Age | Commit message (Collapse) | Author |
|
|
|
|
|
|
|
gives better results than those platforms on avg
Test cases changed to take ~1s each on my i7 laptop
|
|
|
|
|
|
|
|
|
|
|
|
|
|
lack of _manual() capability.
|
|
|
|
|
|
Testing manually works OK. All that's left is top-level volk.c and Automake wrangling.
|
|
conditional linking yet.
|
|
Conflicts:
volk/Makefile.common
volk/lib/qa_utils.cc
|
|
|
|
In volk, these are now actual files instead of symlinks
|
|
|
|
|
|
|
|
* jblum/mergeme/volk/remove_posix_memalign:
volk: simplify the get new method for the aligned pool
volk: replace posix_memalign with something cross platform
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
like in volk.c.
The first run of each function does the rank_archs call. Subsequent calls proceed with no overhead.
volk_init is still being generated but not used at all.
|
|
|
|
when -mavx is used, gcc will promote SSE instructions to their AVX equivalents. while this is a benefit on AVX architectures,
it makes it really difficult to selectively enable code based on architecture at runtime. GCC suggests compiling different
files with different flags and loading them at runtime. GCC can take a flying leap.
|
|
straight from python -> .h
|
|
This reverts commit d364316fa2e047890af09c7bdeed776d6391b3db.
|
|
access to the static stuff
|
|
|
|
|
|
Configure- and compile-time checks modified to only compile architectures that the compiler will handle.
This means that volk will compile every arch that your gcc will compile, no matter if your CPU can execute the instructions or not.
This lets you cross-compile volk.
volk_rank_archs will be deferred to runtime. This has not been done yet.
|
|
|
|
|
|
Conflicts:
volk/include/volk/Makefile.am
volk/lib/Makefile.am
volk/lib/qa_volk.cc
|
|
Distcheck passes for me if liborc is installed or not.
|
|
|
|
|
|
|
|
|
|
breakage when adding new volk kernels.
|
|
|
|
Note: need to convert this to new naming standard.
|
|
|