* Copyright (c) 2002-2003, Yannick Verschueren
* Copyright (c) 2003-2007, Francois-Olivier Devaux and Antonin Descampe
* Copyright (c) 2005, Herve Drolon, FreeImage Team
+ * Copyright (c) 2007, Jonathan Ballard <dzonatas@dzonux.net>
+ * Copyright (c) 2007, Callum Lerwick <seg@haxxed.com>
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* POSSIBILITY OF SUCH DAMAGE.
*/
-/*
- * NOTE:
- * This is a modified version of the openjpeg dwt.c file.
- * Average speed improvement compared to the original file (measured on
- * my own machine, a P4 running at 3.0 GHz):
- * 5x3 wavelets about 2 times faster
- * 9x7 wavelets about 3 times faster
- * for both, encoding and decoding.
- *
- * The better performance is caused by doing the 1-dimensional DWT
- * within a temporary buffer where the data can be accessed sequential
- * for both directions, horizontal and vertical. The 2d vertical DWT was
- * the major bottleneck in the former version.
- *
- * I have also removed the "Add Patrick" part because it is not longer
- * needed.
- *
- * 6/6/2005
- * -Ive (aka Reiner Wahler)
- * mail: ive@lilysoft.com
- */
+#ifdef __SSE__
+#include <xmmintrin.h>
+#endif
#include "opj_includes.h"
/** @defgroup DWT DWT - Implementation of a discrete wavelet transform */
/*@{*/
+#define WS(i) v->mem[(i)*2]
+#define WD(i) v->mem[(1+(i)*2)]
+
+/** @name Local data structures */
+/*@{*/
+
+typedef struct dwt_local {
+ int* mem;
+ int dn;
+ int sn;
+ int cas;
+} dwt_t;
+
+typedef union {
+ float f[4];
+} v4;
+
+typedef struct v4dwt_local {
+ v4* wavelet ;
+ int dn ;
+ int sn ;
+ int cas ;
+} v4dwt_t ;
+
+static const float dwt_alpha = 1.586134342f; // 12994
+static const float dwt_beta = 0.052980118f; // 434
+static const float dwt_gamma = -0.882911075f; // -7233
+static const float dwt_delta = -0.443506852f; // -3633
+
+static const float K = 1.230174105f; // 10078
+/* FIXME: What is this constant? */
+static const float c13318 = 1.625732422f;
+
+/*@}*/
+
+/**
+Virtual function type for wavelet transform in 1-D
+*/
+typedef void (*DWT1DFN)(dwt_t* v);
+
/** @name Local static functions */
/*@{*/
/**
Inverse lazy transform (horizontal)
*/
-static void dwt_interleave_h(int *a, int *b, int dn, int sn, int cas);
+static void dwt_interleave_h(dwt_t* h, int *a);
/**
Inverse lazy transform (vertical)
*/
-static void dwt_interleave_v(int *a, int *b, int dn, int sn, int x, int cas);
+static void dwt_interleave_v(dwt_t* v, int *a, int x);
/**
-Forward 5-3 wavelet tranform in 1-D
+Forward 5-3 wavelet transform in 1-D
*/
static void dwt_encode_1(int *a, int dn, int sn, int cas);
/**
-Inverse 5-3 wavelet tranform in 1-D
+Inverse 5-3 wavelet transform in 1-D
*/
-static void dwt_decode_1(int *a, int dn, int sn, int cas);
+static void dwt_decode_1(dwt_t *v);
/**
Forward 9-7 wavelet transform in 1-D
*/
static void dwt_encode_1_real(int *a, int dn, int sn, int cas);
/**
-Inverse 9-7 wavelet transform in 1-D
+Explicit calculation of the Quantization Stepsizes
*/
-static void dwt_decode_1_real(int *a, int dn, int sn, int cas);
+static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno_stepsize);
/**
-FIXME : comment ???
+Inverse wavelet transform in 2-D.
*/
-static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno_stepsize);
+static void dwt_decode_tile(opj_tcd_tilecomp_t* tilec, int i, DWT1DFN fn);
/*@}*/
/* <summary> */
/* Inverse lazy transform (horizontal). */
/* </summary> */
-static void dwt_interleave_h(int *a, int *b, int dn, int sn, int cas) {
- int i;
- int *ai = NULL;
- int *bi = NULL;
- ai = a;
- bi = b + cas;
- for (i = 0; i < sn; i++) {
- *bi = *ai;
- bi += 2;
- ai++;
+static void dwt_interleave_h(dwt_t* h, int *a) {
+ int *ai = a;
+ int *bi = h->mem + h->cas;
+ int i = h->sn;
+ while( i-- ) {
+ *bi = *(ai++);
+ bi += 2;
}
- ai = a + sn;
- bi = b + 1 - cas;
- for (i = 0; i < dn; i++) {
- *bi = *ai;
+ ai = a + h->sn;
+ bi = h->mem + 1 - h->cas;
+ i = h->dn ;
+ while( i-- ) {
+ *bi = *(ai++);
bi += 2;
- ai++;
}
}
/* <summary> */
/* Inverse lazy transform (vertical). */
/* </summary> */
-static void dwt_interleave_v(int *a, int *b, int dn, int sn, int x, int cas) {
- int i;
- int *ai = NULL;
- int *bi = NULL;
- ai = a;
- bi = b + cas;
- for (i = 0; i < sn; i++) {
+static void dwt_interleave_v(dwt_t* v, int *a, int x) {
+ int *ai = a;
+ int *bi = v->mem + v->cas;
+ int i = v->sn;
+ while( i-- ) {
*bi = *ai;
bi += 2;
ai += x;
}
- ai = a + (sn * x);
- bi = b + 1 - cas;
- for (i = 0; i < dn; i++) {
+ ai = a + (v->sn * x);
+ bi = v->mem + 1 - v->cas;
+ i = v->dn ;
+ while( i-- ) {
*bi = *ai;
bi += 2;
ai += x;
/* <summary> */
-/* Forward 5-3 wavelet tranform in 1-D. */
+/* Forward 5-3 wavelet transform in 1-D. */
/* </summary> */
static void dwt_encode_1(int *a, int dn, int sn, int cas) {
int i;
}
/* <summary> */
-/* Inverse 5-3 wavelet tranform in 1-D. */
+/* Inverse 5-3 wavelet transform in 1-D. */
/* </summary> */
-static void dwt_decode_1(int *a, int dn, int sn, int cas) {
+static void dwt_decode_1_(int *a, int dn, int sn, int cas) {
int i;
if (!cas) {
}
}
+/* <summary> */
+/* Inverse 5-3 wavelet transform in 1-D. */
+/* </summary> */
+static void dwt_decode_1(dwt_t *v) {
+ dwt_decode_1_(v->mem, v->dn, v->sn, v->cas);
+}
+
/* <summary> */
/* Forward 9-7 wavelet transform in 1-D. */
/* </summary> */
}
}
-/* <summary> */
-/* Inverse 9-7 wavelet transform in 1-D. */
-/* </summary> */
-static void dwt_decode_1_real(int *a, int dn, int sn, int cas) {
- int i;
- if (!cas) {
- if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */
- for (i = 0; i < sn; i++)
- S(i) = fix_mul(S(i), 10078); /* 10076 */
- for (i = 0; i < dn; i++)
- D(i) = fix_mul(D(i), 13318); /* 13320 */
- for (i = 0; i < sn; i++)
- S(i) -= fix_mul(D_(i - 1) + D_(i), 3633);
- for (i = 0; i < dn; i++)
- D(i) -= fix_mul(S_(i) + S_(i + 1), 7233);
- for (i = 0; i < sn; i++)
- S(i) += fix_mul(D_(i - 1) + D_(i), 434);
- for (i = 0; i < dn; i++)
- D(i) += fix_mul(S_(i) + S_(i + 1), 12994); /* 12993 */
- }
- } else {
- if ((sn > 0) || (dn > 1)) { /* NEW : CASE ONE ELEMENT */
- for (i = 0; i < sn; i++)
- D(i) = fix_mul(D(i), 10078); /* 10076 */
- for (i = 0; i < dn; i++)
- S(i) = fix_mul(S(i), 13318); /* 13320 */
- for (i = 0; i < sn; i++)
- D(i) -= fix_mul(SS_(i) + SS_(i + 1), 3633);
- for (i = 0; i < dn; i++)
- S(i) -= fix_mul(DD_(i) + DD_(i - 1), 7233);
- for (i = 0; i < sn; i++)
- D(i) += fix_mul(SS_(i) + SS_(i + 1), 434);
- for (i = 0; i < dn; i++)
- S(i) += fix_mul(DD_(i) + DD_(i - 1), 12994); /* 12993 */
- }
- }
-}
-
static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno_stepsize) {
int p, n;
p = int_floorlog2(stepsize) - 13;
*/
/* <summary> */
-/* Forward 5-3 wavelet tranform in 2-D. */
+/* Forward 5-3 wavelet transform in 2-D. */
/* </summary> */
void dwt_encode(opj_tcd_tilecomp_t * tilec) {
int i, j, k;
for (i = 0; i < l; i++) {
int rw; /* width of the resolution level computed */
- int rh; /* heigth of the resolution level computed */
+ int rh; /* height of the resolution level computed */
int rw1; /* width of the resolution level once lower than computed one */
int rh1; /* height of the resolution level once lower than computed one */
int cas_col; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
/* <summary> */
-/* Inverse 5-3 wavelet tranform in 2-D. */
+/* Inverse 5-3 wavelet transform in 2-D. */
/* </summary> */
-void dwt_decode(opj_tcd_tilecomp_t * tilec, int stop) {
- int i, j, k;
- int *a = NULL;
- int *aj = NULL;
- int *bj = NULL;
- int w, l;
-
- w = tilec->x1-tilec->x0;
- l = tilec->numresolutions-1;
- a = tilec->data;
-
- for (i = l - 1; i >= stop; i--) {
- int rw; /* width of the resolution level computed */
- int rh; /* heigth of the resolution level computed */
- int rw1; /* width of the resolution level once lower than computed one */
- int rh1; /* height of the resolution level once lower than computed one */
- int cas_col; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
- int cas_row; /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering */
- int dn, sn;
-
- rw = tilec->resolutions[l - i].x1 - tilec->resolutions[l - i].x0;
- rh = tilec->resolutions[l - i].y1 - tilec->resolutions[l - i].y0;
- rw1= tilec->resolutions[l - i - 1].x1 - tilec->resolutions[l - i - 1].x0;
- rh1= tilec->resolutions[l - i - 1].y1 - tilec->resolutions[l - i - 1].y0;
-
- cas_row = tilec->resolutions[l - i].x0 % 2;
- cas_col = tilec->resolutions[l - i].y0 % 2;
-
- sn = rw1;
- dn = rw - rw1;
- bj = (int*)opj_malloc(rw * sizeof(int));
- for (j = 0; j < rh; j++) {
- aj = a + j*w;
- dwt_interleave_h(aj, bj, dn, sn, cas_row);
- dwt_decode_1(bj, dn, sn, cas_row);
- for (k = 0; k < rw; k++) aj[k] = bj[k];
- }
- opj_free(bj);
-
- sn = rh1;
- dn = rh - rh1;
- bj = (int*)opj_malloc(rh * sizeof(int));
- for (j = 0; j < rw; j++) {
- aj = a + j;
- dwt_interleave_v(aj, bj, dn, sn, w, cas_col);
- dwt_decode_1(bj, dn, sn, cas_col);
- for (k = 0; k < rh; k++) aj[k * w] = bj[k];
- }
- opj_free(bj);
- }
+void dwt_decode(opj_tcd_tilecomp_t* tilec, int numres) {
+ dwt_decode_tile(tilec, numres, &dwt_decode_1);
}
for (i = 0; i < l; i++) {
int rw; /* width of the resolution level computed */
- int rh; /* heigth of the resolution level computed */
+ int rh; /* height of the resolution level computed */
int rw1; /* width of the resolution level once lower than computed one */
int rh1; /* height of the resolution level once lower than computed one */
int cas_col; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
}
-/* <summary> */
-/* Inverse 9-7 wavelet transform in 2-D. */
-/* </summary> */
-void dwt_decode_real(opj_tcd_tilecomp_t * tilec, int stop) {
- int i, j, k;
- int *a = NULL;
- int *aj = NULL;
- int *bj = NULL;
- int w, l;
-
- w = tilec->x1-tilec->x0;
- l = tilec->numresolutions-1;
- a = tilec->data;
-
- for (i = l-1; i >= stop; i--) {
- int rw; /* width of the resolution level computed */
- int rh; /* heigth of the resolution level computed */
- int rw1; /* width of the resolution level once lower than computed one */
- int rh1; /* height of the resolution level once lower than computed one */
- int cas_col; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
- int cas_row; /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering */
- int dn, sn;
-
- rw = tilec->resolutions[l - i].x1 - tilec->resolutions[l - i].x0;
- rh = tilec->resolutions[l - i].y1 - tilec->resolutions[l - i].y0;
- rw1= tilec->resolutions[l - i - 1].x1 - tilec->resolutions[l - i - 1].x0;
- rh1= tilec->resolutions[l - i - 1].y1 - tilec->resolutions[l - i - 1].y0;
-
- cas_col = tilec->resolutions[l - i].x0 % 2; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
- cas_row = tilec->resolutions[l - i].y0 % 2; /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering */
-
- sn = rw1;
- dn = rw-rw1;
- bj = (int*)opj_malloc(rw * sizeof(int));
- for (j = 0; j < rh; j++) {
- aj = a + j * w;
- dwt_interleave_h(aj, bj, dn, sn, cas_col);
- dwt_decode_1_real(bj, dn, sn, cas_col);
- for (k = 0; k < rw; k++) aj[k] = bj[k];
- }
- opj_free(bj);
-
- sn = rh1;
- dn = rh-rh1;
- bj = (int*)opj_malloc(rh * sizeof(int));
- for (j = 0; j < rw; j++) {
- aj = a + j;
- dwt_interleave_v(aj, bj, dn, sn, w, cas_row);
- dwt_decode_1_real(bj, dn, sn, cas_row);
- for (k = 0; k < rh; k++) aj[k * w] = bj[k];
- }
- opj_free(bj);
- }
-}
-
-
/* <summary> */
/* Get gain of 9-7 wavelet transform. */
/* </summary> */
dwt_encode_stepsize((int) floor(stepsize * 8192.0), prec + gain, &tccp->stepsizes[bandno]);
}
}
+
+
+/* <summary> */
+/* Determine maximum computed resolution level for inverse wavelet transform */
+/* </summary> */
+static int dwt_decode_max_resolution(opj_tcd_resolution_t* restrict r, int i) {
+ int mr = 1;
+ int w;
+ while( --i ) {
+ r++;
+ if( mr < ( w = r->x1 - r->x0 ) )
+ mr = w ;
+ if( mr < ( w = r->y1 - r->y0 ) )
+ mr = w ;
+ }
+ return mr ;
+}
+
+
+/* <summary> */
+/* Inverse wavelet transform in 2-D. */
+/* </summary> */
+static void dwt_decode_tile(opj_tcd_tilecomp_t* tilec, int numres, DWT1DFN dwt_1D) {
+ dwt_t h;
+ dwt_t v;
+
+ opj_tcd_resolution_t* tr = tilec->resolutions;
+
+ int rw = tr->x1 - tr->x0; /* width of the resolution level computed */
+ int rh = tr->y1 - tr->y0; /* height of the resolution level computed */
+
+ int w = tilec->x1 - tilec->x0;
+
+ h.mem = opj_aligned_malloc(dwt_decode_max_resolution(tr, numres) * sizeof(int));
+ v.mem = h.mem;
+
+ while( --numres) {
+ int * restrict tiledp = tilec->data;
+ int j;
+
+ ++tr;
+ h.sn = rw;
+ v.sn = rh;
+
+ rw = tr->x1 - tr->x0;
+ rh = tr->y1 - tr->y0;
+
+ h.dn = rw - h.sn;
+ h.cas = tr->x0 % 2;
+
+ for(j = 0; j < rh; ++j) {
+ dwt_interleave_h(&h, &tiledp[j*w]);
+ (dwt_1D)(&h);
+ memcpy(&tiledp[j*w], h.mem, rw * sizeof(int));
+ }
+
+ v.dn = rh - v.sn;
+ v.cas = tr->y0 % 2;
+
+ for(j = 0; j < rw; ++j){
+ int k;
+ dwt_interleave_v(&v, &tiledp[j], w);
+ (dwt_1D)(&v);
+ for(k = 0; k < rh; ++k) {
+ tiledp[k * w + j] = v.mem[k];
+ }
+ }
+ }
+ opj_aligned_free(h.mem);
+}
+
+static void v4dwt_interleave_h(v4dwt_t* restrict w, float* restrict a, int x, int size){
+ float* restrict bi = (float*) (w->wavelet + w->cas);
+ int count = w->sn;
+ int i, k;
+ for(k = 0; k < 2; ++k){
+ if (count + 3 * x < size && ((size_t) a & 0x0f) == 0 && ((size_t) bi & 0x0f) == 0 && (x & 0x0f) == 0) {
+ /* Fast code path */
+ for(i = 0; i < count; ++i){
+ int j = i;
+ bi[i*8 ] = a[j];
+ j += x;
+ bi[i*8 + 1] = a[j];
+ j += x;
+ bi[i*8 + 2] = a[j];
+ j += x;
+ bi[i*8 + 3] = a[j];
+ }
+ } else {
+ /* Slow code path */
+ for(i = 0; i < count; ++i){
+ int j = i;
+ bi[i*8 ] = a[j];
+ j += x;
+ if(j > size) continue;
+ bi[i*8 + 1] = a[j];
+ j += x;
+ if(j > size) continue;
+ bi[i*8 + 2] = a[j];
+ j += x;
+ if(j > size) continue;
+ bi[i*8 + 3] = a[j];
+ }
+ }
+ bi = (float*) (w->wavelet + 1 - w->cas);
+ a += w->sn;
+ size -= w->sn;
+ count = w->dn;
+ }
+}
+
+static void v4dwt_interleave_v(v4dwt_t* restrict v , float* restrict a , int x){
+ v4* restrict bi = v->wavelet + v->cas;
+ int i;
+ for(i = 0; i < v->sn; ++i){
+ memcpy(&bi[i*2], &a[i*x], 4 * sizeof(float));
+ }
+ a += v->sn * x;
+ bi = v->wavelet + 1 - v->cas;
+ for(i = 0; i < v->dn; ++i){
+ memcpy(&bi[i*2], &a[i*x], 4 * sizeof(float));
+ }
+}
+
+#ifdef __SSE__
+
+static void v4dwt_decode_step1_sse(v4* w, int count, const __m128 c){
+ __m128* restrict vw = (__m128*) w;
+ int i;
+ /* 4x unrolled loop */
+ for(i = 0; i < count >> 2; ++i){
+ *vw = _mm_mul_ps(*vw, c);
+ vw += 2;
+ *vw = _mm_mul_ps(*vw, c);
+ vw += 2;
+ *vw = _mm_mul_ps(*vw, c);
+ vw += 2;
+ *vw = _mm_mul_ps(*vw, c);
+ vw += 2;
+ }
+ count &= 3;
+ for(i = 0; i < count; ++i){
+ *vw = _mm_mul_ps(*vw, c);
+ vw += 2;
+ }
+}
+
+static void v4dwt_decode_step2_sse(v4* l, v4* w, int k, int m, __m128 c){
+ __m128* restrict vl = (__m128*) l;
+ __m128* restrict vw = (__m128*) w;
+ int i;
+ __m128 tmp1, tmp2, tmp3;
+ tmp1 = vl[0];
+ for(i = 0; i < m; ++i){
+ tmp2 = vw[-1];
+ tmp3 = vw[ 0];
+ vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c));
+ tmp1 = tmp3;
+ vw += 2;
+ }
+ vl = vw - 2;
+ if(m >= k){
+ return;
+ }
+ c = _mm_add_ps(c, c);
+ c = _mm_mul_ps(c, vl[0]);
+ for(; m < k; ++m){
+ __m128 tmp = vw[-1];
+ vw[-1] = _mm_add_ps(tmp, c);
+ vw += 2;
+ }
+}
+
+#else
+
+static void v4dwt_decode_step1(v4* w, int count, const float c){
+ float* restrict fw = (float*) w;
+ int i;
+ for(i = 0; i < count; ++i){
+ float tmp1 = fw[i*8 ];
+ float tmp2 = fw[i*8 + 1];
+ float tmp3 = fw[i*8 + 2];
+ float tmp4 = fw[i*8 + 3];
+ fw[i*8 ] = tmp1 * c;
+ fw[i*8 + 1] = tmp2 * c;
+ fw[i*8 + 2] = tmp3 * c;
+ fw[i*8 + 3] = tmp4 * c;
+ }
+}
+
+static void v4dwt_decode_step2(v4* l, v4* w, int k, int m, float c){
+ float* restrict fl = (float*) l;
+ float* restrict fw = (float*) w;
+ int i;
+ for(i = 0; i < m; ++i){
+ float tmp1_1 = fl[0];
+ float tmp1_2 = fl[1];
+ float tmp1_3 = fl[2];
+ float tmp1_4 = fl[3];
+ float tmp2_1 = fw[-4];
+ float tmp2_2 = fw[-3];
+ float tmp2_3 = fw[-2];
+ float tmp2_4 = fw[-1];
+ float tmp3_1 = fw[0];
+ float tmp3_2 = fw[1];
+ float tmp3_3 = fw[2];
+ float tmp3_4 = fw[3];
+ fw[-4] = tmp2_1 + ((tmp1_1 + tmp3_1) * c);
+ fw[-3] = tmp2_2 + ((tmp1_2 + tmp3_2) * c);
+ fw[-2] = tmp2_3 + ((tmp1_3 + tmp3_3) * c);
+ fw[-1] = tmp2_4 + ((tmp1_4 + tmp3_4) * c);
+ fl = fw;
+ fw += 8;
+ }
+ if(m < k){
+ float c1;
+ float c2;
+ float c3;
+ float c4;
+ c += c;
+ c1 = fl[0] * c;
+ c2 = fl[1] * c;
+ c3 = fl[2] * c;
+ c4 = fl[3] * c;
+ for(; m < k; ++m){
+ float tmp1 = fw[-4];
+ float tmp2 = fw[-3];
+ float tmp3 = fw[-2];
+ float tmp4 = fw[-1];
+ fw[-4] = tmp1 + c1;
+ fw[-3] = tmp2 + c2;
+ fw[-2] = tmp3 + c3;
+ fw[-1] = tmp4 + c4;
+ fw += 8;
+ }
+ }
+}
+
+#endif
+
+/* <summary> */
+/* Inverse 9-7 wavelet transform in 1-D. */
+/* </summary> */
+static void v4dwt_decode(v4dwt_t* restrict dwt){
+ int a, b;
+ if(dwt->cas == 0) {
+ if(!((dwt->dn > 0) || (dwt->sn > 1))){
+ return;
+ }
+ a = 0;
+ b = 1;
+ }else{
+ if(!((dwt->sn > 0) || (dwt->dn > 1))) {
+ return;
+ }
+ a = 1;
+ b = 0;
+ }
+#ifdef __SSE__
+ v4dwt_decode_step1_sse(dwt->wavelet+a, dwt->sn, _mm_set1_ps(K));
+ v4dwt_decode_step1_sse(dwt->wavelet+b, dwt->dn, _mm_set1_ps(c13318));
+ v4dwt_decode_step2_sse(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), _mm_set1_ps(dwt_delta));
+ v4dwt_decode_step2_sse(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), _mm_set1_ps(dwt_gamma));
+ v4dwt_decode_step2_sse(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), _mm_set1_ps(dwt_beta));
+ v4dwt_decode_step2_sse(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), _mm_set1_ps(dwt_alpha));
+#else
+ v4dwt_decode_step1(dwt->wavelet+a, dwt->sn, K);
+ v4dwt_decode_step1(dwt->wavelet+b, dwt->dn, c13318);
+ v4dwt_decode_step2(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), dwt_delta);
+ v4dwt_decode_step2(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), dwt_gamma);
+ v4dwt_decode_step2(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), dwt_beta);
+ v4dwt_decode_step2(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), dwt_alpha);
+#endif
+}
+
+/* <summary> */
+/* Inverse 9-7 wavelet transform in 2-D. */
+/* </summary> */
+void dwt_decode_real(opj_tcd_tilecomp_t* restrict tilec, int numres){
+ v4dwt_t h;
+ v4dwt_t v;
+
+ opj_tcd_resolution_t* res = tilec->resolutions;
+
+ int rw = res->x1 - res->x0; /* width of the resolution level computed */
+ int rh = res->y1 - res->y0; /* height of the resolution level computed */
+
+ int w = tilec->x1 - tilec->x0;
+
+ h.wavelet = (v4*) opj_aligned_malloc((dwt_decode_max_resolution(res, numres)+5) * sizeof(v4));
+ v.wavelet = h.wavelet;
+
+ while( --numres) {
+ float * restrict aj = (float*) tilec->data;
+ int bufsize = (tilec->x1 - tilec->x0) * (tilec->y1 - tilec->y0);
+ int j;
+
+ h.sn = rw;
+ v.sn = rh;
+
+ ++res;
+
+ rw = res->x1 - res->x0; /* width of the resolution level computed */
+ rh = res->y1 - res->y0; /* height of the resolution level computed */
+
+ h.dn = rw - h.sn;
+ h.cas = res->x0 % 2;
+
+ for(j = rh; j > 3; j -= 4){
+ int k;
+ v4dwt_interleave_h(&h, aj, w, bufsize);
+ v4dwt_decode(&h);
+ for(k = rw; --k >= 0;){
+ aj[k ] = h.wavelet[k].f[0];
+ aj[k+w ] = h.wavelet[k].f[1];
+ aj[k+w*2] = h.wavelet[k].f[2];
+ aj[k+w*3] = h.wavelet[k].f[3];
+ }
+ aj += w*4;
+ bufsize -= w*4;
+ }
+ if (rh & 0x03) {
+ int k;
+ j = rh & 0x03;
+ v4dwt_interleave_h(&h, aj, w, bufsize);
+ v4dwt_decode(&h);
+ for(k = rw; --k >= 0;){
+ switch(j) {
+ case 3: aj[k+w*2] = h.wavelet[k].f[2];
+ case 2: aj[k+w ] = h.wavelet[k].f[1];
+ case 1: aj[k ] = h.wavelet[k].f[0];
+ }
+ }
+ }
+
+ v.dn = rh - v.sn;
+ v.cas = res->y0 % 2;
+
+ aj = (float*) tilec->data;
+ for(j = rw; j > 3; j -= 4){
+ int k;
+ v4dwt_interleave_v(&v, aj, w);
+ v4dwt_decode(&v);
+ for(k = 0; k < rh; ++k){
+ memcpy(&aj[k*w], &v.wavelet[k], 4 * sizeof(float));
+ }
+ aj += 4;
+ }
+ if (rw & 0x03){
+ int k;
+ j = rw & 0x03;
+ v4dwt_interleave_v(&v, aj, w);
+ v4dwt_decode(&v);
+ for(k = 0; k < rh; ++k){
+ memcpy(&aj[k*w], &v.wavelet[k], j * sizeof(float));
+ }
+ }
+ }
+
+ opj_aligned_free(h.wavelet);
+}
+
projects / openjpeg.git / blobdiff