correctly no matter how the "gpio_defaults_block.mag" and
"gl/gpio_defaults_block.v" are defined. Previously it assumed
that they both defined all bits as zero, which was not the case
for the layout. Now both define bit value 0x0402 and the script
can flip bits either direction as needed in both verilog and
layout
it was configured for the caravel_pico SoC, with the housekeeping
SPI able to access the upper 256-word section of the memory if the
CSB bit in the housekeeping control register is cleared. This
testbench tests both access through housekeeping and access directly
from the SoC through the memory-mapped address.
architecture type passed to gcc as the value to the '-march='
option an environment variable, setting that environment variable
to "rv32imc" by default, and overriding it with "rv32ic" specifically
for the new caravel_pico without the multiplier and divider option,
on testbenches "mem" and "storage" which both have multiplies in the
C code.
caravel_pico repository. The issue is that each SoC implementation
defines its own modules and therefore needs its own includes. The
implication is that this file now needs to exist in every SoC
implementation's verilog/gl/ directory.
connections to the core side VCCD1 and VSSD1 on the clamped3 pads.
Also added scripts for running LVS on chip_io to the mag/ directory,
and revised the scripts so that they will only re-run extraction if
there is no netlist file in the mag/ directory.
gate-level simulations from running. Corrected caravan_netlists.v,
which did not have the same change made yesterday to caravel_netlists.v
for the DLL.
knowledge and which went undetected since before MPW-one. Modified
the "pll" and "sysctrl" testbenches so that they run and measure
something useful. Both exercise the clock monitoring on GPIO
outputs functions. The PLL test also runs the digital locked
loop (behavioral verilog). The PLL test overlaps sysctrl, but
"pll" cannot be run on gate level verilog, whereas "sysctrl" can.
(input for monitoring) changed from being connected directly to the
user project (where it shouldn't be) to the same signal on the input
side of the management protect block (where it should be). This is
functionally the same. Checked for any other signals connected
directly from the user project to any block other than mgmt_protect,
didn't find any (good). Modified the gate-level netlists and top-level
layouts for caravel and caravan with the corresponding change. This
was the only change affecting layout. Also: Revised the "pll"
testbench. This is still ongoing work. Also: Fixed the way the
pins on I/O pads are declared in chip_io.v, mprj_io.v, and pads.v, so
that it isn't so bizarre. Most of this change is functionally
agnostic (just a change in the way the ifdefs work), but did fix an
incorrect ifdef that causes the whole user power domain to be broken.
which is now done through setting an environment variable to point to the
location of the management SoC wrapper. Added the missing user project
wrappers to the GL directory (copied from the original caravel repository),
and also the GL version of chip_io_alt. Modified the caravan_netlists and
caravel_netlists files to import the correct list of gate level netlists,
which has been reduced by moving "include" statements for components of the
management SoC into the management SoC repository (e.g., caravel_pico).
the change in the implementation of the serial loader, which split the load
signal out as a separate bit, and therefore had a separate bit-bang entry).
to housekeeping and the management core wrapper to separate the
wb_cyc_i signal and connect to new signal hk_cyc_o on the
management core. Also: Fixed a dangling input (user_clock) on
the housekeeping (minor error caused by the earlier refactoring
and unnoticed because there is no testbench covering that
function).
blocks; that is, there are special versions of the block for the
first 6 GPIO pins. That should allow the GL netlists to simulate,
although the end goal is to have the gen_gpio_defaults.py script
modify the GL netlists to exactly match the configuration, as is
done for the .mag layouts.
add a separate signal for the houskeeping wb_cyc_i wishbone signal,
instead of combining it with the user project's wb_cyc_i. This
change makes it compatible with the LiteX implementation of the
wishbone bus.
to resolve the problem with the typo that caused the propagated
GPIO serial load, reset, and clock signals to get scrambled on the
user2 side. Caravel is now LVS clean again (Caravan needs layout
work).
verilog for both caravel and caravan. Hand-edited the RTL and GL netlists to
correct this; still need to correct the layouts. The error causes the user1
side clock, load, and reset buffers to drive the user2 side as well as the user1
side, making a huge mess of the routing. Will route this by hand.
Fixed rstb_h, which was being input to low-voltage blocks. (2)
Fixed flash_csb_ieb_core and flash_clk_ieb_core, which were not
output from housekeeping as they should be; the solution was
to tie the INP_DIS lines low at the pad by connecting them to
the TIE_LO_ESD line. This should probably be addressed in
housekeeping but would change the current pinout.
problems that had been fixed recently in caravel, and which cause
the caravan testbench to break, but which were not noticed; (2)
corrected the count of gpio_control_block modules, which was one
off, with two of them overlapping (not sure how that even passes
simulation, but it did); (3) fixed a power connection in the
caravel chip_io, which should have caused chip_io to fail LVS,
so apparently LVS was not run on chip_io. . .
got removed from its net by a chang in the LEF view of an I/O pad, and
a lack of declaration of an array to attach to pwr_ctrl_out in the
verilog, which is valid verilog but netgen can't know the bus size
without the no-connect net being declared. The remaining issue has to
do with separation of ground domains in the mgmt_protect block.
(clock -> clock_core in caravel_clocking, VPWR -> vccd_core and
VGND -> vssd_core in the instances of modules that were pulled from
the management SoC to the top level).
longer in the PDK but have been folded into larger library files.
With the most recent push to open_pdks to fix an error in the I/O
verilog library, the verilog testbenches once again pass.
deleted "resetb_core_h" port label. Corrected the chip_io and chip_io_alt
verilog RTL files to replace the user area power supply clamp cells with
the new clamped3 cell from open_pdks.
three parameterized values used in the RTL verilog. Modified the
"user_defines.v" file to create verilog definitions that match the C-style
definitions from "defs.h", for convenience/simplicity.
fixed by moving from after the managment protect to before it, but
an inversion of the signal was missed, leading to an incorrect
wb_rst_i passed to housekeeping. (2) Revised the method to load
the serial GPIO data chain from a 2-pin, I2C-like method to a
more straightforward 3-pin method with separate reset, clock, and
load pins. The load pin propagates through the chaing like the
other two. Added a bit-bang signal for the load signal as well.
(3) Added an implied buffer after the data output of the GPIO
control block to ensure that the data arrives at the next control
block after the clock, to prevent hold violations.
incorrectly assigned to the clock on the user side of the managment
protect block, causing it to be undefined when the user area power
supply is down. The "hkspi_power" testbench which tests using the
housekeeping SPI while the user area power is grounded now works
correctly.
around the entire padframe. The vccd1 and vccd2 domains are local
to their respective pads, and any bus routing must be done inside
the padframe. This means that all pads operate on global vddio for
3.3V as before, but also global vccd for 1.8V. The user 1.8V voltage
domain only goes as far as the input to the GPIO control block.
to areas not part of the management SoC, including chip_io,
mgmt_protect, and mprj_ctrl and sysctrl_wb (which are now part of
housekeeping). The two housekeeping tests fail and need to be
debugged.
management SoC to the management protect block, and updated the
documentation to include the new wishbone interface for those
signals, and the additional signals on the management SoC pinout.
an unconnected wishbone bus (unconnected inputs). Added the missing
signals for the user IRQ enables to management protect (which have
to come from the management SoC).
implement the arbitration between SPI and back-door. The back-door
access flags when it is going to do a read or write, and the SPI
can have an invalid read or fail a write if the SPI is too fast,
but the wishbone access should be valid. As long as the SPI is
much slower than the core clock (say, 1MHz) then there should be no
contention, which means that contention can always be avoided simply
by slowing the SPI signaling down.
status) so that between rdstb and wrstb, the SPI signals when it is
about to read or write a byte. The back-door wishbone interface then
stalls the CPU during these periods. That allows the CPU to continue
running while the SPI is being accessed without data collisions and
without having to stall for the entire time CSB is held low.
Because SCK is asynchronous to the clock, rare collisions are still
possible; this is not expected to be an issue but might be worth
investigating.
via programming. The values for each of the GPIOs at power-up are
defined in the "user_defines.v" file. For the verilog, they are
applied as parameters. For the layout, they will need to be
separately defined cells for each of the GPIOs, or at least for
each set of unique default values.
since the original one was folded into the sysctrl testbench, but that
testbench no longer uses the SPI master. Moved the SPI master from being
an overlay of the housekeeping SPI to occupying GPIO pins 32 to 35.
Made GPIO 35 a bidirectional pin like 36 and 37 so that the output enable
from the SPI master can be used.
storage. Not all of these pass simulation checks. Added back the
bit-bang control of the GPIO programming. Added back the read-only
interface between the housekeeping module and the SRAM 2nd port.
Revised the memory map text document to reflect the addition of the
SRAM ports. There is not yet a testbench for the SRAM read-only
interface.
the 26 domain (now dedicated to the housekeeping module), with
2e0... now 261... and 2f0... now 262... Although this is not
strictly backwards-compatible, the addresses in defs.h have been
modified so that C code remains valid with a recompile.
Tim Edwards
manarabdelaty
jeffdi