Reporting topology requires producer to be available, which creates a
deadlock during delayed init data structures construction.
Calling GetProducer() results in a call to GetProfilerThreadData(),
which in turn calls GetProfilerData() to construct its thread local
variable. However, at this point we already are calling
GetProfilerData() (to construct the profiler itself). This would result
in an incorrect double construction of data, but the code already
prevents this by allowing init code to be entered only once. Hence the
deadlock.
Currently this is a non-issue, as no platform which can report CPU
topology needs to use delayed init.
Queue calibration requires queue access during profiler construction. This
in turn requires construction of profiler data block, *which at this point
is underway*, because the profiler is being constructed.
But rdtscp is serializing!
No, it's not. Quoting the Intel Instruction Set Reference:
"The RDTSCP instruction is not a serializing instruction, but it does
wait until all previous instructions have executed and all previous
loads are globally visible. But it does not wait for previous stores to
be globally visible, and subsequent instructions may begin execution
before the read operation is performed.",
"The RDTSC instruction is not a serializing instruction. It does not
necessarily wait until all previous instructions have been executed
before reading the counter. Similarly, subsequent instructions may begin
execution before the read operation is performed."
So, the difference is in waiting for prior instructions to finish
executing. Notice that even in the rdtscp case, execution of the
following instructions may commence before time measurement is finished
and data stores may be still pending.
But, you may say, Intel in its "How to Benchmark Code Execution Times"
document shows that using rdtscp is superior to rdstc. Well, not
exactly. What they do show is that when a *single function* is
considered, there are ways to measure its execution time with little to
no error.
This is not what Tracy is doing.
In our case there is no way to determine absolute "this is before" and
"this is after" points of a zone, as we probably already are inside
another zone. Stopping the CPU execution, so that a deeply nested zone
may be measured with great precision, will skew the measurements of all
parent zones.
And this is not what we want to measure, anyway. We are not interested
in how a *single function* behaves, but how a *whole program* behaves.
The out-of-order CPU behavior may influence the measurements? Good! We
are interested in that. We want to see *how* the code is really
executed. How is *stopping* the CPU to make a timer read an appropriate
thing to do, when we want to see how a program is performing?
At least that's the theory.
And besides all that, the profiling overhead is now reduced.
Statistics for a one-minute trace:
Capture tool | Running time | Running regions
---------------+--------------+-----------------
cat | 25.11 s | 392,300
tracy_systrace | 10.41 s | 12,249
This reverts commit b32e8fa24e.
Apparently it is possible to receive non-uniform data in alpha channel, which
breaks the original assumption about not needing the mask. This seemed to be a
problem only on 32 bit NEON implementation of DXT1 compression. Other
implementations handle such data without degradation of visual output.
("The queue" is per-thread partial queue here.)
This fixes a problem where one thread writes to the queue, then is
terminated, making the (partially filled) queue available for other
threads to recycle. If another thread re-owns the queue, it will change
the associated thread id, while part of the queue was filled by the
original thread. This obviously created invalid data during dequeue.
The fix makes the recycling process check not only for queue inactivity
(which is marked when the original thread terminates), but also if the
queue is empty, preventing mixing data from different threads.
The monotonic raw clock has the same accuracy as reading cntvct registers, but
using clock_gettime() has a measurable impact on queueing time (135 us vs
83 us).
This change is needed to enable ftrace time readings on ARM linux, which
doesn't provide any way to get raw cntvct readings, like x86-tsc on x86.
