mirror/tracy
1
0
Fork 0
mirror of https://github.com/wolfpld/tracy.git synced 2024-11-30 09:14:36 +00:00
Code Issues Packages Projects Releases Wiki Activity
4,325 Commits 2 Branches 28 Tags 44 MiB
a8e8a4a167
Commit Graph

754 Commits

Author SHA1 Message Date
Bartosz Taudul
37eef59d54 Implement reading sys time on BSD. 2019-11-21 20:41:57 +01:00
Bartosz Taudul
c7a22cc1ff Use libbacktrace on BSD. 2019-11-21 20:41:57 +01:00
Bartosz Taudul
bd7b0a8197 Support callstack capture on BSD. 2019-11-21 02:34:42 +01:00
Bartosz Taudul
c79449a6a1 Get proper program name on BSD. 2019-11-21 02:16:12 +01:00
Bartosz Taudul
7940977dba Report physical memory size on BSD. 2019-11-21 02:14:08 +01:00
Bartosz Taudul
3854ae11b2 Revert "Remove dead code." 
This reverts commit a36b73f745.
 2019-11-17 17:38:02 +01:00
Bartosz Taudul
a36b73f745 Remove dead code. 2019-11-16 18:34:05 +01:00
Bartosz Taudul
8286b0b72f Plumbing for message call stacks. 2019-11-14 23:40:41 +01:00
Bartosz Taudul
0befc75f83 Fix conflicts with X.h. 2019-11-14 18:24:29 +01:00
Bartosz Taudul
655864eb7c Enable crash handler on cygwin. 
Crash is properly recorded, but the profiler hangs while waiting for
shutdown finish.
 2019-11-07 19:20:13 +01:00
Bartosz Taudul
3fd74a92f9 Native threads are used on mingw. 2019-11-07 19:02:54 +01:00
Bartosz Taudul
351e220d30 Don't calculate queue delay if delayed init is used. 
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.
 2019-06-19 17:29:04 +02:00
Bartosz Taudul
c98f1f0b6b Make sure profiler is initialized only once in delayed init scenario. 2019-06-19 17:28:18 +02:00
Bartosz Taudul
d4f58ddaf3 Use native windows threads on cygwin, mingw. 2019-11-06 01:42:14 +01:00
Bartosz Taudul
ca198e44d3 Remove dead code from concurrentqueue. 2019-11-05 21:40:52 +01:00
Bartosz Taudul
b5590ed197 Include <mutex> for std::once. 2019-11-05 21:40:35 +01:00
Bartosz Taudul
3e9bb80217 More header cleanup. 2019-11-05 20:15:53 +01:00
Bartosz Taudul
6bbf273581 Partial header inclusion cleanup. 2019-11-05 20:09:40 +01:00
Bartosz Taudul
907574e637 Allow remote plot configuration. 2019-11-05 17:45:19 +01:00
Bartosz Taudul
f34609fd9b Set per-cpu kernel buffer size to 512 KB. 
The default setting was causing events to be lost on Android.
 2019-11-03 21:52:20 +01:00
Bartosz Taudul
b8d459d48b Use proper string size (for consistency). 
On Android code path this value is ignored.
 2019-11-03 21:51:49 +01:00
Bartosz Taudul
ca0fae33d1 Remove obsolete assert. 
Before-terminate-events now include events that have time delta
processing, with no memory to free.
 2019-11-01 20:10:24 +01:00
Bartosz Taudul
1f0c18882c Don't collect sys time after application has exited. 2019-10-29 23:05:14 +01:00
Bartosz Taudul
0f2503d334 Send time deltas in GPU time events. 2019-10-25 19:52:01 +02:00
Bartosz Taudul
8fa5188176 Send delta times for context switches. 2019-10-25 19:13:11 +02:00
Bartosz Taudul
25b3cdc1ee Send thread wakeups when handling disconnect request. 2019-10-25 18:22:42 +02:00
Bartosz Taudul
04b132b6e2 Check if requested data size doesn't overflow buffer. 2019-10-24 21:22:22 +02:00
Bartosz Taudul
ba61a9ed84 Transfer time deltas, not absolute times. 
This change significantly reduces network bandwidth requirements.

Implemented for:
- CPU zones,
- GPU zones,
- locks,
- plots,
- memory events.
 2019-10-24 00:06:41 +02:00
Bartosz Taudul
cf88265304 Full 64-bit register is set by rdtsc. 2019-10-21 01:13:55 +02:00
Bartosz Taudul
07b66cd4ab Move fake source location out of loop. 2019-10-20 22:18:05 +02:00
Bartosz Taudul
909503403b Simplify delay calibration. 2019-10-20 22:13:29 +02:00
Bartosz Taudul
c774534b47 Use rdtsc instead of rdtscp. 
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.
 2019-10-20 20:52:33 +02:00
Bartosz Taudul
30fc2f02ab Omit calculation of on-stack variable address. 2019-10-20 19:42:29 +02:00
Bartosz Taudul
c3870f8837 Use proper type. 2019-10-10 20:30:08 +02:00
Bartosz Taudul
707f113bda Add missing NOMINMAX definitions. 2019-10-10 20:29:06 +02:00
Bartosz Taudul
7cf3608493 Avoid unused variables. 2019-10-05 02:11:45 +02:00
Bartosz Taudul
e481b5ba22 Add missing thread sent indication. 2019-10-04 19:18:47 +02:00
Bartosz Taudul
9e1935f070 Make C API symbols visible across dlls. 2019-10-03 22:39:26 +02:00
Bartosz Taudul
130365f4ff Inject tracy_systrace into filesystem and use instead of cat. 
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
 2019-09-27 15:51:29 +02:00
Bartosz Taudul
3dba4088ee Embed precompiled tracy_systrace for android. 2019-09-27 15:50:58 +02:00
Bartosz Taudul
e13cbf52fd Allow changing tracy port in client. 2019-09-21 15:11:15 +02:00
Bartosz Taudul
a221f121ba Extract lock state handling to a separate context class. 2019-09-21 14:55:14 +02:00
Bartosz Taudul
37661fd2ee Fix 32 bit NEON version of DXT1 compression. 
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.
 2019-09-03 21:37:07 +02:00
Bartosz Taudul
7a6564feae Only recycle producers, if there's no data in queue. 
("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.
 2019-08-30 14:28:44 +02:00
Bartosz Taudul
00b26c1acf Fix TRACY_NO_SYSTEM_TRACING. 2019-08-26 18:02:10 +02:00
Bartosz Taudul
fbeee3cf61 Fix (?) invalid function pointer signature. 2019-08-26 17:59:58 +02:00
Bartosz Taudul
78127dc357 System threads only allow limited information queries. 2019-08-25 00:33:22 +02:00
Bartosz Taudul
deb59b4c38 Somehow fix event ordering. 2019-08-24 01:43:55 +02:00
Bartosz Taudul
1e74a89924 Check if there's data to read from kernel. 
Reading from kernel pipe, while being a blocking operation, spin locks the
thread.
 2019-08-24 01:06:21 +02:00
Bartosz Taudul
8f6e94d75c Sleep if sys trace pipe buffer underruns. 2019-08-24 00:42:00 +02:00
Bartosz Taudul
2d50d07438 Allow completely disabling system tracing. 2019-08-21 01:16:25 +02:00
Bartosz Taudul
0cbb853945 Add missing SetThreadName() calls. 2019-08-20 16:23:00 +02:00
Bartosz Taudul
332262dd84 Shorter thread names. 2019-08-20 16:22:54 +02:00
Bartosz Taudul
247acd03ee Kernel tracing on android. 2019-08-20 15:49:40 +02:00
Bartosz Taudul
e427d67347 Don't bail out if unimportant variables are not available. 2019-08-20 12:19:05 +02:00
Bartosz Taudul
bfda30be0b Use su on android to set tracing variables. 2019-08-20 12:18:46 +02:00
Bartosz Taudul
9d87a8394d Add missing getline() implementation for android API < 18. 2019-08-19 15:26:09 +02:00
Bartosz Taudul
9be6f4a414 Fix typo. 2019-08-19 13:03:37 +02:00
Bartosz Taudul
d209bb4d01 Add missing function pointer checks. 2019-08-19 12:47:27 +02:00
Bartosz Taudul
20e8a5ecc8 Create tid to pid mapping. 2019-08-17 22:32:41 +02:00
Bartosz Taudul
678e942e9f Transfer PID of profiled program. 2019-08-17 22:19:04 +02:00
Bartosz Taudul
77c636c3fd Retrieve module name for threads with no names on windows. 2019-08-17 21:24:40 +02:00
Bartosz Taudul
f7589bde02 Trace thread wakeups on linux. 2019-08-17 17:18:11 +02:00
Bartosz Taudul
414f903cc5 Collect thread wakeup data. 2019-08-17 17:05:29 +02:00
Bartosz Taudul
e9080bdbcd Hardcode windows PID 4 as "System". 2019-08-17 03:44:47 +02:00
Bartosz Taudul
40eb8a5a03 Proper check for invalid handle. 2019-08-17 03:44:11 +02:00
Bartosz Taudul
6c1dd8eaec Cast thread handle to DWORD. 2019-08-16 21:21:37 +02:00
Bartosz Taudul
d7104c752a Cygwin compat layer. 2019-08-16 21:16:04 +02:00
Bartosz Taudul
819ef2a82b External process/thread name retrieval on linux. 2019-08-16 21:00:42 +02:00
Bartosz Taudul
e975c4d7bf Also retrieve external thread names. 2019-08-16 19:49:16 +02:00
Bartosz Taudul
fe7f56b022 Implement retrieval of external process names. 2019-08-16 19:22:23 +02:00
Bartosz Taudul
83fddd9aa6 Fix unicode builds. 2019-08-16 13:09:27 +02:00
Bartosz Taudul
9d5240c597 Mutable char array is required here due to shit API design. 2019-08-16 13:03:20 +02:00
Bartosz Taudul
14a373a3b8 Add number of CPU cores to host info. 2019-08-15 02:28:35 +02:00
Bartosz Taudul
69077e4e6f Finish sending context switches during disconnect. 2019-08-14 23:06:13 +02:00
Bartosz Taudul
6dc79cf14e Cosmetics. 2019-08-14 23:05:58 +02:00
Bartosz Taudul
c0b524d8de Add a separate method for clearing serial queue. 2019-08-14 22:39:12 +02:00
Bartosz Taudul
71b54dd48a Always collect thread names. 
This fixes an issue when a thread was destroyed before its name could be
retrieved.
 2019-08-14 16:52:04 +02:00
Bartosz Taudul
5e199d1ab3 Support ftrace on ARM. 2019-08-14 16:28:54 +02:00
Bartosz Taudul
5fbb811f5d Degrade ARM timer to monotonic raw clock. 
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.
 2019-08-14 16:19:02 +02:00
Bartosz Taudul
42865d7c7b Don't set x86-tsc clock on non-x86 platforms. 2019-08-14 15:14:36 +02:00
Bartosz Taudul
54a9132bb5 Skip context switch events in on demand mode, if no connection. 2019-08-14 15:09:33 +02:00
Bartosz Taudul
602c38c6c0 Allow checking timer implementation. 2019-08-14 14:35:44 +02:00
Bartosz Taudul
3988b56c92 Capture context switches on linux. 2019-08-14 13:56:15 +02:00
Bartosz Taudul
92b6da7cc2 SetThreadName() only works on the current thread. 
This breaking change is required, because kernel trace facilities use
kernel thread ids, which are inaccessible from the pthread_t level.
 2019-08-14 02:22:45 +02:00
Bartosz Taudul
73cbf2eead Use windows thread ids on cygwin. 2019-08-13 16:22:58 +02:00
Bartosz Taudul
b313e46139 Keep event trace properties to terminate trace on exit. 2019-08-13 13:10:37 +02:00
Bartosz Taudul
90d26cb1b6 Collect and send context switch events. 2019-08-13 02:35:32 +02:00
Bartosz Taudul
fe0f1aea07 Add system tracing skeleton. 2019-08-12 23:05:34 +02:00
Bartosz Taudul
8aa0be39d5 Drop support for CPU id queries. 2019-08-12 23:05:34 +02:00
Bartosz Taudul
d6f32a0839 Serialize lock processing. 
This makes is much easier to process on the server and opens new
optimization possibilities. It also fixes theoretical problems, which
may be caused by invalid ordering of events with the same timestamp.
 2019-08-12 13:51:01 +02:00
Bartosz Taudul
0431c03556 Add serial queue interface. 2019-08-12 13:27:15 +02:00
Bartosz Taudul
4d2c7899ab Allow skipping invariant TSC check. 2019-08-08 19:21:39 +02:00
Bartosz Taudul
3a221dafde Display error messages on console, if available. 2019-08-08 19:18:05 +02:00
Bartosz Taudul
aada588129 Proper buffer reset. 2019-08-04 17:48:19 +02:00
Rokas Kupstys
b391e4c21a Fix multiple build errors when compiling with MinGW. 2019-08-04 15:49:46 +03:00
Bartosz Taudul
12969ee497 Track thread context. 
This change exploits the fact that events are processed in batches
originating from a single thread. A single message changing thread
context is enough to handle multiple messages, as opposed to inclusion
of thread identifier in each message.
 2019-08-02 20:18:08 +02:00
Bartosz Taudul
a4e7a341c0 Proper handling of disconnect request. 2019-08-01 23:14:09 +02:00
Bartosz Taudul
ca3571fd2b Still more. 2019-07-30 01:30:31 +02:00
Bartosz Taudul
47423e6263 And more. 2019-07-30 01:29:13 +02:00
Bartosz Taudul
d3783ae359 Remove magic template syntax. 2019-07-30 01:28:21 +02:00
Bartosz Taudul
9c28b82954 RPMallocInit and RPMallocThreadInit are identical. 2019-07-30 01:15:14 +02:00
Bartosz Taudul
a6a3f45810 Fill in thread id during dequeue, not during enqueue. 2019-07-30 01:15:14 +02:00
Bartosz Taudul
142ef53b42 Dequeue items from a single thread. 2019-07-29 23:44:08 +02:00
Bartosz Taudul
c7f769c52b Allow dequeuing from a single producer, retrieving thread id. 2019-07-29 23:29:30 +02:00
Bartosz Taudul
6cad76ae67 Store thread id in queue producer. 2019-07-29 23:13:06 +02:00
Bartosz Taudul
7ae9a28e32 Drop BlockingConcurrentQueue. 2019-07-29 22:58:13 +02:00
Bartosz Taudul
480a427e07 No need to hash thread ids anymore. 2019-07-29 22:36:04 +02:00
Bartosz Taudul
c60af95053 Remove unused const. 2019-07-29 22:33:32 +02:00
Bartosz Taudul
2d42abf552 Remove CannoAlloc functions. 2019-07-29 22:31:32 +02:00
Bartosz Taudul
b142860c8d More implicit producer removal. 2019-07-29 22:29:39 +02:00
Bartosz Taudul
db6eceb1a6 Producers must be explicit. 2019-07-29 22:25:28 +02:00
Bartosz Taudul
89928fde7b Queue must be always able to alloc. 2019-07-29 22:13:16 +02:00
Bartosz Taudul
a03734afa6 Remove more debug code. 2019-07-29 22:01:06 +02:00
Bartosz Taudul
e9a0145cd5 Remove MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX. 2019-07-29 21:56:53 +02:00
Bartosz Taudul
b496f1ff90 Remove MOODYCAMEL_QUEUE_INTERNAL_DEBUG. 2019-07-29 21:52:49 +02:00
Bartosz Taudul
beaadc3a56 Remove always disabled MCDBGQ_TRACKMEM code. 2019-07-29 21:51:29 +02:00
Bartosz Taudul
82a4a6d9cc Add tracy_ prefix to concurrentqueue.h file name. 2019-07-29 21:47:50 +02:00
Bartosz Taudul
276d764141 Fix cygwin. 2019-07-26 00:02:57 +02:00
Bartosz Taudul
36de7b2cc7 Fix incomplete headers. 2019-07-25 23:41:42 +02:00
Bartosz Taudul
e659220602 Use generic std::call_once() on other platforms. 2019-07-25 23:30:47 +02:00
Bartosz Taudul
d31d1f5946 Detect and report clang-cl. 2019-07-25 19:03:58 +02:00
Bartosz Taudul
092e830264 Use shifts instead of const vector and. 2019-07-22 19:56:47 +02:00
Bartosz Taudul
178dc9eba7 Combine block data directly in AVX registers. 2019-07-20 14:52:34 +02:00
Bartosz Taudul
a6300ef7d1 Ditto on ARM. 2019-07-19 22:13:56 +02:00
Bartosz Taudul
dc49f2f76a Move DXT1 index conversion to server. 2019-07-19 21:46:58 +02:00
Bartosz Taudul
11ba77ced5 Use pthread_once() to initialize rpmalloc on linux. 2019-07-19 20:15:56 +02:00
Bartosz Taudul
4c28593031 Fix races in rpmalloc initialization. 
Ensure rpmalloc_thread_initialize() int worker threads is called only after
rpmalloc_initialize() was called on the main profiler thread.
 2019-07-19 19:25:27 +02:00
Bartosz Taudul
cef8124247 Replace or with addition to enable usra instruction. 2019-07-19 01:40:27 +02:00
Bartosz Taudul
fd4689a6e2 Don't perform unnecessary ands. 2019-07-19 01:19:52 +02:00
Bartosz Taudul
f65373ece7 Replace two packs with one shuffle. 2019-07-13 20:01:12 +02:00
Bartosz Taudul
fc83f97ad3 Same for AVX/SSE. 2019-07-13 19:34:08 +02:00
Bartosz Taudul
62a167541c No need to mask out indices. 2019-07-13 19:07:25 +02:00
Alex
0c5ea710b0 Merged in z33ky/tracy/const-frame-image (pull request #37) 
Constify frame-image pointer in API.
 2019-07-13 13:09:21 +00:00
Bartosz Taudul
7bb9549e84 ARM64 specific NEON implementation of DXT1 compression. 2019-07-13 14:31:33 +02:00
Alexander 'z33ky' Hirsch
c6e8dc8d63 Constify frame-image pointer in API. 2019-07-13 12:33:55 +02:00
Bartosz Taudul
60d2384a6a Allow sending application information messages. 2019-07-12 18:34:46 +02:00
Bartosz Taudul
a1ce5fc1f6 Add include for built-in __get_cpuid() on gcc/clang. 2019-07-10 02:09:19 +02:00
Bartosz Taudul
c164a70b9d Check for rdstcp/invariant tsc support. 2019-07-10 02:04:14 +02:00
Bartosz Taudul
c0670848d2 Reuse variable. 2019-07-08 02:08:06 +02:00
Bartosz Taudul
17dbbe67de Remove dependency on range subtraction. 2019-07-08 00:14:36 +02:00
Bartosz Taudul
af1bd3e1fa Faster horizontal add. 2019-07-07 23:57:23 +02:00
Bartosz Taudul
b32e8fa24e Ditto for NEON. 2019-07-06 00:18:53 +02:00
Bartosz Taudul
d236d4b70f Ditto for AVX2. 2019-07-06 00:05:32 +02:00
Bartosz Taudul
f62b21c21d Masking alpha out is not needed. 
We assume that alpha value is constant for the whole image. The range
calculation is max - min, so alpha zeroes out. The color normalization
to range is color - min, so alpha also zeroes out here.
 2019-07-05 23:58:19 +02:00
Bartosz Taudul
03189a30b8 Two ands less in NEON DXT1 compression. 2019-07-05 18:37:25 +02:00
Bartosz Taudul
275d992cb1 Two ands less in AVX2 DXT1 compression. 2019-07-05 18:22:42 +02:00
Bartosz Taudul
c89358d6b9 Two ands less in SSE DXT1 compression. 2019-07-05 18:17:50 +02:00
Bartosz Taudul
5bfc62f1bf iOS device name decoding. 2019-06-19 09:59:46 +02:00
Bartosz Taudul
59b4f84ce5 Display unknown implementer, part as hex values. 2019-07-03 21:18:17 +02:00