Mercurial > libavcodec.hg
annotate dsputil.h @ 5560:2e63bd8e894d libavcodec
Simplify and speedup code, reduce needed headroom by 2 bits in the 3rd
vertical lifting step of the SSE2 code (untested)
author | michael |
---|---|
date | Mon, 20 Aug 2007 23:59:20 +0000 |
parents | c16a59ef6a86 |
children | 204273dd0349 |
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; |
0 | 40 |
474
11dbd00682fc
avoid name clash with libjpeg - added missing externs
bellard
parents:
429
diff
changeset
|
41 void fdct_ifast (DCTELEM *data); |
1571
aa4dc16c0f18
* adding integer/floating point AAN implementations for DCT 2-4-8
romansh
parents:
1567
diff
changeset
|
42 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
|
43 void ff_jpeg_fdct_islow (DCTELEM *data); |
1567 | 44 void ff_fdct248_islow (DCTELEM *data); |
0 | 45 |
46 void j_rev_dct (DCTELEM *data); | |
2256 | 47 void j_rev_dct4 (DCTELEM *data); |
2257 | 48 void j_rev_dct2 (DCTELEM *data); |
2259 | 49 void j_rev_dct1 (DCTELEM *data); |
0 | 50 |
687
9abb13c21fbe
fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents:
675
diff
changeset
|
51 void ff_fdct_mmx(DCTELEM *block); |
1565 | 52 void ff_fdct_mmx2(DCTELEM *block); |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
53 void ff_fdct_sse2(DCTELEM *block); |
0 | 54 |
2755 | 55 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
|
56 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
|
57 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
|
58 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
|
59 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
|
60 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
|
61 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
62 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
|
63 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
|
64 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
|
65 |
34 | 66 /* encoding scans */ |
1064 | 67 extern const uint8_t ff_alternate_horizontal_scan[64]; |
68 extern const uint8_t ff_alternate_vertical_scan[64]; | |
69 extern const uint8_t ff_zigzag_direct[64]; | |
1567 | 70 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
|
71 |
0 | 72 /* pixel operations */ |
2090 | 73 #define MAX_NEG_CROP 1024 |
0 | 74 |
75 /* temporary */ | |
4179 | 76 extern uint32_t ff_squareTbl[512]; |
4176 | 77 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP]; |
0 | 78 |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
1864
diff
changeset
|
79 /* VP3 DSP functions */ |
2693 | 80 void ff_vp3_idct_c(DCTELEM *block/* align 16*/); |
81 void ff_vp3_idct_put_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); | |
82 void ff_vp3_idct_add_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); | |
0 | 83 |
3245 | 84 /* 1/2^n downscaling functions from imgconvert.c */ |
85 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
86 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
87 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
88 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
89 | |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3245
diff
changeset
|
90 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
|
91 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
|
92 |
675 | 93 /* minimum alignment rules ;) |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
94 if u notice errors in the align stuff, need more alignment for some asm code for some cpu |
675 | 95 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ... |
96 | |
97 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible) | |
98 i (michael) didnt check them, these are just the alignents which i think could be reached easily ... | |
99 | |
100 !future video codecs might need functions with less strict alignment | |
101 */ | |
0 | 102 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
103 /* |
1064 | 104 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size); |
105 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride); | |
106 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
107 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size); | |
296 | 108 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
|
109 */ |
0 | 110 |
111 /* add and put pixel (decoding) */ | |
675 | 112 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16 |
1709 | 113 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4 |
1064 | 114 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
|
115 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 | 116 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride); |
1168 | 117 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 | 118 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset); |
3029 | 119 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset); |
0 | 120 |
984 | 121 #define DEF_OLD_QPEL(name)\ |
1064 | 122 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ |
123 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\ | |
124 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
|
125 |
984 | 126 DEF_OLD_QPEL(qpel16_mc11_old_c) |
127 DEF_OLD_QPEL(qpel16_mc31_old_c) | |
128 DEF_OLD_QPEL(qpel16_mc12_old_c) | |
129 DEF_OLD_QPEL(qpel16_mc32_old_c) | |
130 DEF_OLD_QPEL(qpel16_mc13_old_c) | |
131 DEF_OLD_QPEL(qpel16_mc33_old_c) | |
132 DEF_OLD_QPEL(qpel8_mc11_old_c) | |
133 DEF_OLD_QPEL(qpel8_mc31_old_c) | |
134 DEF_OLD_QPEL(qpel8_mc12_old_c) | |
135 DEF_OLD_QPEL(qpel8_mc32_old_c) | |
136 DEF_OLD_QPEL(qpel8_mc13_old_c) | |
137 DEF_OLD_QPEL(qpel8_mc33_old_c) | |
651 | 138 |
139 #define CALL_2X_PIXELS(a, b, n)\ | |
140 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\ | |
141 b(block , pixels , line_size, h);\ | |
142 b(block+n, pixels+n, line_size, h);\ | |
143 } | |
255 | 144 |
0 | 145 /* motion estimation */ |
1709 | 146 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2 |
5127 | 147 // although currently h<4 is not used as functions with width <8 are neither used nor implemented |
1708 | 148 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 | 149 |
1168 | 150 |
3198
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
151 // for snow slices |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
152 typedef struct slice_buffer_s slice_buffer; |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
153 |
1102 | 154 /** |
155 * DSPContext. | |
156 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
157 typedef struct DSPContext { |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
158 /* pixel ops : interface with DCT */ |
1064 | 159 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size); |
160 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride); | |
161 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
|
162 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); |
1064 | 163 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size); |
2763 | 164 void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size); |
165 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
|
166 int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/); |
1136 | 167 /** |
168 * translational global motion compensation. | |
169 */ | |
1064 | 170 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder); |
1136 | 171 /** |
172 * global motion compensation. | |
173 */ | |
1064 | 174 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy, |
2979 | 175 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
|
176 void (*clear_blocks)(DCTELEM *blocks/*align 16*/); |
1064 | 177 int (*pix_sum)(uint8_t * pix, int line_size); |
178 int (*pix_norm1)(uint8_t * pix, int line_size); | |
1708 | 179 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4 |
2967 | 180 |
1729 | 181 me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */ |
182 me_cmp_func sse[5]; | |
183 me_cmp_func hadamard8_diff[5]; | |
184 me_cmp_func dct_sad[5]; | |
185 me_cmp_func quant_psnr[5]; | |
186 me_cmp_func bit[5]; | |
187 me_cmp_func rd[5]; | |
188 me_cmp_func vsad[5]; | |
189 me_cmp_func vsse[5]; | |
2065
9e4bebc39ade
noise preserving sum of squares comparission function
michael
parents:
2045
diff
changeset
|
190 me_cmp_func nsse[5]; |
2184 | 191 me_cmp_func w53[5]; |
192 me_cmp_func w97[5]; | |
2382 | 193 me_cmp_func dct_max[5]; |
3010
533c6386eca9
8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents:
2979
diff
changeset
|
194 me_cmp_func dct264_sad[5]; |
936 | 195 |
1708 | 196 me_cmp_func me_pre_cmp[5]; |
197 me_cmp_func me_cmp[5]; | |
198 me_cmp_func me_sub_cmp[5]; | |
199 me_cmp_func mb_cmp[5]; | |
1729 | 200 me_cmp_func ildct_cmp[5]; //only width 16 used |
2382 | 201 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
|
202 |
5255 | 203 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2, |
204 int size); | |
4749 | 205 |
1136 | 206 /** |
207 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
4751 | 208 * this is an array[4][4] of motion compensation functions for 4 |
1713 | 209 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 210 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 211 * @param block destination where the result is stored |
212 * @param pixels source | |
213 * @param line_size number of bytes in a horizontal line of block | |
214 * @param h height | |
215 */ | |
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
|
216 op_pixels_func put_pixels_tab[4][4]; |
1136 | 217 |
218 /** | |
219 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
2967 | 220 * This is an array[4][4] of motion compensation functions for 4 |
1713 | 221 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 222 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 223 * @param block destination into which the result is averaged (a+b+1)>>1 |
224 * @param pixels source | |
225 * @param line_size number of bytes in a horizontal line of block | |
226 * @param h height | |
227 */ | |
1319 | 228 op_pixels_func avg_pixels_tab[4][4]; |
1136 | 229 |
230 /** | |
231 * Halfpel motion compensation with no rounding (a+b)>>1. | |
4751 | 232 * this is an array[2][4] of motion compensation functions for 2 |
1225 | 233 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 234 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 235 * @param block destination where the result is stored |
236 * @param pixels source | |
237 * @param line_size number of bytes in a horizontal line of block | |
238 * @param h height | |
239 */ | |
2075 | 240 op_pixels_func put_no_rnd_pixels_tab[4][4]; |
1136 | 241 |
242 /** | |
243 * Halfpel motion compensation with no rounding (a+b)>>1. | |
4751 | 244 * this is an array[2][4] of motion compensation functions for 2 |
1225 | 245 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 246 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 247 * @param block destination into which the result is averaged (a+b)>>1 |
248 * @param pixels source | |
249 * @param line_size number of bytes in a horizontal line of block | |
250 * @param h height | |
251 */ | |
2075 | 252 op_pixels_func avg_no_rnd_pixels_tab[4][4]; |
2967 | 253 |
1864 | 254 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 | 255 |
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
|
256 /** |
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 * Thirdpel motion compensation with rounding (a+b+1)>>1. |
4751 | 258 * this is an array[12] of motion compensation functions for the 9 thirdpe |
259 * 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
|
260 * *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
|
261 * @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
|
262 * @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
|
263 * @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
|
264 * @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
|
265 */ |
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 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
1319 | 267 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
268 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
269 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
|
270 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
|
271 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
|
272 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16]; |
936 | 273 qpel_mc_func put_mspel_pixels_tab[8]; |
2967 | 274 |
1168 | 275 /** |
4751 | 276 * h264 Chroma MC |
1168 | 277 */ |
278 h264_chroma_mc_func put_h264_chroma_pixels_tab[3]; | |
3663 | 279 /* This is really one func used in VC-1 decoding */ |
280 h264_chroma_mc_func put_no_rnd_h264_chroma_pixels_tab[3]; | |
1168 | 281 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
|
282 |
3020
c75fb0747e74
use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents:
3010
diff
changeset
|
283 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
|
284 qpel_mc_func avg_h264_qpel_pixels_tab[4][16]; |
2967 | 285 |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3746
diff
changeset
|
286 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
|
287 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
|
288 |
2415 | 289 h264_weight_func weight_h264_pixels_tab[10]; |
290 h264_biweight_func biweight_h264_pixels_tab[10]; | |
2967 | 291 |
3395
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
292 /* AVS specific */ |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
293 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
|
294 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
|
295 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
|
296 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
|
297 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
|
298 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
|
299 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
|
300 |
1708 | 301 me_cmp_func pix_abs[2][4]; |
2967 | 302 |
866 | 303 /* huffyuv specific */ |
304 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w); | |
936 | 305 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w); |
1527 | 306 /** |
307 * subtract huffyuv's variant of median prediction | |
308 * note, this might read from src1[-1], src2[-1] | |
309 */ | |
310 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top); | |
1273 | 311 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w); |
2633 | 312 |
2707
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
313 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
|
314 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
|
315 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
|
316 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
|
317 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
|
318 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
|
319 // 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
|
320 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
|
321 int bidir, int edges, int step, int mask_mv0, int mask_mv1); |
2967 | 322 |
1644 | 323 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale); |
324 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale); | |
325 | |
2045 | 326 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
|
327 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
328 /* 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
|
329 void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize); |
3574 | 330 /* 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
|
331 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
|
332 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
|
333 /* 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
|
334 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
|
335 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
336 /* 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
|
337 * 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
|
338 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
|
339 |
1092 | 340 /* (I)DCT */ |
341 void (*fdct)(DCTELEM *block/* align 16*/); | |
1567 | 342 void (*fdct248)(DCTELEM *block/* align 16*/); |
2967 | 343 |
1324
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
344 /* IDCT really*/ |
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
345 void (*idct)(DCTELEM *block/* align 16*/); |
2967 | 346 |
1102 | 347 /** |
1104 | 348 * block -> idct -> clip to unsigned 8 bit -> dest. |
1102 | 349 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...) |
1111 | 350 * @param line_size size in bytes of a horizotal line of dest |
1102 | 351 */ |
1092 | 352 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
2967 | 353 |
1102 | 354 /** |
355 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest. | |
1111 | 356 * @param line_size size in bytes of a horizotal line of dest |
1102 | 357 */ |
1092 | 358 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
2967 | 359 |
1102 | 360 /** |
1104 | 361 * idct input permutation. |
1241 | 362 * several optimized IDCTs need a permutated input (relative to the normal order of the reference |
363 * IDCT) | |
364 * this permutation must be performed before the idct_put/add, note, normally this can be merged | |
365 * with the zigzag/alternate scan<br> | |
1102 | 366 * an example to avoid confusion: |
367 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...) | |
368 * - (x -> referece dct -> reference idct -> x) | |
369 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x) | |
370 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...) | |
371 */ | |
1092 | 372 uint8_t idct_permutation[64]; |
373 int idct_permutation_type; | |
374 #define FF_NO_IDCT_PERM 1 | |
375 #define FF_LIBMPEG2_IDCT_PERM 2 | |
376 #define FF_SIMPLE_IDCT_PERM 3 | |
377 #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
|
378 #define FF_PARTTRANS_IDCT_PERM 5 |
1092 | 379 |
1784 | 380 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale); |
381 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale); | |
382 #define BASIS_SHIFT 16 | |
383 #define RECON_SHIFT 6 | |
2967 | 384 |
4268 | 385 /* h264 functions */ |
2272
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
386 void (*h264_idct_add)(uint8_t *dst, DCTELEM *block, int stride); |
2755 | 387 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
|
388 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
|
389 void (*h264_idct8_dc_add)(uint8_t *dst, DCTELEM *block, int stride); |
4279 | 390 void (*h264_dct)(DCTELEM block[4][4]); |
3198
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
391 |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
392 /* snow wavelet */ |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
393 void (*vertical_compose97i)(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
394 void (*horizontal_compose97i)(DWTELEM *b, int width); |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
4311
diff
changeset
|
395 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
|
396 |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3198
diff
changeset
|
397 void (*prefetch)(void *mem, int stride, int h); |
3245 | 398 |
399 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
3526 | 400 |
401 /* vc1 functions */ | |
402 void (*vc1_inv_trans_8x8)(DCTELEM *b); | |
403 void (*vc1_inv_trans_8x4)(DCTELEM *b, int n); | |
404 void (*vc1_inv_trans_4x8)(DCTELEM *b, int n); | |
405 void (*vc1_inv_trans_4x4)(DCTELEM *b, int n); | |
4239 | 406 void (*vc1_v_overlap)(uint8_t* src, int stride); |
407 void (*vc1_h_overlap)(uint8_t* src, int stride); | |
3526 | 408 /* put 8x8 block with bicubic interpolation and quarterpel precision |
409 * last argument is actually round value instead of height | |
410 */ | |
411 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
|
412 } DSPContext; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
413 |
4197 | 414 void dsputil_static_init(void); |
1092 | 415 void dsputil_init(DSPContext* p, AVCodecContext *avctx); |
0 | 416 |
4281
de525a2b41db
ff_check_alignment to warn the user about a missaligned stack
michael
parents:
4279
diff
changeset
|
417 int ff_check_alignment(void); |
de525a2b41db
ff_check_alignment to warn the user about a missaligned stack
michael
parents:
4279
diff
changeset
|
418 |
764 | 419 /** |
420 * permute block according to permuatation. | |
421 * @param last last non zero element in scantable order | |
422 */ | |
1064 | 423 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last); |
34 | 424 |
1729 | 425 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type); |
426 | |
2979 | 427 #define BYTE_VEC32(c) ((c)*0x01010101UL) |
1264 | 428 |
429 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b) | |
430 { | |
431 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
432 } | |
433 | |
434 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b) | |
435 { | |
436 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
437 } | |
438 | |
2184 | 439 static inline int get_penalty_factor(int lambda, int lambda2, int type){ |
440 switch(type&0xFF){ | |
441 default: | |
442 case FF_CMP_SAD: | |
443 return lambda>>FF_LAMBDA_SHIFT; | |
444 case FF_CMP_DCT: | |
445 return (3*lambda)>>(FF_LAMBDA_SHIFT+1); | |
446 case FF_CMP_W53: | |
447 return (4*lambda)>>(FF_LAMBDA_SHIFT); | |
448 case FF_CMP_W97: | |
449 return (2*lambda)>>(FF_LAMBDA_SHIFT); | |
450 case FF_CMP_SATD: | |
3010
533c6386eca9
8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents:
2979
diff
changeset
|
451 case FF_CMP_DCT264: |
2184 | 452 return (2*lambda)>>FF_LAMBDA_SHIFT; |
453 case FF_CMP_RD: | |
454 case FF_CMP_PSNR: | |
455 case FF_CMP_SSE: | |
456 case FF_CMP_NSSE: | |
457 return lambda2>>FF_LAMBDA_SHIFT; | |
458 case FF_CMP_BIT: | |
459 return 1; | |
460 } | |
461 } | |
462 | |
1102 | 463 /** |
1104 | 464 * Empty mmx state. |
1102 | 465 * this must be called between any dsp function and float/double code. |
466 * for example sin(); dsp->idct_put(); emms_c(); cos() | |
467 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
468 #define emms_c() |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
469 |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
470 /* should be defined by architectures supporting |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
471 one or more MultiMedia extension */ |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
472 int mm_support(void); |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
473 |
5149 | 474 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx); |
475 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx); | |
476 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx); | |
477 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx); | |
478 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx); | |
479 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx); | |
480 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx); | |
481 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx); | |
482 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx); | |
483 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
484 #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
|
485 |
62 | 486 #if defined(HAVE_MMX) |
0 | 487 |
862 | 488 #undef emms_c |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
489 |
0 | 490 #define MM_MMX 0x0001 /* standard MMX */ |
491 #define MM_3DNOW 0x0004 /* AMD 3DNOW */ | |
492 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */ | |
493 #define MM_SSE 0x0008 /* SSE functions */ | |
494 #define MM_SSE2 0x0010 /* PIV SSE2 functions */ | |
2388 | 495 #define MM_3DNOWEXT 0x0020 /* AMD 3DNowExt */ |
3279
647a677c00a4
Remove unused and unsupported Cyrix's "Extended MMX",
gpoirier
parents:
3248
diff
changeset
|
496 #define MM_SSE3 0x0040 /* Prescott SSE3 functions */ |
4311
f9325f2bc762
Add SSSE3 (Core2 aka Conroe/Merom/Woodcrester new instructions) detection
gpoirier
parents:
4281
diff
changeset
|
497 #define MM_SSSE3 0x0080 /* Conroe SSSE3 functions */ |
0 | 498 |
4197 | 499 extern int mm_flags; |
500 | |
1064 | 501 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
502 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
|
503 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
0 | 504 |
505 static inline void emms(void) | |
506 { | |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
507 __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
|
508 } |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
509 |
936 | 510 |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
511 #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
|
512 {\ |
4197 | 513 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
|
514 emms();\ |
0 | 515 } |
516 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
517 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
3089 | 518 |
2324 | 519 #define STRIDE_ALIGN 8 |
0 | 520 |
1092 | 521 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx); |
1065 | 522 |
62 | 523 #elif defined(ARCH_ARMV4L) |
524 | |
525 /* 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
|
526 line optimizations */ |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
527 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(4, t, v) |
2324 | 528 #define STRIDE_ALIGN 4 |
62 | 529 |
2776
930e56f92c57
IWMMXT configure support + runtime selection patch by (Gildas Bazin, gbazin : altern org)
michael
parents:
2763
diff
changeset
|
530 #define MM_IWMMXT 0x0100 /* XScale IWMMXT */ |
930e56f92c57
IWMMXT configure support + runtime selection patch by (Gildas Bazin, gbazin : altern org)
michael
parents:
2763
diff
changeset
|
531 |
4197 | 532 extern int mm_flags; |
533 | |
88 | 534 #elif defined(HAVE_MLIB) |
535 | |
536 /* 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
|
537 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 538 #define STRIDE_ALIGN 8 |
88 | 539 |
1959
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
540 #elif defined(ARCH_SPARC) |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
541 |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
542 /* 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
|
543 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 544 #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
|
545 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
546 #elif defined(ARCH_ALPHA) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
547 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
548 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 549 #define STRIDE_ALIGN 8 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
550 |
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
|
551 #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
|
552 |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
553 #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
|
554 |
4197 | 555 extern int mm_flags; |
556 | |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
557 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN) |
1653 | 558 #define pixel altivec_pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
559 #include <altivec.h> |
1653 | 560 #undef pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
561 #endif |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
562 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
563 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v) |
2324 | 564 #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
|
565 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
566 #elif defined(HAVE_MMI) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
567 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
568 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v) |
2324 | 569 #define STRIDE_ALIGN 16 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
570 |
1259
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
571 #elif defined(ARCH_SH4) |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
572 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
573 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 574 #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
|
575 |
3728 | 576 #elif defined(ARCH_BFIN) |
577 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
578 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
3728 | 579 #define STRIDE_ALIGN 8 |
580 | |
0 | 581 #else |
582 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
583 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 584 #define STRIDE_ALIGN 8 |
0 | 585 |
586 #endif | |
587 | |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
588 /* PSNR */ |
1064 | 589 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
|
590 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
|
591 AVCodecContext *avctx); |
781 | 592 |
593 /* FFT computation */ | |
594 | |
595 /* NOTE: soon integer code will be added, so you must use the | |
596 FFTSample type */ | |
597 typedef float FFTSample; | |
598 | |
3555 | 599 struct MDCTContext; |
600 | |
781 | 601 typedef struct FFTComplex { |
602 FFTSample re, im; | |
603 } FFTComplex; | |
604 | |
605 typedef struct FFTContext { | |
606 int nbits; | |
607 int inverse; | |
608 uint16_t *revtab; | |
609 FFTComplex *exptab; | |
610 FFTComplex *exptab1; /* only used by SSE code */ | |
611 void (*fft_calc)(struct FFTContext *s, FFTComplex *z); | |
3555 | 612 void (*imdct_calc)(struct MDCTContext *s, FFTSample *output, |
613 const FFTSample *input, FFTSample *tmp); | |
781 | 614 } FFTContext; |
615 | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
616 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
|
617 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
|
618 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
|
619 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z); |
3175 | 620 void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z); |
621 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
|
622 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
|
623 |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
624 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z) |
781 | 625 { |
626 s->fft_calc(s, z); | |
627 } | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
628 void ff_fft_end(FFTContext *s); |
781 | 629 |
630 /* MDCT computation */ | |
631 | |
632 typedef struct MDCTContext { | |
633 int n; /* size of MDCT (i.e. number of input data * 2) */ | |
634 int nbits; /* n = 2^nbits */ | |
635 /* pre/post rotation tables */ | |
636 FFTSample *tcos; | |
637 FFTSample *tsin; | |
638 FFTContext fft; | |
639 } MDCTContext; | |
640 | |
794 | 641 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
|
642 void ff_imdct_calc(MDCTContext *s, FFTSample *output, |
781 | 643 const FFTSample *input, FFTSample *tmp); |
3555 | 644 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output, |
645 const FFTSample *input, FFTSample *tmp); | |
3746 | 646 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, |
647 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
|
648 void ff_mdct_calc(MDCTContext *s, FFTSample *out, |
781 | 649 const FFTSample *input, FFTSample *tmp); |
794 | 650 void ff_mdct_end(MDCTContext *s); |
781 | 651 |
1708 | 652 #define WARPER8_16(name8, name16)\ |
653 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
654 return name8(s, dst , src , stride, h)\ | |
655 +name8(s, dst+8 , src+8 , stride, h);\ | |
656 } | |
657 | |
658 #define WARPER8_16_SQ(name8, name16)\ | |
659 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
660 int score=0;\ | |
661 score +=name8(s, dst , src , stride, 8);\ | |
662 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
663 if(h==16){\ | |
664 dst += 8*stride;\ | |
665 src += 8*stride;\ | |
666 score +=name8(s, dst , src , stride, 8);\ | |
667 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
668 }\ | |
669 return score;\ | |
936 | 670 } |
671 | |
4240 | 672 |
673 static inline void copy_block2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
674 { | |
675 int i; | |
676 for(i=0; i<h; i++) | |
677 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
678 AV_WN16(dst , AV_RN16(src )); |
4240 | 679 dst+=dstStride; |
680 src+=srcStride; | |
681 } | |
682 } | |
683 | |
684 static inline void copy_block4(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
685 { | |
686 int i; | |
687 for(i=0; i<h; i++) | |
688 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
689 AV_WN32(dst , AV_RN32(src )); |
4240 | 690 dst+=dstStride; |
691 src+=srcStride; | |
692 } | |
693 } | |
694 | |
695 static inline void copy_block8(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
696 { | |
697 int i; | |
698 for(i=0; i<h; i++) | |
699 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
700 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
|
701 AV_WN32(dst+4 , AV_RN32(src+4 )); |
4240 | 702 dst+=dstStride; |
703 src+=srcStride; | |
704 } | |
705 } | |
706 | |
707 static inline void copy_block9(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
708 { | |
709 int i; | |
710 for(i=0; i<h; i++) | |
711 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
712 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
|
713 AV_WN32(dst+4 , AV_RN32(src+4 )); |
4240 | 714 dst[8]= src[8]; |
715 dst+=dstStride; | |
716 src+=srcStride; | |
717 } | |
718 } | |
719 | |
720 static inline void copy_block16(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
721 { | |
722 int i; | |
723 for(i=0; i<h; i++) | |
724 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
725 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
|
726 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
|
727 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
|
728 AV_WN32(dst+12, AV_RN32(src+12)); |
4240 | 729 dst+=dstStride; |
730 src+=srcStride; | |
731 } | |
732 } | |
733 | |
734 static inline void copy_block17(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
735 { | |
736 int i; | |
737 for(i=0; i<h; i++) | |
738 { | |
5520
c16a59ef6a86
* renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents:
5291
diff
changeset
|
739 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
|
740 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
|
741 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
|
742 AV_WN32(dst+12, AV_RN32(src+12)); |
4240 | 743 dst[16]= src[16]; |
744 dst+=dstStride; | |
745 src+=srcStride; | |
746 } | |
747 } | |
748 | |
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
|
749 #endif |