A lot of optimizations in DFG assume a DAG, but the more things are
representable, the more likely it is that a small cyclic sub-graph is
present in an otherwise very large graph that is mostly acyclic. In
order to avoid loosing optimization opportunities, we explicitly extract
the cyclic sub-graphs (which are the strongly connected components +
anything feeing them, up to variable boundaries) and treat them
separately. This enables optimization of the remaining input.
This change introduces a custom reference-counting pointer class that
allows creating such pointers from 'this'. This lets us keep the
receiver object around even if all references to it outside of a class
method no longer exist. Useful for coroutine methods, which may outlive
all external references to the object.
The deletion of objects is deferred until the next time slot. This is to
make clearing the triggered flag on named events in classes safe
(otherwise freed memory could be accessed).
Added DfgVertexVariadic to represent DFG vetices with a varying number
of source operands. Converted DfgVar to be a variadic vertex, with each
driver corresponding to a fixed range of bits in the packed variable.
This allows us to handle AstSel on the LHS of assignments. Also added
support for AstConcat on the LHS by selecting into the RHS as
appropriate.
This improves OpenTitan ST speed by ~13%
This is only a debugging aid at this point, so compile out of the
release build. This reduces peak memory consumption by 4-5%. We still
keep the global counters to detect the tree have changed, to avoid
unnecessary dumps.
Multiple tricks to reduce the size of class FileLine from 72 to 40
bytes:
- Reduce file name index from 32 to 16 bits. This still allows 64K
unique input files, which is hopefully enough.
- Intern message/warning enable bitset and use a 16-bit index, again
allowing 64K unique sets which is hopefully enough.
- Put the m_waive flag into the sign bit of one of the line numbers.
- Use explicit reference counting to avoid overhead of shared_ptr.
Added assertions to ensure interned data fits within it's index space.
This saves ~5-10% peak memory consumption at no measurable run-time cost
on various designs.
Added a new data-flow graph (DFG) based combinational logic optimizer.
The capabilities of this covers a combination of V3Const and V3Gate, but
is also more capable of transforming combinational logic into simplified
forms and more.
This entail adding a new internal representation, `DfgGraph`, and
appropriate `astToDfg` and `dfgToAst` conversion functions. The graph
represents some of the combinational equations (~continuous assignments)
in a module, and for the duration of the DFG passes, it takes over the
role of AstModule. A bulk of the Dfg vertices represent expressions.
These vertex classes, and the corresponding conversions to/from AST are
mostly auto-generated by astgen, together with a DfgVVisitor that can be
used for dynamic dispatch based on vertex (operation) types.
The resulting combinational logic graph (a `DfgGraph`) is then optimized
in various ways. Currently we perform common sub-expression elimination,
variable inlining, and some specific peephole optimizations, but there
is scope for more optimizations in the future using the same
representation. The optimizer is run directly before and after inlining.
The pre inline pass can operate on smaller graphs and hence converges
faster, but still has a chance of substantially reducing the size of the
logic on some designs, making inlining both faster and less memory
intensive. The post inline pass can then optimize across the inlined
module boundaries. No optimization is performed across a module
boundary.
For debugging purposes, each peephole optimization can be disabled
individually via the -fno-dfg-peepnole-<OPT> option, where <OPT> is one
of the optimizations listed in V3DfgPeephole.h, for example
-fno-dfg-peephole-remove-not-not.
The peephole patterns currently implemented were mostly picked based on
the design that inspired this work, and on that design the optimizations
yields ~30% single threaded speedup, and ~50% speedup on 4 threads. As
you can imagine not having to haul around redundant combinational
networks in the rest of the compilation pipeline also helps with memory
consumption, and up to 30% peak memory usage of Verilator was observed
on the same design.
Gains on other arbitrary designs are smaller (and can be improved by
analyzing those designs). For example OpenTitan gains between 1-15%
speedup depending on build type.
- Rename `--dump-treei` option to `--dumpi-tree`, which itself is now a
special case of `--dumpi-<tag>` where tag can be a magic word, or a
filename
- Control dumping via static `dump*()` functions, analogous to `debug()`
- Make dumping independent of the value of `debug()` (so dumping always
works even without the debug flag)
- Add separate `--dumpi-graph` for dumping V3Graphs, which is again a
special case of `--dumpi-<tag>`
- Alias `--dump-<tag>` to `--dumpi-<tag> 3` as before
Use astgen to generate a more thorough version of AstNode::checkTree,
which checks that operands are or consistent structure and type, as
described in the @astgen op directives. Also change checkTree to always
run when --debug-check is given.
Fix discovered fallout.
Introduce the @astgen directives parsed by astgen, currently used for
the generation child node (operand) accessors. Please see the updated
internal documentation for details.
Introduce the @astgen directives parsed by astgen, currently used for
the generation child node (operand) accessors. Please see the updated
internal documentation for details.
This approach reduced total time of V3Undriven stage from 34,2s to 2,5s
in design containing almost 400 000 unused variables.
Signed-off-by: Kamil Rakoczy <krakoczy@antmicro.com>
Generate type specific static overloads of Ast<Node>::addNext, which
return the correct sub-type of the 'this' they were invoked on.
Also remove AstNode::addNextNull, which is now only used in the parser,
implement in verilog.y directly as a template function.
- Move DType representations into V3AstNodeDType.h
- Move AstNodeMath and subclasses into V3AstNodeMath.h
- Move any other AstNode subtypes into V3AstNodeOther.h
- Fix up out-of-order definitions via inline methods and implementations
in V3Inlines.h and V3AstNodes.cpp
- Enforce declaration order of AstNode subtypes via astgen,
which will now fail when definitions are mis-ordered.
Rely less on strings and represent AstNode classes as a 'class Node',
with all associated properties kept together, rather than distributed
over multiple dictionaries or constructed at retrieval time.
No functional change intended.
Small fixup patch so the 'ico' and 'act' scheduling sections could be
ordered as multi-threaded. However, we still only order these single
threaded at the moment (but switching them to multi-threaded now works).
Before this change, some forked processes were being inlined in
`V3Timing` because they contained no `CAwait`s. This only works under
the assumption that no `CAwait`s will be added there later, which is not
true, as a function called by a forked process could be turned into a
coroutine later. The call would be wrapped in a new `CAwait`, but the
process itself would have already been inlined at this point.
This commit moves the inlining to `transformForks` in `V3SchedTiming`,
which is called at a point when all `CAwait`s are already in place.
Signed-off-by: Krzysztof Bieganski <kbieganski@antmicro.com>
The recent patch to defer substitutions on V3Gate crashes on circular
logic that has cycle length >= 3 with all inlineable signals (cycle
length 2 is detected correctly and is not inlined). Fix by stopping
recursion at the loop-back edge.
Fixes#3543
This is detritus from when V3TraceDecl used to run after V3Gate, today
V3TraceDecl runs before V3Gate and this value has no function at all.
No functional change intended.
dynamic_cast is not free. Replace obvious instances (where the result is
unconditionally dereferenced) with static_cast in contexts with
performance implications.
Replace std::set<SiblingMC> with V3Lists to keep track of SiblingMCs
associated with MTasks, use a std::set<LogicMTask*> for ensuring
uniqueness. This yields a bit more speed in PartContraction.
- Use modern C++
- Implement OrderLogicVertex->LogicMTask map with
OrderLogicVertex::userp(), insteas of std::unordered_map
- Simplify data structures
- Simplify code and assert properties
No functional change.
Refactor ProcessMoveBuildGraph utilizing the fact that OrderGraph is a
bipartite graph, also remove unnecessary unordered_map and distribute
variable domain map. No functional change.
Adds timing support to Verilator. It makes it possible to use delays,
event controls within processes (not just at the start), wait
statements, and forks.
Building a design with those constructs requires a compiler that
supports C++20 coroutines (GCC 10, Clang 5).
The basic idea is to have processes and tasks with delays/event controls
implemented as C++20 coroutines. This allows us to suspend and resume
them at any time.
There are five main runtime classes responsible for managing suspended
coroutines:
* `VlCoroutineHandle`, a wrapper over C++20's `std::coroutine_handle`
with move semantics and automatic cleanup.
* `VlDelayScheduler`, for coroutines suspended by delays. It resumes
them at a proper simulation time.
* `VlTriggerScheduler`, for coroutines suspended by event controls. It
resumes them if its corresponding trigger was set.
* `VlForkSync`, used for syncing `fork..join` and `fork..join_any`
blocks.
* `VlCoroutine`, the return type of all verilated coroutines. It allows
for suspending a stack of coroutines (normally, C++ coroutines are
stackless).
There is a new visitor in `V3Timing.cpp` which:
* scales delays according to the timescale,
* simplifies intra-assignment timing controls and net delays into
regular timing controls and assignments,
* simplifies wait statements into loops with event controls,
* marks processes and tasks with timing controls in them as
suspendable,
* creates delay, trigger scheduler, and fork sync variables,
* transforms timing controls and fork joins into C++ awaits
There are new functions in `V3SchedTiming.cpp` (used by `V3Sched.cpp`)
that integrate static scheduling with timing. This involves providing
external domains for variables, so that the necessary combinational
logic gets triggered after coroutine resumption, as well as statements
that need to be injected into the design eval function to perform this
resumption at the correct time.
There is also a function that transforms forked processes into separate
functions.
See the comments in `verilated_timing.h`, `verilated_timing.cpp`,
`V3Timing.cpp`, and `V3SchedTiming.cpp`, as well as the internals
documentation for more details.
Signed-off-by: Krzysztof Bieganski <kbieganski@antmicro.com>
Various optimizations to speed up MTasks coarsening (which is the long
pole in the multi-threaded scheduling of very large designs).
The biggest impact ones:
- Use efficient hand written Pairing Heaps for implementing priority
queues and the scoreboard, instead of the old SortByValueMap. This
helps us avoid having to sort a lot of merge candidates that we will
never actually consider and helps a lot in performance.
- Remove unnecessary associative containers and store data structures
(the heap nodes in particular) directly in the object they relate to.
This eliminates a huge amount of lookups and helps a lot in
performance.
- Distribute storage for SiblingMC instances into the LogicMTask
instances, and combine with the sibling maps. This again eliminates
hash table lookups and makes storage structures smaller.
- Remove some now bidirectional edge maps, keep only the forward map.
There are also some other smaller optimizations:
- Replaced more unnecessary dynamic_casts with static_casts
- Templated some functions/classes to reduce the number of static
branches in loops.
- Improves sorting of edges for sibling candidate creation
- Various micro-optimizations here and there
This speeds up MTask coarsening by 3.8x on a large design, which
translates to a 2.5x speedup of the ordering pass in multi-threaded
mode. (Combined with the earlier optimizations, ordering is now 3x
faster.)
Due to the elimination of a lot of the auxiliary data structures, and
ensuring a minimal size for the necessary ones, memory consumption of
the MTask coarsening is also reduced (measured up to 4.4x reduction
though the accuracy of this is low).
The algorithm is identical except for minor alterations of the order
some candidates are added or removed, this can cause perturbation in the
output due to tied scores being broken based on IDs.
Various optimizations to speed up MTasks coarsening (which is the long
pole in the multi-threaded scheduling of very large designs).
The biggest impact ones:
- Use efficient hand written Pairing Heaps for implementing priority
queues and the scoreboard, instead of the old SortByValueMap. This
helps us avoid having to sort a lot of merge candidates that we will
never actually consider and helps a lot in performance.
- Remove unnecessary associative containers and store data structures
(the heap nodes in particular) directly in the object they relate to.
This eliminates a huge amount of lookups and helps a lot in
performance.
- Distribute storage for SiblingMC instances into the LogicMTask
instances, and combine with the sibling maps. This again eliminates
hash table lookups and makes storage structures smaller.
- Remove some now bidirectional edge maps, keep only the forward map.
There are also some other smaller optimizations:
- Replaced more unnecessary dynamic_casts with static_casts
- Templated some functions/classes to reduce the number of static
branches in loops.
- Improves sorting of edges for sibling candidate creation
- Various micro-optimizations here and there
This speeds up MTask coarsening by 3.8x on a large design, which
translates to a 2.5x speedup of the ordering pass in multi-threaded
mode. (Combined with the earlier optimizations, ordering is now 3x
faster.)
Due to the elimination of a lot of the auxiliary data structures, and
ensuring a minimal size for the necessary ones, memory consumption of
the MTask coarsening is also reduced (measured up to 4.4x reduction
though the accuracy of this is low).
The algorithm is identical except for minor alterations of the order
some candidates are added or removed, this can cause perturbation in the
output due to tied scores being broken based on IDs.
While keeping the client code abstract in PartPropagateCp is nice for
testing, there is performance to be had removing the abstraction. As
this code dominates in scheduling large designs, we eliminate the
abstraction and re-work the testing to use the actual LogicMTask and
MTaskEdge graph types. No functional change intended.
Instead of deleting then re-allocating MTaskEdge instances when merging
two MTasks, just redirect the edged of the donor MTask to the recipient
MTask. This is both faster as it avoids an allocation and a deletion,
together with one update of the sibling maps, and also makes the
algorithm more stable due to MergeCandidate IDs being stable and
allocated up front for all MTaskEdges, before any SiblingMCs are
allocated.
Perturbations in output are expected as the IDs used to break ties
between merge candidates with equal costs are not updated when
redirecting an edge (on purpose). The relinking of only one end of the
graph edges also perturbs the order in which they are enumerated, which
does change candidate opportunities when the number of edges is larger
than PART_SIBLING_EDGE_LIMIT. Confirmed output is identical when
IDs are updated and edges are updated to appear in their original order.
The critical path propagation used to rely on a pointer comparison to
break equal scoring critical path updates. Use the corresponding mtask
ids instead, which is deterministic across invocations.
siblingPairFromRelatives gathers neighbours of a vertex, and sorts them.
It then takes the N best nodes, and creates sibling merge candidates
from them. We now use the unadjusted cost instead of the step cost of
the vertices when sorting. This is both faster as we need not do the
log-space rounding to compute stepCost, and will also make similar but
yet cheaper nodes appear closer to the front as we don't lose precision
in rounding, hence they are more likely to be entered as merge
candidates. Note that when creating the merge candidate, we still use
the stepCost, so it's purpose of reducing the propagation of critical
path updates is maintained in full. In summary, this should make both
Verilator and the generated model very slightly faster, at least in
theory, and I have observed minor improvement in places.
GraphStreamUnordered used to be GraphStream<std::less<const
V3GraphVertex*>>, but a lot of performance improvements can be had by a
specialized implementation, so added a highly optimized one. This helps
a lot with --debug-partition.
Fix compile error for queue method usage, if it is the
first statement in a block of code, and the return
value is not used. Example:
> if (foo)
> void'(bar.pop_front());
* Tests: Add a test to reproduce #3399
* Fix#3399. When reading an inout port in a module, it should refer the
original inout port, not the generated MODTEMP.
Keep a single std::set of key/value pairs, and a single unordered_map
from key to iterators into the set. Also improve some of the accessing
mechanisms using modern C++. This speeds up multi-threaded ordering by
about 10%.
Wilson Snyder
github action
Marcel Chang
Kanad Kanhere
Ryszard Rozak
Geza Lore
Krzysztof Bieganski
Kamil Rakoczy
Mariusz Glebocki
Yu-Sheng Lin
Gustav Svensk
Aleksander Kiryk
Arkadiusz Kozdra
Varun Koyyalagunta
Ryszard Rozak
Mostafa Gamal
Drew Ranck
Yutetsu TAKATSUKASA
Kamil Rakoczy