Mercurial > libavcodec.hg
annotate dsputil.h @ 5660:cd95f83ac235 libavcodec
spatial_decomposition_count != 5 bugfix
author | michael |
---|---|
date | Sun, 09 Sep 2007 12:24:45 +0000 |
parents | 8691652d9dce |
children | ec04964a1d1a |
rev | line source |
---|---|
429 | 1 /* |
2 * DSP utils | |
3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard. | |
1739
07a484280a82
copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents:
1729
diff
changeset
|
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at> |
429 | 5 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
6 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
7 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
8 * FFmpeg is free software; you can redistribute it and/or |
429 | 9 * modify it under the terms of the GNU Lesser General Public |
10 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
11 * version 2.1 of the License, or (at your option) any later version. |
429 | 12 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
13 * FFmpeg is distributed in the hope that it will be useful, |
429 | 14 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 * Lesser General Public License for more details. | |
17 * | |
18 * You should have received a copy of the GNU Lesser General Public | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
19 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
3029
diff
changeset
|
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
429 | 21 */ |
1102 | 22 |
23 /** | |
24 * @file dsputil.h | |
1106 | 25 * DSP utils. |
1213 | 26 * note, many functions in here may use MMX which trashes the FPU state, it is |
27 * absolutely necessary to call emms_c() between dsp & float/double code | |
1102 | 28 */ |
29 | |
0 | 30 #ifndef DSPUTIL_H |
31 #define DSPUTIL_H | |
32 | |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
33 #include "avcodec.h" |
0 | 34 |
1102 | 35 |
255 | 36 //#define DEBUG |
0 | 37 /* dct code */ |
38 typedef short DCTELEM; | |
3198
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
39 typedef int DWTELEM; |
5587
3ae03eacbe9f
use 16bit IDWT (a SIMD implementation of it should be >2x faster then with
michael
parents:
5573
diff
changeset
|
40 typedef short IDWTELEM; |
0 | 41 |
474
11dbd00682fc
avoid name clash with libjpeg - added missing externs
bellard
parents:
429
diff
changeset
|
42 void fdct_ifast (DCTELEM *data); |
1571
aa4dc16c0f18
* adding integer/floating point AAN implementations for DCT 2-4-8
romansh
parents:
1567
diff
changeset
|
43 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
|
44 void ff_jpeg_fdct_islow (DCTELEM *data); |
1567 | 45 void ff_fdct248_islow (DCTELEM *data); |
0 | 46 |
47 void j_rev_dct (DCTELEM *data); | |
2256 | 48 void j_rev_dct4 (DCTELEM *data); |
2257 | 49 void j_rev_dct2 (DCTELEM *data); |
2259 | 50 void j_rev_dct1 (DCTELEM *data); |
0 | 51 |
687
9abb13c21fbe
fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents:
675
diff
changeset
|
52 void ff_fdct_mmx(DCTELEM *block); |
1565 | 53 void ff_fdct_mmx2(DCTELEM *block); |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
54 void ff_fdct_sse2(DCTELEM *block); |
0 | 55 |
2755 | 56 void ff_h264_idct8_add_c(uint8_t *dst, DCTELEM *block, int stride); |
2272
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
57 void ff_h264_idct_add_c(uint8_t *dst, DCTELEM *block, int stride); |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
58 void ff_h264_idct8_dc_add_c(uint8_t *dst, DCTELEM *block, int stride); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
59 void ff_h264_idct_dc_add_c(uint8_t *dst, DCTELEM *block, int stride); |
2272
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
60 void ff_h264_lowres_idct_add_c(uint8_t *dst, int stride, DCTELEM *block); |
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
61 void ff_h264_lowres_idct_put_c(uint8_t *dst, int stride, DCTELEM *block); |
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
62 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
63 void ff_vector_fmul_add_add_c(float *dst, const float *src0, const float *src1, |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
64 const float *src2, int src3, int blocksize, int step); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
65 void ff_float_to_int16_c(int16_t *dst, const float *src, int len); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
66 |
34 | 67 /* encoding scans */ |
1064 | 68 extern const uint8_t ff_alternate_horizontal_scan[64]; |
69 extern const uint8_t ff_alternate_vertical_scan[64]; | |
70 extern const uint8_t ff_zigzag_direct[64]; | |
1567 | 71 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
|
72 |
0 | 73 /* pixel operations */ |
2090 | 74 #define MAX_NEG_CROP 1024 |
0 | 75 |
76 /* temporary */ | |
4179 | 77 extern uint32_t ff_squareTbl[512]; |
4176 | 78 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP]; |
0 | 79 |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
1864
diff
changeset
|
80 /* VP3 DSP functions */ |
2693 | 81 void ff_vp3_idct_c(DCTELEM *block/* align 16*/); |
82 void ff_vp3_idct_put_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); | |
83 void ff_vp3_idct_add_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); | |
0 | 84 |
3245 | 85 /* 1/2^n downscaling functions from imgconvert.c */ |
86 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
87 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
88 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
89 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
90 | |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3245
diff
changeset
|
91 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy, |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3245
diff
changeset
|
92 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3245
diff
changeset
|
93 |
675 | 94 /* minimum alignment rules ;) |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
95 if u notice errors in the align stuff, need more alignment for some asm code for some cpu |
675 | 96 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ... |
97 | |
98 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible) | |
99 i (michael) didnt check them, these are just the alignents which i think could be reached easily ... | |
100 | |
101 !future video codecs might need functions with less strict alignment | |
102 */ | |
0 | 103 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
104 /* |
1064 | 105 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size); |
106 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride); | |
107 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
108 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
296 | 109 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
|
110 */ |
0 | 111 |
112 /* add and put pixel (decoding) */ | |
675 | 113 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16 |
1709 | 114 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4 |
1064 | 115 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
|
116 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 | 117 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride); |
1168 | 118 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y); |
2415 | 119 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset); |
3029 | 120 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset); |
0 | 121 |
984 | 122 #define DEF_OLD_QPEL(name)\ |
1064 | 123 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ |
124 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ | |
125 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
|
126 |
984 | 127 DEF_OLD_QPEL(qpel16_mc11_old_c) |
128 DEF_OLD_QPEL(qpel16_mc31_old_c) | |
129 DEF_OLD_QPEL(qpel16_mc12_old_c) | |
130 DEF_OLD_QPEL(qpel16_mc32_old_c) | |
131 DEF_OLD_QPEL(qpel16_mc13_old_c) | |
132 DEF_OLD_QPEL(qpel16_mc33_old_c) | |
133 DEF_OLD_QPEL(qpel8_mc11_old_c) | |
134 DEF_OLD_QPEL(qpel8_mc31_old_c) | |
135 DEF_OLD_QPEL(qpel8_mc12_old_c) | |
136 DEF_OLD_QPEL(qpel8_mc32_old_c) | |
137 DEF_OLD_QPEL(qpel8_mc13_old_c) | |
138 DEF_OLD_QPEL(qpel8_mc33_old_c) | |
651 | 139 |
140 #define CALL_2X_PIXELS(a, b, n)\ | |
141 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\ | |
142 b(block , pixels , line_size, h);\ | |
143 b(block+n, pixels+n, line_size, h);\ | |
144 } | |
255 | 145 |
0 | 146 /* motion estimation */ |
1709 | 147 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2 |
5127 | 148 // although currently h<4 is not used as functions with width <8 are neither used nor implemented |
1708 | 149 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 | 150 |
1168 | 151 |
3198
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
152 // for snow slices |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
153 typedef struct slice_buffer_s slice_buffer; |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
154 |
1102 | 155 /** |
156 * DSPContext. | |
157 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
158 typedef struct DSPContext { |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
159 /* pixel ops : interface with DCT */ |
1064 | 160 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size); |
161 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride); | |
162 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); | |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
163 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); |
1064 | 164 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); |
2763 | 165 void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size); |
166 void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size); | |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4962
diff
changeset
|
167 int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/); |
1136 | 168 /** |
169 * translational global motion compensation. | |
170 */ | |
1064 | 171 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder); |
1136 | 172 /** |
173 * global motion compensation. | |
174 */ | |
1064 | 175 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy, |
2979 | 176 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height); |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
177 void (*clear_blocks)(DCTELEM *blocks/*align 16*/); |
1064 | 178 int (*pix_sum)(uint8_t * pix, int line_size); |
179 int (*pix_norm1)(uint8_t * pix, int line_size); | |
1708 | 180 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4 |
2967 | 181 |
1729 | 182 me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */ |
183 me_cmp_func sse[5]; | |
184 me_cmp_func hadamard8_diff[5]; | |
185 me_cmp_func dct_sad[5]; | |
186 me_cmp_func quant_psnr[5]; | |
187 me_cmp_func bit[5]; | |
188 me_cmp_func rd[5]; | |
189 me_cmp_func vsad[5]; | |
190 me_cmp_func vsse[5]; | |
2065
9e4bebc39ade
noise preserving sum of squares comparission function
michael
parents:
2045
diff
changeset
|
191 me_cmp_func nsse[5]; |
2184 | 192 me_cmp_func w53[5]; |
193 me_cmp_func w97[5]; | |
2382 | 194 me_cmp_func dct_max[5]; |
3010
533c6386eca9
8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents:
2979
diff
changeset
|
195 me_cmp_func dct264_sad[5]; |
936 | 196 |
1708 | 197 me_cmp_func me_pre_cmp[5]; |
198 me_cmp_func me_cmp[5]; | |
199 me_cmp_func me_sub_cmp[5]; | |
200 me_cmp_func mb_cmp[5]; | |
1729 | 201 me_cmp_func ildct_cmp[5]; //only width 16 used |
2382 | 202 me_cmp_func frame_skip_cmp[5]; //only width 8 used |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
203 |
5255 | 204 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2, |
205 int size); | |
4749 | 206 |
1136 | 207 /** |
208 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
4751 | 209 * this is an array[4][4] of motion compensation functions for 4 |
1713 | 210 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 211 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 212 * @param block destination where the result is stored |
213 * @param pixels source | |
214 * @param line_size number of bytes in a horizontal line of block | |
215 * @param h height | |
216 */ | |
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
|
217 op_pixels_func put_pixels_tab[4][4]; |
1136 | 218 |
219 /** | |
220 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
2967 | 221 * This is an array[4][4] of motion compensation functions for 4 |
1713 | 222 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 223 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 224 * @param block destination into which the result is averaged (a+b+1)>>1 |
225 * @param pixels source | |
226 * @param line_size number of bytes in a horizontal line of block | |
227 * @param h height | |
228 */ | |
1319 | 229 op_pixels_func avg_pixels_tab[4][4]; |
1136 | 230 |
231 /** | |
232 * Halfpel motion compensation with no rounding (a+b)>>1. | |
4751 | 233 * this is an array[2][4] of motion compensation functions for 2 |
1225 | 234 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 235 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 236 * @param block destination where the result is stored |
237 * @param pixels source | |
238 * @param line_size number of bytes in a horizontal line of block | |
239 * @param h height | |
240 */ | |
2075 | 241 op_pixels_func put_no_rnd_pixels_tab[4][4]; |
1136 | 242 |
243 /** | |
244 * Halfpel motion compensation with no rounding (a+b)>>1. | |
4751 | 245 * this is an array[2][4] of motion compensation functions for 2 |
1225 | 246 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 247 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 248 * @param block destination into which the result is averaged (a+b)>>1 |
249 * @param pixels source | |
250 * @param line_size number of bytes in a horizontal line of block | |
251 * @param h height | |
252 */ | |
2075 | 253 op_pixels_func avg_no_rnd_pixels_tab[4][4]; |
2967 | 254 |
1864 | 255 void (*put_no_rnd_pixels_l2[2])(uint8_t *block/*align width (8 or 16)*/, const uint8_t *a/*align 1*/, const uint8_t *b/*align 1*/, int line_size, int h); |
2967 | 256 |
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
|
257 /** |
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
|
258 * Thirdpel motion compensation with rounding (a+b+1)>>1. |
4751 | 259 * this is an array[12] of motion compensation functions for the 9 thirdpe |
260 * positions<br> | |
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
|
261 * *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
|
262 * @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
|
263 * @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
|
264 * @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
|
265 * @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
|
266 */ |
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
|
267 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
1319 | 268 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
269 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
270 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
|
271 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
|
272 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
|
273 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16]; |
936 | 274 qpel_mc_func put_mspel_pixels_tab[8]; |
2967 | 275 |
1168 | 276 /** |
4751 | 277 * h264 Chroma MC |
1168 | 278 */ |
279 h264_chroma_mc_func put_h264_chroma_pixels_tab[3]; | |
3663 | 280 /* This is really one func used in VC-1 decoding */ |
281 h264_chroma_mc_func put_no_rnd_h264_chroma_pixels_tab[3]; | |
1168 | 282 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
|
283 |
3020
c75fb0747e74
use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents:
3010
diff
changeset
|
284 qpel_mc_func put_h264_qpel_pixels_tab[4][16]; |
c75fb0747e74
use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents:
3010
diff
changeset
|
285 qpel_mc_func avg_h264_qpel_pixels_tab[4][16]; |
2967 | 286 |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3746
diff
changeset
|
287 qpel_mc_func put_2tap_qpel_pixels_tab[4][16]; |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3746
diff
changeset
|
288 qpel_mc_func avg_2tap_qpel_pixels_tab[4][16]; |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3746
diff
changeset
|
289 |
2415 | 290 h264_weight_func weight_h264_pixels_tab[10]; |
291 h264_biweight_func biweight_h264_pixels_tab[10]; | |
2967 | 292 |
3395
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
293 /* AVS specific */ |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
294 qpel_mc_func put_cavs_qpel_pixels_tab[2][16]; |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
295 qpel_mc_func avg_cavs_qpel_pixels_tab[2][16]; |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
296 void (*cavs_filter_lv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2); |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
297 void (*cavs_filter_lh)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2); |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
298 void (*cavs_filter_cv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2); |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
299 void (*cavs_filter_ch)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2); |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
300 void (*cavs_idct8_add)(uint8_t *dst, DCTELEM *block, int stride); |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
301 |
1708 | 302 me_cmp_func pix_abs[2][4]; |
2967 | 303 |
866 | 304 /* huffyuv specific */ |
305 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w); | |
936 | 306 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w); |
1527 | 307 /** |
308 * subtract huffyuv's variant of median prediction | |
309 * note, this might read from src1[-1], src2[-1] | |
310 */ | |
311 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top); | |
1273 | 312 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w); |
2633 | 313 |
2707
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
314 void (*h264_v_loop_filter_luma)(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
315 void (*h264_h_loop_filter_luma)(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
316 void (*h264_v_loop_filter_chroma)(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
317 void (*h264_h_loop_filter_chroma)(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0); |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
318 void (*h264_v_loop_filter_chroma_intra)(uint8_t *pix, int stride, int alpha, int beta); |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
319 void (*h264_h_loop_filter_chroma_intra)(uint8_t *pix, int stride, int alpha, int beta); |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3574
diff
changeset
|
320 // h264_loop_filter_strength: simd only. the C version is inlined in h264.c |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3574
diff
changeset
|
321 void (*h264_loop_filter_strength)(int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2], |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3574
diff
changeset
|
322 int bidir, int edges, int step, int mask_mv0, int mask_mv1); |
2967 | 323 |
1644 | 324 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale); |
325 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale); | |
326 | |
2045 | 327 void (*h261_loop_filter)(uint8_t *src, int stride); |
2044
b6f2add2511e
h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents:
2024
diff
changeset
|
328 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
329 /* assume len is a multiple of 4, and arrays are 16-byte aligned */ |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3526
diff
changeset
|
330 void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize); |
3574 | 331 /* assume len is a multiple of 8, and arrays are 16-byte aligned */ |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
332 void (*vector_fmul)(float *dst, const float *src, int len); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
333 void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
334 /* assume len is a multiple of 8, and src arrays are 16-byte aligned */ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
335 void (*vector_fmul_add_add)(float *dst, const float *src0, const float *src1, const float *src2, int src3, int len, int step); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
336 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
337 /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767] |
3660
7e1ee254a3ee
Align the input buffer in ffplay, introduce a public macro for aligned declarations
lu_zero
parents:
3656
diff
changeset
|
338 * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */ |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
339 void (*float_to_int16)(int16_t *dst, const float *src, int len); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3526
diff
changeset
|
340 |
1092 | 341 /* (I)DCT */ |
342 void (*fdct)(DCTELEM *block/* align 16*/); | |
1567 | 343 void (*fdct248)(DCTELEM *block/* align 16*/); |
2967 | 344 |
1324
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
345 /* IDCT really*/ |
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
346 void (*idct)(DCTELEM *block/* align 16*/); |
2967 | 347 |
1102 | 348 /** |
1104 | 349 * block -> idct -> clip to unsigned 8 bit -> dest. |
1102 | 350 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...) |
1111 | 351 * @param line_size size in bytes of a horizotal line of dest |
1102 | 352 */ |
1092 | 353 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
2967 | 354 |
1102 | 355 /** |
356 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest. | |
1111 | 357 * @param line_size size in bytes of a horizotal line of dest |
1102 | 358 */ |
1092 | 359 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
2967 | 360 |
1102 | 361 /** |
1104 | 362 * idct input permutation. |
1241 | 363 * several optimized IDCTs need a permutated input (relative to the normal order of the reference |
364 * IDCT) | |
365 * this permutation must be performed before the idct_put/add, note, normally this can be merged | |
366 * with the zigzag/alternate scan<br> | |
1102 | 367 * an example to avoid confusion: |
368 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...) | |
369 * - (x -> referece dct -> reference idct -> x) | |
370 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x) | |
371 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...) | |
372 */ | |
1092 | 373 uint8_t idct_permutation[64]; |
374 int idct_permutation_type; | |
375 #define FF_NO_IDCT_PERM 1 | |
376 #define FF_LIBMPEG2_IDCT_PERM 2 | |
377 #define FF_SIMPLE_IDCT_PERM 3 | |
378 #define FF_TRANSPOSE_IDCT_PERM 4 | |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2693
diff
changeset
|
379 #define FF_PARTTRANS_IDCT_PERM 5 |
1092 | 380 |
1784 | 381 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale); |
382 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale); | |
383 #define BASIS_SHIFT 16 | |
384 #define RECON_SHIFT 6 | |
2967 | 385 |
4268 | 386 /* h264 functions */ |
2272
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
387 void (*h264_idct_add)(uint8_t *dst, DCTELEM *block, int stride); |
2755 | 388 void (*h264_idct8_add)(uint8_t *dst, DCTELEM *block, int stride); |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
389 void (*h264_idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
390 void (*h264_idct8_dc_add)(uint8_t *dst, DCTELEM *block, int stride); |
4279 | 391 void (*h264_dct)(DCTELEM block[4][4]); |
3198
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
392 |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
393 /* snow wavelet */ |
5587
3ae03eacbe9f
use 16bit IDWT (a SIMD implementation of it should be >2x faster then with
michael
parents:
5573
diff
changeset
|
394 void (*vertical_compose97i)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width); |
3ae03eacbe9f
use 16bit IDWT (a SIMD implementation of it should be >2x faster then with
michael
parents:
5573
diff
changeset
|
395 void (*horizontal_compose97i)(IDWTELEM *b, int width); |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
4311
diff
changeset
|
396 void (*inner_add_yblock)(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8); |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3198
diff
changeset
|
397 |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3198
diff
changeset
|
398 void (*prefetch)(void *mem, int stride, int h); |
3245 | 399 |
400 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
3526 | 401 |
402 /* vc1 functions */ | |
403 void (*vc1_inv_trans_8x8)(DCTELEM *b); | |
404 void (*vc1_inv_trans_8x4)(DCTELEM *b, int n); | |
405 void (*vc1_inv_trans_4x8)(DCTELEM *b, int n); | |
406 void (*vc1_inv_trans_4x4)(DCTELEM *b, int n); | |
4239 | 407 void (*vc1_v_overlap)(uint8_t* src, int stride); |
408 void (*vc1_h_overlap)(uint8_t* src, int stride); | |
3526 | 409 /* put 8x8 block with bicubic interpolation and quarterpel precision |
410 * last argument is actually round value instead of height | |
411 */ | |
412 op_pixels_func put_vc1_mspel_pixels_tab[16]; | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
413 } DSPContext; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
414 |
4197 | 415 void dsputil_static_init(void); |
1092 | 416 void dsputil_init(DSPContext* p, AVCodecContext *avctx); |
0 | 417 |
4281
de525a2b41db
ff_check_alignment to warn the user about a missaligned stack
michael
parents:
4279
diff
changeset
|
418 int ff_check_alignment(void); |
de525a2b41db
ff_check_alignment to warn the user about a missaligned stack
michael
parents:
4279
diff
changeset
|
419 |
764 | 420 /** |
421 * permute block according to permuatation. | |
422 * @param last last non zero element in scantable order | |
423 */ | |
1064 | 424 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last); |
34 | 425 |
1729 | 426 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type); |
427 | |
2979 | 428 #define BYTE_VEC32(c) ((c)*0x01010101UL) |
1264 | 429 |
430 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b) | |
431 { | |
432 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
433 } | |
434 | |
435 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b) | |
436 { | |
437 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
438 } | |
439 | |
2184 | 440 static inline int get_penalty_factor(int lambda, int lambda2, int type){ |
441 switch(type&0xFF){ | |
442 default: | |
443 case FF_CMP_SAD: | |
444 return lambda>>FF_LAMBDA_SHIFT; | |
445 case FF_CMP_DCT: | |
446 return (3*lambda)>>(FF_LAMBDA_SHIFT+1); | |
447 case FF_CMP_W53: | |
448 return (4*lambda)>>(FF_LAMBDA_SHIFT); | |
449 case FF_CMP_W97: | |
450 return (2*lambda)>>(FF_LAMBDA_SHIFT); | |
451 case FF_CMP_SATD: | |
3010
533c6386eca9
8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents:
2979
diff
changeset
|
452 case FF_CMP_DCT264: |
2184 | 453 return (2*lambda)>>FF_LAMBDA_SHIFT; |
454 case FF_CMP_RD: | |
455 case FF_CMP_PSNR: | |
456 case FF_CMP_SSE: | |
457 case FF_CMP_NSSE: | |
458 return lambda2>>FF_LAMBDA_SHIFT; | |
459 case FF_CMP_BIT: | |
460 return 1; | |
461 } | |
462 } | |
463 | |
1102 | 464 /** |
1104 | 465 * Empty mmx state. |
1102 | 466 * this must be called between any dsp function and float/double code. |
467 * for example sin(); dsp->idct_put(); emms_c(); cos() | |
468 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
469 #define emms_c() |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
470 |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
471 /* should be defined by architectures supporting |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
472 one or more MultiMedia extension */ |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
473 int mm_support(void); |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
474 |
5149 | 475 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx); |
476 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx); | |
477 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx); | |
478 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx); | |
479 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx); | |
480 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx); | |
481 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx); | |
482 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx); | |
483 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx); | |
484 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
485 #define DECLARE_ALIGNED_16(t, v) DECLARE_ALIGNED(16, t, v) |
1974
8c5489b2cf3e
move __align16 some place where non-MMX machines can see it
melanson
parents:
1972
diff
changeset
|
486 |
62 | 487 #if defined(HAVE_MMX) |
0 | 488 |
862 | 489 #undef emms_c |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
490 |
0 | 491 #define MM_MMX 0x0001 /* standard MMX */ |
492 #define MM_3DNOW 0x0004 /* AMD 3DNOW */ | |
493 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */ | |
494 #define MM_SSE 0x0008 /* SSE functions */ | |
495 #define MM_SSE2 0x0010 /* PIV SSE2 functions */ | |
2388 | 496 #define MM_3DNOWEXT 0x0020 /* AMD 3DNowExt */ |
3279
647a677c00a4
Remove unused and unsupported Cyrix's "Extended MMX",
gpoirier
parents:
3248
diff
changeset
|
497 #define MM_SSE3 0x0040 /* Prescott SSE3 functions */ |
4311
f9325f2bc762
Add SSSE3 (Core2 aka Conroe/Merom/Woodcrester new instructions) detection
gpoirier
parents:
4281
diff
changeset
|
498 #define MM_SSSE3 0x0080 /* Conroe SSSE3 functions */ |
0 | 499 |
4197 | 500 extern int mm_flags; |
501 | |
1064 | 502 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
503 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); | |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
504 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
0 | 505 |
506 static inline void emms(void) | |
507 { | |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
508 __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
|
509 } |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
510 |
936 | 511 |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
512 #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
|
513 {\ |
4197 | 514 if (mm_flags & MM_MMX)\ |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
515 emms();\ |
0 | 516 } |
517 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
518 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
3089 | 519 |
2324 | 520 #define STRIDE_ALIGN 8 |
0 | 521 |
1092 | 522 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx); |
1065 | 523 |
62 | 524 #elif defined(ARCH_ARMV4L) |
525 | |
526 /* This is to use 4 bytes read to the IDCT pointers for some 'zero' | |
1974
8c5489b2cf3e
move __align16 some place where non-MMX machines can see it
melanson
parents:
1972
diff
changeset
|
527 line optimizations */ |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
528 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(4, t, v) |
2324 | 529 #define STRIDE_ALIGN 4 |
62 | 530 |
2776
930e56f92c57
IWMMXT configure support + runtime selection patch by (Gildas Bazin, gbazin : altern org)
michael
parents:
2763
diff
changeset
|
531 #define MM_IWMMXT 0x0100 /* XScale IWMMXT */ |
930e56f92c57
IWMMXT configure support + runtime selection patch by (Gildas Bazin, gbazin : altern org)
michael
parents:
2763
diff
changeset
|
532 |
4197 | 533 extern int mm_flags; |
534 | |
88 | 535 #elif defined(HAVE_MLIB) |
536 | |
537 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */ | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
538 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 539 #define STRIDE_ALIGN 8 |
88 | 540 |
1959
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
541 #elif defined(ARCH_SPARC) |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
542 |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
543 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */ |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
544 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 545 #define STRIDE_ALIGN 8 |
1959
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
546 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
547 #elif defined(ARCH_ALPHA) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
548 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
549 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 550 #define STRIDE_ALIGN 8 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
551 |
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
|
552 #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
|
553 |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
554 #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
|
555 |
4197 | 556 extern int mm_flags; |
557 | |
5604
8691652d9dce
Do not wrongly mark Apple gcc workarounds as being specific to Darwin.
diego
parents:
5587
diff
changeset
|
558 #if defined(HAVE_ALTIVEC) && !defined(__APPLE_CC__) |
1653 | 559 #define pixel altivec_pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
560 #include <altivec.h> |
1653 | 561 #undef pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
562 #endif |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
563 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
564 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v) |
2324 | 565 #define STRIDE_ALIGN 16 |
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
|
566 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
567 #elif defined(HAVE_MMI) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
568 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
569 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v) |
2324 | 570 #define STRIDE_ALIGN 16 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
571 |
1259
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
572 #elif defined(ARCH_SH4) |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
573 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
574 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 575 #define STRIDE_ALIGN 8 |
1259
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
576 |
3728 | 577 #elif defined(ARCH_BFIN) |
578 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
579 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
3728 | 580 #define STRIDE_ALIGN 8 |
581 | |
0 | 582 #else |
583 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
584 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 585 #define STRIDE_ALIGN 8 |
0 | 586 |
587 #endif | |
588 | |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
589 /* PSNR */ |
1064 | 590 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
|
591 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
|
592 AVCodecContext *avctx); |
781 | 593 |
594 /* FFT computation */ | |
595 | |
596 /* NOTE: soon integer code will be added, so you must use the | |
597 FFTSample type */ | |
598 typedef float FFTSample; | |
599 | |
3555 | 600 struct MDCTContext; |
601 | |
781 | 602 typedef struct FFTComplex { |
603 FFTSample re, im; | |
604 } FFTComplex; | |
605 | |
606 typedef struct FFTContext { | |
607 int nbits; | |
608 int inverse; | |
609 uint16_t *revtab; | |
610 FFTComplex *exptab; | |
611 FFTComplex *exptab1; /* only used by SSE code */ | |
612 void (*fft_calc)(struct FFTContext *s, FFTComplex *z); | |
3555 | 613 void (*imdct_calc)(struct MDCTContext *s, FFTSample *output, |
614 const FFTSample *input, FFTSample *tmp); | |
781 | 615 } FFTContext; |
616 | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
617 int ff_fft_init(FFTContext *s, int nbits, int inverse); |
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
618 void ff_fft_permute(FFTContext *s, FFTComplex *z); |
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
619 void ff_fft_calc_c(FFTContext *s, FFTComplex *z); |
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
620 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z); |
3175 | 621 void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z); |
622 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z); | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
623 void ff_fft_calc_altivec(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
|
624 |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
625 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z) |
781 | 626 { |
627 s->fft_calc(s, z); | |
628 } | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
629 void ff_fft_end(FFTContext *s); |
781 | 630 |
631 /* MDCT computation */ | |
632 | |
633 typedef struct MDCTContext { | |
634 int n; /* size of MDCT (i.e. number of input data * 2) */ | |
635 int nbits; /* n = 2^nbits */ | |
636 /* pre/post rotation tables */ | |
637 FFTSample *tcos; | |
638 FFTSample *tsin; | |
639 FFTContext fft; | |
640 } MDCTContext; | |
641 | |
794 | 642 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
|
643 void ff_imdct_calc(MDCTContext *s, FFTSample *output, |
781 | 644 const FFTSample *input, FFTSample *tmp); |
3555 | 645 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output, |
646 const FFTSample *input, FFTSample *tmp); | |
3746 | 647 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, |
648 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
|
649 void ff_mdct_calc(MDCTContext *s, FFTSample *out, |
781 | 650 const FFTSample *input, FFTSample *tmp); |
794 | 651 void ff_mdct_end(MDCTContext *s); |
781 | 652 |
1708 | 653 #define WARPER8_16(name8, name16)\ |
654 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
655 return name8(s, dst , src , stride, h)\ | |
656 +name8(s, dst+8 , src+8 , stride, h);\ | |
657 } | |
658 | |
659 #define WARPER8_16_SQ(name8, name16)\ | |
660 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
661 int score=0;\ | |
662 score +=name8(s, dst , src , stride, 8);\ | |
663 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
664 if(h==16){\ | |
665 dst += 8*stride;\ | |
666 src += 8*stride;\ | |
667 score +=name8(s, dst , src , stride, 8);\ | |
668 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
669 }\ | |
670 return score;\ | |
936 | 671 } |
672 | |
4240 | 673 |
674 static inline void copy_block2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
675 { | |
676 int i; | |
677 for(i=0; i<h; i++) | |
678 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
679 AV_WN16(dst , AV_RN16(src )); |
4240 | 680 dst+=dstStride; |
681 src+=srcStride; | |
682 } | |
683 } | |
684 | |
685 static inline void copy_block4(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
686 { | |
687 int i; | |
688 for(i=0; i<h; i++) | |
689 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
690 AV_WN32(dst , AV_RN32(src )); |
4240 | 691 dst+=dstStride; |
692 src+=srcStride; | |
693 } | |
694 } | |
695 | |
696 static inline void copy_block8(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
697 { | |
698 int i; | |
699 for(i=0; i<h; i++) | |
700 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
701 AV_WN32(dst , AV_RN32(src )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
702 AV_WN32(dst+4 , AV_RN32(src+4 )); |
4240 | 703 dst+=dstStride; |
704 src+=srcStride; | |
705 } | |
706 } | |
707 | |
708 static inline void copy_block9(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
709 { | |
710 int i; | |
711 for(i=0; i<h; i++) | |
712 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
713 AV_WN32(dst , AV_RN32(src )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
714 AV_WN32(dst+4 , AV_RN32(src+4 )); |
4240 | 715 dst[8]= src[8]; |
716 dst+=dstStride; | |
717 src+=srcStride; | |
718 } | |
719 } | |
720 | |
721 static inline void copy_block16(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
722 { | |
723 int i; | |
724 for(i=0; i<h; i++) | |
725 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
726 AV_WN32(dst , AV_RN32(src )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
727 AV_WN32(dst+4 , AV_RN32(src+4 )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
728 AV_WN32(dst+8 , AV_RN32(src+8 )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
729 AV_WN32(dst+12, AV_RN32(src+12)); |
4240 | 730 dst+=dstStride; |
731 src+=srcStride; | |
732 } | |
733 } | |
734 | |
735 static inline void copy_block17(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
736 { | |
737 int i; | |
738 for(i=0; i<h; i++) | |
739 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
740 AV_WN32(dst , AV_RN32(src )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
741 AV_WN32(dst+4 , AV_RN32(src+4 )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
742 AV_WN32(dst+8 , AV_RN32(src+8 )); |
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
743 AV_WN32(dst+12, AV_RN32(src+12)); |
4240 | 744 dst[16]= src[16]; |
745 dst+=dstStride; | |
746 src+=srcStride; | |
747 } | |
748 } | |
749 | |
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
|
750 #endif |