• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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@dzonux.net>
15  * Copyright (c) 2007, Callum Lerwick <seg@haxxed.com>
16  * Copyright (c) 2017, IntoPIX SA <support@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 #define NB_ELTS_V8  8
89 
90 typedef union {
91     OPJ_FLOAT32 f[NB_ELTS_V8];
92 } opj_v8_t;
93 
94 typedef struct v8dwt_local {
95     opj_v8_t*   wavelet ;
96     OPJ_INT32       dn ;  /* number of elements in high pass band */
97     OPJ_INT32       sn ;  /* number of elements in low pass band */
98     OPJ_INT32       cas ; /* 0 = start on even coord, 1 = start on odd coord */
99     OPJ_UINT32      win_l_x0; /* start coord in low pass band */
100     OPJ_UINT32      win_l_x1; /* end coord in low pass band */
101     OPJ_UINT32      win_h_x0; /* start coord in high pass band */
102     OPJ_UINT32      win_h_x1; /* end coord in high pass band */
103 } opj_v8dwt_t ;
104 
105 /* From table F.4 from the standard */
106 static const OPJ_FLOAT32 opj_dwt_alpha =  -1.586134342f;
107 static const OPJ_FLOAT32 opj_dwt_beta  =  -0.052980118f;
108 static const OPJ_FLOAT32 opj_dwt_gamma = 0.882911075f;
109 static const OPJ_FLOAT32 opj_dwt_delta = 0.443506852f;
110 
111 static const OPJ_FLOAT32 opj_K      = 1.230174105f;
112 static const OPJ_FLOAT32 opj_invK   = (OPJ_FLOAT32)(1.0 / 1.230174105);
113 
114 /*@}*/
115 
116 /** @name Local static functions */
117 /*@{*/
118 
119 /**
120 Forward lazy transform (horizontal)
121 */
122 static void opj_dwt_deinterleave_h(const OPJ_INT32 * OPJ_RESTRICT a,
123                                    OPJ_INT32 * OPJ_RESTRICT b,
124                                    OPJ_INT32 dn,
125                                    OPJ_INT32 sn, OPJ_INT32 cas);
126 
127 /**
128 Forward 9-7 wavelet transform in 1-D
129 */
130 static void opj_dwt_encode_1_real(void *a, OPJ_INT32 dn, OPJ_INT32 sn,
131                                   OPJ_INT32 cas);
132 /**
133 Explicit calculation of the Quantization Stepsizes
134 */
135 static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
136                                     opj_stepsize_t *bandno_stepsize);
137 /**
138 Inverse wavelet transform in 2-D.
139 */
140 static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
141                                     const opj_tcd_tilecomp_t* tilec, OPJ_UINT32 i);
142 
143 static OPJ_BOOL opj_dwt_decode_partial_tile(
144     opj_tcd_tilecomp_t* tilec,
145     OPJ_UINT32 numres);
146 
147 /* Forward transform, for the vertical pass, processing cols columns */
148 /* where cols <= NB_ELTS_V8 */
149 /* Where void* is a OPJ_INT32* for 5x3 and OPJ_FLOAT32* for 9x7 */
150 typedef void (*opj_encode_and_deinterleave_v_fnptr_type)(
151     void *array,
152     void *tmp,
153     OPJ_UINT32 height,
154     OPJ_BOOL even,
155     OPJ_UINT32 stride_width,
156     OPJ_UINT32 cols);
157 
158 /* Where void* is a OPJ_INT32* for 5x3 and OPJ_FLOAT32* for 9x7 */
159 typedef void (*opj_encode_and_deinterleave_h_one_row_fnptr_type)(
160     void *row,
161     void *tmp,
162     OPJ_UINT32 width,
163     OPJ_BOOL even);
164 
165 static OPJ_BOOL opj_dwt_encode_procedure(opj_thread_pool_t* tp,
166         opj_tcd_tilecomp_t * tilec,
167         opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v,
168         opj_encode_and_deinterleave_h_one_row_fnptr_type
169         p_encode_and_deinterleave_h_one_row);
170 
171 static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r,
172         OPJ_UINT32 i);
173 
174 /* <summary>                             */
175 /* Inverse 9-7 wavelet transform in 1-D. */
176 /* </summary>                            */
177 
178 /*@}*/
179 
180 /*@}*/
181 
182 #define IDX_S(i) (i)*2
183 #define IDX_D(i) 1 + (i)* 2
184 #define UNDERFLOW_SN(i) ((i) >= sn&&sn>0)
185 #define UNDERFLOW_DN(i) ((i) >= dn&&dn>0)
186 #define OVERFLOW_S(i) (IDX_S(i) >= a_count)
187 #define OVERFLOW_D(i) (IDX_D(i) >= a_count)
188 
189 #define OPJ_S(i) a[IDX_S(i)]
190 #define OPJ_D(i) a[IDX_D(i)]
191 #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)))
192 #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)))
193 /* new */
194 #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)))
195 #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)))
196 
197 /* <summary>                                                              */
198 /* This table contains the norms of the 5-3 wavelets for different bands. */
199 /* </summary>                                                             */
200 /* FIXME! the array should really be extended up to 33 resolution levels */
201 /* See https://github.com/uclouvain/openjpeg/issues/493 */
202 static const OPJ_FLOAT64 opj_dwt_norms[4][10] = {
203     {1.000, 1.500, 2.750, 5.375, 10.68, 21.34, 42.67, 85.33, 170.7, 341.3},
204     {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9},
205     {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9},
206     {.7186, .9218, 1.586, 3.043, 6.019, 12.01, 24.00, 47.97, 95.93}
207 };
208 
209 /* <summary>                                                              */
210 /* This table contains the norms of the 9-7 wavelets for different bands. */
211 /* </summary>                                                             */
212 /* FIXME! the array should really be extended up to 33 resolution levels */
213 /* See https://github.com/uclouvain/openjpeg/issues/493 */
214 static const OPJ_FLOAT64 opj_dwt_norms_real[4][10] = {
215     {1.000, 1.965, 4.177, 8.403, 16.90, 33.84, 67.69, 135.3, 270.6, 540.9},
216     {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0},
217     {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0},
218     {2.080, 3.865, 8.307, 17.18, 34.71, 69.59, 139.3, 278.6, 557.2}
219 };
220 
221 /*
222 ==========================================================
223    local functions
224 ==========================================================
225 */
226 
227 /* <summary>                             */
228 /* Forward lazy transform (horizontal).  */
229 /* </summary>                            */
opj_dwt_deinterleave_h(const OPJ_INT32 * OPJ_RESTRICT a,OPJ_INT32 * OPJ_RESTRICT b,OPJ_INT32 dn,OPJ_INT32 sn,OPJ_INT32 cas)230 static void opj_dwt_deinterleave_h(const OPJ_INT32 * OPJ_RESTRICT a,
231                                    OPJ_INT32 * OPJ_RESTRICT b,
232                                    OPJ_INT32 dn,
233                                    OPJ_INT32 sn, OPJ_INT32 cas)
234 {
235     OPJ_INT32 i;
236     OPJ_INT32 * OPJ_RESTRICT l_dest = b;
237     const OPJ_INT32 * OPJ_RESTRICT l_src = a + cas;
238 
239     for (i = 0; i < sn; ++i) {
240         *l_dest++ = *l_src;
241         l_src += 2;
242     }
243 
244     l_dest = b + sn;
245     l_src = a + 1 - cas;
246 
247     for (i = 0; i < dn; ++i)  {
248         *l_dest++ = *l_src;
249         l_src += 2;
250     }
251 }
252 
253 #ifdef STANDARD_SLOW_VERSION
254 /* <summary>                             */
255 /* Inverse lazy transform (horizontal).  */
256 /* </summary>                            */
opj_dwt_interleave_h(const opj_dwt_t * h,OPJ_INT32 * a)257 static void opj_dwt_interleave_h(const opj_dwt_t* h, OPJ_INT32 *a)
258 {
259     const OPJ_INT32 *ai = a;
260     OPJ_INT32 *bi = h->mem + h->cas;
261     OPJ_INT32  i    = h->sn;
262     while (i--) {
263         *bi = *(ai++);
264         bi += 2;
265     }
266     ai  = a + h->sn;
267     bi  = h->mem + 1 - h->cas;
268     i   = h->dn ;
269     while (i--) {
270         *bi = *(ai++);
271         bi += 2;
272     }
273 }
274 
275 /* <summary>                             */
276 /* Inverse lazy transform (vertical).    */
277 /* </summary>                            */
opj_dwt_interleave_v(const opj_dwt_t * v,OPJ_INT32 * a,OPJ_INT32 x)278 static void opj_dwt_interleave_v(const opj_dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x)
279 {
280     const OPJ_INT32 *ai = a;
281     OPJ_INT32 *bi = v->mem + v->cas;
282     OPJ_INT32  i = v->sn;
283     while (i--) {
284         *bi = *ai;
285         bi += 2;
286         ai += x;
287     }
288     ai = a + (v->sn * (OPJ_SIZE_T)x);
289     bi = v->mem + 1 - v->cas;
290     i = v->dn ;
291     while (i--) {
292         *bi = *ai;
293         bi += 2;
294         ai += x;
295     }
296 }
297 
298 #endif /* STANDARD_SLOW_VERSION */
299 
300 #ifdef STANDARD_SLOW_VERSION
301 /* <summary>                            */
302 /* Inverse 5-3 wavelet transform in 1-D. */
303 /* </summary>                           */
opj_dwt_decode_1_(OPJ_INT32 * a,OPJ_SIZE_T a_count,OPJ_INT32 dn,OPJ_INT32 sn,OPJ_INT32 cas)304 static void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_SIZE_T a_count, OPJ_INT32 dn,
305                               OPJ_INT32 sn, OPJ_INT32 cas)
306 {
307     OPJ_INT32 i;
308 
309     if (!cas) {
310         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
311             for (i = 0; i < sn; i++) {
312                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
313             }
314             for (i = 0; i < dn; i++) {
315                 OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
316             }
317         }
318     } else {
319         if (!sn  && dn == 1) {        /* NEW :  CASE ONE ELEMENT */
320             OPJ_S(0) /= 2;
321         } else {
322             for (i = 0; i < sn; i++) {
323                 OPJ_D(i) -= (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2;
324             }
325             for (i = 0; i < dn; i++) {
326                 OPJ_S(i) += (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1;
327             }
328         }
329     }
330 }
331 
opj_dwt_decode_1(const opj_dwt_t * v)332 static void opj_dwt_decode_1(const opj_dwt_t *v)
333 {
334     opj_dwt_decode_1_(v->mem, v->mem_count, v->dn, v->sn, v->cas);
335 }
336 
337 #endif /* STANDARD_SLOW_VERSION */
338 
339 #if !defined(STANDARD_SLOW_VERSION)
opj_idwt53_h_cas0(OPJ_INT32 * tmp,const OPJ_INT32 sn,const OPJ_INT32 len,OPJ_INT32 * tiledp)340 static void  opj_idwt53_h_cas0(OPJ_INT32* tmp,
341                                const OPJ_INT32 sn,
342                                const OPJ_INT32 len,
343                                OPJ_INT32* tiledp)
344 {
345     OPJ_INT32 i, j;
346     const OPJ_INT32* in_even = &tiledp[0];
347     const OPJ_INT32* in_odd = &tiledp[sn];
348 
349 #ifdef TWO_PASS_VERSION
350     /* For documentation purpose: performs lifting in two iterations, */
351     /* but without explicit interleaving */
352 
353     assert(len > 1);
354 
355     /* Even */
356     tmp[0] = in_even[0] - ((in_odd[0] + 1) >> 1);
357     for (i = 2, j = 0; i <= len - 2; i += 2, j++) {
358         tmp[i] = in_even[j + 1] - ((in_odd[j] + in_odd[j + 1] + 2) >> 2);
359     }
360     if (len & 1) { /* if len is odd */
361         tmp[len - 1] = in_even[(len - 1) / 2] - ((in_odd[(len - 2) / 2] + 1) >> 1);
362     }
363 
364     /* Odd */
365     for (i = 1, j = 0; i < len - 1; i += 2, j++) {
366         tmp[i] = in_odd[j] + ((tmp[i - 1] + tmp[i + 1]) >> 1);
367     }
368     if (!(len & 1)) { /* if len is even */
369         tmp[len - 1] = in_odd[(len - 1) / 2] + tmp[len - 2];
370     }
371 #else
372     OPJ_INT32 d1c, d1n, s1n, s0c, s0n;
373 
374     assert(len > 1);
375 
376     /* Improved version of the TWO_PASS_VERSION: */
377     /* Performs lifting in one single iteration. Saves memory */
378     /* accesses and explicit interleaving. */
379     s1n = in_even[0];
380     d1n = in_odd[0];
381     s0n = s1n - ((d1n + 1) >> 1);
382 
383     for (i = 0, j = 1; i < (len - 3); i += 2, j++) {
384         d1c = d1n;
385         s0c = s0n;
386 
387         s1n = in_even[j];
388         d1n = in_odd[j];
389 
390         s0n = s1n - ((d1c + d1n + 2) >> 2);
391 
392         tmp[i  ] = s0c;
393         tmp[i + 1] = opj_int_add_no_overflow(d1c, opj_int_add_no_overflow(s0c,
394                                              s0n) >> 1);
395     }
396 
397     tmp[i] = s0n;
398 
399     if (len & 1) {
400         tmp[len - 1] = in_even[(len - 1) / 2] - ((d1n + 1) >> 1);
401         tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1);
402     } else {
403         tmp[len - 1] = d1n + s0n;
404     }
405 #endif
406     memcpy(tiledp, tmp, (OPJ_UINT32)len * sizeof(OPJ_INT32));
407 }
408 
opj_idwt53_h_cas1(OPJ_INT32 * tmp,const OPJ_INT32 sn,const OPJ_INT32 len,OPJ_INT32 * tiledp)409 static void  opj_idwt53_h_cas1(OPJ_INT32* tmp,
410                                const OPJ_INT32 sn,
411                                const OPJ_INT32 len,
412                                OPJ_INT32* tiledp)
413 {
414     OPJ_INT32 i, j;
415     const OPJ_INT32* in_even = &tiledp[sn];
416     const OPJ_INT32* in_odd = &tiledp[0];
417 
418 #ifdef TWO_PASS_VERSION
419     /* For documentation purpose: performs lifting in two iterations, */
420     /* but without explicit interleaving */
421 
422     assert(len > 2);
423 
424     /* Odd */
425     for (i = 1, j = 0; i < len - 1; i += 2, j++) {
426         tmp[i] = in_odd[j] - ((in_even[j] + in_even[j + 1] + 2) >> 2);
427     }
428     if (!(len & 1)) {
429         tmp[len - 1] = in_odd[len / 2 - 1] - ((in_even[len / 2 - 1] + 1) >> 1);
430     }
431 
432     /* Even */
433     tmp[0] = in_even[0] + tmp[1];
434     for (i = 2, j = 1; i < len - 1; i += 2, j++) {
435         tmp[i] = in_even[j] + ((tmp[i + 1] + tmp[i - 1]) >> 1);
436     }
437     if (len & 1) {
438         tmp[len - 1] = in_even[len / 2] + tmp[len - 2];
439     }
440 #else
441     OPJ_INT32 s1, s2, dc, dn;
442 
443     assert(len > 2);
444 
445     /* Improved version of the TWO_PASS_VERSION: */
446     /* Performs lifting in one single iteration. Saves memory */
447     /* accesses and explicit interleaving. */
448 
449     s1 = in_even[1];
450     dc = in_odd[0] - ((in_even[0] + s1 + 2) >> 2);
451     tmp[0] = in_even[0] + dc;
452 
453     for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
454 
455         s2 = in_even[j + 1];
456 
457         dn = in_odd[j] - ((s1 + s2 + 2) >> 2);
458         tmp[i  ] = dc;
459         tmp[i + 1] = opj_int_add_no_overflow(s1, opj_int_add_no_overflow(dn, dc) >> 1);
460 
461         dc = dn;
462         s1 = s2;
463     }
464 
465     tmp[i] = dc;
466 
467     if (!(len & 1)) {
468         dn = in_odd[len / 2 - 1] - ((s1 + 1) >> 1);
469         tmp[len - 2] = s1 + ((dn + dc) >> 1);
470         tmp[len - 1] = dn;
471     } else {
472         tmp[len - 1] = s1 + dc;
473     }
474 #endif
475     memcpy(tiledp, tmp, (OPJ_UINT32)len * sizeof(OPJ_INT32));
476 }
477 
478 
479 #endif /* !defined(STANDARD_SLOW_VERSION) */
480 
481 /* <summary>                            */
482 /* Inverse 5-3 wavelet transform in 1-D for one row. */
483 /* </summary>                           */
484 /* Performs interleave, inverse wavelet transform and copy back to buffer */
opj_idwt53_h(const opj_dwt_t * dwt,OPJ_INT32 * tiledp)485 static void opj_idwt53_h(const opj_dwt_t *dwt,
486                          OPJ_INT32* tiledp)
487 {
488 #ifdef STANDARD_SLOW_VERSION
489     /* For documentation purpose */
490     opj_dwt_interleave_h(dwt, tiledp);
491     opj_dwt_decode_1(dwt);
492     memcpy(tiledp, dwt->mem, (OPJ_UINT32)(dwt->sn + dwt->dn) * sizeof(OPJ_INT32));
493 #else
494     const OPJ_INT32 sn = dwt->sn;
495     const OPJ_INT32 len = sn + dwt->dn;
496     if (dwt->cas == 0) { /* Left-most sample is on even coordinate */
497         if (len > 1) {
498             opj_idwt53_h_cas0(dwt->mem, sn, len, tiledp);
499         } else {
500             /* Unmodified value */
501         }
502     } else { /* Left-most sample is on odd coordinate */
503         if (len == 1) {
504             tiledp[0] /= 2;
505         } else if (len == 2) {
506             OPJ_INT32* out = dwt->mem;
507             const OPJ_INT32* in_even = &tiledp[sn];
508             const OPJ_INT32* in_odd = &tiledp[0];
509             out[1] = in_odd[0] - ((in_even[0] + 1) >> 1);
510             out[0] = in_even[0] + out[1];
511             memcpy(tiledp, dwt->mem, (OPJ_UINT32)len * sizeof(OPJ_INT32));
512         } else if (len > 2) {
513             opj_idwt53_h_cas1(dwt->mem, sn, len, tiledp);
514         }
515     }
516 #endif
517 }
518 
519 #if (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION)
520 
521 /* Conveniency macros to improve the readability of the formulas */
522 #if __AVX2__
523 #define VREG        __m256i
524 #define LOAD_CST(x) _mm256_set1_epi32(x)
525 #define LOAD(x)     _mm256_load_si256((const VREG*)(x))
526 #define LOADU(x)    _mm256_loadu_si256((const VREG*)(x))
527 #define STORE(x,y)  _mm256_store_si256((VREG*)(x),(y))
528 #define STOREU(x,y) _mm256_storeu_si256((VREG*)(x),(y))
529 #define ADD(x,y)    _mm256_add_epi32((x),(y))
530 #define SUB(x,y)    _mm256_sub_epi32((x),(y))
531 #define SAR(x,y)    _mm256_srai_epi32((x),(y))
532 #else
533 #define VREG        __m128i
534 #define LOAD_CST(x) _mm_set1_epi32(x)
535 #define LOAD(x)     _mm_load_si128((const VREG*)(x))
536 #define LOADU(x)    _mm_loadu_si128((const VREG*)(x))
537 #define STORE(x,y)  _mm_store_si128((VREG*)(x),(y))
538 #define STOREU(x,y) _mm_storeu_si128((VREG*)(x),(y))
539 #define ADD(x,y)    _mm_add_epi32((x),(y))
540 #define SUB(x,y)    _mm_sub_epi32((x),(y))
541 #define SAR(x,y)    _mm_srai_epi32((x),(y))
542 #endif
543 #define ADD3(x,y,z) ADD(ADD(x,y),z)
544 
545 static
opj_idwt53_v_final_memcpy(OPJ_INT32 * tiledp_col,const OPJ_INT32 * tmp,OPJ_INT32 len,OPJ_SIZE_T stride)546 void opj_idwt53_v_final_memcpy(OPJ_INT32* tiledp_col,
547                                const OPJ_INT32* tmp,
548                                OPJ_INT32 len,
549                                OPJ_SIZE_T stride)
550 {
551     OPJ_INT32 i;
552     for (i = 0; i < len; ++i) {
553         /* A memcpy(&tiledp_col[i * stride + 0],
554                     &tmp[PARALLEL_COLS_53 * i + 0],
555                     PARALLEL_COLS_53 * sizeof(OPJ_INT32))
556            would do but would be a tiny bit slower.
557            We can take here advantage of our knowledge of alignment */
558         STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + 0],
559                LOAD(&tmp[PARALLEL_COLS_53 * i + 0]));
560         STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + VREG_INT_COUNT],
561                LOAD(&tmp[PARALLEL_COLS_53 * i + VREG_INT_COUNT]));
562     }
563 }
564 
565 /** Vertical inverse 5x3 wavelet transform for 8 columns in SSE2, or
566  * 16 in AVX2, when top-most pixel is on even coordinate */
opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2(OPJ_INT32 * tmp,const OPJ_INT32 sn,const OPJ_INT32 len,OPJ_INT32 * tiledp_col,const OPJ_SIZE_T stride)567 static void opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2(
568     OPJ_INT32* tmp,
569     const OPJ_INT32 sn,
570     const OPJ_INT32 len,
571     OPJ_INT32* tiledp_col,
572     const OPJ_SIZE_T stride)
573 {
574     const OPJ_INT32* in_even = &tiledp_col[0];
575     const OPJ_INT32* in_odd = &tiledp_col[(OPJ_SIZE_T)sn * stride];
576 
577     OPJ_INT32 i;
578     OPJ_SIZE_T j;
579     VREG d1c_0, d1n_0, s1n_0, s0c_0, s0n_0;
580     VREG d1c_1, d1n_1, s1n_1, s0c_1, s0n_1;
581     const VREG two = LOAD_CST(2);
582 
583     assert(len > 1);
584 #if __AVX2__
585     assert(PARALLEL_COLS_53 == 16);
586     assert(VREG_INT_COUNT == 8);
587 #else
588     assert(PARALLEL_COLS_53 == 8);
589     assert(VREG_INT_COUNT == 4);
590 #endif
591 
592     /* Note: loads of input even/odd values must be done in a unaligned */
593     /* fashion. But stores in tmp can be done with aligned store, since */
594     /* the temporary buffer is properly aligned */
595     assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
596 
597     s1n_0 = LOADU(in_even + 0);
598     s1n_1 = LOADU(in_even + VREG_INT_COUNT);
599     d1n_0 = LOADU(in_odd);
600     d1n_1 = LOADU(in_odd + VREG_INT_COUNT);
601 
602     /* s0n = s1n - ((d1n + 1) >> 1); <==> */
603     /* s0n = s1n - ((d1n + d1n + 2) >> 2); */
604     s0n_0 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2));
605     s0n_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2));
606 
607     for (i = 0, j = 1; i < (len - 3); i += 2, j++) {
608         d1c_0 = d1n_0;
609         s0c_0 = s0n_0;
610         d1c_1 = d1n_1;
611         s0c_1 = s0n_1;
612 
613         s1n_0 = LOADU(in_even + j * stride);
614         s1n_1 = LOADU(in_even + j * stride + VREG_INT_COUNT);
615         d1n_0 = LOADU(in_odd + j * stride);
616         d1n_1 = LOADU(in_odd + j * stride + VREG_INT_COUNT);
617 
618         /*s0n = s1n - ((d1c + d1n + 2) >> 2);*/
619         s0n_0 = SUB(s1n_0, SAR(ADD3(d1c_0, d1n_0, two), 2));
620         s0n_1 = SUB(s1n_1, SAR(ADD3(d1c_1, d1n_1, two), 2));
621 
622         STORE(tmp + PARALLEL_COLS_53 * (i + 0), s0c_0);
623         STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0c_1);
624 
625         /* d1c + ((s0c + s0n) >> 1) */
626         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0,
627               ADD(d1c_0, SAR(ADD(s0c_0, s0n_0), 1)));
628         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT,
629               ADD(d1c_1, SAR(ADD(s0c_1, s0n_1), 1)));
630     }
631 
632     STORE(tmp + PARALLEL_COLS_53 * (i + 0) + 0, s0n_0);
633     STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0n_1);
634 
635     if (len & 1) {
636         VREG tmp_len_minus_1;
637         s1n_0 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride);
638         /* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */
639         tmp_len_minus_1 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2));
640         STORE(tmp + PARALLEL_COLS_53 * (len - 1), tmp_len_minus_1);
641         /* d1n + ((s0n + tmp_len_minus_1) >> 1) */
642         STORE(tmp + PARALLEL_COLS_53 * (len - 2),
643               ADD(d1n_0, SAR(ADD(s0n_0, tmp_len_minus_1), 1)));
644 
645         s1n_1 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride + VREG_INT_COUNT);
646         /* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */
647         tmp_len_minus_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2));
648         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
649               tmp_len_minus_1);
650         /* d1n + ((s0n + tmp_len_minus_1) >> 1) */
651         STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT,
652               ADD(d1n_1, SAR(ADD(s0n_1, tmp_len_minus_1), 1)));
653 
654 
655     } else {
656         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0,
657               ADD(d1n_0, s0n_0));
658         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
659               ADD(d1n_1, s0n_1));
660     }
661 
662     opj_idwt53_v_final_memcpy(tiledp_col, tmp, len, stride);
663 }
664 
665 
666 /** Vertical inverse 5x3 wavelet transform for 8 columns in SSE2, or
667  * 16 in AVX2, when top-most pixel is on odd coordinate */
opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2(OPJ_INT32 * tmp,const OPJ_INT32 sn,const OPJ_INT32 len,OPJ_INT32 * tiledp_col,const OPJ_SIZE_T stride)668 static void opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2(
669     OPJ_INT32* tmp,
670     const OPJ_INT32 sn,
671     const OPJ_INT32 len,
672     OPJ_INT32* tiledp_col,
673     const OPJ_SIZE_T stride)
674 {
675     OPJ_INT32 i;
676     OPJ_SIZE_T j;
677 
678     VREG s1_0, s2_0, dc_0, dn_0;
679     VREG s1_1, s2_1, dc_1, dn_1;
680     const VREG two = LOAD_CST(2);
681 
682     const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
683     const OPJ_INT32* in_odd = &tiledp_col[0];
684 
685     assert(len > 2);
686 #if __AVX2__
687     assert(PARALLEL_COLS_53 == 16);
688     assert(VREG_INT_COUNT == 8);
689 #else
690     assert(PARALLEL_COLS_53 == 8);
691     assert(VREG_INT_COUNT == 4);
692 #endif
693 
694     /* Note: loads of input even/odd values must be done in a unaligned */
695     /* fashion. But stores in tmp can be done with aligned store, since */
696     /* the temporary buffer is properly aligned */
697     assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
698 
699     s1_0 = LOADU(in_even + stride);
700     /* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */
701     dc_0 = SUB(LOADU(in_odd + 0),
702                SAR(ADD3(LOADU(in_even + 0), s1_0, two), 2));
703     STORE(tmp + PARALLEL_COLS_53 * 0, ADD(LOADU(in_even + 0), dc_0));
704 
705     s1_1 = LOADU(in_even + stride + VREG_INT_COUNT);
706     /* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */
707     dc_1 = SUB(LOADU(in_odd + VREG_INT_COUNT),
708                SAR(ADD3(LOADU(in_even + VREG_INT_COUNT), s1_1, two), 2));
709     STORE(tmp + PARALLEL_COLS_53 * 0 + VREG_INT_COUNT,
710           ADD(LOADU(in_even + VREG_INT_COUNT), dc_1));
711 
712     for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
713 
714         s2_0 = LOADU(in_even + (j + 1) * stride);
715         s2_1 = LOADU(in_even + (j + 1) * stride + VREG_INT_COUNT);
716 
717         /* dn = in_odd[j * stride] - ((s1 + s2 + 2) >> 2); */
718         dn_0 = SUB(LOADU(in_odd + j * stride),
719                    SAR(ADD3(s1_0, s2_0, two), 2));
720         dn_1 = SUB(LOADU(in_odd + j * stride + VREG_INT_COUNT),
721                    SAR(ADD3(s1_1, s2_1, two), 2));
722 
723         STORE(tmp + PARALLEL_COLS_53 * i, dc_0);
724         STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1);
725 
726         /* tmp[i + 1] = s1 + ((dn + dc) >> 1); */
727         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0,
728               ADD(s1_0, SAR(ADD(dn_0, dc_0), 1)));
729         STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT,
730               ADD(s1_1, SAR(ADD(dn_1, dc_1), 1)));
731 
732         dc_0 = dn_0;
733         s1_0 = s2_0;
734         dc_1 = dn_1;
735         s1_1 = s2_1;
736     }
737     STORE(tmp + PARALLEL_COLS_53 * i, dc_0);
738     STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1);
739 
740     if (!(len & 1)) {
741         /*dn = in_odd[(len / 2 - 1) * stride] - ((s1 + 1) >> 1); */
742         dn_0 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride),
743                    SAR(ADD3(s1_0, s1_0, two), 2));
744         dn_1 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride + VREG_INT_COUNT),
745                    SAR(ADD3(s1_1, s1_1, two), 2));
746 
747         /* tmp[len - 2] = s1 + ((dn + dc) >> 1); */
748         STORE(tmp + PARALLEL_COLS_53 * (len - 2) + 0,
749               ADD(s1_0, SAR(ADD(dn_0, dc_0), 1)));
750         STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT,
751               ADD(s1_1, SAR(ADD(dn_1, dc_1), 1)));
752 
753         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, dn_0);
754         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, dn_1);
755     } else {
756         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, ADD(s1_0, dc_0));
757         STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
758               ADD(s1_1, dc_1));
759     }
760 
761     opj_idwt53_v_final_memcpy(tiledp_col, tmp, len, stride);
762 }
763 
764 #undef VREG
765 #undef LOAD_CST
766 #undef LOADU
767 #undef LOAD
768 #undef STORE
769 #undef STOREU
770 #undef ADD
771 #undef ADD3
772 #undef SUB
773 #undef SAR
774 
775 #endif /* (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION) */
776 
777 #if !defined(STANDARD_SLOW_VERSION)
778 /** Vertical inverse 5x3 wavelet transform for one column, when top-most
779  * pixel is on even coordinate */
opj_idwt3_v_cas0(OPJ_INT32 * tmp,const OPJ_INT32 sn,const OPJ_INT32 len,OPJ_INT32 * tiledp_col,const OPJ_SIZE_T stride)780 static void opj_idwt3_v_cas0(OPJ_INT32* tmp,
781                              const OPJ_INT32 sn,
782                              const OPJ_INT32 len,
783                              OPJ_INT32* tiledp_col,
784                              const OPJ_SIZE_T stride)
785 {
786     OPJ_INT32 i, j;
787     OPJ_INT32 d1c, d1n, s1n, s0c, s0n;
788 
789     assert(len > 1);
790 
791     /* Performs lifting in one single iteration. Saves memory */
792     /* accesses and explicit interleaving. */
793 
794     s1n = tiledp_col[0];
795     d1n = tiledp_col[(OPJ_SIZE_T)sn * stride];
796     s0n = s1n - ((d1n + 1) >> 1);
797 
798     for (i = 0, j = 0; i < (len - 3); i += 2, j++) {
799         d1c = d1n;
800         s0c = s0n;
801 
802         s1n = tiledp_col[(OPJ_SIZE_T)(j + 1) * stride];
803         d1n = tiledp_col[(OPJ_SIZE_T)(sn + j + 1) * stride];
804 
805         s0n = opj_int_sub_no_overflow(s1n,
806                                       opj_int_add_no_overflow(opj_int_add_no_overflow(d1c, d1n), 2) >> 2);
807 
808         tmp[i  ] = s0c;
809         tmp[i + 1] = opj_int_add_no_overflow(d1c, opj_int_add_no_overflow(s0c,
810                                              s0n) >> 1);
811     }
812 
813     tmp[i] = s0n;
814 
815     if (len & 1) {
816         tmp[len - 1] =
817             tiledp_col[(OPJ_SIZE_T)((len - 1) / 2) * stride] -
818             ((d1n + 1) >> 1);
819         tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1);
820     } else {
821         tmp[len - 1] = d1n + s0n;
822     }
823 
824     for (i = 0; i < len; ++i) {
825         tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i];
826     }
827 }
828 
829 
830 /** Vertical inverse 5x3 wavelet transform for one column, when top-most
831  * pixel is on odd coordinate */
opj_idwt3_v_cas1(OPJ_INT32 * tmp,const OPJ_INT32 sn,const OPJ_INT32 len,OPJ_INT32 * tiledp_col,const OPJ_SIZE_T stride)832 static void opj_idwt3_v_cas1(OPJ_INT32* tmp,
833                              const OPJ_INT32 sn,
834                              const OPJ_INT32 len,
835                              OPJ_INT32* tiledp_col,
836                              const OPJ_SIZE_T stride)
837 {
838     OPJ_INT32 i, j;
839     OPJ_INT32 s1, s2, dc, dn;
840     const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
841     const OPJ_INT32* in_odd = &tiledp_col[0];
842 
843     assert(len > 2);
844 
845     /* Performs lifting in one single iteration. Saves memory */
846     /* accesses and explicit interleaving. */
847 
848     s1 = in_even[stride];
849     dc = in_odd[0] - ((in_even[0] + s1 + 2) >> 2);
850     tmp[0] = in_even[0] + dc;
851     for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
852 
853         s2 = in_even[(OPJ_SIZE_T)(j + 1) * stride];
854 
855         dn = in_odd[(OPJ_SIZE_T)j * stride] - ((s1 + s2 + 2) >> 2);
856         tmp[i  ] = dc;
857         tmp[i + 1] = s1 + ((dn + dc) >> 1);
858 
859         dc = dn;
860         s1 = s2;
861     }
862     tmp[i] = dc;
863     if (!(len & 1)) {
864         dn = in_odd[(OPJ_SIZE_T)(len / 2 - 1) * stride] - ((s1 + 1) >> 1);
865         tmp[len - 2] = s1 + ((dn + dc) >> 1);
866         tmp[len - 1] = dn;
867     } else {
868         tmp[len - 1] = s1 + dc;
869     }
870 
871     for (i = 0; i < len; ++i) {
872         tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i];
873     }
874 }
875 #endif /* !defined(STANDARD_SLOW_VERSION) */
876 
877 /* <summary>                            */
878 /* Inverse vertical 5-3 wavelet transform in 1-D for several columns. */
879 /* </summary>                           */
880 /* Performs interleave, inverse wavelet transform and copy back to buffer */
opj_idwt53_v(const opj_dwt_t * dwt,OPJ_INT32 * tiledp_col,OPJ_SIZE_T stride,OPJ_INT32 nb_cols)881 static void opj_idwt53_v(const opj_dwt_t *dwt,
882                          OPJ_INT32* tiledp_col,
883                          OPJ_SIZE_T stride,
884                          OPJ_INT32 nb_cols)
885 {
886 #ifdef STANDARD_SLOW_VERSION
887     /* For documentation purpose */
888     OPJ_INT32 k, c;
889     for (c = 0; c < nb_cols; c ++) {
890         opj_dwt_interleave_v(dwt, tiledp_col + c, stride);
891         opj_dwt_decode_1(dwt);
892         for (k = 0; k < dwt->sn + dwt->dn; ++k) {
893             tiledp_col[c + k * stride] = dwt->mem[k];
894         }
895     }
896 #else
897     const OPJ_INT32 sn = dwt->sn;
898     const OPJ_INT32 len = sn + dwt->dn;
899     if (dwt->cas == 0) {
900         /* If len == 1, unmodified value */
901 
902 #if (defined(__SSE2__) || defined(__AVX2__))
903         if (len > 1 && nb_cols == PARALLEL_COLS_53) {
904             /* Same as below general case, except that thanks to SSE2/AVX2 */
905             /* we can efficiently process 8/16 columns in parallel */
906             opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2(dwt->mem, sn, len, tiledp_col, stride);
907             return;
908         }
909 #endif
910         if (len > 1) {
911             OPJ_INT32 c;
912             for (c = 0; c < nb_cols; c++, tiledp_col++) {
913                 opj_idwt3_v_cas0(dwt->mem, sn, len, tiledp_col, stride);
914             }
915             return;
916         }
917     } else {
918         if (len == 1) {
919             OPJ_INT32 c;
920             for (c = 0; c < nb_cols; c++, tiledp_col++) {
921                 tiledp_col[0] /= 2;
922             }
923             return;
924         }
925 
926         if (len == 2) {
927             OPJ_INT32 c;
928             OPJ_INT32* out = dwt->mem;
929             for (c = 0; c < nb_cols; c++, tiledp_col++) {
930                 OPJ_INT32 i;
931                 const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
932                 const OPJ_INT32* in_odd = &tiledp_col[0];
933 
934                 out[1] = in_odd[0] - ((in_even[0] + 1) >> 1);
935                 out[0] = in_even[0] + out[1];
936 
937                 for (i = 0; i < len; ++i) {
938                     tiledp_col[(OPJ_SIZE_T)i * stride] = out[i];
939                 }
940             }
941 
942             return;
943         }
944 
945 #if (defined(__SSE2__) || defined(__AVX2__))
946         if (len > 2 && nb_cols == PARALLEL_COLS_53) {
947             /* Same as below general case, except that thanks to SSE2/AVX2 */
948             /* we can efficiently process 8/16 columns in parallel */
949             opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2(dwt->mem, sn, len, tiledp_col, stride);
950             return;
951         }
952 #endif
953         if (len > 2) {
954             OPJ_INT32 c;
955             for (c = 0; c < nb_cols; c++, tiledp_col++) {
956                 opj_idwt3_v_cas1(dwt->mem, sn, len, tiledp_col, stride);
957             }
958             return;
959         }
960     }
961 #endif
962 }
963 
964 #if 0
965 static void opj_dwt_encode_step1(OPJ_FLOAT32* fw,
966                                  OPJ_UINT32 end,
967                                  const OPJ_FLOAT32 c)
968 {
969     OPJ_UINT32 i = 0;
970     for (; i < end; ++i) {
971         fw[0] *= c;
972         fw += 2;
973     }
974 }
975 #else
opj_dwt_encode_step1_combined(OPJ_FLOAT32 * fw,OPJ_UINT32 iters_c1,OPJ_UINT32 iters_c2,const OPJ_FLOAT32 c1,const OPJ_FLOAT32 c2)976 static void opj_dwt_encode_step1_combined(OPJ_FLOAT32* fw,
977         OPJ_UINT32 iters_c1,
978         OPJ_UINT32 iters_c2,
979         const OPJ_FLOAT32 c1,
980         const OPJ_FLOAT32 c2)
981 {
982     OPJ_UINT32 i = 0;
983     const OPJ_UINT32 iters_common =  opj_uint_min(iters_c1, iters_c2);
984     assert((((OPJ_SIZE_T)fw) & 0xf) == 0);
985     assert(opj_int_abs((OPJ_INT32)iters_c1 - (OPJ_INT32)iters_c2) <= 1);
986     for (; i + 3 < iters_common; i += 4) {
987 #ifdef __SSE__
988         const __m128 vcst = _mm_set_ps(c2, c1, c2, c1);
989         *(__m128*)fw = _mm_mul_ps(*(__m128*)fw, vcst);
990         *(__m128*)(fw + 4) = _mm_mul_ps(*(__m128*)(fw + 4), vcst);
991 #else
992         fw[0] *= c1;
993         fw[1] *= c2;
994         fw[2] *= c1;
995         fw[3] *= c2;
996         fw[4] *= c1;
997         fw[5] *= c2;
998         fw[6] *= c1;
999         fw[7] *= c2;
1000 #endif
1001         fw += 8;
1002     }
1003     for (; i < iters_common; i++) {
1004         fw[0] *= c1;
1005         fw[1] *= c2;
1006         fw += 2;
1007     }
1008     if (i < iters_c1) {
1009         fw[0] *= c1;
1010     } else if (i < iters_c2) {
1011         fw[1] *= c2;
1012     }
1013 }
1014 
1015 #endif
1016 
opj_dwt_encode_step2(OPJ_FLOAT32 * fl,OPJ_FLOAT32 * fw,OPJ_UINT32 end,OPJ_UINT32 m,OPJ_FLOAT32 c)1017 static void opj_dwt_encode_step2(OPJ_FLOAT32* fl, OPJ_FLOAT32* fw,
1018                                  OPJ_UINT32 end,
1019                                  OPJ_UINT32 m,
1020                                  OPJ_FLOAT32 c)
1021 {
1022     OPJ_UINT32 i;
1023     OPJ_UINT32 imax = opj_uint_min(end, m);
1024     if (imax > 0) {
1025         fw[-1] += (fl[0] + fw[0]) * c;
1026         fw += 2;
1027         i = 1;
1028         for (; i + 3 < imax; i += 4) {
1029             fw[-1] += (fw[-2] + fw[0]) * c;
1030             fw[1] += (fw[0] + fw[2]) * c;
1031             fw[3] += (fw[2] + fw[4]) * c;
1032             fw[5] += (fw[4] + fw[6]) * c;
1033             fw += 8;
1034         }
1035         for (; i < imax; ++i) {
1036             fw[-1] += (fw[-2] + fw[0]) * c;
1037             fw += 2;
1038         }
1039     }
1040     if (m < end) {
1041         assert(m + 1 == end);
1042         fw[-1] += (2 * fw[-2]) * c;
1043     }
1044 }
1045 
opj_dwt_encode_1_real(void * aIn,OPJ_INT32 dn,OPJ_INT32 sn,OPJ_INT32 cas)1046 static void opj_dwt_encode_1_real(void *aIn, OPJ_INT32 dn, OPJ_INT32 sn,
1047                                   OPJ_INT32 cas)
1048 {
1049     OPJ_FLOAT32* w = (OPJ_FLOAT32*)aIn;
1050     OPJ_INT32 a, b;
1051     assert(dn + sn > 1);
1052     if (cas == 0) {
1053         a = 0;
1054         b = 1;
1055     } else {
1056         a = 1;
1057         b = 0;
1058     }
1059     opj_dwt_encode_step2(w + a, w + b + 1,
1060                          (OPJ_UINT32)dn,
1061                          (OPJ_UINT32)opj_int_min(dn, sn - b),
1062                          opj_dwt_alpha);
1063     opj_dwt_encode_step2(w + b, w + a + 1,
1064                          (OPJ_UINT32)sn,
1065                          (OPJ_UINT32)opj_int_min(sn, dn - a),
1066                          opj_dwt_beta);
1067     opj_dwt_encode_step2(w + a, w + b + 1,
1068                          (OPJ_UINT32)dn,
1069                          (OPJ_UINT32)opj_int_min(dn, sn - b),
1070                          opj_dwt_gamma);
1071     opj_dwt_encode_step2(w + b, w + a + 1,
1072                          (OPJ_UINT32)sn,
1073                          (OPJ_UINT32)opj_int_min(sn, dn - a),
1074                          opj_dwt_delta);
1075 #if 0
1076     opj_dwt_encode_step1(w + b, (OPJ_UINT32)dn,
1077                          opj_K);
1078     opj_dwt_encode_step1(w + a, (OPJ_UINT32)sn,
1079                          opj_invK);
1080 #else
1081     if (a == 0) {
1082         opj_dwt_encode_step1_combined(w,
1083                                       (OPJ_UINT32)sn,
1084                                       (OPJ_UINT32)dn,
1085                                       opj_invK,
1086                                       opj_K);
1087     } else {
1088         opj_dwt_encode_step1_combined(w,
1089                                       (OPJ_UINT32)dn,
1090                                       (OPJ_UINT32)sn,
1091                                       opj_K,
1092                                       opj_invK);
1093     }
1094 #endif
1095 }
1096 
opj_dwt_encode_stepsize(OPJ_INT32 stepsize,OPJ_INT32 numbps,opj_stepsize_t * bandno_stepsize)1097 static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
1098                                     opj_stepsize_t *bandno_stepsize)
1099 {
1100     OPJ_INT32 p, n;
1101     p = opj_int_floorlog2(stepsize) - 13;
1102     n = 11 - opj_int_floorlog2(stepsize);
1103     bandno_stepsize->mant = (n < 0 ? stepsize >> -n : stepsize << n) & 0x7ff;
1104     bandno_stepsize->expn = numbps - p;
1105 }
1106 
1107 /*
1108 ==========================================================
1109    DWT interface
1110 ==========================================================
1111 */
1112 
1113 /** Process one line for the horizontal pass of the 5x3 forward transform */
1114 static
opj_dwt_encode_and_deinterleave_h_one_row(void * rowIn,void * tmpIn,OPJ_UINT32 width,OPJ_BOOL even)1115 void opj_dwt_encode_and_deinterleave_h_one_row(void* rowIn,
1116         void* tmpIn,
1117         OPJ_UINT32 width,
1118         OPJ_BOOL even)
1119 {
1120     OPJ_INT32* OPJ_RESTRICT row = (OPJ_INT32*)rowIn;
1121     OPJ_INT32* OPJ_RESTRICT tmp = (OPJ_INT32*)tmpIn;
1122     const OPJ_INT32 sn = (OPJ_INT32)((width + (even ? 1 : 0)) >> 1);
1123     const OPJ_INT32 dn = (OPJ_INT32)(width - (OPJ_UINT32)sn);
1124 
1125     if (even) {
1126         if (width > 1) {
1127             OPJ_INT32 i;
1128             for (i = 0; i < sn - 1; i++) {
1129                 tmp[sn + i] = row[2 * i + 1] - ((row[(i) * 2] + row[(i + 1) * 2]) >> 1);
1130             }
1131             if ((width % 2) == 0) {
1132                 tmp[sn + i] = row[2 * i + 1] - row[(i) * 2];
1133             }
1134             row[0] += (tmp[sn] + tmp[sn] + 2) >> 2;
1135             for (i = 1; i < dn; i++) {
1136                 row[i] = row[2 * i] + ((tmp[sn + (i - 1)] + tmp[sn + i] + 2) >> 2);
1137             }
1138             if ((width % 2) == 1) {
1139                 row[i] = row[2 * i] + ((tmp[sn + (i - 1)] + tmp[sn + (i - 1)] + 2) >> 2);
1140             }
1141             memcpy(row + sn, tmp + sn, (OPJ_SIZE_T)dn * sizeof(OPJ_INT32));
1142         }
1143     } else {
1144         if (width == 1) {
1145             row[0] *= 2;
1146         } else {
1147             OPJ_INT32 i;
1148             tmp[sn + 0] = row[0] - row[1];
1149             for (i = 1; i < sn; i++) {
1150                 tmp[sn + i] = row[2 * i] - ((row[2 * i + 1] + row[2 * (i - 1) + 1]) >> 1);
1151             }
1152             if ((width % 2) == 1) {
1153                 tmp[sn + i] = row[2 * i] - row[2 * (i - 1) + 1];
1154             }
1155 
1156             for (i = 0; i < dn - 1; i++) {
1157                 row[i] = row[2 * i + 1] + ((tmp[sn + i] + tmp[sn + i + 1] + 2) >> 2);
1158             }
1159             if ((width % 2) == 0) {
1160                 row[i] = row[2 * i + 1] + ((tmp[sn + i] + tmp[sn + i] + 2) >> 2);
1161             }
1162             memcpy(row + sn, tmp + sn, (OPJ_SIZE_T)dn * sizeof(OPJ_INT32));
1163         }
1164     }
1165 }
1166 
1167 /** Process one line for the horizontal pass of the 9x7 forward transform */
1168 static
opj_dwt_encode_and_deinterleave_h_one_row_real(void * rowIn,void * tmpIn,OPJ_UINT32 width,OPJ_BOOL even)1169 void opj_dwt_encode_and_deinterleave_h_one_row_real(void* rowIn,
1170         void* tmpIn,
1171         OPJ_UINT32 width,
1172         OPJ_BOOL even)
1173 {
1174     OPJ_FLOAT32* OPJ_RESTRICT row = (OPJ_FLOAT32*)rowIn;
1175     OPJ_FLOAT32* OPJ_RESTRICT tmp = (OPJ_FLOAT32*)tmpIn;
1176     const OPJ_INT32 sn = (OPJ_INT32)((width + (even ? 1 : 0)) >> 1);
1177     const OPJ_INT32 dn = (OPJ_INT32)(width - (OPJ_UINT32)sn);
1178     if (width == 1) {
1179         return;
1180     }
1181     memcpy(tmp, row, width * sizeof(OPJ_FLOAT32));
1182     opj_dwt_encode_1_real(tmp, dn, sn, even ? 0 : 1);
1183     opj_dwt_deinterleave_h((OPJ_INT32 * OPJ_RESTRICT)tmp,
1184                            (OPJ_INT32 * OPJ_RESTRICT)row,
1185                            dn, sn, even ? 0 : 1);
1186 }
1187 
1188 typedef struct {
1189     opj_dwt_t h;
1190     OPJ_UINT32 rw; /* Width of the resolution to process */
1191     OPJ_UINT32 w; /* Width of tiledp */
1192     OPJ_INT32 * OPJ_RESTRICT tiledp;
1193     OPJ_UINT32 min_j;
1194     OPJ_UINT32 max_j;
1195     opj_encode_and_deinterleave_h_one_row_fnptr_type p_function;
1196 } opj_dwt_encode_h_job_t;
1197 
opj_dwt_encode_h_func(void * user_data,opj_tls_t * tls)1198 static void opj_dwt_encode_h_func(void* user_data, opj_tls_t* tls)
1199 {
1200     OPJ_UINT32 j;
1201     opj_dwt_encode_h_job_t* job;
1202     (void)tls;
1203 
1204     job = (opj_dwt_encode_h_job_t*)user_data;
1205     for (j = job->min_j; j < job->max_j; j++) {
1206         OPJ_INT32* OPJ_RESTRICT aj = job->tiledp + j * job->w;
1207         (*job->p_function)(aj, job->h.mem, job->rw,
1208                            job->h.cas == 0 ? OPJ_TRUE : OPJ_FALSE);
1209     }
1210 
1211     opj_aligned_free(job->h.mem);
1212     opj_free(job);
1213 }
1214 
1215 typedef struct {
1216     opj_dwt_t v;
1217     OPJ_UINT32 rh;
1218     OPJ_UINT32 w;
1219     OPJ_INT32 * OPJ_RESTRICT tiledp;
1220     OPJ_UINT32 min_j;
1221     OPJ_UINT32 max_j;
1222     opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v;
1223 } opj_dwt_encode_v_job_t;
1224 
opj_dwt_encode_v_func(void * user_data,opj_tls_t * tls)1225 static void opj_dwt_encode_v_func(void* user_data, opj_tls_t* tls)
1226 {
1227     OPJ_UINT32 j;
1228     opj_dwt_encode_v_job_t* job;
1229     (void)tls;
1230 
1231     job = (opj_dwt_encode_v_job_t*)user_data;
1232     for (j = job->min_j; j + NB_ELTS_V8 - 1 < job->max_j; j += NB_ELTS_V8) {
1233         (*job->p_encode_and_deinterleave_v)(job->tiledp + j,
1234                                             job->v.mem,
1235                                             job->rh,
1236                                             job->v.cas == 0,
1237                                             job->w,
1238                                             NB_ELTS_V8);
1239     }
1240     if (j < job->max_j) {
1241         (*job->p_encode_and_deinterleave_v)(job->tiledp + j,
1242                                             job->v.mem,
1243                                             job->rh,
1244                                             job->v.cas == 0,
1245                                             job->w,
1246                                             job->max_j - j);
1247     }
1248 
1249     opj_aligned_free(job->v.mem);
1250     opj_free(job);
1251 }
1252 
1253 /** Fetch up to cols <= NB_ELTS_V8 for each line, and put them in tmpOut */
1254 /* that has a NB_ELTS_V8 interleave factor. */
opj_dwt_fetch_cols_vertical_pass(const void * arrayIn,void * tmpOut,OPJ_UINT32 height,OPJ_UINT32 stride_width,OPJ_UINT32 cols)1255 static void opj_dwt_fetch_cols_vertical_pass(const void *arrayIn,
1256         void *tmpOut,
1257         OPJ_UINT32 height,
1258         OPJ_UINT32 stride_width,
1259         OPJ_UINT32 cols)
1260 {
1261     const OPJ_INT32* OPJ_RESTRICT array = (const OPJ_INT32 * OPJ_RESTRICT)arrayIn;
1262     OPJ_INT32* OPJ_RESTRICT tmp = (OPJ_INT32 * OPJ_RESTRICT)tmpOut;
1263     if (cols == NB_ELTS_V8) {
1264         OPJ_UINT32 k;
1265         for (k = 0; k < height; ++k) {
1266             memcpy(tmp + NB_ELTS_V8 * k,
1267                    array + k * stride_width,
1268                    NB_ELTS_V8 * sizeof(OPJ_INT32));
1269         }
1270     } else {
1271         OPJ_UINT32 k;
1272         for (k = 0; k < height; ++k) {
1273             OPJ_UINT32 c;
1274             for (c = 0; c < cols; c++) {
1275                 tmp[NB_ELTS_V8 * k + c] = array[c + k * stride_width];
1276             }
1277             for (; c < NB_ELTS_V8; c++) {
1278                 tmp[NB_ELTS_V8 * k + c] = 0;
1279             }
1280         }
1281     }
1282 }
1283 
1284 /* Deinterleave result of forward transform, where cols <= NB_ELTS_V8 */
1285 /* and src contains NB_ELTS_V8 consecutive values for up to NB_ELTS_V8 */
1286 /* columns. */
opj_dwt_deinterleave_v_cols(const OPJ_INT32 * OPJ_RESTRICT src,OPJ_INT32 * OPJ_RESTRICT dst,OPJ_INT32 dn,OPJ_INT32 sn,OPJ_UINT32 stride_width,OPJ_INT32 cas,OPJ_UINT32 cols)1287 static INLINE void opj_dwt_deinterleave_v_cols(
1288     const OPJ_INT32 * OPJ_RESTRICT src,
1289     OPJ_INT32 * OPJ_RESTRICT dst,
1290     OPJ_INT32 dn,
1291     OPJ_INT32 sn,
1292     OPJ_UINT32 stride_width,
1293     OPJ_INT32 cas,
1294     OPJ_UINT32 cols)
1295 {
1296     OPJ_INT32 k;
1297     OPJ_INT32 i = sn;
1298     OPJ_INT32 * OPJ_RESTRICT l_dest = dst;
1299     const OPJ_INT32 * OPJ_RESTRICT l_src = src + cas * NB_ELTS_V8;
1300     OPJ_UINT32 c;
1301 
1302     for (k = 0; k < 2; k++) {
1303         while (i--) {
1304             if (cols == NB_ELTS_V8) {
1305                 memcpy(l_dest, l_src, NB_ELTS_V8 * sizeof(OPJ_INT32));
1306             } else {
1307                 c = 0;
1308                 switch (cols) {
1309                 case 7:
1310                     l_dest[c] = l_src[c];
1311                     c++; /* fallthru */
1312                 case 6:
1313                     l_dest[c] = l_src[c];
1314                     c++; /* fallthru */
1315                 case 5:
1316                     l_dest[c] = l_src[c];
1317                     c++; /* fallthru */
1318                 case 4:
1319                     l_dest[c] = l_src[c];
1320                     c++; /* fallthru */
1321                 case 3:
1322                     l_dest[c] = l_src[c];
1323                     c++; /* fallthru */
1324                 case 2:
1325                     l_dest[c] = l_src[c];
1326                     c++; /* fallthru */
1327                 default:
1328                     l_dest[c] = l_src[c];
1329                     break;
1330                 }
1331             }
1332             l_dest += stride_width;
1333             l_src += 2 * NB_ELTS_V8;
1334         }
1335 
1336         l_dest = dst + (OPJ_SIZE_T)sn * (OPJ_SIZE_T)stride_width;
1337         l_src = src + (1 - cas) * NB_ELTS_V8;
1338         i = dn;
1339     }
1340 }
1341 
1342 
1343 /* Forward 5-3 transform, for the vertical pass, processing cols columns */
1344 /* where cols <= NB_ELTS_V8 */
opj_dwt_encode_and_deinterleave_v(void * arrayIn,void * tmpIn,OPJ_UINT32 height,OPJ_BOOL even,OPJ_UINT32 stride_width,OPJ_UINT32 cols)1345 static void opj_dwt_encode_and_deinterleave_v(
1346     void *arrayIn,
1347     void *tmpIn,
1348     OPJ_UINT32 height,
1349     OPJ_BOOL even,
1350     OPJ_UINT32 stride_width,
1351     OPJ_UINT32 cols)
1352 {
1353     OPJ_INT32* OPJ_RESTRICT array = (OPJ_INT32 * OPJ_RESTRICT)arrayIn;
1354     OPJ_INT32* OPJ_RESTRICT tmp = (OPJ_INT32 * OPJ_RESTRICT)tmpIn;
1355     const OPJ_UINT32 sn = (height + (even ? 1 : 0)) >> 1;
1356     const OPJ_UINT32 dn = height - sn;
1357 
1358     opj_dwt_fetch_cols_vertical_pass(arrayIn, tmpIn, height, stride_width, cols);
1359 
1360 #define OPJ_Sc(i) tmp[(i)*2* NB_ELTS_V8 + c]
1361 #define OPJ_Dc(i) tmp[((1+(i)*2))* NB_ELTS_V8 + c]
1362 
1363 #ifdef __SSE2__
1364     if (height == 1) {
1365         if (!even) {
1366             OPJ_UINT32 c;
1367             for (c = 0; c < NB_ELTS_V8; c++) {
1368                 tmp[c] *= 2;
1369             }
1370         }
1371     } else if (even) {
1372         OPJ_UINT32 c;
1373         OPJ_UINT32 i;
1374         i = 0;
1375         if (i + 1 < sn) {
1376             __m128i xmm_Si_0 = *(const __m128i*)(tmp + 4 * 0);
1377             __m128i xmm_Si_1 = *(const __m128i*)(tmp + 4 * 1);
1378             for (; i + 1 < sn; i++) {
1379                 __m128i xmm_Sip1_0 = *(const __m128i*)(tmp +
1380                                                        (i + 1) * 2 * NB_ELTS_V8 + 4 * 0);
1381                 __m128i xmm_Sip1_1 = *(const __m128i*)(tmp +
1382                                                        (i + 1) * 2 * NB_ELTS_V8 + 4 * 1);
1383                 __m128i xmm_Di_0 = *(const __m128i*)(tmp +
1384                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
1385                 __m128i xmm_Di_1 = *(const __m128i*)(tmp +
1386                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
1387                 xmm_Di_0 = _mm_sub_epi32(xmm_Di_0,
1388                                          _mm_srai_epi32(_mm_add_epi32(xmm_Si_0, xmm_Sip1_0), 1));
1389                 xmm_Di_1 = _mm_sub_epi32(xmm_Di_1,
1390                                          _mm_srai_epi32(_mm_add_epi32(xmm_Si_1, xmm_Sip1_1), 1));
1391                 *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 0) =  xmm_Di_0;
1392                 *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 1) =  xmm_Di_1;
1393                 xmm_Si_0 = xmm_Sip1_0;
1394                 xmm_Si_1 = xmm_Sip1_1;
1395             }
1396         }
1397         if (((height) % 2) == 0) {
1398             for (c = 0; c < NB_ELTS_V8; c++) {
1399                 OPJ_Dc(i) -= OPJ_Sc(i);
1400             }
1401         }
1402         for (c = 0; c < NB_ELTS_V8; c++) {
1403             OPJ_Sc(0) += (OPJ_Dc(0) + OPJ_Dc(0) + 2) >> 2;
1404         }
1405         i = 1;
1406         if (i < dn) {
1407             __m128i xmm_Dim1_0 = *(const __m128i*)(tmp + (1 +
1408                                                    (i - 1) * 2) * NB_ELTS_V8 + 4 * 0);
1409             __m128i xmm_Dim1_1 = *(const __m128i*)(tmp + (1 +
1410                                                    (i - 1) * 2) * NB_ELTS_V8 + 4 * 1);
1411             const __m128i xmm_two = _mm_set1_epi32(2);
1412             for (; i < dn; i++) {
1413                 __m128i xmm_Di_0 = *(const __m128i*)(tmp +
1414                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
1415                 __m128i xmm_Di_1 = *(const __m128i*)(tmp +
1416                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
1417                 __m128i xmm_Si_0 = *(const __m128i*)(tmp +
1418                                                      (i * 2) * NB_ELTS_V8 + 4 * 0);
1419                 __m128i xmm_Si_1 = *(const __m128i*)(tmp +
1420                                                      (i * 2) * NB_ELTS_V8 + 4 * 1);
1421                 xmm_Si_0 = _mm_add_epi32(xmm_Si_0,
1422                                          _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Dim1_0, xmm_Di_0), xmm_two), 2));
1423                 xmm_Si_1 = _mm_add_epi32(xmm_Si_1,
1424                                          _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Dim1_1, xmm_Di_1), xmm_two), 2));
1425                 *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Si_0;
1426                 *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Si_1;
1427                 xmm_Dim1_0 = xmm_Di_0;
1428                 xmm_Dim1_1 = xmm_Di_1;
1429             }
1430         }
1431         if (((height) % 2) == 1) {
1432             for (c = 0; c < NB_ELTS_V8; c++) {
1433                 OPJ_Sc(i) += (OPJ_Dc(i - 1) + OPJ_Dc(i - 1) + 2) >> 2;
1434             }
1435         }
1436     } else {
1437         OPJ_UINT32 c;
1438         OPJ_UINT32 i;
1439         for (c = 0; c < NB_ELTS_V8; c++) {
1440             OPJ_Sc(0) -= OPJ_Dc(0);
1441         }
1442         i = 1;
1443         if (i < sn) {
1444             __m128i xmm_Dim1_0 = *(const __m128i*)(tmp + (1 +
1445                                                    (i - 1) * 2) * NB_ELTS_V8 + 4 * 0);
1446             __m128i xmm_Dim1_1 = *(const __m128i*)(tmp + (1 +
1447                                                    (i - 1) * 2) * NB_ELTS_V8 + 4 * 1);
1448             for (; i < sn; i++) {
1449                 __m128i xmm_Di_0 = *(const __m128i*)(tmp +
1450                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
1451                 __m128i xmm_Di_1 = *(const __m128i*)(tmp +
1452                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
1453                 __m128i xmm_Si_0 = *(const __m128i*)(tmp +
1454                                                      (i * 2) * NB_ELTS_V8 + 4 * 0);
1455                 __m128i xmm_Si_1 = *(const __m128i*)(tmp +
1456                                                      (i * 2) * NB_ELTS_V8 + 4 * 1);
1457                 xmm_Si_0 = _mm_sub_epi32(xmm_Si_0,
1458                                          _mm_srai_epi32(_mm_add_epi32(xmm_Di_0, xmm_Dim1_0), 1));
1459                 xmm_Si_1 = _mm_sub_epi32(xmm_Si_1,
1460                                          _mm_srai_epi32(_mm_add_epi32(xmm_Di_1, xmm_Dim1_1), 1));
1461                 *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Si_0;
1462                 *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Si_1;
1463                 xmm_Dim1_0 = xmm_Di_0;
1464                 xmm_Dim1_1 = xmm_Di_1;
1465             }
1466         }
1467         if (((height) % 2) == 1) {
1468             for (c = 0; c < NB_ELTS_V8; c++) {
1469                 OPJ_Sc(i) -= OPJ_Dc(i - 1);
1470             }
1471         }
1472         i = 0;
1473         if (i + 1 < dn) {
1474             __m128i xmm_Si_0 = *((const __m128i*)(tmp + 4 * 0));
1475             __m128i xmm_Si_1 = *((const __m128i*)(tmp + 4 * 1));
1476             const __m128i xmm_two = _mm_set1_epi32(2);
1477             for (; i + 1 < dn; i++) {
1478                 __m128i xmm_Sip1_0 = *(const __m128i*)(tmp +
1479                                                        (i + 1) * 2 * NB_ELTS_V8 + 4 * 0);
1480                 __m128i xmm_Sip1_1 = *(const __m128i*)(tmp +
1481                                                        (i + 1) * 2 * NB_ELTS_V8 + 4 * 1);
1482                 __m128i xmm_Di_0 = *(const __m128i*)(tmp +
1483                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
1484                 __m128i xmm_Di_1 = *(const __m128i*)(tmp +
1485                                                      (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
1486                 xmm_Di_0 = _mm_add_epi32(xmm_Di_0,
1487                                          _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Si_0, xmm_Sip1_0), xmm_two), 2));
1488                 xmm_Di_1 = _mm_add_epi32(xmm_Di_1,
1489                                          _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Si_1, xmm_Sip1_1), xmm_two), 2));
1490                 *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Di_0;
1491                 *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Di_1;
1492                 xmm_Si_0 = xmm_Sip1_0;
1493                 xmm_Si_1 = xmm_Sip1_1;
1494             }
1495         }
1496         if (((height) % 2) == 0) {
1497             for (c = 0; c < NB_ELTS_V8; c++) {
1498                 OPJ_Dc(i) += (OPJ_Sc(i) + OPJ_Sc(i) + 2) >> 2;
1499             }
1500         }
1501     }
1502 #else
1503     if (even) {
1504         OPJ_UINT32 c;
1505         if (height > 1) {
1506             OPJ_UINT32 i;
1507             for (i = 0; i + 1 < sn; i++) {
1508                 for (c = 0; c < NB_ELTS_V8; c++) {
1509                     OPJ_Dc(i) -= (OPJ_Sc(i) + OPJ_Sc(i + 1)) >> 1;
1510                 }
1511             }
1512             if (((height) % 2) == 0) {
1513                 for (c = 0; c < NB_ELTS_V8; c++) {
1514                     OPJ_Dc(i) -= OPJ_Sc(i);
1515                 }
1516             }
1517             for (c = 0; c < NB_ELTS_V8; c++) {
1518                 OPJ_Sc(0) += (OPJ_Dc(0) + OPJ_Dc(0) + 2) >> 2;
1519             }
1520             for (i = 1; i < dn; i++) {
1521                 for (c = 0; c < NB_ELTS_V8; c++) {
1522                     OPJ_Sc(i) += (OPJ_Dc(i - 1) + OPJ_Dc(i) + 2) >> 2;
1523                 }
1524             }
1525             if (((height) % 2) == 1) {
1526                 for (c = 0; c < NB_ELTS_V8; c++) {
1527                     OPJ_Sc(i) += (OPJ_Dc(i - 1) + OPJ_Dc(i - 1) + 2) >> 2;
1528                 }
1529             }
1530         }
1531     } else {
1532         OPJ_UINT32 c;
1533         if (height == 1) {
1534             for (c = 0; c < NB_ELTS_V8; c++) {
1535                 OPJ_Sc(0) *= 2;
1536             }
1537         } else {
1538             OPJ_UINT32 i;
1539             for (c = 0; c < NB_ELTS_V8; c++) {
1540                 OPJ_Sc(0) -= OPJ_Dc(0);
1541             }
1542             for (i = 1; i < sn; i++) {
1543                 for (c = 0; c < NB_ELTS_V8; c++) {
1544                     OPJ_Sc(i) -= (OPJ_Dc(i) + OPJ_Dc(i - 1)) >> 1;
1545                 }
1546             }
1547             if (((height) % 2) == 1) {
1548                 for (c = 0; c < NB_ELTS_V8; c++) {
1549                     OPJ_Sc(i) -= OPJ_Dc(i - 1);
1550                 }
1551             }
1552             for (i = 0; i + 1 < dn; i++) {
1553                 for (c = 0; c < NB_ELTS_V8; c++) {
1554                     OPJ_Dc(i) += (OPJ_Sc(i) + OPJ_Sc(i + 1) + 2) >> 2;
1555                 }
1556             }
1557             if (((height) % 2) == 0) {
1558                 for (c = 0; c < NB_ELTS_V8; c++) {
1559                     OPJ_Dc(i) += (OPJ_Sc(i) + OPJ_Sc(i) + 2) >> 2;
1560                 }
1561             }
1562         }
1563     }
1564 #endif
1565 
1566     if (cols == NB_ELTS_V8) {
1567         opj_dwt_deinterleave_v_cols(tmp, array, (OPJ_INT32)dn, (OPJ_INT32)sn,
1568                                     stride_width, even ? 0 : 1, NB_ELTS_V8);
1569     } else {
1570         opj_dwt_deinterleave_v_cols(tmp, array, (OPJ_INT32)dn, (OPJ_INT32)sn,
1571                                     stride_width, even ? 0 : 1, cols);
1572     }
1573 }
1574 
opj_v8dwt_encode_step1(OPJ_FLOAT32 * fw,OPJ_UINT32 end,const OPJ_FLOAT32 cst)1575 static void opj_v8dwt_encode_step1(OPJ_FLOAT32* fw,
1576                                    OPJ_UINT32 end,
1577                                    const OPJ_FLOAT32 cst)
1578 {
1579     OPJ_UINT32 i;
1580 #ifdef __SSE__
1581     __m128* vw = (__m128*) fw;
1582     const __m128 vcst = _mm_set1_ps(cst);
1583     for (i = 0; i < end; ++i) {
1584         vw[0] = _mm_mul_ps(vw[0], vcst);
1585         vw[1] = _mm_mul_ps(vw[1], vcst);
1586         vw += 2 * (NB_ELTS_V8 * sizeof(OPJ_FLOAT32) / sizeof(__m128));
1587     }
1588 #else
1589     OPJ_UINT32 c;
1590     for (i = 0; i < end; ++i) {
1591         for (c = 0; c < NB_ELTS_V8; c++) {
1592             fw[i * 2 * NB_ELTS_V8 + c] *= cst;
1593         }
1594     }
1595 #endif
1596 }
1597 
opj_v8dwt_encode_step2(OPJ_FLOAT32 * fl,OPJ_FLOAT32 * fw,OPJ_UINT32 end,OPJ_UINT32 m,OPJ_FLOAT32 cst)1598 static void opj_v8dwt_encode_step2(OPJ_FLOAT32* fl, OPJ_FLOAT32* fw,
1599                                    OPJ_UINT32 end,
1600                                    OPJ_UINT32 m,
1601                                    OPJ_FLOAT32 cst)
1602 {
1603     OPJ_UINT32 i;
1604     OPJ_UINT32 imax = opj_uint_min(end, m);
1605 #ifdef __SSE__
1606     __m128* vw = (__m128*) fw;
1607     __m128 vcst = _mm_set1_ps(cst);
1608     if (imax > 0) {
1609         __m128* vl = (__m128*) fl;
1610         vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(_mm_add_ps(vl[0], vw[0]), vcst));
1611         vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(_mm_add_ps(vl[1], vw[1]), vcst));
1612         vw += 2 * (NB_ELTS_V8 * sizeof(OPJ_FLOAT32) / sizeof(__m128));
1613         i = 1;
1614 
1615         for (; i < imax; ++i) {
1616             vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(_mm_add_ps(vw[-4], vw[0]), vcst));
1617             vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(_mm_add_ps(vw[-3], vw[1]), vcst));
1618             vw += 2 * (NB_ELTS_V8 * sizeof(OPJ_FLOAT32) / sizeof(__m128));
1619         }
1620     }
1621     if (m < end) {
1622         assert(m + 1 == end);
1623         vcst = _mm_add_ps(vcst, vcst);
1624         vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(vw[-4], vcst));
1625         vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(vw[-3], vcst));
1626     }
1627 #else
1628     OPJ_INT32 c;
1629     if (imax > 0) {
1630         for (c = 0; c < NB_ELTS_V8; c++) {
1631             fw[-1 * NB_ELTS_V8 + c] += (fl[0 * NB_ELTS_V8 + c] + fw[0 * NB_ELTS_V8 + c]) *
1632                                        cst;
1633         }
1634         fw += 2 * NB_ELTS_V8;
1635         i = 1;
1636         for (; i < imax; ++i) {
1637             for (c = 0; c < NB_ELTS_V8; c++) {
1638                 fw[-1 * NB_ELTS_V8 + c] += (fw[-2 * NB_ELTS_V8 + c] + fw[0 * NB_ELTS_V8 + c]) *
1639                                            cst;
1640             }
1641             fw += 2 * NB_ELTS_V8;
1642         }
1643     }
1644     if (m < end) {
1645         assert(m + 1 == end);
1646         for (c = 0; c < NB_ELTS_V8; c++) {
1647             fw[-1 * NB_ELTS_V8 + c] += (2 * fw[-2 * NB_ELTS_V8 + c]) * cst;
1648         }
1649     }
1650 #endif
1651 }
1652 
1653 /* Forward 9-7 transform, for the vertical pass, processing cols columns */
1654 /* where cols <= NB_ELTS_V8 */
opj_dwt_encode_and_deinterleave_v_real(void * arrayIn,void * tmpIn,OPJ_UINT32 height,OPJ_BOOL even,OPJ_UINT32 stride_width,OPJ_UINT32 cols)1655 static void opj_dwt_encode_and_deinterleave_v_real(
1656     void *arrayIn,
1657     void *tmpIn,
1658     OPJ_UINT32 height,
1659     OPJ_BOOL even,
1660     OPJ_UINT32 stride_width,
1661     OPJ_UINT32 cols)
1662 {
1663     OPJ_FLOAT32* OPJ_RESTRICT array = (OPJ_FLOAT32 * OPJ_RESTRICT)arrayIn;
1664     OPJ_FLOAT32* OPJ_RESTRICT tmp = (OPJ_FLOAT32 * OPJ_RESTRICT)tmpIn;
1665     const OPJ_INT32 sn = (OPJ_INT32)((height + (even ? 1 : 0)) >> 1);
1666     const OPJ_INT32 dn = (OPJ_INT32)(height - (OPJ_UINT32)sn);
1667     OPJ_INT32 a, b;
1668 
1669     if (height == 1) {
1670         return;
1671     }
1672 
1673     opj_dwt_fetch_cols_vertical_pass(arrayIn, tmpIn, height, stride_width, cols);
1674 
1675     if (even) {
1676         a = 0;
1677         b = 1;
1678     } else {
1679         a = 1;
1680         b = 0;
1681     }
1682     opj_v8dwt_encode_step2(tmp + a * NB_ELTS_V8,
1683                            tmp + (b + 1) * NB_ELTS_V8,
1684                            (OPJ_UINT32)dn,
1685                            (OPJ_UINT32)opj_int_min(dn, sn - b),
1686                            opj_dwt_alpha);
1687     opj_v8dwt_encode_step2(tmp + b * NB_ELTS_V8,
1688                            tmp + (a + 1) * NB_ELTS_V8,
1689                            (OPJ_UINT32)sn,
1690                            (OPJ_UINT32)opj_int_min(sn, dn - a),
1691                            opj_dwt_beta);
1692     opj_v8dwt_encode_step2(tmp + a * NB_ELTS_V8,
1693                            tmp + (b + 1) * NB_ELTS_V8,
1694                            (OPJ_UINT32)dn,
1695                            (OPJ_UINT32)opj_int_min(dn, sn - b),
1696                            opj_dwt_gamma);
1697     opj_v8dwt_encode_step2(tmp + b * NB_ELTS_V8,
1698                            tmp + (a + 1) * NB_ELTS_V8,
1699                            (OPJ_UINT32)sn,
1700                            (OPJ_UINT32)opj_int_min(sn, dn - a),
1701                            opj_dwt_delta);
1702     opj_v8dwt_encode_step1(tmp + b * NB_ELTS_V8, (OPJ_UINT32)dn,
1703                            opj_K);
1704     opj_v8dwt_encode_step1(tmp + a * NB_ELTS_V8, (OPJ_UINT32)sn,
1705                            opj_invK);
1706 
1707 
1708     if (cols == NB_ELTS_V8) {
1709         opj_dwt_deinterleave_v_cols((OPJ_INT32*)tmp,
1710                                     (OPJ_INT32*)array,
1711                                     (OPJ_INT32)dn, (OPJ_INT32)sn,
1712                                     stride_width, even ? 0 : 1, NB_ELTS_V8);
1713     } else {
1714         opj_dwt_deinterleave_v_cols((OPJ_INT32*)tmp,
1715                                     (OPJ_INT32*)array,
1716                                     (OPJ_INT32)dn, (OPJ_INT32)sn,
1717                                     stride_width, even ? 0 : 1, cols);
1718     }
1719 }
1720 
1721 
1722 /* <summary>                            */
1723 /* Forward 5-3 wavelet transform in 2-D. */
1724 /* </summary>                           */
opj_dwt_encode_procedure(opj_thread_pool_t * tp,opj_tcd_tilecomp_t * tilec,opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v,opj_encode_and_deinterleave_h_one_row_fnptr_type p_encode_and_deinterleave_h_one_row)1725 static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_thread_pool_t* tp,
1726         opj_tcd_tilecomp_t * tilec,
1727         opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v,
1728         opj_encode_and_deinterleave_h_one_row_fnptr_type
1729         p_encode_and_deinterleave_h_one_row)
1730 {
1731     OPJ_INT32 i;
1732     OPJ_INT32 *bj = 00;
1733     OPJ_UINT32 w;
1734     OPJ_INT32 l;
1735 
1736     OPJ_SIZE_T l_data_size;
1737 
1738     opj_tcd_resolution_t * l_cur_res = 0;
1739     opj_tcd_resolution_t * l_last_res = 0;
1740     const int num_threads = opj_thread_pool_get_thread_count(tp);
1741     OPJ_INT32 * OPJ_RESTRICT tiledp = tilec->data;
1742 
1743     w = (OPJ_UINT32)(tilec->x1 - tilec->x0);
1744     l = (OPJ_INT32)tilec->numresolutions - 1;
1745 
1746     l_cur_res = tilec->resolutions + l;
1747     l_last_res = l_cur_res - 1;
1748 
1749     l_data_size = opj_dwt_max_resolution(tilec->resolutions, tilec->numresolutions);
1750     /* overflow check */
1751     if (l_data_size > (SIZE_MAX / (NB_ELTS_V8 * sizeof(OPJ_INT32)))) {
1752         /* FIXME event manager error callback */
1753         return OPJ_FALSE;
1754     }
1755     l_data_size *= NB_ELTS_V8 * sizeof(OPJ_INT32);
1756     bj = (OPJ_INT32*)opj_aligned_32_malloc(l_data_size);
1757     /* l_data_size is equal to 0 when numresolutions == 1 but bj is not used */
1758     /* in that case, so do not error out */
1759     if (l_data_size != 0 && ! bj) {
1760         return OPJ_FALSE;
1761     }
1762     i = l;
1763 
1764     while (i--) {
1765         OPJ_UINT32 j;
1766         OPJ_UINT32 rw;           /* width of the resolution level computed   */
1767         OPJ_UINT32 rh;           /* height of the resolution level computed  */
1768         OPJ_UINT32
1769         rw1;      /* width of the resolution level once lower than computed one                                       */
1770         OPJ_UINT32
1771         rh1;      /* height of the resolution level once lower than computed one                                      */
1772         OPJ_INT32 cas_col;  /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */
1773         OPJ_INT32 cas_row;  /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering   */
1774         OPJ_INT32 dn, sn;
1775 
1776         rw  = (OPJ_UINT32)(l_cur_res->x1 - l_cur_res->x0);
1777         rh  = (OPJ_UINT32)(l_cur_res->y1 - l_cur_res->y0);
1778         rw1 = (OPJ_UINT32)(l_last_res->x1 - l_last_res->x0);
1779         rh1 = (OPJ_UINT32)(l_last_res->y1 - l_last_res->y0);
1780 
1781         cas_row = l_cur_res->x0 & 1;
1782         cas_col = l_cur_res->y0 & 1;
1783 
1784         sn = (OPJ_INT32)rh1;
1785         dn = (OPJ_INT32)(rh - rh1);
1786 
1787         /* Perform vertical pass */
1788         if (num_threads <= 1 || rw < 2 * NB_ELTS_V8) {
1789             for (j = 0; j + NB_ELTS_V8 - 1 < rw; j += NB_ELTS_V8) {
1790                 p_encode_and_deinterleave_v(tiledp + j,
1791                                             bj,
1792                                             rh,
1793                                             cas_col == 0,
1794                                             w,
1795                                             NB_ELTS_V8);
1796             }
1797             if (j < rw) {
1798                 p_encode_and_deinterleave_v(tiledp + j,
1799                                             bj,
1800                                             rh,
1801                                             cas_col == 0,
1802                                             w,
1803                                             rw - j);
1804             }
1805         }  else {
1806             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
1807             OPJ_UINT32 step_j;
1808 
1809             if (rw < num_jobs) {
1810                 num_jobs = rw;
1811             }
1812             step_j = ((rw / num_jobs) / NB_ELTS_V8) * NB_ELTS_V8;
1813 
1814             for (j = 0; j < num_jobs; j++) {
1815                 opj_dwt_encode_v_job_t* job;
1816 
1817                 job = (opj_dwt_encode_v_job_t*) opj_malloc(sizeof(opj_dwt_encode_v_job_t));
1818                 if (!job) {
1819                     opj_thread_pool_wait_completion(tp, 0);
1820                     opj_aligned_free(bj);
1821                     return OPJ_FALSE;
1822                 }
1823                 job->v.mem = (OPJ_INT32*)opj_aligned_32_malloc(l_data_size);
1824                 if (!job->v.mem) {
1825                     opj_thread_pool_wait_completion(tp, 0);
1826                     opj_free(job);
1827                     opj_aligned_free(bj);
1828                     return OPJ_FALSE;
1829                 }
1830                 job->v.dn = dn;
1831                 job->v.sn = sn;
1832                 job->v.cas = cas_col;
1833                 job->rh = rh;
1834                 job->w = w;
1835                 job->tiledp = tiledp;
1836                 job->min_j = j * step_j;
1837                 job->max_j = (j + 1 == num_jobs) ? rw : (j + 1) * step_j;
1838                 job->p_encode_and_deinterleave_v = p_encode_and_deinterleave_v;
1839                 opj_thread_pool_submit_job(tp, opj_dwt_encode_v_func, job);
1840             }
1841             opj_thread_pool_wait_completion(tp, 0);
1842         }
1843 
1844         sn = (OPJ_INT32)rw1;
1845         dn = (OPJ_INT32)(rw - rw1);
1846 
1847         /* Perform horizontal pass */
1848         if (num_threads <= 1 || rh <= 1) {
1849             for (j = 0; j < rh; j++) {
1850                 OPJ_INT32* OPJ_RESTRICT aj = tiledp + j * w;
1851                 (*p_encode_and_deinterleave_h_one_row)(aj, bj, rw,
1852                                                        cas_row == 0 ? OPJ_TRUE : OPJ_FALSE);
1853             }
1854         }  else {
1855             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
1856             OPJ_UINT32 step_j;
1857 
1858             if (rh < num_jobs) {
1859                 num_jobs = rh;
1860             }
1861             step_j = (rh / num_jobs);
1862 
1863             for (j = 0; j < num_jobs; j++) {
1864                 opj_dwt_encode_h_job_t* job;
1865 
1866                 job = (opj_dwt_encode_h_job_t*) opj_malloc(sizeof(opj_dwt_encode_h_job_t));
1867                 if (!job) {
1868                     opj_thread_pool_wait_completion(tp, 0);
1869                     opj_aligned_free(bj);
1870                     return OPJ_FALSE;
1871                 }
1872                 job->h.mem = (OPJ_INT32*)opj_aligned_32_malloc(l_data_size);
1873                 if (!job->h.mem) {
1874                     opj_thread_pool_wait_completion(tp, 0);
1875                     opj_free(job);
1876                     opj_aligned_free(bj);
1877                     return OPJ_FALSE;
1878                 }
1879                 job->h.dn = dn;
1880                 job->h.sn = sn;
1881                 job->h.cas = cas_row;
1882                 job->rw = rw;
1883                 job->w = w;
1884                 job->tiledp = tiledp;
1885                 job->min_j = j * step_j;
1886                 job->max_j = (j + 1U) * step_j; /* this can overflow */
1887                 if (j == (num_jobs - 1U)) {  /* this will take care of the overflow */
1888                     job->max_j = rh;
1889                 }
1890                 job->p_function = p_encode_and_deinterleave_h_one_row;
1891                 opj_thread_pool_submit_job(tp, opj_dwt_encode_h_func, job);
1892             }
1893             opj_thread_pool_wait_completion(tp, 0);
1894         }
1895 
1896         l_cur_res = l_last_res;
1897 
1898         --l_last_res;
1899     }
1900 
1901     opj_aligned_free(bj);
1902     return OPJ_TRUE;
1903 }
1904 
1905 /* Forward 5-3 wavelet transform in 2-D. */
1906 /* </summary>                           */
opj_dwt_encode(opj_tcd_t * p_tcd,opj_tcd_tilecomp_t * tilec)1907 OPJ_BOOL opj_dwt_encode(opj_tcd_t *p_tcd,
1908                         opj_tcd_tilecomp_t * tilec)
1909 {
1910     return opj_dwt_encode_procedure(p_tcd->thread_pool, tilec,
1911                                     opj_dwt_encode_and_deinterleave_v,
1912                                     opj_dwt_encode_and_deinterleave_h_one_row);
1913 }
1914 
1915 /* <summary>                            */
1916 /* Inverse 5-3 wavelet transform in 2-D. */
1917 /* </summary>                           */
opj_dwt_decode(opj_tcd_t * p_tcd,opj_tcd_tilecomp_t * tilec,OPJ_UINT32 numres)1918 OPJ_BOOL opj_dwt_decode(opj_tcd_t *p_tcd, opj_tcd_tilecomp_t* tilec,
1919                         OPJ_UINT32 numres)
1920 {
1921     if (p_tcd->whole_tile_decoding) {
1922         return opj_dwt_decode_tile(p_tcd->thread_pool, tilec, numres);
1923     } else {
1924         return opj_dwt_decode_partial_tile(tilec, numres);
1925     }
1926 }
1927 
1928 /* <summary>                */
1929 /* Get norm of 5-3 wavelet. */
1930 /* </summary>               */
opj_dwt_getnorm(OPJ_UINT32 level,OPJ_UINT32 orient)1931 OPJ_FLOAT64 opj_dwt_getnorm(OPJ_UINT32 level, OPJ_UINT32 orient)
1932 {
1933     /* FIXME ! This is just a band-aid to avoid a buffer overflow */
1934     /* but the array should really be extended up to 33 resolution levels */
1935     /* See https://github.com/uclouvain/openjpeg/issues/493 */
1936     if (orient == 0 && level >= 10) {
1937         level = 9;
1938     } else if (orient > 0 && level >= 9) {
1939         level = 8;
1940     }
1941     return opj_dwt_norms[orient][level];
1942 }
1943 
1944 /* <summary>                             */
1945 /* Forward 9-7 wavelet transform in 2-D. */
1946 /* </summary>                            */
opj_dwt_encode_real(opj_tcd_t * p_tcd,opj_tcd_tilecomp_t * tilec)1947 OPJ_BOOL opj_dwt_encode_real(opj_tcd_t *p_tcd,
1948                              opj_tcd_tilecomp_t * tilec)
1949 {
1950     return opj_dwt_encode_procedure(p_tcd->thread_pool, tilec,
1951                                     opj_dwt_encode_and_deinterleave_v_real,
1952                                     opj_dwt_encode_and_deinterleave_h_one_row_real);
1953 }
1954 
1955 /* <summary>                */
1956 /* Get norm of 9-7 wavelet. */
1957 /* </summary>               */
opj_dwt_getnorm_real(OPJ_UINT32 level,OPJ_UINT32 orient)1958 OPJ_FLOAT64 opj_dwt_getnorm_real(OPJ_UINT32 level, OPJ_UINT32 orient)
1959 {
1960     /* FIXME ! This is just a band-aid to avoid a buffer overflow */
1961     /* but the array should really be extended up to 33 resolution levels */
1962     /* See https://github.com/uclouvain/openjpeg/issues/493 */
1963     if (orient == 0 && level >= 10) {
1964         level = 9;
1965     } else if (orient > 0 && level >= 9) {
1966         level = 8;
1967     }
1968     return opj_dwt_norms_real[orient][level];
1969 }
1970 
opj_dwt_calc_explicit_stepsizes(opj_tccp_t * tccp,OPJ_UINT32 prec)1971 void opj_dwt_calc_explicit_stepsizes(opj_tccp_t * tccp, OPJ_UINT32 prec)
1972 {
1973     OPJ_UINT32 numbands, bandno;
1974     numbands = 3 * tccp->numresolutions - 2;
1975     for (bandno = 0; bandno < numbands; bandno++) {
1976         OPJ_FLOAT64 stepsize;
1977         OPJ_UINT32 resno, level, orient, gain;
1978 
1979         resno = (bandno == 0) ? 0 : ((bandno - 1) / 3 + 1);
1980         orient = (bandno == 0) ? 0 : ((bandno - 1) % 3 + 1);
1981         level = tccp->numresolutions - 1 - resno;
1982         gain = (tccp->qmfbid == 0) ? 0 : ((orient == 0) ? 0 : (((orient == 1) ||
1983                                           (orient == 2)) ? 1 : 2));
1984         if (tccp->qntsty == J2K_CCP_QNTSTY_NOQNT) {
1985             stepsize = 1.0;
1986         } else {
1987             OPJ_FLOAT64 norm = opj_dwt_getnorm_real(level, orient);
1988             stepsize = (1 << (gain)) / norm;
1989         }
1990         opj_dwt_encode_stepsize((OPJ_INT32) floor(stepsize * 8192.0),
1991                                 (OPJ_INT32)(prec + gain), &tccp->stepsizes[bandno]);
1992     }
1993 }
1994 
1995 /* <summary>                             */
1996 /* Determine maximum computed resolution level for inverse wavelet transform */
1997 /* </summary>                            */
opj_dwt_max_resolution(opj_tcd_resolution_t * OPJ_RESTRICT r,OPJ_UINT32 i)1998 static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r,
1999         OPJ_UINT32 i)
2000 {
2001     OPJ_UINT32 mr   = 0;
2002     OPJ_UINT32 w;
2003     while (--i) {
2004         ++r;
2005         if (mr < (w = (OPJ_UINT32)(r->x1 - r->x0))) {
2006             mr = w ;
2007         }
2008         if (mr < (w = (OPJ_UINT32)(r->y1 - r->y0))) {
2009             mr = w ;
2010         }
2011     }
2012     return mr ;
2013 }
2014 
2015 typedef struct {
2016     opj_dwt_t h;
2017     OPJ_UINT32 rw;
2018     OPJ_UINT32 w;
2019     OPJ_INT32 * OPJ_RESTRICT tiledp;
2020     OPJ_UINT32 min_j;
2021     OPJ_UINT32 max_j;
2022 } opj_dwt_decode_h_job_t;
2023 
opj_dwt_decode_h_func(void * user_data,opj_tls_t * tls)2024 static void opj_dwt_decode_h_func(void* user_data, opj_tls_t* tls)
2025 {
2026     OPJ_UINT32 j;
2027     opj_dwt_decode_h_job_t* job;
2028     (void)tls;
2029 
2030     job = (opj_dwt_decode_h_job_t*)user_data;
2031     for (j = job->min_j; j < job->max_j; j++) {
2032         opj_idwt53_h(&job->h, &job->tiledp[j * job->w]);
2033     }
2034 
2035     opj_aligned_free(job->h.mem);
2036     opj_free(job);
2037 }
2038 
2039 typedef struct {
2040     opj_dwt_t v;
2041     OPJ_UINT32 rh;
2042     OPJ_UINT32 w;
2043     OPJ_INT32 * OPJ_RESTRICT tiledp;
2044     OPJ_UINT32 min_j;
2045     OPJ_UINT32 max_j;
2046 } opj_dwt_decode_v_job_t;
2047 
opj_dwt_decode_v_func(void * user_data,opj_tls_t * tls)2048 static void opj_dwt_decode_v_func(void* user_data, opj_tls_t* tls)
2049 {
2050     OPJ_UINT32 j;
2051     opj_dwt_decode_v_job_t* job;
2052     (void)tls;
2053 
2054     job = (opj_dwt_decode_v_job_t*)user_data;
2055     for (j = job->min_j; j + PARALLEL_COLS_53 <= job->max_j;
2056             j += PARALLEL_COLS_53) {
2057         opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w,
2058                      PARALLEL_COLS_53);
2059     }
2060     if (j < job->max_j)
2061         opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w,
2062                      (OPJ_INT32)(job->max_j - j));
2063 
2064     opj_aligned_free(job->v.mem);
2065     opj_free(job);
2066 }
2067 
2068 
2069 /* <summary>                            */
2070 /* Inverse wavelet transform in 2-D.    */
2071 /* </summary>                           */
opj_dwt_decode_tile(opj_thread_pool_t * tp,const opj_tcd_tilecomp_t * tilec,OPJ_UINT32 numres)2072 static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
2073         const opj_tcd_tilecomp_t* tilec, OPJ_UINT32 numres)
2074 {
2075     opj_dwt_t h;
2076     opj_dwt_t v;
2077 
2078     opj_tcd_resolution_t* tr = tilec->resolutions;
2079 
2080     OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 -
2081                                  tr->x0);  /* width of the resolution level computed */
2082     OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
2083                                  tr->y0);  /* height of the resolution level computed */
2084 
2085     OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions -
2086                                                                1].x1 -
2087                                 tilec->resolutions[tilec->minimum_num_resolutions - 1].x0);
2088     OPJ_SIZE_T h_mem_size;
2089     int num_threads;
2090 
2091     if (numres == 1U) {
2092         return OPJ_TRUE;
2093     }
2094     num_threads = opj_thread_pool_get_thread_count(tp);
2095     h.mem_count = opj_dwt_max_resolution(tr, numres);
2096     /* overflow check */
2097     if (h.mem_count > (SIZE_MAX / PARALLEL_COLS_53 / sizeof(OPJ_INT32))) {
2098         /* FIXME event manager error callback */
2099         return OPJ_FALSE;
2100     }
2101     /* We need PARALLEL_COLS_53 times the height of the array, */
2102     /* since for the vertical pass */
2103     /* we process PARALLEL_COLS_53 columns at a time */
2104     h_mem_size = h.mem_count * PARALLEL_COLS_53 * sizeof(OPJ_INT32);
2105     h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
2106     if (! h.mem) {
2107         /* FIXME event manager error callback */
2108         return OPJ_FALSE;
2109     }
2110 
2111     v.mem_count = h.mem_count;
2112     v.mem = h.mem;
2113 
2114     while (--numres) {
2115         OPJ_INT32 * OPJ_RESTRICT tiledp = tilec->data;
2116         OPJ_UINT32 j;
2117 
2118         ++tr;
2119         h.sn = (OPJ_INT32)rw;
2120         v.sn = (OPJ_INT32)rh;
2121 
2122         rw = (OPJ_UINT32)(tr->x1 - tr->x0);
2123         rh = (OPJ_UINT32)(tr->y1 - tr->y0);
2124 
2125         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
2126         h.cas = tr->x0 % 2;
2127 
2128         if (num_threads <= 1 || rh <= 1) {
2129             for (j = 0; j < rh; ++j) {
2130                 opj_idwt53_h(&h, &tiledp[(OPJ_SIZE_T)j * w]);
2131             }
2132         } else {
2133             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
2134             OPJ_UINT32 step_j;
2135 
2136             if (rh < num_jobs) {
2137                 num_jobs = rh;
2138             }
2139             step_j = (rh / num_jobs);
2140 
2141             for (j = 0; j < num_jobs; j++) {
2142                 opj_dwt_decode_h_job_t* job;
2143 
2144                 job = (opj_dwt_decode_h_job_t*) opj_malloc(sizeof(opj_dwt_decode_h_job_t));
2145                 if (!job) {
2146                     /* It would be nice to fallback to single thread case, but */
2147                     /* unfortunately some jobs may be launched and have modified */
2148                     /* tiledp, so it is not practical to recover from that error */
2149                     /* FIXME event manager error callback */
2150                     opj_thread_pool_wait_completion(tp, 0);
2151                     opj_aligned_free(h.mem);
2152                     return OPJ_FALSE;
2153                 }
2154                 job->h = h;
2155                 job->rw = rw;
2156                 job->w = w;
2157                 job->tiledp = tiledp;
2158                 job->min_j = j * step_j;
2159                 job->max_j = (j + 1U) * step_j; /* this can overflow */
2160                 if (j == (num_jobs - 1U)) {  /* this will take care of the overflow */
2161                     job->max_j = rh;
2162                 }
2163                 job->h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
2164                 if (!job->h.mem) {
2165                     /* FIXME event manager error callback */
2166                     opj_thread_pool_wait_completion(tp, 0);
2167                     opj_free(job);
2168                     opj_aligned_free(h.mem);
2169                     return OPJ_FALSE;
2170                 }
2171                 opj_thread_pool_submit_job(tp, opj_dwt_decode_h_func, job);
2172             }
2173             opj_thread_pool_wait_completion(tp, 0);
2174         }
2175 
2176         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
2177         v.cas = tr->y0 % 2;
2178 
2179         if (num_threads <= 1 || rw <= 1) {
2180             for (j = 0; j + PARALLEL_COLS_53 <= rw;
2181                     j += PARALLEL_COLS_53) {
2182                 opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, PARALLEL_COLS_53);
2183             }
2184             if (j < rw) {
2185                 opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, (OPJ_INT32)(rw - j));
2186             }
2187         } else {
2188             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
2189             OPJ_UINT32 step_j;
2190 
2191             if (rw < num_jobs) {
2192                 num_jobs = rw;
2193             }
2194             step_j = (rw / num_jobs);
2195 
2196             for (j = 0; j < num_jobs; j++) {
2197                 opj_dwt_decode_v_job_t* job;
2198 
2199                 job = (opj_dwt_decode_v_job_t*) opj_malloc(sizeof(opj_dwt_decode_v_job_t));
2200                 if (!job) {
2201                     /* It would be nice to fallback to single thread case, but */
2202                     /* unfortunately some jobs may be launched and have modified */
2203                     /* tiledp, so it is not practical to recover from that error */
2204                     /* FIXME event manager error callback */
2205                     opj_thread_pool_wait_completion(tp, 0);
2206                     opj_aligned_free(v.mem);
2207                     return OPJ_FALSE;
2208                 }
2209                 job->v = v;
2210                 job->rh = rh;
2211                 job->w = w;
2212                 job->tiledp = tiledp;
2213                 job->min_j = j * step_j;
2214                 job->max_j = (j + 1U) * step_j; /* this can overflow */
2215                 if (j == (num_jobs - 1U)) {  /* this will take care of the overflow */
2216                     job->max_j = rw;
2217                 }
2218                 job->v.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
2219                 if (!job->v.mem) {
2220                     /* FIXME event manager error callback */
2221                     opj_thread_pool_wait_completion(tp, 0);
2222                     opj_free(job);
2223                     opj_aligned_free(v.mem);
2224                     return OPJ_FALSE;
2225                 }
2226                 opj_thread_pool_submit_job(tp, opj_dwt_decode_v_func, job);
2227             }
2228             opj_thread_pool_wait_completion(tp, 0);
2229         }
2230     }
2231     opj_aligned_free(h.mem);
2232     return OPJ_TRUE;
2233 }
2234 
opj_dwt_interleave_partial_h(OPJ_INT32 * dest,OPJ_INT32 cas,opj_sparse_array_int32_t * sa,OPJ_UINT32 sa_line,OPJ_UINT32 sn,OPJ_UINT32 win_l_x0,OPJ_UINT32 win_l_x1,OPJ_UINT32 win_h_x0,OPJ_UINT32 win_h_x1)2235 static void opj_dwt_interleave_partial_h(OPJ_INT32 *dest,
2236         OPJ_INT32 cas,
2237         opj_sparse_array_int32_t* sa,
2238         OPJ_UINT32 sa_line,
2239         OPJ_UINT32 sn,
2240         OPJ_UINT32 win_l_x0,
2241         OPJ_UINT32 win_l_x1,
2242         OPJ_UINT32 win_h_x0,
2243         OPJ_UINT32 win_h_x1)
2244 {
2245     OPJ_BOOL ret;
2246     ret = opj_sparse_array_int32_read(sa,
2247                                       win_l_x0, sa_line,
2248                                       win_l_x1, sa_line + 1,
2249                                       dest + cas + 2 * win_l_x0,
2250                                       2, 0, OPJ_TRUE);
2251     assert(ret);
2252     ret = opj_sparse_array_int32_read(sa,
2253                                       sn + win_h_x0, sa_line,
2254                                       sn + win_h_x1, sa_line + 1,
2255                                       dest + 1 - cas + 2 * win_h_x0,
2256                                       2, 0, OPJ_TRUE);
2257     assert(ret);
2258     OPJ_UNUSED(ret);
2259 }
2260 
2261 
opj_dwt_interleave_partial_v(OPJ_INT32 * dest,OPJ_INT32 cas,opj_sparse_array_int32_t * sa,OPJ_UINT32 sa_col,OPJ_UINT32 nb_cols,OPJ_UINT32 sn,OPJ_UINT32 win_l_y0,OPJ_UINT32 win_l_y1,OPJ_UINT32 win_h_y0,OPJ_UINT32 win_h_y1)2262 static void opj_dwt_interleave_partial_v(OPJ_INT32 *dest,
2263         OPJ_INT32 cas,
2264         opj_sparse_array_int32_t* sa,
2265         OPJ_UINT32 sa_col,
2266         OPJ_UINT32 nb_cols,
2267         OPJ_UINT32 sn,
2268         OPJ_UINT32 win_l_y0,
2269         OPJ_UINT32 win_l_y1,
2270         OPJ_UINT32 win_h_y0,
2271         OPJ_UINT32 win_h_y1)
2272 {
2273     OPJ_BOOL ret;
2274     ret  = opj_sparse_array_int32_read(sa,
2275                                        sa_col, win_l_y0,
2276                                        sa_col + nb_cols, win_l_y1,
2277                                        dest + cas * 4 + 2 * 4 * win_l_y0,
2278                                        1, 2 * 4, OPJ_TRUE);
2279     assert(ret);
2280     ret = opj_sparse_array_int32_read(sa,
2281                                       sa_col, sn + win_h_y0,
2282                                       sa_col + nb_cols, sn + win_h_y1,
2283                                       dest + (1 - cas) * 4 + 2 * 4 * win_h_y0,
2284                                       1, 2 * 4, OPJ_TRUE);
2285     assert(ret);
2286     OPJ_UNUSED(ret);
2287 }
2288 
opj_dwt_decode_partial_1(OPJ_INT32 * a,OPJ_SIZE_T a_count,OPJ_INT32 dn,OPJ_INT32 sn,OPJ_INT32 cas,OPJ_INT32 win_l_x0,OPJ_INT32 win_l_x1,OPJ_INT32 win_h_x0,OPJ_INT32 win_h_x1)2289 static void opj_dwt_decode_partial_1(OPJ_INT32 *a, OPJ_SIZE_T a_count,
2290                                      OPJ_INT32 dn, OPJ_INT32 sn,
2291                                      OPJ_INT32 cas,
2292                                      OPJ_INT32 win_l_x0,
2293                                      OPJ_INT32 win_l_x1,
2294                                      OPJ_INT32 win_h_x0,
2295                                      OPJ_INT32 win_h_x1)
2296 {
2297     OPJ_INT32 i;
2298 
2299     if (!cas) {
2300         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
2301 
2302             /* Naive version is :
2303             for (i = win_l_x0; i < i_max; i++) {
2304                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
2305             }
2306             for (i = win_h_x0; i < win_h_x1; i++) {
2307                 OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
2308             }
2309             but the compiler doesn't manage to unroll it to avoid bound
2310             checking in OPJ_S_ and OPJ_D_ macros
2311             */
2312 
2313             i = win_l_x0;
2314             if (i < win_l_x1) {
2315                 OPJ_INT32 i_max;
2316 
2317                 /* Left-most case */
2318                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
2319                 i ++;
2320 
2321                 i_max = win_l_x1;
2322                 if (i_max > dn) {
2323                     i_max = dn;
2324                 }
2325                 for (; i < i_max; i++) {
2326                     /* No bound checking */
2327                     OPJ_S(i) -= (OPJ_D(i - 1) + OPJ_D(i) + 2) >> 2;
2328                 }
2329                 for (; i < win_l_x1; i++) {
2330                     /* Right-most case */
2331                     OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
2332                 }
2333             }
2334 
2335             i = win_h_x0;
2336             if (i < win_h_x1) {
2337                 OPJ_INT32 i_max = win_h_x1;
2338                 if (i_max >= sn) {
2339                     i_max = sn - 1;
2340                 }
2341                 for (; i < i_max; i++) {
2342                     /* No bound checking */
2343                     OPJ_D(i) += (OPJ_S(i) + OPJ_S(i + 1)) >> 1;
2344                 }
2345                 for (; i < win_h_x1; i++) {
2346                     /* Right-most case */
2347                     OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
2348                 }
2349             }
2350         }
2351     } else {
2352         if (!sn  && dn == 1) {        /* NEW :  CASE ONE ELEMENT */
2353             OPJ_S(0) /= 2;
2354         } else {
2355             for (i = win_l_x0; i < win_l_x1; i++) {
2356                 OPJ_D(i) = opj_int_sub_no_overflow(OPJ_D(i),
2357                                                    opj_int_add_no_overflow(opj_int_add_no_overflow(OPJ_SS_(i), OPJ_SS_(i + 1)),
2358                                                            2) >> 2);
2359             }
2360             for (i = win_h_x0; i < win_h_x1; i++) {
2361                 OPJ_S(i) = opj_int_add_no_overflow(OPJ_S(i),
2362                                                    opj_int_add_no_overflow(OPJ_DD_(i), OPJ_DD_(i - 1)) >> 1);
2363             }
2364         }
2365     }
2366 }
2367 
2368 #define OPJ_S_off(i,off) a[(OPJ_UINT32)(i)*2*4+off]
2369 #define OPJ_D_off(i,off) a[(1+(OPJ_UINT32)(i)*2)*4+off]
2370 #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)))
2371 #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)))
2372 #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)))
2373 #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)))
2374 
opj_dwt_decode_partial_1_parallel(OPJ_INT32 * a,OPJ_UINT32 nb_cols,OPJ_INT32 dn,OPJ_INT32 sn,OPJ_INT32 cas,OPJ_INT32 win_l_x0,OPJ_INT32 win_l_x1,OPJ_INT32 win_h_x0,OPJ_INT32 win_h_x1)2375 static void opj_dwt_decode_partial_1_parallel(OPJ_INT32 *a,
2376         OPJ_UINT32 nb_cols,
2377         OPJ_INT32 dn, OPJ_INT32 sn,
2378         OPJ_INT32 cas,
2379         OPJ_INT32 win_l_x0,
2380         OPJ_INT32 win_l_x1,
2381         OPJ_INT32 win_h_x0,
2382         OPJ_INT32 win_h_x1)
2383 {
2384     OPJ_INT32 i;
2385     OPJ_UINT32 off;
2386 
2387     (void)nb_cols;
2388 
2389     if (!cas) {
2390         if ((dn > 0) || (sn > 1)) { /* NEW :  CASE ONE ELEMENT */
2391 
2392             /* Naive version is :
2393             for (i = win_l_x0; i < i_max; i++) {
2394                 OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
2395             }
2396             for (i = win_h_x0; i < win_h_x1; i++) {
2397                 OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
2398             }
2399             but the compiler doesn't manage to unroll it to avoid bound
2400             checking in OPJ_S_ and OPJ_D_ macros
2401             */
2402 
2403             i = win_l_x0;
2404             if (i < win_l_x1) {
2405                 OPJ_INT32 i_max;
2406 
2407                 /* Left-most case */
2408                 for (off = 0; off < 4; off++) {
2409                     OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2;
2410                 }
2411                 i ++;
2412 
2413                 i_max = win_l_x1;
2414                 if (i_max > dn) {
2415                     i_max = dn;
2416                 }
2417 
2418 #ifdef __SSE2__
2419                 if (i + 1 < i_max) {
2420                     const __m128i two = _mm_set1_epi32(2);
2421                     __m128i Dm1 = _mm_load_si128((__m128i * const)(a + 4 + (i - 1) * 8));
2422                     for (; i + 1 < i_max; i += 2) {
2423                         /* No bound checking */
2424                         __m128i S = _mm_load_si128((__m128i * const)(a + i * 8));
2425                         __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8));
2426                         __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8));
2427                         __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8));
2428                         S = _mm_sub_epi32(S,
2429                                           _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(Dm1, D), two), 2));
2430                         S1 = _mm_sub_epi32(S1,
2431                                            _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(D, D1), two), 2));
2432                         _mm_store_si128((__m128i*)(a + i * 8), S);
2433                         _mm_store_si128((__m128i*)(a + (i + 1) * 8), S1);
2434                         Dm1 = D1;
2435                     }
2436                 }
2437 #endif
2438 
2439                 for (; i < i_max; i++) {
2440                     /* No bound checking */
2441                     for (off = 0; off < 4; off++) {
2442                         OPJ_S_off(i, off) -= (OPJ_D_off(i - 1, off) + OPJ_D_off(i, off) + 2) >> 2;
2443                     }
2444                 }
2445                 for (; i < win_l_x1; i++) {
2446                     /* Right-most case */
2447                     for (off = 0; off < 4; off++) {
2448                         OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2;
2449                     }
2450                 }
2451             }
2452 
2453             i = win_h_x0;
2454             if (i < win_h_x1) {
2455                 OPJ_INT32 i_max = win_h_x1;
2456                 if (i_max >= sn) {
2457                     i_max = sn - 1;
2458                 }
2459 
2460 #ifdef __SSE2__
2461                 if (i + 1 < i_max) {
2462                     __m128i S =  _mm_load_si128((__m128i * const)(a + i * 8));
2463                     for (; i + 1 < i_max; i += 2) {
2464                         /* No bound checking */
2465                         __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8));
2466                         __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8));
2467                         __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8));
2468                         __m128i S2 = _mm_load_si128((__m128i * const)(a + (i + 2) * 8));
2469                         D = _mm_add_epi32(D, _mm_srai_epi32(_mm_add_epi32(S, S1), 1));
2470                         D1 = _mm_add_epi32(D1, _mm_srai_epi32(_mm_add_epi32(S1, S2), 1));
2471                         _mm_store_si128((__m128i*)(a + 4 + i * 8), D);
2472                         _mm_store_si128((__m128i*)(a + 4 + (i + 1) * 8), D1);
2473                         S = S2;
2474                     }
2475                 }
2476 #endif
2477 
2478                 for (; i < i_max; i++) {
2479                     /* No bound checking */
2480                     for (off = 0; off < 4; off++) {
2481                         OPJ_D_off(i, off) += (OPJ_S_off(i, off) + OPJ_S_off(i + 1, off)) >> 1;
2482                     }
2483                 }
2484                 for (; i < win_h_x1; i++) {
2485                     /* Right-most case */
2486                     for (off = 0; off < 4; off++) {
2487                         OPJ_D_off(i, off) += (OPJ_S__off(i, off) + OPJ_S__off(i + 1, off)) >> 1;
2488                     }
2489                 }
2490             }
2491         }
2492     } else {
2493         if (!sn  && dn == 1) {        /* NEW :  CASE ONE ELEMENT */
2494             for (off = 0; off < 4; off++) {
2495                 OPJ_S_off(0, off) /= 2;
2496             }
2497         } else {
2498             for (i = win_l_x0; i < win_l_x1; i++) {
2499                 for (off = 0; off < 4; off++) {
2500                     OPJ_D_off(i, off) = opj_int_sub_no_overflow(
2501                                             OPJ_D_off(i, off),
2502                                             opj_int_add_no_overflow(
2503                                                 opj_int_add_no_overflow(OPJ_SS__off(i, off), OPJ_SS__off(i + 1, off)), 2) >> 2);
2504                 }
2505             }
2506             for (i = win_h_x0; i < win_h_x1; i++) {
2507                 for (off = 0; off < 4; off++) {
2508                     OPJ_S_off(i, off) = opj_int_add_no_overflow(
2509                                             OPJ_S_off(i, off),
2510                                             opj_int_add_no_overflow(OPJ_DD__off(i, off), OPJ_DD__off(i - 1, off)) >> 1);
2511                 }
2512             }
2513         }
2514     }
2515 }
2516 
opj_dwt_get_band_coordinates(opj_tcd_tilecomp_t * tilec,OPJ_UINT32 resno,OPJ_UINT32 bandno,OPJ_UINT32 tcx0,OPJ_UINT32 tcy0,OPJ_UINT32 tcx1,OPJ_UINT32 tcy1,OPJ_UINT32 * tbx0,OPJ_UINT32 * tby0,OPJ_UINT32 * tbx1,OPJ_UINT32 * tby1)2517 static void opj_dwt_get_band_coordinates(opj_tcd_tilecomp_t* tilec,
2518         OPJ_UINT32 resno,
2519         OPJ_UINT32 bandno,
2520         OPJ_UINT32 tcx0,
2521         OPJ_UINT32 tcy0,
2522         OPJ_UINT32 tcx1,
2523         OPJ_UINT32 tcy1,
2524         OPJ_UINT32* tbx0,
2525         OPJ_UINT32* tby0,
2526         OPJ_UINT32* tbx1,
2527         OPJ_UINT32* tby1)
2528 {
2529     /* Compute number of decomposition for this band. See table F-1 */
2530     OPJ_UINT32 nb = (resno == 0) ?
2531                     tilec->numresolutions - 1 :
2532                     tilec->numresolutions - resno;
2533     /* Map above tile-based coordinates to sub-band-based coordinates per */
2534     /* equation B-15 of the standard */
2535     OPJ_UINT32 x0b = bandno & 1;
2536     OPJ_UINT32 y0b = bandno >> 1;
2537     if (tbx0) {
2538         *tbx0 = (nb == 0) ? tcx0 :
2539                 (tcx0 <= (1U << (nb - 1)) * x0b) ? 0 :
2540                 opj_uint_ceildivpow2(tcx0 - (1U << (nb - 1)) * x0b, nb);
2541     }
2542     if (tby0) {
2543         *tby0 = (nb == 0) ? tcy0 :
2544                 (tcy0 <= (1U << (nb - 1)) * y0b) ? 0 :
2545                 opj_uint_ceildivpow2(tcy0 - (1U << (nb - 1)) * y0b, nb);
2546     }
2547     if (tbx1) {
2548         *tbx1 = (nb == 0) ? tcx1 :
2549                 (tcx1 <= (1U << (nb - 1)) * x0b) ? 0 :
2550                 opj_uint_ceildivpow2(tcx1 - (1U << (nb - 1)) * x0b, nb);
2551     }
2552     if (tby1) {
2553         *tby1 = (nb == 0) ? tcy1 :
2554                 (tcy1 <= (1U << (nb - 1)) * y0b) ? 0 :
2555                 opj_uint_ceildivpow2(tcy1 - (1U << (nb - 1)) * y0b, nb);
2556     }
2557 }
2558 
opj_dwt_segment_grow(OPJ_UINT32 filter_width,OPJ_UINT32 max_size,OPJ_UINT32 * start,OPJ_UINT32 * end)2559 static void opj_dwt_segment_grow(OPJ_UINT32 filter_width,
2560                                  OPJ_UINT32 max_size,
2561                                  OPJ_UINT32* start,
2562                                  OPJ_UINT32* end)
2563 {
2564     *start = opj_uint_subs(*start, filter_width);
2565     *end = opj_uint_adds(*end, filter_width);
2566     *end = opj_uint_min(*end, max_size);
2567 }
2568 
2569 
opj_dwt_init_sparse_array(opj_tcd_tilecomp_t * tilec,OPJ_UINT32 numres)2570 static opj_sparse_array_int32_t* opj_dwt_init_sparse_array(
2571     opj_tcd_tilecomp_t* tilec,
2572     OPJ_UINT32 numres)
2573 {
2574     opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]);
2575     OPJ_UINT32 w = (OPJ_UINT32)(tr_max->x1 - tr_max->x0);
2576     OPJ_UINT32 h = (OPJ_UINT32)(tr_max->y1 - tr_max->y0);
2577     OPJ_UINT32 resno, bandno, precno, cblkno;
2578     opj_sparse_array_int32_t* sa = opj_sparse_array_int32_create(
2579                                        w, h, opj_uint_min(w, 64), opj_uint_min(h, 64));
2580     if (sa == NULL) {
2581         return NULL;
2582     }
2583 
2584     for (resno = 0; resno < numres; ++resno) {
2585         opj_tcd_resolution_t* res = &tilec->resolutions[resno];
2586 
2587         for (bandno = 0; bandno < res->numbands; ++bandno) {
2588             opj_tcd_band_t* band = &res->bands[bandno];
2589 
2590             for (precno = 0; precno < res->pw * res->ph; ++precno) {
2591                 opj_tcd_precinct_t* precinct = &band->precincts[precno];
2592                 for (cblkno = 0; cblkno < precinct->cw * precinct->ch; ++cblkno) {
2593                     opj_tcd_cblk_dec_t* cblk = &precinct->cblks.dec[cblkno];
2594                     if (cblk->decoded_data != NULL) {
2595                         OPJ_UINT32 x = (OPJ_UINT32)(cblk->x0 - band->x0);
2596                         OPJ_UINT32 y = (OPJ_UINT32)(cblk->y0 - band->y0);
2597                         OPJ_UINT32 cblk_w = (OPJ_UINT32)(cblk->x1 - cblk->x0);
2598                         OPJ_UINT32 cblk_h = (OPJ_UINT32)(cblk->y1 - cblk->y0);
2599 
2600                         if (band->bandno & 1) {
2601                             opj_tcd_resolution_t* pres = &tilec->resolutions[resno - 1];
2602                             x += (OPJ_UINT32)(pres->x1 - pres->x0);
2603                         }
2604                         if (band->bandno & 2) {
2605                             opj_tcd_resolution_t* pres = &tilec->resolutions[resno - 1];
2606                             y += (OPJ_UINT32)(pres->y1 - pres->y0);
2607                         }
2608 
2609                         if (!opj_sparse_array_int32_write(sa, x, y,
2610                                                           x + cblk_w, y + cblk_h,
2611                                                           cblk->decoded_data,
2612                                                           1, cblk_w, OPJ_TRUE)) {
2613                             opj_sparse_array_int32_free(sa);
2614                             return NULL;
2615                         }
2616                     }
2617                 }
2618             }
2619         }
2620     }
2621 
2622     return sa;
2623 }
2624 
2625 
opj_dwt_decode_partial_tile(opj_tcd_tilecomp_t * tilec,OPJ_UINT32 numres)2626 static OPJ_BOOL opj_dwt_decode_partial_tile(
2627     opj_tcd_tilecomp_t* tilec,
2628     OPJ_UINT32 numres)
2629 {
2630     opj_sparse_array_int32_t* sa;
2631     opj_dwt_t h;
2632     opj_dwt_t v;
2633     OPJ_UINT32 resno;
2634     /* This value matches the maximum left/right extension given in tables */
2635     /* F.2 and F.3 of the standard. */
2636     const OPJ_UINT32 filter_width = 2U;
2637 
2638     opj_tcd_resolution_t* tr = tilec->resolutions;
2639     opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]);
2640 
2641     OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 -
2642                                  tr->x0);  /* width of the resolution level computed */
2643     OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
2644                                  tr->y0);  /* height of the resolution level computed */
2645 
2646     OPJ_SIZE_T h_mem_size;
2647 
2648     /* Compute the intersection of the area of interest, expressed in tile coordinates */
2649     /* with the tile coordinates */
2650     OPJ_UINT32 win_tcx0 = tilec->win_x0;
2651     OPJ_UINT32 win_tcy0 = tilec->win_y0;
2652     OPJ_UINT32 win_tcx1 = tilec->win_x1;
2653     OPJ_UINT32 win_tcy1 = tilec->win_y1;
2654 
2655     if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) {
2656         return OPJ_TRUE;
2657     }
2658 
2659     sa = opj_dwt_init_sparse_array(tilec, numres);
2660     if (sa == NULL) {
2661         return OPJ_FALSE;
2662     }
2663 
2664     if (numres == 1U) {
2665         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
2666                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
2667                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
2668                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
2669                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
2670                        tilec->data_win,
2671                        1, tr_max->win_x1 - tr_max->win_x0,
2672                        OPJ_TRUE);
2673         assert(ret);
2674         OPJ_UNUSED(ret);
2675         opj_sparse_array_int32_free(sa);
2676         return OPJ_TRUE;
2677     }
2678     h.mem_count = opj_dwt_max_resolution(tr, numres);
2679     /* overflow check */
2680     /* in vertical pass, we process 4 columns at a time */
2681     if (h.mem_count > (SIZE_MAX / (4 * sizeof(OPJ_INT32)))) {
2682         /* FIXME event manager error callback */
2683         opj_sparse_array_int32_free(sa);
2684         return OPJ_FALSE;
2685     }
2686 
2687     h_mem_size = h.mem_count * 4 * sizeof(OPJ_INT32);
2688     h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
2689     if (! h.mem) {
2690         /* FIXME event manager error callback */
2691         opj_sparse_array_int32_free(sa);
2692         return OPJ_FALSE;
2693     }
2694 
2695     v.mem_count = h.mem_count;
2696     v.mem = h.mem;
2697 
2698     for (resno = 1; resno < numres; resno ++) {
2699         OPJ_UINT32 i, j;
2700         /* Window of interest subband-based coordinates */
2701         OPJ_UINT32 win_ll_x0, win_ll_y0, win_ll_x1, win_ll_y1;
2702         OPJ_UINT32 win_hl_x0, win_hl_x1;
2703         OPJ_UINT32 win_lh_y0, win_lh_y1;
2704         /* Window of interest tile-resolution-based coordinates */
2705         OPJ_UINT32 win_tr_x0, win_tr_x1, win_tr_y0, win_tr_y1;
2706         /* Tile-resolution subband-based coordinates */
2707         OPJ_UINT32 tr_ll_x0, tr_ll_y0, tr_hl_x0, tr_lh_y0;
2708 
2709         ++tr;
2710 
2711         h.sn = (OPJ_INT32)rw;
2712         v.sn = (OPJ_INT32)rh;
2713 
2714         rw = (OPJ_UINT32)(tr->x1 - tr->x0);
2715         rh = (OPJ_UINT32)(tr->y1 - tr->y0);
2716 
2717         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
2718         h.cas = tr->x0 % 2;
2719 
2720         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
2721         v.cas = tr->y0 % 2;
2722 
2723         /* Get the subband coordinates for the window of interest */
2724         /* LL band */
2725         opj_dwt_get_band_coordinates(tilec, resno, 0,
2726                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
2727                                      &win_ll_x0, &win_ll_y0,
2728                                      &win_ll_x1, &win_ll_y1);
2729 
2730         /* HL band */
2731         opj_dwt_get_band_coordinates(tilec, resno, 1,
2732                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
2733                                      &win_hl_x0, NULL, &win_hl_x1, NULL);
2734 
2735         /* LH band */
2736         opj_dwt_get_band_coordinates(tilec, resno, 2,
2737                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
2738                                      NULL, &win_lh_y0, NULL, &win_lh_y1);
2739 
2740         /* Beware: band index for non-LL0 resolution are 0=HL, 1=LH and 2=HH */
2741         tr_ll_x0 = (OPJ_UINT32)tr->bands[1].x0;
2742         tr_ll_y0 = (OPJ_UINT32)tr->bands[0].y0;
2743         tr_hl_x0 = (OPJ_UINT32)tr->bands[0].x0;
2744         tr_lh_y0 = (OPJ_UINT32)tr->bands[1].y0;
2745 
2746         /* Subtract the origin of the bands for this tile, to the subwindow */
2747         /* of interest band coordinates, so as to get them relative to the */
2748         /* tile */
2749         win_ll_x0 = opj_uint_subs(win_ll_x0, tr_ll_x0);
2750         win_ll_y0 = opj_uint_subs(win_ll_y0, tr_ll_y0);
2751         win_ll_x1 = opj_uint_subs(win_ll_x1, tr_ll_x0);
2752         win_ll_y1 = opj_uint_subs(win_ll_y1, tr_ll_y0);
2753         win_hl_x0 = opj_uint_subs(win_hl_x0, tr_hl_x0);
2754         win_hl_x1 = opj_uint_subs(win_hl_x1, tr_hl_x0);
2755         win_lh_y0 = opj_uint_subs(win_lh_y0, tr_lh_y0);
2756         win_lh_y1 = opj_uint_subs(win_lh_y1, tr_lh_y0);
2757 
2758         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.sn, &win_ll_x0, &win_ll_x1);
2759         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.dn, &win_hl_x0, &win_hl_x1);
2760 
2761         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.sn, &win_ll_y0, &win_ll_y1);
2762         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.dn, &win_lh_y0, &win_lh_y1);
2763 
2764         /* Compute the tile-resolution-based coordinates for the window of interest */
2765         if (h.cas == 0) {
2766             win_tr_x0 = opj_uint_min(2 * win_ll_x0, 2 * win_hl_x0 + 1);
2767             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_ll_x1, 2 * win_hl_x1 + 1), rw);
2768         } else {
2769             win_tr_x0 = opj_uint_min(2 * win_hl_x0, 2 * win_ll_x0 + 1);
2770             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_hl_x1, 2 * win_ll_x1 + 1), rw);
2771         }
2772 
2773         if (v.cas == 0) {
2774             win_tr_y0 = opj_uint_min(2 * win_ll_y0, 2 * win_lh_y0 + 1);
2775             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_ll_y1, 2 * win_lh_y1 + 1), rh);
2776         } else {
2777             win_tr_y0 = opj_uint_min(2 * win_lh_y0, 2 * win_ll_y0 + 1);
2778             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_lh_y1, 2 * win_ll_y1 + 1), rh);
2779         }
2780 
2781         for (j = 0; j < rh; ++j) {
2782             if ((j >= win_ll_y0 && j < win_ll_y1) ||
2783                     (j >= win_lh_y0 + (OPJ_UINT32)v.sn && j < win_lh_y1 + (OPJ_UINT32)v.sn)) {
2784 
2785                 /* Avoids dwt.c:1584:44 (in opj_dwt_decode_partial_1): runtime error: */
2786                 /* signed integer overflow: -1094795586 + -1094795586 cannot be represented in type 'int' */
2787                 /* on opj_decompress -i  ../../openjpeg/MAPA.jp2 -o out.tif -d 0,0,256,256 */
2788                 /* This is less extreme than memsetting the whole buffer to 0 */
2789                 /* although we could potentially do better with better handling of edge conditions */
2790                 if (win_tr_x1 >= 1 && win_tr_x1 < rw) {
2791                     h.mem[win_tr_x1 - 1] = 0;
2792                 }
2793                 if (win_tr_x1 < rw) {
2794                     h.mem[win_tr_x1] = 0;
2795                 }
2796 
2797                 opj_dwt_interleave_partial_h(h.mem,
2798                                              h.cas,
2799                                              sa,
2800                                              j,
2801                                              (OPJ_UINT32)h.sn,
2802                                              win_ll_x0,
2803                                              win_ll_x1,
2804                                              win_hl_x0,
2805                                              win_hl_x1);
2806                 opj_dwt_decode_partial_1(h.mem, h.mem_count, h.dn, h.sn, h.cas,
2807                                          (OPJ_INT32)win_ll_x0,
2808                                          (OPJ_INT32)win_ll_x1,
2809                                          (OPJ_INT32)win_hl_x0,
2810                                          (OPJ_INT32)win_hl_x1);
2811                 if (!opj_sparse_array_int32_write(sa,
2812                                                   win_tr_x0, j,
2813                                                   win_tr_x1, j + 1,
2814                                                   h.mem + win_tr_x0,
2815                                                   1, 0, OPJ_TRUE)) {
2816                     /* FIXME event manager error callback */
2817                     opj_sparse_array_int32_free(sa);
2818                     opj_aligned_free(h.mem);
2819                     return OPJ_FALSE;
2820                 }
2821             }
2822         }
2823 
2824         for (i = win_tr_x0; i < win_tr_x1;) {
2825             OPJ_UINT32 nb_cols = opj_uint_min(4U, win_tr_x1 - i);
2826             opj_dwt_interleave_partial_v(v.mem,
2827                                          v.cas,
2828                                          sa,
2829                                          i,
2830                                          nb_cols,
2831                                          (OPJ_UINT32)v.sn,
2832                                          win_ll_y0,
2833                                          win_ll_y1,
2834                                          win_lh_y0,
2835                                          win_lh_y1);
2836             opj_dwt_decode_partial_1_parallel(v.mem, nb_cols, v.dn, v.sn, v.cas,
2837                                               (OPJ_INT32)win_ll_y0,
2838                                               (OPJ_INT32)win_ll_y1,
2839                                               (OPJ_INT32)win_lh_y0,
2840                                               (OPJ_INT32)win_lh_y1);
2841             if (!opj_sparse_array_int32_write(sa,
2842                                               i, win_tr_y0,
2843                                               i + nb_cols, win_tr_y1,
2844                                               v.mem + 4 * win_tr_y0,
2845                                               1, 4, OPJ_TRUE)) {
2846                 /* FIXME event manager error callback */
2847                 opj_sparse_array_int32_free(sa);
2848                 opj_aligned_free(h.mem);
2849                 return OPJ_FALSE;
2850             }
2851 
2852             i += nb_cols;
2853         }
2854     }
2855     opj_aligned_free(h.mem);
2856 
2857     {
2858         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
2859                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
2860                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
2861                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
2862                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
2863                        tilec->data_win,
2864                        1, tr_max->win_x1 - tr_max->win_x0,
2865                        OPJ_TRUE);
2866         assert(ret);
2867         OPJ_UNUSED(ret);
2868     }
2869     opj_sparse_array_int32_free(sa);
2870     return OPJ_TRUE;
2871 }
2872 
opj_v8dwt_interleave_h(opj_v8dwt_t * OPJ_RESTRICT dwt,OPJ_FLOAT32 * OPJ_RESTRICT a,OPJ_UINT32 width,OPJ_UINT32 remaining_height)2873 static void opj_v8dwt_interleave_h(opj_v8dwt_t* OPJ_RESTRICT dwt,
2874                                    OPJ_FLOAT32* OPJ_RESTRICT a,
2875                                    OPJ_UINT32 width,
2876                                    OPJ_UINT32 remaining_height)
2877 {
2878     OPJ_FLOAT32* OPJ_RESTRICT bi = (OPJ_FLOAT32*)(dwt->wavelet + dwt->cas);
2879     OPJ_UINT32 i, k;
2880     OPJ_UINT32 x0 = dwt->win_l_x0;
2881     OPJ_UINT32 x1 = dwt->win_l_x1;
2882 
2883     for (k = 0; k < 2; ++k) {
2884         if (remaining_height >= NB_ELTS_V8 && ((OPJ_SIZE_T) a & 0x0f) == 0 &&
2885                 ((OPJ_SIZE_T) bi & 0x0f) == 0) {
2886             /* Fast code path */
2887             for (i = x0; i < x1; ++i) {
2888                 OPJ_UINT32 j = i;
2889                 OPJ_FLOAT32* OPJ_RESTRICT dst = bi + i * 2 * NB_ELTS_V8;
2890                 dst[0] = a[j];
2891                 j += width;
2892                 dst[1] = a[j];
2893                 j += width;
2894                 dst[2] = a[j];
2895                 j += width;
2896                 dst[3] = a[j];
2897                 j += width;
2898                 dst[4] = a[j];
2899                 j += width;
2900                 dst[5] = a[j];
2901                 j += width;
2902                 dst[6] = a[j];
2903                 j += width;
2904                 dst[7] = a[j];
2905             }
2906         } else {
2907             /* Slow code path */
2908             for (i = x0; i < x1; ++i) {
2909                 OPJ_UINT32 j = i;
2910                 OPJ_FLOAT32* OPJ_RESTRICT dst = bi + i * 2 * NB_ELTS_V8;
2911                 dst[0] = a[j];
2912                 j += width;
2913                 if (remaining_height == 1) {
2914                     continue;
2915                 }
2916                 dst[1] = a[j];
2917                 j += width;
2918                 if (remaining_height == 2) {
2919                     continue;
2920                 }
2921                 dst[2] = a[j];
2922                 j += width;
2923                 if (remaining_height == 3) {
2924                     continue;
2925                 }
2926                 dst[3] = a[j];
2927                 j += width;
2928                 if (remaining_height == 4) {
2929                     continue;
2930                 }
2931                 dst[4] = a[j];
2932                 j += width;
2933                 if (remaining_height == 5) {
2934                     continue;
2935                 }
2936                 dst[5] = a[j];
2937                 j += width;
2938                 if (remaining_height == 6) {
2939                     continue;
2940                 }
2941                 dst[6] = a[j];
2942                 j += width;
2943                 if (remaining_height == 7) {
2944                     continue;
2945                 }
2946                 dst[7] = a[j];
2947             }
2948         }
2949 
2950         bi = (OPJ_FLOAT32*)(dwt->wavelet + 1 - dwt->cas);
2951         a += dwt->sn;
2952         x0 = dwt->win_h_x0;
2953         x1 = dwt->win_h_x1;
2954     }
2955 }
2956 
opj_v8dwt_interleave_partial_h(opj_v8dwt_t * dwt,opj_sparse_array_int32_t * sa,OPJ_UINT32 sa_line,OPJ_UINT32 remaining_height)2957 static void opj_v8dwt_interleave_partial_h(opj_v8dwt_t* dwt,
2958         opj_sparse_array_int32_t* sa,
2959         OPJ_UINT32 sa_line,
2960         OPJ_UINT32 remaining_height)
2961 {
2962     OPJ_UINT32 i;
2963     for (i = 0; i < remaining_height; i++) {
2964         OPJ_BOOL ret;
2965         ret = opj_sparse_array_int32_read(sa,
2966                                           dwt->win_l_x0, sa_line + i,
2967                                           dwt->win_l_x1, sa_line + i + 1,
2968                                           /* Nasty cast from float* to int32* */
2969                                           (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0) + i,
2970                                           2 * NB_ELTS_V8, 0, OPJ_TRUE);
2971         assert(ret);
2972         ret = opj_sparse_array_int32_read(sa,
2973                                           (OPJ_UINT32)dwt->sn + dwt->win_h_x0, sa_line + i,
2974                                           (OPJ_UINT32)dwt->sn + dwt->win_h_x1, sa_line + i + 1,
2975                                           /* Nasty cast from float* to int32* */
2976                                           (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0) + i,
2977                                           2 * NB_ELTS_V8, 0, OPJ_TRUE);
2978         assert(ret);
2979         OPJ_UNUSED(ret);
2980     }
2981 }
2982 
opj_v8dwt_interleave_v(opj_v8dwt_t * OPJ_RESTRICT dwt,OPJ_FLOAT32 * OPJ_RESTRICT a,OPJ_UINT32 width,OPJ_UINT32 nb_elts_read)2983 static INLINE void opj_v8dwt_interleave_v(opj_v8dwt_t* OPJ_RESTRICT dwt,
2984         OPJ_FLOAT32* OPJ_RESTRICT a,
2985         OPJ_UINT32 width,
2986         OPJ_UINT32 nb_elts_read)
2987 {
2988     opj_v8_t* OPJ_RESTRICT bi = dwt->wavelet + dwt->cas;
2989     OPJ_UINT32 i;
2990 
2991     for (i = dwt->win_l_x0; i < dwt->win_l_x1; ++i) {
2992         memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width],
2993                (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32));
2994     }
2995 
2996     a += (OPJ_UINT32)dwt->sn * (OPJ_SIZE_T)width;
2997     bi = dwt->wavelet + 1 - dwt->cas;
2998 
2999     for (i = dwt->win_h_x0; i < dwt->win_h_x1; ++i) {
3000         memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width],
3001                (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32));
3002     }
3003 }
3004 
opj_v8dwt_interleave_partial_v(opj_v8dwt_t * OPJ_RESTRICT dwt,opj_sparse_array_int32_t * sa,OPJ_UINT32 sa_col,OPJ_UINT32 nb_elts_read)3005 static void opj_v8dwt_interleave_partial_v(opj_v8dwt_t* OPJ_RESTRICT dwt,
3006         opj_sparse_array_int32_t* sa,
3007         OPJ_UINT32 sa_col,
3008         OPJ_UINT32 nb_elts_read)
3009 {
3010     OPJ_BOOL ret;
3011     ret = opj_sparse_array_int32_read(sa,
3012                                       sa_col, dwt->win_l_x0,
3013                                       sa_col + nb_elts_read, dwt->win_l_x1,
3014                                       (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0),
3015                                       1, 2 * NB_ELTS_V8, OPJ_TRUE);
3016     assert(ret);
3017     ret = opj_sparse_array_int32_read(sa,
3018                                       sa_col, (OPJ_UINT32)dwt->sn + dwt->win_h_x0,
3019                                       sa_col + nb_elts_read, (OPJ_UINT32)dwt->sn + dwt->win_h_x1,
3020                                       (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0),
3021                                       1, 2 * NB_ELTS_V8, OPJ_TRUE);
3022     assert(ret);
3023     OPJ_UNUSED(ret);
3024 }
3025 
3026 #ifdef __SSE__
3027 
opj_v8dwt_decode_step1_sse(opj_v8_t * w,OPJ_UINT32 start,OPJ_UINT32 end,const __m128 c)3028 static void opj_v8dwt_decode_step1_sse(opj_v8_t* w,
3029                                        OPJ_UINT32 start,
3030                                        OPJ_UINT32 end,
3031                                        const __m128 c)
3032 {
3033     __m128* OPJ_RESTRICT vw = (__m128*) w;
3034     OPJ_UINT32 i = start;
3035     /* To be adapted if NB_ELTS_V8 changes */
3036     vw += 4 * start;
3037     /* Note: attempt at loop unrolling x2 doesn't help */
3038     for (; i < end; ++i, vw += 4) {
3039         vw[0] = _mm_mul_ps(vw[0], c);
3040         vw[1] = _mm_mul_ps(vw[1], c);
3041     }
3042 }
3043 
opj_v8dwt_decode_step2_sse(opj_v8_t * l,opj_v8_t * w,OPJ_UINT32 start,OPJ_UINT32 end,OPJ_UINT32 m,__m128 c)3044 static void opj_v8dwt_decode_step2_sse(opj_v8_t* l, opj_v8_t* w,
3045                                        OPJ_UINT32 start,
3046                                        OPJ_UINT32 end,
3047                                        OPJ_UINT32 m,
3048                                        __m128 c)
3049 {
3050     __m128* OPJ_RESTRICT vl = (__m128*) l;
3051     __m128* OPJ_RESTRICT vw = (__m128*) w;
3052     /* To be adapted if NB_ELTS_V8 changes */
3053     OPJ_UINT32 i;
3054     OPJ_UINT32 imax = opj_uint_min(end, m);
3055     if (start == 0) {
3056         if (imax >= 1) {
3057             vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(_mm_add_ps(vl[0], vw[0]), c));
3058             vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(_mm_add_ps(vl[1], vw[1]), c));
3059             vw += 4;
3060             start = 1;
3061         }
3062     } else {
3063         vw += start * 4;
3064     }
3065 
3066     i = start;
3067     /* Note: attempt at loop unrolling x2 doesn't help */
3068     for (; i < imax; ++i) {
3069         vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(_mm_add_ps(vw[-4], vw[0]), c));
3070         vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(_mm_add_ps(vw[-3], vw[1]), c));
3071         vw += 4;
3072     }
3073     if (m < end) {
3074         assert(m + 1 == end);
3075         c = _mm_add_ps(c, c);
3076         vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(c, vw[-4]));
3077         vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(c, vw[-3]));
3078     }
3079 }
3080 
3081 #else
3082 
opj_v8dwt_decode_step1(opj_v8_t * w,OPJ_UINT32 start,OPJ_UINT32 end,const OPJ_FLOAT32 c)3083 static void opj_v8dwt_decode_step1(opj_v8_t* w,
3084                                    OPJ_UINT32 start,
3085                                    OPJ_UINT32 end,
3086                                    const OPJ_FLOAT32 c)
3087 {
3088     OPJ_FLOAT32* OPJ_RESTRICT fw = (OPJ_FLOAT32*) w;
3089     OPJ_UINT32 i;
3090     /* To be adapted if NB_ELTS_V8 changes */
3091     for (i = start; i < end; ++i) {
3092         fw[i * 2 * 8    ] = fw[i * 2 * 8    ] * c;
3093         fw[i * 2 * 8 + 1] = fw[i * 2 * 8 + 1] * c;
3094         fw[i * 2 * 8 + 2] = fw[i * 2 * 8 + 2] * c;
3095         fw[i * 2 * 8 + 3] = fw[i * 2 * 8 + 3] * c;
3096         fw[i * 2 * 8 + 4] = fw[i * 2 * 8 + 4] * c;
3097         fw[i * 2 * 8 + 5] = fw[i * 2 * 8 + 5] * c;
3098         fw[i * 2 * 8 + 6] = fw[i * 2 * 8 + 6] * c;
3099         fw[i * 2 * 8 + 7] = fw[i * 2 * 8 + 7] * c;
3100     }
3101 }
3102 
opj_v8dwt_decode_step2(opj_v8_t * l,opj_v8_t * w,OPJ_UINT32 start,OPJ_UINT32 end,OPJ_UINT32 m,OPJ_FLOAT32 c)3103 static void opj_v8dwt_decode_step2(opj_v8_t* l, opj_v8_t* w,
3104                                    OPJ_UINT32 start,
3105                                    OPJ_UINT32 end,
3106                                    OPJ_UINT32 m,
3107                                    OPJ_FLOAT32 c)
3108 {
3109     OPJ_FLOAT32* fl = (OPJ_FLOAT32*) l;
3110     OPJ_FLOAT32* fw = (OPJ_FLOAT32*) w;
3111     OPJ_UINT32 i;
3112     OPJ_UINT32 imax = opj_uint_min(end, m);
3113     if (start > 0) {
3114         fw += 2 * NB_ELTS_V8 * start;
3115         fl = fw - 2 * NB_ELTS_V8;
3116     }
3117     /* To be adapted if NB_ELTS_V8 changes */
3118     for (i = start; i < imax; ++i) {
3119         fw[-8] = fw[-8] + ((fl[0] + fw[0]) * c);
3120         fw[-7] = fw[-7] + ((fl[1] + fw[1]) * c);
3121         fw[-6] = fw[-6] + ((fl[2] + fw[2]) * c);
3122         fw[-5] = fw[-5] + ((fl[3] + fw[3]) * c);
3123         fw[-4] = fw[-4] + ((fl[4] + fw[4]) * c);
3124         fw[-3] = fw[-3] + ((fl[5] + fw[5]) * c);
3125         fw[-2] = fw[-2] + ((fl[6] + fw[6]) * c);
3126         fw[-1] = fw[-1] + ((fl[7] + fw[7]) * c);
3127         fl = fw;
3128         fw += 2 * NB_ELTS_V8;
3129     }
3130     if (m < end) {
3131         assert(m + 1 == end);
3132         c += c;
3133         fw[-8] = fw[-8] + fl[0] * c;
3134         fw[-7] = fw[-7] + fl[1] * c;
3135         fw[-6] = fw[-6] + fl[2] * c;
3136         fw[-5] = fw[-5] + fl[3] * c;
3137         fw[-4] = fw[-4] + fl[4] * c;
3138         fw[-3] = fw[-3] + fl[5] * c;
3139         fw[-2] = fw[-2] + fl[6] * c;
3140         fw[-1] = fw[-1] + fl[7] * c;
3141     }
3142 }
3143 
3144 #endif
3145 
3146 /* <summary>                             */
3147 /* Inverse 9-7 wavelet transform in 1-D. */
3148 /* </summary>                            */
opj_v8dwt_decode(opj_v8dwt_t * OPJ_RESTRICT dwt)3149 static void opj_v8dwt_decode(opj_v8dwt_t* OPJ_RESTRICT dwt)
3150 {
3151     OPJ_INT32 a, b;
3152     /* BUG_WEIRD_TWO_INVK (look for this identifier in tcd.c) */
3153     /* Historic value for 2 / opj_invK */
3154     /* Normally, we should use invK, but if we do so, we have failures in the */
3155     /* conformance test, due to MSE and peak errors significantly higher than */
3156     /* accepted value */
3157     /* Due to using two_invK instead of invK, we have to compensate in tcd.c */
3158     /* the computation of the stepsize for the non LL subbands */
3159     const float two_invK = 1.625732422f;
3160     if (dwt->cas == 0) {
3161         if (!((dwt->dn > 0) || (dwt->sn > 1))) {
3162             return;
3163         }
3164         a = 0;
3165         b = 1;
3166     } else {
3167         if (!((dwt->sn > 0) || (dwt->dn > 1))) {
3168             return;
3169         }
3170         a = 1;
3171         b = 0;
3172     }
3173 #ifdef __SSE__
3174     opj_v8dwt_decode_step1_sse(dwt->wavelet + a, dwt->win_l_x0, dwt->win_l_x1,
3175                                _mm_set1_ps(opj_K));
3176     opj_v8dwt_decode_step1_sse(dwt->wavelet + b, dwt->win_h_x0, dwt->win_h_x1,
3177                                _mm_set1_ps(two_invK));
3178     opj_v8dwt_decode_step2_sse(dwt->wavelet + b, dwt->wavelet + a + 1,
3179                                dwt->win_l_x0, dwt->win_l_x1,
3180                                (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
3181                                _mm_set1_ps(-opj_dwt_delta));
3182     opj_v8dwt_decode_step2_sse(dwt->wavelet + a, dwt->wavelet + b + 1,
3183                                dwt->win_h_x0, dwt->win_h_x1,
3184                                (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
3185                                _mm_set1_ps(-opj_dwt_gamma));
3186     opj_v8dwt_decode_step2_sse(dwt->wavelet + b, dwt->wavelet + a + 1,
3187                                dwt->win_l_x0, dwt->win_l_x1,
3188                                (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
3189                                _mm_set1_ps(-opj_dwt_beta));
3190     opj_v8dwt_decode_step2_sse(dwt->wavelet + a, dwt->wavelet + b + 1,
3191                                dwt->win_h_x0, dwt->win_h_x1,
3192                                (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
3193                                _mm_set1_ps(-opj_dwt_alpha));
3194 #else
3195     opj_v8dwt_decode_step1(dwt->wavelet + a, dwt->win_l_x0, dwt->win_l_x1,
3196                            opj_K);
3197     opj_v8dwt_decode_step1(dwt->wavelet + b, dwt->win_h_x0, dwt->win_h_x1,
3198                            two_invK);
3199     opj_v8dwt_decode_step2(dwt->wavelet + b, dwt->wavelet + a + 1,
3200                            dwt->win_l_x0, dwt->win_l_x1,
3201                            (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
3202                            -opj_dwt_delta);
3203     opj_v8dwt_decode_step2(dwt->wavelet + a, dwt->wavelet + b + 1,
3204                            dwt->win_h_x0, dwt->win_h_x1,
3205                            (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
3206                            -opj_dwt_gamma);
3207     opj_v8dwt_decode_step2(dwt->wavelet + b, dwt->wavelet + a + 1,
3208                            dwt->win_l_x0, dwt->win_l_x1,
3209                            (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a),
3210                            -opj_dwt_beta);
3211     opj_v8dwt_decode_step2(dwt->wavelet + a, dwt->wavelet + b + 1,
3212                            dwt->win_h_x0, dwt->win_h_x1,
3213                            (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b),
3214                            -opj_dwt_alpha);
3215 #endif
3216 }
3217 
3218 typedef struct {
3219     opj_v8dwt_t h;
3220     OPJ_UINT32 rw;
3221     OPJ_UINT32 w;
3222     OPJ_FLOAT32 * OPJ_RESTRICT aj;
3223     OPJ_UINT32 nb_rows;
3224 } opj_dwt97_decode_h_job_t;
3225 
opj_dwt97_decode_h_func(void * user_data,opj_tls_t * tls)3226 static void opj_dwt97_decode_h_func(void* user_data, opj_tls_t* tls)
3227 {
3228     OPJ_UINT32 j;
3229     opj_dwt97_decode_h_job_t* job;
3230     OPJ_FLOAT32 * OPJ_RESTRICT aj;
3231     OPJ_UINT32 w;
3232     (void)tls;
3233 
3234     job = (opj_dwt97_decode_h_job_t*)user_data;
3235     w = job->w;
3236 
3237     assert((job->nb_rows % NB_ELTS_V8) == 0);
3238 
3239     aj = job->aj;
3240     for (j = 0; j + NB_ELTS_V8 <= job->nb_rows; j += NB_ELTS_V8) {
3241         OPJ_UINT32 k;
3242         opj_v8dwt_interleave_h(&job->h, aj, job->w, NB_ELTS_V8);
3243         opj_v8dwt_decode(&job->h);
3244 
3245         /* To be adapted if NB_ELTS_V8 changes */
3246         for (k = 0; k < job->rw; k++) {
3247             aj[k      ] = job->h.wavelet[k].f[0];
3248             aj[k + (OPJ_SIZE_T)w  ] = job->h.wavelet[k].f[1];
3249             aj[k + (OPJ_SIZE_T)w * 2] = job->h.wavelet[k].f[2];
3250             aj[k + (OPJ_SIZE_T)w * 3] = job->h.wavelet[k].f[3];
3251         }
3252         for (k = 0; k < job->rw; k++) {
3253             aj[k + (OPJ_SIZE_T)w * 4] = job->h.wavelet[k].f[4];
3254             aj[k + (OPJ_SIZE_T)w * 5] = job->h.wavelet[k].f[5];
3255             aj[k + (OPJ_SIZE_T)w * 6] = job->h.wavelet[k].f[6];
3256             aj[k + (OPJ_SIZE_T)w * 7] = job->h.wavelet[k].f[7];
3257         }
3258 
3259         aj += w * NB_ELTS_V8;
3260     }
3261 
3262     opj_aligned_free(job->h.wavelet);
3263     opj_free(job);
3264 }
3265 
3266 
3267 typedef struct {
3268     opj_v8dwt_t v;
3269     OPJ_UINT32 rh;
3270     OPJ_UINT32 w;
3271     OPJ_FLOAT32 * OPJ_RESTRICT aj;
3272     OPJ_UINT32 nb_columns;
3273 } opj_dwt97_decode_v_job_t;
3274 
opj_dwt97_decode_v_func(void * user_data,opj_tls_t * tls)3275 static void opj_dwt97_decode_v_func(void* user_data, opj_tls_t* tls)
3276 {
3277     OPJ_UINT32 j;
3278     opj_dwt97_decode_v_job_t* job;
3279     OPJ_FLOAT32 * OPJ_RESTRICT aj;
3280     (void)tls;
3281 
3282     job = (opj_dwt97_decode_v_job_t*)user_data;
3283 
3284     assert((job->nb_columns % NB_ELTS_V8) == 0);
3285 
3286     aj = job->aj;
3287     for (j = 0; j + NB_ELTS_V8 <= job->nb_columns; j += NB_ELTS_V8) {
3288         OPJ_UINT32 k;
3289 
3290         opj_v8dwt_interleave_v(&job->v, aj, job->w, NB_ELTS_V8);
3291         opj_v8dwt_decode(&job->v);
3292 
3293         for (k = 0; k < job->rh; ++k) {
3294             memcpy(&aj[k * (OPJ_SIZE_T)job->w], &job->v.wavelet[k],
3295                    NB_ELTS_V8 * sizeof(OPJ_FLOAT32));
3296         }
3297         aj += NB_ELTS_V8;
3298     }
3299 
3300     opj_aligned_free(job->v.wavelet);
3301     opj_free(job);
3302 }
3303 
3304 
3305 /* <summary>                             */
3306 /* Inverse 9-7 wavelet transform in 2-D. */
3307 /* </summary>                            */
3308 static
opj_dwt_decode_tile_97(opj_thread_pool_t * tp,opj_tcd_tilecomp_t * OPJ_RESTRICT tilec,OPJ_UINT32 numres)3309 OPJ_BOOL opj_dwt_decode_tile_97(opj_thread_pool_t* tp,
3310                                 opj_tcd_tilecomp_t* OPJ_RESTRICT tilec,
3311                                 OPJ_UINT32 numres)
3312 {
3313     opj_v8dwt_t h;
3314     opj_v8dwt_t v;
3315 
3316     opj_tcd_resolution_t* res = tilec->resolutions;
3317 
3318     OPJ_UINT32 rw = (OPJ_UINT32)(res->x1 -
3319                                  res->x0);    /* width of the resolution level computed */
3320     OPJ_UINT32 rh = (OPJ_UINT32)(res->y1 -
3321                                  res->y0);    /* height of the resolution level computed */
3322 
3323     OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions -
3324                                                                1].x1 -
3325                                 tilec->resolutions[tilec->minimum_num_resolutions - 1].x0);
3326 
3327     OPJ_SIZE_T l_data_size;
3328     const int num_threads = opj_thread_pool_get_thread_count(tp);
3329 
3330     if (numres == 1) {
3331         return OPJ_TRUE;
3332     }
3333 
3334     l_data_size = opj_dwt_max_resolution(res, numres);
3335     /* overflow check */
3336     if (l_data_size > (SIZE_MAX / sizeof(opj_v8_t))) {
3337         /* FIXME event manager error callback */
3338         return OPJ_FALSE;
3339     }
3340     h.wavelet = (opj_v8_t*) opj_aligned_malloc(l_data_size * sizeof(opj_v8_t));
3341     if (!h.wavelet) {
3342         /* FIXME event manager error callback */
3343         return OPJ_FALSE;
3344     }
3345     v.wavelet = h.wavelet;
3346 
3347     while (--numres) {
3348         OPJ_FLOAT32 * OPJ_RESTRICT aj = (OPJ_FLOAT32*) tilec->data;
3349         OPJ_UINT32 j;
3350 
3351         h.sn = (OPJ_INT32)rw;
3352         v.sn = (OPJ_INT32)rh;
3353 
3354         ++res;
3355 
3356         rw = (OPJ_UINT32)(res->x1 -
3357                           res->x0);   /* width of the resolution level computed */
3358         rh = (OPJ_UINT32)(res->y1 -
3359                           res->y0);   /* height of the resolution level computed */
3360 
3361         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
3362         h.cas = res->x0 % 2;
3363 
3364         h.win_l_x0 = 0;
3365         h.win_l_x1 = (OPJ_UINT32)h.sn;
3366         h.win_h_x0 = 0;
3367         h.win_h_x1 = (OPJ_UINT32)h.dn;
3368 
3369         if (num_threads <= 1 || rh < 2 * NB_ELTS_V8) {
3370             for (j = 0; j + (NB_ELTS_V8 - 1) < rh; j += NB_ELTS_V8) {
3371                 OPJ_UINT32 k;
3372                 opj_v8dwt_interleave_h(&h, aj, w, NB_ELTS_V8);
3373                 opj_v8dwt_decode(&h);
3374 
3375                 /* To be adapted if NB_ELTS_V8 changes */
3376                 for (k = 0; k < rw; k++) {
3377                     aj[k      ] = h.wavelet[k].f[0];
3378                     aj[k + (OPJ_SIZE_T)w  ] = h.wavelet[k].f[1];
3379                     aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2];
3380                     aj[k + (OPJ_SIZE_T)w * 3] = h.wavelet[k].f[3];
3381                 }
3382                 for (k = 0; k < rw; k++) {
3383                     aj[k + (OPJ_SIZE_T)w * 4] = h.wavelet[k].f[4];
3384                     aj[k + (OPJ_SIZE_T)w * 5] = h.wavelet[k].f[5];
3385                     aj[k + (OPJ_SIZE_T)w * 6] = h.wavelet[k].f[6];
3386                     aj[k + (OPJ_SIZE_T)w * 7] = h.wavelet[k].f[7];
3387                 }
3388 
3389                 aj += w * NB_ELTS_V8;
3390             }
3391         } else {
3392             OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
3393             OPJ_UINT32 step_j;
3394 
3395             if ((rh / NB_ELTS_V8) < num_jobs) {
3396                 num_jobs = rh / NB_ELTS_V8;
3397             }
3398             step_j = ((rh / num_jobs) / NB_ELTS_V8) * NB_ELTS_V8;
3399             for (j = 0; j < num_jobs; j++) {
3400                 opj_dwt97_decode_h_job_t* job;
3401 
3402                 job = (opj_dwt97_decode_h_job_t*) opj_malloc(sizeof(opj_dwt97_decode_h_job_t));
3403                 if (!job) {
3404                     opj_thread_pool_wait_completion(tp, 0);
3405                     opj_aligned_free(h.wavelet);
3406                     return OPJ_FALSE;
3407                 }
3408                 job->h.wavelet = (opj_v8_t*)opj_aligned_malloc(l_data_size * sizeof(opj_v8_t));
3409                 if (!job->h.wavelet) {
3410                     opj_thread_pool_wait_completion(tp, 0);
3411                     opj_free(job);
3412                     opj_aligned_free(h.wavelet);
3413                     return OPJ_FALSE;
3414                 }
3415                 job->h.dn = h.dn;
3416                 job->h.sn = h.sn;
3417                 job->h.cas = h.cas;
3418                 job->h.win_l_x0 = h.win_l_x0;
3419                 job->h.win_l_x1 = h.win_l_x1;
3420                 job->h.win_h_x0 = h.win_h_x0;
3421                 job->h.win_h_x1 = h.win_h_x1;
3422                 job->rw = rw;
3423                 job->w = w;
3424                 job->aj = aj;
3425                 job->nb_rows = (j + 1 == num_jobs) ? (rh & (OPJ_UINT32)~
3426                                                       (NB_ELTS_V8 - 1)) - j * step_j : step_j;
3427                 aj += w * job->nb_rows;
3428                 opj_thread_pool_submit_job(tp, opj_dwt97_decode_h_func, job);
3429             }
3430             opj_thread_pool_wait_completion(tp, 0);
3431             j = rh & (OPJ_UINT32)~(NB_ELTS_V8 - 1);
3432         }
3433 
3434         if (j < rh) {
3435             OPJ_UINT32 k;
3436             opj_v8dwt_interleave_h(&h, aj, w, rh - j);
3437             opj_v8dwt_decode(&h);
3438             for (k = 0; k < rw; k++) {
3439                 OPJ_UINT32 l;
3440                 for (l = 0; l < rh - j; l++) {
3441                     aj[k + (OPJ_SIZE_T)w  * l ] = h.wavelet[k].f[l];
3442                 }
3443             }
3444         }
3445 
3446         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
3447         v.cas = res->y0 % 2;
3448         v.win_l_x0 = 0;
3449         v.win_l_x1 = (OPJ_UINT32)v.sn;
3450         v.win_h_x0 = 0;
3451         v.win_h_x1 = (OPJ_UINT32)v.dn;
3452 
3453         aj = (OPJ_FLOAT32*) tilec->data;
3454         if (num_threads <= 1 || rw < 2 * NB_ELTS_V8) {
3455             for (j = rw; j > (NB_ELTS_V8 - 1); j -= NB_ELTS_V8) {
3456                 OPJ_UINT32 k;
3457 
3458                 opj_v8dwt_interleave_v(&v, aj, w, NB_ELTS_V8);
3459                 opj_v8dwt_decode(&v);
3460 
3461                 for (k = 0; k < rh; ++k) {
3462                     memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k], NB_ELTS_V8 * sizeof(OPJ_FLOAT32));
3463                 }
3464                 aj += NB_ELTS_V8;
3465             }
3466         } else {
3467             /* "bench_dwt -I" shows that scaling is poor, likely due to RAM
3468                 transfer being the limiting factor. So limit the number of
3469                 threads.
3470              */
3471             OPJ_UINT32 num_jobs = opj_uint_max((OPJ_UINT32)num_threads / 2, 2U);
3472             OPJ_UINT32 step_j;
3473 
3474             if ((rw / NB_ELTS_V8) < num_jobs) {
3475                 num_jobs = rw / NB_ELTS_V8;
3476             }
3477             step_j = ((rw / num_jobs) / NB_ELTS_V8) * NB_ELTS_V8;
3478             for (j = 0; j < num_jobs; j++) {
3479                 opj_dwt97_decode_v_job_t* job;
3480 
3481                 job = (opj_dwt97_decode_v_job_t*) opj_malloc(sizeof(opj_dwt97_decode_v_job_t));
3482                 if (!job) {
3483                     opj_thread_pool_wait_completion(tp, 0);
3484                     opj_aligned_free(h.wavelet);
3485                     return OPJ_FALSE;
3486                 }
3487                 job->v.wavelet = (opj_v8_t*)opj_aligned_malloc(l_data_size * sizeof(opj_v8_t));
3488                 if (!job->v.wavelet) {
3489                     opj_thread_pool_wait_completion(tp, 0);
3490                     opj_free(job);
3491                     opj_aligned_free(h.wavelet);
3492                     return OPJ_FALSE;
3493                 }
3494                 job->v.dn = v.dn;
3495                 job->v.sn = v.sn;
3496                 job->v.cas = v.cas;
3497                 job->v.win_l_x0 = v.win_l_x0;
3498                 job->v.win_l_x1 = v.win_l_x1;
3499                 job->v.win_h_x0 = v.win_h_x0;
3500                 job->v.win_h_x1 = v.win_h_x1;
3501                 job->rh = rh;
3502                 job->w = w;
3503                 job->aj = aj;
3504                 job->nb_columns = (j + 1 == num_jobs) ? (rw & (OPJ_UINT32)~
3505                                   (NB_ELTS_V8 - 1)) - j * step_j : step_j;
3506                 aj += job->nb_columns;
3507                 opj_thread_pool_submit_job(tp, opj_dwt97_decode_v_func, job);
3508             }
3509             opj_thread_pool_wait_completion(tp, 0);
3510         }
3511 
3512         if (rw & (NB_ELTS_V8 - 1)) {
3513             OPJ_UINT32 k;
3514 
3515             j = rw & (NB_ELTS_V8 - 1);
3516 
3517             opj_v8dwt_interleave_v(&v, aj, w, j);
3518             opj_v8dwt_decode(&v);
3519 
3520             for (k = 0; k < rh; ++k) {
3521                 memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k],
3522                        (OPJ_SIZE_T)j * sizeof(OPJ_FLOAT32));
3523             }
3524         }
3525     }
3526 
3527     opj_aligned_free(h.wavelet);
3528     return OPJ_TRUE;
3529 }
3530 
3531 static
opj_dwt_decode_partial_97(opj_tcd_tilecomp_t * OPJ_RESTRICT tilec,OPJ_UINT32 numres)3532 OPJ_BOOL opj_dwt_decode_partial_97(opj_tcd_tilecomp_t* OPJ_RESTRICT tilec,
3533                                    OPJ_UINT32 numres)
3534 {
3535     opj_sparse_array_int32_t* sa;
3536     opj_v8dwt_t h;
3537     opj_v8dwt_t v;
3538     OPJ_UINT32 resno;
3539     /* This value matches the maximum left/right extension given in tables */
3540     /* F.2 and F.3 of the standard. Note: in opj_tcd_is_subband_area_of_interest() */
3541     /* we currently use 3. */
3542     const OPJ_UINT32 filter_width = 4U;
3543 
3544     opj_tcd_resolution_t* tr = tilec->resolutions;
3545     opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]);
3546 
3547     OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 -
3548                                  tr->x0);    /* width of the resolution level computed */
3549     OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
3550                                  tr->y0);    /* height of the resolution level computed */
3551 
3552     OPJ_SIZE_T l_data_size;
3553 
3554     /* Compute the intersection of the area of interest, expressed in tile coordinates */
3555     /* with the tile coordinates */
3556     OPJ_UINT32 win_tcx0 = tilec->win_x0;
3557     OPJ_UINT32 win_tcy0 = tilec->win_y0;
3558     OPJ_UINT32 win_tcx1 = tilec->win_x1;
3559     OPJ_UINT32 win_tcy1 = tilec->win_y1;
3560 
3561     if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) {
3562         return OPJ_TRUE;
3563     }
3564 
3565     sa = opj_dwt_init_sparse_array(tilec, numres);
3566     if (sa == NULL) {
3567         return OPJ_FALSE;
3568     }
3569 
3570     if (numres == 1U) {
3571         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
3572                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
3573                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
3574                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
3575                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
3576                        tilec->data_win,
3577                        1, tr_max->win_x1 - tr_max->win_x0,
3578                        OPJ_TRUE);
3579         assert(ret);
3580         OPJ_UNUSED(ret);
3581         opj_sparse_array_int32_free(sa);
3582         return OPJ_TRUE;
3583     }
3584 
3585     l_data_size = opj_dwt_max_resolution(tr, numres);
3586     /* overflow check */
3587     if (l_data_size > (SIZE_MAX / sizeof(opj_v8_t))) {
3588         /* FIXME event manager error callback */
3589         opj_sparse_array_int32_free(sa);
3590         return OPJ_FALSE;
3591     }
3592     h.wavelet = (opj_v8_t*) opj_aligned_malloc(l_data_size * sizeof(opj_v8_t));
3593     if (!h.wavelet) {
3594         /* FIXME event manager error callback */
3595         opj_sparse_array_int32_free(sa);
3596         return OPJ_FALSE;
3597     }
3598     v.wavelet = h.wavelet;
3599 
3600     for (resno = 1; resno < numres; resno ++) {
3601         OPJ_UINT32 j;
3602         /* Window of interest subband-based coordinates */
3603         OPJ_UINT32 win_ll_x0, win_ll_y0, win_ll_x1, win_ll_y1;
3604         OPJ_UINT32 win_hl_x0, win_hl_x1;
3605         OPJ_UINT32 win_lh_y0, win_lh_y1;
3606         /* Window of interest tile-resolution-based coordinates */
3607         OPJ_UINT32 win_tr_x0, win_tr_x1, win_tr_y0, win_tr_y1;
3608         /* Tile-resolution subband-based coordinates */
3609         OPJ_UINT32 tr_ll_x0, tr_ll_y0, tr_hl_x0, tr_lh_y0;
3610 
3611         ++tr;
3612 
3613         h.sn = (OPJ_INT32)rw;
3614         v.sn = (OPJ_INT32)rh;
3615 
3616         rw = (OPJ_UINT32)(tr->x1 - tr->x0);
3617         rh = (OPJ_UINT32)(tr->y1 - tr->y0);
3618 
3619         h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn);
3620         h.cas = tr->x0 % 2;
3621 
3622         v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn);
3623         v.cas = tr->y0 % 2;
3624 
3625         /* Get the subband coordinates for the window of interest */
3626         /* LL band */
3627         opj_dwt_get_band_coordinates(tilec, resno, 0,
3628                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
3629                                      &win_ll_x0, &win_ll_y0,
3630                                      &win_ll_x1, &win_ll_y1);
3631 
3632         /* HL band */
3633         opj_dwt_get_band_coordinates(tilec, resno, 1,
3634                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
3635                                      &win_hl_x0, NULL, &win_hl_x1, NULL);
3636 
3637         /* LH band */
3638         opj_dwt_get_band_coordinates(tilec, resno, 2,
3639                                      win_tcx0, win_tcy0, win_tcx1, win_tcy1,
3640                                      NULL, &win_lh_y0, NULL, &win_lh_y1);
3641 
3642         /* Beware: band index for non-LL0 resolution are 0=HL, 1=LH and 2=HH */
3643         tr_ll_x0 = (OPJ_UINT32)tr->bands[1].x0;
3644         tr_ll_y0 = (OPJ_UINT32)tr->bands[0].y0;
3645         tr_hl_x0 = (OPJ_UINT32)tr->bands[0].x0;
3646         tr_lh_y0 = (OPJ_UINT32)tr->bands[1].y0;
3647 
3648         /* Subtract the origin of the bands for this tile, to the subwindow */
3649         /* of interest band coordinates, so as to get them relative to the */
3650         /* tile */
3651         win_ll_x0 = opj_uint_subs(win_ll_x0, tr_ll_x0);
3652         win_ll_y0 = opj_uint_subs(win_ll_y0, tr_ll_y0);
3653         win_ll_x1 = opj_uint_subs(win_ll_x1, tr_ll_x0);
3654         win_ll_y1 = opj_uint_subs(win_ll_y1, tr_ll_y0);
3655         win_hl_x0 = opj_uint_subs(win_hl_x0, tr_hl_x0);
3656         win_hl_x1 = opj_uint_subs(win_hl_x1, tr_hl_x0);
3657         win_lh_y0 = opj_uint_subs(win_lh_y0, tr_lh_y0);
3658         win_lh_y1 = opj_uint_subs(win_lh_y1, tr_lh_y0);
3659 
3660         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.sn, &win_ll_x0, &win_ll_x1);
3661         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.dn, &win_hl_x0, &win_hl_x1);
3662 
3663         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.sn, &win_ll_y0, &win_ll_y1);
3664         opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.dn, &win_lh_y0, &win_lh_y1);
3665 
3666         /* Compute the tile-resolution-based coordinates for the window of interest */
3667         if (h.cas == 0) {
3668             win_tr_x0 = opj_uint_min(2 * win_ll_x0, 2 * win_hl_x0 + 1);
3669             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_ll_x1, 2 * win_hl_x1 + 1), rw);
3670         } else {
3671             win_tr_x0 = opj_uint_min(2 * win_hl_x0, 2 * win_ll_x0 + 1);
3672             win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_hl_x1, 2 * win_ll_x1 + 1), rw);
3673         }
3674 
3675         if (v.cas == 0) {
3676             win_tr_y0 = opj_uint_min(2 * win_ll_y0, 2 * win_lh_y0 + 1);
3677             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_ll_y1, 2 * win_lh_y1 + 1), rh);
3678         } else {
3679             win_tr_y0 = opj_uint_min(2 * win_lh_y0, 2 * win_ll_y0 + 1);
3680             win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_lh_y1, 2 * win_ll_y1 + 1), rh);
3681         }
3682 
3683         h.win_l_x0 = win_ll_x0;
3684         h.win_l_x1 = win_ll_x1;
3685         h.win_h_x0 = win_hl_x0;
3686         h.win_h_x1 = win_hl_x1;
3687         for (j = 0; j + (NB_ELTS_V8 - 1) < rh; j += NB_ELTS_V8) {
3688             if ((j + (NB_ELTS_V8 - 1) >= win_ll_y0 && j < win_ll_y1) ||
3689                     (j + (NB_ELTS_V8 - 1) >= win_lh_y0 + (OPJ_UINT32)v.sn &&
3690                      j < win_lh_y1 + (OPJ_UINT32)v.sn)) {
3691                 opj_v8dwt_interleave_partial_h(&h, sa, j, opj_uint_min(NB_ELTS_V8, rh - j));
3692                 opj_v8dwt_decode(&h);
3693                 if (!opj_sparse_array_int32_write(sa,
3694                                                   win_tr_x0, j,
3695                                                   win_tr_x1, j + NB_ELTS_V8,
3696                                                   (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0],
3697                                                   NB_ELTS_V8, 1, OPJ_TRUE)) {
3698                     /* FIXME event manager error callback */
3699                     opj_sparse_array_int32_free(sa);
3700                     opj_aligned_free(h.wavelet);
3701                     return OPJ_FALSE;
3702                 }
3703             }
3704         }
3705 
3706         if (j < rh &&
3707                 ((j + (NB_ELTS_V8 - 1) >= win_ll_y0 && j < win_ll_y1) ||
3708                  (j + (NB_ELTS_V8 - 1) >= win_lh_y0 + (OPJ_UINT32)v.sn &&
3709                   j < win_lh_y1 + (OPJ_UINT32)v.sn))) {
3710             opj_v8dwt_interleave_partial_h(&h, sa, j, rh - j);
3711             opj_v8dwt_decode(&h);
3712             if (!opj_sparse_array_int32_write(sa,
3713                                               win_tr_x0, j,
3714                                               win_tr_x1, rh,
3715                                               (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0],
3716                                               NB_ELTS_V8, 1, OPJ_TRUE)) {
3717                 /* FIXME event manager error callback */
3718                 opj_sparse_array_int32_free(sa);
3719                 opj_aligned_free(h.wavelet);
3720                 return OPJ_FALSE;
3721             }
3722         }
3723 
3724         v.win_l_x0 = win_ll_y0;
3725         v.win_l_x1 = win_ll_y1;
3726         v.win_h_x0 = win_lh_y0;
3727         v.win_h_x1 = win_lh_y1;
3728         for (j = win_tr_x0; j < win_tr_x1; j += NB_ELTS_V8) {
3729             OPJ_UINT32 nb_elts = opj_uint_min(NB_ELTS_V8, win_tr_x1 - j);
3730 
3731             opj_v8dwt_interleave_partial_v(&v, sa, j, nb_elts);
3732             opj_v8dwt_decode(&v);
3733 
3734             if (!opj_sparse_array_int32_write(sa,
3735                                               j, win_tr_y0,
3736                                               j + nb_elts, win_tr_y1,
3737                                               (OPJ_INT32*)&h.wavelet[win_tr_y0].f[0],
3738                                               1, NB_ELTS_V8, OPJ_TRUE)) {
3739                 /* FIXME event manager error callback */
3740                 opj_sparse_array_int32_free(sa);
3741                 opj_aligned_free(h.wavelet);
3742                 return OPJ_FALSE;
3743             }
3744         }
3745     }
3746 
3747     {
3748         OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
3749                        tr_max->win_x0 - (OPJ_UINT32)tr_max->x0,
3750                        tr_max->win_y0 - (OPJ_UINT32)tr_max->y0,
3751                        tr_max->win_x1 - (OPJ_UINT32)tr_max->x0,
3752                        tr_max->win_y1 - (OPJ_UINT32)tr_max->y0,
3753                        tilec->data_win,
3754                        1, tr_max->win_x1 - tr_max->win_x0,
3755                        OPJ_TRUE);
3756         assert(ret);
3757         OPJ_UNUSED(ret);
3758     }
3759     opj_sparse_array_int32_free(sa);
3760 
3761     opj_aligned_free(h.wavelet);
3762     return OPJ_TRUE;
3763 }
3764 
3765 
opj_dwt_decode_real(opj_tcd_t * p_tcd,opj_tcd_tilecomp_t * OPJ_RESTRICT tilec,OPJ_UINT32 numres)3766 OPJ_BOOL opj_dwt_decode_real(opj_tcd_t *p_tcd,
3767                              opj_tcd_tilecomp_t* OPJ_RESTRICT tilec,
3768                              OPJ_UINT32 numres)
3769 {
3770     if (p_tcd->whole_tile_decoding) {
3771         return opj_dwt_decode_tile_97(p_tcd->thread_pool, tilec, numres);
3772     } else {
3773         return opj_dwt_decode_partial_97(tilec, numres);
3774     }
3775 }
3776