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.
__gcov_flush was a private function and was removed from later GCC
versions (at least from 11.2.0, possibly earlier). Replace with the
documented public __gcov_dump.
VCD tracing is now parallelized using the same thread pool as the model.
We achieve this by breaking the top level trace functions into multiple
top level functions (as many as --threads), and after emitting the time
stamp to the VCD file on the main thread, we execute the tracing
functions in parallel on the same thread pool as the model (which we
pass to the trace file during registration), tracing into a secondary
per thread buffer. The main thread will then stitch (memcpy) the buffers
together into the output file.
This makes the `--trace-threads` option redundant with `--trace`, which
now only affects `--trace-fst`. FST tracing uses the previous offloading
scheme.
This obviously helps a lot in VCD tracing performance, and I have seen
better than Amdahl speedup, namely I get 3.9x on XiangShan 4T (2.7x on
OpenTitan 4T).
The --prof-threads option has been split into two independent options:
1. --prof-exec, for collecting verilator_gantt and other execution
related profiling data, and
2. --prof-pgo, for collecting data needed for PGO
The implementation of execution profiling is extricated from
VlThreadPool and is now a separate class VlExecutionProfiler. This means
--prof-exec can now be used for single-threaded models (though it does
not measure a lot of things just yet). For consistency VerilatedProfiler
is renamed VlPgoProfiler. Both VlExecutionProfiler and VlPgoProfiler are
in verilated_profiler.{h/cpp}, but can be used completely independently.
Also re-worked the execution profile format so it now only emits events
without holding onto any temporaries. This is in preparation for some
future optimizations that would be hindered by the introduction of function
locals via AstText.
Also removed the Barrier event. Clearing the profile buffers is not
notably more expensive as the profiling records are trivially
destructible.
* Add VL_ATTR_NO_SANITIZE_ALIGN macro to disable alignment check of ubsan
* Mark a function VL_ATTR_NO_SANITIZE_ALIGN because the function is intentionally using unaligned access for the sake of performance.
Co-authored-by: Wilson Snyder <wsnyder@wsnyder.org>
Use SIMD intrinsics to render VCD traces.
I have measured 10-40% single threaded performance increase with VCD
tracing on SweRV EH1 and lowRISC Ibex using SSE2 intrinsics to render
the trace. Also helps a tiny bit with FST, but now almost all of the FST
overhead is in the FST library.
I have reworked the tracing routines to use more precisely sized
arguments. The nice thing about this is that the performance without the
intrinsics is pretty much the same as it was before, as we do at most 2x
as much work as necessary, but in exchange there are no data dependent
branches at all.
