Remove DECLARE_ALIGNED_{8,16} macros
[ffmpeg.git] / libavcodec / x86 / vc1dsp_mmx.c
1 /*
2  * VC-1 and WMV3 - DSP functions MMX-optimized
3  * Copyright (c) 2007 Christophe GISQUET <christophe.gisquet@free.fr>
4  *
5  * Permission is hereby granted, free of charge, to any person
6  * obtaining a copy of this software and associated documentation
7  * files (the "Software"), to deal in the Software without
8  * restriction, including without limitation the rights to use,
9  * copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following
12  * conditions:
13  *
14  * The above copyright notice and this permission notice shall be
15  * included in all copies or substantial portions of the Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
19  * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
20  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
21  * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
22  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
24  * OTHER DEALINGS IN THE SOFTWARE.
25  */
26
27 #include "libavutil/x86_cpu.h"
28 #include "libavcodec/dsputil.h"
29 #include "dsputil_mmx.h"
30
31 #define OP_PUT(S,D)
32 #define OP_AVG(S,D) "pavgb " #S ", " #D " \n\t"
33
34 /** Add rounder from mm7 to mm3 and pack result at destination */
35 #define NORMALIZE_MMX(SHIFT)                                    \
36      "paddw     %%mm7, %%mm3           \n\t" /* +bias-r */      \
37      "paddw     %%mm7, %%mm4           \n\t" /* +bias-r */      \
38      "psraw     "SHIFT", %%mm3         \n\t"                    \
39      "psraw     "SHIFT", %%mm4         \n\t"
40
41 #define TRANSFER_DO_PACK(OP)                    \
42      "packuswb  %%mm4, %%mm3           \n\t"    \
43      OP((%2), %%mm3)                            \
44      "movq      %%mm3, (%2)            \n\t"
45
46 #define TRANSFER_DONT_PACK(OP)                  \
47      OP(0(%2), %%mm3)                           \
48      OP(8(%2), %%mm4)                           \
49      "movq      %%mm3, 0(%2)           \n\t"    \
50      "movq      %%mm4, 8(%2)           \n\t"
51
52 /** @see MSPEL_FILTER13_CORE for use as UNPACK macro */
53 #define DO_UNPACK(reg)  "punpcklbw %%mm0, " reg "\n\t"
54 #define DONT_UNPACK(reg)
55
56 /** Compute the rounder 32-r or 8-r and unpacks it to mm7 */
57 #define LOAD_ROUNDER_MMX(ROUND)                 \
58      "movd      "ROUND", %%mm7         \n\t"    \
59      "punpcklwd %%mm7, %%mm7           \n\t"    \
60      "punpckldq %%mm7, %%mm7           \n\t"
61
62 #define SHIFT2_LINE(OFF, R0,R1,R2,R3)           \
63     "paddw     %%mm"#R2", %%mm"#R1"    \n\t"    \
64     "movd      (%0,%3), %%mm"#R0"      \n\t"    \
65     "pmullw    %%mm6, %%mm"#R1"        \n\t"    \
66     "punpcklbw %%mm0, %%mm"#R0"        \n\t"    \
67     "movd      (%0,%2), %%mm"#R3"      \n\t"    \
68     "psubw     %%mm"#R0", %%mm"#R1"    \n\t"    \
69     "punpcklbw %%mm0, %%mm"#R3"        \n\t"    \
70     "paddw     %%mm7, %%mm"#R1"        \n\t"    \
71     "psubw     %%mm"#R3", %%mm"#R1"    \n\t"    \
72     "psraw     %4, %%mm"#R1"           \n\t"    \
73     "movq      %%mm"#R1", "#OFF"(%1)   \n\t"    \
74     "add       %2, %0                  \n\t"
75
76 DECLARE_ALIGNED(16, const uint64_t, ff_pw_9) = 0x0009000900090009ULL;
77
78 /** Sacrifying mm6 allows to pipeline loads from src */
79 static void vc1_put_ver_16b_shift2_mmx(int16_t *dst,
80                                        const uint8_t *src, x86_reg stride,
81                                        int rnd, int64_t shift)
82 {
83     __asm__ volatile(
84         "mov       $3, %%"REG_c"           \n\t"
85         LOAD_ROUNDER_MMX("%5")
86         "movq      "MANGLE(ff_pw_9)", %%mm6 \n\t"
87         "1:                                \n\t"
88         "movd      (%0), %%mm2             \n\t"
89         "add       %2, %0                  \n\t"
90         "movd      (%0), %%mm3             \n\t"
91         "punpcklbw %%mm0, %%mm2            \n\t"
92         "punpcklbw %%mm0, %%mm3            \n\t"
93         SHIFT2_LINE(  0, 1, 2, 3, 4)
94         SHIFT2_LINE( 24, 2, 3, 4, 1)
95         SHIFT2_LINE( 48, 3, 4, 1, 2)
96         SHIFT2_LINE( 72, 4, 1, 2, 3)
97         SHIFT2_LINE( 96, 1, 2, 3, 4)
98         SHIFT2_LINE(120, 2, 3, 4, 1)
99         SHIFT2_LINE(144, 3, 4, 1, 2)
100         SHIFT2_LINE(168, 4, 1, 2, 3)
101         "sub       %6, %0                  \n\t"
102         "add       $8, %1                  \n\t"
103         "dec       %%"REG_c"               \n\t"
104         "jnz 1b                            \n\t"
105         : "+r"(src), "+r"(dst)
106         : "r"(stride), "r"(-2*stride),
107           "m"(shift), "m"(rnd), "r"(9*stride-4)
108         : "%"REG_c, "memory"
109     );
110 }
111
112 /**
113  * Data is already unpacked, so some operations can directly be made from
114  * memory.
115  */
116 #define VC1_HOR_16b_SHIFT2(OP, OPNAME)\
117 static void OPNAME ## vc1_hor_16b_shift2_mmx(uint8_t *dst, x86_reg stride,\
118                                              const int16_t *src, int rnd)\
119 {\
120     int h = 8;\
121 \
122     src -= 1;\
123     rnd -= (-1+9+9-1)*1024; /* Add -1024 bias */\
124     __asm__ volatile(\
125         LOAD_ROUNDER_MMX("%4")\
126         "movq      "MANGLE(ff_pw_128)", %%mm6\n\t"\
127         "movq      "MANGLE(ff_pw_9)", %%mm5 \n\t"\
128         "1:                                \n\t"\
129         "movq      2*0+0(%1), %%mm1        \n\t"\
130         "movq      2*0+8(%1), %%mm2        \n\t"\
131         "movq      2*1+0(%1), %%mm3        \n\t"\
132         "movq      2*1+8(%1), %%mm4        \n\t"\
133         "paddw     2*3+0(%1), %%mm1        \n\t"\
134         "paddw     2*3+8(%1), %%mm2        \n\t"\
135         "paddw     2*2+0(%1), %%mm3        \n\t"\
136         "paddw     2*2+8(%1), %%mm4        \n\t"\
137         "pmullw    %%mm5, %%mm3            \n\t"\
138         "pmullw    %%mm5, %%mm4            \n\t"\
139         "psubw     %%mm1, %%mm3            \n\t"\
140         "psubw     %%mm2, %%mm4            \n\t"\
141         NORMALIZE_MMX("$7")\
142         /* Remove bias */\
143         "paddw     %%mm6, %%mm3            \n\t"\
144         "paddw     %%mm6, %%mm4            \n\t"\
145         TRANSFER_DO_PACK(OP)\
146         "add       $24, %1                 \n\t"\
147         "add       %3, %2                  \n\t"\
148         "decl      %0                      \n\t"\
149         "jnz 1b                            \n\t"\
150         : "+r"(h), "+r" (src),  "+r" (dst)\
151         : "r"(stride), "m"(rnd)\
152         : "memory"\
153     );\
154 }
155
156 VC1_HOR_16b_SHIFT2(OP_PUT, put_)
157 VC1_HOR_16b_SHIFT2(OP_AVG, avg_)
158
159
160 /**
161  * Purely vertical or horizontal 1/2 shift interpolation.
162  * Sacrify mm6 for *9 factor.
163  */
164 #define VC1_SHIFT2(OP, OPNAME)\
165 static void OPNAME ## vc1_shift2_mmx(uint8_t *dst, const uint8_t *src,\
166                                      x86_reg stride, int rnd, x86_reg offset)\
167 {\
168     rnd = 8-rnd;\
169     __asm__ volatile(\
170         "mov       $8, %%"REG_c"           \n\t"\
171         LOAD_ROUNDER_MMX("%5")\
172         "movq      "MANGLE(ff_pw_9)", %%mm6\n\t"\
173         "1:                                \n\t"\
174         "movd      0(%0   ), %%mm3         \n\t"\
175         "movd      4(%0   ), %%mm4         \n\t"\
176         "movd      0(%0,%2), %%mm1         \n\t"\
177         "movd      4(%0,%2), %%mm2         \n\t"\
178         "add       %2, %0                  \n\t"\
179         "punpcklbw %%mm0, %%mm3            \n\t"\
180         "punpcklbw %%mm0, %%mm4            \n\t"\
181         "punpcklbw %%mm0, %%mm1            \n\t"\
182         "punpcklbw %%mm0, %%mm2            \n\t"\
183         "paddw     %%mm1, %%mm3            \n\t"\
184         "paddw     %%mm2, %%mm4            \n\t"\
185         "movd      0(%0,%3), %%mm1         \n\t"\
186         "movd      4(%0,%3), %%mm2         \n\t"\
187         "pmullw    %%mm6, %%mm3            \n\t" /* 0,9,9,0*/\
188         "pmullw    %%mm6, %%mm4            \n\t" /* 0,9,9,0*/\
189         "punpcklbw %%mm0, %%mm1            \n\t"\
190         "punpcklbw %%mm0, %%mm2            \n\t"\
191         "psubw     %%mm1, %%mm3            \n\t" /*-1,9,9,0*/\
192         "psubw     %%mm2, %%mm4            \n\t" /*-1,9,9,0*/\
193         "movd      0(%0,%2), %%mm1         \n\t"\
194         "movd      4(%0,%2), %%mm2         \n\t"\
195         "punpcklbw %%mm0, %%mm1            \n\t"\
196         "punpcklbw %%mm0, %%mm2            \n\t"\
197         "psubw     %%mm1, %%mm3            \n\t" /*-1,9,9,-1*/\
198         "psubw     %%mm2, %%mm4            \n\t" /*-1,9,9,-1*/\
199         NORMALIZE_MMX("$4")\
200         "packuswb  %%mm4, %%mm3            \n\t"\
201         OP((%1), %%mm3)\
202         "movq      %%mm3, (%1)             \n\t"\
203         "add       %6, %0                  \n\t"\
204         "add       %4, %1                  \n\t"\
205         "dec       %%"REG_c"               \n\t"\
206         "jnz 1b                            \n\t"\
207         : "+r"(src),  "+r"(dst)\
208         : "r"(offset), "r"(-2*offset), "g"(stride), "m"(rnd),\
209           "g"(stride-offset)\
210         : "%"REG_c, "memory"\
211     );\
212 }
213
214 VC1_SHIFT2(OP_PUT, put_)
215 VC1_SHIFT2(OP_AVG, avg_)
216
217 /**
218  * Filter coefficients made global to allow access by all 1 or 3 quarter shift
219  * interpolation functions.
220  */
221 DECLARE_ASM_CONST(16, uint64_t, ff_pw_53) = 0x0035003500350035ULL;
222 DECLARE_ASM_CONST(16, uint64_t, ff_pw_18) = 0x0012001200120012ULL;
223
224 /**
225  * Core of the 1/4 and 3/4 shift bicubic interpolation.
226  *
227  * @param UNPACK  Macro unpacking arguments from 8 to 16bits (can be empty).
228  * @param MOVQ    "movd 1" or "movq 2", if data read is already unpacked.
229  * @param A1      Address of 1st tap (beware of unpacked/packed).
230  * @param A2      Address of 2nd tap
231  * @param A3      Address of 3rd tap
232  * @param A4      Address of 4th tap
233  */
234 #define MSPEL_FILTER13_CORE(UNPACK, MOVQ, A1, A2, A3, A4)       \
235      MOVQ "*0+"A1", %%mm1       \n\t"                           \
236      MOVQ "*4+"A1", %%mm2       \n\t"                           \
237      UNPACK("%%mm1")                                            \
238      UNPACK("%%mm2")                                            \
239      "pmullw    "MANGLE(ff_pw_3)", %%mm1\n\t"                   \
240      "pmullw    "MANGLE(ff_pw_3)", %%mm2\n\t"                   \
241      MOVQ "*0+"A2", %%mm3       \n\t"                           \
242      MOVQ "*4+"A2", %%mm4       \n\t"                           \
243      UNPACK("%%mm3")                                            \
244      UNPACK("%%mm4")                                            \
245      "pmullw    %%mm6, %%mm3    \n\t" /* *18 */                 \
246      "pmullw    %%mm6, %%mm4    \n\t" /* *18 */                 \
247      "psubw     %%mm1, %%mm3    \n\t" /* 18,-3 */               \
248      "psubw     %%mm2, %%mm4    \n\t" /* 18,-3 */               \
249      MOVQ "*0+"A4", %%mm1       \n\t"                           \
250      MOVQ "*4+"A4", %%mm2       \n\t"                           \
251      UNPACK("%%mm1")                                            \
252      UNPACK("%%mm2")                                            \
253      "psllw     $2, %%mm1       \n\t" /* 4* */                  \
254      "psllw     $2, %%mm2       \n\t" /* 4* */                  \
255      "psubw     %%mm1, %%mm3    \n\t" /* -4,18,-3 */            \
256      "psubw     %%mm2, %%mm4    \n\t" /* -4,18,-3 */            \
257      MOVQ "*0+"A3", %%mm1       \n\t"                           \
258      MOVQ "*4+"A3", %%mm2       \n\t"                           \
259      UNPACK("%%mm1")                                            \
260      UNPACK("%%mm2")                                            \
261      "pmullw    %%mm5, %%mm1    \n\t" /* *53 */                 \
262      "pmullw    %%mm5, %%mm2    \n\t" /* *53 */                 \
263      "paddw     %%mm1, %%mm3    \n\t" /* 4,53,18,-3 */          \
264      "paddw     %%mm2, %%mm4    \n\t" /* 4,53,18,-3 */
265
266 /**
267  * Macro to build the vertical 16bits version of vc1_put_shift[13].
268  * Here, offset=src_stride. Parameters passed A1 to A4 must use
269  * %3 (src_stride) and %4 (3*src_stride).
270  *
271  * @param  NAME   Either 1 or 3
272  * @see MSPEL_FILTER13_CORE for information on A1->A4
273  */
274 #define MSPEL_FILTER13_VER_16B(NAME, A1, A2, A3, A4)                    \
275 static void                                                             \
276 vc1_put_ver_16b_ ## NAME ## _mmx(int16_t *dst, const uint8_t *src,      \
277                                  x86_reg src_stride,                   \
278                                  int rnd, int64_t shift)                \
279 {                                                                       \
280     int h = 8;                                                          \
281     src -= src_stride;                                                  \
282     __asm__ volatile(                                                       \
283         LOAD_ROUNDER_MMX("%5")                                          \
284         "movq      "MANGLE(ff_pw_53)", %%mm5\n\t"                       \
285         "movq      "MANGLE(ff_pw_18)", %%mm6\n\t"                       \
286         ASMALIGN(3)                                                     \
287         "1:                        \n\t"                                \
288         MSPEL_FILTER13_CORE(DO_UNPACK, "movd  1", A1, A2, A3, A4)       \
289         NORMALIZE_MMX("%6")                                             \
290         TRANSFER_DONT_PACK(OP_PUT)                                      \
291         /* Last 3 (in fact 4) bytes on the line */                      \
292         "movd      8+"A1", %%mm1   \n\t"                                \
293         DO_UNPACK("%%mm1")                                              \
294         "movq      %%mm1, %%mm3    \n\t"                                \
295         "paddw     %%mm1, %%mm1    \n\t"                                \
296         "paddw     %%mm3, %%mm1    \n\t" /* 3* */                       \
297         "movd      8+"A2", %%mm3   \n\t"                                \
298         DO_UNPACK("%%mm3")                                              \
299         "pmullw    %%mm6, %%mm3    \n\t" /* *18 */                      \
300         "psubw     %%mm1, %%mm3    \n\t" /*18,-3 */                     \
301         "movd      8+"A3", %%mm1   \n\t"                                \
302         DO_UNPACK("%%mm1")                                              \
303         "pmullw    %%mm5, %%mm1    \n\t" /* *53 */                      \
304         "paddw     %%mm1, %%mm3    \n\t" /*53,18,-3 */                  \
305         "movd      8+"A4", %%mm1   \n\t"                                \
306         DO_UNPACK("%%mm1")                                              \
307         "psllw     $2, %%mm1       \n\t" /* 4* */                       \
308         "psubw     %%mm1, %%mm3    \n\t"                                \
309         "paddw     %%mm7, %%mm3    \n\t"                                \
310         "psraw     %6, %%mm3       \n\t"                                \
311         "movq      %%mm3, 16(%2)   \n\t"                                \
312         "add       %3, %1          \n\t"                                \
313         "add       $24, %2         \n\t"                                \
314         "decl      %0              \n\t"                                \
315         "jnz 1b                    \n\t"                                \
316         : "+r"(h), "+r" (src),  "+r" (dst)                              \
317         : "r"(src_stride), "r"(3*src_stride),                           \
318           "m"(rnd), "m"(shift)                                          \
319         : "memory"                                                      \
320     );                                                                  \
321 }
322
323 /**
324  * Macro to build the horizontal 16bits version of vc1_put_shift[13].
325  * Here, offset=16bits, so parameters passed A1 to A4 should be simple.
326  *
327  * @param  NAME   Either 1 or 3
328  * @see MSPEL_FILTER13_CORE for information on A1->A4
329  */
330 #define MSPEL_FILTER13_HOR_16B(NAME, A1, A2, A3, A4, OP, OPNAME)        \
331 static void                                                             \
332 OPNAME ## vc1_hor_16b_ ## NAME ## _mmx(uint8_t *dst, x86_reg stride,    \
333                                  const int16_t *src, int rnd)           \
334 {                                                                       \
335     int h = 8;                                                          \
336     src -= 1;                                                           \
337     rnd -= (-4+58+13-3)*256; /* Add -256 bias */                        \
338     __asm__ volatile(                                                       \
339         LOAD_ROUNDER_MMX("%4")                                          \
340         "movq      "MANGLE(ff_pw_18)", %%mm6   \n\t"                    \
341         "movq      "MANGLE(ff_pw_53)", %%mm5   \n\t"                    \
342         ASMALIGN(3)                                                     \
343         "1:                        \n\t"                                \
344         MSPEL_FILTER13_CORE(DONT_UNPACK, "movq 2", A1, A2, A3, A4)      \
345         NORMALIZE_MMX("$7")                                             \
346         /* Remove bias */                                               \
347         "paddw     "MANGLE(ff_pw_128)", %%mm3  \n\t"                    \
348         "paddw     "MANGLE(ff_pw_128)", %%mm4  \n\t"                    \
349         TRANSFER_DO_PACK(OP)                                            \
350         "add       $24, %1         \n\t"                                \
351         "add       %3, %2          \n\t"                                \
352         "decl      %0              \n\t"                                \
353         "jnz 1b                    \n\t"                                \
354         : "+r"(h), "+r" (src),  "+r" (dst)                              \
355         : "r"(stride), "m"(rnd)                                         \
356         : "memory"                                                      \
357     );                                                                  \
358 }
359
360 /**
361  * Macro to build the 8bits, any direction, version of vc1_put_shift[13].
362  * Here, offset=src_stride. Parameters passed A1 to A4 must use
363  * %3 (offset) and %4 (3*offset).
364  *
365  * @param  NAME   Either 1 or 3
366  * @see MSPEL_FILTER13_CORE for information on A1->A4
367  */
368 #define MSPEL_FILTER13_8B(NAME, A1, A2, A3, A4, OP, OPNAME)             \
369 static void                                                             \
370 OPNAME ## vc1_## NAME ## _mmx(uint8_t *dst, const uint8_t *src,         \
371                         x86_reg stride, int rnd, x86_reg offset)      \
372 {                                                                       \
373     int h = 8;                                                          \
374     src -= offset;                                                      \
375     rnd = 32-rnd;                                                       \
376     __asm__ volatile (                                                      \
377         LOAD_ROUNDER_MMX("%6")                                          \
378         "movq      "MANGLE(ff_pw_53)", %%mm5       \n\t"                \
379         "movq      "MANGLE(ff_pw_18)", %%mm6       \n\t"                \
380         ASMALIGN(3)                                                     \
381         "1:                        \n\t"                                \
382         MSPEL_FILTER13_CORE(DO_UNPACK, "movd   1", A1, A2, A3, A4)      \
383         NORMALIZE_MMX("$6")                                             \
384         TRANSFER_DO_PACK(OP)                                            \
385         "add       %5, %1          \n\t"                                \
386         "add       %5, %2          \n\t"                                \
387         "decl      %0              \n\t"                                \
388         "jnz 1b                    \n\t"                                \
389         : "+r"(h), "+r" (src),  "+r" (dst)                              \
390         : "r"(offset), "r"(3*offset), "g"(stride), "m"(rnd)             \
391         : "memory"                                                      \
392     );                                                                  \
393 }
394
395 /** 1/4 shift bicubic interpolation */
396 MSPEL_FILTER13_8B     (shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )", OP_PUT, put_)
397 MSPEL_FILTER13_8B     (shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )", OP_AVG, avg_)
398 MSPEL_FILTER13_VER_16B(shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )")
399 MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)", OP_PUT, put_)
400 MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)", OP_AVG, avg_)
401
402 /** 3/4 shift bicubic interpolation */
403 MSPEL_FILTER13_8B     (shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )", OP_PUT, put_)
404 MSPEL_FILTER13_8B     (shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )", OP_AVG, avg_)
405 MSPEL_FILTER13_VER_16B(shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )")
406 MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)", OP_PUT, put_)
407 MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)", OP_AVG, avg_)
408
409 typedef void (*vc1_mspel_mc_filter_ver_16bits)(int16_t *dst, const uint8_t *src, x86_reg src_stride, int rnd, int64_t shift);
410 typedef void (*vc1_mspel_mc_filter_hor_16bits)(uint8_t *dst, x86_reg dst_stride, const int16_t *src, int rnd);
411 typedef void (*vc1_mspel_mc_filter_8bits)(uint8_t *dst, const uint8_t *src, x86_reg stride, int rnd, x86_reg offset);
412
413 /**
414  * Interpolates fractional pel values by applying proper vertical then
415  * horizontal filter.
416  *
417  * @param  dst     Destination buffer for interpolated pels.
418  * @param  src     Source buffer.
419  * @param  stride  Stride for both src and dst buffers.
420  * @param  hmode   Horizontal filter (expressed in quarter pixels shift).
421  * @param  hmode   Vertical filter.
422  * @param  rnd     Rounding bias.
423  */
424 #define VC1_MSPEL_MC(OP)\
425 static void OP ## vc1_mspel_mc(uint8_t *dst, const uint8_t *src, int stride,\
426                                int hmode, int vmode, int rnd)\
427 {\
428     static const vc1_mspel_mc_filter_ver_16bits vc1_put_shift_ver_16bits[] =\
429          { NULL, vc1_put_ver_16b_shift1_mmx, vc1_put_ver_16b_shift2_mmx, vc1_put_ver_16b_shift3_mmx };\
430     static const vc1_mspel_mc_filter_hor_16bits vc1_put_shift_hor_16bits[] =\
431          { NULL, OP ## vc1_hor_16b_shift1_mmx, OP ## vc1_hor_16b_shift2_mmx, OP ## vc1_hor_16b_shift3_mmx };\
432     static const vc1_mspel_mc_filter_8bits vc1_put_shift_8bits[] =\
433          { NULL, OP ## vc1_shift1_mmx, OP ## vc1_shift2_mmx, OP ## vc1_shift3_mmx };\
434 \
435     __asm__ volatile(\
436         "pxor %%mm0, %%mm0         \n\t"\
437         ::: "memory"\
438     );\
439 \
440     if (vmode) { /* Vertical filter to apply */\
441         if (hmode) { /* Horizontal filter to apply, output to tmp */\
442             static const int shift_value[] = { 0, 5, 1, 5 };\
443             int              shift = (shift_value[hmode]+shift_value[vmode])>>1;\
444             int              r;\
445             DECLARE_ALIGNED(16, int16_t, tmp)[12*8];\
446 \
447             r = (1<<(shift-1)) + rnd-1;\
448             vc1_put_shift_ver_16bits[vmode](tmp, src-1, stride, r, shift);\
449 \
450             vc1_put_shift_hor_16bits[hmode](dst, stride, tmp+1, 64-rnd);\
451             return;\
452         }\
453         else { /* No horizontal filter, output 8 lines to dst */\
454             vc1_put_shift_8bits[vmode](dst, src, stride, 1-rnd, stride);\
455             return;\
456         }\
457     }\
458 \
459     /* Horizontal mode with no vertical mode */\
460     vc1_put_shift_8bits[hmode](dst, src, stride, rnd, 1);\
461 }
462
463 VC1_MSPEL_MC(put_)
464 VC1_MSPEL_MC(avg_)
465
466 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd);
467 void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd);
468
469 /** Macro to ease bicubic filter interpolation functions declarations */
470 #define DECLARE_FUNCTION(a, b)                                          \
471 static void put_vc1_mspel_mc ## a ## b ## _mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { \
472      put_vc1_mspel_mc(dst, src, stride, a, b, rnd);                     \
473 }\
474 static void avg_vc1_mspel_mc ## a ## b ## _mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) { \
475      avg_vc1_mspel_mc(dst, src, stride, a, b, rnd);                     \
476 }
477
478 DECLARE_FUNCTION(0, 1)
479 DECLARE_FUNCTION(0, 2)
480 DECLARE_FUNCTION(0, 3)
481
482 DECLARE_FUNCTION(1, 0)
483 DECLARE_FUNCTION(1, 1)
484 DECLARE_FUNCTION(1, 2)
485 DECLARE_FUNCTION(1, 3)
486
487 DECLARE_FUNCTION(2, 0)
488 DECLARE_FUNCTION(2, 1)
489 DECLARE_FUNCTION(2, 2)
490 DECLARE_FUNCTION(2, 3)
491
492 DECLARE_FUNCTION(3, 0)
493 DECLARE_FUNCTION(3, 1)
494 DECLARE_FUNCTION(3, 2)
495 DECLARE_FUNCTION(3, 3)
496
497 static void vc1_inv_trans_4x4_dc_mmx2(uint8_t *dest, int linesize, DCTELEM *block)
498 {
499     int dc = block[0];
500     dc = (17 * dc +  4) >> 3;
501     dc = (17 * dc + 64) >> 7;
502     __asm__ volatile(
503         "movd          %0, %%mm0 \n\t"
504         "pshufw $0, %%mm0, %%mm0 \n\t"
505         "pxor       %%mm1, %%mm1 \n\t"
506         "psubw      %%mm0, %%mm1 \n\t"
507         "packuswb   %%mm0, %%mm0 \n\t"
508         "packuswb   %%mm1, %%mm1 \n\t"
509         ::"r"(dc)
510     );
511     __asm__ volatile(
512         "movd          %0, %%mm2 \n\t"
513         "movd          %1, %%mm3 \n\t"
514         "movd          %2, %%mm4 \n\t"
515         "movd          %3, %%mm5 \n\t"
516         "paddusb    %%mm0, %%mm2 \n\t"
517         "paddusb    %%mm0, %%mm3 \n\t"
518         "paddusb    %%mm0, %%mm4 \n\t"
519         "paddusb    %%mm0, %%mm5 \n\t"
520         "psubusb    %%mm1, %%mm2 \n\t"
521         "psubusb    %%mm1, %%mm3 \n\t"
522         "psubusb    %%mm1, %%mm4 \n\t"
523         "psubusb    %%mm1, %%mm5 \n\t"
524         "movd       %%mm2, %0    \n\t"
525         "movd       %%mm3, %1    \n\t"
526         "movd       %%mm4, %2    \n\t"
527         "movd       %%mm5, %3    \n\t"
528         :"+m"(*(uint32_t*)(dest+0*linesize)),
529          "+m"(*(uint32_t*)(dest+1*linesize)),
530          "+m"(*(uint32_t*)(dest+2*linesize)),
531          "+m"(*(uint32_t*)(dest+3*linesize))
532     );
533 }
534
535 static void vc1_inv_trans_4x8_dc_mmx2(uint8_t *dest, int linesize, DCTELEM *block)
536 {
537     int dc = block[0];
538     dc = (17 * dc +  4) >> 3;
539     dc = (12 * dc + 64) >> 7;
540     __asm__ volatile(
541         "movd          %0, %%mm0 \n\t"
542         "pshufw $0, %%mm0, %%mm0 \n\t"
543         "pxor       %%mm1, %%mm1 \n\t"
544         "psubw      %%mm0, %%mm1 \n\t"
545         "packuswb   %%mm0, %%mm0 \n\t"
546         "packuswb   %%mm1, %%mm1 \n\t"
547         ::"r"(dc)
548     );
549     __asm__ volatile(
550         "movd          %0, %%mm2 \n\t"
551         "movd          %1, %%mm3 \n\t"
552         "movd          %2, %%mm4 \n\t"
553         "movd          %3, %%mm5 \n\t"
554         "paddusb    %%mm0, %%mm2 \n\t"
555         "paddusb    %%mm0, %%mm3 \n\t"
556         "paddusb    %%mm0, %%mm4 \n\t"
557         "paddusb    %%mm0, %%mm5 \n\t"
558         "psubusb    %%mm1, %%mm2 \n\t"
559         "psubusb    %%mm1, %%mm3 \n\t"
560         "psubusb    %%mm1, %%mm4 \n\t"
561         "psubusb    %%mm1, %%mm5 \n\t"
562         "movd       %%mm2, %0    \n\t"
563         "movd       %%mm3, %1    \n\t"
564         "movd       %%mm4, %2    \n\t"
565         "movd       %%mm5, %3    \n\t"
566         :"+m"(*(uint32_t*)(dest+0*linesize)),
567          "+m"(*(uint32_t*)(dest+1*linesize)),
568          "+m"(*(uint32_t*)(dest+2*linesize)),
569          "+m"(*(uint32_t*)(dest+3*linesize))
570     );
571     dest += 4*linesize;
572     __asm__ volatile(
573         "movd          %0, %%mm2 \n\t"
574         "movd          %1, %%mm3 \n\t"
575         "movd          %2, %%mm4 \n\t"
576         "movd          %3, %%mm5 \n\t"
577         "paddusb    %%mm0, %%mm2 \n\t"
578         "paddusb    %%mm0, %%mm3 \n\t"
579         "paddusb    %%mm0, %%mm4 \n\t"
580         "paddusb    %%mm0, %%mm5 \n\t"
581         "psubusb    %%mm1, %%mm2 \n\t"
582         "psubusb    %%mm1, %%mm3 \n\t"
583         "psubusb    %%mm1, %%mm4 \n\t"
584         "psubusb    %%mm1, %%mm5 \n\t"
585         "movd       %%mm2, %0    \n\t"
586         "movd       %%mm3, %1    \n\t"
587         "movd       %%mm4, %2    \n\t"
588         "movd       %%mm5, %3    \n\t"
589         :"+m"(*(uint32_t*)(dest+0*linesize)),
590          "+m"(*(uint32_t*)(dest+1*linesize)),
591          "+m"(*(uint32_t*)(dest+2*linesize)),
592          "+m"(*(uint32_t*)(dest+3*linesize))
593     );
594 }
595
596 static void vc1_inv_trans_8x4_dc_mmx2(uint8_t *dest, int linesize, DCTELEM *block)
597 {
598     int dc = block[0];
599     dc = ( 3 * dc +  1) >> 1;
600     dc = (17 * dc + 64) >> 7;
601     __asm__ volatile(
602         "movd          %0, %%mm0 \n\t"
603         "pshufw $0, %%mm0, %%mm0 \n\t"
604         "pxor       %%mm1, %%mm1 \n\t"
605         "psubw      %%mm0, %%mm1 \n\t"
606         "packuswb   %%mm0, %%mm0 \n\t"
607         "packuswb   %%mm1, %%mm1 \n\t"
608         ::"r"(dc)
609     );
610     __asm__ volatile(
611         "movq          %0, %%mm2 \n\t"
612         "movq          %1, %%mm3 \n\t"
613         "movq          %2, %%mm4 \n\t"
614         "movq          %3, %%mm5 \n\t"
615         "paddusb    %%mm0, %%mm2 \n\t"
616         "paddusb    %%mm0, %%mm3 \n\t"
617         "paddusb    %%mm0, %%mm4 \n\t"
618         "paddusb    %%mm0, %%mm5 \n\t"
619         "psubusb    %%mm1, %%mm2 \n\t"
620         "psubusb    %%mm1, %%mm3 \n\t"
621         "psubusb    %%mm1, %%mm4 \n\t"
622         "psubusb    %%mm1, %%mm5 \n\t"
623         "movq       %%mm2, %0    \n\t"
624         "movq       %%mm3, %1    \n\t"
625         "movq       %%mm4, %2    \n\t"
626         "movq       %%mm5, %3    \n\t"
627         :"+m"(*(uint32_t*)(dest+0*linesize)),
628          "+m"(*(uint32_t*)(dest+1*linesize)),
629          "+m"(*(uint32_t*)(dest+2*linesize)),
630          "+m"(*(uint32_t*)(dest+3*linesize))
631     );
632 }
633
634 static void vc1_inv_trans_8x8_dc_mmx2(uint8_t *dest, int linesize, DCTELEM *block)
635 {
636     int dc = block[0];
637     dc = (3 * dc +  1) >> 1;
638     dc = (3 * dc + 16) >> 5;
639     __asm__ volatile(
640         "movd          %0, %%mm0 \n\t"
641         "pshufw $0, %%mm0, %%mm0 \n\t"
642         "pxor       %%mm1, %%mm1 \n\t"
643         "psubw      %%mm0, %%mm1 \n\t"
644         "packuswb   %%mm0, %%mm0 \n\t"
645         "packuswb   %%mm1, %%mm1 \n\t"
646         ::"r"(dc)
647     );
648     __asm__ volatile(
649         "movq          %0, %%mm2 \n\t"
650         "movq          %1, %%mm3 \n\t"
651         "movq          %2, %%mm4 \n\t"
652         "movq          %3, %%mm5 \n\t"
653         "paddusb    %%mm0, %%mm2 \n\t"
654         "paddusb    %%mm0, %%mm3 \n\t"
655         "paddusb    %%mm0, %%mm4 \n\t"
656         "paddusb    %%mm0, %%mm5 \n\t"
657         "psubusb    %%mm1, %%mm2 \n\t"
658         "psubusb    %%mm1, %%mm3 \n\t"
659         "psubusb    %%mm1, %%mm4 \n\t"
660         "psubusb    %%mm1, %%mm5 \n\t"
661         "movq       %%mm2, %0    \n\t"
662         "movq       %%mm3, %1    \n\t"
663         "movq       %%mm4, %2    \n\t"
664         "movq       %%mm5, %3    \n\t"
665         :"+m"(*(uint32_t*)(dest+0*linesize)),
666          "+m"(*(uint32_t*)(dest+1*linesize)),
667          "+m"(*(uint32_t*)(dest+2*linesize)),
668          "+m"(*(uint32_t*)(dest+3*linesize))
669     );
670     dest += 4*linesize;
671     __asm__ volatile(
672         "movq          %0, %%mm2 \n\t"
673         "movq          %1, %%mm3 \n\t"
674         "movq          %2, %%mm4 \n\t"
675         "movq          %3, %%mm5 \n\t"
676         "paddusb    %%mm0, %%mm2 \n\t"
677         "paddusb    %%mm0, %%mm3 \n\t"
678         "paddusb    %%mm0, %%mm4 \n\t"
679         "paddusb    %%mm0, %%mm5 \n\t"
680         "psubusb    %%mm1, %%mm2 \n\t"
681         "psubusb    %%mm1, %%mm3 \n\t"
682         "psubusb    %%mm1, %%mm4 \n\t"
683         "psubusb    %%mm1, %%mm5 \n\t"
684         "movq       %%mm2, %0    \n\t"
685         "movq       %%mm3, %1    \n\t"
686         "movq       %%mm4, %2    \n\t"
687         "movq       %%mm5, %3    \n\t"
688         :"+m"(*(uint32_t*)(dest+0*linesize)),
689          "+m"(*(uint32_t*)(dest+1*linesize)),
690          "+m"(*(uint32_t*)(dest+2*linesize)),
691          "+m"(*(uint32_t*)(dest+3*linesize))
692     );
693 }
694
695 void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx) {
696     mm_flags = mm_support();
697
698     dsp->put_vc1_mspel_pixels_tab[ 0] = ff_put_vc1_mspel_mc00_mmx;
699     dsp->put_vc1_mspel_pixels_tab[ 4] = put_vc1_mspel_mc01_mmx;
700     dsp->put_vc1_mspel_pixels_tab[ 8] = put_vc1_mspel_mc02_mmx;
701     dsp->put_vc1_mspel_pixels_tab[12] = put_vc1_mspel_mc03_mmx;
702
703     dsp->put_vc1_mspel_pixels_tab[ 1] = put_vc1_mspel_mc10_mmx;
704     dsp->put_vc1_mspel_pixels_tab[ 5] = put_vc1_mspel_mc11_mmx;
705     dsp->put_vc1_mspel_pixels_tab[ 9] = put_vc1_mspel_mc12_mmx;
706     dsp->put_vc1_mspel_pixels_tab[13] = put_vc1_mspel_mc13_mmx;
707
708     dsp->put_vc1_mspel_pixels_tab[ 2] = put_vc1_mspel_mc20_mmx;
709     dsp->put_vc1_mspel_pixels_tab[ 6] = put_vc1_mspel_mc21_mmx;
710     dsp->put_vc1_mspel_pixels_tab[10] = put_vc1_mspel_mc22_mmx;
711     dsp->put_vc1_mspel_pixels_tab[14] = put_vc1_mspel_mc23_mmx;
712
713     dsp->put_vc1_mspel_pixels_tab[ 3] = put_vc1_mspel_mc30_mmx;
714     dsp->put_vc1_mspel_pixels_tab[ 7] = put_vc1_mspel_mc31_mmx;
715     dsp->put_vc1_mspel_pixels_tab[11] = put_vc1_mspel_mc32_mmx;
716     dsp->put_vc1_mspel_pixels_tab[15] = put_vc1_mspel_mc33_mmx;
717
718     if (mm_flags & FF_MM_MMX2){
719         dsp->avg_vc1_mspel_pixels_tab[ 0] = ff_avg_vc1_mspel_mc00_mmx2;
720         dsp->avg_vc1_mspel_pixels_tab[ 4] = avg_vc1_mspel_mc01_mmx2;
721         dsp->avg_vc1_mspel_pixels_tab[ 8] = avg_vc1_mspel_mc02_mmx2;
722         dsp->avg_vc1_mspel_pixels_tab[12] = avg_vc1_mspel_mc03_mmx2;
723
724         dsp->avg_vc1_mspel_pixels_tab[ 1] = avg_vc1_mspel_mc10_mmx2;
725         dsp->avg_vc1_mspel_pixels_tab[ 5] = avg_vc1_mspel_mc11_mmx2;
726         dsp->avg_vc1_mspel_pixels_tab[ 9] = avg_vc1_mspel_mc12_mmx2;
727         dsp->avg_vc1_mspel_pixels_tab[13] = avg_vc1_mspel_mc13_mmx2;
728
729         dsp->avg_vc1_mspel_pixels_tab[ 2] = avg_vc1_mspel_mc20_mmx2;
730         dsp->avg_vc1_mspel_pixels_tab[ 6] = avg_vc1_mspel_mc21_mmx2;
731         dsp->avg_vc1_mspel_pixels_tab[10] = avg_vc1_mspel_mc22_mmx2;
732         dsp->avg_vc1_mspel_pixels_tab[14] = avg_vc1_mspel_mc23_mmx2;
733
734         dsp->avg_vc1_mspel_pixels_tab[ 3] = avg_vc1_mspel_mc30_mmx2;
735         dsp->avg_vc1_mspel_pixels_tab[ 7] = avg_vc1_mspel_mc31_mmx2;
736         dsp->avg_vc1_mspel_pixels_tab[11] = avg_vc1_mspel_mc32_mmx2;
737         dsp->avg_vc1_mspel_pixels_tab[15] = avg_vc1_mspel_mc33_mmx2;
738
739         dsp->vc1_inv_trans_8x8_dc = vc1_inv_trans_8x8_dc_mmx2;
740         dsp->vc1_inv_trans_4x8_dc = vc1_inv_trans_4x8_dc_mmx2;
741         dsp->vc1_inv_trans_8x4_dc = vc1_inv_trans_8x4_dc_mmx2;
742         dsp->vc1_inv_trans_4x4_dc = vc1_inv_trans_4x4_dc_mmx2;
743     }
744 }