Home | History | Annotate | Download | only in libopenjpeg20
      1 /*
      2  * The copyright in this software is being made available under the 2-clauses
      3  * BSD License, included below. This software may be subject to other third
      4  * party and contributor rights, including patent rights, and no such rights
      5  * are granted under this license.
      6  *
      7  * Copyright (c) 2002-2014, Universite catholique de Louvain (UCL), Belgium
      8  * Copyright (c) 2002-2014, Professor Benoit Macq
      9  * Copyright (c) 2001-2003, David Janssens
     10  * Copyright (c) 2002-2003, Yannick Verschueren
     11  * Copyright (c) 2003-2007, Francois-Olivier Devaux
     12  * Copyright (c) 2003-2014, Antonin Descampe
     13  * Copyright (c) 2005, Herve Drolon, FreeImage Team
     14  * Copyright (c) 2007, Jonathan Ballard <dzonatas (at) dzonux.net>
     15  * Copyright (c) 2007, Callum Lerwick <seg (at) haxxed.com>
     16  * Copyright (c) 2017, IntoPIX SA <support (at) intopix.com>
     17  * All rights reserved.
     18  *
     19  * Redistribution and use in source and binary forms, with or without
     20  * modification, are permitted provided that the following conditions
     21  * are met:
     22  * 1. Redistributions of source code must retain the above copyright
     23  *    notice, this list of conditions and the following disclaimer.
     24  * 2. Redistributions in binary form must reproduce the above copyright
     25  *    notice, this list of conditions and the following disclaimer in the
     26  *    documentation and/or other materials provided with the distribution.
     27  *
     28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS `AS IS'
     29  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
     32  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     33  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     34  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     35  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     36  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     37  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     38  * POSSIBILITY OF SUCH DAMAGE.
     39  */
     40 
     41 #include <assert.h>
     42 
     43 #define OPJ_SKIP_POISON
     44 #include "opj_includes.h"
     45 
     46 #ifdef __SSE__
     47 #include <xmmintrin.h>
     48 #endif
     49 #ifdef __SSE2__
     50 #include <emmintrin.h>
     51 #endif
     52 #ifdef __SSSE3__
     53 #include <tmmintrin.h>
     54 #endif
     55 #ifdef __AVX2__
     56 #include <immintrin.h>
     57 #endif
     58 
     59 #if defined(__GNUC__)
     60 #pragma GCC poison malloc calloc realloc free
     61 #endif
     62 
     63 /** @defgroup DWT DWT - Implementation of a discrete wavelet transform */
     64 /*@{*/
     65 
     66 #ifdef __AVX2__
     67 /** Number of int32 values in a AVX2 register */
     68 #define VREG_INT_COUNT       8
     69 #else
     70 /** Number of int32 values in a SSE2 register */
     71 #define VREG_INT_COUNT       4
     72 #endif
     73 
     74 /** Number of columns that we can process in parallel in the vertical pass */
     75 #define PARALLEL_COLS_53     (2*VREG_INT_COUNT)
     76 
     77 /** @name Local data structures */
     78 /*@{*/
     79 
     80 typedef struct dwt_local {
     81     OPJ_INT32* mem;
     82     OPJ_SIZE_T mem_count;
     83     OPJ_INT32 dn;   /* number of elements in high pass band */
     84     OPJ_INT32 sn;   /* number of elements in low pass band */
     85     OPJ_INT32 cas;  /* 0 = start on even coord, 1 = start on odd coord */
     86 } opj_dwt_t;
     87 
     88 typedef union {
     89     OPJ_FLOAT32 f[4];
     90 } opj_v4_t;
     91 
     92 typedef struct v4dwt_local {
     93     opj_v4_t*   wavelet ;
     94     OPJ_INT32       dn ;  /* number of elements in high pass band */
     95     OPJ_INT32       sn ;  /* number of elements in low pass band */
     96     OPJ_INT32       cas ; /* 0 = start on even coord, 1 = start on odd coord */
     97     OPJ_UINT32      win_l_x0; /* start coord in low pass band */
     98     OPJ_UINT32      win_l_x1; /* end coord in low pass band */
     99     OPJ_UINT32      win_h_x0; /* start coord in high pass band */
    100     OPJ_UINT32      win_h_x1; /* end coord in high pass band */
    101 } opj_v4dwt_t ;
    102 
    103 static const OPJ_FLOAT32 opj_dwt_alpha =  1.586134342f; /*  12994 */
    104 static const OPJ_FLOAT32 opj_dwt_beta  =  0.052980118f; /*    434 */
    105 static const OPJ_FLOAT32 opj_dwt_gamma = -0.882911075f; /*  -7233 */
    106 static const OPJ_FLOAT32 opj_dwt_delta = -0.443506852f; /*  -3633 */
    107 
    108 static const OPJ_FLOAT32 opj_K      = 1.230174105f; /*  10078 */
    109 static const OPJ_FLOAT32 opj_c13318 = 1.625732422f;
    110 
    111 /*@}*/
    112 
    113 /**
    114 Virtual function type for wavelet transform in 1-D
    115 */
    116 typedef void (*DWT1DFN)(const opj_dwt_t* v);
    117 
    118 /** @name Local static functions */
    119 /*@{*/
    120 
    121 /**
    122 Forward lazy transform (horizontal)
    123 */
    124 static void opj_dwt_deinterleave_h(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn,
    125                                    OPJ_INT32 sn, OPJ_INT32 cas);
    126 /**
    127 Forward lazy transform (vertical)
    128 */
    129 static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn,
    130                                    OPJ_INT32 sn, OPJ_INT32 x, OPJ_INT32 cas);
    131 /**
    132 Forward 5-3 wavelet transform in 1-D
    133 */
    134 static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_SIZE_T a_count, OPJ_INT32 dn,
    135     OPJ_INT32 sn, OPJ_INT32 cas);
    136 /**
    137 Forward 9-7 wavelet transform in 1-D
    138 */
    139 static void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_SIZE_T a_count,
    140     OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas);
    141 
    142 
    143 /**
    144 Explicit calculation of the Quantization Stepsizes
    145 */
    146 static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
    147                                     opj_stepsize_t *bandno_stepsize);
    148 /**
    149 Inverse wavelet transform in 2-D.
    150 */
    151 static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
    152                                     const opj_tcd_tilecomp_t* tilec, OPJ_UINT32 i);
    153 
    154 static OPJ_BOOL opj_dwt_decode_partial_tile(
    155     opj_tcd_tilecomp_t* tilec,
    156     OPJ_UINT32 numres);
    157 
    158 static OPJ_BOOL opj_dwt_encode_procedure(const opj_tcd_tilecomp_t * tilec,
    159         void(*p_function)(OPJ_INT32 *, OPJ_SIZE_T, OPJ_INT32, OPJ_INT32, OPJ_INT32));
    160 
    161 static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r,
    162         OPJ_UINT32 i);
    163 
    164 /* <summary>                             */
    165 /* Inverse 9-7 wavelet transform in 1-D. */
    166 /* </summary>                            */
    167 static void opj_v4dwt_decode(opj_v4dwt_t* OPJ_RESTRICT dwt);
    168 
    169 static void opj_v4dwt_interleave_h(opj_v4dwt_t* OPJ_RESTRICT dwt,
    170                                    OPJ_FLOAT32* OPJ_RESTRICT a,
    171                                    OPJ_UINT32 width,
    172                                    OPJ_UINT32 remaining_height);
    173 
    174 static void opj_v4dwt_interleave_v(opj_v4dwt_t* OPJ_RESTRICT dwt,
    175                                    OPJ_FLOAT32* OPJ_RESTRICT a,
    176                                    OPJ_UINT32 width,
    177                                    OPJ_UINT32 nb_elts_read);
    178 
    179 #ifdef __SSE__
    180 static void opj_v4dwt_decode_step1_sse(opj_v4_t* w,
    181                                        OPJ_UINT32 start,
    182                                        OPJ_UINT32 end,
    183                                        const __m128 c);
    184 
    185 static void opj_v4dwt_decode_step2_sse(opj_v4_t* l, opj_v4_t* w,
    186                                        OPJ_UINT32 start,
    187                                        OPJ_UINT32 end,
    188                                        OPJ_UINT32 m, __m128 c);
    189 
    190 #else
    191 static void opj_v4dwt_decode_step1(opj_v4_t* w,
    192                                    OPJ_UINT32 start,
    193                                    OPJ_UINT32 end,
    194                                    const OPJ_FLOAT32 c);
    195 
    196 static void opj_v4dwt_decode_step2(opj_v4_t* l, opj_v4_t* w,
    197                                    OPJ_UINT32 start,
    198                                    OPJ_UINT32 end,
    199                                    OPJ_UINT32 m,
    200                                    OPJ_FLOAT32 c);
    201 
    202 #endif
    203 
    204 /*@}*/
    205 
    206 /*@}*/
    207 
    208 #define IDX_S(i) (i)*2
    209 #define IDX_D(i) 1 + (i)* 2
    210 #define UNDERFLOW_SN(i) ((i) >= sn&&sn>0)
    211 #define UNDERFLOW_DN(i) ((i) >= dn&&dn>0)
    212 #define OVERFLOW_S(i) (IDX_S(i) >= a_count)
    213 #define OVERFLOW_D(i) (IDX_D(i) >= a_count)
    214 
    215 #define OPJ_S(i) a[IDX_S(i)]
    216 #define OPJ_D(i) a[IDX_D(i)]
    217 #define OPJ_S_(i) ((i)<0 ? OPJ_S(0) : (UNDERFLOW_SN(i) ? OPJ_S(sn - 1) : OVERFLOW_S(i) ? OPJ_S(i - 1) : OPJ_S(i)))
    218 #define OPJ_D_(i) ((i)<0 ? OPJ_D(0) : (UNDERFLOW_DN(i) ? OPJ_D(dn - 1) : OVERFLOW_D(i) ? OPJ_D(i - 1) : OPJ_D(i)))
    219 /* new */
    220 #define OPJ_SS_(i) ((i)<0 ? OPJ_S(0) : (UNDERFLOW_DN(i) ? OPJ_S(dn - 1) : OVERFLOW_S(i) ? OPJ_S(i - 1) : OPJ_S(i)))
    221 #define OPJ_DD_(i) ((i)<0 ? OPJ_D(0) : (UNDERFLOW_SN(i) ? OPJ_D(sn - 1) : OVERFLOW_D(i) ? OPJ_D(i - 1) : OPJ_D(i)))
    222 
    223 /* <summary>                                                              */
    224 /* This table contains the norms of the 5-3 wavelets for different bands. */
    225 /* </summary>                                                             */
    226 /* FIXME! the array should really be extended up to 33 resolution levels */
    227 /* See https://github.com/uclouvain/openjpeg/issues/493 */
    228 static const OPJ_FLOAT64 opj_dwt_norms[4][10] = {
    229     {1.000, 1.500, 2.750, 5.375, 10.68, 21.34, 42.67, 85.33, 170.7, 341.3},
    230     {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9},
    231     {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9},
    232     {.7186, .9218, 1.586, 3.043, 6.019, 12.01, 24.00, 47.97, 95.93}
    233 };
    234 
    235 /* <summary>                                                              */
    236 /* This table contains the norms of the 9-7 wavelets for different bands. */
    237 /* </summary>                                                             */
    238 /* FIXME! the array should really be extended up to 33 resolution levels */
    239 /* See https://github.com/uclouvain/openjpeg/issues/493 */
    240 static const OPJ_FLOAT64 opj_dwt_norms_real[4][10] = {
    241     {1.000, 1.965, 4.177, 8.403, 16.90, 33.84, 67.69, 135.3, 270.6, 540.9},
    242     {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0},
    243     {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0},
    244     {2.080, 3.865, 8.307, 17.18, 34.71, 69.59, 139.3, 278.6, 557.2}
    245 };
    246 
    247 /*
    248 ==========================================================
    249    local functions
    250 ==========================================================
    251 */
    252 
    253 /* <summary>                             */
    254 /* Forward lazy transform (horizontal).  */
    255 /* </summary>                            */
    256 static void opj_dwt_deinterleave_h(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn,
    257                                    OPJ_INT32 sn, OPJ_INT32 cas)
    258 {
    259     OPJ_INT32 i;
    260     OPJ_INT32 * l_dest = b;
    261     OPJ_INT32 * l_src = a + cas;
    262 
    263     for (i = 0; i < sn; ++i) {
    264         *l_dest++ = *l_src;
    265         l_src += 2;
    266     }
    267 
    268     l_dest = b + sn;
    269     l_src = a + 1 - cas;
    270 
    271     for (i = 0; i < dn; ++i)  {
    272         *l_dest++ = *l_src;
    273         l_src += 2;
    274     }
    275 }
    276 
    277 /* <summary>                             */
    278 /* Forward lazy transform (vertical).    */
    279 /* </summary>                            */
    280 static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn,
    281                                    OPJ_INT32 sn, OPJ_INT32 x, OPJ_INT32 cas)
    282 {
    283     OPJ_INT32 i = sn;
    284     OPJ_INT32 * l_dest = b;
    285     OPJ_INT32 * l_src = a + cas;
    286 
    287     while (i--) {
    288         *l_dest = *l_src;
    289         l_dest += x;
    290         l_src += 2;
    291     } /* b[i*x]=a[2*i+cas]; */
    292 
    293     l_dest = b + (OPJ_SIZE_T)sn * (OPJ_SIZE_T)x;
    294     l_src = a + 1 - cas;
    295 
    296     i = dn;
    297     while (i--) {
    298         *l_dest = *l_src;
    299         l_dest += x;
    300         l_src += 2;
    301     } /*b[(sn+i)*x]=a[(2*i+1-cas)];*/
    302 }
    303 
    304 #ifdef STANDARD_SLOW_VERSION
    305 /* <summary>                             */
    306 /* Inverse lazy transform (horizontal).  */
    307 /* </summary>                            */
    308 static void opj_dwt_interleave_h(const opj_dwt_t* h, OPJ_INT32 *a)
    309 {
    310     OPJ_INT32 *ai = a;
    311     OPJ_INT32 *bi = h->mem + h->cas;
    312     OPJ_INT32  i    = h->sn;
    313     while (i--) {
    314         *bi = *(ai++);
    315         bi += 2;
    316     }
    317     ai  = a + h->sn;
    318     bi  = h->mem + 1 - h->cas;
    319     i   = h->dn ;
    320     while (i--) {
    321         *bi = *(ai++);
    322         bi += 2;
    323     }
    324 }
    325 
    326 /* <summary>                             */
    327 /* Inverse lazy transform (vertical).    */
    328 /* </summary>                            */
    329 static void opj_dwt_interleave_v(const opj_dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x)
    330 {
    331     OPJ_INT32 *ai = a;
    332     OPJ_INT32 *bi = v->mem + v->cas;
    333     OPJ_INT32  i = v->sn;
    334     while (i--) {
    335         *bi = *ai;
    336         bi += 2;
    337         ai += x;
    338     }
    339     ai = a + (v->sn * (OPJ_SIZE_T)x);
    340     bi = v->mem + 1 - v->cas;
    341     i = v->dn ;
    342     while (i--) {
    343         *bi = *ai;
    344         bi += 2;
    345         ai += x;
    346     }
    347 }
    348 
    349 #endif /* STANDARD_SLOW_VERSION */
    350 
    351 /* <summary>                            */
    352 /* Forward 5-3 wavelet transform in 1-D. */
    353 /* </summary>                           */
    354 static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_SIZE_T a_count, OPJ_INT32 dn,
    355                              OPJ_INT32 sn, OPJ_INT32 cas)
    356 {
    357     OPJ_INT32 i;
    358 
    359     if (!cas) {
    360         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
    361             for (i = 0; i < dn; i++) {
    362                 OPJ_D(i) -= (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
    363             }
    364             for (i = 0; i < sn; i++) {
    365                 OPJ_S(i) += (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
    366             }
    367         }
    368     } else {
    369         if (!sn && dn == 1) {       /* NEW :  CASE ONE ELEMENT */
    370             OPJ_S(0) *= 2;
    371         } else {
    372             for (i = 0; i < dn; i++) {
    373                 OPJ_S(i) -= (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1;
    374             }
    375             for (i = 0; i < sn; i++) {
    376                 OPJ_D(i) += (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2;
    377             }
    378         }
    379     }
    380 }
    381 
    382 #ifdef STANDARD_SLOW_VERSION
    383 /* <summary>                            */
    384 /* Inverse 5-3 wavelet transform in 1-D. */
    385 /* </summary>                           */
    386 static void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_SIZE_T a_count, OPJ_INT32 dn,
    387                               OPJ_INT32 sn, OPJ_INT32 cas)
    388 {
    389     OPJ_INT32 i;
    390 
    391     if (!cas) {
    392         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
    393             for (i = 0; i < sn; i++) {
    394                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
    395             }
    396             for (i = 0; i < dn; i++) {
    397                 OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
    398             }
    399         }
    400     } else {
    401         if (!sn  && dn == 1) {        /* NEW :  CASE ONE ELEMENT */
    402             OPJ_S(0) /= 2;
    403         } else {
    404             for (i = 0; i < sn; i++) {
    405                 OPJ_D(i) -= (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2;
    406             }
    407             for (i = 0; i < dn; i++) {
    408                 OPJ_S(i) += (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1;
    409             }
    410         }
    411     }
    412 }
    413 
    414 static void opj_dwt_decode_1(const opj_dwt_t *v)
    415 {
    416     opj_dwt_decode_1_(v->mem, v->mem_count, v->dn, v->sn, v->cas);
    417 }
    418 
    419 #endif /* STANDARD_SLOW_VERSION */
    420 
    421 #if !defined(STANDARD_SLOW_VERSION)
    422 static void  opj_idwt53_h_cas0(OPJ_INT32* tmp,
    423                                const OPJ_INT32 sn,
    424                                const OPJ_INT32 len,
    425                                OPJ_INT32* tiledp)
    426 {
    427     OPJ_INT32 i, j;
    428     const OPJ_INT32* in_even = &tiledp[0];
    429     const OPJ_INT32* in_odd = &tiledp[sn];
    430 
    431 #ifdef TWO_PASS_VERSION
    432     /* For documentation purpose: performs lifting in two iterations, */
    433     /* but without explicit interleaving */
    434 
    435     assert(len > 1);
    436 
    437     /* Even */
    438     tmp[0] = in_even[0] - ((in_odd[0] + 1) >> 1);
    439     for (i = 2, j = 0; i <= len - 2; i += 2, j++) {
    440         tmp[i] = in_even[j + 1] - ((in_odd[j] + in_odd[j + 1] + 2) >> 2);
    441     }
    442     if (len & 1) { /* if len is odd */
    443         tmp[len - 1] = in_even[(len - 1) / 2] - ((in_odd[(len - 2) / 2] + 1) >> 1);
    444     }
    445 
    446     /* Odd */
    447     for (i = 1, j = 0; i < len - 1; i += 2, j++) {
    448         tmp[i] = in_odd[j] + ((tmp[i - 1] + tmp[i + 1]) >> 1);
    449     }
    450     if (!(len & 1)) { /* if len is even */
    451         tmp[len - 1] = in_odd[(len - 1) / 2] + tmp[len - 2];
    452     }
    453 #else
    454     OPJ_INT32 d1c, d1n, s1n, s0c, s0n;
    455 
    456     assert(len > 1);
    457 
    458     /* Improved version of the TWO_PASS_VERSION: */
    459     /* Performs lifting in one single iteration. Saves memory */
    460     /* accesses and explicit interleaving. */
    461     s1n = in_even[0];
    462     d1n = in_odd[0];
    463     s0n = s1n - ((d1n + 1) >> 1);
    464 
    465     for (i = 0, j = 1; i < (len - 3); i += 2, j++) {
    466         d1c = d1n;
    467         s0c = s0n;
    468 
    469         s1n = in_even[j];
    470         d1n = in_odd[j];
    471 
    472         s0n = s1n - ((d1c + d1n + 2) >> 2);
    473 
    474         tmp[i  ] = s0c;
    475         tmp[i + 1] = d1c + ((s0c + s0n) >> 1);
    476     }
    477 
    478     tmp[i] = s0n;
    479 
    480     if (len & 1) {
    481         tmp[len - 1] = in_even[(len - 1) / 2] - ((d1n + 1) >> 1);
    482         tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1);
    483     } else {
    484         tmp[len - 1] = d1n + s0n;
    485     }
    486 #endif
    487     memcpy(tiledp, tmp, (OPJ_UINT32)len * sizeof(OPJ_INT32));
    488 }
    489 
    490 static void  opj_idwt53_h_cas1(OPJ_INT32* tmp,
    491                                const OPJ_INT32 sn,
    492                                const OPJ_INT32 len,
    493                                OPJ_INT32* tiledp)
    494 {
    495     OPJ_INT32 i, j;
    496     const OPJ_INT32* in_even = &tiledp[sn];
    497     const OPJ_INT32* in_odd = &tiledp[0];
    498 
    499 #ifdef TWO_PASS_VERSION
    500     /* For documentation purpose: performs lifting in two iterations, */
    501     /* but without explicit interleaving */
    502 
    503     assert(len > 2);
    504 
    505     /* Odd */
    506     for (i = 1, j = 0; i < len - 1; i += 2, j++) {
    507         tmp[i] = in_odd[j] - ((in_even[j] + in_even[j + 1] + 2) >> 2);
    508     }
    509     if (!(len & 1)) {
    510         tmp[len - 1] = in_odd[len / 2 - 1] - ((in_even[len / 2 - 1] + 1) >> 1);
    511     }
    512 
    513     /* Even */
    514     tmp[0] = in_even[0] + tmp[1];
    515     for (i = 2, j = 1; i < len - 1; i += 2, j++) {
    516         tmp[i] = in_even[j] + ((tmp[i + 1] + tmp[i - 1]) >> 1);
    517     }
    518     if (len & 1) {
    519         tmp[len - 1] = in_even[len / 2] + tmp[len - 2];
    520     }
    521 #else
    522     OPJ_INT32 s1, s2, dc, dn;
    523 
    524     assert(len > 2);
    525 
    526     /* Improved version of the TWO_PASS_VERSION: */
    527     /* Performs lifting in one single iteration. Saves memory */
    528     /* accesses and explicit interleaving. */
    529 
    530     s1 = in_even[1];
    531     dc = in_odd[0] - ((in_even[0] + s1 + 2) >> 2);
    532     tmp[0] = in_even[0] + dc;
    533 
    534     for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
    535 
    536         s2 = in_even[j + 1];
    537 
    538         dn = in_odd[j] - ((s1 + s2 + 2) >> 2);
    539         tmp[i  ] = dc;
    540         tmp[i + 1] = s1 + ((dn + dc) >> 1);
    541 
    542         dc = dn;
    543         s1 = s2;
    544     }
    545 
    546     tmp[i] = dc;
    547 
    548     if (!(len & 1)) {
    549         dn = in_odd[len / 2 - 1] - ((s1 + 1) >> 1);
    550         tmp[len - 2] = s1 + ((dn + dc) >> 1);
    551         tmp[len - 1] = dn;
    552     } else {
    553         tmp[len - 1] = s1 + dc;
    554     }
    555 #endif
    556     memcpy(tiledp, tmp, (OPJ_UINT32)len * sizeof(OPJ_INT32));
    557 }
    558 
    559 
    560 #endif /* !defined(STANDARD_SLOW_VERSION) */
    561 
    562 /* <summary>                            */
    563 /* Inverse 5-3 wavelet transform in 1-D for one row. */
    564 /* </summary>                           */
    565 /* Performs interleave, inverse wavelet transform and copy back to buffer */
    566 static void opj_idwt53_h(const opj_dwt_t *dwt,
    567                          OPJ_INT32* tiledp)
    568 {
    569 #ifdef STANDARD_SLOW_VERSION
    570     /* For documentation purpose */
    571     opj_dwt_interleave_h(dwt, tiledp);
    572     opj_dwt_decode_1(dwt);
    573     memcpy(tiledp, dwt->mem, (OPJ_UINT32)(dwt->sn + dwt->dn) * sizeof(OPJ_INT32));
    574 #else
    575     const OPJ_INT32 sn = dwt->sn;
    576     const OPJ_INT32 len = sn + dwt->dn;
    577     if (dwt->cas == 0) { /* Left-most sample is on even coordinate */
    578         if (len > 1) {
    579             opj_idwt53_h_cas0(dwt->mem, sn, len, tiledp);
    580         } else {
    581             /* Unmodified value */
    582         }
    583     } else { /* Left-most sample is on odd coordinate */
    584         if (len == 1) {
    585             tiledp[0] /= 2;
    586         } else if (len == 2) {
    587             OPJ_INT32* out = dwt->mem;
    588             const OPJ_INT32* in_even = &tiledp[sn];
    589             const OPJ_INT32* in_odd = &tiledp[0];
    590             out[1] = in_odd[0] - ((in_even[0] + 1) >> 1);
    591             out[0] = in_even[0] + out[1];
    592             memcpy(tiledp, dwt->mem, (OPJ_UINT32)len * sizeof(OPJ_INT32));
    593         } else if (len > 2) {
    594             opj_idwt53_h_cas1(dwt->mem, sn, len, tiledp);
    595         }
    596     }
    597 #endif
    598 }
    599 
    600 #if (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION)
    601 
    602 /* Conveniency macros to improve the readabilty of the formulas */
    603 #if __AVX2__
    604 #define VREG        __m256i
    605 #define LOAD_CST(x) _mm256_set1_epi32(x)
    606 #define LOAD(x)     _mm256_load_si256((const VREG*)(x))
    607 #define LOADU(x)    _mm256_loadu_si256((const VREG*)(x))
    608 #define STORE(x,y)  _mm256_store_si256((VREG*)(x),(y))
    609 #define STOREU(x,y) _mm256_storeu_si256((VREG*)(x),(y))
    610 #define ADD(x,y)    _mm256_add_epi32((x),(y))
    611 #define SUB(x,y)    _mm256_sub_epi32((x),(y))
    612 #define SAR(x,y)    _mm256_srai_epi32((x),(y))
    613 #else
    614 #define VREG        __m128i
    615 #define LOAD_CST(x) _mm_set1_epi32(x)
    616 #define LOAD(x)     _mm_load_si128((const VREG*)(x))
    617 #define LOADU(x)    _mm_loadu_si128((const VREG*)(x))
    618 #define STORE(x,y)  _mm_store_si128((VREG*)(x),(y))
    619 #define STOREU(x,y) _mm_storeu_si128((VREG*)(x),(y))
    620 #define ADD(x,y)    _mm_add_epi32((x),(y))
    621 #define SUB(x,y)    _mm_sub_epi32((x),(y))
    622 #define SAR(x,y)    _mm_srai_epi32((x),(y))
    623 #endif
    624 #define ADD3(x,y,z) ADD(ADD(x,y),z)
    625 
    626 static
    627 void opj_idwt53_v_final_memcpy(OPJ_INT32* tiledp_col,
    628                                const OPJ_INT32* tmp,
    629                                OPJ_INT32 len,
    630                                OPJ_SIZE_T stride)
    631 {
    632     OPJ_INT32 i;
    633     for (i = 0; i < len; ++i) {
    634         /* A memcpy(&tiledp_col[i * stride + 0],
    635                     &tmp[PARALLEL_COLS_53 * i + 0],
    636                     PARALLEL_COLS_53 * sizeof(OPJ_INT32))
    637            would do but would be a tiny bit slower.
    638            We can take here advantage of our knowledge of alignment */
    639         STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + 0],
    640                LOAD(&tmp[PARALLEL_COLS_53 * i + 0]));
    641         STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + VREG_INT_COUNT],
    642                LOAD(&tmp[PARALLEL_COLS_53 * i + VREG_INT_COUNT]));
    643     }
    644 }
    645 
    646 /** Vertical inverse 5x3 wavelet transform for 8 columns in SSE2, or
    647  * 16 in AVX2, when top-most pixel is on even coordinate */
    648 static void opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2(
    649     OPJ_INT32* tmp,
    650     const OPJ_INT32 sn,
    651     const OPJ_INT32 len,
    652     OPJ_INT32* tiledp_col,
    653     const OPJ_SIZE_T stride)
    654 {
    655     const OPJ_INT32* in_even = &tiledp_col[0];
    656     const OPJ_INT32* in_odd = &tiledp_col[(OPJ_SIZE_T)sn * stride];
    657 
    658     OPJ_INT32 i;
    659     OPJ_SIZE_T j;
    660     VREG d1c_0, d1n_0, s1n_0, s0c_0, s0n_0;
    661     VREG d1c_1, d1n_1, s1n_1, s0c_1, s0n_1;
    662     const VREG two = LOAD_CST(2);
    663 
    664     assert(len > 1);
    665 #if __AVX2__
    666     assert(PARALLEL_COLS_53 == 16);
    667     assert(VREG_INT_COUNT == 8);
    668 #else
    669     assert(PARALLEL_COLS_53 == 8);
    670     assert(VREG_INT_COUNT == 4);
    671 #endif
    672 
    673     /* Note: loads of input even/odd values must be done in a unaligned */
    674     /* fashion. But stores in tmp can be done with aligned store, since */
    675     /* the temporary buffer is properly aligned */
    676     assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
    677 
    678     s1n_0 = LOADU(in_even + 0);
    679     s1n_1 = LOADU(in_even + VREG_INT_COUNT);
    680     d1n_0 = LOADU(in_odd);
    681     d1n_1 = LOADU(in_odd + VREG_INT_COUNT);
    682 
    683     /* s0n = s1n - ((d1n + 1) >> 1); <==> */
    684     /* s0n = s1n - ((d1n + d1n + 2) >> 2); */
    685     s0n_0 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2));
    686     s0n_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2));
    687 
    688     for (i = 0, j = 1; i < (len - 3); i += 2, j++) {
    689         d1c_0 = d1n_0;
    690         s0c_0 = s0n_0;
    691         d1c_1 = d1n_1;
    692         s0c_1 = s0n_1;
    693 
    694         s1n_0 = LOADU(in_even + j * stride);
    695         s1n_1 = LOADU(in_even + j * stride + VREG_INT_COUNT);
    696         d1n_0 = LOADU(in_odd + j * stride);
    697         d1n_1 = LOADU(in_odd + j * stride + VREG_INT_COUNT);
    698 
    699         /*s0n = s1n - ((d1c + d1n + 2) >> 2);*/
    700         s0n_0 = SUB(s1n_0, SAR(ADD3(d1c_0, d1n_0, two), 2));
    701         s0n_1 = SUB(s1n_1, SAR(ADD3(d1c_1, d1n_1, two), 2));
    702 
    703         STORE(tmp + PARALLEL_COLS_53 * (i + 0), s0c_0);
    704         STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0c_1);
    705 
    706         /* d1c + ((s0c + s0n) >> 1) */
    707         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0,
    708               ADD(d1c_0, SAR(ADD(s0c_0, s0n_0), 1)));
    709         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT,
    710               ADD(d1c_1, SAR(ADD(s0c_1, s0n_1), 1)));
    711     }
    712 
    713     STORE(tmp + PARALLEL_COLS_53 * (i + 0) + 0, s0n_0);
    714     STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0n_1);
    715 
    716     if (len & 1) {
    717         VREG tmp_len_minus_1;
    718         s1n_0 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride);
    719         /* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */
    720         tmp_len_minus_1 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2));
    721         STORE(tmp + PARALLEL_COLS_53 * (len - 1), tmp_len_minus_1);
    722         /* d1n + ((s0n + tmp_len_minus_1) >> 1) */
    723         STORE(tmp + PARALLEL_COLS_53 * (len - 2),
    724               ADD(d1n_0, SAR(ADD(s0n_0, tmp_len_minus_1), 1)));
    725 
    726         s1n_1 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride + VREG_INT_COUNT);
    727         /* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */
    728         tmp_len_minus_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2));
    729         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
    730               tmp_len_minus_1);
    731         /* d1n + ((s0n + tmp_len_minus_1) >> 1) */
    732         STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT,
    733               ADD(d1n_1, SAR(ADD(s0n_1, tmp_len_minus_1), 1)));
    734 
    735 
    736     } else {
    737         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0,
    738               ADD(d1n_0, s0n_0));
    739         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
    740               ADD(d1n_1, s0n_1));
    741     }
    742 
    743     opj_idwt53_v_final_memcpy(tiledp_col, tmp, len, stride);
    744 }
    745 
    746 
    747 /** Vertical inverse 5x3 wavelet transform for 8 columns in SSE2, or
    748  * 16 in AVX2, when top-most pixel is on odd coordinate */
    749 static void opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2(
    750     OPJ_INT32* tmp,
    751     const OPJ_INT32 sn,
    752     const OPJ_INT32 len,
    753     OPJ_INT32* tiledp_col,
    754     const OPJ_SIZE_T stride)
    755 {
    756     OPJ_INT32 i;
    757     OPJ_SIZE_T j;
    758 
    759     VREG s1_0, s2_0, dc_0, dn_0;
    760     VREG s1_1, s2_1, dc_1, dn_1;
    761     const VREG two = LOAD_CST(2);
    762 
    763     const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
    764     const OPJ_INT32* in_odd = &tiledp_col[0];
    765 
    766     assert(len > 2);
    767 #if __AVX2__
    768     assert(PARALLEL_COLS_53 == 16);
    769     assert(VREG_INT_COUNT == 8);
    770 #else
    771     assert(PARALLEL_COLS_53 == 8);
    772     assert(VREG_INT_COUNT == 4);
    773 #endif
    774 
    775     /* Note: loads of input even/odd values must be done in a unaligned */
    776     /* fashion. But stores in tmp can be done with aligned store, since */
    777     /* the temporary buffer is properly aligned */
    778     assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
    779 
    780     s1_0 = LOADU(in_even + stride);
    781     /* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */
    782     dc_0 = SUB(LOADU(in_odd + 0),
    783                SAR(ADD3(LOADU(in_even + 0), s1_0, two), 2));
    784     STORE(tmp + PARALLEL_COLS_53 * 0, ADD(LOADU(in_even + 0), dc_0));
    785 
    786     s1_1 = LOADU(in_even + stride + VREG_INT_COUNT);
    787     /* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */
    788     dc_1 = SUB(LOADU(in_odd + VREG_INT_COUNT),
    789                SAR(ADD3(LOADU(in_even + VREG_INT_COUNT), s1_1, two), 2));
    790     STORE(tmp + PARALLEL_COLS_53 * 0 + VREG_INT_COUNT,
    791           ADD(LOADU(in_even + VREG_INT_COUNT), dc_1));
    792 
    793     for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
    794 
    795         s2_0 = LOADU(in_even + (j + 1) * stride);
    796         s2_1 = LOADU(in_even + (j + 1) * stride + VREG_INT_COUNT);
    797 
    798         /* dn = in_odd[j * stride] - ((s1 + s2 + 2) >> 2); */
    799         dn_0 = SUB(LOADU(in_odd + j * stride),
    800                    SAR(ADD3(s1_0, s2_0, two), 2));
    801         dn_1 = SUB(LOADU(in_odd + j * stride + VREG_INT_COUNT),
    802                    SAR(ADD3(s1_1, s2_1, two), 2));
    803 
    804         STORE(tmp + PARALLEL_COLS_53 * i, dc_0);
    805         STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1);
    806 
    807         /* tmp[i + 1] = s1 + ((dn + dc) >> 1); */
    808         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0,
    809               ADD(s1_0, SAR(ADD(dn_0, dc_0), 1)));
    810         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT,
    811               ADD(s1_1, SAR(ADD(dn_1, dc_1), 1)));
    812 
    813         dc_0 = dn_0;
    814         s1_0 = s2_0;
    815         dc_1 = dn_1;
    816         s1_1 = s2_1;
    817     }
    818     STORE(tmp + PARALLEL_COLS_53 * i, dc_0);
    819     STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1);
    820 
    821     if (!(len & 1)) {
    822         /*dn = in_odd[(len / 2 - 1) * stride] - ((s1 + 1) >> 1); */
    823         dn_0 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride),
    824                    SAR(ADD3(s1_0, s1_0, two), 2));
    825         dn_1 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride + VREG_INT_COUNT),
    826                    SAR(ADD3(s1_1, s1_1, two), 2));
    827 
    828         /* tmp[len - 2] = s1 + ((dn + dc) >> 1); */
    829         STORE(tmp + PARALLEL_COLS_53 * (len - 2) + 0,
    830               ADD(s1_0, SAR(ADD(dn_0, dc_0), 1)));
    831         STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT,
    832               ADD(s1_1, SAR(ADD(dn_1, dc_1), 1)));
    833 
    834         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, dn_0);
    835         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, dn_1);
    836     } else {
    837         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, ADD(s1_0, dc_0));
    838         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
    839               ADD(s1_1, dc_1));
    840     }
    841 
    842     opj_idwt53_v_final_memcpy(tiledp_col, tmp, len, stride);
    843 }
    844 
    845 #undef VREG
    846 #undef LOAD_CST
    847 #undef LOADU
    848 #undef LOAD
    849 #undef STORE
    850 #undef STOREU
    851 #undef ADD
    852 #undef ADD3
    853 #undef SUB
    854 #undef SAR
    855 
    856 #endif /* (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION) */
    857 
    858 #if !defined(STANDARD_SLOW_VERSION)
    859 /** Vertical inverse 5x3 wavelet transform for one column, when top-most
    860  * pixel is on even coordinate */
    861 static void opj_idwt3_v_cas0(OPJ_INT32* tmp,
    862                              const OPJ_INT32 sn,
    863                              const OPJ_INT32 len,
    864                              OPJ_INT32* tiledp_col,
    865                              const OPJ_SIZE_T stride)
    866 {
    867     OPJ_INT32 i, j;
    868     OPJ_INT32 d1c, d1n, s1n, s0c, s0n;
    869 
    870     assert(len > 1);
    871 
    872     /* Performs lifting in one single iteration. Saves memory */
    873     /* accesses and explicit interleaving. */
    874 
    875     s1n = tiledp_col[0];
    876     d1n = tiledp_col[(OPJ_SIZE_T)sn * stride];
    877     s0n = s1n - ((d1n + 1) >> 1);
    878 
    879     for (i = 0, j = 0; i < (len - 3); i += 2, j++) {
    880         d1c = d1n;
    881         s0c = s0n;
    882 
    883         s1n = tiledp_col[(OPJ_SIZE_T)(j + 1) * stride];
    884         d1n = tiledp_col[(OPJ_SIZE_T)(sn + j + 1) * stride];
    885 
    886         s0n = s1n - ((d1c + d1n + 2) >> 2);
    887 
    888         tmp[i  ] = s0c;
    889         tmp[i + 1] = d1c + ((s0c + s0n) >> 1);
    890     }
    891 
    892     tmp[i] = s0n;
    893 
    894     if (len & 1) {
    895         tmp[len - 1] =
    896             tiledp_col[(OPJ_SIZE_T)((len - 1) / 2) * stride] -
    897             ((d1n + 1) >> 1);
    898         tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1);
    899     } else {
    900         tmp[len - 1] = d1n + s0n;
    901     }
    902 
    903     for (i = 0; i < len; ++i) {
    904         tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i];
    905     }
    906 }
    907 
    908 
    909 /** Vertical inverse 5x3 wavelet transform for one column, when top-most
    910  * pixel is on odd coordinate */
    911 static void opj_idwt3_v_cas1(OPJ_INT32* tmp,
    912                              const OPJ_INT32 sn,
    913                              const OPJ_INT32 len,
    914                              OPJ_INT32* tiledp_col,
    915                              const OPJ_SIZE_T stride)
    916 {
    917     OPJ_INT32 i, j;
    918     OPJ_INT32 s1, s2, dc, dn;
    919     const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
    920     const OPJ_INT32* in_odd = &tiledp_col[0];
    921 
    922     assert(len > 2);
    923 
    924     /* Performs lifting in one single iteration. Saves memory */
    925     /* accesses and explicit interleaving. */
    926 
    927     s1 = in_even[stride];
    928     dc = in_odd[0] - ((in_even[0] + s1 + 2) >> 2);
    929     tmp[0] = in_even[0] + dc;
    930     for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
    931 
    932         s2 = in_even[(OPJ_SIZE_T)(j + 1) * stride];
    933 
    934         dn = in_odd[(OPJ_SIZE_T)j * stride] - ((s1 + s2 + 2) >> 2);
    935         tmp[i  ] = dc;
    936         tmp[i + 1] = s1 + ((dn + dc) >> 1);
    937 
    938         dc = dn;
    939         s1 = s2;
    940     }
    941     tmp[i] = dc;
    942     if (!(len & 1)) {
    943         dn = in_odd[(OPJ_SIZE_T)(len / 2 - 1) * stride] - ((s1 + 1) >> 1);
    944         tmp[len - 2] = s1 + ((dn + dc) >> 1);
    945         tmp[len - 1] = dn;
    946     } else {
    947         tmp[len - 1] = s1 + dc;
    948     }
    949 
    950     for (i = 0; i < len; ++i) {
    951         tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i];
    952     }
    953 }
    954 #endif /* !defined(STANDARD_SLOW_VERSION) */
    955 
    956 /* <summary>                            */
    957 /* Inverse vertical 5-3 wavelet transform in 1-D for several columns. */
    958 /* </summary>                           */
    959 /* Performs interleave, inverse wavelet transform and copy back to buffer */
    960 static void opj_idwt53_v(const opj_dwt_t *dwt,
    961                          OPJ_INT32* tiledp_col,
    962                          OPJ_SIZE_T stride,
    963                          OPJ_INT32 nb_cols)
    964 {
    965 #ifdef STANDARD_SLOW_VERSION
    966     /* For documentation purpose */
    967     OPJ_INT32 k, c;
    968     for (c = 0; c < nb_cols; c ++) {
    969         opj_dwt_interleave_v(dwt, tiledp_col + c, stride);
    970         opj_dwt_decode_1(dwt);
    971         for (k = 0; k < dwt->sn + dwt->dn; ++k) {
    972             tiledp_col[c + k * stride] = dwt->mem[k];
    973         }
    974     }
    975 #else
    976     const OPJ_INT32 sn = dwt->sn;
    977     const OPJ_INT32 len = sn + dwt->dn;
    978     if (dwt->cas == 0) {
    979         /* If len == 1, unmodified value */
    980 
    981 #if (defined(__SSE2__) || defined(__AVX2__))
    982         if (len > 1 && nb_cols == PARALLEL_COLS_53) {
    983             /* Same as below general case, except that thanks to SSE2/AVX2 */
    984             /* we can efficently process 8/16 columns in parallel */
    985             opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2(dwt->mem, sn, len, tiledp_col, stride);
    986             return;
    987         }
    988 #endif
    989         if (len > 1) {
    990             OPJ_INT32 c;
    991             for (c = 0; c < nb_cols; c++, tiledp_col++) {
    992                 opj_idwt3_v_cas0(dwt->mem, sn, len, tiledp_col, stride);
    993             }
    994             return;
    995         }
    996     } else {
    997         if (len == 1) {
    998             OPJ_INT32 c;
    999             for (c = 0; c < nb_cols; c++, tiledp_col++) {
   1000                 tiledp_col[0] /= 2;
   1001             }
   1002             return;
   1003         }
   1004 
   1005         if (len == 2) {
   1006             OPJ_INT32 c;
   1007             OPJ_INT32* out = dwt->mem;
   1008             for (c = 0; c < nb_cols; c++, tiledp_col++) {
   1009                 OPJ_INT32 i;
   1010                 const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
   1011                 const OPJ_INT32* in_odd = &tiledp_col[0];
   1012 
   1013                 out[1] = in_odd[0] - ((in_even[0] + 1) >> 1);
   1014                 out[0] = in_even[0] + out[1];
   1015 
   1016                 for (i = 0; i < len; ++i) {
   1017                     tiledp_col[(OPJ_SIZE_T)i * stride] = out[i];
   1018                 }
   1019             }
   1020 
   1021             return;
   1022         }
   1023 
   1024 #if (defined(__SSE2__) || defined(__AVX2__))
   1025         if (len > 2 && nb_cols == PARALLEL_COLS_53) {
   1026             /* Same as below general case, except that thanks to SSE2/AVX2 */
   1027             /* we can efficently process 8/16 columns in parallel */
   1028             opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2(dwt->mem, sn, len, tiledp_col, stride);
   1029             return;
   1030         }
   1031 #endif
   1032         if (len > 2) {
   1033             OPJ_INT32 c;
   1034             for (c = 0; c < nb_cols; c++, tiledp_col++) {
   1035                 opj_idwt3_v_cas1(dwt->mem, sn, len, tiledp_col, stride);
   1036             }
   1037             return;
   1038         }
   1039     }
   1040 #endif
   1041 }
   1042 
   1043 
   1044 /* <summary>                             */
   1045 /* Forward 9-7 wavelet transform in 1-D. */
   1046 /* </summary>                            */
   1047 static void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_SIZE_T a_count,
   1048                                   OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas)
   1049 {
   1050     OPJ_INT32 i;
   1051     if (!cas) {
   1052         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
   1053             for (i = 0; i < dn; i++) {
   1054                 OPJ_D(i) -= opj_int_fix_mul(OPJ_S_(i) + OPJ_S_(i + 1), 12993);
   1055             }
   1056             for (i = 0; i < sn; i++) {
   1057                 OPJ_S(i) -= opj_int_fix_mul(OPJ_D_(i - 1) + OPJ_D_(i), 434);
   1058             }
   1059             for (i = 0; i < dn; i++) {
   1060                 OPJ_D(i) += opj_int_fix_mul(OPJ_S_(i) + OPJ_S_(i + 1), 7233);
   1061             }
   1062             for (i = 0; i < sn; i++) {
   1063                 OPJ_S(i) += opj_int_fix_mul(OPJ_D_(i - 1) + OPJ_D_(i), 3633);
   1064             }
   1065             for (i = 0; i < dn; i++) {
   1066                 OPJ_D(i) = opj_int_fix_mul(OPJ_D(i), 5038);    /*5038 */
   1067             }
   1068             for (i = 0; i < sn; i++) {
   1069                 OPJ_S(i) = opj_int_fix_mul(OPJ_S(i), 6659);    /*6660 */
   1070             }
   1071         }
   1072     } else {
   1073         if ((sn > 0) || (dn > 1)) { /* NEW :  CASE ONE ELEMENT */
   1074             for (i = 0; i < dn; i++) {
   1075                 OPJ_S(i) -= opj_int_fix_mul(OPJ_DD_(i) + OPJ_DD_(i - 1), 12993);
   1076             }
   1077             for (i = 0; i < sn; i++) {
   1078                 OPJ_D(i) -= opj_int_fix_mul(OPJ_SS_(i) + OPJ_SS_(i + 1), 434);
   1079             }
   1080             for (i = 0; i < dn; i++) {
   1081                 OPJ_S(i) += opj_int_fix_mul(OPJ_DD_(i) + OPJ_DD_(i - 1), 7233);
   1082             }
   1083             for (i = 0; i < sn; i++) {
   1084                 OPJ_D(i) += opj_int_fix_mul(OPJ_SS_(i) + OPJ_SS_(i + 1), 3633);
   1085             }
   1086             for (i = 0; i < dn; i++) {
   1087                 OPJ_S(i) = opj_int_fix_mul(OPJ_S(i), 5038);    /*5038 */
   1088             }
   1089             for (i = 0; i < sn; i++) {
   1090                 OPJ_D(i) = opj_int_fix_mul(OPJ_D(i), 6659);    /*6660 */
   1091             }
   1092         }
   1093     }
   1094 }
   1095 
   1096 static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
   1097                                     opj_stepsize_t *bandno_stepsize)
   1098 {
   1099     OPJ_INT32 p, n;
   1100     p = opj_int_floorlog2(stepsize) - 13;
   1101     n = 11 - opj_int_floorlog2(stepsize);
   1102     bandno_stepsize->mant = (n < 0 ? stepsize >> -n : stepsize << n) & 0x7ff;
   1103     bandno_stepsize->expn = numbps - p;
   1104 }
   1105 
   1106 /*
   1107 ==========================================================
   1108    DWT interface
   1109 ==========================================================
   1110 */
   1111 
   1112 
   1113 /* <summary>                            */
   1114 /* Forward 5-3 wavelet transform in 2-D. */
   1115 /* </summary>                           */
   1116 static INLINE OPJ_BOOL opj_dwt_encode_procedure(const opj_tcd_tilecomp_t * tilec,
   1117         void(*p_function)(OPJ_INT32 *, OPJ_SIZE_T, OPJ_INT32, OPJ_INT32, OPJ_INT32))
   1118 {
   1119     OPJ_INT32 i, j, k;
   1120     OPJ_INT32 *a = 00;
   1121     OPJ_INT32 *aj = 00;
   1122     OPJ_INT32 *bj = 00;
   1123     OPJ_INT32 w, l;
   1124 
   1125     OPJ_INT32 rw;           /* width of the resolution level computed   */
   1126     OPJ_INT32 rh;           /* height of the resolution level computed  */
   1127     OPJ_SIZE_T l_data_count;
   1128     OPJ_SIZE_T l_data_size;
   1129 
   1130     opj_tcd_resolution_t * l_cur_res = 0;
   1131     opj_tcd_resolution_t * l_last_res = 0;
   1132 
   1133     w = tilec->x1 - tilec->x0;
   1134     l = (OPJ_INT32)tilec->numresolutions - 1;
   1135     a = tilec->data;
   1136 
   1137     l_cur_res = tilec->resolutions + l;
   1138     l_last_res = l_cur_res - 1;
   1139 
   1140     l_data_count = opj_dwt_max_resolution(tilec->resolutions, tilec->numresolutions);
   1141     /* overflow check */
   1142     if (l_data_count > (SIZE_MAX / sizeof(OPJ_INT32))) {
   1143         /* FIXME event manager error callback */
   1144         return OPJ_FALSE;
   1145     }
   1146     l_data_size = l_data_count * sizeof(OPJ_INT32);
   1147     bj = (OPJ_INT32*)opj_malloc(l_data_size);
   1148     /* l_data_size is equal to 0 when numresolutions == 1 but bj is not used */
   1149     /* in that case, so do not error out */
   1150     if (l_data_size != 0 && ! bj) {
   1151         return OPJ_FALSE;
   1152     }
   1153     i = l;
   1154 
   1155     while (i--) {
   1156         OPJ_INT32 rw1;      /* width of the resolution level once lower than computed one                                       */
   1157         OPJ_INT32 rh1;      /* height of the resolution level once lower than computed one                                      */
   1158         OPJ_INT32 cas_col;  /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
   1159         OPJ_INT32 cas_row;  /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering   */
   1160         OPJ_INT32 dn, sn;
   1161 
   1162         rw  = l_cur_res->x1 - l_cur_res->x0;
   1163         rh  = l_cur_res->y1 - l_cur_res->y0;
   1164         rw1 = l_last_res->x1 - l_last_res->x0;
   1165         rh1 = l_last_res->y1 - l_last_res->y0;
   1166 
   1167         cas_row = l_cur_res->x0 & 1;
   1168         cas_col = l_cur_res->y0 & 1;
   1169 
   1170         sn = rh1;
   1171         dn = rh - rh1;
   1172         for (j = 0; j < rw; ++j) {
   1173             aj = a + j;
   1174             for (k = 0; k < rh; ++k) {
   1175                 bj[k] = aj[k * w];
   1176             }
   1177 
   1178             (*p_function) (bj, l_data_count, dn, sn, cas_col);
   1179 
   1180             opj_dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col);
   1181         }
   1182 
   1183         sn = rw1;
   1184         dn = rw - rw1;
   1185 
   1186         for (j = 0; j < rh; j++) {
   1187             aj = a + j * w;
   1188             for (k = 0; k < rw; k++) {
   1189                 bj[k] = aj[k];
   1190             }
   1191             (*p_function) (bj, l_data_count, dn, sn, cas_row);
   1192             opj_dwt_deinterleave_h(bj, aj, dn, sn, cas_row);
   1193         }
   1194 
   1195         l_cur_res = l_last_res;
   1196 
   1197         --l_last_res;
   1198     }
   1199 
   1200     opj_free(bj);
   1201     return OPJ_TRUE;
   1202 }
   1203 
   1204 /* Forward 5-3 wavelet transform in 2-D. */
   1205 /* </summary>                           */
   1206 OPJ_BOOL opj_dwt_encode(opj_tcd_tilecomp_t * tilec)
   1207 {
   1208     return opj_dwt_encode_procedure(tilec, opj_dwt_encode_1);
   1209 }
   1210 
   1211 /* <summary>                            */
   1212 /* Inverse 5-3 wavelet transform in 2-D. */
   1213 /* </summary>                           */
   1214 OPJ_BOOL opj_dwt_decode(opj_tcd_t *p_tcd, opj_tcd_tilecomp_t* tilec,
   1215                         OPJ_UINT32 numres)
   1216 {
   1217     if (p_tcd->whole_tile_decoding) {
   1218         return opj_dwt_decode_tile(p_tcd->thread_pool, tilec, numres);
   1219     } else {
   1220         return opj_dwt_decode_partial_tile(tilec, numres);
   1221     }
   1222 }
   1223 
   1224 
   1225 /* <summary>                          */
   1226 /* Get gain of 5-3 wavelet transform. */
   1227 /* </summary>                         */
   1228 OPJ_UINT32 opj_dwt_getgain(OPJ_UINT32 orient)
   1229 {
   1230     if (orient == 0) {
   1231         return 0;
   1232     }
   1233     if (orient == 1 || orient == 2) {
   1234         return 1;
   1235     }
   1236     return 2;
   1237 }
   1238 
   1239 /* <summary>                */
   1240 /* Get norm of 5-3 wavelet. */
   1241 /* </summary>               */
   1242 OPJ_FLOAT64 opj_dwt_getnorm(OPJ_UINT32 level, OPJ_UINT32 orient)
   1243 {
   1244     /* FIXME ! This is just a band-aid to avoid a buffer overflow */
   1245     /* but the array should really be extended up to 33 resolution levels */
   1246     /* See https://github.com/uclouvain/openjpeg/issues/493 */
   1247     if (orient == 0 && level >= 10) {
   1248         level = 9;
   1249     } else if (orient > 0 && level >= 9) {
   1250         level = 8;
   1251     }
   1252     return opj_dwt_norms[orient][level];
   1253 }
   1254 
   1255 /* <summary>                             */
   1256 /* Forward 9-7 wavelet transform in 2-D. */
   1257 /* </summary>                            */
   1258 OPJ_BOOL opj_dwt_encode_real(opj_tcd_tilecomp_t * tilec)
   1259 {
   1260     return opj_dwt_encode_procedure(tilec, opj_dwt_encode_1_real);
   1261 }
   1262 
   1263 /* <summary>                          */
   1264 /* Get gain of 9-7 wavelet transform. */
   1265 /* </summary>                         */
   1266 OPJ_UINT32 opj_dwt_getgain_real(OPJ_UINT32 orient)
   1267 {
   1268     (void)orient;
   1269     return 0;
   1270 }
   1271 
   1272 /* <summary>                */
   1273 /* Get norm of 9-7 wavelet. */
   1274 /* </summary>               */
   1275 OPJ_FLOAT64 opj_dwt_getnorm_real(OPJ_UINT32 level, OPJ_UINT32 orient)
   1276 {
   1277     /* FIXME ! This is just a band-aid to avoid a buffer overflow */
   1278     /* but the array should really be extended up to 33 resolution levels */
   1279     /* See https://github.com/uclouvain/openjpeg/issues/493 */
   1280     if (orient == 0 && level >= 10) {
   1281         level = 9;
   1282     } else if (orient > 0 && level >= 9) {
   1283         level = 8;
   1284     }
   1285     return opj_dwt_norms_real[orient][level];
   1286 }
   1287 
   1288 void opj_dwt_calc_explicit_stepsizes(opj_tccp_t * tccp, OPJ_UINT32 prec)
   1289 {
   1290     OPJ_UINT32 numbands, bandno;
   1291     numbands = 3 * tccp->numresolutions - 2;
   1292     for (bandno = 0; bandno < numbands; bandno++) {
   1293         OPJ_FLOAT64 stepsize;
   1294         OPJ_UINT32 resno, level, orient, gain;
   1295 
   1296         resno = (bandno == 0) ? 0 : ((bandno - 1) / 3 + 1);
   1297         orient = (bandno == 0) ? 0 : ((bandno - 1) % 3 + 1);
   1298         level = tccp->numresolutions - 1 - resno;
   1299         gain = (tccp->qmfbid == 0) ? 0 : ((orient == 0) ? 0 : (((orient == 1) ||
   1300                                           (orient == 2)) ? 1 : 2));
   1301         if (tccp->qntsty == J2K_CCP_QNTSTY_NOQNT) {
   1302             stepsize = 1.0;
   1303         } else {
   1304             OPJ_FLOAT64 norm = opj_dwt_norms_real[orient][level];
   1305             stepsize = (1 << (gain)) / norm;
   1306         }
   1307         opj_dwt_encode_stepsize((OPJ_INT32) floor(stepsize * 8192.0),
   1308                                 (OPJ_INT32)(prec + gain), &tccp->stepsizes[bandno]);
   1309     }
   1310 }
   1311 
   1312 /* <summary>                             */
   1313 /* Determine maximum computed resolution level for inverse wavelet transform */
   1314 /* </summary>                            */
   1315 static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r,
   1316         OPJ_UINT32 i)
   1317 {
   1318     OPJ_UINT32 mr   = 0;
   1319     OPJ_UINT32 w;
   1320     while (--i) {
   1321         ++r;
   1322         if (mr < (w = (OPJ_UINT32)(r->x1 - r->x0))) {
   1323             mr = w ;
   1324         }
   1325         if (mr < (w = (OPJ_UINT32)(r->y1 - r->y0))) {
   1326             mr = w ;
   1327         }
   1328     }
   1329     return mr ;
   1330 }
   1331 
   1332 typedef struct {
   1333     opj_dwt_t h;
   1334     OPJ_UINT32 rw;
   1335     OPJ_UINT32 w;
   1336     OPJ_INT32 * OPJ_RESTRICT tiledp;
   1337     OPJ_UINT32 min_j;
   1338     OPJ_UINT32 max_j;
   1339 } opj_dwd_decode_h_job_t;
   1340 
   1341 static void opj_dwt_decode_h_func(void* user_data, opj_tls_t* tls)
   1342 {
   1343     OPJ_UINT32 j;
   1344     opj_dwd_decode_h_job_t* job;
   1345     (void)tls;
   1346 
   1347     job = (opj_dwd_decode_h_job_t*)user_data;
   1348     for (j = job->min_j; j < job->max_j; j++) {
   1349         opj_idwt53_h(&job->h, &job->tiledp[j * job->w]);
   1350     }
   1351 
   1352     opj_aligned_free(job->h.mem);
   1353     opj_free(job);
   1354 }
   1355 
   1356 typedef struct {
   1357     opj_dwt_t v;
   1358     OPJ_UINT32 rh;
   1359     OPJ_UINT32 w;
   1360     OPJ_INT32 * OPJ_RESTRICT tiledp;
   1361     OPJ_UINT32 min_j;
   1362     OPJ_UINT32 max_j;
   1363 } opj_dwd_decode_v_job_t;
   1364 
   1365 static void opj_dwt_decode_v_func(void* user_data, opj_tls_t* tls)
   1366 {
   1367     OPJ_UINT32 j;
   1368     opj_dwd_decode_v_job_t* job;
   1369     (void)tls;
   1370 
   1371     job = (opj_dwd_decode_v_job_t*)user_data;
   1372     for (j = job->min_j; j + PARALLEL_COLS_53 <= job->max_j;
   1373             j += PARALLEL_COLS_53) {
   1374         opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w,
   1375                      PARALLEL_COLS_53);
   1376     }
   1377     if (j < job->max_j)
   1378         opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w,
   1379                      (OPJ_INT32)(job->max_j - j));
   1380 
   1381     opj_aligned_free(job->v.mem);
   1382     opj_free(job);
   1383 }
   1384 
   1385 
   1386 /* <summary>                            */
   1387 /* Inverse wavelet transform in 2-D.    */
   1388 /* </summary>                           */
   1389 static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
   1390         const opj_tcd_tilecomp_t* tilec, OPJ_UINT32 numres)
   1391 {
   1392     opj_dwt_t h;
   1393     opj_dwt_t v;
   1394 
   1395     opj_tcd_resolution_t* tr = tilec->resolutions;
   1396 
   1397     OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 -
   1398                                  tr->x0);  /* width of the resolution level computed */
   1399     OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
   1400                                  tr->y0);  /* height of the resolution level computed */
   1401 
   1402     OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions -
   1403                                                                1].x1 -
   1404                                 tilec->resolutions[tilec->minimum_num_resolutions - 1].x0);
   1405     OPJ_SIZE_T h_mem_size;
   1406     int num_threads;
   1407 
   1408     if (numres == 1U) {
   1409         return OPJ_TRUE;
   1410     }
   1411     num_threads = opj_thread_pool_get_thread_count(tp);
   1412     h.mem_count = opj_dwt_max_resolution(tr, numres);
   1413     /* overflow check */
   1414     if (h.mem_count > (SIZE_MAX / PARALLEL_COLS_53 / sizeof(OPJ_INT32))) {
   1415         /* FIXME event manager error callback */
   1416         return OPJ_FALSE;
   1417     }
   1418     /* We need PARALLEL_COLS_53 times the height of the array, */
   1419     /* since for the vertical pass */
   1420     /* we process PARALLEL_COLS_53 columns at a time */
   1421     h_mem_size = h.mem_count * PARALLEL_COLS_53 * sizeof(OPJ_INT32);
   1422     h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
   1423     if (! h.mem) {
   1424         /* FIXME event manager error callback */
   1425         return OPJ_FALSE;
   1426     }
   1427 
   1428     v.mem_count = h.mem_count;
   1429     v.mem = h.mem;
   1430 
   1431     while (--numres) {
   1432         OPJ_INT32 * OPJ_RESTRICT tiledp = tilec->data;
   1433         OPJ_UINT32 j;
   1434 
   1435         ++tr;
   1436         h.sn = (OPJ_INT32)rw;
   1437         v.sn = (OPJ_INT32)rh;
   1438 
   1439         rw = (OPJ_UINT32)(tr->x1 - tr->x0);
   1440         rh = (OPJ_UINT32)(tr->y1 - tr->y0);
   1441 
   1442         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
   1443         h.cas = tr->x0 % 2;
   1444 
   1445         if (num_threads <= 1 || rh <= 1) {
   1446             for (j = 0; j < rh; ++j) {
   1447                 opj_idwt53_h(&h, &tiledp[(OPJ_SIZE_T)j * w]);
   1448             }
   1449         } else {
   1450             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
   1451             OPJ_UINT32 step_j;
   1452 
   1453             if (rh < num_jobs) {
   1454                 num_jobs = rh;
   1455             }
   1456             step_j = (rh / num_jobs);
   1457 
   1458             for (j = 0; j < num_jobs; j++) {
   1459                 opj_dwd_decode_h_job_t* job;
   1460 
   1461                 job = (opj_dwd_decode_h_job_t*) opj_malloc(sizeof(opj_dwd_decode_h_job_t));
   1462                 if (!job) {
   1463                     /* It would be nice to fallback to single thread case, but */
   1464                     /* unfortunately some jobs may be launched and have modified */
   1465                     /* tiledp, so it is not practical to recover from that error */
   1466                     /* FIXME event manager error callback */
   1467                     opj_thread_pool_wait_completion(tp, 0);
   1468                     opj_aligned_free(h.mem);
   1469                     return OPJ_FALSE;
   1470                 }
   1471                 job->h = h;
   1472                 job->rw = rw;
   1473                 job->w = w;
   1474                 job->tiledp = tiledp;
   1475                 job->min_j = j * step_j;
   1476                 job->max_j = (j + 1U) * step_j; /* this can overflow */
   1477                 if (j == (num_jobs - 1U)) {  /* this will take care of the overflow */
   1478                     job->max_j = rh;
   1479                 }
   1480                 job->h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
   1481                 if (!job->h.mem) {
   1482                     /* FIXME event manager error callback */
   1483                     opj_thread_pool_wait_completion(tp, 0);
   1484                     opj_free(job);
   1485                     opj_aligned_free(h.mem);
   1486                     return OPJ_FALSE;
   1487                 }
   1488                 opj_thread_pool_submit_job(tp, opj_dwt_decode_h_func, job);
   1489             }
   1490             opj_thread_pool_wait_completion(tp, 0);
   1491         }
   1492 
   1493         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
   1494         v.cas = tr->y0 % 2;
   1495 
   1496         if (num_threads <= 1 || rw <= 1) {
   1497             for (j = 0; j + PARALLEL_COLS_53 <= rw;
   1498                     j += PARALLEL_COLS_53) {
   1499                 opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, PARALLEL_COLS_53);
   1500             }
   1501             if (j < rw) {
   1502                 opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, (OPJ_INT32)(rw - j));
   1503             }
   1504         } else {
   1505             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
   1506             OPJ_UINT32 step_j;
   1507 
   1508             if (rw < num_jobs) {
   1509                 num_jobs = rw;
   1510             }
   1511             step_j = (rw / num_jobs);
   1512 
   1513             for (j = 0; j < num_jobs; j++) {
   1514                 opj_dwd_decode_v_job_t* job;
   1515 
   1516                 job = (opj_dwd_decode_v_job_t*) opj_malloc(sizeof(opj_dwd_decode_v_job_t));
   1517                 if (!job) {
   1518                     /* It would be nice to fallback to single thread case, but */
   1519                     /* unfortunately some jobs may be launched and have modified */
   1520                     /* tiledp, so it is not practical to recover from that error */
   1521                     /* FIXME event manager error callback */
   1522                     opj_thread_pool_wait_completion(tp, 0);
   1523                     opj_aligned_free(v.mem);
   1524                     return OPJ_FALSE;
   1525                 }
   1526                 job->v = v;
   1527                 job->rh = rh;
   1528                 job->w = w;
   1529                 job->tiledp = tiledp;
   1530                 job->min_j = j * step_j;
   1531                 job->max_j = (j + 1U) * step_j; /* this can overflow */
   1532                 if (j == (num_jobs - 1U)) {  /* this will take care of the overflow */
   1533                     job->max_j = rw;
   1534                 }
   1535                 job->v.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
   1536                 if (!job->v.mem) {
   1537                     /* FIXME event manager error callback */
   1538                     opj_thread_pool_wait_completion(tp, 0);
   1539                     opj_free(job);
   1540                     opj_aligned_free(v.mem);
   1541                     return OPJ_FALSE;
   1542                 }
   1543                 opj_thread_pool_submit_job(tp, opj_dwt_decode_v_func, job);
   1544             }
   1545             opj_thread_pool_wait_completion(tp, 0);
   1546         }
   1547     }
   1548     opj_aligned_free(h.mem);
   1549     return OPJ_TRUE;
   1550 }
   1551 
   1552 static void opj_dwt_interleave_partial_h(OPJ_INT32 *dest,
   1553         OPJ_INT32 cas,
   1554         opj_sparse_array_int32_t* sa,
   1555         OPJ_UINT32 sa_line,
   1556         OPJ_UINT32 sn,
   1557         OPJ_UINT32 win_l_x0,
   1558         OPJ_UINT32 win_l_x1,
   1559         OPJ_UINT32 win_h_x0,
   1560         OPJ_UINT32 win_h_x1)
   1561 {
   1562     OPJ_BOOL ret;
   1563     ret = opj_sparse_array_int32_read(sa,
   1564                                       win_l_x0, sa_line,
   1565                                       win_l_x1, sa_line + 1,
   1566                                       dest + cas + 2 * win_l_x0,
   1567                                       2, 0, OPJ_TRUE);
   1568     assert(ret);
   1569     ret = opj_sparse_array_int32_read(sa,
   1570                                       sn + win_h_x0, sa_line,
   1571                                       sn + win_h_x1, sa_line + 1,
   1572                                       dest + 1 - cas + 2 * win_h_x0,
   1573                                       2, 0, OPJ_TRUE);
   1574     assert(ret);
   1575     OPJ_UNUSED(ret);
   1576 }
   1577 
   1578 
   1579 static void opj_dwt_interleave_partial_v(OPJ_INT32 *dest,
   1580         OPJ_INT32 cas,
   1581         opj_sparse_array_int32_t* sa,
   1582         OPJ_UINT32 sa_col,
   1583         OPJ_UINT32 nb_cols,
   1584         OPJ_UINT32 sn,
   1585         OPJ_UINT32 win_l_y0,
   1586         OPJ_UINT32 win_l_y1,
   1587         OPJ_UINT32 win_h_y0,
   1588         OPJ_UINT32 win_h_y1)
   1589 {
   1590     OPJ_BOOL ret;
   1591     ret  = opj_sparse_array_int32_read(sa,
   1592                                        sa_col, win_l_y0,
   1593                                        sa_col + nb_cols, win_l_y1,
   1594                                        dest + cas * 4 + 2 * 4 * win_l_y0,
   1595                                        1, 2 * 4, OPJ_TRUE);
   1596     assert(ret);
   1597     ret = opj_sparse_array_int32_read(sa,
   1598                                       sa_col, sn + win_h_y0,
   1599                                       sa_col + nb_cols, sn + win_h_y1,
   1600                                       dest + (1 - cas) * 4 + 2 * 4 * win_h_y0,
   1601                                       1, 2 * 4, OPJ_TRUE);
   1602     assert(ret);
   1603     OPJ_UNUSED(ret);
   1604 }
   1605 
   1606 static void opj_dwt_decode_partial_1(OPJ_INT32 *a, OPJ_SIZE_T a_count,
   1607                                      OPJ_INT32 dn, OPJ_INT32 sn,
   1608                                      OPJ_INT32 cas,
   1609                                      OPJ_INT32 win_l_x0,
   1610                                      OPJ_INT32 win_l_x1,
   1611                                      OPJ_INT32 win_h_x0,
   1612                                      OPJ_INT32 win_h_x1)
   1613 {
   1614     OPJ_INT32 i;
   1615 
   1616     if (!cas) {
   1617         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
   1618 
   1619             /* Naive version is :
   1620             for (i = win_l_x0; i < i_max; i++) {
   1621                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
   1622             }
   1623             for (i = win_h_x0; i < win_h_x1; i++) {
   1624                 OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
   1625             }
   1626             but the compiler doesn't manage to unroll it to avoid bound
   1627             checking in OPJ_S_ and OPJ_D_ macros
   1628             */
   1629 
   1630             i = win_l_x0;
   1631             if (i < win_l_x1) {
   1632                 OPJ_INT32 i_max;
   1633 
   1634                 /* Left-most case */
   1635                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
   1636                 i ++;
   1637 
   1638                 i_max = win_l_x1;
   1639                 if (i_max > dn) {
   1640                     i_max = dn;
   1641                 }
   1642                 for (; i < i_max; i++) {
   1643                     /* No bound checking */
   1644                     OPJ_S(i) -= (OPJ_D(i - 1) + OPJ_D(i) + 2) >> 2;
   1645                 }
   1646                 for (; i < win_l_x1; i++) {
   1647                     /* Right-most case */
   1648                     OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
   1649                 }
   1650             }
   1651 
   1652             i = win_h_x0;
   1653             if (i < win_h_x1) {
   1654                 OPJ_INT32 i_max = win_h_x1;
   1655                 if (i_max >= sn) {
   1656                     i_max = sn - 1;
   1657                 }
   1658                 for (; i < i_max; i++) {
   1659                     /* No bound checking */
   1660                     OPJ_D(i) += (OPJ_S(i) + OPJ_S(i + 1)) >> 1;
   1661                 }
   1662                 for (; i < win_h_x1; i++) {
   1663                     /* Right-most case */
   1664                     OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
   1665                 }
   1666             }
   1667         }
   1668     } else {
   1669         if (!sn  && dn == 1) {        /* NEW :  CASE ONE ELEMENT */
   1670             OPJ_S(0) /= 2;
   1671         } else {
   1672             for (i = win_l_x0; i < win_l_x1; i++) {
   1673                 OPJ_D(i) -= (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2;
   1674             }
   1675             for (i = win_h_x0; i < win_h_x1; i++) {
   1676                 OPJ_S(i) += (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1;
   1677             }
   1678         }
   1679     }
   1680 }
   1681 
   1682 #define OPJ_S_off(i,off) a[(OPJ_UINT32)(i)*2*4+off]
   1683 #define OPJ_D_off(i,off) a[(1+(OPJ_UINT32)(i)*2)*4+off]
   1684 #define OPJ_S__off(i,off) ((i)<0?OPJ_S_off(0,off):((i)>=sn?OPJ_S_off(sn-1,off):OPJ_S_off(i,off)))
   1685 #define OPJ_D__off(i,off) ((i)<0?OPJ_D_off(0,off):((i)>=dn?OPJ_D_off(dn-1,off):OPJ_D_off(i,off)))
   1686 #define OPJ_SS__off(i,off) ((i)<0?OPJ_S_off(0,off):((i)>=dn?OPJ_S_off(dn-1,off):OPJ_S_off(i,off)))
   1687 #define OPJ_DD__off(i,off) ((i)<0?OPJ_D_off(0,off):((i)>=sn?OPJ_D_off(sn-1,off):OPJ_D_off(i,off)))
   1688 
   1689 static void opj_dwt_decode_partial_1_parallel(OPJ_INT32 *a,
   1690         OPJ_UINT32 nb_cols,
   1691         OPJ_INT32 dn, OPJ_INT32 sn,
   1692         OPJ_INT32 cas,
   1693         OPJ_INT32 win_l_x0,
   1694         OPJ_INT32 win_l_x1,
   1695         OPJ_INT32 win_h_x0,
   1696         OPJ_INT32 win_h_x1)
   1697 {
   1698     OPJ_INT32 i;
   1699     OPJ_UINT32 off;
   1700 
   1701     (void)nb_cols;
   1702 
   1703     if (!cas) {
   1704         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
   1705 
   1706             /* Naive version is :
   1707             for (i = win_l_x0; i < i_max; i++) {
   1708                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
   1709             }
   1710             for (i = win_h_x0; i < win_h_x1; i++) {
   1711                 OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
   1712             }
   1713             but the compiler doesn't manage to unroll it to avoid bound
   1714             checking in OPJ_S_ and OPJ_D_ macros
   1715             */
   1716 
   1717             i = win_l_x0;
   1718             if (i < win_l_x1) {
   1719                 OPJ_INT32 i_max;
   1720 
   1721                 /* Left-most case */
   1722                 for (off = 0; off < 4; off++) {
   1723                     OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2;
   1724                 }
   1725                 i ++;
   1726 
   1727                 i_max = win_l_x1;
   1728                 if (i_max > dn) {
   1729                     i_max = dn;
   1730                 }
   1731 
   1732 #ifdef __SSE2__
   1733                 if (i + 1 < i_max) {
   1734                     const __m128i two = _mm_set1_epi32(2);
   1735                     __m128i Dm1 = _mm_load_si128((__m128i * const)(a + 4 + (i - 1) * 8));
   1736                     for (; i + 1 < i_max; i += 2) {
   1737                         /* No bound checking */
   1738                         __m128i S = _mm_load_si128((__m128i * const)(a + i * 8));
   1739                         __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8));
   1740                         __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8));
   1741                         __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8));
   1742                         S = _mm_sub_epi32(S,
   1743                                           _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(Dm1, D), two), 2));
   1744                         S1 = _mm_sub_epi32(S1,
   1745                                            _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(D, D1), two), 2));
   1746                         _mm_store_si128((__m128i*)(a + i * 8), S);
   1747                         _mm_store_si128((__m128i*)(a + (i + 1) * 8), S1);
   1748                         Dm1 = D1;
   1749                     }
   1750                 }
   1751 #endif
   1752 
   1753                 for (; i < i_max; i++) {
   1754                     /* No bound checking */
   1755                     for (off = 0; off < 4; off++) {
   1756                         OPJ_S_off(i, off) -= (OPJ_D_off(i - 1, off) + OPJ_D_off(i, off) + 2) >> 2;
   1757                     }
   1758                 }
   1759                 for (; i < win_l_x1; i++) {
   1760                     /* Right-most case */
   1761                     for (off = 0; off < 4; off++) {
   1762                         OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2;
   1763                     }
   1764                 }
   1765             }
   1766 
   1767             i = win_h_x0;
   1768             if (i < win_h_x1) {
   1769                 OPJ_INT32 i_max = win_h_x1;
   1770                 if (i_max >= sn) {
   1771                     i_max = sn - 1;
   1772                 }
   1773 
   1774 #ifdef __SSE2__
   1775                 if (i + 1 < i_max) {
   1776                     __m128i S =  _mm_load_si128((__m128i * const)(a + i * 8));
   1777                     for (; i + 1 < i_max; i += 2) {
   1778                         /* No bound checking */
   1779                         __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8));
   1780                         __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8));
   1781                         __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8));
   1782                         __m128i S2 = _mm_load_si128((__m128i * const)(a + (i + 2) * 8));
   1783                         D = _mm_add_epi32(D, _mm_srai_epi32(_mm_add_epi32(S, S1), 1));
   1784                         D1 = _mm_add_epi32(D1, _mm_srai_epi32(_mm_add_epi32(S1, S2), 1));
   1785                         _mm_store_si128((__m128i*)(a + 4 + i * 8), D);
   1786                         _mm_store_si128((__m128i*)(a + 4 + (i + 1) * 8), D1);
   1787                         S = S2;
   1788                     }
   1789                 }
   1790 #endif
   1791 
   1792                 for (; i < i_max; i++) {
   1793                     /* No bound checking */
   1794                     for (off = 0; off < 4; off++) {
   1795                         OPJ_D_off(i, off) += (OPJ_S_off(i, off) + OPJ_S_off(i + 1, off)) >> 1;
   1796                     }
   1797                 }
   1798                 for (; i < win_h_x1; i++) {
   1799                     /* Right-most case */
   1800                     for (off = 0; off < 4; off++) {
   1801                         OPJ_D_off(i, off) += (OPJ_S__off(i, off) + OPJ_S__off(i + 1, off)) >> 1;
   1802                     }
   1803                 }
   1804             }
   1805         }
   1806     } else {
   1807         if (!sn  && dn == 1) {        /* NEW :  CASE ONE ELEMENT */
   1808             for (off = 0; off < 4; off++) {
   1809                 OPJ_S_off(0, off) /= 2;
   1810             }
   1811         } else {
   1812             for (i = win_l_x0; i < win_l_x1; i++) {
   1813                 for (off = 0; off < 4; off++) {
   1814                     OPJ_D_off(i, off) -= (OPJ_SS__off(i, off) + OPJ_SS__off(i + 1, off) + 2) >> 2;
   1815                 }
   1816             }
   1817             for (i = win_h_x0; i < win_h_x1; i++) {
   1818                 for (off = 0; off < 4; off++) {
   1819                     OPJ_S_off(i, off) += (OPJ_DD__off(i, off) + OPJ_DD__off(i - 1, off)) >> 1;
   1820                 }
   1821             }
   1822         }
   1823     }
   1824 }
   1825 
   1826 static void opj_dwt_get_band_coordinates(opj_tcd_tilecomp_t* tilec,
   1827         OPJ_UINT32 resno,
   1828         OPJ_UINT32 bandno,
   1829         OPJ_UINT32 tcx0,
   1830         OPJ_UINT32 tcy0,
   1831         OPJ_UINT32 tcx1,
   1832         OPJ_UINT32 tcy1,
   1833         OPJ_UINT32* tbx0,
   1834         OPJ_UINT32* tby0,
   1835         OPJ_UINT32* tbx1,
   1836         OPJ_UINT32* tby1)
   1837 {
   1838     /* Compute number of decomposition for this band. See table F-1 */
   1839     OPJ_UINT32 nb = (resno == 0) ?
   1840                     tilec->numresolutions - 1 :
   1841                     tilec->numresolutions - resno;
   1842     /* Map above tile-based coordinates to sub-band-based coordinates per */
   1843     /* equation B-15 of the standard */
   1844     OPJ_UINT32 x0b = bandno & 1;
   1845     OPJ_UINT32 y0b = bandno >> 1;
   1846     if (tbx0) {
   1847         *tbx0 = (nb == 0) ? tcx0 :
   1848                 (tcx0 <= (1U << (nb - 1)) * x0b) ? 0 :
   1849                 opj_uint_ceildivpow2(tcx0 - (1U << (nb - 1)) * x0b, nb);
   1850     }
   1851     if (tby0) {
   1852         *tby0 = (nb == 0) ? tcy0 :
   1853                 (tcy0 <= (1U << (nb - 1)) * y0b) ? 0 :
   1854                 opj_uint_ceildivpow2(tcy0 - (1U << (nb - 1)) * y0b, nb);
   1855     }
   1856     if (tbx1) {
   1857         *tbx1 = (nb == 0) ? tcx1 :
   1858                 (tcx1 <= (1U << (nb - 1)) * x0b) ? 0 :
   1859                 opj_uint_ceildivpow2(tcx1 - (1U << (nb - 1)) * x0b, nb);
   1860     }
   1861     if (tby1) {
   1862         *tby1 = (nb == 0) ? tcy1 :
   1863                 (tcy1 <= (1U << (nb - 1)) * y0b) ? 0 :
   1864                 opj_uint_ceildivpow2(tcy1 - (1U << (nb - 1)) * y0b, nb);
   1865     }
   1866 }
   1867 
   1868 static void opj_dwt_segment_grow(OPJ_UINT32 filter_width,
   1869                                  OPJ_UINT32 max_size,
   1870                                  OPJ_UINT32* start,
   1871                                  OPJ_UINT32* end)
   1872 {
   1873     *start = opj_uint_subs(*start, filter_width);
   1874     *end = opj_uint_adds(*end, filter_width);
   1875     *end = opj_uint_min(*end, max_size);
   1876 }
   1877 
   1878 
   1879 static opj_sparse_array_int32_t* opj_dwt_init_sparse_array(
   1880     opj_tcd_tilecomp_t* tilec,
   1881     OPJ_UINT32 numres)
   1882 {
   1883     opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]);
   1884     OPJ_UINT32 w = (OPJ_UINT32)(tr_max->x1 - tr_max->x0);
   1885     OPJ_UINT32 h = (OPJ_UINT32)(tr_max->y1 - tr_max->y0);
   1886     OPJ_UINT32 resno, bandno, precno, cblkno;
   1887     opj_sparse_array_int32_t* sa = opj_sparse_array_int32_create(
   1888                                        w, h, opj_uint_min(w, 64), opj_uint_min(h, 64));
   1889     if (sa == NULL) {
   1890         return NULL;
   1891     }
   1892 
   1893     for (resno = 0; resno < numres; ++resno) {
   1894         opj_tcd_resolution_t* res = &tilec->resolutions[resno];
   1895 
   1896         for (bandno = 0; bandno < res->numbands; ++bandno) {
   1897             opj_tcd_band_t* band = &res->bands[bandno];
   1898 
   1899             for (precno = 0; precno < res->pw * res->ph; ++precno) {
   1900                 opj_tcd_precinct_t* precinct = &band->precincts[precno];
   1901                 for (cblkno = 0; cblkno < precinct->cw * precinct->ch; ++cblkno) {
   1902                     opj_tcd_cblk_dec_t* cblk = &precinct->cblks.dec[cblkno];
   1903                     if (cblk->decoded_data != NULL) {
   1904                         OPJ_UINT32 x = (OPJ_UINT32)(cblk->x0 - band->x0);
   1905                         OPJ_UINT32 y = (OPJ_UINT32)(cblk->y0 - band->y0);
   1906                         OPJ_UINT32 cblk_w = (OPJ_UINT32)(cblk->x1 - cblk->x0);
   1907                         OPJ_UINT32 cblk_h = (OPJ_UINT32)(cblk->y1 - cblk->y0);
   1908 
   1909                         if (band->bandno & 1) {
   1910                             opj_tcd_resolution_t* pres = &tilec->resolutions[resno - 1];
   1911                             x += (OPJ_UINT32)(pres->x1 - pres->x0);
   1912                         }
   1913                         if (band->bandno & 2) {
   1914                             opj_tcd_resolution_t* pres = &tilec->resolutions[resno - 1];
   1915                             y += (OPJ_UINT32)(pres->y1 - pres->y0);
   1916                         }
   1917 
   1918                         if (!opj_sparse_array_int32_write(sa, x, y,
   1919                                                           x + cblk_w, y + cblk_h,
   1920                                                           cblk->decoded_data,
   1921                                                           1, cblk_w, OPJ_TRUE)) {
   1922                             opj_sparse_array_int32_free(sa);
   1923                             return NULL;
   1924                         }
   1925                     }
   1926                 }
   1927             }
   1928         }
   1929     }
   1930 
   1931     return sa;
   1932 }
   1933 
   1934 
   1935 static OPJ_BOOL opj_dwt_decode_partial_tile(
   1936     opj_tcd_tilecomp_t* tilec,
   1937     OPJ_UINT32 numres)
   1938 {
   1939     opj_sparse_array_int32_t* sa;
   1940     opj_dwt_t h;
   1941     opj_dwt_t v;
   1942     OPJ_UINT32 resno;
   1943     /* This value matches the maximum left/right extension given in tables */
   1944     /* F.2 and F.3 of the standard. */
   1945     const OPJ_UINT32 filter_width = 2U;
   1946 
   1947     opj_tcd_resolution_t* tr = tilec->resolutions;
   1948     opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]);
   1949 
   1950     OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 -
   1951                                  tr->x0);  /* width of the resolution level computed */
   1952     OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
   1953                                  tr->y0);  /* height of the resolution level computed */
   1954 
   1955     OPJ_SIZE_T h_mem_size;
   1956 
   1957     /* Compute the intersection of the area of interest, expressed in tile coordinates */
   1958     /* with the tile coordinates */
   1959     OPJ_UINT32 win_tcx0 = tilec->win_x0;
   1960     OPJ_UINT32 win_tcy0 = tilec->win_y0;
   1961     OPJ_UINT32 win_tcx1 = tilec->win_x1;
   1962     OPJ_UINT32 win_tcy1 = tilec->win_y1;
   1963 
   1964     if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) {
   1965         return OPJ_TRUE;
   1966     }
   1967 
   1968     sa = opj_dwt_init_sparse_array(tilec, numres);
   1969     if (sa == NULL) {
   1970         return OPJ_FALSE;
   1971     }
   1972 
   1973     if (numres == 1U) {
   1974         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
   1975                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
   1976                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
   1977                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
   1978                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
   1979                        tilec->data_win,
   1980                        1, tr_max->win_x1 - tr_max->win_x0,
   1981                        OPJ_TRUE);
   1982         assert(ret);
   1983         OPJ_UNUSED(ret);
   1984         opj_sparse_array_int32_free(sa);
   1985         return OPJ_TRUE;
   1986     }
   1987     h.mem_count = opj_dwt_max_resolution(tr, numres);
   1988     /* overflow check */
   1989     /* in vertical pass, we process 4 columns at a time */
   1990     if (h.mem_count > (SIZE_MAX / (4 * sizeof(OPJ_INT32)))) {
   1991         /* FIXME event manager error callback */
   1992         opj_sparse_array_int32_free(sa);
   1993         return OPJ_FALSE;
   1994     }
   1995 
   1996     h_mem_size = h.mem_count * 4 * sizeof(OPJ_INT32);
   1997     h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
   1998     if (! h.mem) {
   1999         /* FIXME event manager error callback */
   2000         opj_sparse_array_int32_free(sa);
   2001         return OPJ_FALSE;
   2002     }
   2003 
   2004     v.mem_count = h.mem_count;
   2005     v.mem = h.mem;
   2006 
   2007     for (resno = 1; resno < numres; resno ++) {
   2008         OPJ_UINT32 i, j;
   2009         /* Window of interest subband-based coordinates */
   2010         OPJ_UINT32 win_ll_x0, win_ll_y0, win_ll_x1, win_ll_y1;
   2011         OPJ_UINT32 win_hl_x0, win_hl_x1;
   2012         OPJ_UINT32 win_lh_y0, win_lh_y1;
   2013         /* Window of interest tile-resolution-based coordinates */
   2014         OPJ_UINT32 win_tr_x0, win_tr_x1, win_tr_y0, win_tr_y1;
   2015         /* Tile-resolution subband-based coordinates */
   2016         OPJ_UINT32 tr_ll_x0, tr_ll_y0, tr_hl_x0, tr_lh_y0;
   2017 
   2018         ++tr;
   2019 
   2020         h.sn = (OPJ_INT32)rw;
   2021         v.sn = (OPJ_INT32)rh;
   2022 
   2023         rw = (OPJ_UINT32)(tr->x1 - tr->x0);
   2024         rh = (OPJ_UINT32)(tr->y1 - tr->y0);
   2025 
   2026         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
   2027         h.cas = tr->x0 % 2;
   2028 
   2029         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
   2030         v.cas = tr->y0 % 2;
   2031 
   2032         /* Get the subband coordinates for the window of interest */
   2033         /* LL band */
   2034         opj_dwt_get_band_coordinates(tilec, resno, 0,
   2035                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
   2036                                      &win_ll_x0, &win_ll_y0,
   2037                                      &win_ll_x1, &win_ll_y1);
   2038 
   2039         /* HL band */
   2040         opj_dwt_get_band_coordinates(tilec, resno, 1,
   2041                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
   2042                                      &win_hl_x0, NULL, &win_hl_x1, NULL);
   2043 
   2044         /* LH band */
   2045         opj_dwt_get_band_coordinates(tilec, resno, 2,
   2046                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
   2047                                      NULL, &win_lh_y0, NULL, &win_lh_y1);
   2048 
   2049         /* Beware: band index for non-LL0 resolution are 0=HL, 1=LH and 2=HH */
   2050         tr_ll_x0 = (OPJ_UINT32)tr->bands[1].x0;
   2051         tr_ll_y0 = (OPJ_UINT32)tr->bands[0].y0;
   2052         tr_hl_x0 = (OPJ_UINT32)tr->bands[0].x0;
   2053         tr_lh_y0 = (OPJ_UINT32)tr->bands[1].y0;
   2054 
   2055         /* Substract the origin of the bands for this tile, to the subwindow */
   2056         /* of interest band coordinates, so as to get them relative to the */
   2057         /* tile */
   2058         win_ll_x0 = opj_uint_subs(win_ll_x0, tr_ll_x0);
   2059         win_ll_y0 = opj_uint_subs(win_ll_y0, tr_ll_y0);
   2060         win_ll_x1 = opj_uint_subs(win_ll_x1, tr_ll_x0);
   2061         win_ll_y1 = opj_uint_subs(win_ll_y1, tr_ll_y0);
   2062         win_hl_x0 = opj_uint_subs(win_hl_x0, tr_hl_x0);
   2063         win_hl_x1 = opj_uint_subs(win_hl_x1, tr_hl_x0);
   2064         win_lh_y0 = opj_uint_subs(win_lh_y0, tr_lh_y0);
   2065         win_lh_y1 = opj_uint_subs(win_lh_y1, tr_lh_y0);
   2066 
   2067         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.sn, &win_ll_x0, &win_ll_x1);
   2068         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.dn, &win_hl_x0, &win_hl_x1);
   2069 
   2070         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.sn, &win_ll_y0, &win_ll_y1);
   2071         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.dn, &win_lh_y0, &win_lh_y1);
   2072 
   2073         /* Compute the tile-resolution-based coordinates for the window of interest */
   2074         if (h.cas == 0) {
   2075             win_tr_x0 = opj_uint_min(2 * win_ll_x0, 2 * win_hl_x0 + 1);
   2076             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_ll_x1, 2 * win_hl_x1 + 1), rw);
   2077         } else {
   2078             win_tr_x0 = opj_uint_min(2 * win_hl_x0, 2 * win_ll_x0 + 1);
   2079             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_hl_x1, 2 * win_ll_x1 + 1), rw);
   2080         }
   2081 
   2082         if (v.cas == 0) {
   2083             win_tr_y0 = opj_uint_min(2 * win_ll_y0, 2 * win_lh_y0 + 1);
   2084             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_ll_y1, 2 * win_lh_y1 + 1), rh);
   2085         } else {
   2086             win_tr_y0 = opj_uint_min(2 * win_lh_y0, 2 * win_ll_y0 + 1);
   2087             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_lh_y1, 2 * win_ll_y1 + 1), rh);
   2088         }
   2089 
   2090         for (j = 0; j < rh; ++j) {
   2091             if ((j >= win_ll_y0 && j < win_ll_y1) ||
   2092                     (j >= win_lh_y0 + (OPJ_UINT32)v.sn && j < win_lh_y1 + (OPJ_UINT32)v.sn)) {
   2093 
   2094                 /* Avoids dwt.c:1584:44 (in opj_dwt_decode_partial_1): runtime error: */
   2095                 /* signed integer overflow: -1094795586 + -1094795586 cannot be represented in type 'int' */
   2096                 /* on opj_decompress -i  ../../openjpeg/MAPA.jp2 -o out.tif -d 0,0,256,256 */
   2097                 /* This is less extreme than memsetting the whole buffer to 0 */
   2098                 /* although we could potentially do better with better handling of edge conditions */
   2099                 if (win_tr_x1 >= 1 && win_tr_x1 < rw) {
   2100                     h.mem[win_tr_x1 - 1] = 0;
   2101                 }
   2102                 if (win_tr_x1 < rw) {
   2103                     h.mem[win_tr_x1] = 0;
   2104                 }
   2105 
   2106                 opj_dwt_interleave_partial_h(h.mem,
   2107                                              h.cas,
   2108                                              sa,
   2109                                              j,
   2110                                              (OPJ_UINT32)h.sn,
   2111                                              win_ll_x0,
   2112                                              win_ll_x1,
   2113                                              win_hl_x0,
   2114                                              win_hl_x1);
   2115                 opj_dwt_decode_partial_1(h.mem, h.mem_count, h.dn, h.sn, h.cas,
   2116                                          (OPJ_INT32)win_ll_x0,
   2117                                          (OPJ_INT32)win_ll_x1,
   2118                                          (OPJ_INT32)win_hl_x0,
   2119                                          (OPJ_INT32)win_hl_x1);
   2120                 if (!opj_sparse_array_int32_write(sa,
   2121                                                   win_tr_x0, j,
   2122                                                   win_tr_x1, j + 1,
   2123                                                   h.mem + win_tr_x0,
   2124                                                   1, 0, OPJ_TRUE)) {
   2125                     /* FIXME event manager error callback */
   2126                     opj_sparse_array_int32_free(sa);
   2127                     opj_aligned_free(h.mem);
   2128                     return OPJ_FALSE;
   2129                 }
   2130             }
   2131         }
   2132 
   2133         for (i = win_tr_x0; i < win_tr_x1;) {
   2134             OPJ_UINT32 nb_cols = opj_uint_min(4U, win_tr_x1 - i);
   2135             opj_dwt_interleave_partial_v(v.mem,
   2136                                          v.cas,
   2137                                          sa,
   2138                                          i,
   2139                                          nb_cols,
   2140                                          (OPJ_UINT32)v.sn,
   2141                                          win_ll_y0,
   2142                                          win_ll_y1,
   2143                                          win_lh_y0,
   2144                                          win_lh_y1);
   2145             opj_dwt_decode_partial_1_parallel(v.mem, nb_cols, v.dn, v.sn, v.cas,
   2146                                               (OPJ_INT32)win_ll_y0,
   2147                                               (OPJ_INT32)win_ll_y1,
   2148                                               (OPJ_INT32)win_lh_y0,
   2149                                               (OPJ_INT32)win_lh_y1);
   2150             if (!opj_sparse_array_int32_write(sa,
   2151                                               i, win_tr_y0,
   2152                                               i + nb_cols, win_tr_y1,
   2153                                               v.mem + 4 * win_tr_y0,
   2154                                               1, 4, OPJ_TRUE)) {
   2155                 /* FIXME event manager error callback */
   2156                 opj_sparse_array_int32_free(sa);
   2157                 opj_aligned_free(h.mem);
   2158                 return OPJ_FALSE;
   2159             }
   2160 
   2161             i += nb_cols;
   2162         }
   2163     }
   2164     opj_aligned_free(h.mem);
   2165 
   2166     {
   2167         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
   2168                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
   2169                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
   2170                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
   2171                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
   2172                        tilec->data_win,
   2173                        1, tr_max->win_x1 - tr_max->win_x0,
   2174                        OPJ_TRUE);
   2175         assert(ret);
   2176         OPJ_UNUSED(ret);
   2177     }
   2178     opj_sparse_array_int32_free(sa);
   2179     return OPJ_TRUE;
   2180 }
   2181 
   2182 static void opj_v4dwt_interleave_h(opj_v4dwt_t* OPJ_RESTRICT dwt,
   2183                                    OPJ_FLOAT32* OPJ_RESTRICT a,
   2184                                    OPJ_UINT32 width,
   2185                                    OPJ_UINT32 remaining_height)
   2186 {
   2187     OPJ_FLOAT32* OPJ_RESTRICT bi = (OPJ_FLOAT32*)(dwt->wavelet + dwt->cas);
   2188     OPJ_UINT32 i, k;
   2189     OPJ_UINT32 x0 = dwt->win_l_x0;
   2190     OPJ_UINT32 x1 = dwt->win_l_x1;
   2191 
   2192     for (k = 0; k < 2; ++k) {
   2193         if (remaining_height >= 4 && ((OPJ_SIZE_T) a & 0x0f) == 0 &&
   2194                 ((OPJ_SIZE_T) bi & 0x0f) == 0 && (width & 0x0f) == 0) {
   2195             /* Fast code path */
   2196             for (i = x0; i < x1; ++i) {
   2197                 OPJ_UINT32 j = i;
   2198                 bi[i * 8    ] = a[j];
   2199                 j += width;
   2200                 bi[i * 8 + 1] = a[j];
   2201                 j += width;
   2202                 bi[i * 8 + 2] = a[j];
   2203                 j += width;
   2204                 bi[i * 8 + 3] = a[j];
   2205             }
   2206         } else {
   2207             /* Slow code path */
   2208             for (i = x0; i < x1; ++i) {
   2209                 OPJ_UINT32 j = i;
   2210                 bi[i * 8    ] = a[j];
   2211                 j += width;
   2212                 if (remaining_height == 1) {
   2213                     continue;
   2214                 }
   2215                 bi[i * 8 + 1] = a[j];
   2216                 j += width;
   2217                 if (remaining_height == 2) {
   2218                     continue;
   2219                 }
   2220                 bi[i * 8 + 2] = a[j];
   2221                 j += width;
   2222                 if (remaining_height == 3) {
   2223                     continue;
   2224                 }
   2225                 bi[i * 8 + 3] = a[j]; /* This one*/
   2226             }
   2227         }
   2228 
   2229         bi = (OPJ_FLOAT32*)(dwt->wavelet + 1 - dwt->cas);
   2230         a += dwt->sn;
   2231         x0 = dwt->win_h_x0;
   2232         x1 = dwt->win_h_x1;
   2233     }
   2234 }
   2235 
   2236 static void opj_v4dwt_interleave_partial_h(opj_v4dwt_t* dwt,
   2237         opj_sparse_array_int32_t* sa,
   2238         OPJ_UINT32 sa_line,
   2239         OPJ_UINT32 remaining_height)
   2240 {
   2241     OPJ_UINT32 i;
   2242     for (i = 0; i < remaining_height; i++) {
   2243         OPJ_BOOL ret;
   2244         ret = opj_sparse_array_int32_read(sa,
   2245                                           dwt->win_l_x0, sa_line + i,
   2246                                           dwt->win_l_x1, sa_line + i + 1,
   2247                                           /* Nasty cast from float* to int32* */
   2248                                           (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0) + i,
   2249                                           8, 0, OPJ_TRUE);
   2250         assert(ret);
   2251         ret = opj_sparse_array_int32_read(sa,
   2252                                           (OPJ_UINT32)dwt->sn + dwt->win_h_x0, sa_line + i,
   2253                                           (OPJ_UINT32)dwt->sn + dwt->win_h_x1, sa_line + i + 1,
   2254                                           /* Nasty cast from float* to int32* */
   2255                                           (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0) + i,
   2256                                           8, 0, OPJ_TRUE);
   2257         assert(ret);
   2258         OPJ_UNUSED(ret);
   2259     }
   2260 }
   2261 
   2262 static void opj_v4dwt_interleave_v(opj_v4dwt_t* OPJ_RESTRICT dwt,
   2263                                    OPJ_FLOAT32* OPJ_RESTRICT a,
   2264                                    OPJ_UINT32 width,
   2265                                    OPJ_UINT32 nb_elts_read)
   2266 {
   2267     opj_v4_t* OPJ_RESTRICT bi = dwt->wavelet + dwt->cas;
   2268     OPJ_UINT32 i;
   2269 
   2270     for (i = dwt->win_l_x0; i < dwt->win_l_x1; ++i) {
   2271         memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width],
   2272                (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32));
   2273     }
   2274 
   2275     a += (OPJ_UINT32)dwt->sn * (OPJ_SIZE_T)width;
   2276     bi = dwt->wavelet + 1 - dwt->cas;
   2277 
   2278     for (i = dwt->win_h_x0; i < dwt->win_h_x1; ++i) {
   2279         memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width],
   2280                (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32));
   2281     }
   2282 }
   2283 
   2284 static void opj_v4dwt_interleave_partial_v(opj_v4dwt_t* OPJ_RESTRICT dwt,
   2285         opj_sparse_array_int32_t* sa,
   2286         OPJ_UINT32 sa_col,
   2287         OPJ_UINT32 nb_elts_read)
   2288 {
   2289     OPJ_BOOL ret;
   2290     ret = opj_sparse_array_int32_read(sa,
   2291                                       sa_col, dwt->win_l_x0,
   2292                                       sa_col + nb_elts_read, dwt->win_l_x1,
   2293                                       (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0),
   2294                                       1, 8, OPJ_TRUE);
   2295     assert(ret);
   2296     ret = opj_sparse_array_int32_read(sa,
   2297                                       sa_col, (OPJ_UINT32)dwt->sn + dwt->win_h_x0,
   2298                                       sa_col + nb_elts_read, (OPJ_UINT32)dwt->sn + dwt->win_h_x1,
   2299                                       (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0),
   2300                                       1, 8, OPJ_TRUE);
   2301     assert(ret);
   2302     OPJ_UNUSED(ret);
   2303 }
   2304 
   2305 #ifdef __SSE__
   2306 
   2307 static void opj_v4dwt_decode_step1_sse(opj_v4_t* w,
   2308                                        OPJ_UINT32 start,
   2309                                        OPJ_UINT32 end,
   2310                                        const __m128 c)
   2311 {
   2312     __m128* OPJ_RESTRICT vw = (__m128*) w;
   2313     OPJ_UINT32 i;
   2314     /* 4x unrolled loop */
   2315     vw += 2 * start;
   2316     for (i = start; i + 3 < end; i += 4, vw += 8) {
   2317         __m128 xmm0 = _mm_mul_ps(vw[0], c);
   2318         __m128 xmm2 = _mm_mul_ps(vw[2], c);
   2319         __m128 xmm4 = _mm_mul_ps(vw[4], c);
   2320         __m128 xmm6 = _mm_mul_ps(vw[6], c);
   2321         vw[0] = xmm0;
   2322         vw[2] = xmm2;
   2323         vw[4] = xmm4;
   2324         vw[6] = xmm6;
   2325     }
   2326     for (; i < end; ++i, vw += 2) {
   2327         vw[0] = _mm_mul_ps(vw[0], c);
   2328     }
   2329 }
   2330 
   2331 static void opj_v4dwt_decode_step2_sse(opj_v4_t* l, opj_v4_t* w,
   2332                                        OPJ_UINT32 start,
   2333                                        OPJ_UINT32 end,
   2334                                        OPJ_UINT32 m,
   2335                                        __m128 c)
   2336 {
   2337     __m128* OPJ_RESTRICT vl = (__m128*) l;
   2338     __m128* OPJ_RESTRICT vw = (__m128*) w;
   2339     OPJ_UINT32 i;
   2340     OPJ_UINT32 imax = opj_uint_min(end, m);
   2341     __m128 tmp1, tmp2, tmp3;
   2342     if (start == 0) {
   2343         tmp1 = vl[0];
   2344     } else {
   2345         vw += start * 2;
   2346         tmp1 = vw[-3];
   2347     }
   2348 
   2349     i = start;
   2350 
   2351     /* 4x loop unrolling */
   2352     for (; i + 3 < imax; i += 4) {
   2353         __m128 tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
   2354         tmp2 = vw[-1];
   2355         tmp3 = vw[ 0];
   2356         tmp4 = vw[ 1];
   2357         tmp5 = vw[ 2];
   2358         tmp6 = vw[ 3];
   2359         tmp7 = vw[ 4];
   2360         tmp8 = vw[ 5];
   2361         tmp9 = vw[ 6];
   2362         vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c));
   2363         vw[ 1] = _mm_add_ps(tmp4, _mm_mul_ps(_mm_add_ps(tmp3, tmp5), c));
   2364         vw[ 3] = _mm_add_ps(tmp6, _mm_mul_ps(_mm_add_ps(tmp5, tmp7), c));
   2365         vw[ 5] = _mm_add_ps(tmp8, _mm_mul_ps(_mm_add_ps(tmp7, tmp9), c));
   2366         tmp1 = tmp9;
   2367         vw += 8;
   2368     }
   2369 
   2370     for (; i < imax; ++i) {
   2371         tmp2 = vw[-1];
   2372         tmp3 = vw[ 0];
   2373         vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c));
   2374         tmp1 = tmp3;
   2375         vw += 2;
   2376     }
   2377     if (m < end) {
   2378         assert(m + 1 == end);
   2379         c = _mm_add_ps(c, c);
   2380         c = _mm_mul_ps(c, vw[-2]);
   2381         vw[-1] = _mm_add_ps(vw[-1], c);
   2382     }
   2383 }
   2384 
   2385 #else
   2386 
   2387 static void opj_v4dwt_decode_step1(opj_v4_t* w,
   2388                                    OPJ_UINT32 start,
   2389                                    OPJ_UINT32 end,
   2390                                    const OPJ_FLOAT32 c)
   2391 {
   2392     OPJ_FLOAT32* OPJ_RESTRICT fw = (OPJ_FLOAT32*) w;
   2393     OPJ_UINT32 i;
   2394     for (i = start; i < end; ++i) {
   2395         OPJ_FLOAT32 tmp1 = fw[i * 8    ];
   2396         OPJ_FLOAT32 tmp2 = fw[i * 8 + 1];
   2397         OPJ_FLOAT32 tmp3 = fw[i * 8 + 2];
   2398         OPJ_FLOAT32 tmp4 = fw[i * 8 + 3];
   2399         fw[i * 8    ] = tmp1 * c;
   2400         fw[i * 8 + 1] = tmp2 * c;
   2401         fw[i * 8 + 2] = tmp3 * c;
   2402         fw[i * 8 + 3] = tmp4 * c;
   2403     }
   2404 }
   2405 
   2406 static void opj_v4dwt_decode_step2(opj_v4_t* l, opj_v4_t* w,
   2407                                    OPJ_UINT32 start,
   2408                                    OPJ_UINT32 end,
   2409                                    OPJ_UINT32 m,
   2410                                    OPJ_FLOAT32 c)
   2411 {
   2412     OPJ_FLOAT32* fl = (OPJ_FLOAT32*) l;
   2413     OPJ_FLOAT32* fw = (OPJ_FLOAT32*) w;
   2414     OPJ_UINT32 i;
   2415     OPJ_UINT32 imax = opj_uint_min(end, m);
   2416     if (start > 0) {
   2417         fw += 8 * start;
   2418         fl = fw - 8;
   2419     }
   2420     for (i = start; i < imax; ++i) {
   2421         OPJ_FLOAT32 tmp1_1 = fl[0];
   2422         OPJ_FLOAT32 tmp1_2 = fl[1];
   2423         OPJ_FLOAT32 tmp1_3 = fl[2];
   2424         OPJ_FLOAT32 tmp1_4 = fl[3];
   2425         OPJ_FLOAT32 tmp2_1 = fw[-4];
   2426         OPJ_FLOAT32 tmp2_2 = fw[-3];
   2427         OPJ_FLOAT32 tmp2_3 = fw[-2];
   2428         OPJ_FLOAT32 tmp2_4 = fw[-1];
   2429         OPJ_FLOAT32 tmp3_1 = fw[0];
   2430         OPJ_FLOAT32 tmp3_2 = fw[1];
   2431         OPJ_FLOAT32 tmp3_3 = fw[2];
   2432         OPJ_FLOAT32 tmp3_4 = fw[3];
   2433         fw[-4] = tmp2_1 + ((tmp1_1 + tmp3_1) * c);
   2434         fw[-3] = tmp2_2 + ((tmp1_2 + tmp3_2) * c);
   2435         fw[-2] = tmp2_3 + ((tmp1_3 + tmp3_3) * c);
   2436         fw[-1] = tmp2_4 + ((tmp1_4 + tmp3_4) * c);
   2437         fl = fw;
   2438         fw += 8;
   2439     }
   2440     if (m < end) {
   2441         assert(m + 1 == end);
   2442         c += c;
   2443         fw[-4] = fw[-4] + fl[0] * c;
   2444         fw[-3] = fw[-3] + fl[1] * c;
   2445         fw[-2] = fw[-2] + fl[2] * c;
   2446         fw[-1] = fw[-1] + fl[3] * c;
   2447     }
   2448 }
   2449 
   2450 #endif
   2451 
   2452 /* <summary>                             */
   2453 /* Inverse 9-7 wavelet transform in 1-D. */
   2454 /* </summary>                            */
   2455 static void opj_v4dwt_decode(opj_v4dwt_t* OPJ_RESTRICT dwt)
   2456 {
   2457     OPJ_INT32 a, b;
   2458     if (dwt->cas == 0) {
   2459         if (!((dwt->dn > 0) || (dwt->sn > 1))) {
   2460             return;
   2461         }
   2462         a = 0;
   2463         b = 1;
   2464     } else {
   2465         if (!((dwt->sn > 0) || (dwt->dn > 1))) {
   2466             return;
   2467         }
   2468         a = 1;
   2469         b = 0;
   2470     }
   2471 #ifdef __SSE__
   2472     opj_v4dwt_decode_step1_sse(dwt->wavelet + a, dwt->win_l_x0, dwt->win_l_x1,
   2473                                _mm_set1_ps(opj_K));
   2474     opj_v4dwt_decode_step1_sse(dwt->wavelet + b, dwt->win_h_x0, dwt->win_h_x1,
   2475                                _mm_set1_ps(opj_c13318));
   2476     opj_v4dwt_decode_step2_sse(dwt->wavelet + b, dwt->wavelet + a + 1,
   2477                                dwt->win_l_x0, dwt->win_l_x1,
   2478                                (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
   2479                                _mm_set1_ps(opj_dwt_delta));
   2480     opj_v4dwt_decode_step2_sse(dwt->wavelet + a, dwt->wavelet + b + 1,
   2481                                dwt->win_h_x0, dwt->win_h_x1,
   2482                                (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
   2483                                _mm_set1_ps(opj_dwt_gamma));
   2484     opj_v4dwt_decode_step2_sse(dwt->wavelet + b, dwt->wavelet + a + 1,
   2485                                dwt->win_l_x0, dwt->win_l_x1,
   2486                                (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
   2487                                _mm_set1_ps(opj_dwt_beta));
   2488     opj_v4dwt_decode_step2_sse(dwt->wavelet + a, dwt->wavelet + b + 1,
   2489                                dwt->win_h_x0, dwt->win_h_x1,
   2490                                (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
   2491                                _mm_set1_ps(opj_dwt_alpha));
   2492 #else
   2493     opj_v4dwt_decode_step1(dwt->wavelet + a, dwt->win_l_x0, dwt->win_l_x1,
   2494                            opj_K);
   2495     opj_v4dwt_decode_step1(dwt->wavelet + b, dwt->win_h_x0, dwt->win_h_x1,
   2496                            opj_c13318);
   2497     opj_v4dwt_decode_step2(dwt->wavelet + b, dwt->wavelet + a + 1,
   2498                            dwt->win_l_x0, dwt->win_l_x1,
   2499                            (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
   2500                            opj_dwt_delta);
   2501     opj_v4dwt_decode_step2(dwt->wavelet + a, dwt->wavelet + b + 1,
   2502                            dwt->win_h_x0, dwt->win_h_x1,
   2503                            (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
   2504                            opj_dwt_gamma);
   2505     opj_v4dwt_decode_step2(dwt->wavelet + b, dwt->wavelet + a + 1,
   2506                            dwt->win_l_x0, dwt->win_l_x1,
   2507                            (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
   2508                            opj_dwt_beta);
   2509     opj_v4dwt_decode_step2(dwt->wavelet + a, dwt->wavelet + b + 1,
   2510                            dwt->win_h_x0, dwt->win_h_x1,
   2511                            (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
   2512                            opj_dwt_alpha);
   2513 #endif
   2514 }
   2515 
   2516 
   2517 /* <summary>                             */
   2518 /* Inverse 9-7 wavelet transform in 2-D. */
   2519 /* </summary>                            */
   2520 static
   2521 OPJ_BOOL opj_dwt_decode_tile_97(opj_tcd_tilecomp_t* OPJ_RESTRICT tilec,
   2522                                 OPJ_UINT32 numres)
   2523 {
   2524     opj_v4dwt_t h;
   2525     opj_v4dwt_t v;
   2526 
   2527     opj_tcd_resolution_t* res = tilec->resolutions;
   2528 
   2529     OPJ_UINT32 rw = (OPJ_UINT32)(res->x1 -
   2530                                  res->x0);    /* width of the resolution level computed */
   2531     OPJ_UINT32 rh = (OPJ_UINT32)(res->y1 -
   2532                                  res->y0);    /* height of the resolution level computed */
   2533 
   2534     OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions -
   2535                                                                1].x1 -
   2536                                 tilec->resolutions[tilec->minimum_num_resolutions - 1].x0);
   2537 
   2538     OPJ_SIZE_T l_data_size;
   2539 
   2540     l_data_size = opj_dwt_max_resolution(res, numres);
   2541     /* overflow check */
   2542     if (l_data_size > (SIZE_MAX - 5U)) {
   2543         /* FIXME event manager error callback */
   2544         return OPJ_FALSE;
   2545     }
   2546     l_data_size += 5U;
   2547     /* overflow check */
   2548     if (l_data_size > (SIZE_MAX / sizeof(opj_v4_t))) {
   2549         /* FIXME event manager error callback */
   2550         return OPJ_FALSE;
   2551     }
   2552     h.wavelet = (opj_v4_t*) opj_aligned_malloc(l_data_size * sizeof(opj_v4_t));
   2553     if (!h.wavelet) {
   2554         /* FIXME event manager error callback */
   2555         return OPJ_FALSE;
   2556     }
   2557     v.wavelet = h.wavelet;
   2558 
   2559     while (--numres) {
   2560         OPJ_FLOAT32 * OPJ_RESTRICT aj = (OPJ_FLOAT32*) tilec->data;
   2561         OPJ_UINT32 j;
   2562 
   2563         h.sn = (OPJ_INT32)rw;
   2564         v.sn = (OPJ_INT32)rh;
   2565 
   2566         ++res;
   2567 
   2568         rw = (OPJ_UINT32)(res->x1 -
   2569                           res->x0);   /* width of the resolution level computed */
   2570         rh = (OPJ_UINT32)(res->y1 -
   2571                           res->y0);   /* height of the resolution level computed */
   2572 
   2573         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
   2574         h.cas = res->x0 % 2;
   2575 
   2576         h.win_l_x0 = 0;
   2577         h.win_l_x1 = (OPJ_UINT32)h.sn;
   2578         h.win_h_x0 = 0;
   2579         h.win_h_x1 = (OPJ_UINT32)h.dn;
   2580         for (j = 0; j + 3 < rh; j += 4) {
   2581             OPJ_UINT32 k;
   2582             opj_v4dwt_interleave_h(&h, aj, w, rh - j);
   2583             opj_v4dwt_decode(&h);
   2584 
   2585             for (k = 0; k < rw; k++) {
   2586                 aj[k      ] = h.wavelet[k].f[0];
   2587                 aj[k + (OPJ_SIZE_T)w  ] = h.wavelet[k].f[1];
   2588                 aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2];
   2589                 aj[k + (OPJ_SIZE_T)w * 3] = h.wavelet[k].f[3];
   2590             }
   2591 
   2592             aj += w * 4;
   2593         }
   2594 
   2595         if (j < rh) {
   2596             OPJ_UINT32 k;
   2597             opj_v4dwt_interleave_h(&h, aj, w, rh - j);
   2598             opj_v4dwt_decode(&h);
   2599             for (k = 0; k < rw; k++) {
   2600                 switch (rh - j) {
   2601                 case 3:
   2602                     aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2];
   2603                 /* FALLTHRU */
   2604                 case 2:
   2605                     aj[k + (OPJ_SIZE_T)w  ] = h.wavelet[k].f[1];
   2606                 /* FALLTHRU */
   2607                 case 1:
   2608                     aj[k] = h.wavelet[k].f[0];
   2609                 }
   2610             }
   2611         }
   2612 
   2613         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
   2614         v.cas = res->y0 % 2;
   2615         v.win_l_x0 = 0;
   2616         v.win_l_x1 = (OPJ_UINT32)v.sn;
   2617         v.win_h_x0 = 0;
   2618         v.win_h_x1 = (OPJ_UINT32)v.dn;
   2619 
   2620         aj = (OPJ_FLOAT32*) tilec->data;
   2621         for (j = rw; j > 3; j -= 4) {
   2622             OPJ_UINT32 k;
   2623 
   2624             opj_v4dwt_interleave_v(&v, aj, w, 4);
   2625             opj_v4dwt_decode(&v);
   2626 
   2627             for (k = 0; k < rh; ++k) {
   2628                 memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k], 4 * sizeof(OPJ_FLOAT32));
   2629             }
   2630             aj += 4;
   2631         }
   2632 
   2633         if (rw & 0x03) {
   2634             OPJ_UINT32 k;
   2635 
   2636             j = rw & 0x03;
   2637 
   2638             opj_v4dwt_interleave_v(&v, aj, w, j);
   2639             opj_v4dwt_decode(&v);
   2640 
   2641             for (k = 0; k < rh; ++k) {
   2642                 memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k],
   2643                        (OPJ_SIZE_T)j * sizeof(OPJ_FLOAT32));
   2644             }
   2645         }
   2646     }
   2647 
   2648     opj_aligned_free(h.wavelet);
   2649     return OPJ_TRUE;
   2650 }
   2651 
   2652 static
   2653 OPJ_BOOL opj_dwt_decode_partial_97(opj_tcd_tilecomp_t* OPJ_RESTRICT tilec,
   2654                                    OPJ_UINT32 numres)
   2655 {
   2656     opj_sparse_array_int32_t* sa;
   2657     opj_v4dwt_t h;
   2658     opj_v4dwt_t v;
   2659     OPJ_UINT32 resno;
   2660     /* This value matches the maximum left/right extension given in tables */
   2661     /* F.2 and F.3 of the standard. Note: in opj_tcd_is_subband_area_of_interest() */
   2662     /* we currently use 3. */
   2663     const OPJ_UINT32 filter_width = 4U;
   2664 
   2665     opj_tcd_resolution_t* tr = tilec->resolutions;
   2666     opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]);
   2667 
   2668     OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 -
   2669                                  tr->x0);    /* width of the resolution level computed */
   2670     OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
   2671                                  tr->y0);    /* height of the resolution level computed */
   2672 
   2673     OPJ_SIZE_T l_data_size;
   2674 
   2675     /* Compute the intersection of the area of interest, expressed in tile coordinates */
   2676     /* with the tile coordinates */
   2677     OPJ_UINT32 win_tcx0 = tilec->win_x0;
   2678     OPJ_UINT32 win_tcy0 = tilec->win_y0;
   2679     OPJ_UINT32 win_tcx1 = tilec->win_x1;
   2680     OPJ_UINT32 win_tcy1 = tilec->win_y1;
   2681 
   2682     if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) {
   2683         return OPJ_TRUE;
   2684     }
   2685 
   2686     sa = opj_dwt_init_sparse_array(tilec, numres);
   2687     if (sa == NULL) {
   2688         return OPJ_FALSE;
   2689     }
   2690 
   2691     if (numres == 1U) {
   2692         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
   2693                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
   2694                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
   2695                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
   2696                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
   2697                        tilec->data_win,
   2698                        1, tr_max->win_x1 - tr_max->win_x0,
   2699                        OPJ_TRUE);
   2700         assert(ret);
   2701         OPJ_UNUSED(ret);
   2702         opj_sparse_array_int32_free(sa);
   2703         return OPJ_TRUE;
   2704     }
   2705 
   2706     l_data_size = opj_dwt_max_resolution(tr, numres);
   2707     /* overflow check */
   2708     if (l_data_size > (SIZE_MAX - 5U)) {
   2709         /* FIXME event manager error callback */
   2710         return OPJ_FALSE;
   2711     }
   2712     l_data_size += 5U;
   2713     /* overflow check */
   2714     if (l_data_size > (SIZE_MAX / sizeof(opj_v4_t))) {
   2715         /* FIXME event manager error callback */
   2716         return OPJ_FALSE;
   2717     }
   2718     h.wavelet = (opj_v4_t*) opj_aligned_malloc(l_data_size * sizeof(opj_v4_t));
   2719     if (!h.wavelet) {
   2720         /* FIXME event manager error callback */
   2721         return OPJ_FALSE;
   2722     }
   2723     v.wavelet = h.wavelet;
   2724 
   2725     for (resno = 1; resno < numres; resno ++) {
   2726         OPJ_UINT32 j;
   2727         /* Window of interest subband-based coordinates */
   2728         OPJ_UINT32 win_ll_x0, win_ll_y0, win_ll_x1, win_ll_y1;
   2729         OPJ_UINT32 win_hl_x0, win_hl_x1;
   2730         OPJ_UINT32 win_lh_y0, win_lh_y1;
   2731         /* Window of interest tile-resolution-based coordinates */
   2732         OPJ_UINT32 win_tr_x0, win_tr_x1, win_tr_y0, win_tr_y1;
   2733         /* Tile-resolution subband-based coordinates */
   2734         OPJ_UINT32 tr_ll_x0, tr_ll_y0, tr_hl_x0, tr_lh_y0;
   2735 
   2736         ++tr;
   2737 
   2738         h.sn = (OPJ_INT32)rw;
   2739         v.sn = (OPJ_INT32)rh;
   2740 
   2741         rw = (OPJ_UINT32)(tr->x1 - tr->x0);
   2742         rh = (OPJ_UINT32)(tr->y1 - tr->y0);
   2743 
   2744         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
   2745         h.cas = tr->x0 % 2;
   2746 
   2747         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
   2748         v.cas = tr->y0 % 2;
   2749 
   2750         /* Get the subband coordinates for the window of interest */
   2751         /* LL band */
   2752         opj_dwt_get_band_coordinates(tilec, resno, 0,
   2753                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
   2754                                      &win_ll_x0, &win_ll_y0,
   2755                                      &win_ll_x1, &win_ll_y1);
   2756 
   2757         /* HL band */
   2758         opj_dwt_get_band_coordinates(tilec, resno, 1,
   2759                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
   2760                                      &win_hl_x0, NULL, &win_hl_x1, NULL);
   2761 
   2762         /* LH band */
   2763         opj_dwt_get_band_coordinates(tilec, resno, 2,
   2764                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
   2765                                      NULL, &win_lh_y0, NULL, &win_lh_y1);
   2766 
   2767         /* Beware: band index for non-LL0 resolution are 0=HL, 1=LH and 2=HH */
   2768         tr_ll_x0 = (OPJ_UINT32)tr->bands[1].x0;
   2769         tr_ll_y0 = (OPJ_UINT32)tr->bands[0].y0;
   2770         tr_hl_x0 = (OPJ_UINT32)tr->bands[0].x0;
   2771         tr_lh_y0 = (OPJ_UINT32)tr->bands[1].y0;
   2772 
   2773         /* Substract the origin of the bands for this tile, to the subwindow */
   2774         /* of interest band coordinates, so as to get them relative to the */
   2775         /* tile */
   2776         win_ll_x0 = opj_uint_subs(win_ll_x0, tr_ll_x0);
   2777         win_ll_y0 = opj_uint_subs(win_ll_y0, tr_ll_y0);
   2778         win_ll_x1 = opj_uint_subs(win_ll_x1, tr_ll_x0);
   2779         win_ll_y1 = opj_uint_subs(win_ll_y1, tr_ll_y0);
   2780         win_hl_x0 = opj_uint_subs(win_hl_x0, tr_hl_x0);
   2781         win_hl_x1 = opj_uint_subs(win_hl_x1, tr_hl_x0);
   2782         win_lh_y0 = opj_uint_subs(win_lh_y0, tr_lh_y0);
   2783         win_lh_y1 = opj_uint_subs(win_lh_y1, tr_lh_y0);
   2784 
   2785         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.sn, &win_ll_x0, &win_ll_x1);
   2786         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.dn, &win_hl_x0, &win_hl_x1);
   2787 
   2788         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.sn, &win_ll_y0, &win_ll_y1);
   2789         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.dn, &win_lh_y0, &win_lh_y1);
   2790 
   2791         /* Compute the tile-resolution-based coordinates for the window of interest */
   2792         if (h.cas == 0) {
   2793             win_tr_x0 = opj_uint_min(2 * win_ll_x0, 2 * win_hl_x0 + 1);
   2794             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_ll_x1, 2 * win_hl_x1 + 1), rw);
   2795         } else {
   2796             win_tr_x0 = opj_uint_min(2 * win_hl_x0, 2 * win_ll_x0 + 1);
   2797             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_hl_x1, 2 * win_ll_x1 + 1), rw);
   2798         }
   2799 
   2800         if (v.cas == 0) {
   2801             win_tr_y0 = opj_uint_min(2 * win_ll_y0, 2 * win_lh_y0 + 1);
   2802             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_ll_y1, 2 * win_lh_y1 + 1), rh);
   2803         } else {
   2804             win_tr_y0 = opj_uint_min(2 * win_lh_y0, 2 * win_ll_y0 + 1);
   2805             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_lh_y1, 2 * win_ll_y1 + 1), rh);
   2806         }
   2807 
   2808         h.win_l_x0 = win_ll_x0;
   2809         h.win_l_x1 = win_ll_x1;
   2810         h.win_h_x0 = win_hl_x0;
   2811         h.win_h_x1 = win_hl_x1;
   2812         for (j = 0; j + 3 < rh; j += 4) {
   2813             if ((j + 3 >= win_ll_y0 && j < win_ll_y1) ||
   2814                     (j + 3 >= win_lh_y0 + (OPJ_UINT32)v.sn &&
   2815                      j < win_lh_y1 + (OPJ_UINT32)v.sn)) {
   2816                 opj_v4dwt_interleave_partial_h(&h, sa, j, opj_uint_min(4U, rh - j));
   2817                 opj_v4dwt_decode(&h);
   2818                 if (!opj_sparse_array_int32_write(sa,
   2819                                                   win_tr_x0, j,
   2820                                                   win_tr_x1, j + 4,
   2821                                                   (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0],
   2822                                                   4, 1, OPJ_TRUE)) {
   2823                     /* FIXME event manager error callback */
   2824                     opj_sparse_array_int32_free(sa);
   2825                     opj_aligned_free(h.wavelet);
   2826                     return OPJ_FALSE;
   2827                 }
   2828             }
   2829         }
   2830 
   2831         if (j < rh &&
   2832                 ((j + 3 >= win_ll_y0 && j < win_ll_y1) ||
   2833                  (j + 3 >= win_lh_y0 + (OPJ_UINT32)v.sn &&
   2834                   j < win_lh_y1 + (OPJ_UINT32)v.sn))) {
   2835             opj_v4dwt_interleave_partial_h(&h, sa, j, rh - j);
   2836             opj_v4dwt_decode(&h);
   2837             if (!opj_sparse_array_int32_write(sa,
   2838                                               win_tr_x0, j,
   2839                                               win_tr_x1, rh,
   2840                                               (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0],
   2841                                               4, 1, OPJ_TRUE)) {
   2842                 /* FIXME event manager error callback */
   2843                 opj_sparse_array_int32_free(sa);
   2844                 opj_aligned_free(h.wavelet);
   2845                 return OPJ_FALSE;
   2846             }
   2847         }
   2848 
   2849         v.win_l_x0 = win_ll_y0;
   2850         v.win_l_x1 = win_ll_y1;
   2851         v.win_h_x0 = win_lh_y0;
   2852         v.win_h_x1 = win_lh_y1;
   2853         for (j = win_tr_x0; j < win_tr_x1; j += 4) {
   2854             OPJ_UINT32 nb_elts = opj_uint_min(4U, win_tr_x1 - j);
   2855 
   2856             opj_v4dwt_interleave_partial_v(&v, sa, j, nb_elts);
   2857             opj_v4dwt_decode(&v);
   2858 
   2859             if (!opj_sparse_array_int32_write(sa,
   2860                                               j, win_tr_y0,
   2861                                               j + nb_elts, win_tr_y1,
   2862                                               (OPJ_INT32*)&h.wavelet[win_tr_y0].f[0],
   2863                                               1, 4, OPJ_TRUE)) {
   2864                 /* FIXME event manager error callback */
   2865                 opj_sparse_array_int32_free(sa);
   2866                 opj_aligned_free(h.wavelet);
   2867                 return OPJ_FALSE;
   2868             }
   2869         }
   2870     }
   2871 
   2872     {
   2873         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
   2874                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
   2875                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
   2876                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
   2877                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
   2878                        tilec->data_win,
   2879                        1, tr_max->win_x1 - tr_max->win_x0,
   2880                        OPJ_TRUE);
   2881         assert(ret);
   2882         OPJ_UNUSED(ret);
   2883     }
   2884     opj_sparse_array_int32_free(sa);
   2885 
   2886     opj_aligned_free(h.wavelet);
   2887     return OPJ_TRUE;
   2888 }
   2889 
   2890 
   2891 OPJ_BOOL opj_dwt_decode_real(opj_tcd_t *p_tcd,
   2892                              opj_tcd_tilecomp_t* OPJ_RESTRICT tilec,
   2893                              OPJ_UINT32 numres)
   2894 {
   2895     if (p_tcd->whole_tile_decoding) {
   2896         return opj_dwt_decode_tile_97(tilec, numres);
   2897     } else {
   2898         return opj_dwt_decode_partial_97(tilec, numres);
   2899     }
   2900 }
   2901