The Clear Linux project is currently focusing on applying function multi-version (FMV) technology to packages where it is detected that AVX instructions can yield a performance improvement. To solve some of the issues involved with supporting FMV in a full Linux distribution, the project provides a patch generator based on vectorization-candidate detection (using the gcc -fopt-info-vec
compiler flag). This tool can provide all the FMV patches that a Linux distribution might use.
Although this patch tool was originally designed to be used for Clear Linux OS, the patch generator can be applied to any project built with gcc.
The following sample code will be used:
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#define MAX 1000000
int a[256], b[256], c[256];
void foo()
{
int i,x;
for (x=0; x<MAX; x++)
{
for (i=0; i<256; i++)
{
a[i] = b[i] + c[i];
}
}
}
int main()
{
foo();
return 0;
}
gcc -O2 -ftree-vectorize -fopt-info-vec example.c -o example &> log.txt
Breaking down the command:
- The
-ftree-vectorize
flag enables vectorization. Note: it is enabled by default at-O3
- The
-fopt-info-vec
generates a report to stdout providing details of code that the compiler was able to vectorize. - The
&> log.txt
pipes the output to a file of the same name.
Upon examing the log.txt file (using cat log.txt
), it can be seen that gcc was able to vectorize parts of the code:
example.c:13:13: note: loop vectorized
example.c:13:13: note: loop vectorized
The patch tool takes the format
perl make-fmv-patch.pl <buildlog> <sourcecode>
After checking out the tool from the repository, run the patch tool using log.txt
as input:
perl make-fmv-patch.pl log.txt .
Upon completion, the patch tool will output log confirming that example.c has been patched:
patching ./example.c @ lines (8)
The command generates the following example.c.patch patch:
--- ./example.c 2019-01-13 00:04:47.646072680 +0000
+++ ./example.c~ 2019-01-13 00:04:57.586072680 +0000
@@ -5,6 +5,7 @@
int a[256], b[256], c[256];
+__attribute__((target_clones("avx2","arch=atom","default")))
void foo()
{
int i,x;
Using the example.c.patch file, apply the patch to example.c:
patch -p1 < example.c.patch
The example.c file should have been patched, and look something like:
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#define MAX 1000000
int a[256], b[256], c[256];
__attribute__((target_clones("avx2","arch=atom","default")))
void foo()
{
int i,x;
for (x=0; x<MAX; x++)
{
for (i=0; i<256; i++)
{
a[i] = b[i] + c[i];
}
}
}
int main()
{
foo();
return 0;
}
Compile the code again (with FMV syntax having been applied to the code):
gcc -O2 -ftree-vectorize -fopt-info-vec example.c -o example
The console output should contain:
example.c:14:13: note: loop vectorized
example.c:14:13: note: loop vectorized
example.c:14:13: note: loop vectorized
Voila!
- GCC v6 compiler or newer
- Perl
- Exuberant ctags:
In order to use
make-fvm-patch.pl
your system will need Exuberant Ctags (https://sourceforge.net/projects/ctags/), this is due to the--c-kinds
ctags' flag usage.
The repostory makes available a tests directory. Users can use the tests directory to conduct their own tests and/or share issues with sample code (for easy reproducability).
cd tests
make
cd ..
Once built, follow the steps outlined in the example above:
perl make-fmv-patch.pl tests/max-build.log tests/
perl make-fmv-patch.pl tests/loop-build.log tests/
then...
- Apply the patches
- Recompile the code
- Run
- Clear Linux: Use the function multi-version patch generator