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 | #define OPJ_WS(i) v->mem[(i)*2] |
67 | #define OPJ_WD(i) v->mem[(1+(i)*2)] |
68 | |
69 | #ifdef __AVX2__ |
70 | /** Number of int32 values in a AVX2 register */ |
71 | #define VREG_INT_COUNT 8 |
72 | #else |
73 | /** Number of int32 values in a SSE2 register */ |
74 | #define VREG_INT_COUNT 4 |
75 | #endif |
76 | |
77 | /** Number of columns that we can process in parallel in the vertical pass */ |
78 | #define PARALLEL_COLS_53 (2*VREG_INT_COUNT) |
79 | |
80 | /** @name Local data structures */ |
81 | /*@{*/ |
82 | |
83 | typedef struct dwt_local { |
84 | OPJ_INT32* mem; |
85 | OPJ_INT32 dn; /* number of elements in high pass band */ |
86 | OPJ_INT32 sn; /* number of elements in low pass band */ |
87 | OPJ_INT32 cas; /* 0 = start on even coord, 1 = start on odd coord */ |
88 | } opj_dwt_t; |
89 | |
90 | typedef union { |
91 | OPJ_FLOAT32 f[4]; |
92 | } opj_v4_t; |
93 | |
94 | typedef struct v4dwt_local { |
95 | opj_v4_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_v4dwt_t ; |
104 | |
105 | static const OPJ_FLOAT32 opj_dwt_alpha = 1.586134342f; /* 12994 */ |
106 | static const OPJ_FLOAT32 opj_dwt_beta = 0.052980118f; /* 434 */ |
107 | static const OPJ_FLOAT32 opj_dwt_gamma = -0.882911075f; /* -7233 */ |
108 | static const OPJ_FLOAT32 opj_dwt_delta = -0.443506852f; /* -3633 */ |
109 | |
110 | static const OPJ_FLOAT32 opj_K = 1.230174105f; /* 10078 */ |
111 | static const OPJ_FLOAT32 opj_c13318 = 1.625732422f; |
112 | |
113 | /*@}*/ |
114 | |
115 | /** |
116 | Virtual function type for wavelet transform in 1-D |
117 | */ |
118 | typedef void (*DWT1DFN)(const opj_dwt_t* v); |
119 | |
120 | /** @name Local static functions */ |
121 | /*@{*/ |
122 | |
123 | /** |
124 | Forward lazy transform (horizontal) |
125 | */ |
126 | static void opj_dwt_deinterleave_h(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn, |
127 | OPJ_INT32 sn, OPJ_INT32 cas); |
128 | /** |
129 | Forward lazy transform (vertical) |
130 | */ |
131 | static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn, |
132 | OPJ_INT32 sn, OPJ_INT32 x, OPJ_INT32 cas); |
133 | /** |
134 | Forward 5-3 wavelet transform in 1-D |
135 | */ |
136 | static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, |
137 | OPJ_INT32 cas); |
138 | /** |
139 | Forward 9-7 wavelet transform in 1-D |
140 | */ |
141 | static void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, |
142 | OPJ_INT32 cas); |
143 | /** |
144 | Explicit calculation of the Quantization Stepsizes |
145 | */ |
146 | static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps, |
147 | opj_stepsize_t *bandno_stepsize); |
148 | /** |
149 | Inverse wavelet transform in 2-D. |
150 | */ |
151 | static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp, |
152 | opj_tcd_tilecomp_t* tilec, OPJ_UINT32 i); |
153 | |
154 | static OPJ_BOOL opj_dwt_decode_partial_tile( |
155 | opj_tcd_tilecomp_t* tilec, |
156 | OPJ_UINT32 numres); |
157 | |
158 | static OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec, |
159 | void (*p_function)(OPJ_INT32 *, OPJ_INT32, OPJ_INT32, OPJ_INT32)); |
160 | |
161 | static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r, |
162 | OPJ_UINT32 i); |
163 | |
164 | /* <summary> */ |
165 | /* Inverse 9-7 wavelet transform in 1-D. */ |
166 | /* </summary> */ |
167 | static void opj_v4dwt_decode(opj_v4dwt_t* OPJ_RESTRICT dwt); |
168 | |
169 | static void opj_v4dwt_interleave_h(opj_v4dwt_t* OPJ_RESTRICT dwt, |
170 | OPJ_FLOAT32* OPJ_RESTRICT a, |
171 | OPJ_UINT32 width, |
172 | OPJ_UINT32 remaining_height); |
173 | |
174 | static void opj_v4dwt_interleave_v(opj_v4dwt_t* OPJ_RESTRICT dwt, |
175 | OPJ_FLOAT32* OPJ_RESTRICT a, |
176 | OPJ_UINT32 width, |
177 | OPJ_UINT32 nb_elts_read); |
178 | |
179 | #ifdef __SSE__ |
180 | static void opj_v4dwt_decode_step1_sse(opj_v4_t* w, |
181 | OPJ_UINT32 start, |
182 | OPJ_UINT32 end, |
183 | const __m128 c); |
184 | |
185 | static void opj_v4dwt_decode_step2_sse(opj_v4_t* l, opj_v4_t* w, |
186 | OPJ_UINT32 start, |
187 | OPJ_UINT32 end, |
188 | OPJ_UINT32 m, __m128 c); |
189 | |
190 | #else |
191 | static void opj_v4dwt_decode_step1(opj_v4_t* w, |
192 | OPJ_UINT32 start, |
193 | OPJ_UINT32 end, |
194 | const OPJ_FLOAT32 c); |
195 | |
196 | static void opj_v4dwt_decode_step2(opj_v4_t* l, opj_v4_t* w, |
197 | OPJ_UINT32 start, |
198 | OPJ_UINT32 end, |
199 | OPJ_UINT32 m, |
200 | OPJ_FLOAT32 c); |
201 | |
202 | #endif |
203 | |
204 | /*@}*/ |
205 | |
206 | /*@}*/ |
207 | |
208 | #define OPJ_S(i) a[(i)*2] |
209 | #define OPJ_D(i) a[(1+(i)*2)] |
210 | #define OPJ_S_(i) ((i)<0?OPJ_S(0):((i)>=sn?OPJ_S(sn-1):OPJ_S(i))) |
211 | #define OPJ_D_(i) ((i)<0?OPJ_D(0):((i)>=dn?OPJ_D(dn-1):OPJ_D(i))) |
212 | /* new */ |
213 | #define OPJ_SS_(i) ((i)<0?OPJ_S(0):((i)>=dn?OPJ_S(dn-1):OPJ_S(i))) |
214 | #define OPJ_DD_(i) ((i)<0?OPJ_D(0):((i)>=sn?OPJ_D(sn-1):OPJ_D(i))) |
215 | |
216 | /* <summary> */ |
217 | /* This table contains the norms of the 5-3 wavelets for different bands. */ |
218 | /* </summary> */ |
219 | /* FIXME! the array should really be extended up to 33 resolution levels */ |
220 | /* See https://github.com/uclouvain/openjpeg/issues/493 */ |
221 | static const OPJ_FLOAT64 opj_dwt_norms[4][10] = { |
222 | {1.000, 1.500, 2.750, 5.375, 10.68, 21.34, 42.67, 85.33, 170.7, 341.3}, |
223 | {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9}, |
224 | {1.038, 1.592, 2.919, 5.703, 11.33, 22.64, 45.25, 90.48, 180.9}, |
225 | {.7186, .9218, 1.586, 3.043, 6.019, 12.01, 24.00, 47.97, 95.93} |
226 | }; |
227 | |
228 | /* <summary> */ |
229 | /* This table contains the norms of the 9-7 wavelets for different bands. */ |
230 | /* </summary> */ |
231 | /* FIXME! the array should really be extended up to 33 resolution levels */ |
232 | /* See https://github.com/uclouvain/openjpeg/issues/493 */ |
233 | static const OPJ_FLOAT64 opj_dwt_norms_real[4][10] = { |
234 | {1.000, 1.965, 4.177, 8.403, 16.90, 33.84, 67.69, 135.3, 270.6, 540.9}, |
235 | {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0}, |
236 | {2.022, 3.989, 8.355, 17.04, 34.27, 68.63, 137.3, 274.6, 549.0}, |
237 | {2.080, 3.865, 8.307, 17.18, 34.71, 69.59, 139.3, 278.6, 557.2} |
238 | }; |
239 | |
240 | /* |
241 | ========================================================== |
242 | local functions |
243 | ========================================================== |
244 | */ |
245 | |
246 | /* <summary> */ |
247 | /* Forward lazy transform (horizontal). */ |
248 | /* </summary> */ |
249 | static void opj_dwt_deinterleave_h(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn, |
250 | OPJ_INT32 sn, OPJ_INT32 cas) |
251 | { |
252 | OPJ_INT32 i; |
253 | OPJ_INT32 * l_dest = b; |
254 | OPJ_INT32 * l_src = a + cas; |
255 | |
256 | for (i = 0; i < sn; ++i) { |
257 | *l_dest++ = *l_src; |
258 | l_src += 2; |
259 | } |
260 | |
261 | l_dest = b + sn; |
262 | l_src = a + 1 - cas; |
263 | |
264 | for (i = 0; i < dn; ++i) { |
265 | *l_dest++ = *l_src; |
266 | l_src += 2; |
267 | } |
268 | } |
269 | |
270 | /* <summary> */ |
271 | /* Forward lazy transform (vertical). */ |
272 | /* </summary> */ |
273 | static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn, |
274 | OPJ_INT32 sn, OPJ_INT32 x, OPJ_INT32 cas) |
275 | { |
276 | OPJ_INT32 i = sn; |
277 | OPJ_INT32 * l_dest = b; |
278 | OPJ_INT32 * l_src = a + cas; |
279 | |
280 | while (i--) { |
281 | *l_dest = *l_src; |
282 | l_dest += x; |
283 | l_src += 2; |
284 | } /* b[i*x]=a[2*i+cas]; */ |
285 | |
286 | l_dest = b + (OPJ_SIZE_T)sn * (OPJ_SIZE_T)x; |
287 | l_src = a + 1 - cas; |
288 | |
289 | i = dn; |
290 | while (i--) { |
291 | *l_dest = *l_src; |
292 | l_dest += x; |
293 | l_src += 2; |
294 | } /*b[(sn+i)*x]=a[(2*i+1-cas)];*/ |
295 | } |
296 | |
297 | #ifdef STANDARD_SLOW_VERSION |
298 | /* <summary> */ |
299 | /* Inverse lazy transform (horizontal). */ |
300 | /* </summary> */ |
301 | static void opj_dwt_interleave_h(const opj_dwt_t* h, OPJ_INT32 *a) |
302 | { |
303 | OPJ_INT32 *ai = a; |
304 | OPJ_INT32 *bi = h->mem + h->cas; |
305 | OPJ_INT32 i = h->sn; |
306 | while (i--) { |
307 | *bi = *(ai++); |
308 | bi += 2; |
309 | } |
310 | ai = a + h->sn; |
311 | bi = h->mem + 1 - h->cas; |
312 | i = h->dn ; |
313 | while (i--) { |
314 | *bi = *(ai++); |
315 | bi += 2; |
316 | } |
317 | } |
318 | |
319 | /* <summary> */ |
320 | /* Inverse lazy transform (vertical). */ |
321 | /* </summary> */ |
322 | static void opj_dwt_interleave_v(const opj_dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x) |
323 | { |
324 | OPJ_INT32 *ai = a; |
325 | OPJ_INT32 *bi = v->mem + v->cas; |
326 | OPJ_INT32 i = v->sn; |
327 | while (i--) { |
328 | *bi = *ai; |
329 | bi += 2; |
330 | ai += x; |
331 | } |
332 | ai = a + (v->sn * (OPJ_SIZE_T)x); |
333 | bi = v->mem + 1 - v->cas; |
334 | i = v->dn ; |
335 | while (i--) { |
336 | *bi = *ai; |
337 | bi += 2; |
338 | ai += x; |
339 | } |
340 | } |
341 | |
342 | #endif /* STANDARD_SLOW_VERSION */ |
343 | |
344 | /* <summary> */ |
345 | /* Forward 5-3 wavelet transform in 1-D. */ |
346 | /* </summary> */ |
347 | static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, |
348 | OPJ_INT32 cas) |
349 | { |
350 | OPJ_INT32 i; |
351 | |
352 | if (!cas) { |
353 | if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ |
354 | for (i = 0; i < dn; i++) { |
355 | OPJ_D(i) -= (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1; |
356 | } |
357 | for (i = 0; i < sn; i++) { |
358 | OPJ_S(i) += (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2; |
359 | } |
360 | } |
361 | } else { |
362 | if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */ |
363 | OPJ_S(0) *= 2; |
364 | } else { |
365 | for (i = 0; i < dn; i++) { |
366 | OPJ_S(i) -= (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1; |
367 | } |
368 | for (i = 0; i < sn; i++) { |
369 | OPJ_D(i) += (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2; |
370 | } |
371 | } |
372 | } |
373 | } |
374 | |
375 | #ifdef STANDARD_SLOW_VERSION |
376 | /* <summary> */ |
377 | /* Inverse 5-3 wavelet transform in 1-D. */ |
378 | /* </summary> */ |
379 | static void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, |
380 | OPJ_INT32 cas) |
381 | { |
382 | OPJ_INT32 i; |
383 | |
384 | if (!cas) { |
385 | if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ |
386 | for (i = 0; i < sn; i++) { |
387 | OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2; |
388 | } |
389 | for (i = 0; i < dn; i++) { |
390 | OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1; |
391 | } |
392 | } |
393 | } else { |
394 | if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */ |
395 | OPJ_S(0) /= 2; |
396 | } else { |
397 | for (i = 0; i < sn; i++) { |
398 | OPJ_D(i) -= (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2; |
399 | } |
400 | for (i = 0; i < dn; i++) { |
401 | OPJ_S(i) += (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1; |
402 | } |
403 | } |
404 | } |
405 | } |
406 | |
407 | static void opj_dwt_decode_1(const opj_dwt_t *v) |
408 | { |
409 | opj_dwt_decode_1_(v->mem, v->dn, v->sn, v->cas); |
410 | } |
411 | |
412 | #endif /* STANDARD_SLOW_VERSION */ |
413 | |
414 | #if !defined(STANDARD_SLOW_VERSION) |
415 | static void opj_idwt53_h_cas0(OPJ_INT32* tmp, |
416 | const OPJ_INT32 sn, |
417 | const OPJ_INT32 len, |
418 | OPJ_INT32* tiledp) |
419 | { |
420 | OPJ_INT32 i, j; |
421 | const OPJ_INT32* in_even = &tiledp[0]; |
422 | const OPJ_INT32* in_odd = &tiledp[sn]; |
423 | |
424 | #ifdef TWO_PASS_VERSION |
425 | /* For documentation purpose: performs lifting in two iterations, */ |
426 | /* but without explicit interleaving */ |
427 | |
428 | assert(len > 1); |
429 | |
430 | /* Even */ |
431 | tmp[0] = in_even[0] - ((in_odd[0] + 1) >> 1); |
432 | for (i = 2, j = 0; i <= len - 2; i += 2, j++) { |
433 | tmp[i] = in_even[j + 1] - ((in_odd[j] + in_odd[j + 1] + 2) >> 2); |
434 | } |
435 | if (len & 1) { /* if len is odd */ |
436 | tmp[len - 1] = in_even[(len - 1) / 2] - ((in_odd[(len - 2) / 2] + 1) >> 1); |
437 | } |
438 | |
439 | /* Odd */ |
440 | for (i = 1, j = 0; i < len - 1; i += 2, j++) { |
441 | tmp[i] = in_odd[j] + ((tmp[i - 1] + tmp[i + 1]) >> 1); |
442 | } |
443 | if (!(len & 1)) { /* if len is even */ |
444 | tmp[len - 1] = in_odd[(len - 1) / 2] + tmp[len - 2]; |
445 | } |
446 | #else |
447 | OPJ_INT32 d1c, d1n, s1n, s0c, s0n; |
448 | |
449 | assert(len > 1); |
450 | |
451 | /* Improved version of the TWO_PASS_VERSION: */ |
452 | /* Performs lifting in one single iteration. Saves memory */ |
453 | /* accesses and explicit interleaving. */ |
454 | s1n = in_even[0]; |
455 | d1n = in_odd[0]; |
456 | s0n = s1n - ((d1n + 1) >> 1); |
457 | |
458 | for (i = 0, j = 1; i < (len - 3); i += 2, j++) { |
459 | d1c = d1n; |
460 | s0c = s0n; |
461 | |
462 | s1n = in_even[j]; |
463 | d1n = in_odd[j]; |
464 | |
465 | s0n = s1n - ((d1c + d1n + 2) >> 2); |
466 | |
467 | tmp[i ] = s0c; |
468 | tmp[i + 1] = d1c + ((s0c + s0n) >> 1); |
469 | } |
470 | |
471 | tmp[i] = s0n; |
472 | |
473 | if (len & 1) { |
474 | tmp[len - 1] = in_even[(len - 1) / 2] - ((d1n + 1) >> 1); |
475 | tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1); |
476 | } else { |
477 | tmp[len - 1] = d1n + s0n; |
478 | } |
479 | #endif |
480 | memcpy(tiledp, tmp, (OPJ_UINT32)len * sizeof(OPJ_INT32)); |
481 | } |
482 | |
483 | static void opj_idwt53_h_cas1(OPJ_INT32* tmp, |
484 | const OPJ_INT32 sn, |
485 | const OPJ_INT32 len, |
486 | OPJ_INT32* tiledp) |
487 | { |
488 | OPJ_INT32 i, j; |
489 | const OPJ_INT32* in_even = &tiledp[sn]; |
490 | const OPJ_INT32* in_odd = &tiledp[0]; |
491 | |
492 | #ifdef TWO_PASS_VERSION |
493 | /* For documentation purpose: performs lifting in two iterations, */ |
494 | /* but without explicit interleaving */ |
495 | |
496 | assert(len > 2); |
497 | |
498 | /* Odd */ |
499 | for (i = 1, j = 0; i < len - 1; i += 2, j++) { |
500 | tmp[i] = in_odd[j] - ((in_even[j] + in_even[j + 1] + 2) >> 2); |
501 | } |
502 | if (!(len & 1)) { |
503 | tmp[len - 1] = in_odd[len / 2 - 1] - ((in_even[len / 2 - 1] + 1) >> 1); |
504 | } |
505 | |
506 | /* Even */ |
507 | tmp[0] = in_even[0] + tmp[1]; |
508 | for (i = 2, j = 1; i < len - 1; i += 2, j++) { |
509 | tmp[i] = in_even[j] + ((tmp[i + 1] + tmp[i - 1]) >> 1); |
510 | } |
511 | if (len & 1) { |
512 | tmp[len - 1] = in_even[len / 2] + tmp[len - 2]; |
513 | } |
514 | #else |
515 | OPJ_INT32 s1, s2, dc, dn; |
516 | |
517 | assert(len > 2); |
518 | |
519 | /* Improved version of the TWO_PASS_VERSION: */ |
520 | /* Performs lifting in one single iteration. Saves memory */ |
521 | /* accesses and explicit interleaving. */ |
522 | |
523 | s1 = in_even[1]; |
524 | dc = in_odd[0] - ((in_even[0] + s1 + 2) >> 2); |
525 | tmp[0] = in_even[0] + dc; |
526 | |
527 | for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) { |
528 | |
529 | s2 = in_even[j + 1]; |
530 | |
531 | dn = in_odd[j] - ((s1 + s2 + 2) >> 2); |
532 | tmp[i ] = dc; |
533 | tmp[i + 1] = s1 + ((dn + dc) >> 1); |
534 | |
535 | dc = dn; |
536 | s1 = s2; |
537 | } |
538 | |
539 | tmp[i] = dc; |
540 | |
541 | if (!(len & 1)) { |
542 | dn = in_odd[len / 2 - 1] - ((s1 + 1) >> 1); |
543 | tmp[len - 2] = s1 + ((dn + dc) >> 1); |
544 | tmp[len - 1] = dn; |
545 | } else { |
546 | tmp[len - 1] = s1 + dc; |
547 | } |
548 | #endif |
549 | memcpy(tiledp, tmp, (OPJ_UINT32)len * sizeof(OPJ_INT32)); |
550 | } |
551 | |
552 | |
553 | #endif /* !defined(STANDARD_SLOW_VERSION) */ |
554 | |
555 | /* <summary> */ |
556 | /* Inverse 5-3 wavelet transform in 1-D for one row. */ |
557 | /* </summary> */ |
558 | /* Performs interleave, inverse wavelet transform and copy back to buffer */ |
559 | static void opj_idwt53_h(const opj_dwt_t *dwt, |
560 | OPJ_INT32* tiledp) |
561 | { |
562 | #ifdef STANDARD_SLOW_VERSION |
563 | /* For documentation purpose */ |
564 | opj_dwt_interleave_h(dwt, tiledp); |
565 | opj_dwt_decode_1(dwt); |
566 | memcpy(tiledp, dwt->mem, (OPJ_UINT32)(dwt->sn + dwt->dn) * sizeof(OPJ_INT32)); |
567 | #else |
568 | const OPJ_INT32 sn = dwt->sn; |
569 | const OPJ_INT32 len = sn + dwt->dn; |
570 | if (dwt->cas == 0) { /* Left-most sample is on even coordinate */ |
571 | if (len > 1) { |
572 | opj_idwt53_h_cas0(dwt->mem, sn, len, tiledp); |
573 | } else { |
574 | /* Unmodified value */ |
575 | } |
576 | } else { /* Left-most sample is on odd coordinate */ |
577 | if (len == 1) { |
578 | tiledp[0] /= 2; |
579 | } else if (len == 2) { |
580 | OPJ_INT32* out = dwt->mem; |
581 | const OPJ_INT32* in_even = &tiledp[sn]; |
582 | const OPJ_INT32* in_odd = &tiledp[0]; |
583 | out[1] = in_odd[0] - ((in_even[0] + 1) >> 1); |
584 | out[0] = in_even[0] + out[1]; |
585 | memcpy(tiledp, dwt->mem, (OPJ_UINT32)len * sizeof(OPJ_INT32)); |
586 | } else if (len > 2) { |
587 | opj_idwt53_h_cas1(dwt->mem, sn, len, tiledp); |
588 | } |
589 | } |
590 | #endif |
591 | } |
592 | |
593 | #if (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION) |
594 | |
595 | /* Conveniency macros to improve the readabilty of the formulas */ |
596 | #if __AVX2__ |
597 | #define VREG __m256i |
598 | #define LOAD_CST(x) _mm256_set1_epi32(x) |
599 | #define LOAD(x) _mm256_load_si256((const VREG*)(x)) |
600 | #define LOADU(x) _mm256_loadu_si256((const VREG*)(x)) |
601 | #define STORE(x,y) _mm256_store_si256((VREG*)(x),(y)) |
602 | #define STOREU(x,y) _mm256_storeu_si256((VREG*)(x),(y)) |
603 | #define ADD(x,y) _mm256_add_epi32((x),(y)) |
604 | #define SUB(x,y) _mm256_sub_epi32((x),(y)) |
605 | #define SAR(x,y) _mm256_srai_epi32((x),(y)) |
606 | #else |
607 | #define VREG __m128i |
608 | #define LOAD_CST(x) _mm_set1_epi32(x) |
609 | #define LOAD(x) _mm_load_si128((const VREG*)(x)) |
610 | #define LOADU(x) _mm_loadu_si128((const VREG*)(x)) |
611 | #define STORE(x,y) _mm_store_si128((VREG*)(x),(y)) |
612 | #define STOREU(x,y) _mm_storeu_si128((VREG*)(x),(y)) |
613 | #define ADD(x,y) _mm_add_epi32((x),(y)) |
614 | #define SUB(x,y) _mm_sub_epi32((x),(y)) |
615 | #define SAR(x,y) _mm_srai_epi32((x),(y)) |
616 | #endif |
617 | #define ADD3(x,y,z) ADD(ADD(x,y),z) |
618 | |
619 | static |
620 | void opj_idwt53_v_final_memcpy(OPJ_INT32* tiledp_col, |
621 | const OPJ_INT32* tmp, |
622 | OPJ_INT32 len, |
623 | OPJ_SIZE_T stride) |
624 | { |
625 | OPJ_INT32 i; |
626 | for (i = 0; i < len; ++i) { |
627 | /* A memcpy(&tiledp_col[i * stride + 0], |
628 | &tmp[PARALLEL_COLS_53 * i + 0], |
629 | PARALLEL_COLS_53 * sizeof(OPJ_INT32)) |
630 | would do but would be a tiny bit slower. |
631 | We can take here advantage of our knowledge of alignment */ |
632 | STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + 0], |
633 | LOAD(&tmp[PARALLEL_COLS_53 * i + 0])); |
634 | STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + VREG_INT_COUNT], |
635 | LOAD(&tmp[PARALLEL_COLS_53 * i + VREG_INT_COUNT])); |
636 | } |
637 | } |
638 | |
639 | /** Vertical inverse 5x3 wavelet transform for 8 columns in SSE2, or |
640 | * 16 in AVX2, when top-most pixel is on even coordinate */ |
641 | static void opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2( |
642 | OPJ_INT32* tmp, |
643 | const OPJ_INT32 sn, |
644 | const OPJ_INT32 len, |
645 | OPJ_INT32* tiledp_col, |
646 | const OPJ_SIZE_T stride) |
647 | { |
648 | const OPJ_INT32* in_even = &tiledp_col[0]; |
649 | const OPJ_INT32* in_odd = &tiledp_col[(OPJ_SIZE_T)sn * stride]; |
650 | |
651 | OPJ_INT32 i; |
652 | OPJ_SIZE_T j; |
653 | VREG d1c_0, d1n_0, s1n_0, s0c_0, s0n_0; |
654 | VREG d1c_1, d1n_1, s1n_1, s0c_1, s0n_1; |
655 | const VREG two = LOAD_CST(2); |
656 | |
657 | assert(len > 1); |
658 | #if __AVX2__ |
659 | assert(PARALLEL_COLS_53 == 16); |
660 | assert(VREG_INT_COUNT == 8); |
661 | #else |
662 | assert(PARALLEL_COLS_53 == 8); |
663 | assert(VREG_INT_COUNT == 4); |
664 | #endif |
665 | |
666 | /* Note: loads of input even/odd values must be done in a unaligned */ |
667 | /* fashion. But stores in tmp can be done with aligned store, since */ |
668 | /* the temporary buffer is properly aligned */ |
669 | assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0); |
670 | |
671 | s1n_0 = LOADU(in_even + 0); |
672 | s1n_1 = LOADU(in_even + VREG_INT_COUNT); |
673 | d1n_0 = LOADU(in_odd); |
674 | d1n_1 = LOADU(in_odd + VREG_INT_COUNT); |
675 | |
676 | /* s0n = s1n - ((d1n + 1) >> 1); <==> */ |
677 | /* s0n = s1n - ((d1n + d1n + 2) >> 2); */ |
678 | s0n_0 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2)); |
679 | s0n_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2)); |
680 | |
681 | for (i = 0, j = 1; i < (len - 3); i += 2, j++) { |
682 | d1c_0 = d1n_0; |
683 | s0c_0 = s0n_0; |
684 | d1c_1 = d1n_1; |
685 | s0c_1 = s0n_1; |
686 | |
687 | s1n_0 = LOADU(in_even + j * stride); |
688 | s1n_1 = LOADU(in_even + j * stride + VREG_INT_COUNT); |
689 | d1n_0 = LOADU(in_odd + j * stride); |
690 | d1n_1 = LOADU(in_odd + j * stride + VREG_INT_COUNT); |
691 | |
692 | /*s0n = s1n - ((d1c + d1n + 2) >> 2);*/ |
693 | s0n_0 = SUB(s1n_0, SAR(ADD3(d1c_0, d1n_0, two), 2)); |
694 | s0n_1 = SUB(s1n_1, SAR(ADD3(d1c_1, d1n_1, two), 2)); |
695 | |
696 | STORE(tmp + PARALLEL_COLS_53 * (i + 0), s0c_0); |
697 | STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0c_1); |
698 | |
699 | /* d1c + ((s0c + s0n) >> 1) */ |
700 | STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0, |
701 | ADD(d1c_0, SAR(ADD(s0c_0, s0n_0), 1))); |
702 | STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT, |
703 | ADD(d1c_1, SAR(ADD(s0c_1, s0n_1), 1))); |
704 | } |
705 | |
706 | STORE(tmp + PARALLEL_COLS_53 * (i + 0) + 0, s0n_0); |
707 | STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0n_1); |
708 | |
709 | if (len & 1) { |
710 | VREG tmp_len_minus_1; |
711 | s1n_0 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride); |
712 | /* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */ |
713 | tmp_len_minus_1 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2)); |
714 | STORE(tmp + PARALLEL_COLS_53 * (len - 1), tmp_len_minus_1); |
715 | /* d1n + ((s0n + tmp_len_minus_1) >> 1) */ |
716 | STORE(tmp + PARALLEL_COLS_53 * (len - 2), |
717 | ADD(d1n_0, SAR(ADD(s0n_0, tmp_len_minus_1), 1))); |
718 | |
719 | s1n_1 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride + VREG_INT_COUNT); |
720 | /* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */ |
721 | tmp_len_minus_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2)); |
722 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, |
723 | tmp_len_minus_1); |
724 | /* d1n + ((s0n + tmp_len_minus_1) >> 1) */ |
725 | STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT, |
726 | ADD(d1n_1, SAR(ADD(s0n_1, tmp_len_minus_1), 1))); |
727 | |
728 | |
729 | } else { |
730 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, |
731 | ADD(d1n_0, s0n_0)); |
732 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, |
733 | ADD(d1n_1, s0n_1)); |
734 | } |
735 | |
736 | opj_idwt53_v_final_memcpy(tiledp_col, tmp, len, stride); |
737 | } |
738 | |
739 | |
740 | /** Vertical inverse 5x3 wavelet transform for 8 columns in SSE2, or |
741 | * 16 in AVX2, when top-most pixel is on odd coordinate */ |
742 | static void opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2( |
743 | OPJ_INT32* tmp, |
744 | const OPJ_INT32 sn, |
745 | const OPJ_INT32 len, |
746 | OPJ_INT32* tiledp_col, |
747 | const OPJ_SIZE_T stride) |
748 | { |
749 | OPJ_INT32 i; |
750 | OPJ_SIZE_T j; |
751 | |
752 | VREG s1_0, s2_0, dc_0, dn_0; |
753 | VREG s1_1, s2_1, dc_1, dn_1; |
754 | const VREG two = LOAD_CST(2); |
755 | |
756 | const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride]; |
757 | const OPJ_INT32* in_odd = &tiledp_col[0]; |
758 | |
759 | assert(len > 2); |
760 | #if __AVX2__ |
761 | assert(PARALLEL_COLS_53 == 16); |
762 | assert(VREG_INT_COUNT == 8); |
763 | #else |
764 | assert(PARALLEL_COLS_53 == 8); |
765 | assert(VREG_INT_COUNT == 4); |
766 | #endif |
767 | |
768 | /* Note: loads of input even/odd values must be done in a unaligned */ |
769 | /* fashion. But stores in tmp can be done with aligned store, since */ |
770 | /* the temporary buffer is properly aligned */ |
771 | assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0); |
772 | |
773 | s1_0 = LOADU(in_even + stride); |
774 | /* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */ |
775 | dc_0 = SUB(LOADU(in_odd + 0), |
776 | SAR(ADD3(LOADU(in_even + 0), s1_0, two), 2)); |
777 | STORE(tmp + PARALLEL_COLS_53 * 0, ADD(LOADU(in_even + 0), dc_0)); |
778 | |
779 | s1_1 = LOADU(in_even + stride + VREG_INT_COUNT); |
780 | /* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */ |
781 | dc_1 = SUB(LOADU(in_odd + VREG_INT_COUNT), |
782 | SAR(ADD3(LOADU(in_even + VREG_INT_COUNT), s1_1, two), 2)); |
783 | STORE(tmp + PARALLEL_COLS_53 * 0 + VREG_INT_COUNT, |
784 | ADD(LOADU(in_even + VREG_INT_COUNT), dc_1)); |
785 | |
786 | for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) { |
787 | |
788 | s2_0 = LOADU(in_even + (j + 1) * stride); |
789 | s2_1 = LOADU(in_even + (j + 1) * stride + VREG_INT_COUNT); |
790 | |
791 | /* dn = in_odd[j * stride] - ((s1 + s2 + 2) >> 2); */ |
792 | dn_0 = SUB(LOADU(in_odd + j * stride), |
793 | SAR(ADD3(s1_0, s2_0, two), 2)); |
794 | dn_1 = SUB(LOADU(in_odd + j * stride + VREG_INT_COUNT), |
795 | SAR(ADD3(s1_1, s2_1, two), 2)); |
796 | |
797 | STORE(tmp + PARALLEL_COLS_53 * i, dc_0); |
798 | STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1); |
799 | |
800 | /* tmp[i + 1] = s1 + ((dn + dc) >> 1); */ |
801 | STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0, |
802 | ADD(s1_0, SAR(ADD(dn_0, dc_0), 1))); |
803 | STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT, |
804 | ADD(s1_1, SAR(ADD(dn_1, dc_1), 1))); |
805 | |
806 | dc_0 = dn_0; |
807 | s1_0 = s2_0; |
808 | dc_1 = dn_1; |
809 | s1_1 = s2_1; |
810 | } |
811 | STORE(tmp + PARALLEL_COLS_53 * i, dc_0); |
812 | STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1); |
813 | |
814 | if (!(len & 1)) { |
815 | /*dn = in_odd[(len / 2 - 1) * stride] - ((s1 + 1) >> 1); */ |
816 | dn_0 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride), |
817 | SAR(ADD3(s1_0, s1_0, two), 2)); |
818 | dn_1 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride + VREG_INT_COUNT), |
819 | SAR(ADD3(s1_1, s1_1, two), 2)); |
820 | |
821 | /* tmp[len - 2] = s1 + ((dn + dc) >> 1); */ |
822 | STORE(tmp + PARALLEL_COLS_53 * (len - 2) + 0, |
823 | ADD(s1_0, SAR(ADD(dn_0, dc_0), 1))); |
824 | STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT, |
825 | ADD(s1_1, SAR(ADD(dn_1, dc_1), 1))); |
826 | |
827 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, dn_0); |
828 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, dn_1); |
829 | } else { |
830 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, ADD(s1_0, dc_0)); |
831 | STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, |
832 | ADD(s1_1, dc_1)); |
833 | } |
834 | |
835 | opj_idwt53_v_final_memcpy(tiledp_col, tmp, len, stride); |
836 | } |
837 | |
838 | #undef VREG |
839 | #undef LOAD_CST |
840 | #undef LOADU |
841 | #undef LOAD |
842 | #undef STORE |
843 | #undef STOREU |
844 | #undef ADD |
845 | #undef ADD3 |
846 | #undef SUB |
847 | #undef SAR |
848 | |
849 | #endif /* (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION) */ |
850 | |
851 | #if !defined(STANDARD_SLOW_VERSION) |
852 | /** Vertical inverse 5x3 wavelet transform for one column, when top-most |
853 | * pixel is on even coordinate */ |
854 | static void opj_idwt3_v_cas0(OPJ_INT32* tmp, |
855 | const OPJ_INT32 sn, |
856 | const OPJ_INT32 len, |
857 | OPJ_INT32* tiledp_col, |
858 | const OPJ_SIZE_T stride) |
859 | { |
860 | OPJ_INT32 i, j; |
861 | OPJ_INT32 d1c, d1n, s1n, s0c, s0n; |
862 | |
863 | assert(len > 1); |
864 | |
865 | /* Performs lifting in one single iteration. Saves memory */ |
866 | /* accesses and explicit interleaving. */ |
867 | |
868 | s1n = tiledp_col[0]; |
869 | d1n = tiledp_col[(OPJ_SIZE_T)sn * stride]; |
870 | s0n = s1n - ((d1n + 1) >> 1); |
871 | |
872 | for (i = 0, j = 0; i < (len - 3); i += 2, j++) { |
873 | d1c = d1n; |
874 | s0c = s0n; |
875 | |
876 | s1n = tiledp_col[(OPJ_SIZE_T)(j + 1) * stride]; |
877 | d1n = tiledp_col[(OPJ_SIZE_T)(sn + j + 1) * stride]; |
878 | |
879 | s0n = s1n - ((d1c + d1n + 2) >> 2); |
880 | |
881 | tmp[i ] = s0c; |
882 | tmp[i + 1] = d1c + ((s0c + s0n) >> 1); |
883 | } |
884 | |
885 | tmp[i] = s0n; |
886 | |
887 | if (len & 1) { |
888 | tmp[len - 1] = |
889 | tiledp_col[(OPJ_SIZE_T)((len - 1) / 2) * stride] - |
890 | ((d1n + 1) >> 1); |
891 | tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1); |
892 | } else { |
893 | tmp[len - 1] = d1n + s0n; |
894 | } |
895 | |
896 | for (i = 0; i < len; ++i) { |
897 | tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i]; |
898 | } |
899 | } |
900 | |
901 | |
902 | /** Vertical inverse 5x3 wavelet transform for one column, when top-most |
903 | * pixel is on odd coordinate */ |
904 | static void opj_idwt3_v_cas1(OPJ_INT32* tmp, |
905 | const OPJ_INT32 sn, |
906 | const OPJ_INT32 len, |
907 | OPJ_INT32* tiledp_col, |
908 | const OPJ_SIZE_T stride) |
909 | { |
910 | OPJ_INT32 i, j; |
911 | OPJ_INT32 s1, s2, dc, dn; |
912 | const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride]; |
913 | const OPJ_INT32* in_odd = &tiledp_col[0]; |
914 | |
915 | assert(len > 2); |
916 | |
917 | /* Performs lifting in one single iteration. Saves memory */ |
918 | /* accesses and explicit interleaving. */ |
919 | |
920 | s1 = in_even[stride]; |
921 | dc = in_odd[0] - ((in_even[0] + s1 + 2) >> 2); |
922 | tmp[0] = in_even[0] + dc; |
923 | for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) { |
924 | |
925 | s2 = in_even[(OPJ_SIZE_T)(j + 1) * stride]; |
926 | |
927 | dn = in_odd[(OPJ_SIZE_T)j * stride] - ((s1 + s2 + 2) >> 2); |
928 | tmp[i ] = dc; |
929 | tmp[i + 1] = s1 + ((dn + dc) >> 1); |
930 | |
931 | dc = dn; |
932 | s1 = s2; |
933 | } |
934 | tmp[i] = dc; |
935 | if (!(len & 1)) { |
936 | dn = in_odd[(OPJ_SIZE_T)(len / 2 - 1) * stride] - ((s1 + 1) >> 1); |
937 | tmp[len - 2] = s1 + ((dn + dc) >> 1); |
938 | tmp[len - 1] = dn; |
939 | } else { |
940 | tmp[len - 1] = s1 + dc; |
941 | } |
942 | |
943 | for (i = 0; i < len; ++i) { |
944 | tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i]; |
945 | } |
946 | } |
947 | #endif /* !defined(STANDARD_SLOW_VERSION) */ |
948 | |
949 | /* <summary> */ |
950 | /* Inverse vertical 5-3 wavelet transform in 1-D for several columns. */ |
951 | /* </summary> */ |
952 | /* Performs interleave, inverse wavelet transform and copy back to buffer */ |
953 | static void opj_idwt53_v(const opj_dwt_t *dwt, |
954 | OPJ_INT32* tiledp_col, |
955 | OPJ_SIZE_T stride, |
956 | OPJ_INT32 nb_cols) |
957 | { |
958 | #ifdef STANDARD_SLOW_VERSION |
959 | /* For documentation purpose */ |
960 | OPJ_INT32 k, c; |
961 | for (c = 0; c < nb_cols; c ++) { |
962 | opj_dwt_interleave_v(dwt, tiledp_col + c, stride); |
963 | opj_dwt_decode_1(dwt); |
964 | for (k = 0; k < dwt->sn + dwt->dn; ++k) { |
965 | tiledp_col[c + k * stride] = dwt->mem[k]; |
966 | } |
967 | } |
968 | #else |
969 | const OPJ_INT32 sn = dwt->sn; |
970 | const OPJ_INT32 len = sn + dwt->dn; |
971 | if (dwt->cas == 0) { |
972 | /* If len == 1, unmodified value */ |
973 | |
974 | #if (defined(__SSE2__) || defined(__AVX2__)) |
975 | if (len > 1 && nb_cols == PARALLEL_COLS_53) { |
976 | /* Same as below general case, except that thanks to SSE2/AVX2 */ |
977 | /* we can efficently process 8/16 columns in parallel */ |
978 | opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2(dwt->mem, sn, len, tiledp_col, stride); |
979 | return; |
980 | } |
981 | #endif |
982 | if (len > 1) { |
983 | OPJ_INT32 c; |
984 | for (c = 0; c < nb_cols; c++, tiledp_col++) { |
985 | opj_idwt3_v_cas0(dwt->mem, sn, len, tiledp_col, stride); |
986 | } |
987 | return; |
988 | } |
989 | } else { |
990 | if (len == 1) { |
991 | OPJ_INT32 c; |
992 | for (c = 0; c < nb_cols; c++, tiledp_col++) { |
993 | tiledp_col[0] /= 2; |
994 | } |
995 | return; |
996 | } |
997 | |
998 | if (len == 2) { |
999 | OPJ_INT32 c; |
1000 | OPJ_INT32* out = dwt->mem; |
1001 | for (c = 0; c < nb_cols; c++, tiledp_col++) { |
1002 | OPJ_INT32 i; |
1003 | const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride]; |
1004 | const OPJ_INT32* in_odd = &tiledp_col[0]; |
1005 | |
1006 | out[1] = in_odd[0] - ((in_even[0] + 1) >> 1); |
1007 | out[0] = in_even[0] + out[1]; |
1008 | |
1009 | for (i = 0; i < len; ++i) { |
1010 | tiledp_col[(OPJ_SIZE_T)i * stride] = out[i]; |
1011 | } |
1012 | } |
1013 | |
1014 | return; |
1015 | } |
1016 | |
1017 | #if (defined(__SSE2__) || defined(__AVX2__)) |
1018 | if (len > 2 && nb_cols == PARALLEL_COLS_53) { |
1019 | /* Same as below general case, except that thanks to SSE2/AVX2 */ |
1020 | /* we can efficently process 8/16 columns in parallel */ |
1021 | opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2(dwt->mem, sn, len, tiledp_col, stride); |
1022 | return; |
1023 | } |
1024 | #endif |
1025 | if (len > 2) { |
1026 | OPJ_INT32 c; |
1027 | for (c = 0; c < nb_cols; c++, tiledp_col++) { |
1028 | opj_idwt3_v_cas1(dwt->mem, sn, len, tiledp_col, stride); |
1029 | } |
1030 | return; |
1031 | } |
1032 | } |
1033 | #endif |
1034 | } |
1035 | |
1036 | |
1037 | /* <summary> */ |
1038 | /* Forward 9-7 wavelet transform in 1-D. */ |
1039 | /* </summary> */ |
1040 | static void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, |
1041 | OPJ_INT32 cas) |
1042 | { |
1043 | OPJ_INT32 i; |
1044 | if (!cas) { |
1045 | if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ |
1046 | for (i = 0; i < dn; i++) { |
1047 | OPJ_D(i) -= opj_int_fix_mul(OPJ_S_(i) + OPJ_S_(i + 1), 12993); |
1048 | } |
1049 | for (i = 0; i < sn; i++) { |
1050 | OPJ_S(i) -= opj_int_fix_mul(OPJ_D_(i - 1) + OPJ_D_(i), 434); |
1051 | } |
1052 | for (i = 0; i < dn; i++) { |
1053 | OPJ_D(i) += opj_int_fix_mul(OPJ_S_(i) + OPJ_S_(i + 1), 7233); |
1054 | } |
1055 | for (i = 0; i < sn; i++) { |
1056 | OPJ_S(i) += opj_int_fix_mul(OPJ_D_(i - 1) + OPJ_D_(i), 3633); |
1057 | } |
1058 | for (i = 0; i < dn; i++) { |
1059 | OPJ_D(i) = opj_int_fix_mul(OPJ_D(i), 5038); /*5038 */ |
1060 | } |
1061 | for (i = 0; i < sn; i++) { |
1062 | OPJ_S(i) = opj_int_fix_mul(OPJ_S(i), 6659); /*6660 */ |
1063 | } |
1064 | } |
1065 | } else { |
1066 | if ((sn > 0) || (dn > 1)) { /* NEW : CASE ONE ELEMENT */ |
1067 | for (i = 0; i < dn; i++) { |
1068 | OPJ_S(i) -= opj_int_fix_mul(OPJ_DD_(i) + OPJ_DD_(i - 1), 12993); |
1069 | } |
1070 | for (i = 0; i < sn; i++) { |
1071 | OPJ_D(i) -= opj_int_fix_mul(OPJ_SS_(i) + OPJ_SS_(i + 1), 434); |
1072 | } |
1073 | for (i = 0; i < dn; i++) { |
1074 | OPJ_S(i) += opj_int_fix_mul(OPJ_DD_(i) + OPJ_DD_(i - 1), 7233); |
1075 | } |
1076 | for (i = 0; i < sn; i++) { |
1077 | OPJ_D(i) += opj_int_fix_mul(OPJ_SS_(i) + OPJ_SS_(i + 1), 3633); |
1078 | } |
1079 | for (i = 0; i < dn; i++) { |
1080 | OPJ_S(i) = opj_int_fix_mul(OPJ_S(i), 5038); /*5038 */ |
1081 | } |
1082 | for (i = 0; i < sn; i++) { |
1083 | OPJ_D(i) = opj_int_fix_mul(OPJ_D(i), 6659); /*6660 */ |
1084 | } |
1085 | } |
1086 | } |
1087 | } |
1088 | |
1089 | static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps, |
1090 | opj_stepsize_t *bandno_stepsize) |
1091 | { |
1092 | OPJ_INT32 p, n; |
1093 | p = opj_int_floorlog2(stepsize) - 13; |
1094 | n = 11 - opj_int_floorlog2(stepsize); |
1095 | bandno_stepsize->mant = (n < 0 ? stepsize >> -n : stepsize << n) & 0x7ff; |
1096 | bandno_stepsize->expn = numbps - p; |
1097 | } |
1098 | |
1099 | /* |
1100 | ========================================================== |
1101 | DWT interface |
1102 | ========================================================== |
1103 | */ |
1104 | |
1105 | |
1106 | /* <summary> */ |
1107 | /* Forward 5-3 wavelet transform in 2-D. */ |
1108 | /* </summary> */ |
1109 | static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec, |
1110 | void (*p_function)(OPJ_INT32 *, OPJ_INT32, OPJ_INT32, OPJ_INT32)) |
1111 | { |
1112 | OPJ_INT32 i, j, k; |
1113 | OPJ_INT32 *a = 00; |
1114 | OPJ_INT32 *aj = 00; |
1115 | OPJ_INT32 *bj = 00; |
1116 | OPJ_INT32 w, l; |
1117 | |
1118 | OPJ_INT32 rw; /* width of the resolution level computed */ |
1119 | OPJ_INT32 rh; /* height of the resolution level computed */ |
1120 | OPJ_SIZE_T l_data_size; |
1121 | |
1122 | opj_tcd_resolution_t * l_cur_res = 0; |
1123 | opj_tcd_resolution_t * l_last_res = 0; |
1124 | |
1125 | w = tilec->x1 - tilec->x0; |
1126 | l = (OPJ_INT32)tilec->numresolutions - 1; |
1127 | a = tilec->data; |
1128 | |
1129 | l_cur_res = tilec->resolutions + l; |
1130 | l_last_res = l_cur_res - 1; |
1131 | |
1132 | l_data_size = opj_dwt_max_resolution(tilec->resolutions, tilec->numresolutions); |
1133 | /* overflow check */ |
1134 | if (l_data_size > (SIZE_MAX / sizeof(OPJ_INT32))) { |
1135 | /* FIXME event manager error callback */ |
1136 | return OPJ_FALSE; |
1137 | } |
1138 | l_data_size *= sizeof(OPJ_INT32); |
1139 | bj = (OPJ_INT32*)opj_malloc(l_data_size); |
1140 | /* l_data_size is equal to 0 when numresolutions == 1 but bj is not used */ |
1141 | /* in that case, so do not error out */ |
1142 | if (l_data_size != 0 && ! bj) { |
1143 | return OPJ_FALSE; |
1144 | } |
1145 | i = l; |
1146 | |
1147 | while (i--) { |
1148 | OPJ_INT32 rw1; /* width of the resolution level once lower than computed one */ |
1149 | OPJ_INT32 rh1; /* height of the resolution level once lower than computed one */ |
1150 | OPJ_INT32 cas_col; /* 0 = non inversion on horizontal filtering 1 = inversion between low-pass and high-pass filtering */ |
1151 | OPJ_INT32 cas_row; /* 0 = non inversion on vertical filtering 1 = inversion between low-pass and high-pass filtering */ |
1152 | OPJ_INT32 dn, sn; |
1153 | |
1154 | rw = l_cur_res->x1 - l_cur_res->x0; |
1155 | rh = l_cur_res->y1 - l_cur_res->y0; |
1156 | rw1 = l_last_res->x1 - l_last_res->x0; |
1157 | rh1 = l_last_res->y1 - l_last_res->y0; |
1158 | |
1159 | cas_row = l_cur_res->x0 & 1; |
1160 | cas_col = l_cur_res->y0 & 1; |
1161 | |
1162 | sn = rh1; |
1163 | dn = rh - rh1; |
1164 | for (j = 0; j < rw; ++j) { |
1165 | aj = a + j; |
1166 | for (k = 0; k < rh; ++k) { |
1167 | bj[k] = aj[k * w]; |
1168 | } |
1169 | |
1170 | (*p_function)(bj, dn, sn, cas_col); |
1171 | |
1172 | opj_dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col); |
1173 | } |
1174 | |
1175 | sn = rw1; |
1176 | dn = rw - rw1; |
1177 | |
1178 | for (j = 0; j < rh; j++) { |
1179 | aj = a + j * w; |
1180 | for (k = 0; k < rw; k++) { |
1181 | bj[k] = aj[k]; |
1182 | } |
1183 | (*p_function)(bj, dn, sn, cas_row); |
1184 | opj_dwt_deinterleave_h(bj, aj, dn, sn, cas_row); |
1185 | } |
1186 | |
1187 | l_cur_res = l_last_res; |
1188 | |
1189 | --l_last_res; |
1190 | } |
1191 | |
1192 | opj_free(bj); |
1193 | return OPJ_TRUE; |
1194 | } |
1195 | |
1196 | /* Forward 5-3 wavelet transform in 2-D. */ |
1197 | /* </summary> */ |
1198 | OPJ_BOOL opj_dwt_encode(opj_tcd_tilecomp_t * tilec) |
1199 | { |
1200 | return opj_dwt_encode_procedure(tilec, opj_dwt_encode_1); |
1201 | } |
1202 | |
1203 | /* <summary> */ |
1204 | /* Inverse 5-3 wavelet transform in 2-D. */ |
1205 | /* </summary> */ |
1206 | OPJ_BOOL opj_dwt_decode(opj_tcd_t *p_tcd, opj_tcd_tilecomp_t* tilec, |
1207 | OPJ_UINT32 numres) |
1208 | { |
1209 | if (p_tcd->whole_tile_decoding) { |
1210 | return opj_dwt_decode_tile(p_tcd->thread_pool, tilec, numres); |
1211 | } else { |
1212 | return opj_dwt_decode_partial_tile(tilec, numres); |
1213 | } |
1214 | } |
1215 | |
1216 | |
1217 | /* <summary> */ |
1218 | /* Get gain of 5-3 wavelet transform. */ |
1219 | /* </summary> */ |
1220 | OPJ_UINT32 opj_dwt_getgain(OPJ_UINT32 orient) |
1221 | { |
1222 | if (orient == 0) { |
1223 | return 0; |
1224 | } |
1225 | if (orient == 1 || orient == 2) { |
1226 | return 1; |
1227 | } |
1228 | return 2; |
1229 | } |
1230 | |
1231 | /* <summary> */ |
1232 | /* Get norm of 5-3 wavelet. */ |
1233 | /* </summary> */ |
1234 | OPJ_FLOAT64 opj_dwt_getnorm(OPJ_UINT32 level, OPJ_UINT32 orient) |
1235 | { |
1236 | /* FIXME ! This is just a band-aid to avoid a buffer overflow */ |
1237 | /* but the array should really be extended up to 33 resolution levels */ |
1238 | /* See https://github.com/uclouvain/openjpeg/issues/493 */ |
1239 | if (orient == 0 && level >= 10) { |
1240 | level = 9; |
1241 | } else if (orient > 0 && level >= 9) { |
1242 | level = 8; |
1243 | } |
1244 | return opj_dwt_norms[orient][level]; |
1245 | } |
1246 | |
1247 | /* <summary> */ |
1248 | /* Forward 9-7 wavelet transform in 2-D. */ |
1249 | /* </summary> */ |
1250 | OPJ_BOOL opj_dwt_encode_real(opj_tcd_tilecomp_t * tilec) |
1251 | { |
1252 | return opj_dwt_encode_procedure(tilec, opj_dwt_encode_1_real); |
1253 | } |
1254 | |
1255 | /* <summary> */ |
1256 | /* Get gain of 9-7 wavelet transform. */ |
1257 | /* </summary> */ |
1258 | OPJ_UINT32 opj_dwt_getgain_real(OPJ_UINT32 orient) |
1259 | { |
1260 | (void)orient; |
1261 | return 0; |
1262 | } |
1263 | |
1264 | /* <summary> */ |
1265 | /* Get norm of 9-7 wavelet. */ |
1266 | /* </summary> */ |
1267 | OPJ_FLOAT64 opj_dwt_getnorm_real(OPJ_UINT32 level, OPJ_UINT32 orient) |
1268 | { |
1269 | /* FIXME ! This is just a band-aid to avoid a buffer overflow */ |
1270 | /* but the array should really be extended up to 33 resolution levels */ |
1271 | /* See https://github.com/uclouvain/openjpeg/issues/493 */ |
1272 | if (orient == 0 && level >= 10) { |
1273 | level = 9; |
1274 | } else if (orient > 0 && level >= 9) { |
1275 | level = 8; |
1276 | } |
1277 | return opj_dwt_norms_real[orient][level]; |
1278 | } |
1279 | |
1280 | void opj_dwt_calc_explicit_stepsizes(opj_tccp_t * tccp, OPJ_UINT32 prec) |
1281 | { |
1282 | OPJ_UINT32 numbands, bandno; |
1283 | numbands = 3 * tccp->numresolutions - 2; |
1284 | for (bandno = 0; bandno < numbands; bandno++) { |
1285 | OPJ_FLOAT64 stepsize; |
1286 | OPJ_UINT32 resno, level, orient, gain; |
1287 | |
1288 | resno = (bandno == 0) ? 0 : ((bandno - 1) / 3 + 1); |
1289 | orient = (bandno == 0) ? 0 : ((bandno - 1) % 3 + 1); |
1290 | level = tccp->numresolutions - 1 - resno; |
1291 | gain = (tccp->qmfbid == 0) ? 0 : ((orient == 0) ? 0 : (((orient == 1) || |
1292 | (orient == 2)) ? 1 : 2)); |
1293 | if (tccp->qntsty == J2K_CCP_QNTSTY_NOQNT) { |
1294 | stepsize = 1.0; |
1295 | } else { |
1296 | OPJ_FLOAT64 norm = opj_dwt_norms_real[orient][level]; |
1297 | stepsize = (1 << (gain)) / norm; |
1298 | } |
1299 | opj_dwt_encode_stepsize((OPJ_INT32) floor(stepsize * 8192.0), |
1300 | (OPJ_INT32)(prec + gain), &tccp->stepsizes[bandno]); |
1301 | } |
1302 | } |
1303 | |
1304 | /* <summary> */ |
1305 | /* Determine maximum computed resolution level for inverse wavelet transform */ |
1306 | /* </summary> */ |
1307 | static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r, |
1308 | OPJ_UINT32 i) |
1309 | { |
1310 | OPJ_UINT32 mr = 0; |
1311 | OPJ_UINT32 w; |
1312 | while (--i) { |
1313 | ++r; |
1314 | if (mr < (w = (OPJ_UINT32)(r->x1 - r->x0))) { |
1315 | mr = w ; |
1316 | } |
1317 | if (mr < (w = (OPJ_UINT32)(r->y1 - r->y0))) { |
1318 | mr = w ; |
1319 | } |
1320 | } |
1321 | return mr ; |
1322 | } |
1323 | |
1324 | typedef struct { |
1325 | opj_dwt_t h; |
1326 | OPJ_UINT32 rw; |
1327 | OPJ_UINT32 w; |
1328 | OPJ_INT32 * OPJ_RESTRICT tiledp; |
1329 | OPJ_UINT32 min_j; |
1330 | OPJ_UINT32 max_j; |
1331 | } opj_dwd_decode_h_job_t; |
1332 | |
1333 | static void opj_dwt_decode_h_func(void* user_data, opj_tls_t* tls) |
1334 | { |
1335 | OPJ_UINT32 j; |
1336 | opj_dwd_decode_h_job_t* job; |
1337 | (void)tls; |
1338 | |
1339 | job = (opj_dwd_decode_h_job_t*)user_data; |
1340 | for (j = job->min_j; j < job->max_j; j++) { |
1341 | opj_idwt53_h(&job->h, &job->tiledp[j * job->w]); |
1342 | } |
1343 | |
1344 | opj_aligned_free(job->h.mem); |
1345 | opj_free(job); |
1346 | } |
1347 | |
1348 | typedef struct { |
1349 | opj_dwt_t v; |
1350 | OPJ_UINT32 rh; |
1351 | OPJ_UINT32 w; |
1352 | OPJ_INT32 * OPJ_RESTRICT tiledp; |
1353 | OPJ_UINT32 min_j; |
1354 | OPJ_UINT32 max_j; |
1355 | } opj_dwd_decode_v_job_t; |
1356 | |
1357 | static void opj_dwt_decode_v_func(void* user_data, opj_tls_t* tls) |
1358 | { |
1359 | OPJ_UINT32 j; |
1360 | opj_dwd_decode_v_job_t* job; |
1361 | (void)tls; |
1362 | |
1363 | job = (opj_dwd_decode_v_job_t*)user_data; |
1364 | for (j = job->min_j; j + PARALLEL_COLS_53 <= job->max_j; |
1365 | j += PARALLEL_COLS_53) { |
1366 | opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w, |
1367 | PARALLEL_COLS_53); |
1368 | } |
1369 | if (j < job->max_j) |
1370 | opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w, |
1371 | (OPJ_INT32)(job->max_j - j)); |
1372 | |
1373 | opj_aligned_free(job->v.mem); |
1374 | opj_free(job); |
1375 | } |
1376 | |
1377 | |
1378 | /* <summary> */ |
1379 | /* Inverse wavelet transform in 2-D. */ |
1380 | /* </summary> */ |
1381 | static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp, |
1382 | opj_tcd_tilecomp_t* tilec, OPJ_UINT32 numres) |
1383 | { |
1384 | opj_dwt_t h; |
1385 | opj_dwt_t v; |
1386 | |
1387 | opj_tcd_resolution_t* tr = tilec->resolutions; |
1388 | |
1389 | OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 - |
1390 | tr->x0); /* width of the resolution level computed */ |
1391 | OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 - |
1392 | tr->y0); /* height of the resolution level computed */ |
1393 | |
1394 | OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions - |
1395 | 1].x1 - |
1396 | tilec->resolutions[tilec->minimum_num_resolutions - 1].x0); |
1397 | OPJ_SIZE_T h_mem_size; |
1398 | int num_threads; |
1399 | |
1400 | if (numres == 1U) { |
1401 | return OPJ_TRUE; |
1402 | } |
1403 | num_threads = opj_thread_pool_get_thread_count(tp); |
1404 | h_mem_size = opj_dwt_max_resolution(tr, numres); |
1405 | /* overflow check */ |
1406 | if (h_mem_size > (SIZE_MAX / PARALLEL_COLS_53 / sizeof(OPJ_INT32))) { |
1407 | /* FIXME event manager error callback */ |
1408 | return OPJ_FALSE; |
1409 | } |
1410 | /* We need PARALLEL_COLS_53 times the height of the array, */ |
1411 | /* since for the vertical pass */ |
1412 | /* we process PARALLEL_COLS_53 columns at a time */ |
1413 | h_mem_size *= PARALLEL_COLS_53 * sizeof(OPJ_INT32); |
1414 | h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size); |
1415 | if (! h.mem) { |
1416 | /* FIXME event manager error callback */ |
1417 | return OPJ_FALSE; |
1418 | } |
1419 | |
1420 | v.mem = h.mem; |
1421 | |
1422 | while (--numres) { |
1423 | OPJ_INT32 * OPJ_RESTRICT tiledp = tilec->data; |
1424 | OPJ_UINT32 j; |
1425 | |
1426 | ++tr; |
1427 | h.sn = (OPJ_INT32)rw; |
1428 | v.sn = (OPJ_INT32)rh; |
1429 | |
1430 | rw = (OPJ_UINT32)(tr->x1 - tr->x0); |
1431 | rh = (OPJ_UINT32)(tr->y1 - tr->y0); |
1432 | |
1433 | h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn); |
1434 | h.cas = tr->x0 % 2; |
1435 | |
1436 | if (num_threads <= 1 || rh <= 1) { |
1437 | for (j = 0; j < rh; ++j) { |
1438 | opj_idwt53_h(&h, &tiledp[(OPJ_SIZE_T)j * w]); |
1439 | } |
1440 | } else { |
1441 | OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads; |
1442 | OPJ_UINT32 step_j; |
1443 | |
1444 | if (rh < num_jobs) { |
1445 | num_jobs = rh; |
1446 | } |
1447 | step_j = (rh / num_jobs); |
1448 | |
1449 | for (j = 0; j < num_jobs; j++) { |
1450 | opj_dwd_decode_h_job_t* job; |
1451 | |
1452 | job = (opj_dwd_decode_h_job_t*) opj_malloc(sizeof(opj_dwd_decode_h_job_t)); |
1453 | if (!job) { |
1454 | /* It would be nice to fallback to single thread case, but */ |
1455 | /* unfortunately some jobs may be launched and have modified */ |
1456 | /* tiledp, so it is not practical to recover from that error */ |
1457 | /* FIXME event manager error callback */ |
1458 | opj_thread_pool_wait_completion(tp, 0); |
1459 | opj_aligned_free(h.mem); |
1460 | return OPJ_FALSE; |
1461 | } |
1462 | job->h = h; |
1463 | job->rw = rw; |
1464 | job->w = w; |
1465 | job->tiledp = tiledp; |
1466 | job->min_j = j * step_j; |
1467 | job->max_j = (j + 1U) * step_j; /* this can overflow */ |
1468 | if (j == (num_jobs - 1U)) { /* this will take care of the overflow */ |
1469 | job->max_j = rh; |
1470 | } |
1471 | job->h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size); |
1472 | if (!job->h.mem) { |
1473 | /* FIXME event manager error callback */ |
1474 | opj_thread_pool_wait_completion(tp, 0); |
1475 | opj_free(job); |
1476 | opj_aligned_free(h.mem); |
1477 | return OPJ_FALSE; |
1478 | } |
1479 | opj_thread_pool_submit_job(tp, opj_dwt_decode_h_func, job); |
1480 | } |
1481 | opj_thread_pool_wait_completion(tp, 0); |
1482 | } |
1483 | |
1484 | v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn); |
1485 | v.cas = tr->y0 % 2; |
1486 | |
1487 | if (num_threads <= 1 || rw <= 1) { |
1488 | for (j = 0; j + PARALLEL_COLS_53 <= rw; |
1489 | j += PARALLEL_COLS_53) { |
1490 | opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, PARALLEL_COLS_53); |
1491 | } |
1492 | if (j < rw) { |
1493 | opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, (OPJ_INT32)(rw - j)); |
1494 | } |
1495 | } else { |
1496 | OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads; |
1497 | OPJ_UINT32 step_j; |
1498 | |
1499 | if (rw < num_jobs) { |
1500 | num_jobs = rw; |
1501 | } |
1502 | step_j = (rw / num_jobs); |
1503 | |
1504 | for (j = 0; j < num_jobs; j++) { |
1505 | opj_dwd_decode_v_job_t* job; |
1506 | |
1507 | job = (opj_dwd_decode_v_job_t*) opj_malloc(sizeof(opj_dwd_decode_v_job_t)); |
1508 | if (!job) { |
1509 | /* It would be nice to fallback to single thread case, but */ |
1510 | /* unfortunately some jobs may be launched and have modified */ |
1511 | /* tiledp, so it is not practical to recover from that error */ |
1512 | /* FIXME event manager error callback */ |
1513 | opj_thread_pool_wait_completion(tp, 0); |
1514 | opj_aligned_free(v.mem); |
1515 | return OPJ_FALSE; |
1516 | } |
1517 | job->v = v; |
1518 | job->rh = rh; |
1519 | job->w = w; |
1520 | job->tiledp = tiledp; |
1521 | job->min_j = j * step_j; |
1522 | job->max_j = (j + 1U) * step_j; /* this can overflow */ |
1523 | if (j == (num_jobs - 1U)) { /* this will take care of the overflow */ |
1524 | job->max_j = rw; |
1525 | } |
1526 | job->v.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size); |
1527 | if (!job->v.mem) { |
1528 | /* FIXME event manager error callback */ |
1529 | opj_thread_pool_wait_completion(tp, 0); |
1530 | opj_free(job); |
1531 | opj_aligned_free(v.mem); |
1532 | return OPJ_FALSE; |
1533 | } |
1534 | opj_thread_pool_submit_job(tp, opj_dwt_decode_v_func, job); |
1535 | } |
1536 | opj_thread_pool_wait_completion(tp, 0); |
1537 | } |
1538 | } |
1539 | opj_aligned_free(h.mem); |
1540 | return OPJ_TRUE; |
1541 | } |
1542 | |
1543 | static void opj_dwt_interleave_partial_h(OPJ_INT32 *dest, |
1544 | OPJ_INT32 cas, |
1545 | opj_sparse_array_int32_t* sa, |
1546 | OPJ_UINT32 sa_line, |
1547 | OPJ_UINT32 sn, |
1548 | OPJ_UINT32 win_l_x0, |
1549 | OPJ_UINT32 win_l_x1, |
1550 | OPJ_UINT32 win_h_x0, |
1551 | OPJ_UINT32 win_h_x1) |
1552 | { |
1553 | OPJ_BOOL ret; |
1554 | ret = opj_sparse_array_int32_read(sa, |
1555 | win_l_x0, sa_line, |
1556 | win_l_x1, sa_line + 1, |
1557 | dest + cas + 2 * win_l_x0, |
1558 | 2, 0, OPJ_TRUE); |
1559 | assert(ret); |
1560 | ret = opj_sparse_array_int32_read(sa, |
1561 | sn + win_h_x0, sa_line, |
1562 | sn + win_h_x1, sa_line + 1, |
1563 | dest + 1 - cas + 2 * win_h_x0, |
1564 | 2, 0, OPJ_TRUE); |
1565 | assert(ret); |
1566 | OPJ_UNUSED(ret); |
1567 | } |
1568 | |
1569 | |
1570 | static void opj_dwt_interleave_partial_v(OPJ_INT32 *dest, |
1571 | OPJ_INT32 cas, |
1572 | opj_sparse_array_int32_t* sa, |
1573 | OPJ_UINT32 sa_col, |
1574 | OPJ_UINT32 nb_cols, |
1575 | OPJ_UINT32 sn, |
1576 | OPJ_UINT32 win_l_y0, |
1577 | OPJ_UINT32 win_l_y1, |
1578 | OPJ_UINT32 win_h_y0, |
1579 | OPJ_UINT32 win_h_y1) |
1580 | { |
1581 | OPJ_BOOL ret; |
1582 | ret = opj_sparse_array_int32_read(sa, |
1583 | sa_col, win_l_y0, |
1584 | sa_col + nb_cols, win_l_y1, |
1585 | dest + cas * 4 + 2 * 4 * win_l_y0, |
1586 | 1, 2 * 4, OPJ_TRUE); |
1587 | assert(ret); |
1588 | ret = opj_sparse_array_int32_read(sa, |
1589 | sa_col, sn + win_h_y0, |
1590 | sa_col + nb_cols, sn + win_h_y1, |
1591 | dest + (1 - cas) * 4 + 2 * 4 * win_h_y0, |
1592 | 1, 2 * 4, OPJ_TRUE); |
1593 | assert(ret); |
1594 | OPJ_UNUSED(ret); |
1595 | } |
1596 | |
1597 | static void opj_dwt_decode_partial_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn, |
1598 | OPJ_INT32 cas, |
1599 | OPJ_INT32 win_l_x0, |
1600 | OPJ_INT32 win_l_x1, |
1601 | OPJ_INT32 win_h_x0, |
1602 | OPJ_INT32 win_h_x1) |
1603 | { |
1604 | OPJ_INT32 i; |
1605 | |
1606 | if (!cas) { |
1607 | if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ |
1608 | |
1609 | /* Naive version is : |
1610 | for (i = win_l_x0; i < i_max; i++) { |
1611 | OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2; |
1612 | } |
1613 | for (i = win_h_x0; i < win_h_x1; i++) { |
1614 | OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1; |
1615 | } |
1616 | but the compiler doesn't manage to unroll it to avoid bound |
1617 | checking in OPJ_S_ and OPJ_D_ macros |
1618 | */ |
1619 | |
1620 | i = win_l_x0; |
1621 | if (i < win_l_x1) { |
1622 | OPJ_INT32 i_max; |
1623 | |
1624 | /* Left-most case */ |
1625 | OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2; |
1626 | i ++; |
1627 | |
1628 | i_max = win_l_x1; |
1629 | if (i_max > dn) { |
1630 | i_max = dn; |
1631 | } |
1632 | for (; i < i_max; i++) { |
1633 | /* No bound checking */ |
1634 | OPJ_S(i) -= (OPJ_D(i - 1) + OPJ_D(i) + 2) >> 2; |
1635 | } |
1636 | for (; i < win_l_x1; i++) { |
1637 | /* Right-most case */ |
1638 | OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2; |
1639 | } |
1640 | } |
1641 | |
1642 | i = win_h_x0; |
1643 | if (i < win_h_x1) { |
1644 | OPJ_INT32 i_max = win_h_x1; |
1645 | if (i_max >= sn) { |
1646 | i_max = sn - 1; |
1647 | } |
1648 | for (; i < i_max; i++) { |
1649 | /* No bound checking */ |
1650 | OPJ_D(i) += (OPJ_S(i) + OPJ_S(i + 1)) >> 1; |
1651 | } |
1652 | for (; i < win_h_x1; i++) { |
1653 | /* Right-most case */ |
1654 | OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1; |
1655 | } |
1656 | } |
1657 | } |
1658 | } else { |
1659 | if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */ |
1660 | OPJ_S(0) /= 2; |
1661 | } else { |
1662 | for (i = win_l_x0; i < win_l_x1; i++) { |
1663 | OPJ_D(i) -= (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2; |
1664 | } |
1665 | for (i = win_h_x0; i < win_h_x1; i++) { |
1666 | OPJ_S(i) += (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1; |
1667 | } |
1668 | } |
1669 | } |
1670 | } |
1671 | |
1672 | #define OPJ_S_off(i,off) a[(OPJ_UINT32)(i)*2*4+off] |
1673 | #define OPJ_D_off(i,off) a[(1+(OPJ_UINT32)(i)*2)*4+off] |
1674 | #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))) |
1675 | #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))) |
1676 | #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))) |
1677 | #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))) |
1678 | |
1679 | static void opj_dwt_decode_partial_1_parallel(OPJ_INT32 *a, |
1680 | OPJ_UINT32 nb_cols, |
1681 | OPJ_INT32 dn, OPJ_INT32 sn, |
1682 | OPJ_INT32 cas, |
1683 | OPJ_INT32 win_l_x0, |
1684 | OPJ_INT32 win_l_x1, |
1685 | OPJ_INT32 win_h_x0, |
1686 | OPJ_INT32 win_h_x1) |
1687 | { |
1688 | OPJ_INT32 i; |
1689 | OPJ_UINT32 off; |
1690 | |
1691 | (void)nb_cols; |
1692 | |
1693 | if (!cas) { |
1694 | if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */ |
1695 | |
1696 | /* Naive version is : |
1697 | for (i = win_l_x0; i < i_max; i++) { |
1698 | OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2; |
1699 | } |
1700 | for (i = win_h_x0; i < win_h_x1; i++) { |
1701 | OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1; |
1702 | } |
1703 | but the compiler doesn't manage to unroll it to avoid bound |
1704 | checking in OPJ_S_ and OPJ_D_ macros |
1705 | */ |
1706 | |
1707 | i = win_l_x0; |
1708 | if (i < win_l_x1) { |
1709 | OPJ_INT32 i_max; |
1710 | |
1711 | /* Left-most case */ |
1712 | for (off = 0; off < 4; off++) { |
1713 | OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2; |
1714 | } |
1715 | i ++; |
1716 | |
1717 | i_max = win_l_x1; |
1718 | if (i_max > dn) { |
1719 | i_max = dn; |
1720 | } |
1721 | |
1722 | #ifdef __SSE2__ |
1723 | if (i + 1 < i_max) { |
1724 | const __m128i two = _mm_set1_epi32(2); |
1725 | __m128i Dm1 = _mm_load_si128((__m128i * const)(a + 4 + (i - 1) * 8)); |
1726 | for (; i + 1 < i_max; i += 2) { |
1727 | /* No bound checking */ |
1728 | __m128i S = _mm_load_si128((__m128i * const)(a + i * 8)); |
1729 | __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8)); |
1730 | __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8)); |
1731 | __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8)); |
1732 | S = _mm_sub_epi32(S, |
1733 | _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(Dm1, D), two), 2)); |
1734 | S1 = _mm_sub_epi32(S1, |
1735 | _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(D, D1), two), 2)); |
1736 | _mm_store_si128((__m128i*)(a + i * 8), S); |
1737 | _mm_store_si128((__m128i*)(a + (i + 1) * 8), S1); |
1738 | Dm1 = D1; |
1739 | } |
1740 | } |
1741 | #endif |
1742 | |
1743 | for (; i < i_max; i++) { |
1744 | /* No bound checking */ |
1745 | for (off = 0; off < 4; off++) { |
1746 | OPJ_S_off(i, off) -= (OPJ_D_off(i - 1, off) + OPJ_D_off(i, off) + 2) >> 2; |
1747 | } |
1748 | } |
1749 | for (; i < win_l_x1; i++) { |
1750 | /* Right-most case */ |
1751 | for (off = 0; off < 4; off++) { |
1752 | OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2; |
1753 | } |
1754 | } |
1755 | } |
1756 | |
1757 | i = win_h_x0; |
1758 | if (i < win_h_x1) { |
1759 | OPJ_INT32 i_max = win_h_x1; |
1760 | if (i_max >= sn) { |
1761 | i_max = sn - 1; |
1762 | } |
1763 | |
1764 | #ifdef __SSE2__ |
1765 | if (i + 1 < i_max) { |
1766 | __m128i S = _mm_load_si128((__m128i * const)(a + i * 8)); |
1767 | for (; i + 1 < i_max; i += 2) { |
1768 | /* No bound checking */ |
1769 | __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8)); |
1770 | __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8)); |
1771 | __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8)); |
1772 | __m128i S2 = _mm_load_si128((__m128i * const)(a + (i + 2) * 8)); |
1773 | D = _mm_add_epi32(D, _mm_srai_epi32(_mm_add_epi32(S, S1), 1)); |
1774 | D1 = _mm_add_epi32(D1, _mm_srai_epi32(_mm_add_epi32(S1, S2), 1)); |
1775 | _mm_store_si128((__m128i*)(a + 4 + i * 8), D); |
1776 | _mm_store_si128((__m128i*)(a + 4 + (i + 1) * 8), D1); |
1777 | S = S2; |
1778 | } |
1779 | } |
1780 | #endif |
1781 | |
1782 | for (; i < i_max; i++) { |
1783 | /* No bound checking */ |
1784 | for (off = 0; off < 4; off++) { |
1785 | OPJ_D_off(i, off) += (OPJ_S_off(i, off) + OPJ_S_off(i + 1, off)) >> 1; |
1786 | } |
1787 | } |
1788 | for (; i < win_h_x1; i++) { |
1789 | /* Right-most case */ |
1790 | for (off = 0; off < 4; off++) { |
1791 | OPJ_D_off(i, off) += (OPJ_S__off(i, off) + OPJ_S__off(i + 1, off)) >> 1; |
1792 | } |
1793 | } |
1794 | } |
1795 | } |
1796 | } else { |
1797 | if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */ |
1798 | for (off = 0; off < 4; off++) { |
1799 | OPJ_S_off(0, off) /= 2; |
1800 | } |
1801 | } else { |
1802 | for (i = win_l_x0; i < win_l_x1; i++) { |
1803 | for (off = 0; off < 4; off++) { |
1804 | OPJ_D_off(i, off) -= (OPJ_SS__off(i, off) + OPJ_SS__off(i + 1, off) + 2) >> 2; |
1805 | } |
1806 | } |
1807 | for (i = win_h_x0; i < win_h_x1; i++) { |
1808 | for (off = 0; off < 4; off++) { |
1809 | OPJ_S_off(i, off) += (OPJ_DD__off(i, off) + OPJ_DD__off(i - 1, off)) >> 1; |
1810 | } |
1811 | } |
1812 | } |
1813 | } |
1814 | } |
1815 | |
1816 | static void opj_dwt_get_band_coordinates(opj_tcd_tilecomp_t* tilec, |
1817 | OPJ_UINT32 resno, |
1818 | OPJ_UINT32 bandno, |
1819 | OPJ_UINT32 tcx0, |
1820 | OPJ_UINT32 tcy0, |
1821 | OPJ_UINT32 tcx1, |
1822 | OPJ_UINT32 tcy1, |
1823 | OPJ_UINT32* tbx0, |
1824 | OPJ_UINT32* tby0, |
1825 | OPJ_UINT32* tbx1, |
1826 | OPJ_UINT32* tby1) |
1827 | { |
1828 | /* Compute number of decomposition for this band. See table F-1 */ |
1829 | OPJ_UINT32 nb = (resno == 0) ? |
1830 | tilec->numresolutions - 1 : |
1831 | tilec->numresolutions - resno; |
1832 | /* Map above tile-based coordinates to sub-band-based coordinates per */ |
1833 | /* equation B-15 of the standard */ |
1834 | OPJ_UINT32 x0b = bandno & 1; |
1835 | OPJ_UINT32 y0b = bandno >> 1; |
1836 | if (tbx0) { |
1837 | *tbx0 = (nb == 0) ? tcx0 : |
1838 | (tcx0 <= (1U << (nb - 1)) * x0b) ? 0 : |
1839 | opj_uint_ceildivpow2(tcx0 - (1U << (nb - 1)) * x0b, nb); |
1840 | } |
1841 | if (tby0) { |
1842 | *tby0 = (nb == 0) ? tcy0 : |
1843 | (tcy0 <= (1U << (nb - 1)) * y0b) ? 0 : |
1844 | opj_uint_ceildivpow2(tcy0 - (1U << (nb - 1)) * y0b, nb); |
1845 | } |
1846 | if (tbx1) { |
1847 | *tbx1 = (nb == 0) ? tcx1 : |
1848 | (tcx1 <= (1U << (nb - 1)) * x0b) ? 0 : |
1849 | opj_uint_ceildivpow2(tcx1 - (1U << (nb - 1)) * x0b, nb); |
1850 | } |
1851 | if (tby1) { |
1852 | *tby1 = (nb == 0) ? tcy1 : |
1853 | (tcy1 <= (1U << (nb - 1)) * y0b) ? 0 : |
1854 | opj_uint_ceildivpow2(tcy1 - (1U << (nb - 1)) * y0b, nb); |
1855 | } |
1856 | } |
1857 | |
1858 | static void opj_dwt_segment_grow(OPJ_UINT32 filter_width, |
1859 | OPJ_UINT32 max_size, |
1860 | OPJ_UINT32* start, |
1861 | OPJ_UINT32* end) |
1862 | { |
1863 | *start = opj_uint_subs(*start, filter_width); |
1864 | *end = opj_uint_adds(*end, filter_width); |
1865 | *end = opj_uint_min(*end, max_size); |
1866 | } |
1867 | |
1868 | |
1869 | static opj_sparse_array_int32_t* opj_dwt_init_sparse_array( |
1870 | opj_tcd_tilecomp_t* tilec, |
1871 | OPJ_UINT32 numres) |
1872 | { |
1873 | opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]); |
1874 | OPJ_UINT32 w = (OPJ_UINT32)(tr_max->x1 - tr_max->x0); |
1875 | OPJ_UINT32 h = (OPJ_UINT32)(tr_max->y1 - tr_max->y0); |
1876 | OPJ_UINT32 resno, bandno, precno, cblkno; |
1877 | opj_sparse_array_int32_t* sa = opj_sparse_array_int32_create( |
1878 | w, h, opj_uint_min(w, 64), opj_uint_min(h, 64)); |
1879 | if (sa == NULL) { |
1880 | return NULL; |
1881 | } |
1882 | |
1883 | for (resno = 0; resno < numres; ++resno) { |
1884 | opj_tcd_resolution_t* res = &tilec->resolutions[resno]; |
1885 | |
1886 | for (bandno = 0; bandno < res->numbands; ++bandno) { |
1887 | opj_tcd_band_t* band = &res->bands[bandno]; |
1888 | |
1889 | for (precno = 0; precno < res->pw * res->ph; ++precno) { |
1890 | opj_tcd_precinct_t* precinct = &band->precincts[precno]; |
1891 | for (cblkno = 0; cblkno < precinct->cw * precinct->ch; ++cblkno) { |
1892 | opj_tcd_cblk_dec_t* cblk = &precinct->cblks.dec[cblkno]; |
1893 | if (cblk->decoded_data != NULL) { |
1894 | OPJ_UINT32 x = (OPJ_UINT32)(cblk->x0 - band->x0); |
1895 | OPJ_UINT32 y = (OPJ_UINT32)(cblk->y0 - band->y0); |
1896 | OPJ_UINT32 cblk_w = (OPJ_UINT32)(cblk->x1 - cblk->x0); |
1897 | OPJ_UINT32 cblk_h = (OPJ_UINT32)(cblk->y1 - cblk->y0); |
1898 | |
1899 | if (band->bandno & 1) { |
1900 | opj_tcd_resolution_t* pres = &tilec->resolutions[resno - 1]; |
1901 | x += (OPJ_UINT32)(pres->x1 - pres->x0); |
1902 | } |
1903 | if (band->bandno & 2) { |
1904 | opj_tcd_resolution_t* pres = &tilec->resolutions[resno - 1]; |
1905 | y += (OPJ_UINT32)(pres->y1 - pres->y0); |
1906 | } |
1907 | |
1908 | if (!opj_sparse_array_int32_write(sa, x, y, |
1909 | x + cblk_w, y + cblk_h, |
1910 | cblk->decoded_data, |
1911 | 1, cblk_w, OPJ_TRUE)) { |
1912 | opj_sparse_array_int32_free(sa); |
1913 | return NULL; |
1914 | } |
1915 | } |
1916 | } |
1917 | } |
1918 | } |
1919 | } |
1920 | |
1921 | return sa; |
1922 | } |
1923 | |
1924 | |
1925 | static OPJ_BOOL opj_dwt_decode_partial_tile( |
1926 | opj_tcd_tilecomp_t* tilec, |
1927 | OPJ_UINT32 numres) |
1928 | { |
1929 | opj_sparse_array_int32_t* sa; |
1930 | opj_dwt_t h; |
1931 | opj_dwt_t v; |
1932 | OPJ_UINT32 resno; |
1933 | /* This value matches the maximum left/right extension given in tables */ |
1934 | /* F.2 and F.3 of the standard. */ |
1935 | const OPJ_UINT32 filter_width = 2U; |
1936 | |
1937 | opj_tcd_resolution_t* tr = tilec->resolutions; |
1938 | opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]); |
1939 | |
1940 | OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 - |
1941 | tr->x0); /* width of the resolution level computed */ |
1942 | OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 - |
1943 | tr->y0); /* height of the resolution level computed */ |
1944 | |
1945 | OPJ_SIZE_T h_mem_size; |
1946 | |
1947 | /* Compute the intersection of the area of interest, expressed in tile coordinates */ |
1948 | /* with the tile coordinates */ |
1949 | OPJ_UINT32 win_tcx0 = tilec->win_x0; |
1950 | OPJ_UINT32 win_tcy0 = tilec->win_y0; |
1951 | OPJ_UINT32 win_tcx1 = tilec->win_x1; |
1952 | OPJ_UINT32 win_tcy1 = tilec->win_y1; |
1953 | |
1954 | if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) { |
1955 | return OPJ_TRUE; |
1956 | } |
1957 | |
1958 | sa = opj_dwt_init_sparse_array(tilec, numres); |
1959 | if (sa == NULL) { |
1960 | return OPJ_FALSE; |
1961 | } |
1962 | |
1963 | if (numres == 1U) { |
1964 | OPJ_BOOL ret = opj_sparse_array_int32_read(sa, |
1965 | tr_max->win_x0 - (OPJ_UINT32)tr_max->x0, |
1966 | tr_max->win_y0 - (OPJ_UINT32)tr_max->y0, |
1967 | tr_max->win_x1 - (OPJ_UINT32)tr_max->x0, |
1968 | tr_max->win_y1 - (OPJ_UINT32)tr_max->y0, |
1969 | tilec->data_win, |
1970 | 1, tr_max->win_x1 - tr_max->win_x0, |
1971 | OPJ_TRUE); |
1972 | assert(ret); |
1973 | OPJ_UNUSED(ret); |
1974 | opj_sparse_array_int32_free(sa); |
1975 | return OPJ_TRUE; |
1976 | } |
1977 | h_mem_size = opj_dwt_max_resolution(tr, numres); |
1978 | /* overflow check */ |
1979 | /* in vertical pass, we process 4 columns at a time */ |
1980 | if (h_mem_size > (SIZE_MAX / (4 * sizeof(OPJ_INT32)))) { |
1981 | /* FIXME event manager error callback */ |
1982 | opj_sparse_array_int32_free(sa); |
1983 | return OPJ_FALSE; |
1984 | } |
1985 | |
1986 | h_mem_size *= 4 * sizeof(OPJ_INT32); |
1987 | h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size); |
1988 | if (! h.mem) { |
1989 | /* FIXME event manager error callback */ |
1990 | opj_sparse_array_int32_free(sa); |
1991 | return OPJ_FALSE; |
1992 | } |
1993 | |
1994 | v.mem = h.mem; |
1995 | |
1996 | for (resno = 1; resno < numres; resno ++) { |
1997 | OPJ_UINT32 i, j; |
1998 | /* Window of interest subband-based coordinates */ |
1999 | OPJ_UINT32 win_ll_x0, win_ll_y0, win_ll_x1, win_ll_y1; |
2000 | OPJ_UINT32 win_hl_x0, win_hl_x1; |
2001 | OPJ_UINT32 win_lh_y0, win_lh_y1; |
2002 | /* Window of interest tile-resolution-based coordinates */ |
2003 | OPJ_UINT32 win_tr_x0, win_tr_x1, win_tr_y0, win_tr_y1; |
2004 | /* Tile-resolution subband-based coordinates */ |
2005 | OPJ_UINT32 tr_ll_x0, tr_ll_y0, tr_hl_x0, tr_lh_y0; |
2006 | |
2007 | ++tr; |
2008 | |
2009 | h.sn = (OPJ_INT32)rw; |
2010 | v.sn = (OPJ_INT32)rh; |
2011 | |
2012 | rw = (OPJ_UINT32)(tr->x1 - tr->x0); |
2013 | rh = (OPJ_UINT32)(tr->y1 - tr->y0); |
2014 | |
2015 | h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn); |
2016 | h.cas = tr->x0 % 2; |
2017 | |
2018 | v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn); |
2019 | v.cas = tr->y0 % 2; |
2020 | |
2021 | /* Get the subband coordinates for the window of interest */ |
2022 | /* LL band */ |
2023 | opj_dwt_get_band_coordinates(tilec, resno, 0, |
2024 | win_tcx0, win_tcy0, win_tcx1, win_tcy1, |
2025 | &win_ll_x0, &win_ll_y0, |
2026 | &win_ll_x1, &win_ll_y1); |
2027 | |
2028 | /* HL band */ |
2029 | opj_dwt_get_band_coordinates(tilec, resno, 1, |
2030 | win_tcx0, win_tcy0, win_tcx1, win_tcy1, |
2031 | &win_hl_x0, NULL, &win_hl_x1, NULL); |
2032 | |
2033 | /* LH band */ |
2034 | opj_dwt_get_band_coordinates(tilec, resno, 2, |
2035 | win_tcx0, win_tcy0, win_tcx1, win_tcy1, |
2036 | NULL, &win_lh_y0, NULL, &win_lh_y1); |
2037 | |
2038 | /* Beware: band index for non-LL0 resolution are 0=HL, 1=LH and 2=HH */ |
2039 | tr_ll_x0 = (OPJ_UINT32)tr->bands[1].x0; |
2040 | tr_ll_y0 = (OPJ_UINT32)tr->bands[0].y0; |
2041 | tr_hl_x0 = (OPJ_UINT32)tr->bands[0].x0; |
2042 | tr_lh_y0 = (OPJ_UINT32)tr->bands[1].y0; |
2043 | |
2044 | /* Substract the origin of the bands for this tile, to the subwindow */ |
2045 | /* of interest band coordinates, so as to get them relative to the */ |
2046 | /* tile */ |
2047 | win_ll_x0 = opj_uint_subs(win_ll_x0, tr_ll_x0); |
2048 | win_ll_y0 = opj_uint_subs(win_ll_y0, tr_ll_y0); |
2049 | win_ll_x1 = opj_uint_subs(win_ll_x1, tr_ll_x0); |
2050 | win_ll_y1 = opj_uint_subs(win_ll_y1, tr_ll_y0); |
2051 | win_hl_x0 = opj_uint_subs(win_hl_x0, tr_hl_x0); |
2052 | win_hl_x1 = opj_uint_subs(win_hl_x1, tr_hl_x0); |
2053 | win_lh_y0 = opj_uint_subs(win_lh_y0, tr_lh_y0); |
2054 | win_lh_y1 = opj_uint_subs(win_lh_y1, tr_lh_y0); |
2055 | |
2056 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.sn, &win_ll_x0, &win_ll_x1); |
2057 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.dn, &win_hl_x0, &win_hl_x1); |
2058 | |
2059 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.sn, &win_ll_y0, &win_ll_y1); |
2060 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.dn, &win_lh_y0, &win_lh_y1); |
2061 | |
2062 | /* Compute the tile-resolution-based coordinates for the window of interest */ |
2063 | if (h.cas == 0) { |
2064 | win_tr_x0 = opj_uint_min(2 * win_ll_x0, 2 * win_hl_x0 + 1); |
2065 | win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_ll_x1, 2 * win_hl_x1 + 1), rw); |
2066 | } else { |
2067 | win_tr_x0 = opj_uint_min(2 * win_hl_x0, 2 * win_ll_x0 + 1); |
2068 | win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_hl_x1, 2 * win_ll_x1 + 1), rw); |
2069 | } |
2070 | |
2071 | if (v.cas == 0) { |
2072 | win_tr_y0 = opj_uint_min(2 * win_ll_y0, 2 * win_lh_y0 + 1); |
2073 | win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_ll_y1, 2 * win_lh_y1 + 1), rh); |
2074 | } else { |
2075 | win_tr_y0 = opj_uint_min(2 * win_lh_y0, 2 * win_ll_y0 + 1); |
2076 | win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_lh_y1, 2 * win_ll_y1 + 1), rh); |
2077 | } |
2078 | |
2079 | for (j = 0; j < rh; ++j) { |
2080 | if ((j >= win_ll_y0 && j < win_ll_y1) || |
2081 | (j >= win_lh_y0 + (OPJ_UINT32)v.sn && j < win_lh_y1 + (OPJ_UINT32)v.sn)) { |
2082 | |
2083 | /* Avoids dwt.c:1584:44 (in opj_dwt_decode_partial_1): runtime error: */ |
2084 | /* signed integer overflow: -1094795586 + -1094795586 cannot be represented in type 'int' */ |
2085 | /* on opj_decompress -i ../../openjpeg/MAPA.jp2 -o out.tif -d 0,0,256,256 */ |
2086 | /* This is less extreme than memsetting the whole buffer to 0 */ |
2087 | /* although we could potentially do better with better handling of edge conditions */ |
2088 | if (win_tr_x1 >= 1 && win_tr_x1 < rw) { |
2089 | h.mem[win_tr_x1 - 1] = 0; |
2090 | } |
2091 | if (win_tr_x1 < rw) { |
2092 | h.mem[win_tr_x1] = 0; |
2093 | } |
2094 | |
2095 | opj_dwt_interleave_partial_h(h.mem, |
2096 | h.cas, |
2097 | sa, |
2098 | j, |
2099 | (OPJ_UINT32)h.sn, |
2100 | win_ll_x0, |
2101 | win_ll_x1, |
2102 | win_hl_x0, |
2103 | win_hl_x1); |
2104 | opj_dwt_decode_partial_1(h.mem, h.dn, h.sn, h.cas, |
2105 | (OPJ_INT32)win_ll_x0, |
2106 | (OPJ_INT32)win_ll_x1, |
2107 | (OPJ_INT32)win_hl_x0, |
2108 | (OPJ_INT32)win_hl_x1); |
2109 | if (!opj_sparse_array_int32_write(sa, |
2110 | win_tr_x0, j, |
2111 | win_tr_x1, j + 1, |
2112 | h.mem + win_tr_x0, |
2113 | 1, 0, OPJ_TRUE)) { |
2114 | /* FIXME event manager error callback */ |
2115 | opj_sparse_array_int32_free(sa); |
2116 | opj_aligned_free(h.mem); |
2117 | return OPJ_FALSE; |
2118 | } |
2119 | } |
2120 | } |
2121 | |
2122 | for (i = win_tr_x0; i < win_tr_x1;) { |
2123 | OPJ_UINT32 nb_cols = opj_uint_min(4U, win_tr_x1 - i); |
2124 | opj_dwt_interleave_partial_v(v.mem, |
2125 | v.cas, |
2126 | sa, |
2127 | i, |
2128 | nb_cols, |
2129 | (OPJ_UINT32)v.sn, |
2130 | win_ll_y0, |
2131 | win_ll_y1, |
2132 | win_lh_y0, |
2133 | win_lh_y1); |
2134 | opj_dwt_decode_partial_1_parallel(v.mem, nb_cols, v.dn, v.sn, v.cas, |
2135 | (OPJ_INT32)win_ll_y0, |
2136 | (OPJ_INT32)win_ll_y1, |
2137 | (OPJ_INT32)win_lh_y0, |
2138 | (OPJ_INT32)win_lh_y1); |
2139 | if (!opj_sparse_array_int32_write(sa, |
2140 | i, win_tr_y0, |
2141 | i + nb_cols, win_tr_y1, |
2142 | v.mem + 4 * win_tr_y0, |
2143 | 1, 4, OPJ_TRUE)) { |
2144 | /* FIXME event manager error callback */ |
2145 | opj_sparse_array_int32_free(sa); |
2146 | opj_aligned_free(h.mem); |
2147 | return OPJ_FALSE; |
2148 | } |
2149 | |
2150 | i += nb_cols; |
2151 | } |
2152 | } |
2153 | opj_aligned_free(h.mem); |
2154 | |
2155 | { |
2156 | OPJ_BOOL ret = opj_sparse_array_int32_read(sa, |
2157 | tr_max->win_x0 - (OPJ_UINT32)tr_max->x0, |
2158 | tr_max->win_y0 - (OPJ_UINT32)tr_max->y0, |
2159 | tr_max->win_x1 - (OPJ_UINT32)tr_max->x0, |
2160 | tr_max->win_y1 - (OPJ_UINT32)tr_max->y0, |
2161 | tilec->data_win, |
2162 | 1, tr_max->win_x1 - tr_max->win_x0, |
2163 | OPJ_TRUE); |
2164 | assert(ret); |
2165 | OPJ_UNUSED(ret); |
2166 | } |
2167 | opj_sparse_array_int32_free(sa); |
2168 | return OPJ_TRUE; |
2169 | } |
2170 | |
2171 | static void opj_v4dwt_interleave_h(opj_v4dwt_t* OPJ_RESTRICT dwt, |
2172 | OPJ_FLOAT32* OPJ_RESTRICT a, |
2173 | OPJ_UINT32 width, |
2174 | OPJ_UINT32 remaining_height) |
2175 | { |
2176 | OPJ_FLOAT32* OPJ_RESTRICT bi = (OPJ_FLOAT32*)(dwt->wavelet + dwt->cas); |
2177 | OPJ_UINT32 i, k; |
2178 | OPJ_UINT32 x0 = dwt->win_l_x0; |
2179 | OPJ_UINT32 x1 = dwt->win_l_x1; |
2180 | |
2181 | for (k = 0; k < 2; ++k) { |
2182 | if (remaining_height >= 4 && ((OPJ_SIZE_T) a & 0x0f) == 0 && |
2183 | ((OPJ_SIZE_T) bi & 0x0f) == 0 && (width & 0x0f) == 0) { |
2184 | /* Fast code path */ |
2185 | for (i = x0; i < x1; ++i) { |
2186 | OPJ_UINT32 j = i; |
2187 | bi[i * 8 ] = a[j]; |
2188 | j += width; |
2189 | bi[i * 8 + 1] = a[j]; |
2190 | j += width; |
2191 | bi[i * 8 + 2] = a[j]; |
2192 | j += width; |
2193 | bi[i * 8 + 3] = a[j]; |
2194 | } |
2195 | } else { |
2196 | /* Slow code path */ |
2197 | for (i = x0; i < x1; ++i) { |
2198 | OPJ_UINT32 j = i; |
2199 | bi[i * 8 ] = a[j]; |
2200 | j += width; |
2201 | if (remaining_height == 1) { |
2202 | continue; |
2203 | } |
2204 | bi[i * 8 + 1] = a[j]; |
2205 | j += width; |
2206 | if (remaining_height == 2) { |
2207 | continue; |
2208 | } |
2209 | bi[i * 8 + 2] = a[j]; |
2210 | j += width; |
2211 | if (remaining_height == 3) { |
2212 | continue; |
2213 | } |
2214 | bi[i * 8 + 3] = a[j]; /* This one*/ |
2215 | } |
2216 | } |
2217 | |
2218 | bi = (OPJ_FLOAT32*)(dwt->wavelet + 1 - dwt->cas); |
2219 | a += dwt->sn; |
2220 | x0 = dwt->win_h_x0; |
2221 | x1 = dwt->win_h_x1; |
2222 | } |
2223 | } |
2224 | |
2225 | static void opj_v4dwt_interleave_partial_h(opj_v4dwt_t* dwt, |
2226 | opj_sparse_array_int32_t* sa, |
2227 | OPJ_UINT32 sa_line, |
2228 | OPJ_UINT32 remaining_height) |
2229 | { |
2230 | OPJ_UINT32 i; |
2231 | for (i = 0; i < remaining_height; i++) { |
2232 | OPJ_BOOL ret; |
2233 | ret = opj_sparse_array_int32_read(sa, |
2234 | dwt->win_l_x0, sa_line + i, |
2235 | dwt->win_l_x1, sa_line + i + 1, |
2236 | /* Nasty cast from float* to int32* */ |
2237 | (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0) + i, |
2238 | 8, 0, OPJ_TRUE); |
2239 | assert(ret); |
2240 | ret = opj_sparse_array_int32_read(sa, |
2241 | (OPJ_UINT32)dwt->sn + dwt->win_h_x0, sa_line + i, |
2242 | (OPJ_UINT32)dwt->sn + dwt->win_h_x1, sa_line + i + 1, |
2243 | /* Nasty cast from float* to int32* */ |
2244 | (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0) + i, |
2245 | 8, 0, OPJ_TRUE); |
2246 | assert(ret); |
2247 | OPJ_UNUSED(ret); |
2248 | } |
2249 | } |
2250 | |
2251 | static void opj_v4dwt_interleave_v(opj_v4dwt_t* OPJ_RESTRICT dwt, |
2252 | OPJ_FLOAT32* OPJ_RESTRICT a, |
2253 | OPJ_UINT32 width, |
2254 | OPJ_UINT32 nb_elts_read) |
2255 | { |
2256 | opj_v4_t* OPJ_RESTRICT bi = dwt->wavelet + dwt->cas; |
2257 | OPJ_UINT32 i; |
2258 | |
2259 | for (i = dwt->win_l_x0; i < dwt->win_l_x1; ++i) { |
2260 | memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width], |
2261 | (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32)); |
2262 | } |
2263 | |
2264 | a += (OPJ_UINT32)dwt->sn * (OPJ_SIZE_T)width; |
2265 | bi = dwt->wavelet + 1 - dwt->cas; |
2266 | |
2267 | for (i = dwt->win_h_x0; i < dwt->win_h_x1; ++i) { |
2268 | memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width], |
2269 | (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32)); |
2270 | } |
2271 | } |
2272 | |
2273 | static void opj_v4dwt_interleave_partial_v(opj_v4dwt_t* OPJ_RESTRICT dwt, |
2274 | opj_sparse_array_int32_t* sa, |
2275 | OPJ_UINT32 sa_col, |
2276 | OPJ_UINT32 nb_elts_read) |
2277 | { |
2278 | OPJ_BOOL ret; |
2279 | ret = opj_sparse_array_int32_read(sa, |
2280 | sa_col, dwt->win_l_x0, |
2281 | sa_col + nb_elts_read, dwt->win_l_x1, |
2282 | (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0), |
2283 | 1, 8, OPJ_TRUE); |
2284 | assert(ret); |
2285 | ret = opj_sparse_array_int32_read(sa, |
2286 | sa_col, (OPJ_UINT32)dwt->sn + dwt->win_h_x0, |
2287 | sa_col + nb_elts_read, (OPJ_UINT32)dwt->sn + dwt->win_h_x1, |
2288 | (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0), |
2289 | 1, 8, OPJ_TRUE); |
2290 | assert(ret); |
2291 | OPJ_UNUSED(ret); |
2292 | } |
2293 | |
2294 | #ifdef __SSE__ |
2295 | |
2296 | static void opj_v4dwt_decode_step1_sse(opj_v4_t* w, |
2297 | OPJ_UINT32 start, |
2298 | OPJ_UINT32 end, |
2299 | const __m128 c) |
2300 | { |
2301 | __m128* OPJ_RESTRICT vw = (__m128*) w; |
2302 | OPJ_UINT32 i; |
2303 | /* 4x unrolled loop */ |
2304 | vw += 2 * start; |
2305 | for (i = start; i + 3 < end; i += 4, vw += 8) { |
2306 | __m128 xmm0 = _mm_mul_ps(vw[0], c); |
2307 | __m128 xmm2 = _mm_mul_ps(vw[2], c); |
2308 | __m128 xmm4 = _mm_mul_ps(vw[4], c); |
2309 | __m128 xmm6 = _mm_mul_ps(vw[6], c); |
2310 | vw[0] = xmm0; |
2311 | vw[2] = xmm2; |
2312 | vw[4] = xmm4; |
2313 | vw[6] = xmm6; |
2314 | } |
2315 | for (; i < end; ++i, vw += 2) { |
2316 | vw[0] = _mm_mul_ps(vw[0], c); |
2317 | } |
2318 | } |
2319 | |
2320 | static void opj_v4dwt_decode_step2_sse(opj_v4_t* l, opj_v4_t* w, |
2321 | OPJ_UINT32 start, |
2322 | OPJ_UINT32 end, |
2323 | OPJ_UINT32 m, |
2324 | __m128 c) |
2325 | { |
2326 | __m128* OPJ_RESTRICT vl = (__m128*) l; |
2327 | __m128* OPJ_RESTRICT vw = (__m128*) w; |
2328 | OPJ_UINT32 i; |
2329 | OPJ_UINT32 imax = opj_uint_min(end, m); |
2330 | __m128 tmp1, tmp2, tmp3; |
2331 | if (start == 0) { |
2332 | tmp1 = vl[0]; |
2333 | } else { |
2334 | vw += start * 2; |
2335 | tmp1 = vw[-3]; |
2336 | } |
2337 | |
2338 | i = start; |
2339 | |
2340 | /* 4x loop unrolling */ |
2341 | for (; i + 3 < imax; i += 4) { |
2342 | __m128 tmp4, tmp5, tmp6, tmp7, tmp8, tmp9; |
2343 | tmp2 = vw[-1]; |
2344 | tmp3 = vw[ 0]; |
2345 | tmp4 = vw[ 1]; |
2346 | tmp5 = vw[ 2]; |
2347 | tmp6 = vw[ 3]; |
2348 | tmp7 = vw[ 4]; |
2349 | tmp8 = vw[ 5]; |
2350 | tmp9 = vw[ 6]; |
2351 | vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c)); |
2352 | vw[ 1] = _mm_add_ps(tmp4, _mm_mul_ps(_mm_add_ps(tmp3, tmp5), c)); |
2353 | vw[ 3] = _mm_add_ps(tmp6, _mm_mul_ps(_mm_add_ps(tmp5, tmp7), c)); |
2354 | vw[ 5] = _mm_add_ps(tmp8, _mm_mul_ps(_mm_add_ps(tmp7, tmp9), c)); |
2355 | tmp1 = tmp9; |
2356 | vw += 8; |
2357 | } |
2358 | |
2359 | for (; i < imax; ++i) { |
2360 | tmp2 = vw[-1]; |
2361 | tmp3 = vw[ 0]; |
2362 | vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c)); |
2363 | tmp1 = tmp3; |
2364 | vw += 2; |
2365 | } |
2366 | if (m < end) { |
2367 | assert(m + 1 == end); |
2368 | c = _mm_add_ps(c, c); |
2369 | c = _mm_mul_ps(c, vw[-2]); |
2370 | vw[-1] = _mm_add_ps(vw[-1], c); |
2371 | } |
2372 | } |
2373 | |
2374 | #else |
2375 | |
2376 | static void opj_v4dwt_decode_step1(opj_v4_t* w, |
2377 | OPJ_UINT32 start, |
2378 | OPJ_UINT32 end, |
2379 | const OPJ_FLOAT32 c) |
2380 | { |
2381 | OPJ_FLOAT32* OPJ_RESTRICT fw = (OPJ_FLOAT32*) w; |
2382 | OPJ_UINT32 i; |
2383 | for (i = start; i < end; ++i) { |
2384 | OPJ_FLOAT32 tmp1 = fw[i * 8 ]; |
2385 | OPJ_FLOAT32 tmp2 = fw[i * 8 + 1]; |
2386 | OPJ_FLOAT32 tmp3 = fw[i * 8 + 2]; |
2387 | OPJ_FLOAT32 tmp4 = fw[i * 8 + 3]; |
2388 | fw[i * 8 ] = tmp1 * c; |
2389 | fw[i * 8 + 1] = tmp2 * c; |
2390 | fw[i * 8 + 2] = tmp3 * c; |
2391 | fw[i * 8 + 3] = tmp4 * c; |
2392 | } |
2393 | } |
2394 | |
2395 | static void opj_v4dwt_decode_step2(opj_v4_t* l, opj_v4_t* w, |
2396 | OPJ_UINT32 start, |
2397 | OPJ_UINT32 end, |
2398 | OPJ_UINT32 m, |
2399 | OPJ_FLOAT32 c) |
2400 | { |
2401 | OPJ_FLOAT32* fl = (OPJ_FLOAT32*) l; |
2402 | OPJ_FLOAT32* fw = (OPJ_FLOAT32*) w; |
2403 | OPJ_UINT32 i; |
2404 | OPJ_UINT32 imax = opj_uint_min(end, m); |
2405 | if (start > 0) { |
2406 | fw += 8 * start; |
2407 | fl = fw - 8; |
2408 | } |
2409 | for (i = start; i < imax; ++i) { |
2410 | OPJ_FLOAT32 tmp1_1 = fl[0]; |
2411 | OPJ_FLOAT32 tmp1_2 = fl[1]; |
2412 | OPJ_FLOAT32 tmp1_3 = fl[2]; |
2413 | OPJ_FLOAT32 tmp1_4 = fl[3]; |
2414 | OPJ_FLOAT32 tmp2_1 = fw[-4]; |
2415 | OPJ_FLOAT32 tmp2_2 = fw[-3]; |
2416 | OPJ_FLOAT32 tmp2_3 = fw[-2]; |
2417 | OPJ_FLOAT32 tmp2_4 = fw[-1]; |
2418 | OPJ_FLOAT32 tmp3_1 = fw[0]; |
2419 | OPJ_FLOAT32 tmp3_2 = fw[1]; |
2420 | OPJ_FLOAT32 tmp3_3 = fw[2]; |
2421 | OPJ_FLOAT32 tmp3_4 = fw[3]; |
2422 | fw[-4] = tmp2_1 + ((tmp1_1 + tmp3_1) * c); |
2423 | fw[-3] = tmp2_2 + ((tmp1_2 + tmp3_2) * c); |
2424 | fw[-2] = tmp2_3 + ((tmp1_3 + tmp3_3) * c); |
2425 | fw[-1] = tmp2_4 + ((tmp1_4 + tmp3_4) * c); |
2426 | fl = fw; |
2427 | fw += 8; |
2428 | } |
2429 | if (m < end) { |
2430 | assert(m + 1 == end); |
2431 | c += c; |
2432 | fw[-4] = fw[-4] + fl[0] * c; |
2433 | fw[-3] = fw[-3] + fl[1] * c; |
2434 | fw[-2] = fw[-2] + fl[2] * c; |
2435 | fw[-1] = fw[-1] + fl[3] * c; |
2436 | } |
2437 | } |
2438 | |
2439 | #endif |
2440 | |
2441 | /* <summary> */ |
2442 | /* Inverse 9-7 wavelet transform in 1-D. */ |
2443 | /* </summary> */ |
2444 | static void opj_v4dwt_decode(opj_v4dwt_t* OPJ_RESTRICT dwt) |
2445 | { |
2446 | OPJ_INT32 a, b; |
2447 | if (dwt->cas == 0) { |
2448 | if (!((dwt->dn > 0) || (dwt->sn > 1))) { |
2449 | return; |
2450 | } |
2451 | a = 0; |
2452 | b = 1; |
2453 | } else { |
2454 | if (!((dwt->sn > 0) || (dwt->dn > 1))) { |
2455 | return; |
2456 | } |
2457 | a = 1; |
2458 | b = 0; |
2459 | } |
2460 | #ifdef __SSE__ |
2461 | opj_v4dwt_decode_step1_sse(dwt->wavelet + a, dwt->win_l_x0, dwt->win_l_x1, |
2462 | _mm_set1_ps(opj_K)); |
2463 | opj_v4dwt_decode_step1_sse(dwt->wavelet + b, dwt->win_h_x0, dwt->win_h_x1, |
2464 | _mm_set1_ps(opj_c13318)); |
2465 | opj_v4dwt_decode_step2_sse(dwt->wavelet + b, dwt->wavelet + a + 1, |
2466 | dwt->win_l_x0, dwt->win_l_x1, |
2467 | (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a), |
2468 | _mm_set1_ps(opj_dwt_delta)); |
2469 | opj_v4dwt_decode_step2_sse(dwt->wavelet + a, dwt->wavelet + b + 1, |
2470 | dwt->win_h_x0, dwt->win_h_x1, |
2471 | (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b), |
2472 | _mm_set1_ps(opj_dwt_gamma)); |
2473 | opj_v4dwt_decode_step2_sse(dwt->wavelet + b, dwt->wavelet + a + 1, |
2474 | dwt->win_l_x0, dwt->win_l_x1, |
2475 | (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a), |
2476 | _mm_set1_ps(opj_dwt_beta)); |
2477 | opj_v4dwt_decode_step2_sse(dwt->wavelet + a, dwt->wavelet + b + 1, |
2478 | dwt->win_h_x0, dwt->win_h_x1, |
2479 | (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b), |
2480 | _mm_set1_ps(opj_dwt_alpha)); |
2481 | #else |
2482 | opj_v4dwt_decode_step1(dwt->wavelet + a, dwt->win_l_x0, dwt->win_l_x1, |
2483 | opj_K); |
2484 | opj_v4dwt_decode_step1(dwt->wavelet + b, dwt->win_h_x0, dwt->win_h_x1, |
2485 | opj_c13318); |
2486 | opj_v4dwt_decode_step2(dwt->wavelet + b, dwt->wavelet + a + 1, |
2487 | dwt->win_l_x0, dwt->win_l_x1, |
2488 | (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a), |
2489 | opj_dwt_delta); |
2490 | opj_v4dwt_decode_step2(dwt->wavelet + a, dwt->wavelet + b + 1, |
2491 | dwt->win_h_x0, dwt->win_h_x1, |
2492 | (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b), |
2493 | opj_dwt_gamma); |
2494 | opj_v4dwt_decode_step2(dwt->wavelet + b, dwt->wavelet + a + 1, |
2495 | dwt->win_l_x0, dwt->win_l_x1, |
2496 | (OPJ_UINT32)opj_int_min(dwt->sn, dwt->dn - a), |
2497 | opj_dwt_beta); |
2498 | opj_v4dwt_decode_step2(dwt->wavelet + a, dwt->wavelet + b + 1, |
2499 | dwt->win_h_x0, dwt->win_h_x1, |
2500 | (OPJ_UINT32)opj_int_min(dwt->dn, dwt->sn - b), |
2501 | opj_dwt_alpha); |
2502 | #endif |
2503 | } |
2504 | |
2505 | |
2506 | /* <summary> */ |
2507 | /* Inverse 9-7 wavelet transform in 2-D. */ |
2508 | /* </summary> */ |
2509 | static |
2510 | OPJ_BOOL opj_dwt_decode_tile_97(opj_tcd_tilecomp_t* OPJ_RESTRICT tilec, |
2511 | OPJ_UINT32 numres) |
2512 | { |
2513 | opj_v4dwt_t h; |
2514 | opj_v4dwt_t v; |
2515 | |
2516 | opj_tcd_resolution_t* res = tilec->resolutions; |
2517 | |
2518 | OPJ_UINT32 rw = (OPJ_UINT32)(res->x1 - |
2519 | res->x0); /* width of the resolution level computed */ |
2520 | OPJ_UINT32 rh = (OPJ_UINT32)(res->y1 - |
2521 | res->y0); /* height of the resolution level computed */ |
2522 | |
2523 | OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions - |
2524 | 1].x1 - |
2525 | tilec->resolutions[tilec->minimum_num_resolutions - 1].x0); |
2526 | |
2527 | OPJ_SIZE_T l_data_size; |
2528 | |
2529 | l_data_size = opj_dwt_max_resolution(res, numres); |
2530 | /* overflow check */ |
2531 | if (l_data_size > (SIZE_MAX - 5U)) { |
2532 | /* FIXME event manager error callback */ |
2533 | return OPJ_FALSE; |
2534 | } |
2535 | l_data_size += 5U; |
2536 | /* overflow check */ |
2537 | if (l_data_size > (SIZE_MAX / sizeof(opj_v4_t))) { |
2538 | /* FIXME event manager error callback */ |
2539 | return OPJ_FALSE; |
2540 | } |
2541 | h.wavelet = (opj_v4_t*) opj_aligned_malloc(l_data_size * sizeof(opj_v4_t)); |
2542 | if (!h.wavelet) { |
2543 | /* FIXME event manager error callback */ |
2544 | return OPJ_FALSE; |
2545 | } |
2546 | v.wavelet = h.wavelet; |
2547 | |
2548 | while (--numres) { |
2549 | OPJ_FLOAT32 * OPJ_RESTRICT aj = (OPJ_FLOAT32*) tilec->data; |
2550 | OPJ_UINT32 j; |
2551 | |
2552 | h.sn = (OPJ_INT32)rw; |
2553 | v.sn = (OPJ_INT32)rh; |
2554 | |
2555 | ++res; |
2556 | |
2557 | rw = (OPJ_UINT32)(res->x1 - |
2558 | res->x0); /* width of the resolution level computed */ |
2559 | rh = (OPJ_UINT32)(res->y1 - |
2560 | res->y0); /* height of the resolution level computed */ |
2561 | |
2562 | h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn); |
2563 | h.cas = res->x0 % 2; |
2564 | |
2565 | h.win_l_x0 = 0; |
2566 | h.win_l_x1 = (OPJ_UINT32)h.sn; |
2567 | h.win_h_x0 = 0; |
2568 | h.win_h_x1 = (OPJ_UINT32)h.dn; |
2569 | for (j = 0; j + 3 < rh; j += 4) { |
2570 | OPJ_UINT32 k; |
2571 | opj_v4dwt_interleave_h(&h, aj, w, rh - j); |
2572 | opj_v4dwt_decode(&h); |
2573 | |
2574 | for (k = 0; k < rw; k++) { |
2575 | aj[k ] = h.wavelet[k].f[0]; |
2576 | aj[k + (OPJ_SIZE_T)w ] = h.wavelet[k].f[1]; |
2577 | aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2]; |
2578 | aj[k + (OPJ_SIZE_T)w * 3] = h.wavelet[k].f[3]; |
2579 | } |
2580 | |
2581 | aj += w * 4; |
2582 | } |
2583 | |
2584 | if (j < rh) { |
2585 | OPJ_UINT32 k; |
2586 | opj_v4dwt_interleave_h(&h, aj, w, rh - j); |
2587 | opj_v4dwt_decode(&h); |
2588 | for (k = 0; k < rw; k++) { |
2589 | switch (rh - j) { |
2590 | case 3: |
2591 | aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2]; |
2592 | /* FALLTHRU */ |
2593 | case 2: |
2594 | aj[k + (OPJ_SIZE_T)w ] = h.wavelet[k].f[1]; |
2595 | /* FALLTHRU */ |
2596 | case 1: |
2597 | aj[k] = h.wavelet[k].f[0]; |
2598 | } |
2599 | } |
2600 | } |
2601 | |
2602 | v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn); |
2603 | v.cas = res->y0 % 2; |
2604 | v.win_l_x0 = 0; |
2605 | v.win_l_x1 = (OPJ_UINT32)v.sn; |
2606 | v.win_h_x0 = 0; |
2607 | v.win_h_x1 = (OPJ_UINT32)v.dn; |
2608 | |
2609 | aj = (OPJ_FLOAT32*) tilec->data; |
2610 | for (j = rw; j > 3; j -= 4) { |
2611 | OPJ_UINT32 k; |
2612 | |
2613 | opj_v4dwt_interleave_v(&v, aj, w, 4); |
2614 | opj_v4dwt_decode(&v); |
2615 | |
2616 | for (k = 0; k < rh; ++k) { |
2617 | memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k], 4 * sizeof(OPJ_FLOAT32)); |
2618 | } |
2619 | aj += 4; |
2620 | } |
2621 | |
2622 | if (rw & 0x03) { |
2623 | OPJ_UINT32 k; |
2624 | |
2625 | j = rw & 0x03; |
2626 | |
2627 | opj_v4dwt_interleave_v(&v, aj, w, j); |
2628 | opj_v4dwt_decode(&v); |
2629 | |
2630 | for (k = 0; k < rh; ++k) { |
2631 | memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k], |
2632 | (OPJ_SIZE_T)j * sizeof(OPJ_FLOAT32)); |
2633 | } |
2634 | } |
2635 | } |
2636 | |
2637 | opj_aligned_free(h.wavelet); |
2638 | return OPJ_TRUE; |
2639 | } |
2640 | |
2641 | static |
2642 | OPJ_BOOL opj_dwt_decode_partial_97(opj_tcd_tilecomp_t* OPJ_RESTRICT tilec, |
2643 | OPJ_UINT32 numres) |
2644 | { |
2645 | opj_sparse_array_int32_t* sa; |
2646 | opj_v4dwt_t h; |
2647 | opj_v4dwt_t v; |
2648 | OPJ_UINT32 resno; |
2649 | /* This value matches the maximum left/right extension given in tables */ |
2650 | /* F.2 and F.3 of the standard. Note: in opj_tcd_is_subband_area_of_interest() */ |
2651 | /* we currently use 3. */ |
2652 | const OPJ_UINT32 filter_width = 4U; |
2653 | |
2654 | opj_tcd_resolution_t* tr = tilec->resolutions; |
2655 | opj_tcd_resolution_t* tr_max = &(tilec->resolutions[numres - 1]); |
2656 | |
2657 | OPJ_UINT32 rw = (OPJ_UINT32)(tr->x1 - |
2658 | tr->x0); /* width of the resolution level computed */ |
2659 | OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 - |
2660 | tr->y0); /* height of the resolution level computed */ |
2661 | |
2662 | OPJ_SIZE_T l_data_size; |
2663 | |
2664 | /* Compute the intersection of the area of interest, expressed in tile coordinates */ |
2665 | /* with the tile coordinates */ |
2666 | OPJ_UINT32 win_tcx0 = tilec->win_x0; |
2667 | OPJ_UINT32 win_tcy0 = tilec->win_y0; |
2668 | OPJ_UINT32 win_tcx1 = tilec->win_x1; |
2669 | OPJ_UINT32 win_tcy1 = tilec->win_y1; |
2670 | |
2671 | if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) { |
2672 | return OPJ_TRUE; |
2673 | } |
2674 | |
2675 | sa = opj_dwt_init_sparse_array(tilec, numres); |
2676 | if (sa == NULL) { |
2677 | return OPJ_FALSE; |
2678 | } |
2679 | |
2680 | if (numres == 1U) { |
2681 | OPJ_BOOL ret = opj_sparse_array_int32_read(sa, |
2682 | tr_max->win_x0 - (OPJ_UINT32)tr_max->x0, |
2683 | tr_max->win_y0 - (OPJ_UINT32)tr_max->y0, |
2684 | tr_max->win_x1 - (OPJ_UINT32)tr_max->x0, |
2685 | tr_max->win_y1 - (OPJ_UINT32)tr_max->y0, |
2686 | tilec->data_win, |
2687 | 1, tr_max->win_x1 - tr_max->win_x0, |
2688 | OPJ_TRUE); |
2689 | assert(ret); |
2690 | OPJ_UNUSED(ret); |
2691 | opj_sparse_array_int32_free(sa); |
2692 | return OPJ_TRUE; |
2693 | } |
2694 | |
2695 | l_data_size = opj_dwt_max_resolution(tr, numres); |
2696 | /* overflow check */ |
2697 | if (l_data_size > (SIZE_MAX - 5U)) { |
2698 | /* FIXME event manager error callback */ |
2699 | return OPJ_FALSE; |
2700 | } |
2701 | l_data_size += 5U; |
2702 | /* overflow check */ |
2703 | if (l_data_size > (SIZE_MAX / sizeof(opj_v4_t))) { |
2704 | /* FIXME event manager error callback */ |
2705 | return OPJ_FALSE; |
2706 | } |
2707 | h.wavelet = (opj_v4_t*) opj_aligned_malloc(l_data_size * sizeof(opj_v4_t)); |
2708 | if (!h.wavelet) { |
2709 | /* FIXME event manager error callback */ |
2710 | return OPJ_FALSE; |
2711 | } |
2712 | v.wavelet = h.wavelet; |
2713 | |
2714 | for (resno = 1; resno < numres; resno ++) { |
2715 | OPJ_UINT32 j; |
2716 | /* Window of interest subband-based coordinates */ |
2717 | OPJ_UINT32 win_ll_x0, win_ll_y0, win_ll_x1, win_ll_y1; |
2718 | OPJ_UINT32 win_hl_x0, win_hl_x1; |
2719 | OPJ_UINT32 win_lh_y0, win_lh_y1; |
2720 | /* Window of interest tile-resolution-based coordinates */ |
2721 | OPJ_UINT32 win_tr_x0, win_tr_x1, win_tr_y0, win_tr_y1; |
2722 | /* Tile-resolution subband-based coordinates */ |
2723 | OPJ_UINT32 tr_ll_x0, tr_ll_y0, tr_hl_x0, tr_lh_y0; |
2724 | |
2725 | ++tr; |
2726 | |
2727 | h.sn = (OPJ_INT32)rw; |
2728 | v.sn = (OPJ_INT32)rh; |
2729 | |
2730 | rw = (OPJ_UINT32)(tr->x1 - tr->x0); |
2731 | rh = (OPJ_UINT32)(tr->y1 - tr->y0); |
2732 | |
2733 | h.dn = (OPJ_INT32)(rw - (OPJ_UINT32)h.sn); |
2734 | h.cas = tr->x0 % 2; |
2735 | |
2736 | v.dn = (OPJ_INT32)(rh - (OPJ_UINT32)v.sn); |
2737 | v.cas = tr->y0 % 2; |
2738 | |
2739 | /* Get the subband coordinates for the window of interest */ |
2740 | /* LL band */ |
2741 | opj_dwt_get_band_coordinates(tilec, resno, 0, |
2742 | win_tcx0, win_tcy0, win_tcx1, win_tcy1, |
2743 | &win_ll_x0, &win_ll_y0, |
2744 | &win_ll_x1, &win_ll_y1); |
2745 | |
2746 | /* HL band */ |
2747 | opj_dwt_get_band_coordinates(tilec, resno, 1, |
2748 | win_tcx0, win_tcy0, win_tcx1, win_tcy1, |
2749 | &win_hl_x0, NULL, &win_hl_x1, NULL); |
2750 | |
2751 | /* LH band */ |
2752 | opj_dwt_get_band_coordinates(tilec, resno, 2, |
2753 | win_tcx0, win_tcy0, win_tcx1, win_tcy1, |
2754 | NULL, &win_lh_y0, NULL, &win_lh_y1); |
2755 | |
2756 | /* Beware: band index for non-LL0 resolution are 0=HL, 1=LH and 2=HH */ |
2757 | tr_ll_x0 = (OPJ_UINT32)tr->bands[1].x0; |
2758 | tr_ll_y0 = (OPJ_UINT32)tr->bands[0].y0; |
2759 | tr_hl_x0 = (OPJ_UINT32)tr->bands[0].x0; |
2760 | tr_lh_y0 = (OPJ_UINT32)tr->bands[1].y0; |
2761 | |
2762 | /* Substract the origin of the bands for this tile, to the subwindow */ |
2763 | /* of interest band coordinates, so as to get them relative to the */ |
2764 | /* tile */ |
2765 | win_ll_x0 = opj_uint_subs(win_ll_x0, tr_ll_x0); |
2766 | win_ll_y0 = opj_uint_subs(win_ll_y0, tr_ll_y0); |
2767 | win_ll_x1 = opj_uint_subs(win_ll_x1, tr_ll_x0); |
2768 | win_ll_y1 = opj_uint_subs(win_ll_y1, tr_ll_y0); |
2769 | win_hl_x0 = opj_uint_subs(win_hl_x0, tr_hl_x0); |
2770 | win_hl_x1 = opj_uint_subs(win_hl_x1, tr_hl_x0); |
2771 | win_lh_y0 = opj_uint_subs(win_lh_y0, tr_lh_y0); |
2772 | win_lh_y1 = opj_uint_subs(win_lh_y1, tr_lh_y0); |
2773 | |
2774 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.sn, &win_ll_x0, &win_ll_x1); |
2775 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)h.dn, &win_hl_x0, &win_hl_x1); |
2776 | |
2777 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.sn, &win_ll_y0, &win_ll_y1); |
2778 | opj_dwt_segment_grow(filter_width, (OPJ_UINT32)v.dn, &win_lh_y0, &win_lh_y1); |
2779 | |
2780 | /* Compute the tile-resolution-based coordinates for the window of interest */ |
2781 | if (h.cas == 0) { |
2782 | win_tr_x0 = opj_uint_min(2 * win_ll_x0, 2 * win_hl_x0 + 1); |
2783 | win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_ll_x1, 2 * win_hl_x1 + 1), rw); |
2784 | } else { |
2785 | win_tr_x0 = opj_uint_min(2 * win_hl_x0, 2 * win_ll_x0 + 1); |
2786 | win_tr_x1 = opj_uint_min(opj_uint_max(2 * win_hl_x1, 2 * win_ll_x1 + 1), rw); |
2787 | } |
2788 | |
2789 | if (v.cas == 0) { |
2790 | win_tr_y0 = opj_uint_min(2 * win_ll_y0, 2 * win_lh_y0 + 1); |
2791 | win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_ll_y1, 2 * win_lh_y1 + 1), rh); |
2792 | } else { |
2793 | win_tr_y0 = opj_uint_min(2 * win_lh_y0, 2 * win_ll_y0 + 1); |
2794 | win_tr_y1 = opj_uint_min(opj_uint_max(2 * win_lh_y1, 2 * win_ll_y1 + 1), rh); |
2795 | } |
2796 | |
2797 | h.win_l_x0 = win_ll_x0; |
2798 | h.win_l_x1 = win_ll_x1; |
2799 | h.win_h_x0 = win_hl_x0; |
2800 | h.win_h_x1 = win_hl_x1; |
2801 | for (j = 0; j + 3 < rh; j += 4) { |
2802 | if ((j + 3 >= win_ll_y0 && j < win_ll_y1) || |
2803 | (j + 3 >= win_lh_y0 + (OPJ_UINT32)v.sn && |
2804 | j < win_lh_y1 + (OPJ_UINT32)v.sn)) { |
2805 | opj_v4dwt_interleave_partial_h(&h, sa, j, opj_uint_min(4U, rh - j)); |
2806 | opj_v4dwt_decode(&h); |
2807 | if (!opj_sparse_array_int32_write(sa, |
2808 | win_tr_x0, j, |
2809 | win_tr_x1, j + 4, |
2810 | (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0], |
2811 | 4, 1, OPJ_TRUE)) { |
2812 | /* FIXME event manager error callback */ |
2813 | opj_sparse_array_int32_free(sa); |
2814 | opj_aligned_free(h.wavelet); |
2815 | return OPJ_FALSE; |
2816 | } |
2817 | } |
2818 | } |
2819 | |
2820 | if (j < rh && |
2821 | ((j + 3 >= win_ll_y0 && j < win_ll_y1) || |
2822 | (j + 3 >= win_lh_y0 + (OPJ_UINT32)v.sn && |
2823 | j < win_lh_y1 + (OPJ_UINT32)v.sn))) { |
2824 | opj_v4dwt_interleave_partial_h(&h, sa, j, rh - j); |
2825 | opj_v4dwt_decode(&h); |
2826 | if (!opj_sparse_array_int32_write(sa, |
2827 | win_tr_x0, j, |
2828 | win_tr_x1, rh, |
2829 | (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0], |
2830 | 4, 1, OPJ_TRUE)) { |
2831 | /* FIXME event manager error callback */ |
2832 | opj_sparse_array_int32_free(sa); |
2833 | opj_aligned_free(h.wavelet); |
2834 | return OPJ_FALSE; |
2835 | } |
2836 | } |
2837 | |
2838 | v.win_l_x0 = win_ll_y0; |
2839 | v.win_l_x1 = win_ll_y1; |
2840 | v.win_h_x0 = win_lh_y0; |
2841 | v.win_h_x1 = win_lh_y1; |
2842 | for (j = win_tr_x0; j < win_tr_x1; j += 4) { |
2843 | OPJ_UINT32 nb_elts = opj_uint_min(4U, win_tr_x1 - j); |
2844 | |
2845 | opj_v4dwt_interleave_partial_v(&v, sa, j, nb_elts); |
2846 | opj_v4dwt_decode(&v); |
2847 | |
2848 | if (!opj_sparse_array_int32_write(sa, |
2849 | j, win_tr_y0, |
2850 | j + nb_elts, win_tr_y1, |
2851 | (OPJ_INT32*)&h.wavelet[win_tr_y0].f[0], |
2852 | 1, 4, OPJ_TRUE)) { |
2853 | /* FIXME event manager error callback */ |
2854 | opj_sparse_array_int32_free(sa); |
2855 | opj_aligned_free(h.wavelet); |
2856 | return OPJ_FALSE; |
2857 | } |
2858 | } |
2859 | } |
2860 | |
2861 | { |
2862 | OPJ_BOOL ret = opj_sparse_array_int32_read(sa, |
2863 | tr_max->win_x0 - (OPJ_UINT32)tr_max->x0, |
2864 | tr_max->win_y0 - (OPJ_UINT32)tr_max->y0, |
2865 | tr_max->win_x1 - (OPJ_UINT32)tr_max->x0, |
2866 | tr_max->win_y1 - (OPJ_UINT32)tr_max->y0, |
2867 | tilec->data_win, |
2868 | 1, tr_max->win_x1 - tr_max->win_x0, |
2869 | OPJ_TRUE); |
2870 | assert(ret); |
2871 | OPJ_UNUSED(ret); |
2872 | } |
2873 | opj_sparse_array_int32_free(sa); |
2874 | |
2875 | opj_aligned_free(h.wavelet); |
2876 | return OPJ_TRUE; |
2877 | } |
2878 | |
2879 | |
2880 | OPJ_BOOL opj_dwt_decode_real(opj_tcd_t *p_tcd, |
2881 | opj_tcd_tilecomp_t* OPJ_RESTRICT tilec, |
2882 | OPJ_UINT32 numres) |
2883 | { |
2884 | if (p_tcd->whole_tile_decoding) { |
2885 | return opj_dwt_decode_tile_97(tilec, numres); |
2886 | } else { |
2887 | return opj_dwt_decode_partial_97(tilec, numres); |
2888 | } |
2889 | } |
2890 | |