liz/rawTherapee
liz/rawTherapee
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Code review and speedup for Amaze Demosaic

Browse Source
This commit is contained in:
heckflosse 2016-01-24 01:44:35 +01:00
parent e5b1abdc3b
commit 2017a0e592
4 changed files with 884 additions and 868 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1697
rtengine/amaze_demosaic_RT.cc
View File

File diff suppressed because it is too large Load Diff

9
rtengine/helpersse2.h
View File

@ -4,7 +4,6 @@
#ifdef __GNUC__
#define INLINE __inline
//#define INLINE __attribute__((always_inline))
#else
#define INLINE inline
#endif
@ -48,20 +47,20 @@ typedef __m128i vint2;
// SSE4.1 => use _mm_blend_ps instead of _mm_set_epi32 and vself
#define STC2VFU(a,v) {\
__m128 TST1V = _mm_loadu_ps(&a);\
__m128 TST2V = _mm_shuffle_ps(v,v,_MM_SHUFFLE( 1,1,0,0 ));\
__m128 TST2V = _mm_unpacklo_ps(v,v);\
_mm_storeu_ps(&a, _mm_blend_ps(TST1V,TST2V,5));\
TST1V = _mm_loadu_ps((&a)+4);\
TST2V = _mm_shuffle_ps(v,v,_MM_SHUFFLE( 3,3,2,2 ));\
TST2V = _mm_unpackhi_ps(v,v);\
_mm_storeu_ps((&a)+4, _mm_blend_ps(TST1V,TST2V,5));\
}
#else
#define STC2VFU(a,v) {\
__m128 TST1V = _mm_loadu_ps(&a);\
__m128 TST2V = _mm_shuffle_ps(v,v,_MM_SHUFFLE( 1,1,0,0 ));\
__m128 TST2V = _mm_unpacklo_ps(v,v);\
vmask cmask = _mm_set_epi32(0xffffffff,0,0xffffffff,0);\
_mm_storeu_ps(&a, vself(cmask,TST1V,TST2V));\
TST1V = _mm_loadu_ps((&a)+4);\
TST2V = _mm_shuffle_ps(v,v,_MM_SHUFFLE( 3,3,2,2 ));\
TST2V = _mm_unpackhi_ps(v,v);\
_mm_storeu_ps((&a)+4, vself(cmask,TST1V,TST2V));\
}
#endif

10
rtengine/rt_math.h
View File

@ -78,5 +78,15 @@ inline const _Tp& max(const _Tp& a, const _Tp& b, const _Tp& c, const _Tp& d)
{
return std::max(d, std::max(c, std::max(a, b)));
}
template<typename _Tp>
inline const _Tp intp(const _Tp a, const _Tp b, const _Tp c) {
// calculate a * b + (1 - a) * c
// following is valid:
// intp(a, b+x, c+x) = vintpf(a, b, c) + x
// intp(a, b*x, c*x) = vintpf(a, b, c) * x
return a * (b-c) + c;
}
}
#endif

36
rtengine/sleefsseavx.c
View File

@ -1316,7 +1316,9 @@ return vmaxf( b, vminf(a,c));
}
static INLINE vfloat ULIMV( vfloat a, vfloat b, vfloat c ){
return vself( vmaskf_lt(b,c), LIMV(a,b,c), LIMV(a,c,b));
// made to clamp a in range [b,c] but in fact it's also the median of a,b,c, which means that the result is independent on order of arguments
// ULIMV(a,b,c) = ULIMV(a,c,b) = ULIMV(b,a,c) = ULIMV(b,c,a) = ULIMV(c,a,b) = ULIMV(c,b,a)
return vmaxf(vminf(a,b), vminf(vmaxf(a,b),c));
}
static INLINE vfloat SQRV(vfloat a){
@ -1324,17 +1326,45 @@ static INLINE vfloat SQRV(vfloat a){
}
static inline void vswap( vmask condition, vfloat &a, vfloat &b) {
// conditional swap the elements of two vfloats
vfloat temp = vself(condition, a, b); // the values which fit to condition
condition = vnotm(condition); // invert the condition
a = vself(condition, a, b); // the values which fit to inverted condition
b = temp;
}
static inline float vhadd( vfloat a )
{
static inline float vhadd( vfloat a ) {
// returns a[0] + a[1] + a[2] + a[3]
a += _mm_movehl_ps(a, a);
return _mm_cvtss_f32(_mm_add_ss(a, _mm_shuffle_ps(a, a, 1)));
}
static INLINE vfloat vmul2f(vfloat a){
// fastest way to multiply by 2
return a + a;
}
static INLINE vfloat vintpf(vfloat a, vfloat b, vfloat c) {
// calculate a * b + (1 - a) * c (interpolate two values)
// following is valid:
// vintpf(a, b+x, c+x) = vintpf(a, b, c) + x
// vintpf(a, b*x, c*x) = vintpf(a, b, c) * x
return a * (b-c) + c;
}
static INLINE vfloat vdup(vfloat a){
// returns { a[0],a[0],a[1],a[1] }
return _mm_unpacklo_ps( a, a );
}
static INLINE vfloat vaddc2vfu(float &a)
{
// loads a[0]..a[7] and returns { a[0]+a[1], a[2]+a[3], a[4]+a[5], a[6]+a[7] }
vfloat a1 = _mm_loadu_ps( &a );
vfloat a2 = _mm_loadu_ps( (&a) + 4 );
return _mm_shuffle_ps(a1,a2,_MM_SHUFFLE( 2,0,2,0 )) + _mm_shuffle_ps(a1,a2,_MM_SHUFFLE( 3,1,3,1 ));
}
#endif // __SSE2__
#endif // SLEEFSSEAVX