2 * Copyright (c) 2005 Zoltan Hidvegi <hzoli -a- hzoli -d- com>,
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * MMX optimized version of (put|avg)_h264_chroma_mc8.
24 * H264_CHROMA_MC8_TMPL must be defined to the desired function name
25 * H264_CHROMA_OP must be defined to empty for put and pavgb/pavgusb for avg
26 * H264_CHROMA_MC8_MV0 must be defined to a (put|avg)_pixels8 function
28 static void H264_CHROMA_MC8_TMPL(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
30 DECLARE_ALIGNED_8(uint64_t, AA);
31 DECLARE_ALIGNED_8(uint64_t, DD);
35 /* no filter needed */
36 H264_CHROMA_MC8_MV0(dst, src, stride, h);
40 assert(x<8 && y<8 && x>=0 && y>=0);
44 /* 1 dimensional filter only */
45 const int dxy = x ? 1 : stride;
50 "punpcklwd %%mm5, %%mm5\n\t"
51 "punpckldq %%mm5, %%mm5\n\t" /* mm5 = B = x */
52 "movq %%mm4, %%mm6\n\t"
53 "pxor %%mm7, %%mm7\n\t"
54 "psubw %%mm5, %%mm4\n\t" /* mm4 = A = 8-x */
55 "psrlw $1, %%mm6\n\t" /* mm6 = 4 */
56 :: "rm"(x+y), "m"(ff_pw_8));
60 /* mm0 = src[0..7], mm1 = src[1..8] */
63 :: "m"(src[0]), "m"(src[dxy]));
66 /* [mm0,mm1] = A * src[0..7] */
67 /* [mm2,mm3] = B * src[1..8] */
68 "movq %%mm0, %%mm1\n\t"
69 "movq %%mm2, %%mm3\n\t"
70 "punpcklbw %%mm7, %%mm0\n\t"
71 "punpckhbw %%mm7, %%mm1\n\t"
72 "punpcklbw %%mm7, %%mm2\n\t"
73 "punpckhbw %%mm7, %%mm3\n\t"
74 "pmullw %%mm4, %%mm0\n\t"
75 "pmullw %%mm4, %%mm1\n\t"
76 "pmullw %%mm5, %%mm2\n\t"
77 "pmullw %%mm5, %%mm3\n\t"
79 /* dst[0..7] = (A * src[0..7] + B * src[1..8] + 4) >> 3 */
80 "paddw %%mm6, %%mm0\n\t"
81 "paddw %%mm6, %%mm1\n\t"
82 "paddw %%mm2, %%mm0\n\t"
83 "paddw %%mm3, %%mm1\n\t"
86 "packuswb %%mm1, %%mm0\n\t"
87 H264_CHROMA_OP(%0, %%mm0)
97 /* general case, bilinear */
98 asm volatile("movd %2, %%mm4\n\t"
100 "punpcklwd %%mm4, %%mm4\n\t"
101 "punpcklwd %%mm6, %%mm6\n\t"
102 "punpckldq %%mm4, %%mm4\n\t" /* mm4 = x words */
103 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = y words */
104 "movq %%mm4, %%mm5\n\t"
105 "pmullw %%mm6, %%mm4\n\t" /* mm4 = x * y */
106 "psllw $3, %%mm5\n\t"
107 "psllw $3, %%mm6\n\t"
108 "movq %%mm5, %%mm7\n\t"
109 "paddw %%mm6, %%mm7\n\t"
110 "movq %%mm4, %1\n\t" /* DD = x * y */
111 "psubw %%mm4, %%mm5\n\t" /* mm5 = B = 8x - xy */
112 "psubw %%mm4, %%mm6\n\t" /* mm6 = C = 8y - xy */
113 "paddw %4, %%mm4\n\t"
114 "psubw %%mm7, %%mm4\n\t" /* mm4 = A = xy - (8x+8y) + 64 */
115 "pxor %%mm7, %%mm7\n\t"
117 : "=m" (AA), "=m" (DD) : "rm" (x), "rm" (y), "m" (ff_pw_64));
120 /* mm0 = src[0..7], mm1 = src[1..8] */
123 : : "m" (src[0]), "m" (src[1]));
129 /* mm2 = A * src[0..3] + B * src[1..4] */
130 /* mm3 = A * src[4..7] + B * src[5..8] */
131 "movq %%mm0, %%mm2\n\t"
132 "movq %%mm1, %%mm3\n\t"
133 "punpckhbw %%mm7, %%mm0\n\t"
134 "punpcklbw %%mm7, %%mm1\n\t"
135 "punpcklbw %%mm7, %%mm2\n\t"
136 "punpckhbw %%mm7, %%mm3\n\t"
137 "pmullw %0, %%mm0\n\t"
138 "pmullw %0, %%mm2\n\t"
139 "pmullw %%mm5, %%mm1\n\t"
140 "pmullw %%mm5, %%mm3\n\t"
141 "paddw %%mm1, %%mm2\n\t"
142 "paddw %%mm0, %%mm3\n\t"
146 /* [mm2,mm3] += C * src[0..7] */
148 "movq %%mm0, %%mm1\n\t"
149 "punpcklbw %%mm7, %%mm0\n\t"
150 "punpckhbw %%mm7, %%mm1\n\t"
151 "pmullw %%mm6, %%mm0\n\t"
152 "pmullw %%mm6, %%mm1\n\t"
153 "paddw %%mm0, %%mm2\n\t"
154 "paddw %%mm1, %%mm3\n\t"
158 /* [mm2,mm3] += D * src[1..8] */
160 "movq %%mm1, %%mm0\n\t"
161 "movq %%mm1, %%mm4\n\t"
162 "punpcklbw %%mm7, %%mm0\n\t"
163 "punpckhbw %%mm7, %%mm4\n\t"
164 "pmullw %2, %%mm0\n\t"
165 "pmullw %2, %%mm4\n\t"
166 "paddw %%mm0, %%mm2\n\t"
167 "paddw %%mm4, %%mm3\n\t"
169 : : "m" (src[0]), "m" (src[1]), "m" (DD));
172 /* dst[0..7] = ([mm2,mm3] + 32) >> 6 */
173 "paddw %1, %%mm2\n\t"
174 "paddw %1, %%mm3\n\t"
175 "psrlw $6, %%mm2\n\t"
176 "psrlw $6, %%mm3\n\t"
177 "packuswb %%mm3, %%mm2\n\t"
178 H264_CHROMA_OP(%0, %%mm2)
180 : "=m" (dst[0]) : "m" (ff_pw_32));
185 static void H264_CHROMA_MC4_TMPL(uint8_t *dst/*align 4*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
188 "pxor %%mm7, %%mm7 \n\t"
189 "movd %5, %%mm2 \n\t"
190 "movd %6, %%mm3 \n\t"
191 "movq "MANGLE(ff_pw_8)", %%mm4\n\t"
192 "movq "MANGLE(ff_pw_8)", %%mm5\n\t"
193 "punpcklwd %%mm2, %%mm2 \n\t"
194 "punpcklwd %%mm3, %%mm3 \n\t"
195 "punpcklwd %%mm2, %%mm2 \n\t"
196 "punpcklwd %%mm3, %%mm3 \n\t"
197 "psubw %%mm2, %%mm4 \n\t"
198 "psubw %%mm3, %%mm5 \n\t"
200 "movd (%1), %%mm0 \n\t"
201 "movd 1(%1), %%mm6 \n\t"
203 "punpcklbw %%mm7, %%mm0 \n\t"
204 "punpcklbw %%mm7, %%mm6 \n\t"
205 "pmullw %%mm4, %%mm0 \n\t"
206 "pmullw %%mm2, %%mm6 \n\t"
207 "paddw %%mm0, %%mm6 \n\t"
210 "movd (%1), %%mm0 \n\t"
211 "movd 1(%1), %%mm1 \n\t"
213 "punpcklbw %%mm7, %%mm0 \n\t"
214 "punpcklbw %%mm7, %%mm1 \n\t"
215 "pmullw %%mm4, %%mm0 \n\t"
216 "pmullw %%mm2, %%mm1 \n\t"
217 "paddw %%mm0, %%mm1 \n\t"
218 "movq %%mm1, %%mm0 \n\t"
219 "pmullw %%mm5, %%mm6 \n\t"
220 "pmullw %%mm3, %%mm1 \n\t"
221 "paddw %4, %%mm6 \n\t"
222 "paddw %%mm6, %%mm1 \n\t"
223 "psrlw $6, %%mm1 \n\t"
224 "packuswb %%mm1, %%mm1 \n\t"
225 H264_CHROMA_OP4((%0), %%mm1, %%mm6)
226 "movd %%mm1, (%0) \n\t"
228 "movd (%1), %%mm6 \n\t"
229 "movd 1(%1), %%mm1 \n\t"
231 "punpcklbw %%mm7, %%mm6 \n\t"
232 "punpcklbw %%mm7, %%mm1 \n\t"
233 "pmullw %%mm4, %%mm6 \n\t"
234 "pmullw %%mm2, %%mm1 \n\t"
235 "paddw %%mm6, %%mm1 \n\t"
236 "movq %%mm1, %%mm6 \n\t"
237 "pmullw %%mm5, %%mm0 \n\t"
238 "pmullw %%mm3, %%mm1 \n\t"
239 "paddw %4, %%mm0 \n\t"
240 "paddw %%mm0, %%mm1 \n\t"
241 "psrlw $6, %%mm1 \n\t"
242 "packuswb %%mm1, %%mm1 \n\t"
243 H264_CHROMA_OP4((%0), %%mm1, %%mm0)
244 "movd %%mm1, (%0) \n\t"
248 : "+r"(dst), "+r"(src), "+r"(h)
249 : "r"((long)stride), "m"(ff_pw_32), "m"(x), "m"(y)
253 #ifdef H264_CHROMA_MC2_TMPL
254 static void H264_CHROMA_MC2_TMPL(uint8_t *dst/*align 2*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
256 int tmp = ((1<<16)-1)*x + 8;
258 int AB= (tmp<<3) - CD;
260 /* mm5 = {A,B,A,B} */
261 /* mm6 = {C,D,C,D} */
264 "punpckldq %%mm5, %%mm5\n\t"
265 "punpckldq %%mm6, %%mm6\n\t"
266 "pxor %%mm7, %%mm7\n\t"
267 /* mm0 = src[0,1,1,2] */
269 "punpcklbw %%mm7, %%mm2\n\t"
270 "pshufw $0x94, %%mm2, %%mm2\n\t"
271 :: "r"(AB), "r"(CD), "m"(src[0]));
277 /* mm1 = A * src[0,1] + B * src[1,2] */
278 "movq %%mm2, %%mm1\n\t"
279 "pmaddwd %%mm5, %%mm1\n\t"
280 /* mm0 = src[0,1,1,2] */
281 "movd (%1), %%mm0\n\t"
282 "punpcklbw %%mm7, %%mm0\n\t"
283 "pshufw $0x94, %%mm0, %%mm0\n\t"
284 /* mm1 += C * src[0,1] + D * src[1,2] */
285 "movq %%mm0, %%mm2\n\t"
286 "pmaddwd %%mm6, %%mm0\n\t"
287 "paddw %3, %%mm1\n\t"
288 "paddw %%mm0, %%mm1\n\t"
289 /* dst[0,1] = pack((mm1 + 32) >> 6) */
290 "psrlw $6, %%mm1\n\t"
291 "packssdw %%mm7, %%mm1\n\t"
292 "packuswb %%mm7, %%mm1\n\t"
293 H264_CHROMA_OP4((%0), %%mm1, %%mm3)
294 "movd %%mm1, %%esi\n\t"
295 "movw %%si, (%0)\n\t"
299 : "+r" (dst), "+r"(src), "+r"(h)
300 : "m" (ff_pw_32), "r"((long)stride)