f11170c0755f6953d90465c859dda095bc3ba2a5
[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/cpu.h"
28 #include "libavutil/mem.h"
29 #include "libavutil/x86/asm.h"
30 #include "libavutil/x86/cpu.h"
31 #include "libavcodec/vc1dsp.h"
32 #include "constants.h"
33 #include "fpel.h"
34 #include "vc1dsp.h"
35
36 #if HAVE_6REGS && HAVE_INLINE_ASM
37
38 #define OP_PUT(S,D)
39 #define OP_AVG(S,D) "pavgb " #S ", " #D " \n\t"
40
41 /** Add rounder from mm7 to mm3 and pack result at destination */
42 #define NORMALIZE_MMX(SHIFT)                                    \
43      "paddw     %%mm7, %%mm3           \n\t" /* +bias-r */      \
44      "paddw     %%mm7, %%mm4           \n\t" /* +bias-r */      \
45      "psraw     "SHIFT", %%mm3         \n\t"                    \
46      "psraw     "SHIFT", %%mm4         \n\t"
47
48 #define TRANSFER_DO_PACK(OP)                    \
49      "packuswb  %%mm4, %%mm3           \n\t"    \
50      OP((%2), %%mm3)                            \
51      "movq      %%mm3, (%2)            \n\t"
52
53 #define TRANSFER_DONT_PACK(OP)                  \
54      OP(0(%2), %%mm3)                           \
55      OP(8(%2), %%mm4)                           \
56      "movq      %%mm3, 0(%2)           \n\t"    \
57      "movq      %%mm4, 8(%2)           \n\t"
58
59 /** @see MSPEL_FILTER13_CORE for use as UNPACK macro */
60 #define DO_UNPACK(reg)  "punpcklbw %%mm0, " reg "\n\t"
61 #define DONT_UNPACK(reg)
62
63 /** Compute the rounder 32-r or 8-r and unpacks it to mm7 */
64 #define LOAD_ROUNDER_MMX(ROUND)                 \
65      "movd      "ROUND", %%mm7         \n\t"    \
66      "punpcklwd %%mm7, %%mm7           \n\t"    \
67      "punpckldq %%mm7, %%mm7           \n\t"
68
69 #define SHIFT2_LINE(OFF, R0,R1,R2,R3)           \
70     "paddw     %%mm"#R2", %%mm"#R1"    \n\t"    \
71     "movd      (%0,%3), %%mm"#R0"      \n\t"    \
72     "pmullw    %%mm6, %%mm"#R1"        \n\t"    \
73     "punpcklbw %%mm0, %%mm"#R0"        \n\t"    \
74     "movd      (%0,%2), %%mm"#R3"      \n\t"    \
75     "psubw     %%mm"#R0", %%mm"#R1"    \n\t"    \
76     "punpcklbw %%mm0, %%mm"#R3"        \n\t"    \
77     "paddw     %%mm7, %%mm"#R1"        \n\t"    \
78     "psubw     %%mm"#R3", %%mm"#R1"    \n\t"    \
79     "psraw     %4, %%mm"#R1"           \n\t"    \
80     "movq      %%mm"#R1", "#OFF"(%1)   \n\t"    \
81     "add       %2, %0                  \n\t"
82
83 /** Sacrifying mm6 allows to pipeline loads from src */
84 static void vc1_put_ver_16b_shift2_mmx(int16_t *dst,
85                                        const uint8_t *src, x86_reg stride,
86                                        int rnd, int64_t shift)
87 {
88     __asm__ volatile(
89         "mov       $3, %%"REG_c"           \n\t"
90         LOAD_ROUNDER_MMX("%5")
91         "movq      "MANGLE(ff_pw_9)", %%mm6 \n\t"
92         "1:                                \n\t"
93         "movd      (%0), %%mm2             \n\t"
94         "add       %2, %0                  \n\t"
95         "movd      (%0), %%mm3             \n\t"
96         "punpcklbw %%mm0, %%mm2            \n\t"
97         "punpcklbw %%mm0, %%mm3            \n\t"
98         SHIFT2_LINE(  0, 1, 2, 3, 4)
99         SHIFT2_LINE( 24, 2, 3, 4, 1)
100         SHIFT2_LINE( 48, 3, 4, 1, 2)
101         SHIFT2_LINE( 72, 4, 1, 2, 3)
102         SHIFT2_LINE( 96, 1, 2, 3, 4)
103         SHIFT2_LINE(120, 2, 3, 4, 1)
104         SHIFT2_LINE(144, 3, 4, 1, 2)
105         SHIFT2_LINE(168, 4, 1, 2, 3)
106         "sub       %6, %0                  \n\t"
107         "add       $8, %1                  \n\t"
108         "dec       %%"REG_c"               \n\t"
109         "jnz 1b                            \n\t"
110         : "+r"(src), "+r"(dst)
111         : "r"(stride), "r"(-2*stride),
112           "m"(shift), "m"(rnd), "r"(9*stride-4)
113           NAMED_CONSTRAINTS_ADD(ff_pw_9)
114         : "%"REG_c, "memory"
115     );
116 }
117
118 /**
119  * Data is already unpacked, so some operations can directly be made from
120  * memory.
121  */
122 #define VC1_HOR_16b_SHIFT2(OP, OPNAME)\
123 static void OPNAME ## vc1_hor_16b_shift2_mmx(uint8_t *dst, x86_reg stride,\
124                                              const int16_t *src, int rnd)\
125 {\
126     int h = 8;\
127 \
128     src -= 1;\
129     rnd -= (-1+9+9-1)*1024; /* Add -1024 bias */\
130     __asm__ volatile(\
131         LOAD_ROUNDER_MMX("%4")\
132         "movq      "MANGLE(ff_pw_128)", %%mm6\n\t"\
133         "movq      "MANGLE(ff_pw_9)", %%mm5 \n\t"\
134         "1:                                \n\t"\
135         "movq      2*0+0(%1), %%mm1        \n\t"\
136         "movq      2*0+8(%1), %%mm2        \n\t"\
137         "movq      2*1+0(%1), %%mm3        \n\t"\
138         "movq      2*1+8(%1), %%mm4        \n\t"\
139         "paddw     2*3+0(%1), %%mm1        \n\t"\
140         "paddw     2*3+8(%1), %%mm2        \n\t"\
141         "paddw     2*2+0(%1), %%mm3        \n\t"\
142         "paddw     2*2+8(%1), %%mm4        \n\t"\
143         "pmullw    %%mm5, %%mm3            \n\t"\
144         "pmullw    %%mm5, %%mm4            \n\t"\
145         "psubw     %%mm1, %%mm3            \n\t"\
146         "psubw     %%mm2, %%mm4            \n\t"\
147         NORMALIZE_MMX("$7")\
148         /* Remove bias */\
149         "paddw     %%mm6, %%mm3            \n\t"\
150         "paddw     %%mm6, %%mm4            \n\t"\
151         TRANSFER_DO_PACK(OP)\
152         "add       $24, %1                 \n\t"\
153         "add       %3, %2                  \n\t"\
154         "decl      %0                      \n\t"\
155         "jnz 1b                            \n\t"\
156         : "+r"(h), "+r" (src),  "+r" (dst)\
157         : "r"(stride), "m"(rnd)\
158           NAMED_CONSTRAINTS_ADD(ff_pw_128,ff_pw_9)\
159         : "memory"\
160     );\
161 }
162
163 VC1_HOR_16b_SHIFT2(OP_PUT, put_)
164 VC1_HOR_16b_SHIFT2(OP_AVG, avg_)
165
166
167 /**
168  * Purely vertical or horizontal 1/2 shift interpolation.
169  * Sacrify mm6 for *9 factor.
170  */
171 #define VC1_SHIFT2(OP, OPNAME)\
172 static void OPNAME ## vc1_shift2_mmx(uint8_t *dst, const uint8_t *src,\
173                                      x86_reg stride, int rnd, x86_reg offset)\
174 {\
175     rnd = 8-rnd;\
176     __asm__ volatile(\
177         "mov       $8, %%"REG_c"           \n\t"\
178         LOAD_ROUNDER_MMX("%5")\
179         "movq      "MANGLE(ff_pw_9)", %%mm6\n\t"\
180         "1:                                \n\t"\
181         "movd      0(%0   ), %%mm3         \n\t"\
182         "movd      4(%0   ), %%mm4         \n\t"\
183         "movd      0(%0,%2), %%mm1         \n\t"\
184         "movd      4(%0,%2), %%mm2         \n\t"\
185         "add       %2, %0                  \n\t"\
186         "punpcklbw %%mm0, %%mm3            \n\t"\
187         "punpcklbw %%mm0, %%mm4            \n\t"\
188         "punpcklbw %%mm0, %%mm1            \n\t"\
189         "punpcklbw %%mm0, %%mm2            \n\t"\
190         "paddw     %%mm1, %%mm3            \n\t"\
191         "paddw     %%mm2, %%mm4            \n\t"\
192         "movd      0(%0,%3), %%mm1         \n\t"\
193         "movd      4(%0,%3), %%mm2         \n\t"\
194         "pmullw    %%mm6, %%mm3            \n\t" /* 0,9,9,0*/\
195         "pmullw    %%mm6, %%mm4            \n\t" /* 0,9,9,0*/\
196         "punpcklbw %%mm0, %%mm1            \n\t"\
197         "punpcklbw %%mm0, %%mm2            \n\t"\
198         "psubw     %%mm1, %%mm3            \n\t" /*-1,9,9,0*/\
199         "psubw     %%mm2, %%mm4            \n\t" /*-1,9,9,0*/\
200         "movd      0(%0,%2), %%mm1         \n\t"\
201         "movd      4(%0,%2), %%mm2         \n\t"\
202         "punpcklbw %%mm0, %%mm1            \n\t"\
203         "punpcklbw %%mm0, %%mm2            \n\t"\
204         "psubw     %%mm1, %%mm3            \n\t" /*-1,9,9,-1*/\
205         "psubw     %%mm2, %%mm4            \n\t" /*-1,9,9,-1*/\
206         NORMALIZE_MMX("$4")\
207         "packuswb  %%mm4, %%mm3            \n\t"\
208         OP((%1), %%mm3)\
209         "movq      %%mm3, (%1)             \n\t"\
210         "add       %6, %0                  \n\t"\
211         "add       %4, %1                  \n\t"\
212         "dec       %%"REG_c"               \n\t"\
213         "jnz 1b                            \n\t"\
214         : "+r"(src),  "+r"(dst)\
215         : "r"(offset), "r"(-2*offset), "g"(stride), "m"(rnd),\
216           "g"(stride-offset)\
217           NAMED_CONSTRAINTS_ADD(ff_pw_9)\
218         : "%"REG_c, "memory"\
219     );\
220 }
221
222 VC1_SHIFT2(OP_PUT, put_)
223 VC1_SHIFT2(OP_AVG, avg_)
224
225 /**
226  * Core of the 1/4 and 3/4 shift bicubic interpolation.
227  *
228  * @param UNPACK  Macro unpacking arguments from 8 to 16bits (can be empty).
229  * @param MOVQ    "movd 1" or "movq 2", if data read is already unpacked.
230  * @param A1      Address of 1st tap (beware of unpacked/packed).
231  * @param A2      Address of 2nd tap
232  * @param A3      Address of 3rd tap
233  * @param A4      Address of 4th tap
234  */
235 #define MSPEL_FILTER13_CORE(UNPACK, MOVQ, A1, A2, A3, A4)       \
236      MOVQ "*0+"A1", %%mm1       \n\t"                           \
237      MOVQ "*4+"A1", %%mm2       \n\t"                           \
238      UNPACK("%%mm1")                                            \
239      UNPACK("%%mm2")                                            \
240      "pmullw    "MANGLE(ff_pw_3)", %%mm1\n\t"                   \
241      "pmullw    "MANGLE(ff_pw_3)", %%mm2\n\t"                   \
242      MOVQ "*0+"A2", %%mm3       \n\t"                           \
243      MOVQ "*4+"A2", %%mm4       \n\t"                           \
244      UNPACK("%%mm3")                                            \
245      UNPACK("%%mm4")                                            \
246      "pmullw    %%mm6, %%mm3    \n\t" /* *18 */                 \
247      "pmullw    %%mm6, %%mm4    \n\t" /* *18 */                 \
248      "psubw     %%mm1, %%mm3    \n\t" /* 18,-3 */               \
249      "psubw     %%mm2, %%mm4    \n\t" /* 18,-3 */               \
250      MOVQ "*0+"A4", %%mm1       \n\t"                           \
251      MOVQ "*4+"A4", %%mm2       \n\t"                           \
252      UNPACK("%%mm1")                                            \
253      UNPACK("%%mm2")                                            \
254      "psllw     $2, %%mm1       \n\t" /* 4* */                  \
255      "psllw     $2, %%mm2       \n\t" /* 4* */                  \
256      "psubw     %%mm1, %%mm3    \n\t" /* -4,18,-3 */            \
257      "psubw     %%mm2, %%mm4    \n\t" /* -4,18,-3 */            \
258      MOVQ "*0+"A3", %%mm1       \n\t"                           \
259      MOVQ "*4+"A3", %%mm2       \n\t"                           \
260      UNPACK("%%mm1")                                            \
261      UNPACK("%%mm2")                                            \
262      "pmullw    %%mm5, %%mm1    \n\t" /* *53 */                 \
263      "pmullw    %%mm5, %%mm2    \n\t" /* *53 */                 \
264      "paddw     %%mm1, %%mm3    \n\t" /* 4,53,18,-3 */          \
265      "paddw     %%mm2, %%mm4    \n\t" /* 4,53,18,-3 */
266
267 /**
268  * Macro to build the vertical 16bits version of vc1_put_shift[13].
269  * Here, offset=src_stride. Parameters passed A1 to A4 must use
270  * %3 (src_stride) and %4 (3*src_stride).
271  *
272  * @param  NAME   Either 1 or 3
273  * @see MSPEL_FILTER13_CORE for information on A1->A4
274  */
275 #define MSPEL_FILTER13_VER_16B(NAME, A1, A2, A3, A4)                    \
276 static void                                                             \
277 vc1_put_ver_16b_ ## NAME ## _mmx(int16_t *dst, const uint8_t *src,      \
278                                  x86_reg src_stride,                   \
279                                  int rnd, int64_t shift)                \
280 {                                                                       \
281     int h = 8;                                                          \
282     src -= src_stride;                                                  \
283     __asm__ volatile(                                                       \
284         LOAD_ROUNDER_MMX("%5")                                          \
285         "movq      "MANGLE(ff_pw_53)", %%mm5\n\t"                       \
286         "movq      "MANGLE(ff_pw_18)", %%mm6\n\t"                       \
287         ".p2align 3                \n\t"                                \
288         "1:                        \n\t"                                \
289         MSPEL_FILTER13_CORE(DO_UNPACK, "movd  1", A1, A2, A3, A4)       \
290         NORMALIZE_MMX("%6")                                             \
291         TRANSFER_DONT_PACK(OP_PUT)                                      \
292         /* Last 3 (in fact 4) bytes on the line */                      \
293         "movd      8+"A1", %%mm1   \n\t"                                \
294         DO_UNPACK("%%mm1")                                              \
295         "movq      %%mm1, %%mm3    \n\t"                                \
296         "paddw     %%mm1, %%mm1    \n\t"                                \
297         "paddw     %%mm3, %%mm1    \n\t" /* 3* */                       \
298         "movd      8+"A2", %%mm3   \n\t"                                \
299         DO_UNPACK("%%mm3")                                              \
300         "pmullw    %%mm6, %%mm3    \n\t" /* *18 */                      \
301         "psubw     %%mm1, %%mm3    \n\t" /*18,-3 */                     \
302         "movd      8+"A3", %%mm1   \n\t"                                \
303         DO_UNPACK("%%mm1")                                              \
304         "pmullw    %%mm5, %%mm1    \n\t" /* *53 */                      \
305         "paddw     %%mm1, %%mm3    \n\t" /*53,18,-3 */                  \
306         "movd      8+"A4", %%mm1   \n\t"                                \
307         DO_UNPACK("%%mm1")                                              \
308         "psllw     $2, %%mm1       \n\t" /* 4* */                       \
309         "psubw     %%mm1, %%mm3    \n\t"                                \
310         "paddw     %%mm7, %%mm3    \n\t"                                \
311         "psraw     %6, %%mm3       \n\t"                                \
312         "movq      %%mm3, 16(%2)   \n\t"                                \
313         "add       %3, %1          \n\t"                                \
314         "add       $24, %2         \n\t"                                \
315         "decl      %0              \n\t"                                \
316         "jnz 1b                    \n\t"                                \
317         : "+r"(h), "+r" (src),  "+r" (dst)                              \
318         : "r"(src_stride), "r"(3*src_stride),                           \
319           "m"(rnd), "m"(shift)                                          \
320           NAMED_CONSTRAINTS_ADD(ff_pw_3,ff_pw_53,ff_pw_18)              \
321         : "memory"                                                      \
322     );                                                                  \
323 }
324
325 /**
326  * Macro to build the horizontal 16bits version of vc1_put_shift[13].
327  * Here, offset=16bits, so parameters passed A1 to A4 should be simple.
328  *
329  * @param  NAME   Either 1 or 3
330  * @see MSPEL_FILTER13_CORE for information on A1->A4
331  */
332 #define MSPEL_FILTER13_HOR_16B(NAME, A1, A2, A3, A4, OP, OPNAME)        \
333 static void                                                             \
334 OPNAME ## vc1_hor_16b_ ## NAME ## _mmx(uint8_t *dst, x86_reg stride,    \
335                                  const int16_t *src, int rnd)           \
336 {                                                                       \
337     int h = 8;                                                          \
338     src -= 1;                                                           \
339     rnd -= (-4+58+13-3)*256; /* Add -256 bias */                        \
340     __asm__ volatile(                                                       \
341         LOAD_ROUNDER_MMX("%4")                                          \
342         "movq      "MANGLE(ff_pw_18)", %%mm6   \n\t"                    \
343         "movq      "MANGLE(ff_pw_53)", %%mm5   \n\t"                    \
344         ".p2align 3                \n\t"                                \
345         "1:                        \n\t"                                \
346         MSPEL_FILTER13_CORE(DONT_UNPACK, "movq 2", A1, A2, A3, A4)      \
347         NORMALIZE_MMX("$7")                                             \
348         /* Remove bias */                                               \
349         "paddw     "MANGLE(ff_pw_128)", %%mm3  \n\t"                    \
350         "paddw     "MANGLE(ff_pw_128)", %%mm4  \n\t"                    \
351         TRANSFER_DO_PACK(OP)                                            \
352         "add       $24, %1         \n\t"                                \
353         "add       %3, %2          \n\t"                                \
354         "decl      %0              \n\t"                                \
355         "jnz 1b                    \n\t"                                \
356         : "+r"(h), "+r" (src),  "+r" (dst)                              \
357         : "r"(stride), "m"(rnd)                                         \
358           NAMED_CONSTRAINTS_ADD(ff_pw_3,ff_pw_18,ff_pw_53,ff_pw_128)    \
359         : "memory"                                                      \
360     );                                                                  \
361 }
362
363 /**
364  * Macro to build the 8bits, any direction, version of vc1_put_shift[13].
365  * Here, offset=src_stride. Parameters passed A1 to A4 must use
366  * %3 (offset) and %4 (3*offset).
367  *
368  * @param  NAME   Either 1 or 3
369  * @see MSPEL_FILTER13_CORE for information on A1->A4
370  */
371 #define MSPEL_FILTER13_8B(NAME, A1, A2, A3, A4, OP, OPNAME)             \
372 static void                                                             \
373 OPNAME ## vc1_## NAME ## _mmx(uint8_t *dst, const uint8_t *src,         \
374                         x86_reg stride, int rnd, x86_reg offset)      \
375 {                                                                       \
376     int h = 8;                                                          \
377     src -= offset;                                                      \
378     rnd = 32-rnd;                                                       \
379     __asm__ volatile (                                                      \
380         LOAD_ROUNDER_MMX("%6")                                          \
381         "movq      "MANGLE(ff_pw_53)", %%mm5       \n\t"                \
382         "movq      "MANGLE(ff_pw_18)", %%mm6       \n\t"                \
383         ".p2align 3                \n\t"                                \
384         "1:                        \n\t"                                \
385         MSPEL_FILTER13_CORE(DO_UNPACK, "movd   1", A1, A2, A3, A4)      \
386         NORMALIZE_MMX("$6")                                             \
387         TRANSFER_DO_PACK(OP)                                            \
388         "add       %5, %1          \n\t"                                \
389         "add       %5, %2          \n\t"                                \
390         "decl      %0              \n\t"                                \
391         "jnz 1b                    \n\t"                                \
392         : "+r"(h), "+r" (src),  "+r" (dst)                              \
393         : "r"(offset), "r"(3*offset), "g"(stride), "m"(rnd)             \
394           NAMED_CONSTRAINTS_ADD(ff_pw_53,ff_pw_18,ff_pw_3)              \
395         : "memory"                                                      \
396     );                                                                  \
397 }
398
399 /** 1/4 shift bicubic interpolation */
400 MSPEL_FILTER13_8B     (shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )", OP_PUT, put_)
401 MSPEL_FILTER13_8B     (shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )", OP_AVG, avg_)
402 MSPEL_FILTER13_VER_16B(shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )")
403 MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)", OP_PUT, put_)
404 MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)", OP_AVG, avg_)
405
406 /** 3/4 shift bicubic interpolation */
407 MSPEL_FILTER13_8B     (shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )", OP_PUT, put_)
408 MSPEL_FILTER13_8B     (shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )", OP_AVG, avg_)
409 MSPEL_FILTER13_VER_16B(shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )")
410 MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)", OP_PUT, put_)
411 MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)", OP_AVG, avg_)
412
413 typedef void (*vc1_mspel_mc_filter_ver_16bits)(int16_t *dst, const uint8_t *src, x86_reg src_stride, int rnd, int64_t shift);
414 typedef void (*vc1_mspel_mc_filter_hor_16bits)(uint8_t *dst, x86_reg dst_stride, const int16_t *src, int rnd);
415 typedef void (*vc1_mspel_mc_filter_8bits)(uint8_t *dst, const uint8_t *src, x86_reg stride, int rnd, x86_reg offset);
416
417 /**
418  * Interpolate fractional pel values by applying proper vertical then
419  * horizontal filter.
420  *
421  * @param  dst     Destination buffer for interpolated pels.
422  * @param  src     Source buffer.
423  * @param  stride  Stride for both src and dst buffers.
424  * @param  hmode   Horizontal filter (expressed in quarter pixels shift).
425  * @param  hmode   Vertical filter.
426  * @param  rnd     Rounding bias.
427  */
428 #define VC1_MSPEL_MC(OP)\
429 static void OP ## vc1_mspel_mc(uint8_t *dst, const uint8_t *src, int stride,\
430                                int hmode, int vmode, int rnd)\
431 {\
432     static const vc1_mspel_mc_filter_ver_16bits vc1_put_shift_ver_16bits[] =\
433          { NULL, vc1_put_ver_16b_shift1_mmx, vc1_put_ver_16b_shift2_mmx, vc1_put_ver_16b_shift3_mmx };\
434     static const vc1_mspel_mc_filter_hor_16bits vc1_put_shift_hor_16bits[] =\
435          { NULL, OP ## vc1_hor_16b_shift1_mmx, OP ## vc1_hor_16b_shift2_mmx, OP ## vc1_hor_16b_shift3_mmx };\
436     static const vc1_mspel_mc_filter_8bits vc1_put_shift_8bits[] =\
437          { NULL, OP ## vc1_shift1_mmx, OP ## vc1_shift2_mmx, OP ## vc1_shift3_mmx };\
438 \
439     __asm__ volatile(\
440         "pxor %%mm0, %%mm0         \n\t"\
441         ::: "memory"\
442     );\
443 \
444     if (vmode) { /* Vertical filter to apply */\
445         if (hmode) { /* Horizontal filter to apply, output to tmp */\
446             static const int shift_value[] = { 0, 5, 1, 5 };\
447             int              shift = (shift_value[hmode]+shift_value[vmode])>>1;\
448             int              r;\
449             LOCAL_ALIGNED(16, int16_t, tmp, [12*8]);\
450 \
451             r = (1<<(shift-1)) + rnd-1;\
452             vc1_put_shift_ver_16bits[vmode](tmp, src-1, stride, r, shift);\
453 \
454             vc1_put_shift_hor_16bits[hmode](dst, stride, tmp+1, 64-rnd);\
455             return;\
456         }\
457         else { /* No horizontal filter, output 8 lines to dst */\
458             vc1_put_shift_8bits[vmode](dst, src, stride, 1-rnd, stride);\
459             return;\
460         }\
461     }\
462 \
463     /* Horizontal mode with no vertical mode */\
464     vc1_put_shift_8bits[hmode](dst, src, stride, rnd, 1);\
465 } \
466 static void OP ## vc1_mspel_mc_16(uint8_t *dst, const uint8_t *src, \
467                                   int stride, int hmode, int vmode, int rnd)\
468 { \
469     OP ## vc1_mspel_mc(dst + 0, src + 0, stride, hmode, vmode, rnd); \
470     OP ## vc1_mspel_mc(dst + 8, src + 8, stride, hmode, vmode, rnd); \
471     dst += 8*stride; src += 8*stride; \
472     OP ## vc1_mspel_mc(dst + 0, src + 0, stride, hmode, vmode, rnd); \
473     OP ## vc1_mspel_mc(dst + 8, src + 8, stride, hmode, vmode, rnd); \
474 }
475
476 VC1_MSPEL_MC(put_)
477 VC1_MSPEL_MC(avg_)
478
479 /** Macro to ease bicubic filter interpolation functions declarations */
480 #define DECLARE_FUNCTION(a, b)                                          \
481 static void put_vc1_mspel_mc ## a ## b ## _mmx(uint8_t *dst,            \
482                                                const uint8_t *src,      \
483                                                ptrdiff_t stride,        \
484                                                int rnd)                 \
485 {                                                                       \
486      put_vc1_mspel_mc(dst, src, stride, a, b, rnd);                     \
487 }\
488 static void avg_vc1_mspel_mc ## a ## b ## _mmxext(uint8_t *dst,         \
489                                                   const uint8_t *src,   \
490                                                   ptrdiff_t stride,     \
491                                                   int rnd)              \
492 {                                                                       \
493      avg_vc1_mspel_mc(dst, src, stride, a, b, rnd);                     \
494 }\
495 static void put_vc1_mspel_mc ## a ## b ## _16_mmx(uint8_t *dst,         \
496                                                   const uint8_t *src,   \
497                                                   ptrdiff_t stride,     \
498                                                   int rnd)              \
499 {                                                                       \
500      put_vc1_mspel_mc_16(dst, src, stride, a, b, rnd);                  \
501 }\
502 static void avg_vc1_mspel_mc ## a ## b ## _16_mmxext(uint8_t *dst,      \
503                                                      const uint8_t *src,\
504                                                      ptrdiff_t stride,  \
505                                                      int rnd)           \
506 {                                                                       \
507      avg_vc1_mspel_mc_16(dst, src, stride, a, b, rnd);                  \
508 }
509
510 DECLARE_FUNCTION(0, 1)
511 DECLARE_FUNCTION(0, 2)
512 DECLARE_FUNCTION(0, 3)
513
514 DECLARE_FUNCTION(1, 0)
515 DECLARE_FUNCTION(1, 1)
516 DECLARE_FUNCTION(1, 2)
517 DECLARE_FUNCTION(1, 3)
518
519 DECLARE_FUNCTION(2, 0)
520 DECLARE_FUNCTION(2, 1)
521 DECLARE_FUNCTION(2, 2)
522 DECLARE_FUNCTION(2, 3)
523
524 DECLARE_FUNCTION(3, 0)
525 DECLARE_FUNCTION(3, 1)
526 DECLARE_FUNCTION(3, 2)
527 DECLARE_FUNCTION(3, 3)
528
529 static void vc1_inv_trans_4x4_dc_mmxext(uint8_t *dest, int linesize,
530                                         int16_t *block)
531 {
532     int dc = block[0];
533     dc = (17 * dc +  4) >> 3;
534     dc = (17 * dc + 64) >> 7;
535     __asm__ volatile(
536         "movd          %0, %%mm0 \n\t"
537         "pshufw $0, %%mm0, %%mm0 \n\t"
538         "pxor       %%mm1, %%mm1 \n\t"
539         "psubw      %%mm0, %%mm1 \n\t"
540         "packuswb   %%mm0, %%mm0 \n\t"
541         "packuswb   %%mm1, %%mm1 \n\t"
542         ::"r"(dc)
543     );
544     __asm__ volatile(
545         "movd          %0, %%mm2 \n\t"
546         "movd          %1, %%mm3 \n\t"
547         "movd          %2, %%mm4 \n\t"
548         "movd          %3, %%mm5 \n\t"
549         "paddusb    %%mm0, %%mm2 \n\t"
550         "paddusb    %%mm0, %%mm3 \n\t"
551         "paddusb    %%mm0, %%mm4 \n\t"
552         "paddusb    %%mm0, %%mm5 \n\t"
553         "psubusb    %%mm1, %%mm2 \n\t"
554         "psubusb    %%mm1, %%mm3 \n\t"
555         "psubusb    %%mm1, %%mm4 \n\t"
556         "psubusb    %%mm1, %%mm5 \n\t"
557         "movd       %%mm2, %0    \n\t"
558         "movd       %%mm3, %1    \n\t"
559         "movd       %%mm4, %2    \n\t"
560         "movd       %%mm5, %3    \n\t"
561         :"+m"(*(uint32_t*)(dest+0*linesize)),
562          "+m"(*(uint32_t*)(dest+1*linesize)),
563          "+m"(*(uint32_t*)(dest+2*linesize)),
564          "+m"(*(uint32_t*)(dest+3*linesize))
565     );
566 }
567
568 static void vc1_inv_trans_4x8_dc_mmxext(uint8_t *dest, int linesize,
569                                         int16_t *block)
570 {
571     int dc = block[0];
572     dc = (17 * dc +  4) >> 3;
573     dc = (12 * dc + 64) >> 7;
574     __asm__ volatile(
575         "movd          %0, %%mm0 \n\t"
576         "pshufw $0, %%mm0, %%mm0 \n\t"
577         "pxor       %%mm1, %%mm1 \n\t"
578         "psubw      %%mm0, %%mm1 \n\t"
579         "packuswb   %%mm0, %%mm0 \n\t"
580         "packuswb   %%mm1, %%mm1 \n\t"
581         ::"r"(dc)
582     );
583     __asm__ volatile(
584         "movd          %0, %%mm2 \n\t"
585         "movd          %1, %%mm3 \n\t"
586         "movd          %2, %%mm4 \n\t"
587         "movd          %3, %%mm5 \n\t"
588         "paddusb    %%mm0, %%mm2 \n\t"
589         "paddusb    %%mm0, %%mm3 \n\t"
590         "paddusb    %%mm0, %%mm4 \n\t"
591         "paddusb    %%mm0, %%mm5 \n\t"
592         "psubusb    %%mm1, %%mm2 \n\t"
593         "psubusb    %%mm1, %%mm3 \n\t"
594         "psubusb    %%mm1, %%mm4 \n\t"
595         "psubusb    %%mm1, %%mm5 \n\t"
596         "movd       %%mm2, %0    \n\t"
597         "movd       %%mm3, %1    \n\t"
598         "movd       %%mm4, %2    \n\t"
599         "movd       %%mm5, %3    \n\t"
600         :"+m"(*(uint32_t*)(dest+0*linesize)),
601          "+m"(*(uint32_t*)(dest+1*linesize)),
602          "+m"(*(uint32_t*)(dest+2*linesize)),
603          "+m"(*(uint32_t*)(dest+3*linesize))
604     );
605     dest += 4*linesize;
606     __asm__ volatile(
607         "movd          %0, %%mm2 \n\t"
608         "movd          %1, %%mm3 \n\t"
609         "movd          %2, %%mm4 \n\t"
610         "movd          %3, %%mm5 \n\t"
611         "paddusb    %%mm0, %%mm2 \n\t"
612         "paddusb    %%mm0, %%mm3 \n\t"
613         "paddusb    %%mm0, %%mm4 \n\t"
614         "paddusb    %%mm0, %%mm5 \n\t"
615         "psubusb    %%mm1, %%mm2 \n\t"
616         "psubusb    %%mm1, %%mm3 \n\t"
617         "psubusb    %%mm1, %%mm4 \n\t"
618         "psubusb    %%mm1, %%mm5 \n\t"
619         "movd       %%mm2, %0    \n\t"
620         "movd       %%mm3, %1    \n\t"
621         "movd       %%mm4, %2    \n\t"
622         "movd       %%mm5, %3    \n\t"
623         :"+m"(*(uint32_t*)(dest+0*linesize)),
624          "+m"(*(uint32_t*)(dest+1*linesize)),
625          "+m"(*(uint32_t*)(dest+2*linesize)),
626          "+m"(*(uint32_t*)(dest+3*linesize))
627     );
628 }
629
630 static void vc1_inv_trans_8x4_dc_mmxext(uint8_t *dest, int linesize,
631                                         int16_t *block)
632 {
633     int dc = block[0];
634     dc = ( 3 * dc +  1) >> 1;
635     dc = (17 * dc + 64) >> 7;
636     __asm__ volatile(
637         "movd          %0, %%mm0 \n\t"
638         "pshufw $0, %%mm0, %%mm0 \n\t"
639         "pxor       %%mm1, %%mm1 \n\t"
640         "psubw      %%mm0, %%mm1 \n\t"
641         "packuswb   %%mm0, %%mm0 \n\t"
642         "packuswb   %%mm1, %%mm1 \n\t"
643         ::"r"(dc)
644     );
645     __asm__ volatile(
646         "movq          %0, %%mm2 \n\t"
647         "movq          %1, %%mm3 \n\t"
648         "movq          %2, %%mm4 \n\t"
649         "movq          %3, %%mm5 \n\t"
650         "paddusb    %%mm0, %%mm2 \n\t"
651         "paddusb    %%mm0, %%mm3 \n\t"
652         "paddusb    %%mm0, %%mm4 \n\t"
653         "paddusb    %%mm0, %%mm5 \n\t"
654         "psubusb    %%mm1, %%mm2 \n\t"
655         "psubusb    %%mm1, %%mm3 \n\t"
656         "psubusb    %%mm1, %%mm4 \n\t"
657         "psubusb    %%mm1, %%mm5 \n\t"
658         "movq       %%mm2, %0    \n\t"
659         "movq       %%mm3, %1    \n\t"
660         "movq       %%mm4, %2    \n\t"
661         "movq       %%mm5, %3    \n\t"
662         :"+m"(*(uint32_t*)(dest+0*linesize)),
663          "+m"(*(uint32_t*)(dest+1*linesize)),
664          "+m"(*(uint32_t*)(dest+2*linesize)),
665          "+m"(*(uint32_t*)(dest+3*linesize))
666     );
667 }
668
669 static void vc1_inv_trans_8x8_dc_mmxext(uint8_t *dest, int linesize,
670                                         int16_t *block)
671 {
672     int dc = block[0];
673     dc = (3 * dc +  1) >> 1;
674     dc = (3 * dc + 16) >> 5;
675     __asm__ volatile(
676         "movd          %0, %%mm0 \n\t"
677         "pshufw $0, %%mm0, %%mm0 \n\t"
678         "pxor       %%mm1, %%mm1 \n\t"
679         "psubw      %%mm0, %%mm1 \n\t"
680         "packuswb   %%mm0, %%mm0 \n\t"
681         "packuswb   %%mm1, %%mm1 \n\t"
682         ::"r"(dc)
683     );
684     __asm__ volatile(
685         "movq          %0, %%mm2 \n\t"
686         "movq          %1, %%mm3 \n\t"
687         "movq          %2, %%mm4 \n\t"
688         "movq          %3, %%mm5 \n\t"
689         "paddusb    %%mm0, %%mm2 \n\t"
690         "paddusb    %%mm0, %%mm3 \n\t"
691         "paddusb    %%mm0, %%mm4 \n\t"
692         "paddusb    %%mm0, %%mm5 \n\t"
693         "psubusb    %%mm1, %%mm2 \n\t"
694         "psubusb    %%mm1, %%mm3 \n\t"
695         "psubusb    %%mm1, %%mm4 \n\t"
696         "psubusb    %%mm1, %%mm5 \n\t"
697         "movq       %%mm2, %0    \n\t"
698         "movq       %%mm3, %1    \n\t"
699         "movq       %%mm4, %2    \n\t"
700         "movq       %%mm5, %3    \n\t"
701         :"+m"(*(uint32_t*)(dest+0*linesize)),
702          "+m"(*(uint32_t*)(dest+1*linesize)),
703          "+m"(*(uint32_t*)(dest+2*linesize)),
704          "+m"(*(uint32_t*)(dest+3*linesize))
705     );
706     dest += 4*linesize;
707     __asm__ volatile(
708         "movq          %0, %%mm2 \n\t"
709         "movq          %1, %%mm3 \n\t"
710         "movq          %2, %%mm4 \n\t"
711         "movq          %3, %%mm5 \n\t"
712         "paddusb    %%mm0, %%mm2 \n\t"
713         "paddusb    %%mm0, %%mm3 \n\t"
714         "paddusb    %%mm0, %%mm4 \n\t"
715         "paddusb    %%mm0, %%mm5 \n\t"
716         "psubusb    %%mm1, %%mm2 \n\t"
717         "psubusb    %%mm1, %%mm3 \n\t"
718         "psubusb    %%mm1, %%mm4 \n\t"
719         "psubusb    %%mm1, %%mm5 \n\t"
720         "movq       %%mm2, %0    \n\t"
721         "movq       %%mm3, %1    \n\t"
722         "movq       %%mm4, %2    \n\t"
723         "movq       %%mm5, %3    \n\t"
724         :"+m"(*(uint32_t*)(dest+0*linesize)),
725          "+m"(*(uint32_t*)(dest+1*linesize)),
726          "+m"(*(uint32_t*)(dest+2*linesize)),
727          "+m"(*(uint32_t*)(dest+3*linesize))
728     );
729 }
730
731 #if HAVE_MMX_EXTERNAL
732 static void put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
733                                    ptrdiff_t stride, int rnd)
734 {
735     ff_put_pixels8_mmx(dst, src, stride, 8);
736 }
737 static void put_vc1_mspel_mc00_16_mmx(uint8_t *dst, const uint8_t *src,
738                                       ptrdiff_t stride, int rnd)
739 {
740     ff_put_pixels16_mmx(dst, src, stride, 16);
741 }
742 static void avg_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
743                                    ptrdiff_t stride, int rnd)
744 {
745     ff_avg_pixels8_mmx(dst, src, stride, 8);
746 }
747 static void avg_vc1_mspel_mc00_16_mmx(uint8_t *dst, const uint8_t *src,
748                                       ptrdiff_t stride, int rnd)
749 {
750     ff_avg_pixels16_mmx(dst, src, stride, 16);
751 }
752 #endif
753
754 #define FN_ASSIGN(OP, X, Y, INSN) \
755     dsp->OP##vc1_mspel_pixels_tab[1][X+4*Y] = OP##vc1_mspel_mc##X##Y##INSN; \
756     dsp->OP##vc1_mspel_pixels_tab[0][X+4*Y] = OP##vc1_mspel_mc##X##Y##_16##INSN
757
758 av_cold void ff_vc1dsp_init_mmx(VC1DSPContext *dsp)
759 {
760 #if HAVE_MMX_EXTERNAL
761     FN_ASSIGN(put_, 0, 0, _mmx);
762     FN_ASSIGN(avg_, 0, 0, _mmx);
763 #endif
764     FN_ASSIGN(put_, 0, 1, _mmx);
765     FN_ASSIGN(put_, 0, 2, _mmx);
766     FN_ASSIGN(put_, 0, 3, _mmx);
767
768     FN_ASSIGN(put_, 1, 0, _mmx);
769     FN_ASSIGN(put_, 1, 1, _mmx);
770     FN_ASSIGN(put_, 1, 2, _mmx);
771     FN_ASSIGN(put_, 1, 3, _mmx);
772
773     FN_ASSIGN(put_, 2, 0, _mmx);
774     FN_ASSIGN(put_, 2, 1, _mmx);
775     FN_ASSIGN(put_, 2, 2, _mmx);
776     FN_ASSIGN(put_, 2, 3, _mmx);
777
778     FN_ASSIGN(put_, 3, 0, _mmx);
779     FN_ASSIGN(put_, 3, 1, _mmx);
780     FN_ASSIGN(put_, 3, 2, _mmx);
781     FN_ASSIGN(put_, 3, 3, _mmx);
782 }
783
784 av_cold void ff_vc1dsp_init_mmxext(VC1DSPContext *dsp)
785 {
786     FN_ASSIGN(avg_, 0, 1, _mmxext);
787     FN_ASSIGN(avg_, 0, 2, _mmxext);
788     FN_ASSIGN(avg_, 0, 3, _mmxext);
789
790     FN_ASSIGN(avg_, 1, 0, _mmxext);
791     FN_ASSIGN(avg_, 1, 1, _mmxext);
792     FN_ASSIGN(avg_, 1, 2, _mmxext);
793     FN_ASSIGN(avg_, 1, 3, _mmxext);
794
795     FN_ASSIGN(avg_, 2, 0, _mmxext);
796     FN_ASSIGN(avg_, 2, 1, _mmxext);
797     FN_ASSIGN(avg_, 2, 2, _mmxext);
798     FN_ASSIGN(avg_, 2, 3, _mmxext);
799
800     FN_ASSIGN(avg_, 3, 0, _mmxext);
801     FN_ASSIGN(avg_, 3, 1, _mmxext);
802     FN_ASSIGN(avg_, 3, 2, _mmxext);
803     FN_ASSIGN(avg_, 3, 3, _mmxext);
804
805     dsp->vc1_inv_trans_8x8_dc = vc1_inv_trans_8x8_dc_mmxext;
806     dsp->vc1_inv_trans_4x8_dc = vc1_inv_trans_4x8_dc_mmxext;
807     dsp->vc1_inv_trans_8x4_dc = vc1_inv_trans_8x4_dc_mmxext;
808     dsp->vc1_inv_trans_4x4_dc = vc1_inv_trans_4x4_dc_mmxext;
809 }
810 #endif /* HAVE_6REGS && HAVE_INLINE_ASM */