(no commit message)

[libreriscv.git] / 3d_gpu / architecture / decoder.mdwn

# Decoder

<http://bugs.libre-riscv.org/show_bug.cgi?id=186>

The decoder is in charge of translating the POWER instruction stream into operations that can be handled by our backend. It will have an extra input bit, set via a MSR, that will switch on GPU instructions.

Source code: <https://git.libre-riscv.org/?p=soc.git;a=tree;f=src/soc/decoder;hb=HEAD>

# POWER

The decoder has been written in python, to parse straight CSV files and other information taken directly from the Power ISA Standards PDF files. This significantly reduces the possibility of manual transcription errors and greatly reduces code size.  Based on Anton Blanchard's excellent microwatt design, these tables are in [[openpower/isatables]] which includes links to download the csv files.

The top level decoder object recursively drops through progressive levels of case statement groups, covering additional portions of the incoming instruction bits.  More on this technique - for which python and nmigen were *specifically* and strategically chosen - is outlined here <http://lists.libre-riscv.org/pipermail/libre-riscv-dev/2020-March/004882.html>

The PowerDecoder2, on encountering for example an ADD
operation, needs to know whether Rc=0/1, whether OE=0/1, whether
RB is to be read, whether an immediate is to be read and so on.
With all of this information being specified in the CSV files, on
a per-instruction basis, it is simply a matter of expanding that
information out into a data structure called Decode2ToExecute1Type.
From there it becomes easily possible for other parts of the processor
to take appropriate action.

* [Decode2ToExecute1Type](https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/decode2execute1.py;hb=HEAD)

## Link to Function Units

The Decoder (PowerDecode2) knows which registers are needed, however what
it does not know is:

* which Register file ports to connect to (this is defined by regspecs)
* the order of those regfile ports (again: defined by regspecs)

Neither do the Phase-aware Function Units (derived from MultiCompUnit)
themselves know anything about the PowerDecoder, and they certainly
do not know when a given instruction will need to tell *them* to read
RA, or RB.  For example: negation of RA only requires one operand,
where add RA, RB requires two.  Who tells whom that information, when
the ALU's job is simply to add, and the Decoder's job is simply to decode?

This is where a special function called "rdflags()" comes into play.
rdflags works closely in conjunction with regspecs and the PowerDecoder2,
in each Function Unit's "pipe\_data.py" file.  It defines the flags that
determine, from current instruction, whether the Function Unit actually
*wants* any given Register Read Ports activated or not.

That dynamically-determined information will then actively disable
(or allow) Register file Read requests (rd.req) on a per-port basis.

Example:

    class ALUInputData(IntegerData):
        regspec = [('INT', 'ra', '0:63'), # RA
                   ('INT', 'rb', '0:63'), # RB/immediate
                   ('XER', 'xer_so', '32'), # XER bit 32: SO
                   ('XER', 'xer_ca', '34,45')] # XER bit 34/45: CA/CA32

This shows us that, for the ALU pipeline, it expects two INTEGER
operands (RA and RB) both 64-bit, and it expects XER SO, CA and CA32
bits.  However this information - as to which operands are required -
is *dynamic*.

Continuing from the OP_ADD example, where inspection of the CSV files
(or the ISA tables) shows that we optionally need xer_so (OE=1),
optionally need xer_ca (Rc=1), and even optionally need RB (add with
immediate), we begin to understand that a dynamic system linking the
PowerDecoder2 information to the Function Units is needed.  This is
where power\_regspec\_map.py comes into play.

    def regspec_decode_read(e, regfile, name):
        if regfile == 'INT':
            # Int register numbering is *unary* encoded
            if name == 'ra': # RA
                return e.read_reg1.ok, 1<<e.read_reg1.data
            if name == 'rb': # RB
                return e.read_reg2.ok, 1<<e.read_reg2.data

Here we can see that, for INTEGER registers, if the Function Unit
has a connection (an incoming operand) named "RA", the tuple returned
contains two crucial pieces of information:

1. The field from PowerDecoder2 which tells us if RA is even actually
  required by this (decoded) instruction
2. The INTEGER Register file read port activation signal (its read-enable
  line-activation) which, if sent to the INTEGER Register file, will
  request the actual register required by this current (decoded) 
  instruction.

Thus we have the *dynamic* information - not hardcoded in RTL but
specified in *python* - encoding both if (first item of tuple) and
what (second item of tuple) each Function Unit receives, and this
for each and every operand.  A corresponding process exists for write,
as well.

* [[architecture/regfile]]
* [CompUnits](https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/fu/compunits/compunits.py;hb=HEAD)
* Example [ALU pipe_data.py specification](https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/fu/alu/pipe_data.py;hb=HEAD)
* [power_regspec_map.py](https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/power_regspec_map.py;hb=HEAD)

## Fixed point instructions

 - addi, addis, mulli - fairly straightforward - extract registers and immediate and translate to the appropriate op
 - addic, addic., subfic - similar to above, but now carry needs to be saved somewhere
 - add[o][.], subf[o][.], adde\*, subfe\*, addze\*, neg\*, mullw\*, divw\* - These are more fun. They need to set the carry (if . is present) and overflow (if o is present) flags, as well as taking in the carry flag for the extended versions.
 - addex - uses the overflow flag as a carry in, and if CY is set to 1, sets overflow like it would carry.
 - cmp, cmpi - sets bits of the selected comparison result register based on whether the comparison result was greater than, less than, or equal to
 - andi., ori, andis., oris, xori, xoris - similar to above, though the and versions set the flags in CR0
 - and\*, or\*, xor\*, nand\*, eqv\*, andc\*, orc\* - similar to the register-register arithmetic instructions above

# Decoder internals

The Decoder uses a class called PowerOp which get instantiated
for every instruction. PowerOp class instantiation has member signals
whose values get set respectively for each instruction.

We use Python Enums to help with common decoder values.
Below is the POWER add insruction.

| opcode       | unit | internal op | in1 | in2 | in3  | out | CR in | CR out | inv A | inv out | cry in | cry out | ldst len | BR | sgn ext | upd | rsrv | 32b | sgn | rc | lk | sgl pipe | comment | form |
|--------------|------|-------------|-----|-----|------|-----|-------|--------|-------|---------|--------|---------|----------|----|---------|-----|------|-----|-----|----|----|----------|---------|------|
| 0b0100001010 | ALU  | OP_ADD      | RA  | RB  | NONE | RT  | 0     | 0      | 0     | 0       | ZERO   | 0       | NONE     | 0  | 0       | 0   | 0    | 0   | 0   | RC | 0  | 0        | add     | XO   |

Here is an example of a toy multiplexer that sets various fields in the
PowerOP signal class to the correct values for the add instruction when
select is set equal to 1.  This should give you a feel for how we work with
enums and PowerOP.

    from nmigen import Module, Elaboratable, Signal, Cat, Mux
    from soc.decoder.power_enums import (Function, Form, InternalOp,
                             In1Sel, In2Sel, In3Sel, OutSel, RC, LdstLen,
                             CryIn, get_csv, single_bit_flags,
                             get_signal_name, default_values)
    from soc.decoder.power_fields import DecodeFields
    from soc.decoder.power_fieldsn import SigDecode, SignalBitRange
    from soc.decoder.power_decoder import PowerOp
    
    class Op_Add_Example(Elaboratable):
        def __init__(self):
            self.select = Signal(reset_less=True)
            self.op_add = PowerOp()
        
        def elaborate(self, platform):
            m = Module()
            op_add = self.op_add
    
            with m.If(self.select == 1):
                m.d.comb += op_add.function_unit.eq(Function.ALU)
                m.d.comb += op_add.form.eq(Form.XO)
                m.d.comb += op_add.internal_op.eq(InternalOp.OP_ADD)
                m.d.comb += op_add.in1_sel.eq(In1Sel.RA)
                m.d.comb += op_add.in2_sel.eq(In2Sel.RB)
                m.d.comb += op_add.in3_sel.eq(In3Sel.NONE)
                m.d.comb += op_add.out_sel.eq(OutSel.RT)
                m.d.comb += op_add.rc_sel.eq(RC.RC)
                m.d.comb += op_add.ldst_len.eq(LdstLen.NONE)
                m.d.comb += op_add.cry_in.eq(CryIn.ZERO)
            
            return m
    
    from nmigen.back import verilog
    verilog_file = "op_add_example.v"
    top = Op_Add_Example()
    f = open(verilog_file, "w")
    verilog = verilog.convert(top, name='top', strip_internal_attrs=True,
                              ports=top.op_add.ports())
    f.write(verilog)
    print(f"Verilog Written to: {verilog_file}")

The [actual POWER9 Decoder](https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/power_decoder2.py;hb=HEAD)
uses this principle, in conjunction with reading the information shown
in the table above from CSV files (as opposed to hardcoding them in
python source).  These [[CSV files|openpower/isatables]],
being machine-readable in a wide variety
of programming languages, are conveniently available for use by
other projects well beyond just this SOC.

This also demonstrates one of the design aspects taken in this project: to
*combine* the power of python's full capabilities in order to create
advanced dynamically generated HDL, rather than (as done with MyHDL)
limit python code to a subset of its full capabilities.

The CSV Files are loaded by
[power_decoder.py](https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/power_decoder.py;hb=HEAD)
and are used to construct a hierarchical cascade of switch statements.  The original code came from
[microwatt](https://github.com/antonblanchard/microwatt/blob/master/decode1.vhdl)
where the original hardcoded cascade can be seen.

The docstring for power_decoder.py gives more details: each level in the hierarchy, just as in the original decode1.vhdl, will take slices of the instruction bitpattern, match against it, and if successful will continue with further subdecoders until a line is met that contains the required Operand Information (a PowerOp) exactly as shown at the top of this page.

In this way, different sections of the instruction are successively decoded (major opcode, then minor opcode, then sub-patterns under those) until the required instruction is fully recognised, and the hierarchical cascade of switch patterns results in a flat interpretation being produced that is useful internally. 

# second explanation / walkthrough

we (manually) extracted the pseudo-code from the v3.0B specification
(took 2 days):
<https://git.libre-soc.org/?p=libreriscv.git;a=blob;f=openpower/isa/fixedlogical.mdwn;hb=HEAD>

then wrote a parser and language translator (aka compiler) to convert
those code-fragments to python:
<https://git.libre-soc.org/?p=soc.git;a=tree;f=src/soc/decoder/pseudo;hb=HEAD>

then went to a lot of trouble over the course of several months to
co-simulate them, update them, and make them accurate according to the
actual spec:
<https://git.libre-soc.org/?p=libreriscv.git;a=blob;f=openpower/isa/fixedarith.mdwn;h=470a833ca2b8a826f5511c4122114583ef169e55;hb=HEAD#l721>

and created a fully-functioning python-based OpenPOWER ISA simulator:
<https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/isa/caller.py;hb=HEAD>

there is absolutely no reason why this language-translator (aka compiler)
here
<https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/pseudo/parser.py;hb=HEAD>

should not be joined by another compiler, targetting c for use inside
the linux kernel or, another compiler which auto-generates c++ for use
inside power-gem5, such that this:
<https://github.com/power-gem5/gem5/blob/cae53531103ebc5bccddf874db85f2659b64000a/src/arch/power/isa/decoder.isa#L1214>

becomes an absolute breeze to update.

note that we maintain a decoder which is based on Microwatt: we extracted
microwatt's decode1.vhdl into CSV files, and parse them in python as
hierarchical recursive data structures:
<https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/power_decoder.py;hb=HEAD>

where the actual CSV files that it reads are here:
<https://git.libre-soc.org/?p=libreriscv.git;a=tree;f=openpower/isatables;hb=HEAD>

this is then combined with *another* table that was extracted from the
OpenPOWER v3.0B PDF:
<https://git.libre-soc.org/?p=libreriscv.git;a=blob;f=openpower/isatables/fields.text;hb=HEAD>

(the parser for that recognises "vertical bars" as being
field-separators):
<https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/power_fields.py;hb=HEAD>

and FINALLY - and this is about the only major piece of code that
actually involves any kind of manual code - again it is based on Microwatt
decode2.vhdl - we put everything together to turn a binary opcode into
"something that needs to be executed":
<https://git.libre-soc.org/?p=soc.git;a=blob;f=src/soc/decoder/power_decoder2.py;hb=HEAD>

so our OpenPOWER simulator is actually based on:

* machine-readable CSV files
* machine-readable Field-Form files
* machine-readable spec-accurate pseudocode files

the only reason we haven't used those to turn it into HDL is because
doing so is a massive research project, where a first pass would be
highly likely to generate sub-optimal HDL