Add x86 assembly for some 10-bit H.264 intra predict functions.
[ffmpeg.git] / libavcodec / x86 / h264_intrapred_init.c
1 /*
2  * Copyright (c) 2010 Jason Garrett-Glaser
3  *
4  * This file is part of Libav.
5  *
6  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/cpu.h"
22 #include "libavcodec/h264pred.h"
23
24 #define PRED4x4(TYPE, DEPTH, OPT) \
25 void ff_pred4x4_ ## TYPE ## _ ## DEPTH ## _ ## OPT (uint8_t *src, const uint8_t *topright, int stride);
26
27 PRED4x4(dc, 10, mmxext)
28 PRED4x4(down_left, 10, sse2)
29 PRED4x4(down_left, 10, avx)
30 PRED4x4(down_right, 10, sse2)
31 PRED4x4(down_right, 10, ssse3)
32 PRED4x4(down_right, 10, avx)
33 PRED4x4(vertical_left, 10, sse2)
34 PRED4x4(vertical_left, 10, avx)
35 PRED4x4(vertical_right, 10, sse2)
36 PRED4x4(vertical_right, 10, ssse3)
37 PRED4x4(vertical_right, 10, avx)
38 PRED4x4(horizontal_up, 10, mmxext)
39 PRED4x4(horizontal_down, 10, sse2)
40 PRED4x4(horizontal_down, 10, ssse3)
41 PRED4x4(horizontal_down, 10, avx)
42
43 #define PRED8x8(TYPE, DEPTH, OPT) \
44 void ff_pred8x8_ ## TYPE ## _ ## DEPTH ## _ ## OPT (uint8_t *src, int stride);
45
46 PRED8x8(vertical, 10, sse2)
47 PRED8x8(horizontal, 10, sse2)
48
49 void ff_pred16x16_vertical_mmx     (uint8_t *src, int stride);
50 void ff_pred16x16_vertical_sse     (uint8_t *src, int stride);
51 void ff_pred16x16_horizontal_mmx   (uint8_t *src, int stride);
52 void ff_pred16x16_horizontal_mmxext(uint8_t *src, int stride);
53 void ff_pred16x16_horizontal_ssse3 (uint8_t *src, int stride);
54 void ff_pred16x16_dc_mmxext        (uint8_t *src, int stride);
55 void ff_pred16x16_dc_sse2          (uint8_t *src, int stride);
56 void ff_pred16x16_dc_ssse3         (uint8_t *src, int stride);
57 void ff_pred16x16_plane_h264_mmx   (uint8_t *src, int stride);
58 void ff_pred16x16_plane_h264_mmx2  (uint8_t *src, int stride);
59 void ff_pred16x16_plane_h264_sse2  (uint8_t *src, int stride);
60 void ff_pred16x16_plane_h264_ssse3 (uint8_t *src, int stride);
61 void ff_pred16x16_plane_rv40_mmx   (uint8_t *src, int stride);
62 void ff_pred16x16_plane_rv40_mmx2  (uint8_t *src, int stride);
63 void ff_pred16x16_plane_rv40_sse2  (uint8_t *src, int stride);
64 void ff_pred16x16_plane_rv40_ssse3 (uint8_t *src, int stride);
65 void ff_pred16x16_plane_svq3_mmx   (uint8_t *src, int stride);
66 void ff_pred16x16_plane_svq3_mmx2  (uint8_t *src, int stride);
67 void ff_pred16x16_plane_svq3_sse2  (uint8_t *src, int stride);
68 void ff_pred16x16_plane_svq3_ssse3 (uint8_t *src, int stride);
69 void ff_pred16x16_tm_vp8_mmx       (uint8_t *src, int stride);
70 void ff_pred16x16_tm_vp8_mmxext    (uint8_t *src, int stride);
71 void ff_pred16x16_tm_vp8_sse2      (uint8_t *src, int stride);
72 void ff_pred8x8_top_dc_mmxext      (uint8_t *src, int stride);
73 void ff_pred8x8_dc_rv40_mmxext     (uint8_t *src, int stride);
74 void ff_pred8x8_dc_mmxext          (uint8_t *src, int stride);
75 void ff_pred8x8_vertical_mmx       (uint8_t *src, int stride);
76 void ff_pred8x8_horizontal_mmx     (uint8_t *src, int stride);
77 void ff_pred8x8_horizontal_mmxext  (uint8_t *src, int stride);
78 void ff_pred8x8_horizontal_ssse3   (uint8_t *src, int stride);
79 void ff_pred8x8_plane_mmx          (uint8_t *src, int stride);
80 void ff_pred8x8_plane_mmx2         (uint8_t *src, int stride);
81 void ff_pred8x8_plane_sse2         (uint8_t *src, int stride);
82 void ff_pred8x8_plane_ssse3        (uint8_t *src, int stride);
83 void ff_pred8x8_tm_vp8_mmx         (uint8_t *src, int stride);
84 void ff_pred8x8_tm_vp8_mmxext      (uint8_t *src, int stride);
85 void ff_pred8x8_tm_vp8_sse2        (uint8_t *src, int stride);
86 void ff_pred8x8_tm_vp8_ssse3       (uint8_t *src, int stride);
87 void ff_pred8x8l_top_dc_mmxext     (uint8_t *src, int has_topleft, int has_topright, int stride);
88 void ff_pred8x8l_top_dc_ssse3      (uint8_t *src, int has_topleft, int has_topright, int stride);
89 void ff_pred8x8l_dc_mmxext         (uint8_t *src, int has_topleft, int has_topright, int stride);
90 void ff_pred8x8l_dc_ssse3          (uint8_t *src, int has_topleft, int has_topright, int stride);
91 void ff_pred8x8l_horizontal_mmxext (uint8_t *src, int has_topleft, int has_topright, int stride);
92 void ff_pred8x8l_horizontal_ssse3  (uint8_t *src, int has_topleft, int has_topright, int stride);
93 void ff_pred8x8l_vertical_mmxext   (uint8_t *src, int has_topleft, int has_topright, int stride);
94 void ff_pred8x8l_vertical_ssse3    (uint8_t *src, int has_topleft, int has_topright, int stride);
95 void ff_pred8x8l_down_left_mmxext  (uint8_t *src, int has_topleft, int has_topright, int stride);
96 void ff_pred8x8l_down_left_sse2    (uint8_t *src, int has_topleft, int has_topright, int stride);
97 void ff_pred8x8l_down_left_ssse3   (uint8_t *src, int has_topleft, int has_topright, int stride);
98 void ff_pred8x8l_down_right_mmxext (uint8_t *src, int has_topleft, int has_topright, int stride);
99 void ff_pred8x8l_down_right_sse2   (uint8_t *src, int has_topleft, int has_topright, int stride);
100 void ff_pred8x8l_down_right_ssse3  (uint8_t *src, int has_topleft, int has_topright, int stride);
101 void ff_pred8x8l_vertical_right_mmxext(uint8_t *src, int has_topleft, int has_topright, int stride);
102 void ff_pred8x8l_vertical_right_sse2(uint8_t *src, int has_topleft, int has_topright, int stride);
103 void ff_pred8x8l_vertical_right_ssse3(uint8_t *src, int has_topleft, int has_topright, int stride);
104 void ff_pred8x8l_vertical_left_sse2(uint8_t *src, int has_topleft, int has_topright, int stride);
105 void ff_pred8x8l_vertical_left_ssse3(uint8_t *src, int has_topleft, int has_topright, int stride);
106 void ff_pred8x8l_horizontal_up_mmxext(uint8_t *src, int has_topleft, int has_topright, int stride);
107 void ff_pred8x8l_horizontal_up_ssse3(uint8_t *src, int has_topleft, int has_topright, int stride);
108 void ff_pred8x8l_horizontal_down_mmxext(uint8_t *src, int has_topleft, int has_topright, int stride);
109 void ff_pred8x8l_horizontal_down_sse2(uint8_t *src, int has_topleft, int has_topright, int stride);
110 void ff_pred8x8l_horizontal_down_ssse3(uint8_t *src, int has_topleft, int has_topright, int stride);
111 void ff_pred4x4_dc_mmxext          (uint8_t *src, const uint8_t *topright, int stride);
112 void ff_pred4x4_down_left_mmxext   (uint8_t *src, const uint8_t *topright, int stride);
113 void ff_pred4x4_down_right_mmxext  (uint8_t *src, const uint8_t *topright, int stride);
114 void ff_pred4x4_vertical_left_mmxext(uint8_t *src, const uint8_t *topright, int stride);
115 void ff_pred4x4_vertical_right_mmxext(uint8_t *src, const uint8_t *topright, int stride);
116 void ff_pred4x4_horizontal_up_mmxext(uint8_t *src, const uint8_t *topright, int stride);
117 void ff_pred4x4_horizontal_down_mmxext(uint8_t *src, const uint8_t *topright, int stride);
118 void ff_pred4x4_tm_vp8_mmx         (uint8_t *src, const uint8_t *topright, int stride);
119 void ff_pred4x4_tm_vp8_mmxext      (uint8_t *src, const uint8_t *topright, int stride);
120 void ff_pred4x4_tm_vp8_ssse3       (uint8_t *src, const uint8_t *topright, int stride);
121 void ff_pred4x4_vertical_vp8_mmxext(uint8_t *src, const uint8_t *topright, int stride);
122
123 void ff_h264_pred_init_x86(H264PredContext *h, int codec_id, const int bit_depth)
124 {
125     int mm_flags = av_get_cpu_flags();
126
127     if (bit_depth == 8) {
128 #if HAVE_YASM
129     if (mm_flags & AV_CPU_FLAG_MMX) {
130         h->pred16x16[VERT_PRED8x8] = ff_pred16x16_vertical_mmx;
131         h->pred16x16[HOR_PRED8x8 ] = ff_pred16x16_horizontal_mmx;
132         h->pred8x8  [VERT_PRED8x8] = ff_pred8x8_vertical_mmx;
133         h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_mmx;
134         if (codec_id == CODEC_ID_VP8) {
135             h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_tm_vp8_mmx;
136             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_mmx;
137             h->pred4x4  [TM_VP8_PRED  ] = ff_pred4x4_tm_vp8_mmx;
138         } else {
139             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_plane_mmx;
140             if (codec_id == CODEC_ID_SVQ3) {
141                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_svq3_mmx;
142             } else if (codec_id == CODEC_ID_RV40) {
143                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_rv40_mmx;
144             } else {
145                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_h264_mmx;
146             }
147         }
148     }
149
150     if (mm_flags & AV_CPU_FLAG_MMX2) {
151         h->pred16x16[HOR_PRED8x8 ] = ff_pred16x16_horizontal_mmxext;
152         h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_mmxext;
153         h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_mmxext;
154         h->pred8x8l [TOP_DC_PRED ] = ff_pred8x8l_top_dc_mmxext;
155         h->pred8x8l [DC_PRED     ] = ff_pred8x8l_dc_mmxext;
156         h->pred8x8l [HOR_PRED    ] = ff_pred8x8l_horizontal_mmxext;
157         h->pred8x8l [VERT_PRED   ] = ff_pred8x8l_vertical_mmxext;
158         h->pred8x8l [DIAG_DOWN_RIGHT_PRED] = ff_pred8x8l_down_right_mmxext;
159         h->pred8x8l [VERT_RIGHT_PRED     ] = ff_pred8x8l_vertical_right_mmxext;
160         h->pred8x8l [HOR_UP_PRED         ] = ff_pred8x8l_horizontal_up_mmxext;
161         h->pred8x8l [DIAG_DOWN_LEFT_PRED ] = ff_pred8x8l_down_left_mmxext;
162         h->pred8x8l [HOR_DOWN_PRED       ] = ff_pred8x8l_horizontal_down_mmxext;
163         h->pred4x4  [DIAG_DOWN_RIGHT_PRED] = ff_pred4x4_down_right_mmxext;
164         h->pred4x4  [VERT_RIGHT_PRED     ] = ff_pred4x4_vertical_right_mmxext;
165         h->pred4x4  [HOR_DOWN_PRED       ] = ff_pred4x4_horizontal_down_mmxext;
166         h->pred4x4  [DC_PRED     ] = ff_pred4x4_dc_mmxext;
167         if (codec_id == CODEC_ID_VP8 || codec_id == CODEC_ID_H264)
168             h->pred4x4  [DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_mmxext;
169         if (codec_id == CODEC_ID_SVQ3 || codec_id == CODEC_ID_H264)
170             h->pred4x4  [VERT_LEFT_PRED      ] = ff_pred4x4_vertical_left_mmxext;
171         if (codec_id != CODEC_ID_RV40) {
172             h->pred4x4  [HOR_UP_PRED         ] = ff_pred4x4_horizontal_up_mmxext;
173         }
174         if (codec_id == CODEC_ID_SVQ3 || codec_id == CODEC_ID_H264) {
175             h->pred8x8  [TOP_DC_PRED8x8      ] = ff_pred8x8_top_dc_mmxext;
176             h->pred8x8  [DC_PRED8x8          ] = ff_pred8x8_dc_mmxext;
177         }
178         if (codec_id == CODEC_ID_VP8) {
179             h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_tm_vp8_mmxext;
180             h->pred8x8  [DC_PRED8x8   ] = ff_pred8x8_dc_rv40_mmxext;
181             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_mmxext;
182             h->pred4x4  [TM_VP8_PRED  ] = ff_pred4x4_tm_vp8_mmxext;
183             h->pred4x4  [VERT_PRED    ] = ff_pred4x4_vertical_vp8_mmxext;
184         } else {
185             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_plane_mmx2;
186             if (codec_id == CODEC_ID_SVQ3) {
187                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_svq3_mmx2;
188             } else if (codec_id == CODEC_ID_RV40) {
189                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_rv40_mmx2;
190             } else {
191                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_h264_mmx2;
192             }
193         }
194     }
195
196     if (mm_flags & AV_CPU_FLAG_SSE) {
197         h->pred16x16[VERT_PRED8x8] = ff_pred16x16_vertical_sse;
198     }
199
200     if (mm_flags & AV_CPU_FLAG_SSE2) {
201         h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_sse2;
202         h->pred8x8l [DIAG_DOWN_LEFT_PRED ] = ff_pred8x8l_down_left_sse2;
203         h->pred8x8l [DIAG_DOWN_RIGHT_PRED] = ff_pred8x8l_down_right_sse2;
204         h->pred8x8l [VERT_RIGHT_PRED     ] = ff_pred8x8l_vertical_right_sse2;
205         h->pred8x8l [VERT_LEFT_PRED      ] = ff_pred8x8l_vertical_left_sse2;
206         h->pred8x8l [HOR_DOWN_PRED       ] = ff_pred8x8l_horizontal_down_sse2;
207         if (codec_id == CODEC_ID_VP8) {
208             h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_tm_vp8_sse2;
209             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_sse2;
210         } else {
211             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_plane_sse2;
212             if (codec_id == CODEC_ID_SVQ3) {
213                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_svq3_sse2;
214             } else if (codec_id == CODEC_ID_RV40) {
215                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_rv40_sse2;
216             } else {
217                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_h264_sse2;
218             }
219         }
220     }
221
222     if (mm_flags & AV_CPU_FLAG_SSSE3) {
223         h->pred16x16[HOR_PRED8x8 ] = ff_pred16x16_horizontal_ssse3;
224         h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_ssse3;
225         h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_ssse3;
226         h->pred8x8l [TOP_DC_PRED ] = ff_pred8x8l_top_dc_ssse3;
227         h->pred8x8l [DC_PRED     ] = ff_pred8x8l_dc_ssse3;
228         h->pred8x8l [HOR_PRED    ] = ff_pred8x8l_horizontal_ssse3;
229         h->pred8x8l [VERT_PRED   ] = ff_pred8x8l_vertical_ssse3;
230         h->pred8x8l [DIAG_DOWN_LEFT_PRED ] = ff_pred8x8l_down_left_ssse3;
231         h->pred8x8l [DIAG_DOWN_RIGHT_PRED] = ff_pred8x8l_down_right_ssse3;
232         h->pred8x8l [VERT_RIGHT_PRED     ] = ff_pred8x8l_vertical_right_ssse3;
233         h->pred8x8l [VERT_LEFT_PRED      ] = ff_pred8x8l_vertical_left_ssse3;
234         h->pred8x8l [HOR_UP_PRED         ] = ff_pred8x8l_horizontal_up_ssse3;
235         h->pred8x8l [HOR_DOWN_PRED       ] = ff_pred8x8l_horizontal_down_ssse3;
236         if (codec_id == CODEC_ID_VP8) {
237             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_ssse3;
238             h->pred4x4  [TM_VP8_PRED  ] = ff_pred4x4_tm_vp8_ssse3;
239         } else {
240             h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_plane_ssse3;
241             if (codec_id == CODEC_ID_SVQ3) {
242                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_svq3_ssse3;
243             } else if (codec_id == CODEC_ID_RV40) {
244                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_rv40_ssse3;
245             } else {
246                 h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_plane_h264_ssse3;
247             }
248         }
249     }
250 #endif
251     } else if (bit_depth == 10) {
252 #if HAVE_YASM
253         if (mm_flags & AV_CPU_FLAG_MMX2) {
254             h->pred4x4[DC_PRED             ] = ff_pred4x4_dc_10_mmxext;
255             h->pred4x4[HOR_UP_PRED         ] = ff_pred4x4_horizontal_up_10_mmxext;
256         }
257         if (mm_flags & AV_CPU_FLAG_SSE2) {
258             h->pred4x4[DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_10_sse2;
259             h->pred4x4[DIAG_DOWN_RIGHT_PRED] = ff_pred4x4_down_right_10_sse2;
260             h->pred4x4[VERT_LEFT_PRED      ] = ff_pred4x4_vertical_left_10_sse2;
261             h->pred4x4[VERT_RIGHT_PRED     ] = ff_pred4x4_vertical_right_10_sse2;
262             h->pred4x4[HOR_DOWN_PRED       ] = ff_pred4x4_horizontal_down_10_sse2;
263
264             h->pred8x8[VERT_PRED8x8        ] = ff_pred8x8_vertical_10_sse2;
265             h->pred8x8[HOR_PRED8x8         ] = ff_pred8x8_horizontal_10_sse2;
266         }
267         if (mm_flags & AV_CPU_FLAG_SSSE3) {
268             h->pred4x4[DIAG_DOWN_RIGHT_PRED] = ff_pred4x4_down_right_10_ssse3;
269             h->pred4x4[VERT_RIGHT_PRED     ] = ff_pred4x4_vertical_right_10_ssse3;
270             h->pred4x4[HOR_DOWN_PRED       ] = ff_pred4x4_horizontal_down_10_ssse3;
271         }
272 #if HAVE_AVX
273         if (mm_flags&AV_CPU_FLAG_AVX) {
274             h->pred4x4[DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_10_avx;
275             h->pred4x4[DIAG_DOWN_RIGHT_PRED] = ff_pred4x4_down_right_10_avx;
276             h->pred4x4[VERT_RIGHT_PRED     ] = ff_pred4x4_vertical_right_10_avx;
277             h->pred4x4[HOR_DOWN_PRED       ] = ff_pred4x4_horizontal_down_10_avx;
278         }
279 #endif /* HAVE_AVX */
280 #endif /* HAVE_YASM */
281     }
282 }