Mercurial > libavcodec.hg
annotate i386/mpegvideo_mmx.c @ 4116:7a08ccdc8550 libavcodec
Use common define for x86_32 and x86_64.
author | diego |
---|---|
date | Wed, 01 Nov 2006 18:34:40 +0000 |
parents | c8c591fe26f8 |
children | 580d2c397251 |
rev | line source |
---|---|
8 | 1 /* |
2 * The simplest mpeg encoder (well, it was the simplest!) | |
429 | 3 * Copyright (c) 2000,2001 Fabrice Bellard. |
8 | 4 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3576
diff
changeset
|
5 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3576
diff
changeset
|
6 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3576
diff
changeset
|
7 * FFmpeg is free software; you can redistribute it and/or |
429 | 8 * modify it under the terms of the GNU Lesser General Public |
9 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3576
diff
changeset
|
10 * version 2.1 of the License, or (at your option) any later version. |
8 | 11 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3576
diff
changeset
|
12 * FFmpeg is distributed in the hope that it will be useful, |
8 | 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
15 * Lesser General Public License for more details. | |
8 | 16 * |
429 | 17 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3576
diff
changeset
|
18 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
8 | 20 * |
21 * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru> | |
325 | 22 * h263, mpeg1, mpeg2 dequantizer & draw_edges by Michael Niedermayer <michaelni@gmx.at> |
8 | 23 */ |
24 | |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
25 #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
|
26 #include "../mpegvideo.h" |
220 | 27 #include "../avcodec.h" |
3398
e0927bc44a10
Move REG_* macros from libavcodec/i386/mmx.h to libavutil/x86_cpu.h
lucabe
parents:
3281
diff
changeset
|
28 #include "x86_cpu.h" |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
29 |
1064 | 30 extern uint8_t zigzag_direct_noperm[64]; |
31 extern uint16_t inv_zigzag_direct16[64]; | |
200 | 32 |
8 | 33 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL; |
34 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL; | |
35 | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
36 |
1689 | 37 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s, |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
38 DCTELEM *block, int n, int qscale) |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
39 { |
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:
2024
diff
changeset
|
40 long level, qmul, qadd, nCoeffs; |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
41 |
706
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
42 qmul = qscale << 1; |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
43 |
1661 | 44 assert(s->block_last_index[n]>=0 || s->h263_aic); |
2967 | 45 |
1689 | 46 if (!s->h263_aic) { |
47 if (n < 4) | |
48 level = block[0] * s->y_dc_scale; | |
49 else | |
50 level = block[0] * s->c_dc_scale; | |
51 qadd = (qscale - 1) | 1; | |
52 }else{ | |
53 qadd = 0; | |
54 level= block[0]; | |
55 } | |
56 if(s->ac_pred) | |
706
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
57 nCoeffs=63; |
1689 | 58 else |
59 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ]; | |
200 | 60 //printf("%d %d ", qmul, qadd); |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
61 asm volatile( |
2979 | 62 "movd %1, %%mm6 \n\t" //qmul |
63 "packssdw %%mm6, %%mm6 \n\t" | |
64 "packssdw %%mm6, %%mm6 \n\t" | |
65 "movd %2, %%mm5 \n\t" //qadd | |
66 "pxor %%mm7, %%mm7 \n\t" | |
67 "packssdw %%mm5, %%mm5 \n\t" | |
68 "packssdw %%mm5, %%mm5 \n\t" | |
69 "psubw %%mm5, %%mm7 \n\t" | |
70 "pxor %%mm4, %%mm4 \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3398
diff
changeset
|
71 ASMALIGN(4) |
2979 | 72 "1: \n\t" |
73 "movq (%0, %3), %%mm0 \n\t" | |
74 "movq 8(%0, %3), %%mm1 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
75 |
2979 | 76 "pmullw %%mm6, %%mm0 \n\t" |
77 "pmullw %%mm6, %%mm1 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
78 |
2979 | 79 "movq (%0, %3), %%mm2 \n\t" |
80 "movq 8(%0, %3), %%mm3 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
81 |
2979 | 82 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 |
83 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
84 |
2979 | 85 "pxor %%mm2, %%mm0 \n\t" |
86 "pxor %%mm3, %%mm1 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
87 |
2979 | 88 "paddw %%mm7, %%mm0 \n\t" |
89 "paddw %%mm7, %%mm1 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
90 |
2979 | 91 "pxor %%mm0, %%mm2 \n\t" |
92 "pxor %%mm1, %%mm3 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
93 |
2979 | 94 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 |
95 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
96 |
2979 | 97 "pandn %%mm2, %%mm0 \n\t" |
98 "pandn %%mm3, %%mm1 \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
99 |
2979 | 100 "movq %%mm0, (%0, %3) \n\t" |
101 "movq %%mm1, 8(%0, %3) \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
102 |
2979 | 103 "add $16, %3 \n\t" |
104 "jng 1b \n\t" | |
105 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs)) | |
106 : "memory" | |
107 ); | |
1689 | 108 block[0]= level; |
109 } | |
110 | |
111 | |
112 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s, | |
113 DCTELEM *block, int n, int qscale) | |
114 { | |
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:
2024
diff
changeset
|
115 long qmul, qadd, nCoeffs; |
1689 | 116 |
117 qmul = qscale << 1; | |
118 qadd = (qscale - 1) | 1; | |
119 | |
120 assert(s->block_last_index[n]>=0 || s->h263_aic); | |
2967 | 121 |
1689 | 122 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ]; |
123 //printf("%d %d ", qmul, qadd); | |
124 asm volatile( | |
2979 | 125 "movd %1, %%mm6 \n\t" //qmul |
126 "packssdw %%mm6, %%mm6 \n\t" | |
127 "packssdw %%mm6, %%mm6 \n\t" | |
128 "movd %2, %%mm5 \n\t" //qadd | |
129 "pxor %%mm7, %%mm7 \n\t" | |
130 "packssdw %%mm5, %%mm5 \n\t" | |
131 "packssdw %%mm5, %%mm5 \n\t" | |
132 "psubw %%mm5, %%mm7 \n\t" | |
133 "pxor %%mm4, %%mm4 \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3398
diff
changeset
|
134 ASMALIGN(4) |
2979 | 135 "1: \n\t" |
136 "movq (%0, %3), %%mm0 \n\t" | |
137 "movq 8(%0, %3), %%mm1 \n\t" | |
1689 | 138 |
2979 | 139 "pmullw %%mm6, %%mm0 \n\t" |
140 "pmullw %%mm6, %%mm1 \n\t" | |
1689 | 141 |
2979 | 142 "movq (%0, %3), %%mm2 \n\t" |
143 "movq 8(%0, %3), %%mm3 \n\t" | |
1689 | 144 |
2979 | 145 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 |
146 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 | |
1689 | 147 |
2979 | 148 "pxor %%mm2, %%mm0 \n\t" |
149 "pxor %%mm3, %%mm1 \n\t" | |
1689 | 150 |
2979 | 151 "paddw %%mm7, %%mm0 \n\t" |
152 "paddw %%mm7, %%mm1 \n\t" | |
1689 | 153 |
2979 | 154 "pxor %%mm0, %%mm2 \n\t" |
155 "pxor %%mm1, %%mm3 \n\t" | |
1689 | 156 |
2979 | 157 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 |
158 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 | |
1689 | 159 |
2979 | 160 "pandn %%mm2, %%mm0 \n\t" |
161 "pandn %%mm3, %%mm1 \n\t" | |
1689 | 162 |
2979 | 163 "movq %%mm0, (%0, %3) \n\t" |
164 "movq %%mm1, 8(%0, %3) \n\t" | |
1689 | 165 |
2979 | 166 "add $16, %3 \n\t" |
167 "jng 1b \n\t" | |
168 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs)) | |
169 : "memory" | |
170 ); | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
171 } |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
172 |
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
173 |
8 | 174 /* |
175 NK: | |
176 Note: looking at PARANOID: | |
177 "enable all paranoid tests for rounding, overflows, etc..." | |
178 | |
179 #ifdef PARANOID | |
180 if (level < -2048 || level > 2047) | |
181 fprintf(stderr, "unquant error %d %d\n", i, level); | |
182 #endif | |
183 We can suppose that result of two multiplications can't be greate of 0xFFFF | |
184 i.e. is 16-bit, so we use here only PMULLW instruction and can avoid | |
185 a complex multiplication. | |
186 ===================================================== | |
187 Full formula for multiplication of 2 integer numbers | |
188 which are represent as high:low words: | |
189 input: value1 = high1:low1 | |
190 value2 = high2:low2 | |
191 output: value3 = value1*value2 | |
192 value3=high3:low3 (on overflow: modulus 2^32 wrap-around) | |
193 this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4 | |
194 but this algorithm will compute only 0x66cb0ce4 | |
195 this limited by 16-bit size of operands | |
196 --------------------------------- | |
197 tlow1 = high1*low2 | |
198 tlow2 = high2*low1 | |
199 tlow1 = tlow1 + tlow2 | |
200 high3:low3 = low1*low2 | |
201 high3 += tlow1 | |
202 */ | |
1689 | 203 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s, |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
204 DCTELEM *block, int n, int qscale) |
8 | 205 { |
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:
2024
diff
changeset
|
206 long nCoeffs; |
1064 | 207 const uint16_t *quant_matrix; |
1689 | 208 int block0; |
706
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
209 |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
210 assert(s->block_last_index[n]>=0); |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
211 |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
212 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1; |
200 | 213 |
2967 | 214 if (n < 4) |
1689 | 215 block0 = block[0] * s->y_dc_scale; |
216 else | |
217 block0 = block[0] * s->c_dc_scale; | |
218 /* XXX: only mpeg1 */ | |
219 quant_matrix = s->intra_matrix; | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
220 asm volatile( |
2979 | 221 "pcmpeqw %%mm7, %%mm7 \n\t" |
222 "psrlw $15, %%mm7 \n\t" | |
223 "movd %2, %%mm6 \n\t" | |
224 "packssdw %%mm6, %%mm6 \n\t" | |
225 "packssdw %%mm6, %%mm6 \n\t" | |
226 "mov %3, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3398
diff
changeset
|
227 ASMALIGN(4) |
2979 | 228 "1: \n\t" |
229 "movq (%0, %%"REG_a"), %%mm0 \n\t" | |
230 "movq 8(%0, %%"REG_a"), %%mm1 \n\t" | |
231 "movq (%1, %%"REG_a"), %%mm4 \n\t" | |
232 "movq 8(%1, %%"REG_a"), %%mm5 \n\t" | |
233 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] | |
234 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] | |
235 "pxor %%mm2, %%mm2 \n\t" | |
236 "pxor %%mm3, %%mm3 \n\t" | |
237 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 | |
238 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 | |
239 "pxor %%mm2, %%mm0 \n\t" | |
240 "pxor %%mm3, %%mm1 \n\t" | |
241 "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) | |
242 "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) | |
243 "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*q | |
244 "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*q | |
245 "pxor %%mm4, %%mm4 \n\t" | |
246 "pxor %%mm5, %%mm5 \n\t" // FIXME slow | |
247 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 | |
248 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 | |
249 "psraw $3, %%mm0 \n\t" | |
250 "psraw $3, %%mm1 \n\t" | |
251 "psubw %%mm7, %%mm0 \n\t" | |
252 "psubw %%mm7, %%mm1 \n\t" | |
253 "por %%mm7, %%mm0 \n\t" | |
254 "por %%mm7, %%mm1 \n\t" | |
255 "pxor %%mm2, %%mm0 \n\t" | |
256 "pxor %%mm3, %%mm1 \n\t" | |
257 "psubw %%mm2, %%mm0 \n\t" | |
258 "psubw %%mm3, %%mm1 \n\t" | |
259 "pandn %%mm0, %%mm4 \n\t" | |
260 "pandn %%mm1, %%mm5 \n\t" | |
261 "movq %%mm4, (%0, %%"REG_a") \n\t" | |
262 "movq %%mm5, 8(%0, %%"REG_a") \n\t" | |
325 | 263 |
2979 | 264 "add $16, %%"REG_a" \n\t" |
265 "js 1b \n\t" | |
266 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs) | |
267 : "%"REG_a, "memory" | |
268 ); | |
1689 | 269 block[0]= block0; |
270 } | |
325 | 271 |
1689 | 272 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s, |
273 DCTELEM *block, int n, int qscale) | |
274 { | |
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:
2024
diff
changeset
|
275 long nCoeffs; |
1689 | 276 const uint16_t *quant_matrix; |
277 | |
278 assert(s->block_last_index[n]>=0); | |
279 | |
280 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1; | |
281 | |
344 | 282 quant_matrix = s->inter_matrix; |
325 | 283 asm volatile( |
2979 | 284 "pcmpeqw %%mm7, %%mm7 \n\t" |
285 "psrlw $15, %%mm7 \n\t" | |
286 "movd %2, %%mm6 \n\t" | |
287 "packssdw %%mm6, %%mm6 \n\t" | |
288 "packssdw %%mm6, %%mm6 \n\t" | |
289 "mov %3, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3398
diff
changeset
|
290 ASMALIGN(4) |
2979 | 291 "1: \n\t" |
292 "movq (%0, %%"REG_a"), %%mm0 \n\t" | |
293 "movq 8(%0, %%"REG_a"), %%mm1 \n\t" | |
294 "movq (%1, %%"REG_a"), %%mm4 \n\t" | |
295 "movq 8(%1, %%"REG_a"), %%mm5 \n\t" | |
296 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] | |
297 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] | |
298 "pxor %%mm2, %%mm2 \n\t" | |
299 "pxor %%mm3, %%mm3 \n\t" | |
300 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 | |
301 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 | |
302 "pxor %%mm2, %%mm0 \n\t" | |
303 "pxor %%mm3, %%mm1 \n\t" | |
304 "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) | |
305 "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) | |
306 "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2 | |
307 "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2 | |
308 "paddw %%mm7, %%mm0 \n\t" // abs(block[i])*2 + 1 | |
309 "paddw %%mm7, %%mm1 \n\t" // abs(block[i])*2 + 1 | |
310 "pmullw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q | |
311 "pmullw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q | |
312 "pxor %%mm4, %%mm4 \n\t" | |
313 "pxor %%mm5, %%mm5 \n\t" // FIXME slow | |
314 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 | |
315 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 | |
316 "psraw $4, %%mm0 \n\t" | |
317 "psraw $4, %%mm1 \n\t" | |
318 "psubw %%mm7, %%mm0 \n\t" | |
319 "psubw %%mm7, %%mm1 \n\t" | |
320 "por %%mm7, %%mm0 \n\t" | |
321 "por %%mm7, %%mm1 \n\t" | |
322 "pxor %%mm2, %%mm0 \n\t" | |
323 "pxor %%mm3, %%mm1 \n\t" | |
324 "psubw %%mm2, %%mm0 \n\t" | |
325 "psubw %%mm3, %%mm1 \n\t" | |
326 "pandn %%mm0, %%mm4 \n\t" | |
327 "pandn %%mm1, %%mm5 \n\t" | |
328 "movq %%mm4, (%0, %%"REG_a") \n\t" | |
329 "movq %%mm5, 8(%0, %%"REG_a") \n\t" | |
325 | 330 |
2979 | 331 "add $16, %%"REG_a" \n\t" |
332 "js 1b \n\t" | |
333 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs) | |
334 : "%"REG_a, "memory" | |
335 ); | |
325 | 336 } |
337 | |
1689 | 338 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s, |
325 | 339 DCTELEM *block, int n, int qscale) |
340 { | |
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:
2024
diff
changeset
|
341 long nCoeffs; |
1064 | 342 const uint16_t *quant_matrix; |
1689 | 343 int block0; |
2967 | 344 |
706
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
345 assert(s->block_last_index[n]>=0); |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
346 |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
347 if(s->alternate_scan) nCoeffs= 63; //FIXME |
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
348 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; |
325 | 349 |
2967 | 350 if (n < 4) |
1689 | 351 block0 = block[0] * s->y_dc_scale; |
352 else | |
353 block0 = block[0] * s->c_dc_scale; | |
354 quant_matrix = s->intra_matrix; | |
325 | 355 asm volatile( |
2979 | 356 "pcmpeqw %%mm7, %%mm7 \n\t" |
357 "psrlw $15, %%mm7 \n\t" | |
358 "movd %2, %%mm6 \n\t" | |
359 "packssdw %%mm6, %%mm6 \n\t" | |
360 "packssdw %%mm6, %%mm6 \n\t" | |
361 "mov %3, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3398
diff
changeset
|
362 ASMALIGN(4) |
2979 | 363 "1: \n\t" |
364 "movq (%0, %%"REG_a"), %%mm0 \n\t" | |
365 "movq 8(%0, %%"REG_a"), %%mm1 \n\t" | |
366 "movq (%1, %%"REG_a"), %%mm4 \n\t" | |
367 "movq 8(%1, %%"REG_a"), %%mm5 \n\t" | |
368 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] | |
369 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] | |
370 "pxor %%mm2, %%mm2 \n\t" | |
371 "pxor %%mm3, %%mm3 \n\t" | |
372 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 | |
373 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 | |
374 "pxor %%mm2, %%mm0 \n\t" | |
375 "pxor %%mm3, %%mm1 \n\t" | |
376 "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) | |
377 "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) | |
378 "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*q | |
379 "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*q | |
380 "pxor %%mm4, %%mm4 \n\t" | |
381 "pxor %%mm5, %%mm5 \n\t" // FIXME slow | |
382 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 | |
383 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 | |
384 "psraw $3, %%mm0 \n\t" | |
385 "psraw $3, %%mm1 \n\t" | |
386 "pxor %%mm2, %%mm0 \n\t" | |
387 "pxor %%mm3, %%mm1 \n\t" | |
388 "psubw %%mm2, %%mm0 \n\t" | |
389 "psubw %%mm3, %%mm1 \n\t" | |
390 "pandn %%mm0, %%mm4 \n\t" | |
391 "pandn %%mm1, %%mm5 \n\t" | |
392 "movq %%mm4, (%0, %%"REG_a") \n\t" | |
393 "movq %%mm5, 8(%0, %%"REG_a") \n\t" | |
145
bd1adece8280
dct_unquantize_h263_mmx() by Michael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
14
diff
changeset
|
394 |
2979 | 395 "add $16, %%"REG_a" \n\t" |
396 "jng 1b \n\t" | |
397 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs) | |
398 : "%"REG_a, "memory" | |
399 ); | |
1689 | 400 block[0]= block0; |
325 | 401 //Note, we dont do mismatch control for intra as errors cannot accumulate |
1689 | 402 } |
325 | 403 |
1689 | 404 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s, |
405 DCTELEM *block, int n, int qscale) | |
406 { | |
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:
2024
diff
changeset
|
407 long nCoeffs; |
1689 | 408 const uint16_t *quant_matrix; |
2967 | 409 |
1689 | 410 assert(s->block_last_index[n]>=0); |
411 | |
412 if(s->alternate_scan) nCoeffs= 63; //FIXME | |
413 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; | |
414 | |
344 | 415 quant_matrix = s->inter_matrix; |
325 | 416 asm volatile( |
2979 | 417 "pcmpeqw %%mm7, %%mm7 \n\t" |
418 "psrlq $48, %%mm7 \n\t" | |
419 "movd %2, %%mm6 \n\t" | |
420 "packssdw %%mm6, %%mm6 \n\t" | |
421 "packssdw %%mm6, %%mm6 \n\t" | |
422 "mov %3, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3398
diff
changeset
|
423 ASMALIGN(4) |
2979 | 424 "1: \n\t" |
425 "movq (%0, %%"REG_a"), %%mm0 \n\t" | |
426 "movq 8(%0, %%"REG_a"), %%mm1 \n\t" | |
427 "movq (%1, %%"REG_a"), %%mm4 \n\t" | |
428 "movq 8(%1, %%"REG_a"), %%mm5 \n\t" | |
429 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] | |
430 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] | |
431 "pxor %%mm2, %%mm2 \n\t" | |
432 "pxor %%mm3, %%mm3 \n\t" | |
433 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 | |
434 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 | |
435 "pxor %%mm2, %%mm0 \n\t" | |
436 "pxor %%mm3, %%mm1 \n\t" | |
437 "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) | |
438 "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) | |
439 "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2 | |
440 "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2 | |
441 "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*2*q | |
442 "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*2*q | |
443 "paddw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q | |
444 "paddw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q | |
445 "pxor %%mm4, %%mm4 \n\t" | |
446 "pxor %%mm5, %%mm5 \n\t" // FIXME slow | |
447 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 | |
448 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 | |
449 "psrlw $4, %%mm0 \n\t" | |
450 "psrlw $4, %%mm1 \n\t" | |
451 "pxor %%mm2, %%mm0 \n\t" | |
452 "pxor %%mm3, %%mm1 \n\t" | |
453 "psubw %%mm2, %%mm0 \n\t" | |
454 "psubw %%mm3, %%mm1 \n\t" | |
455 "pandn %%mm0, %%mm4 \n\t" | |
456 "pandn %%mm1, %%mm5 \n\t" | |
457 "pxor %%mm4, %%mm7 \n\t" | |
458 "pxor %%mm5, %%mm7 \n\t" | |
459 "movq %%mm4, (%0, %%"REG_a") \n\t" | |
460 "movq %%mm5, 8(%0, %%"REG_a") \n\t" | |
325 | 461 |
2979 | 462 "add $16, %%"REG_a" \n\t" |
463 "jng 1b \n\t" | |
464 "movd 124(%0, %3), %%mm0 \n\t" | |
465 "movq %%mm7, %%mm6 \n\t" | |
466 "psrlq $32, %%mm7 \n\t" | |
467 "pxor %%mm6, %%mm7 \n\t" | |
468 "movq %%mm7, %%mm6 \n\t" | |
469 "psrlq $16, %%mm7 \n\t" | |
470 "pxor %%mm6, %%mm7 \n\t" | |
471 "pslld $31, %%mm7 \n\t" | |
472 "psrlq $15, %%mm7 \n\t" | |
473 "pxor %%mm7, %%mm0 \n\t" | |
474 "movd %%mm0, 124(%0, %3) \n\t" | |
2967 | 475 |
2979 | 476 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (-2*nCoeffs) |
477 : "%"REG_a, "memory" | |
478 ); | |
8 | 479 } |
480 | |
2967 | 481 /* draw the edges of width 'w' of an image of size width, height |
206 | 482 this mmx version can only handle w==8 || w==16 */ |
1064 | 483 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w) |
206 | 484 { |
1064 | 485 uint8_t *ptr, *last_line; |
206 | 486 int i; |
487 | |
488 last_line = buf + (height - 1) * wrap; | |
489 /* left and right */ | |
490 ptr = buf; | |
491 if(w==8) | |
492 { | |
2979 | 493 asm volatile( |
494 "1: \n\t" | |
495 "movd (%0), %%mm0 \n\t" | |
496 "punpcklbw %%mm0, %%mm0 \n\t" | |
497 "punpcklwd %%mm0, %%mm0 \n\t" | |
498 "punpckldq %%mm0, %%mm0 \n\t" | |
499 "movq %%mm0, -8(%0) \n\t" | |
500 "movq -8(%0, %2), %%mm1 \n\t" | |
501 "punpckhbw %%mm1, %%mm1 \n\t" | |
502 "punpckhwd %%mm1, %%mm1 \n\t" | |
503 "punpckhdq %%mm1, %%mm1 \n\t" | |
504 "movq %%mm1, (%0, %2) \n\t" | |
505 "add %1, %0 \n\t" | |
506 "cmp %3, %0 \n\t" | |
507 " jb 1b \n\t" | |
508 : "+r" (ptr) | |
509 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height) | |
510 ); | |
206 | 511 } |
512 else | |
513 { | |
2979 | 514 asm volatile( |
515 "1: \n\t" | |
516 "movd (%0), %%mm0 \n\t" | |
517 "punpcklbw %%mm0, %%mm0 \n\t" | |
518 "punpcklwd %%mm0, %%mm0 \n\t" | |
519 "punpckldq %%mm0, %%mm0 \n\t" | |
520 "movq %%mm0, -8(%0) \n\t" | |
521 "movq %%mm0, -16(%0) \n\t" | |
522 "movq -8(%0, %2), %%mm1 \n\t" | |
523 "punpckhbw %%mm1, %%mm1 \n\t" | |
524 "punpckhwd %%mm1, %%mm1 \n\t" | |
525 "punpckhdq %%mm1, %%mm1 \n\t" | |
526 "movq %%mm1, (%0, %2) \n\t" | |
527 "movq %%mm1, 8(%0, %2) \n\t" | |
528 "add %1, %0 \n\t" | |
529 "cmp %3, %0 \n\t" | |
530 " jb 1b \n\t" | |
531 : "+r" (ptr) | |
532 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height) | |
533 ); | |
206 | 534 } |
2967 | 535 |
206 | 536 for(i=0;i<w;i+=4) { |
537 /* top and bottom (and hopefully also the corners) */ | |
2979 | 538 ptr= buf - (i + 1) * wrap - w; |
539 asm volatile( | |
540 "1: \n\t" | |
541 "movq (%1, %0), %%mm0 \n\t" | |
542 "movq %%mm0, (%0) \n\t" | |
543 "movq %%mm0, (%0, %2) \n\t" | |
544 "movq %%mm0, (%0, %2, 2) \n\t" | |
545 "movq %%mm0, (%0, %3) \n\t" | |
546 "add $8, %0 \n\t" | |
547 "cmp %4, %0 \n\t" | |
548 " jb 1b \n\t" | |
549 : "+r" (ptr) | |
550 : "r" ((long)buf - (long)ptr - w), "r" ((long)-wrap), "r" ((long)-wrap*3), "r" (ptr+width+2*w) | |
551 ); | |
552 ptr= last_line + (i + 1) * wrap - w; | |
553 asm volatile( | |
554 "1: \n\t" | |
555 "movq (%1, %0), %%mm0 \n\t" | |
556 "movq %%mm0, (%0) \n\t" | |
557 "movq %%mm0, (%0, %2) \n\t" | |
558 "movq %%mm0, (%0, %2, 2) \n\t" | |
559 "movq %%mm0, (%0, %3) \n\t" | |
560 "add $8, %0 \n\t" | |
561 "cmp %4, %0 \n\t" | |
562 " jb 1b \n\t" | |
563 : "+r" (ptr) | |
564 : "r" ((long)last_line - (long)ptr - w), "r" ((long)wrap), "r" ((long)wrap*3), "r" (ptr+width+2*w) | |
565 ); | |
206 | 566 } |
567 } | |
568 | |
1719 | 569 static void denoise_dct_mmx(MpegEncContext *s, DCTELEM *block){ |
570 const int intra= s->mb_intra; | |
571 int *sum= s->dct_error_sum[intra]; | |
572 uint16_t *offset= s->dct_offset[intra]; | |
573 | |
574 s->dct_count[intra]++; | |
575 | |
576 asm volatile( | |
2979 | 577 "pxor %%mm7, %%mm7 \n\t" |
578 "1: \n\t" | |
579 "pxor %%mm0, %%mm0 \n\t" | |
580 "pxor %%mm1, %%mm1 \n\t" | |
581 "movq (%0), %%mm2 \n\t" | |
582 "movq 8(%0), %%mm3 \n\t" | |
583 "pcmpgtw %%mm2, %%mm0 \n\t" | |
584 "pcmpgtw %%mm3, %%mm1 \n\t" | |
585 "pxor %%mm0, %%mm2 \n\t" | |
586 "pxor %%mm1, %%mm3 \n\t" | |
587 "psubw %%mm0, %%mm2 \n\t" | |
588 "psubw %%mm1, %%mm3 \n\t" | |
589 "movq %%mm2, %%mm4 \n\t" | |
590 "movq %%mm3, %%mm5 \n\t" | |
591 "psubusw (%2), %%mm2 \n\t" | |
592 "psubusw 8(%2), %%mm3 \n\t" | |
593 "pxor %%mm0, %%mm2 \n\t" | |
594 "pxor %%mm1, %%mm3 \n\t" | |
595 "psubw %%mm0, %%mm2 \n\t" | |
596 "psubw %%mm1, %%mm3 \n\t" | |
597 "movq %%mm2, (%0) \n\t" | |
598 "movq %%mm3, 8(%0) \n\t" | |
599 "movq %%mm4, %%mm2 \n\t" | |
600 "movq %%mm5, %%mm3 \n\t" | |
601 "punpcklwd %%mm7, %%mm4 \n\t" | |
602 "punpckhwd %%mm7, %%mm2 \n\t" | |
603 "punpcklwd %%mm7, %%mm5 \n\t" | |
604 "punpckhwd %%mm7, %%mm3 \n\t" | |
605 "paddd (%1), %%mm4 \n\t" | |
606 "paddd 8(%1), %%mm2 \n\t" | |
607 "paddd 16(%1), %%mm5 \n\t" | |
608 "paddd 24(%1), %%mm3 \n\t" | |
609 "movq %%mm4, (%1) \n\t" | |
610 "movq %%mm2, 8(%1) \n\t" | |
611 "movq %%mm5, 16(%1) \n\t" | |
612 "movq %%mm3, 24(%1) \n\t" | |
613 "add $16, %0 \n\t" | |
614 "add $32, %1 \n\t" | |
615 "add $16, %2 \n\t" | |
616 "cmp %3, %0 \n\t" | |
617 " jb 1b \n\t" | |
1719 | 618 : "+r" (block), "+r" (sum), "+r" (offset) |
619 : "r"(block+64) | |
620 ); | |
621 } | |
622 | |
1720
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
623 static void denoise_dct_sse2(MpegEncContext *s, DCTELEM *block){ |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
624 const int intra= s->mb_intra; |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
625 int *sum= s->dct_error_sum[intra]; |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
626 uint16_t *offset= s->dct_offset[intra]; |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
627 |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
628 s->dct_count[intra]++; |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
629 |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
630 asm volatile( |
2979 | 631 "pxor %%xmm7, %%xmm7 \n\t" |
632 "1: \n\t" | |
633 "pxor %%xmm0, %%xmm0 \n\t" | |
634 "pxor %%xmm1, %%xmm1 \n\t" | |
635 "movdqa (%0), %%xmm2 \n\t" | |
636 "movdqa 16(%0), %%xmm3 \n\t" | |
637 "pcmpgtw %%xmm2, %%xmm0 \n\t" | |
638 "pcmpgtw %%xmm3, %%xmm1 \n\t" | |
639 "pxor %%xmm0, %%xmm2 \n\t" | |
640 "pxor %%xmm1, %%xmm3 \n\t" | |
641 "psubw %%xmm0, %%xmm2 \n\t" | |
642 "psubw %%xmm1, %%xmm3 \n\t" | |
643 "movdqa %%xmm2, %%xmm4 \n\t" | |
644 "movdqa %%xmm3, %%xmm5 \n\t" | |
645 "psubusw (%2), %%xmm2 \n\t" | |
646 "psubusw 16(%2), %%xmm3 \n\t" | |
647 "pxor %%xmm0, %%xmm2 \n\t" | |
648 "pxor %%xmm1, %%xmm3 \n\t" | |
649 "psubw %%xmm0, %%xmm2 \n\t" | |
650 "psubw %%xmm1, %%xmm3 \n\t" | |
651 "movdqa %%xmm2, (%0) \n\t" | |
652 "movdqa %%xmm3, 16(%0) \n\t" | |
653 "movdqa %%xmm4, %%xmm6 \n\t" | |
654 "movdqa %%xmm5, %%xmm0 \n\t" | |
655 "punpcklwd %%xmm7, %%xmm4 \n\t" | |
656 "punpckhwd %%xmm7, %%xmm6 \n\t" | |
657 "punpcklwd %%xmm7, %%xmm5 \n\t" | |
658 "punpckhwd %%xmm7, %%xmm0 \n\t" | |
659 "paddd (%1), %%xmm4 \n\t" | |
660 "paddd 16(%1), %%xmm6 \n\t" | |
661 "paddd 32(%1), %%xmm5 \n\t" | |
662 "paddd 48(%1), %%xmm0 \n\t" | |
663 "movdqa %%xmm4, (%1) \n\t" | |
664 "movdqa %%xmm6, 16(%1) \n\t" | |
665 "movdqa %%xmm5, 32(%1) \n\t" | |
666 "movdqa %%xmm0, 48(%1) \n\t" | |
667 "add $32, %0 \n\t" | |
668 "add $64, %1 \n\t" | |
669 "add $32, %2 \n\t" | |
670 "cmp %3, %0 \n\t" | |
671 " jb 1b \n\t" | |
1720
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
672 : "+r" (block), "+r" (sum), "+r" (offset) |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
673 : "r"(block+64) |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
674 ); |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
675 } |
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
676 |
220 | 677 #undef HAVE_MMX2 |
678 #define RENAME(a) a ## _MMX | |
1565 | 679 #define RENAMEl(a) a ## _mmx |
220 | 680 #include "mpegvideo_mmx_template.c" |
681 | |
682 #define HAVE_MMX2 | |
683 #undef RENAME | |
1597 | 684 #undef RENAMEl |
220 | 685 #define RENAME(a) a ## _MMX2 |
1565 | 686 #define RENAMEl(a) a ## _mmx2 |
220 | 687 #include "mpegvideo_mmx_template.c" |
206 | 688 |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
689 #undef RENAME |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
690 #undef RENAMEl |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
691 #define RENAME(a) a ## _SSE2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
692 #define RENAMEl(a) a ## _sse2 |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
693 #include "mpegvideo_mmx_template.c" |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
694 |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
695 void MPV_common_init_mmx(MpegEncContext *s) |
8 | 696 { |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
697 if (mm_flags & MM_MMX) { |
706
e65798d228ea
idct permutation cleanup, idct can be selected per context now
michaelni
parents:
687
diff
changeset
|
698 const int dct_algo = s->avctx->dct_algo; |
2967 | 699 |
1689 | 700 s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx; |
701 s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx; | |
702 s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx; | |
703 s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx; | |
3281
7fac25904a8b
missmatch control for mpeg2 intra dequantization if bitexact=1
michael
parents:
3036
diff
changeset
|
704 if(!(s->flags & CODEC_FLAG_BITEXACT)) |
7fac25904a8b
missmatch control for mpeg2 intra dequantization if bitexact=1
michael
parents:
3036
diff
changeset
|
705 s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx; |
1689 | 706 s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx; |
312 | 707 |
350
6ebbecc10063
- Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents:
344
diff
changeset
|
708 draw_edges = draw_edges_mmx; |
2967 | 709 |
1720
96a86bd1e0d5
denoise_dct_sse2() patch by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1719
diff
changeset
|
710 if (mm_flags & MM_SSE2) { |
2979 | 711 s->denoise_dct= denoise_dct_sse2; |
712 } else { | |
713 s->denoise_dct= denoise_dct_mmx; | |
714 } | |
220 | 715 |
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:
620
diff
changeset
|
716 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){ |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
717 if(mm_flags & MM_SSE2){ |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
718 s->dct_quantize= dct_quantize_SSE2; |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1720
diff
changeset
|
719 } else if(mm_flags & MM_MMXEXT){ |
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:
620
diff
changeset
|
720 s->dct_quantize= dct_quantize_MMX2; |
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:
620
diff
changeset
|
721 } else { |
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:
620
diff
changeset
|
722 s->dct_quantize= dct_quantize_MMX; |
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:
620
diff
changeset
|
723 } |
350
6ebbecc10063
- Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents:
344
diff
changeset
|
724 } |
14
8ceb13af9cb6
renamed - use of s->dct_unquantize function pointer - SHOULD add faster h263 mmx specific unquantization stuff
glantau
parents:
8
diff
changeset
|
725 } |
8 | 726 } |