annotate i386/mpegvideo_mmx_template.c @ 2558:2b01396ab483 libavcodec

optimize & simplify
author michael
date Wed, 09 Mar 2005 23:48:17 +0000
parents 5e5cf598a48b
children 41315d0120b3
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
1 /*
429
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
2 * MPEG video MMX templates
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
3 *
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
4 * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
5 *
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
6 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
8 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
9 * version 2 of the License, or (at your option) any later version.
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
10 *
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
11 * This library is distributed in the hope that it will be useful,
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
14 * Lesser General Public License for more details.
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
15 *
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
17 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
718a22dc121f license/copyright change
glantau
parents: 350
diff changeset
19 */
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
20 #undef SPREADW
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
21 #undef PMAXW
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
22 #ifdef HAVE_MMX2
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
23 #define SPREADW(a) "pshufw $0, " #a ", " #a " \n\t"
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
24 #define PMAXW(a,b) "pmaxsw " #a ", " #b " \n\t"
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
25
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
26 #else
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
27 #define SPREADW(a) \
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
28 "punpcklwd " #a ", " #a " \n\t"\
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
29 "punpcklwd " #a ", " #a " \n\t"
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
30 #define PMAXW(a,b) \
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
31 "psubusw " #a ", " #b " \n\t"\
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
32 "paddw " #a ", " #b " \n\t"
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
33 #endif
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
34
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
35 static int RENAME(dct_quantize)(MpegEncContext *s,
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
36 DCTELEM *block, int n,
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
37 int qscale, int *overflow)
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
38 {
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
39 long last_non_zero_p1;
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
40 int level=0, q; //=0 is cuz gcc says uninitalized ...
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1013
diff changeset
41 const uint16_t *qmat, *bias;
1403
bd42cc06ceb7 fixing thread safety, this may cause alignment problms (if it does than the stuff should be moved to MpegEncContext)
michaelni
parents: 1251
diff changeset
42 __align8 int16_t temp_block[64];
bd42cc06ceb7 fixing thread safety, this may cause alignment problms (if it does than the stuff should be moved to MpegEncContext)
michaelni
parents: 1251
diff changeset
43
bd42cc06ceb7 fixing thread safety, this may cause alignment problms (if it does than the stuff should be moved to MpegEncContext)
michaelni
parents: 1251
diff changeset
44 assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
45
625
bb6a69f9d409 slow but accurate integer dct from IJG (should be ok with the LGPL as the old DCT is the fast integer DCT from IJG)
michaelni
parents: 599
diff changeset
46 //s->fdct (block);
1565
1a9a63f59849 minor mmx2 optimization if the dct
michael
parents: 1554
diff changeset
47 RENAMEl(ff_fdct) (block); //cant be anything else ...
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
48
1597
4c9165372ab3 noise reduction of dct coefficients
michael
parents: 1565
diff changeset
49 if(s->dct_error_sum)
1719
4e72fb256b25 denoise_dct_mmx()
michael
parents: 1597
diff changeset
50 s->denoise_dct(s, block);
1597
4c9165372ab3 noise reduction of dct coefficients
michael
parents: 1565
diff changeset
51
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
52 if (s->mb_intra) {
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
53 int dummy;
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
54 if (n < 4)
641
eaa9ef2e2557 fixing encoding at high QPs (was broken after the dct accuracy increase due to too small inverse table)
michaelni
parents: 635
diff changeset
55 q = s->y_dc_scale;
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
56 else
641
eaa9ef2e2557 fixing encoding at high QPs (was broken after the dct accuracy increase due to too small inverse table)
michaelni
parents: 635
diff changeset
57 q = s->c_dc_scale;
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
58 /* note: block[0] is assumed to be positive */
350
6ebbecc10063 - Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents: 344
diff changeset
59 if (!s->h263_aic) {
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
60 #if 1
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
61 asm volatile (
2115
4ea05f23730b dc_precission=11 fix
michael
parents: 1719
diff changeset
62 "mul %%ecx \n\t"
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
63 : "=d" (level), "=a"(dummy)
641
eaa9ef2e2557 fixing encoding at high QPs (was broken after the dct accuracy increase due to too small inverse table)
michaelni
parents: 635
diff changeset
64 : "a" ((block[0]>>2) + q), "c" (inverse[q<<1])
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
65 );
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
66 #else
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
67 asm volatile (
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
68 "xorl %%edx, %%edx \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
69 "divw %%cx \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
70 "movzwl %%ax, %%eax \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
71 : "=a" (level)
641
eaa9ef2e2557 fixing encoding at high QPs (was broken after the dct accuracy increase due to too small inverse table)
michaelni
parents: 635
diff changeset
72 : "a" ((block[0]>>2) + q), "c" (q<<1)
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
73 : "%edx"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
74 );
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
75 #endif
350
6ebbecc10063 - Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents: 344
diff changeset
76 } else
6ebbecc10063 - Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents: 344
diff changeset
77 /* For AIC we skip quant/dequant of INTRADC */
641
eaa9ef2e2557 fixing encoding at high QPs (was broken after the dct accuracy increase due to too small inverse table)
michaelni
parents: 635
diff changeset
78 level = (block[0] + 4)>>3;
350
6ebbecc10063 - Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents: 344
diff changeset
79
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
80 block[0]=0; //avoid fake overflow
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
81 // temp_block[0] = (block[0] + (q >> 1)) / q;
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
82 last_non_zero_p1 = 1;
1554
d736e24bf303 move mmx quantizer matrixes out of MpegEncContext (23k -> 7k) (no meassureable slowdown)
michael
parents: 1403
diff changeset
83 bias = s->q_intra_matrix16[qscale][1];
d736e24bf303 move mmx quantizer matrixes out of MpegEncContext (23k -> 7k) (no meassureable slowdown)
michael
parents: 1403
diff changeset
84 qmat = s->q_intra_matrix16[qscale][0];
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
85 } else {
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
86 last_non_zero_p1 = 0;
1554
d736e24bf303 move mmx quantizer matrixes out of MpegEncContext (23k -> 7k) (no meassureable slowdown)
michael
parents: 1403
diff changeset
87 bias = s->q_inter_matrix16[qscale][1];
d736e24bf303 move mmx quantizer matrixes out of MpegEncContext (23k -> 7k) (no meassureable slowdown)
michael
parents: 1403
diff changeset
88 qmat = s->q_inter_matrix16[qscale][0];
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
89 }
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
90
2327
5e5cf598a48b H.261 encoder by (Maarten Daniels <maarten dot daniels at luc dot ac dot be>)
michael
parents: 2293
diff changeset
91 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
92
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
93 asm volatile(
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
94 "movd %%"REG_a", %%mm3 \n\t" // last_non_zero_p1
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
95 SPREADW(%%mm3)
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
96 "pxor %%mm7, %%mm7 \n\t" // 0
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
97 "pxor %%mm4, %%mm4 \n\t" // 0
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
98 "movq (%2), %%mm5 \n\t" // qmat[0]
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
99 "pxor %%mm6, %%mm6 \n\t"
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
100 "psubw (%3), %%mm6 \n\t" // -bias[0]
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
101 "mov $-128, %%"REG_a" \n\t"
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
102 ".balign 16 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
103 "1: \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
104 "pxor %%mm1, %%mm1 \n\t" // 0
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
105 "movq (%1, %%"REG_a"), %%mm0 \n\t" // block[i]
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
106 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
107 "pxor %%mm1, %%mm0 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
108 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
109 "psubusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
110 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
111 "por %%mm0, %%mm4 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
112 "pxor %%mm1, %%mm0 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
113 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
114 "movq %%mm0, (%5, %%"REG_a") \n\t"
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
115 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
116 "movq (%4, %%"REG_a"), %%mm1 \n\t"
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
117 "movq %%mm7, (%1, %%"REG_a") \n\t" // 0
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
118 "pandn %%mm1, %%mm0 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
119 PMAXW(%%mm0, %%mm3)
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
120 "add $8, %%"REG_a" \n\t"
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
121 " js 1b \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
122 "movq %%mm3, %%mm0 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
123 "psrlq $32, %%mm3 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
124 PMAXW(%%mm0, %%mm3)
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
125 "movq %%mm3, %%mm0 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
126 "psrlq $16, %%mm3 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
127 PMAXW(%%mm0, %%mm3)
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
128 "movd %%mm3, %%"REG_a" \n\t"
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
129 "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
130 : "+a" (last_non_zero_p1)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
131 : "r" (block+64), "r" (qmat), "r" (bias),
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
132 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
133 );
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
134 // note the asm is split cuz gcc doesnt like that many operands ...
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
135 asm volatile(
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
136 "movd %1, %%mm1 \n\t" // max_qcoeff
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
137 SPREADW(%%mm1)
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
138 "psubusw %%mm1, %%mm4 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
139 "packuswb %%mm4, %%mm4 \n\t"
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
140 "movd %%mm4, %0 \n\t" // *overflow
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
141 : "=g" (*overflow)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
142 : "g" (s->max_qcoeff)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
143 );
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
144 }else{ // FMT_H263
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
145 asm volatile(
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
146 "movd %%"REG_a", %%mm3 \n\t" // last_non_zero_p1
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
147 SPREADW(%%mm3)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
148 "pxor %%mm7, %%mm7 \n\t" // 0
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
149 "pxor %%mm4, %%mm4 \n\t" // 0
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
150 "mov $-128, %%"REG_a" \n\t"
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
151 ".balign 16 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
152 "1: \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
153 "pxor %%mm1, %%mm1 \n\t" // 0
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
154 "movq (%1, %%"REG_a"), %%mm0 \n\t" // block[i]
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
155 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
156 "pxor %%mm1, %%mm0 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
157 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
158 "movq (%3, %%"REG_a"), %%mm6 \n\t" // bias[0]
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
159 "paddusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
160 "movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
161 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
162 "por %%mm0, %%mm4 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
163 "pxor %%mm1, %%mm0 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
164 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
165 "movq %%mm0, (%5, %%"REG_a") \n\t"
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
166 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
167 "movq (%4, %%"REG_a"), %%mm1 \n\t"
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
168 "movq %%mm7, (%1, %%"REG_a") \n\t" // 0
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
169 "pandn %%mm1, %%mm0 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
170 PMAXW(%%mm0, %%mm3)
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
171 "add $8, %%"REG_a" \n\t"
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
172 " js 1b \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
173 "movq %%mm3, %%mm0 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
174 "psrlq $32, %%mm3 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
175 PMAXW(%%mm0, %%mm3)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
176 "movq %%mm3, %%mm0 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
177 "psrlq $16, %%mm3 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
178 PMAXW(%%mm0, %%mm3)
2293
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
179 "movd %%mm3, %%"REG_a" \n\t"
15cfba1b97b5 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents: 2115
diff changeset
180 "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
1251
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
181 : "+a" (last_non_zero_p1)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
182 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
183 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
184 );
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
185 // note the asm is split cuz gcc doesnt like that many operands ...
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
186 asm volatile(
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
187 "movd %1, %%mm1 \n\t" // max_qcoeff
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
188 SPREADW(%%mm1)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
189 "psubusw %%mm1, %%mm4 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
190 "packuswb %%mm4, %%mm4 \n\t"
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
191 "movd %%mm4, %0 \n\t" // *overflow
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
192 : "=g" (*overflow)
afdd177080c9 * oops - reversed last commit
kabi
parents: 1248
diff changeset
193 : "g" (s->max_qcoeff)
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
194 );
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
195 }
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
196
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
197 if(s->mb_intra) block[0]= level;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
198 else block[0]= temp_block[0];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
199
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
200 if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
201 if(last_non_zero_p1 <= 1) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
202 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
203 block[0x20] = temp_block[0x10];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
204 if(last_non_zero_p1 <= 4) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
205 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
206 block[0x09] = temp_block[0x03];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
207 if(last_non_zero_p1 <= 7) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
208 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
209 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
210 if(last_non_zero_p1 <= 11) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
211 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
212 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
213 block[0x0C] = temp_block[0x05];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
214 if(last_non_zero_p1 <= 16) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
215 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
216 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
217 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
218 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
219 if(last_non_zero_p1 <= 24) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
220 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
221 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
222 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
223 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
224 if(last_non_zero_p1 <= 32) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
225 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
226 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
227 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
228 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
229 if(last_non_zero_p1 <= 40) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
230 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
231 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
232 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
233 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
234 if(last_non_zero_p1 <= 48) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
235 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
236 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
237 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
238 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
239 if(last_non_zero_p1 <= 56) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
240 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
241 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
242 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
243 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
244 }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
245 if(last_non_zero_p1 <= 1) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
246 block[0x04] = temp_block[0x01];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
247 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
248 if(last_non_zero_p1 <= 4) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
249 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
250 block[0x05] = temp_block[0x03];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
251 if(last_non_zero_p1 <= 7) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
252 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
253 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
254 if(last_non_zero_p1 <= 11) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
255 block[0x1C] = temp_block[0x19];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
256 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
257 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
258 if(last_non_zero_p1 <= 16) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
259 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
260 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
261 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
262 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
263 if(last_non_zero_p1 <= 24) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
264 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
265 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
266 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
267 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
268 if(last_non_zero_p1 <= 32) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
269 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
270 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
271 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
272 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
273 if(last_non_zero_p1 <= 40) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
274 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
275 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
276 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
277 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
278 if(last_non_zero_p1 <= 48) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
279 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
280 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
281 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
282 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
283 if(last_non_zero_p1 <= 56) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
284 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
285 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
286 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
287 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
288 }else{
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
289 if(last_non_zero_p1 <= 1) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
290 block[0x01] = temp_block[0x01];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
291 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
292 if(last_non_zero_p1 <= 4) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
293 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
294 block[0x03] = temp_block[0x03];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
295 if(last_non_zero_p1 <= 7) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
296 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
297 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
298 if(last_non_zero_p1 <= 11) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
299 block[0x19] = temp_block[0x19];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
300 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
301 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
302 if(last_non_zero_p1 <= 16) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
303 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
304 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
305 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
306 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
307 if(last_non_zero_p1 <= 24) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
308 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
309 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
310 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
311 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
312 if(last_non_zero_p1 <= 32) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
313 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
314 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
315 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
316 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
317 if(last_non_zero_p1 <= 40) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
318 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
319 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
320 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
321 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
322 if(last_non_zero_p1 <= 48) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
323 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
324 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
325 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
326 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
327 if(last_non_zero_p1 <= 56) goto end;
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
328 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
329 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
330 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
331 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
332 }
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 687
diff changeset
333 end:
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
334 /*
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
335 for(i=0; i<last_non_zero_p1; i++)
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
336 {
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
337 int j= zigzag_direct_noperm[i];
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
338 block[block_permute_op(j)]= temp_block[j];
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
339 }
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
340 */
344
9f6071a87e17 fixed msmpeg4 infinite loop if buggy stream
michaelni
parents: 230
diff changeset
341
220
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
342 return last_non_zero_p1 - 1;
0b234715e205 (commit by michael)
arpi_esp
parents:
diff changeset
343 }