1 /*
2 * MPEG video MMX templates
3 *
4 * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 #include <stdint.h>
24
25 #include "libavutil/internal.h"
26 #include "libavutil/mem_internal.h"
27 #include "libavutil/x86/asm.h"
28 #include "libavcodec/mpegutils.h"
29 #include "libavcodec/mpegvideo.h"
30 #include "fdct.h"
31
32 #undef MMREG_WIDTH
33 #undef MM
34 #undef MOVQ
35 #undef SPREADW
36 #undef PMAXW
37 #undef PMAX
38 #undef SAVE_SIGN
39 #undef RESTORE_SIGN
40
41 #if COMPILE_TEMPLATE_SSE2
42 #define MMREG_WIDTH "16"
43 #define MM "%%xmm"
44 #define MOVQ "movdqa"
45 #define SPREADW(a) \
46 "pshuflw $0, "a", "a" \n\t"\
47 "punpcklwd "a", "a" \n\t"
48 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
49 #define PMAX(a,b) \
50 "movhlps "a", "b" \n\t"\
51 PMAXW(b, a)\
52 "pshuflw $0x0E, "a", "b" \n\t"\
53 PMAXW(b, a)\
54 "pshuflw $0x01, "a", "b" \n\t"\
55 PMAXW(b, a)
56 #else
57 #define MMREG_WIDTH "8"
58 #define MM "%%mm"
59 #define MOVQ "movq"
60 #define SPREADW(a) \
61 "punpcklwd "a", "a" \n\t"\
62 "punpcklwd "a", "a" \n\t"
63 #define PMAXW(a,b) \
64 "psubusw "a", "b" \n\t"\
65 "paddw "a", "b" \n\t"
66 #define PMAX(a,b) \
67 "movq "a", "b" \n\t"\
68 "psrlq $32, "a" \n\t"\
69 PMAXW(b, a)\
70 "movq "a", "b" \n\t"\
71 "psrlq $16, "a" \n\t"\
72 PMAXW(b, a)
73
74 #endif
75
76 #if COMPILE_TEMPLATE_SSSE3
77 #define SAVE_SIGN(a,b) \
78 "movdqa "b", "a" \n\t"\
79 "pabsw "b", "b" \n\t"
80 #define RESTORE_SIGN(a,b) \
81 "psignw "a", "b" \n\t"
82 #else
83 #define SAVE_SIGN(a,b) \
84 "pxor "a", "a" \n\t"\
85 "pcmpgtw "b", "a" \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
86 "pxor "a", "b" \n\t"\
87 "psubw "a", "b" \n\t" /* ABS(block[i]) */
88 #define RESTORE_SIGN(a,b) \
89 "pxor "a", "b" \n\t"\
90 "psubw "a", "b" \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
91 #endif
92
RENAME(dct_quantize)93 static int RENAME(dct_quantize)(MpegEncContext *s,
94 int16_t *block, int n,
95 int qscale, int *overflow)
96 {
97 x86_reg last_non_zero_p1;
98 int level=0, q; //=0 is because gcc says uninitialized ...
99 const uint16_t *qmat, *bias;
100 LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
101
102 av_assert2((7&(uintptr_t)(&temp_block[0])) == 0); //did gcc align it correctly?
103
104 //s->fdct (block);
105 RENAME_FDCT(ff_fdct)(block); // cannot be anything else ...
106
107 if(s->dct_error_sum)
108 s->denoise_dct(s, block);
109
110 if (s->mb_intra) {
111 int dummy;
112 if (n < 4){
113 q = s->y_dc_scale;
114 bias = s->q_intra_matrix16[qscale][1];
115 qmat = s->q_intra_matrix16[qscale][0];
116 }else{
117 q = s->c_dc_scale;
118 bias = s->q_chroma_intra_matrix16[qscale][1];
119 qmat = s->q_chroma_intra_matrix16[qscale][0];
120 }
121 /* note: block[0] is assumed to be positive */
122 if (!s->h263_aic) {
123 __asm__ volatile (
124 "mul %%ecx \n\t"
125 : "=d" (level), "=a"(dummy)
126 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
127 );
128 } else
129 /* For AIC we skip quant/dequant of INTRADC */
130 level = (block[0] + 4)>>3;
131
132 block[0]=0; //avoid fake overflow
133 // temp_block[0] = (block[0] + (q >> 1)) / q;
134 last_non_zero_p1 = 1;
135 } else {
136 last_non_zero_p1 = 0;
137 bias = s->q_inter_matrix16[qscale][1];
138 qmat = s->q_inter_matrix16[qscale][0];
139 }
140
141 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
142
143 __asm__ volatile(
144 "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
145 SPREADW(MM"3")
146 "pxor "MM"7, "MM"7 \n\t" // 0
147 "pxor "MM"4, "MM"4 \n\t" // 0
148 MOVQ" (%2), "MM"5 \n\t" // qmat[0]
149 "pxor "MM"6, "MM"6 \n\t"
150 "psubw (%3), "MM"6 \n\t" // -bias[0]
151 "mov $-128, %%"FF_REG_a" \n\t"
152 ".p2align 4 \n\t"
153 "1: \n\t"
154 MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
155 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
156 "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
157 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
158 "por "MM"0, "MM"4 \n\t"
159 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
160 MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
161 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
162 MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
163 MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
164 "pandn "MM"1, "MM"0 \n\t"
165 PMAXW(MM"0", MM"3")
166 "add $"MMREG_WIDTH", %%"FF_REG_a" \n\t"
167 " js 1b \n\t"
168 PMAX(MM"3", MM"0")
169 "movd "MM"3, %%"FF_REG_a" \n\t"
170 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
171 : "+a" (last_non_zero_p1)
172 : "r" (block+64), "r" (qmat), "r" (bias),
173 "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
174 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
175 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
176 );
177 }else{ // FMT_H263
178 __asm__ volatile(
179 "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
180 SPREADW(MM"3")
181 "pxor "MM"7, "MM"7 \n\t" // 0
182 "pxor "MM"4, "MM"4 \n\t" // 0
183 "mov $-128, %%"FF_REG_a" \n\t"
184 ".p2align 4 \n\t"
185 "1: \n\t"
186 MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
187 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
188 MOVQ" (%3, %%"FF_REG_a"), "MM"6 \n\t" // bias[0]
189 "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
190 MOVQ" (%2, %%"FF_REG_a"), "MM"5 \n\t" // qmat[i]
191 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
192 "por "MM"0, "MM"4 \n\t"
193 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
194 MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
195 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
196 MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
197 MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
198 "pandn "MM"1, "MM"0 \n\t"
199 PMAXW(MM"0", MM"3")
200 "add $"MMREG_WIDTH", %%"FF_REG_a" \n\t"
201 " js 1b \n\t"
202 PMAX(MM"3", MM"0")
203 "movd "MM"3, %%"FF_REG_a" \n\t"
204 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
205 : "+a" (last_non_zero_p1)
206 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
207 "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
208 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
209 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
210 );
211 }
212 __asm__ volatile(
213 "movd %1, "MM"1 \n\t" // max_qcoeff
214 SPREADW(MM"1")
215 "psubusw "MM"1, "MM"4 \n\t"
216 "packuswb "MM"4, "MM"4 \n\t"
217 #if COMPILE_TEMPLATE_SSE2
218 "packsswb "MM"4, "MM"4 \n\t"
219 #endif
220 "movd "MM"4, %0 \n\t" // *overflow
221 : "=g" (*overflow)
222 : "g" (s->max_qcoeff)
223 );
224
225 if(s->mb_intra) block[0]= level;
226 else block[0]= temp_block[0];
227
228 if (s->idsp.perm_type == FF_IDCT_PERM_SIMPLE) {
229 if(last_non_zero_p1 <= 1) goto end;
230 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
231 block[0x20] = temp_block[0x10];
232 if(last_non_zero_p1 <= 4) goto end;
233 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
234 block[0x09] = temp_block[0x03];
235 if(last_non_zero_p1 <= 7) goto end;
236 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
237 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
238 if(last_non_zero_p1 <= 11) goto end;
239 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
240 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
241 block[0x0C] = temp_block[0x05];
242 if(last_non_zero_p1 <= 16) goto end;
243 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
244 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
245 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
246 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
247 if(last_non_zero_p1 <= 24) goto end;
248 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
249 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
250 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
251 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
252 if(last_non_zero_p1 <= 32) goto end;
253 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
254 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
255 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
256 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
257 if(last_non_zero_p1 <= 40) goto end;
258 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
259 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
260 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
261 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
262 if(last_non_zero_p1 <= 48) goto end;
263 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
264 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
265 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
266 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
267 if(last_non_zero_p1 <= 56) goto end;
268 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
269 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
270 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
271 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
272 }else if(s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2){
273 if(last_non_zero_p1 <= 1) goto end;
274 block[0x04] = temp_block[0x01];
275 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
276 if(last_non_zero_p1 <= 4) goto end;
277 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
278 block[0x05] = temp_block[0x03];
279 if(last_non_zero_p1 <= 7) goto end;
280 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
281 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
282 if(last_non_zero_p1 <= 11) goto end;
283 block[0x1C] = temp_block[0x19];
284 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
285 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
286 if(last_non_zero_p1 <= 16) goto end;
287 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
288 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
289 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
290 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
291 if(last_non_zero_p1 <= 24) goto end;
292 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
293 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
294 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
295 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
296 if(last_non_zero_p1 <= 32) goto end;
297 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
298 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
299 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
300 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
301 if(last_non_zero_p1 <= 40) goto end;
302 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
303 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
304 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
305 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
306 if(last_non_zero_p1 <= 48) goto end;
307 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
308 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
309 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
310 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
311 if(last_non_zero_p1 <= 56) goto end;
312 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
313 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
314 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
315 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
316 } else if (s->idsp.perm_type == FF_IDCT_PERM_NONE) {
317 if(last_non_zero_p1 <= 1) goto end;
318 block[0x01] = temp_block[0x01];
319 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
320 if(last_non_zero_p1 <= 4) goto end;
321 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
322 block[0x03] = temp_block[0x03];
323 if(last_non_zero_p1 <= 7) goto end;
324 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
325 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
326 if(last_non_zero_p1 <= 11) goto end;
327 block[0x19] = temp_block[0x19];
328 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
329 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
330 if(last_non_zero_p1 <= 16) goto end;
331 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
332 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
333 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
334 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
335 if(last_non_zero_p1 <= 24) goto end;
336 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
337 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
338 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
339 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
340 if(last_non_zero_p1 <= 32) goto end;
341 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
342 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
343 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
344 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
345 if(last_non_zero_p1 <= 40) goto end;
346 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
347 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
348 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
349 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
350 if(last_non_zero_p1 <= 48) goto end;
351 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
352 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
353 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
354 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
355 if(last_non_zero_p1 <= 56) goto end;
356 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
357 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
358 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
359 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
360 } else if (s->idsp.perm_type == FF_IDCT_PERM_TRANSPOSE) {
361 if(last_non_zero_p1 <= 1) goto end;
362 block[0x08] = temp_block[0x01];
363 block[0x01] = temp_block[0x08]; block[0x02] = temp_block[0x10];
364 if(last_non_zero_p1 <= 4) goto end;
365 block[0x09] = temp_block[0x09]; block[0x10] = temp_block[0x02];
366 block[0x18] = temp_block[0x03];
367 if(last_non_zero_p1 <= 7) goto end;
368 block[0x11] = temp_block[0x0A]; block[0x0A] = temp_block[0x11];
369 block[0x03] = temp_block[0x18]; block[0x04] = temp_block[0x20];
370 if(last_non_zero_p1 <= 11) goto end;
371 block[0x0B] = temp_block[0x19];
372 block[0x12] = temp_block[0x12]; block[0x19] = temp_block[0x0B];
373 block[0x20] = temp_block[0x04]; block[0x28] = temp_block[0x05];
374 if(last_non_zero_p1 <= 16) goto end;
375 block[0x21] = temp_block[0x0C]; block[0x1A] = temp_block[0x13];
376 block[0x13] = temp_block[0x1A]; block[0x0C] = temp_block[0x21];
377 block[0x05] = temp_block[0x28]; block[0x06] = temp_block[0x30];
378 block[0x0D] = temp_block[0x29]; block[0x14] = temp_block[0x22];
379 if(last_non_zero_p1 <= 24) goto end;
380 block[0x1B] = temp_block[0x1B]; block[0x22] = temp_block[0x14];
381 block[0x29] = temp_block[0x0D]; block[0x30] = temp_block[0x06];
382 block[0x38] = temp_block[0x07]; block[0x31] = temp_block[0x0E];
383 block[0x2A] = temp_block[0x15]; block[0x23] = temp_block[0x1C];
384 if(last_non_zero_p1 <= 32) goto end;
385 block[0x1C] = temp_block[0x23]; block[0x15] = temp_block[0x2A];
386 block[0x0E] = temp_block[0x31]; block[0x07] = temp_block[0x38];
387 block[0x0F] = temp_block[0x39]; block[0x16] = temp_block[0x32];
388 block[0x1D] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
389 if(last_non_zero_p1 <= 40) goto end;
390 block[0x2B] = temp_block[0x1D]; block[0x32] = temp_block[0x16];
391 block[0x39] = temp_block[0x0F]; block[0x3A] = temp_block[0x17];
392 block[0x33] = temp_block[0x1E]; block[0x2C] = temp_block[0x25];
393 block[0x25] = temp_block[0x2C]; block[0x1E] = temp_block[0x33];
394 if(last_non_zero_p1 <= 48) goto end;
395 block[0x17] = temp_block[0x3A]; block[0x1F] = temp_block[0x3B];
396 block[0x26] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
397 block[0x34] = temp_block[0x26]; block[0x3B] = temp_block[0x1F];
398 block[0x3C] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
399 if(last_non_zero_p1 <= 56) goto end;
400 block[0x2E] = temp_block[0x35]; block[0x27] = temp_block[0x3C];
401 block[0x2F] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
402 block[0x3D] = temp_block[0x2F]; block[0x3E] = temp_block[0x37];
403 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
404 } else {
405 av_log(s, AV_LOG_DEBUG, "s->idsp.perm_type: %d\n",
406 (int)s->idsp.perm_type);
407 av_assert0(s->idsp.perm_type == FF_IDCT_PERM_NONE ||
408 s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2 ||
409 s->idsp.perm_type == FF_IDCT_PERM_SIMPLE ||
410 s->idsp.perm_type == FF_IDCT_PERM_TRANSPOSE);
411 }
412 end:
413 return last_non_zero_p1 - 1;
414 }
415