Skip to content

Latest commit

 

History

History
397 lines (306 loc) · 13.5 KB

README.md

File metadata and controls

397 lines (306 loc) · 13.5 KB
Raw

Status

Chapters are checked off as they are (a) implemented and passes the supplied tests.

✅ Chapter 1 ✅ Chapter 2 ✅ Chapter 3 ✅ Chapter 4 ✅ Chapter 5 ✅ Chapter 6 ✅ Chapter 7 ✅ Chapter 8

Significant Implementation Differences

Hardware RAM

We do not use the RAM modules built in the project, opting to use single port RAM built into the iCE40 FPGA. This is far more efficient use of resources.

Software Pipelining

Due to the characteristics of the CPU a value loaded using an A-instruction doesn't appear in register A (aka addressM) until the next clock edge. For arithematics this is fine - the ALU will see the value previous loaded and compute the right values. For memory access however the single port RAM IP core being used will not update its on the same clock edge in which addressM is updated. It will only do so on the next edge. In other words:

  1. edge-1: addressM takes on new value, RAM output is undefined because on edge transition of the value at the address present immediately prior is clocked into the RAM.
  2. edge-2: the RAM output is now valid and reflects the value at addressM

As such every memory read needs to be preceeded by a nop (0). This is the pipeline equivalent of inserting a bubble into a pipelined CPU but we are doing it in software.

Writes have a similar issue. When outM, addressM and writeM are asserted on the same clock edge no write happens b/c their value immediately prior to transition is what matters. So it takes another nop for the write to commit.

Update 2020-04-29

I "optimised" memory access by shifting the RAM clock 1/2 period (i.e. ram_clk = ~clk). This allows addressM to be present at the rise edge of the RAM clock. This means a sequence like

@R0
D=M

No longer needs a nop. However there is no getting around the fact that each write takes an additional cycle. Additionally when M is on the RHS and the LHS, e.g. M=M+1 a nop before and after is still required.

Inserting these nops is time consuming and prone to error. tools/assembler.py will, by default, insert nops for you so programs supplied by the course can be used as-is without modification provided they are assembled using our assembler. Note that I wrote the assembler before I realised it was project 6.

eXtended Register

Project 05x extends the HACK platform to implement an additional W register. This extended platform is called HACKx and it extends the C-instruction from:

1 1  1 a c1 c2 c3 c4 c5 c6 d1 d2 d3 j1 j2 j3
⬇️
1 w d4 a c1 c2 c3 c4 c5 c6 d1 d2 d3 j1 j2 j3

Where

  • /w a forms a 2-bit vector that selects the x input of the ALU from A, inM or W corresponding to 00, 01, 10. 11 is undefined.
  • d4 works like d1..d3 but inverted. When it is UNSET the W register takes on the ALU output on the next clock.

The interpretation of w and d4 has been chosen so that programs intended for the HACK platform can run on the HACKx platform without modification. Yay backward compatibility!

I named the register W as a homage to the W register found in PIC microcontrollers.

Assembler

Macros

Macros are implemented in two places:

  1. asm.py used by vm2asm.py
  2. assembler.py

asm.py Macros

These macros are intended to help with implementing the stack virtual machine.

  1. $inc_sp, $dec_sp: increments and decrements stack pointer (SP)
  2. $load_sp, $save_sp: loads SP into A and saves A into SP
  3. $_: replaced with a unique identifier, e.g. $_LABEL can be used multiple times and each time will result in unique label, e.g. 001_LABEL, 002_LABEL.

assembler.py Macros

These macros are intended to help with direct assembly programming

  1. $const <name> <value>: inserts the symbol <name> into the symbole table with the specified integer <value>
  2. $call <label>: pushes the return address at the top of the stack then performs a jump to the specified label
  3. $return: pop the return address at the top of the stack and jumps to it
  4. $this: replaced with the most recent func_ label, e.g. func_FOO
  5. $copy_mm <dst> <src>: copies content of memory at src into dest
  6. $copy_mv <dst> <value>: copies value into memory address dest
  7. $if_var_goto <var> <dest>: jumps to dest if *var contains non-zero value
  8. $if_M_goto <dest>: jumps to dest if M contains non-zero value
  9. $if_A_goto <dest>: jumps to dest if A contains non-zero value
  10. $if_D_goto <dest>: jumps to dest if D contains non-zero value

The $call/$return macros allows for nested function calls, e.g.

(SUB_INC_D)
  D=D+1
  $return

(SUB_INC_M_AND_D)
  M=M+1
  $call SUB_INC_D

(SUB_MAIN)
  @SCREEN
  $call SUB_INC_M_AND_D

Valid Symbol Characters

The course calls for $ to be a valid character, however I accidentally used it for the macro system so it cannot be a valid identifier any more.

Numeric Constants

Our implementation of the assembler (tools/assembler.py) accepts hex and binary constants in the form of:

  • 0xNNNN for hexadecimal constants
  • 0bNNNN for binary constants

Take a page from verilog's book hexadecimal and binary constants can be spaced out using underscore(_) to improve readability.

The assembler fully supports the HACKx platform. To ensure HACK compatible machine code is emitted specify -C on the commandline. This will cause the assembler to error when it encounters use of the W register.

The assembler can optionally annotate the machine code output with the corresponding source block and the PC value if the -A option is given. This is useful during debugging.

T0-T3 Registers

Some of the R0-R15 registers serve dual purpose and only R13-15 is actually available for general use. To avoid having to remember which registers can be used freely R13-15 can also be addressed as T0-T3.

Optimisations

When -O<opt> is specified the assembler will perform some simple optimisations. <opt> can be one of:

  • all: perform all optimisations
  • loads: remove redundant loads where two (or more) A-instructions loading the same value will be reduced to one iff the A register is not modified in between
  • consec_nops: consecutive NOP (0) instructions will be collapsed into one
  • unneeded_nops: unneeded NOP (0) instructions will be removed. A NOP is unneeded if the next instructions following memory write doesn't access memory.

VM Translator

Our implementation of the vm-to-asm translator (tools/vm2asm.py) is capable of generating assembly for the HACKx and HACK platform with the HACKx platform being the default target. To run tests VM programs from the course specify -C when invoking vm2asm.py. It is not necessary to also specify -C to the assembler b/c vm2asm.py will not use the W register.

Like the assembler vm2asm.py will produce annotated assembly if -A is given.

Direct Segment Manipulation

Following the stack model religiously means incrementing a value looks like this:

push local 0
push constant 1
add
pop local 0

This sequence ultimately results in no change in the stack pointer value (2 push and 2 pops) and since incrementing by 1 doesn't require another operand we could have manipulated the value using M=M+1. This is true for other 1-operand operations our ALU is capable of, e.g. !, M-1 etc.

The VM translator implemented here supports the following direct segment manipulation commands:

  • s_inc <segment> <index>: increments the segment value directly
  • s_dec <segment> <index>: decrements the segment value directly
  • s_neg <segment> <index>: negates (-x) the segment value directly
  • s_not <segment> <index>: binary not's (~x) the segment value directly
  • s_set <segment> <index>: sets all bits of the segment value directly
  • s_clear <segment> <inex>: clears all bits of the segment value directly

Optimisations

When targetting the HACKx platform the VM translator will use the W register as a dedicated stack pointer register. This significantly reduces the number of memory access commands.

Chapters and Projects

Chapters 1-5 are implemented in their respective projects/ folder. Chapters 6 onwards live in tools/ b/c that makes the mose sense to me.

Running Tests

I have issue using the supplied software as-is, often getting opaque errors about "Expression expected on line 0". So I would often I edit the .tst files to remove the load commands and only preserve the RAM setup commands. I would then mostly manually verify the content of RAM is as expected against the .cmp files.

Development Environment

apio

This project uses apio: https://github.com/FPGAwars/apio . I have a custom fork which enables nicer testbenches: https://github.com/freespace/apio

Useful Commands

  1. apio clean: removes build artefacts
  2. apio sim -t Top_tb.v: run simulation using Top_tb.v
  3. apio build: builds and generates the necessary binary files for uploading

Generally I also use apio clean with sim or build, e.g.

apio clean; apio sim -t Top_tb.v

This ensures a clean build which is helpful in avoiding staleness or strange behaviours when switching between git branches etc.

Firmware Generation

To generate firmware.hack for use with Top_tb.v from one of the test .asm programs run something like the following in the same directory as Top_tb.v:

python ../../../tools/assembler.py -A -O all < memory_access_test.asm > firmware.hack

gtkwave Setup

  1. Install using brew cask install gtkwave
  2. Install Switch perl module into system dir: sudo cpan install Switch
  3. If required fix up permissions in /Library/Perl, e.g.
sudo find /Library/Perl -type d -exec chmod a+rx {} \;
sudo find /Library/Perl -type f -exec chmod a+r {} \;
  1. Replace /usr/local/bin/gtkwave with symlink to /Applications/gtkwave.app/Contents/Resources/bin/gtkwave

When you type gtkwave in the commandline it should launch the app.

Testbench

Testbench Template

  1. Define the DUMPSTR macro:
`define DUMPSTR(x) `"x.vcd`"
  1. Define the simulation output files:
$dumpfile(`DUMPSTR(`VCD_OUTPUT));
$dumpvars(0, <testbench_name>);

Where <testbench_name> is something like Nand_tb.

Running Testbenches

apio sim will run the first test bench it finds (alphabetic sort, ends in _tb.v).

In my branch of apio (https://github.com/freespace/apio) you can use

apio sim -t <testbench.v>

which will run the specified testbench file. This allows us to have more than 1 testbench per module per project.

Testbench Gotchas

  • When using if-statements with wires a delay is required for the wire value to update otherwise nothing seems to happen

    a = 1;
b = 1;
#10;
if (y != 1) begin
  $display("FAILED for input 11");
end

  • Assigning constants to a wire will mean it can only take on the assigned value or x, e.g.

    wire a = 0;
a = 1;
// value of a is now 'x'
a = 0;
// value of a is now 0

Importing Sources from Other Projects

On *nix use the following script to pull in all verilog files created in previous projects

find  .. -type f -name '*.v' ! -name '*_tb.v' -maxdepth 2 -exec ln -s {} . \;

This confuses git b/c git doesn't know about symlinks and thinks there are new files for it to track. Fix it with:

find . -type l | sed -e s'/^\.\///g' >> .gitignore

iCE40 UltraPlus

Technology Library

The iCE40 UltraPlus comes primitives, like the block RAM, which can be instantiated directly. They are documented in:

http://www.latticesemi.com/~/media/LatticeSemi/Documents/TechnicalBriefs/SBTICETechnologyLibrary201608.pdf

Title of the document is "LATTICE ICE Technology Library" should the URL become invalid in the future.

Yosys implements some of the primitives, e.g. SB_RAM40_4K, which are defined in

https://github.com/YosysHQ/yosys/blob/master/techlibs/ice40/cells_sim.v

Not all primitives are implemented b/c @cliffordwolf believes that it is better to let the tooling figure things out. (See YosysHQ/yosys#423).

N.B. I am using the word primitive in the same way lattice is using it. @cliffordwolf would call them macros in so far as all RAM primitives other than SB_RAM40_4K is constructed using SB_RAM40_4K.

Technical Notes

Troubleshotting

ERROR: IO 'video_sync' is unconstrained in PCF (override this error with --pcf-allow-unconstrained)

  • You have an unused module which defines input/outputs. Remove the module should remove the error

ERROR: Unable to place cell 'ROM.5.0.0_RAM', no Bels remaining of type 'ICESTORM_RAM'

  • The ROM size is too large. The icebreaker has 30 EBR units of 256x16 size. This allows a maximum of 256*30 = 7680 instructions in the ROM. Use apio build --verbose-pnr to get usage statistics.