled the events can be ignored.
This is required by the free running event counter that can catch
invalid events during startup cased for example by an invalid link clock.
If the lane looses synchronization due invalid characters or disparity
error the lane alignment monitor checks random input which results in
irrelevant reporting of frame alignment error.
If all lanes are synchronized (CGS state machine is in DATA phase) for long
enough therefore the link is also synchronized/DATA phase reset the error
counter since the accumulated values during INIT/CHECK are irrelevant.
These errors are handled by the per-lane CGS state machine.
All errors accumulated during INIT/CHECK phase of CGS are relevant only
if the link is unable to reach the DATA phase.
The link stays in DATA phase unless software resets it,
so all errors accumulated during the DATA phase are relevant.
The previous implementation of the de-glitch only delayed the assertion
of the SYNC phase by 64 clock cycles with the DEGLITCH state but if meanwhile
one of the lanes got into a bad state cgs_ready de-asserted the state machine
continued to go SYNCHRONIZED (DATA) state.
This change extends the required number of consecutive cycles while all lanes
must stay in data phase before moving the link to SYNCHRONIZED state from 8 to 256;
This increases the reliability of link bring-up without needing extra
link restarts from software side.
Add statistics for :
- number of link enable events
- number of interrupt events (regardless of mask)
0x0B0 0x2C0 Stats Control Register
[0] - Write 1 to clear stat registers
0x0B1 0x2C4 Link Enable Stat Register
[15:0] Number of times the link was enabled from power-on or from last
stat clear
0x0B4 0x2D0 IRQ Stat Register 0
[31:16] IRQ 1 counter
[15:0] IRQ 0 counter
0x0B5 0x2D4 IRQ Stat Register 1
[31:16] IRQ 3 counter
[15:0] IRQ 2 counter
0x0B6 0x2D8 IRQ Stat Register 2
[31:16] IRQ 5 counter
[15:0] IRQ 4 counter
0x0B7 0x2DC IRQ Stat Register 3
[31:16] IRQ 7 counter
[15:0] IRQ 6 counter
Quartus Standard 19.1 throw a critical warning for registers that have
different reset and initial power-up level.
Do not initialize those registers so we can get rid of the warning.
Define both AXI4 Memory Mapped and microprocessor interface for the
reigster map, then activate/deactive one of it in fucntion of the memory
interface type parameter.
Define the missing status_sync interface, which should be connected to
the offload.
Context switching with a parameter is not a good idea. The simulator
may evaluate both branch of the IF statement, even though the inactive
branch may not be valid.
Use if..generate to make the code more robust for both synthesizers and
simulators.
Converting from RGB to YCbCr takes one less stage than converting
from YCbCr to RGB color space.
Moving extra delay stage(5), of the sync signals to a particular
YCbCr to RGB color space conversion case.
Define INPUT_PIPELINE parameter, which can be used to activate the
REGISTER_INPUTS parameter of the PHY. This parameter will add an
additional register stage into the incoming parallel data stream.
It can be used to relax the timing margin between the PHY and Link modules.
This patch contains an initial effort to support the Stratix 10
architecture in our JESD204 framework.
Several instances were updated, doing simple context switching using the
DEVICE_FAMILY system parameter:
- xcvr_reset_control
- lane PLL (ATX PLL)
- link PLL (fPLL)
- native XCVR instance
Apart from the slightly different parameters of the instances above,
there were small differences at the reconfiguration Avalon_MM interface.
The link_pll_reset_control is required just for Arria10, so in case of
Stratix10 it isn't instantiated.
In Stratix 10 architecture there are several additional ports of the
xcvr_reset_control module that must be connected to the native XCVR
instance or tied to GND.
The following xcvr_reset_control ports were defined and connected to the
XCVR:
- rx|tx_analogreset_stat
- rx|tx_digitalreset_stat
- pll_select
If dac_valid is not a constant '1' it gets synchronized with the
dac_data_sync signal. This causes that dac_valid never asserts while
dac_data_sync is high, this way skipping the phase initialization.
ADRV9001 interfacing IP supports the following modes on Xilinx devices:
A B C D E F G H
CSSI__1-lane 1 32 80 80 2.5 SDR 8
CSSI__1-lane 1 32 160 80 5 DDR 4
CSSI__4-lane 4 8 80 80 10 SDR 2
CSSI__4-lane 4 8 160 80 20 DDR 1
LSSI__1-lane 1 32 983.04 491.52 30.72 DDR 4
LSSI__2-lane 2 16 983.04 491.52 61.44 DDR 2
Columns description:
A - SSI Modes
B - Data Lanes Per Channel
C - Serialization factor Per data lane
D - Max data lane rate(MHz)
E - Max Clock rate (MHz)
F - Max Sample Rate for I/Q (MHz)
G - Data Type
H - DDS Rate
CSSI - CMOS Source Synchronous Interface
LSSI - LVDS Source Synchronous Interface
Intel devices supports only CSSI modes.
De-assert dac_rst together with an updated control set.
This allows writing the control registers before releasing the reset.
This is important at start-up when stable set of controls is required.
De-assert adc_rst together with an updated control set.
This allows writing the control registers before releasing the reset.
This is important at start-up when stable set of controls is required.
Allow monitoring of non-PN patterns which have zeros in it.
e.g. nible-ramp, full range ramp.
Singular zeros got ignored if not out of sync, while OOS_THRESHOLD
consecutive zeros or non-matching data asserts the out of sync line.
Fix the *_ip.tcl scripts for axi_spi_engine and spi_engine_offload
module.
In case of a bool parameters the value_format and value properties must
be set for both user and hdl paramters. If not, in the generated verilog
code the tool will use "true" or "false" strings, instead of 0 or 1.
The input data path has a delay section that compensates for the ADC path delay.
By using a Dynamic Shift Registers coding style we can improve/change the
resource utilization on m2k:
Before After Resources
LUT 10097 10048 48 (0.28%)
LUTRAM 516 540 -24 (-0.4%)
FF 15285 14803 482 (1.37%)
The number of delay taps in the LA data path can be controlled manually, from
the regmap or automatically, according to the axi_adc_decimate's rate.
Moreover, because the rate is configure by software, and the time of
initialization, is different for the ADC path and LA path. There is an
uncertainty of plus/minus one sample between the two. Because ADC and LA
paths share the same clock we can easily synchronize the two paths. We
can't use reset, because the rate generation mechanism is different
between the two. So the ADC path is used as master valid generator and we
can use it to drive the LA path.
The synchronization is done by setting the rate source bit. This
mechanism can only be used if the desired rate for both path is equal,
including oversampling fom ADC decimation.
Adds information on:
- Log 2 of interface data widths in bits
- Interface type (0 - Axi MemoryMap, 1 - AXI Stream, 2 - FIFO ) .
Lets the driver discover interface widths and interface type settings,
this will deprecate the corresponding device tree properties.
This is useful in case of parametrized projects where the width of
the datapath is changing. This change will allow the use of a generic
device tree node.
Updated version to 4.3.a
Optimize the oversampling mechanism.
The behavior of the axi_dac_interpolate was changing if a debug module was
added to the core.
The current code has a better utilization and reliability.
When using an oversampling of 2 for axi_dac_interpolate the rate was
the same as with oversampling by 1(bypassing).
This commit removes the bypass for the ratio of 2.
For projects where the clock ratio between the sampling clock and core clock
is higher than 2, the ad_dds generates a number of samples equal with
the clock ratio. There is a phase offset between the samples, proportional
with the requested DDS frequency.
In scenarios where the DDS out frequency is closer to the upper
limit(Nyquist) and/or the clock ratio is also greater than 2 and the
dac_data_sync reminds low for an extended period of time, the DAC will
receive at each core clock period, a number of samples equal with the
clock ratio and with an amplitude influenced by the DDS out frequency.
In most cases similar with a sawtooth signal.
With this commit we ensures that samples received by the DAC are 0 for
the period where dac_data_sync signal is high. Only when the signal
transitions to low, the phase accumulator is initialized and the phase
information is passed to the phase to amplitude converter.
Another issue can appear when the sync signal is too short; less then
CLK_RATIO * clock cycles. Because the phase accumulator will not
synchronize at all stages, the final result will be a random combination of
sine-waves. Added a minimum sync pulse after the dac_data_sync is set
low.
When frame alignment error monitoring is enabled and error threshold is met
at least for one lane, generate an interrupt so software can reset the link and
do further bring-up steps.
Add support for RX frame alignment character checking when scrambling is enabled and
for link reset on misalignment.
Add support for xcelium simulator to jesd204/tb
The Pattern generator is part of the axi_logic_analyzer core.
The trigger signal can be internal (Oscilloscope or Logic Analyzer) or
external(TI or TO pins).
The sdo_enabled and sdi_enabled control lines are generated from the
current state of the CMD bus.
In case of a delayed SDI latching the sdi_enabled can be deasserted at
the moment of the last valid bit, losing the generation of the sdi_data_valid
signal, which eventually cause a data loss, or even deadlock on software driver.
To make the logic mode robust, latch the value of the CMD[9:8] at every
transfer command. Doing so the sdo_enabled and sdi_enabled control lines will
store the last active transfer command state and they will be
independent of the current state of the CMD bus. This way we can add
longer time delay to the SDI latching if it's necessary.
Having the same name for dac and adc TPLs creates conflict in the
address segment naming having random names associated to the segments.
This causes difficulties during scripting of the project in test bench
mode.
The value of the HDL parameter NUM_OF_SDI can be read out from the
register at address 0x0C. The same register contains the value of the
DATA_WIDTH.
The register has the following bit layout:
[15: 0] DATA_WIDTH
[23:16] NUM_OF_SDI
[31:24] 8'b0
Forward the offload's sync_id to the register map, by defining an
additional AXI stream interface between the offload and axi_spi_engine.
The last sync_id of the offload module can read out from the
register 0x00C4. It also can generate and interrupt if the irq mask is
configured accordingly.
There is a major compatibility issue between 2019.1 and 2019.2.
The file system_top.hdf got a different file extention. This will
cause a compilation failer in the end of the build. To save time
and fail earlier, upgrade the version mismatch message to ERROR.
If user still wants to build a branch with different tool version
the variable ADI_IGNORE_VERSION_CHECK should be set to 1.
The external synchronization signal should be synchronous with the
dac clock. Synchronization will be done on the rising edge of the signal.
The control bit is self clearing. Status bit shows that the synchronization
is armed but the synchronization signal has not yet been received
Added EXT_SYNC parameter to be able to keep the dac_sync original
behavior
The external synchronization signal should be synchronous with the
adc clock. Synchronization will be done on the rising edge of the signal.
The control bit is self clearing. Status bit shows that the synchronization
is armed but the synchronization signal has not yet been received. While
the synchronization mechanism is armed, the adc_rst output signal is set
The current format should allow for the SYSREF signal to be used as
synchronous capture start, but will need to be disabled before the
synchronization mechanism is armed
The commit 9ab88f1200 introduced a new
feature for the execution module, which provides the possibility to
delay the SDI line latch with one or more core clock cycle. Unfortunatly
the implementation was not correct and the SDI line was latched at the
wrong time.
This patch fix the aligment of the shift register and the SDI_DELAY parameter,
to latch the SDI line of the physical interface at the right time.
Improve the description of the feature.
In some cases, the Vivado version can contain other characters than just
numbers. One such example is after applying the patch of AR# 71948,
which makes `version -short` return something like `2018.3_AR71948`.
This patch changes the version check to ignore anything after the first
two components of the version.
Add definition for new ultrascale device packages.
The package information is used by software for xcvr calibration.
At the moment, the factors that are influencing the calibration for the new
packages are not clear.
The previous mechanism was "probing" the DMAs for valid data. Better said,
each interpolation channel enabled it's DMA until a valid data was received,
then it disabled the DMA read and waited for the adjacent channel(DMA) to
receive a valid data. Only when for both channels had valid data on the
DMAs interfaces was the transmission started. This added an undesired and
redundant complexity to the interpolation channels. Furthermore, for continuous
transmission, using the above mechanism lead to a fixed phase(sample)
shift between the two channels at each start.
By using the streaming mechanism the interface is simplified and the
above problems are solved.
For Intel projects:
In cases where the clock of source synchronous interface is not routed
through a clock capable pin the DPA receive mode can't be used. Instead
the clock will be routed through a clock buffer from an IO to the clock
tree and from there to the IOPLL.
implemented mux for temp reading either from internal or external
source; updated regmap; added param to identify source for temp
information; updated tacho measurements; added AVG_POW param used
for tacho measuremet average useful for simulations; defaults for
tacho measurements changed to params and added registers; added
prescaler for fsm control, FSM updated; changed register write
process; connected INTERNAL_SYSMONE to regmap, value can now be
read by software;
parameters with same names were duplicated with transceiver specific
names due different default values.
This does not scales very well.
Use same name for the parameters as for other parameters and do the
default value handling in the IP configuration layer.
In order to help timing closure on multi SLR FPGAs add a pipeline stage
between the link layer and physical layer. This will add a fixed amount
of delay to the overall latency.
Laszlo Nagy
Istvan Csomortani
stefan.raus
Aaron Holtzman
Matt Blanton
sergiu arpadi
Lars-Peter Clausen
AndreiGrozav
Adrian Costina
Josh Blum
Stanca Pop
Rodrigo Alencar
Mathias Tausen
Maxim