Copyright | (C) 2013-2016 University of Twente 2017 Google Inc. 2019 Myrtle Software Ltd |
---|---|
License | BSD2 (see the file LICENSE) |
Maintainer | Christiaan Baaij <christiaan.baaij@gmail.com> |
Safe Haskell | Safe |
Language | Haskell2010 |
Extensions |
|
Self-synchronizing circuits based on data-flow principles.
Synopsis
- newtype DataFlow dom iEn oEn i o = DF {}
- liftDF :: (Signal dom i -> Signal dom Bool -> Signal dom Bool -> (Signal dom o, Signal dom Bool, Signal dom Bool)) -> DataFlow dom Bool Bool i o
- pureDF :: (i -> o) -> DataFlow dom Bool Bool i o
- mealyDF :: (KnownDomain dom, NFDataX s) => Clock dom -> Reset dom -> Enable dom -> (s -> i -> (s, o)) -> s -> DataFlow dom Bool Bool i o
- mooreDF :: (KnownDomain dom, NFDataX s) => Clock dom -> Reset dom -> Enable dom -> (s -> i -> s) -> (s -> o) -> s -> DataFlow dom Bool Bool i o
- fifoDF :: forall addrSize m n a dom. (KnownDomain dom, NFDataX a, KnownNat addrSize, KnownNat n, KnownNat m, (m + n) ~ (2 ^ addrSize)) => Clock dom -> Reset dom -> Enable dom -> SNat (m + n) -> Vec m a -> DataFlow dom Bool Bool a a
- idDF :: DataFlow dom en en a a
- seqDF :: DataFlow dom aEn bEn a b -> DataFlow dom bEn cEn b c -> DataFlow dom aEn cEn a c
- firstDF :: DataFlow dom aEn bEn a b -> DataFlow dom (aEn, cEn) (bEn, cEn) (a, c) (b, c)
- swapDF :: DataFlow dom (aEn, bEn) (bEn, aEn) (a, b) (b, a)
- secondDF :: DataFlow dom aEn bEn a b -> DataFlow dom (cEn, aEn) (cEn, bEn) (c, a) (c, b)
- parDF :: DataFlow dom aEn bEn a b -> DataFlow dom cEn dEn c d -> DataFlow dom (aEn, cEn) (bEn, dEn) (a, c) (b, d)
- parNDF :: KnownNat n => Vec n (DataFlow dom aEn bEn a b) -> DataFlow dom (Vec n aEn) (Vec n bEn) (Vec n a) (Vec n b)
- loopDF :: (KnownDomain dom, NFDataX d, KnownNat m, KnownNat n, KnownNat addrSize, (m + n) ~ (2 ^ addrSize)) => Clock dom -> Reset dom -> Enable dom -> SNat (m + n) -> Vec m d -> DataFlow dom (Bool, Bool) (Bool, Bool) (a, d) (b, d) -> DataFlow dom Bool Bool a b
- loopDF_nobuf :: DataFlow dom (Bool, Bool) (Bool, Bool) (a, d) (b, d) -> DataFlow dom Bool Bool a b
- class LockStep a b where
Data types
newtype DataFlow dom iEn oEn i o Source #
Dataflow circuit with bidirectional synchronization channels.
In the forward direction we assert validity of the data. In the backward
direction we assert that the circuit is ready to receive new data. A circuit
adhering to the DataFlow
type should:
- Not consume data when validity is deasserted.
- Only update its output when readiness is asserted.
The DataFlow
type is defined as:
newtype DataFlow' dom iEn oEn i o = DF { df ::Signal
dom i -- Incoming data ->Signal
dom iEn -- Flagged with valid bitsiEn
. ->Signal
dom oEn -- Incoming back-pressure, ready edge. -> (Signal
dom o -- Outgoing data. ,Signal
dom oEn -- Flagged with valid bitsoEn
. ,Signal
dom iEn -- Outgoing back-pressure, ready edge. ) }
where:
dom
is the domain to which the circuit is synchronized.iEn
is the type of the bidirectional incoming synchronization channel.oEn
is the type of the bidirectional outgoing synchronization channel.i
is the incoming data type.o
is the outgoing data type.
We define several composition operators for our DataFlow
circuits:
seqDF
sequential composition.parDF
parallel composition.loopDF
add a feedback arc.lockStep
proceed in lock-step.
When you look at the types of the above operators it becomes clear why we parametrize in the types of the synchronization channels.
Creating DataFlow circuits
liftDF :: (Signal dom i -> Signal dom Bool -> Signal dom Bool -> (Signal dom o, Signal dom Bool, Signal dom Bool)) -> DataFlow dom Bool Bool i o Source #
Dataflow circuit synchronized to the systemClockGen
.
type DataFlow iEn oEn i o = DataFlow' systemClockGen iEn oEn i o
Create a DataFlow
circuit from a circuit description with the appropriate
type:
Signal
dom i -- Incoming data. ->Signal
dom Bool -- Flagged with a single valid bit. ->Signal
dom Bool -- Incoming back-pressure, ready bit. -> (Signal
dom o -- Outgoing data. ,Signal
dom oEn -- Flagged with a single valid bit. ,Signal
dom iEn -- Outgoing back-pressure, ready bit. )
A circuit adhering to the DataFlow
type should:
- Not consume data when validity is deasserted.
- Only update its output when readiness is asserted.
pureDF :: (i -> o) -> DataFlow dom Bool Bool i o Source #
Create a DataFlow
circuit where the given function f
operates on the
data, and the synchronization channels are passed unaltered.
mealyDF :: (KnownDomain dom, NFDataX s) => Clock dom -> Reset dom -> Enable dom -> (s -> i -> (s, o)) -> s -> DataFlow dom Bool Bool i o Source #
Create a DataFlow
circuit from a Mealy machine description as those of
Clash.Prelude.Mealy
mooreDF :: (KnownDomain dom, NFDataX s) => Clock dom -> Reset dom -> Enable dom -> (s -> i -> s) -> (s -> o) -> s -> DataFlow dom Bool Bool i o Source #
Create a DataFlow
circuit from a Moore machine description as those of
Clash.Prelude.Moore
:: (KnownDomain dom, NFDataX a, KnownNat addrSize, KnownNat n, KnownNat m, (m + n) ~ (2 ^ addrSize)) | |
=> Clock dom | |
-> Reset dom | |
-> Enable dom | |
-> SNat (m + n) | Depth of the FIFO buffer. Must be a power of two. |
-> Vec m a | Initial content. Can be smaller than the size of the
FIFO. Empty spaces are initialized with |
-> DataFlow dom Bool Bool a a |
Composition combinators
seqDF :: DataFlow dom aEn bEn a b -> DataFlow dom bEn cEn b c -> DataFlow dom aEn cEn a c Source #
Sequential composition of two DataFlow
circuits.
firstDF :: DataFlow dom aEn bEn a b -> DataFlow dom (aEn, cEn) (bEn, cEn) (a, c) (b, c) Source #
Apply the circuit to the first halve of the communication channels, leave the second halve unchanged.
swapDF :: DataFlow dom (aEn, bEn) (bEn, aEn) (a, b) (b, a) Source #
Swap the two communication channels.
secondDF :: DataFlow dom aEn bEn a b -> DataFlow dom (cEn, aEn) (cEn, bEn) (c, a) (c, b) Source #
Apply the circuit to the second halve of the communication channels, leave the first halve unchanged.
parDF :: DataFlow dom aEn bEn a b -> DataFlow dom cEn dEn c d -> DataFlow dom (aEn, cEn) (bEn, dEn) (a, c) (b, d) Source #
Compose two DataFlow
circuits in parallel.
parNDF :: KnownNat n => Vec n (DataFlow dom aEn bEn a b) -> DataFlow dom (Vec n aEn) (Vec n bEn) (Vec n a) (Vec n b) Source #
Compose n DataFlow
circuits in parallel.
:: (KnownDomain dom, NFDataX d, KnownNat m, KnownNat n, KnownNat addrSize, (m + n) ~ (2 ^ addrSize)) | |
=> Clock dom | |
-> Reset dom | |
-> Enable dom | |
-> SNat (m + n) | Depth of the FIFO buffer. Must be a power of two |
-> Vec m d | Initial content of the FIFO buffer. Can be smaller than the size of the
FIFO. Empty spaces are initialized with |
-> DataFlow dom (Bool, Bool) (Bool, Bool) (a, d) (b, d) | |
-> DataFlow dom Bool Bool a b |
Feed back the second halve of the communication channel. The feedback loop
is buffered by a fifoDF
circuit.
So given a circuit h with two synchronization channels:
h :: DataFlow
(Bool,Bool) (Bool,Bool) (a,d) (b,d)
Feeding back the d part (including its synchronization channels) results in:
loopDF
d4 Nil h
When you have a circuit h'
, with only a single synchronization channel:
h' :: DataFlow
Bool Bool (a,d) (b,d)
and you want to compose h' in a feedback loop, the following will not work:
f `` (
seqDF
loopDF
d4 Nil h') `` g
seqDF
The circuits f
, h
, and g
, must operate in lock-step because the h'
circuit only has a single synchronization channel. Consequently, there
should only be progress when all three circuits are producing valid data
and all three circuits are ready to receive new data. We need to compose
h' with the lockStep
and stepLock
functions to achieve the lock-step
operation.
f `` (
seqDF
lockStep
``
seqDF
loopDF
d4 Nil h' ``
seqDF
stepLock
) `` g
seqDF
loopDF_nobuf :: DataFlow dom (Bool, Bool) (Bool, Bool) (a, d) (b, d) -> DataFlow dom Bool Bool a b Source #
Feed back the second halve of the communication channel. Unlike loopDF
,
the feedback loop is not buffered.
Lock-Step operation
class LockStep a b where Source #
Reduce or extend the synchronization granularity of parallel compositions.
lockStep :: DataFlow dom a Bool b b Source #
Reduce the synchronization granularity to a single Bool
ean value.
Given:
f ::DataFlow
Bool Bool a b g ::DataFlow
Bool Bool c d h ::DataFlow
Bool Bool (b,d) (p,q)
We cannot simply write:
(f `` g) `
parDF
` h
seqDF
because, f `parDF` g
, has type,
,
which does not match the expected synchronization granularity of DataFlow
(Bool,Bool) (Bool,Bool) (a,c) (b,d)h
. We
need a circuit in between that has the type:
DataFlow
(Bool,Bool) Bool (b,d) (b,d)
Simply &&
-ing the valid signals in the forward direction, and
duplicating the ready signal in the backward direction is however not
enough. We also need to make sure that f
does not update its output when
g
's output is invalid and visa versa, as h
can only consume its input
when both f
and g
are producing valid data. g
's ready port is hence
only asserted when h
is ready and f
is producing valid data. And f
's
ready port is only asserted when h
is ready and g
is producing valid
data. f
and g
will hence be proceeding in lock-step.
The lockStep
function ensures that all synchronization signals are
properly connected:
(f `` g) `
parDF
`
seqDF
lockStep
`` h
seqDF
Note 1: ensure that the components that you are synchronizing have
buffered/delayed ready
and valid
signals, or lockStep
has the
potential to introduce combinational loops. You can do this by placing
fifoDF
s on the parallel channels. Extending the above example, you would
write:
((f ``
seqDF
fifoDF
d4 Nil) `` (g `
parDF
`
seqDF
fifoDF
d4 Nil)) ``
seqDF
lockStep
`` h
seqDF
Note 2: lockStep
works for arbitrarily nested tuples. That is:
p ::DataFlow
Bool Bool ((b,d),d) z q ::DataFlow
((Bool,Bool),Bool) ((Bool,Bool),Bool) ((a,c),c) ((b,d),d) q = f `` g `
parDF
` g r = q `
parDF
`
seqDF
lockStep
`` p
seqDF
Does the right thing.
stepLock :: DataFlow dom Bool a b b Source #
Extend the synchronization granularity from a single Bool
ean value.
Given:
f ::DataFlow
Bool Bool a b g ::DataFlow
Bool Bool c d h ::DataFlow
Bool Bool (p,q) (a,c)
We cannot simply write:
h `` (f `
seqDF
` g)
parDF
because, f `parDF` g
, has type,
,
which does not match the expected synchronization granularity of DataFlow
(Bool,Bool) (Bool,Bool) (a,c) (b,d)h
. We
need a circuit in between that has the type:
DataFlow
Bool (Bool,Bool) (a,c) (a,c)
Simply &&
-ing the ready signals in the backward direction, and
duplicating the valid signal in the forward direction is however not
enough. We need to make sure that f
does not consume values when g
is
not ready and visa versa, because h
cannot update the values of its
output tuple independently. f
's valid port is hence only asserted when
h
is valid and g
is ready to receive new values. g
's valid port is
only asserted when h
is valid and f
is ready to receive new values.
f
and g
will hence be proceeding in lock-step.
The stepLock
function ensures that all synchronization signals are
properly connected:
h ``
seqDF
stepLock
`` (f `
seqDF
` g)
parDF
Note 1: ensure that the components that you are synchronizing have
buffered/delayed ready
and valid
signals, or stepLock
has the
potential to introduce combinational loops. You can do this by placing
fifoDF
s on the parallel channels. Extending the above example, you would
write:
h ``
seqDF
stepLock
`` ((
seqDF
fifoDF
d4 Nil `` f) `
seqDF
` (
parDF
fifoDF
d4 Nil `` g))
seqDF
Note 2: stepLock
works for arbitrarily nested tuples. That is:
p ::DataFlow
Bool Bool z ((a,c),c) q ::DataFlow
((Bool,Bool),Bool) ((Bool,Bool),Bool) ((a,c),c) ((b,d),d) q = f `` g `
parDF
` g r = p `
parDF
`
seqDF
stepLock
`` q
seqDF
Does the right thing.