Mercurial > libavcodec.hg
annotate x86/h264dsp_mmx.c @ 12528:f4caff0e1e10 libavcodec
In dv decoder, set sample aspect aspect ratio, fix issue #1612
author | bcoudurier |
---|---|
date | Mon, 27 Sep 2010 21:31:29 +0000 |
parents | 58bd8cb86f5d |
children |
rev | line source |
---|---|
8430 | 1 /* |
2 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt | |
3 * | |
4 * This file is part of FFmpeg. | |
5 * | |
6 * FFmpeg is free software; you can redistribute it and/or | |
7 * modify it under the terms of the GNU Lesser General Public | |
8 * License as published by the Free Software Foundation; either | |
9 * version 2.1 of the License, or (at your option) any later version. | |
10 * | |
11 * FFmpeg is distributed in the hope that it will be useful, | |
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 * Lesser General Public License for more details. | |
15 * | |
16 * You should have received a copy of the GNU Lesser General Public | |
17 * License along with FFmpeg; if not, write to the Free Software | |
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
19 */ | |
20 | |
12475
9fef0a8ddd63
Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents:
12456
diff
changeset
|
21 #include "libavutil/cpu.h" |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
22 #include "libavutil/x86_cpu.h" |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
23 #include "libavcodec/h264dsp.h" |
8430 | 24 #include "dsputil_mmx.h" |
25 | |
11369 | 26 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_3_1 ) = 0x0103010301030103ULL; |
8430 | 27 |
28 /***********************************/ | |
29 /* IDCT */ | |
30 | |
12492
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
31 void ff_h264_idct_add_mmx (uint8_t *dst, int16_t *block, int stride); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
32 void ff_h264_idct8_add_mmx (uint8_t *dst, int16_t *block, int stride); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
33 void ff_h264_idct8_add_sse2 (uint8_t *dst, int16_t *block, int stride); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
34 void ff_h264_idct_dc_add_mmx2 (uint8_t *dst, int16_t *block, int stride); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
35 void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride); |
8430 | 36 |
12492
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
37 void ff_h264_idct_add16_mmx (uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
38 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
39 void ff_h264_idct8_add4_mmx (uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
40 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
41 void ff_h264_idct_add16_mmx2 (uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
42 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
43 void ff_h264_idct_add16intra_mmx (uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
44 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
45 void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
46 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
47 void ff_h264_idct8_add4_mmx2 (uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
48 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
49 void ff_h264_idct8_add4_sse2 (uint8_t *dst, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
50 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
51 void ff_h264_idct_add8_mmx (uint8_t **dest, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
52 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
53 void ff_h264_idct_add8_mmx2 (uint8_t **dest, const int *block_offset, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
54 DCTELEM *block, int stride, const uint8_t nnzc[6*8]); |
8430 | 55 |
12492
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
56 void ff_h264_idct_add16_sse2 (uint8_t *dst, const int *block_offset, DCTELEM *block, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
57 int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
58 void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
59 int stride, const uint8_t nnzc[6*8]); |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
60 void ff_h264_idct_add8_sse2 (uint8_t **dest, const int *block_offset, DCTELEM *block, |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
61 int stride, const uint8_t nnzc[6*8]); |
8510
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
62 |
8430 | 63 /***********************************/ |
64 /* deblocking */ | |
65 | |
66 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2], | |
67 int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) { | |
68 int dir; | |
69 __asm__ volatile( | |
11017 | 70 "movq %0, %%mm7 \n" |
71 "movq %1, %%mm6 \n" | |
72 ::"m"(ff_pb_1), "m"(ff_pb_3) | |
8430 | 73 ); |
74 if(field) | |
75 __asm__ volatile( | |
11017 | 76 "movq %0, %%mm6 \n" |
77 ::"m"(ff_pb_3_1) | |
8430 | 78 ); |
11017 | 79 __asm__ volatile( |
80 "movq %%mm6, %%mm5 \n" | |
81 "paddb %%mm5, %%mm5 \n" | |
82 :); | |
8430 | 83 |
84 // could do a special case for dir==0 && edges==1, but it only reduces the | |
85 // average filter time by 1.2% | |
86 for( dir=1; dir>=0; dir-- ) { | |
10953 | 87 const x86_reg d_idx = dir ? -8 : -1; |
8430 | 88 const int mask_mv = dir ? mask_mv1 : mask_mv0; |
11369 | 89 DECLARE_ALIGNED(8, const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL; |
11017 | 90 int b_idx, edge; |
8430 | 91 for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) { |
92 __asm__ volatile( | |
93 "pand %0, %%mm0 \n\t" | |
94 ::"m"(mask_dir) | |
95 ); | |
96 if(!(mask_mv & edge)) { | |
11017 | 97 if(bidir) { |
8430 | 98 __asm__ volatile( |
11017 | 99 "movd (%1,%0), %%mm2 \n" |
100 "punpckldq 40(%1,%0), %%mm2 \n" // { ref0[bn], ref1[bn] } | |
101 "pshufw $0x44, (%1), %%mm0 \n" // { ref0[b], ref0[b] } | |
102 "pshufw $0x44, 40(%1), %%mm1 \n" // { ref1[b], ref1[b] } | |
103 "pshufw $0x4E, %%mm2, %%mm3 \n" | |
104 "psubb %%mm2, %%mm0 \n" // { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] } | |
105 "psubb %%mm3, %%mm1 \n" // { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] } | |
106 "1: \n" | |
107 "por %%mm1, %%mm0 \n" | |
108 "movq (%2,%0,4), %%mm1 \n" | |
109 "movq 8(%2,%0,4), %%mm2 \n" | |
110 "movq %%mm1, %%mm3 \n" | |
111 "movq %%mm2, %%mm4 \n" | |
112 "psubw (%2), %%mm1 \n" | |
113 "psubw 8(%2), %%mm2 \n" | |
114 "psubw 160(%2), %%mm3 \n" | |
115 "psubw 168(%2), %%mm4 \n" | |
116 "packsswb %%mm2, %%mm1 \n" | |
117 "packsswb %%mm4, %%mm3 \n" | |
118 "paddb %%mm6, %%mm1 \n" | |
119 "paddb %%mm6, %%mm3 \n" | |
120 "psubusb %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit | |
121 "psubusb %%mm5, %%mm3 \n" | |
122 "packsswb %%mm3, %%mm1 \n" | |
123 "add $40, %0 \n" | |
124 "cmp $40, %0 \n" | |
125 "jl 1b \n" | |
126 "sub $80, %0 \n" | |
127 "pshufw $0x4E, %%mm1, %%mm1 \n" | |
128 "por %%mm1, %%mm0 \n" | |
129 "pshufw $0x4E, %%mm0, %%mm1 \n" | |
130 "pminub %%mm1, %%mm0 \n" | |
131 ::"r"(d_idx), | |
132 "r"(ref[0]+b_idx), | |
133 "r"(mv[0]+b_idx) | |
8430 | 134 ); |
11017 | 135 } else { |
10938
2a50b786b888
Fix h264_loop_filter_strength_mmx2() so it works with b frames.
michael
parents:
10937
diff
changeset
|
136 __asm__ volatile( |
11017 | 137 "movd (%1), %%mm0 \n" |
138 "psubb (%1,%0), %%mm0 \n" // ref[b] != ref[bn] | |
139 "movq (%2), %%mm1 \n" | |
140 "movq 8(%2), %%mm2 \n" | |
141 "psubw (%2,%0,4), %%mm1 \n" | |
142 "psubw 8(%2,%0,4), %%mm2 \n" | |
143 "packsswb %%mm2, %%mm1 \n" | |
144 "paddb %%mm6, %%mm1 \n" | |
145 "psubusb %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit | |
146 "packsswb %%mm1, %%mm1 \n" | |
147 "por %%mm1, %%mm0 \n" | |
148 ::"r"(d_idx), | |
149 "r"(ref[0]+b_idx), | |
150 "r"(mv[0]+b_idx) | |
10938
2a50b786b888
Fix h264_loop_filter_strength_mmx2() so it works with b frames.
michael
parents:
10937
diff
changeset
|
151 ); |
2a50b786b888
Fix h264_loop_filter_strength_mmx2() so it works with b frames.
michael
parents:
10937
diff
changeset
|
152 } |
8430 | 153 } |
154 __asm__ volatile( | |
11017 | 155 "movd %0, %%mm1 \n" |
156 "por %1, %%mm1 \n" // nnz[b] || nnz[bn] | |
8430 | 157 ::"m"(nnz[b_idx]), |
158 "m"(nnz[b_idx+d_idx]) | |
159 ); | |
160 __asm__ volatile( | |
11017 | 161 "pminub %%mm7, %%mm1 \n" |
162 "pminub %%mm7, %%mm0 \n" | |
163 "psllw $1, %%mm1 \n" | |
164 "pxor %%mm2, %%mm2 \n" | |
165 "pmaxub %%mm0, %%mm1 \n" | |
166 "punpcklbw %%mm2, %%mm1 \n" | |
167 "movq %%mm1, %0 \n" | |
8430 | 168 :"=m"(*bS[dir][edge]) |
169 ::"memory" | |
170 ); | |
171 } | |
172 edges = 4; | |
173 step = 1; | |
174 } | |
175 __asm__ volatile( | |
176 "movq (%0), %%mm0 \n\t" | |
177 "movq 8(%0), %%mm1 \n\t" | |
178 "movq 16(%0), %%mm2 \n\t" | |
179 "movq 24(%0), %%mm3 \n\t" | |
180 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4) | |
181 "movq %%mm0, (%0) \n\t" | |
182 "movq %%mm3, 8(%0) \n\t" | |
183 "movq %%mm4, 16(%0) \n\t" | |
184 "movq %%mm2, 24(%0) \n\t" | |
185 ::"r"(bS[0]) | |
186 :"memory" | |
187 ); | |
188 } | |
189 | |
12454
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
190 #define LF_FUNC(DIR, TYPE, OPT) \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
191 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
192 int alpha, int beta, int8_t *tc0); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
193 #define LF_IFUNC(DIR, TYPE, OPT) \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
194 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
195 int alpha, int beta); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
196 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
197 LF_FUNC (h, chroma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
198 LF_IFUNC(h, chroma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
199 LF_FUNC (v, chroma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
200 LF_IFUNC(v, chroma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
201 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
202 LF_FUNC (h, luma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
203 LF_IFUNC(h, luma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
204 #if HAVE_YASM && ARCH_X86_32 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
205 LF_FUNC (v8, luma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
206 static void ff_x264_deblock_v_luma_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
207 { |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
208 if((tc0[0] & tc0[1]) >= 0) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
209 ff_x264_deblock_v8_luma_mmxext(pix+0, stride, alpha, beta, tc0); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
210 if((tc0[2] & tc0[3]) >= 0) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
211 ff_x264_deblock_v8_luma_mmxext(pix+8, stride, alpha, beta, tc0+2); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
212 } |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
213 LF_IFUNC(v8, luma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
214 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
215 { |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
216 ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
217 ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
218 } |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
219 #endif |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
220 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
221 LF_FUNC (h, luma, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
222 LF_IFUNC(h, luma_intra, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
223 LF_FUNC (v, luma, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
224 LF_IFUNC(v, luma_intra, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
225 |
8430 | 226 /***********************************/ |
227 /* weighted prediction */ | |
228 | |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
229 #define H264_WEIGHT(W, H, OPT) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
230 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
231 int stride, int log2_denom, int weight, int offset); |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
232 |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
233 #define H264_BIWEIGHT(W, H, OPT) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
234 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
235 uint8_t *src, int stride, int log2_denom, int weightd, \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
236 int weights, int offset); |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
237 |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
238 #define H264_BIWEIGHT_MMX(W,H) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
239 H264_WEIGHT (W, H, mmx2) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
240 H264_BIWEIGHT(W, H, mmx2) |
8430 | 241 |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
242 #define H264_BIWEIGHT_MMX_SSE(W,H) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
243 H264_BIWEIGHT_MMX(W, H) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
244 H264_WEIGHT (W, H, sse2) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
245 H264_BIWEIGHT (W, H, sse2) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
246 H264_BIWEIGHT (W, H, ssse3) |
8430 | 247 |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
248 H264_BIWEIGHT_MMX_SSE(16, 16) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
249 H264_BIWEIGHT_MMX_SSE(16, 8) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
250 H264_BIWEIGHT_MMX_SSE( 8, 16) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
251 H264_BIWEIGHT_MMX_SSE( 8, 8) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
252 H264_BIWEIGHT_MMX_SSE( 8, 4) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
253 H264_BIWEIGHT_MMX ( 4, 8) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
254 H264_BIWEIGHT_MMX ( 4, 4) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
255 H264_BIWEIGHT_MMX ( 4, 2) |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
256 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
257 void ff_h264dsp_init_x86(H264DSPContext *c) |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
258 { |
12475
9fef0a8ddd63
Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents:
12456
diff
changeset
|
259 int mm_flags = av_get_cpu_flags(); |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
260 |
12492
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
261 if (mm_flags & AV_CPU_FLAG_MMX2) { |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
262 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
263 } |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
264 #if HAVE_YASM |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
265 if (mm_flags & AV_CPU_FLAG_MMX) { |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
266 c->h264_idct_dc_add= |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
267 c->h264_idct_add= ff_h264_idct_add_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
268 c->h264_idct8_dc_add= |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
269 c->h264_idct8_add= ff_h264_idct8_add_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
270 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
271 c->h264_idct_add16 = ff_h264_idct_add16_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
272 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
273 c->h264_idct_add8 = ff_h264_idct_add8_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
274 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
275 |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
276 if (mm_flags & AV_CPU_FLAG_MMX2) { |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
277 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
278 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
279 c->h264_idct_add16 = ff_h264_idct_add16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
280 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
281 c->h264_idct_add8 = ff_h264_idct_add8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
282 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
283 |
12454
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
284 c->h264_v_loop_filter_chroma= ff_x264_deblock_v_chroma_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
285 c->h264_h_loop_filter_chroma= ff_x264_deblock_h_chroma_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
286 c->h264_v_loop_filter_chroma_intra= ff_x264_deblock_v_chroma_intra_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
287 c->h264_h_loop_filter_chroma_intra= ff_x264_deblock_h_chroma_intra_mmxext; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
288 #if ARCH_X86_32 |
12454
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
289 c->h264_v_loop_filter_luma= ff_x264_deblock_v_luma_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
290 c->h264_h_loop_filter_luma= ff_x264_deblock_h_luma_mmxext; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
291 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
292 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
293 #endif |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
294 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
295 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
296 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
297 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
298 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
299 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
300 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
301 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
302 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
303 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
304 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
305 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
306 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
307 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
308 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
309 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
310 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
311 |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
312 if (mm_flags&AV_CPU_FLAG_SSE2) { |
12492
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
313 c->h264_idct8_add = ff_h264_idct8_add_sse2; |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
314 c->h264_idct8_add4= ff_h264_idct8_add4_sse2; |
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
315 |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
316 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
317 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
318 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
319 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
320 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
321 |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
322 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
323 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
324 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2; |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
325 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
326 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
327 |
12505
b2d560458a72
x86: disable SSE functions using stack when stack is not aligned
mru
parents:
12502
diff
changeset
|
328 #if HAVE_ALIGNED_STACK |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
329 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
330 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
331 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
332 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2; |
12505
b2d560458a72
x86: disable SSE functions using stack when stack is not aligned
mru
parents:
12502
diff
changeset
|
333 #endif |
12502
6bc14239edfb
x86: remove hack disabling sse2 h264 loop filter with 32-bit icc
mru
parents:
12492
diff
changeset
|
334 |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
335 c->h264_idct_add16 = ff_h264_idct_add16_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
336 c->h264_idct_add8 = ff_h264_idct_add8_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
337 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
338 } |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
339 if (mm_flags&AV_CPU_FLAG_SSSE3) { |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
340 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
341 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
342 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3; |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
343 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
344 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3; |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
345 } |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
346 } |
12492
58a960d6e34c
Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from
rbultje
parents:
12480
diff
changeset
|
347 } |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
348 #endif |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
349 } |