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
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
|
# Benchmarking `CompCert` and GCC
## Compiling `CompCert`
The first step to benchmark `CompCert` is to compile it - the `INSTALL.md` instructions of the project root folder should guide you on installing it.
For the benchmarks to work, the compiler `ccomp` should be on your `$PATH`, with the runtime libraries installed correctly (with a successful `make install` on the project root directory).
## Using the harness
`rules.mk` contains generic rules to compile with `gcc` and `ccomp`, with different optimizations, and producing different binaries. It also produces a `measures.csv` file containing the different timings given by the bench.
Up to 5 different sets of optimizations per compiler can be used.
To use this `rules.mk`, create a folder, put inside all the .c/.h source files, and write a Makefile resembling:
```make
TARGET=float_mat
MEASURES="c1 c2 c3 c4 c5 c6 c7 c8"
include ../rules.mk
```
This is all that is required to write, the `rules.mk` handles everything.
There is the possibility to define some variables to fine tune what you want. For instance, `ALL_CFILES` describes the .c source files whose objects are to be linked.
Here is an exhaustive list of the variables:
- `TARGET`: name of the binary to produce
- `MEASURES`: list of the different timings. This supposes that the program
prints something of the form `c3 cycles: 44131`.
- `ALL_CFILES`: list of .c files to compile. By default, `$(wildcard *.c)`
- `CLOCK`: `basename` of the clock file to compile. Default `../clock`
- `ALL_CFLAGS`: `cflags` that are to be included for all compilers
- `ALL_GCCFLAGS`: same, but GCC specific
- `ALL_CCOMPFLAGS`: same, but `ccomp` specific
- `K1C_CC`: GCC compiler (default `k1-cos-gcc`)
- `K1C_CCOMP`: `CompCert` compiler (default `ccomp`)
- `EXECUTE_CYCLES`: running command (default is `k1-cluster --syscall=libstd_scalls.so --cycle-based --`)
- `EXECUTE_ARGS`: execution arguments. You can use a macro `__BASE__` which expands to the name of the binary being executed.
- `GCCiFLAGS` with `i` from 0 to 4: the wanted optimizations. If one of these flags is empty, nothing is done. Same for `CCOMPiFLAGS`. Look at `rules.mk` to see the default values. You might find something like this:
# You can define up to GCC4FLAGS and CCOMP4FLAGS
GCC0FLAGS?=
GCC1FLAGS?=$(ALL_GCCFLAGS) -O1
GCC2FLAGS?=$(ALL_GCCFLAGS) -O2
GCC3FLAGS?=$(ALL_GCCFLAGS) -O3
GCC4FLAGS?=
CCOMP0FLAGS?=
CCOMP1FLAGS?=$(ALL_CCOMPFLAGS) -fno-postpass
CCOMP2FLAGS?=$(ALL_CCOMPFLAGS)
CCOMP3FLAGS?=
CCOMP4FLAGS?=
# Prefix names
GCC0PREFIX?=
GCC1PREFIX?=.gcc.o1
GCC2PREFIX?=.gcc.o2
GCC3PREFIX?=.gcc.o3
GCC4PREFIX?=
CCOMP0PREFIX?=
CCOMP1PREFIX?=.ccomp.o1
CCOMP2PREFIX?=.ccomp.o2
CCOMP3PREFIX?=
CCOMP4PREFIX?=
The `PREFIX` are the prefixes to add to the secondary produced files (assembly, object, executable, ..). You should be careful that if a `FLAGS` is set, then the according `PREFIX` should be set as well.
Assembly files are generated in `asm/`, objects in `obj/`, binaries in `bin/` and outputs in `out/`.
To compile and execute all the benches : `make` while in the `monniaux` directory (without any `-j` flag). Doing so will compile CompCert, install it, and then proceed to execute each bench.
To compile and/or execute a single bench, `cd` to the bench directory, then:
- `make` for compiling the bench
- `make run` for running it
You can use `-j` flag when in a single bench directory.
## Individual scripts
If you want to run the building and running scripts individually without having to use the `Makefile` from `test/monniaux`, you can run the `build_benches.sh` script which builds each bench using all the available cores on your machine.
Once the benches are built, you can then run `run_benches.sh file.csv` where `file.csv` is where you want to store the timings of the benchmarks. `run_benches.sh` also uses all the available cores of your machine.
## Adding timings to a benchmark
If you just add a benchmark without any timing function, the resulting `measures.csv` file will be empty for lack of timing output.
To add a timing, you must use the functions whose prototypes are in `clock.h`
#include "../clock.h"
/* ... */
clock_prepare();
/* ... */
clock_start();
/* .. computations .. */
clock_stop();
/* ... */
print_total_clock(); // print to stdout
printerr_total_clock(); // print to stderr
If the benchmark doesn't use `stdout` in a binary way you can use `print_total_clock()`. However, some benchmarks like `jpeg-6b` print their binary content to `stdout`, which then messes up the `grep` command when attempting to use it to extract the cycles from `stdout`.
The solution is then to use `printerr_total_clock()` which will print the cycles to `stderr`, and use `EXECUTE_ARGS` ressembling this:
EXECUTE_ARGS=-dct int -outfile __BASE__.jpg testimg.ppm 2> __BASE__.out
`__BASE__` is a macro that gets expanded to the base name - that is, the `TARGET` concatenated with one of the `GCCiPREFIX` or `CCOMPiPREFIX`. For instance, in `jpeg-6b`, `__BASE__` could be `jpeg-6b.ccomp.o2`.
|