diff options
| author | Julian Taylor <jtaylor.debian@googlemail.com> | 2016-11-29 00:19:21 +0100 |
|---|---|---|
| committer | Julian Taylor <jtaylor.debian@googlemail.com> | 2017-01-12 17:17:07 +0100 |
| commit | f0f7ad80f2ef2d7525965dfe27c0e2ab68647197 (patch) | |
| tree | 885b69a2c89ca924395da8f908f5ba2c48383864 /numpy/core/setup_common.py | |
| parent | 7e6091c9a3fc4536ccbadb337e88650b2c901313 (diff) | |
| download | numpy-f0f7ad80f2ef2d7525965dfe27c0e2ab68647197.tar.gz | |
ENH: vectorize packbits with SSE2
SSE2 has a special instruction to pack bytes into bits,
available as the intrinsic _mm_movemask_epi8. It is significantly faster
than the per byte loop currently being used.
Unfortunately packbits is bitwise "big endian", the most significant bit
is the first in the input byte while _mm_movemask_epi is little endian
so we need to byteswap the input first. But it is still about 8-10 times
faster than the scalar code.
Diffstat (limited to 'numpy/core/setup_common.py')
| -rw-r--r-- | numpy/core/setup_common.py | 2 |
1 files changed, 2 insertions, 0 deletions
diff --git a/numpy/core/setup_common.py b/numpy/core/setup_common.py index 18066d991..596b3996c 100644 --- a/numpy/core/setup_common.py +++ b/numpy/core/setup_common.py @@ -130,6 +130,8 @@ OPTIONAL_INTRINSICS = [("__builtin_isnan", '5.'), # broken on OSX 10.11, make sure its not optimized away ("volatile int r = __builtin_cpu_supports", '"sse"', "stdio.h", "__BUILTIN_CPU_SUPPORTS"), + # MMX only needed for icc, but some clangs don't have it + ("_m_from_int64", '0', "emmintrin.h"), ("_mm_load_ps", '(float*)0', "xmmintrin.h"), # SSE ("_mm_prefetch", '(float*)0, _MM_HINT_NTA', "xmmintrin.h"), # SSE |