Cesar Strauss [Sat, 1 Oct 2022 22:00:55 +0000 (19:00 -0300)]
Leave shadow / die proof for last
Cesar Strauss [Sat, 1 Oct 2022 14:48:19 +0000 (11:48 -0300)]
Start of formal proof of MultiCompUnit
Create a "random" ALU, controlled by the formal engine.
Add an "issue" transaction counter and ask the fomal engine to
give an example.
Jacob Lifshay [Tue, 16 Aug 2022 06:43:13 +0000 (23:43 -0700)]
change goldschmidt_div_sqrt to use nmutil.plain_data rather than dataclasses
Luke Kenneth Casson Leighton [Sun, 14 Aug 2022 21:09:13 +0000 (22:09 +0100)]
grr not a yield function
Luke Kenneth Casson Leighton [Sun, 14 Aug 2022 18:50:46 +0000 (19:50 +0100)]
add get_fpregs stub function to HDLstate
Luke Kenneth Casson Leighton [Wed, 6 Jul 2022 13:22:54 +0000 (14:22 +0100)]
update pinmux submodule, rename to "fabric"
Luke Kenneth Casson Leighton [Wed, 6 Jul 2022 08:11:47 +0000 (09:11 +0100)]
add fabric compatibility mode
Luke Kenneth Casson Leighton [Tue, 5 Jul 2022 13:20:11 +0000 (14:20 +0100)]
MulOutputData was only 64-bit output not 128-bit
meaning that mulhd is truncated (always zero)
https://bugs.libre-soc.org/show_bug.cgi?id=855
Luke Kenneth Casson Leighton [Mon, 4 Jul 2022 15:12:46 +0000 (16:12 +0100)]
add signal for resetting trap internal state (kaivb cache)
Luke Kenneth Casson Leighton [Mon, 4 Jul 2022 15:07:29 +0000 (16:07 +0100)]
set msr_o.data not msr_o Record in trap main_stage.py
Luke Kenneth Casson Leighton [Sun, 26 Jun 2022 11:57:33 +0000 (12:57 +0100)]
adapt TRAP function in main state pipeline to put KAIVB
into top bits
Luke Kenneth Casson Leighton [Sun, 26 Jun 2022 11:38:52 +0000 (12:38 +0100)]
store KAIVB SPR 850 in TRAP Pipeline
by redirecting OP_MTSPR and OP_MFSPR to it
https://bugs.libre-soc.org/show_bug.cgi?id=859
Luke Kenneth Casson Leighton [Sun, 26 Jun 2022 11:11:42 +0000 (12:11 +0100)]
reduce icache/dcache TLB sizes
Luke Kenneth Casson Leighton [Sun, 26 Jun 2022 11:11:14 +0000 (12:11 +0100)]
update trap test_pipe_caller.py to use up-to-date test method
Luke Kenneth Casson Leighton [Sun, 26 Jun 2022 11:10:35 +0000 (12:10 +0100)]
missing module argument to TestRunner execute
Luke Kenneth Casson Leighton [Fri, 17 Jun 2022 14:13:11 +0000 (15:13 +0100)]
convert trap test_pipe_caller.py to consistent format
(like alu test_pipe_caller.py)
Andrey Miroshnikov [Mon, 23 May 2022 22:17:07 +0000 (22:17 +0000)]
Change usage of WB sel for individual control
Luke Kenneth Casson Leighton [Sun, 1 May 2022 17:56:27 +0000 (18:56 +0100)]
split out front of div into separate stage, still too long combinatorial
blocks
Luke Kenneth Casson Leighton [Sat, 30 Apr 2022 21:20:02 +0000 (22:20 +0100)]
add missing module
Luke Kenneth Casson Leighton [Sat, 30 Apr 2022 21:17:58 +0000 (22:17 +0100)]
split off CR0/XER production in DIV Function Unit into separate stage
due to massive combinatorial chains
Luke Kenneth Casson Leighton [Sat, 30 Apr 2022 12:52:00 +0000 (13:52 +0100)]
clear out DEC in core.cur_state.dec due to spurious interrupt.
this is slightly complicated. the STATE regfile contains pc, msr, svstate,
dec, and tb, being a reflection of CoreState. reading from STATE regfile
is on a one-clock delay. the DEC/TB FSM needs to decrement DEC and increment
TB, by reading from the STATE regfile and then writing a new value.
of course, the SPR pipeline has to get a word in edgeways as well.
but...
the complication comes in that it is the PowerDecoder2 which receives
a *cached* copy of DEC, and this cached copy is what has (up until now)
been out-of-date with what is in the STATE regfile.
the hack-job-solution is to zero-out the cached copy when the SPR pipeline
writes a new value to DEC. the DEC/TB FSM will then rewrite a correct
value into it.
given that PowerDecoder2 only uses the MSB of DEC (and the EE bit of MSR)
to determine whether to fire an interrupt, this should be perfectly fine.
Cesar Strauss [Sat, 30 Apr 2022 11:08:27 +0000 (08:08 -0300)]
Implement transparent read port option on the XOR wrapper SRAM
Add a bypass path when simultaneously reading and writing to the same
address.
There needs to be an independent Mux for each write lane, since we may not be
writing on all lanes, necessarily.
Jacob Lifshay [Fri, 29 Apr 2022 18:56:49 +0000 (11:56 -0700)]
fix waay-too-precise error requirements
reduces 128x64->64-bit divider to 220k cells rather than 700k cells
Luke Kenneth Casson Leighton [Fri, 29 Apr 2022 09:32:24 +0000 (10:32 +0100)]
add option to set small cache sizes in
issuer_verilog.py
Jacob Lifshay [Fri, 29 Apr 2022 06:05:08 +0000 (23:05 -0700)]
add comment
Jacob Lifshay [Fri, 29 Apr 2022 06:04:05 +0000 (23:04 -0700)]
fix so HDL works for 5, 8, 16, 32, and 64-bits.
Jacob Lifshay [Fri, 29 Apr 2022 05:40:32 +0000 (22:40 -0700)]
HDL works for io_width=5
Cesar Strauss [Thu, 28 Apr 2022 22:37:20 +0000 (19:37 -0300)]
Test simultaneous transparent reads and partial writes
Should catch some corner cases.
Jacob Lifshay [Thu, 28 Apr 2022 09:26:09 +0000 (02:26 -0700)]
add docs for clz
Jacob Lifshay [Thu, 28 Apr 2022 09:19:01 +0000 (02:19 -0700)]
add WIP HDL version of goldschmidt division -- it's currently broken
Luke Kenneth Casson Leighton [Thu, 28 Apr 2022 08:21:42 +0000 (09:21 +0100)]
notes added to setup.py - absolute paranoia is needed on
what dependencies the HDL has. dropping USD 16 million on 7nm Mask Charges
you absolutely cannot have arbitrary software downloaded off the internet
from external sources.
Jacob Lifshay [Thu, 28 Apr 2022 05:52:41 +0000 (22:52 -0700)]
move GoldschmidtDivState
Jacob Lifshay [Thu, 28 Apr 2022 05:50:42 +0000 (22:50 -0700)]
add FIXME comments
Jacob Lifshay [Thu, 28 Apr 2022 05:39:08 +0000 (22:39 -0700)]
add the goldschmidt sqrt/rsqrt algorithm, still need code to calculate good parameters
Jacob Lifshay [Wed, 27 Apr 2022 23:21:46 +0000 (16:21 -0700)]
switch cached-property dependency to using libre-soc's git repo
Jacob Lifshay [Wed, 27 Apr 2022 06:50:37 +0000 (23:50 -0700)]
improved goldschmidt division algorithm parameter optimization algorithm
Jacob Lifshay [Wed, 27 Apr 2022 04:57:40 +0000 (21:57 -0700)]
split out non-derived params into separate class without all the accuracy checking stuff.
Jacob Lifshay [Wed, 27 Apr 2022 04:52:08 +0000 (21:52 -0700)]
split out n_hat as separate property
Jacob Lifshay [Wed, 27 Apr 2022 04:41:05 +0000 (21:41 -0700)]
add default_cost_fn
Jacob Lifshay [Wed, 27 Apr 2022 04:36:23 +0000 (21:36 -0700)]
move GoldschmidtDivParams.get to bottom of class
Jacob Lifshay [Wed, 27 Apr 2022 04:27:29 +0000 (21:27 -0700)]
rename _goldschmidt_div_ops to GoldschmidtDivState.__make_ops
Jacob Lifshay [Tue, 26 Apr 2022 03:56:30 +0000 (20:56 -0700)]
goldschmidt division works! still needs better parameter selection tho...
Jacob Lifshay [Tue, 26 Apr 2022 02:00:01 +0000 (19:00 -0700)]
fix goofed __init__.py file name
Jacob Lifshay [Mon, 25 Apr 2022 08:44:31 +0000 (01:44 -0700)]
working on goldschmidt_div_sqrt.py
Jacob Lifshay [Mon, 25 Apr 2022 08:44:13 +0000 (01:44 -0700)]
add cached_property dependency
Jacob Lifshay [Sat, 23 Apr 2022 02:28:37 +0000 (19:28 -0700)]
working on goldschmidt division algorithm
Luke Kenneth Casson Leighton [Fri, 22 Apr 2022 13:13:21 +0000 (14:13 +0100)]
whitespace
Jacob Lifshay [Fri, 22 Apr 2022 07:58:17 +0000 (00:58 -0700)]
add WIP goldschmidt division algorithm
Cesar Strauss [Sun, 17 Apr 2022 22:27:58 +0000 (19:27 -0300)]
Implement a 1W/1R register file, XOR style
Test case seems to be working.
Total 6x 1RW SRAMs were used, but no DFF RAM was needed.
Still needs a bypass for a transparent read option.
Cesar Strauss [Sun, 17 Apr 2022 20:19:32 +0000 (17:19 -0300)]
Formal proof of pseudo 1W/2R SRAM
Mostly taken from the proof of the 1W/1R SRAM. Sorry for all the
duplication, they are all variations of the same theme.
Cesar Strauss [Sun, 17 Apr 2022 18:44:23 +0000 (15:44 -0300)]
Add transparent option for the full read port
Cesar Strauss [Sun, 17 Apr 2022 18:13:52 +0000 (15:13 -0300)]
Implement a pseudo 1W/2R memory
This is intended for a XOR-based 6x1RW implementation of a 1R/1W
register file.
It costs 3 x 1RW memory blocks.
Luke Kenneth Casson Leighton [Sat, 16 Apr 2022 22:03:27 +0000 (23:03 +0100)]
reduce dcache/icache number of ways, to fit into ECP5 45k resource
Tobias Platen [Sat, 16 Apr 2022 20:50:29 +0000 (22:50 +0200)]
part two of issuer_fix: read pspec.microwatt_old and pspec.microwatt_debug
Cesar Strauss [Sat, 16 Apr 2022 20:44:30 +0000 (17:44 -0300)]
Check non-transparent 1W/1R SRAM wrapper
Add special case for the non-transparent case, where a simultaneous write
returns the old value of the holding register.
This completes the proof for this 1W/1R SRAM wrapper
Cesar Strauss [Sat, 16 Apr 2022 20:36:04 +0000 (17:36 -0300)]
Enable read port for non-transparent memories
For some reason, the reset value (=1) of the port enable is not recognized
by the formal engine, even if the simulation is fine with it.
Tobias Platen [Sat, 16 Apr 2022 20:23:14 +0000 (22:23 +0200)]
Merge ssh://git.libre-riscv.org:922/soc
Tobias Platen [Sat, 16 Apr 2022 20:21:34 +0000 (22:21 +0200)]
part one of issuer_fix: add parameter to issuer_verilog.py
Cesar Strauss [Sat, 16 Apr 2022 17:33:04 +0000 (14:33 -0300)]
Add port declarations to the SRAM wrappers
Cesar Strauss [Sat, 16 Apr 2022 17:25:14 +0000 (14:25 -0300)]
Change write lane signal from one-hot to binary
It's simpler to write, and actually runs faster, somehow...
Luke Kenneth Casson Leighton [Sat, 16 Apr 2022 16:21:05 +0000 (17:21 +0100)]
whoops, WBASyncBridge ack signal not wired up!
Luke Kenneth Casson Leighton [Sat, 16 Apr 2022 15:26:33 +0000 (16:26 +0100)]
select width is data_width // data granularity.
(common mistake to make, confused the hell out of me too)
Cesar Strauss [Sat, 16 Apr 2022 14:33:08 +0000 (11:33 -0300)]
Synchronize LVT state, completing the induction proof
Cesar Strauss [Sat, 16 Apr 2022 14:29:05 +0000 (11:29 -0300)]
Sync proof state with downstream memories
Luke Kenneth Casson Leighton [Sat, 16 Apr 2022 13:51:47 +0000 (14:51 +0100)]
put the old microwatt compatibility back
PLEASE DO NOT CHANGE THIS WITHOUT CONSULTATION
Luke Kenneth Casson Leighton [Sat, 16 Apr 2022 12:33:33 +0000 (13:33 +0100)]
blegh.
* add in a new PRE_IDLE state into issue FSM (to allow pc_at_reset to
get a word in edgeways)
* set nia to pc_at_reset on core_rst (not 0x0)
* add in a terrible hack where the STATE regfile entry for PC is
shifted down by one.
absolutely no frickin idea why that is needed
Cesar Strauss [Fri, 15 Apr 2022 19:43:56 +0000 (16:43 -0300)]
Complete moving the induction support into the DUT
That way, we pass the state down, to be checked by single code at
the lowest level, instead of having to check all variations of the
lower level states, at the top level.
Had to increment the proof depth, for some reason.
Cesar Strauss [Fri, 15 Apr 2022 19:36:05 +0000 (16:36 -0300)]
Fix incorrect signal widths
dbg_data is the width of a write lane (granularity) and dbg_wrote is a
single bit.
Cesar Strauss [Fri, 15 Apr 2022 16:55:00 +0000 (13:55 -0300)]
Move part of formal proof to the implementation
For induction, some state has to be synchronized with the
device under test. Make the device itself responsible for checking
this, to avoid exposing its internal workings, and making the proof
more generic.
This is done for PhasedDualPortRegfile, the rest will follow.
Luke Kenneth Casson Leighton [Thu, 14 Apr 2022 13:13:23 +0000 (14:13 +0100)]
add option Spec to XICS ICP/ICS to be able to activate (in a horrible
hack way) make_wb_layout to add stall
this is done by spotting "spec.microwatt_compat" adds stall
blegh
Luke Kenneth Casson Leighton [Thu, 14 Apr 2022 13:12:36 +0000 (14:12 +0100)]
move IRQLine out because that makes soc dependent on LambdaSOC
Luke Kenneth Casson Leighton [Thu, 14 Apr 2022 11:47:24 +0000 (12:47 +0100)]
80 char limit, remove creation of stall from ack/cyc, it has to be
done in the parent because that is what knows the usage
Raptor Engineering Development Team [Thu, 14 Apr 2022 00:57:20 +0000 (19:57 -0500)]
wb_async: Allow different feature fields for master/slave busses
Raptor Engineering Development Team [Thu, 14 Apr 2022 00:56:47 +0000 (19:56 -0500)]
Add separate memory clock register to SYSCON
Tobias Platen [Tue, 12 Apr 2022 18:37:00 +0000 (20:37 +0200)]
issuer.py: add microwatt_old and microwatt_debug options
Raptor Engineering Development Team [Mon, 11 Apr 2022 19:31:02 +0000 (14:31 -0500)]
Separate core and nest clocks in Microwatt SYSCON
Raptor Engineering Development Team [Mon, 11 Apr 2022 19:30:39 +0000 (14:30 -0500)]
Add initial wrapper for Wishbone asynchronous bridge module
Cesar Strauss [Sun, 10 Apr 2022 15:21:14 +0000 (12:21 -0300)]
Begin a formal proof of the LVT-based 1W/1R wrapper
Start with a Bounded Model Check, on the transparent case.
Cesar Strauss [Sun, 10 Apr 2022 12:24:19 +0000 (09:24 -0300)]
Implement 1W/1R with a transparent (or not) read port.
Seems that it's just a matter of the underlying memories being
transparent (or not).
Cesar Strauss [Sat, 9 Apr 2022 21:15:52 +0000 (18:15 -0300)]
Implement a true 1W/1R memory from 1RW blocks
Uses four phased memories, multiplexers, and FFRAM for the LVT.
Only non-transparent reads, for now.
Test case is included, seems to work.
Luke Kenneth Casson Leighton [Sat, 9 Apr 2022 12:36:12 +0000 (13:36 +0100)]
add a new make target for setting coldboot firmware at 0xfff0_0000
put reset values of pc and msr into Issuer explicitly on reset
Luke Kenneth Casson Leighton [Fri, 8 Apr 2022 20:10:19 +0000 (21:10 +0100)]
syntax error
Luke Kenneth Casson Leighton [Fri, 8 Apr 2022 20:10:00 +0000 (21:10 +0100)]
add dram to SysCon
Luke Kenneth Casson Leighton [Fri, 8 Apr 2022 17:41:30 +0000 (18:41 +0100)]
add SPI offset to microwatt syscon
Luke Kenneth Casson Leighton [Wed, 6 Apr 2022 11:29:03 +0000 (12:29 +0100)]
only add clock-settings on ECP5 due to special SPI clock handling
(Tercel QSPI)
Luke Kenneth Casson Leighton [Mon, 4 Apr 2022 21:32:32 +0000 (22:32 +0100)]
add tempfile to uart16550 wrapper which defines DATA_BUS_WIDTH_8
Luke Kenneth Casson Leighton [Mon, 4 Apr 2022 20:41:51 +0000 (21:41 +0100)]
disable sphinx verilg-diagrams for now
Luke Kenneth Casson Leighton [Mon, 4 Apr 2022 19:13:41 +0000 (20:13 +0100)]
allow direction-setting on each of dq0-3 in Tercel QSPI
Luke Kenneth Casson Leighton [Sun, 3 Apr 2022 23:52:27 +0000 (00:52 +0100)]
cant stand the practice of putting docstrings *after* the code they
document. a trick to get them to go on one line: semi-colons
Cesar Strauss [Sun, 3 Apr 2022 22:12:15 +0000 (19:12 -0300)]
Extend the proof to a non-transparent port
For the non-transparent case, have to distinguish the case of a
simultaneous write, where the expected value is the previous
value of the holding register, not the new.
Cesar Strauss [Sun, 3 Apr 2022 19:31:13 +0000 (16:31 -0300)]
Run formal proof on both types (even/odd) of phased SRAMs
Cesar Strauss [Sun, 3 Apr 2022 19:10:16 +0000 (16:10 -0300)]
Complete the formal proof of the pseudo dual port SRAM
Have to ensure that the holding register is in sync with the memory
contents, to avoid false positives due to unreachable states.
Since we are adding assertions, we can only strengthen the proof, even
if it does become more complex.
Cesar Strauss [Sun, 3 Apr 2022 18:50:43 +0000 (15:50 -0300)]
Implement a debug port on the pseudo 1W/1R SRAM
Exposes the debug ports of the underlying memories.
This is needed to assist the induction proof.
Cesar Strauss [Sun, 3 Apr 2022 18:24:16 +0000 (15:24 -0300)]
Formal proof of the phased write dual port memory wrapper
Similar to the proof of the 1RW memory block, but read and writes are
done on dedicated ports.
For now, only transparent regfile is tested.
To complete the formal verification, an induction proof will follow.
Luke Kenneth Casson Leighton [Sun, 3 Apr 2022 16:00:18 +0000 (17:00 +0100)]
correct default to zero string not zero int
Luke Kenneth Casson Leighton [Sun, 3 Apr 2022 10:30:44 +0000 (11:30 +0100)]
add alternative pc_reset argument to issuer_verilog.py
which propagates right the way down to core.py
next to msr_reset it is now possible to set the pc_reset value.
these actually have to go into the regfile as initial values,
which will be fun for an ASIC
Luke Kenneth Casson Leighton [Sun, 3 Apr 2022 10:29:33 +0000 (11:29 +0100)]
fix some of instantiation errors in opencores_ethmac.py
Raptor Engineering Development Team [Sat, 2 Apr 2022 21:52:49 +0000 (16:52 -0500)]
Fix opencores EthMAC module wiring
Cesar Strauss [Sat, 2 Apr 2022 20:46:29 +0000 (17:46 -0300)]
Implement transparent read ports on the phased write SRAM
Add a multiplexer to select the write memory instead of the read
memory, in case both port addresses coincide.
Cesar Strauss [Fri, 1 Apr 2022 09:47:01 +0000 (06:47 -0300)]
Implement and test a "phased write port" memory
It has one read and one write port, but the writes can only happen every
two cycles.
It uses two 1RW memory blocks.
For the moment, it's not transparent: a simultaneous read and write from
the same address returns the old value. Next step is implementing a
transparent read port.