custom trc: another small speedup, #5025

This commit is contained in:
heckflosse
2018-11-25 16:21:47 +01:00
parent c9817e369c
commit a3549a6247

View File

@@ -443,29 +443,26 @@ void ImProcFunctions::workingtrc(Imagefloat* src, Imagefloat* dst, int cw, int c
{
dx = dz = 1.0;
}
double toxyz[3][3] = {
const float toxyz[3][3] = {
{
(wprof[0][0] / (dx * (normalizeIn ? 65535.0 : 1.0))), //I have suppressed / Color::D50x
(wprof[0][1] / (dx * (normalizeIn ? 65535.0 : 1.0))),
(wprof[0][2] / (dx * (normalizeIn ? 65535.0 : 1.0)))
static_cast<float>(wprof[0][0] / (dx * (normalizeIn ? 65535.0 : 1.0))), //I have suppressed / Color::D50x
static_cast<float>(wprof[0][1] / (dx * (normalizeIn ? 65535.0 : 1.0))),
static_cast<float>(wprof[0][2] / (dx * (normalizeIn ? 65535.0 : 1.0)))
}, {
(wprof[1][0] / (normalizeIn ? 65535.0 : 1.0)),
(wprof[1][1] / (normalizeIn ? 65535.0 : 1.0)),
(wprof[1][2] / (normalizeIn ? 65535.0 : 1.0))
static_cast<float>(wprof[1][0] / (normalizeIn ? 65535.0 : 1.0)),
static_cast<float>(wprof[1][1] / (normalizeIn ? 65535.0 : 1.0)),
static_cast<float>(wprof[1][2] / (normalizeIn ? 65535.0 : 1.0))
}, {
(wprof[2][0] / (dz * (normalizeIn ? 65535.0 : 1.0))), //I have suppressed / Color::D50z
(wprof[2][1] / (dz * (normalizeIn ? 65535.0 : 1.0))),
(wprof[2][2] / (dz * (normalizeIn ? 65535.0 : 1.0)))
static_cast<float>(wprof[2][0] / (dz * (normalizeIn ? 65535.0 : 1.0))), //I have suppressed / Color::D50z
static_cast<float>(wprof[2][1] / (dz * (normalizeIn ? 65535.0 : 1.0))),
static_cast<float>(wprof[2][2] / (dz * (normalizeIn ? 65535.0 : 1.0)))
}
};
double pwr;
double ts;
ts = slpos;
double pwr = 1.0 / gampos;
double ts = slpos;
int five = mul;
pwr = 1.0 / gampos;
if (gampos < 1.0) {
pwr = gampos;
@@ -474,7 +471,7 @@ void ImProcFunctions::workingtrc(Imagefloat* src, Imagefloat* dst, int cw, int c
}
// int select_temp = 1; //5003K
const double eps = 0.000000001; // not divide by zero
constexpr double eps = 0.000000001; // not divide by zero
enum class ColorTemp {
D50 = 5003, // for Widegamut, ProPhoto Best, Beta -> D50
@@ -484,10 +481,8 @@ void ImProcFunctions::workingtrc(Imagefloat* src, Imagefloat* dst, int cw, int c
};
ColorTemp temp = ColorTemp::D50;
cmsHPROFILE oprofdef;
float p[6]; //primaries
if (true) {
//primaries for 10 working profiles ==> output profiles
if (profile == "WideGamut") {
p[0] = 0.7350; //Widegamut primaries
@@ -608,14 +603,28 @@ void ImProcFunctions::workingtrc(Imagefloat* src, Imagefloat* dst, int cw, int c
}
GammaTRC[0] = GammaTRC[1] = GammaTRC[2] = cmsBuildParametricToneCurve(NULL, five, gammaParams);//5 = more smoother than 4
oprofdef = cmsCreateRGBProfile(&xyD, &Primaries, GammaTRC);
const cmsHPROFILE oprofdef = cmsCreateRGBProfile(&xyD, &Primaries, GammaTRC);
cmsFreeToneCurve(GammaTRC[0]);
}
if (oprofdef) {
#pragma omp parallel for if (multiThread)
cmsUInt32Number flags = cmsFLAGS_NOOPTIMIZE | cmsFLAGS_NOCACHE;
cmsHPROFILE iprof = ICCStore::getInstance()->getXYZProfile();
lcmsMutex->lock();
// cmsHTRANSFORM hTransform = cmsCreateTransform(iprof, TYPE_RGB_16, oprofdef, TYPE_RGB_16, params->icm.outputIntent, cmsFLAGS_NOOPTIMIZE | cmsFLAGS_NOCACHE);
cmsHTRANSFORM hTransform = cmsCreateTransform(iprof, TYPE_RGB_FLT, oprofdef, TYPE_RGB_FLT, params->icm.outputIntent, flags);
lcmsMutex->unlock();
#ifdef _OPENMP
#pragma omp parallel if (multiThread)
#endif
{
AlignedBuffer<float> pBuf(cw * 3);
#ifdef _OPENMP
#pragma omp for schedule(dynamic, 16)
#endif
for (int i = 0; i < ch; i++) {
float *p = pBuf.data;
float* rr = src->r(i);
float* rg = src->g(i);
float* rb = src->b(i);
@@ -629,28 +638,24 @@ void ImProcFunctions::workingtrc(Imagefloat* src, Imagefloat* dst, int cw, int c
float g1 = rg[j];
float b1 = rb[j];
xa[j] = toxyz[0][0] * r1 + toxyz[0][1] * g1 + toxyz[0][2] * b1;
ya[j] = toxyz[1][0] * r1 + toxyz[1][1] * g1 + toxyz[1][2] * b1;
za[j] = toxyz[2][0] * r1 + toxyz[2][1] * g1 + toxyz[2][2] * b1;
*(p++) = toxyz[0][0] * r1 + toxyz[0][1] * g1 + toxyz[0][2] * b1;
*(p++) = toxyz[1][0] * r1 + toxyz[1][1] * g1 + toxyz[1][2] * b1;
*(p++) = toxyz[2][0] * r1 + toxyz[2][1] * g1 + toxyz[2][2] * b1;
}
cmsDoTransform (hTransform, pBuf.data, pBuf.data, cw);
p = pBuf.data;
for (int x = 0; x < cw; x++) {
*(xa++) = *(p++);
*(ya++) = *(p++);
*(za++) = *(p++);
}
}
}
cmsUInt32Number flags = cmsFLAGS_NOOPTIMIZE | cmsFLAGS_NOCACHE;
lcmsMutex->lock();
cmsHPROFILE iprof = ICCStore::getInstance()->getXYZProfile();
// cmsHTRANSFORM hTransform = cmsCreateTransform(iprof, TYPE_RGB_16, oprofdef, TYPE_RGB_16, params->icm.outputIntent, cmsFLAGS_NOOPTIMIZE | cmsFLAGS_NOCACHE);
cmsHTRANSFORM hTransform = cmsCreateTransform(iprof, TYPE_RGB_FLT, oprofdef, TYPE_RGB_FLT, params->icm.outputIntent, flags);
lcmsMutex->unlock();
dst->ExecCMSTransform(hTransform);
cmsDeleteTransform(hTransform);
if (normalizeOut) {
dst->normalizeFloatTo65535();
}
}
}