Mercurial > libavcodec.hg
annotate x86/dsputil_mmx.c @ 12414:3fc4c625b6f3 libavcodec
Remove global mm_flags variable
author | mru |
---|---|
date | Tue, 24 Aug 2010 17:47:05 +0000 |
parents | 09a31ef6ed58 |
children | 9f06475db098 |
rev | line source |
---|---|
8430 | 1 /* |
2 * MMX optimized DSP utils | |
8629
04423b2f6e0b
cosmetics: Remove pointless period after copyright statement non-sentences.
diego
parents:
8596
diff
changeset
|
3 * Copyright (c) 2000, 2001 Fabrice Bellard |
8430 | 4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at> |
5 * | |
6 * This file is part of FFmpeg. | |
7 * | |
8 * FFmpeg is free software; you can redistribute it and/or | |
9 * modify it under the terms of the GNU Lesser General Public | |
10 * License as published by the Free Software Foundation; either | |
11 * version 2.1 of the License, or (at your option) any later version. | |
12 * | |
13 * FFmpeg is distributed in the hope that it will be useful, | |
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 * Lesser General Public License for more details. | |
17 * | |
18 * You should have received a copy of the GNU Lesser General Public | |
19 * License along with FFmpeg; if not, write to the Free Software | |
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 * | |
22 * MMX optimization by Nick Kurshev <nickols_k@mail.ru> | |
23 */ | |
24 | |
25 #include "libavutil/x86_cpu.h" | |
26 #include "libavcodec/dsputil.h" | |
11499 | 27 #include "libavcodec/h264dsp.h" |
8430 | 28 #include "libavcodec/mpegvideo.h" |
29 #include "libavcodec/simple_idct.h" | |
30 #include "dsputil_mmx.h" | |
31 #include "vp3dsp_mmx.h" | |
32 #include "vp3dsp_sse2.h" | |
8817 | 33 #include "vp6dsp_mmx.h" |
8818 | 34 #include "vp6dsp_sse2.h" |
8430 | 35 #include "idct_xvid.h" |
36 | |
37 //#undef NDEBUG | |
38 //#include <assert.h> | |
39 | |
40 /* pixel operations */ | |
11369 | 41 DECLARE_ALIGNED(8, const uint64_t, ff_bone) = 0x0101010101010101ULL; |
42 DECLARE_ALIGNED(8, const uint64_t, ff_wtwo) = 0x0002000200020002ULL; | |
8430 | 43 |
11369 | 44 DECLARE_ALIGNED(16, const uint64_t, ff_pdw_80000000)[2] = |
8430 | 45 {0x8000000080000000ULL, 0x8000000080000000ULL}; |
46 | |
11369 | 47 DECLARE_ALIGNED(8, const uint64_t, ff_pw_3 ) = 0x0003000300030003ULL; |
12143 | 48 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_4 ) = {0x0004000400040004ULL, 0x0004000400040004ULL}; |
11369 | 49 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_5 ) = {0x0005000500050005ULL, 0x0005000500050005ULL}; |
50 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_8 ) = {0x0008000800080008ULL, 0x0008000800080008ULL}; | |
12205
d38e8565ba05
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents:
12168
diff
changeset
|
51 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_9 ) = {0x0009000900090009ULL, 0x0009000900090009ULL}; |
11369 | 52 DECLARE_ALIGNED(8, const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL; |
53 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_16 ) = {0x0010001000100010ULL, 0x0010001000100010ULL}; | |
12205
d38e8565ba05
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents:
12168
diff
changeset
|
54 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_18 ) = {0x0012001200120012ULL, 0x0012001200120012ULL}; |
11369 | 55 DECLARE_ALIGNED(8, const uint64_t, ff_pw_20 ) = 0x0014001400140014ULL; |
12205
d38e8565ba05
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents:
12168
diff
changeset
|
56 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_27 ) = {0x001B001B001B001BULL, 0x001B001B001B001BULL}; |
11369 | 57 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_28 ) = {0x001C001C001C001CULL, 0x001C001C001C001CULL}; |
58 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL}; | |
59 DECLARE_ALIGNED(8, const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL; | |
12206 | 60 DECLARE_ALIGNED(8, const uint64_t, ff_pw_53 ) = 0x0035003500350035ULL; |
12205
d38e8565ba05
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents:
12168
diff
changeset
|
61 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_63 ) = {0x003F003F003F003FULL, 0x003F003F003F003FULL}; |
11369 | 62 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_64 ) = {0x0040004000400040ULL, 0x0040004000400040ULL}; |
63 DECLARE_ALIGNED(8, const uint64_t, ff_pw_96 ) = 0x0060006000600060ULL; | |
64 DECLARE_ALIGNED(8, const uint64_t, ff_pw_128) = 0x0080008000800080ULL; | |
65 DECLARE_ALIGNED(8, const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL; | |
8430 | 66 |
12168
b246b214c2e9
VP8 H/V inner loopfilter MMX/MMXEXT/SSE2 optimizations.
rbultje
parents:
12143
diff
changeset
|
67 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_1 ) = {0x0101010101010101ULL, 0x0101010101010101ULL}; |
11951
afee30fe8c26
16x16 and 8x8c x86 SIMD intra pred functions for VP8 and H.264
darkshikari
parents:
11826
diff
changeset
|
68 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_3 ) = {0x0303030303030303ULL, 0x0303030303030303ULL}; |
12086
d780ae746855
Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents:
11981
diff
changeset
|
69 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_4 ) = {0x0404040404040404ULL, 0x0404040404040404ULL}; |
11369 | 70 DECLARE_ALIGNED(8, const uint64_t, ff_pb_7 ) = 0x0707070707070707ULL; |
71 DECLARE_ALIGNED(8, const uint64_t, ff_pb_1F ) = 0x1F1F1F1F1F1F1F1FULL; | |
72 DECLARE_ALIGNED(8, const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL; | |
12086
d780ae746855
Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents:
11981
diff
changeset
|
73 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_80 ) = {0x8080808080808080ULL, 0x8080808080808080ULL}; |
11369 | 74 DECLARE_ALIGNED(8, const uint64_t, ff_pb_81 ) = 0x8181818181818181ULL; |
75 DECLARE_ALIGNED(8, const uint64_t, ff_pb_A1 ) = 0xA1A1A1A1A1A1A1A1ULL; | |
12086
d780ae746855
Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents:
11981
diff
changeset
|
76 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_F8 ) = {0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL}; |
11369 | 77 DECLARE_ALIGNED(8, const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL; |
12086
d780ae746855
Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).
rbultje
parents:
11981
diff
changeset
|
78 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_FE ) = {0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL}; |
8430 | 79 |
11369 | 80 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 }; |
81 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 }; | |
8430 | 82 |
83 #define JUMPALIGN() __asm__ volatile (ASMALIGN(3)::) | |
84 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%" #regd ", %%" #regd ::) | |
85 | |
86 #define MOVQ_BFE(regd) \ | |
87 __asm__ volatile ( \ | |
88 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ | |
89 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
90 | |
91 #ifndef PIC | |
92 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone)) | |
93 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo)) | |
94 #else | |
95 // for shared library it's better to use this way for accessing constants | |
96 // pcmpeqd -> -1 | |
97 #define MOVQ_BONE(regd) \ | |
98 __asm__ volatile ( \ | |
99 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
100 "psrlw $15, %%" #regd " \n\t" \ | |
101 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
102 | |
103 #define MOVQ_WTWO(regd) \ | |
104 __asm__ volatile ( \ | |
105 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
106 "psrlw $15, %%" #regd " \n\t" \ | |
107 "psllw $1, %%" #regd " \n\t"::) | |
108 | |
109 #endif | |
110 | |
111 // using regr as temporary and for the output result | |
112 // first argument is unmodifed and second is trashed | |
113 // regfe is supposed to contain 0xfefefefefefefefe | |
114 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
115 "movq " #rega ", " #regr " \n\t"\ | |
116 "pand " #regb ", " #regr " \n\t"\ | |
117 "pxor " #rega ", " #regb " \n\t"\ | |
118 "pand " #regfe "," #regb " \n\t"\ | |
119 "psrlq $1, " #regb " \n\t"\ | |
120 "paddb " #regb ", " #regr " \n\t" | |
121 | |
122 #define PAVGB_MMX(rega, regb, regr, regfe) \ | |
123 "movq " #rega ", " #regr " \n\t"\ | |
124 "por " #regb ", " #regr " \n\t"\ | |
125 "pxor " #rega ", " #regb " \n\t"\ | |
126 "pand " #regfe "," #regb " \n\t"\ | |
127 "psrlq $1, " #regb " \n\t"\ | |
128 "psubb " #regb ", " #regr " \n\t" | |
129 | |
130 // mm6 is supposed to contain 0xfefefefefefefefe | |
131 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ | |
132 "movq " #rega ", " #regr " \n\t"\ | |
133 "movq " #regc ", " #regp " \n\t"\ | |
134 "pand " #regb ", " #regr " \n\t"\ | |
135 "pand " #regd ", " #regp " \n\t"\ | |
136 "pxor " #rega ", " #regb " \n\t"\ | |
137 "pxor " #regc ", " #regd " \n\t"\ | |
138 "pand %%mm6, " #regb " \n\t"\ | |
139 "pand %%mm6, " #regd " \n\t"\ | |
140 "psrlq $1, " #regb " \n\t"\ | |
141 "psrlq $1, " #regd " \n\t"\ | |
142 "paddb " #regb ", " #regr " \n\t"\ | |
143 "paddb " #regd ", " #regp " \n\t" | |
144 | |
145 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ | |
146 "movq " #rega ", " #regr " \n\t"\ | |
147 "movq " #regc ", " #regp " \n\t"\ | |
148 "por " #regb ", " #regr " \n\t"\ | |
149 "por " #regd ", " #regp " \n\t"\ | |
150 "pxor " #rega ", " #regb " \n\t"\ | |
151 "pxor " #regc ", " #regd " \n\t"\ | |
152 "pand %%mm6, " #regb " \n\t"\ | |
153 "pand %%mm6, " #regd " \n\t"\ | |
154 "psrlq $1, " #regd " \n\t"\ | |
155 "psrlq $1, " #regb " \n\t"\ | |
156 "psubb " #regb ", " #regr " \n\t"\ | |
157 "psubb " #regd ", " #regp " \n\t" | |
158 | |
159 /***********************************/ | |
160 /* MMX no rounding */ | |
161 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx | |
162 #define SET_RND MOVQ_WONE | |
163 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) | |
164 #define PAVGB(a, b, c, e) PAVGB_MMX_NO_RND(a, b, c, e) | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
165 #define OP_AVG(a, b, c, e) PAVGB_MMX(a, b, c, e) |
8430 | 166 |
167 #include "dsputil_mmx_rnd_template.c" | |
168 | |
169 #undef DEF | |
170 #undef SET_RND | |
171 #undef PAVGBP | |
172 #undef PAVGB | |
173 /***********************************/ | |
174 /* MMX rounding */ | |
175 | |
176 #define DEF(x, y) x ## _ ## y ##_mmx | |
177 #define SET_RND MOVQ_WTWO | |
178 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) | |
179 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e) | |
180 | |
181 #include "dsputil_mmx_rnd_template.c" | |
182 | |
183 #undef DEF | |
184 #undef SET_RND | |
185 #undef PAVGBP | |
186 #undef PAVGB | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
187 #undef OP_AVG |
8430 | 188 |
189 /***********************************/ | |
190 /* 3Dnow specific */ | |
191 | |
192 #define DEF(x) x ## _3dnow | |
193 #define PAVGB "pavgusb" | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
194 #define OP_AVG PAVGB |
8430 | 195 |
196 #include "dsputil_mmx_avg_template.c" | |
197 | |
198 #undef DEF | |
199 #undef PAVGB | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
200 #undef OP_AVG |
8430 | 201 |
202 /***********************************/ | |
203 /* MMX2 specific */ | |
204 | |
205 #define DEF(x) x ## _mmx2 | |
206 | |
207 /* Introduced only in MMX2 set */ | |
208 #define PAVGB "pavgb" | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
209 #define OP_AVG PAVGB |
8430 | 210 |
211 #include "dsputil_mmx_avg_template.c" | |
212 | |
213 #undef DEF | |
214 #undef PAVGB | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
215 #undef OP_AVG |
8430 | 216 |
217 #define put_no_rnd_pixels16_mmx put_pixels16_mmx | |
218 #define put_no_rnd_pixels8_mmx put_pixels8_mmx | |
219 #define put_pixels16_mmx2 put_pixels16_mmx | |
220 #define put_pixels8_mmx2 put_pixels8_mmx | |
221 #define put_pixels4_mmx2 put_pixels4_mmx | |
222 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx | |
223 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx | |
224 #define put_pixels16_3dnow put_pixels16_mmx | |
225 #define put_pixels8_3dnow put_pixels8_mmx | |
226 #define put_pixels4_3dnow put_pixels4_mmx | |
227 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx | |
228 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx | |
229 | |
230 /***********************************/ | |
231 /* standard MMX */ | |
232 | |
233 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) | |
234 { | |
235 const DCTELEM *p; | |
236 uint8_t *pix; | |
237 | |
238 /* read the pixels */ | |
239 p = block; | |
240 pix = pixels; | |
241 /* unrolled loop */ | |
242 __asm__ volatile( | |
243 "movq %3, %%mm0 \n\t" | |
244 "movq 8%3, %%mm1 \n\t" | |
245 "movq 16%3, %%mm2 \n\t" | |
246 "movq 24%3, %%mm3 \n\t" | |
247 "movq 32%3, %%mm4 \n\t" | |
248 "movq 40%3, %%mm5 \n\t" | |
249 "movq 48%3, %%mm6 \n\t" | |
250 "movq 56%3, %%mm7 \n\t" | |
251 "packuswb %%mm1, %%mm0 \n\t" | |
252 "packuswb %%mm3, %%mm2 \n\t" | |
253 "packuswb %%mm5, %%mm4 \n\t" | |
254 "packuswb %%mm7, %%mm6 \n\t" | |
255 "movq %%mm0, (%0) \n\t" | |
256 "movq %%mm2, (%0, %1) \n\t" | |
257 "movq %%mm4, (%0, %1, 2) \n\t" | |
258 "movq %%mm6, (%0, %2) \n\t" | |
259 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "m"(*p) | |
260 :"memory"); | |
261 pix += line_size*4; | |
262 p += 32; | |
263 | |
264 // if here would be an exact copy of the code above | |
265 // compiler would generate some very strange code | |
266 // thus using "r" | |
267 __asm__ volatile( | |
268 "movq (%3), %%mm0 \n\t" | |
269 "movq 8(%3), %%mm1 \n\t" | |
270 "movq 16(%3), %%mm2 \n\t" | |
271 "movq 24(%3), %%mm3 \n\t" | |
272 "movq 32(%3), %%mm4 \n\t" | |
273 "movq 40(%3), %%mm5 \n\t" | |
274 "movq 48(%3), %%mm6 \n\t" | |
275 "movq 56(%3), %%mm7 \n\t" | |
276 "packuswb %%mm1, %%mm0 \n\t" | |
277 "packuswb %%mm3, %%mm2 \n\t" | |
278 "packuswb %%mm5, %%mm4 \n\t" | |
279 "packuswb %%mm7, %%mm6 \n\t" | |
280 "movq %%mm0, (%0) \n\t" | |
281 "movq %%mm2, (%0, %1) \n\t" | |
282 "movq %%mm4, (%0, %1, 2) \n\t" | |
283 "movq %%mm6, (%0, %2) \n\t" | |
284 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "r"(p) | |
285 :"memory"); | |
286 } | |
287 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
10766
diff
changeset
|
288 DECLARE_ASM_CONST(8, uint8_t, ff_vector128)[8] = |
8430 | 289 { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 }; |
290 | |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
291 #define put_signed_pixels_clamped_mmx_half(off) \ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
292 "movq "#off"(%2), %%mm1 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
293 "movq 16+"#off"(%2), %%mm2 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
294 "movq 32+"#off"(%2), %%mm3 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
295 "movq 48+"#off"(%2), %%mm4 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
296 "packsswb 8+"#off"(%2), %%mm1 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
297 "packsswb 24+"#off"(%2), %%mm2 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
298 "packsswb 40+"#off"(%2), %%mm3 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
299 "packsswb 56+"#off"(%2), %%mm4 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
300 "paddb %%mm0, %%mm1 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
301 "paddb %%mm0, %%mm2 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
302 "paddb %%mm0, %%mm3 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
303 "paddb %%mm0, %%mm4 \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
304 "movq %%mm1, (%0) \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
305 "movq %%mm2, (%0, %3) \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
306 "movq %%mm3, (%0, %3, 2) \n\t"\ |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
307 "movq %%mm4, (%0, %1) \n\t" |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
308 |
8430 | 309 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
310 { | |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
311 x86_reg line_skip = line_size; |
9341
06532529c428
Mark line_skip3 asm argument as output-only instead of using av_uninit.
reimar
parents:
9340
diff
changeset
|
312 x86_reg line_skip3; |
8430 | 313 |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
314 __asm__ volatile ( |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
315 "movq "MANGLE(ff_vector128)", %%mm0 \n\t" |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
316 "lea (%3, %3, 2), %1 \n\t" |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
317 put_signed_pixels_clamped_mmx_half(0) |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
318 "lea (%0, %3, 4), %0 \n\t" |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
319 put_signed_pixels_clamped_mmx_half(64) |
9341
06532529c428
Mark line_skip3 asm argument as output-only instead of using av_uninit.
reimar
parents:
9340
diff
changeset
|
320 :"+&r" (pixels), "=&r" (line_skip3) |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
321 :"r" (block), "r"(line_skip) |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
322 :"memory"); |
8430 | 323 } |
324 | |
325 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) | |
326 { | |
327 const DCTELEM *p; | |
328 uint8_t *pix; | |
329 int i; | |
330 | |
331 /* read the pixels */ | |
332 p = block; | |
333 pix = pixels; | |
334 MOVQ_ZERO(mm7); | |
335 i = 4; | |
336 do { | |
337 __asm__ volatile( | |
338 "movq (%2), %%mm0 \n\t" | |
339 "movq 8(%2), %%mm1 \n\t" | |
340 "movq 16(%2), %%mm2 \n\t" | |
341 "movq 24(%2), %%mm3 \n\t" | |
342 "movq %0, %%mm4 \n\t" | |
343 "movq %1, %%mm6 \n\t" | |
344 "movq %%mm4, %%mm5 \n\t" | |
345 "punpcklbw %%mm7, %%mm4 \n\t" | |
346 "punpckhbw %%mm7, %%mm5 \n\t" | |
347 "paddsw %%mm4, %%mm0 \n\t" | |
348 "paddsw %%mm5, %%mm1 \n\t" | |
349 "movq %%mm6, %%mm5 \n\t" | |
350 "punpcklbw %%mm7, %%mm6 \n\t" | |
351 "punpckhbw %%mm7, %%mm5 \n\t" | |
352 "paddsw %%mm6, %%mm2 \n\t" | |
353 "paddsw %%mm5, %%mm3 \n\t" | |
354 "packuswb %%mm1, %%mm0 \n\t" | |
355 "packuswb %%mm3, %%mm2 \n\t" | |
356 "movq %%mm0, %0 \n\t" | |
357 "movq %%mm2, %1 \n\t" | |
358 :"+m"(*pix), "+m"(*(pix+line_size)) | |
359 :"r"(p) | |
360 :"memory"); | |
361 pix += line_size*2; | |
362 p += 16; | |
363 } while (--i); | |
364 } | |
365 | |
366 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
367 { | |
368 __asm__ volatile( | |
369 "lea (%3, %3), %%"REG_a" \n\t" | |
370 ASMALIGN(3) | |
371 "1: \n\t" | |
372 "movd (%1), %%mm0 \n\t" | |
373 "movd (%1, %3), %%mm1 \n\t" | |
374 "movd %%mm0, (%2) \n\t" | |
375 "movd %%mm1, (%2, %3) \n\t" | |
376 "add %%"REG_a", %1 \n\t" | |
377 "add %%"REG_a", %2 \n\t" | |
378 "movd (%1), %%mm0 \n\t" | |
379 "movd (%1, %3), %%mm1 \n\t" | |
380 "movd %%mm0, (%2) \n\t" | |
381 "movd %%mm1, (%2, %3) \n\t" | |
382 "add %%"REG_a", %1 \n\t" | |
383 "add %%"REG_a", %2 \n\t" | |
384 "subl $4, %0 \n\t" | |
385 "jnz 1b \n\t" | |
386 : "+g"(h), "+r" (pixels), "+r" (block) | |
387 : "r"((x86_reg)line_size) | |
388 : "%"REG_a, "memory" | |
389 ); | |
390 } | |
391 | |
392 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
393 { | |
394 __asm__ volatile( | |
395 "lea (%3, %3), %%"REG_a" \n\t" | |
396 ASMALIGN(3) | |
397 "1: \n\t" | |
398 "movq (%1), %%mm0 \n\t" | |
399 "movq (%1, %3), %%mm1 \n\t" | |
400 "movq %%mm0, (%2) \n\t" | |
401 "movq %%mm1, (%2, %3) \n\t" | |
402 "add %%"REG_a", %1 \n\t" | |
403 "add %%"REG_a", %2 \n\t" | |
404 "movq (%1), %%mm0 \n\t" | |
405 "movq (%1, %3), %%mm1 \n\t" | |
406 "movq %%mm0, (%2) \n\t" | |
407 "movq %%mm1, (%2, %3) \n\t" | |
408 "add %%"REG_a", %1 \n\t" | |
409 "add %%"REG_a", %2 \n\t" | |
410 "subl $4, %0 \n\t" | |
411 "jnz 1b \n\t" | |
412 : "+g"(h), "+r" (pixels), "+r" (block) | |
413 : "r"((x86_reg)line_size) | |
414 : "%"REG_a, "memory" | |
415 ); | |
416 } | |
417 | |
418 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
419 { | |
420 __asm__ volatile( | |
421 "lea (%3, %3), %%"REG_a" \n\t" | |
422 ASMALIGN(3) | |
423 "1: \n\t" | |
424 "movq (%1), %%mm0 \n\t" | |
425 "movq 8(%1), %%mm4 \n\t" | |
426 "movq (%1, %3), %%mm1 \n\t" | |
427 "movq 8(%1, %3), %%mm5 \n\t" | |
428 "movq %%mm0, (%2) \n\t" | |
429 "movq %%mm4, 8(%2) \n\t" | |
430 "movq %%mm1, (%2, %3) \n\t" | |
431 "movq %%mm5, 8(%2, %3) \n\t" | |
432 "add %%"REG_a", %1 \n\t" | |
433 "add %%"REG_a", %2 \n\t" | |
434 "movq (%1), %%mm0 \n\t" | |
435 "movq 8(%1), %%mm4 \n\t" | |
436 "movq (%1, %3), %%mm1 \n\t" | |
437 "movq 8(%1, %3), %%mm5 \n\t" | |
438 "movq %%mm0, (%2) \n\t" | |
439 "movq %%mm4, 8(%2) \n\t" | |
440 "movq %%mm1, (%2, %3) \n\t" | |
441 "movq %%mm5, 8(%2, %3) \n\t" | |
442 "add %%"REG_a", %1 \n\t" | |
443 "add %%"REG_a", %2 \n\t" | |
444 "subl $4, %0 \n\t" | |
445 "jnz 1b \n\t" | |
446 : "+g"(h), "+r" (pixels), "+r" (block) | |
447 : "r"((x86_reg)line_size) | |
448 : "%"REG_a, "memory" | |
449 ); | |
450 } | |
451 | |
452 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
453 { | |
454 __asm__ volatile( | |
455 "1: \n\t" | |
456 "movdqu (%1), %%xmm0 \n\t" | |
457 "movdqu (%1,%3), %%xmm1 \n\t" | |
458 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
459 "movdqu (%1,%4), %%xmm3 \n\t" | |
460 "movdqa %%xmm0, (%2) \n\t" | |
461 "movdqa %%xmm1, (%2,%3) \n\t" | |
462 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
463 "movdqa %%xmm3, (%2,%4) \n\t" | |
464 "subl $4, %0 \n\t" | |
465 "lea (%1,%3,4), %1 \n\t" | |
466 "lea (%2,%3,4), %2 \n\t" | |
467 "jnz 1b \n\t" | |
468 : "+g"(h), "+r" (pixels), "+r" (block) | |
469 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size) | |
470 : "memory" | |
471 ); | |
472 } | |
473 | |
474 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
475 { | |
476 __asm__ volatile( | |
477 "1: \n\t" | |
478 "movdqu (%1), %%xmm0 \n\t" | |
479 "movdqu (%1,%3), %%xmm1 \n\t" | |
480 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
481 "movdqu (%1,%4), %%xmm3 \n\t" | |
482 "pavgb (%2), %%xmm0 \n\t" | |
483 "pavgb (%2,%3), %%xmm1 \n\t" | |
484 "pavgb (%2,%3,2), %%xmm2 \n\t" | |
485 "pavgb (%2,%4), %%xmm3 \n\t" | |
486 "movdqa %%xmm0, (%2) \n\t" | |
487 "movdqa %%xmm1, (%2,%3) \n\t" | |
488 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
489 "movdqa %%xmm3, (%2,%4) \n\t" | |
490 "subl $4, %0 \n\t" | |
491 "lea (%1,%3,4), %1 \n\t" | |
492 "lea (%2,%3,4), %2 \n\t" | |
493 "jnz 1b \n\t" | |
494 : "+g"(h), "+r" (pixels), "+r" (block) | |
495 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size) | |
496 : "memory" | |
497 ); | |
498 } | |
499 | |
500 #define CLEAR_BLOCKS(name,n) \ | |
501 static void name(DCTELEM *blocks)\ | |
502 {\ | |
503 __asm__ volatile(\ | |
504 "pxor %%mm7, %%mm7 \n\t"\ | |
505 "mov %1, %%"REG_a" \n\t"\ | |
506 "1: \n\t"\ | |
507 "movq %%mm7, (%0, %%"REG_a") \n\t"\ | |
508 "movq %%mm7, 8(%0, %%"REG_a") \n\t"\ | |
509 "movq %%mm7, 16(%0, %%"REG_a") \n\t"\ | |
510 "movq %%mm7, 24(%0, %%"REG_a") \n\t"\ | |
511 "add $32, %%"REG_a" \n\t"\ | |
512 " js 1b \n\t"\ | |
513 : : "r" (((uint8_t *)blocks)+128*n),\ | |
514 "i" (-128*n)\ | |
515 : "%"REG_a\ | |
516 );\ | |
517 } | |
518 CLEAR_BLOCKS(clear_blocks_mmx, 6) | |
519 CLEAR_BLOCKS(clear_block_mmx, 1) | |
520 | |
521 static void clear_block_sse(DCTELEM *block) | |
522 { | |
523 __asm__ volatile( | |
524 "xorps %%xmm0, %%xmm0 \n" | |
525 "movaps %%xmm0, (%0) \n" | |
526 "movaps %%xmm0, 16(%0) \n" | |
527 "movaps %%xmm0, 32(%0) \n" | |
528 "movaps %%xmm0, 48(%0) \n" | |
529 "movaps %%xmm0, 64(%0) \n" | |
530 "movaps %%xmm0, 80(%0) \n" | |
531 "movaps %%xmm0, 96(%0) \n" | |
532 "movaps %%xmm0, 112(%0) \n" | |
533 :: "r"(block) | |
534 : "memory" | |
535 ); | |
536 } | |
537 | |
9861 | 538 static void clear_blocks_sse(DCTELEM *blocks) |
539 {\ | |
540 __asm__ volatile( | |
541 "xorps %%xmm0, %%xmm0 \n" | |
542 "mov %1, %%"REG_a" \n" | |
543 "1: \n" | |
544 "movaps %%xmm0, (%0, %%"REG_a") \n" | |
545 "movaps %%xmm0, 16(%0, %%"REG_a") \n" | |
546 "movaps %%xmm0, 32(%0, %%"REG_a") \n" | |
547 "movaps %%xmm0, 48(%0, %%"REG_a") \n" | |
548 "movaps %%xmm0, 64(%0, %%"REG_a") \n" | |
549 "movaps %%xmm0, 80(%0, %%"REG_a") \n" | |
550 "movaps %%xmm0, 96(%0, %%"REG_a") \n" | |
551 "movaps %%xmm0, 112(%0, %%"REG_a") \n" | |
552 "add $128, %%"REG_a" \n" | |
553 " js 1b \n" | |
554 : : "r" (((uint8_t *)blocks)+128*6), | |
555 "i" (-128*6) | |
556 : "%"REG_a | |
557 ); | |
558 } | |
559 | |
8430 | 560 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){ |
561 x86_reg i=0; | |
562 __asm__ volatile( | |
563 "jmp 2f \n\t" | |
564 "1: \n\t" | |
565 "movq (%1, %0), %%mm0 \n\t" | |
566 "movq (%2, %0), %%mm1 \n\t" | |
567 "paddb %%mm0, %%mm1 \n\t" | |
568 "movq %%mm1, (%2, %0) \n\t" | |
569 "movq 8(%1, %0), %%mm0 \n\t" | |
570 "movq 8(%2, %0), %%mm1 \n\t" | |
571 "paddb %%mm0, %%mm1 \n\t" | |
572 "movq %%mm1, 8(%2, %0) \n\t" | |
573 "add $16, %0 \n\t" | |
574 "2: \n\t" | |
575 "cmp %3, %0 \n\t" | |
576 " js 1b \n\t" | |
577 : "+r" (i) | |
578 : "r"(src), "r"(dst), "r"((x86_reg)w-15) | |
579 ); | |
580 for(; i<w; i++) | |
581 dst[i+0] += src[i+0]; | |
582 } | |
583 | |
584 static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){ | |
585 x86_reg i=0; | |
586 __asm__ volatile( | |
587 "jmp 2f \n\t" | |
588 "1: \n\t" | |
589 "movq (%2, %0), %%mm0 \n\t" | |
590 "movq 8(%2, %0), %%mm1 \n\t" | |
591 "paddb (%3, %0), %%mm0 \n\t" | |
592 "paddb 8(%3, %0), %%mm1 \n\t" | |
593 "movq %%mm0, (%1, %0) \n\t" | |
594 "movq %%mm1, 8(%1, %0) \n\t" | |
595 "add $16, %0 \n\t" | |
596 "2: \n\t" | |
597 "cmp %4, %0 \n\t" | |
598 " js 1b \n\t" | |
599 : "+r" (i) | |
600 : "r"(dst), "r"(src1), "r"(src2), "r"((x86_reg)w-15) | |
601 ); | |
602 for(; i<w; i++) | |
603 dst[i] = src1[i] + src2[i]; | |
604 } | |
605 | |
8798
a5c8210814d7
Add check whether the compiler/assembler supports 10 or more operands.
diego
parents:
8760
diff
changeset
|
606 #if HAVE_7REGS && HAVE_TEN_OPERANDS |
10431 | 607 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top) { |
8760 | 608 x86_reg w2 = -w; |
609 x86_reg x; | |
610 int l = *left & 0xff; | |
611 int tl = *left_top & 0xff; | |
612 int t; | |
613 __asm__ volatile( | |
614 "mov %7, %3 \n" | |
615 "1: \n" | |
616 "movzx (%3,%4), %2 \n" | |
617 "mov %2, %k3 \n" | |
618 "sub %b1, %b3 \n" | |
619 "add %b0, %b3 \n" | |
620 "mov %2, %1 \n" | |
621 "cmp %0, %2 \n" | |
622 "cmovg %0, %2 \n" | |
623 "cmovg %1, %0 \n" | |
624 "cmp %k3, %0 \n" | |
625 "cmovg %k3, %0 \n" | |
626 "mov %7, %3 \n" | |
627 "cmp %2, %0 \n" | |
628 "cmovl %2, %0 \n" | |
629 "add (%6,%4), %b0 \n" | |
630 "mov %b0, (%5,%4) \n" | |
631 "inc %4 \n" | |
632 "jl 1b \n" | |
633 :"+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2) | |
634 :"r"(dst+w), "r"(diff+w), "rm"(top+w) | |
635 ); | |
636 *left = l; | |
637 *left_top = tl; | |
638 } | |
639 #endif | |
640 | |
8430 | 641 #define H263_LOOP_FILTER \ |
642 "pxor %%mm7, %%mm7 \n\t"\ | |
643 "movq %0, %%mm0 \n\t"\ | |
644 "movq %0, %%mm1 \n\t"\ | |
645 "movq %3, %%mm2 \n\t"\ | |
646 "movq %3, %%mm3 \n\t"\ | |
647 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
648 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
649 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
650 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
651 "psubw %%mm2, %%mm0 \n\t"\ | |
652 "psubw %%mm3, %%mm1 \n\t"\ | |
653 "movq %1, %%mm2 \n\t"\ | |
654 "movq %1, %%mm3 \n\t"\ | |
655 "movq %2, %%mm4 \n\t"\ | |
656 "movq %2, %%mm5 \n\t"\ | |
657 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
658 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
659 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
660 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
661 "psubw %%mm2, %%mm4 \n\t"\ | |
662 "psubw %%mm3, %%mm5 \n\t"\ | |
663 "psllw $2, %%mm4 \n\t"\ | |
664 "psllw $2, %%mm5 \n\t"\ | |
665 "paddw %%mm0, %%mm4 \n\t"\ | |
666 "paddw %%mm1, %%mm5 \n\t"\ | |
667 "pxor %%mm6, %%mm6 \n\t"\ | |
668 "pcmpgtw %%mm4, %%mm6 \n\t"\ | |
669 "pcmpgtw %%mm5, %%mm7 \n\t"\ | |
670 "pxor %%mm6, %%mm4 \n\t"\ | |
671 "pxor %%mm7, %%mm5 \n\t"\ | |
672 "psubw %%mm6, %%mm4 \n\t"\ | |
673 "psubw %%mm7, %%mm5 \n\t"\ | |
674 "psrlw $3, %%mm4 \n\t"\ | |
675 "psrlw $3, %%mm5 \n\t"\ | |
676 "packuswb %%mm5, %%mm4 \n\t"\ | |
677 "packsswb %%mm7, %%mm6 \n\t"\ | |
678 "pxor %%mm7, %%mm7 \n\t"\ | |
679 "movd %4, %%mm2 \n\t"\ | |
680 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
681 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
682 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
683 "psubusb %%mm4, %%mm2 \n\t"\ | |
684 "movq %%mm2, %%mm3 \n\t"\ | |
685 "psubusb %%mm4, %%mm3 \n\t"\ | |
686 "psubb %%mm3, %%mm2 \n\t"\ | |
687 "movq %1, %%mm3 \n\t"\ | |
688 "movq %2, %%mm4 \n\t"\ | |
689 "pxor %%mm6, %%mm3 \n\t"\ | |
690 "pxor %%mm6, %%mm4 \n\t"\ | |
691 "paddusb %%mm2, %%mm3 \n\t"\ | |
692 "psubusb %%mm2, %%mm4 \n\t"\ | |
693 "pxor %%mm6, %%mm3 \n\t"\ | |
694 "pxor %%mm6, %%mm4 \n\t"\ | |
695 "paddusb %%mm2, %%mm2 \n\t"\ | |
696 "packsswb %%mm1, %%mm0 \n\t"\ | |
697 "pcmpgtb %%mm0, %%mm7 \n\t"\ | |
698 "pxor %%mm7, %%mm0 \n\t"\ | |
699 "psubb %%mm7, %%mm0 \n\t"\ | |
700 "movq %%mm0, %%mm1 \n\t"\ | |
701 "psubusb %%mm2, %%mm0 \n\t"\ | |
702 "psubb %%mm0, %%mm1 \n\t"\ | |
703 "pand %5, %%mm1 \n\t"\ | |
704 "psrlw $2, %%mm1 \n\t"\ | |
705 "pxor %%mm7, %%mm1 \n\t"\ | |
706 "psubb %%mm7, %%mm1 \n\t"\ | |
707 "movq %0, %%mm5 \n\t"\ | |
708 "movq %3, %%mm6 \n\t"\ | |
709 "psubb %%mm1, %%mm5 \n\t"\ | |
710 "paddb %%mm1, %%mm6 \n\t" | |
711 | |
712 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale){ | |
10749
5cca4b6c459d
Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents:
10645
diff
changeset
|
713 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { |
8430 | 714 const int strength= ff_h263_loop_filter_strength[qscale]; |
715 | |
716 __asm__ volatile( | |
717 | |
718 H263_LOOP_FILTER | |
719 | |
720 "movq %%mm3, %1 \n\t" | |
721 "movq %%mm4, %2 \n\t" | |
722 "movq %%mm5, %0 \n\t" | |
723 "movq %%mm6, %3 \n\t" | |
724 : "+m" (*(uint64_t*)(src - 2*stride)), | |
725 "+m" (*(uint64_t*)(src - 1*stride)), | |
726 "+m" (*(uint64_t*)(src + 0*stride)), | |
727 "+m" (*(uint64_t*)(src + 1*stride)) | |
728 : "g" (2*strength), "m"(ff_pb_FC) | |
729 ); | |
730 } | |
731 } | |
732 | |
733 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ | |
734 __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ... | |
735 "movd %4, %%mm0 \n\t" | |
736 "movd %5, %%mm1 \n\t" | |
737 "movd %6, %%mm2 \n\t" | |
738 "movd %7, %%mm3 \n\t" | |
739 "punpcklbw %%mm1, %%mm0 \n\t" | |
740 "punpcklbw %%mm3, %%mm2 \n\t" | |
741 "movq %%mm0, %%mm1 \n\t" | |
742 "punpcklwd %%mm2, %%mm0 \n\t" | |
743 "punpckhwd %%mm2, %%mm1 \n\t" | |
744 "movd %%mm0, %0 \n\t" | |
745 "punpckhdq %%mm0, %%mm0 \n\t" | |
746 "movd %%mm0, %1 \n\t" | |
747 "movd %%mm1, %2 \n\t" | |
748 "punpckhdq %%mm1, %%mm1 \n\t" | |
749 "movd %%mm1, %3 \n\t" | |
750 | |
751 : "=m" (*(uint32_t*)(dst + 0*dst_stride)), | |
752 "=m" (*(uint32_t*)(dst + 1*dst_stride)), | |
753 "=m" (*(uint32_t*)(dst + 2*dst_stride)), | |
754 "=m" (*(uint32_t*)(dst + 3*dst_stride)) | |
755 : "m" (*(uint32_t*)(src + 0*src_stride)), | |
756 "m" (*(uint32_t*)(src + 1*src_stride)), | |
757 "m" (*(uint32_t*)(src + 2*src_stride)), | |
758 "m" (*(uint32_t*)(src + 3*src_stride)) | |
759 ); | |
760 } | |
761 | |
762 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){ | |
10749
5cca4b6c459d
Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents:
10645
diff
changeset
|
763 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { |
8430 | 764 const int strength= ff_h263_loop_filter_strength[qscale]; |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
10766
diff
changeset
|
765 DECLARE_ALIGNED(8, uint64_t, temp)[4]; |
8430 | 766 uint8_t *btemp= (uint8_t*)temp; |
767 | |
768 src -= 2; | |
769 | |
770 transpose4x4(btemp , src , 8, stride); | |
771 transpose4x4(btemp+4, src + 4*stride, 8, stride); | |
772 __asm__ volatile( | |
773 H263_LOOP_FILTER // 5 3 4 6 | |
774 | |
775 : "+m" (temp[0]), | |
776 "+m" (temp[1]), | |
777 "+m" (temp[2]), | |
778 "+m" (temp[3]) | |
779 : "g" (2*strength), "m"(ff_pb_FC) | |
780 ); | |
781 | |
782 __asm__ volatile( | |
783 "movq %%mm5, %%mm1 \n\t" | |
784 "movq %%mm4, %%mm0 \n\t" | |
785 "punpcklbw %%mm3, %%mm5 \n\t" | |
786 "punpcklbw %%mm6, %%mm4 \n\t" | |
787 "punpckhbw %%mm3, %%mm1 \n\t" | |
788 "punpckhbw %%mm6, %%mm0 \n\t" | |
789 "movq %%mm5, %%mm3 \n\t" | |
790 "movq %%mm1, %%mm6 \n\t" | |
791 "punpcklwd %%mm4, %%mm5 \n\t" | |
792 "punpcklwd %%mm0, %%mm1 \n\t" | |
793 "punpckhwd %%mm4, %%mm3 \n\t" | |
794 "punpckhwd %%mm0, %%mm6 \n\t" | |
795 "movd %%mm5, (%0) \n\t" | |
796 "punpckhdq %%mm5, %%mm5 \n\t" | |
797 "movd %%mm5, (%0,%2) \n\t" | |
798 "movd %%mm3, (%0,%2,2) \n\t" | |
799 "punpckhdq %%mm3, %%mm3 \n\t" | |
800 "movd %%mm3, (%0,%3) \n\t" | |
801 "movd %%mm1, (%1) \n\t" | |
802 "punpckhdq %%mm1, %%mm1 \n\t" | |
803 "movd %%mm1, (%1,%2) \n\t" | |
804 "movd %%mm6, (%1,%2,2) \n\t" | |
805 "punpckhdq %%mm6, %%mm6 \n\t" | |
806 "movd %%mm6, (%1,%3) \n\t" | |
807 :: "r" (src), | |
808 "r" (src + 4*stride), | |
809 "r" ((x86_reg) stride ), | |
810 "r" ((x86_reg)(3*stride)) | |
811 ); | |
812 } | |
813 } | |
814 | |
815 /* draw the edges of width 'w' of an image of size width, height | |
816 this mmx version can only handle w==8 || w==16 */ | |
817 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w) | |
818 { | |
819 uint8_t *ptr, *last_line; | |
820 int i; | |
821 | |
822 last_line = buf + (height - 1) * wrap; | |
823 /* left and right */ | |
824 ptr = buf; | |
825 if(w==8) | |
826 { | |
827 __asm__ volatile( | |
828 "1: \n\t" | |
829 "movd (%0), %%mm0 \n\t" | |
830 "punpcklbw %%mm0, %%mm0 \n\t" | |
831 "punpcklwd %%mm0, %%mm0 \n\t" | |
832 "punpckldq %%mm0, %%mm0 \n\t" | |
833 "movq %%mm0, -8(%0) \n\t" | |
834 "movq -8(%0, %2), %%mm1 \n\t" | |
835 "punpckhbw %%mm1, %%mm1 \n\t" | |
836 "punpckhwd %%mm1, %%mm1 \n\t" | |
837 "punpckhdq %%mm1, %%mm1 \n\t" | |
838 "movq %%mm1, (%0, %2) \n\t" | |
839 "add %1, %0 \n\t" | |
840 "cmp %3, %0 \n\t" | |
841 " jb 1b \n\t" | |
842 : "+r" (ptr) | |
843 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height) | |
844 ); | |
845 } | |
846 else | |
847 { | |
848 __asm__ volatile( | |
849 "1: \n\t" | |
850 "movd (%0), %%mm0 \n\t" | |
851 "punpcklbw %%mm0, %%mm0 \n\t" | |
852 "punpcklwd %%mm0, %%mm0 \n\t" | |
853 "punpckldq %%mm0, %%mm0 \n\t" | |
854 "movq %%mm0, -8(%0) \n\t" | |
855 "movq %%mm0, -16(%0) \n\t" | |
856 "movq -8(%0, %2), %%mm1 \n\t" | |
857 "punpckhbw %%mm1, %%mm1 \n\t" | |
858 "punpckhwd %%mm1, %%mm1 \n\t" | |
859 "punpckhdq %%mm1, %%mm1 \n\t" | |
860 "movq %%mm1, (%0, %2) \n\t" | |
861 "movq %%mm1, 8(%0, %2) \n\t" | |
862 "add %1, %0 \n\t" | |
863 "cmp %3, %0 \n\t" | |
864 " jb 1b \n\t" | |
865 : "+r" (ptr) | |
866 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height) | |
867 ); | |
868 } | |
869 | |
870 for(i=0;i<w;i+=4) { | |
871 /* top and bottom (and hopefully also the corners) */ | |
872 ptr= buf - (i + 1) * wrap - w; | |
873 __asm__ volatile( | |
874 "1: \n\t" | |
875 "movq (%1, %0), %%mm0 \n\t" | |
876 "movq %%mm0, (%0) \n\t" | |
877 "movq %%mm0, (%0, %2) \n\t" | |
878 "movq %%mm0, (%0, %2, 2) \n\t" | |
879 "movq %%mm0, (%0, %3) \n\t" | |
880 "add $8, %0 \n\t" | |
881 "cmp %4, %0 \n\t" | |
882 " jb 1b \n\t" | |
883 : "+r" (ptr) | |
884 : "r" ((x86_reg)buf - (x86_reg)ptr - w), "r" ((x86_reg)-wrap), "r" ((x86_reg)-wrap*3), "r" (ptr+width+2*w) | |
885 ); | |
886 ptr= last_line + (i + 1) * wrap - w; | |
887 __asm__ volatile( | |
888 "1: \n\t" | |
889 "movq (%1, %0), %%mm0 \n\t" | |
890 "movq %%mm0, (%0) \n\t" | |
891 "movq %%mm0, (%0, %2) \n\t" | |
892 "movq %%mm0, (%0, %2, 2) \n\t" | |
893 "movq %%mm0, (%0, %3) \n\t" | |
894 "add $8, %0 \n\t" | |
895 "cmp %4, %0 \n\t" | |
896 " jb 1b \n\t" | |
897 : "+r" (ptr) | |
898 : "r" ((x86_reg)last_line - (x86_reg)ptr - w), "r" ((x86_reg)wrap), "r" ((x86_reg)wrap*3), "r" (ptr+width+2*w) | |
899 ); | |
900 } | |
901 } | |
902 | |
903 #define PAETH(cpu, abs3)\ | |
904 static void add_png_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\ | |
905 {\ | |
906 x86_reg i = -bpp;\ | |
907 x86_reg end = w-3;\ | |
908 __asm__ volatile(\ | |
909 "pxor %%mm7, %%mm7 \n"\ | |
910 "movd (%1,%0), %%mm0 \n"\ | |
911 "movd (%2,%0), %%mm1 \n"\ | |
912 "punpcklbw %%mm7, %%mm0 \n"\ | |
913 "punpcklbw %%mm7, %%mm1 \n"\ | |
914 "add %4, %0 \n"\ | |
915 "1: \n"\ | |
916 "movq %%mm1, %%mm2 \n"\ | |
917 "movd (%2,%0), %%mm1 \n"\ | |
918 "movq %%mm2, %%mm3 \n"\ | |
919 "punpcklbw %%mm7, %%mm1 \n"\ | |
920 "movq %%mm2, %%mm4 \n"\ | |
921 "psubw %%mm1, %%mm3 \n"\ | |
922 "psubw %%mm0, %%mm4 \n"\ | |
923 "movq %%mm3, %%mm5 \n"\ | |
924 "paddw %%mm4, %%mm5 \n"\ | |
925 abs3\ | |
926 "movq %%mm4, %%mm6 \n"\ | |
927 "pminsw %%mm5, %%mm6 \n"\ | |
928 "pcmpgtw %%mm6, %%mm3 \n"\ | |
929 "pcmpgtw %%mm5, %%mm4 \n"\ | |
930 "movq %%mm4, %%mm6 \n"\ | |
931 "pand %%mm3, %%mm4 \n"\ | |
932 "pandn %%mm3, %%mm6 \n"\ | |
933 "pandn %%mm0, %%mm3 \n"\ | |
934 "movd (%3,%0), %%mm0 \n"\ | |
935 "pand %%mm1, %%mm6 \n"\ | |
936 "pand %%mm4, %%mm2 \n"\ | |
937 "punpcklbw %%mm7, %%mm0 \n"\ | |
938 "movq %6, %%mm5 \n"\ | |
939 "paddw %%mm6, %%mm0 \n"\ | |
940 "paddw %%mm2, %%mm3 \n"\ | |
941 "paddw %%mm3, %%mm0 \n"\ | |
942 "pand %%mm5, %%mm0 \n"\ | |
943 "movq %%mm0, %%mm3 \n"\ | |
944 "packuswb %%mm3, %%mm3 \n"\ | |
945 "movd %%mm3, (%1,%0) \n"\ | |
946 "add %4, %0 \n"\ | |
947 "cmp %5, %0 \n"\ | |
948 "jle 1b \n"\ | |
949 :"+r"(i)\ | |
950 :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\ | |
951 "m"(ff_pw_255)\ | |
952 :"memory"\ | |
953 );\ | |
954 } | |
955 | |
956 #define ABS3_MMX2\ | |
957 "psubw %%mm5, %%mm7 \n"\ | |
958 "pmaxsw %%mm7, %%mm5 \n"\ | |
959 "pxor %%mm6, %%mm6 \n"\ | |
960 "pxor %%mm7, %%mm7 \n"\ | |
961 "psubw %%mm3, %%mm6 \n"\ | |
962 "psubw %%mm4, %%mm7 \n"\ | |
963 "pmaxsw %%mm6, %%mm3 \n"\ | |
964 "pmaxsw %%mm7, %%mm4 \n"\ | |
965 "pxor %%mm7, %%mm7 \n" | |
966 | |
967 #define ABS3_SSSE3\ | |
968 "pabsw %%mm3, %%mm3 \n"\ | |
969 "pabsw %%mm4, %%mm4 \n"\ | |
970 "pabsw %%mm5, %%mm5 \n" | |
971 | |
972 PAETH(mmx2, ABS3_MMX2) | |
8590 | 973 #if HAVE_SSSE3 |
8430 | 974 PAETH(ssse3, ABS3_SSSE3) |
975 #endif | |
976 | |
977 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ | |
978 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ | |
979 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ | |
980 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ | |
981 "movq "#in7", " #m3 " \n\t" /* d */\ | |
982 "movq "#in0", %%mm5 \n\t" /* D */\ | |
983 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
984 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
985 "movq "#in1", %%mm5 \n\t" /* C */\ | |
986 "movq "#in2", %%mm6 \n\t" /* B */\ | |
987 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
988 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
989 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
990 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
991 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ | |
992 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ | |
993 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
994 "psraw $5, %%mm5 \n\t"\ | |
995 "packuswb %%mm5, %%mm5 \n\t"\ | |
996 OP(%%mm5, out, %%mm7, d) | |
997 | |
998 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ | |
999 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1000 uint64_t temp;\ | |
1001 \ | |
1002 __asm__ volatile(\ | |
1003 "pxor %%mm7, %%mm7 \n\t"\ | |
1004 "1: \n\t"\ | |
1005 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1006 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1007 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1008 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1009 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1010 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1011 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1012 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1013 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1014 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1015 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1016 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1017 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1018 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1019 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1020 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1021 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1022 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1023 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1024 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1025 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1026 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1027 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1028 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1029 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1030 "paddw %6, %%mm6 \n\t"\ | |
1031 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1032 "psraw $5, %%mm0 \n\t"\ | |
1033 "movq %%mm0, %5 \n\t"\ | |
1034 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ | |
1035 \ | |
1036 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ | |
1037 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
1038 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
1039 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
1040 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
1041 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
1042 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
1043 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
1044 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
1045 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1046 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1047 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
1048 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
1049 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
1050 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
1051 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1052 "paddw %%mm2, %%mm1 \n\t" /* a */\ | |
1053 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
1054 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1055 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ | |
1056 "paddw %6, %%mm1 \n\t"\ | |
1057 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ | |
1058 "psraw $5, %%mm3 \n\t"\ | |
1059 "movq %5, %%mm1 \n\t"\ | |
1060 "packuswb %%mm3, %%mm1 \n\t"\ | |
1061 OP_MMX2(%%mm1, (%1),%%mm4, q)\ | |
1062 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ | |
1063 \ | |
1064 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ | |
1065 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
1066 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
1067 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
1068 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
1069 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
1070 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
1071 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
1072 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
1073 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1074 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
1075 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
1076 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
1077 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ | |
1078 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ | |
1079 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
1080 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
1081 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
1082 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
1083 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
1084 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
1085 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ | |
1086 "paddw %6, %%mm0 \n\t"\ | |
1087 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1088 "psraw $5, %%mm0 \n\t"\ | |
1089 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ | |
1090 \ | |
1091 "paddw %%mm5, %%mm3 \n\t" /* a */\ | |
1092 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
1093 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
1094 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
1095 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
1096 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
1097 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
1098 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
1099 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
1100 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ | |
1101 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
1102 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1103 "paddw %6, %%mm4 \n\t"\ | |
1104 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ | |
1105 "psraw $5, %%mm4 \n\t"\ | |
1106 "packuswb %%mm4, %%mm0 \n\t"\ | |
1107 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ | |
1108 \ | |
1109 "add %3, %0 \n\t"\ | |
1110 "add %4, %1 \n\t"\ | |
1111 "decl %2 \n\t"\ | |
1112 " jnz 1b \n\t"\ | |
1113 : "+a"(src), "+c"(dst), "+D"(h)\ | |
1114 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ | |
1115 : "memory"\ | |
1116 );\ | |
1117 }\ | |
1118 \ | |
1119 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1120 int i;\ | |
1121 int16_t temp[16];\ | |
1122 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1123 for(i=0; i<h; i++)\ | |
1124 {\ | |
1125 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1126 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1127 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1128 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
1129 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1130 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
1131 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
1132 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
1133 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
1134 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
1135 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
1136 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
1137 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
1138 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
1139 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
1140 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
1141 __asm__ volatile(\ | |
1142 "movq (%0), %%mm0 \n\t"\ | |
1143 "movq 8(%0), %%mm1 \n\t"\ | |
1144 "paddw %2, %%mm0 \n\t"\ | |
1145 "paddw %2, %%mm1 \n\t"\ | |
1146 "psraw $5, %%mm0 \n\t"\ | |
1147 "psraw $5, %%mm1 \n\t"\ | |
1148 "packuswb %%mm1, %%mm0 \n\t"\ | |
1149 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ | |
1150 "movq 16(%0), %%mm0 \n\t"\ | |
1151 "movq 24(%0), %%mm1 \n\t"\ | |
1152 "paddw %2, %%mm0 \n\t"\ | |
1153 "paddw %2, %%mm1 \n\t"\ | |
1154 "psraw $5, %%mm0 \n\t"\ | |
1155 "psraw $5, %%mm1 \n\t"\ | |
1156 "packuswb %%mm1, %%mm0 \n\t"\ | |
1157 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ | |
1158 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
1159 : "memory"\ | |
1160 );\ | |
1161 dst+=dstStride;\ | |
1162 src+=srcStride;\ | |
1163 }\ | |
1164 }\ | |
1165 \ | |
1166 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1167 __asm__ volatile(\ | |
1168 "pxor %%mm7, %%mm7 \n\t"\ | |
1169 "1: \n\t"\ | |
1170 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1171 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1172 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1173 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1174 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1175 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1176 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1177 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1178 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1179 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1180 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1181 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1182 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1183 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1184 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1185 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1186 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1187 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1188 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1189 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1190 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1191 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1192 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1193 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1194 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1195 "paddw %5, %%mm6 \n\t"\ | |
1196 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1197 "psraw $5, %%mm0 \n\t"\ | |
1198 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ | |
1199 \ | |
1200 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ | |
1201 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
1202 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
1203 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
1204 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
1205 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
1206 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
1207 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
1208 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
1209 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1210 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1211 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1212 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1213 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1214 "paddw %5, %%mm1 \n\t"\ | |
1215 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ | |
1216 "psraw $5, %%mm3 \n\t"\ | |
1217 "packuswb %%mm3, %%mm0 \n\t"\ | |
1218 OP_MMX2(%%mm0, (%1), %%mm4, q)\ | |
1219 \ | |
1220 "add %3, %0 \n\t"\ | |
1221 "add %4, %1 \n\t"\ | |
1222 "decl %2 \n\t"\ | |
1223 " jnz 1b \n\t"\ | |
1224 : "+a"(src), "+c"(dst), "+d"(h)\ | |
1225 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER)\ | |
1226 : "memory"\ | |
1227 );\ | |
1228 }\ | |
1229 \ | |
1230 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1231 int i;\ | |
1232 int16_t temp[8];\ | |
1233 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1234 for(i=0; i<h; i++)\ | |
1235 {\ | |
1236 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1237 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1238 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1239 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
1240 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1241 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
1242 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
1243 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
1244 __asm__ volatile(\ | |
1245 "movq (%0), %%mm0 \n\t"\ | |
1246 "movq 8(%0), %%mm1 \n\t"\ | |
1247 "paddw %2, %%mm0 \n\t"\ | |
1248 "paddw %2, %%mm1 \n\t"\ | |
1249 "psraw $5, %%mm0 \n\t"\ | |
1250 "psraw $5, %%mm1 \n\t"\ | |
1251 "packuswb %%mm1, %%mm0 \n\t"\ | |
1252 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ | |
1253 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
1254 :"memory"\ | |
1255 );\ | |
1256 dst+=dstStride;\ | |
1257 src+=srcStride;\ | |
1258 }\ | |
1259 } | |
1260 | |
1261 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
1262 \ | |
1263 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1264 uint64_t temp[17*4];\ | |
1265 uint64_t *temp_ptr= temp;\ | |
1266 int count= 17;\ | |
1267 \ | |
1268 /*FIXME unroll */\ | |
1269 __asm__ volatile(\ | |
1270 "pxor %%mm7, %%mm7 \n\t"\ | |
1271 "1: \n\t"\ | |
1272 "movq (%0), %%mm0 \n\t"\ | |
1273 "movq (%0), %%mm1 \n\t"\ | |
1274 "movq 8(%0), %%mm2 \n\t"\ | |
1275 "movq 8(%0), %%mm3 \n\t"\ | |
1276 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1277 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1278 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1279 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1280 "movq %%mm0, (%1) \n\t"\ | |
1281 "movq %%mm1, 17*8(%1) \n\t"\ | |
1282 "movq %%mm2, 2*17*8(%1) \n\t"\ | |
1283 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
1284 "add $8, %1 \n\t"\ | |
1285 "add %3, %0 \n\t"\ | |
1286 "decl %2 \n\t"\ | |
1287 " jnz 1b \n\t"\ | |
1288 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ | |
1289 : "r" ((x86_reg)srcStride)\ | |
1290 : "memory"\ | |
1291 );\ | |
1292 \ | |
1293 temp_ptr= temp;\ | |
1294 count=4;\ | |
1295 \ | |
1296 /*FIXME reorder for speed */\ | |
1297 __asm__ volatile(\ | |
1298 /*"pxor %%mm7, %%mm7 \n\t"*/\ | |
1299 "1: \n\t"\ | |
1300 "movq (%0), %%mm0 \n\t"\ | |
1301 "movq 8(%0), %%mm1 \n\t"\ | |
1302 "movq 16(%0), %%mm2 \n\t"\ | |
1303 "movq 24(%0), %%mm3 \n\t"\ | |
1304 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ | |
1305 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
1306 "add %4, %1 \n\t"\ | |
1307 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ | |
1308 \ | |
1309 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ | |
1310 "add %4, %1 \n\t"\ | |
1311 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ | |
1312 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
1313 "add %4, %1 \n\t"\ | |
1314 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ | |
1315 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
1316 "add %4, %1 \n\t"\ | |
1317 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ | |
1318 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
1319 "add %4, %1 \n\t"\ | |
1320 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ | |
1321 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
1322 "add %4, %1 \n\t"\ | |
1323 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ | |
1324 \ | |
1325 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ | |
1326 "add %4, %1 \n\t" \ | |
1327 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ | |
1328 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
1329 \ | |
1330 "add $136, %0 \n\t"\ | |
1331 "add %6, %1 \n\t"\ | |
1332 "decl %2 \n\t"\ | |
1333 " jnz 1b \n\t"\ | |
1334 \ | |
1335 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ | |
1336 : "r"((x86_reg)dstStride), "r"(2*(x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(x86_reg)dstStride)\ | |
1337 :"memory"\ | |
1338 );\ | |
1339 }\ | |
1340 \ | |
1341 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1342 uint64_t temp[9*2];\ | |
1343 uint64_t *temp_ptr= temp;\ | |
1344 int count= 9;\ | |
1345 \ | |
1346 /*FIXME unroll */\ | |
1347 __asm__ volatile(\ | |
1348 "pxor %%mm7, %%mm7 \n\t"\ | |
1349 "1: \n\t"\ | |
1350 "movq (%0), %%mm0 \n\t"\ | |
1351 "movq (%0), %%mm1 \n\t"\ | |
1352 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1353 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1354 "movq %%mm0, (%1) \n\t"\ | |
1355 "movq %%mm1, 9*8(%1) \n\t"\ | |
1356 "add $8, %1 \n\t"\ | |
1357 "add %3, %0 \n\t"\ | |
1358 "decl %2 \n\t"\ | |
1359 " jnz 1b \n\t"\ | |
1360 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ | |
1361 : "r" ((x86_reg)srcStride)\ | |
1362 : "memory"\ | |
1363 );\ | |
1364 \ | |
1365 temp_ptr= temp;\ | |
1366 count=2;\ | |
1367 \ | |
1368 /*FIXME reorder for speed */\ | |
1369 __asm__ volatile(\ | |
1370 /*"pxor %%mm7, %%mm7 \n\t"*/\ | |
1371 "1: \n\t"\ | |
1372 "movq (%0), %%mm0 \n\t"\ | |
1373 "movq 8(%0), %%mm1 \n\t"\ | |
1374 "movq 16(%0), %%mm2 \n\t"\ | |
1375 "movq 24(%0), %%mm3 \n\t"\ | |
1376 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ | |
1377 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
1378 "add %4, %1 \n\t"\ | |
1379 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ | |
1380 \ | |
1381 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ | |
1382 "add %4, %1 \n\t"\ | |
1383 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ | |
1384 \ | |
1385 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ | |
1386 "add %4, %1 \n\t"\ | |
1387 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ | |
1388 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
1389 \ | |
1390 "add $72, %0 \n\t"\ | |
1391 "add %6, %1 \n\t"\ | |
1392 "decl %2 \n\t"\ | |
1393 " jnz 1b \n\t"\ | |
1394 \ | |
1395 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ | |
1396 : "r"((x86_reg)dstStride), "r"(2*(x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(x86_reg)dstStride)\ | |
1397 : "memory"\ | |
1398 );\ | |
1399 }\ | |
1400 \ | |
1401 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ | |
1402 OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\ | |
1403 }\ | |
1404 \ | |
1405 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1406 uint64_t temp[8];\ | |
1407 uint8_t * const half= (uint8_t*)temp;\ | |
1408 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
1409 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ | |
1410 }\ | |
1411 \ | |
1412 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1413 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ | |
1414 }\ | |
1415 \ | |
1416 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1417 uint64_t temp[8];\ | |
1418 uint8_t * const half= (uint8_t*)temp;\ | |
1419 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
1420 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\ | |
1421 }\ | |
1422 \ | |
1423 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1424 uint64_t temp[8];\ | |
1425 uint8_t * const half= (uint8_t*)temp;\ | |
1426 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ | |
1427 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ | |
1428 }\ | |
1429 \ | |
1430 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1431 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1432 }\ | |
1433 \ | |
1434 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1435 uint64_t temp[8];\ | |
1436 uint8_t * const half= (uint8_t*)temp;\ | |
1437 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ | |
1438 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\ | |
1439 }\ | |
1440 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1441 uint64_t half[8 + 9];\ | |
1442 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1443 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1444 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1445 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ | |
1446 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1447 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ | |
1448 }\ | |
1449 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1450 uint64_t half[8 + 9];\ | |
1451 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1452 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1453 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1454 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ | |
1455 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1456 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ | |
1457 }\ | |
1458 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1459 uint64_t half[8 + 9];\ | |
1460 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1461 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1462 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1463 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ | |
1464 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1465 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ | |
1466 }\ | |
1467 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1468 uint64_t half[8 + 9];\ | |
1469 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1470 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1471 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1472 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ | |
1473 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1474 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ | |
1475 }\ | |
1476 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1477 uint64_t half[8 + 9];\ | |
1478 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1479 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1480 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1481 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1482 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ | |
1483 }\ | |
1484 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1485 uint64_t half[8 + 9];\ | |
1486 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1487 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1488 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1489 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1490 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ | |
1491 }\ | |
1492 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1493 uint64_t half[8 + 9];\ | |
1494 uint8_t * const halfH= ((uint8_t*)half);\ | |
1495 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1496 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ | |
1497 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
1498 }\ | |
1499 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1500 uint64_t half[8 + 9];\ | |
1501 uint8_t * const halfH= ((uint8_t*)half);\ | |
1502 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1503 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ | |
1504 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
1505 }\ | |
1506 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1507 uint64_t half[9];\ | |
1508 uint8_t * const halfH= ((uint8_t*)half);\ | |
1509 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1510 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
1511 }\ | |
1512 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ | |
1513 OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\ | |
1514 }\ | |
1515 \ | |
1516 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1517 uint64_t temp[32];\ | |
1518 uint8_t * const half= (uint8_t*)temp;\ | |
1519 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
1520 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ | |
1521 }\ | |
1522 \ | |
1523 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1524 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ | |
1525 }\ | |
1526 \ | |
1527 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1528 uint64_t temp[32];\ | |
1529 uint8_t * const half= (uint8_t*)temp;\ | |
1530 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
1531 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\ | |
1532 }\ | |
1533 \ | |
1534 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1535 uint64_t temp[32];\ | |
1536 uint8_t * const half= (uint8_t*)temp;\ | |
1537 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ | |
1538 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ | |
1539 }\ | |
1540 \ | |
1541 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1542 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1543 }\ | |
1544 \ | |
1545 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1546 uint64_t temp[32];\ | |
1547 uint8_t * const half= (uint8_t*)temp;\ | |
1548 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ | |
1549 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\ | |
1550 }\ | |
1551 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1552 uint64_t half[16*2 + 17*2];\ | |
1553 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1554 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1555 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1556 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ | |
1557 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1558 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ | |
1559 }\ | |
1560 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1561 uint64_t half[16*2 + 17*2];\ | |
1562 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1563 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1564 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1565 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ | |
1566 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1567 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ | |
1568 }\ | |
1569 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1570 uint64_t half[16*2 + 17*2];\ | |
1571 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1572 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1573 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1574 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ | |
1575 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1576 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ | |
1577 }\ | |
1578 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1579 uint64_t half[16*2 + 17*2];\ | |
1580 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1581 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1582 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1583 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ | |
1584 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1585 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ | |
1586 }\ | |
1587 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1588 uint64_t half[16*2 + 17*2];\ | |
1589 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1590 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1591 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1592 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1593 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ | |
1594 }\ | |
1595 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1596 uint64_t half[16*2 + 17*2];\ | |
1597 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1598 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1599 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1600 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1601 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ | |
1602 }\ | |
1603 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1604 uint64_t half[17*2];\ | |
1605 uint8_t * const halfH= ((uint8_t*)half);\ | |
1606 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1607 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ | |
1608 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
1609 }\ | |
1610 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1611 uint64_t half[17*2];\ | |
1612 uint8_t * const halfH= ((uint8_t*)half);\ | |
1613 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1614 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ | |
1615 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
1616 }\ | |
1617 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1618 uint64_t half[17*2];\ | |
1619 uint8_t * const halfH= ((uint8_t*)half);\ | |
1620 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1621 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
1622 } | |
1623 | |
1624 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" | |
1625 #define AVG_3DNOW_OP(a,b,temp, size) \ | |
1626 "mov" #size " " #b ", " #temp " \n\t"\ | |
1627 "pavgusb " #temp ", " #a " \n\t"\ | |
1628 "mov" #size " " #a ", " #b " \n\t" | |
1629 #define AVG_MMX2_OP(a,b,temp, size) \ | |
1630 "mov" #size " " #b ", " #temp " \n\t"\ | |
1631 "pavgb " #temp ", " #a " \n\t"\ | |
1632 "mov" #size " " #a ", " #b " \n\t" | |
1633 | |
1634 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
1635 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
1636 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
1637 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
1638 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
1639 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
1640 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) | |
1641 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) | |
1642 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) | |
1643 | |
1644 /***********************************/ | |
1645 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */ | |
1646 | |
1647 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\ | |
1648 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1649 OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\ | |
1650 } | |
1651 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\ | |
1652 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1653 OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\ | |
1654 } | |
1655 | |
1656 #define QPEL_2TAP(OPNAME, SIZE, MMX)\ | |
1657 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\ | |
1658 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\ | |
1659 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\ | |
1660 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\ | |
1661 OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\ | |
1662 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\ | |
1663 OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\ | |
1664 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\ | |
1665 OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\ | |
1666 static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1667 OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\ | |
1668 }\ | |
1669 static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1670 OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\ | |
1671 }\ | |
1672 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0, 1, 0)\ | |
1673 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1, -1, 0)\ | |
1674 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0, stride, 0)\ | |
1675 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride, -stride, 0)\ | |
1676 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0, stride, 1)\ | |
1677 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1, stride, -1)\ | |
1678 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride, -stride, 1)\ | |
1679 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\ | |
1680 | |
1681 QPEL_2TAP(put_, 16, mmx2) | |
1682 QPEL_2TAP(avg_, 16, mmx2) | |
1683 QPEL_2TAP(put_, 8, mmx2) | |
1684 QPEL_2TAP(avg_, 8, mmx2) | |
1685 QPEL_2TAP(put_, 16, 3dnow) | |
1686 QPEL_2TAP(avg_, 16, 3dnow) | |
1687 QPEL_2TAP(put_, 8, 3dnow) | |
1688 QPEL_2TAP(avg_, 8, 3dnow) | |
1689 | |
1690 | |
1691 #if 0 | |
8527
f8bf438c6000
Add missing 'void' keyword to parameterless function declarations.
diego
parents:
8519
diff
changeset
|
1692 static void just_return(void) { return; } |
8430 | 1693 #endif |
1694 | |
1695 static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy, | |
1696 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){ | |
1697 const int w = 8; | |
1698 const int ix = ox>>(16+shift); | |
1699 const int iy = oy>>(16+shift); | |
1700 const int oxs = ox>>4; | |
1701 const int oys = oy>>4; | |
1702 const int dxxs = dxx>>4; | |
1703 const int dxys = dxy>>4; | |
1704 const int dyxs = dyx>>4; | |
1705 const int dyys = dyy>>4; | |
1706 const uint16_t r4[4] = {r,r,r,r}; | |
1707 const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys}; | |
1708 const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys}; | |
1709 const uint64_t shift2 = 2*shift; | |
1710 uint8_t edge_buf[(h+1)*stride]; | |
1711 int x, y; | |
1712 | |
1713 const int dxw = (dxx-(1<<(16+shift)))*(w-1); | |
1714 const int dyh = (dyy-(1<<(16+shift)))*(h-1); | |
1715 const int dxh = dxy*(h-1); | |
1716 const int dyw = dyx*(w-1); | |
1717 if( // non-constant fullpel offset (3% of blocks) | |
1718 ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) | | |
1719 (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift) | |
1720 // uses more than 16 bits of subpel mv (only at huge resolution) | |
1721 || (dxx|dxy|dyx|dyy)&15 ) | |
1722 { | |
1723 //FIXME could still use mmx for some of the rows | |
1724 ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height); | |
1725 return; | |
1726 } | |
1727 | |
1728 src += ix + iy*stride; | |
1729 if( (unsigned)ix >= width-w || | |
1730 (unsigned)iy >= height-h ) | |
1731 { | |
1732 ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height); | |
1733 src = edge_buf; | |
1734 } | |
1735 | |
1736 __asm__ volatile( | |
1737 "movd %0, %%mm6 \n\t" | |
1738 "pxor %%mm7, %%mm7 \n\t" | |
1739 "punpcklwd %%mm6, %%mm6 \n\t" | |
1740 "punpcklwd %%mm6, %%mm6 \n\t" | |
1741 :: "r"(1<<shift) | |
1742 ); | |
1743 | |
1744 for(x=0; x<w; x+=4){ | |
1745 uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0), | |
1746 oxs - dxys + dxxs*(x+1), | |
1747 oxs - dxys + dxxs*(x+2), | |
1748 oxs - dxys + dxxs*(x+3) }; | |
1749 uint16_t dy4[4] = { oys - dyys + dyxs*(x+0), | |
1750 oys - dyys + dyxs*(x+1), | |
1751 oys - dyys + dyxs*(x+2), | |
1752 oys - dyys + dyxs*(x+3) }; | |
1753 | |
1754 for(y=0; y<h; y++){ | |
1755 __asm__ volatile( | |
1756 "movq %0, %%mm4 \n\t" | |
1757 "movq %1, %%mm5 \n\t" | |
1758 "paddw %2, %%mm4 \n\t" | |
1759 "paddw %3, %%mm5 \n\t" | |
1760 "movq %%mm4, %0 \n\t" | |
1761 "movq %%mm5, %1 \n\t" | |
1762 "psrlw $12, %%mm4 \n\t" | |
1763 "psrlw $12, %%mm5 \n\t" | |
1764 : "+m"(*dx4), "+m"(*dy4) | |
1765 : "m"(*dxy4), "m"(*dyy4) | |
1766 ); | |
1767 | |
1768 __asm__ volatile( | |
1769 "movq %%mm6, %%mm2 \n\t" | |
1770 "movq %%mm6, %%mm1 \n\t" | |
1771 "psubw %%mm4, %%mm2 \n\t" | |
1772 "psubw %%mm5, %%mm1 \n\t" | |
1773 "movq %%mm2, %%mm0 \n\t" | |
1774 "movq %%mm4, %%mm3 \n\t" | |
1775 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy) | |
1776 "pmullw %%mm5, %%mm3 \n\t" // dx*dy | |
1777 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy | |
1778 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy) | |
1779 | |
1780 "movd %4, %%mm5 \n\t" | |
1781 "movd %3, %%mm4 \n\t" | |
1782 "punpcklbw %%mm7, %%mm5 \n\t" | |
1783 "punpcklbw %%mm7, %%mm4 \n\t" | |
1784 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy | |
1785 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy | |
1786 | |
1787 "movd %2, %%mm5 \n\t" | |
1788 "movd %1, %%mm4 \n\t" | |
1789 "punpcklbw %%mm7, %%mm5 \n\t" | |
1790 "punpcklbw %%mm7, %%mm4 \n\t" | |
1791 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy) | |
1792 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy) | |
1793 "paddw %5, %%mm1 \n\t" | |
1794 "paddw %%mm3, %%mm2 \n\t" | |
1795 "paddw %%mm1, %%mm0 \n\t" | |
1796 "paddw %%mm2, %%mm0 \n\t" | |
1797 | |
1798 "psrlw %6, %%mm0 \n\t" | |
1799 "packuswb %%mm0, %%mm0 \n\t" | |
1800 "movd %%mm0, %0 \n\t" | |
1801 | |
1802 : "=m"(dst[x+y*stride]) | |
1803 : "m"(src[0]), "m"(src[1]), | |
1804 "m"(src[stride]), "m"(src[stride+1]), | |
1805 "m"(*r4), "m"(shift2) | |
1806 ); | |
1807 src += stride; | |
1808 } | |
1809 src += 4-h*stride; | |
1810 } | |
1811 } | |
1812 | |
1813 #define PREFETCH(name, op) \ | |
1814 static void name(void *mem, int stride, int h){\ | |
1815 const uint8_t *p= mem;\ | |
1816 do{\ | |
1817 __asm__ volatile(#op" %0" :: "m"(*p));\ | |
1818 p+= stride;\ | |
1819 }while(--h);\ | |
1820 } | |
1821 PREFETCH(prefetch_mmx2, prefetcht0) | |
1822 PREFETCH(prefetch_3dnow, prefetch) | |
1823 #undef PREFETCH | |
1824 | |
1825 #include "h264dsp_mmx.c" | |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
1826 #include "rv40dsp_mmx.c" |
8430 | 1827 |
1828 /* CAVS specific */ | |
1829 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1830 put_pixels8_mmx(dst, src, stride, 8); | |
1831 } | |
1832 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1833 avg_pixels8_mmx(dst, src, stride, 8); | |
1834 } | |
1835 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1836 put_pixels16_mmx(dst, src, stride, 16); | |
1837 } | |
1838 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1839 avg_pixels16_mmx(dst, src, stride, 16); | |
1840 } | |
1841 | |
1842 /* VC1 specific */ | |
1843 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { | |
1844 put_pixels8_mmx(dst, src, stride, 8); | |
1845 } | |
9441 | 1846 void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) { |
1847 avg_pixels8_mmx2(dst, src, stride, 8); | |
1848 } | |
8430 | 1849 |
1850 /* XXX: those functions should be suppressed ASAP when all IDCTs are | |
1851 converted */ | |
8590 | 1852 #if CONFIG_GPL |
8430 | 1853 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) |
1854 { | |
1855 ff_mmx_idct (block); | |
1856 put_pixels_clamped_mmx(block, dest, line_size); | |
1857 } | |
1858 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1859 { | |
1860 ff_mmx_idct (block); | |
1861 add_pixels_clamped_mmx(block, dest, line_size); | |
1862 } | |
1863 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1864 { | |
1865 ff_mmxext_idct (block); | |
1866 put_pixels_clamped_mmx(block, dest, line_size); | |
1867 } | |
1868 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1869 { | |
1870 ff_mmxext_idct (block); | |
1871 add_pixels_clamped_mmx(block, dest, line_size); | |
1872 } | |
1873 #endif | |
1874 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1875 { | |
1876 ff_idct_xvid_mmx (block); | |
1877 put_pixels_clamped_mmx(block, dest, line_size); | |
1878 } | |
1879 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1880 { | |
1881 ff_idct_xvid_mmx (block); | |
1882 add_pixels_clamped_mmx(block, dest, line_size); | |
1883 } | |
1884 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1885 { | |
1886 ff_idct_xvid_mmx2 (block); | |
1887 put_pixels_clamped_mmx(block, dest, line_size); | |
1888 } | |
1889 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1890 { | |
1891 ff_idct_xvid_mmx2 (block); | |
1892 add_pixels_clamped_mmx(block, dest, line_size); | |
1893 } | |
1894 | |
1895 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize) | |
1896 { | |
1897 int i; | |
1898 __asm__ volatile("pxor %%mm7, %%mm7":); | |
1899 for(i=0; i<blocksize; i+=2) { | |
1900 __asm__ volatile( | |
1901 "movq %0, %%mm0 \n\t" | |
1902 "movq %1, %%mm1 \n\t" | |
1903 "movq %%mm0, %%mm2 \n\t" | |
1904 "movq %%mm1, %%mm3 \n\t" | |
1905 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0 | |
1906 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0 | |
1907 "pslld $31, %%mm2 \n\t" // keep only the sign bit | |
1908 "pxor %%mm2, %%mm1 \n\t" | |
1909 "movq %%mm3, %%mm4 \n\t" | |
1910 "pand %%mm1, %%mm3 \n\t" | |
1911 "pandn %%mm1, %%mm4 \n\t" | |
1912 "pfadd %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m))) | |
1913 "pfsub %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m))) | |
1914 "movq %%mm3, %1 \n\t" | |
1915 "movq %%mm0, %0 \n\t" | |
1916 :"+m"(mag[i]), "+m"(ang[i]) | |
1917 ::"memory" | |
1918 ); | |
1919 } | |
1920 __asm__ volatile("femms"); | |
1921 } | |
1922 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize) | |
1923 { | |
1924 int i; | |
1925 | |
1926 __asm__ volatile( | |
1927 "movaps %0, %%xmm5 \n\t" | |
1928 ::"m"(ff_pdw_80000000[0]) | |
1929 ); | |
1930 for(i=0; i<blocksize; i+=4) { | |
1931 __asm__ volatile( | |
1932 "movaps %0, %%xmm0 \n\t" | |
1933 "movaps %1, %%xmm1 \n\t" | |
1934 "xorps %%xmm2, %%xmm2 \n\t" | |
1935 "xorps %%xmm3, %%xmm3 \n\t" | |
1936 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0 | |
1937 "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0 | |
1938 "andps %%xmm5, %%xmm2 \n\t" // keep only the sign bit | |
1939 "xorps %%xmm2, %%xmm1 \n\t" | |
1940 "movaps %%xmm3, %%xmm4 \n\t" | |
1941 "andps %%xmm1, %%xmm3 \n\t" | |
1942 "andnps %%xmm1, %%xmm4 \n\t" | |
1943 "addps %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m))) | |
1944 "subps %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m))) | |
1945 "movaps %%xmm3, %1 \n\t" | |
1946 "movaps %%xmm0, %0 \n\t" | |
1947 :"+m"(mag[i]), "+m"(ang[i]) | |
1948 ::"memory" | |
1949 ); | |
1950 } | |
1951 } | |
1952 | |
1953 #define IF1(x) x | |
1954 #define IF0(x) | |
1955 | |
1956 #define MIX5(mono,stereo)\ | |
1957 __asm__ volatile(\ | |
1958 "movss 0(%2), %%xmm5 \n"\ | |
1959 "movss 8(%2), %%xmm6 \n"\ | |
1960 "movss 24(%2), %%xmm7 \n"\ | |
1961 "shufps $0, %%xmm5, %%xmm5 \n"\ | |
1962 "shufps $0, %%xmm6, %%xmm6 \n"\ | |
1963 "shufps $0, %%xmm7, %%xmm7 \n"\ | |
1964 "1: \n"\ | |
1965 "movaps (%0,%1), %%xmm0 \n"\ | |
1966 "movaps 0x400(%0,%1), %%xmm1 \n"\ | |
1967 "movaps 0x800(%0,%1), %%xmm2 \n"\ | |
1968 "movaps 0xc00(%0,%1), %%xmm3 \n"\ | |
1969 "movaps 0x1000(%0,%1), %%xmm4 \n"\ | |
1970 "mulps %%xmm5, %%xmm0 \n"\ | |
1971 "mulps %%xmm6, %%xmm1 \n"\ | |
1972 "mulps %%xmm5, %%xmm2 \n"\ | |
1973 "mulps %%xmm7, %%xmm3 \n"\ | |
1974 "mulps %%xmm7, %%xmm4 \n"\ | |
1975 stereo("addps %%xmm1, %%xmm0 \n")\ | |
1976 "addps %%xmm1, %%xmm2 \n"\ | |
1977 "addps %%xmm3, %%xmm0 \n"\ | |
1978 "addps %%xmm4, %%xmm2 \n"\ | |
1979 mono("addps %%xmm2, %%xmm0 \n")\ | |
1980 "movaps %%xmm0, (%0,%1) \n"\ | |
1981 stereo("movaps %%xmm2, 0x400(%0,%1) \n")\ | |
1982 "add $16, %0 \n"\ | |
1983 "jl 1b \n"\ | |
1984 :"+&r"(i)\ | |
1985 :"r"(samples[0]+len), "r"(matrix)\ | |
1986 :"memory"\ | |
1987 ); | |
1988 | |
1989 #define MIX_MISC(stereo)\ | |
1990 __asm__ volatile(\ | |
1991 "1: \n"\ | |
1992 "movaps (%3,%0), %%xmm0 \n"\ | |
1993 stereo("movaps %%xmm0, %%xmm1 \n")\ | |
1994 "mulps %%xmm6, %%xmm0 \n"\ | |
1995 stereo("mulps %%xmm7, %%xmm1 \n")\ | |
1996 "lea 1024(%3,%0), %1 \n"\ | |
1997 "mov %5, %2 \n"\ | |
1998 "2: \n"\ | |
1999 "movaps (%1), %%xmm2 \n"\ | |
2000 stereo("movaps %%xmm2, %%xmm3 \n")\ | |
2001 "mulps (%4,%2), %%xmm2 \n"\ | |
2002 stereo("mulps 16(%4,%2), %%xmm3 \n")\ | |
2003 "addps %%xmm2, %%xmm0 \n"\ | |
2004 stereo("addps %%xmm3, %%xmm1 \n")\ | |
2005 "add $1024, %1 \n"\ | |
2006 "add $32, %2 \n"\ | |
2007 "jl 2b \n"\ | |
2008 "movaps %%xmm0, (%3,%0) \n"\ | |
2009 stereo("movaps %%xmm1, 1024(%3,%0) \n")\ | |
2010 "add $16, %0 \n"\ | |
2011 "jl 1b \n"\ | |
2012 :"+&r"(i), "=&r"(j), "=&r"(k)\ | |
2013 :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\ | |
2014 :"memory"\ | |
2015 ); | |
2016 | |
2017 static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len) | |
2018 { | |
2019 int (*matrix_cmp)[2] = (int(*)[2])matrix; | |
2020 intptr_t i,j,k; | |
2021 | |
2022 i = -len*sizeof(float); | |
2023 if(in_ch == 5 && out_ch == 2 && !(matrix_cmp[0][1]|matrix_cmp[2][0]|matrix_cmp[3][1]|matrix_cmp[4][0]|(matrix_cmp[1][0]^matrix_cmp[1][1])|(matrix_cmp[0][0]^matrix_cmp[2][1]))) { | |
2024 MIX5(IF0,IF1); | |
2025 } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) { | |
2026 MIX5(IF1,IF0); | |
2027 } else { | |
11369 | 2028 DECLARE_ALIGNED(16, float, matrix_simd)[in_ch][2][4]; |
8430 | 2029 j = 2*in_ch*sizeof(float); |
2030 __asm__ volatile( | |
2031 "1: \n" | |
2032 "sub $8, %0 \n" | |
2033 "movss (%2,%0), %%xmm6 \n" | |
2034 "movss 4(%2,%0), %%xmm7 \n" | |
2035 "shufps $0, %%xmm6, %%xmm6 \n" | |
2036 "shufps $0, %%xmm7, %%xmm7 \n" | |
2037 "movaps %%xmm6, (%1,%0,4) \n" | |
2038 "movaps %%xmm7, 16(%1,%0,4) \n" | |
2039 "jg 1b \n" | |
2040 :"+&r"(j) | |
2041 :"r"(matrix_simd), "r"(matrix) | |
2042 :"memory" | |
2043 ); | |
2044 if(out_ch == 2) { | |
2045 MIX_MISC(IF1); | |
2046 } else { | |
2047 MIX_MISC(IF0); | |
2048 } | |
2049 } | |
2050 } | |
2051 | |
2052 static void vector_fmul_3dnow(float *dst, const float *src, int len){ | |
2053 x86_reg i = (len-4)*4; | |
2054 __asm__ volatile( | |
2055 "1: \n\t" | |
2056 "movq (%1,%0), %%mm0 \n\t" | |
2057 "movq 8(%1,%0), %%mm1 \n\t" | |
2058 "pfmul (%2,%0), %%mm0 \n\t" | |
2059 "pfmul 8(%2,%0), %%mm1 \n\t" | |
2060 "movq %%mm0, (%1,%0) \n\t" | |
2061 "movq %%mm1, 8(%1,%0) \n\t" | |
2062 "sub $16, %0 \n\t" | |
2063 "jge 1b \n\t" | |
2064 "femms \n\t" | |
2065 :"+r"(i) | |
2066 :"r"(dst), "r"(src) | |
2067 :"memory" | |
2068 ); | |
2069 } | |
2070 static void vector_fmul_sse(float *dst, const float *src, int len){ | |
2071 x86_reg i = (len-8)*4; | |
2072 __asm__ volatile( | |
2073 "1: \n\t" | |
2074 "movaps (%1,%0), %%xmm0 \n\t" | |
2075 "movaps 16(%1,%0), %%xmm1 \n\t" | |
2076 "mulps (%2,%0), %%xmm0 \n\t" | |
2077 "mulps 16(%2,%0), %%xmm1 \n\t" | |
2078 "movaps %%xmm0, (%1,%0) \n\t" | |
2079 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2080 "sub $32, %0 \n\t" | |
2081 "jge 1b \n\t" | |
2082 :"+r"(i) | |
2083 :"r"(dst), "r"(src) | |
2084 :"memory" | |
2085 ); | |
2086 } | |
2087 | |
2088 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){ | |
2089 x86_reg i = len*4-16; | |
2090 __asm__ volatile( | |
2091 "1: \n\t" | |
2092 "pswapd 8(%1), %%mm0 \n\t" | |
2093 "pswapd (%1), %%mm1 \n\t" | |
2094 "pfmul (%3,%0), %%mm0 \n\t" | |
2095 "pfmul 8(%3,%0), %%mm1 \n\t" | |
2096 "movq %%mm0, (%2,%0) \n\t" | |
2097 "movq %%mm1, 8(%2,%0) \n\t" | |
2098 "add $16, %1 \n\t" | |
2099 "sub $16, %0 \n\t" | |
2100 "jge 1b \n\t" | |
2101 :"+r"(i), "+r"(src1) | |
2102 :"r"(dst), "r"(src0) | |
2103 ); | |
2104 __asm__ volatile("femms"); | |
2105 } | |
2106 static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){ | |
2107 x86_reg i = len*4-32; | |
2108 __asm__ volatile( | |
2109 "1: \n\t" | |
2110 "movaps 16(%1), %%xmm0 \n\t" | |
2111 "movaps (%1), %%xmm1 \n\t" | |
2112 "shufps $0x1b, %%xmm0, %%xmm0 \n\t" | |
2113 "shufps $0x1b, %%xmm1, %%xmm1 \n\t" | |
2114 "mulps (%3,%0), %%xmm0 \n\t" | |
2115 "mulps 16(%3,%0), %%xmm1 \n\t" | |
2116 "movaps %%xmm0, (%2,%0) \n\t" | |
2117 "movaps %%xmm1, 16(%2,%0) \n\t" | |
2118 "add $32, %1 \n\t" | |
2119 "sub $32, %0 \n\t" | |
2120 "jge 1b \n\t" | |
2121 :"+r"(i), "+r"(src1) | |
2122 :"r"(dst), "r"(src0) | |
2123 ); | |
2124 } | |
2125 | |
10300
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2126 static void vector_fmul_add_3dnow(float *dst, const float *src0, const float *src1, |
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2127 const float *src2, int len){ |
8430 | 2128 x86_reg i = (len-4)*4; |
10301 | 2129 __asm__ volatile( |
2130 "1: \n\t" | |
2131 "movq (%2,%0), %%mm0 \n\t" | |
2132 "movq 8(%2,%0), %%mm1 \n\t" | |
2133 "pfmul (%3,%0), %%mm0 \n\t" | |
2134 "pfmul 8(%3,%0), %%mm1 \n\t" | |
2135 "pfadd (%4,%0), %%mm0 \n\t" | |
2136 "pfadd 8(%4,%0), %%mm1 \n\t" | |
2137 "movq %%mm0, (%1,%0) \n\t" | |
2138 "movq %%mm1, 8(%1,%0) \n\t" | |
2139 "sub $16, %0 \n\t" | |
2140 "jge 1b \n\t" | |
2141 :"+r"(i) | |
2142 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
2143 :"memory" | |
2144 ); | |
8430 | 2145 __asm__ volatile("femms"); |
2146 } | |
10300
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2147 static void vector_fmul_add_sse(float *dst, const float *src0, const float *src1, |
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2148 const float *src2, int len){ |
8430 | 2149 x86_reg i = (len-8)*4; |
10301 | 2150 __asm__ volatile( |
2151 "1: \n\t" | |
2152 "movaps (%2,%0), %%xmm0 \n\t" | |
2153 "movaps 16(%2,%0), %%xmm1 \n\t" | |
2154 "mulps (%3,%0), %%xmm0 \n\t" | |
2155 "mulps 16(%3,%0), %%xmm1 \n\t" | |
2156 "addps (%4,%0), %%xmm0 \n\t" | |
2157 "addps 16(%4,%0), %%xmm1 \n\t" | |
2158 "movaps %%xmm0, (%1,%0) \n\t" | |
2159 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2160 "sub $32, %0 \n\t" | |
2161 "jge 1b \n\t" | |
2162 :"+r"(i) | |
2163 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
2164 :"memory" | |
2165 ); | |
8430 | 2166 } |
2167 | |
2168 static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1, | |
2169 const float *win, float add_bias, int len){ | |
8590 | 2170 #if HAVE_6REGS |
8430 | 2171 if(add_bias == 0){ |
2172 x86_reg i = -len*4; | |
2173 x86_reg j = len*4-8; | |
2174 __asm__ volatile( | |
2175 "1: \n" | |
2176 "pswapd (%5,%1), %%mm1 \n" | |
2177 "movq (%5,%0), %%mm0 \n" | |
2178 "pswapd (%4,%1), %%mm5 \n" | |
2179 "movq (%3,%0), %%mm4 \n" | |
2180 "movq %%mm0, %%mm2 \n" | |
2181 "movq %%mm1, %%mm3 \n" | |
2182 "pfmul %%mm4, %%mm2 \n" // src0[len+i]*win[len+i] | |
2183 "pfmul %%mm5, %%mm3 \n" // src1[ j]*win[len+j] | |
2184 "pfmul %%mm4, %%mm1 \n" // src0[len+i]*win[len+j] | |
2185 "pfmul %%mm5, %%mm0 \n" // src1[ j]*win[len+i] | |
2186 "pfadd %%mm3, %%mm2 \n" | |
2187 "pfsub %%mm0, %%mm1 \n" | |
2188 "pswapd %%mm2, %%mm2 \n" | |
2189 "movq %%mm1, (%2,%0) \n" | |
2190 "movq %%mm2, (%2,%1) \n" | |
2191 "sub $8, %1 \n" | |
2192 "add $8, %0 \n" | |
2193 "jl 1b \n" | |
2194 "femms \n" | |
2195 :"+r"(i), "+r"(j) | |
2196 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len) | |
2197 ); | |
2198 }else | |
2199 #endif | |
2200 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len); | |
2201 } | |
2202 | |
2203 static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1, | |
2204 const float *win, float add_bias, int len){ | |
8590 | 2205 #if HAVE_6REGS |
8430 | 2206 if(add_bias == 0){ |
2207 x86_reg i = -len*4; | |
2208 x86_reg j = len*4-16; | |
2209 __asm__ volatile( | |
2210 "1: \n" | |
2211 "movaps (%5,%1), %%xmm1 \n" | |
2212 "movaps (%5,%0), %%xmm0 \n" | |
2213 "movaps (%4,%1), %%xmm5 \n" | |
2214 "movaps (%3,%0), %%xmm4 \n" | |
2215 "shufps $0x1b, %%xmm1, %%xmm1 \n" | |
2216 "shufps $0x1b, %%xmm5, %%xmm5 \n" | |
2217 "movaps %%xmm0, %%xmm2 \n" | |
2218 "movaps %%xmm1, %%xmm3 \n" | |
2219 "mulps %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i] | |
2220 "mulps %%xmm5, %%xmm3 \n" // src1[ j]*win[len+j] | |
2221 "mulps %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j] | |
2222 "mulps %%xmm5, %%xmm0 \n" // src1[ j]*win[len+i] | |
2223 "addps %%xmm3, %%xmm2 \n" | |
2224 "subps %%xmm0, %%xmm1 \n" | |
2225 "shufps $0x1b, %%xmm2, %%xmm2 \n" | |
2226 "movaps %%xmm1, (%2,%0) \n" | |
2227 "movaps %%xmm2, (%2,%1) \n" | |
2228 "sub $16, %1 \n" | |
2229 "add $16, %0 \n" | |
2230 "jl 1b \n" | |
2231 :"+r"(i), "+r"(j) | |
2232 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len) | |
2233 ); | |
2234 }else | |
2235 #endif | |
2236 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len); | |
2237 } | |
2238 | |
2239 static void int32_to_float_fmul_scalar_sse(float *dst, const int *src, float mul, int len) | |
2240 { | |
2241 x86_reg i = -4*len; | |
2242 __asm__ volatile( | |
2243 "movss %3, %%xmm4 \n" | |
2244 "shufps $0, %%xmm4, %%xmm4 \n" | |
2245 "1: \n" | |
2246 "cvtpi2ps (%2,%0), %%xmm0 \n" | |
2247 "cvtpi2ps 8(%2,%0), %%xmm1 \n" | |
2248 "cvtpi2ps 16(%2,%0), %%xmm2 \n" | |
2249 "cvtpi2ps 24(%2,%0), %%xmm3 \n" | |
2250 "movlhps %%xmm1, %%xmm0 \n" | |
2251 "movlhps %%xmm3, %%xmm2 \n" | |
2252 "mulps %%xmm4, %%xmm0 \n" | |
2253 "mulps %%xmm4, %%xmm2 \n" | |
2254 "movaps %%xmm0, (%1,%0) \n" | |
2255 "movaps %%xmm2, 16(%1,%0) \n" | |
2256 "add $32, %0 \n" | |
2257 "jl 1b \n" | |
2258 :"+r"(i) | |
2259 :"r"(dst+len), "r"(src+len), "m"(mul) | |
2260 ); | |
2261 } | |
2262 | |
2263 static void int32_to_float_fmul_scalar_sse2(float *dst, const int *src, float mul, int len) | |
2264 { | |
2265 x86_reg i = -4*len; | |
2266 __asm__ volatile( | |
2267 "movss %3, %%xmm4 \n" | |
2268 "shufps $0, %%xmm4, %%xmm4 \n" | |
2269 "1: \n" | |
2270 "cvtdq2ps (%2,%0), %%xmm0 \n" | |
2271 "cvtdq2ps 16(%2,%0), %%xmm1 \n" | |
2272 "mulps %%xmm4, %%xmm0 \n" | |
2273 "mulps %%xmm4, %%xmm1 \n" | |
2274 "movaps %%xmm0, (%1,%0) \n" | |
2275 "movaps %%xmm1, 16(%1,%0) \n" | |
2276 "add $32, %0 \n" | |
2277 "jl 1b \n" | |
2278 :"+r"(i) | |
2279 :"r"(dst+len), "r"(src+len), "m"(mul) | |
2280 ); | |
2281 } | |
2282 | |
10105 | 2283 static void vector_clipf_sse(float *dst, const float *src, float min, float max, |
10104
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2284 int len) |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2285 { |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2286 x86_reg i = (len-16)*4; |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2287 __asm__ volatile( |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2288 "movss %3, %%xmm4 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2289 "movss %4, %%xmm5 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2290 "shufps $0, %%xmm4, %%xmm4 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2291 "shufps $0, %%xmm5, %%xmm5 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2292 "1: \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2293 "movaps (%2,%0), %%xmm0 \n\t" // 3/1 on intel |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2294 "movaps 16(%2,%0), %%xmm1 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2295 "movaps 32(%2,%0), %%xmm2 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2296 "movaps 48(%2,%0), %%xmm3 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2297 "maxps %%xmm4, %%xmm0 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2298 "maxps %%xmm4, %%xmm1 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2299 "maxps %%xmm4, %%xmm2 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2300 "maxps %%xmm4, %%xmm3 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2301 "minps %%xmm5, %%xmm0 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2302 "minps %%xmm5, %%xmm1 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2303 "minps %%xmm5, %%xmm2 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2304 "minps %%xmm5, %%xmm3 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2305 "movaps %%xmm0, (%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2306 "movaps %%xmm1, 16(%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2307 "movaps %%xmm2, 32(%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2308 "movaps %%xmm3, 48(%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2309 "sub $64, %0 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2310 "jge 1b \n\t" |
10107
3b61bc6ce377
Mark "i" parameter of vector_clipf_sse() as early-clobber
vitor
parents:
10105
diff
changeset
|
2311 :"+&r"(i) |
10104
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2312 :"r"(dst), "r"(src), "m"(min), "m"(max) |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2313 :"memory" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2314 ); |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2315 } |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2316 |
8430 | 2317 static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){ |
2318 x86_reg reglen = len; | |
2319 // not bit-exact: pf2id uses different rounding than C and SSE | |
2320 __asm__ volatile( | |
2321 "add %0 , %0 \n\t" | |
2322 "lea (%2,%0,2) , %2 \n\t" | |
2323 "add %0 , %1 \n\t" | |
2324 "neg %0 \n\t" | |
2325 "1: \n\t" | |
2326 "pf2id (%2,%0,2) , %%mm0 \n\t" | |
2327 "pf2id 8(%2,%0,2) , %%mm1 \n\t" | |
2328 "pf2id 16(%2,%0,2) , %%mm2 \n\t" | |
2329 "pf2id 24(%2,%0,2) , %%mm3 \n\t" | |
2330 "packssdw %%mm1 , %%mm0 \n\t" | |
2331 "packssdw %%mm3 , %%mm2 \n\t" | |
2332 "movq %%mm0 , (%1,%0) \n\t" | |
2333 "movq %%mm2 , 8(%1,%0) \n\t" | |
2334 "add $16 , %0 \n\t" | |
2335 " js 1b \n\t" | |
2336 "femms \n\t" | |
2337 :"+r"(reglen), "+r"(dst), "+r"(src) | |
2338 ); | |
2339 } | |
2340 static void float_to_int16_sse(int16_t *dst, const float *src, long len){ | |
2341 x86_reg reglen = len; | |
2342 __asm__ volatile( | |
2343 "add %0 , %0 \n\t" | |
2344 "lea (%2,%0,2) , %2 \n\t" | |
2345 "add %0 , %1 \n\t" | |
2346 "neg %0 \n\t" | |
2347 "1: \n\t" | |
2348 "cvtps2pi (%2,%0,2) , %%mm0 \n\t" | |
2349 "cvtps2pi 8(%2,%0,2) , %%mm1 \n\t" | |
2350 "cvtps2pi 16(%2,%0,2) , %%mm2 \n\t" | |
2351 "cvtps2pi 24(%2,%0,2) , %%mm3 \n\t" | |
2352 "packssdw %%mm1 , %%mm0 \n\t" | |
2353 "packssdw %%mm3 , %%mm2 \n\t" | |
2354 "movq %%mm0 , (%1,%0) \n\t" | |
2355 "movq %%mm2 , 8(%1,%0) \n\t" | |
2356 "add $16 , %0 \n\t" | |
2357 " js 1b \n\t" | |
2358 "emms \n\t" | |
2359 :"+r"(reglen), "+r"(dst), "+r"(src) | |
2360 ); | |
2361 } | |
2362 | |
2363 static void float_to_int16_sse2(int16_t *dst, const float *src, long len){ | |
2364 x86_reg reglen = len; | |
2365 __asm__ volatile( | |
2366 "add %0 , %0 \n\t" | |
2367 "lea (%2,%0,2) , %2 \n\t" | |
2368 "add %0 , %1 \n\t" | |
2369 "neg %0 \n\t" | |
2370 "1: \n\t" | |
2371 "cvtps2dq (%2,%0,2) , %%xmm0 \n\t" | |
2372 "cvtps2dq 16(%2,%0,2) , %%xmm1 \n\t" | |
2373 "packssdw %%xmm1 , %%xmm0 \n\t" | |
2374 "movdqa %%xmm0 , (%1,%0) \n\t" | |
2375 "add $16 , %0 \n\t" | |
2376 " js 1b \n\t" | |
2377 :"+r"(reglen), "+r"(dst), "+r"(src) | |
2378 ); | |
2379 } | |
2380 | |
2381 void ff_float_to_int16_interleave6_sse(int16_t *dst, const float **src, int len); | |
2382 void ff_float_to_int16_interleave6_3dnow(int16_t *dst, const float **src, int len); | |
2383 void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len); | |
11981 | 2384 int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2, int order, int shift); |
2385 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order, int shift); | |
2386 int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul); | |
2387 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul); | |
2388 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul); | |
10431 | 2389 void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top); |
2390 int ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left); | |
2391 int ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left); | |
8430 | 2392 void ff_x264_deblock_v_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); |
2393 void ff_x264_deblock_h_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); | |
10645 | 2394 void ff_x264_deblock_h_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta); |
2395 void ff_x264_deblock_v_luma_intra_sse2(uint8_t *pix, int stride, int alpha, int beta); | |
2396 void ff_x264_deblock_h_luma_intra_sse2(uint8_t *pix, int stride, int alpha, int beta); | |
2397 | |
2398 #if HAVE_YASM && ARCH_X86_32 | |
8430 | 2399 void ff_x264_deblock_v8_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta); |
2400 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta) | |
2401 { | |
2402 ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta); | |
2403 ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta); | |
2404 } | |
10645 | 2405 #elif !HAVE_YASM |
8430 | 2406 #define ff_float_to_int16_interleave6_sse(a,b,c) float_to_int16_interleave_misc_sse(a,b,c,6) |
2407 #define ff_float_to_int16_interleave6_3dnow(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6) | |
2408 #define ff_float_to_int16_interleave6_3dn2(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6) | |
2409 #endif | |
2410 #define ff_float_to_int16_interleave6_sse2 ff_float_to_int16_interleave6_sse | |
2411 | |
2412 #define FLOAT_TO_INT16_INTERLEAVE(cpu, body) \ | |
2413 /* gcc pessimizes register allocation if this is in the same function as float_to_int16_interleave_sse2*/\ | |
2414 static av_noinline void float_to_int16_interleave_misc_##cpu(int16_t *dst, const float **src, long len, int channels){\ | |
11369 | 2415 DECLARE_ALIGNED(16, int16_t, tmp)[len];\ |
8430 | 2416 int i,j,c;\ |
2417 for(c=0; c<channels; c++){\ | |
2418 float_to_int16_##cpu(tmp, src[c], len);\ | |
2419 for(i=0, j=c; i<len; i++, j+=channels)\ | |
2420 dst[j] = tmp[i];\ | |
2421 }\ | |
2422 }\ | |
2423 \ | |
2424 static void float_to_int16_interleave_##cpu(int16_t *dst, const float **src, long len, int channels){\ | |
2425 if(channels==1)\ | |
2426 float_to_int16_##cpu(dst, src[0], len);\ | |
2427 else if(channels==2){\ | |
2428 x86_reg reglen = len; \ | |
2429 const float *src0 = src[0];\ | |
2430 const float *src1 = src[1];\ | |
2431 __asm__ volatile(\ | |
2432 "shl $2, %0 \n"\ | |
2433 "add %0, %1 \n"\ | |
2434 "add %0, %2 \n"\ | |
2435 "add %0, %3 \n"\ | |
2436 "neg %0 \n"\ | |
2437 body\ | |
2438 :"+r"(reglen), "+r"(dst), "+r"(src0), "+r"(src1)\ | |
2439 );\ | |
2440 }else if(channels==6){\ | |
2441 ff_float_to_int16_interleave6_##cpu(dst, src, len);\ | |
2442 }else\ | |
2443 float_to_int16_interleave_misc_##cpu(dst, src, len, channels);\ | |
2444 } | |
2445 | |
2446 FLOAT_TO_INT16_INTERLEAVE(3dnow, | |
2447 "1: \n" | |
2448 "pf2id (%2,%0), %%mm0 \n" | |
2449 "pf2id 8(%2,%0), %%mm1 \n" | |
2450 "pf2id (%3,%0), %%mm2 \n" | |
2451 "pf2id 8(%3,%0), %%mm3 \n" | |
2452 "packssdw %%mm1, %%mm0 \n" | |
2453 "packssdw %%mm3, %%mm2 \n" | |
2454 "movq %%mm0, %%mm1 \n" | |
2455 "punpcklwd %%mm2, %%mm0 \n" | |
2456 "punpckhwd %%mm2, %%mm1 \n" | |
2457 "movq %%mm0, (%1,%0)\n" | |
2458 "movq %%mm1, 8(%1,%0)\n" | |
2459 "add $16, %0 \n" | |
2460 "js 1b \n" | |
2461 "femms \n" | |
2462 ) | |
2463 | |
2464 FLOAT_TO_INT16_INTERLEAVE(sse, | |
2465 "1: \n" | |
2466 "cvtps2pi (%2,%0), %%mm0 \n" | |
2467 "cvtps2pi 8(%2,%0), %%mm1 \n" | |
2468 "cvtps2pi (%3,%0), %%mm2 \n" | |
2469 "cvtps2pi 8(%3,%0), %%mm3 \n" | |
2470 "packssdw %%mm1, %%mm0 \n" | |
2471 "packssdw %%mm3, %%mm2 \n" | |
2472 "movq %%mm0, %%mm1 \n" | |
2473 "punpcklwd %%mm2, %%mm0 \n" | |
2474 "punpckhwd %%mm2, %%mm1 \n" | |
2475 "movq %%mm0, (%1,%0)\n" | |
2476 "movq %%mm1, 8(%1,%0)\n" | |
2477 "add $16, %0 \n" | |
2478 "js 1b \n" | |
2479 "emms \n" | |
2480 ) | |
2481 | |
2482 FLOAT_TO_INT16_INTERLEAVE(sse2, | |
2483 "1: \n" | |
2484 "cvtps2dq (%2,%0), %%xmm0 \n" | |
2485 "cvtps2dq (%3,%0), %%xmm1 \n" | |
2486 "packssdw %%xmm1, %%xmm0 \n" | |
2487 "movhlps %%xmm0, %%xmm1 \n" | |
2488 "punpcklwd %%xmm1, %%xmm0 \n" | |
2489 "movdqa %%xmm0, (%1,%0) \n" | |
2490 "add $16, %0 \n" | |
2491 "js 1b \n" | |
2492 ) | |
2493 | |
2494 static void float_to_int16_interleave_3dn2(int16_t *dst, const float **src, long len, int channels){ | |
2495 if(channels==6) | |
2496 ff_float_to_int16_interleave6_3dn2(dst, src, len); | |
2497 else | |
2498 float_to_int16_interleave_3dnow(dst, src, len, channels); | |
2499 } | |
2500 | |
10964 | 2501 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order); |
2502 | |
8430 | 2503 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
2504 { | |
12414 | 2505 int mm_flags = mm_support(); |
8430 | 2506 |
2507 if (avctx->dsp_mask) { | |
2508 if (avctx->dsp_mask & FF_MM_FORCE) | |
2509 mm_flags |= (avctx->dsp_mask & 0xffff); | |
2510 else | |
2511 mm_flags &= ~(avctx->dsp_mask & 0xffff); | |
2512 } | |
2513 | |
2514 #if 0 | |
2515 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:"); | |
2516 if (mm_flags & FF_MM_MMX) | |
2517 av_log(avctx, AV_LOG_INFO, " mmx"); | |
9342
7f594601d5e9
Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents:
9341
diff
changeset
|
2518 if (mm_flags & FF_MM_MMX2) |
7f594601d5e9
Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents:
9341
diff
changeset
|
2519 av_log(avctx, AV_LOG_INFO, " mmx2"); |
8430 | 2520 if (mm_flags & FF_MM_3DNOW) |
2521 av_log(avctx, AV_LOG_INFO, " 3dnow"); | |
2522 if (mm_flags & FF_MM_SSE) | |
2523 av_log(avctx, AV_LOG_INFO, " sse"); | |
2524 if (mm_flags & FF_MM_SSE2) | |
2525 av_log(avctx, AV_LOG_INFO, " sse2"); | |
2526 av_log(avctx, AV_LOG_INFO, "\n"); | |
2527 #endif | |
2528 | |
2529 if (mm_flags & FF_MM_MMX) { | |
2530 const int idct_algo= avctx->idct_algo; | |
2531 | |
2532 if(avctx->lowres==0){ | |
2533 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
2534 c->idct_put= ff_simple_idct_put_mmx; | |
2535 c->idct_add= ff_simple_idct_add_mmx; | |
2536 c->idct = ff_simple_idct_mmx; | |
2537 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM; | |
8590 | 2538 #if CONFIG_GPL |
8430 | 2539 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){ |
9342
7f594601d5e9
Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents:
9341
diff
changeset
|
2540 if(mm_flags & FF_MM_MMX2){ |
8430 | 2541 c->idct_put= ff_libmpeg2mmx2_idct_put; |
2542 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
2543 c->idct = ff_mmxext_idct; | |
2544 }else{ | |
2545 c->idct_put= ff_libmpeg2mmx_idct_put; | |
2546 c->idct_add= ff_libmpeg2mmx_idct_add; | |
2547 c->idct = ff_mmx_idct; | |
2548 } | |
2549 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM; | |
2550 #endif | |
9975
d6d7e8d4a04d
Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents:
9959
diff
changeset
|
2551 }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) && |
8430 | 2552 idct_algo==FF_IDCT_VP3){ |
2553 if(mm_flags & FF_MM_SSE2){ | |
2554 c->idct_put= ff_vp3_idct_put_sse2; | |
2555 c->idct_add= ff_vp3_idct_add_sse2; | |
2556 c->idct = ff_vp3_idct_sse2; | |
2557 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2558 }else{ | |
2559 c->idct_put= ff_vp3_idct_put_mmx; | |
2560 c->idct_add= ff_vp3_idct_add_mmx; | |
2561 c->idct = ff_vp3_idct_mmx; | |
2562 c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM; | |
2563 } | |
2564 }else if(idct_algo==FF_IDCT_CAVS){ | |
2565 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2566 }else if(idct_algo==FF_IDCT_XVIDMMX){ | |
2567 if(mm_flags & FF_MM_SSE2){ | |
2568 c->idct_put= ff_idct_xvid_sse2_put; | |
2569 c->idct_add= ff_idct_xvid_sse2_add; | |
2570 c->idct = ff_idct_xvid_sse2; | |
2571 c->idct_permutation_type= FF_SSE2_IDCT_PERM; | |
9342
7f594601d5e9
Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents:
9341
diff
changeset
|
2572 }else if(mm_flags & FF_MM_MMX2){ |
8430 | 2573 c->idct_put= ff_idct_xvid_mmx2_put; |
2574 c->idct_add= ff_idct_xvid_mmx2_add; | |
2575 c->idct = ff_idct_xvid_mmx2; | |
2576 }else{ | |
2577 c->idct_put= ff_idct_xvid_mmx_put; | |
2578 c->idct_add= ff_idct_xvid_mmx_add; | |
2579 c->idct = ff_idct_xvid_mmx; | |
2580 } | |
2581 } | |
2582 } | |
2583 | |
2584 c->put_pixels_clamped = put_pixels_clamped_mmx; | |
2585 c->put_signed_pixels_clamped = put_signed_pixels_clamped_mmx; | |
2586 c->add_pixels_clamped = add_pixels_clamped_mmx; | |
2587 c->clear_block = clear_block_mmx; | |
2588 c->clear_blocks = clear_blocks_mmx; | |
10766
78c2be62260a
Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents:
10749
diff
changeset
|
2589 if ((mm_flags & FF_MM_SSE) && |
78c2be62260a
Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents:
10749
diff
changeset
|
2590 !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){ |
78c2be62260a
Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents:
10749
diff
changeset
|
2591 /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */ |
9861 | 2592 c->clear_block = clear_block_sse; |
2593 c->clear_blocks = clear_blocks_sse; | |
2594 } | |
8430 | 2595 |
2596 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \ | |
2597 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \ | |
2598 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \ | |
2599 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \ | |
2600 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU | |
2601 | |
2602 SET_HPEL_FUNCS(put, 0, 16, mmx); | |
2603 SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx); | |
2604 SET_HPEL_FUNCS(avg, 0, 16, mmx); | |
2605 SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx); | |
2606 SET_HPEL_FUNCS(put, 1, 8, mmx); | |
2607 SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx); | |
2608 SET_HPEL_FUNCS(avg, 1, 8, mmx); | |
2609 SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx); | |
2610 | |
2611 c->gmc= gmc_mmx; | |
2612 | |
2613 c->add_bytes= add_bytes_mmx; | |
2614 c->add_bytes_l2= add_bytes_l2_mmx; | |
2615 | |
2616 c->draw_edges = draw_edges_mmx; | |
2617 | |
10749
5cca4b6c459d
Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents:
10645
diff
changeset
|
2618 if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { |
8430 | 2619 c->h263_v_loop_filter= h263_v_loop_filter_mmx; |
2620 c->h263_h_loop_filter= h263_h_loop_filter_mmx; | |
2621 } | |
2622 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_rnd; | |
2623 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_mmx; | |
9439
ef3a7b711cc0
Rename put_no_rnd_h264_chroma* to reflect its usage in VC1 only
conrad
parents:
9342
diff
changeset
|
2624 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_vc1_chroma_mc8_mmx_nornd; |
8430 | 2625 |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2626 c->put_rv40_chroma_pixels_tab[0]= put_rv40_chroma_mc8_mmx; |
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2627 c->put_rv40_chroma_pixels_tab[1]= put_rv40_chroma_mc4_mmx; |
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2628 |
8817 | 2629 if (CONFIG_VP6_DECODER) { |
2630 c->vp6_filter_diag4 = ff_vp6_filter_diag4_mmx; | |
2631 } | |
2632 | |
9342
7f594601d5e9
Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents:
9341
diff
changeset
|
2633 if (mm_flags & FF_MM_MMX2) { |
8430 | 2634 c->prefetch = prefetch_mmx2; |
2635 | |
2636 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2; | |
2637 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; | |
2638 | |
2639 c->avg_pixels_tab[0][0] = avg_pixels16_mmx2; | |
2640 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2; | |
2641 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; | |
2642 | |
2643 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2; | |
2644 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; | |
2645 | |
2646 c->avg_pixels_tab[1][0] = avg_pixels8_mmx2; | |
2647 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2; | |
2648 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; | |
2649 | |
2650 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2651 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
2652 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
2653 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
2654 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
2655 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
2656 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; | |
2657 | |
9975
d6d7e8d4a04d
Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents:
9959
diff
changeset
|
2658 if (CONFIG_VP3_DECODER) { |
8430 | 2659 c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2; |
2660 c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2; | |
2661 } | |
2662 } | |
11637 | 2663 if (CONFIG_VP3_DECODER) { |
2664 c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2; | |
2665 } | |
8430 | 2666 |
11826
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2667 if (CONFIG_VP3_DECODER |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2668 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) { |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2669 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_mmx2; |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2670 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_mmx2; |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2671 } |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2672 |
8430 | 2673 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \ |
2674 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \ | |
2675 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \ | |
2676 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \ | |
2677 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \ | |
2678 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \ | |
2679 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \ | |
2680 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \ | |
2681 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \ | |
2682 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \ | |
2683 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \ | |
2684 c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \ | |
2685 c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \ | |
2686 c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \ | |
2687 c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \ | |
2688 c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \ | |
2689 c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU | |
2690 | |
2691 SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2); | |
2692 SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2); | |
2693 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2); | |
2694 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2); | |
2695 SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2); | |
2696 SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2); | |
2697 | |
2698 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2); | |
2699 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2); | |
2700 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2); | |
2701 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2); | |
2702 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2); | |
2703 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2); | |
2704 | |
2705 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2); | |
2706 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2); | |
2707 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2); | |
2708 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2); | |
2709 | |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2710 c->avg_rv40_chroma_pixels_tab[0]= avg_rv40_chroma_mc8_mmx2; |
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2711 c->avg_rv40_chroma_pixels_tab[1]= avg_rv40_chroma_mc4_mmx2; |
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2712 |
9440 | 2713 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_vc1_chroma_mc8_mmx2_nornd; |
2714 | |
8430 | 2715 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2_rnd; |
2716 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2; | |
2717 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2; | |
2718 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2; | |
2719 | |
8760 | 2720 #if HAVE_YASM |
2721 c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2; | |
2722 #endif | |
8798
a5c8210814d7
Add check whether the compiler/assembler supports 10 or more operands.
diego
parents:
8760
diff
changeset
|
2723 #if HAVE_7REGS && HAVE_TEN_OPERANDS |
8760 | 2724 if( mm_flags&FF_MM_3DNOW ) |
2725 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov; | |
2726 #endif | |
2727 | |
9995
3141f69e3905
Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents:
9975
diff
changeset
|
2728 if (CONFIG_VC1_DECODER) |
8430 | 2729 ff_vc1dsp_init_mmx(c, avctx); |
2730 | |
2731 c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2; | |
2732 } else if (mm_flags & FF_MM_3DNOW) { | |
2733 c->prefetch = prefetch_3dnow; | |
2734 | |
2735 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow; | |
2736 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; | |
2737 | |
2738 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow; | |
2739 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow; | |
2740 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; | |
2741 | |
2742 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow; | |
2743 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow; | |
2744 | |
2745 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow; | |
2746 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow; | |
2747 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; | |
2748 | |
2749 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2750 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
2751 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
2752 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
2753 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
2754 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
2755 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
2756 } | |
2757 | |
11826
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2758 if (CONFIG_VP3_DECODER |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2759 && (avctx->codec_id == CODEC_ID_VP3 || avctx->codec_id == CODEC_ID_THEORA)) { |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2760 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_exact_3dnow; |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2761 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_exact_3dnow; |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2762 } |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2763 |
8430 | 2764 SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow); |
2765 SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow); | |
2766 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow); | |
2767 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow); | |
2768 SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow); | |
2769 SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow); | |
2770 | |
2771 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow); | |
2772 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow); | |
2773 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow); | |
2774 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow); | |
2775 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow); | |
2776 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow); | |
2777 | |
2778 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow); | |
2779 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow); | |
2780 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow); | |
2781 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow); | |
2782 | |
2783 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_3dnow_rnd; | |
2784 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow; | |
2785 | |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2786 c->avg_rv40_chroma_pixels_tab[0]= avg_rv40_chroma_mc8_3dnow; |
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2787 c->avg_rv40_chroma_pixels_tab[1]= avg_rv40_chroma_mc4_3dnow; |
8430 | 2788 } |
2789 | |
2790 | |
2791 #define H264_QPEL_FUNCS(x, y, CPU)\ | |
2792 c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\ | |
2793 c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\ | |
2794 c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\ | |
2795 c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU; | |
2796 if((mm_flags & FF_MM_SSE2) && !(mm_flags & FF_MM_3DNOW)){ | |
2797 // these functions are slower than mmx on AMD, but faster on Intel | |
2798 c->put_pixels_tab[0][0] = put_pixels16_sse2; | |
2799 c->avg_pixels_tab[0][0] = avg_pixels16_sse2; | |
2800 H264_QPEL_FUNCS(0, 0, sse2); | |
2801 } | |
2802 if(mm_flags & FF_MM_SSE2){ | |
2803 H264_QPEL_FUNCS(0, 1, sse2); | |
2804 H264_QPEL_FUNCS(0, 2, sse2); | |
2805 H264_QPEL_FUNCS(0, 3, sse2); | |
2806 H264_QPEL_FUNCS(1, 1, sse2); | |
2807 H264_QPEL_FUNCS(1, 2, sse2); | |
2808 H264_QPEL_FUNCS(1, 3, sse2); | |
2809 H264_QPEL_FUNCS(2, 1, sse2); | |
2810 H264_QPEL_FUNCS(2, 2, sse2); | |
2811 H264_QPEL_FUNCS(2, 3, sse2); | |
2812 H264_QPEL_FUNCS(3, 1, sse2); | |
2813 H264_QPEL_FUNCS(3, 2, sse2); | |
2814 H264_QPEL_FUNCS(3, 3, sse2); | |
8818 | 2815 |
2816 if (CONFIG_VP6_DECODER) { | |
2817 c->vp6_filter_diag4 = ff_vp6_filter_diag4_sse2; | |
2818 } | |
8430 | 2819 } |
8590 | 2820 #if HAVE_SSSE3 |
8430 | 2821 if(mm_flags & FF_MM_SSSE3){ |
2822 H264_QPEL_FUNCS(1, 0, ssse3); | |
2823 H264_QPEL_FUNCS(1, 1, ssse3); | |
2824 H264_QPEL_FUNCS(1, 2, ssse3); | |
2825 H264_QPEL_FUNCS(1, 3, ssse3); | |
2826 H264_QPEL_FUNCS(2, 0, ssse3); | |
2827 H264_QPEL_FUNCS(2, 1, ssse3); | |
2828 H264_QPEL_FUNCS(2, 2, ssse3); | |
2829 H264_QPEL_FUNCS(2, 3, ssse3); | |
2830 H264_QPEL_FUNCS(3, 0, ssse3); | |
2831 H264_QPEL_FUNCS(3, 1, ssse3); | |
2832 H264_QPEL_FUNCS(3, 2, ssse3); | |
2833 H264_QPEL_FUNCS(3, 3, ssse3); | |
9439
ef3a7b711cc0
Rename put_no_rnd_h264_chroma* to reflect its usage in VC1 only
conrad
parents:
9342
diff
changeset
|
2834 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_vc1_chroma_mc8_ssse3_nornd; |
9440 | 2835 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_vc1_chroma_mc8_ssse3_nornd; |
8430 | 2836 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_ssse3_rnd; |
2837 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_ssse3_rnd; | |
2838 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_ssse3; | |
2839 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_ssse3; | |
2840 c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3; | |
10430 | 2841 #if HAVE_YASM |
2842 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3; | |
2843 if (mm_flags & FF_MM_SSE4) // not really sse4, just slow on Conroe | |
2844 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4; | |
2845 #endif | |
8430 | 2846 } |
2847 #endif | |
2848 | |
2849 if(mm_flags & FF_MM_3DNOW){ | |
2850 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow; | |
2851 c->vector_fmul = vector_fmul_3dnow; | |
2852 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2853 c->float_to_int16 = float_to_int16_3dnow; | |
2854 c->float_to_int16_interleave = float_to_int16_interleave_3dnow; | |
2855 } | |
2856 } | |
2857 if(mm_flags & FF_MM_3DNOWEXT){ | |
2858 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; | |
2859 c->vector_fmul_window = vector_fmul_window_3dnow2; | |
2860 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2861 c->float_to_int16_interleave = float_to_int16_interleave_3dn2; | |
2862 } | |
2863 } | |
10633 | 2864 if(mm_flags & FF_MM_MMX2){ |
2865 #if HAVE_YASM | |
2866 c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2; | |
10644 | 2867 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2; |
10633 | 2868 #endif |
2869 } | |
8430 | 2870 if(mm_flags & FF_MM_SSE){ |
2871 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse; | |
2872 c->ac3_downmix = ac3_downmix_sse; | |
2873 c->vector_fmul = vector_fmul_sse; | |
2874 c->vector_fmul_reverse = vector_fmul_reverse_sse; | |
10300
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2875 c->vector_fmul_add = vector_fmul_add_sse; |
8430 | 2876 c->vector_fmul_window = vector_fmul_window_sse; |
2877 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse; | |
10104
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2878 c->vector_clipf = vector_clipf_sse; |
8430 | 2879 c->float_to_int16 = float_to_int16_sse; |
2880 c->float_to_int16_interleave = float_to_int16_interleave_sse; | |
10964 | 2881 #if HAVE_YASM |
2882 c->scalarproduct_float = ff_scalarproduct_float_sse; | |
2883 #endif | |
8430 | 2884 } |
2885 if(mm_flags & FF_MM_3DNOW) | |
10300
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2886 c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse |
8430 | 2887 if(mm_flags & FF_MM_SSE2){ |
2888 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2; | |
2889 c->float_to_int16 = float_to_int16_sse2; | |
2890 c->float_to_int16_interleave = float_to_int16_interleave_sse2; | |
10633 | 2891 #if HAVE_YASM |
2892 c->scalarproduct_int16 = ff_scalarproduct_int16_sse2; | |
10644 | 2893 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2; |
10633 | 2894 #endif |
8430 | 2895 } |
10644 | 2896 if((mm_flags & FF_MM_SSSE3) && !(mm_flags & (FF_MM_SSE42|FF_MM_3DNOW)) && HAVE_YASM) // cachesplit |
2897 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3; | |
8430 | 2898 } |
2899 | |
8596
68e959302527
replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents:
8590
diff
changeset
|
2900 if (CONFIG_ENCODERS) |
8430 | 2901 dsputilenc_init_mmx(c, avctx); |
2902 | |
2903 #if 0 | |
2904 // for speed testing | |
2905 get_pixels = just_return; | |
2906 put_pixels_clamped = just_return; | |
2907 add_pixels_clamped = just_return; | |
2908 | |
2909 pix_abs16x16 = just_return; | |
2910 pix_abs16x16_x2 = just_return; | |
2911 pix_abs16x16_y2 = just_return; | |
2912 pix_abs16x16_xy2 = just_return; | |
2913 | |
2914 put_pixels_tab[0] = just_return; | |
2915 put_pixels_tab[1] = just_return; | |
2916 put_pixels_tab[2] = just_return; | |
2917 put_pixels_tab[3] = just_return; | |
2918 | |
2919 put_no_rnd_pixels_tab[0] = just_return; | |
2920 put_no_rnd_pixels_tab[1] = just_return; | |
2921 put_no_rnd_pixels_tab[2] = just_return; | |
2922 put_no_rnd_pixels_tab[3] = just_return; | |
2923 | |
2924 avg_pixels_tab[0] = just_return; | |
2925 avg_pixels_tab[1] = just_return; | |
2926 avg_pixels_tab[2] = just_return; | |
2927 avg_pixels_tab[3] = just_return; | |
2928 | |
2929 avg_no_rnd_pixels_tab[0] = just_return; | |
2930 avg_no_rnd_pixels_tab[1] = just_return; | |
2931 avg_no_rnd_pixels_tab[2] = just_return; | |
2932 avg_no_rnd_pixels_tab[3] = just_return; | |
2933 | |
2934 //av_fdct = just_return; | |
2935 //ff_idct = just_return; | |
2936 #endif | |
2937 } | |
11499 | 2938 |
2939 #if CONFIG_H264DSP | |
2940 void ff_h264dsp_init_x86(H264DSPContext *c) | |
2941 { | |
12414 | 2942 int mm_flags = mm_support(); |
11499 | 2943 |
2944 if (mm_flags & FF_MM_MMX) { | |
2945 c->h264_idct_dc_add= | |
2946 c->h264_idct_add= ff_h264_idct_add_mmx; | |
2947 c->h264_idct8_dc_add= | |
2948 c->h264_idct8_add= ff_h264_idct8_add_mmx; | |
2949 | |
2950 c->h264_idct_add16 = ff_h264_idct_add16_mmx; | |
2951 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx; | |
2952 c->h264_idct_add8 = ff_h264_idct_add8_mmx; | |
2953 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx; | |
2954 | |
2955 if (mm_flags & FF_MM_MMX2) { | |
2956 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2; | |
2957 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; | |
2958 c->h264_idct_add16 = ff_h264_idct_add16_mmx2; | |
2959 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx2; | |
2960 c->h264_idct_add8 = ff_h264_idct_add8_mmx2; | |
2961 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2; | |
2962 | |
2963 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2; | |
2964 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2; | |
2965 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2; | |
2966 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2; | |
2967 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_mmx2; | |
2968 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_mmx2; | |
2969 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; | |
2970 | |
2971 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2; | |
2972 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2; | |
2973 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2; | |
2974 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2; | |
2975 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2; | |
2976 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2; | |
2977 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2; | |
2978 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2; | |
2979 | |
2980 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2; | |
2981 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2; | |
2982 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2; | |
2983 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2; | |
2984 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2; | |
2985 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2; | |
2986 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2; | |
2987 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2; | |
2988 } | |
2989 if(mm_flags & FF_MM_SSE2){ | |
2990 c->h264_idct8_add = ff_h264_idct8_add_sse2; | |
2991 c->h264_idct8_add4= ff_h264_idct8_add4_sse2; | |
2992 } | |
2993 | |
12213 | 2994 #if HAVE_YASM |
11499 | 2995 if (mm_flags & FF_MM_MMX2){ |
2996 #if ARCH_X86_32 | |
2997 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext; | |
2998 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext; | |
2999 #endif | |
3000 if( mm_flags&FF_MM_SSE2 ){ | |
12366 | 3001 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2; |
3002 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2; | |
11499 | 3003 #if ARCH_X86_64 || !defined(__ICC) || __ICC > 1110 |
3004 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2; | |
3005 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2; | |
3006 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2; | |
3007 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2; | |
3008 #endif | |
12213 | 3009 #if CONFIG_GPL |
11499 | 3010 c->h264_idct_add16 = ff_h264_idct_add16_sse2; |
3011 c->h264_idct_add8 = ff_h264_idct_add8_sse2; | |
3012 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2; | |
12213 | 3013 #endif |
11499 | 3014 } |
12366 | 3015 if ( mm_flags&FF_MM_SSSE3 ){ |
3016 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3; | |
3017 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3; | |
3018 } | |
11499 | 3019 } |
3020 #endif | |
3021 } | |
3022 } | |
3023 #endif /* CONFIG_H264DSP */ |