[Oberon] Oberon FPGA hardware point of view

[Chris Burrows]

> -----Original Message-----
> From: Oberon [mailto:oberon-bounces at lists.inf.ethz.ch] On Behalf Of
> Walter Gallegos
> Sent: Wednesday, 8 August 2018 4:04 AM
> To: oberon at lists.inf.ethz.ch; Walter Gallegos
> Subject: [Oberon] Oberon FPGA hardware point of view
> 
> A FPGA hardware designer point of view;
> 
> In some projects (all my projects) the CPU executes software as
> coprocessing; in parallel but outside the main data flow of hardware
> DSP; hardware is faster and more efficient than software DSP.
> 
> On this scenery, the memory map could change from one project to
> another project. So, hardware/software designers need certain degree
> of freedom to access memory mapped areas.
> 
> I propose two modifications :
> 
> 1/ Add memory mapped variables
> 
> VAR [label] : [type] AT [address]
> VAR [label] : ARRAY [size] OF [type] AT [address]
> 

On the surface this sounds OK but in reality the sheer number of addresses
involved led us to use a different approach.

Our technique follows Paul Reed's recent advice here i.e. separate all of
the hardware specific details from everything else into the lowest level
modules. This has worked out really well for us. It avoids the software
maintenance nightmare of having to track down / modify hard-coded addresses
and other hardware-specific details scattered throughout a system.

With the Astrobe for ARM Cortex-M3, M4 and M7 Oberon systems we were faced
with the prospect of having to maintain hundreds of memory-mapped addresses
for similar, but different, memory-mapped peripherals for more than 60
different types of microcontroller from two different manufacturers (NXP and
STMicroelectronics). You might expect to have a couple of different sets of
common definitions for each manufacturer but we ended up needing 11
altogether. Not as bad as 60 perhaps but I'm convinced the designers could
have done a lot better at eliminating inconsistencies. 

The scheme we have used is this: There is a single module, called MCU.mod,
for each family of microcontrollers e.g. LPC176x (NXP Cortex-M3), STM32F7
(STM Cortex-M7) etc. There are eleven of these, all with the same name but
stored in a folder named after the microcontroller family. 

The *key* feature is: the *only* items contained in the MCU module are CONST
declarations. 

Each named constant represents a peripheral register address. Much of the
time just the base address of each peripheral is different from one MCU to
another so we take advantage of the fact that CONSTs can contain arithmetic
expressions. We can then specify a single base address for each device and
the other related registers are common offsets from that base. 

e.g. 

for UART0:

  U0Base* = 04000C000H;
  U0RBR* = U0Base+000H;
  U0THR* = U0Base+000H;
  U0DLL* = U0Base+000H;
  U0DLM* = U0Base+004H;
  ...

For UART2:

  U2Base* = 040098000H;
  U2RBR* = U2Base+000H;
  U2THR* = U2Base+000H;
  U2DLL* = U2Base+000H;
  U2DLM* = U2Base+004H;
  ...
  ...

RBR, THR, DLL, DLM etc. are the names used for each UART function exactly as
used by NXP in their programming reference manual. In case you are
wondering, yes - RBR, THR and DLL all map to the same absolute address.

Now, just to be different, the corresponding base address for the LPC1347
family is U0Base* = 040008000H. If that wasn't bad enough, sometimes the
relative offset addresses are different as well! 

There are more than 500 of these definitions in the LPC176x version of
MCU.mod and there are still a number of peripherals that we have not yet
included.

Now that we have isolated what is *different* in MCU.mod, we can then
implement a common hardware-interface module (e.g. Serial.mod) which uses
these constant definitions, with their generic names, when implementing the
(still hardware-specific) functions:

  IMPORT MCU, SYSTEM;

  PROCEDURE PutCh*(ch: CHAR);
  BEGIN
    REPEAT UNTIL SYSTEM.BIT(ULSR, 5);
    SYSTEM.PUT(UTHR, ch)
  END PutCh;


The next level up in the module hierarchy is the familiar common
*hardware-independent* 'Out' module. This works with all the different
microcontrollers providing functions Out.Char, Out.String. It includes the
statement:

  IMPORT Serial;

And the functions just call PutCh in different ways.

The mechanism that we use to specify which particular MCU.mod and Serial.mod
files are actually used when we compile an application which targets a
particular family of microcontrollers is to associate the application with a
configuration file containing mcu-specific 'search paths'. An extract from
the map file for an application called 'Info' shows the consequences:

LPC1769:
MCU             D:\AstrobeM3-v6.4\Lib\LPC1769\MCU.arm
Out             D:\AstrobeM3-v6.4\Lib\General\Out.arm
Serial          D:\AstrobeM3-v6.4\Lib\LPC1769\Serial.arm
Info            D:\AstrobeM3-v6.4\Examples\General\Info.arm

LPC1347:
MCU             D:\AstrobeM3-v6.4\Lib\LPC1347\MCU.arm
Out             D:\AstrobeM3-v6.4\Lib\General\Out.arm
Serial          D:\AstrobeM3-v6.4\Lib\LPC1347\Serial.arm
Info            D:\AstrobeM3\Release\Examples\General\Info.arm

Most of the time the only functions we need to use with these CONST
addresses are SYSTEM.PUT to write a value, SYSTEM.GET to read a value and
SYSTEM.BIT to test the value of a single bit. 

If multi-byte data needs to be efficiently read or written to an Oberon
ARRAY or RECORD, SYSTEM.ADR, SYSTEM.COPY and SYSTEM.VAL (or ARRAY OF BYTE
parameters) are the Oberon features that can be used. Once the conversion
has been done from a byte-stream to an application-specific Oberon data
structure at the hardware interface the Oberon data structure can be used
naturally in the application from then on. It only needs to be converted
back to a byte-stream when the hardware needs to be updated. 

Regards,
Chris Burrows
CFB Software
http://www.astrobe.com