Mercurial > libavcodec.hg
annotate x86/dsputil_mmx.c @ 12205:d38e8565ba05 libavcodec
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
and chroma (width=8).
author | rbultje |
---|---|
date | Tue, 20 Jul 2010 22:58:56 +0000 |
parents | b246b214c2e9 |
children | fe243bb5ef61 |
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 int mm_flags; /* multimedia extension flags */ | |
41 | |
42 /* pixel operations */ | |
11369 | 43 DECLARE_ALIGNED(8, const uint64_t, ff_bone) = 0x0101010101010101ULL; |
44 DECLARE_ALIGNED(8, const uint64_t, ff_wtwo) = 0x0002000200020002ULL; | |
8430 | 45 |
11369 | 46 DECLARE_ALIGNED(16, const uint64_t, ff_pdw_80000000)[2] = |
8430 | 47 {0x8000000080000000ULL, 0x8000000080000000ULL}; |
48 | |
11369 | 49 DECLARE_ALIGNED(8, const uint64_t, ff_pw_3 ) = 0x0003000300030003ULL; |
12143 | 50 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_4 ) = {0x0004000400040004ULL, 0x0004000400040004ULL}; |
11369 | 51 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_5 ) = {0x0005000500050005ULL, 0x0005000500050005ULL}; |
52 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
|
53 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_9 ) = {0x0009000900090009ULL, 0x0009000900090009ULL}; |
11369 | 54 DECLARE_ALIGNED(8, const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL; |
55 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
|
56 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_18 ) = {0x0012001200120012ULL, 0x0012001200120012ULL}; |
11369 | 57 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
|
58 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_27 ) = {0x001B001B001B001BULL, 0x001B001B001B001BULL}; |
11369 | 59 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_28 ) = {0x001C001C001C001CULL, 0x001C001C001C001CULL}; |
60 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL}; | |
61 DECLARE_ALIGNED(8, const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL; | |
12205
d38e8565ba05
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16)
rbultje
parents:
12168
diff
changeset
|
62 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_63 ) = {0x003F003F003F003FULL, 0x003F003F003F003FULL}; |
11369 | 63 DECLARE_ALIGNED(16, const xmm_reg, ff_pw_64 ) = {0x0040004000400040ULL, 0x0040004000400040ULL}; |
64 DECLARE_ALIGNED(8, const uint64_t, ff_pw_96 ) = 0x0060006000600060ULL; | |
65 DECLARE_ALIGNED(8, const uint64_t, ff_pw_128) = 0x0080008000800080ULL; | |
66 DECLARE_ALIGNED(8, const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL; | |
8430 | 67 |
12168
b246b214c2e9
VP8 H/V inner loopfilter MMX/MMXEXT/SSE2 optimizations.
rbultje
parents:
12143
diff
changeset
|
68 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
|
69 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
|
70 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_4 ) = {0x0404040404040404ULL, 0x0404040404040404ULL}; |
11369 | 71 DECLARE_ALIGNED(8, const uint64_t, ff_pb_7 ) = 0x0707070707070707ULL; |
72 DECLARE_ALIGNED(8, const uint64_t, ff_pb_1F ) = 0x1F1F1F1F1F1F1F1FULL; | |
73 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
|
74 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_80 ) = {0x8080808080808080ULL, 0x8080808080808080ULL}; |
11369 | 75 DECLARE_ALIGNED(8, const uint64_t, ff_pb_81 ) = 0x8181818181818181ULL; |
76 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
|
77 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_F8 ) = {0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL}; |
11369 | 78 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
|
79 DECLARE_ALIGNED(16, const xmm_reg, ff_pb_FE ) = {0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL}; |
8430 | 80 |
11369 | 81 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 }; |
82 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 }; | |
8430 | 83 |
84 #define JUMPALIGN() __asm__ volatile (ASMALIGN(3)::) | |
85 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%" #regd ", %%" #regd ::) | |
86 | |
87 #define MOVQ_BFE(regd) \ | |
88 __asm__ volatile ( \ | |
89 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ | |
90 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
91 | |
92 #ifndef PIC | |
93 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone)) | |
94 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo)) | |
95 #else | |
96 // for shared library it's better to use this way for accessing constants | |
97 // pcmpeqd -> -1 | |
98 #define MOVQ_BONE(regd) \ | |
99 __asm__ volatile ( \ | |
100 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
101 "psrlw $15, %%" #regd " \n\t" \ | |
102 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
103 | |
104 #define MOVQ_WTWO(regd) \ | |
105 __asm__ volatile ( \ | |
106 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
107 "psrlw $15, %%" #regd " \n\t" \ | |
108 "psllw $1, %%" #regd " \n\t"::) | |
109 | |
110 #endif | |
111 | |
112 // using regr as temporary and for the output result | |
113 // first argument is unmodifed and second is trashed | |
114 // regfe is supposed to contain 0xfefefefefefefefe | |
115 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
116 "movq " #rega ", " #regr " \n\t"\ | |
117 "pand " #regb ", " #regr " \n\t"\ | |
118 "pxor " #rega ", " #regb " \n\t"\ | |
119 "pand " #regfe "," #regb " \n\t"\ | |
120 "psrlq $1, " #regb " \n\t"\ | |
121 "paddb " #regb ", " #regr " \n\t" | |
122 | |
123 #define PAVGB_MMX(rega, regb, regr, regfe) \ | |
124 "movq " #rega ", " #regr " \n\t"\ | |
125 "por " #regb ", " #regr " \n\t"\ | |
126 "pxor " #rega ", " #regb " \n\t"\ | |
127 "pand " #regfe "," #regb " \n\t"\ | |
128 "psrlq $1, " #regb " \n\t"\ | |
129 "psubb " #regb ", " #regr " \n\t" | |
130 | |
131 // mm6 is supposed to contain 0xfefefefefefefefe | |
132 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ | |
133 "movq " #rega ", " #regr " \n\t"\ | |
134 "movq " #regc ", " #regp " \n\t"\ | |
135 "pand " #regb ", " #regr " \n\t"\ | |
136 "pand " #regd ", " #regp " \n\t"\ | |
137 "pxor " #rega ", " #regb " \n\t"\ | |
138 "pxor " #regc ", " #regd " \n\t"\ | |
139 "pand %%mm6, " #regb " \n\t"\ | |
140 "pand %%mm6, " #regd " \n\t"\ | |
141 "psrlq $1, " #regb " \n\t"\ | |
142 "psrlq $1, " #regd " \n\t"\ | |
143 "paddb " #regb ", " #regr " \n\t"\ | |
144 "paddb " #regd ", " #regp " \n\t" | |
145 | |
146 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ | |
147 "movq " #rega ", " #regr " \n\t"\ | |
148 "movq " #regc ", " #regp " \n\t"\ | |
149 "por " #regb ", " #regr " \n\t"\ | |
150 "por " #regd ", " #regp " \n\t"\ | |
151 "pxor " #rega ", " #regb " \n\t"\ | |
152 "pxor " #regc ", " #regd " \n\t"\ | |
153 "pand %%mm6, " #regb " \n\t"\ | |
154 "pand %%mm6, " #regd " \n\t"\ | |
155 "psrlq $1, " #regd " \n\t"\ | |
156 "psrlq $1, " #regb " \n\t"\ | |
157 "psubb " #regb ", " #regr " \n\t"\ | |
158 "psubb " #regd ", " #regp " \n\t" | |
159 | |
160 /***********************************/ | |
161 /* MMX no rounding */ | |
162 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx | |
163 #define SET_RND MOVQ_WONE | |
164 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) | |
165 #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
|
166 #define OP_AVG(a, b, c, e) PAVGB_MMX(a, b, c, e) |
8430 | 167 |
168 #include "dsputil_mmx_rnd_template.c" | |
169 | |
170 #undef DEF | |
171 #undef SET_RND | |
172 #undef PAVGBP | |
173 #undef PAVGB | |
174 /***********************************/ | |
175 /* MMX rounding */ | |
176 | |
177 #define DEF(x, y) x ## _ ## y ##_mmx | |
178 #define SET_RND MOVQ_WTWO | |
179 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) | |
180 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e) | |
181 | |
182 #include "dsputil_mmx_rnd_template.c" | |
183 | |
184 #undef DEF | |
185 #undef SET_RND | |
186 #undef PAVGBP | |
187 #undef PAVGB | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
188 #undef OP_AVG |
8430 | 189 |
190 /***********************************/ | |
191 /* 3Dnow specific */ | |
192 | |
193 #define DEF(x) x ## _3dnow | |
194 #define PAVGB "pavgusb" | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
195 #define OP_AVG PAVGB |
8430 | 196 |
197 #include "dsputil_mmx_avg_template.c" | |
198 | |
199 #undef DEF | |
200 #undef PAVGB | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
201 #undef OP_AVG |
8430 | 202 |
203 /***********************************/ | |
204 /* MMX2 specific */ | |
205 | |
206 #define DEF(x) x ## _mmx2 | |
207 | |
208 /* Introduced only in MMX2 set */ | |
209 #define PAVGB "pavgb" | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
210 #define OP_AVG PAVGB |
8430 | 211 |
212 #include "dsputil_mmx_avg_template.c" | |
213 | |
214 #undef DEF | |
215 #undef PAVGB | |
9445
41245484dc0b
avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents:
9441
diff
changeset
|
216 #undef OP_AVG |
8430 | 217 |
218 #define put_no_rnd_pixels16_mmx put_pixels16_mmx | |
219 #define put_no_rnd_pixels8_mmx put_pixels8_mmx | |
220 #define put_pixels16_mmx2 put_pixels16_mmx | |
221 #define put_pixels8_mmx2 put_pixels8_mmx | |
222 #define put_pixels4_mmx2 put_pixels4_mmx | |
223 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx | |
224 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx | |
225 #define put_pixels16_3dnow put_pixels16_mmx | |
226 #define put_pixels8_3dnow put_pixels8_mmx | |
227 #define put_pixels4_3dnow put_pixels4_mmx | |
228 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx | |
229 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx | |
230 | |
231 /***********************************/ | |
232 /* standard MMX */ | |
233 | |
234 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) | |
235 { | |
236 const DCTELEM *p; | |
237 uint8_t *pix; | |
238 | |
239 /* read the pixels */ | |
240 p = block; | |
241 pix = pixels; | |
242 /* unrolled loop */ | |
243 __asm__ volatile( | |
244 "movq %3, %%mm0 \n\t" | |
245 "movq 8%3, %%mm1 \n\t" | |
246 "movq 16%3, %%mm2 \n\t" | |
247 "movq 24%3, %%mm3 \n\t" | |
248 "movq 32%3, %%mm4 \n\t" | |
249 "movq 40%3, %%mm5 \n\t" | |
250 "movq 48%3, %%mm6 \n\t" | |
251 "movq 56%3, %%mm7 \n\t" | |
252 "packuswb %%mm1, %%mm0 \n\t" | |
253 "packuswb %%mm3, %%mm2 \n\t" | |
254 "packuswb %%mm5, %%mm4 \n\t" | |
255 "packuswb %%mm7, %%mm6 \n\t" | |
256 "movq %%mm0, (%0) \n\t" | |
257 "movq %%mm2, (%0, %1) \n\t" | |
258 "movq %%mm4, (%0, %1, 2) \n\t" | |
259 "movq %%mm6, (%0, %2) \n\t" | |
260 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "m"(*p) | |
261 :"memory"); | |
262 pix += line_size*4; | |
263 p += 32; | |
264 | |
265 // if here would be an exact copy of the code above | |
266 // compiler would generate some very strange code | |
267 // thus using "r" | |
268 __asm__ volatile( | |
269 "movq (%3), %%mm0 \n\t" | |
270 "movq 8(%3), %%mm1 \n\t" | |
271 "movq 16(%3), %%mm2 \n\t" | |
272 "movq 24(%3), %%mm3 \n\t" | |
273 "movq 32(%3), %%mm4 \n\t" | |
274 "movq 40(%3), %%mm5 \n\t" | |
275 "movq 48(%3), %%mm6 \n\t" | |
276 "movq 56(%3), %%mm7 \n\t" | |
277 "packuswb %%mm1, %%mm0 \n\t" | |
278 "packuswb %%mm3, %%mm2 \n\t" | |
279 "packuswb %%mm5, %%mm4 \n\t" | |
280 "packuswb %%mm7, %%mm6 \n\t" | |
281 "movq %%mm0, (%0) \n\t" | |
282 "movq %%mm2, (%0, %1) \n\t" | |
283 "movq %%mm4, (%0, %1, 2) \n\t" | |
284 "movq %%mm6, (%0, %2) \n\t" | |
285 ::"r" (pix), "r" ((x86_reg)line_size), "r" ((x86_reg)line_size*3), "r"(p) | |
286 :"memory"); | |
287 } | |
288 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
10766
diff
changeset
|
289 DECLARE_ASM_CONST(8, uint8_t, ff_vector128)[8] = |
8430 | 290 { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 }; |
291 | |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
292 #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
|
293 "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
|
294 "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
|
295 "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
|
296 "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
|
297 "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
|
298 "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
|
299 "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
|
300 "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
|
301 "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
|
302 "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
|
303 "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
|
304 "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
|
305 "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
|
306 "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
|
307 "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
|
308 "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
|
309 |
8430 | 310 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
311 { | |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
312 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
|
313 x86_reg line_skip3; |
8430 | 314 |
9337
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
315 __asm__ volatile ( |
a0d54042ea37
Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents:
8818
diff
changeset
|
316 "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
|
317 "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
|
318 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
|
319 "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
|
320 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
|
321 :"+&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
|
322 :"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
|
323 :"memory"); |
8430 | 324 } |
325 | |
326 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) | |
327 { | |
328 const DCTELEM *p; | |
329 uint8_t *pix; | |
330 int i; | |
331 | |
332 /* read the pixels */ | |
333 p = block; | |
334 pix = pixels; | |
335 MOVQ_ZERO(mm7); | |
336 i = 4; | |
337 do { | |
338 __asm__ volatile( | |
339 "movq (%2), %%mm0 \n\t" | |
340 "movq 8(%2), %%mm1 \n\t" | |
341 "movq 16(%2), %%mm2 \n\t" | |
342 "movq 24(%2), %%mm3 \n\t" | |
343 "movq %0, %%mm4 \n\t" | |
344 "movq %1, %%mm6 \n\t" | |
345 "movq %%mm4, %%mm5 \n\t" | |
346 "punpcklbw %%mm7, %%mm4 \n\t" | |
347 "punpckhbw %%mm7, %%mm5 \n\t" | |
348 "paddsw %%mm4, %%mm0 \n\t" | |
349 "paddsw %%mm5, %%mm1 \n\t" | |
350 "movq %%mm6, %%mm5 \n\t" | |
351 "punpcklbw %%mm7, %%mm6 \n\t" | |
352 "punpckhbw %%mm7, %%mm5 \n\t" | |
353 "paddsw %%mm6, %%mm2 \n\t" | |
354 "paddsw %%mm5, %%mm3 \n\t" | |
355 "packuswb %%mm1, %%mm0 \n\t" | |
356 "packuswb %%mm3, %%mm2 \n\t" | |
357 "movq %%mm0, %0 \n\t" | |
358 "movq %%mm2, %1 \n\t" | |
359 :"+m"(*pix), "+m"(*(pix+line_size)) | |
360 :"r"(p) | |
361 :"memory"); | |
362 pix += line_size*2; | |
363 p += 16; | |
364 } while (--i); | |
365 } | |
366 | |
367 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
368 { | |
369 __asm__ volatile( | |
370 "lea (%3, %3), %%"REG_a" \n\t" | |
371 ASMALIGN(3) | |
372 "1: \n\t" | |
373 "movd (%1), %%mm0 \n\t" | |
374 "movd (%1, %3), %%mm1 \n\t" | |
375 "movd %%mm0, (%2) \n\t" | |
376 "movd %%mm1, (%2, %3) \n\t" | |
377 "add %%"REG_a", %1 \n\t" | |
378 "add %%"REG_a", %2 \n\t" | |
379 "movd (%1), %%mm0 \n\t" | |
380 "movd (%1, %3), %%mm1 \n\t" | |
381 "movd %%mm0, (%2) \n\t" | |
382 "movd %%mm1, (%2, %3) \n\t" | |
383 "add %%"REG_a", %1 \n\t" | |
384 "add %%"REG_a", %2 \n\t" | |
385 "subl $4, %0 \n\t" | |
386 "jnz 1b \n\t" | |
387 : "+g"(h), "+r" (pixels), "+r" (block) | |
388 : "r"((x86_reg)line_size) | |
389 : "%"REG_a, "memory" | |
390 ); | |
391 } | |
392 | |
393 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
394 { | |
395 __asm__ volatile( | |
396 "lea (%3, %3), %%"REG_a" \n\t" | |
397 ASMALIGN(3) | |
398 "1: \n\t" | |
399 "movq (%1), %%mm0 \n\t" | |
400 "movq (%1, %3), %%mm1 \n\t" | |
401 "movq %%mm0, (%2) \n\t" | |
402 "movq %%mm1, (%2, %3) \n\t" | |
403 "add %%"REG_a", %1 \n\t" | |
404 "add %%"REG_a", %2 \n\t" | |
405 "movq (%1), %%mm0 \n\t" | |
406 "movq (%1, %3), %%mm1 \n\t" | |
407 "movq %%mm0, (%2) \n\t" | |
408 "movq %%mm1, (%2, %3) \n\t" | |
409 "add %%"REG_a", %1 \n\t" | |
410 "add %%"REG_a", %2 \n\t" | |
411 "subl $4, %0 \n\t" | |
412 "jnz 1b \n\t" | |
413 : "+g"(h), "+r" (pixels), "+r" (block) | |
414 : "r"((x86_reg)line_size) | |
415 : "%"REG_a, "memory" | |
416 ); | |
417 } | |
418 | |
419 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
420 { | |
421 __asm__ volatile( | |
422 "lea (%3, %3), %%"REG_a" \n\t" | |
423 ASMALIGN(3) | |
424 "1: \n\t" | |
425 "movq (%1), %%mm0 \n\t" | |
426 "movq 8(%1), %%mm4 \n\t" | |
427 "movq (%1, %3), %%mm1 \n\t" | |
428 "movq 8(%1, %3), %%mm5 \n\t" | |
429 "movq %%mm0, (%2) \n\t" | |
430 "movq %%mm4, 8(%2) \n\t" | |
431 "movq %%mm1, (%2, %3) \n\t" | |
432 "movq %%mm5, 8(%2, %3) \n\t" | |
433 "add %%"REG_a", %1 \n\t" | |
434 "add %%"REG_a", %2 \n\t" | |
435 "movq (%1), %%mm0 \n\t" | |
436 "movq 8(%1), %%mm4 \n\t" | |
437 "movq (%1, %3), %%mm1 \n\t" | |
438 "movq 8(%1, %3), %%mm5 \n\t" | |
439 "movq %%mm0, (%2) \n\t" | |
440 "movq %%mm4, 8(%2) \n\t" | |
441 "movq %%mm1, (%2, %3) \n\t" | |
442 "movq %%mm5, 8(%2, %3) \n\t" | |
443 "add %%"REG_a", %1 \n\t" | |
444 "add %%"REG_a", %2 \n\t" | |
445 "subl $4, %0 \n\t" | |
446 "jnz 1b \n\t" | |
447 : "+g"(h), "+r" (pixels), "+r" (block) | |
448 : "r"((x86_reg)line_size) | |
449 : "%"REG_a, "memory" | |
450 ); | |
451 } | |
452 | |
453 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
454 { | |
455 __asm__ volatile( | |
456 "1: \n\t" | |
457 "movdqu (%1), %%xmm0 \n\t" | |
458 "movdqu (%1,%3), %%xmm1 \n\t" | |
459 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
460 "movdqu (%1,%4), %%xmm3 \n\t" | |
461 "movdqa %%xmm0, (%2) \n\t" | |
462 "movdqa %%xmm1, (%2,%3) \n\t" | |
463 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
464 "movdqa %%xmm3, (%2,%4) \n\t" | |
465 "subl $4, %0 \n\t" | |
466 "lea (%1,%3,4), %1 \n\t" | |
467 "lea (%2,%3,4), %2 \n\t" | |
468 "jnz 1b \n\t" | |
469 : "+g"(h), "+r" (pixels), "+r" (block) | |
470 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size) | |
471 : "memory" | |
472 ); | |
473 } | |
474 | |
475 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
476 { | |
477 __asm__ volatile( | |
478 "1: \n\t" | |
479 "movdqu (%1), %%xmm0 \n\t" | |
480 "movdqu (%1,%3), %%xmm1 \n\t" | |
481 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
482 "movdqu (%1,%4), %%xmm3 \n\t" | |
483 "pavgb (%2), %%xmm0 \n\t" | |
484 "pavgb (%2,%3), %%xmm1 \n\t" | |
485 "pavgb (%2,%3,2), %%xmm2 \n\t" | |
486 "pavgb (%2,%4), %%xmm3 \n\t" | |
487 "movdqa %%xmm0, (%2) \n\t" | |
488 "movdqa %%xmm1, (%2,%3) \n\t" | |
489 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
490 "movdqa %%xmm3, (%2,%4) \n\t" | |
491 "subl $4, %0 \n\t" | |
492 "lea (%1,%3,4), %1 \n\t" | |
493 "lea (%2,%3,4), %2 \n\t" | |
494 "jnz 1b \n\t" | |
495 : "+g"(h), "+r" (pixels), "+r" (block) | |
496 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size) | |
497 : "memory" | |
498 ); | |
499 } | |
500 | |
501 #define CLEAR_BLOCKS(name,n) \ | |
502 static void name(DCTELEM *blocks)\ | |
503 {\ | |
504 __asm__ volatile(\ | |
505 "pxor %%mm7, %%mm7 \n\t"\ | |
506 "mov %1, %%"REG_a" \n\t"\ | |
507 "1: \n\t"\ | |
508 "movq %%mm7, (%0, %%"REG_a") \n\t"\ | |
509 "movq %%mm7, 8(%0, %%"REG_a") \n\t"\ | |
510 "movq %%mm7, 16(%0, %%"REG_a") \n\t"\ | |
511 "movq %%mm7, 24(%0, %%"REG_a") \n\t"\ | |
512 "add $32, %%"REG_a" \n\t"\ | |
513 " js 1b \n\t"\ | |
514 : : "r" (((uint8_t *)blocks)+128*n),\ | |
515 "i" (-128*n)\ | |
516 : "%"REG_a\ | |
517 );\ | |
518 } | |
519 CLEAR_BLOCKS(clear_blocks_mmx, 6) | |
520 CLEAR_BLOCKS(clear_block_mmx, 1) | |
521 | |
522 static void clear_block_sse(DCTELEM *block) | |
523 { | |
524 __asm__ volatile( | |
525 "xorps %%xmm0, %%xmm0 \n" | |
526 "movaps %%xmm0, (%0) \n" | |
527 "movaps %%xmm0, 16(%0) \n" | |
528 "movaps %%xmm0, 32(%0) \n" | |
529 "movaps %%xmm0, 48(%0) \n" | |
530 "movaps %%xmm0, 64(%0) \n" | |
531 "movaps %%xmm0, 80(%0) \n" | |
532 "movaps %%xmm0, 96(%0) \n" | |
533 "movaps %%xmm0, 112(%0) \n" | |
534 :: "r"(block) | |
535 : "memory" | |
536 ); | |
537 } | |
538 | |
9861 | 539 static void clear_blocks_sse(DCTELEM *blocks) |
540 {\ | |
541 __asm__ volatile( | |
542 "xorps %%xmm0, %%xmm0 \n" | |
543 "mov %1, %%"REG_a" \n" | |
544 "1: \n" | |
545 "movaps %%xmm0, (%0, %%"REG_a") \n" | |
546 "movaps %%xmm0, 16(%0, %%"REG_a") \n" | |
547 "movaps %%xmm0, 32(%0, %%"REG_a") \n" | |
548 "movaps %%xmm0, 48(%0, %%"REG_a") \n" | |
549 "movaps %%xmm0, 64(%0, %%"REG_a") \n" | |
550 "movaps %%xmm0, 80(%0, %%"REG_a") \n" | |
551 "movaps %%xmm0, 96(%0, %%"REG_a") \n" | |
552 "movaps %%xmm0, 112(%0, %%"REG_a") \n" | |
553 "add $128, %%"REG_a" \n" | |
554 " js 1b \n" | |
555 : : "r" (((uint8_t *)blocks)+128*6), | |
556 "i" (-128*6) | |
557 : "%"REG_a | |
558 ); | |
559 } | |
560 | |
8430 | 561 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){ |
562 x86_reg i=0; | |
563 __asm__ volatile( | |
564 "jmp 2f \n\t" | |
565 "1: \n\t" | |
566 "movq (%1, %0), %%mm0 \n\t" | |
567 "movq (%2, %0), %%mm1 \n\t" | |
568 "paddb %%mm0, %%mm1 \n\t" | |
569 "movq %%mm1, (%2, %0) \n\t" | |
570 "movq 8(%1, %0), %%mm0 \n\t" | |
571 "movq 8(%2, %0), %%mm1 \n\t" | |
572 "paddb %%mm0, %%mm1 \n\t" | |
573 "movq %%mm1, 8(%2, %0) \n\t" | |
574 "add $16, %0 \n\t" | |
575 "2: \n\t" | |
576 "cmp %3, %0 \n\t" | |
577 " js 1b \n\t" | |
578 : "+r" (i) | |
579 : "r"(src), "r"(dst), "r"((x86_reg)w-15) | |
580 ); | |
581 for(; i<w; i++) | |
582 dst[i+0] += src[i+0]; | |
583 } | |
584 | |
585 static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){ | |
586 x86_reg i=0; | |
587 __asm__ volatile( | |
588 "jmp 2f \n\t" | |
589 "1: \n\t" | |
590 "movq (%2, %0), %%mm0 \n\t" | |
591 "movq 8(%2, %0), %%mm1 \n\t" | |
592 "paddb (%3, %0), %%mm0 \n\t" | |
593 "paddb 8(%3, %0), %%mm1 \n\t" | |
594 "movq %%mm0, (%1, %0) \n\t" | |
595 "movq %%mm1, 8(%1, %0) \n\t" | |
596 "add $16, %0 \n\t" | |
597 "2: \n\t" | |
598 "cmp %4, %0 \n\t" | |
599 " js 1b \n\t" | |
600 : "+r" (i) | |
601 : "r"(dst), "r"(src1), "r"(src2), "r"((x86_reg)w-15) | |
602 ); | |
603 for(; i<w; i++) | |
604 dst[i] = src1[i] + src2[i]; | |
605 } | |
606 | |
8798
a5c8210814d7
Add check whether the compiler/assembler supports 10 or more operands.
diego
parents:
8760
diff
changeset
|
607 #if HAVE_7REGS && HAVE_TEN_OPERANDS |
10431 | 608 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 | 609 x86_reg w2 = -w; |
610 x86_reg x; | |
611 int l = *left & 0xff; | |
612 int tl = *left_top & 0xff; | |
613 int t; | |
614 __asm__ volatile( | |
615 "mov %7, %3 \n" | |
616 "1: \n" | |
617 "movzx (%3,%4), %2 \n" | |
618 "mov %2, %k3 \n" | |
619 "sub %b1, %b3 \n" | |
620 "add %b0, %b3 \n" | |
621 "mov %2, %1 \n" | |
622 "cmp %0, %2 \n" | |
623 "cmovg %0, %2 \n" | |
624 "cmovg %1, %0 \n" | |
625 "cmp %k3, %0 \n" | |
626 "cmovg %k3, %0 \n" | |
627 "mov %7, %3 \n" | |
628 "cmp %2, %0 \n" | |
629 "cmovl %2, %0 \n" | |
630 "add (%6,%4), %b0 \n" | |
631 "mov %b0, (%5,%4) \n" | |
632 "inc %4 \n" | |
633 "jl 1b \n" | |
634 :"+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2) | |
635 :"r"(dst+w), "r"(diff+w), "rm"(top+w) | |
636 ); | |
637 *left = l; | |
638 *left_top = tl; | |
639 } | |
640 #endif | |
641 | |
8430 | 642 #define H263_LOOP_FILTER \ |
643 "pxor %%mm7, %%mm7 \n\t"\ | |
644 "movq %0, %%mm0 \n\t"\ | |
645 "movq %0, %%mm1 \n\t"\ | |
646 "movq %3, %%mm2 \n\t"\ | |
647 "movq %3, %%mm3 \n\t"\ | |
648 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
649 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
650 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
651 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
652 "psubw %%mm2, %%mm0 \n\t"\ | |
653 "psubw %%mm3, %%mm1 \n\t"\ | |
654 "movq %1, %%mm2 \n\t"\ | |
655 "movq %1, %%mm3 \n\t"\ | |
656 "movq %2, %%mm4 \n\t"\ | |
657 "movq %2, %%mm5 \n\t"\ | |
658 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
659 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
660 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
661 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
662 "psubw %%mm2, %%mm4 \n\t"\ | |
663 "psubw %%mm3, %%mm5 \n\t"\ | |
664 "psllw $2, %%mm4 \n\t"\ | |
665 "psllw $2, %%mm5 \n\t"\ | |
666 "paddw %%mm0, %%mm4 \n\t"\ | |
667 "paddw %%mm1, %%mm5 \n\t"\ | |
668 "pxor %%mm6, %%mm6 \n\t"\ | |
669 "pcmpgtw %%mm4, %%mm6 \n\t"\ | |
670 "pcmpgtw %%mm5, %%mm7 \n\t"\ | |
671 "pxor %%mm6, %%mm4 \n\t"\ | |
672 "pxor %%mm7, %%mm5 \n\t"\ | |
673 "psubw %%mm6, %%mm4 \n\t"\ | |
674 "psubw %%mm7, %%mm5 \n\t"\ | |
675 "psrlw $3, %%mm4 \n\t"\ | |
676 "psrlw $3, %%mm5 \n\t"\ | |
677 "packuswb %%mm5, %%mm4 \n\t"\ | |
678 "packsswb %%mm7, %%mm6 \n\t"\ | |
679 "pxor %%mm7, %%mm7 \n\t"\ | |
680 "movd %4, %%mm2 \n\t"\ | |
681 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
682 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
683 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
684 "psubusb %%mm4, %%mm2 \n\t"\ | |
685 "movq %%mm2, %%mm3 \n\t"\ | |
686 "psubusb %%mm4, %%mm3 \n\t"\ | |
687 "psubb %%mm3, %%mm2 \n\t"\ | |
688 "movq %1, %%mm3 \n\t"\ | |
689 "movq %2, %%mm4 \n\t"\ | |
690 "pxor %%mm6, %%mm3 \n\t"\ | |
691 "pxor %%mm6, %%mm4 \n\t"\ | |
692 "paddusb %%mm2, %%mm3 \n\t"\ | |
693 "psubusb %%mm2, %%mm4 \n\t"\ | |
694 "pxor %%mm6, %%mm3 \n\t"\ | |
695 "pxor %%mm6, %%mm4 \n\t"\ | |
696 "paddusb %%mm2, %%mm2 \n\t"\ | |
697 "packsswb %%mm1, %%mm0 \n\t"\ | |
698 "pcmpgtb %%mm0, %%mm7 \n\t"\ | |
699 "pxor %%mm7, %%mm0 \n\t"\ | |
700 "psubb %%mm7, %%mm0 \n\t"\ | |
701 "movq %%mm0, %%mm1 \n\t"\ | |
702 "psubusb %%mm2, %%mm0 \n\t"\ | |
703 "psubb %%mm0, %%mm1 \n\t"\ | |
704 "pand %5, %%mm1 \n\t"\ | |
705 "psrlw $2, %%mm1 \n\t"\ | |
706 "pxor %%mm7, %%mm1 \n\t"\ | |
707 "psubb %%mm7, %%mm1 \n\t"\ | |
708 "movq %0, %%mm5 \n\t"\ | |
709 "movq %3, %%mm6 \n\t"\ | |
710 "psubb %%mm1, %%mm5 \n\t"\ | |
711 "paddb %%mm1, %%mm6 \n\t" | |
712 | |
713 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
|
714 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { |
8430 | 715 const int strength= ff_h263_loop_filter_strength[qscale]; |
716 | |
717 __asm__ volatile( | |
718 | |
719 H263_LOOP_FILTER | |
720 | |
721 "movq %%mm3, %1 \n\t" | |
722 "movq %%mm4, %2 \n\t" | |
723 "movq %%mm5, %0 \n\t" | |
724 "movq %%mm6, %3 \n\t" | |
725 : "+m" (*(uint64_t*)(src - 2*stride)), | |
726 "+m" (*(uint64_t*)(src - 1*stride)), | |
727 "+m" (*(uint64_t*)(src + 0*stride)), | |
728 "+m" (*(uint64_t*)(src + 1*stride)) | |
729 : "g" (2*strength), "m"(ff_pb_FC) | |
730 ); | |
731 } | |
732 } | |
733 | |
734 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ | |
735 __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ... | |
736 "movd %4, %%mm0 \n\t" | |
737 "movd %5, %%mm1 \n\t" | |
738 "movd %6, %%mm2 \n\t" | |
739 "movd %7, %%mm3 \n\t" | |
740 "punpcklbw %%mm1, %%mm0 \n\t" | |
741 "punpcklbw %%mm3, %%mm2 \n\t" | |
742 "movq %%mm0, %%mm1 \n\t" | |
743 "punpcklwd %%mm2, %%mm0 \n\t" | |
744 "punpckhwd %%mm2, %%mm1 \n\t" | |
745 "movd %%mm0, %0 \n\t" | |
746 "punpckhdq %%mm0, %%mm0 \n\t" | |
747 "movd %%mm0, %1 \n\t" | |
748 "movd %%mm1, %2 \n\t" | |
749 "punpckhdq %%mm1, %%mm1 \n\t" | |
750 "movd %%mm1, %3 \n\t" | |
751 | |
752 : "=m" (*(uint32_t*)(dst + 0*dst_stride)), | |
753 "=m" (*(uint32_t*)(dst + 1*dst_stride)), | |
754 "=m" (*(uint32_t*)(dst + 2*dst_stride)), | |
755 "=m" (*(uint32_t*)(dst + 3*dst_stride)) | |
756 : "m" (*(uint32_t*)(src + 0*src_stride)), | |
757 "m" (*(uint32_t*)(src + 1*src_stride)), | |
758 "m" (*(uint32_t*)(src + 2*src_stride)), | |
759 "m" (*(uint32_t*)(src + 3*src_stride)) | |
760 ); | |
761 } | |
762 | |
763 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
|
764 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { |
8430 | 765 const int strength= ff_h263_loop_filter_strength[qscale]; |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
10766
diff
changeset
|
766 DECLARE_ALIGNED(8, uint64_t, temp)[4]; |
8430 | 767 uint8_t *btemp= (uint8_t*)temp; |
768 | |
769 src -= 2; | |
770 | |
771 transpose4x4(btemp , src , 8, stride); | |
772 transpose4x4(btemp+4, src + 4*stride, 8, stride); | |
773 __asm__ volatile( | |
774 H263_LOOP_FILTER // 5 3 4 6 | |
775 | |
776 : "+m" (temp[0]), | |
777 "+m" (temp[1]), | |
778 "+m" (temp[2]), | |
779 "+m" (temp[3]) | |
780 : "g" (2*strength), "m"(ff_pb_FC) | |
781 ); | |
782 | |
783 __asm__ volatile( | |
784 "movq %%mm5, %%mm1 \n\t" | |
785 "movq %%mm4, %%mm0 \n\t" | |
786 "punpcklbw %%mm3, %%mm5 \n\t" | |
787 "punpcklbw %%mm6, %%mm4 \n\t" | |
788 "punpckhbw %%mm3, %%mm1 \n\t" | |
789 "punpckhbw %%mm6, %%mm0 \n\t" | |
790 "movq %%mm5, %%mm3 \n\t" | |
791 "movq %%mm1, %%mm6 \n\t" | |
792 "punpcklwd %%mm4, %%mm5 \n\t" | |
793 "punpcklwd %%mm0, %%mm1 \n\t" | |
794 "punpckhwd %%mm4, %%mm3 \n\t" | |
795 "punpckhwd %%mm0, %%mm6 \n\t" | |
796 "movd %%mm5, (%0) \n\t" | |
797 "punpckhdq %%mm5, %%mm5 \n\t" | |
798 "movd %%mm5, (%0,%2) \n\t" | |
799 "movd %%mm3, (%0,%2,2) \n\t" | |
800 "punpckhdq %%mm3, %%mm3 \n\t" | |
801 "movd %%mm3, (%0,%3) \n\t" | |
802 "movd %%mm1, (%1) \n\t" | |
803 "punpckhdq %%mm1, %%mm1 \n\t" | |
804 "movd %%mm1, (%1,%2) \n\t" | |
805 "movd %%mm6, (%1,%2,2) \n\t" | |
806 "punpckhdq %%mm6, %%mm6 \n\t" | |
807 "movd %%mm6, (%1,%3) \n\t" | |
808 :: "r" (src), | |
809 "r" (src + 4*stride), | |
810 "r" ((x86_reg) stride ), | |
811 "r" ((x86_reg)(3*stride)) | |
812 ); | |
813 } | |
814 } | |
815 | |
816 /* draw the edges of width 'w' of an image of size width, height | |
817 this mmx version can only handle w==8 || w==16 */ | |
818 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w) | |
819 { | |
820 uint8_t *ptr, *last_line; | |
821 int i; | |
822 | |
823 last_line = buf + (height - 1) * wrap; | |
824 /* left and right */ | |
825 ptr = buf; | |
826 if(w==8) | |
827 { | |
828 __asm__ volatile( | |
829 "1: \n\t" | |
830 "movd (%0), %%mm0 \n\t" | |
831 "punpcklbw %%mm0, %%mm0 \n\t" | |
832 "punpcklwd %%mm0, %%mm0 \n\t" | |
833 "punpckldq %%mm0, %%mm0 \n\t" | |
834 "movq %%mm0, -8(%0) \n\t" | |
835 "movq -8(%0, %2), %%mm1 \n\t" | |
836 "punpckhbw %%mm1, %%mm1 \n\t" | |
837 "punpckhwd %%mm1, %%mm1 \n\t" | |
838 "punpckhdq %%mm1, %%mm1 \n\t" | |
839 "movq %%mm1, (%0, %2) \n\t" | |
840 "add %1, %0 \n\t" | |
841 "cmp %3, %0 \n\t" | |
842 " jb 1b \n\t" | |
843 : "+r" (ptr) | |
844 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height) | |
845 ); | |
846 } | |
847 else | |
848 { | |
849 __asm__ volatile( | |
850 "1: \n\t" | |
851 "movd (%0), %%mm0 \n\t" | |
852 "punpcklbw %%mm0, %%mm0 \n\t" | |
853 "punpcklwd %%mm0, %%mm0 \n\t" | |
854 "punpckldq %%mm0, %%mm0 \n\t" | |
855 "movq %%mm0, -8(%0) \n\t" | |
856 "movq %%mm0, -16(%0) \n\t" | |
857 "movq -8(%0, %2), %%mm1 \n\t" | |
858 "punpckhbw %%mm1, %%mm1 \n\t" | |
859 "punpckhwd %%mm1, %%mm1 \n\t" | |
860 "punpckhdq %%mm1, %%mm1 \n\t" | |
861 "movq %%mm1, (%0, %2) \n\t" | |
862 "movq %%mm1, 8(%0, %2) \n\t" | |
863 "add %1, %0 \n\t" | |
864 "cmp %3, %0 \n\t" | |
865 " jb 1b \n\t" | |
866 : "+r" (ptr) | |
867 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height) | |
868 ); | |
869 } | |
870 | |
871 for(i=0;i<w;i+=4) { | |
872 /* top and bottom (and hopefully also the corners) */ | |
873 ptr= buf - (i + 1) * wrap - w; | |
874 __asm__ volatile( | |
875 "1: \n\t" | |
876 "movq (%1, %0), %%mm0 \n\t" | |
877 "movq %%mm0, (%0) \n\t" | |
878 "movq %%mm0, (%0, %2) \n\t" | |
879 "movq %%mm0, (%0, %2, 2) \n\t" | |
880 "movq %%mm0, (%0, %3) \n\t" | |
881 "add $8, %0 \n\t" | |
882 "cmp %4, %0 \n\t" | |
883 " jb 1b \n\t" | |
884 : "+r" (ptr) | |
885 : "r" ((x86_reg)buf - (x86_reg)ptr - w), "r" ((x86_reg)-wrap), "r" ((x86_reg)-wrap*3), "r" (ptr+width+2*w) | |
886 ); | |
887 ptr= last_line + (i + 1) * wrap - w; | |
888 __asm__ volatile( | |
889 "1: \n\t" | |
890 "movq (%1, %0), %%mm0 \n\t" | |
891 "movq %%mm0, (%0) \n\t" | |
892 "movq %%mm0, (%0, %2) \n\t" | |
893 "movq %%mm0, (%0, %2, 2) \n\t" | |
894 "movq %%mm0, (%0, %3) \n\t" | |
895 "add $8, %0 \n\t" | |
896 "cmp %4, %0 \n\t" | |
897 " jb 1b \n\t" | |
898 : "+r" (ptr) | |
899 : "r" ((x86_reg)last_line - (x86_reg)ptr - w), "r" ((x86_reg)wrap), "r" ((x86_reg)wrap*3), "r" (ptr+width+2*w) | |
900 ); | |
901 } | |
902 } | |
903 | |
904 #define PAETH(cpu, abs3)\ | |
905 static void add_png_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\ | |
906 {\ | |
907 x86_reg i = -bpp;\ | |
908 x86_reg end = w-3;\ | |
909 __asm__ volatile(\ | |
910 "pxor %%mm7, %%mm7 \n"\ | |
911 "movd (%1,%0), %%mm0 \n"\ | |
912 "movd (%2,%0), %%mm1 \n"\ | |
913 "punpcklbw %%mm7, %%mm0 \n"\ | |
914 "punpcklbw %%mm7, %%mm1 \n"\ | |
915 "add %4, %0 \n"\ | |
916 "1: \n"\ | |
917 "movq %%mm1, %%mm2 \n"\ | |
918 "movd (%2,%0), %%mm1 \n"\ | |
919 "movq %%mm2, %%mm3 \n"\ | |
920 "punpcklbw %%mm7, %%mm1 \n"\ | |
921 "movq %%mm2, %%mm4 \n"\ | |
922 "psubw %%mm1, %%mm3 \n"\ | |
923 "psubw %%mm0, %%mm4 \n"\ | |
924 "movq %%mm3, %%mm5 \n"\ | |
925 "paddw %%mm4, %%mm5 \n"\ | |
926 abs3\ | |
927 "movq %%mm4, %%mm6 \n"\ | |
928 "pminsw %%mm5, %%mm6 \n"\ | |
929 "pcmpgtw %%mm6, %%mm3 \n"\ | |
930 "pcmpgtw %%mm5, %%mm4 \n"\ | |
931 "movq %%mm4, %%mm6 \n"\ | |
932 "pand %%mm3, %%mm4 \n"\ | |
933 "pandn %%mm3, %%mm6 \n"\ | |
934 "pandn %%mm0, %%mm3 \n"\ | |
935 "movd (%3,%0), %%mm0 \n"\ | |
936 "pand %%mm1, %%mm6 \n"\ | |
937 "pand %%mm4, %%mm2 \n"\ | |
938 "punpcklbw %%mm7, %%mm0 \n"\ | |
939 "movq %6, %%mm5 \n"\ | |
940 "paddw %%mm6, %%mm0 \n"\ | |
941 "paddw %%mm2, %%mm3 \n"\ | |
942 "paddw %%mm3, %%mm0 \n"\ | |
943 "pand %%mm5, %%mm0 \n"\ | |
944 "movq %%mm0, %%mm3 \n"\ | |
945 "packuswb %%mm3, %%mm3 \n"\ | |
946 "movd %%mm3, (%1,%0) \n"\ | |
947 "add %4, %0 \n"\ | |
948 "cmp %5, %0 \n"\ | |
949 "jle 1b \n"\ | |
950 :"+r"(i)\ | |
951 :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\ | |
952 "m"(ff_pw_255)\ | |
953 :"memory"\ | |
954 );\ | |
955 } | |
956 | |
957 #define ABS3_MMX2\ | |
958 "psubw %%mm5, %%mm7 \n"\ | |
959 "pmaxsw %%mm7, %%mm5 \n"\ | |
960 "pxor %%mm6, %%mm6 \n"\ | |
961 "pxor %%mm7, %%mm7 \n"\ | |
962 "psubw %%mm3, %%mm6 \n"\ | |
963 "psubw %%mm4, %%mm7 \n"\ | |
964 "pmaxsw %%mm6, %%mm3 \n"\ | |
965 "pmaxsw %%mm7, %%mm4 \n"\ | |
966 "pxor %%mm7, %%mm7 \n" | |
967 | |
968 #define ABS3_SSSE3\ | |
969 "pabsw %%mm3, %%mm3 \n"\ | |
970 "pabsw %%mm4, %%mm4 \n"\ | |
971 "pabsw %%mm5, %%mm5 \n" | |
972 | |
973 PAETH(mmx2, ABS3_MMX2) | |
8590 | 974 #if HAVE_SSSE3 |
8430 | 975 PAETH(ssse3, ABS3_SSSE3) |
976 #endif | |
977 | |
978 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ | |
979 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ | |
980 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ | |
981 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ | |
982 "movq "#in7", " #m3 " \n\t" /* d */\ | |
983 "movq "#in0", %%mm5 \n\t" /* D */\ | |
984 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
985 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
986 "movq "#in1", %%mm5 \n\t" /* C */\ | |
987 "movq "#in2", %%mm6 \n\t" /* B */\ | |
988 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
989 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
990 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
991 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
992 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ | |
993 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ | |
994 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
995 "psraw $5, %%mm5 \n\t"\ | |
996 "packuswb %%mm5, %%mm5 \n\t"\ | |
997 OP(%%mm5, out, %%mm7, d) | |
998 | |
999 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ | |
1000 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1001 uint64_t temp;\ | |
1002 \ | |
1003 __asm__ volatile(\ | |
1004 "pxor %%mm7, %%mm7 \n\t"\ | |
1005 "1: \n\t"\ | |
1006 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1007 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1008 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1009 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1010 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1011 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1012 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1013 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1014 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1015 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1016 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1017 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1018 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1019 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1020 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1021 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1022 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1023 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1024 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1025 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1026 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1027 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1028 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1029 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1030 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1031 "paddw %6, %%mm6 \n\t"\ | |
1032 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1033 "psraw $5, %%mm0 \n\t"\ | |
1034 "movq %%mm0, %5 \n\t"\ | |
1035 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ | |
1036 \ | |
1037 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ | |
1038 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
1039 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
1040 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
1041 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
1042 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
1043 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
1044 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
1045 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
1046 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1047 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1048 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
1049 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
1050 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
1051 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
1052 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1053 "paddw %%mm2, %%mm1 \n\t" /* a */\ | |
1054 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
1055 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1056 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ | |
1057 "paddw %6, %%mm1 \n\t"\ | |
1058 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ | |
1059 "psraw $5, %%mm3 \n\t"\ | |
1060 "movq %5, %%mm1 \n\t"\ | |
1061 "packuswb %%mm3, %%mm1 \n\t"\ | |
1062 OP_MMX2(%%mm1, (%1),%%mm4, q)\ | |
1063 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ | |
1064 \ | |
1065 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ | |
1066 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
1067 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
1068 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
1069 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
1070 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
1071 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
1072 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
1073 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
1074 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1075 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
1076 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
1077 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
1078 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ | |
1079 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ | |
1080 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
1081 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
1082 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
1083 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
1084 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
1085 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
1086 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ | |
1087 "paddw %6, %%mm0 \n\t"\ | |
1088 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1089 "psraw $5, %%mm0 \n\t"\ | |
1090 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ | |
1091 \ | |
1092 "paddw %%mm5, %%mm3 \n\t" /* a */\ | |
1093 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
1094 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
1095 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
1096 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
1097 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
1098 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
1099 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
1100 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
1101 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ | |
1102 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
1103 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1104 "paddw %6, %%mm4 \n\t"\ | |
1105 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ | |
1106 "psraw $5, %%mm4 \n\t"\ | |
1107 "packuswb %%mm4, %%mm0 \n\t"\ | |
1108 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ | |
1109 \ | |
1110 "add %3, %0 \n\t"\ | |
1111 "add %4, %1 \n\t"\ | |
1112 "decl %2 \n\t"\ | |
1113 " jnz 1b \n\t"\ | |
1114 : "+a"(src), "+c"(dst), "+D"(h)\ | |
1115 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ | |
1116 : "memory"\ | |
1117 );\ | |
1118 }\ | |
1119 \ | |
1120 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1121 int i;\ | |
1122 int16_t temp[16];\ | |
1123 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1124 for(i=0; i<h; i++)\ | |
1125 {\ | |
1126 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1127 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1128 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1129 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
1130 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1131 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
1132 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
1133 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
1134 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
1135 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
1136 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
1137 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
1138 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
1139 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
1140 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
1141 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
1142 __asm__ volatile(\ | |
1143 "movq (%0), %%mm0 \n\t"\ | |
1144 "movq 8(%0), %%mm1 \n\t"\ | |
1145 "paddw %2, %%mm0 \n\t"\ | |
1146 "paddw %2, %%mm1 \n\t"\ | |
1147 "psraw $5, %%mm0 \n\t"\ | |
1148 "psraw $5, %%mm1 \n\t"\ | |
1149 "packuswb %%mm1, %%mm0 \n\t"\ | |
1150 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ | |
1151 "movq 16(%0), %%mm0 \n\t"\ | |
1152 "movq 24(%0), %%mm1 \n\t"\ | |
1153 "paddw %2, %%mm0 \n\t"\ | |
1154 "paddw %2, %%mm1 \n\t"\ | |
1155 "psraw $5, %%mm0 \n\t"\ | |
1156 "psraw $5, %%mm1 \n\t"\ | |
1157 "packuswb %%mm1, %%mm0 \n\t"\ | |
1158 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ | |
1159 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
1160 : "memory"\ | |
1161 );\ | |
1162 dst+=dstStride;\ | |
1163 src+=srcStride;\ | |
1164 }\ | |
1165 }\ | |
1166 \ | |
1167 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1168 __asm__ volatile(\ | |
1169 "pxor %%mm7, %%mm7 \n\t"\ | |
1170 "1: \n\t"\ | |
1171 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1172 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1173 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1174 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1175 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1176 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1177 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1178 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1179 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1180 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1181 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1182 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1183 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1184 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1185 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1186 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1187 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1188 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1189 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1190 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1191 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1192 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1193 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1194 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1195 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1196 "paddw %5, %%mm6 \n\t"\ | |
1197 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1198 "psraw $5, %%mm0 \n\t"\ | |
1199 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ | |
1200 \ | |
1201 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ | |
1202 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
1203 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
1204 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
1205 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
1206 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
1207 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
1208 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
1209 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
1210 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1211 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1212 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1213 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1214 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1215 "paddw %5, %%mm1 \n\t"\ | |
1216 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ | |
1217 "psraw $5, %%mm3 \n\t"\ | |
1218 "packuswb %%mm3, %%mm0 \n\t"\ | |
1219 OP_MMX2(%%mm0, (%1), %%mm4, q)\ | |
1220 \ | |
1221 "add %3, %0 \n\t"\ | |
1222 "add %4, %1 \n\t"\ | |
1223 "decl %2 \n\t"\ | |
1224 " jnz 1b \n\t"\ | |
1225 : "+a"(src), "+c"(dst), "+d"(h)\ | |
1226 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER)\ | |
1227 : "memory"\ | |
1228 );\ | |
1229 }\ | |
1230 \ | |
1231 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1232 int i;\ | |
1233 int16_t temp[8];\ | |
1234 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1235 for(i=0; i<h; i++)\ | |
1236 {\ | |
1237 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1238 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1239 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1240 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
1241 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1242 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
1243 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
1244 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
1245 __asm__ volatile(\ | |
1246 "movq (%0), %%mm0 \n\t"\ | |
1247 "movq 8(%0), %%mm1 \n\t"\ | |
1248 "paddw %2, %%mm0 \n\t"\ | |
1249 "paddw %2, %%mm1 \n\t"\ | |
1250 "psraw $5, %%mm0 \n\t"\ | |
1251 "psraw $5, %%mm1 \n\t"\ | |
1252 "packuswb %%mm1, %%mm0 \n\t"\ | |
1253 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ | |
1254 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
1255 :"memory"\ | |
1256 );\ | |
1257 dst+=dstStride;\ | |
1258 src+=srcStride;\ | |
1259 }\ | |
1260 } | |
1261 | |
1262 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
1263 \ | |
1264 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1265 uint64_t temp[17*4];\ | |
1266 uint64_t *temp_ptr= temp;\ | |
1267 int count= 17;\ | |
1268 \ | |
1269 /*FIXME unroll */\ | |
1270 __asm__ volatile(\ | |
1271 "pxor %%mm7, %%mm7 \n\t"\ | |
1272 "1: \n\t"\ | |
1273 "movq (%0), %%mm0 \n\t"\ | |
1274 "movq (%0), %%mm1 \n\t"\ | |
1275 "movq 8(%0), %%mm2 \n\t"\ | |
1276 "movq 8(%0), %%mm3 \n\t"\ | |
1277 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1278 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1279 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1280 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1281 "movq %%mm0, (%1) \n\t"\ | |
1282 "movq %%mm1, 17*8(%1) \n\t"\ | |
1283 "movq %%mm2, 2*17*8(%1) \n\t"\ | |
1284 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
1285 "add $8, %1 \n\t"\ | |
1286 "add %3, %0 \n\t"\ | |
1287 "decl %2 \n\t"\ | |
1288 " jnz 1b \n\t"\ | |
1289 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ | |
1290 : "r" ((x86_reg)srcStride)\ | |
1291 : "memory"\ | |
1292 );\ | |
1293 \ | |
1294 temp_ptr= temp;\ | |
1295 count=4;\ | |
1296 \ | |
1297 /*FIXME reorder for speed */\ | |
1298 __asm__ volatile(\ | |
1299 /*"pxor %%mm7, %%mm7 \n\t"*/\ | |
1300 "1: \n\t"\ | |
1301 "movq (%0), %%mm0 \n\t"\ | |
1302 "movq 8(%0), %%mm1 \n\t"\ | |
1303 "movq 16(%0), %%mm2 \n\t"\ | |
1304 "movq 24(%0), %%mm3 \n\t"\ | |
1305 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ | |
1306 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
1307 "add %4, %1 \n\t"\ | |
1308 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ | |
1309 \ | |
1310 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ | |
1311 "add %4, %1 \n\t"\ | |
1312 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ | |
1313 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
1314 "add %4, %1 \n\t"\ | |
1315 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ | |
1316 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
1317 "add %4, %1 \n\t"\ | |
1318 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ | |
1319 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
1320 "add %4, %1 \n\t"\ | |
1321 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ | |
1322 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
1323 "add %4, %1 \n\t"\ | |
1324 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ | |
1325 \ | |
1326 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ | |
1327 "add %4, %1 \n\t" \ | |
1328 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ | |
1329 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
1330 \ | |
1331 "add $136, %0 \n\t"\ | |
1332 "add %6, %1 \n\t"\ | |
1333 "decl %2 \n\t"\ | |
1334 " jnz 1b \n\t"\ | |
1335 \ | |
1336 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ | |
1337 : "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)\ | |
1338 :"memory"\ | |
1339 );\ | |
1340 }\ | |
1341 \ | |
1342 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1343 uint64_t temp[9*2];\ | |
1344 uint64_t *temp_ptr= temp;\ | |
1345 int count= 9;\ | |
1346 \ | |
1347 /*FIXME unroll */\ | |
1348 __asm__ volatile(\ | |
1349 "pxor %%mm7, %%mm7 \n\t"\ | |
1350 "1: \n\t"\ | |
1351 "movq (%0), %%mm0 \n\t"\ | |
1352 "movq (%0), %%mm1 \n\t"\ | |
1353 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1354 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1355 "movq %%mm0, (%1) \n\t"\ | |
1356 "movq %%mm1, 9*8(%1) \n\t"\ | |
1357 "add $8, %1 \n\t"\ | |
1358 "add %3, %0 \n\t"\ | |
1359 "decl %2 \n\t"\ | |
1360 " jnz 1b \n\t"\ | |
1361 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ | |
1362 : "r" ((x86_reg)srcStride)\ | |
1363 : "memory"\ | |
1364 );\ | |
1365 \ | |
1366 temp_ptr= temp;\ | |
1367 count=2;\ | |
1368 \ | |
1369 /*FIXME reorder for speed */\ | |
1370 __asm__ volatile(\ | |
1371 /*"pxor %%mm7, %%mm7 \n\t"*/\ | |
1372 "1: \n\t"\ | |
1373 "movq (%0), %%mm0 \n\t"\ | |
1374 "movq 8(%0), %%mm1 \n\t"\ | |
1375 "movq 16(%0), %%mm2 \n\t"\ | |
1376 "movq 24(%0), %%mm3 \n\t"\ | |
1377 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ | |
1378 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
1379 "add %4, %1 \n\t"\ | |
1380 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ | |
1381 \ | |
1382 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ | |
1383 "add %4, %1 \n\t"\ | |
1384 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ | |
1385 \ | |
1386 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ | |
1387 "add %4, %1 \n\t"\ | |
1388 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ | |
1389 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
1390 \ | |
1391 "add $72, %0 \n\t"\ | |
1392 "add %6, %1 \n\t"\ | |
1393 "decl %2 \n\t"\ | |
1394 " jnz 1b \n\t"\ | |
1395 \ | |
1396 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ | |
1397 : "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)\ | |
1398 : "memory"\ | |
1399 );\ | |
1400 }\ | |
1401 \ | |
1402 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ | |
1403 OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\ | |
1404 }\ | |
1405 \ | |
1406 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1407 uint64_t temp[8];\ | |
1408 uint8_t * const half= (uint8_t*)temp;\ | |
1409 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
1410 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ | |
1411 }\ | |
1412 \ | |
1413 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1414 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ | |
1415 }\ | |
1416 \ | |
1417 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1418 uint64_t temp[8];\ | |
1419 uint8_t * const half= (uint8_t*)temp;\ | |
1420 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
1421 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\ | |
1422 }\ | |
1423 \ | |
1424 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1425 uint64_t temp[8];\ | |
1426 uint8_t * const half= (uint8_t*)temp;\ | |
1427 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ | |
1428 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ | |
1429 }\ | |
1430 \ | |
1431 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1432 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1433 }\ | |
1434 \ | |
1435 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1436 uint64_t temp[8];\ | |
1437 uint8_t * const half= (uint8_t*)temp;\ | |
1438 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ | |
1439 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\ | |
1440 }\ | |
1441 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1442 uint64_t half[8 + 9];\ | |
1443 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1444 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1445 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1446 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ | |
1447 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1448 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ | |
1449 }\ | |
1450 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1451 uint64_t half[8 + 9];\ | |
1452 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1453 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1454 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1455 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ | |
1456 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1457 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ | |
1458 }\ | |
1459 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1460 uint64_t half[8 + 9];\ | |
1461 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1462 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1463 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1464 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ | |
1465 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1466 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ | |
1467 }\ | |
1468 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1469 uint64_t half[8 + 9];\ | |
1470 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1471 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1472 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1473 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ | |
1474 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1475 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ | |
1476 }\ | |
1477 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1478 uint64_t half[8 + 9];\ | |
1479 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1480 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1481 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1482 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1483 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ | |
1484 }\ | |
1485 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1486 uint64_t half[8 + 9];\ | |
1487 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1488 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1489 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1490 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ | |
1491 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ | |
1492 }\ | |
1493 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1494 uint64_t half[8 + 9];\ | |
1495 uint8_t * const halfH= ((uint8_t*)half);\ | |
1496 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1497 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ | |
1498 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
1499 }\ | |
1500 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1501 uint64_t half[8 + 9];\ | |
1502 uint8_t * const halfH= ((uint8_t*)half);\ | |
1503 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1504 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ | |
1505 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
1506 }\ | |
1507 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1508 uint64_t half[9];\ | |
1509 uint8_t * const halfH= ((uint8_t*)half);\ | |
1510 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1511 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
1512 }\ | |
1513 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ | |
1514 OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\ | |
1515 }\ | |
1516 \ | |
1517 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1518 uint64_t temp[32];\ | |
1519 uint8_t * const half= (uint8_t*)temp;\ | |
1520 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
1521 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ | |
1522 }\ | |
1523 \ | |
1524 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1525 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ | |
1526 }\ | |
1527 \ | |
1528 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1529 uint64_t temp[32];\ | |
1530 uint8_t * const half= (uint8_t*)temp;\ | |
1531 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
1532 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\ | |
1533 }\ | |
1534 \ | |
1535 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1536 uint64_t temp[32];\ | |
1537 uint8_t * const half= (uint8_t*)temp;\ | |
1538 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ | |
1539 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ | |
1540 }\ | |
1541 \ | |
1542 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1543 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1544 }\ | |
1545 \ | |
1546 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1547 uint64_t temp[32];\ | |
1548 uint8_t * const half= (uint8_t*)temp;\ | |
1549 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ | |
1550 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\ | |
1551 }\ | |
1552 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1553 uint64_t half[16*2 + 17*2];\ | |
1554 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1555 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1556 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1557 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ | |
1558 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1559 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ | |
1560 }\ | |
1561 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1562 uint64_t half[16*2 + 17*2];\ | |
1563 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1564 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1565 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1566 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ | |
1567 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1568 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ | |
1569 }\ | |
1570 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1571 uint64_t half[16*2 + 17*2];\ | |
1572 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1573 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1574 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1575 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ | |
1576 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1577 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ | |
1578 }\ | |
1579 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1580 uint64_t half[16*2 + 17*2];\ | |
1581 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1582 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1583 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1584 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ | |
1585 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1586 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ | |
1587 }\ | |
1588 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1589 uint64_t half[16*2 + 17*2];\ | |
1590 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1591 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1592 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1593 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1594 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ | |
1595 }\ | |
1596 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1597 uint64_t half[16*2 + 17*2];\ | |
1598 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1599 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1600 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1601 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ | |
1602 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ | |
1603 }\ | |
1604 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1605 uint64_t half[17*2];\ | |
1606 uint8_t * const halfH= ((uint8_t*)half);\ | |
1607 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1608 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ | |
1609 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
1610 }\ | |
1611 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1612 uint64_t half[17*2];\ | |
1613 uint8_t * const halfH= ((uint8_t*)half);\ | |
1614 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1615 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ | |
1616 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
1617 }\ | |
1618 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1619 uint64_t half[17*2];\ | |
1620 uint8_t * const halfH= ((uint8_t*)half);\ | |
1621 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1622 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
1623 } | |
1624 | |
1625 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" | |
1626 #define AVG_3DNOW_OP(a,b,temp, size) \ | |
1627 "mov" #size " " #b ", " #temp " \n\t"\ | |
1628 "pavgusb " #temp ", " #a " \n\t"\ | |
1629 "mov" #size " " #a ", " #b " \n\t" | |
1630 #define AVG_MMX2_OP(a,b,temp, size) \ | |
1631 "mov" #size " " #b ", " #temp " \n\t"\ | |
1632 "pavgb " #temp ", " #a " \n\t"\ | |
1633 "mov" #size " " #a ", " #b " \n\t" | |
1634 | |
1635 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
1636 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
1637 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
1638 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
1639 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
1640 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
1641 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) | |
1642 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) | |
1643 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) | |
1644 | |
1645 /***********************************/ | |
1646 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */ | |
1647 | |
1648 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\ | |
1649 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1650 OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\ | |
1651 } | |
1652 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\ | |
1653 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1654 OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\ | |
1655 } | |
1656 | |
1657 #define QPEL_2TAP(OPNAME, SIZE, MMX)\ | |
1658 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\ | |
1659 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\ | |
1660 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\ | |
1661 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\ | |
1662 OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\ | |
1663 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\ | |
1664 OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\ | |
1665 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\ | |
1666 OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\ | |
1667 static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1668 OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\ | |
1669 }\ | |
1670 static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1671 OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\ | |
1672 }\ | |
1673 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0, 1, 0)\ | |
1674 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1, -1, 0)\ | |
1675 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0, stride, 0)\ | |
1676 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride, -stride, 0)\ | |
1677 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0, stride, 1)\ | |
1678 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1, stride, -1)\ | |
1679 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride, -stride, 1)\ | |
1680 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\ | |
1681 | |
1682 QPEL_2TAP(put_, 16, mmx2) | |
1683 QPEL_2TAP(avg_, 16, mmx2) | |
1684 QPEL_2TAP(put_, 8, mmx2) | |
1685 QPEL_2TAP(avg_, 8, mmx2) | |
1686 QPEL_2TAP(put_, 16, 3dnow) | |
1687 QPEL_2TAP(avg_, 16, 3dnow) | |
1688 QPEL_2TAP(put_, 8, 3dnow) | |
1689 QPEL_2TAP(avg_, 8, 3dnow) | |
1690 | |
1691 | |
1692 #if 0 | |
8527
f8bf438c6000
Add missing 'void' keyword to parameterless function declarations.
diego
parents:
8519
diff
changeset
|
1693 static void just_return(void) { return; } |
8430 | 1694 #endif |
1695 | |
1696 static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy, | |
1697 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){ | |
1698 const int w = 8; | |
1699 const int ix = ox>>(16+shift); | |
1700 const int iy = oy>>(16+shift); | |
1701 const int oxs = ox>>4; | |
1702 const int oys = oy>>4; | |
1703 const int dxxs = dxx>>4; | |
1704 const int dxys = dxy>>4; | |
1705 const int dyxs = dyx>>4; | |
1706 const int dyys = dyy>>4; | |
1707 const uint16_t r4[4] = {r,r,r,r}; | |
1708 const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys}; | |
1709 const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys}; | |
1710 const uint64_t shift2 = 2*shift; | |
1711 uint8_t edge_buf[(h+1)*stride]; | |
1712 int x, y; | |
1713 | |
1714 const int dxw = (dxx-(1<<(16+shift)))*(w-1); | |
1715 const int dyh = (dyy-(1<<(16+shift)))*(h-1); | |
1716 const int dxh = dxy*(h-1); | |
1717 const int dyw = dyx*(w-1); | |
1718 if( // non-constant fullpel offset (3% of blocks) | |
1719 ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) | | |
1720 (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift) | |
1721 // uses more than 16 bits of subpel mv (only at huge resolution) | |
1722 || (dxx|dxy|dyx|dyy)&15 ) | |
1723 { | |
1724 //FIXME could still use mmx for some of the rows | |
1725 ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height); | |
1726 return; | |
1727 } | |
1728 | |
1729 src += ix + iy*stride; | |
1730 if( (unsigned)ix >= width-w || | |
1731 (unsigned)iy >= height-h ) | |
1732 { | |
1733 ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height); | |
1734 src = edge_buf; | |
1735 } | |
1736 | |
1737 __asm__ volatile( | |
1738 "movd %0, %%mm6 \n\t" | |
1739 "pxor %%mm7, %%mm7 \n\t" | |
1740 "punpcklwd %%mm6, %%mm6 \n\t" | |
1741 "punpcklwd %%mm6, %%mm6 \n\t" | |
1742 :: "r"(1<<shift) | |
1743 ); | |
1744 | |
1745 for(x=0; x<w; x+=4){ | |
1746 uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0), | |
1747 oxs - dxys + dxxs*(x+1), | |
1748 oxs - dxys + dxxs*(x+2), | |
1749 oxs - dxys + dxxs*(x+3) }; | |
1750 uint16_t dy4[4] = { oys - dyys + dyxs*(x+0), | |
1751 oys - dyys + dyxs*(x+1), | |
1752 oys - dyys + dyxs*(x+2), | |
1753 oys - dyys + dyxs*(x+3) }; | |
1754 | |
1755 for(y=0; y<h; y++){ | |
1756 __asm__ volatile( | |
1757 "movq %0, %%mm4 \n\t" | |
1758 "movq %1, %%mm5 \n\t" | |
1759 "paddw %2, %%mm4 \n\t" | |
1760 "paddw %3, %%mm5 \n\t" | |
1761 "movq %%mm4, %0 \n\t" | |
1762 "movq %%mm5, %1 \n\t" | |
1763 "psrlw $12, %%mm4 \n\t" | |
1764 "psrlw $12, %%mm5 \n\t" | |
1765 : "+m"(*dx4), "+m"(*dy4) | |
1766 : "m"(*dxy4), "m"(*dyy4) | |
1767 ); | |
1768 | |
1769 __asm__ volatile( | |
1770 "movq %%mm6, %%mm2 \n\t" | |
1771 "movq %%mm6, %%mm1 \n\t" | |
1772 "psubw %%mm4, %%mm2 \n\t" | |
1773 "psubw %%mm5, %%mm1 \n\t" | |
1774 "movq %%mm2, %%mm0 \n\t" | |
1775 "movq %%mm4, %%mm3 \n\t" | |
1776 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy) | |
1777 "pmullw %%mm5, %%mm3 \n\t" // dx*dy | |
1778 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy | |
1779 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy) | |
1780 | |
1781 "movd %4, %%mm5 \n\t" | |
1782 "movd %3, %%mm4 \n\t" | |
1783 "punpcklbw %%mm7, %%mm5 \n\t" | |
1784 "punpcklbw %%mm7, %%mm4 \n\t" | |
1785 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy | |
1786 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy | |
1787 | |
1788 "movd %2, %%mm5 \n\t" | |
1789 "movd %1, %%mm4 \n\t" | |
1790 "punpcklbw %%mm7, %%mm5 \n\t" | |
1791 "punpcklbw %%mm7, %%mm4 \n\t" | |
1792 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy) | |
1793 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy) | |
1794 "paddw %5, %%mm1 \n\t" | |
1795 "paddw %%mm3, %%mm2 \n\t" | |
1796 "paddw %%mm1, %%mm0 \n\t" | |
1797 "paddw %%mm2, %%mm0 \n\t" | |
1798 | |
1799 "psrlw %6, %%mm0 \n\t" | |
1800 "packuswb %%mm0, %%mm0 \n\t" | |
1801 "movd %%mm0, %0 \n\t" | |
1802 | |
1803 : "=m"(dst[x+y*stride]) | |
1804 : "m"(src[0]), "m"(src[1]), | |
1805 "m"(src[stride]), "m"(src[stride+1]), | |
1806 "m"(*r4), "m"(shift2) | |
1807 ); | |
1808 src += stride; | |
1809 } | |
1810 src += 4-h*stride; | |
1811 } | |
1812 } | |
1813 | |
1814 #define PREFETCH(name, op) \ | |
1815 static void name(void *mem, int stride, int h){\ | |
1816 const uint8_t *p= mem;\ | |
1817 do{\ | |
1818 __asm__ volatile(#op" %0" :: "m"(*p));\ | |
1819 p+= stride;\ | |
1820 }while(--h);\ | |
1821 } | |
1822 PREFETCH(prefetch_mmx2, prefetcht0) | |
1823 PREFETCH(prefetch_3dnow, prefetch) | |
1824 #undef PREFETCH | |
1825 | |
1826 #include "h264dsp_mmx.c" | |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
1827 #include "rv40dsp_mmx.c" |
8430 | 1828 |
1829 /* CAVS specific */ | |
1830 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1831 put_pixels8_mmx(dst, src, stride, 8); | |
1832 } | |
1833 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1834 avg_pixels8_mmx(dst, src, stride, 8); | |
1835 } | |
1836 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1837 put_pixels16_mmx(dst, src, stride, 16); | |
1838 } | |
1839 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1840 avg_pixels16_mmx(dst, src, stride, 16); | |
1841 } | |
1842 | |
1843 /* VC1 specific */ | |
1844 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { | |
1845 put_pixels8_mmx(dst, src, stride, 8); | |
1846 } | |
9441 | 1847 void ff_avg_vc1_mspel_mc00_mmx2(uint8_t *dst, const uint8_t *src, int stride, int rnd) { |
1848 avg_pixels8_mmx2(dst, src, stride, 8); | |
1849 } | |
8430 | 1850 |
1851 /* XXX: those functions should be suppressed ASAP when all IDCTs are | |
1852 converted */ | |
8590 | 1853 #if CONFIG_GPL |
8430 | 1854 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) |
1855 { | |
1856 ff_mmx_idct (block); | |
1857 put_pixels_clamped_mmx(block, dest, line_size); | |
1858 } | |
1859 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1860 { | |
1861 ff_mmx_idct (block); | |
1862 add_pixels_clamped_mmx(block, dest, line_size); | |
1863 } | |
1864 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1865 { | |
1866 ff_mmxext_idct (block); | |
1867 put_pixels_clamped_mmx(block, dest, line_size); | |
1868 } | |
1869 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1870 { | |
1871 ff_mmxext_idct (block); | |
1872 add_pixels_clamped_mmx(block, dest, line_size); | |
1873 } | |
1874 #endif | |
1875 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1876 { | |
1877 ff_idct_xvid_mmx (block); | |
1878 put_pixels_clamped_mmx(block, dest, line_size); | |
1879 } | |
1880 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1881 { | |
1882 ff_idct_xvid_mmx (block); | |
1883 add_pixels_clamped_mmx(block, dest, line_size); | |
1884 } | |
1885 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1886 { | |
1887 ff_idct_xvid_mmx2 (block); | |
1888 put_pixels_clamped_mmx(block, dest, line_size); | |
1889 } | |
1890 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1891 { | |
1892 ff_idct_xvid_mmx2 (block); | |
1893 add_pixels_clamped_mmx(block, dest, line_size); | |
1894 } | |
1895 | |
1896 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize) | |
1897 { | |
1898 int i; | |
1899 __asm__ volatile("pxor %%mm7, %%mm7":); | |
1900 for(i=0; i<blocksize; i+=2) { | |
1901 __asm__ volatile( | |
1902 "movq %0, %%mm0 \n\t" | |
1903 "movq %1, %%mm1 \n\t" | |
1904 "movq %%mm0, %%mm2 \n\t" | |
1905 "movq %%mm1, %%mm3 \n\t" | |
1906 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0 | |
1907 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0 | |
1908 "pslld $31, %%mm2 \n\t" // keep only the sign bit | |
1909 "pxor %%mm2, %%mm1 \n\t" | |
1910 "movq %%mm3, %%mm4 \n\t" | |
1911 "pand %%mm1, %%mm3 \n\t" | |
1912 "pandn %%mm1, %%mm4 \n\t" | |
1913 "pfadd %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m))) | |
1914 "pfsub %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m))) | |
1915 "movq %%mm3, %1 \n\t" | |
1916 "movq %%mm0, %0 \n\t" | |
1917 :"+m"(mag[i]), "+m"(ang[i]) | |
1918 ::"memory" | |
1919 ); | |
1920 } | |
1921 __asm__ volatile("femms"); | |
1922 } | |
1923 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize) | |
1924 { | |
1925 int i; | |
1926 | |
1927 __asm__ volatile( | |
1928 "movaps %0, %%xmm5 \n\t" | |
1929 ::"m"(ff_pdw_80000000[0]) | |
1930 ); | |
1931 for(i=0; i<blocksize; i+=4) { | |
1932 __asm__ volatile( | |
1933 "movaps %0, %%xmm0 \n\t" | |
1934 "movaps %1, %%xmm1 \n\t" | |
1935 "xorps %%xmm2, %%xmm2 \n\t" | |
1936 "xorps %%xmm3, %%xmm3 \n\t" | |
1937 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0 | |
1938 "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0 | |
1939 "andps %%xmm5, %%xmm2 \n\t" // keep only the sign bit | |
1940 "xorps %%xmm2, %%xmm1 \n\t" | |
1941 "movaps %%xmm3, %%xmm4 \n\t" | |
1942 "andps %%xmm1, %%xmm3 \n\t" | |
1943 "andnps %%xmm1, %%xmm4 \n\t" | |
1944 "addps %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m))) | |
1945 "subps %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m))) | |
1946 "movaps %%xmm3, %1 \n\t" | |
1947 "movaps %%xmm0, %0 \n\t" | |
1948 :"+m"(mag[i]), "+m"(ang[i]) | |
1949 ::"memory" | |
1950 ); | |
1951 } | |
1952 } | |
1953 | |
1954 #define IF1(x) x | |
1955 #define IF0(x) | |
1956 | |
1957 #define MIX5(mono,stereo)\ | |
1958 __asm__ volatile(\ | |
1959 "movss 0(%2), %%xmm5 \n"\ | |
1960 "movss 8(%2), %%xmm6 \n"\ | |
1961 "movss 24(%2), %%xmm7 \n"\ | |
1962 "shufps $0, %%xmm5, %%xmm5 \n"\ | |
1963 "shufps $0, %%xmm6, %%xmm6 \n"\ | |
1964 "shufps $0, %%xmm7, %%xmm7 \n"\ | |
1965 "1: \n"\ | |
1966 "movaps (%0,%1), %%xmm0 \n"\ | |
1967 "movaps 0x400(%0,%1), %%xmm1 \n"\ | |
1968 "movaps 0x800(%0,%1), %%xmm2 \n"\ | |
1969 "movaps 0xc00(%0,%1), %%xmm3 \n"\ | |
1970 "movaps 0x1000(%0,%1), %%xmm4 \n"\ | |
1971 "mulps %%xmm5, %%xmm0 \n"\ | |
1972 "mulps %%xmm6, %%xmm1 \n"\ | |
1973 "mulps %%xmm5, %%xmm2 \n"\ | |
1974 "mulps %%xmm7, %%xmm3 \n"\ | |
1975 "mulps %%xmm7, %%xmm4 \n"\ | |
1976 stereo("addps %%xmm1, %%xmm0 \n")\ | |
1977 "addps %%xmm1, %%xmm2 \n"\ | |
1978 "addps %%xmm3, %%xmm0 \n"\ | |
1979 "addps %%xmm4, %%xmm2 \n"\ | |
1980 mono("addps %%xmm2, %%xmm0 \n")\ | |
1981 "movaps %%xmm0, (%0,%1) \n"\ | |
1982 stereo("movaps %%xmm2, 0x400(%0,%1) \n")\ | |
1983 "add $16, %0 \n"\ | |
1984 "jl 1b \n"\ | |
1985 :"+&r"(i)\ | |
1986 :"r"(samples[0]+len), "r"(matrix)\ | |
1987 :"memory"\ | |
1988 ); | |
1989 | |
1990 #define MIX_MISC(stereo)\ | |
1991 __asm__ volatile(\ | |
1992 "1: \n"\ | |
1993 "movaps (%3,%0), %%xmm0 \n"\ | |
1994 stereo("movaps %%xmm0, %%xmm1 \n")\ | |
1995 "mulps %%xmm6, %%xmm0 \n"\ | |
1996 stereo("mulps %%xmm7, %%xmm1 \n")\ | |
1997 "lea 1024(%3,%0), %1 \n"\ | |
1998 "mov %5, %2 \n"\ | |
1999 "2: \n"\ | |
2000 "movaps (%1), %%xmm2 \n"\ | |
2001 stereo("movaps %%xmm2, %%xmm3 \n")\ | |
2002 "mulps (%4,%2), %%xmm2 \n"\ | |
2003 stereo("mulps 16(%4,%2), %%xmm3 \n")\ | |
2004 "addps %%xmm2, %%xmm0 \n"\ | |
2005 stereo("addps %%xmm3, %%xmm1 \n")\ | |
2006 "add $1024, %1 \n"\ | |
2007 "add $32, %2 \n"\ | |
2008 "jl 2b \n"\ | |
2009 "movaps %%xmm0, (%3,%0) \n"\ | |
2010 stereo("movaps %%xmm1, 1024(%3,%0) \n")\ | |
2011 "add $16, %0 \n"\ | |
2012 "jl 1b \n"\ | |
2013 :"+&r"(i), "=&r"(j), "=&r"(k)\ | |
2014 :"r"(samples[0]+len), "r"(matrix_simd+in_ch), "g"((intptr_t)-32*(in_ch-1))\ | |
2015 :"memory"\ | |
2016 ); | |
2017 | |
2018 static void ac3_downmix_sse(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len) | |
2019 { | |
2020 int (*matrix_cmp)[2] = (int(*)[2])matrix; | |
2021 intptr_t i,j,k; | |
2022 | |
2023 i = -len*sizeof(float); | |
2024 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]))) { | |
2025 MIX5(IF0,IF1); | |
2026 } else if(in_ch == 5 && out_ch == 1 && matrix_cmp[0][0]==matrix_cmp[2][0] && matrix_cmp[3][0]==matrix_cmp[4][0]) { | |
2027 MIX5(IF1,IF0); | |
2028 } else { | |
11369 | 2029 DECLARE_ALIGNED(16, float, matrix_simd)[in_ch][2][4]; |
8430 | 2030 j = 2*in_ch*sizeof(float); |
2031 __asm__ volatile( | |
2032 "1: \n" | |
2033 "sub $8, %0 \n" | |
2034 "movss (%2,%0), %%xmm6 \n" | |
2035 "movss 4(%2,%0), %%xmm7 \n" | |
2036 "shufps $0, %%xmm6, %%xmm6 \n" | |
2037 "shufps $0, %%xmm7, %%xmm7 \n" | |
2038 "movaps %%xmm6, (%1,%0,4) \n" | |
2039 "movaps %%xmm7, 16(%1,%0,4) \n" | |
2040 "jg 1b \n" | |
2041 :"+&r"(j) | |
2042 :"r"(matrix_simd), "r"(matrix) | |
2043 :"memory" | |
2044 ); | |
2045 if(out_ch == 2) { | |
2046 MIX_MISC(IF1); | |
2047 } else { | |
2048 MIX_MISC(IF0); | |
2049 } | |
2050 } | |
2051 } | |
2052 | |
2053 static void vector_fmul_3dnow(float *dst, const float *src, int len){ | |
2054 x86_reg i = (len-4)*4; | |
2055 __asm__ volatile( | |
2056 "1: \n\t" | |
2057 "movq (%1,%0), %%mm0 \n\t" | |
2058 "movq 8(%1,%0), %%mm1 \n\t" | |
2059 "pfmul (%2,%0), %%mm0 \n\t" | |
2060 "pfmul 8(%2,%0), %%mm1 \n\t" | |
2061 "movq %%mm0, (%1,%0) \n\t" | |
2062 "movq %%mm1, 8(%1,%0) \n\t" | |
2063 "sub $16, %0 \n\t" | |
2064 "jge 1b \n\t" | |
2065 "femms \n\t" | |
2066 :"+r"(i) | |
2067 :"r"(dst), "r"(src) | |
2068 :"memory" | |
2069 ); | |
2070 } | |
2071 static void vector_fmul_sse(float *dst, const float *src, int len){ | |
2072 x86_reg i = (len-8)*4; | |
2073 __asm__ volatile( | |
2074 "1: \n\t" | |
2075 "movaps (%1,%0), %%xmm0 \n\t" | |
2076 "movaps 16(%1,%0), %%xmm1 \n\t" | |
2077 "mulps (%2,%0), %%xmm0 \n\t" | |
2078 "mulps 16(%2,%0), %%xmm1 \n\t" | |
2079 "movaps %%xmm0, (%1,%0) \n\t" | |
2080 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2081 "sub $32, %0 \n\t" | |
2082 "jge 1b \n\t" | |
2083 :"+r"(i) | |
2084 :"r"(dst), "r"(src) | |
2085 :"memory" | |
2086 ); | |
2087 } | |
2088 | |
2089 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){ | |
2090 x86_reg i = len*4-16; | |
2091 __asm__ volatile( | |
2092 "1: \n\t" | |
2093 "pswapd 8(%1), %%mm0 \n\t" | |
2094 "pswapd (%1), %%mm1 \n\t" | |
2095 "pfmul (%3,%0), %%mm0 \n\t" | |
2096 "pfmul 8(%3,%0), %%mm1 \n\t" | |
2097 "movq %%mm0, (%2,%0) \n\t" | |
2098 "movq %%mm1, 8(%2,%0) \n\t" | |
2099 "add $16, %1 \n\t" | |
2100 "sub $16, %0 \n\t" | |
2101 "jge 1b \n\t" | |
2102 :"+r"(i), "+r"(src1) | |
2103 :"r"(dst), "r"(src0) | |
2104 ); | |
2105 __asm__ volatile("femms"); | |
2106 } | |
2107 static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){ | |
2108 x86_reg i = len*4-32; | |
2109 __asm__ volatile( | |
2110 "1: \n\t" | |
2111 "movaps 16(%1), %%xmm0 \n\t" | |
2112 "movaps (%1), %%xmm1 \n\t" | |
2113 "shufps $0x1b, %%xmm0, %%xmm0 \n\t" | |
2114 "shufps $0x1b, %%xmm1, %%xmm1 \n\t" | |
2115 "mulps (%3,%0), %%xmm0 \n\t" | |
2116 "mulps 16(%3,%0), %%xmm1 \n\t" | |
2117 "movaps %%xmm0, (%2,%0) \n\t" | |
2118 "movaps %%xmm1, 16(%2,%0) \n\t" | |
2119 "add $32, %1 \n\t" | |
2120 "sub $32, %0 \n\t" | |
2121 "jge 1b \n\t" | |
2122 :"+r"(i), "+r"(src1) | |
2123 :"r"(dst), "r"(src0) | |
2124 ); | |
2125 } | |
2126 | |
10300
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2127 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
|
2128 const float *src2, int len){ |
8430 | 2129 x86_reg i = (len-4)*4; |
10301 | 2130 __asm__ volatile( |
2131 "1: \n\t" | |
2132 "movq (%2,%0), %%mm0 \n\t" | |
2133 "movq 8(%2,%0), %%mm1 \n\t" | |
2134 "pfmul (%3,%0), %%mm0 \n\t" | |
2135 "pfmul 8(%3,%0), %%mm1 \n\t" | |
2136 "pfadd (%4,%0), %%mm0 \n\t" | |
2137 "pfadd 8(%4,%0), %%mm1 \n\t" | |
2138 "movq %%mm0, (%1,%0) \n\t" | |
2139 "movq %%mm1, 8(%1,%0) \n\t" | |
2140 "sub $16, %0 \n\t" | |
2141 "jge 1b \n\t" | |
2142 :"+r"(i) | |
2143 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
2144 :"memory" | |
2145 ); | |
8430 | 2146 __asm__ volatile("femms"); |
2147 } | |
10300
4d1b9ca628fc
Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents:
10107
diff
changeset
|
2148 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
|
2149 const float *src2, int len){ |
8430 | 2150 x86_reg i = (len-8)*4; |
10301 | 2151 __asm__ volatile( |
2152 "1: \n\t" | |
2153 "movaps (%2,%0), %%xmm0 \n\t" | |
2154 "movaps 16(%2,%0), %%xmm1 \n\t" | |
2155 "mulps (%3,%0), %%xmm0 \n\t" | |
2156 "mulps 16(%3,%0), %%xmm1 \n\t" | |
2157 "addps (%4,%0), %%xmm0 \n\t" | |
2158 "addps 16(%4,%0), %%xmm1 \n\t" | |
2159 "movaps %%xmm0, (%1,%0) \n\t" | |
2160 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2161 "sub $32, %0 \n\t" | |
2162 "jge 1b \n\t" | |
2163 :"+r"(i) | |
2164 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
2165 :"memory" | |
2166 ); | |
8430 | 2167 } |
2168 | |
2169 static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1, | |
2170 const float *win, float add_bias, int len){ | |
8590 | 2171 #if HAVE_6REGS |
8430 | 2172 if(add_bias == 0){ |
2173 x86_reg i = -len*4; | |
2174 x86_reg j = len*4-8; | |
2175 __asm__ volatile( | |
2176 "1: \n" | |
2177 "pswapd (%5,%1), %%mm1 \n" | |
2178 "movq (%5,%0), %%mm0 \n" | |
2179 "pswapd (%4,%1), %%mm5 \n" | |
2180 "movq (%3,%0), %%mm4 \n" | |
2181 "movq %%mm0, %%mm2 \n" | |
2182 "movq %%mm1, %%mm3 \n" | |
2183 "pfmul %%mm4, %%mm2 \n" // src0[len+i]*win[len+i] | |
2184 "pfmul %%mm5, %%mm3 \n" // src1[ j]*win[len+j] | |
2185 "pfmul %%mm4, %%mm1 \n" // src0[len+i]*win[len+j] | |
2186 "pfmul %%mm5, %%mm0 \n" // src1[ j]*win[len+i] | |
2187 "pfadd %%mm3, %%mm2 \n" | |
2188 "pfsub %%mm0, %%mm1 \n" | |
2189 "pswapd %%mm2, %%mm2 \n" | |
2190 "movq %%mm1, (%2,%0) \n" | |
2191 "movq %%mm2, (%2,%1) \n" | |
2192 "sub $8, %1 \n" | |
2193 "add $8, %0 \n" | |
2194 "jl 1b \n" | |
2195 "femms \n" | |
2196 :"+r"(i), "+r"(j) | |
2197 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len) | |
2198 ); | |
2199 }else | |
2200 #endif | |
2201 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len); | |
2202 } | |
2203 | |
2204 static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1, | |
2205 const float *win, float add_bias, int len){ | |
8590 | 2206 #if HAVE_6REGS |
8430 | 2207 if(add_bias == 0){ |
2208 x86_reg i = -len*4; | |
2209 x86_reg j = len*4-16; | |
2210 __asm__ volatile( | |
2211 "1: \n" | |
2212 "movaps (%5,%1), %%xmm1 \n" | |
2213 "movaps (%5,%0), %%xmm0 \n" | |
2214 "movaps (%4,%1), %%xmm5 \n" | |
2215 "movaps (%3,%0), %%xmm4 \n" | |
2216 "shufps $0x1b, %%xmm1, %%xmm1 \n" | |
2217 "shufps $0x1b, %%xmm5, %%xmm5 \n" | |
2218 "movaps %%xmm0, %%xmm2 \n" | |
2219 "movaps %%xmm1, %%xmm3 \n" | |
2220 "mulps %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i] | |
2221 "mulps %%xmm5, %%xmm3 \n" // src1[ j]*win[len+j] | |
2222 "mulps %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j] | |
2223 "mulps %%xmm5, %%xmm0 \n" // src1[ j]*win[len+i] | |
2224 "addps %%xmm3, %%xmm2 \n" | |
2225 "subps %%xmm0, %%xmm1 \n" | |
2226 "shufps $0x1b, %%xmm2, %%xmm2 \n" | |
2227 "movaps %%xmm1, (%2,%0) \n" | |
2228 "movaps %%xmm2, (%2,%1) \n" | |
2229 "sub $16, %1 \n" | |
2230 "add $16, %0 \n" | |
2231 "jl 1b \n" | |
2232 :"+r"(i), "+r"(j) | |
2233 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len) | |
2234 ); | |
2235 }else | |
2236 #endif | |
2237 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len); | |
2238 } | |
2239 | |
2240 static void int32_to_float_fmul_scalar_sse(float *dst, const int *src, float mul, int len) | |
2241 { | |
2242 x86_reg i = -4*len; | |
2243 __asm__ volatile( | |
2244 "movss %3, %%xmm4 \n" | |
2245 "shufps $0, %%xmm4, %%xmm4 \n" | |
2246 "1: \n" | |
2247 "cvtpi2ps (%2,%0), %%xmm0 \n" | |
2248 "cvtpi2ps 8(%2,%0), %%xmm1 \n" | |
2249 "cvtpi2ps 16(%2,%0), %%xmm2 \n" | |
2250 "cvtpi2ps 24(%2,%0), %%xmm3 \n" | |
2251 "movlhps %%xmm1, %%xmm0 \n" | |
2252 "movlhps %%xmm3, %%xmm2 \n" | |
2253 "mulps %%xmm4, %%xmm0 \n" | |
2254 "mulps %%xmm4, %%xmm2 \n" | |
2255 "movaps %%xmm0, (%1,%0) \n" | |
2256 "movaps %%xmm2, 16(%1,%0) \n" | |
2257 "add $32, %0 \n" | |
2258 "jl 1b \n" | |
2259 :"+r"(i) | |
2260 :"r"(dst+len), "r"(src+len), "m"(mul) | |
2261 ); | |
2262 } | |
2263 | |
2264 static void int32_to_float_fmul_scalar_sse2(float *dst, const int *src, float mul, int len) | |
2265 { | |
2266 x86_reg i = -4*len; | |
2267 __asm__ volatile( | |
2268 "movss %3, %%xmm4 \n" | |
2269 "shufps $0, %%xmm4, %%xmm4 \n" | |
2270 "1: \n" | |
2271 "cvtdq2ps (%2,%0), %%xmm0 \n" | |
2272 "cvtdq2ps 16(%2,%0), %%xmm1 \n" | |
2273 "mulps %%xmm4, %%xmm0 \n" | |
2274 "mulps %%xmm4, %%xmm1 \n" | |
2275 "movaps %%xmm0, (%1,%0) \n" | |
2276 "movaps %%xmm1, 16(%1,%0) \n" | |
2277 "add $32, %0 \n" | |
2278 "jl 1b \n" | |
2279 :"+r"(i) | |
2280 :"r"(dst+len), "r"(src+len), "m"(mul) | |
2281 ); | |
2282 } | |
2283 | |
10105 | 2284 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
|
2285 int len) |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2286 { |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2287 x86_reg i = (len-16)*4; |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2288 __asm__ volatile( |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2289 "movss %3, %%xmm4 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2290 "movss %4, %%xmm5 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2291 "shufps $0, %%xmm4, %%xmm4 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2292 "shufps $0, %%xmm5, %%xmm5 \n" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2293 "1: \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2294 "movaps (%2,%0), %%xmm0 \n\t" // 3/1 on intel |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2295 "movaps 16(%2,%0), %%xmm1 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2296 "movaps 32(%2,%0), %%xmm2 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2297 "movaps 48(%2,%0), %%xmm3 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2298 "maxps %%xmm4, %%xmm0 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2299 "maxps %%xmm4, %%xmm1 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2300 "maxps %%xmm4, %%xmm2 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2301 "maxps %%xmm4, %%xmm3 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2302 "minps %%xmm5, %%xmm0 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2303 "minps %%xmm5, %%xmm1 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2304 "minps %%xmm5, %%xmm2 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2305 "minps %%xmm5, %%xmm3 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2306 "movaps %%xmm0, (%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2307 "movaps %%xmm1, 16(%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2308 "movaps %%xmm2, 32(%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2309 "movaps %%xmm3, 48(%1,%0) \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2310 "sub $64, %0 \n\t" |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2311 "jge 1b \n\t" |
10107
3b61bc6ce377
Mark "i" parameter of vector_clipf_sse() as early-clobber
vitor
parents:
10105
diff
changeset
|
2312 :"+&r"(i) |
10104
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2313 :"r"(dst), "r"(src), "m"(min), "m"(max) |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2314 :"memory" |
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 } |
0fa3d21b317e
SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents:
9995
diff
changeset
|
2317 |
8430 | 2318 static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){ |
2319 x86_reg reglen = len; | |
2320 // not bit-exact: pf2id uses different rounding than C and SSE | |
2321 __asm__ volatile( | |
2322 "add %0 , %0 \n\t" | |
2323 "lea (%2,%0,2) , %2 \n\t" | |
2324 "add %0 , %1 \n\t" | |
2325 "neg %0 \n\t" | |
2326 "1: \n\t" | |
2327 "pf2id (%2,%0,2) , %%mm0 \n\t" | |
2328 "pf2id 8(%2,%0,2) , %%mm1 \n\t" | |
2329 "pf2id 16(%2,%0,2) , %%mm2 \n\t" | |
2330 "pf2id 24(%2,%0,2) , %%mm3 \n\t" | |
2331 "packssdw %%mm1 , %%mm0 \n\t" | |
2332 "packssdw %%mm3 , %%mm2 \n\t" | |
2333 "movq %%mm0 , (%1,%0) \n\t" | |
2334 "movq %%mm2 , 8(%1,%0) \n\t" | |
2335 "add $16 , %0 \n\t" | |
2336 " js 1b \n\t" | |
2337 "femms \n\t" | |
2338 :"+r"(reglen), "+r"(dst), "+r"(src) | |
2339 ); | |
2340 } | |
2341 static void float_to_int16_sse(int16_t *dst, const float *src, long len){ | |
2342 x86_reg reglen = len; | |
2343 __asm__ volatile( | |
2344 "add %0 , %0 \n\t" | |
2345 "lea (%2,%0,2) , %2 \n\t" | |
2346 "add %0 , %1 \n\t" | |
2347 "neg %0 \n\t" | |
2348 "1: \n\t" | |
2349 "cvtps2pi (%2,%0,2) , %%mm0 \n\t" | |
2350 "cvtps2pi 8(%2,%0,2) , %%mm1 \n\t" | |
2351 "cvtps2pi 16(%2,%0,2) , %%mm2 \n\t" | |
2352 "cvtps2pi 24(%2,%0,2) , %%mm3 \n\t" | |
2353 "packssdw %%mm1 , %%mm0 \n\t" | |
2354 "packssdw %%mm3 , %%mm2 \n\t" | |
2355 "movq %%mm0 , (%1,%0) \n\t" | |
2356 "movq %%mm2 , 8(%1,%0) \n\t" | |
2357 "add $16 , %0 \n\t" | |
2358 " js 1b \n\t" | |
2359 "emms \n\t" | |
2360 :"+r"(reglen), "+r"(dst), "+r"(src) | |
2361 ); | |
2362 } | |
2363 | |
2364 static void float_to_int16_sse2(int16_t *dst, const float *src, long len){ | |
2365 x86_reg reglen = len; | |
2366 __asm__ volatile( | |
2367 "add %0 , %0 \n\t" | |
2368 "lea (%2,%0,2) , %2 \n\t" | |
2369 "add %0 , %1 \n\t" | |
2370 "neg %0 \n\t" | |
2371 "1: \n\t" | |
2372 "cvtps2dq (%2,%0,2) , %%xmm0 \n\t" | |
2373 "cvtps2dq 16(%2,%0,2) , %%xmm1 \n\t" | |
2374 "packssdw %%xmm1 , %%xmm0 \n\t" | |
2375 "movdqa %%xmm0 , (%1,%0) \n\t" | |
2376 "add $16 , %0 \n\t" | |
2377 " js 1b \n\t" | |
2378 :"+r"(reglen), "+r"(dst), "+r"(src) | |
2379 ); | |
2380 } | |
2381 | |
2382 void ff_float_to_int16_interleave6_sse(int16_t *dst, const float **src, int len); | |
2383 void ff_float_to_int16_interleave6_3dnow(int16_t *dst, const float **src, int len); | |
2384 void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len); | |
11981 | 2385 int32_t ff_scalarproduct_int16_mmx2(const int16_t *v1, const int16_t *v2, int order, int shift); |
2386 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2, int order, int shift); | |
2387 int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul); | |
2388 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul); | |
2389 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul); | |
10431 | 2390 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); |
2391 int ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left); | |
2392 int ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left); | |
8430 | 2393 void ff_x264_deblock_v_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); |
2394 void ff_x264_deblock_h_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); | |
10645 | 2395 void ff_x264_deblock_h_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta); |
2396 void ff_x264_deblock_v_luma_intra_sse2(uint8_t *pix, int stride, int alpha, int beta); | |
2397 void ff_x264_deblock_h_luma_intra_sse2(uint8_t *pix, int stride, int alpha, int beta); | |
2398 | |
2399 #if HAVE_YASM && ARCH_X86_32 | |
8430 | 2400 void ff_x264_deblock_v8_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta); |
2401 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta) | |
2402 { | |
2403 ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta); | |
2404 ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta); | |
2405 } | |
10645 | 2406 #elif !HAVE_YASM |
8430 | 2407 #define ff_float_to_int16_interleave6_sse(a,b,c) float_to_int16_interleave_misc_sse(a,b,c,6) |
2408 #define ff_float_to_int16_interleave6_3dnow(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6) | |
2409 #define ff_float_to_int16_interleave6_3dn2(a,b,c) float_to_int16_interleave_misc_3dnow(a,b,c,6) | |
2410 #endif | |
2411 #define ff_float_to_int16_interleave6_sse2 ff_float_to_int16_interleave6_sse | |
2412 | |
2413 #define FLOAT_TO_INT16_INTERLEAVE(cpu, body) \ | |
2414 /* gcc pessimizes register allocation if this is in the same function as float_to_int16_interleave_sse2*/\ | |
2415 static av_noinline void float_to_int16_interleave_misc_##cpu(int16_t *dst, const float **src, long len, int channels){\ | |
11369 | 2416 DECLARE_ALIGNED(16, int16_t, tmp)[len];\ |
8430 | 2417 int i,j,c;\ |
2418 for(c=0; c<channels; c++){\ | |
2419 float_to_int16_##cpu(tmp, src[c], len);\ | |
2420 for(i=0, j=c; i<len; i++, j+=channels)\ | |
2421 dst[j] = tmp[i];\ | |
2422 }\ | |
2423 }\ | |
2424 \ | |
2425 static void float_to_int16_interleave_##cpu(int16_t *dst, const float **src, long len, int channels){\ | |
2426 if(channels==1)\ | |
2427 float_to_int16_##cpu(dst, src[0], len);\ | |
2428 else if(channels==2){\ | |
2429 x86_reg reglen = len; \ | |
2430 const float *src0 = src[0];\ | |
2431 const float *src1 = src[1];\ | |
2432 __asm__ volatile(\ | |
2433 "shl $2, %0 \n"\ | |
2434 "add %0, %1 \n"\ | |
2435 "add %0, %2 \n"\ | |
2436 "add %0, %3 \n"\ | |
2437 "neg %0 \n"\ | |
2438 body\ | |
2439 :"+r"(reglen), "+r"(dst), "+r"(src0), "+r"(src1)\ | |
2440 );\ | |
2441 }else if(channels==6){\ | |
2442 ff_float_to_int16_interleave6_##cpu(dst, src, len);\ | |
2443 }else\ | |
2444 float_to_int16_interleave_misc_##cpu(dst, src, len, channels);\ | |
2445 } | |
2446 | |
2447 FLOAT_TO_INT16_INTERLEAVE(3dnow, | |
2448 "1: \n" | |
2449 "pf2id (%2,%0), %%mm0 \n" | |
2450 "pf2id 8(%2,%0), %%mm1 \n" | |
2451 "pf2id (%3,%0), %%mm2 \n" | |
2452 "pf2id 8(%3,%0), %%mm3 \n" | |
2453 "packssdw %%mm1, %%mm0 \n" | |
2454 "packssdw %%mm3, %%mm2 \n" | |
2455 "movq %%mm0, %%mm1 \n" | |
2456 "punpcklwd %%mm2, %%mm0 \n" | |
2457 "punpckhwd %%mm2, %%mm1 \n" | |
2458 "movq %%mm0, (%1,%0)\n" | |
2459 "movq %%mm1, 8(%1,%0)\n" | |
2460 "add $16, %0 \n" | |
2461 "js 1b \n" | |
2462 "femms \n" | |
2463 ) | |
2464 | |
2465 FLOAT_TO_INT16_INTERLEAVE(sse, | |
2466 "1: \n" | |
2467 "cvtps2pi (%2,%0), %%mm0 \n" | |
2468 "cvtps2pi 8(%2,%0), %%mm1 \n" | |
2469 "cvtps2pi (%3,%0), %%mm2 \n" | |
2470 "cvtps2pi 8(%3,%0), %%mm3 \n" | |
2471 "packssdw %%mm1, %%mm0 \n" | |
2472 "packssdw %%mm3, %%mm2 \n" | |
2473 "movq %%mm0, %%mm1 \n" | |
2474 "punpcklwd %%mm2, %%mm0 \n" | |
2475 "punpckhwd %%mm2, %%mm1 \n" | |
2476 "movq %%mm0, (%1,%0)\n" | |
2477 "movq %%mm1, 8(%1,%0)\n" | |
2478 "add $16, %0 \n" | |
2479 "js 1b \n" | |
2480 "emms \n" | |
2481 ) | |
2482 | |
2483 FLOAT_TO_INT16_INTERLEAVE(sse2, | |
2484 "1: \n" | |
2485 "cvtps2dq (%2,%0), %%xmm0 \n" | |
2486 "cvtps2dq (%3,%0), %%xmm1 \n" | |
2487 "packssdw %%xmm1, %%xmm0 \n" | |
2488 "movhlps %%xmm0, %%xmm1 \n" | |
2489 "punpcklwd %%xmm1, %%xmm0 \n" | |
2490 "movdqa %%xmm0, (%1,%0) \n" | |
2491 "add $16, %0 \n" | |
2492 "js 1b \n" | |
2493 ) | |
2494 | |
2495 static void float_to_int16_interleave_3dn2(int16_t *dst, const float **src, long len, int channels){ | |
2496 if(channels==6) | |
2497 ff_float_to_int16_interleave6_3dn2(dst, src, len); | |
2498 else | |
2499 float_to_int16_interleave_3dnow(dst, src, len, channels); | |
2500 } | |
2501 | |
10964 | 2502 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order); |
2503 | |
8430 | 2504 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
2505 { | |
2506 mm_flags = mm_support(); | |
2507 | |
2508 if (avctx->dsp_mask) { | |
2509 if (avctx->dsp_mask & FF_MM_FORCE) | |
2510 mm_flags |= (avctx->dsp_mask & 0xffff); | |
2511 else | |
2512 mm_flags &= ~(avctx->dsp_mask & 0xffff); | |
2513 } | |
2514 | |
2515 #if 0 | |
2516 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:"); | |
2517 if (mm_flags & FF_MM_MMX) | |
2518 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
|
2519 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
|
2520 av_log(avctx, AV_LOG_INFO, " mmx2"); |
8430 | 2521 if (mm_flags & FF_MM_3DNOW) |
2522 av_log(avctx, AV_LOG_INFO, " 3dnow"); | |
2523 if (mm_flags & FF_MM_SSE) | |
2524 av_log(avctx, AV_LOG_INFO, " sse"); | |
2525 if (mm_flags & FF_MM_SSE2) | |
2526 av_log(avctx, AV_LOG_INFO, " sse2"); | |
2527 av_log(avctx, AV_LOG_INFO, "\n"); | |
2528 #endif | |
2529 | |
2530 if (mm_flags & FF_MM_MMX) { | |
2531 const int idct_algo= avctx->idct_algo; | |
2532 | |
2533 if(avctx->lowres==0){ | |
2534 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
2535 c->idct_put= ff_simple_idct_put_mmx; | |
2536 c->idct_add= ff_simple_idct_add_mmx; | |
2537 c->idct = ff_simple_idct_mmx; | |
2538 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM; | |
8590 | 2539 #if CONFIG_GPL |
8430 | 2540 }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
|
2541 if(mm_flags & FF_MM_MMX2){ |
8430 | 2542 c->idct_put= ff_libmpeg2mmx2_idct_put; |
2543 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
2544 c->idct = ff_mmxext_idct; | |
2545 }else{ | |
2546 c->idct_put= ff_libmpeg2mmx_idct_put; | |
2547 c->idct_add= ff_libmpeg2mmx_idct_add; | |
2548 c->idct = ff_mmx_idct; | |
2549 } | |
2550 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM; | |
2551 #endif | |
9975
d6d7e8d4a04d
Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents:
9959
diff
changeset
|
2552 }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) && |
8430 | 2553 idct_algo==FF_IDCT_VP3){ |
2554 if(mm_flags & FF_MM_SSE2){ | |
2555 c->idct_put= ff_vp3_idct_put_sse2; | |
2556 c->idct_add= ff_vp3_idct_add_sse2; | |
2557 c->idct = ff_vp3_idct_sse2; | |
2558 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2559 }else{ | |
2560 c->idct_put= ff_vp3_idct_put_mmx; | |
2561 c->idct_add= ff_vp3_idct_add_mmx; | |
2562 c->idct = ff_vp3_idct_mmx; | |
2563 c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM; | |
2564 } | |
2565 }else if(idct_algo==FF_IDCT_CAVS){ | |
2566 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2567 }else if(idct_algo==FF_IDCT_XVIDMMX){ | |
2568 if(mm_flags & FF_MM_SSE2){ | |
2569 c->idct_put= ff_idct_xvid_sse2_put; | |
2570 c->idct_add= ff_idct_xvid_sse2_add; | |
2571 c->idct = ff_idct_xvid_sse2; | |
2572 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
|
2573 }else if(mm_flags & FF_MM_MMX2){ |
8430 | 2574 c->idct_put= ff_idct_xvid_mmx2_put; |
2575 c->idct_add= ff_idct_xvid_mmx2_add; | |
2576 c->idct = ff_idct_xvid_mmx2; | |
2577 }else{ | |
2578 c->idct_put= ff_idct_xvid_mmx_put; | |
2579 c->idct_add= ff_idct_xvid_mmx_add; | |
2580 c->idct = ff_idct_xvid_mmx; | |
2581 } | |
2582 } | |
2583 } | |
2584 | |
2585 c->put_pixels_clamped = put_pixels_clamped_mmx; | |
2586 c->put_signed_pixels_clamped = put_signed_pixels_clamped_mmx; | |
2587 c->add_pixels_clamped = add_pixels_clamped_mmx; | |
2588 c->clear_block = clear_block_mmx; | |
2589 c->clear_blocks = clear_blocks_mmx; | |
10766
78c2be62260a
Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents:
10749
diff
changeset
|
2590 if ((mm_flags & FF_MM_SSE) && |
78c2be62260a
Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents:
10749
diff
changeset
|
2591 !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){ |
78c2be62260a
Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents:
10749
diff
changeset
|
2592 /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */ |
9861 | 2593 c->clear_block = clear_block_sse; |
2594 c->clear_blocks = clear_blocks_sse; | |
2595 } | |
8430 | 2596 |
2597 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \ | |
2598 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \ | |
2599 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \ | |
2600 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \ | |
2601 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU | |
2602 | |
2603 SET_HPEL_FUNCS(put, 0, 16, mmx); | |
2604 SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx); | |
2605 SET_HPEL_FUNCS(avg, 0, 16, mmx); | |
2606 SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx); | |
2607 SET_HPEL_FUNCS(put, 1, 8, mmx); | |
2608 SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx); | |
2609 SET_HPEL_FUNCS(avg, 1, 8, mmx); | |
2610 SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx); | |
2611 | |
2612 c->gmc= gmc_mmx; | |
2613 | |
2614 c->add_bytes= add_bytes_mmx; | |
2615 c->add_bytes_l2= add_bytes_l2_mmx; | |
2616 | |
2617 c->draw_edges = draw_edges_mmx; | |
2618 | |
10749
5cca4b6c459d
Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents:
10645
diff
changeset
|
2619 if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { |
8430 | 2620 c->h263_v_loop_filter= h263_v_loop_filter_mmx; |
2621 c->h263_h_loop_filter= h263_h_loop_filter_mmx; | |
2622 } | |
2623 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_rnd; | |
2624 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
|
2625 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_vc1_chroma_mc8_mmx_nornd; |
8430 | 2626 |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2627 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
|
2628 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
|
2629 |
8817 | 2630 if (CONFIG_VP6_DECODER) { |
2631 c->vp6_filter_diag4 = ff_vp6_filter_diag4_mmx; | |
2632 } | |
2633 | |
9342
7f594601d5e9
Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents:
9341
diff
changeset
|
2634 if (mm_flags & FF_MM_MMX2) { |
8430 | 2635 c->prefetch = prefetch_mmx2; |
2636 | |
2637 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2; | |
2638 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; | |
2639 | |
2640 c->avg_pixels_tab[0][0] = avg_pixels16_mmx2; | |
2641 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2; | |
2642 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; | |
2643 | |
2644 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2; | |
2645 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; | |
2646 | |
2647 c->avg_pixels_tab[1][0] = avg_pixels8_mmx2; | |
2648 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2; | |
2649 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; | |
2650 | |
2651 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2652 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
2653 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
2654 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
2655 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
2656 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
2657 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; | |
2658 | |
9975
d6d7e8d4a04d
Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents:
9959
diff
changeset
|
2659 if (CONFIG_VP3_DECODER) { |
8430 | 2660 c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2; |
2661 c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2; | |
2662 } | |
2663 } | |
11637 | 2664 if (CONFIG_VP3_DECODER) { |
2665 c->vp3_idct_dc_add = ff_vp3_idct_dc_add_mmx2; | |
2666 } | |
8430 | 2667 |
11826
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2668 if (CONFIG_VP3_DECODER |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2669 && (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
|
2670 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
|
2671 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
|
2672 } |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2673 |
8430 | 2674 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \ |
2675 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \ | |
2676 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \ | |
2677 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \ | |
2678 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \ | |
2679 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \ | |
2680 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \ | |
2681 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \ | |
2682 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \ | |
2683 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \ | |
2684 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \ | |
2685 c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \ | |
2686 c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \ | |
2687 c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \ | |
2688 c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \ | |
2689 c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \ | |
2690 c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU | |
2691 | |
2692 SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2); | |
2693 SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2); | |
2694 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2); | |
2695 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2); | |
2696 SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2); | |
2697 SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2); | |
2698 | |
2699 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2); | |
2700 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2); | |
2701 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2); | |
2702 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2); | |
2703 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2); | |
2704 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2); | |
2705 | |
2706 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2); | |
2707 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2); | |
2708 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2); | |
2709 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2); | |
2710 | |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2711 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
|
2712 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
|
2713 |
9440 | 2714 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_vc1_chroma_mc8_mmx2_nornd; |
2715 | |
8430 | 2716 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2_rnd; |
2717 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2; | |
2718 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2; | |
2719 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2; | |
2720 | |
8760 | 2721 #if HAVE_YASM |
2722 c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2; | |
2723 #endif | |
8798
a5c8210814d7
Add check whether the compiler/assembler supports 10 or more operands.
diego
parents:
8760
diff
changeset
|
2724 #if HAVE_7REGS && HAVE_TEN_OPERANDS |
8760 | 2725 if( mm_flags&FF_MM_3DNOW ) |
2726 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov; | |
2727 #endif | |
2728 | |
8596
68e959302527
replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents:
8590
diff
changeset
|
2729 if (CONFIG_CAVS_DECODER) |
8430 | 2730 ff_cavsdsp_init_mmx2(c, avctx); |
2731 | |
9995
3141f69e3905
Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents:
9975
diff
changeset
|
2732 if (CONFIG_VC1_DECODER) |
8430 | 2733 ff_vc1dsp_init_mmx(c, avctx); |
2734 | |
2735 c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2; | |
2736 } else if (mm_flags & FF_MM_3DNOW) { | |
2737 c->prefetch = prefetch_3dnow; | |
2738 | |
2739 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow; | |
2740 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; | |
2741 | |
2742 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow; | |
2743 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow; | |
2744 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; | |
2745 | |
2746 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow; | |
2747 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow; | |
2748 | |
2749 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow; | |
2750 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow; | |
2751 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; | |
2752 | |
2753 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2754 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
2755 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
2756 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
2757 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
2758 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
2759 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
2760 } | |
2761 | |
11826
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2762 if (CONFIG_VP3_DECODER |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2763 && (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
|
2764 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
|
2765 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
|
2766 } |
11c5a87497d3
Add bitexact versions of put_no_rnd_pixels8 _x2 and _y2 for vp3/theora
conrad
parents:
11637
diff
changeset
|
2767 |
8430 | 2768 SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow); |
2769 SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow); | |
2770 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow); | |
2771 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow); | |
2772 SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow); | |
2773 SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow); | |
2774 | |
2775 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow); | |
2776 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow); | |
2777 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow); | |
2778 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow); | |
2779 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow); | |
2780 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow); | |
2781 | |
2782 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow); | |
2783 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow); | |
2784 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow); | |
2785 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow); | |
2786 | |
2787 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_3dnow_rnd; | |
2788 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow; | |
2789 | |
8519
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2790 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
|
2791 c->avg_rv40_chroma_pixels_tab[1]= avg_rv40_chroma_mc4_3dnow; |
cc64e1343397
Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents:
8510
diff
changeset
|
2792 |
8596
68e959302527
replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents:
8590
diff
changeset
|
2793 if (CONFIG_CAVS_DECODER) |
8430 | 2794 ff_cavsdsp_init_3dnow(c, avctx); |
2795 } | |
2796 | |
2797 | |
2798 #define H264_QPEL_FUNCS(x, y, CPU)\ | |
2799 c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\ | |
2800 c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\ | |
2801 c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\ | |
2802 c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU; | |
2803 if((mm_flags & FF_MM_SSE2) && !(mm_flags & FF_MM_3DNOW)){ | |
2804 // these functions are slower than mmx on AMD, but faster on Intel | |
2805 c->put_pixels_tab[0][0] = put_pixels16_sse2; | |
2806 c->avg_pixels_tab[0][0] = avg_pixels16_sse2; | |
2807 H264_QPEL_FUNCS(0, 0, sse2); | |
2808 } | |
2809 if(mm_flags & FF_MM_SSE2){ | |
2810 H264_QPEL_FUNCS(0, 1, sse2); | |
2811 H264_QPEL_FUNCS(0, 2, sse2); | |
2812 H264_QPEL_FUNCS(0, 3, sse2); | |
2813 H264_QPEL_FUNCS(1, 1, sse2); | |
2814 H264_QPEL_FUNCS(1, 2, sse2); | |
2815 H264_QPEL_FUNCS(1, 3, sse2); | |
2816 H264_QPEL_FUNCS(2, 1, sse2); | |
2817 H264_QPEL_FUNCS(2, 2, sse2); | |
2818 H264_QPEL_FUNCS(2, 3, sse2); | |
2819 H264_QPEL_FUNCS(3, 1, sse2); | |
2820 H264_QPEL_FUNCS(3, 2, sse2); | |
2821 H264_QPEL_FUNCS(3, 3, sse2); | |
8818 | 2822 |
2823 if (CONFIG_VP6_DECODER) { | |
2824 c->vp6_filter_diag4 = ff_vp6_filter_diag4_sse2; | |
2825 } | |
8430 | 2826 } |
8590 | 2827 #if HAVE_SSSE3 |
8430 | 2828 if(mm_flags & FF_MM_SSSE3){ |
2829 H264_QPEL_FUNCS(1, 0, ssse3); | |
2830 H264_QPEL_FUNCS(1, 1, ssse3); | |
2831 H264_QPEL_FUNCS(1, 2, ssse3); | |
2832 H264_QPEL_FUNCS(1, 3, ssse3); | |
2833 H264_QPEL_FUNCS(2, 0, ssse3); | |
2834 H264_QPEL_FUNCS(2, 1, ssse3); | |
2835 H264_QPEL_FUNCS(2, 2, ssse3); | |
2836 H264_QPEL_FUNCS(2, 3, ssse3); | |
2837 H264_QPEL_FUNCS(3, 0, ssse3); | |
2838 H264_QPEL_FUNCS(3, 1, ssse3); | |
2839 H264_QPEL_FUNCS(3, 2, ssse3); | |
2840 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
|
2841 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_vc1_chroma_mc8_ssse3_nornd; |
9440 | 2842 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_vc1_chroma_mc8_ssse3_nornd; |
8430 | 2843 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_ssse3_rnd; |
2844 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_ssse3_rnd; | |
2845 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_ssse3; | |
2846 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_ssse3; | |
2847 c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3; | |
10430 | 2848 #if HAVE_YASM |
2849 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3; | |
2850 if (mm_flags & FF_MM_SSE4) // not really sse4, just slow on Conroe | |
2851 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4; | |
2852 #endif | |
8430 | 2853 } |
2854 #endif | |
2855 | |
2856 if(mm_flags & FF_MM_3DNOW){ | |
2857 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow; | |
2858 c->vector_fmul = vector_fmul_3dnow; | |
2859 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2860 c->float_to_int16 = float_to_int16_3dnow; | |
2861 c->float_to_int16_interleave = float_to_int16_interleave_3dnow; | |
2862 } | |
2863 } | |
2864 if(mm_flags & FF_MM_3DNOWEXT){ | |
2865 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; | |
2866 c->vector_fmul_window = vector_fmul_window_3dnow2; | |
2867 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2868 c->float_to_int16_interleave = float_to_int16_interleave_3dn2; | |
2869 } | |
2870 } | |
10633 | 2871 if(mm_flags & FF_MM_MMX2){ |
2872 #if HAVE_YASM | |
2873 c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2; | |
10644 | 2874 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmx2; |
10633 | 2875 #endif |
2876 } | |
8430 | 2877 if(mm_flags & FF_MM_SSE){ |
2878 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse; | |
2879 c->ac3_downmix = ac3_downmix_sse; | |
2880 c->vector_fmul = vector_fmul_sse; | |
2881 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
|
2882 c->vector_fmul_add = vector_fmul_add_sse; |
8430 | 2883 c->vector_fmul_window = vector_fmul_window_sse; |
2884 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
|
2885 c->vector_clipf = vector_clipf_sse; |
8430 | 2886 c->float_to_int16 = float_to_int16_sse; |
2887 c->float_to_int16_interleave = float_to_int16_interleave_sse; | |
10964 | 2888 #if HAVE_YASM |
2889 c->scalarproduct_float = ff_scalarproduct_float_sse; | |
2890 #endif | |
8430 | 2891 } |
2892 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
|
2893 c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse |
8430 | 2894 if(mm_flags & FF_MM_SSE2){ |
2895 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2; | |
2896 c->float_to_int16 = float_to_int16_sse2; | |
2897 c->float_to_int16_interleave = float_to_int16_interleave_sse2; | |
10633 | 2898 #if HAVE_YASM |
2899 c->scalarproduct_int16 = ff_scalarproduct_int16_sse2; | |
10644 | 2900 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2; |
10633 | 2901 #endif |
8430 | 2902 } |
10644 | 2903 if((mm_flags & FF_MM_SSSE3) && !(mm_flags & (FF_MM_SSE42|FF_MM_3DNOW)) && HAVE_YASM) // cachesplit |
2904 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3; | |
8430 | 2905 } |
2906 | |
8596
68e959302527
replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents:
8590
diff
changeset
|
2907 if (CONFIG_ENCODERS) |
8430 | 2908 dsputilenc_init_mmx(c, avctx); |
2909 | |
2910 #if 0 | |
2911 // for speed testing | |
2912 get_pixels = just_return; | |
2913 put_pixels_clamped = just_return; | |
2914 add_pixels_clamped = just_return; | |
2915 | |
2916 pix_abs16x16 = just_return; | |
2917 pix_abs16x16_x2 = just_return; | |
2918 pix_abs16x16_y2 = just_return; | |
2919 pix_abs16x16_xy2 = just_return; | |
2920 | |
2921 put_pixels_tab[0] = just_return; | |
2922 put_pixels_tab[1] = just_return; | |
2923 put_pixels_tab[2] = just_return; | |
2924 put_pixels_tab[3] = just_return; | |
2925 | |
2926 put_no_rnd_pixels_tab[0] = just_return; | |
2927 put_no_rnd_pixels_tab[1] = just_return; | |
2928 put_no_rnd_pixels_tab[2] = just_return; | |
2929 put_no_rnd_pixels_tab[3] = just_return; | |
2930 | |
2931 avg_pixels_tab[0] = just_return; | |
2932 avg_pixels_tab[1] = just_return; | |
2933 avg_pixels_tab[2] = just_return; | |
2934 avg_pixels_tab[3] = just_return; | |
2935 | |
2936 avg_no_rnd_pixels_tab[0] = just_return; | |
2937 avg_no_rnd_pixels_tab[1] = just_return; | |
2938 avg_no_rnd_pixels_tab[2] = just_return; | |
2939 avg_no_rnd_pixels_tab[3] = just_return; | |
2940 | |
2941 //av_fdct = just_return; | |
2942 //ff_idct = just_return; | |
2943 #endif | |
2944 } | |
11499 | 2945 |
2946 #if CONFIG_H264DSP | |
2947 void ff_h264dsp_init_x86(H264DSPContext *c) | |
2948 { | |
2949 mm_flags = mm_support(); | |
2950 | |
2951 if (mm_flags & FF_MM_MMX) { | |
2952 c->h264_idct_dc_add= | |
2953 c->h264_idct_add= ff_h264_idct_add_mmx; | |
2954 c->h264_idct8_dc_add= | |
2955 c->h264_idct8_add= ff_h264_idct8_add_mmx; | |
2956 | |
2957 c->h264_idct_add16 = ff_h264_idct_add16_mmx; | |
2958 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx; | |
2959 c->h264_idct_add8 = ff_h264_idct_add8_mmx; | |
2960 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx; | |
2961 | |
2962 if (mm_flags & FF_MM_MMX2) { | |
2963 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2; | |
2964 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; | |
2965 c->h264_idct_add16 = ff_h264_idct_add16_mmx2; | |
2966 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx2; | |
2967 c->h264_idct_add8 = ff_h264_idct_add8_mmx2; | |
2968 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2; | |
2969 | |
2970 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2; | |
2971 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2; | |
2972 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2; | |
2973 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2; | |
2974 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_mmx2; | |
2975 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_mmx2; | |
2976 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; | |
2977 | |
2978 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2; | |
2979 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2; | |
2980 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2; | |
2981 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2; | |
2982 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2; | |
2983 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2; | |
2984 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2; | |
2985 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2; | |
2986 | |
2987 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2; | |
2988 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2; | |
2989 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2; | |
2990 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2; | |
2991 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2; | |
2992 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2; | |
2993 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2; | |
2994 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2; | |
2995 } | |
2996 if(mm_flags & FF_MM_SSE2){ | |
2997 c->h264_idct8_add = ff_h264_idct8_add_sse2; | |
2998 c->h264_idct8_add4= ff_h264_idct8_add4_sse2; | |
2999 } | |
3000 | |
3001 #if CONFIG_GPL && HAVE_YASM | |
3002 if (mm_flags & FF_MM_MMX2){ | |
3003 #if ARCH_X86_32 | |
3004 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext; | |
3005 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext; | |
3006 #endif | |
3007 if( mm_flags&FF_MM_SSE2 ){ | |
3008 #if ARCH_X86_64 || !defined(__ICC) || __ICC > 1110 | |
3009 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2; | |
3010 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2; | |
3011 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2; | |
3012 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2; | |
3013 #endif | |
3014 c->h264_idct_add16 = ff_h264_idct_add16_sse2; | |
3015 c->h264_idct_add8 = ff_h264_idct_add8_sse2; | |
3016 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2; | |
3017 } | |
3018 } | |
3019 #endif | |
3020 } | |
3021 } | |
3022 #endif /* CONFIG_H264DSP */ |