-
Notifications
You must be signed in to change notification settings - Fork 4
/
README
78 lines (49 loc) · 2.84 KB
/
README
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
Paris/PPCG/Polyhedral/PRL Runtime Library (PRL)
===============================================
Compiling
----------
./autogen.sh
./configure
make
The user needs to provide the path to the OpenCL header files
and the path to the OpenCL library. This can be done for example
by adding CFLAGS="-I<PATH_TO_OPENCL_HEADERS> -L<PATH_TO_OPENCL_LIB>"
when calling configure. For example:
./configure CFLAGS="-I/opt/AMDAPP/include/ -L/opt/AMDAPP/lib/x86_64/"
Usage
-----
Either use pencilcc/penciltool
- or -
add -lprl_opencl to the linker line. When compiling, add 'prl/include' to the header search path (or install them to the default header search path).
The library will initialize on its first use.
Profiling
---------
This version of the runtime library can profile PENCIL programs automatically. Note that this implementation is not really prepared to be used from multiple threads.
### Ad-hoc Profiling (stats)
This can be enabled by setting the environment variables
PRL_DUMP_CPU=1
PRL_DUMP_GPU=1
PRL_DUMP_ALL=1
sets both of them. The will output statistics to stdout.
PRL_DUMP_CPU will print how long calls to the OpenCL API took on the CPU. PRL_DUMP_GPU prints the durations of tasks on the GPU as reported by OpenCL itself. It is printed as summary when the program ends. PRL_TRACE_GPU will print the duration of every OpenCL queue item.
### Benchmarking (timings)
Profiling a single function call can be unreliable do to noise and one-time effects. Therefore there is also a mechanism to measure a piece of code multiple times. The easiest way to do this is to create a new program that call the function
prl_prof_benchmark(timed_func, user, init_callback, init_user, finit_callback, finit_user)
which does basically this:
prl_prof_reset();
for (auto i = 0; i < PRL_PROF_DRY_RUNS; ++i) {
if (init_callback) (*init_callback)(init_user);
(*timed_func)(user);
if (finit_callback) (*finit_callback)(finit_user);
}
for (auto i = 0; i < PRL_PROF_RUNS; ++i) {
if (init_callback) (*init_callback)(init_user);
prl_prof_start();
(*timed_func)(user);
prl_prof_stop();
if (finit_callback) (*finit_callback)(finit_user);
}
prl_prof_dump();
PRL_PROF_RUNS and PRL_PROF_DRY_RUNS are environment variables that can be set. The defaults are 1 for dry runs and 10 for timed runs. Dry runs allow eliminating first-time effect like cold memory cashes and compiling the OpenCL program.
The execution time between 'prl_prof_start()' and 'prl_prof_stop()' is measured. Since it is executed multiple times, it records the time of every execution. The data is printed to stdout on 'prl_prof_dump()'. It prints the median times and relative standard deviation of all profiling items.
It might not be possible to use the 'prl_timings' function in you application. In this case, one can do it manually by following the structure of the prl_timings snippet.