LCMS performance optimizations

This commit is contained in:
Oliver Duis 2011-04-09 17:49:44 +02:00
parent 31791b268f
commit 6024cb817a
6 changed files with 84 additions and 44 deletions

View File

@ -271,3 +271,10 @@ Image16::tofloat() const
} }
return imgfloat; return imgfloat;
} }
// Parallized transformation; create transform with cmsFLAGS_NOCACHE!
void Image16::ExecCMSTransform(cmsHTRANSFORM hTransform) {
#pragma omp parallel for
for (int i=0; i<height; i++)
cmsDoTransform(hTransform, data + 3*i*rowstride, data + 3*i*rowstride, rowstride);
}

View File

@ -87,6 +87,8 @@ class Image16 : public ImageIO, public IImage16 {
virtual unsigned short** getRPlane () { return r; } virtual unsigned short** getRPlane () { return r; }
virtual unsigned short** getGPlane () { return g; } virtual unsigned short** getGPlane () { return g; }
virtual unsigned short** getBPlane () { return b; } virtual unsigned short** getBPlane () { return b; }
void ExecCMSTransform(cmsHTRANSFORM hTransform);
}; };
}; };
#endif #endif

View File

@ -21,6 +21,7 @@
#include <image8.h> #include <image8.h>
#include <string.h> #include <string.h>
#include <rtengine.h> #include <rtengine.h>
#include <mytime.h>
using namespace rtengine; using namespace rtengine;
@ -234,4 +235,9 @@ Imagefloat::to16() const
return img16; return img16;
} }
// Parallized transformation; create transform with cmsFLAGS_NOCACHE!
void Imagefloat::ExecCMSTransform(cmsHTRANSFORM hTransform) {
#pragma omp parallel for
for (int i=0; i<height; i++)
cmsDoTransform(hTransform, data + 3*i*rowstride, data + 3*i*rowstride, rowstride);
}

View File

@ -89,6 +89,8 @@ class Imagefloat : public ImageIO, public IImagefloat {
virtual float** getRPlane () { return r; } virtual float** getRPlane () { return r; }
virtual float** getGPlane () { return g; } virtual float** getGPlane () { return g; }
virtual float** getBPlane () { return b; } virtual float** getBPlane () { return b; }
void ExecCMSTransform(cmsHTRANSFORM hTransform);
}; };
}; };
#endif #endif

View File

@ -1587,7 +1587,8 @@ void RawImageSource::colorSpaceConversion (Imagefloat* im, ColorManagementParams
im->b[i][j] = (newb); im->b[i][j] = (newb);
} }
} else {// use supplied input profile } else {
// use supplied input profile
// color space transform is expecting data in the range (0,1) // color space transform is expecting data in the range (0,1)
for ( int h = 0; h < im->height; ++h ) for ( int h = 0; h < im->height; ++h )
for ( int w = 0; w < im->width; ++w ) { for ( int w = 0; w < im->width; ++w ) {
@ -1597,13 +1598,18 @@ void RawImageSource::colorSpaceConversion (Imagefloat* im, ColorManagementParams
} }
out = iccStore->workingSpace (cmp.working); out = iccStore->workingSpace (cmp.working);
// out = iccStore->workingSpaceGamma (wProfile); // out = iccStore->workingSpaceGamma (wProfile);
lcmsMutex->lock (); lcmsMutex->lock ();
cmsHTRANSFORM hTransform = cmsCreateTransform (in, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), out, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), settings->colorimetricIntent, 0); cmsHTRANSFORM hTransform = cmsCreateTransform (in, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), out, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), settings->colorimetricIntent,
cmsFLAGS_NOCACHE ); // NOCACHE is important for thread safety
lcmsMutex->unlock (); lcmsMutex->unlock ();
if (hTransform) {//there is an input profile
if (hTransform) {
// there is an input profile
if (cmp.gammaOnInput) { if (cmp.gammaOnInput) {
float gd = pow (2.0, defgain); float gd = pow (2.0, defgain);
defgain = 0.0; // Writeback defgain to be 0.0 defgain = 0.0; // Writeback defgain to be 0.0
#pragma omp parallel for #pragma omp parallel for
for (int i=0; i<im->height; i++) for (int i=0; i<im->height; i++)
for (int j=0; j<im->width; j++) { for (int j=0; j<im->width; j++) {
@ -1613,21 +1619,28 @@ void RawImageSource::colorSpaceConversion (Imagefloat* im, ColorManagementParams
im->b[i][j] = CurveFactory::gamma (CLIP(gd*im->b[i][j])); im->b[i][j] = CurveFactory::gamma (CLIP(gd*im->b[i][j]));
} }
} }
cmsDoTransform (hTransform, im->data, im->data, im->planestride);
} else {//create the profile im->ExecCMSTransform(hTransform);
} else {
// create the profile from camera
lcmsMutex->lock (); lcmsMutex->lock ();
hTransform = cmsCreateTransform (camprofile, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), out, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), settings->colorimetricIntent, cmsFLAGS_NOOPTIMIZE); hTransform = cmsCreateTransform (camprofile, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), out, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), settings->colorimetricIntent,
cmsFLAGS_NOOPTIMIZE | cmsFLAGS_NOCACHE ); // NOCACHE is important for thread safety
lcmsMutex->unlock (); lcmsMutex->unlock ();
cmsDoTransform (hTransform, im->data, im->data, im->planestride);
im->ExecCMSTransform(hTransform);
} }
cmsDeleteTransform(hTransform);
// restore normalization to the range (0,65535) // restore normalization to the range (0,65535)
#pragma omp parallel for
for ( int h = 0; h < im->height; ++h ) for ( int h = 0; h < im->height; ++h )
for ( int w = 0; w < im->width; ++w ) { for ( int w = 0; w < im->width; ++w ) {
im->r[h][w] *= 65535.0 ; im->r[h][w] *= 65535.0 ;
im->g[h][w] *= 65535.0 ; im->g[h][w] *= 65535.0 ;
im->b[h][w] *= 65535.0 ; im->b[h][w] *= 65535.0 ;
} }
cmsDeleteTransform(hTransform);
} }
t3.set (); t3.set ();
// printf ("ICM TIME: %d\n", t3.etime(t1)); // printf ("ICM TIME: %d\n", t3.etime(t1));
@ -1642,9 +1655,8 @@ void RawImageSource::colorSpaceConversion16 (Image16* im, ColorManagementParams
if (cmp.input == "(none)") if (cmp.input == "(none)")
return; return;
MyTime t1, t2, t3; //MyTime t1, t2, t3;
//t1.set ();
t1.set ();
cmsHPROFILE in; cmsHPROFILE in;
cmsHPROFILE out; cmsHPROFILE out;
@ -1697,12 +1709,14 @@ void RawImageSource::colorSpaceConversion16 (Image16* im, ColorManagementParams
out = iccStore->workingSpace (cmp.working); out = iccStore->workingSpace (cmp.working);
// out = iccStore->workingSpaceGamma (wProfile); // out = iccStore->workingSpaceGamma (wProfile);
lcmsMutex->lock (); lcmsMutex->lock ();
cmsHTRANSFORM hTransform = cmsCreateTransform (in, TYPE_RGB_16_PLANAR, out, TYPE_RGB_16_PLANAR, settings->colorimetricIntent, 0); cmsHTRANSFORM hTransform = cmsCreateTransform (in, TYPE_RGB_16_PLANAR, out, TYPE_RGB_16_PLANAR, settings->colorimetricIntent, cmsFLAGS_NOCACHE); // NOCACHE is important for thread safety
lcmsMutex->unlock (); lcmsMutex->unlock ();
if (hTransform) { if (hTransform) {
if (cmp.gammaOnInput) { if (cmp.gammaOnInput) {
float gd = pow (2.0, defgain); float gd = pow (2.0, defgain);
defgain = 0.0; defgain = 0.0;
#pragma omp parallel for #pragma omp parallel for
for (int i=0; i<im->height; i++) for (int i=0; i<im->height; i++)
for (int j=0; j<im->width; j++) { for (int j=0; j<im->width; j++) {
@ -1711,17 +1725,21 @@ void RawImageSource::colorSpaceConversion16 (Image16* im, ColorManagementParams
im->b[i][j] = CurveFactory::gamma ((gd*im->b[i][j])); im->b[i][j] = CurveFactory::gamma ((gd*im->b[i][j]));
} }
} }
cmsDoTransform (hTransform, im->data, im->data, im->planestride);
im->ExecCMSTransform(hTransform);
} }
else { else {
lcmsMutex->lock (); lcmsMutex->lock ();
hTransform = cmsCreateTransform (camprofile, TYPE_RGB_16_PLANAR, out, TYPE_RGB_16_PLANAR, settings->colorimetricIntent, 0); hTransform = cmsCreateTransform (camprofile, TYPE_RGB_16_PLANAR, out, TYPE_RGB_16_PLANAR, settings->colorimetricIntent, cmsFLAGS_NOCACHE);
lcmsMutex->unlock (); lcmsMutex->unlock ();
cmsDoTransform (hTransform, im->data, im->data, im->planestride);
im->ExecCMSTransform(hTransform);
} }
cmsDeleteTransform(hTransform); cmsDeleteTransform(hTransform);
} }
t3.set ();
//t3.set ();
// printf ("ICM TIME: %d\n", t3.etime(t1)); // printf ("ICM TIME: %d\n", t3.etime(t1));
} }

View File

@ -315,9 +315,12 @@ void StdImageSource::colorSpaceConversion (Imagefloat* im, ColorManagementParams
if (cmp.input!="(none)") { if (cmp.input!="(none)") {
lcmsMutex->lock (); lcmsMutex->lock ();
cmsHTRANSFORM hTransform = cmsCreateTransform (in, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), out, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), settings->colorimetricIntent, cmsFLAGS_NOOPTIMIZE); cmsHTRANSFORM hTransform = cmsCreateTransform (in, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), out, (FLOAT_SH(1)|COLORSPACE_SH(PT_RGB)|CHANNELS_SH(3)|BYTES_SH(4)|PLANAR_SH(1)), settings->colorimetricIntent,
cmsFLAGS_NOOPTIMIZE | cmsFLAGS_NOCACHE);
lcmsMutex->unlock (); lcmsMutex->unlock ();
cmsDoTransform (hTransform, im->data, im->data, im->planestride);
im->ExecCMSTransform(hTransform);
cmsDeleteTransform(hTransform); cmsDeleteTransform(hTransform);
} }
} }
@ -350,9 +353,11 @@ void StdImageSource::colorSpaceConversion16 (Image16* im, ColorManagementParams
if (cmp.input!="(none)") { if (cmp.input!="(none)") {
lcmsMutex->lock (); lcmsMutex->lock ();
cmsHTRANSFORM hTransform = cmsCreateTransform (in, TYPE_RGB_16_PLANAR, out, TYPE_RGB_16_PLANAR, settings->colorimetricIntent, 0); cmsHTRANSFORM hTransform = cmsCreateTransform (in, TYPE_RGB_16_PLANAR, out, TYPE_RGB_16_PLANAR, settings->colorimetricIntent, cmsFLAGS_NOCACHE);
lcmsMutex->unlock (); lcmsMutex->unlock ();
cmsDoTransform (hTransform, im->data, im->data, im->planestride);
im->ExecCMSTransform(hTransform);
cmsDeleteTransform(hTransform); cmsDeleteTransform(hTransform);
} }
} }