Optimization for Output-Resize Method Lanczos, Issue 2248
This commit is contained in:
Ingo
@@ -19,11 +19,11 @@
|
||||
|
||||
#include "improcfun.h"
|
||||
#include "rt_math.h"
|
||||
|
||||
#ifdef _OPENMP
|
||||
#include <omp.h>
|
||||
#endif
|
||||
|
||||
#include "sleef.c"
|
||||
#include "opthelper.h"
|
||||
//#define PROFILE
|
||||
|
||||
#ifdef PROFILE
|
||||
@@ -40,7 +40,7 @@ static inline float Lanc(float x, float a)
|
||||
return 0.0f;
|
||||
else {
|
||||
x = static_cast<float>(M_PI) * x;
|
||||
return sinf(x) * sinf(x / a) / (x * x / a);
|
||||
return a * xsinf(x) * xsinf(x / a) / (x * x);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -51,6 +51,8 @@ static void Lanczos(const Image16* src, Image16* dst, float scale)
|
||||
const float sc = min(scale, 1.0f);
|
||||
const int support = static_cast<int>(2.0f * a / sc) + 1;
|
||||
|
||||
#pragma omp parallel
|
||||
{
|
||||
// storage for precomputed parameters for horisontal interpolation
|
||||
float * wwh = new float[support * dst->width];
|
||||
int * jj0 = new int[dst->width];
|
||||
@@ -90,9 +92,8 @@ static void Lanczos(const Image16* src, Image16* dst, float scale)
|
||||
w[k] /= ws;
|
||||
}
|
||||
}
|
||||
|
||||
// Phase 2: do actual interpolation
|
||||
|
||||
#pragma omp for
|
||||
for (int i = 0; i < dst->height; i++) {
|
||||
|
||||
// y coord of the center of pixel on src image
|
||||
@@ -166,9 +167,174 @@ static void Lanczos(const Image16* src, Image16* dst, float scale)
|
||||
delete[] lg;
|
||||
delete[] lb;
|
||||
}
|
||||
}
|
||||
|
||||
// this function is implemented for future use and not tested yet
|
||||
SSEFUNCTION static void Lanczos(const LabImage* src, LabImage* dst, float scale)
|
||||
{
|
||||
const float delta = 1.0f / scale;
|
||||
const float a = 3.0f;
|
||||
const float sc = min(scale, 1.0f);
|
||||
const int support = static_cast<int>(2.0f * a / sc) + 1;
|
||||
|
||||
#pragma omp parallel
|
||||
{
|
||||
// storage for precomputed parameters for horisontal interpolation
|
||||
float * wwh = new float[support * dst->W];
|
||||
int * jj0 = new int[dst->W];
|
||||
int * jj1 = new int[dst->W];
|
||||
|
||||
// temporal storage for vertically-interpolated row of pixels
|
||||
float * lL = new float[src->W];
|
||||
float * la = new float[src->W];
|
||||
float * lb = new float[src->W];
|
||||
|
||||
// Phase 1: precompute coefficients for horisontal interpolation
|
||||
|
||||
for (int j = 0; j < dst->W; j++) {
|
||||
|
||||
// x coord of the center of pixel on src image
|
||||
float x0 = (static_cast<float>(j) + 0.5f) * delta - 0.5f;
|
||||
|
||||
// weights for interpolation in horisontal direction
|
||||
float * w = wwh + j * support;
|
||||
|
||||
// sum of weights used for normalization
|
||||
float ws = 0.0f;
|
||||
|
||||
jj0[j] = max(0, static_cast<int>(floorf(x0 - a / sc)) + 1);
|
||||
jj1[j] = min(src->W, static_cast<int>(floorf(x0 + a / sc)) + 1);
|
||||
|
||||
// calculate weights
|
||||
for (int jj = jj0[j]; jj < jj1[j]; jj++) {
|
||||
int k = jj - jj0[j];
|
||||
float z = sc * (x0 - static_cast<float>(jj));
|
||||
w[k] = Lanc(z, a);
|
||||
ws += w[k];
|
||||
}
|
||||
|
||||
// normalize weights
|
||||
for (int k = 0; k < support; k++) {
|
||||
w[k] /= ws;
|
||||
}
|
||||
}
|
||||
// Phase 2: do actual interpolation
|
||||
#pragma omp for
|
||||
for (int i = 0; i < dst->H; i++) {
|
||||
|
||||
// y coord of the center of pixel on src image
|
||||
float y0 = (static_cast<float>(i) + 0.5f) * delta - 0.5f;
|
||||
|
||||
// weights for interpolation in y direction
|
||||
float w[support];
|
||||
|
||||
// sum of weights used for normalization
|
||||
float ws= 0.0f;
|
||||
|
||||
int ii0 = max(0, static_cast<int>(floorf(y0 - a / sc)) + 1);
|
||||
int ii1 = min(src->H, static_cast<int>(floorf(y0 + a / sc)) + 1);
|
||||
|
||||
// calculate weights for vertical interpolation
|
||||
for (int ii = ii0; ii < ii1; ii++) {
|
||||
int k = ii - ii0;
|
||||
float z = sc * (y0 - static_cast<float>(ii));
|
||||
w[k] = Lanc(z, a);
|
||||
ws += w[k];
|
||||
}
|
||||
|
||||
// normalize weights
|
||||
for (int k = 0; k < support; k++) {
|
||||
w[k] /= ws;
|
||||
}
|
||||
|
||||
// Do vertical interpolation. Store results.
|
||||
#ifdef __SSE299__ // actually disabled. To enable replace __SSE299__ by __SSE2__
|
||||
__m128 Lv,av,bv,wkv;
|
||||
int j;
|
||||
for (j = 0; j < src->W-3; j+=4) {
|
||||
Lv = _mm_setzero_ps();
|
||||
av = _mm_setzero_ps();
|
||||
bv = _mm_setzero_ps();
|
||||
|
||||
for (int ii = ii0; ii < ii1; ii++) {
|
||||
int k = ii - ii0;
|
||||
wkv = LVFU(w[k]);
|
||||
Lv += wkv * LVFU(src->L[ii][j]);
|
||||
av += wkv * LVFU(src->a[ii][j]);
|
||||
bv += wkv * LVFU(src->b[ii][j]);
|
||||
}
|
||||
|
||||
_mm_storeu_ps(&lL[j],Lv);
|
||||
_mm_storeu_ps(&la[j],av);
|
||||
_mm_storeu_ps(&lb[j],bv);
|
||||
}
|
||||
for (; j < src->W; j++) {
|
||||
|
||||
float L = 0.0f, a = 0.0f, b = 0.0f;
|
||||
|
||||
for (int ii = ii0; ii < ii1; ii++) {
|
||||
int k = ii - ii0;
|
||||
|
||||
L += w[k] * src->L[ii][j];
|
||||
a += w[k] * src->a[ii][j];
|
||||
b += w[k] * src->b[ii][j];
|
||||
}
|
||||
|
||||
lL[j] = L;
|
||||
la[j] = a;
|
||||
lb[j] = b;
|
||||
}
|
||||
|
||||
#else
|
||||
for (int j = 0; j < src->W; j++) {
|
||||
|
||||
float L = 0.0f, a = 0.0f, b = 0.0f;
|
||||
|
||||
for (int ii = ii0; ii < ii1; ii++) {
|
||||
int k = ii - ii0;
|
||||
|
||||
L += w[k] * src->L[ii][j];
|
||||
a += w[k] * src->a[ii][j];
|
||||
b += w[k] * src->b[ii][j];
|
||||
}
|
||||
|
||||
lL[j] = L;
|
||||
la[j] = a;
|
||||
lb[j] = b;
|
||||
}
|
||||
#endif
|
||||
// Do horizontal interpolation
|
||||
for(int j = 0; j < dst->W; j++) {
|
||||
|
||||
float * wh = wwh + support * j;
|
||||
|
||||
float L = 0.0f, a = 0.0f, b = 0.0f;
|
||||
|
||||
for (int jj = jj0[j]; jj < jj1[j]; jj++) {
|
||||
int k = jj - jj0[j];
|
||||
|
||||
L += wh[k] * lL[jj];
|
||||
a += wh[k] * la[jj];
|
||||
b += wh[k] * lb[jj];
|
||||
}
|
||||
|
||||
dst->L[i][j] = L;
|
||||
dst->a[i][j] = a;
|
||||
dst->b[i][j] = b;
|
||||
}
|
||||
}
|
||||
|
||||
delete[] wwh;
|
||||
delete[] jj0;
|
||||
delete[] jj1;
|
||||
delete[] lL;
|
||||
delete[] la;
|
||||
delete[] lb;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
void ImProcFunctions::resize (Image16* src, Image16* dst, float dScale) {
|
||||
|
||||
#ifdef PROFILE
|
||||
time_t t1 = clock();
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user