rpitx/csdr
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

fp16: disable SSE if we're on ARM.

Browse source
This commit is contained in:
ha7ilm 2016-04-03 19:47:05 +02:00
parent 8a2618d2ab
commit 906d91e84a
2 changed files with 6 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
fp16.c
View file

@ -416,6 +416,7 @@ FP16 approx_float_to_half(FP32 f)
return o; return o;
} }
#ifndef NEON_OPTS
// round-half-up (same as ISPC) // round-half-up (same as ISPC)
__m128i float_to_half_SSE2(__m128 f) __m128i float_to_half_SSE2(__m128 f)
{ {
@ -547,6 +548,7 @@ __m128i approx_float_to_half_SSE2(__m128 f)
#undef CONSTF #undef CONSTF
} }
#endif
// from fox toolkit float->half code (which "approx" variants match) // from fox toolkit float->half code (which "approx" variants match)
static uint basetable[512]; static uint basetable[512];

4
fp16.h
View file

@ -106,7 +106,9 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include <stdio.h> #include <stdio.h>
#include <stdlib.h> #include <stdlib.h>
#ifndef NEON_OPTS
#include <emmintrin.h> #include <emmintrin.h>
#endif
typedef unsigned int uint; typedef unsigned int uint;
@ -144,9 +146,11 @@ FP16 float_to_half_fast2(FP32 f);
FP16 float_to_half_fast3(FP32 f); FP16 float_to_half_fast3(FP32 f);
FP16 float_to_half_fast3_rtne(FP32 f); FP16 float_to_half_fast3_rtne(FP32 f);
FP16 approx_float_to_half(FP32 f); FP16 approx_float_to_half(FP32 f);
#ifndef NEON_OPTS
__m128i float_to_half_SSE2(__m128 f); __m128i float_to_half_SSE2(__m128 f);
__m128i float_to_half_rtne_SSE2(__m128 f); __m128i float_to_half_rtne_SSE2(__m128 f);
__m128i approx_float_to_half_SSE2(__m128 f); __m128i approx_float_to_half_SSE2(__m128 f);
#endif
void fp16_generatetables(); void fp16_generatetables();
uint float_to_half_foxtk(uint f); uint float_to_half_foxtk(uint f);
FP32 half_to_float(FP16 h); FP32 half_to_float(FP16 h);