Mercurial > libavcodec.hg
annotate dsputil.h @ 1267:85b71f9f7450 libavcodec
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
author | michaelni |
---|---|
date | Thu, 15 May 2003 23:30:03 +0000 |
parents | 2fa34e615c76 |
children | a979fab41ed8 |
rev | line source |
---|---|
429 | 1 /* |
2 * DSP utils | |
3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard. | |
4 * | |
5 * This library is free software; you can redistribute it and/or | |
6 * modify it under the terms of the GNU Lesser General Public | |
7 * License as published by the Free Software Foundation; either | |
8 * version 2 of the License, or (at your option) any later version. | |
9 * | |
10 * This library is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 * Lesser General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU Lesser General Public | |
16 * License along with this library; if not, write to the Free Software | |
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
18 */ | |
1102 | 19 |
20 /** | |
21 * @file dsputil.h | |
1106 | 22 * DSP utils. |
1213 | 23 * note, many functions in here may use MMX which trashes the FPU state, it is |
24 * absolutely necessary to call emms_c() between dsp & float/double code | |
1102 | 25 */ |
26 | |
0 | 27 #ifndef DSPUTIL_H |
28 #define DSPUTIL_H | |
29 | |
30 #include "common.h" | |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
31 #include "avcodec.h" |
0 | 32 |
1102 | 33 |
255 | 34 //#define DEBUG |
0 | 35 /* dct code */ |
36 typedef short DCTELEM; | |
1008 | 37 //typedef int DCTELEM; |
0 | 38 |
474
11dbd00682fc
avoid name clash with libjpeg - added missing externs
bellard
parents:
429
diff
changeset
|
39 void fdct_ifast (DCTELEM *data); |
625
bb6a69f9d409
slow but accurate integer dct from IJG (should be ok with the LGPL as the old DCT is the fast integer DCT from IJG)
michaelni
parents:
623
diff
changeset
|
40 void ff_jpeg_fdct_islow (DCTELEM *data); |
0 | 41 |
42 void j_rev_dct (DCTELEM *data); | |
43 | |
687
9abb13c21fbe
fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents:
675
diff
changeset
|
44 void ff_fdct_mmx(DCTELEM *block); |
0 | 45 |
34 | 46 /* encoding scans */ |
1064 | 47 extern const uint8_t ff_alternate_horizontal_scan[64]; |
48 extern const uint8_t ff_alternate_vertical_scan[64]; | |
49 extern const uint8_t ff_zigzag_direct[64]; | |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
50 |
0 | 51 /* pixel operations */ |
52 #define MAX_NEG_CROP 384 | |
53 | |
54 /* temporary */ | |
1064 | 55 extern uint32_t squareTbl[512]; |
56 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP]; | |
0 | 57 |
58 | |
675 | 59 /* minimum alignment rules ;) |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
60 if u notice errors in the align stuff, need more alignment for some asm code for some cpu |
675 | 61 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ... |
62 | |
63 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible) | |
64 i (michael) didnt check them, these are just the alignents which i think could be reached easily ... | |
65 | |
66 !future video codecs might need functions with less strict alignment | |
67 */ | |
0 | 68 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
69 /* |
1064 | 70 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size); |
71 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride); | |
72 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
73 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
296 | 74 void clear_blocks_c(DCTELEM *blocks); |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
75 */ |
0 | 76 |
77 /* add and put pixel (decoding) */ | |
675 | 78 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16 |
1064 | 79 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h); |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
80 typedef void (*tpel_mc_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int w, int h); |
1064 | 81 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride); |
1168 | 82 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y); |
0 | 83 |
984 | 84 #define DEF_OLD_QPEL(name)\ |
1064 | 85 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ |
86 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ | |
87 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride); | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
88 |
984 | 89 DEF_OLD_QPEL(qpel16_mc11_old_c) |
90 DEF_OLD_QPEL(qpel16_mc31_old_c) | |
91 DEF_OLD_QPEL(qpel16_mc12_old_c) | |
92 DEF_OLD_QPEL(qpel16_mc32_old_c) | |
93 DEF_OLD_QPEL(qpel16_mc13_old_c) | |
94 DEF_OLD_QPEL(qpel16_mc33_old_c) | |
95 DEF_OLD_QPEL(qpel8_mc11_old_c) | |
96 DEF_OLD_QPEL(qpel8_mc31_old_c) | |
97 DEF_OLD_QPEL(qpel8_mc12_old_c) | |
98 DEF_OLD_QPEL(qpel8_mc32_old_c) | |
99 DEF_OLD_QPEL(qpel8_mc13_old_c) | |
100 DEF_OLD_QPEL(qpel8_mc33_old_c) | |
651 | 101 |
102 #define CALL_2X_PIXELS(a, b, n)\ | |
103 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\ | |
104 b(block , pixels , line_size, h);\ | |
105 b(block+n, pixels+n, line_size, h);\ | |
106 } | |
255 | 107 |
0 | 108 /* motion estimation */ |
109 | |
1064 | 110 typedef int (*op_pixels_abs_func)(uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/; |
936 | 111 |
1064 | 112 typedef int (*me_cmp_func)(void /*MpegEncContext*/ *s, uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/; |
936 | 113 |
1168 | 114 |
1102 | 115 /** |
116 * DSPContext. | |
117 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
118 typedef struct DSPContext { |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
119 /* pixel ops : interface with DCT */ |
1064 | 120 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size); |
121 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride); | |
122 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); | |
123 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); | |
1136 | 124 /** |
125 * translational global motion compensation. | |
126 */ | |
1064 | 127 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder); |
1136 | 128 /** |
129 * global motion compensation. | |
130 */ | |
1064 | 131 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy, |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
132 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height); |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
133 void (*clear_blocks)(DCTELEM *blocks/*align 16*/); |
1064 | 134 int (*pix_sum)(uint8_t * pix, int line_size); |
135 int (*pix_norm1)(uint8_t * pix, int line_size); | |
936 | 136 me_cmp_func sad[2]; /* identical to pix_absAxA except additional void * */ |
137 me_cmp_func sse[2]; | |
138 me_cmp_func hadamard8_diff[2]; | |
139 me_cmp_func dct_sad[2]; | |
140 me_cmp_func quant_psnr[2]; | |
1007 | 141 me_cmp_func bit[2]; |
142 me_cmp_func rd[2]; | |
936 | 143 int (*hadamard8_abs )(uint8_t *src, int stride, int mean); |
144 | |
954 | 145 me_cmp_func me_pre_cmp[11]; |
936 | 146 me_cmp_func me_cmp[11]; |
147 me_cmp_func me_sub_cmp[11]; | |
148 me_cmp_func mb_cmp[11]; | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
149 |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
150 /* maybe create an array for 16/8/4/2 functions */ |
1136 | 151 /** |
152 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
153 * this is an array[4][4] of motion compensation funcions for 4 |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
154 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br> |
1213 | 155 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 156 * @param block destination where the result is stored |
157 * @param pixels source | |
158 * @param line_size number of bytes in a horizontal line of block | |
159 * @param h height | |
160 */ | |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
161 op_pixels_func put_pixels_tab[4][4]; |
1136 | 162 |
163 /** | |
164 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
1225 | 165 * this is an array[2][4] of motion compensation funcions for 2 |
166 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> | |
1213 | 167 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 168 * @param block destination into which the result is averaged (a+b+1)>>1 |
169 * @param pixels source | |
170 * @param line_size number of bytes in a horizontal line of block | |
171 * @param h height | |
172 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
173 op_pixels_func avg_pixels_tab[2][4]; |
1136 | 174 |
175 /** | |
176 * Halfpel motion compensation with no rounding (a+b)>>1. | |
1225 | 177 * this is an array[2][4] of motion compensation funcions for 2 |
178 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> | |
1213 | 179 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 180 * @param block destination where the result is stored |
181 * @param pixels source | |
182 * @param line_size number of bytes in a horizontal line of block | |
183 * @param h height | |
184 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
185 op_pixels_func put_no_rnd_pixels_tab[2][4]; |
1136 | 186 |
187 /** | |
188 * Halfpel motion compensation with no rounding (a+b)>>1. | |
1225 | 189 * this is an array[2][4] of motion compensation funcions for 2 |
190 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> | |
1213 | 191 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 192 * @param block destination into which the result is averaged (a+b)>>1 |
193 * @param pixels source | |
194 * @param line_size number of bytes in a horizontal line of block | |
195 * @param h height | |
196 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
197 op_pixels_func avg_no_rnd_pixels_tab[2][4]; |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
198 |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
199 /** |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
200 * Thirdpel motion compensation with rounding (a+b+1)>>1. |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
201 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br> |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
202 * *pixels_tab[ xthirdpel + 4*ythirdpel ] |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
203 * @param block destination where the result is stored |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
204 * @param pixels source |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
205 * @param line_size number of bytes in a horizontal line of block |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
206 * @param h height |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
207 */ |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
208 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
209 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
210 qpel_mc_func put_qpel_pixels_tab[2][16]; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
211 qpel_mc_func avg_qpel_pixels_tab[2][16]; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
212 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16]; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
213 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16]; |
936 | 214 qpel_mc_func put_mspel_pixels_tab[8]; |
1168 | 215 |
216 /** | |
217 * h264 Chram MC | |
218 */ | |
219 h264_chroma_mc_func put_h264_chroma_pixels_tab[3]; | |
220 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3]; | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
221 |
1168 | 222 qpel_mc_func put_h264_qpel_pixels_tab[3][16]; |
223 qpel_mc_func avg_h264_qpel_pixels_tab[3][16]; | |
224 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
225 op_pixels_abs_func pix_abs16x16; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
226 op_pixels_abs_func pix_abs16x16_x2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
227 op_pixels_abs_func pix_abs16x16_y2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
228 op_pixels_abs_func pix_abs16x16_xy2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
229 op_pixels_abs_func pix_abs8x8; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
230 op_pixels_abs_func pix_abs8x8_x2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
231 op_pixels_abs_func pix_abs8x8_y2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
232 op_pixels_abs_func pix_abs8x8_xy2; |
866 | 233 |
234 /* huffyuv specific */ | |
235 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w); | |
936 | 236 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w); |
1092 | 237 |
238 /* (I)DCT */ | |
239 void (*fdct)(DCTELEM *block/* align 16*/); | |
1102 | 240 |
241 /** | |
1104 | 242 * block -> idct -> clip to unsigned 8 bit -> dest. |
1102 | 243 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...) |
1111 | 244 * @param line_size size in bytes of a horizotal line of dest |
1102 | 245 */ |
1092 | 246 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
1102 | 247 |
248 /** | |
249 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest. | |
1111 | 250 * @param line_size size in bytes of a horizotal line of dest |
1102 | 251 */ |
1092 | 252 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
1102 | 253 |
254 /** | |
1104 | 255 * idct input permutation. |
1241 | 256 * several optimized IDCTs need a permutated input (relative to the normal order of the reference |
257 * IDCT) | |
258 * this permutation must be performed before the idct_put/add, note, normally this can be merged | |
259 * with the zigzag/alternate scan<br> | |
1102 | 260 * an example to avoid confusion: |
261 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...) | |
262 * - (x -> referece dct -> reference idct -> x) | |
263 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x) | |
264 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...) | |
265 */ | |
1092 | 266 uint8_t idct_permutation[64]; |
267 int idct_permutation_type; | |
268 #define FF_NO_IDCT_PERM 1 | |
269 #define FF_LIBMPEG2_IDCT_PERM 2 | |
270 #define FF_SIMPLE_IDCT_PERM 3 | |
271 #define FF_TRANSPOSE_IDCT_PERM 4 | |
272 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
273 } DSPContext; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
274 |
1201 | 275 void dsputil_static_init(void); |
1092 | 276 void dsputil_init(DSPContext* p, AVCodecContext *avctx); |
0 | 277 |
764 | 278 /** |
279 * permute block according to permuatation. | |
280 * @param last last non zero element in scantable order | |
281 */ | |
1064 | 282 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last); |
34 | 283 |
1264 | 284 #define BYTE_VEC32(c) ((c)*0x01010101UL) |
285 | |
286 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b) | |
287 { | |
288 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
289 } | |
290 | |
291 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b) | |
292 { | |
293 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
294 } | |
295 | |
1102 | 296 /** |
1104 | 297 * Empty mmx state. |
1102 | 298 * this must be called between any dsp function and float/double code. |
299 * for example sin(); dsp->idct_put(); emms_c(); cos() | |
300 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
301 #define emms_c() |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
302 |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
303 /* should be defined by architectures supporting |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
304 one or more MultiMedia extension */ |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
305 int mm_support(void); |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
306 |
62 | 307 #if defined(HAVE_MMX) |
0 | 308 |
862 | 309 #undef emms_c |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
310 |
0 | 311 #define MM_MMX 0x0001 /* standard MMX */ |
312 #define MM_3DNOW 0x0004 /* AMD 3DNOW */ | |
313 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */ | |
314 #define MM_SSE 0x0008 /* SSE functions */ | |
315 #define MM_SSE2 0x0010 /* PIV SSE2 functions */ | |
316 | |
317 extern int mm_flags; | |
318 | |
1064 | 319 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
320 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); | |
0 | 321 |
322 static inline void emms(void) | |
323 { | |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
324 __asm __volatile ("emms;":::"memory"); |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
325 } |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
326 |
936 | 327 |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
328 #define emms_c() \ |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
329 {\ |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
330 if (mm_flags & MM_MMX)\ |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
331 emms();\ |
0 | 332 } |
333 | |
334 #define __align8 __attribute__ ((aligned (8))) | |
335 | |
1092 | 336 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx); |
337 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx); | |
1065 | 338 |
62 | 339 #elif defined(ARCH_ARMV4L) |
340 | |
341 /* This is to use 4 bytes read to the IDCT pointers for some 'zero' | |
342 line ptimizations */ | |
343 #define __align8 __attribute__ ((aligned (4))) | |
344 | |
1092 | 345 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx); |
62 | 346 |
88 | 347 #elif defined(HAVE_MLIB) |
348 | |
349 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */ | |
350 #define __align8 __attribute__ ((aligned (8))) | |
351 | |
1092 | 352 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx); |
88 | 353 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
354 #elif defined(ARCH_ALPHA) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
355 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
356 #define __align8 __attribute__ ((aligned (8))) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
357 |
1092 | 358 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx); |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
359 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
612
diff
changeset
|
360 #elif defined(ARCH_POWERPC) |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
612
diff
changeset
|
361 |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
362 #define MM_ALTIVEC 0x0001 /* standard AltiVec */ |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
363 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
364 extern int mm_flags; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
365 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
366 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN) |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
367 #include <altivec.h> |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
368 #endif |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
369 |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
612
diff
changeset
|
370 #define __align8 __attribute__ ((aligned (16))) |
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
612
diff
changeset
|
371 |
1092 | 372 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
612
diff
changeset
|
373 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
374 #elif defined(HAVE_MMI) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
375 |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
376 #define __align8 __attribute__ ((aligned (16))) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
377 |
1092 | 378 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx); |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
379 |
1259
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
380 #elif defined(ARCH_SH4) |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
381 |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
382 #define __align8 __attribute__ ((aligned (8))) |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
383 |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
384 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx); |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
385 |
0 | 386 #else |
387 | |
388 #define __align8 | |
389 | |
390 #endif | |
391 | |
493
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
392 #ifdef __GNUC__ |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
393 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
394 struct unaligned_64 { uint64_t l; } __attribute__((packed)); |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
395 struct unaligned_32 { uint32_t l; } __attribute__((packed)); |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
396 struct unaligned_16 { uint16_t l; } __attribute__((packed)); |
493
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
397 |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
398 #define LD16(a) (((const struct unaligned_16 *) (a))->l) |
493
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
399 #define LD32(a) (((const struct unaligned_32 *) (a))->l) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
400 #define LD64(a) (((const struct unaligned_64 *) (a))->l) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
401 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
402 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
403 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
404 #else /* __GNUC__ */ |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
405 |
1267
85b71f9f7450
moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents:
1264
diff
changeset
|
406 #define LD16(a) (*((uint16_t*)(a))) |
493
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
407 #define LD32(a) (*((uint32_t*)(a))) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
408 #define LD64(a) (*((uint64_t*)(a))) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
409 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
410 #define ST32(a, b) *((uint32_t*)(a)) = (b) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
411 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
412 #endif /* !__GNUC__ */ |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
413 |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
414 /* PSNR */ |
1064 | 415 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3], |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
416 int orig_linesize[3], int coded_linesize, |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
417 AVCodecContext *avctx); |
781 | 418 |
419 /* FFT computation */ | |
420 | |
421 /* NOTE: soon integer code will be added, so you must use the | |
422 FFTSample type */ | |
423 typedef float FFTSample; | |
424 | |
425 typedef struct FFTComplex { | |
426 FFTSample re, im; | |
427 } FFTComplex; | |
428 | |
429 typedef struct FFTContext { | |
430 int nbits; | |
431 int inverse; | |
432 uint16_t *revtab; | |
433 FFTComplex *exptab; | |
434 FFTComplex *exptab1; /* only used by SSE code */ | |
435 void (*fft_calc)(struct FFTContext *s, FFTComplex *z); | |
436 } FFTContext; | |
437 | |
438 int fft_init(FFTContext *s, int nbits, int inverse); | |
439 void fft_permute(FFTContext *s, FFTComplex *z); | |
440 void fft_calc_c(FFTContext *s, FFTComplex *z); | |
441 void fft_calc_sse(FFTContext *s, FFTComplex *z); | |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
954
diff
changeset
|
442 void fft_calc_altivec(FFTContext *s, FFTComplex *z); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
954
diff
changeset
|
443 |
781 | 444 static inline void fft_calc(FFTContext *s, FFTComplex *z) |
445 { | |
446 s->fft_calc(s, z); | |
447 } | |
448 void fft_end(FFTContext *s); | |
449 | |
450 /* MDCT computation */ | |
451 | |
452 typedef struct MDCTContext { | |
453 int n; /* size of MDCT (i.e. number of input data * 2) */ | |
454 int nbits; /* n = 2^nbits */ | |
455 /* pre/post rotation tables */ | |
456 FFTSample *tcos; | |
457 FFTSample *tsin; | |
458 FFTContext fft; | |
459 } MDCTContext; | |
460 | |
794 | 461 int ff_mdct_init(MDCTContext *s, int nbits, int inverse); |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
462 void ff_imdct_calc(MDCTContext *s, FFTSample *output, |
781 | 463 const FFTSample *input, FFTSample *tmp); |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
464 void ff_mdct_calc(MDCTContext *s, FFTSample *out, |
781 | 465 const FFTSample *input, FFTSample *tmp); |
794 | 466 void ff_mdct_end(MDCTContext *s); |
781 | 467 |
936 | 468 #define WARPER88_1616(name8, name16)\ |
469 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride){\ | |
470 return name8(s, dst , src , stride)\ | |
471 +name8(s, dst+8 , src+8 , stride)\ | |
472 +name8(s, dst +8*stride, src +8*stride, stride)\ | |
473 +name8(s, dst+8+8*stride, src+8+8*stride, stride);\ | |
474 } | |
475 | |
838
b78812db886f
lrintf detection (based upon a patch by Franois Revol <revol at free dot fr>)
michaelni
parents:
802
diff
changeset
|
476 #ifndef HAVE_LRINTF |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
477 /* XXX: add ISOC specific test to avoid specific BSD testing. */ |
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
478 /* better than nothing implementation. */ |
802 | 479 /* btw, rintf() is existing on fbsd too -- alex */ |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
480 static inline long int lrintf(float x) |
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
481 { |
1040
998d5035b15b
win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents:
1033
diff
changeset
|
482 #ifdef CONFIG_WIN32 |
998d5035b15b
win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents:
1033
diff
changeset
|
483 /* XXX: incorrect, but make it compile */ |
998d5035b15b
win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents:
1033
diff
changeset
|
484 return (int)(x); |
998d5035b15b
win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents:
1033
diff
changeset
|
485 #else |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
486 return (int)(rint(x)); |
1040
998d5035b15b
win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents:
1033
diff
changeset
|
487 #endif |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
488 } |
0 | 489 #endif |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
490 |
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
491 #endif |