| 12c5a998 | 13-Dec-2024 |
klin02 <[email protected]> |
submodule(utility), transforms: collect XSLogs to SimTop.LogPerfEndpoint
XSLog depends on LogPerfCtrl declared at Top Module. Previous we
annotate such signal as dontTouch, and accessed through Hier
submodule(utility), transforms: collect XSLogs to SimTop.LogPerfEndpoint
XSLog depends on LogPerfCtrl declared at Top Module. Previous we
annotate such signal as dontTouch, and accessed through Hierarchical
name like SimTop.xx by dummy LogPerfHelper.
However, as XSLog is called in many spaces in DUT, which are not
visible to each other, especailly some in WhenContext. XS will generate
thousands of LogPerfHelper to get same LogPerfCtrl. Too many module
instantiations greately slow down compilation speed, especailly in
Palladium (more than 5 times slower than same DUT without Log).
This change collect all XSLog to SimTop.LogPerfEndpoint, with
LogPerfCtrl directly passed by IO. Some tips as follows:
1. Not call XSLog inside whenContext. To collect XSLogs, we should
access Cond and Data from other module, but data in WhenContext is
not accessible even through tap. Use XSLog(cond, pable) instead of
when(cond) {XSLog(pable)}. We also add chisel Internal API currentWhen
to check that.
2. Generate Hierarchical Module path through FIRRTL transforms.
Sometimes we want to append module path for better debugging.
XSCompatibility add a hacky way to use Chisel internal API to get
tag of current Module. Then we will replace these tag with
path during ChiselStage. Note path can only be acessed after circuit
elaboration.
3. Register and invoke caller of XSPerf and related object. As XSPerf
depends on LogPerfCtrl such as dump. We should deferred apply() until
collect. So we regirster collect() method when firstly apply XSLog,
then XSLog will automatically call XSPerf.collect() method during
collection. Note deferred apply is called in another module, so original
module tag should be recorded for path generation.
4. Concat XSLogs with same condition. Too many fwrites in same module
will cause UPOPTTHREADS warning with 16-threads Verilator. Consider
many XSLogs have same condition (especailly XSPerfs), we reuse same
condition and concat their printables to reduce fwrites. Note we also
limit size of concatation to 1000 to avoid segmentation fault caused
by too long printf.
show more ...
|