Mercurial > libavcodec.hg
annotate x86/idct_mmx.c @ 12492:58a960d6e34c libavcodec
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now
coded in asm instead of C, this is (depending on the function) up to 50%
faster for cases where gcc didn't do a great job at looping.
Since h264_idct_add8() is now faster than the manual loop setup in h264.c,
in-asm idct calling can now be enabled for chroma as well (see r16207). For
MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does
the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%.
author | rbultje |
---|---|
date | Tue, 14 Sep 2010 13:36:26 +0000 |
parents | be9129c5503e |
children |
rev | line source |
---|---|
8430 | 1 /* |
2 * idct_mmx.c | |
3 * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca> | |
4 * | |
5 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder. | |
6 * See http://libmpeg2.sourceforge.net/ for updates. | |
7 * | |
8 * mpeg2dec is free software; you can redistribute it and/or modify | |
9 * it under the terms of the GNU General Public License as published by | |
10 * the Free Software Foundation; either version 2 of the License, or | |
11 * (at your option) any later version. | |
12 * | |
13 * mpeg2dec is distributed in the hope that it will be useful, | |
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 * GNU General Public License for more details. | |
17 * | |
18 * You should have received a copy of the GNU General Public License | |
19 * along with mpeg2dec; if not, write to the Free Software | |
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 */ | |
22 | |
23 #include "libavutil/common.h" | |
24 #include "libavcodec/dsputil.h" | |
25 | |
12296 | 26 #include "libavutil/x86_cpu.h" |
11381 | 27 #include "dsputil_mmx.h" |
8430 | 28 |
29 #define ROW_SHIFT 11 | |
30 #define COL_SHIFT 6 | |
31 | |
32 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT))) | |
33 #define rounder(bias) {round (bias), round (bias)} | |
34 | |
35 | |
36 #if 0 | |
37 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */ | |
38 static inline void idct_row (int16_t * row, int offset, | |
39 int16_t * table, int32_t * rounder) | |
40 { | |
41 int C1, C2, C3, C4, C5, C6, C7; | |
42 int a0, a1, a2, a3, b0, b1, b2, b3; | |
43 | |
44 row += offset; | |
45 | |
46 C1 = table[1]; | |
47 C2 = table[2]; | |
48 C3 = table[3]; | |
49 C4 = table[4]; | |
50 C5 = table[5]; | |
51 C6 = table[6]; | |
52 C7 = table[7]; | |
53 | |
54 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder; | |
55 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder; | |
56 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder; | |
57 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder; | |
58 | |
59 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7]; | |
60 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7]; | |
61 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7]; | |
62 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7]; | |
63 | |
64 row[0] = (a0 + b0) >> ROW_SHIFT; | |
65 row[1] = (a1 + b1) >> ROW_SHIFT; | |
66 row[2] = (a2 + b2) >> ROW_SHIFT; | |
67 row[3] = (a3 + b3) >> ROW_SHIFT; | |
68 row[4] = (a3 - b3) >> ROW_SHIFT; | |
69 row[5] = (a2 - b2) >> ROW_SHIFT; | |
70 row[6] = (a1 - b1) >> ROW_SHIFT; | |
71 row[7] = (a0 - b0) >> ROW_SHIFT; | |
72 } | |
73 #endif | |
74 | |
75 | |
76 /* MMXEXT row IDCT */ | |
77 | |
78 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \ | |
79 c4, c6, c4, c6, \ | |
80 c1, c3, -c1, -c5, \ | |
81 c5, c7, c3, -c7, \ | |
82 c4, -c6, c4, -c6, \ | |
83 -c4, c2, c4, -c2, \ | |
84 c5, -c1, c3, -c1, \ | |
85 c7, c3, c7, -c5 } | |
86 | |
87 static inline void mmxext_row_head (int16_t * const row, const int offset, | |
88 const int16_t * const table) | |
89 { | |
12296 | 90 __asm__ volatile( |
91 "movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */ | |
8430 | 92 |
12296 | 93 "movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */ |
94 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */ | |
8430 | 95 |
12296 | 96 "movq (%1), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */ |
97 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */ | |
8430 | 98 |
12296 | 99 "movq 8(%1), %%mm4 \n\t" /* mm4 = C6 C4 C6 C4 */ |
100 "pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */ | |
8430 | 101 |
12296 | 102 "pshufw $0x4e, %%mm2, %%mm2 \n\t" /* mm2 = x2 x0 x6 x4 */ |
103 :: "r" ((row+offset)), "r" (table) | |
104 ); | |
8430 | 105 } |
106 | |
107 static inline void mmxext_row (const int16_t * const table, | |
108 const int32_t * const rounder) | |
109 { | |
12296 | 110 __asm__ volatile ( |
111 "movq 16(%0), %%mm1 \n\t" /* mm1 = -C5 -C1 C3 C1 */ | |
112 "pmaddwd %%mm2, %%mm4 \n\t" /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */ | |
8430 | 113 |
12296 | 114 "pmaddwd 32(%0), %%mm0 \n\t" /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */ |
115 "pshufw $0x4e, %%mm6, %%mm6 \n\t" /* mm6 = x3 x1 x7 x5 */ | |
8430 | 116 |
12296 | 117 "movq 24(%0), %%mm7 \n\t" /* mm7 = -C7 C3 C7 C5 */ |
118 "pmaddwd %%mm5, %%mm1 \n\t" /* mm1= -C1*x5-C5*x7 C1*x1+C3*x3 */ | |
119 | |
120 "paddd (%1), %%mm3 \n\t" /* mm3 += rounder */ | |
121 "pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */ | |
8430 | 122 |
12296 | 123 "pmaddwd 40(%0), %%mm2 \n\t" /* mm2= C4*x0-C2*x2 -C4*x4+C2*x6 */ |
124 "paddd %%mm4, %%mm3 \n\t" /* mm3 = a1 a0 + rounder */ | |
8430 | 125 |
12296 | 126 "pmaddwd 48(%0), %%mm5 \n\t" /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */ |
127 "movq %%mm3, %%mm4 \n\t" /* mm4 = a1 a0 + rounder */ | |
8430 | 128 |
12296 | 129 "pmaddwd 56(%0), %%mm6 \n\t" /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */ |
130 "paddd %%mm7, %%mm1 \n\t" /* mm1 = b1 b0 */ | |
8430 | 131 |
12296 | 132 "paddd (%1), %%mm0 \n\t" /* mm0 += rounder */ |
133 "psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */ | |
8430 | 134 |
12296 | 135 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */ |
136 "paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */ | |
8430 | 137 |
12296 | 138 "paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */ |
139 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */ | |
8430 | 140 |
12296 | 141 "paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */ |
142 "movq %%mm0, %%mm4 \n\t" /* mm4 = a3 a2 + rounder */ | |
8430 | 143 |
12296 | 144 "paddd %%mm5, %%mm0 \n\t" /* mm0 = a3+b3 a2+b2 + rounder */ |
145 "psubd %%mm5, %%mm4 \n\t" /* mm4 = a3-b3 a2-b2 + rounder */ | |
146 : : "r" (table), "r" (rounder)); | |
8430 | 147 } |
148 | |
149 static inline void mmxext_row_tail (int16_t * const row, const int store) | |
150 { | |
12296 | 151 __asm__ volatile ( |
152 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */ | |
8430 | 153 |
12296 | 154 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */ |
8430 | 155 |
12296 | 156 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */ |
157 | |
158 "packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */ | |
8430 | 159 |
12296 | 160 "movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */ |
161 "pshufw $0xb1, %%mm4, %%mm4 \n\t" /* mm4 = y7 y6 y5 y4 */ | |
8430 | 162 |
12296 | 163 /* slot */ |
8430 | 164 |
12296 | 165 "movq %%mm4, 8(%0) \n\t" /* save y7 y6 y5 y4 */ |
166 :: "r" (row+store) | |
167 ); | |
8430 | 168 } |
169 | |
170 static inline void mmxext_row_mid (int16_t * const row, const int store, | |
171 const int offset, | |
172 const int16_t * const table) | |
173 { | |
12296 | 174 __asm__ volatile ( |
175 "movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */ | |
176 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */ | |
8430 | 177 |
12296 | 178 "movq 8(%0,%1), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */ |
179 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */ | |
8430 | 180 |
12296 | 181 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */ |
182 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */ | |
183 | |
184 "packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */ | |
185 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */ | |
8430 | 186 |
12296 | 187 "movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */ |
188 "pshufw $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */ | |
8430 | 189 |
12296 | 190 "movq (%3), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */ |
191 "movq %%mm4, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */ | |
8430 | 192 |
12296 | 193 "pmaddwd %%mm0, %%mm3 \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */ |
8430 | 194 |
12296 | 195 "movq 8(%3), %%mm4 \n\t" /* mm4 = C6 C4 C6 C4 */ |
196 "pshufw $0x4e, %%mm2, %%mm2\n\t" /* mm2 = x2 x0 x6 x4 */ | |
12297 | 197 :: "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table) |
12296 | 198 ); |
8430 | 199 } |
200 | |
201 | |
202 /* MMX row IDCT */ | |
203 | |
204 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \ | |
205 c4, c6, -c4, -c2, \ | |
206 c1, c3, c3, -c7, \ | |
207 c5, c7, -c1, -c5, \ | |
208 c4, -c6, c4, -c2, \ | |
209 -c4, c2, c4, -c6, \ | |
210 c5, -c1, c7, -c5, \ | |
211 c7, c3, c3, -c1 } | |
212 | |
213 static inline void mmx_row_head (int16_t * const row, const int offset, | |
214 const int16_t * const table) | |
215 { | |
12296 | 216 __asm__ volatile ( |
217 "movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */ | |
8430 | 218 |
12296 | 219 "movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */ |
220 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */ | |
8430 | 221 |
12296 | 222 "movq (%1), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */ |
223 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */ | |
8430 | 224 |
12296 | 225 "punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */ |
226 | |
227 "movq 8(%1), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */ | |
228 "pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */ | |
8430 | 229 |
12296 | 230 "movq 16(%1), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */ |
231 "punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */ | |
232 :: "r" ((row+offset)), "r" (table) | |
233 ); | |
8430 | 234 } |
235 | |
236 static inline void mmx_row (const int16_t * const table, | |
237 const int32_t * const rounder) | |
238 { | |
12296 | 239 __asm__ volatile ( |
240 "pmaddwd %%mm2, %%mm4 \n\t" /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */ | |
241 "punpckldq %%mm5, %%mm5 \n\t" /* mm5 = x3 x1 x3 x1 */ | |
8430 | 242 |
12296 | 243 "pmaddwd 32(%0), %%mm0 \n\t" /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */ |
244 "punpckhdq %%mm6, %%mm6 \n\t" /* mm6 = x7 x5 x7 x5 */ | |
8430 | 245 |
12296 | 246 "movq 24(%0), %%mm7 \n\t" /* mm7 = -C5 -C1 C7 C5 */ |
247 "pmaddwd %%mm5, %%mm1 \n\t" /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */ | |
248 | |
249 "paddd (%1), %%mm3 \n\t" /* mm3 += rounder */ | |
250 "pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */ | |
8430 | 251 |
12296 | 252 "pmaddwd 40(%0), %%mm2 \n\t" /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */ |
253 "paddd %%mm4, %%mm3 \n\t" /* mm3 = a1 a0 + rounder */ | |
8430 | 254 |
12296 | 255 "pmaddwd 48(%0), %%mm5 \n\t" /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */ |
256 "movq %%mm3, %%mm4 \n\t" /* mm4 = a1 a0 + rounder */ | |
8430 | 257 |
12296 | 258 "pmaddwd 56(%0), %%mm6 \n\t" /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */ |
259 "paddd %%mm7, %%mm1 \n\t" /* mm1 = b1 b0 */ | |
8430 | 260 |
12296 | 261 "paddd (%1), %%mm0 \n\t" /* mm0 += rounder */ |
262 "psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */ | |
8430 | 263 |
12296 | 264 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */ |
265 "paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */ | |
8430 | 266 |
12296 | 267 "paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */ |
268 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */ | |
8430 | 269 |
12296 | 270 "paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */ |
271 "movq %%mm0, %%mm7 \n\t" /* mm7 = a3 a2 + rounder */ | |
8430 | 272 |
12296 | 273 "paddd %%mm5, %%mm0 \n\t" /* mm0 = a3+b3 a2+b2 + rounder */ |
274 "psubd %%mm5, %%mm7 \n\t" /* mm7 = a3-b3 a2-b2 + rounder */ | |
275 :: "r" (table), "r" (rounder) | |
276 ); | |
8430 | 277 } |
278 | |
279 static inline void mmx_row_tail (int16_t * const row, const int store) | |
280 { | |
12296 | 281 __asm__ volatile ( |
282 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */ | |
8430 | 283 |
12296 | 284 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */ |
8430 | 285 |
12296 | 286 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */ |
8430 | 287 |
12296 | 288 "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */ |
289 | |
290 "movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */ | |
291 "movq %%mm7, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */ | |
8430 | 292 |
12296 | 293 "pslld $16, %%mm7 \n\t" /* mm7 = y7 0 y5 0 */ |
8430 | 294 |
12296 | 295 "psrld $16, %%mm4 \n\t" /* mm4 = 0 y6 0 y4 */ |
296 | |
297 "por %%mm4, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */ | |
8430 | 298 |
12296 | 299 /* slot */ |
8430 | 300 |
12296 | 301 "movq %%mm7, 8(%0) \n\t" /* save y7 y6 y5 y4 */ |
302 :: "r" (row+store) | |
303 ); | |
8430 | 304 } |
305 | |
306 static inline void mmx_row_mid (int16_t * const row, const int store, | |
307 const int offset, const int16_t * const table) | |
308 { | |
309 | |
12296 | 310 __asm__ volatile ( |
311 "movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */ | |
312 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */ | |
313 | |
314 "movq 8(%0,%1), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */ | |
315 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */ | |
8430 | 316 |
12296 | 317 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */ |
318 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */ | |
8430 | 319 |
12296 | 320 "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */ |
321 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */ | |
8430 | 322 |
12296 | 323 "movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */ |
324 "movq %%mm7, %%mm1 \n\t" /* mm1 = y6 y7 y4 y5 */ | |
8430 | 325 |
12296 | 326 "punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */ |
327 "psrld $16, %%mm7 \n\t" /* mm7 = 0 y6 0 y4 */ | |
8430 | 328 |
12296 | 329 "movq (%3), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */ |
330 "pslld $16, %%mm1 \n\t" /* mm1 = y7 0 y5 0 */ | |
8430 | 331 |
12296 | 332 "movq 8(%3), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */ |
333 "por %%mm1, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */ | |
8430 | 334 |
12296 | 335 "movq 16(%3), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */ |
336 "punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */ | |
8430 | 337 |
12296 | 338 "movq %%mm7, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */ |
339 "pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */ | |
12297 | 340 : : "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table) |
12296 | 341 ); |
8430 | 342 } |
343 | |
344 | |
345 #if 0 | |
346 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */ | |
347 static inline void idct_col (int16_t * col, int offset) | |
348 { | |
349 /* multiplication - as implemented on mmx */ | |
350 #define F(c,x) (((c) * (x)) >> 16) | |
351 | |
352 /* saturation - it helps us handle torture test cases */ | |
353 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x)) | |
354 | |
355 int16_t x0, x1, x2, x3, x4, x5, x6, x7; | |
356 int16_t y0, y1, y2, y3, y4, y5, y6, y7; | |
357 int16_t a0, a1, a2, a3, b0, b1, b2, b3; | |
358 int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12; | |
359 | |
360 col += offset; | |
361 | |
362 x0 = col[0*8]; | |
363 x1 = col[1*8]; | |
364 x2 = col[2*8]; | |
365 x3 = col[3*8]; | |
366 x4 = col[4*8]; | |
367 x5 = col[5*8]; | |
368 x6 = col[6*8]; | |
369 x7 = col[7*8]; | |
370 | |
371 u04 = S (x0 + x4); | |
372 v04 = S (x0 - x4); | |
373 u26 = S (F (T2, x6) + x2); | |
374 v26 = S (F (T2, x2) - x6); | |
375 | |
376 a0 = S (u04 + u26); | |
377 a1 = S (v04 + v26); | |
378 a2 = S (v04 - v26); | |
379 a3 = S (u04 - u26); | |
380 | |
381 u17 = S (F (T1, x7) + x1); | |
382 v17 = S (F (T1, x1) - x7); | |
383 u35 = S (F (T3, x5) + x3); | |
384 v35 = S (F (T3, x3) - x5); | |
385 | |
386 b0 = S (u17 + u35); | |
387 b3 = S (v17 - v35); | |
388 u12 = S (u17 - u35); | |
389 v12 = S (v17 + v35); | |
390 u12 = S (2 * F (C4, u12)); | |
391 v12 = S (2 * F (C4, v12)); | |
392 b1 = S (u12 + v12); | |
393 b2 = S (u12 - v12); | |
394 | |
395 y0 = S (a0 + b0) >> COL_SHIFT; | |
396 y1 = S (a1 + b1) >> COL_SHIFT; | |
397 y2 = S (a2 + b2) >> COL_SHIFT; | |
398 y3 = S (a3 + b3) >> COL_SHIFT; | |
399 | |
400 y4 = S (a3 - b3) >> COL_SHIFT; | |
401 y5 = S (a2 - b2) >> COL_SHIFT; | |
402 y6 = S (a1 - b1) >> COL_SHIFT; | |
403 y7 = S (a0 - b0) >> COL_SHIFT; | |
404 | |
405 col[0*8] = y0; | |
406 col[1*8] = y1; | |
407 col[2*8] = y2; | |
408 col[3*8] = y3; | |
409 col[4*8] = y4; | |
410 col[5*8] = y5; | |
411 col[6*8] = y6; | |
412 col[7*8] = y7; | |
413 } | |
414 #endif | |
415 | |
416 | |
417 /* MMX column IDCT */ | |
418 static inline void idct_col (int16_t * const col, const int offset) | |
419 { | |
420 #define T1 13036 | |
421 #define T2 27146 | |
422 #define T3 43790 | |
423 #define C4 23170 | |
424 | |
12296 | 425 DECLARE_ALIGNED(8, static const short, t1_vector)[] = { |
426 T1,T1,T1,T1, | |
427 T2,T2,T2,T2, | |
428 T3,T3,T3,T3, | |
429 C4,C4,C4,C4 | |
430 }; | |
8430 | 431 |
432 /* column code adapted from Peter Gubanov */ | |
433 /* http://www.elecard.com/peter/idct.shtml */ | |
434 | |
12296 | 435 __asm__ volatile ( |
436 "movq (%0), %%mm0 \n\t" /* mm0 = T1 */ | |
8430 | 437 |
12296 | 438 "movq 2*8(%1), %%mm1 \n\t" /* mm1 = x1 */ |
439 "movq %%mm0, %%mm2 \n\t" /* mm2 = T1 */ | |
8430 | 440 |
12296 | 441 "movq 7*2*8(%1), %%mm4 \n\t" /* mm4 = x7 */ |
442 "pmulhw %%mm1, %%mm0 \n\t" /* mm0 = T1*x1 */ | |
8430 | 443 |
12296 | 444 "movq 16(%0), %%mm5 \n\t" /* mm5 = T3 */ |
445 "pmulhw %%mm4, %%mm2 \n\t" /* mm2 = T1*x7 */ | |
8430 | 446 |
12296 | 447 "movq 2*5*8(%1), %%mm6 \n\t" /* mm6 = x5 */ |
448 "movq %%mm5, %%mm7 \n\t" /* mm7 = T3-1 */ | |
8430 | 449 |
12296 | 450 "movq 3*8*2(%1), %%mm3 \n\t" /* mm3 = x3 */ |
451 "psubsw %%mm4, %%mm0 \n\t" /* mm0 = v17 */ | |
8430 | 452 |
12296 | 453 "movq 8(%0), %%mm4 \n\t" /* mm4 = T2 */ |
454 "pmulhw %%mm3, %%mm5 \n\t" /* mm5 = (T3-1)*x3 */ | |
8430 | 455 |
12296 | 456 "paddsw %%mm2, %%mm1 \n\t" /* mm1 = u17 */ |
457 "pmulhw %%mm6, %%mm7 \n\t" /* mm7 = (T3-1)*x5 */ | |
458 | |
459 /* slot */ | |
8430 | 460 |
12296 | 461 "movq %%mm4, %%mm2 \n\t" /* mm2 = T2 */ |
462 "paddsw %%mm3, %%mm5 \n\t" /* mm5 = T3*x3 */ | |
8430 | 463 |
12296 | 464 "pmulhw 2*8*2(%1), %%mm4 \n\t" /* mm4 = T2*x2 */ |
465 "paddsw %%mm6, %%mm7 \n\t" /* mm7 = T3*x5 */ | |
8430 | 466 |
12296 | 467 "psubsw %%mm6, %%mm5 \n\t" /* mm5 = v35 */ |
468 "paddsw %%mm3, %%mm7 \n\t" /* mm7 = u35 */ | |
8430 | 469 |
12296 | 470 "movq 6*8*2(%1), %%mm3 \n\t" /* mm3 = x6 */ |
471 "movq %%mm0, %%mm6 \n\t" /* mm6 = v17 */ | |
8430 | 472 |
12296 | 473 "pmulhw %%mm3, %%mm2 \n\t" /* mm2 = T2*x6 */ |
474 "psubsw %%mm5, %%mm0 \n\t" /* mm0 = b3 */ | |
8430 | 475 |
12296 | 476 "psubsw %%mm3, %%mm4 \n\t" /* mm4 = v26 */ |
477 "paddsw %%mm6, %%mm5 \n\t" /* mm5 = v12 */ | |
8430 | 478 |
12296 | 479 "movq %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */ |
480 "movq %%mm1, %%mm6 \n\t" /* mm6 = u17 */ | |
8430 | 481 |
12296 | 482 "paddsw 2*8*2(%1), %%mm2 \n\t" /* mm2 = u26 */ |
483 "paddsw %%mm7, %%mm6 \n\t" /* mm6 = b0 */ | |
8430 | 484 |
12296 | 485 "psubsw %%mm7, %%mm1 \n\t" /* mm1 = u12 */ |
486 "movq %%mm1, %%mm7 \n\t" /* mm7 = u12 */ | |
8430 | 487 |
12296 | 488 "movq 0*8(%1), %%mm3 \n\t" /* mm3 = x0 */ |
489 "paddsw %%mm5, %%mm1 \n\t" /* mm1 = u12+v12 */ | |
8430 | 490 |
12296 | 491 "movq 24(%0), %%mm0 \n\t" /* mm0 = C4/2 */ |
492 "psubsw %%mm5, %%mm7 \n\t" /* mm7 = u12-v12 */ | |
8430 | 493 |
12296 | 494 "movq %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */ |
495 "pmulhw %%mm0, %%mm1 \n\t" /* mm1 = b1/2 */ | |
8430 | 496 |
12296 | 497 "movq %%mm4, %%mm6 \n\t" /* mm6 = v26 */ |
498 "pmulhw %%mm0, %%mm7 \n\t" /* mm7 = b2/2 */ | |
8430 | 499 |
12296 | 500 "movq 4*8*2(%1), %%mm5 \n\t" /* mm5 = x4 */ |
501 "movq %%mm3, %%mm0 \n\t" /* mm0 = x0 */ | |
8430 | 502 |
12296 | 503 "psubsw %%mm5, %%mm3 \n\t" /* mm3 = v04 */ |
504 "paddsw %%mm5, %%mm0 \n\t" /* mm0 = u04 */ | |
8430 | 505 |
12296 | 506 "paddsw %%mm3, %%mm4 \n\t" /* mm4 = a1 */ |
507 "movq %%mm0, %%mm5 \n\t" /* mm5 = u04 */ | |
8430 | 508 |
12296 | 509 "psubsw %%mm6, %%mm3 \n\t" /* mm3 = a2 */ |
510 "paddsw %%mm2, %%mm5 \n\t" /* mm5 = a0 */ | |
8430 | 511 |
12296 | 512 "paddsw %%mm1, %%mm1 \n\t" /* mm1 = b1 */ |
513 "psubsw %%mm2, %%mm0 \n\t" /* mm0 = a3 */ | |
8430 | 514 |
12296 | 515 "paddsw %%mm7, %%mm7 \n\t" /* mm7 = b2 */ |
516 "movq %%mm3, %%mm2 \n\t" /* mm2 = a2 */ | |
8430 | 517 |
12296 | 518 "movq %%mm4, %%mm6 \n\t" /* mm6 = a1 */ |
519 "paddsw %%mm7, %%mm3 \n\t" /* mm3 = a2+b2 */ | |
8430 | 520 |
12296 | 521 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */ |
522 "paddsw %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */ | |
8430 | 523 |
12296 | 524 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */ |
525 "psubsw %%mm1, %%mm6 \n\t" /* mm6 = a1-b1 */ | |
8430 | 526 |
12296 | 527 "movq 5*8*2(%1), %%mm1 \n\t" /* mm1 = b0 */ |
528 "psubsw %%mm7, %%mm2 \n\t" /* mm2 = a2-b2 */ | |
8430 | 529 |
12296 | 530 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */ |
531 "movq %%mm5, %%mm7 \n\t" /* mm7 = a0 */ | |
8430 | 532 |
12296 | 533 "movq %%mm4, 1*8*2(%1)\n\t" /* save y1 */ |
534 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */ | |
8430 | 535 |
12296 | 536 "movq %%mm3, 2*8*2(%1)\n\t" /* save y2 */ |
537 "paddsw %%mm1, %%mm5 \n\t" /* mm5 = a0+b0 */ | |
8430 | 538 |
12296 | 539 "movq 3*8*2(%1), %%mm4 \n\t" /* mm4 = b3 */ |
540 "psubsw %%mm1, %%mm7 \n\t" /* mm7 = a0-b0 */ | |
8430 | 541 |
12296 | 542 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */ |
543 "movq %%mm0, %%mm3 \n\t" /* mm3 = a3 */ | |
8430 | 544 |
12296 | 545 "movq %%mm2, 5*8*2(%1)\n\t" /* save y5 */ |
546 "psubsw %%mm4, %%mm3 \n\t" /* mm3 = a3-b3 */ | |
8430 | 547 |
12296 | 548 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */ |
549 "paddsw %%mm0, %%mm4 \n\t" /* mm4 = a3+b3 */ | |
8430 | 550 |
12296 | 551 "movq %%mm5, 0*8*2(%1)\n\t" /* save y0 */ |
552 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */ | |
8430 | 553 |
12296 | 554 "movq %%mm6, 6*8*2(%1)\n\t" /* save y6 */ |
555 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */ | |
556 | |
557 "movq %%mm7, 7*8*2(%1)\n\t" /* save y7 */ | |
8430 | 558 |
12296 | 559 "movq %%mm3, 4*8*2(%1)\n\t" /* save y4 */ |
8430 | 560 |
12296 | 561 "movq %%mm4, 3*8*2(%1)\n\t" /* save y3 */ |
562 :: "r" (t1_vector), "r" (col+offset) | |
563 ); | |
8430 | 564 |
565 #undef T1 | |
566 #undef T2 | |
567 #undef T3 | |
568 #undef C4 | |
569 } | |
570 | |
571 | |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
572 DECLARE_ALIGNED(8, static const int32_t, rounder0)[] = |
8430 | 573 rounder ((1 << (COL_SHIFT - 1)) - 0.5); |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
574 DECLARE_ALIGNED(8, static const int32_t, rounder4)[] = rounder (0); |
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
575 DECLARE_ALIGNED(8, static const int32_t, rounder1)[] = |
8430 | 576 rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
577 DECLARE_ALIGNED(8, static const int32_t, rounder7)[] = |
8430 | 578 rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
579 DECLARE_ALIGNED(8, static const int32_t, rounder2)[] = |
8430 | 580 rounder (0.60355339059); /* C2 * (C6+C2)/2 */ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
581 DECLARE_ALIGNED(8, static const int32_t, rounder6)[] = |
8430 | 582 rounder (-0.25); /* C2 * (C6-C2)/2 */ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
583 DECLARE_ALIGNED(8, static const int32_t, rounder3)[] = |
8430 | 584 rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
585 DECLARE_ALIGNED(8, static const int32_t, rounder5)[] = |
8430 | 586 rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */ |
587 | |
588 #undef COL_SHIFT | |
589 #undef ROW_SHIFT | |
590 | |
591 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \ | |
592 void idct (int16_t * const block) \ | |
593 { \ | |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
594 DECLARE_ALIGNED(16, static const int16_t, table04)[] = \ |
8430 | 595 table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
596 DECLARE_ALIGNED(16, static const int16_t, table17)[] = \ |
8430 | 597 table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
598 DECLARE_ALIGNED(16, static const int16_t, table26)[] = \ |
8430 | 599 table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \ |
11508
7be32921237f
Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents:
11381
diff
changeset
|
600 DECLARE_ALIGNED(16, static const int16_t, table35)[] = \ |
8430 | 601 table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \ |
602 \ | |
603 idct_row_head (block, 0*8, table04); \ | |
604 idct_row (table04, rounder0); \ | |
605 idct_row_mid (block, 0*8, 4*8, table04); \ | |
606 idct_row (table04, rounder4); \ | |
607 idct_row_mid (block, 4*8, 1*8, table17); \ | |
608 idct_row (table17, rounder1); \ | |
609 idct_row_mid (block, 1*8, 7*8, table17); \ | |
610 idct_row (table17, rounder7); \ | |
611 idct_row_mid (block, 7*8, 2*8, table26); \ | |
612 idct_row (table26, rounder2); \ | |
613 idct_row_mid (block, 2*8, 6*8, table26); \ | |
614 idct_row (table26, rounder6); \ | |
615 idct_row_mid (block, 6*8, 3*8, table35); \ | |
616 idct_row (table35, rounder3); \ | |
617 idct_row_mid (block, 3*8, 5*8, table35); \ | |
618 idct_row (table35, rounder5); \ | |
619 idct_row_tail (block, 5*8); \ | |
620 \ | |
621 idct_col (block, 0); \ | |
622 idct_col (block, 4); \ | |
623 } | |
624 | |
625 declare_idct (ff_mmxext_idct, mmxext_table, | |
626 mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid) | |
627 | |
628 declare_idct (ff_mmx_idct, mmx_table, | |
629 mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid) | |
630 |