Mercurial > libavcodec.hg

comparison i386/dsputil_h264_template_mmx.c @ 2922:d772011258ec libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx. 2-4% overall speedup.
author lorenm
date Thu, 27 Oct 2005 06:45:29 +0000
parents a49f140179e9
children 0b546eab515d
comparison
equal deleted inserted replaced
2921:d22a3556292a 2922:d772011258ec
1 /* 1 /*
2 * Copyright (c) 2005 Zoltan Hidvegi <hzoli -a- hzoli -d- com> 2 * Copyright (c) 2005 Zoltan Hidvegi <hzoli -a- hzoli -d- com>,
3 * Loren Merritt
3 * 4 *
4 * This library is free software; you can redistribute it and/or 5 * This library is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public 6 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either 7 * License as published by the Free Software Foundation; either
7 * version 2 of the License, or (at your option) any later version. 8 * version 2 of the License, or (at your option) any later version.
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 */ 18 */
18 19
19 /** 20 /**
20 * MMX optimized version of (put|avg)_h264_chroma_mc8. 21 * MMX optimized version of (put|avg)_h264_chroma_mc8.
21 * H264_CHROMA_MC8_TMPL must be defined to the desired function name and 22 * H264_CHROMA_MC8_TMPL must be defined to the desired function name
22 * H264_CHROMA_OP must be defined to empty for put and pavgb/pavgusb for avg. 23 * H264_CHROMA_OP must be defined to empty for put and pavgb/pavgusb for avg
24 * H264_CHROMA_MC8_MV0 must be defined to a (put|avg)_pixels8 function
23 */ 25 */
24 static void H264_CHROMA_MC8_TMPL(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) 26 static void H264_CHROMA_MC8_TMPL(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
25 { 27 {
26 uint64_t AA __align8; 28 uint64_t AA __align8;
27 uint64_t DD __align8; 29 uint64_t DD __align8;
28 unsigned long srcos = (long)src & 7;
29 uint64_t sh1 __align8 = srcos * 8;
30 uint64_t sh2 __align8 = 56 - sh1;
31 int i; 30 int i;
32 31
32 if(y==0 && x==0) {
33 /* no filter needed */
34 H264_CHROMA_MC8_MV0(dst, src, stride, h);
35 return;
36 }
37
33 assert(x<8 && y<8 && x>=0 && y>=0); 38 assert(x<8 && y<8 && x>=0 && y>=0);
34 39
35 asm volatile("movd %1, %%mm4\n\t" 40 if(y==0)
36 "movd %2, %%mm6\n\t" 41 {
42 /* horizontal filter only */
43 asm volatile("movd %0, %%mm5\n\t"
44 "punpcklwd %%mm5, %%mm5\n\t"
45 "punpckldq %%mm5, %%mm5\n\t" /* mm5 = B = x */
46 "movq %1, %%mm4\n\t"
47 "pxor %%mm7, %%mm7\n\t"
48 "psubw %%mm5, %%mm4\n\t" /* mm4 = A = 8-x */
49 : : "rm" (x), "m" (ff_pw_8));
50
51 for(i=0; i<h; i++) {
52 asm volatile(
53 /* mm0 = src[0..7], mm1 = src[1..8] */
54 "movq %0, %%mm0\n\t"
55 "movq %1, %%mm1\n\t"
56 : : "m" (src[0]), "m" (src[1]));
57
58 asm volatile(
59 /* [mm2,mm3] = A * src[0..7] */
60 "movq %%mm0, %%mm2\n\t"
61 "punpcklbw %%mm7, %%mm2\n\t"
62 "pmullw %%mm4, %%mm2\n\t"
63 "movq %%mm0, %%mm3\n\t"
64 "punpckhbw %%mm7, %%mm3\n\t"
65 "pmullw %%mm4, %%mm3\n\t"
66
67 /* [mm2,mm3] += B * src[1..8] */
68 "movq %%mm1, %%mm0\n\t"
69 "punpcklbw %%mm7, %%mm0\n\t"
70 "pmullw %%mm5, %%mm0\n\t"
71 "punpckhbw %%mm7, %%mm1\n\t"
72 "pmullw %%mm5, %%mm1\n\t"
73 "paddw %%mm0, %%mm2\n\t"
74 "paddw %%mm1, %%mm3\n\t"
75
76 /* dst[0..7] = pack(([mm2,mm3] + 32) >> 6) */
77 "paddw %1, %%mm2\n\t"
78 "paddw %1, %%mm3\n\t"
79 "psrlw $3, %%mm2\n\t"
80 "psrlw $3, %%mm3\n\t"
81 "packuswb %%mm3, %%mm2\n\t"
82 H264_CHROMA_OP(%0, %%mm2)
83 "movq %%mm2, %0\n\t"
84 : "=m" (dst[0]) : "m" (ff_pw_4));
85
86 src += stride;
87 dst += stride;
88 }
89 return;
90 }
91
92 if(x==0)
93 {
94 /* vertical filter only */
95 asm volatile("movd %0, %%mm6\n\t"
96 "punpcklwd %%mm6, %%mm6\n\t"
97 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = C = y */
98 "movq %1, %%mm4\n\t"
99 "pxor %%mm7, %%mm7\n\t"
100 "psubw %%mm6, %%mm4\n\t" /* mm4 = A = 8-y */
101 : : "rm" (y), "m" (ff_pw_8));
102
103 asm volatile(
104 /* mm0 = src[0..7] */
105 "movq %0, %%mm0\n\t"
106 : : "m" (src[0]));
107
108 for(i=0; i<h; i++) {
109 asm volatile(
110 /* [mm2,mm3] = A * src[0..7] */
111 "movq %mm0, %mm2\n\t"
112 "punpcklbw %mm7, %mm2\n\t"
113 "pmullw %mm4, %mm2\n\t"
114 "movq %mm0, %mm3\n\t"
115 "punpckhbw %mm7, %mm3\n\t"
116 "pmullw %mm4, %mm3\n\t");
117
118 src += stride;
119 asm volatile(
120 /* mm0 = src[0..7] */
121 "movq %0, %%mm0\n\t"
122 : : "m" (src[0]));
123
124 asm volatile(
125 /* [mm2,mm3] += C * src[0..7] */
126 "movq %mm0, %mm1\n\t"
127 "punpcklbw %mm7, %mm1\n\t"
128 "pmullw %mm6, %mm1\n\t"
129 "paddw %mm1, %mm2\n\t"
130 "movq %mm0, %mm5\n\t"
131 "punpckhbw %mm7, %mm5\n\t"
132 "pmullw %mm6, %mm5\n\t"
133 "paddw %mm5, %mm3\n\t");
134
135 asm volatile(
136 /* dst[0..7] = pack(([mm2,mm3] + 32) >> 6) */
137 "paddw %1, %%mm2\n\t"
138 "paddw %1, %%mm3\n\t"
139 "psrlw $3, %%mm2\n\t"
140 "psrlw $3, %%mm3\n\t"
141 "packuswb %%mm3, %%mm2\n\t"
142 H264_CHROMA_OP(%0, %%mm2)
143 "movq %%mm2, %0\n\t"
144 : "=m" (dst[0]) : "m" (ff_pw_4));
145
146 dst += stride;
147 }
148 return;
149 }
150
151 /* general case, bilinear */
152 asm volatile("movd %2, %%mm4\n\t"
153 "movd %3, %%mm6\n\t"
37 "punpcklwd %%mm4, %%mm4\n\t" 154 "punpcklwd %%mm4, %%mm4\n\t"
38 "punpcklwd %%mm6, %%mm6\n\t" 155 "punpcklwd %%mm6, %%mm6\n\t"
39 "punpckldq %%mm4, %%mm4\n\t" /* mm4 = x words */ 156 "punpckldq %%mm4, %%mm4\n\t" /* mm4 = x words */
40 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = y words */ 157 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = y words */
41 "movq %%mm4, %%mm5\n\t" 158 "movq %%mm4, %%mm5\n\t"
42 "pmullw %%mm6, %%mm4\n\t" /* mm4 = x * y */ 159 "pmullw %%mm6, %%mm4\n\t" /* mm4 = x * y */
43 "psllw $3, %%mm5\n\t" 160 "psllw $3, %%mm5\n\t"
44 "psllw $3, %%mm6\n\t" 161 "psllw $3, %%mm6\n\t"
45 "movq %%mm5, %%mm7\n\t" 162 "movq %%mm5, %%mm7\n\t"
46 "paddw %%mm6, %%mm7\n\t" 163 "paddw %%mm6, %%mm7\n\t"
47 "movq %%mm4, %0\n\t" /* DD = x * y */ 164 "movq %%mm4, %1\n\t" /* DD = x * y */
48 "psubw %%mm4, %%mm5\n\t" /* mm5 = B = 8x - xy */ 165 "psubw %%mm4, %%mm5\n\t" /* mm5 = B = 8x - xy */
49 "psubw %%mm4, %%mm6\n\t" /* mm6 = C = 8y - xy */ 166 "psubw %%mm4, %%mm6\n\t" /* mm6 = C = 8y - xy */
50 "paddw %3, %%mm4\n\t" 167 "paddw %4, %%mm4\n\t"
51 "psubw %%mm7, %%mm4\n\t" /* mm4 = A = xy - (8x+8y) + 64 */ 168 "psubw %%mm7, %%mm4\n\t" /* mm4 = A = xy - (8x+8y) + 64 */
52 "pxor %%mm7, %%mm7\n\t" 169 "pxor %%mm7, %%mm7\n\t"
53 : "=m" (DD) : "rm" (x), "rm" (y), "m" (ff_pw_64)); 170 "movq %%mm4, %0\n\t"
54 171 : "=m" (AA), "=m" (DD) : "rm" (x), "rm" (y), "m" (ff_pw_64));
55 asm volatile("movq %%mm4, %0" : "=m" (AA)); 172
56
57 src -= srcos;
58 asm volatile( 173 asm volatile(
59 /* mm0 = src[0..7], mm1 = src[1..8] */ 174 /* mm0 = src[0..7], mm1 = src[1..8] */
60 "movq %0, %%mm1\n\t" 175 "movq %0, %%mm0\n\t"
61 "movq %1, %%mm0\n\t" 176 "movq %1, %%mm1\n\t"
62 "psrlq %2, %%mm1\n\t" 177 : : "m" (src[0]), "m" (src[1]));
63 "psllq %3, %%mm0\n\t"
64 "movq %%mm0, %%mm4\n\t"
65 "psllq $8, %%mm0\n\t"
66 "por %%mm1, %%mm0\n\t"
67 "psrlq $8, %%mm1\n\t"
68 "por %%mm4, %%mm1\n\t"
69 : : "m" (src[0]), "m" (src[8]), "m" (sh1), "m" (sh2));
70 178
71 for(i=0; i<h; i++) { 179 for(i=0; i<h; i++) {
72 asm volatile( 180 asm volatile(
73 /* [mm2,mm3] = A * src[0..7] */ 181 /* [mm2,mm3] = A * src[0..7] */
74 "movq %%mm0, %%mm2\n\t" 182 "movq %%mm0, %%mm2\n\t"
89 : : "m" (AA)); 197 : : "m" (AA));
90 198
91 src += stride; 199 src += stride;
92 asm volatile( 200 asm volatile(
93 /* mm0 = src[0..7], mm1 = src[1..8] */ 201 /* mm0 = src[0..7], mm1 = src[1..8] */
94 "movq %0, %%mm1\n\t" 202 "movq %0, %%mm0\n\t"
95 "movq %1, %%mm0\n\t" 203 "movq %1, %%mm1\n\t"
96 "psrlq %2, %%mm1\n\t" 204 : : "m" (src[0]), "m" (src[1]));
97 "psllq %3, %%mm0\n\t"
98 "movq %%mm0, %%mm4\n\t"
99 "psllq $8, %%mm0\n\t"
100 "por %%mm1, %%mm0\n\t"
101 "psrlq $8, %%mm1\n\t"
102 "por %%mm4, %%mm1\n\t"
103 : : "m" (src[0]), "m" (src[8]), "m" (sh1), "m" (sh2));
104 205
105 asm volatile( 206 asm volatile(
106 /* [mm2,mm3] += C * src[0..7] */ 207 /* [mm2,mm3] += C * src[0..7] */
107 "movq %mm0, %mm4\n\t" 208 "movq %mm0, %mm4\n\t"
108 "punpcklbw %mm7, %mm4\n\t" 209 "punpcklbw %mm7, %mm4\n\t"
136 "movq %%mm2, %0\n\t" 237 "movq %%mm2, %0\n\t"
137 : "=m" (dst[0]) : "m" (ff_pw_32)); 238 : "=m" (dst[0]) : "m" (ff_pw_32));
138 dst+= stride; 239 dst+= stride;
139 } 240 }
140 } 241 }
242
243 static void H264_CHROMA_MC4_TMPL(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y)
244 {
245 uint64_t AA __align8;
246 uint64_t DD __align8;
247 int i;
248
249 /* no special case for mv=(0,0) in 4x*, since it's much less common than in 8x*.
250 * could still save a few cycles, but maybe not worth the complexity. */
251
252 assert(x<8 && y<8 && x>=0 && y>=0);
253
254 asm volatile("movd %2, %%mm4\n\t"
255 "movd %3, %%mm6\n\t"
256 "punpcklwd %%mm4, %%mm4\n\t"
257 "punpcklwd %%mm6, %%mm6\n\t"
258 "punpckldq %%mm4, %%mm4\n\t" /* mm4 = x words */
259 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = y words */
260 "movq %%mm4, %%mm5\n\t"
261 "pmullw %%mm6, %%mm4\n\t" /* mm4 = x * y */
262 "psllw $3, %%mm5\n\t"
263 "psllw $3, %%mm6\n\t"
264 "movq %%mm5, %%mm7\n\t"
265 "paddw %%mm6, %%mm7\n\t"
266 "movq %%mm4, %1\n\t" /* DD = x * y */
267 "psubw %%mm4, %%mm5\n\t" /* mm5 = B = 8x - xy */
268 "psubw %%mm4, %%mm6\n\t" /* mm6 = C = 8y - xy */
269 "paddw %4, %%mm4\n\t"
270 "psubw %%mm7, %%mm4\n\t" /* mm4 = A = xy - (8x+8y) + 64 */
271 "pxor %%mm7, %%mm7\n\t"
272 "movq %%mm4, %0\n\t"
273 : "=m" (AA), "=m" (DD) : "rm" (x), "rm" (y), "m" (ff_pw_64));
274
275 asm volatile(
276 /* mm0 = src[0..3], mm1 = src[1..4] */
277 "movd %0, %%mm0\n\t"
278 "movd %1, %%mm1\n\t"
279 "punpcklbw %%mm7, %%mm0\n\t"
280 "punpcklbw %%mm7, %%mm1\n\t"
281 : : "m" (src[0]), "m" (src[1]));
282
283 for(i=0; i<h; i++) {
284 asm volatile(
285 /* mm2 = A * src[0..3] + B * src[1..4] */
286 "movq %%mm0, %%mm2\n\t"
287 "pmullw %0, %%mm2\n\t"
288 "pmullw %%mm5, %%mm1\n\t"
289 "paddw %%mm1, %%mm2\n\t"
290 : : "m" (AA));
291
292 src += stride;
293 asm volatile(
294 /* mm0 = src[0..3], mm1 = src[1..4] */
295 "movd %0, %%mm0\n\t"
296 "movd %1, %%mm1\n\t"
297 "punpcklbw %%mm7, %%mm0\n\t"
298 "punpcklbw %%mm7, %%mm1\n\t"
299 : : "m" (src[0]), "m" (src[1]));
300
301 asm volatile(
302 /* mm2 += C * src[0..3] + D * src[1..4] */
303 "movq %%mm0, %%mm3\n\t"
304 "movq %%mm1, %%mm4\n\t"
305 "pmullw %%mm6, %%mm3\n\t"
306 "pmullw %0, %%mm4\n\t"
307 "paddw %%mm3, %%mm2\n\t"
308 "paddw %%mm4, %%mm2\n\t"
309 : : "m" (DD));
310
311 asm volatile(
312 /* dst[0..3] = pack((mm2 + 32) >> 6) */
313 "paddw %1, %%mm2\n\t"
314 "psrlw $6, %%mm2\n\t"
315 "packuswb %%mm7, %%mm2\n\t"
316 H264_CHROMA_OP4(%0, %%mm2, %%mm3)
317 "movd %%mm2, %0\n\t"
318 : "=m" (dst[0]) : "m" (ff_pw_32));
319 dst += stride;
320 }
321 }