X-Git-Url: https://main.carlh.net/gitweb/?a=blobdiff_plain;f=libopenjpeg%2Fdwt.c;h=94586eb53046d304693f8c2e0b413d4bee851528;hb=8363a6ab1e031bb4b2e40a92e56efd40fdab1aa1;hp=161512aad0dab0092f94a589733c5d1afe998cbd;hpb=4bb067642f155b3cc054f77e1f08b171491a8354;p=openjpeg.git diff --git a/libopenjpeg/dwt.c b/libopenjpeg/dwt.c index 161512aa..94586eb5 100644 --- a/libopenjpeg/dwt.c +++ b/libopenjpeg/dwt.c @@ -5,6 +5,8 @@ * 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 + * Copyright (c) 2007, Callum Lerwick * All rights reserved. * * Redistribution and use in source and binary forms, with or without @@ -29,6 +31,9 @@ * POSSIBILITY OF SUCH DAMAGE. */ +#ifdef __SSE__ +#include +#endif #include "opj_includes.h" @@ -41,12 +46,31 @@ /** @name Local data structures */ /*@{*/ -typedef struct dwt_local{ - int * mem ; - int dn ; - int sn ; - int cas ; - } dwt_t ; +typedef struct dwt_local { + OPJ_INT32* mem; + OPJ_INT32 dn; + OPJ_INT32 sn; + OPJ_INT32 cas; +} dwt_t; + +typedef union { + OPJ_FLOAT32 f[4]; +} v4; + +typedef struct v4dwt_local { + v4* wavelet ; + OPJ_INT32 dn ; + OPJ_INT32 sn ; + OPJ_INT32 cas ; +} v4dwt_t ; + +static const OPJ_FLOAT32 opj_dwt_alpha = 1.586134342f; /* 12994 */ +static const OPJ_FLOAT32 opj_dwt_beta = 0.052980118f; /* 434 */ +static const OPJ_FLOAT32 opj_dwt_gamma = -0.882911075f; /* -7233 */ +static const OPJ_FLOAT32 opj_dwt_delta = -0.443506852f; /* -3633 */ + +static const OPJ_FLOAT32 opj_K = 1.230174105f; /* 10078 */ +static const OPJ_FLOAT32 opj_c13318 = 1.625732422f; /*@}*/ @@ -61,43 +85,64 @@ typedef void (*DWT1DFN)(dwt_t* v); /** Forward lazy transform (horizontal) */ -static void dwt_deinterleave_h(int *a, int *b, int dn, int sn, int cas); +static void opj_dwt_deinterleave_h(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas); /** Forward lazy transform (vertical) */ -static void dwt_deinterleave_v(int *a, int *b, int dn, int sn, int x, int cas); +static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 x, OPJ_INT32 cas); /** Inverse lazy transform (horizontal) */ -static void dwt_interleave_h(dwt_t* h, int *a); +static void opj_dwt_interleave_h(dwt_t* h, OPJ_INT32 *a); /** Inverse lazy transform (vertical) */ -static void dwt_interleave_v(dwt_t* v, int *a, int x); +static void opj_dwt_interleave_v(dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x); /** Forward 5-3 wavelet transform in 1-D */ -static void dwt_encode_1(int *a, int dn, int sn, int cas); +static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas); /** Inverse 5-3 wavelet transform in 1-D */ -static void dwt_decode_1(dwt_t *v); +static void opj_dwt_decode_1(dwt_t *v); +static void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas); /** Forward 9-7 wavelet transform in 1-D */ -static void dwt_encode_1_real(int *a, int dn, int sn, int cas); +static void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas); /** -Inverse 9-7 wavelet transform in 1-D +Explicit calculation of the Quantization Stepsizes */ -static void dwt_decode_1_real(dwt_t *v); -/** -FIXME : comment ??? -*/ -static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno_stepsize); +static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps, opj_stepsize_t *bandno_stepsize); /** Inverse wavelet transform in 2-D. */ -static void dwt_decode_tile(opj_tcd_tilecomp_t * tilec, int stop , DWT1DFN fn); +static opj_bool opj_dwt_decode_tile(opj_tcd_tilecomp_v2_t* tilec, OPJ_UINT32 i, DWT1DFN fn); + +static opj_bool opj_dwt_encode_procedure( opj_tcd_tilecomp_v2_t * tilec, + void (*p_function)(OPJ_INT32 *, OPJ_INT32,OPJ_INT32,OPJ_INT32) ); + +static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_v2_t* restrict r, OPJ_UINT32 i); + +/* */ +/* Inverse 9-7 wavelet transform in 1-D. */ +/* */ +static void opj_v4dwt_decode(v4dwt_t* restrict dwt); + +static void opj_v4dwt_interleave_h(v4dwt_t* restrict w, OPJ_FLOAT32* restrict a, OPJ_INT32 x, OPJ_INT32 size); + +static void opj_v4dwt_interleave_v(v4dwt_t* restrict v , OPJ_FLOAT32* restrict a , OPJ_INT32 x, OPJ_INT32 nb_elts_read); + +#ifdef __SSE__ +static void opj_v4dwt_decode_step1_sse(v4* w, OPJ_INT32 count, const __m128 c); + +static void opj_v4dwt_decode_step2_sse(v4* l, v4* w, OPJ_INT32 k, OPJ_INT32 m, __m128 c); +#endif + +static void opj_v4dwt_decode_step1(v4* w, OPJ_INT32 count, const OPJ_FLOAT32 c); +static void opj_v4dwt_decode_step2(v4* l, v4* w, OPJ_INT32 k, OPJ_INT32 m, OPJ_FLOAT32 c); + /*@}*/ @@ -114,7 +159,7 @@ static void dwt_decode_tile(opj_tcd_tilecomp_t * tilec, int stop , DWT1DFN fn); /* */ /* This table contains the norms of the 5-3 wavelets for different bands. */ /* */ -static const double dwt_norms[4][10] = { +static const OPJ_FLOAT64 opj_dwt_norms[4][10] = { {1.000, 1.500, 2.750, 5.375, 10.68, 21.34, 42.67, 85.33, 170.7, 341.3}, {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9}, {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9}, @@ -124,7 +169,7 @@ static const double dwt_norms[4][10] = { /* */ /* This table contains the norms of the 9-7 wavelets for different bands. */ /* */ -static const double dwt_norms_real[4][10] = { +static const OPJ_FLOAT64 opj_dwt_norms_real[4][10] = { {1.000, 1.965, 4.177, 8.403, 16.90, 33.84, 67.69, 135.3, 270.6, 540.9}, {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0}, {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0}, @@ -140,28 +185,57 @@ static const double dwt_norms_real[4][10] = { /* */ /* Forward lazy transform (horizontal). */ /* */ -static void dwt_deinterleave_h(int *a, int *b, int dn, int sn, int cas) { - int i; - for (i=0; i */ /* Forward lazy transform (vertical). */ /* */ -static void dwt_deinterleave_v(int *a, int *b, int dn, int sn, int x, int cas) { - int i; - for (i=0; i */ /* Inverse lazy transform (horizontal). */ /* */ -static void dwt_interleave_h(dwt_t* h, int *a) { - int *ai = a; - int *bi = h->mem + h->cas; - int i = h->sn; +void opj_dwt_interleave_h(dwt_t* h, OPJ_INT32 *a) { + OPJ_INT32 *ai = a; + OPJ_INT32 *bi = h->mem + h->cas; + OPJ_INT32 i = h->sn; while( i-- ) { *bi = *(ai++); bi += 2; @@ -178,10 +252,10 @@ static void dwt_interleave_h(dwt_t* h, int *a) { /* */ /* Inverse lazy transform (vertical). */ /* */ -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; +void opj_dwt_interleave_v(dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x) { + OPJ_INT32 *ai = a; + OPJ_INT32 *bi = v->mem + v->cas; + OPJ_INT32 i = v->sn; while( i-- ) { *bi = *ai; bi += 2; @@ -201,8 +275,8 @@ static void dwt_interleave_v(dwt_t* v, int *a, int x) { /* */ /* Forward 5-3 wavelet transform in 1-D. */ /* */ -static void dwt_encode_1(int *a, int dn, int sn, int cas) { - int i; +void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas) { + OPJ_INT32 i; if (!cas) { if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ @@ -222,8 +296,8 @@ static void dwt_encode_1(int *a, int dn, int sn, int cas) { /* */ /* Inverse 5-3 wavelet transform in 1-D. */ /* */ -static void dwt_decode_1_(int *a, int dn, int sn, int cas) { - int i; +void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas) { + OPJ_INT32 i; if (!cas) { if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ @@ -243,15 +317,15 @@ static void dwt_decode_1_(int *a, int dn, int sn, int cas) { /* */ /* Inverse 5-3 wavelet transform in 1-D. */ /* */ -static void dwt_decode_1(dwt_t *v) { - dwt_decode_1_(v->mem, v->dn, v->sn, v->cas); +void opj_dwt_decode_1(dwt_t *v) { + opj_dwt_decode_1_(v->mem, v->dn, v->sn, v->cas); } /* */ /* Forward 9-7 wavelet transform in 1-D. */ /* */ -static void dwt_encode_1_real(int *a, int dn, int sn, int cas) { - int i; +void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas) { + OPJ_INT32 i; if (!cas) { if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ for (i = 0; i < dn; i++) @@ -285,104 +359,8 @@ static void dwt_encode_1_real(int *a, int dn, int sn, int cas) { } } -static void dwt_decode_sm(dwt_t* v, int k, int n, int x) { - int m = k > n ? n : k; - int l = v->mem[1]; //D(0); - int j; - int i; - for (i = 0; i < m; i++) { - j = l; - WS(i) -= fix_mul( ( l = WD(i) ) + j , x); - } - if( i < k ) { - l = fix_mul( l + l , x ); - for (; i < k; i++) - WS(i) -= l; - } -} - -static void dwt_decode_sp(dwt_t* v, int k, int n, int x) { - int m = k > n ? n : k; - int l = v->mem[1]; //D(0); - int j; - int i; - for (i = 0; i < m; i++) { - j = l; - WS(i) += fix_mul( ( l = WD(i) ) + j , x); - } - if( i < k ) { - l = fix_mul( l + l , x ); - for (; i < k; i++) - WS(i) += l; - } -} - -static void dwt_decode_dm(dwt_t* v, int k, int n, int x) { - int m = k >= n ? n-1 : k; - int l = v->mem[0]; //S(0); - int i; - int j; - for (i = 0; i < m; i++) { - j = l; - WD(i) -= fix_mul( ( l = WS(i+1) ) + j , x); - } - if( i < k ) { - l = fix_mul( l + l , x ); - for (; i < k; i++) - WD(i) -= l; - } -} - -static void dwt_decode_dp(dwt_t* v, int k, int n, int x) { - int m = k >= n ? n-1 : k; - int l = v->mem[0]; //S(0); - int i; - int j; - for (i = 0; i < m; i++) { - j = l; - WD(i) += fix_mul( ( l = WS(i+1) ) + j , x); - } - - if( i < k ) { - l = fix_mul( l + l , x ); - for (; i < k; i++) - WD(i) += l; - } -} - - -/* */ -/* Inverse 9-7 wavelet transform in 1-D. */ -/* */ -static void dwt_decode_1_real(dwt_t* v) { - int i; - if (!v->cas) { - if ((v->dn > 0) || (v->sn > 1)) { /* NEW : CASE ONE ELEMENT */ - for (i = 0; i < v->sn; i++) - WS(i) = fix_mul(WS(i), 10078); /* 10076 */ - for (i = 0; i < v->dn; i++) - WD(i) = fix_mul(WD(i), 13318); /* 13320 */ - dwt_decode_sm(v, v->sn, v->dn, 3633); - dwt_decode_dm(v, v->dn, v->sn, 7233); - dwt_decode_sp(v, v->sn, v->dn, 434); - dwt_decode_dp(v, v->dn, v->sn, 12994); - } - } else { - if ((v->sn > 0) || (v->dn > 1)) { /* NEW : CASE ONE ELEMENT */ - for (i = 0; i < v->sn; i++) - WD(i) = fix_mul(WD(i), 10078); /* 10076 */ - for (i = 0; i < v->dn; i++) - WS(i) = fix_mul(WS(i), 13318); /* 13320 */ - dwt_decode_dm(v, v->sn, v->dn, 3633); - dwt_decode_sm(v, v->dn, v->sn, 7233); - dwt_decode_dp(v, v->sn, v->dn, 434); - dwt_decode_sp(v, v->dn, v->sn, 12994); - } - } -} - -static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno_stepsize) { - int p, n; +void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps, opj_stepsize_t *bandno_stepsize) { + OPJ_INT32 p, n; p = int_floorlog2(stepsize) - 13; n = 11 - int_floorlog2(stepsize); bandno_stepsize->mant = (n < 0 ? stepsize >> -n : stepsize << n) & 0x7ff; @@ -395,74 +373,108 @@ static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno ========================================================== */ + /* */ /* Forward 5-3 wavelet transform in 2-D. */ /* */ -void dwt_encode(opj_tcd_tilecomp_t * tilec) { - int i, j, k; - int *a = NULL; - int *aj = NULL; - int *bj = NULL; - int w, l; - +INLINE opj_bool opj_dwt_encode_procedure(opj_tcd_tilecomp_v2_t * tilec,void (*p_function)(OPJ_INT32 *, OPJ_INT32,OPJ_INT32,OPJ_INT32) ) +{ + OPJ_INT32 i, j, k; + OPJ_INT32 *a = 00; + OPJ_INT32 *aj = 00; + OPJ_INT32 *bj = 00; + OPJ_INT32 w, l; + + OPJ_INT32 rw; /* width of the resolution level computed */ + OPJ_INT32 rh; /* height of the resolution level computed */ + OPJ_INT32 l_data_size; + + opj_tcd_resolution_v2_t * l_cur_res = 0; + opj_tcd_resolution_v2_t * l_last_res = 0; + w = tilec->x1-tilec->x0; l = tilec->numresolutions-1; a = tilec->data; - - for (i = 0; i < l; i++) { - int rw; /* width 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 */ - 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; - + + l_cur_res = tilec->resolutions + l; + l_last_res = l_cur_res - 1; + + rw = l_cur_res->x1 - l_cur_res->x0; + rh = l_cur_res->y1 - l_cur_res->y0; + + l_data_size = opj_dwt_max_resolution( tilec->resolutions,tilec->numresolutions) * sizeof(OPJ_INT32); + bj = (OPJ_INT32*)opj_malloc(l_data_size); + if (! bj) { + return OPJ_FALSE; + } + i = l; + + while (i--) { + OPJ_INT32 rw1; /* width of the resolution level once lower than computed one */ + OPJ_INT32 rh1; /* height of the resolution level once lower than computed one */ + OPJ_INT32 cas_col; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */ + OPJ_INT32 cas_row; /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering */ + OPJ_INT32 dn, sn; + + rw = l_cur_res->x1 - l_cur_res->x0; + rh = l_cur_res->y1 - l_cur_res->y0; + rw1 = l_last_res->x1 - l_last_res->x0; + rh1 = l_last_res->y1 - l_last_res->y0; + + cas_row = l_cur_res->x0 & 1; + cas_col = l_cur_res->y0 & 1; + sn = rh1; dn = rh - rh1; - bj = (int*)opj_malloc(rh * sizeof(int)); - for (j = 0; j < rw; j++) { + for (j = 0; j < rw; ++j) { aj = a + j; - for (k = 0; k < rh; k++) bj[k] = aj[k*w]; - dwt_encode_1(bj, dn, sn, cas_col); - dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col); + for (k = 0; k < rh; ++k) { + bj[k] = aj[k*w]; + } + + (*p_function) (bj, dn, sn, cas_col); + + opj_dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col); } - opj_free(bj); - + sn = rw1; dn = rw - rw1; - bj = (int*)opj_malloc(rw * sizeof(int)); + for (j = 0; j < rh; j++) { aj = a + j * w; for (k = 0; k < rw; k++) bj[k] = aj[k]; - dwt_encode_1(bj, dn, sn, cas_row); - dwt_deinterleave_h(bj, aj, dn, sn, cas_row); + (*p_function) (bj, dn, sn, cas_row); + opj_dwt_deinterleave_h(bj, aj, dn, sn, cas_row); } - opj_free(bj); + + l_cur_res = l_last_res; + + --l_last_res; } + + opj_free(bj); + return OPJ_TRUE; } +/* Forward 5-3 wavelet transform in 2-D. */ +/* */ +opj_bool opj_dwt_encode(opj_tcd_tilecomp_v2_t * tilec) +{ + return opj_dwt_encode_procedure(tilec,opj_dwt_encode_1); +} /* */ /* Inverse 5-3 wavelet transform in 2-D. */ /* */ -void dwt_decode(opj_tcd_tilecomp_t * tilec, int stop) { - dwt_decode_tile(tilec, stop, &dwt_decode_1); +opj_bool opj_dwt_decode(opj_tcd_tilecomp_v2_t* tilec, OPJ_UINT32 numres) { + return opj_dwt_decode_tile(tilec, numres, &opj_dwt_decode_1); } /* */ /* Get gain of 5-3 wavelet transform. */ /* */ -int dwt_getgain(int orient) { +OPJ_UINT32 opj_dwt_getgain(OPJ_UINT32 orient) { if (orient == 0) return 0; if (orient == 1 || orient == 2) @@ -474,78 +486,28 @@ int dwt_getgain(int orient) { /* Get norm of 5-3 wavelet. */ /* */ double dwt_getnorm(int level, int orient) { - return dwt_norms[orient][level]; + return opj_dwt_norms[orient][level]; } -/* */ -/* Forward 9-7 wavelet transform in 2-D. */ -/* */ - -void dwt_encode_real(opj_tcd_tilecomp_t * tilec) { - 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 = 0; i < l; i++) { - int rw; /* width 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 */ - 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 = rh1; - dn = rh - rh1; - bj = (int*)opj_malloc(rh * sizeof(int)); - for (j = 0; j < rw; j++) { - aj = a + j; - for (k = 0; k < rh; k++) bj[k] = aj[k*w]; - dwt_encode_1_real(bj, dn, sn, cas_col); - dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col); - } - opj_free(bj); - - sn = rw1; - dn = rw - rw1; - bj = (int*)opj_malloc(rw * sizeof(int)); - for (j = 0; j < rh; j++) { - aj = a + j * w; - for (k = 0; k < rw; k++) bj[k] = aj[k]; - dwt_encode_1_real(bj, dn, sn, cas_row); - dwt_deinterleave_h(bj, aj, dn, sn, cas_row); - } - opj_free(bj); - } +/* */ +/* Get norm of 5-3 wavelet. */ +/* */ +OPJ_FLOAT64 opj_dwt_getnorm(OPJ_UINT32 level, OPJ_UINT32 orient) { + return opj_dwt_norms[orient][level]; } - /* */ -/* Inverse 9-7 wavelet transform in 2-D. */ +/* Forward 9-7 wavelet transform in 2-D. */ /* */ -void dwt_decode_real(opj_tcd_tilecomp_t * tilec, int stop) { - dwt_decode_tile(tilec, stop, dwt_decode_1_real); +opj_bool opj_dwt_encode_real(opj_tcd_tilecomp_v2_t * tilec) +{ + return opj_dwt_encode_procedure(tilec,opj_dwt_encode_1_real); } - /* */ /* Get gain of 9-7 wavelet transform. */ /* */ -int dwt_getgain_real(int orient) { +OPJ_UINT32 opj_dwt_getgain_real(OPJ_UINT32 orient) { (void)orient; return 0; } @@ -554,15 +516,22 @@ int dwt_getgain_real(int orient) { /* Get norm of 9-7 wavelet. */ /* */ double dwt_getnorm_real(int level, int orient) { - return dwt_norms_real[orient][level]; + return opj_dwt_norms_real[orient][level]; } -void dwt_calc_explicit_stepsizes(opj_tccp_t * tccp, int prec) { - int numbands, bandno; +/* */ +/* Get norm of 9-7 wavelet. */ +/* */ +OPJ_FLOAT64 opj_dwt_getnorm_real(OPJ_UINT32 level, OPJ_UINT32 orient) { + return opj_dwt_norms_real[orient][level]; +} + +void opj_dwt_calc_explicit_stepsizes(opj_tccp_t * tccp, OPJ_UINT32 prec) { + OPJ_UINT32 numbands, bandno; numbands = 3 * tccp->numresolutions - 2; for (bandno = 0; bandno < numbands; bandno++) { - double stepsize; - int resno, level, orient, gain; + OPJ_FLOAT64 stepsize; + OPJ_UINT32 resno, level, orient, gain; resno = (bandno == 0) ? 0 : ((bandno - 1) / 3 + 1); orient = (bandno == 0) ? 0 : ((bandno - 1) % 3 + 1); @@ -571,18 +540,18 @@ void dwt_calc_explicit_stepsizes(opj_tccp_t * tccp, int prec) { if (tccp->qntsty == J2K_CCP_QNTSTY_NOQNT) { stepsize = 1.0; } else { - double norm = dwt_norms_real[orient][level]; + OPJ_FLOAT64 norm = opj_dwt_norms_real[orient][level]; stepsize = (1 << (gain)) / norm; } - dwt_encode_stepsize((int) floor(stepsize * 8192.0), prec + gain, &tccp->stepsizes[bandno]); + opj_dwt_encode_stepsize((OPJ_INT32) floor(stepsize * 8192.0), prec + gain, &tccp->stepsizes[bandno]); } } - +#ifdef OPJ_V1 /* */ /* Determine maximum computed resolution level for inverse wavelet transform */ /* */ -static int dwt_decode_max_resolution(opj_tcd_resolution_t* r, int i) { +static int dwt_decode_max_resolution(opj_tcd_resolution_t* restrict r, int i) { int mr = 1; int w; while( --i ) { @@ -594,67 +563,499 @@ static int dwt_decode_max_resolution(opj_tcd_resolution_t* r, int i) { } return mr ; } +#endif +/* */ +/* Determine maximum computed resolution level for inverse wavelet transform */ +/* */ +static OPJ_UINT32 dwt_max_resolution(opj_tcd_resolution_t* restrict r, OPJ_UINT32 i) { + OPJ_UINT32 mr = 0; + OPJ_UINT32 w; + while( --i ) { + ++r; + if( mr < ( w = r->x1 - r->x0 ) ) + mr = w ; + if( mr < ( w = r->y1 - r->y0 ) ) + mr = w ; + } + return mr ; +} +/* */ +/* Determine maximum computed resolution level for inverse wavelet transform */ +/* */ +OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_v2_t* restrict r, OPJ_UINT32 i) { + OPJ_UINT32 mr = 0; + OPJ_UINT32 w; + while( --i ) { + ++r; + if( mr < ( w = r->x1 - r->x0 ) ) + mr = w ; + if( mr < ( w = r->y1 - r->y0 ) ) + mr = w ; + } + return mr ; +} /* */ /* Inverse wavelet transform in 2-D. */ /* */ -static void dwt_decode_tile(opj_tcd_tilecomp_t * tilec, int stop, DWT1DFN dwt_1D) { - opj_tcd_resolution_t* tr; - int i, j, k; - int *a = NULL; - int *aj = NULL; - int w; //, l; - int rw; /* width of the resolution level computed */ - int rh; /* height of the resolution level computed */ +opj_bool opj_dwt_decode_tile(opj_tcd_tilecomp_v2_t* tilec, OPJ_UINT32 numres, DWT1DFN dwt_1D) { dwt_t h; dwt_t v; - - if( 1 > ( i = tilec->numresolutions - stop ) ) - return ; - tr = tilec->resolutions; + opj_tcd_resolution_v2_t* tr = tilec->resolutions; - w = tilec->x1-tilec->x0; - a = tilec->data; + OPJ_UINT32 rw = tr->x1 - tr->x0; /* width of the resolution level computed */ + OPJ_UINT32 rh = tr->y1 - tr->y0; /* height of the resolution level computed */ + + OPJ_UINT32 w = tilec->x1 - tilec->x0; + + h.mem = (OPJ_INT32*) + opj_aligned_malloc(opj_dwt_max_resolution(tr, numres) * sizeof(OPJ_INT32)); + if + (! h.mem) + { + return OPJ_FALSE; + } - h.mem = (int*) opj_aligned_malloc(dwt_decode_max_resolution(tr, i) * sizeof(int)); v.mem = h.mem; - rw = tr->x1 - tr->x0; - rh = tr->y1 - tr->y0; + while( --numres) { + OPJ_INT32 * restrict tiledp = tilec->data; + OPJ_UINT32 j; - while( --i ) { - tr++; + ++tr; h.sn = rw; v.sn = rh; - h.dn = ( rw = tr->x1 - tr->x0 ) - h.sn; - v.dn = ( rh = tr->y1 - tr->y0 ) - v.sn; - + + rw = tr->x1 - tr->x0; + rh = tr->y1 - tr->y0; + + h.dn = rw - h.sn; h.cas = tr->x0 % 2; - v.cas = tr->y0 % 2; - aj = a; - j = rh; - while( j-- ) { - dwt_interleave_h(&h, aj); + for(j = 0; j < rh; ++j) { + opj_dwt_interleave_h(&h, &tiledp[j*w]); (dwt_1D)(&h); - k = rw; - while( k-- ) - aj[k] = h.mem[k]; - aj += w; + memcpy(&tiledp[j*w], h.mem, rw * sizeof(OPJ_INT32)); } - aj = a; - j = rw; - while( j-- ) { - dwt_interleave_v(&v, aj, w); + v.dn = rh - v.sn; + v.cas = tr->y0 % 2; + + for(j = 0; j < rw; ++j){ + OPJ_UINT32 k; + opj_dwt_interleave_v(&v, &tiledp[j], w); (dwt_1D)(&v); - k = rh; - while( k-- ) - aj[k * w] = v.mem[k]; - aj++; + for(k = 0; k < rh; ++k) { + tiledp[k * w + j] = v.mem[k]; + } } } opj_aligned_free(h.mem); + return OPJ_TRUE; +} + +void opj_v4dwt_interleave_h(v4dwt_t* restrict w, OPJ_FLOAT32* restrict a, OPJ_INT32 x, OPJ_INT32 size){ + OPJ_FLOAT32* restrict bi = (OPJ_FLOAT32*) (w->wavelet + w->cas); + OPJ_INT32 count = w->sn; + OPJ_INT32 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){ + OPJ_INT32 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){ + OPJ_INT32 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]; /* This one*/ + } + } + + bi = (float*) (w->wavelet + 1 - w->cas); + a += w->sn; + size -= w->sn; + count = w->dn; + } +} + +void opj_v4dwt_interleave_v(v4dwt_t* restrict v , OPJ_FLOAT32* restrict a , OPJ_INT32 x, OPJ_INT32 nb_elts_read){ + v4* restrict bi = v->wavelet + v->cas; + OPJ_INT32 i; + + for(i = 0; i < v->sn; ++i){ + memcpy(&bi[i*2], &a[i*x], nb_elts_read * sizeof(OPJ_FLOAT32)); + } + + a += v->sn * x; + bi = v->wavelet + 1 - v->cas; + + for(i = 0; i < v->dn; ++i){ + memcpy(&bi[i*2], &a[i*x], nb_elts_read * sizeof(OPJ_FLOAT32)); + } +} + +#ifdef __SSE__ + +void opj_v4dwt_decode_step1_sse(v4* w, int count, const __m128 c){ + __m128* restrict vw = (__m128*) w; + OPJ_INT32 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; + } +} + +void opj_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 + +void opj_v4dwt_decode_step1(v4* w, OPJ_INT32 count, const OPJ_FLOAT32 c) +{ + OPJ_FLOAT32* restrict fw = (OPJ_FLOAT32*) w; + OPJ_INT32 i; + for(i = 0; i < count; ++i){ + OPJ_FLOAT32 tmp1 = fw[i*8 ]; + OPJ_FLOAT32 tmp2 = fw[i*8 + 1]; + OPJ_FLOAT32 tmp3 = fw[i*8 + 2]; + OPJ_FLOAT32 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; + } +} + +void opj_v4dwt_decode_step2(v4* l, v4* w, OPJ_INT32 k, OPJ_INT32 m, OPJ_FLOAT32 c) +{ + OPJ_FLOAT32* restrict fl = (OPJ_FLOAT32*) l; + OPJ_FLOAT32* restrict fw = (OPJ_FLOAT32*) w; + int i; + for(i = 0; i < m; ++i){ + OPJ_FLOAT32 tmp1_1 = fl[0]; + OPJ_FLOAT32 tmp1_2 = fl[1]; + OPJ_FLOAT32 tmp1_3 = fl[2]; + OPJ_FLOAT32 tmp1_4 = fl[3]; + OPJ_FLOAT32 tmp2_1 = fw[-4]; + OPJ_FLOAT32 tmp2_2 = fw[-3]; + OPJ_FLOAT32 tmp2_3 = fw[-2]; + OPJ_FLOAT32 tmp2_4 = fw[-1]; + OPJ_FLOAT32 tmp3_1 = fw[0]; + OPJ_FLOAT32 tmp3_2 = fw[1]; + OPJ_FLOAT32 tmp3_3 = fw[2]; + OPJ_FLOAT32 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){ + OPJ_FLOAT32 c1; + OPJ_FLOAT32 c2; + OPJ_FLOAT32 c3; + OPJ_FLOAT32 c4; + c += c; + c1 = fl[0] * c; + c2 = fl[1] * c; + c3 = fl[2] * c; + c4 = fl[3] * c; + for(; m < k; ++m){ + OPJ_FLOAT32 tmp1 = fw[-4]; + OPJ_FLOAT32 tmp2 = fw[-3]; + OPJ_FLOAT32 tmp3 = fw[-2]; + OPJ_FLOAT32 tmp4 = fw[-1]; + fw[-4] = tmp1 + c1; + fw[-3] = tmp2 + c2; + fw[-2] = tmp3 + c3; + fw[-1] = tmp4 + c4; + fw += 8; + } + } +} + +#endif + +/* */ +/* Inverse 9-7 wavelet transform in 1-D. */ +/* */ +void opj_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__ + opj_v4dwt_decode_step1_sse(dwt->wavelet+a, dwt->sn, _mm_set1_ps(opj_K)); + opj_v4dwt_decode_step1_sse(dwt->wavelet+b, dwt->dn, _mm_set1_ps(opj_c13318)); + opj_v4dwt_decode_step2_sse(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), _mm_set1_ps(opj_dwt_delta)); + opj_v4dwt_decode_step2_sse(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), _mm_set1_ps(opj_dwt_gamma)); + opj_v4dwt_decode_step2_sse(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), _mm_set1_ps(opj_dwt_beta)); + opj_v4dwt_decode_step2_sse(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), _mm_set1_ps(opj_dwt_alpha)); +#else + opj_v4dwt_decode_step1(dwt->wavelet+a, dwt->sn, opj_K); + opj_v4dwt_decode_step1(dwt->wavelet+b, dwt->dn, opj_c13318); + opj_v4dwt_decode_step2(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), opj_dwt_delta); + opj_v4dwt_decode_step2(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), opj_dwt_gamma); + opj_v4dwt_decode_step2(dwt->wavelet+b, dwt->wavelet+a+1, dwt->sn, int_min(dwt->sn, dwt->dn-a), opj_dwt_beta); + opj_v4dwt_decode_step2(dwt->wavelet+a, dwt->wavelet+b+1, dwt->dn, int_min(dwt->dn, dwt->sn-b), opj_dwt_alpha); +#endif +} + + +/* KEEP TRUNK VERSION + return type of v2 because rev557 */ +/* */ +/* Inverse 9-7 wavelet transform in 2-D. */ +/* */ +/* V1 void dwt_decode_real(opj_tcd_tilecomp_t* restrict tilec, int numres){ */ +opj_bool 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_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; + opj_v4dwt_interleave_h(&h, aj, w, bufsize); + opj_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; + opj_v4dwt_interleave_h(&h, aj, w, bufsize); + opj_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; + opj_v4dwt_interleave_v(&v, aj, w, 4); + opj_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; + opj_v4dwt_interleave_v(&v, aj, w, j); + opj_v4dwt_decode(&v); + for(k = 0; k < rh; ++k){ + memcpy(&aj[k*w], &v.wavelet[k], j * sizeof(float)); + } + } + } + + opj_aligned_free(h.wavelet); + return OPJ_TRUE; +} + + +/* */ +/* Inverse 9-7 wavelet transform in 2-D. */ +/* */ +opj_bool dwt_decode_real_v2(opj_tcd_tilecomp_v2_t* restrict tilec, OPJ_UINT32 numres) +{ + v4dwt_t h; + v4dwt_t v; + + opj_tcd_resolution_v2_t* res = tilec->resolutions; + + OPJ_UINT32 rw = res->x1 - res->x0; /* width of the resolution level computed */ + OPJ_UINT32 rh = res->y1 - res->y0; /* height of the resolution level computed */ + + OPJ_UINT32 w = tilec->x1 - tilec->x0; + + h.wavelet = (v4*) opj_aligned_malloc((opj_dwt_max_resolution(res, numres)+5) * sizeof(v4)); + v.wavelet = h.wavelet; + + while( --numres) { + OPJ_FLOAT32 * restrict aj = (OPJ_FLOAT32*) tilec->data; + OPJ_UINT32 bufsize = (tilec->x1 - tilec->x0) * (tilec->y1 - tilec->y0); + OPJ_INT32 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) { + OPJ_INT32 k; + opj_v4dwt_interleave_h(&h, aj, w, bufsize); + opj_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) { + OPJ_INT32 k; + j = rh & 0x03; + opj_v4dwt_interleave_h(&h, aj, w, bufsize); + opj_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 = (OPJ_FLOAT32*) tilec->data; + for(j = rw; j > 3; j -= 4){ + OPJ_UINT32 k; + + opj_v4dwt_interleave_v(&v, aj, w, 4); + opj_v4dwt_decode(&v); + + for(k = 0; k < rh; ++k){ + memcpy(&aj[k*w], &v.wavelet[k], 4 * sizeof(OPJ_FLOAT32)); + } + aj += 4; + } + + if (rw & 0x03){ + OPJ_UINT32 k; + + j = rw & 0x03; + + opj_v4dwt_interleave_v(&v, aj, w, j); + opj_v4dwt_decode(&v); + + for(k = 0; k < rh; ++k){ + memcpy(&aj[k*w], &v.wavelet[k], j * sizeof(OPJ_FLOAT32)); + } + } + } + + opj_aligned_free(h.wavelet); + return OPJ_TRUE; }