This is a large refactor of the MTask graph and coarsening algorithm, in
prep for fixing the bug described in #7913, it can also improve the
resulting multi-threaded schedule.
Two major changes:
OrderMTaskGraph now maintains the critical paths of the MTasks through
mutation. There are 2 ways to mutate the graph, which are done via
methods on the graph itself: adding an edge (used during construction,
and will be used later during fixing data hazards), or merging an MTask
into another (used during contraction). All critical path measures are
automatically updated and propagated on any mutation, so no external
algorithm needs to maintain them explicitly.
The merge candidate scoreboard used during contraction is simplified to
remove deferral of updated scores. This simplifies the code and results
in a greedily more optimal schedule. (The previous tranched rescore was
an optimization to work around the previous std::set based scoreboard,
however since the algorithm now uses an efficient PairingHeap,
verilation time is not impacted by the more accurate scoring, while
yielding better results).
Combining these two into a single patch as the code is highly
interdependent and any one change without the other would be just a
noisy transit point with unclear performance implications. Together it
should be a clear improvement.
Also added a stronger validation step run with '--debug-partition',
which checks all invariants throughout the algorithms.
Prep for fixing test added in #7913.
This is a large scale no functional change refactor, however, MT output
is perturbed as tied scores will be broken differently due to ordering
changes (still deterministic).
Split multi-threaded scheduling out of the monolithic
V3OrderParallel.cpp, into relatively independent parts, simplify the
data structures, and drop redundant or unused code.
New translation units:
- V3OrderMTaskGraph.h/.cpp: OrderMTaskGraph, the graph of LogicMTask
vertices and MTaskEdge edges. LogicMTask and MTaskEdge no longer
depend on the coarsening algorithm's merge candidate types;
per-algorithm auxiliary data is attached externally via the vertex and
edge user pointers.
- V3OrderMTaskFixHazards.cpp: data hazard fixup, was FixDataHazards.
- V3OrderMTaskContraction.cpp: graph coarsening, was Partitioner
together with PropagateCp and the merge candidate types.
- V3OrderParallel.cpp: now just the partitioning driver and ExecMTask
graph construction.
Data structure changes:
- Delete V3Scoreboard.h/.cpp. The generic template had a single user, now
a file-local MergeCandidateScoreboard in V3OrderMTaskContraction.cpp.
- Merge candidates are now MergeCandidate/SiblingMC/EdgeMC, distinguished
by a bit in the candidate id rather than by a vtable, and allocated by
the scoreboard, which owns their lifetime. This removes the multiple
inheritance previously used by MTaskEdge.
Move `hashGraphDebug` which prints the hash of a graph's shape for debugging
to generic `V3Graph::hashGraphDebug`.
Removed (can be added back later):
- Unnecesasry self tests that force special data stucture requirements.
- Per stage --stats output under --debug. (Final figures still reported.)
- Various debug dumps
Rewrite with much less running around in the templates. Use private
methods only + friend functions that do the actual type check. This
avoids cppcheck warnings.
The goal here is to use as single ordering heuristic (which can be
improved later) within MTasks as we do for serial code ordering. The
heuristic itself is factored out into the new OrderMoveGraphSerializer.
This also yields slightly nicer ordering than the previously use
GraphStream, so we end up with fewer trigger (domain) conditionals in
the MTasks, this can be worth a few percent speedup.
This has the somewhat nice side-effect of reusing OrderMoveGraphVertex
for both serial and parallel mode, so MTaskMoveGraphVertex can be
removed.
Serial mode yields identical output.
- 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
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.
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.
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.
This is a simple debugging aid to allow constraining the graph layout
via GraphViz rank directives. Note this is not related in any way to the
vertex 'rank' attribute used by some of the graph algorithms.
No functional change.
- More efficient comparison by pre-computing sorting keys.
- Remove work items in algorithms known to be redundant earlier.
This greatly reduces data structure sizes.
- Use V3GraphVertex->user() for state tracking instead of unordered_map
while both of these are constant time, they do add up.
- In `makeMinSpanningTree`, instead of batch inserting outgoing edges of
each visited vertex into an ordered set, keep an ordered set of sorted
vectors of edges. This reduces the size of the ordered set
significantly (it is now O(V) rather than O(E), and as the subject
graph is a complete graph, V ~ sqrt(E), so this is a significant gain).
- Use a vector + sorting in `perfectMatching` instead of an ordered set.
This is faster on large working sets.
This yields 3.8x speedup on the variable order pass and overall 14%
verilation speed gain on a large design.
Geza Lore
Wilson Snyder
Geza Lore
Mariusz Glebocki
Anthony Donlon
Kamil Rakoczy