With add_memory_region, user needs to provide the memory origin, which should not be needed since
could be retrieved from mem_map and prevent automatic allocation which is already possible for csr
and interrupts.
New add_mem_region method now allows both: defining the memory origin in mem_map (which will then
be used) or let the SoC builder automatically find and allocate a memory region.
Hard IP blocks are fixed in location, so long/deep combinational paths routing to multiple hard IP blocks can lead to timing closure problems.
XADC and MMCM DRPs currently have their DEN pins triggered by the ".re" output of a CSR. This is asynchronously derived from a fairly complicated set of logic that involves a logic path that goes all the way back through the cache and arbitration mechanisms of the wishbone bus. On more complex designs, this is leading to a failure of timing closure for these paths, because the hard IP blocks can be located in disparate portions of the chip which "pulls" the logic cluster in opposite directions in an attempt to absorb the routing delays to these IP blocks, leading to non-optimal placement for everything else and thus timing closure problems.
This pull request proposes that we add a pipeline delay on these critical paths. This delays the commit of the data to the DRP by one cycle, but greatly relieves timing because the pipeline register can be placed close to the cluster of logic that computes addresses, caching, and arbitration, allowing for the routing slack to the hard IP blocks to be absorbed by the path between the pipe register and the hard IP block.
In general, this shouldn't be a problem because the algorithm to program the DRP is to hit the write or read CSR, and then poll the drdy bit until it is asserted (so the process is already pretty slow). The MMCM in particular should have almost no impact, because MMCM updates are infrequent and the subsequent lock time of the MMCM is pretty long. The XADC is potentially more problematic because it can produce data at up to 1MSPS; but if sysclk is around 100MHz, adding 10ns to the read latency is relatively small compared to the theoretical maximum data rate of one every 1,000ns.
Note that the xadc patch requires introducing a bit of logic into the non-DRP path. This is because without explicitly putting an "if" statement around the logic, you fall back to the non-blocking semantics of the verilog operator, which ultimately leads to a pretty hefty combinational path. By having a default "if" that should get optimized out when DRP is not enabled, when the DRP path /is/ enabled the synthesizer knows it can safely push the async signal into a simple mux as opposed to worrying about enforcing the non-blocking operator semantics to get the desired result.
Revert to treating SDRAM_DFII_PIX_[RD|WR]DATA CSRs as arrays
of bytes, but use the new uintX_t array accessors for improved
legibility, and to avoid unnecessary byteswapping.
Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
Implement CSR accessors for all standard integer types and
combinations of subregister alignments (32 or 64 bit) and
sizes (i.e., csr_data_width 8, 16, or 32).
Rename accessors to better reflect the size of the register
being accessed, and correspondingly update the generation
of "csr.h" in "export.py".
Additionally, provide read/write accessors that superimpose arrays
of standard unsigned C types over a CSR register (which may itself
be spread across multiple subregisters).
Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
A bridged/crossover UART can now just be created by:
- passing uart_name="stream" to SoCCore/SoCSDRAM.
- adding a crossover UART core to the design:
# UART Crossover (over Wishbone Bridge
from litex.soc.cores.uart import UART
self.submodules.uart_xover = UART(tx_fifo_depth=2, rx_fifo_depth=2)
self.add_csr("uart_xover")
self.comb += self.uart.connect(self.uart_xover)
This version of the UART adds a second, compatible UART after
the first. This maintians software compatibility, and allows a
program running on the other side of the litex bridge to act as
a terminal emulator by manually reading and writing the second
UART.
Signed-off-by: Sean Cross <sean@xobs.io>
For the "P" side of the analog channels, actually, connecting
a digital line to them has "no meaning". The docs say that
either you connect an analog pin to a pad, or vivado "ties it off
appropriately". I wish it were the case that tying a pin to 0 or 1
would actually connect it to a power or ground, because it means
that even in unipolar mode you have to burn two pins to break out
the signal of interest *and* the ground reference analog pad
(I thought I could just connect it to "0" and the pin would be
grounded, but that doesn't happen -- it's just ignored if it's
not wired to a pad).
For the pad specifier, is it OK to leave it with an optional
argument of analog_pads=None? I tried assigning to the
self.analog property after instantiation, but this doesn't
seem to work, the default values are preferred. It looks like
if you don't want to do the analog_pads= optional argument
the other way to do it would be to add code on the instiating
module that tampers with the properties of the instance directly,
but I think that's sort of ugly.
Also, I noticed you stripped out the layout specifier for
the analog_pads. I thought it would be nice to provide that
in the file, so the caller doesn't have to infer what the
pad layout is by reading the code...what's the motivation for
removing that?
Similarly to how CSRBank subregisters are aligned to the CPU word
width (see commit f4770219f), ensure SRAM word_bits are also aligned
to the CPU word width.
Additionally, fix the MMPTR() macro to access CSR subregisters as
CPU word (unsigned long) sized slices.
This fixes the functionality of the 'ident' bios command on 64-bit
CPUs (e.g., Rocket).
Signed-off-by: Gabriel Somlo <gsomlo@gmail.com>
The design is backward-compatible in functionality for users
who don't want to use DRP. That is, on power on, the XADC will
scan the supply and temperature and store them in CSRs.
If drp_enable is set, the scanning stops, and the XADC is now
controlled by the DRP bus.
Wher drp_enable is reset, the XADC may return to an auto-sample
mode, but only if the internal registers are configured to do this.
If you return to drp_enable without, for example, turning on
the continuous sequence and setting which channels to check,
the results will be unpredictable (mostly either it'll scan just
once and stop, or it'll not scan all the channels, depending on
the register settings).
At this point, the backward compatibility was confirmed in testing,
the DRP API is still a work in progress as the application this
is being developed for needs to support fun stuff like real time
sampling of signals to a buffer.
Down the road, this block may have to be modified again to support an
output FIFO, so we're not railing the CPU trying to do real time
sampling of ADC data. This will probably be added as a True/False flag
of some sort in the parameter list, because the FIFO will be expensive
as far as BRAM goes to implement and applications that don't need the
FIFO buffer can probably use that BRAM for better things.
