Mercurial > libavcodec.hg
annotate dsputil.h @ 1709:39a209b0c82c libavcodec
info about h and tpel
author | michael |
---|---|
date | Tue, 30 Dec 2003 19:19:45 +0000 |
parents | dea5b2946999 |
children | 49e46374baa4 |
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; | |
37 | |
474
11dbd00682fc
avoid name clash with libjpeg - added missing externs
bellard
parents:
429
diff
changeset
|
38 void fdct_ifast (DCTELEM *data); |
1571
aa4dc16c0f18
* adding integer/floating point AAN implementations for DCT 2-4-8
romansh
parents:
1567
diff
changeset
|
39 void fdct_ifast248 (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); |
1567 | 41 void ff_fdct248_islow (DCTELEM *data); |
0 | 42 |
43 void j_rev_dct (DCTELEM *data); | |
44 | |
687
9abb13c21fbe
fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents:
675
diff
changeset
|
45 void ff_fdct_mmx(DCTELEM *block); |
1565 | 46 void ff_fdct_mmx2(DCTELEM *block); |
0 | 47 |
34 | 48 /* encoding scans */ |
1064 | 49 extern const uint8_t ff_alternate_horizontal_scan[64]; |
50 extern const uint8_t ff_alternate_vertical_scan[64]; | |
51 extern const uint8_t ff_zigzag_direct[64]; | |
1567 | 52 extern const uint8_t ff_zigzag248_direct[64]; |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
53 |
0 | 54 /* pixel operations */ |
55 #define MAX_NEG_CROP 384 | |
56 | |
57 /* temporary */ | |
1064 | 58 extern uint32_t squareTbl[512]; |
59 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP]; | |
0 | 60 |
61 | |
675 | 62 /* minimum alignment rules ;) |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
63 if u notice errors in the align stuff, need more alignment for some asm code for some cpu |
675 | 64 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ... |
65 | |
66 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible) | |
67 i (michael) didnt check them, these are just the alignents which i think could be reached easily ... | |
68 | |
69 !future video codecs might need functions with less strict alignment | |
70 */ | |
0 | 71 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
72 /* |
1064 | 73 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size); |
74 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride); | |
75 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
76 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
296 | 77 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
|
78 */ |
0 | 79 |
80 /* add and put pixel (decoding) */ | |
675 | 81 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16 |
1709 | 82 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4 |
1064 | 83 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
|
84 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 | 85 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride); |
1168 | 86 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 | 87 |
984 | 88 #define DEF_OLD_QPEL(name)\ |
1064 | 89 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ |
90 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ | |
91 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
|
92 |
984 | 93 DEF_OLD_QPEL(qpel16_mc11_old_c) |
94 DEF_OLD_QPEL(qpel16_mc31_old_c) | |
95 DEF_OLD_QPEL(qpel16_mc12_old_c) | |
96 DEF_OLD_QPEL(qpel16_mc32_old_c) | |
97 DEF_OLD_QPEL(qpel16_mc13_old_c) | |
98 DEF_OLD_QPEL(qpel16_mc33_old_c) | |
99 DEF_OLD_QPEL(qpel8_mc11_old_c) | |
100 DEF_OLD_QPEL(qpel8_mc31_old_c) | |
101 DEF_OLD_QPEL(qpel8_mc12_old_c) | |
102 DEF_OLD_QPEL(qpel8_mc32_old_c) | |
103 DEF_OLD_QPEL(qpel8_mc13_old_c) | |
104 DEF_OLD_QPEL(qpel8_mc33_old_c) | |
651 | 105 |
106 #define CALL_2X_PIXELS(a, b, n)\ | |
107 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\ | |
108 b(block , pixels , line_size, h);\ | |
109 b(block+n, pixels+n, line_size, h);\ | |
110 } | |
255 | 111 |
0 | 112 /* motion estimation */ |
1709 | 113 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2 |
114 // allthough currently h<4 is not used as functions with width <8 are not used and neither implemented | |
1708 | 115 typedef int (*me_cmp_func)(void /*MpegEncContext*/ *s, uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size, int h)/* __attribute__ ((const))*/; |
936 | 116 |
1168 | 117 |
1102 | 118 /** |
119 * DSPContext. | |
120 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
121 typedef struct DSPContext { |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
122 /* pixel ops : interface with DCT */ |
1064 | 123 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size); |
124 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride); | |
125 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); | |
126 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); | |
1136 | 127 /** |
128 * translational global motion compensation. | |
129 */ | |
1064 | 130 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder); |
1136 | 131 /** |
132 * global motion compensation. | |
133 */ | |
1064 | 134 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
|
135 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
|
136 void (*clear_blocks)(DCTELEM *blocks/*align 16*/); |
1064 | 137 int (*pix_sum)(uint8_t * pix, int line_size); |
138 int (*pix_norm1)(uint8_t * pix, int line_size); | |
1708 | 139 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4 |
140 | |
141 me_cmp_func sad[4]; /* identical to pix_absAxA except additional void * */ | |
142 me_cmp_func sse[4]; | |
143 me_cmp_func hadamard8_diff[4]; | |
144 me_cmp_func dct_sad[4]; | |
145 me_cmp_func quant_psnr[4]; | |
146 me_cmp_func bit[4]; | |
147 me_cmp_func rd[4]; | |
936 | 148 int (*hadamard8_abs )(uint8_t *src, int stride, int mean); |
149 | |
1708 | 150 me_cmp_func me_pre_cmp[5]; |
151 me_cmp_func me_cmp[5]; | |
152 me_cmp_func me_sub_cmp[5]; | |
153 me_cmp_func mb_cmp[5]; | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
154 |
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
|
155 /* maybe create an array for 16/8/4/2 functions */ |
1136 | 156 /** |
157 * 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
|
158 * 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
|
159 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br> |
1213 | 160 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 161 * @param block destination where the result is stored |
162 * @param pixels source | |
163 * @param line_size number of bytes in a horizontal line of block | |
164 * @param h height | |
165 */ | |
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
|
166 op_pixels_func put_pixels_tab[4][4]; |
1136 | 167 |
168 /** | |
169 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
1320 | 170 * This is an array[4][4] of motion compensation functions for 4 |
1319 | 171 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br> |
1213 | 172 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 173 * @param block destination into which the result is averaged (a+b+1)>>1 |
174 * @param pixels source | |
175 * @param line_size number of bytes in a horizontal line of block | |
176 * @param h height | |
177 */ | |
1319 | 178 op_pixels_func avg_pixels_tab[4][4]; |
1136 | 179 |
180 /** | |
181 * Halfpel motion compensation with no rounding (a+b)>>1. | |
1225 | 182 * this is an array[2][4] of motion compensation funcions for 2 |
183 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> | |
1213 | 184 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 185 * @param block destination where the result is stored |
186 * @param pixels source | |
187 * @param line_size number of bytes in a horizontal line of block | |
188 * @param h height | |
189 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
190 op_pixels_func put_no_rnd_pixels_tab[2][4]; |
1136 | 191 |
192 /** | |
193 * Halfpel motion compensation with no rounding (a+b)>>1. | |
1225 | 194 * this is an array[2][4] of motion compensation funcions for 2 |
195 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> | |
1213 | 196 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 197 * @param block destination into which the result is averaged (a+b)>>1 |
198 * @param pixels source | |
199 * @param line_size number of bytes in a horizontal line of block | |
200 * @param h height | |
201 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
202 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
|
203 |
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 /** |
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 * 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
|
206 * 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
|
207 * *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
|
208 * @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
|
209 * @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
|
210 * @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
|
211 * @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
|
212 */ |
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
|
213 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
1319 | 214 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
215 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
216 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
|
217 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
|
218 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
|
219 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16]; |
936 | 220 qpel_mc_func put_mspel_pixels_tab[8]; |
1168 | 221 |
222 /** | |
223 * h264 Chram MC | |
224 */ | |
225 h264_chroma_mc_func put_h264_chroma_pixels_tab[3]; | |
226 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
|
227 |
1168 | 228 qpel_mc_func put_h264_qpel_pixels_tab[3][16]; |
229 qpel_mc_func avg_h264_qpel_pixels_tab[3][16]; | |
230 | |
1708 | 231 me_cmp_func pix_abs[2][4]; |
866 | 232 |
233 /* huffyuv specific */ | |
234 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w); | |
936 | 235 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w); |
1527 | 236 /** |
237 * subtract huffyuv's variant of median prediction | |
238 * note, this might read from src1[-1], src2[-1] | |
239 */ | |
240 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top); | |
1273 | 241 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w); |
1092 | 242 |
1644 | 243 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale); |
244 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale); | |
245 | |
1092 | 246 /* (I)DCT */ |
247 void (*fdct)(DCTELEM *block/* align 16*/); | |
1567 | 248 void (*fdct248)(DCTELEM *block/* align 16*/); |
1102 | 249 |
1324
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
250 /* IDCT really*/ |
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
251 void (*idct)(DCTELEM *block/* align 16*/); |
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
252 |
1102 | 253 /** |
1104 | 254 * block -> idct -> clip to unsigned 8 bit -> dest. |
1102 | 255 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...) |
1111 | 256 * @param line_size size in bytes of a horizotal line of dest |
1102 | 257 */ |
1092 | 258 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
1102 | 259 |
260 /** | |
261 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest. | |
1111 | 262 * @param line_size size in bytes of a horizotal line of dest |
1102 | 263 */ |
1092 | 264 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
1102 | 265 |
266 /** | |
1104 | 267 * idct input permutation. |
1241 | 268 * several optimized IDCTs need a permutated input (relative to the normal order of the reference |
269 * IDCT) | |
270 * this permutation must be performed before the idct_put/add, note, normally this can be merged | |
271 * with the zigzag/alternate scan<br> | |
1102 | 272 * an example to avoid confusion: |
273 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...) | |
274 * - (x -> referece dct -> reference idct -> x) | |
275 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x) | |
276 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...) | |
277 */ | |
1092 | 278 uint8_t idct_permutation[64]; |
279 int idct_permutation_type; | |
280 #define FF_NO_IDCT_PERM 1 | |
281 #define FF_LIBMPEG2_IDCT_PERM 2 | |
282 #define FF_SIMPLE_IDCT_PERM 3 | |
283 #define FF_TRANSPOSE_IDCT_PERM 4 | |
284 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
285 } DSPContext; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
286 |
1201 | 287 void dsputil_static_init(void); |
1092 | 288 void dsputil_init(DSPContext* p, AVCodecContext *avctx); |
0 | 289 |
764 | 290 /** |
291 * permute block according to permuatation. | |
292 * @param last last non zero element in scantable order | |
293 */ | |
1064 | 294 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last); |
34 | 295 |
1264 | 296 #define BYTE_VEC32(c) ((c)*0x01010101UL) |
297 | |
298 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b) | |
299 { | |
300 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
301 } | |
302 | |
303 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b) | |
304 { | |
305 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
306 } | |
307 | |
1102 | 308 /** |
1104 | 309 * Empty mmx state. |
1102 | 310 * this must be called between any dsp function and float/double code. |
311 * for example sin(); dsp->idct_put(); emms_c(); cos() | |
312 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
313 #define emms_c() |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
314 |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
315 /* should be defined by architectures supporting |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
316 one or more MultiMedia extension */ |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
317 int mm_support(void); |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
318 |
62 | 319 #if defined(HAVE_MMX) |
0 | 320 |
862 | 321 #undef emms_c |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
322 |
0 | 323 #define MM_MMX 0x0001 /* standard MMX */ |
324 #define MM_3DNOW 0x0004 /* AMD 3DNOW */ | |
325 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */ | |
326 #define MM_SSE 0x0008 /* SSE functions */ | |
327 #define MM_SSE2 0x0010 /* PIV SSE2 functions */ | |
328 | |
329 extern int mm_flags; | |
330 | |
1064 | 331 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
332 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); | |
0 | 333 |
334 static inline void emms(void) | |
335 { | |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
336 __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
|
337 } |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
338 |
936 | 339 |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
340 #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
|
341 {\ |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
342 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
|
343 emms();\ |
0 | 344 } |
345 | |
346 #define __align8 __attribute__ ((aligned (8))) | |
347 | |
1092 | 348 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx); |
349 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx); | |
1065 | 350 |
62 | 351 #elif defined(ARCH_ARMV4L) |
352 | |
353 /* This is to use 4 bytes read to the IDCT pointers for some 'zero' | |
354 line ptimizations */ | |
355 #define __align8 __attribute__ ((aligned (4))) | |
356 | |
1092 | 357 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx); |
62 | 358 |
88 | 359 #elif defined(HAVE_MLIB) |
360 | |
361 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */ | |
362 #define __align8 __attribute__ ((aligned (8))) | |
363 | |
1092 | 364 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx); |
88 | 365 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
366 #elif defined(ARCH_ALPHA) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
367 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
368 #define __align8 __attribute__ ((aligned (8))) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
369 |
1092 | 370 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
|
371 |
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
|
372 #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
|
373 |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
374 #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
|
375 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
376 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
|
377 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
378 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN) |
1653 | 379 #define pixel altivec_pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
380 #include <altivec.h> |
1653 | 381 #undef pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
382 #endif |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
383 |
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
|
384 #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
|
385 |
1092 | 386 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
|
387 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
388 #elif defined(HAVE_MMI) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
389 |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
390 #define __align8 __attribute__ ((aligned (16))) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
391 |
1092 | 392 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
|
393 |
1259
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
394 #elif defined(ARCH_SH4) |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
395 |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
396 #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
|
397 |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
398 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
|
399 |
0 | 400 #else |
401 | |
402 #define __align8 | |
403 | |
404 #endif | |
405 | |
493
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
406 #ifdef __GNUC__ |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
407 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
408 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
|
409 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
|
410 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
|
411 |
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
|
412 #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
|
413 #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
|
414 #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
|
415 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
416 #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
|
417 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
418 #else /* __GNUC__ */ |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
419 |
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
|
420 #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
|
421 #define LD32(a) (*((uint32_t*)(a))) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
422 #define LD64(a) (*((uint64_t*)(a))) |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
423 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
424 #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
|
425 |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
426 #endif /* !__GNUC__ */ |
873b9075d853
move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents:
480
diff
changeset
|
427 |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
428 /* PSNR */ |
1064 | 429 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
|
430 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
|
431 AVCodecContext *avctx); |
781 | 432 |
433 /* FFT computation */ | |
434 | |
435 /* NOTE: soon integer code will be added, so you must use the | |
436 FFTSample type */ | |
437 typedef float FFTSample; | |
438 | |
439 typedef struct FFTComplex { | |
440 FFTSample re, im; | |
441 } FFTComplex; | |
442 | |
443 typedef struct FFTContext { | |
444 int nbits; | |
445 int inverse; | |
446 uint16_t *revtab; | |
447 FFTComplex *exptab; | |
448 FFTComplex *exptab1; /* only used by SSE code */ | |
449 void (*fft_calc)(struct FFTContext *s, FFTComplex *z); | |
450 } FFTContext; | |
451 | |
452 int fft_init(FFTContext *s, int nbits, int inverse); | |
453 void fft_permute(FFTContext *s, FFTComplex *z); | |
454 void fft_calc_c(FFTContext *s, FFTComplex *z); | |
455 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
|
456 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
|
457 |
781 | 458 static inline void fft_calc(FFTContext *s, FFTComplex *z) |
459 { | |
460 s->fft_calc(s, z); | |
461 } | |
462 void fft_end(FFTContext *s); | |
463 | |
464 /* MDCT computation */ | |
465 | |
466 typedef struct MDCTContext { | |
467 int n; /* size of MDCT (i.e. number of input data * 2) */ | |
468 int nbits; /* n = 2^nbits */ | |
469 /* pre/post rotation tables */ | |
470 FFTSample *tcos; | |
471 FFTSample *tsin; | |
472 FFTContext fft; | |
473 } MDCTContext; | |
474 | |
794 | 475 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
|
476 void ff_imdct_calc(MDCTContext *s, FFTSample *output, |
781 | 477 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
|
478 void ff_mdct_calc(MDCTContext *s, FFTSample *out, |
781 | 479 const FFTSample *input, FFTSample *tmp); |
794 | 480 void ff_mdct_end(MDCTContext *s); |
781 | 481 |
1708 | 482 #define WARPER8_16(name8, name16)\ |
483 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
484 return name8(s, dst , src , stride, h)\ | |
485 +name8(s, dst+8 , src+8 , stride, h);\ | |
486 } | |
487 | |
488 #define WARPER8_16_SQ(name8, name16)\ | |
489 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
490 int score=0;\ | |
491 score +=name8(s, dst , src , stride, 8);\ | |
492 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
493 if(h==16){\ | |
494 dst += 8*stride;\ | |
495 src += 8*stride;\ | |
496 score +=name8(s, dst , src , stride, 8);\ | |
497 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
498 }\ | |
499 return score;\ | |
936 | 500 } |
501 | |
838
b78812db886f
lrintf detection (based upon a patch by Franois Revol <revol at free dot fr>)
michaelni
parents:
802
diff
changeset
|
502 #ifndef HAVE_LRINTF |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
503 /* 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
|
504 /* better than nothing implementation. */ |
802 | 505 /* 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
|
506 static inline long int lrintf(float x) |
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
507 { |
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
|
508 #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
|
509 /* 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
|
510 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
|
511 #else |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
512 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
|
513 #endif |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
514 } |
0 | 515 #endif |
796
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
516 |
8a5b70c68fbd
added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents:
794
diff
changeset
|
517 #endif |