2 * Chinese AVS video (AVS1-P2, JiZhun profile) decoder.
3 * Copyright (c) 2006 Stefan Gehrer <stefan.gehrer@gmx.de>
5 * MMX-optimized DSP functions, based on H.264 optimizations by
6 * Michael Niedermayer and Loren Merritt
8 * This file is part of Libav.
10 * Libav is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
15 * Libav is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with Libav; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 #include "libavutil/attributes.h"
26 #include "libavutil/common.h"
27 #include "libavutil/cpu.h"
28 #include "libavutil/x86/asm.h"
29 #include "libavutil/x86/cpu.h"
30 #include "libavcodec/cavsdsp.h"
31 #include "dsputil_mmx.h"
36 /* in/out: mma=mma+mmb, mmb=mmb-mma */
37 #define SUMSUB_BA( a, b ) \
38 "paddw "#b", "#a" \n\t"\
39 "paddw "#b", "#b" \n\t"\
40 "psubw "#a", "#b" \n\t"
42 /*****************************************************************************
46 ****************************************************************************/
48 static inline void cavs_idct8_1d(int16_t *block, uint64_t bias)
51 "movq 112(%0), %%mm4 \n\t" /* mm4 = src7 */
52 "movq 16(%0), %%mm5 \n\t" /* mm5 = src1 */
53 "movq 80(%0), %%mm2 \n\t" /* mm2 = src5 */
54 "movq 48(%0), %%mm7 \n\t" /* mm7 = src3 */
55 "movq %%mm4, %%mm0 \n\t"
56 "movq %%mm5, %%mm3 \n\t"
57 "movq %%mm2, %%mm6 \n\t"
58 "movq %%mm7, %%mm1 \n\t"
60 "paddw %%mm4, %%mm4 \n\t" /* mm4 = 2*src7 */
61 "paddw %%mm3, %%mm3 \n\t" /* mm3 = 2*src1 */
62 "paddw %%mm6, %%mm6 \n\t" /* mm6 = 2*src5 */
63 "paddw %%mm1, %%mm1 \n\t" /* mm1 = 2*src3 */
64 "paddw %%mm4, %%mm0 \n\t" /* mm0 = 3*src7 */
65 "paddw %%mm3, %%mm5 \n\t" /* mm5 = 3*src1 */
66 "paddw %%mm6, %%mm2 \n\t" /* mm2 = 3*src5 */
67 "paddw %%mm1, %%mm7 \n\t" /* mm7 = 3*src3 */
68 "psubw %%mm4, %%mm5 \n\t" /* mm5 = 3*src1 - 2*src7 = a0 */
69 "paddw %%mm6, %%mm7 \n\t" /* mm7 = 3*src3 + 2*src5 = a1 */
70 "psubw %%mm2, %%mm1 \n\t" /* mm1 = 2*src3 - 3*src5 = a2 */
71 "paddw %%mm0, %%mm3 \n\t" /* mm3 = 2*src1 + 3*src7 = a3 */
73 "movq %%mm5, %%mm4 \n\t"
74 "movq %%mm7, %%mm6 \n\t"
75 "movq %%mm3, %%mm0 \n\t"
76 "movq %%mm1, %%mm2 \n\t"
77 SUMSUB_BA( %%mm7, %%mm5 ) /* mm7 = a0 + a1 mm5 = a0 - a1 */
78 "paddw %%mm3, %%mm7 \n\t" /* mm7 = a0 + a1 + a3 */
79 "paddw %%mm1, %%mm5 \n\t" /* mm5 = a0 - a1 + a2 */
80 "paddw %%mm7, %%mm7 \n\t"
81 "paddw %%mm5, %%mm5 \n\t"
82 "paddw %%mm6, %%mm7 \n\t" /* mm7 = b4 */
83 "paddw %%mm4, %%mm5 \n\t" /* mm5 = b5 */
85 SUMSUB_BA( %%mm1, %%mm3 ) /* mm1 = a3 + a2 mm3 = a3 - a2 */
86 "psubw %%mm1, %%mm4 \n\t" /* mm4 = a0 - a2 - a3 */
87 "movq %%mm4, %%mm1 \n\t" /* mm1 = a0 - a2 - a3 */
88 "psubw %%mm6, %%mm3 \n\t" /* mm3 = a3 - a2 - a1 */
89 "paddw %%mm1, %%mm1 \n\t"
90 "paddw %%mm3, %%mm3 \n\t"
91 "psubw %%mm2, %%mm1 \n\t" /* mm1 = b7 */
92 "paddw %%mm0, %%mm3 \n\t" /* mm3 = b6 */
94 "movq 32(%0), %%mm2 \n\t" /* mm2 = src2 */
95 "movq 96(%0), %%mm6 \n\t" /* mm6 = src6 */
96 "movq %%mm2, %%mm4 \n\t"
97 "movq %%mm6, %%mm0 \n\t"
98 "psllw $2, %%mm4 \n\t" /* mm4 = 4*src2 */
99 "psllw $2, %%mm6 \n\t" /* mm6 = 4*src6 */
100 "paddw %%mm4, %%mm2 \n\t" /* mm2 = 5*src2 */
101 "paddw %%mm6, %%mm0 \n\t" /* mm0 = 5*src6 */
102 "paddw %%mm2, %%mm2 \n\t"
103 "paddw %%mm0, %%mm0 \n\t"
104 "psubw %%mm0, %%mm4 \n\t" /* mm4 = 4*src2 - 10*src6 = a7 */
105 "paddw %%mm2, %%mm6 \n\t" /* mm6 = 4*src6 + 10*src2 = a6 */
107 "movq (%0), %%mm2 \n\t" /* mm2 = src0 */
108 "movq 64(%0), %%mm0 \n\t" /* mm0 = src4 */
109 SUMSUB_BA( %%mm0, %%mm2 ) /* mm0 = src0+src4 mm2 = src0-src4 */
110 "psllw $3, %%mm0 \n\t"
111 "psllw $3, %%mm2 \n\t"
112 "paddw %1, %%mm0 \n\t" /* add rounding bias */
113 "paddw %1, %%mm2 \n\t" /* add rounding bias */
115 SUMSUB_BA( %%mm6, %%mm0 ) /* mm6 = a4 + a6 mm0 = a4 - a6 */
116 SUMSUB_BA( %%mm4, %%mm2 ) /* mm4 = a5 + a7 mm2 = a5 - a7 */
117 SUMSUB_BA( %%mm7, %%mm6 ) /* mm7 = dst0 mm6 = dst7 */
118 SUMSUB_BA( %%mm5, %%mm4 ) /* mm5 = dst1 mm4 = dst6 */
119 SUMSUB_BA( %%mm3, %%mm2 ) /* mm3 = dst2 mm2 = dst5 */
120 SUMSUB_BA( %%mm1, %%mm0 ) /* mm1 = dst3 mm0 = dst4 */
121 :: "r"(block), "m"(bias)
125 static void cavs_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride)
128 DECLARE_ALIGNED(8, int16_t, b2)[64];
131 DECLARE_ALIGNED(8, uint64_t, tmp);
133 cavs_idct8_1d(block+4*i, ff_pw_4.a);
136 "psraw $3, %%mm7 \n\t"
137 "psraw $3, %%mm6 \n\t"
138 "psraw $3, %%mm5 \n\t"
139 "psraw $3, %%mm4 \n\t"
140 "psraw $3, %%mm3 \n\t"
141 "psraw $3, %%mm2 \n\t"
142 "psraw $3, %%mm1 \n\t"
143 "psraw $3, %%mm0 \n\t"
144 "movq %%mm7, %0 \n\t"
145 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 )
146 "movq %%mm0, 8(%1) \n\t"
147 "movq %%mm6, 24(%1) \n\t"
148 "movq %%mm7, 40(%1) \n\t"
149 "movq %%mm4, 56(%1) \n\t"
150 "movq %0, %%mm7 \n\t"
151 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 )
152 "movq %%mm7, (%1) \n\t"
153 "movq %%mm1, 16(%1) \n\t"
154 "movq %%mm0, 32(%1) \n\t"
155 "movq %%mm3, 48(%1) \n\t"
163 cavs_idct8_1d(b2+4*i, ff_pw_64.a);
166 "psraw $7, %%mm7 \n\t"
167 "psraw $7, %%mm6 \n\t"
168 "psraw $7, %%mm5 \n\t"
169 "psraw $7, %%mm4 \n\t"
170 "psraw $7, %%mm3 \n\t"
171 "psraw $7, %%mm2 \n\t"
172 "psraw $7, %%mm1 \n\t"
173 "psraw $7, %%mm0 \n\t"
174 "movq %%mm7, (%0) \n\t"
175 "movq %%mm5, 16(%0) \n\t"
176 "movq %%mm3, 32(%0) \n\t"
177 "movq %%mm1, 48(%0) \n\t"
178 "movq %%mm0, 64(%0) \n\t"
179 "movq %%mm2, 80(%0) \n\t"
180 "movq %%mm4, 96(%0) \n\t"
181 "movq %%mm6, 112(%0) \n\t"
187 ff_add_pixels_clamped_mmx(b2, dst, stride);
190 #endif /* HAVE_MMX_INLINE */
192 #if (HAVE_MMXEXT_INLINE || HAVE_AMD3DNOW_INLINE)
194 /*****************************************************************************
196 * motion compensation
198 ****************************************************************************/
200 /* vertical filter [-1 -2 96 42 -7 0] */
201 #define QPEL_CAVSV1(A,B,C,D,E,F,OP,MUL2) \
202 "movd (%0), "#F" \n\t"\
203 "movq "#C", %%mm6 \n\t"\
204 "pmullw %5, %%mm6 \n\t"\
205 "movq "#D", %%mm7 \n\t"\
206 "pmullw "MANGLE(MUL2)", %%mm7\n\t"\
207 "psllw $3, "#E" \n\t"\
208 "psubw "#E", %%mm6 \n\t"\
209 "psraw $3, "#E" \n\t"\
210 "paddw %%mm7, %%mm6 \n\t"\
211 "paddw "#E", %%mm6 \n\t"\
212 "paddw "#B", "#B" \n\t"\
213 "pxor %%mm7, %%mm7 \n\t"\
215 "punpcklbw %%mm7, "#F" \n\t"\
216 "psubw "#B", %%mm6 \n\t"\
217 "psraw $1, "#B" \n\t"\
218 "psubw "#A", %%mm6 \n\t"\
219 "paddw %4, %%mm6 \n\t"\
220 "psraw $7, %%mm6 \n\t"\
221 "packuswb %%mm6, %%mm6 \n\t"\
222 OP(%%mm6, (%1), A, d) \
225 /* vertical filter [ 0 -1 5 5 -1 0] */
226 #define QPEL_CAVSV2(A,B,C,D,E,F,OP,MUL2) \
227 "movd (%0), "#F" \n\t"\
228 "movq "#C", %%mm6 \n\t"\
229 "paddw "#D", %%mm6 \n\t"\
230 "pmullw %5, %%mm6 \n\t"\
232 "punpcklbw %%mm7, "#F" \n\t"\
233 "psubw "#B", %%mm6 \n\t"\
234 "psubw "#E", %%mm6 \n\t"\
235 "paddw %4, %%mm6 \n\t"\
236 "psraw $3, %%mm6 \n\t"\
237 "packuswb %%mm6, %%mm6 \n\t"\
238 OP(%%mm6, (%1), A, d) \
241 /* vertical filter [ 0 -7 42 96 -2 -1] */
242 #define QPEL_CAVSV3(A,B,C,D,E,F,OP,MUL2) \
243 "movd (%0), "#F" \n\t"\
244 "movq "#C", %%mm6 \n\t"\
245 "pmullw "MANGLE(MUL2)", %%mm6\n\t"\
246 "movq "#D", %%mm7 \n\t"\
247 "pmullw %5, %%mm7 \n\t"\
248 "psllw $3, "#B" \n\t"\
249 "psubw "#B", %%mm6 \n\t"\
250 "psraw $3, "#B" \n\t"\
251 "paddw %%mm7, %%mm6 \n\t"\
252 "paddw "#B", %%mm6 \n\t"\
253 "paddw "#E", "#E" \n\t"\
254 "pxor %%mm7, %%mm7 \n\t"\
256 "punpcklbw %%mm7, "#F" \n\t"\
257 "psubw "#E", %%mm6 \n\t"\
258 "psraw $1, "#E" \n\t"\
259 "psubw "#F", %%mm6 \n\t"\
260 "paddw %4, %%mm6 \n\t"\
261 "psraw $7, %%mm6 \n\t"\
262 "packuswb %%mm6, %%mm6 \n\t"\
263 OP(%%mm6, (%1), A, d) \
267 #define QPEL_CAVSVNUM(VOP,OP,ADD,MUL1,MUL2)\
273 "pxor %%mm7, %%mm7 \n\t"\
274 "movd (%0), %%mm0 \n\t"\
276 "movd (%0), %%mm1 \n\t"\
278 "movd (%0), %%mm2 \n\t"\
280 "movd (%0), %%mm3 \n\t"\
282 "movd (%0), %%mm4 \n\t"\
284 "punpcklbw %%mm7, %%mm0 \n\t"\
285 "punpcklbw %%mm7, %%mm1 \n\t"\
286 "punpcklbw %%mm7, %%mm2 \n\t"\
287 "punpcklbw %%mm7, %%mm3 \n\t"\
288 "punpcklbw %%mm7, %%mm4 \n\t"\
289 VOP(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP, MUL2)\
290 VOP(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP, MUL2)\
291 VOP(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP, MUL2)\
292 VOP(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP, MUL2)\
293 VOP(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP, MUL2)\
294 VOP(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP, MUL2)\
295 VOP(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP, MUL2)\
296 VOP(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP, MUL2)\
298 : "+a"(src), "+c"(dst)\
299 : "S"((x86_reg)srcStride), "r"((x86_reg)dstStride), "m"(ADD), "m"(MUL1)\
304 VOP(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP, MUL2)\
305 VOP(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP, MUL2)\
306 VOP(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP, MUL2)\
307 VOP(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP, MUL2)\
308 VOP(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP, MUL2)\
309 VOP(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP, MUL2)\
310 VOP(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP, MUL2)\
311 VOP(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP, MUL2)\
313 : "+a"(src), "+c"(dst)\
314 : "S"((x86_reg)srcStride), "r"((x86_reg)dstStride), "m"(ADD), "m"(MUL1)\
318 src += 4-(h+5)*srcStride;\
319 dst += 4-h*dstStride;\
322 #define QPEL_CAVS(OPNAME, OP, MMX)\
323 static void OPNAME ## cavs_qpel8_h_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
326 "pxor %%mm7, %%mm7 \n\t"\
327 "movq %5, %%mm6 \n\t"\
329 "movq (%0), %%mm0 \n\t"\
330 "movq 1(%0), %%mm2 \n\t"\
331 "movq %%mm0, %%mm1 \n\t"\
332 "movq %%mm2, %%mm3 \n\t"\
333 "punpcklbw %%mm7, %%mm0 \n\t"\
334 "punpckhbw %%mm7, %%mm1 \n\t"\
335 "punpcklbw %%mm7, %%mm2 \n\t"\
336 "punpckhbw %%mm7, %%mm3 \n\t"\
337 "paddw %%mm2, %%mm0 \n\t"\
338 "paddw %%mm3, %%mm1 \n\t"\
339 "pmullw %%mm6, %%mm0 \n\t"\
340 "pmullw %%mm6, %%mm1 \n\t"\
341 "movq -1(%0), %%mm2 \n\t"\
342 "movq 2(%0), %%mm4 \n\t"\
343 "movq %%mm2, %%mm3 \n\t"\
344 "movq %%mm4, %%mm5 \n\t"\
345 "punpcklbw %%mm7, %%mm2 \n\t"\
346 "punpckhbw %%mm7, %%mm3 \n\t"\
347 "punpcklbw %%mm7, %%mm4 \n\t"\
348 "punpckhbw %%mm7, %%mm5 \n\t"\
349 "paddw %%mm4, %%mm2 \n\t"\
350 "paddw %%mm3, %%mm5 \n\t"\
351 "psubw %%mm2, %%mm0 \n\t"\
352 "psubw %%mm5, %%mm1 \n\t"\
353 "movq %6, %%mm5 \n\t"\
354 "paddw %%mm5, %%mm0 \n\t"\
355 "paddw %%mm5, %%mm1 \n\t"\
356 "psraw $3, %%mm0 \n\t"\
357 "psraw $3, %%mm1 \n\t"\
358 "packuswb %%mm1, %%mm0 \n\t"\
359 OP(%%mm0, (%1),%%mm5, q) \
364 : "+a"(src), "+c"(dst), "+m"(h)\
365 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_4)\
370 static inline void OPNAME ## cavs_qpel8or16_v1_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
371 QPEL_CAVSVNUM(QPEL_CAVSV1,OP,ff_pw_64,ff_pw_96,ff_pw_42) \
374 static inline void OPNAME ## cavs_qpel8or16_v2_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
375 QPEL_CAVSVNUM(QPEL_CAVSV2,OP,ff_pw_4,ff_pw_5,ff_pw_5) \
378 static inline void OPNAME ## cavs_qpel8or16_v3_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
379 QPEL_CAVSVNUM(QPEL_CAVSV3,OP,ff_pw_64,ff_pw_96,ff_pw_42) \
382 static void OPNAME ## cavs_qpel8_v1_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
383 OPNAME ## cavs_qpel8or16_v1_ ## MMX(dst , src , dstStride, srcStride, 8);\
385 static void OPNAME ## cavs_qpel16_v1_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
386 OPNAME ## cavs_qpel8or16_v1_ ## MMX(dst , src , dstStride, srcStride, 16);\
387 OPNAME ## cavs_qpel8or16_v1_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
390 static void OPNAME ## cavs_qpel8_v2_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
391 OPNAME ## cavs_qpel8or16_v2_ ## MMX(dst , src , dstStride, srcStride, 8);\
393 static void OPNAME ## cavs_qpel16_v2_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
394 OPNAME ## cavs_qpel8or16_v2_ ## MMX(dst , src , dstStride, srcStride, 16);\
395 OPNAME ## cavs_qpel8or16_v2_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
398 static void OPNAME ## cavs_qpel8_v3_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
399 OPNAME ## cavs_qpel8or16_v3_ ## MMX(dst , src , dstStride, srcStride, 8);\
401 static void OPNAME ## cavs_qpel16_v3_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
402 OPNAME ## cavs_qpel8or16_v3_ ## MMX(dst , src , dstStride, srcStride, 16);\
403 OPNAME ## cavs_qpel8or16_v3_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
406 static void OPNAME ## cavs_qpel16_h_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
407 OPNAME ## cavs_qpel8_h_ ## MMX(dst , src , dstStride, srcStride);\
408 OPNAME ## cavs_qpel8_h_ ## MMX(dst+8, src+8, dstStride, srcStride);\
411 OPNAME ## cavs_qpel8_h_ ## MMX(dst , src , dstStride, srcStride);\
412 OPNAME ## cavs_qpel8_h_ ## MMX(dst+8, src+8, dstStride, srcStride);\
415 #define CAVS_MC(OPNAME, SIZE, MMX) \
416 static void OPNAME ## cavs_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, ptrdiff_t stride)\
418 OPNAME ## cavs_qpel ## SIZE ## _h_ ## MMX(dst, src, stride, stride);\
421 static void OPNAME ## cavs_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, ptrdiff_t stride)\
423 OPNAME ## cavs_qpel ## SIZE ## _v1_ ## MMX(dst, src, stride, stride);\
426 static void OPNAME ## cavs_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, ptrdiff_t stride)\
428 OPNAME ## cavs_qpel ## SIZE ## _v2_ ## MMX(dst, src, stride, stride);\
431 static void OPNAME ## cavs_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, ptrdiff_t stride)\
433 OPNAME ## cavs_qpel ## SIZE ## _v3_ ## MMX(dst, src, stride, stride);\
436 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t"
437 #define AVG_3DNOW_OP(a,b,temp, size) \
438 "mov" #size " " #b ", " #temp " \n\t"\
439 "pavgusb " #temp ", " #a " \n\t"\
440 "mov" #size " " #a ", " #b " \n\t"
441 #define AVG_MMXEXT_OP(a, b, temp, size) \
442 "mov" #size " " #b ", " #temp " \n\t"\
443 "pavgb " #temp ", " #a " \n\t"\
444 "mov" #size " " #a ", " #b " \n\t"
446 #endif /* (HAVE_MMXEXT_INLINE || HAVE_AMD3DNOW_INLINE) */
449 static av_cold void cavsdsp_init_mmx(CAVSDSPContext *c,
450 AVCodecContext *avctx)
452 c->put_cavs_qpel_pixels_tab[0][0] = ff_put_cavs_qpel16_mc00_mmx;
453 c->put_cavs_qpel_pixels_tab[1][0] = ff_put_cavs_qpel8_mc00_mmx;
454 c->avg_cavs_qpel_pixels_tab[0][0] = ff_avg_cavs_qpel16_mc00_mmx;
455 c->avg_cavs_qpel_pixels_tab[1][0] = ff_avg_cavs_qpel8_mc00_mmx;
457 c->cavs_idct8_add = cavs_idct8_add_mmx;
458 c->idct_perm = FF_TRANSPOSE_IDCT_PERM;
460 #endif /* HAVE_MMX_INLINE */
462 #define DSPFUNC(PFX, IDX, NUM, EXT) \
463 c->PFX ## _cavs_qpel_pixels_tab[IDX][ 2] = PFX ## _cavs_qpel ## NUM ## _mc20_ ## EXT; \
464 c->PFX ## _cavs_qpel_pixels_tab[IDX][ 4] = PFX ## _cavs_qpel ## NUM ## _mc01_ ## EXT; \
465 c->PFX ## _cavs_qpel_pixels_tab[IDX][ 8] = PFX ## _cavs_qpel ## NUM ## _mc02_ ## EXT; \
466 c->PFX ## _cavs_qpel_pixels_tab[IDX][12] = PFX ## _cavs_qpel ## NUM ## _mc03_ ## EXT; \
468 #if HAVE_MMXEXT_INLINE
469 QPEL_CAVS(put_, PUT_OP, mmxext)
470 QPEL_CAVS(avg_, AVG_MMXEXT_OP, mmxext)
472 CAVS_MC(put_, 8, mmxext)
473 CAVS_MC(put_, 16, mmxext)
474 CAVS_MC(avg_, 8, mmxext)
475 CAVS_MC(avg_, 16, mmxext)
477 static av_cold void cavsdsp_init_mmxext(CAVSDSPContext *c,
478 AVCodecContext *avctx)
480 DSPFUNC(put, 0, 16, mmxext);
481 DSPFUNC(put, 1, 8, mmxext);
482 DSPFUNC(avg, 0, 16, mmxext);
483 DSPFUNC(avg, 1, 8, mmxext);
485 #endif /* HAVE_MMXEXT_INLINE */
487 #if HAVE_AMD3DNOW_INLINE
488 QPEL_CAVS(put_, PUT_OP, 3dnow)
489 QPEL_CAVS(avg_, AVG_3DNOW_OP, 3dnow)
491 CAVS_MC(put_, 8, 3dnow)
492 CAVS_MC(put_, 16,3dnow)
493 CAVS_MC(avg_, 8, 3dnow)
494 CAVS_MC(avg_, 16,3dnow)
496 static av_cold void cavsdsp_init_3dnow(CAVSDSPContext *c,
497 AVCodecContext *avctx)
499 DSPFUNC(put, 0, 16, 3dnow);
500 DSPFUNC(put, 1, 8, 3dnow);
501 DSPFUNC(avg, 0, 16, 3dnow);
502 DSPFUNC(avg, 1, 8, 3dnow);
504 #endif /* HAVE_AMD3DNOW_INLINE */
506 av_cold void ff_cavsdsp_init_x86(CAVSDSPContext *c, AVCodecContext *avctx)
509 int mm_flags = av_get_cpu_flags();
511 if (mm_flags & AV_CPU_FLAG_MMX)
512 cavsdsp_init_mmx(c, avctx);
513 #endif /* HAVE_MMX_INLINE */
514 #if HAVE_MMXEXT_INLINE
515 if (mm_flags & AV_CPU_FLAG_MMXEXT)
516 cavsdsp_init_mmxext(c, avctx);
517 #endif /* HAVE_MMXEXT_INLINE */
518 #if HAVE_AMD3DNOW_INLINE
519 if (mm_flags & AV_CPU_FLAG_3DNOW)
520 cavsdsp_init_3dnow(c, avctx);
521 #endif /* HAVE_AMD3DNOW_INLINE */