Mercurial > libavcodec.hg
annotate i386/mpegvideo_mmx.c @ 200:6ab301aaa652 libavcodec
(commit by michael)
dequantizers skip trailing zeros
msmpeg4 non-intra decoder has its dequantizer "build in" now
author | arpi_esp |
---|---|
date | Sun, 13 Jan 2002 04:59:37 +0000 |
parents | acbd3bc999b3 |
children | 994aa8623443 |
rev | line source |
---|---|
8 | 1 /* |
2 * The simplest mpeg encoder (well, it was the simplest!) | |
3 * Copyright (c) 2000,2001 Gerard Lantau. | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
18 * | |
19 * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru> | |
200 | 20 * h263 dequantizer by Michael Niedermayer <michaelni@gmx.at> |
8 | 21 */ |
22 | |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
23 #include "../dsputil.h" |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
24 #include "../mpegvideo.h" |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
25 |
200 | 26 extern UINT8 zigzag_end[64]; |
27 | |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
28 #if 0 |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
29 |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
30 /* XXX: GL: I don't understand why this function needs optimization |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
31 (it is called only once per frame!), so I disabled it */ |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
32 |
8 | 33 void MPV_frame_start(MpegEncContext *s) |
34 { | |
35 if (s->pict_type == B_TYPE) { | |
36 __asm __volatile( | |
37 "movl (%1), %%eax\n\t" | |
38 "movl 4(%1), %%edx\n\t" | |
39 "movl 8(%1), %%ecx\n\t" | |
40 "movl %%eax, (%0)\n\t" | |
41 "movl %%edx, 4(%0)\n\t" | |
42 "movl %%ecx, 8(%0)\n\t" | |
43 : | |
44 :"r"(s->current_picture), "r"(s->aux_picture) | |
45 :"eax","edx","ecx","memory"); | |
46 } else { | |
47 /* swap next and last */ | |
48 __asm __volatile( | |
49 "movl (%1), %%eax\n\t" | |
50 "movl 4(%1), %%edx\n\t" | |
51 "movl 8(%1), %%ecx\n\t" | |
52 "xchgl (%0), %%eax\n\t" | |
53 "xchgl 4(%0), %%edx\n\t" | |
54 "xchgl 8(%0), %%ecx\n\t" | |
55 "movl %%eax, (%1)\n\t" | |
56 "movl %%edx, 4(%1)\n\t" | |
57 "movl %%ecx, 8(%1)\n\t" | |
58 "movl %%eax, (%2)\n\t" | |
59 "movl %%edx, 4(%2)\n\t" | |
60 "movl %%ecx, 8(%2)\n\t" | |
61 : | |
62 :"r"(s->last_picture), "r"(s->next_picture), "r"(s->current_picture) | |
63 :"eax","edx","ecx","memory"); | |
64 } | |
65 } | |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
66 #endif |
8 | 67 |
68 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL; | |
69 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL; | |
70 | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
71 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
72 static void dct_unquantize_h263_mmx(MpegEncContext *s, |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
73 DCTELEM *block, int n, int qscale) |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
74 { |
200 | 75 int i, level, qmul, qadd, nCoeffs; |
76 | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
77 qmul = s->qscale << 1; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
78 qadd = (s->qscale - 1) | 1; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
79 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
80 if (s->mb_intra) { |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
81 if (n < 4) |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
82 block[0] = block[0] * s->y_dc_scale; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
83 else |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
84 block[0] = block[0] * s->c_dc_scale; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
85 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
86 for(i=1; i<8; i++) { |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
87 level = block[i]; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
88 if (level) { |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
89 if (level < 0) { |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
90 level = level * qmul - qadd; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
91 } else { |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
92 level = level * qmul + qadd; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
93 } |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
94 block[i] = level; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
95 } |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
96 } |
200 | 97 nCoeffs=64; |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
98 } else { |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
99 i = 0; |
200 | 100 nCoeffs= zigzag_end[ s->block_last_index[n] ]; |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
101 } |
200 | 102 //printf("%d %d ", qmul, qadd); |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
103 asm volatile( |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
104 "movd %1, %%mm6 \n\t" //qmul |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
105 "packssdw %%mm6, %%mm6 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
106 "packssdw %%mm6, %%mm6 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
107 "movd %2, %%mm5 \n\t" //qadd |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
108 "pxor %%mm7, %%mm7 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
109 "packssdw %%mm5, %%mm5 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
110 "packssdw %%mm5, %%mm5 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
111 "psubw %%mm5, %%mm7 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
112 "pxor %%mm4, %%mm4 \n\t" |
153 | 113 ".balign 16\n\t" |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
114 "1: \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
115 "movq (%0, %3), %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
116 "movq 8(%0, %3), %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
117 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
118 "pmullw %%mm6, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
119 "pmullw %%mm6, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
120 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
121 "movq (%0, %3), %%mm2 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
122 "movq 8(%0, %3), %%mm3 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
123 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
124 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
125 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
126 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
127 "pxor %%mm2, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
128 "pxor %%mm3, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
129 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
130 "paddw %%mm7, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
131 "paddw %%mm7, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
132 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
133 "pxor %%mm0, %%mm2 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
134 "pxor %%mm1, %%mm3 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
135 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
136 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
137 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
138 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
139 "pandn %%mm2, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
140 "pandn %%mm3, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
141 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
142 "movq %%mm0, (%0, %3) \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
143 "movq %%mm1, 8(%0, %3) \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
144 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
145 "addl $16, %3 \n\t" |
200 | 146 "js 1b \n\t" |
147 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(i-nCoeffs)) | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
148 : "memory" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
149 ); |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
150 } |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
151 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
152 |
8 | 153 /* |
154 NK: | |
155 Note: looking at PARANOID: | |
156 "enable all paranoid tests for rounding, overflows, etc..." | |
157 | |
158 #ifdef PARANOID | |
159 if (level < -2048 || level > 2047) | |
160 fprintf(stderr, "unquant error %d %d\n", i, level); | |
161 #endif | |
162 We can suppose that result of two multiplications can't be greate of 0xFFFF | |
163 i.e. is 16-bit, so we use here only PMULLW instruction and can avoid | |
164 a complex multiplication. | |
165 ===================================================== | |
166 Full formula for multiplication of 2 integer numbers | |
167 which are represent as high:low words: | |
168 input: value1 = high1:low1 | |
169 value2 = high2:low2 | |
170 output: value3 = value1*value2 | |
171 value3=high3:low3 (on overflow: modulus 2^32 wrap-around) | |
172 this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4 | |
173 but this algorithm will compute only 0x66cb0ce4 | |
174 this limited by 16-bit size of operands | |
175 --------------------------------- | |
176 tlow1 = high1*low2 | |
177 tlow2 = high2*low1 | |
178 tlow1 = tlow1 + tlow2 | |
179 high3:low3 = low1*low2 | |
180 high3 += tlow1 | |
181 */ | |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
182 static void dct_unquantize_mpeg1_mmx(MpegEncContext *s, |
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
183 DCTELEM *block, int n, int qscale) |
8 | 184 { |
200 | 185 int i, level, nCoeffs; |
8 | 186 const UINT16 *quant_matrix; |
200 | 187 |
188 if(s->alternate_scan) nCoeffs= 64; | |
189 else nCoeffs= nCoeffs= zigzag_end[ s->block_last_index[n] ]; | |
190 | |
8 | 191 if (s->mb_intra) { |
192 if (n < 4) | |
193 block[0] = block[0] * s->y_dc_scale; | |
194 else | |
195 block[0] = block[0] * s->c_dc_scale; | |
200 | 196 /* isnt used anymore (we have a h263 unquantizer since some time) |
197 if (s->out_format == FMT_H263) { | |
8 | 198 i = 1; |
199 goto unquant_even; | |
200 | 200 }*/ |
8 | 201 /* XXX: only mpeg1 */ |
202 quant_matrix = s->intra_matrix; | |
203 i=1; | |
204 /* Align on 4 elements boundary */ | |
205 while(i&3) | |
206 { | |
207 level = block[i]; | |
208 if (level) { | |
209 if (level < 0) level = -level; | |
210 level = (int)(level * qscale * quant_matrix[i]) >> 3; | |
211 level = (level - 1) | 1; | |
212 if (block[i] < 0) level = -level; | |
213 block[i] = level; | |
214 } | |
215 i++; | |
216 } | |
217 __asm __volatile( | |
218 "movd %0, %%mm6\n\t" /* mm6 = qscale | 0 */ | |
219 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = qscale | qscale */ | |
220 "movq %2, %%mm4\n\t" | |
221 "movq %%mm6, %%mm7\n\t" | |
222 "movq %1, %%mm5\n\t" | |
223 "packssdw %%mm6, %%mm7\n\t" /* mm7 = qscale | qscale | qscale | qscale */ | |
224 "pxor %%mm6, %%mm6\n\t" | |
225 ::"g"(qscale),"m"(mm_wone),"m"(mm_wabs):"memory"); | |
200 | 226 for(;i<nCoeffs;i+=4) { |
8 | 227 __asm __volatile( |
228 "movq %1, %%mm0\n\t" | |
229 "movq %%mm7, %%mm1\n\t" | |
230 "movq %%mm0, %%mm2\n\t" | |
231 "movq %%mm0, %%mm3\n\t" | |
232 "pcmpgtw %%mm6, %%mm2\n\t" | |
233 "pmullw %2, %%mm1\n\t" | |
234 "pandn %%mm4, %%mm2\n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
235 "por %%mm5, %%mm2\n\t" |
8 | 236 "pmullw %%mm2, %%mm0\n\t" /* mm0 = abs(block[i]). */ |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
237 |
8 | 238 "pcmpeqw %%mm6, %%mm3\n\t" |
239 "pmullw %%mm0, %%mm1\n\t" | |
240 "psraw $3, %%mm1\n\t" | |
241 "psubw %%mm5, %%mm1\n\t" /* block[i] --; */ | |
242 "pandn %%mm4, %%mm3\n\t" /* fake of pcmpneqw : mm0 != 0 then mm1 = -1 */ | |
243 "por %%mm5, %%mm1\n\t" /* block[i] |= 1 */ | |
244 "pmullw %%mm2, %%mm1\n\t" /* change signs again */ | |
245 | |
246 "pand %%mm3, %%mm1\n\t" /* nullify if was zero */ | |
247 "movq %%mm1, %0" | |
248 :"=m"(block[i]) | |
249 :"m"(block[i]), "m"(quant_matrix[i]) | |
250 :"memory"); | |
251 } | |
252 } else { | |
253 i = 0; | |
254 unquant_even: | |
255 quant_matrix = s->non_intra_matrix; | |
256 /* Align on 4 elements boundary */ | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
257 while(i&7) |
8 | 258 { |
259 level = block[i]; | |
260 if (level) { | |
261 if (level < 0) level = -level; | |
262 level = (((level << 1) + 1) * qscale * | |
263 ((int) quant_matrix[i])) >> 4; | |
264 level = (level - 1) | 1; | |
265 if(block[i] < 0) level = -level; | |
266 block[i] = level; | |
267 } | |
268 i++; | |
269 } | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
270 asm volatile( |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
271 "pcmpeqw %%mm7, %%mm7 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
272 "psrlw $15, %%mm7 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
273 "movd %2, %%mm6 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
274 "packssdw %%mm6, %%mm6 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
275 "packssdw %%mm6, %%mm6 \n\t" |
153 | 276 ".balign 16\n\t" |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
277 "1: \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
278 "movq (%0, %3), %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
279 "movq 8(%0, %3), %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
280 "movq (%1, %3), %%mm4 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
281 "movq 8(%1, %3), %%mm5 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
282 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
283 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
284 "pxor %%mm2, %%mm2 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
285 "pxor %%mm3, %%mm3 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
286 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
287 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
288 "pxor %%mm2, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
289 "pxor %%mm3, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
290 "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
291 "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
292 "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
293 "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
294 "paddw %%mm7, %%mm0 \n\t" // abs(block[i])*2 + 1 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
295 "paddw %%mm7, %%mm1 \n\t" // abs(block[i])*2 + 1 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
296 "pmullw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
297 "pmullw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
298 "pxor %%mm4, %%mm4 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
299 "pxor %%mm5, %%mm5 \n\t" // FIXME slow |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
300 "pcmpeqw (%0, %3), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
301 "pcmpeqw 8(%0, %3), %%mm5 \n\t" // block[i] == 0 ? -1 : 0 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
302 "psraw $4, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
303 "psraw $4, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
304 "psubw %%mm7, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
305 "psubw %%mm7, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
306 "por %%mm7, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
307 "por %%mm7, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
308 "pxor %%mm2, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
309 "pxor %%mm3, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
310 "psubw %%mm2, %%mm0 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
311 "psubw %%mm3, %%mm1 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
312 "pandn %%mm0, %%mm4 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
313 "pandn %%mm1, %%mm5 \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
314 "movq %%mm4, (%0, %3) \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
315 "movq %%mm5, 8(%0, %3) \n\t" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
316 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
317 "addl $16, %3 \n\t" |
200 | 318 "js 1b \n\t" |
319 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (2*(i-nCoeffs)) | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
320 : "memory" |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
321 ); |
8 | 322 } |
323 } | |
324 | |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
325 void MPV_common_init_mmx(MpegEncContext *s) |
8 | 326 { |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
327 if (mm_flags & MM_MMX) { |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
328 if (s->out_format == FMT_H263) |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
329 s->dct_unquantize = dct_unquantize_h263_mmx; |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
330 else |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
331 s->dct_unquantize = dct_unquantize_mpeg1_mmx; |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
332 } |
8 | 333 } |