CSR_BASE is currently defined twice. Once in mem.h as the base
of the CSR region in the SoC address space, and once in csr.h
as the base address for all CSRs.
This fixes two issues with those definitions:
- The mem.h one is unconditional which prevents an external
redefinition (which is useful under some circumstances such as
when using an address decoder outside of LiteX with a standalone
core).
- The csr.h one is actually the origin of the first CSR region
rather than the origin of the CSR region in the SoC space. They
are usually the same ... unless you don't have CSR bank 0 in
which case the csr.h one becomes different. This causes conflicts
with the mem.h definition and breaks projects using a standalone
cores.
The first one is fixed by adding the #ifndef/#endif around the
definition of the memory regions, the second one by passing the
csr_base to use to get_csr_header()
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When running in sim, those delays can take a *long* time, which
isn't always necessary with the simulated litedram PHY.
This allows system.h to optionally set CONFIG_SIM_DISABLE_DELAYS
which causes cdelay to do nothing.
This is especially useful when using a verilated litedram inside
a bigger/slower simulated design as to not spend a huge amount
of time going through the initializations.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Having common cpu variants/extensions has no real additional value since we are supporting
very various CPUs where minimal/standard/full have different meanings. Checking against
common variants/extensions has also cause more issues recently when adding new CPUs than
the additional value it was supported to provide.
So let's just simplify things: a CPU provide the supported variants and we just check
against that.
This commit:
* adds the support for I2S standard mode,
* extends I2S left justified mode,
* allows to configure sample size for tx/rx in 1-32 bits range,
* implements I2S master mode,
* allows to concatenate channels or used the padded mode.
This required to rework the FSM.
Have all the new compound accessors be written in terms of the simple
ones and fix how CSR_ACCCESORS_DEFINED can be used to override the
simple ones but keep the definitions of the other ones around.
This *should* also also fix incorrect multiple accesses done
by 64-bit CPUs to 32-bit CSR busses, and make the accessors not
depend on CONFIG_CSR_ALIGNMENT being the same as sizeof(unsigned long)*8
In addition, the generated csr.h now will include system.h
always when with_access_functions is True. This guarantees that the
higher level accessors are defined. The extern prototypes for the
simple accessors when CSR_ACCCESORS_DEFINED are removed and system.h
is responsible for providing them. It is also added to hw/common.h
This allows system.h to set CSR_ACCCESORS_DEFINED when necessary, in
which case it's responsible for both declaring and defining the simple
accessors. That way, it can make them inline rather than forcing them
to be extern which at least on microwatt saves spaces.
One can continue to use -DCSR_ACCCESORS_DEFINED but in that case a system.h
will have to be provided with at least the extern definitions.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
- do the CSR alignment update only if CPU is not CPUNone.
- revert PointToPoint interconnect when 1 master and 1 slave since this will
break others use cases and will prevent mapping slave to a specific location.
It's probably better to let the synthesis tools optimize the 1:1 mapping directly.
- add with_soc_interconnect parameter to add_sdram that defaults to True. When
set to False, only the LiteDRAMCore will be instantiated and interconnect with
the SoC will not be added.
This should also improve Wishbone timings.
Tested on iCEBreaker:
./icebreaker.py --cpu-type=None --uart-name=uartbone --csr-csv=csr.csv --build --flash
With the following script:
#!/usr/bin/env python3
import sys
from litex import RemoteClient
wb = RemoteClient()
wb.open()
# # #
print("scratch: 0x{:08x}".format(wb.regs.ctrl_scratch.read()))
errors = 0
for i in range(2):
for j in range(32):
wb.write(wb.mems.sram.base + 4*j, i + j)
for j in range(32):
if wb.read(wb.mems.sram.base + 4*j) != (i + j):
errors += 1
print("sram errors: {:d}".format(errors))
# # #
wb.close()
Currently, we create a wishbone interface with the default address
width (30 bits) for the bridge. Instead, create an interface that
has the same number of address bits as the CSR bus.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When creating standalone IP cores such as standalone LiteDRAM without
a CPU, the CSR are presented externally via a wishbone with just enough
address bits to access individual CSRs (14), and no address decoding
otherwise. It is expected that the design using such core will have
its own address decoder gating cyc/stb.
However, such a design might still need to use LiteX code such as
the sdram init code, which relies on the generated csr.h. Thus we
want to be able to control the CSR base address used by that generated
csr.h.
This could be handled instead by having the "host" code provide
modified csr_{read,write}_simple() that include the necessary base
address. However, such an approach would make things complicated
if the design includes multiple such standalone cores with separate
CSR busses (such as LiteDRAM and LiteEth).
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When creating a standalone LiteDRAM core with no CPU, there is
no need to create a wishbone slave to LiteDRAM interface.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This creates a lot of useless churn in the resulting verilog. Instead
use a point to point interconnect in that case.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
