/src/libhevc/common/ihevc_itrans_res.c
Line | Count | Source |
1 | | /****************************************************************************** |
2 | | * |
3 | | * Copyright (C) 2012 Ittiam Systems Pvt Ltd, Bangalore |
4 | | * |
5 | | * Licensed under the Apache License, Version 2.0 (the "License"); |
6 | | * you may not use this file except in compliance with the License. |
7 | | * You may obtain a copy of the License at: |
8 | | * |
9 | | * http://www.apache.org/licenses/LICENSE-2.0 |
10 | | * |
11 | | * Unless required by applicable law or agreed to in writing, software |
12 | | * distributed under the License is distributed on an "AS IS" BASIS, |
13 | | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
14 | | * See the License for the specific language governing permissions and |
15 | | * limitations under the License. |
16 | | * |
17 | | ******************************************************************************/ |
18 | | /** |
19 | | ******************************************************************************* |
20 | | * @file |
21 | | * ihevc_itrans_res.c |
22 | | * |
23 | | * @brief |
24 | | * Contains function definitions for inverse transform |
25 | | * |
26 | | * @author |
27 | | * 100470 |
28 | | * |
29 | | * @par List of Functions: |
30 | | * - ihevc_itrans_res_4x4_ttype1() |
31 | | * - ihevc_itrans_res_4x4() |
32 | | * - ihevcd_itrans_res_dc() |
33 | | * - ihevc_itrans_res_8x8() |
34 | | * - ihevc_itrans_res_16x16() |
35 | | * - ihevc_itrans_res_32x32() |
36 | | * - ihevc_res_4x4_rotate() |
37 | | * - ihevc_res_nxn_copy() |
38 | | * - ihevc_res_nxn_rdpcm_horz() |
39 | | * - ihevc_res_nxn_rdpcm_vert() |
40 | | * |
41 | | * @remarks |
42 | | * None |
43 | | * |
44 | | ******************************************************************************* |
45 | | */ |
46 | | |
47 | | #include <stdio.h> |
48 | | #include <string.h> |
49 | | |
50 | | #include "ihevc_typedefs.h" |
51 | | #include "ihevc_macros.h" |
52 | | #include "ihevc_platform_macros.h" |
53 | | #include "ihevc_defs.h" |
54 | | #include "ihevc_trans_tables.h" |
55 | | #include "ihevc_func_selector.h" |
56 | | #include "ihevc_trans_macros.h" |
57 | | #include "ihevc_itrans_res.h" |
58 | | |
59 | | |
60 | | void ihevc_itrans_res_4x4_ttype1(WORD16 *pi2_src, |
61 | | WORD16 *pi2_tmp, |
62 | | WORD16 *pi2_dst, |
63 | | WORD32 src_strd, |
64 | | WORD32 dst_strd, |
65 | | WORD32 zero_cols, |
66 | | WORD32 zero_rows) |
67 | 0 | { |
68 | 0 | WORD32 i, c[4]; |
69 | 0 | WORD32 add; |
70 | 0 | WORD32 shift; |
71 | 0 | WORD16 *pi2_tmp_orig; |
72 | 0 | WORD32 trans_size; |
73 | 0 | UNUSED(zero_rows); |
74 | 0 | trans_size = TRANS_SIZE_4; |
75 | |
|
76 | 0 | pi2_tmp_orig = pi2_tmp; |
77 | | |
78 | | /* Inverse Transform 1st stage */ |
79 | 0 | shift = IT_SHIFT_STAGE_1; |
80 | 0 | add = 1 << (shift - 1); |
81 | |
|
82 | 0 | for(i = 0; i < trans_size; i++) |
83 | 0 | { |
84 | | /* Checking for Zero Cols */ |
85 | 0 | if((zero_cols & 1) == 1) |
86 | 0 | { |
87 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
88 | 0 | } |
89 | 0 | else |
90 | 0 | { |
91 | | // Intermediate Variables |
92 | 0 | c[0] = pi2_src[0] + pi2_src[2 * src_strd]; |
93 | 0 | c[1] = pi2_src[2 * src_strd] + pi2_src[3 * src_strd]; |
94 | 0 | c[2] = pi2_src[0] - pi2_src[3 * src_strd]; |
95 | 0 | c[3] = 74 * pi2_src[src_strd]; |
96 | |
|
97 | 0 | pi2_tmp[0] = |
98 | 0 | CLIP_S16((29 * c[0] + 55 * c[1] + c[3] + add) >> shift); |
99 | 0 | pi2_tmp[1] = |
100 | 0 | CLIP_S16((55 * c[2] - 29 * c[1] + c[3] + add) >> shift); |
101 | 0 | pi2_tmp[2] = |
102 | 0 | CLIP_S16((74 * (pi2_src[0] - pi2_src[2 * src_strd] + pi2_src[3 * src_strd]) + add) >> shift); |
103 | 0 | pi2_tmp[3] = |
104 | 0 | CLIP_S16((55 * c[0] + 29 * c[2] - c[3] + add) >> shift); |
105 | 0 | } |
106 | 0 | pi2_src++; |
107 | 0 | pi2_tmp += trans_size; |
108 | 0 | zero_cols = zero_cols >> 1; |
109 | 0 | } |
110 | |
|
111 | 0 | pi2_tmp = pi2_tmp_orig; |
112 | | |
113 | | /* Inverse Transform 2nd stage */ |
114 | 0 | shift = IT_SHIFT_STAGE_2; |
115 | 0 | add = 1 << (shift - 1); |
116 | |
|
117 | 0 | for(i = 0; i < trans_size; i++) |
118 | 0 | { |
119 | 0 | WORD32 itrans_out; |
120 | | // Intermediate Variables |
121 | 0 | c[0] = pi2_tmp[0] + pi2_tmp[2 * trans_size]; |
122 | 0 | c[1] = pi2_tmp[2 * trans_size] + pi2_tmp[3 * trans_size]; |
123 | 0 | c[2] = pi2_tmp[0] - pi2_tmp[3 * trans_size]; |
124 | 0 | c[3] = 74 * pi2_tmp[trans_size]; |
125 | |
|
126 | 0 | pi2_dst[0] = |
127 | 0 | CLIP_S16((29 * c[0] + 55 * c[1] + c[3] + add) >> shift); |
128 | |
|
129 | 0 | pi2_dst[1] = |
130 | 0 | CLIP_S16((55 * c[2] - 29 * c[1] + c[3] + add) >> shift); |
131 | |
|
132 | 0 | pi2_dst[2] = |
133 | 0 | CLIP_S16((74 * (pi2_tmp[0] - pi2_tmp[2 * trans_size] + pi2_tmp[3 * trans_size]) + add) >> shift); |
134 | |
|
135 | 0 | pi2_dst[3] = |
136 | 0 | CLIP_S16((55 * c[0] + 29 * c[2] - c[3] + add) >> shift); |
137 | |
|
138 | 0 | pi2_tmp++; |
139 | 0 | pi2_dst += dst_strd; |
140 | 0 | } |
141 | 0 | } |
142 | | |
143 | | |
144 | | void ihevc_itrans_res_4x4(WORD16 *pi2_src, |
145 | | WORD16 *pi2_tmp, |
146 | | WORD16 *pi2_dst, |
147 | | WORD32 src_strd, |
148 | | WORD32 dst_strd, |
149 | | WORD32 zero_cols, |
150 | | WORD32 zero_rows) |
151 | | |
152 | 0 | { |
153 | 0 | WORD32 j; |
154 | 0 | WORD32 e[2], o[2]; |
155 | 0 | WORD32 add; |
156 | 0 | WORD32 shift; |
157 | 0 | WORD16 *pi2_tmp_orig; |
158 | 0 | WORD32 trans_size; |
159 | 0 | UNUSED(zero_rows); |
160 | 0 | trans_size = TRANS_SIZE_4; |
161 | |
|
162 | 0 | pi2_tmp_orig = pi2_tmp; |
163 | | |
164 | | /* Inverse Transform 1st stage */ |
165 | 0 | shift = IT_SHIFT_STAGE_1; |
166 | 0 | add = 1 << (shift - 1); |
167 | |
|
168 | 0 | for(j = 0; j < trans_size; j++) |
169 | 0 | { |
170 | | /* Checking for Zero Cols */ |
171 | 0 | if((zero_cols & 1) == 1) |
172 | 0 | { |
173 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
174 | 0 | } |
175 | 0 | else |
176 | 0 | { |
177 | | |
178 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
179 | 0 | o[0] = g_ai2_ihevc_trans_4[1][0] * pi2_src[src_strd] |
180 | 0 | + g_ai2_ihevc_trans_4[3][0] * pi2_src[3 * src_strd]; |
181 | 0 | o[1] = g_ai2_ihevc_trans_4[1][1] * pi2_src[src_strd] |
182 | 0 | + g_ai2_ihevc_trans_4[3][1] * pi2_src[3 * src_strd]; |
183 | 0 | e[0] = g_ai2_ihevc_trans_4[0][0] * pi2_src[0] |
184 | 0 | + g_ai2_ihevc_trans_4[2][0] * pi2_src[2 * src_strd]; |
185 | 0 | e[1] = g_ai2_ihevc_trans_4[0][1] * pi2_src[0] |
186 | 0 | + g_ai2_ihevc_trans_4[2][1] * pi2_src[2 * src_strd]; |
187 | |
|
188 | 0 | pi2_tmp[0] = |
189 | 0 | CLIP_S16(((e[0] + o[0] + add) >> shift)); |
190 | 0 | pi2_tmp[1] = |
191 | 0 | CLIP_S16(((e[1] + o[1] + add) >> shift)); |
192 | 0 | pi2_tmp[2] = |
193 | 0 | CLIP_S16(((e[1] - o[1] + add) >> shift)); |
194 | 0 | pi2_tmp[3] = |
195 | 0 | CLIP_S16(((e[0] - o[0] + add) >> shift)); |
196 | |
|
197 | 0 | } |
198 | 0 | pi2_src++; |
199 | 0 | pi2_tmp += trans_size; |
200 | 0 | zero_cols = zero_cols >> 1; |
201 | 0 | } |
202 | |
|
203 | 0 | pi2_tmp = pi2_tmp_orig; |
204 | | |
205 | | /* Inverse Transform 2nd stage */ |
206 | 0 | shift = IT_SHIFT_STAGE_2; |
207 | 0 | add = 1 << (shift - 1); |
208 | |
|
209 | 0 | for(j = 0; j < trans_size; j++) |
210 | 0 | { |
211 | 0 | WORD32 itrans_out; |
212 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
213 | 0 | o[0] = g_ai2_ihevc_trans_4[1][0] * pi2_tmp[trans_size] |
214 | 0 | + g_ai2_ihevc_trans_4[3][0] * pi2_tmp[3 * trans_size]; |
215 | 0 | o[1] = g_ai2_ihevc_trans_4[1][1] * pi2_tmp[trans_size] |
216 | 0 | + g_ai2_ihevc_trans_4[3][1] * pi2_tmp[3 * trans_size]; |
217 | 0 | e[0] = g_ai2_ihevc_trans_4[0][0] * pi2_tmp[0] |
218 | 0 | + g_ai2_ihevc_trans_4[2][0] * pi2_tmp[2 * trans_size]; |
219 | 0 | e[1] = g_ai2_ihevc_trans_4[0][1] * pi2_tmp[0] |
220 | 0 | + g_ai2_ihevc_trans_4[2][1] * pi2_tmp[2 * trans_size]; |
221 | |
|
222 | 0 | pi2_dst[0] = |
223 | 0 | CLIP_S16(((e[0] + o[0] + add) >> shift)); |
224 | |
|
225 | 0 | pi2_dst[1] = |
226 | 0 | CLIP_S16(((e[1] + o[1] + add) >> shift)); |
227 | |
|
228 | 0 | pi2_dst[2] = |
229 | 0 | CLIP_S16(((e[1] - o[1] + add) >> shift)); |
230 | |
|
231 | 0 | pi2_dst[3] = |
232 | 0 | CLIP_S16(((e[0] - o[0] + add) >> shift)); |
233 | |
|
234 | 0 | pi2_tmp++; |
235 | 0 | pi2_dst += dst_strd; |
236 | |
|
237 | 0 | } |
238 | 0 | } |
239 | | |
240 | | |
241 | | void ihevcd_itrans_res_dc(WORD16 *pi2_dst, |
242 | | WORD32 dst_strd, |
243 | | WORD32 log2_trans_size, |
244 | | WORD16 i2_coeff_value) |
245 | 0 | { |
246 | 0 | WORD32 row, col; |
247 | 0 | WORD32 add, shift; |
248 | 0 | WORD32 dc_value, quant_out; |
249 | 0 | WORD32 trans_size; |
250 | |
|
251 | 0 | trans_size = (1 << log2_trans_size); |
252 | |
|
253 | 0 | quant_out = i2_coeff_value; |
254 | |
|
255 | 0 | shift = IT_SHIFT_STAGE_1; |
256 | 0 | add = 1 << (shift - 1); |
257 | 0 | dc_value = CLIP_S16((quant_out * 64 + add) >> shift); |
258 | 0 | shift = IT_SHIFT_STAGE_2; |
259 | 0 | add = 1 << (shift - 1); |
260 | 0 | dc_value = CLIP_S16((dc_value * 64 + add) >> shift); |
261 | |
|
262 | 0 | for(row = 0; row < trans_size; row++) |
263 | 0 | for(col = 0; col < trans_size; col++) |
264 | 0 | pi2_dst[row * dst_strd + col] = dc_value; |
265 | |
|
266 | 0 | } |
267 | | |
268 | | |
269 | | void ihevc_itrans_res_8x8(WORD16 *pi2_src, |
270 | | WORD16 *pi2_tmp, |
271 | | WORD16 *pi2_dst, |
272 | | WORD32 src_strd, |
273 | | WORD32 dst_strd, |
274 | | WORD32 zero_cols, |
275 | | WORD32 zero_rows) |
276 | 0 | { |
277 | 0 | WORD32 j, k; |
278 | 0 | WORD32 e[4], o[4]; |
279 | 0 | WORD32 ee[2], eo[2]; |
280 | 0 | WORD32 add; |
281 | 0 | WORD32 shift; |
282 | 0 | WORD16 *pi2_tmp_orig; |
283 | 0 | WORD32 trans_size; |
284 | 0 | WORD32 zero_rows_2nd_stage = zero_cols; |
285 | 0 | WORD32 row_limit_2nd_stage; |
286 | |
|
287 | 0 | trans_size = TRANS_SIZE_8; |
288 | |
|
289 | 0 | pi2_tmp_orig = pi2_tmp; |
290 | |
|
291 | 0 | if((zero_cols & 0xF0) == 0xF0) |
292 | 0 | row_limit_2nd_stage = 4; |
293 | 0 | else |
294 | 0 | row_limit_2nd_stage = TRANS_SIZE_8; |
295 | | |
296 | |
|
297 | 0 | if((zero_rows & 0xF0) == 0xF0) /* First 4 rows of input are non-zero */ |
298 | 0 | { |
299 | | /************************************************************************************************/ |
300 | | /**********************************START - IT_RECON_8x8******************************************/ |
301 | | /************************************************************************************************/ |
302 | | |
303 | | /* Inverse Transform 1st stage */ |
304 | 0 | shift = IT_SHIFT_STAGE_1; |
305 | 0 | add = 1 << (shift - 1); |
306 | |
|
307 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
308 | 0 | { |
309 | | /* Checking for Zero Cols */ |
310 | 0 | if((zero_cols & 1) == 1) |
311 | 0 | { |
312 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
313 | 0 | } |
314 | 0 | else |
315 | 0 | { |
316 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
317 | 0 | for(k = 0; k < 4; k++) |
318 | 0 | { |
319 | 0 | o[k] = g_ai2_ihevc_trans_8[1][k] * pi2_src[src_strd] |
320 | 0 | + g_ai2_ihevc_trans_8[3][k] |
321 | 0 | * pi2_src[3 * src_strd]; |
322 | 0 | } |
323 | 0 | eo[0] = g_ai2_ihevc_trans_8[2][0] * pi2_src[2 * src_strd]; |
324 | 0 | eo[1] = g_ai2_ihevc_trans_8[2][1] * pi2_src[2 * src_strd]; |
325 | 0 | ee[0] = g_ai2_ihevc_trans_8[0][0] * pi2_src[0]; |
326 | 0 | ee[1] = g_ai2_ihevc_trans_8[0][1] * pi2_src[0]; |
327 | | |
328 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
329 | 0 | e[0] = ee[0] + eo[0]; |
330 | 0 | e[3] = ee[0] - eo[0]; |
331 | 0 | e[1] = ee[1] + eo[1]; |
332 | 0 | e[2] = ee[1] - eo[1]; |
333 | 0 | for(k = 0; k < 4; k++) |
334 | 0 | { |
335 | 0 | pi2_tmp[k] = |
336 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
337 | 0 | pi2_tmp[k + 4] = |
338 | 0 | CLIP_S16(((e[3 - k] - o[3 - k] + add) >> shift)); |
339 | 0 | } |
340 | 0 | } |
341 | 0 | pi2_src++; |
342 | 0 | pi2_tmp += trans_size; |
343 | 0 | zero_cols = zero_cols >> 1; |
344 | 0 | } |
345 | |
|
346 | 0 | pi2_tmp = pi2_tmp_orig; |
347 | | |
348 | | /* Inverse Transform 2nd stage */ |
349 | 0 | shift = IT_SHIFT_STAGE_2; |
350 | 0 | add = 1 << (shift - 1); |
351 | 0 | if((zero_rows_2nd_stage & 0xF0) == 0xF0) /* First 4 rows of output of 1st stage are non-zero */ |
352 | 0 | { |
353 | 0 | for(j = 0; j < trans_size; j++) |
354 | 0 | { |
355 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
356 | 0 | for(k = 0; k < 4; k++) |
357 | 0 | { |
358 | 0 | o[k] = g_ai2_ihevc_trans_8[1][k] * pi2_tmp[trans_size] |
359 | 0 | + g_ai2_ihevc_trans_8[3][k] * pi2_tmp[3 * trans_size]; |
360 | 0 | } |
361 | 0 | eo[0] = g_ai2_ihevc_trans_8[2][0] * pi2_tmp[2 * trans_size]; |
362 | 0 | eo[1] = g_ai2_ihevc_trans_8[2][1] * pi2_tmp[2 * trans_size]; |
363 | 0 | ee[0] = g_ai2_ihevc_trans_8[0][0] * pi2_tmp[0]; |
364 | 0 | ee[1] = g_ai2_ihevc_trans_8[0][1] * pi2_tmp[0]; |
365 | | |
366 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
367 | 0 | e[0] = ee[0] + eo[0]; |
368 | 0 | e[3] = ee[0] - eo[0]; |
369 | 0 | e[1] = ee[1] + eo[1]; |
370 | 0 | e[2] = ee[1] - eo[1]; |
371 | 0 | for(k = 0; k < 4; k++) |
372 | 0 | { |
373 | 0 | pi2_dst[k] = |
374 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
375 | |
|
376 | 0 | pi2_dst[k + 4] = |
377 | 0 | CLIP_S16(((e[3 - k] - o[3 - k] + add) >> shift)); |
378 | 0 | } |
379 | 0 | pi2_tmp++; |
380 | 0 | pi2_dst += dst_strd; |
381 | 0 | } |
382 | 0 | } |
383 | 0 | else /* All rows of output of 1st stage are non-zero */ |
384 | 0 | { |
385 | 0 | for(j = 0; j < trans_size; j++) |
386 | 0 | { |
387 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
388 | 0 | for(k = 0; k < 4; k++) |
389 | 0 | { |
390 | 0 | o[k] = g_ai2_ihevc_trans_8[1][k] * pi2_tmp[trans_size] |
391 | 0 | + g_ai2_ihevc_trans_8[3][k] |
392 | 0 | * pi2_tmp[3 * trans_size] |
393 | 0 | + g_ai2_ihevc_trans_8[5][k] |
394 | 0 | * pi2_tmp[5 * trans_size] |
395 | 0 | + g_ai2_ihevc_trans_8[7][k] |
396 | 0 | * pi2_tmp[7 * trans_size]; |
397 | 0 | } |
398 | |
|
399 | 0 | eo[0] = g_ai2_ihevc_trans_8[2][0] * pi2_tmp[2 * trans_size] |
400 | 0 | + g_ai2_ihevc_trans_8[6][0] * pi2_tmp[6 * trans_size]; |
401 | 0 | eo[1] = g_ai2_ihevc_trans_8[2][1] * pi2_tmp[2 * trans_size] |
402 | 0 | + g_ai2_ihevc_trans_8[6][1] * pi2_tmp[6 * trans_size]; |
403 | 0 | ee[0] = g_ai2_ihevc_trans_8[0][0] * pi2_tmp[0] |
404 | 0 | + g_ai2_ihevc_trans_8[4][0] * pi2_tmp[4 * trans_size]; |
405 | 0 | ee[1] = g_ai2_ihevc_trans_8[0][1] * pi2_tmp[0] |
406 | 0 | + g_ai2_ihevc_trans_8[4][1] * pi2_tmp[4 * trans_size]; |
407 | | |
408 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
409 | 0 | e[0] = ee[0] + eo[0]; |
410 | 0 | e[3] = ee[0] - eo[0]; |
411 | 0 | e[1] = ee[1] + eo[1]; |
412 | 0 | e[2] = ee[1] - eo[1]; |
413 | 0 | for(k = 0; k < 4; k++) |
414 | 0 | { |
415 | 0 | pi2_dst[k] = |
416 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
417 | |
|
418 | 0 | pi2_dst[k + 4] = |
419 | 0 | CLIP_S16(((e[3 - k] - o[3 - k] + add) >> shift)); |
420 | 0 | } |
421 | 0 | pi2_tmp++; |
422 | 0 | pi2_dst += dst_strd; |
423 | 0 | } |
424 | 0 | } |
425 | | /************************************************************************************************/ |
426 | | /************************************END - IT_RECON_8x8******************************************/ |
427 | | /************************************************************************************************/ |
428 | 0 | } |
429 | 0 | else /* All rows of input are non-zero */ |
430 | 0 | { |
431 | | /************************************************************************************************/ |
432 | | /**********************************START - IT_RECON_8x8******************************************/ |
433 | | /************************************************************************************************/ |
434 | | |
435 | | /* Inverse Transform 1st stage */ |
436 | 0 | shift = IT_SHIFT_STAGE_1; |
437 | 0 | add = 1 << (shift - 1); |
438 | |
|
439 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
440 | 0 | { |
441 | | /* Checking for Zero Cols */ |
442 | 0 | if((zero_cols & 1) == 1) |
443 | 0 | { |
444 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
445 | 0 | } |
446 | 0 | else |
447 | 0 | { |
448 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
449 | 0 | for(k = 0; k < 4; k++) |
450 | 0 | { |
451 | 0 | o[k] = g_ai2_ihevc_trans_8[1][k] * pi2_src[src_strd] |
452 | 0 | + g_ai2_ihevc_trans_8[3][k] |
453 | 0 | * pi2_src[3 * src_strd] |
454 | 0 | + g_ai2_ihevc_trans_8[5][k] |
455 | 0 | * pi2_src[5 * src_strd] |
456 | 0 | + g_ai2_ihevc_trans_8[7][k] |
457 | 0 | * pi2_src[7 * src_strd]; |
458 | 0 | } |
459 | |
|
460 | 0 | eo[0] = g_ai2_ihevc_trans_8[2][0] * pi2_src[2 * src_strd] |
461 | 0 | + g_ai2_ihevc_trans_8[6][0] * pi2_src[6 * src_strd]; |
462 | 0 | eo[1] = g_ai2_ihevc_trans_8[2][1] * pi2_src[2 * src_strd] |
463 | 0 | + g_ai2_ihevc_trans_8[6][1] * pi2_src[6 * src_strd]; |
464 | 0 | ee[0] = g_ai2_ihevc_trans_8[0][0] * pi2_src[0] |
465 | 0 | + g_ai2_ihevc_trans_8[4][0] * pi2_src[4 * src_strd]; |
466 | 0 | ee[1] = g_ai2_ihevc_trans_8[0][1] * pi2_src[0] |
467 | 0 | + g_ai2_ihevc_trans_8[4][1] * pi2_src[4 * src_strd]; |
468 | | |
469 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
470 | 0 | e[0] = ee[0] + eo[0]; |
471 | 0 | e[3] = ee[0] - eo[0]; |
472 | 0 | e[1] = ee[1] + eo[1]; |
473 | 0 | e[2] = ee[1] - eo[1]; |
474 | 0 | for(k = 0; k < 4; k++) |
475 | 0 | { |
476 | 0 | pi2_tmp[k] = |
477 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
478 | 0 | pi2_tmp[k + 4] = |
479 | 0 | CLIP_S16(((e[3 - k] - o[3 - k] + add) >> shift)); |
480 | 0 | } |
481 | 0 | } |
482 | 0 | pi2_src++; |
483 | 0 | pi2_tmp += trans_size; |
484 | 0 | zero_cols = zero_cols >> 1; |
485 | 0 | } |
486 | |
|
487 | 0 | pi2_tmp = pi2_tmp_orig; |
488 | | |
489 | | /* Inverse Transform 2nd stage */ |
490 | 0 | shift = IT_SHIFT_STAGE_2; |
491 | 0 | add = 1 << (shift - 1); |
492 | 0 | if((zero_rows_2nd_stage & 0xF0) == 0xF0) /* First 4 rows of output of 1st stage are non-zero */ |
493 | 0 | { |
494 | 0 | for(j = 0; j < trans_size; j++) |
495 | 0 | { |
496 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
497 | 0 | for(k = 0; k < 4; k++) |
498 | 0 | { |
499 | 0 | o[k] = g_ai2_ihevc_trans_8[1][k] * pi2_tmp[trans_size] |
500 | 0 | + g_ai2_ihevc_trans_8[3][k] * pi2_tmp[3 * trans_size]; |
501 | 0 | } |
502 | 0 | eo[0] = g_ai2_ihevc_trans_8[2][0] * pi2_tmp[2 * trans_size]; |
503 | 0 | eo[1] = g_ai2_ihevc_trans_8[2][1] * pi2_tmp[2 * trans_size]; |
504 | 0 | ee[0] = g_ai2_ihevc_trans_8[0][0] * pi2_tmp[0]; |
505 | 0 | ee[1] = g_ai2_ihevc_trans_8[0][1] * pi2_tmp[0]; |
506 | | |
507 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
508 | 0 | e[0] = ee[0] + eo[0]; |
509 | 0 | e[3] = ee[0] - eo[0]; |
510 | 0 | e[1] = ee[1] + eo[1]; |
511 | 0 | e[2] = ee[1] - eo[1]; |
512 | 0 | for(k = 0; k < 4; k++) |
513 | 0 | { |
514 | 0 | pi2_dst[k] = |
515 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
516 | |
|
517 | 0 | pi2_dst[k + 4] = |
518 | 0 | CLIP_S16(((e[3 - k] - o[3 - k] + add) >> shift)); |
519 | 0 | } |
520 | 0 | pi2_tmp++; |
521 | 0 | pi2_dst += dst_strd; |
522 | 0 | } |
523 | 0 | } |
524 | 0 | else /* All rows of output of 1st stage are non-zero */ |
525 | 0 | { |
526 | 0 | for(j = 0; j < trans_size; j++) |
527 | 0 | { |
528 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
529 | 0 | for(k = 0; k < 4; k++) |
530 | 0 | { |
531 | 0 | o[k] = g_ai2_ihevc_trans_8[1][k] * pi2_tmp[trans_size] |
532 | 0 | + g_ai2_ihevc_trans_8[3][k] |
533 | 0 | * pi2_tmp[3 * trans_size] |
534 | 0 | + g_ai2_ihevc_trans_8[5][k] |
535 | 0 | * pi2_tmp[5 * trans_size] |
536 | 0 | + g_ai2_ihevc_trans_8[7][k] |
537 | 0 | * pi2_tmp[7 * trans_size]; |
538 | 0 | } |
539 | |
|
540 | 0 | eo[0] = g_ai2_ihevc_trans_8[2][0] * pi2_tmp[2 * trans_size] |
541 | 0 | + g_ai2_ihevc_trans_8[6][0] * pi2_tmp[6 * trans_size]; |
542 | 0 | eo[1] = g_ai2_ihevc_trans_8[2][1] * pi2_tmp[2 * trans_size] |
543 | 0 | + g_ai2_ihevc_trans_8[6][1] * pi2_tmp[6 * trans_size]; |
544 | 0 | ee[0] = g_ai2_ihevc_trans_8[0][0] * pi2_tmp[0] |
545 | 0 | + g_ai2_ihevc_trans_8[4][0] * pi2_tmp[4 * trans_size]; |
546 | 0 | ee[1] = g_ai2_ihevc_trans_8[0][1] * pi2_tmp[0] |
547 | 0 | + g_ai2_ihevc_trans_8[4][1] * pi2_tmp[4 * trans_size]; |
548 | | |
549 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
550 | 0 | e[0] = ee[0] + eo[0]; |
551 | 0 | e[3] = ee[0] - eo[0]; |
552 | 0 | e[1] = ee[1] + eo[1]; |
553 | 0 | e[2] = ee[1] - eo[1]; |
554 | 0 | for(k = 0; k < 4; k++) |
555 | 0 | { |
556 | 0 | pi2_dst[k] = |
557 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
558 | |
|
559 | 0 | pi2_dst[k + 4] = |
560 | 0 | CLIP_S16(((e[3 - k] - o[3 - k] + add) >> shift)); |
561 | 0 | } |
562 | 0 | pi2_tmp++; |
563 | 0 | pi2_dst += dst_strd; |
564 | 0 | } |
565 | 0 | } |
566 | | /************************************************************************************************/ |
567 | | /************************************END - IT_RECON_8x8******************************************/ |
568 | | /************************************************************************************************/ |
569 | 0 | } |
570 | 0 | } |
571 | | |
572 | | |
573 | | void ihevc_itrans_res_16x16(WORD16 *pi2_src, |
574 | | WORD16 *pi2_tmp, |
575 | | WORD16 *pi2_dst, |
576 | | WORD32 src_strd, |
577 | | WORD32 dst_strd, |
578 | | WORD32 zero_cols, |
579 | | WORD32 zero_rows) |
580 | 0 | { |
581 | 0 | WORD32 j, k; |
582 | 0 | WORD32 e[8], o[8]; |
583 | 0 | WORD32 ee[4], eo[4]; |
584 | 0 | WORD32 eee[2], eeo[2]; |
585 | 0 | WORD32 add; |
586 | 0 | WORD32 shift; |
587 | 0 | WORD16 *pi2_tmp_orig; |
588 | 0 | WORD32 trans_size; |
589 | 0 | WORD32 zero_rows_2nd_stage = zero_cols; |
590 | 0 | WORD32 row_limit_2nd_stage; |
591 | |
|
592 | 0 | if((zero_cols & 0xFFF0) == 0xFFF0) |
593 | 0 | row_limit_2nd_stage = 4; |
594 | 0 | else if((zero_cols & 0xFF00) == 0xFF00) |
595 | 0 | row_limit_2nd_stage = 8; |
596 | 0 | else |
597 | 0 | row_limit_2nd_stage = TRANS_SIZE_16; |
598 | |
|
599 | 0 | trans_size = TRANS_SIZE_16; |
600 | 0 | pi2_tmp_orig = pi2_tmp; |
601 | 0 | if((zero_rows & 0xFFF0) == 0xFFF0) /* First 4 rows of input are non-zero */ |
602 | 0 | { |
603 | | /* Inverse Transform 1st stage */ |
604 | | /************************************************************************************************/ |
605 | | /**********************************START - IT_RECON_16x16****************************************/ |
606 | | /************************************************************************************************/ |
607 | |
|
608 | 0 | shift = IT_SHIFT_STAGE_1; |
609 | 0 | add = 1 << (shift - 1); |
610 | |
|
611 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
612 | 0 | { |
613 | | /* Checking for Zero Cols */ |
614 | 0 | if((zero_cols & 1) == 1) |
615 | 0 | { |
616 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
617 | 0 | } |
618 | 0 | else |
619 | 0 | { |
620 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
621 | 0 | for(k = 0; k < 8; k++) |
622 | 0 | { |
623 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_src[src_strd] |
624 | 0 | + g_ai2_ihevc_trans_16[3][k] |
625 | 0 | * pi2_src[3 * src_strd]; |
626 | 0 | } |
627 | 0 | for(k = 0; k < 4; k++) |
628 | 0 | { |
629 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_src[2 * src_strd]; |
630 | 0 | } |
631 | 0 | eeo[0] = 0; |
632 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_src[0]; |
633 | 0 | eeo[1] = 0; |
634 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_src[0]; |
635 | | |
636 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
637 | 0 | for(k = 0; k < 2; k++) |
638 | 0 | { |
639 | 0 | ee[k] = eee[k] + eeo[k]; |
640 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
641 | 0 | } |
642 | 0 | for(k = 0; k < 4; k++) |
643 | 0 | { |
644 | 0 | e[k] = ee[k] + eo[k]; |
645 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
646 | 0 | } |
647 | 0 | for(k = 0; k < 8; k++) |
648 | 0 | { |
649 | 0 | pi2_tmp[k] = |
650 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
651 | 0 | pi2_tmp[k + 8] = |
652 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
653 | 0 | } |
654 | 0 | } |
655 | 0 | pi2_src++; |
656 | 0 | pi2_tmp += trans_size; |
657 | 0 | zero_cols = zero_cols >> 1; |
658 | 0 | } |
659 | |
|
660 | 0 | pi2_tmp = pi2_tmp_orig; |
661 | | |
662 | | /* Inverse Transform 2nd stage */ |
663 | 0 | shift = IT_SHIFT_STAGE_2; |
664 | 0 | add = 1 << (shift - 1); |
665 | |
|
666 | 0 | if((zero_rows_2nd_stage & 0xFFF0) == 0xFFF0) /* First 4 rows of output of 1st stage are non-zero */ |
667 | 0 | { |
668 | 0 | for(j = 0; j < trans_size; j++) |
669 | 0 | { |
670 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
671 | 0 | for(k = 0; k < 8; k++) |
672 | 0 | { |
673 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
674 | 0 | + g_ai2_ihevc_trans_16[3][k] |
675 | 0 | * pi2_tmp[3 * trans_size]; |
676 | 0 | } |
677 | 0 | for(k = 0; k < 4; k++) |
678 | 0 | { |
679 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size]; |
680 | 0 | } |
681 | 0 | eeo[0] = 0; |
682 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0]; |
683 | 0 | eeo[1] = 0; |
684 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0]; |
685 | | |
686 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
687 | 0 | for(k = 0; k < 2; k++) |
688 | 0 | { |
689 | 0 | ee[k] = eee[k] + eeo[k]; |
690 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
691 | 0 | } |
692 | 0 | for(k = 0; k < 4; k++) |
693 | 0 | { |
694 | 0 | e[k] = ee[k] + eo[k]; |
695 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
696 | 0 | } |
697 | 0 | for(k = 0; k < 8; k++) |
698 | 0 | { |
699 | 0 | pi2_dst[k] = |
700 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
701 | |
|
702 | 0 | pi2_dst[k + 8] = |
703 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
704 | 0 | } |
705 | 0 | pi2_tmp++; |
706 | 0 | pi2_dst += dst_strd; |
707 | 0 | } |
708 | 0 | } |
709 | 0 | else if((zero_rows_2nd_stage & 0xFF00) == 0xFF00) /* First 4 rows of output of 1st stage are non-zero */ |
710 | 0 | { |
711 | 0 | for(j = 0; j < trans_size; j++) |
712 | 0 | { |
713 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
714 | 0 | for(k = 0; k < 8; k++) |
715 | 0 | { |
716 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
717 | 0 | + g_ai2_ihevc_trans_16[3][k] |
718 | 0 | * pi2_tmp[3 * trans_size] |
719 | 0 | + g_ai2_ihevc_trans_16[5][k] |
720 | 0 | * pi2_tmp[5 * trans_size] |
721 | 0 | + g_ai2_ihevc_trans_16[7][k] |
722 | 0 | * pi2_tmp[7 * trans_size]; |
723 | 0 | } |
724 | 0 | for(k = 0; k < 4; k++) |
725 | 0 | { |
726 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size] |
727 | 0 | + g_ai2_ihevc_trans_16[6][k] |
728 | 0 | * pi2_tmp[6 * trans_size]; |
729 | 0 | } |
730 | 0 | eeo[0] = g_ai2_ihevc_trans_16[4][0] * pi2_tmp[4 * trans_size]; |
731 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0]; |
732 | 0 | eeo[1] = g_ai2_ihevc_trans_16[4][1] * pi2_tmp[4 * trans_size]; |
733 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0]; |
734 | | |
735 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
736 | 0 | for(k = 0; k < 2; k++) |
737 | 0 | { |
738 | 0 | ee[k] = eee[k] + eeo[k]; |
739 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
740 | 0 | } |
741 | 0 | for(k = 0; k < 4; k++) |
742 | 0 | { |
743 | 0 | e[k] = ee[k] + eo[k]; |
744 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
745 | 0 | } |
746 | 0 | for(k = 0; k < 8; k++) |
747 | 0 | { |
748 | 0 | pi2_dst[k] = |
749 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
750 | |
|
751 | 0 | pi2_dst[k + 8] = |
752 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
753 | 0 | } |
754 | 0 | pi2_tmp++; |
755 | 0 | pi2_dst += dst_strd; |
756 | 0 | } |
757 | 0 | } |
758 | 0 | else /* All rows of output of 1st stage are non-zero */ |
759 | 0 | { |
760 | 0 | for(j = 0; j < trans_size; j++) |
761 | 0 | { |
762 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
763 | 0 | for(k = 0; k < 8; k++) |
764 | 0 | { |
765 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
766 | 0 | + g_ai2_ihevc_trans_16[3][k] |
767 | 0 | * pi2_tmp[3 * trans_size] |
768 | 0 | + g_ai2_ihevc_trans_16[5][k] |
769 | 0 | * pi2_tmp[5 * trans_size] |
770 | 0 | + g_ai2_ihevc_trans_16[7][k] |
771 | 0 | * pi2_tmp[7 * trans_size] |
772 | 0 | + g_ai2_ihevc_trans_16[9][k] |
773 | 0 | * pi2_tmp[9 * trans_size] |
774 | 0 | + g_ai2_ihevc_trans_16[11][k] |
775 | 0 | * pi2_tmp[11 * trans_size] |
776 | 0 | + g_ai2_ihevc_trans_16[13][k] |
777 | 0 | * pi2_tmp[13 * trans_size] |
778 | 0 | + g_ai2_ihevc_trans_16[15][k] |
779 | 0 | * pi2_tmp[15 * trans_size]; |
780 | 0 | } |
781 | 0 | for(k = 0; k < 4; k++) |
782 | 0 | { |
783 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size] |
784 | 0 | + g_ai2_ihevc_trans_16[6][k] |
785 | 0 | * pi2_tmp[6 * trans_size] |
786 | 0 | + g_ai2_ihevc_trans_16[10][k] |
787 | 0 | * pi2_tmp[10 * trans_size] |
788 | 0 | + g_ai2_ihevc_trans_16[14][k] |
789 | 0 | * pi2_tmp[14 * trans_size]; |
790 | 0 | } |
791 | 0 | eeo[0] = |
792 | 0 | g_ai2_ihevc_trans_16[4][0] * pi2_tmp[4 * trans_size] |
793 | 0 | + g_ai2_ihevc_trans_16[12][0] |
794 | 0 | * pi2_tmp[12 |
795 | 0 | * trans_size]; |
796 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0] |
797 | 0 | + g_ai2_ihevc_trans_16[8][0] * pi2_tmp[8 * trans_size]; |
798 | 0 | eeo[1] = |
799 | 0 | g_ai2_ihevc_trans_16[4][1] * pi2_tmp[4 * trans_size] |
800 | 0 | + g_ai2_ihevc_trans_16[12][1] |
801 | 0 | * pi2_tmp[12 |
802 | 0 | * trans_size]; |
803 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0] |
804 | 0 | + g_ai2_ihevc_trans_16[8][1] * pi2_tmp[8 * trans_size]; |
805 | | |
806 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
807 | 0 | for(k = 0; k < 2; k++) |
808 | 0 | { |
809 | 0 | ee[k] = eee[k] + eeo[k]; |
810 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
811 | 0 | } |
812 | 0 | for(k = 0; k < 4; k++) |
813 | 0 | { |
814 | 0 | e[k] = ee[k] + eo[k]; |
815 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
816 | 0 | } |
817 | 0 | for(k = 0; k < 8; k++) |
818 | 0 | { |
819 | 0 | pi2_dst[k] = |
820 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
821 | |
|
822 | 0 | pi2_dst[k + 8] = |
823 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
824 | 0 | } |
825 | 0 | pi2_tmp++; |
826 | 0 | pi2_dst += dst_strd; |
827 | 0 | } |
828 | 0 | } |
829 | | /************************************************************************************************/ |
830 | | /************************************END - IT_RECON_16x16****************************************/ |
831 | | /************************************************************************************************/ |
832 | 0 | } |
833 | 0 | else if((zero_rows & 0xFF00) == 0xFF00) /* First 8 rows of input are non-zero */ |
834 | 0 | { |
835 | | /* Inverse Transform 1st stage */ |
836 | | /************************************************************************************************/ |
837 | | /**********************************START - IT_RECON_16x16****************************************/ |
838 | | /************************************************************************************************/ |
839 | |
|
840 | 0 | shift = IT_SHIFT_STAGE_1; |
841 | 0 | add = 1 << (shift - 1); |
842 | |
|
843 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
844 | 0 | { |
845 | | /* Checking for Zero Cols */ |
846 | 0 | if((zero_cols & 1) == 1) |
847 | 0 | { |
848 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
849 | 0 | } |
850 | 0 | else |
851 | 0 | { |
852 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
853 | 0 | for(k = 0; k < 8; k++) |
854 | 0 | { |
855 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_src[src_strd] |
856 | 0 | + g_ai2_ihevc_trans_16[3][k] |
857 | 0 | * pi2_src[3 * src_strd] |
858 | 0 | + g_ai2_ihevc_trans_16[5][k] |
859 | 0 | * pi2_src[5 * src_strd] |
860 | 0 | + g_ai2_ihevc_trans_16[7][k] |
861 | 0 | * pi2_src[7 * src_strd]; |
862 | 0 | } |
863 | 0 | for(k = 0; k < 4; k++) |
864 | 0 | { |
865 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_src[2 * src_strd] |
866 | 0 | + g_ai2_ihevc_trans_16[6][k] |
867 | 0 | * pi2_src[6 * src_strd]; |
868 | 0 | } |
869 | 0 | eeo[0] = g_ai2_ihevc_trans_16[4][0] * pi2_src[4 * src_strd]; |
870 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_src[0]; |
871 | 0 | eeo[1] = g_ai2_ihevc_trans_16[4][1] * pi2_src[4 * src_strd]; |
872 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_src[0]; |
873 | | |
874 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
875 | 0 | for(k = 0; k < 2; k++) |
876 | 0 | { |
877 | 0 | ee[k] = eee[k] + eeo[k]; |
878 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
879 | 0 | } |
880 | 0 | for(k = 0; k < 4; k++) |
881 | 0 | { |
882 | 0 | e[k] = ee[k] + eo[k]; |
883 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
884 | 0 | } |
885 | 0 | for(k = 0; k < 8; k++) |
886 | 0 | { |
887 | 0 | pi2_tmp[k] = |
888 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
889 | 0 | pi2_tmp[k + 8] = |
890 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
891 | 0 | } |
892 | 0 | } |
893 | 0 | pi2_src++; |
894 | 0 | pi2_tmp += trans_size; |
895 | 0 | zero_cols = zero_cols >> 1; |
896 | 0 | } |
897 | |
|
898 | 0 | pi2_tmp = pi2_tmp_orig; |
899 | | |
900 | | /* Inverse Transform 2nd stage */ |
901 | 0 | shift = IT_SHIFT_STAGE_2; |
902 | 0 | add = 1 << (shift - 1); |
903 | |
|
904 | 0 | if((zero_rows_2nd_stage & 0xFFF0) == 0xFFF0) /* First 4 rows of output of 1st stage are non-zero */ |
905 | 0 | { |
906 | 0 | for(j = 0; j < trans_size; j++) |
907 | 0 | { |
908 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
909 | 0 | for(k = 0; k < 8; k++) |
910 | 0 | { |
911 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
912 | 0 | + g_ai2_ihevc_trans_16[3][k] |
913 | 0 | * pi2_tmp[3 * trans_size]; |
914 | 0 | } |
915 | 0 | for(k = 0; k < 4; k++) |
916 | 0 | { |
917 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size]; |
918 | 0 | } |
919 | 0 | eeo[0] = 0; |
920 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0]; |
921 | 0 | eeo[1] = 0; |
922 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0]; |
923 | | |
924 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
925 | 0 | for(k = 0; k < 2; k++) |
926 | 0 | { |
927 | 0 | ee[k] = eee[k] + eeo[k]; |
928 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
929 | 0 | } |
930 | 0 | for(k = 0; k < 4; k++) |
931 | 0 | { |
932 | 0 | e[k] = ee[k] + eo[k]; |
933 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
934 | 0 | } |
935 | 0 | for(k = 0; k < 8; k++) |
936 | 0 | { |
937 | 0 | pi2_dst[k] = |
938 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
939 | |
|
940 | 0 | pi2_dst[k + 8] = |
941 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
942 | 0 | } |
943 | 0 | pi2_tmp++; |
944 | 0 | pi2_dst += dst_strd; |
945 | 0 | } |
946 | 0 | } |
947 | 0 | else if((zero_rows_2nd_stage & 0xFF00) == 0xFF00) /* First 4 rows of output of 1st stage are non-zero */ |
948 | 0 | { |
949 | 0 | for(j = 0; j < trans_size; j++) |
950 | 0 | { |
951 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
952 | 0 | for(k = 0; k < 8; k++) |
953 | 0 | { |
954 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
955 | 0 | + g_ai2_ihevc_trans_16[3][k] |
956 | 0 | * pi2_tmp[3 * trans_size] |
957 | 0 | + g_ai2_ihevc_trans_16[5][k] |
958 | 0 | * pi2_tmp[5 * trans_size] |
959 | 0 | + g_ai2_ihevc_trans_16[7][k] |
960 | 0 | * pi2_tmp[7 * trans_size]; |
961 | 0 | } |
962 | 0 | for(k = 0; k < 4; k++) |
963 | 0 | { |
964 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size] |
965 | 0 | + g_ai2_ihevc_trans_16[6][k] |
966 | 0 | * pi2_tmp[6 * trans_size]; |
967 | 0 | } |
968 | 0 | eeo[0] = g_ai2_ihevc_trans_16[4][0] * pi2_tmp[4 * trans_size]; |
969 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0]; |
970 | 0 | eeo[1] = g_ai2_ihevc_trans_16[4][1] * pi2_tmp[4 * trans_size]; |
971 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0]; |
972 | | |
973 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
974 | 0 | for(k = 0; k < 2; k++) |
975 | 0 | { |
976 | 0 | ee[k] = eee[k] + eeo[k]; |
977 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
978 | 0 | } |
979 | 0 | for(k = 0; k < 4; k++) |
980 | 0 | { |
981 | 0 | e[k] = ee[k] + eo[k]; |
982 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
983 | 0 | } |
984 | 0 | for(k = 0; k < 8; k++) |
985 | 0 | { |
986 | 0 | pi2_dst[k] = |
987 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
988 | |
|
989 | 0 | pi2_dst[k + 8] = |
990 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
991 | 0 | } |
992 | 0 | pi2_tmp++; |
993 | 0 | pi2_dst += dst_strd; |
994 | 0 | } |
995 | 0 | } |
996 | 0 | else /* All rows of output of 1st stage are non-zero */ |
997 | 0 | { |
998 | 0 | for(j = 0; j < trans_size; j++) |
999 | 0 | { |
1000 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1001 | 0 | for(k = 0; k < 8; k++) |
1002 | 0 | { |
1003 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
1004 | 0 | + g_ai2_ihevc_trans_16[3][k] |
1005 | 0 | * pi2_tmp[3 * trans_size] |
1006 | 0 | + g_ai2_ihevc_trans_16[5][k] |
1007 | 0 | * pi2_tmp[5 * trans_size] |
1008 | 0 | + g_ai2_ihevc_trans_16[7][k] |
1009 | 0 | * pi2_tmp[7 * trans_size] |
1010 | 0 | + g_ai2_ihevc_trans_16[9][k] |
1011 | 0 | * pi2_tmp[9 * trans_size] |
1012 | 0 | + g_ai2_ihevc_trans_16[11][k] |
1013 | 0 | * pi2_tmp[11 * trans_size] |
1014 | 0 | + g_ai2_ihevc_trans_16[13][k] |
1015 | 0 | * pi2_tmp[13 * trans_size] |
1016 | 0 | + g_ai2_ihevc_trans_16[15][k] |
1017 | 0 | * pi2_tmp[15 * trans_size]; |
1018 | 0 | } |
1019 | 0 | for(k = 0; k < 4; k++) |
1020 | 0 | { |
1021 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size] |
1022 | 0 | + g_ai2_ihevc_trans_16[6][k] |
1023 | 0 | * pi2_tmp[6 * trans_size] |
1024 | 0 | + g_ai2_ihevc_trans_16[10][k] |
1025 | 0 | * pi2_tmp[10 * trans_size] |
1026 | 0 | + g_ai2_ihevc_trans_16[14][k] |
1027 | 0 | * pi2_tmp[14 * trans_size]; |
1028 | 0 | } |
1029 | 0 | eeo[0] = |
1030 | 0 | g_ai2_ihevc_trans_16[4][0] * pi2_tmp[4 * trans_size] |
1031 | 0 | + g_ai2_ihevc_trans_16[12][0] |
1032 | 0 | * pi2_tmp[12 |
1033 | 0 | * trans_size]; |
1034 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0] |
1035 | 0 | + g_ai2_ihevc_trans_16[8][0] * pi2_tmp[8 * trans_size]; |
1036 | 0 | eeo[1] = |
1037 | 0 | g_ai2_ihevc_trans_16[4][1] * pi2_tmp[4 * trans_size] |
1038 | 0 | + g_ai2_ihevc_trans_16[12][1] |
1039 | 0 | * pi2_tmp[12 |
1040 | 0 | * trans_size]; |
1041 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0] |
1042 | 0 | + g_ai2_ihevc_trans_16[8][1] * pi2_tmp[8 * trans_size]; |
1043 | | |
1044 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1045 | 0 | for(k = 0; k < 2; k++) |
1046 | 0 | { |
1047 | 0 | ee[k] = eee[k] + eeo[k]; |
1048 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
1049 | 0 | } |
1050 | 0 | for(k = 0; k < 4; k++) |
1051 | 0 | { |
1052 | 0 | e[k] = ee[k] + eo[k]; |
1053 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
1054 | 0 | } |
1055 | 0 | for(k = 0; k < 8; k++) |
1056 | 0 | { |
1057 | 0 | pi2_dst[k] = |
1058 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1059 | |
|
1060 | 0 | pi2_dst[k + 8] = |
1061 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
1062 | 0 | } |
1063 | 0 | pi2_tmp++; |
1064 | 0 | pi2_dst += dst_strd; |
1065 | 0 | } |
1066 | 0 | } |
1067 | | /************************************************************************************************/ |
1068 | | /************************************END - IT_RECON_16x16****************************************/ |
1069 | | /************************************************************************************************/ |
1070 | 0 | } |
1071 | 0 | else /* All rows of input are non-zero */ |
1072 | 0 | { |
1073 | | /* Inverse Transform 1st stage */ |
1074 | | /************************************************************************************************/ |
1075 | | /**********************************START - IT_RECON_16x16****************************************/ |
1076 | | /************************************************************************************************/ |
1077 | |
|
1078 | 0 | shift = IT_SHIFT_STAGE_1; |
1079 | 0 | add = 1 << (shift - 1); |
1080 | |
|
1081 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
1082 | 0 | { |
1083 | | /* Checking for Zero Cols */ |
1084 | 0 | if((zero_cols & 1) == 1) |
1085 | 0 | { |
1086 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
1087 | 0 | } |
1088 | 0 | else |
1089 | 0 | { |
1090 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1091 | 0 | for(k = 0; k < 8; k++) |
1092 | 0 | { |
1093 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_src[src_strd] |
1094 | 0 | + g_ai2_ihevc_trans_16[3][k] |
1095 | 0 | * pi2_src[3 * src_strd] |
1096 | 0 | + g_ai2_ihevc_trans_16[5][k] |
1097 | 0 | * pi2_src[5 * src_strd] |
1098 | 0 | + g_ai2_ihevc_trans_16[7][k] |
1099 | 0 | * pi2_src[7 * src_strd] |
1100 | 0 | + g_ai2_ihevc_trans_16[9][k] |
1101 | 0 | * pi2_src[9 * src_strd] |
1102 | 0 | + g_ai2_ihevc_trans_16[11][k] |
1103 | 0 | * pi2_src[11 * src_strd] |
1104 | 0 | + g_ai2_ihevc_trans_16[13][k] |
1105 | 0 | * pi2_src[13 * src_strd] |
1106 | 0 | + g_ai2_ihevc_trans_16[15][k] |
1107 | 0 | * pi2_src[15 * src_strd]; |
1108 | 0 | } |
1109 | 0 | for(k = 0; k < 4; k++) |
1110 | 0 | { |
1111 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_src[2 * src_strd] |
1112 | 0 | + g_ai2_ihevc_trans_16[6][k] |
1113 | 0 | * pi2_src[6 * src_strd] |
1114 | 0 | + g_ai2_ihevc_trans_16[10][k] |
1115 | 0 | * pi2_src[10 * src_strd] |
1116 | 0 | + g_ai2_ihevc_trans_16[14][k] |
1117 | 0 | * pi2_src[14 * src_strd]; |
1118 | 0 | } |
1119 | 0 | eeo[0] = g_ai2_ihevc_trans_16[4][0] * pi2_src[4 * src_strd] |
1120 | 0 | + g_ai2_ihevc_trans_16[12][0] |
1121 | 0 | * pi2_src[12 * src_strd]; |
1122 | 0 | eee[0] = |
1123 | 0 | g_ai2_ihevc_trans_16[0][0] * pi2_src[0] |
1124 | 0 | + g_ai2_ihevc_trans_16[8][0] |
1125 | 0 | * pi2_src[8 |
1126 | 0 | * src_strd]; |
1127 | 0 | eeo[1] = g_ai2_ihevc_trans_16[4][1] * pi2_src[4 * src_strd] |
1128 | 0 | + g_ai2_ihevc_trans_16[12][1] |
1129 | 0 | * pi2_src[12 * src_strd]; |
1130 | 0 | eee[1] = |
1131 | 0 | g_ai2_ihevc_trans_16[0][1] * pi2_src[0] |
1132 | 0 | + g_ai2_ihevc_trans_16[8][1] |
1133 | 0 | * pi2_src[8 |
1134 | 0 | * src_strd]; |
1135 | | |
1136 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1137 | 0 | for(k = 0; k < 2; k++) |
1138 | 0 | { |
1139 | 0 | ee[k] = eee[k] + eeo[k]; |
1140 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
1141 | 0 | } |
1142 | 0 | for(k = 0; k < 4; k++) |
1143 | 0 | { |
1144 | 0 | e[k] = ee[k] + eo[k]; |
1145 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
1146 | 0 | } |
1147 | 0 | for(k = 0; k < 8; k++) |
1148 | 0 | { |
1149 | 0 | pi2_tmp[k] = |
1150 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1151 | 0 | pi2_tmp[k + 8] = |
1152 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
1153 | 0 | } |
1154 | 0 | } |
1155 | 0 | pi2_src++; |
1156 | 0 | pi2_tmp += trans_size; |
1157 | 0 | zero_cols = zero_cols >> 1; |
1158 | 0 | } |
1159 | |
|
1160 | 0 | pi2_tmp = pi2_tmp_orig; |
1161 | | |
1162 | | /* Inverse Transform 2nd stage */ |
1163 | 0 | shift = IT_SHIFT_STAGE_2; |
1164 | 0 | add = 1 << (shift - 1); |
1165 | |
|
1166 | 0 | if((zero_rows_2nd_stage & 0xFFF0) == 0xFFF0) /* First 4 rows of output of 1st stage are non-zero */ |
1167 | 0 | { |
1168 | 0 | for(j = 0; j < trans_size; j++) |
1169 | 0 | { |
1170 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1171 | 0 | for(k = 0; k < 8; k++) |
1172 | 0 | { |
1173 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
1174 | 0 | + g_ai2_ihevc_trans_16[3][k] |
1175 | 0 | * pi2_tmp[3 * trans_size]; |
1176 | 0 | } |
1177 | 0 | for(k = 0; k < 4; k++) |
1178 | 0 | { |
1179 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size]; |
1180 | 0 | } |
1181 | 0 | eeo[0] = 0; |
1182 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0]; |
1183 | 0 | eeo[1] = 0; |
1184 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0]; |
1185 | | |
1186 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1187 | 0 | for(k = 0; k < 2; k++) |
1188 | 0 | { |
1189 | 0 | ee[k] = eee[k] + eeo[k]; |
1190 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
1191 | 0 | } |
1192 | 0 | for(k = 0; k < 4; k++) |
1193 | 0 | { |
1194 | 0 | e[k] = ee[k] + eo[k]; |
1195 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
1196 | 0 | } |
1197 | 0 | for(k = 0; k < 8; k++) |
1198 | 0 | { |
1199 | 0 | pi2_dst[k] = |
1200 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1201 | |
|
1202 | 0 | pi2_dst[k + 8] = |
1203 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
1204 | 0 | } |
1205 | 0 | pi2_tmp++; |
1206 | 0 | pi2_dst += dst_strd; |
1207 | 0 | } |
1208 | 0 | } |
1209 | 0 | else if((zero_rows_2nd_stage & 0xFF00) == 0xFF00) /* First 4 rows of output of 1st stage are non-zero */ |
1210 | 0 | { |
1211 | 0 | for(j = 0; j < trans_size; j++) |
1212 | 0 | { |
1213 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1214 | 0 | for(k = 0; k < 8; k++) |
1215 | 0 | { |
1216 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
1217 | 0 | + g_ai2_ihevc_trans_16[3][k] |
1218 | 0 | * pi2_tmp[3 * trans_size] |
1219 | 0 | + g_ai2_ihevc_trans_16[5][k] |
1220 | 0 | * pi2_tmp[5 * trans_size] |
1221 | 0 | + g_ai2_ihevc_trans_16[7][k] |
1222 | 0 | * pi2_tmp[7 * trans_size]; |
1223 | 0 | } |
1224 | 0 | for(k = 0; k < 4; k++) |
1225 | 0 | { |
1226 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size] |
1227 | 0 | + g_ai2_ihevc_trans_16[6][k] |
1228 | 0 | * pi2_tmp[6 * trans_size]; |
1229 | 0 | } |
1230 | 0 | eeo[0] = g_ai2_ihevc_trans_16[4][0] * pi2_tmp[4 * trans_size]; |
1231 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0]; |
1232 | 0 | eeo[1] = g_ai2_ihevc_trans_16[4][1] * pi2_tmp[4 * trans_size]; |
1233 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0]; |
1234 | | |
1235 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1236 | 0 | for(k = 0; k < 2; k++) |
1237 | 0 | { |
1238 | 0 | ee[k] = eee[k] + eeo[k]; |
1239 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
1240 | 0 | } |
1241 | 0 | for(k = 0; k < 4; k++) |
1242 | 0 | { |
1243 | 0 | e[k] = ee[k] + eo[k]; |
1244 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
1245 | 0 | } |
1246 | 0 | for(k = 0; k < 8; k++) |
1247 | 0 | { |
1248 | 0 | pi2_dst[k] = |
1249 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1250 | |
|
1251 | 0 | pi2_dst[k + 8] = |
1252 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
1253 | 0 | } |
1254 | 0 | pi2_tmp++; |
1255 | 0 | pi2_dst += dst_strd; |
1256 | 0 | } |
1257 | 0 | } |
1258 | 0 | else /* All rows of output of 1st stage are non-zero */ |
1259 | 0 | { |
1260 | 0 | for(j = 0; j < trans_size; j++) |
1261 | 0 | { |
1262 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1263 | 0 | for(k = 0; k < 8; k++) |
1264 | 0 | { |
1265 | 0 | o[k] = g_ai2_ihevc_trans_16[1][k] * pi2_tmp[trans_size] |
1266 | 0 | + g_ai2_ihevc_trans_16[3][k] |
1267 | 0 | * pi2_tmp[3 * trans_size] |
1268 | 0 | + g_ai2_ihevc_trans_16[5][k] |
1269 | 0 | * pi2_tmp[5 * trans_size] |
1270 | 0 | + g_ai2_ihevc_trans_16[7][k] |
1271 | 0 | * pi2_tmp[7 * trans_size] |
1272 | 0 | + g_ai2_ihevc_trans_16[9][k] |
1273 | 0 | * pi2_tmp[9 * trans_size] |
1274 | 0 | + g_ai2_ihevc_trans_16[11][k] |
1275 | 0 | * pi2_tmp[11 * trans_size] |
1276 | 0 | + g_ai2_ihevc_trans_16[13][k] |
1277 | 0 | * pi2_tmp[13 * trans_size] |
1278 | 0 | + g_ai2_ihevc_trans_16[15][k] |
1279 | 0 | * pi2_tmp[15 * trans_size]; |
1280 | 0 | } |
1281 | 0 | for(k = 0; k < 4; k++) |
1282 | 0 | { |
1283 | 0 | eo[k] = g_ai2_ihevc_trans_16[2][k] * pi2_tmp[2 * trans_size] |
1284 | 0 | + g_ai2_ihevc_trans_16[6][k] |
1285 | 0 | * pi2_tmp[6 * trans_size] |
1286 | 0 | + g_ai2_ihevc_trans_16[10][k] |
1287 | 0 | * pi2_tmp[10 * trans_size] |
1288 | 0 | + g_ai2_ihevc_trans_16[14][k] |
1289 | 0 | * pi2_tmp[14 * trans_size]; |
1290 | 0 | } |
1291 | 0 | eeo[0] = |
1292 | 0 | g_ai2_ihevc_trans_16[4][0] * pi2_tmp[4 * trans_size] |
1293 | 0 | + g_ai2_ihevc_trans_16[12][0] |
1294 | 0 | * pi2_tmp[12 |
1295 | 0 | * trans_size]; |
1296 | 0 | eee[0] = g_ai2_ihevc_trans_16[0][0] * pi2_tmp[0] |
1297 | 0 | + g_ai2_ihevc_trans_16[8][0] * pi2_tmp[8 * trans_size]; |
1298 | 0 | eeo[1] = |
1299 | 0 | g_ai2_ihevc_trans_16[4][1] * pi2_tmp[4 * trans_size] |
1300 | 0 | + g_ai2_ihevc_trans_16[12][1] |
1301 | 0 | * pi2_tmp[12 |
1302 | 0 | * trans_size]; |
1303 | 0 | eee[1] = g_ai2_ihevc_trans_16[0][1] * pi2_tmp[0] |
1304 | 0 | + g_ai2_ihevc_trans_16[8][1] * pi2_tmp[8 * trans_size]; |
1305 | | |
1306 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1307 | 0 | for(k = 0; k < 2; k++) |
1308 | 0 | { |
1309 | 0 | ee[k] = eee[k] + eeo[k]; |
1310 | 0 | ee[k + 2] = eee[1 - k] - eeo[1 - k]; |
1311 | 0 | } |
1312 | 0 | for(k = 0; k < 4; k++) |
1313 | 0 | { |
1314 | 0 | e[k] = ee[k] + eo[k]; |
1315 | 0 | e[k + 4] = ee[3 - k] - eo[3 - k]; |
1316 | 0 | } |
1317 | 0 | for(k = 0; k < 8; k++) |
1318 | 0 | { |
1319 | 0 | pi2_dst[k] = |
1320 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1321 | |
|
1322 | 0 | pi2_dst[k + 8] = |
1323 | 0 | CLIP_S16(((e[7 - k] - o[7 - k] + add) >> shift)); |
1324 | 0 | } |
1325 | 0 | pi2_tmp++; |
1326 | 0 | pi2_dst += dst_strd; |
1327 | 0 | } |
1328 | 0 | } |
1329 | | /************************************************************************************************/ |
1330 | | /************************************END - IT_RECON_16x16****************************************/ |
1331 | | /************************************************************************************************/ |
1332 | 0 | } |
1333 | |
|
1334 | 0 | } |
1335 | | |
1336 | | |
1337 | | void ihevc_itrans_res_32x32(WORD16 *pi2_src, |
1338 | | WORD16 *pi2_tmp, |
1339 | | WORD16 *pi2_dst, |
1340 | | WORD32 src_strd, |
1341 | | WORD32 dst_strd, |
1342 | | WORD32 zero_cols, |
1343 | | WORD32 zero_rows) |
1344 | 0 | { |
1345 | 0 | WORD32 j, k; |
1346 | 0 | WORD32 e[16], o[16]; |
1347 | 0 | WORD32 ee[8], eo[8]; |
1348 | 0 | WORD32 eee[4], eeo[4]; |
1349 | 0 | WORD32 eeee[2], eeeo[2]; |
1350 | 0 | WORD32 add; |
1351 | 0 | WORD32 shift; |
1352 | 0 | WORD16 *pi2_tmp_orig; |
1353 | 0 | WORD32 trans_size; |
1354 | 0 | WORD32 zero_rows_2nd_stage = zero_cols; |
1355 | 0 | WORD32 row_limit_2nd_stage; |
1356 | |
|
1357 | 0 | trans_size = TRANS_SIZE_32; |
1358 | 0 | pi2_tmp_orig = pi2_tmp; |
1359 | |
|
1360 | 0 | if((zero_cols & 0xFFFFFFF0) == 0xFFFFFFF0) |
1361 | 0 | row_limit_2nd_stage = 4; |
1362 | 0 | else if((zero_cols & 0xFFFFFF00) == 0xFFFFFF00) |
1363 | 0 | row_limit_2nd_stage = 8; |
1364 | 0 | else |
1365 | 0 | row_limit_2nd_stage = TRANS_SIZE_32; |
1366 | |
|
1367 | 0 | if((zero_rows & 0xFFFFFFF0) == 0xFFFFFFF0) /* First 4 rows of input are non-zero */ |
1368 | 0 | { |
1369 | | /************************************************************************************************/ |
1370 | | /**********************************START - IT_RECON_32x32****************************************/ |
1371 | | /************************************************************************************************/ |
1372 | | /* Inverse Transform 1st stage */ |
1373 | 0 | shift = IT_SHIFT_STAGE_1; |
1374 | 0 | add = 1 << (shift - 1); |
1375 | |
|
1376 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
1377 | 0 | { |
1378 | | /* Checking for Zero Cols */ |
1379 | 0 | if((zero_cols & 1) == 1) |
1380 | 0 | { |
1381 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
1382 | 0 | } |
1383 | 0 | else |
1384 | 0 | { |
1385 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1386 | 0 | for(k = 0; k < 16; k++) |
1387 | 0 | { |
1388 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_src[src_strd] |
1389 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1390 | 0 | * pi2_src[3 * src_strd]; |
1391 | 0 | } |
1392 | 0 | for(k = 0; k < 8; k++) |
1393 | 0 | { |
1394 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_src[2 * src_strd]; |
1395 | 0 | } |
1396 | | // for(k = 0; k < 4; k++) |
1397 | 0 | { |
1398 | 0 | eeo[0] = 0; |
1399 | 0 | eeo[1] = 0; |
1400 | 0 | eeo[2] = 0; |
1401 | 0 | eeo[3] = 0; |
1402 | 0 | } |
1403 | 0 | eeeo[0] = 0; |
1404 | 0 | eeeo[1] = 0; |
1405 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_src[0]; |
1406 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_src[0]; |
1407 | | |
1408 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1409 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1410 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1411 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1412 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1413 | 0 | for(k = 0; k < 4; k++) |
1414 | 0 | { |
1415 | 0 | ee[k] = eee[k] + eeo[k]; |
1416 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1417 | 0 | } |
1418 | 0 | for(k = 0; k < 8; k++) |
1419 | 0 | { |
1420 | 0 | e[k] = ee[k] + eo[k]; |
1421 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1422 | 0 | } |
1423 | 0 | for(k = 0; k < 16; k++) |
1424 | 0 | { |
1425 | 0 | pi2_tmp[k] = |
1426 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1427 | 0 | pi2_tmp[k + 16] = |
1428 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1429 | 0 | } |
1430 | 0 | } |
1431 | 0 | pi2_src++; |
1432 | 0 | pi2_tmp += trans_size; |
1433 | 0 | zero_cols = zero_cols >> 1; |
1434 | 0 | } |
1435 | |
|
1436 | 0 | pi2_tmp = pi2_tmp_orig; |
1437 | | |
1438 | | /* Inverse Transform 2nd stage */ |
1439 | 0 | shift = IT_SHIFT_STAGE_2; |
1440 | 0 | add = 1 << (shift - 1); |
1441 | 0 | if((zero_rows_2nd_stage & 0xFFFFFFF0) == 0xFFFFFFF0) /* First 4 rows of output of 1st stage are non-zero */ |
1442 | 0 | { |
1443 | 0 | for(j = 0; j < trans_size; j++) |
1444 | 0 | { |
1445 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1446 | 0 | for(k = 0; k < 16; k++) |
1447 | 0 | { |
1448 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
1449 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1450 | 0 | * pi2_tmp[3 * trans_size]; |
1451 | 0 | } |
1452 | 0 | for(k = 0; k < 8; k++) |
1453 | 0 | { |
1454 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size]; |
1455 | 0 | } |
1456 | | // for(k = 0; k < 4; k++) |
1457 | 0 | { |
1458 | 0 | eeo[0] = 0; |
1459 | 0 | eeo[1] = 0; |
1460 | 0 | eeo[2] = 0; |
1461 | 0 | eeo[3] = 0; |
1462 | 0 | } |
1463 | 0 | eeeo[0] = 0; |
1464 | 0 | eeeo[1] = 0; |
1465 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0]; |
1466 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0]; |
1467 | | |
1468 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1469 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1470 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1471 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1472 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1473 | 0 | for(k = 0; k < 4; k++) |
1474 | 0 | { |
1475 | 0 | ee[k] = eee[k] + eeo[k]; |
1476 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1477 | 0 | } |
1478 | 0 | for(k = 0; k < 8; k++) |
1479 | 0 | { |
1480 | 0 | e[k] = ee[k] + eo[k]; |
1481 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1482 | 0 | } |
1483 | 0 | for(k = 0; k < 16; k++) |
1484 | 0 | { |
1485 | 0 | pi2_dst[k] = |
1486 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1487 | |
|
1488 | 0 | pi2_dst[k + 16] = |
1489 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1490 | 0 | } |
1491 | 0 | pi2_tmp++; |
1492 | 0 | pi2_dst += dst_strd; |
1493 | 0 | } |
1494 | 0 | } |
1495 | 0 | else if((zero_rows_2nd_stage & 0xFFFFFF00) == 0xFFFFFF00) /* First 8 rows of output of 1st stage are non-zero */ |
1496 | 0 | { |
1497 | 0 | for(j = 0; j < trans_size; j++) |
1498 | 0 | { |
1499 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1500 | 0 | for(k = 0; k < 16; k++) |
1501 | 0 | { |
1502 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
1503 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1504 | 0 | * pi2_tmp[3 * trans_size] |
1505 | 0 | + g_ai2_ihevc_trans_32[5][k] |
1506 | 0 | * pi2_tmp[5 * trans_size] |
1507 | 0 | + g_ai2_ihevc_trans_32[7][k] |
1508 | 0 | * pi2_tmp[7 * trans_size]; |
1509 | 0 | } |
1510 | 0 | for(k = 0; k < 8; k++) |
1511 | 0 | { |
1512 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size] |
1513 | 0 | + g_ai2_ihevc_trans_32[6][k] |
1514 | 0 | * pi2_tmp[6 * trans_size]; |
1515 | 0 | } |
1516 | 0 | for(k = 0; k < 4; k++) |
1517 | 0 | { |
1518 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_tmp[4 * trans_size]; |
1519 | 0 | } |
1520 | 0 | eeeo[0] = 0; |
1521 | 0 | eeeo[1] = 0; |
1522 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0]; |
1523 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0]; |
1524 | | |
1525 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1526 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1527 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1528 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1529 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1530 | 0 | for(k = 0; k < 4; k++) |
1531 | 0 | { |
1532 | 0 | ee[k] = eee[k] + eeo[k]; |
1533 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1534 | 0 | } |
1535 | 0 | for(k = 0; k < 8; k++) |
1536 | 0 | { |
1537 | 0 | e[k] = ee[k] + eo[k]; |
1538 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1539 | 0 | } |
1540 | 0 | for(k = 0; k < 16; k++) |
1541 | 0 | { |
1542 | 0 | pi2_dst[k] = |
1543 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1544 | |
|
1545 | 0 | pi2_dst[k + 16] = |
1546 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1547 | 0 | } |
1548 | 0 | pi2_tmp++; |
1549 | 0 | pi2_dst += dst_strd; |
1550 | 0 | } |
1551 | 0 | } |
1552 | 0 | else /* All rows of output of 1st stage are non-zero */ |
1553 | 0 | { |
1554 | 0 | for(j = 0; j < trans_size; j++) |
1555 | 0 | { |
1556 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1557 | 0 | for(k = 0; k < 16; k++) |
1558 | 0 | { |
1559 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
1560 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1561 | 0 | * pi2_tmp[3 * trans_size] |
1562 | 0 | + g_ai2_ihevc_trans_32[5][k] |
1563 | 0 | * pi2_tmp[5 * trans_size] |
1564 | 0 | + g_ai2_ihevc_trans_32[7][k] |
1565 | 0 | * pi2_tmp[7 * trans_size] |
1566 | 0 | + g_ai2_ihevc_trans_32[9][k] |
1567 | 0 | * pi2_tmp[9 * trans_size] |
1568 | 0 | + g_ai2_ihevc_trans_32[11][k] |
1569 | 0 | * pi2_tmp[11 * trans_size] |
1570 | 0 | + g_ai2_ihevc_trans_32[13][k] |
1571 | 0 | * pi2_tmp[13 * trans_size] |
1572 | 0 | + g_ai2_ihevc_trans_32[15][k] |
1573 | 0 | * pi2_tmp[15 * trans_size] |
1574 | 0 | + g_ai2_ihevc_trans_32[17][k] |
1575 | 0 | * pi2_tmp[17 * trans_size] |
1576 | 0 | + g_ai2_ihevc_trans_32[19][k] |
1577 | 0 | * pi2_tmp[19 * trans_size] |
1578 | 0 | + g_ai2_ihevc_trans_32[21][k] |
1579 | 0 | * pi2_tmp[21 * trans_size] |
1580 | 0 | + g_ai2_ihevc_trans_32[23][k] |
1581 | 0 | * pi2_tmp[23 * trans_size] |
1582 | 0 | + g_ai2_ihevc_trans_32[25][k] |
1583 | 0 | * pi2_tmp[25 * trans_size] |
1584 | 0 | + g_ai2_ihevc_trans_32[27][k] |
1585 | 0 | * pi2_tmp[27 * trans_size] |
1586 | 0 | + g_ai2_ihevc_trans_32[29][k] |
1587 | 0 | * pi2_tmp[29 * trans_size] |
1588 | 0 | + g_ai2_ihevc_trans_32[31][k] |
1589 | 0 | * pi2_tmp[31 * trans_size]; |
1590 | 0 | } |
1591 | 0 | for(k = 0; k < 8; k++) |
1592 | 0 | { |
1593 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size] |
1594 | 0 | + g_ai2_ihevc_trans_32[6][k] |
1595 | 0 | * pi2_tmp[6 * trans_size] |
1596 | 0 | + g_ai2_ihevc_trans_32[10][k] |
1597 | 0 | * pi2_tmp[10 * trans_size] |
1598 | 0 | + g_ai2_ihevc_trans_32[14][k] |
1599 | 0 | * pi2_tmp[14 * trans_size] |
1600 | 0 | + g_ai2_ihevc_trans_32[18][k] |
1601 | 0 | * pi2_tmp[18 * trans_size] |
1602 | 0 | + g_ai2_ihevc_trans_32[22][k] |
1603 | 0 | * pi2_tmp[22 * trans_size] |
1604 | 0 | + g_ai2_ihevc_trans_32[26][k] |
1605 | 0 | * pi2_tmp[26 * trans_size] |
1606 | 0 | + g_ai2_ihevc_trans_32[30][k] |
1607 | 0 | * pi2_tmp[30 * trans_size]; |
1608 | 0 | } |
1609 | 0 | for(k = 0; k < 4; k++) |
1610 | 0 | { |
1611 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_tmp[4 * trans_size] |
1612 | 0 | + g_ai2_ihevc_trans_32[12][k] |
1613 | 0 | * pi2_tmp[12 * trans_size] |
1614 | 0 | + g_ai2_ihevc_trans_32[20][k] |
1615 | 0 | * pi2_tmp[20 * trans_size] |
1616 | 0 | + g_ai2_ihevc_trans_32[28][k] |
1617 | 0 | * pi2_tmp[28 * trans_size]; |
1618 | 0 | } |
1619 | 0 | eeeo[0] = |
1620 | 0 | g_ai2_ihevc_trans_32[8][0] * pi2_tmp[8 * trans_size] |
1621 | 0 | + g_ai2_ihevc_trans_32[24][0] |
1622 | 0 | * pi2_tmp[24 |
1623 | 0 | * trans_size]; |
1624 | 0 | eeeo[1] = |
1625 | 0 | g_ai2_ihevc_trans_32[8][1] * pi2_tmp[8 * trans_size] |
1626 | 0 | + g_ai2_ihevc_trans_32[24][1] |
1627 | 0 | * pi2_tmp[24 |
1628 | 0 | * trans_size]; |
1629 | 0 | eeee[0] = |
1630 | 0 | g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0] |
1631 | 0 | + g_ai2_ihevc_trans_32[16][0] |
1632 | 0 | * pi2_tmp[16 |
1633 | 0 | * trans_size]; |
1634 | 0 | eeee[1] = |
1635 | 0 | g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0] |
1636 | 0 | + g_ai2_ihevc_trans_32[16][1] |
1637 | 0 | * pi2_tmp[16 |
1638 | 0 | * trans_size]; |
1639 | | |
1640 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1641 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1642 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1643 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1644 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1645 | 0 | for(k = 0; k < 4; k++) |
1646 | 0 | { |
1647 | 0 | ee[k] = eee[k] + eeo[k]; |
1648 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1649 | 0 | } |
1650 | 0 | for(k = 0; k < 8; k++) |
1651 | 0 | { |
1652 | 0 | e[k] = ee[k] + eo[k]; |
1653 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1654 | 0 | } |
1655 | 0 | for(k = 0; k < 16; k++) |
1656 | 0 | { |
1657 | 0 | pi2_dst[k] = |
1658 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1659 | |
|
1660 | 0 | pi2_dst[k + 16] = |
1661 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1662 | 0 | } |
1663 | 0 | pi2_tmp++; |
1664 | 0 | pi2_dst += dst_strd; |
1665 | 0 | } |
1666 | 0 | } |
1667 | | /************************************************************************************************/ |
1668 | | /************************************END - IT_RECON_32x32****************************************/ |
1669 | | /************************************************************************************************/ |
1670 | 0 | } |
1671 | 0 | else if((zero_rows & 0xFFFFFF00) == 0xFFFFFF00) /* First 8 rows of input are non-zero */ |
1672 | 0 | { |
1673 | | /************************************************************************************************/ |
1674 | | /**********************************START - IT_RECON_32x32****************************************/ |
1675 | | /************************************************************************************************/ |
1676 | | /* Inverse Transform 1st stage */ |
1677 | 0 | shift = IT_SHIFT_STAGE_1; |
1678 | 0 | add = 1 << (shift - 1); |
1679 | |
|
1680 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
1681 | 0 | { |
1682 | | /* Checking for Zero Cols */ |
1683 | 0 | if((zero_cols & 1) == 1) |
1684 | 0 | { |
1685 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
1686 | 0 | } |
1687 | 0 | else |
1688 | 0 | { |
1689 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1690 | 0 | for(k = 0; k < 16; k++) |
1691 | 0 | { |
1692 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_src[src_strd] |
1693 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1694 | 0 | * pi2_src[3 * src_strd] |
1695 | 0 | + g_ai2_ihevc_trans_32[5][k] |
1696 | 0 | * pi2_src[5 * src_strd] |
1697 | 0 | + g_ai2_ihevc_trans_32[7][k] |
1698 | 0 | * pi2_src[7 * src_strd]; |
1699 | 0 | } |
1700 | 0 | for(k = 0; k < 8; k++) |
1701 | 0 | { |
1702 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_src[2 * src_strd] |
1703 | 0 | + g_ai2_ihevc_trans_32[6][k] |
1704 | 0 | * pi2_src[6 * src_strd]; |
1705 | 0 | } |
1706 | 0 | for(k = 0; k < 4; k++) |
1707 | 0 | { |
1708 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_src[4 * src_strd]; |
1709 | 0 | } |
1710 | 0 | eeeo[0] = 0; |
1711 | 0 | eeeo[1] = 0; |
1712 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_src[0]; |
1713 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_src[0]; |
1714 | | |
1715 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1716 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1717 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1718 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1719 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1720 | 0 | for(k = 0; k < 4; k++) |
1721 | 0 | { |
1722 | 0 | ee[k] = eee[k] + eeo[k]; |
1723 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1724 | 0 | } |
1725 | 0 | for(k = 0; k < 8; k++) |
1726 | 0 | { |
1727 | 0 | e[k] = ee[k] + eo[k]; |
1728 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1729 | 0 | } |
1730 | 0 | for(k = 0; k < 16; k++) |
1731 | 0 | { |
1732 | 0 | pi2_tmp[k] = |
1733 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1734 | 0 | pi2_tmp[k + 16] = |
1735 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1736 | 0 | } |
1737 | 0 | } |
1738 | 0 | pi2_src++; |
1739 | 0 | pi2_tmp += trans_size; |
1740 | 0 | zero_cols = zero_cols >> 1; |
1741 | 0 | } |
1742 | |
|
1743 | 0 | pi2_tmp = pi2_tmp_orig; |
1744 | | |
1745 | | /* Inverse Transform 2nd stage */ |
1746 | 0 | shift = IT_SHIFT_STAGE_2; |
1747 | 0 | add = 1 << (shift - 1); |
1748 | 0 | if((zero_rows_2nd_stage & 0xFFFFFFF0) == 0xFFFFFFF0) /* First 4 rows of output of 1st stage are non-zero */ |
1749 | 0 | { |
1750 | 0 | for(j = 0; j < trans_size; j++) |
1751 | 0 | { |
1752 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1753 | 0 | for(k = 0; k < 16; k++) |
1754 | 0 | { |
1755 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
1756 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1757 | 0 | * pi2_tmp[3 * trans_size]; |
1758 | 0 | } |
1759 | 0 | for(k = 0; k < 8; k++) |
1760 | 0 | { |
1761 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size]; |
1762 | 0 | } |
1763 | | // for(k = 0; k < 4; k++) |
1764 | 0 | { |
1765 | 0 | eeo[0] = 0; |
1766 | 0 | eeo[1] = 0; |
1767 | 0 | eeo[2] = 0; |
1768 | 0 | eeo[3] = 0; |
1769 | 0 | } |
1770 | 0 | eeeo[0] = 0; |
1771 | 0 | eeeo[1] = 0; |
1772 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0]; |
1773 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0]; |
1774 | | |
1775 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1776 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1777 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1778 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1779 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1780 | 0 | for(k = 0; k < 4; k++) |
1781 | 0 | { |
1782 | 0 | ee[k] = eee[k] + eeo[k]; |
1783 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1784 | 0 | } |
1785 | 0 | for(k = 0; k < 8; k++) |
1786 | 0 | { |
1787 | 0 | e[k] = ee[k] + eo[k]; |
1788 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1789 | 0 | } |
1790 | 0 | for(k = 0; k < 16; k++) |
1791 | 0 | { |
1792 | 0 | pi2_dst[k] = |
1793 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1794 | |
|
1795 | 0 | pi2_dst[k + 16] = |
1796 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1797 | 0 | } |
1798 | 0 | pi2_tmp++; |
1799 | 0 | pi2_dst += dst_strd; |
1800 | 0 | } |
1801 | 0 | } |
1802 | 0 | else if((zero_rows_2nd_stage & 0xFFFFFF00) == 0xFFFFFF00) /* First 8 rows of output of 1st stage are non-zero */ |
1803 | 0 | { |
1804 | 0 | for(j = 0; j < trans_size; j++) |
1805 | 0 | { |
1806 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1807 | 0 | for(k = 0; k < 16; k++) |
1808 | 0 | { |
1809 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
1810 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1811 | 0 | * pi2_tmp[3 * trans_size] |
1812 | 0 | + g_ai2_ihevc_trans_32[5][k] |
1813 | 0 | * pi2_tmp[5 * trans_size] |
1814 | 0 | + g_ai2_ihevc_trans_32[7][k] |
1815 | 0 | * pi2_tmp[7 * trans_size]; |
1816 | 0 | } |
1817 | 0 | for(k = 0; k < 8; k++) |
1818 | 0 | { |
1819 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size] |
1820 | 0 | + g_ai2_ihevc_trans_32[6][k] |
1821 | 0 | * pi2_tmp[6 * trans_size]; |
1822 | 0 | } |
1823 | 0 | for(k = 0; k < 4; k++) |
1824 | 0 | { |
1825 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_tmp[4 * trans_size]; |
1826 | 0 | } |
1827 | 0 | eeeo[0] = 0; |
1828 | 0 | eeeo[1] = 0; |
1829 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0]; |
1830 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0]; |
1831 | | |
1832 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1833 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1834 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1835 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1836 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1837 | 0 | for(k = 0; k < 4; k++) |
1838 | 0 | { |
1839 | 0 | ee[k] = eee[k] + eeo[k]; |
1840 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1841 | 0 | } |
1842 | 0 | for(k = 0; k < 8; k++) |
1843 | 0 | { |
1844 | 0 | e[k] = ee[k] + eo[k]; |
1845 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1846 | 0 | } |
1847 | 0 | for(k = 0; k < 16; k++) |
1848 | 0 | { |
1849 | 0 | pi2_dst[k] = |
1850 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1851 | |
|
1852 | 0 | pi2_dst[k + 16] = |
1853 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1854 | 0 | } |
1855 | 0 | pi2_tmp++; |
1856 | 0 | pi2_dst += dst_strd; |
1857 | 0 | } |
1858 | 0 | } |
1859 | 0 | else /* All rows of output of 1st stage are non-zero */ |
1860 | 0 | { |
1861 | 0 | for(j = 0; j < trans_size; j++) |
1862 | 0 | { |
1863 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1864 | 0 | for(k = 0; k < 16; k++) |
1865 | 0 | { |
1866 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
1867 | 0 | + g_ai2_ihevc_trans_32[3][k] |
1868 | 0 | * pi2_tmp[3 * trans_size] |
1869 | 0 | + g_ai2_ihevc_trans_32[5][k] |
1870 | 0 | * pi2_tmp[5 * trans_size] |
1871 | 0 | + g_ai2_ihevc_trans_32[7][k] |
1872 | 0 | * pi2_tmp[7 * trans_size] |
1873 | 0 | + g_ai2_ihevc_trans_32[9][k] |
1874 | 0 | * pi2_tmp[9 * trans_size] |
1875 | 0 | + g_ai2_ihevc_trans_32[11][k] |
1876 | 0 | * pi2_tmp[11 * trans_size] |
1877 | 0 | + g_ai2_ihevc_trans_32[13][k] |
1878 | 0 | * pi2_tmp[13 * trans_size] |
1879 | 0 | + g_ai2_ihevc_trans_32[15][k] |
1880 | 0 | * pi2_tmp[15 * trans_size] |
1881 | 0 | + g_ai2_ihevc_trans_32[17][k] |
1882 | 0 | * pi2_tmp[17 * trans_size] |
1883 | 0 | + g_ai2_ihevc_trans_32[19][k] |
1884 | 0 | * pi2_tmp[19 * trans_size] |
1885 | 0 | + g_ai2_ihevc_trans_32[21][k] |
1886 | 0 | * pi2_tmp[21 * trans_size] |
1887 | 0 | + g_ai2_ihevc_trans_32[23][k] |
1888 | 0 | * pi2_tmp[23 * trans_size] |
1889 | 0 | + g_ai2_ihevc_trans_32[25][k] |
1890 | 0 | * pi2_tmp[25 * trans_size] |
1891 | 0 | + g_ai2_ihevc_trans_32[27][k] |
1892 | 0 | * pi2_tmp[27 * trans_size] |
1893 | 0 | + g_ai2_ihevc_trans_32[29][k] |
1894 | 0 | * pi2_tmp[29 * trans_size] |
1895 | 0 | + g_ai2_ihevc_trans_32[31][k] |
1896 | 0 | * pi2_tmp[31 * trans_size]; |
1897 | 0 | } |
1898 | 0 | for(k = 0; k < 8; k++) |
1899 | 0 | { |
1900 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size] |
1901 | 0 | + g_ai2_ihevc_trans_32[6][k] |
1902 | 0 | * pi2_tmp[6 * trans_size] |
1903 | 0 | + g_ai2_ihevc_trans_32[10][k] |
1904 | 0 | * pi2_tmp[10 * trans_size] |
1905 | 0 | + g_ai2_ihevc_trans_32[14][k] |
1906 | 0 | * pi2_tmp[14 * trans_size] |
1907 | 0 | + g_ai2_ihevc_trans_32[18][k] |
1908 | 0 | * pi2_tmp[18 * trans_size] |
1909 | 0 | + g_ai2_ihevc_trans_32[22][k] |
1910 | 0 | * pi2_tmp[22 * trans_size] |
1911 | 0 | + g_ai2_ihevc_trans_32[26][k] |
1912 | 0 | * pi2_tmp[26 * trans_size] |
1913 | 0 | + g_ai2_ihevc_trans_32[30][k] |
1914 | 0 | * pi2_tmp[30 * trans_size]; |
1915 | 0 | } |
1916 | 0 | for(k = 0; k < 4; k++) |
1917 | 0 | { |
1918 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_tmp[4 * trans_size] |
1919 | 0 | + g_ai2_ihevc_trans_32[12][k] |
1920 | 0 | * pi2_tmp[12 * trans_size] |
1921 | 0 | + g_ai2_ihevc_trans_32[20][k] |
1922 | 0 | * pi2_tmp[20 * trans_size] |
1923 | 0 | + g_ai2_ihevc_trans_32[28][k] |
1924 | 0 | * pi2_tmp[28 * trans_size]; |
1925 | 0 | } |
1926 | 0 | eeeo[0] = |
1927 | 0 | g_ai2_ihevc_trans_32[8][0] * pi2_tmp[8 * trans_size] |
1928 | 0 | + g_ai2_ihevc_trans_32[24][0] |
1929 | 0 | * pi2_tmp[24 |
1930 | 0 | * trans_size]; |
1931 | 0 | eeeo[1] = |
1932 | 0 | g_ai2_ihevc_trans_32[8][1] * pi2_tmp[8 * trans_size] |
1933 | 0 | + g_ai2_ihevc_trans_32[24][1] |
1934 | 0 | * pi2_tmp[24 |
1935 | 0 | * trans_size]; |
1936 | 0 | eeee[0] = |
1937 | 0 | g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0] |
1938 | 0 | + g_ai2_ihevc_trans_32[16][0] |
1939 | 0 | * pi2_tmp[16 |
1940 | 0 | * trans_size]; |
1941 | 0 | eeee[1] = |
1942 | 0 | g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0] |
1943 | 0 | + g_ai2_ihevc_trans_32[16][1] |
1944 | 0 | * pi2_tmp[16 |
1945 | 0 | * trans_size]; |
1946 | | |
1947 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
1948 | 0 | eee[0] = eeee[0] + eeeo[0]; |
1949 | 0 | eee[3] = eeee[0] - eeeo[0]; |
1950 | 0 | eee[1] = eeee[1] + eeeo[1]; |
1951 | 0 | eee[2] = eeee[1] - eeeo[1]; |
1952 | 0 | for(k = 0; k < 4; k++) |
1953 | 0 | { |
1954 | 0 | ee[k] = eee[k] + eeo[k]; |
1955 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
1956 | 0 | } |
1957 | 0 | for(k = 0; k < 8; k++) |
1958 | 0 | { |
1959 | 0 | e[k] = ee[k] + eo[k]; |
1960 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
1961 | 0 | } |
1962 | 0 | for(k = 0; k < 16; k++) |
1963 | 0 | { |
1964 | 0 | pi2_dst[k] = |
1965 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
1966 | |
|
1967 | 0 | pi2_dst[k + 16] = |
1968 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
1969 | 0 | } |
1970 | 0 | pi2_tmp++; |
1971 | 0 | pi2_dst += dst_strd; |
1972 | 0 | } |
1973 | 0 | } |
1974 | | /************************************************************************************************/ |
1975 | | /************************************END - IT_RECON_32x32****************************************/ |
1976 | | /************************************************************************************************/ |
1977 | 0 | } |
1978 | 0 | else /* All rows of input are non-zero */ |
1979 | 0 | { |
1980 | | /************************************************************************************************/ |
1981 | | /**********************************START - IT_RECON_32x32****************************************/ |
1982 | | /************************************************************************************************/ |
1983 | | /* Inverse Transform 1st stage */ |
1984 | 0 | shift = IT_SHIFT_STAGE_1; |
1985 | 0 | add = 1 << (shift - 1); |
1986 | |
|
1987 | 0 | for(j = 0; j < row_limit_2nd_stage; j++) |
1988 | 0 | { |
1989 | | /* Checking for Zero Cols */ |
1990 | 0 | if((zero_cols & 1) == 1) |
1991 | 0 | { |
1992 | 0 | memset(pi2_tmp, 0, trans_size * sizeof(WORD16)); |
1993 | 0 | } |
1994 | 0 | else |
1995 | 0 | { |
1996 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
1997 | 0 | for(k = 0; k < 16; k++) |
1998 | 0 | { |
1999 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_src[src_strd] |
2000 | 0 | + g_ai2_ihevc_trans_32[3][k] |
2001 | 0 | * pi2_src[3 * src_strd] |
2002 | 0 | + g_ai2_ihevc_trans_32[5][k] |
2003 | 0 | * pi2_src[5 * src_strd] |
2004 | 0 | + g_ai2_ihevc_trans_32[7][k] |
2005 | 0 | * pi2_src[7 * src_strd] |
2006 | 0 | + g_ai2_ihevc_trans_32[9][k] |
2007 | 0 | * pi2_src[9 * src_strd] |
2008 | 0 | + g_ai2_ihevc_trans_32[11][k] |
2009 | 0 | * pi2_src[11 * src_strd] |
2010 | 0 | + g_ai2_ihevc_trans_32[13][k] |
2011 | 0 | * pi2_src[13 * src_strd] |
2012 | 0 | + g_ai2_ihevc_trans_32[15][k] |
2013 | 0 | * pi2_src[15 * src_strd] |
2014 | 0 | + g_ai2_ihevc_trans_32[17][k] |
2015 | 0 | * pi2_src[17 * src_strd] |
2016 | 0 | + g_ai2_ihevc_trans_32[19][k] |
2017 | 0 | * pi2_src[19 * src_strd] |
2018 | 0 | + g_ai2_ihevc_trans_32[21][k] |
2019 | 0 | * pi2_src[21 * src_strd] |
2020 | 0 | + g_ai2_ihevc_trans_32[23][k] |
2021 | 0 | * pi2_src[23 * src_strd] |
2022 | 0 | + g_ai2_ihevc_trans_32[25][k] |
2023 | 0 | * pi2_src[25 * src_strd] |
2024 | 0 | + g_ai2_ihevc_trans_32[27][k] |
2025 | 0 | * pi2_src[27 * src_strd] |
2026 | 0 | + g_ai2_ihevc_trans_32[29][k] |
2027 | 0 | * pi2_src[29 * src_strd] |
2028 | 0 | + g_ai2_ihevc_trans_32[31][k] |
2029 | 0 | * pi2_src[31 * src_strd]; |
2030 | 0 | } |
2031 | 0 | for(k = 0; k < 8; k++) |
2032 | 0 | { |
2033 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_src[2 * src_strd] |
2034 | 0 | + g_ai2_ihevc_trans_32[6][k] |
2035 | 0 | * pi2_src[6 * src_strd] |
2036 | 0 | + g_ai2_ihevc_trans_32[10][k] |
2037 | 0 | * pi2_src[10 * src_strd] |
2038 | 0 | + g_ai2_ihevc_trans_32[14][k] |
2039 | 0 | * pi2_src[14 * src_strd] |
2040 | 0 | + g_ai2_ihevc_trans_32[18][k] |
2041 | 0 | * pi2_src[18 * src_strd] |
2042 | 0 | + g_ai2_ihevc_trans_32[22][k] |
2043 | 0 | * pi2_src[22 * src_strd] |
2044 | 0 | + g_ai2_ihevc_trans_32[26][k] |
2045 | 0 | * pi2_src[26 * src_strd] |
2046 | 0 | + g_ai2_ihevc_trans_32[30][k] |
2047 | 0 | * pi2_src[30 * src_strd]; |
2048 | 0 | } |
2049 | 0 | for(k = 0; k < 4; k++) |
2050 | 0 | { |
2051 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_src[4 * src_strd] |
2052 | 0 | + g_ai2_ihevc_trans_32[12][k] |
2053 | 0 | * pi2_src[12 * src_strd] |
2054 | 0 | + g_ai2_ihevc_trans_32[20][k] |
2055 | 0 | * pi2_src[20 * src_strd] |
2056 | 0 | + g_ai2_ihevc_trans_32[28][k] |
2057 | 0 | * pi2_src[28 * src_strd]; |
2058 | 0 | } |
2059 | 0 | eeeo[0] = g_ai2_ihevc_trans_32[8][0] * pi2_src[8 * src_strd] |
2060 | 0 | + g_ai2_ihevc_trans_32[24][0] |
2061 | 0 | * pi2_src[24 * src_strd]; |
2062 | 0 | eeeo[1] = g_ai2_ihevc_trans_32[8][1] * pi2_src[8 * src_strd] |
2063 | 0 | + g_ai2_ihevc_trans_32[24][1] |
2064 | 0 | * pi2_src[24 * src_strd]; |
2065 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_src[0] |
2066 | 0 | + g_ai2_ihevc_trans_32[16][0] |
2067 | 0 | * pi2_src[16 * src_strd]; |
2068 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_src[0] |
2069 | 0 | + g_ai2_ihevc_trans_32[16][1] |
2070 | 0 | * pi2_src[16 * src_strd]; |
2071 | | |
2072 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
2073 | 0 | eee[0] = eeee[0] + eeeo[0]; |
2074 | 0 | eee[3] = eeee[0] - eeeo[0]; |
2075 | 0 | eee[1] = eeee[1] + eeeo[1]; |
2076 | 0 | eee[2] = eeee[1] - eeeo[1]; |
2077 | 0 | for(k = 0; k < 4; k++) |
2078 | 0 | { |
2079 | 0 | ee[k] = eee[k] + eeo[k]; |
2080 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
2081 | 0 | } |
2082 | 0 | for(k = 0; k < 8; k++) |
2083 | 0 | { |
2084 | 0 | e[k] = ee[k] + eo[k]; |
2085 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
2086 | 0 | } |
2087 | 0 | for(k = 0; k < 16; k++) |
2088 | 0 | { |
2089 | 0 | pi2_tmp[k] = |
2090 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
2091 | 0 | pi2_tmp[k + 16] = |
2092 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
2093 | 0 | } |
2094 | 0 | } |
2095 | 0 | pi2_src++; |
2096 | 0 | pi2_tmp += trans_size; |
2097 | 0 | zero_cols = zero_cols >> 1; |
2098 | 0 | } |
2099 | |
|
2100 | 0 | pi2_tmp = pi2_tmp_orig; |
2101 | | |
2102 | | /* Inverse Transform 2nd stage */ |
2103 | 0 | shift = IT_SHIFT_STAGE_2; |
2104 | 0 | add = 1 << (shift - 1); |
2105 | 0 | if((zero_rows_2nd_stage & 0xFFFFFFF0) == 0xFFFFFFF0) /* First 4 rows of output of 1st stage are non-zero */ |
2106 | 0 | { |
2107 | 0 | for(j = 0; j < trans_size; j++) |
2108 | 0 | { |
2109 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
2110 | 0 | for(k = 0; k < 16; k++) |
2111 | 0 | { |
2112 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
2113 | 0 | + g_ai2_ihevc_trans_32[3][k] |
2114 | 0 | * pi2_tmp[3 * trans_size]; |
2115 | 0 | } |
2116 | 0 | for(k = 0; k < 8; k++) |
2117 | 0 | { |
2118 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size]; |
2119 | 0 | } |
2120 | | // for(k = 0; k < 4; k++) |
2121 | 0 | { |
2122 | 0 | eeo[0] = 0; |
2123 | 0 | eeo[1] = 0; |
2124 | 0 | eeo[2] = 0; |
2125 | 0 | eeo[3] = 0; |
2126 | 0 | } |
2127 | 0 | eeeo[0] = 0; |
2128 | 0 | eeeo[1] = 0; |
2129 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0]; |
2130 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0]; |
2131 | | |
2132 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
2133 | 0 | eee[0] = eeee[0] + eeeo[0]; |
2134 | 0 | eee[3] = eeee[0] - eeeo[0]; |
2135 | 0 | eee[1] = eeee[1] + eeeo[1]; |
2136 | 0 | eee[2] = eeee[1] - eeeo[1]; |
2137 | 0 | for(k = 0; k < 4; k++) |
2138 | 0 | { |
2139 | 0 | ee[k] = eee[k] + eeo[k]; |
2140 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
2141 | 0 | } |
2142 | 0 | for(k = 0; k < 8; k++) |
2143 | 0 | { |
2144 | 0 | e[k] = ee[k] + eo[k]; |
2145 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
2146 | 0 | } |
2147 | 0 | for(k = 0; k < 16; k++) |
2148 | 0 | { |
2149 | 0 | pi2_dst[k] = |
2150 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
2151 | |
|
2152 | 0 | pi2_dst[k + 16] = |
2153 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
2154 | 0 | } |
2155 | 0 | pi2_tmp++; |
2156 | 0 | pi2_dst += dst_strd; |
2157 | 0 | } |
2158 | 0 | } |
2159 | 0 | else if((zero_rows_2nd_stage & 0xFFFFFF00) == 0xFFFFFF00) /* First 8 rows of output of 1st stage are non-zero */ |
2160 | 0 | { |
2161 | 0 | for(j = 0; j < trans_size; j++) |
2162 | 0 | { |
2163 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
2164 | 0 | for(k = 0; k < 16; k++) |
2165 | 0 | { |
2166 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
2167 | 0 | + g_ai2_ihevc_trans_32[3][k] |
2168 | 0 | * pi2_tmp[3 * trans_size] |
2169 | 0 | + g_ai2_ihevc_trans_32[5][k] |
2170 | 0 | * pi2_tmp[5 * trans_size] |
2171 | 0 | + g_ai2_ihevc_trans_32[7][k] |
2172 | 0 | * pi2_tmp[7 * trans_size]; |
2173 | 0 | } |
2174 | 0 | for(k = 0; k < 8; k++) |
2175 | 0 | { |
2176 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size] |
2177 | 0 | + g_ai2_ihevc_trans_32[6][k] |
2178 | 0 | * pi2_tmp[6 * trans_size]; |
2179 | 0 | } |
2180 | 0 | for(k = 0; k < 4; k++) |
2181 | 0 | { |
2182 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_tmp[4 * trans_size]; |
2183 | 0 | } |
2184 | 0 | eeeo[0] = 0; |
2185 | 0 | eeeo[1] = 0; |
2186 | 0 | eeee[0] = g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0]; |
2187 | 0 | eeee[1] = g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0]; |
2188 | | |
2189 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
2190 | 0 | eee[0] = eeee[0] + eeeo[0]; |
2191 | 0 | eee[3] = eeee[0] - eeeo[0]; |
2192 | 0 | eee[1] = eeee[1] + eeeo[1]; |
2193 | 0 | eee[2] = eeee[1] - eeeo[1]; |
2194 | 0 | for(k = 0; k < 4; k++) |
2195 | 0 | { |
2196 | 0 | ee[k] = eee[k] + eeo[k]; |
2197 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
2198 | 0 | } |
2199 | 0 | for(k = 0; k < 8; k++) |
2200 | 0 | { |
2201 | 0 | e[k] = ee[k] + eo[k]; |
2202 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
2203 | 0 | } |
2204 | 0 | for(k = 0; k < 16; k++) |
2205 | 0 | { |
2206 | 0 | pi2_dst[k] = |
2207 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
2208 | |
|
2209 | 0 | pi2_dst[k + 16] = |
2210 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
2211 | 0 | } |
2212 | 0 | pi2_tmp++; |
2213 | 0 | pi2_dst += dst_strd; |
2214 | 0 | } |
2215 | 0 | } |
2216 | 0 | else /* All rows of output of 1st stage are non-zero */ |
2217 | 0 | { |
2218 | 0 | for(j = 0; j < trans_size; j++) |
2219 | 0 | { |
2220 | | /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */ |
2221 | 0 | for(k = 0; k < 16; k++) |
2222 | 0 | { |
2223 | 0 | o[k] = g_ai2_ihevc_trans_32[1][k] * pi2_tmp[trans_size] |
2224 | 0 | + g_ai2_ihevc_trans_32[3][k] |
2225 | 0 | * pi2_tmp[3 * trans_size] |
2226 | 0 | + g_ai2_ihevc_trans_32[5][k] |
2227 | 0 | * pi2_tmp[5 * trans_size] |
2228 | 0 | + g_ai2_ihevc_trans_32[7][k] |
2229 | 0 | * pi2_tmp[7 * trans_size] |
2230 | 0 | + g_ai2_ihevc_trans_32[9][k] |
2231 | 0 | * pi2_tmp[9 * trans_size] |
2232 | 0 | + g_ai2_ihevc_trans_32[11][k] |
2233 | 0 | * pi2_tmp[11 * trans_size] |
2234 | 0 | + g_ai2_ihevc_trans_32[13][k] |
2235 | 0 | * pi2_tmp[13 * trans_size] |
2236 | 0 | + g_ai2_ihevc_trans_32[15][k] |
2237 | 0 | * pi2_tmp[15 * trans_size] |
2238 | 0 | + g_ai2_ihevc_trans_32[17][k] |
2239 | 0 | * pi2_tmp[17 * trans_size] |
2240 | 0 | + g_ai2_ihevc_trans_32[19][k] |
2241 | 0 | * pi2_tmp[19 * trans_size] |
2242 | 0 | + g_ai2_ihevc_trans_32[21][k] |
2243 | 0 | * pi2_tmp[21 * trans_size] |
2244 | 0 | + g_ai2_ihevc_trans_32[23][k] |
2245 | 0 | * pi2_tmp[23 * trans_size] |
2246 | 0 | + g_ai2_ihevc_trans_32[25][k] |
2247 | 0 | * pi2_tmp[25 * trans_size] |
2248 | 0 | + g_ai2_ihevc_trans_32[27][k] |
2249 | 0 | * pi2_tmp[27 * trans_size] |
2250 | 0 | + g_ai2_ihevc_trans_32[29][k] |
2251 | 0 | * pi2_tmp[29 * trans_size] |
2252 | 0 | + g_ai2_ihevc_trans_32[31][k] |
2253 | 0 | * pi2_tmp[31 * trans_size]; |
2254 | 0 | } |
2255 | 0 | for(k = 0; k < 8; k++) |
2256 | 0 | { |
2257 | 0 | eo[k] = g_ai2_ihevc_trans_32[2][k] * pi2_tmp[2 * trans_size] |
2258 | 0 | + g_ai2_ihevc_trans_32[6][k] |
2259 | 0 | * pi2_tmp[6 * trans_size] |
2260 | 0 | + g_ai2_ihevc_trans_32[10][k] |
2261 | 0 | * pi2_tmp[10 * trans_size] |
2262 | 0 | + g_ai2_ihevc_trans_32[14][k] |
2263 | 0 | * pi2_tmp[14 * trans_size] |
2264 | 0 | + g_ai2_ihevc_trans_32[18][k] |
2265 | 0 | * pi2_tmp[18 * trans_size] |
2266 | 0 | + g_ai2_ihevc_trans_32[22][k] |
2267 | 0 | * pi2_tmp[22 * trans_size] |
2268 | 0 | + g_ai2_ihevc_trans_32[26][k] |
2269 | 0 | * pi2_tmp[26 * trans_size] |
2270 | 0 | + g_ai2_ihevc_trans_32[30][k] |
2271 | 0 | * pi2_tmp[30 * trans_size]; |
2272 | 0 | } |
2273 | 0 | for(k = 0; k < 4; k++) |
2274 | 0 | { |
2275 | 0 | eeo[k] = g_ai2_ihevc_trans_32[4][k] * pi2_tmp[4 * trans_size] |
2276 | 0 | + g_ai2_ihevc_trans_32[12][k] |
2277 | 0 | * pi2_tmp[12 * trans_size] |
2278 | 0 | + g_ai2_ihevc_trans_32[20][k] |
2279 | 0 | * pi2_tmp[20 * trans_size] |
2280 | 0 | + g_ai2_ihevc_trans_32[28][k] |
2281 | 0 | * pi2_tmp[28 * trans_size]; |
2282 | 0 | } |
2283 | 0 | eeeo[0] = |
2284 | 0 | g_ai2_ihevc_trans_32[8][0] * pi2_tmp[8 * trans_size] |
2285 | 0 | + g_ai2_ihevc_trans_32[24][0] |
2286 | 0 | * pi2_tmp[24 |
2287 | 0 | * trans_size]; |
2288 | 0 | eeeo[1] = |
2289 | 0 | g_ai2_ihevc_trans_32[8][1] * pi2_tmp[8 * trans_size] |
2290 | 0 | + g_ai2_ihevc_trans_32[24][1] |
2291 | 0 | * pi2_tmp[24 |
2292 | 0 | * trans_size]; |
2293 | 0 | eeee[0] = |
2294 | 0 | g_ai2_ihevc_trans_32[0][0] * pi2_tmp[0] |
2295 | 0 | + g_ai2_ihevc_trans_32[16][0] |
2296 | 0 | * pi2_tmp[16 |
2297 | 0 | * trans_size]; |
2298 | 0 | eeee[1] = |
2299 | 0 | g_ai2_ihevc_trans_32[0][1] * pi2_tmp[0] |
2300 | 0 | + g_ai2_ihevc_trans_32[16][1] |
2301 | 0 | * pi2_tmp[16 |
2302 | 0 | * trans_size]; |
2303 | | |
2304 | | /* Combining e and o terms at each hierarchy levels to calculate the final spatial domain vector */ |
2305 | 0 | eee[0] = eeee[0] + eeeo[0]; |
2306 | 0 | eee[3] = eeee[0] - eeeo[0]; |
2307 | 0 | eee[1] = eeee[1] + eeeo[1]; |
2308 | 0 | eee[2] = eeee[1] - eeeo[1]; |
2309 | 0 | for(k = 0; k < 4; k++) |
2310 | 0 | { |
2311 | 0 | ee[k] = eee[k] + eeo[k]; |
2312 | 0 | ee[k + 4] = eee[3 - k] - eeo[3 - k]; |
2313 | 0 | } |
2314 | 0 | for(k = 0; k < 8; k++) |
2315 | 0 | { |
2316 | 0 | e[k] = ee[k] + eo[k]; |
2317 | 0 | e[k + 8] = ee[7 - k] - eo[7 - k]; |
2318 | 0 | } |
2319 | 0 | for(k = 0; k < 16; k++) |
2320 | 0 | { |
2321 | 0 | pi2_dst[k] = |
2322 | 0 | CLIP_S16(((e[k] + o[k] + add) >> shift)); |
2323 | |
|
2324 | 0 | pi2_dst[k + 16] = |
2325 | 0 | CLIP_S16(((e[15 - k] - o[15 - k] + add) >> shift)); |
2326 | 0 | } |
2327 | 0 | pi2_tmp++; |
2328 | 0 | pi2_dst += dst_strd; |
2329 | 0 | } |
2330 | 0 | } |
2331 | | /************************************************************************************************/ |
2332 | | /************************************END - IT_RECON_32x32****************************************/ |
2333 | | /************************************************************************************************/ |
2334 | 0 | } |
2335 | 0 | } |
2336 | | |
2337 | | |
2338 | | void ihevc_res_4x4_rotate(WORD16 *pi2_src, |
2339 | | WORD16 *pi2_dst, |
2340 | | WORD32 src_strd, |
2341 | | WORD32 dst_strd, |
2342 | | WORD32 zero_cols) |
2343 | 0 | { |
2344 | 0 | WORD32 i, j; |
2345 | 0 | WORD32 trans_size; |
2346 | |
|
2347 | 0 | trans_size = TRANS_SIZE_4; |
2348 | |
|
2349 | 0 | WORD32 offset = trans_size * src_strd - 1; |
2350 | |
|
2351 | 0 | zero_cols = gau4_ihevcd_4_bit_reverse[zero_cols & 0xF]; |
2352 | |
|
2353 | 0 | for(i = 0; i < trans_size; i++) |
2354 | 0 | { |
2355 | | /* Checking for Zero Cols */ |
2356 | 0 | if((zero_cols & 1) == 1) |
2357 | 0 | { |
2358 | 0 | for(j = 0; j < trans_size; j++) |
2359 | 0 | { |
2360 | 0 | pi2_dst[j * dst_strd] = 0; |
2361 | 0 | } |
2362 | 0 | } |
2363 | 0 | else |
2364 | 0 | { |
2365 | 0 | for(j = 0; j < trans_size; j++) |
2366 | 0 | { |
2367 | 0 | pi2_dst[j * dst_strd] = pi2_src[offset - (j * src_strd + i)]; |
2368 | 0 | } |
2369 | 0 | } |
2370 | 0 | pi2_dst++; |
2371 | 0 | zero_cols = zero_cols >> 1; |
2372 | 0 | } |
2373 | 0 | } |
2374 | | |
2375 | | |
2376 | | void ihevc_res_nxn_copy(WORD16 *pi2_src, |
2377 | | WORD16 *pi2_dst, |
2378 | | WORD32 src_strd, |
2379 | | WORD32 dst_strd, |
2380 | | WORD32 trans_size, |
2381 | | WORD32 zero_cols) |
2382 | 0 | { |
2383 | |
|
2384 | 0 | WORD32 i, j; |
2385 | |
|
2386 | 0 | for(i = 0; i < trans_size; i++) |
2387 | 0 | { |
2388 | | /* Checking for Zero Cols */ |
2389 | 0 | if((zero_cols & 1) == 1) |
2390 | 0 | { |
2391 | 0 | for(j = 0; j < trans_size; j++) |
2392 | 0 | { |
2393 | 0 | pi2_dst[j * dst_strd] = 0; |
2394 | 0 | } |
2395 | 0 | } |
2396 | 0 | else |
2397 | 0 | { |
2398 | 0 | for(j = 0; j < trans_size; j++) |
2399 | 0 | { |
2400 | 0 | pi2_dst[j * dst_strd] = pi2_src[(j * src_strd + i)]; |
2401 | 0 | } |
2402 | 0 | } |
2403 | 0 | pi2_dst++; |
2404 | 0 | zero_cols = zero_cols >> 1; |
2405 | 0 | } |
2406 | 0 | } |
2407 | | |
2408 | | |
2409 | | void ihevc_res_nxn_rdpcm_horz(WORD16 *pi2_src, |
2410 | | WORD16 *pi2_dst, |
2411 | | WORD32 src_strd, |
2412 | | WORD32 dst_strd, |
2413 | | WORD32 trans_size, |
2414 | | WORD32 zero_cols) |
2415 | 0 | { |
2416 | 0 | WORD32 i, j; |
2417 | | |
2418 | | /* Checking for Zero Cols */ |
2419 | 0 | if((zero_cols & 1) == 1) |
2420 | 0 | { |
2421 | 0 | for(j = 0; j < trans_size; j++) |
2422 | 0 | { |
2423 | 0 | pi2_dst[j * dst_strd] = 0; |
2424 | 0 | } |
2425 | 0 | } |
2426 | 0 | else |
2427 | 0 | { |
2428 | 0 | for(j = 0; j < trans_size; j++) |
2429 | 0 | { |
2430 | 0 | pi2_dst[j * dst_strd] = pi2_src[j * src_strd]; |
2431 | 0 | } |
2432 | 0 | } |
2433 | 0 | pi2_dst++; |
2434 | 0 | zero_cols >>= 1; |
2435 | |
|
2436 | 0 | for(i = 1; i < trans_size; i++) |
2437 | 0 | { |
2438 | | /* Checking for Zero Cols */ |
2439 | 0 | if((zero_cols & 1) == 1) |
2440 | 0 | { |
2441 | 0 | for(j = 0; j < trans_size; j++) |
2442 | 0 | { |
2443 | 0 | pi2_dst[j * dst_strd] = pi2_dst[j * dst_strd - 1]; |
2444 | 0 | } |
2445 | 0 | } |
2446 | 0 | else |
2447 | 0 | { |
2448 | 0 | for(j = 0; j < trans_size; j++) |
2449 | 0 | { |
2450 | 0 | pi2_dst[j * dst_strd] = pi2_src[j * src_strd + i] + pi2_dst[j * dst_strd - 1]; |
2451 | 0 | } |
2452 | 0 | } |
2453 | 0 | pi2_dst++; |
2454 | 0 | zero_cols >>= 1; |
2455 | 0 | } |
2456 | 0 | } |
2457 | | |
2458 | | |
2459 | | void ihevc_res_nxn_rdpcm_vert(WORD16 *pi2_src, |
2460 | | WORD16 *pi2_dst, |
2461 | | WORD32 src_strd, |
2462 | | WORD32 dst_strd, |
2463 | | WORD32 trans_size, |
2464 | | WORD32 zero_cols) |
2465 | 0 | { |
2466 | 0 | WORD32 i, j; |
2467 | |
|
2468 | 0 | for(i = 0; i < trans_size; i++) |
2469 | 0 | { |
2470 | | /* Checking for Zero Cols */ |
2471 | 0 | if((zero_cols & 1) == 1) |
2472 | 0 | { |
2473 | 0 | for(j = 0; j < trans_size; j++) |
2474 | 0 | { |
2475 | 0 | pi2_dst[j * dst_strd] = 0; |
2476 | 0 | } |
2477 | 0 | } |
2478 | 0 | else |
2479 | 0 | { |
2480 | 0 | WORD16 acc = pi2_src[i]; |
2481 | |
|
2482 | 0 | pi2_dst[0] = acc; |
2483 | 0 | for(j = 1; j < trans_size; j++) |
2484 | 0 | { |
2485 | 0 | acc += pi2_src[j * src_strd + i]; |
2486 | 0 | pi2_dst[j * dst_strd] = acc; |
2487 | 0 | } |
2488 | 0 | } |
2489 | 0 | pi2_dst++; |
2490 | 0 | zero_cols = zero_cols >> 1; |
2491 | 0 | } |
2492 | 0 | } |
2493 | | |