Commit Graph

180 Commits

Author SHA1 Message Date
Bartosz Taudul
11cf650be6 Fix GPU queries ordering.
With multithreaded Vulkan rendering it is possible that GPU time queries
will be sent in a different order than the originating CPU queries were
made. This commit changes the in-order queue to a map of queries,
waiting to be resolved.
2018-06-22 16:37:54 +02:00
Bartosz Taudul
af0c64c888 Remove GPU resync support.
The whole concept is not really reliable. And it forces CPU to GPU sync,
which is bad.
2018-06-22 16:34:51 +02:00
Bartosz Taudul
cd5ca3e754 Don't use hash table to store 256 pointers. 2018-06-22 15:14:44 +02:00
Bartosz Taudul
3a885bb8fd Support callstack collection for OpenGL GPU zones. 2018-06-22 02:13:35 +02:00
Bartosz Taudul
35dc2f796e Process GpuZoneBeginCallstack queue event. 2018-06-22 01:56:32 +02:00
Bartosz Taudul
4992ae6b39 Take callstack field in ZoneEvent into account in save/load. 2018-06-22 01:30:08 +02:00
Bartosz Taudul
5e01a8ead9 Process callstack queue event. 2018-06-22 01:15:49 +02:00
Bartosz Taudul
205a4e4ca2 Add callstack index to ZoneEvent. 2018-06-22 01:11:03 +02:00
Bartosz Taudul
978e168cbd Handle ZoneBeginCallstack queue event.
This is identical to ZoneBegin handling, but requires some additional
bookkeeping to account for the incoming callstack information.
2018-06-22 01:07:25 +02:00
Bartosz Taudul
973eab2b4a Fix typo. 2018-06-20 23:42:00 +02:00
Bartosz Taudul
2a618c90d5 Properly save compressed thread in GPU events. 2018-06-20 23:12:49 +02:00
Bartosz Taudul
7912807133 Wait for transfer of pending callback frames. 2018-06-20 14:57:48 +02:00
Bartosz Taudul
60395c85e0 Wait for pending callstacks. 2018-06-20 14:54:08 +02:00
Bartosz Taudul
9a5329b97d Save and load callstack frames. 2018-06-20 01:59:25 +02:00
Bartosz Taudul
e56ee377f4 Fix off-by-one. 2018-06-20 01:54:27 +02:00
Bartosz Taudul
88b1955a5a Filename in callstack frame is not a persistent pointer. 2018-06-20 01:26:05 +02:00
Bartosz Taudul
4000f27e15 Stack frame accessor. 2018-06-20 01:18:59 +02:00
Bartosz Taudul
0c0afa5ac7 Process callstack frames. 2018-06-20 01:07:09 +02:00
Bartosz Taudul
203744cdd9 Callstack frame queries. 2018-06-20 00:25:26 +02:00
Bartosz Taudul
06f34052a5 Have to track callstacks of both alloc and free. 2018-06-19 22:08:47 +02:00
Bartosz Taudul
0de279005b Load saved callstack payload. 2018-06-19 22:05:15 +02:00
Bartosz Taudul
14b71e988b Properly skip memory event data. 2018-06-19 22:05:15 +02:00
Bartosz Taudul
4033d74479 Callstack payload index 0 is invalid. 2018-06-19 22:05:15 +02:00
Bartosz Taudul
b6e71dd909 Load memory event callstack index. 2018-06-19 21:51:06 +02:00
Bartosz Taudul
7c1333ce2f Save callstack payload. 2018-06-19 21:39:52 +02:00
Bartosz Taudul
2940230fcf Save callstack index in memory events. 2018-06-19 21:39:42 +02:00
Bartosz Taudul
77db91253b Assign callstack idx to memory event. 2018-06-19 21:34:36 +02:00
Bartosz Taudul
c28465aa7c Store unique callstack payloads. 2018-06-19 21:16:02 +02:00
Bartosz Taudul
cbc9ede3ca No-op callstack payload handling. 2018-06-19 19:31:16 +02:00
Bartosz Taudul
6a63d09a49 Don't check for each type, if range check is possible. 2018-06-19 19:31:16 +02:00
Bartosz Taudul
e51eef3dcd Process memory events with callstack. 2018-06-19 18:52:45 +02:00
Bartosz Taudul
59dc55002b Callstack ptr in server data structures.
Will be probably reduced to 32-bit index later on.
2018-06-19 18:52:10 +02:00
Bartosz Taudul
bb0631585c Store thread id of GPU events. 2018-06-17 19:07:07 +02:00
Bartosz Taudul
cfd7ac3957 Map compressed thread id 0 to real thread id 0. 2018-06-17 19:03:06 +02:00
Bartosz Taudul
d5a4c693d8 Take GPU timestamp period into account. 2018-06-17 18:49:56 +02:00
Bartosz Taudul
dcd6cac078 Save GPU timestamp period.
Bump file version to 0.3.2.
2018-06-17 18:27:42 +02:00
Bartosz Taudul
2be1d1d2b2 Use proper type. 2018-06-07 13:30:46 +02:00
Bartosz Taudul
b7930f67da Calculate total self time of zones. 2018-06-06 00:39:22 +02:00
Bartosz Taudul
53aea660c8 Store thread id in MessageData. 2018-05-25 21:10:38 +02:00
Bartosz Taudul
bb0246730f Don't save MessageData padding.
This requires file version bump to 0.3.1.
2018-05-25 21:10:38 +02:00
Bartosz Taudul
312c20b0bc Fallback to pdqsort if parallel STL is not available. 2018-05-12 22:41:18 +02:00
Bartosz Taudul
920bfc8c82 Parallelize (big) sorts in worker. 2018-05-08 01:40:22 +02:00
Bartosz Taudul
dbc963d55c Drop template argument from std::lock_guard. 2018-05-08 01:25:16 +02:00
Bartosz Taudul
3768ed5dd7 Don't reconstruct mem plot if there's no mem event data. 2018-05-04 16:08:16 +02:00
Bartosz Taudul
e7ffe288e6 One less FileWrite::Write() call. 2018-05-04 15:11:19 +02:00
Bartosz Taudul
e058bb34c1 CompressThread body must be available. 2018-05-03 18:43:51 +02:00
Bartosz Taudul
b18841aa75 Store ordered list of memory frees. 2018-05-02 17:59:50 +02:00
Bartosz Taudul
754e79b443 Setup memory plot pointer on dump load. 2018-05-02 17:18:52 +02:00
Bartosz Taudul
7266a979c3 Omit stack. 2018-05-01 02:13:49 +02:00
Bartosz Taudul
8beb1c1a39 Add thread compression cache.
Observation: calls to CompressThread() are likely to be repeated with
the same value. Exploit that by storing last query and its result.
2018-05-01 01:29:25 +02:00
Bartosz Taudul
ec58aa4ce1 Don't increase vector size in each iteration. 2018-04-30 13:57:12 +02:00
Bartosz Taudul
553e3ca38b Optimize mem plot reconstruction loop. 2018-04-30 13:45:36 +02:00
Bartosz Taudul
76f0c8fafe Sort source location zones on a separate thread. 2018-04-30 03:54:09 +02:00
Bartosz Taudul
63e4f6fa04 Directly store values. 2018-04-30 03:30:19 +02:00
Bartosz Taudul
e5cb241c19 Optimize creation of vector of frees. 2018-04-29 13:40:47 +02:00
Bartosz Taudul
3eb73b8d43 Move memory plot reconstruction to a background thread. 2018-04-29 13:40:04 +02:00
Bartosz Taudul
bc84ebc338 Read/write LockEvent data in one go. 2018-04-29 03:41:58 +02:00
Bartosz Taudul
c5133e0b4e Walk lockmap timeline pointer. 2018-04-29 03:41:58 +02:00
Bartosz Taudul
9769cc4d7d Read/write most of MemEvent in one go. 2018-04-29 03:41:58 +02:00
Bartosz Taudul
d5f0f0939d No need to track min memory usage.
At least if client instrumentation was not broken and the data makes
sense.
2018-04-29 02:57:20 +02:00
Bartosz Taudul
7fdc6f5453 Zero as initial max value is fine too. 2018-04-29 02:56:23 +02:00
Bartosz Taudul
723f98d24b Overflow checks are not needed. 2018-04-29 02:47:25 +02:00
Bartosz Taudul
b06f445de9 Don't use stack to write two values... 2018-04-29 02:32:20 +02:00
Bartosz Taudul
333d3a92c8 Perform memory usage calculation on doubles. 2018-04-29 02:29:06 +02:00
Bartosz Taudul
aceaed25b9 Walk plot data pointer. 2018-04-29 02:11:47 +02:00
Bartosz Taudul
868fbace5a Don't compress thread twice, if it's the same. 2018-04-29 02:04:51 +02:00
Bartosz Taudul
fdaebc2bd8 No need to perform space check here. 2018-04-29 01:38:54 +02:00
Bartosz Taudul
d64f0390da Don't use std::sort. 2018-04-29 01:23:30 +02:00
Bartosz Taudul
7df7bf1745 Begin memory plot with no memory usage. 2018-04-28 16:26:45 +02:00
Bartosz Taudul
a0b8ed2e50 Restore memory plot when loading data dump. 2018-04-28 16:26:45 +02:00
Bartosz Taudul
d8bfe7de2e Create memory plot based on memory alloc/free events. 2018-04-28 15:49:12 +02:00
Bartosz Taudul
cd34ed6968 Two plot types: user and memory.
Only user plots are saved in a dump file.
2018-04-28 15:48:05 +02:00
Bartosz Taudul
1fb47899b2 Fix skipping lock data with new dump version. 2018-04-22 01:26:51 +02:00
Bartosz Taudul
436cd2b6cf Drop '###Profiler' from capture name. 2018-04-21 23:29:28 +02:00
Bartosz Taudul
d1e185e176 Cleanup message data. 2018-04-21 20:36:33 +02:00
Bartosz Taudul
4cd9cf5dd9 Cleanup zone data. 2018-04-21 20:34:29 +02:00
Bartosz Taudul
0de5bcacaf Free plot data. 2018-04-21 20:12:16 +02:00
Bartosz Taudul
dda25cf66a Cosmetics. 2018-04-21 20:11:59 +02:00
Bartosz Taudul
cb298893e7 Fix skipping lock data. 2018-04-21 16:02:36 +02:00
Bartosz Taudul
121cced681 Don't save unneeded lock data.
Store only the minimal lock information required and calculate lock
counts, wait lists, etc. at load time.
2018-04-21 15:42:08 +02:00
Bartosz Taudul
a63f214964 Use static assert where static assert is due. 2018-04-21 14:47:15 +02:00
Bartosz Taudul
36efe96e9d Throw exception when trying to open unsupported dump version. 2018-04-21 14:18:42 +02:00
Bartosz Taudul
d9fd1ce74a Add dump file header. 2018-04-21 13:45:48 +02:00
Bartosz Taudul
84fd351fba Allow partial load of data from dump. 2018-04-20 16:03:09 +02:00
Bartosz Taudul
3df7c70f99 Optimize mem alloc processing. 2018-04-10 16:06:01 +02:00
Bartosz Taudul
be50fb26b5 Remove useless assert. 2018-04-10 14:37:17 +02:00
Bartosz Taudul
4e1dbb3973 Fix lock announce processing. 2018-04-09 14:28:40 +02:00
Bartosz Taudul
bf99bff87d Store MemEvents directly in the vector. 2018-04-03 14:17:51 +02:00
Bartosz Taudul
821b08fbe4 Thread compression state is not preserved. 2018-04-02 14:52:36 +02:00
Bartosz Taudul
1fa943d109 Save/load memory data. 2018-04-02 02:05:39 +02:00
Bartosz Taudul
52f59c90bf Track memory usage. 2018-04-02 00:00:49 +02:00
Bartosz Taudul
a574f98f0c Memory events are now serialized. 2018-04-01 20:13:01 +02:00
Bartosz Taudul
b12375815c Broken memory events processing. 2018-04-01 02:03:34 +02:00
Bartosz Taudul
991fc6bd95 Memory allocations tracker. 2018-03-31 21:56:05 +02:00
Bartosz Taudul
225423bd21 Cosmetics. 2018-03-24 14:42:48 +01:00
Bartosz Taudul
a9e1a9bddb Calculate total time spent in source location.
This simple solution doesn't handle recursion at all.
2018-03-24 14:24:30 +01:00
Bartosz Taudul
fea0234a60 Change zone end "-1" comparisons to "0" comparisons. 2018-03-24 02:00:20 +01:00
Bartosz Taudul
6a4e58b545 Force inline compress/decompress thread id. 2018-03-24 01:31:58 +01:00
Bartosz Taudul
69b49f527d Inline GetZoneEndDirect(). 2018-03-23 02:06:44 +01:00
Bartosz Taudul
6e6addfa81 Use pdqsort. 2018-03-20 19:19:07 +01:00