Mercurial > libavcodec.hg
annotate dsputil.h @ 5131:757efe20a3c8 libavcodec
Blackfin dct_quantize_bfin routine
2x performance boost in performing quantization for mpeg encoding
author | mhoffman |
---|---|
date | Tue, 12 Jun 2007 19:13:50 +0000 |
parents | 4dbe6578f811 |
children | b7bae4ce8b07 |
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 |
4749 | 203 int (*ssd_int8_vs_int16)(int8_t *pix1, int16_t *pix2, int size); |
204 | |
1136 | 205 /** |
206 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
4751 | 207 * this is an array[4][4] of motion compensation functions for 4 |
1713 | 208 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 209 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 210 * @param block destination where the result is stored |
211 * @param pixels source | |
212 * @param line_size number of bytes in a horizontal line of block | |
213 * @param h height | |
214 */ | |
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
|
215 op_pixels_func put_pixels_tab[4][4]; |
1136 | 216 |
217 /** | |
218 * Halfpel motion compensation with rounding (a+b+1)>>1. | |
2967 | 219 * This is an array[4][4] of motion compensation functions for 4 |
1713 | 220 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 221 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 222 * @param block destination into which the result is averaged (a+b+1)>>1 |
223 * @param pixels source | |
224 * @param line_size number of bytes in a horizontal line of block | |
225 * @param h height | |
226 */ | |
1319 | 227 op_pixels_func avg_pixels_tab[4][4]; |
1136 | 228 |
229 /** | |
230 * Halfpel motion compensation with no rounding (a+b)>>1. | |
4751 | 231 * this is an array[2][4] of motion compensation functions for 2 |
1225 | 232 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 233 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 234 * @param block destination where the result is stored |
235 * @param pixels source | |
236 * @param line_size number of bytes in a horizontal line of block | |
237 * @param h height | |
238 */ | |
2075 | 239 op_pixels_func put_no_rnd_pixels_tab[4][4]; |
1136 | 240 |
241 /** | |
242 * Halfpel motion compensation with no rounding (a+b)>>1. | |
4751 | 243 * this is an array[2][4] of motion compensation functions for 2 |
1225 | 244 * horizontal blocksizes (8,16) and the 4 halfpel positions<br> |
1213 | 245 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ] |
1136 | 246 * @param block destination into which the result is averaged (a+b)>>1 |
247 * @param pixels source | |
248 * @param line_size number of bytes in a horizontal line of block | |
249 * @param h height | |
250 */ | |
2075 | 251 op_pixels_func avg_no_rnd_pixels_tab[4][4]; |
2967 | 252 |
1864 | 253 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 | 254 |
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
|
255 /** |
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 * Thirdpel motion compensation with rounding (a+b+1)>>1. |
4751 | 257 * this is an array[12] of motion compensation functions for the 9 thirdpe |
258 * 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
|
259 * *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
|
260 * @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
|
261 * @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
|
262 * @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
|
263 * @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
|
264 */ |
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 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
1319 | 266 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width? |
267 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
268 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
|
269 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
|
270 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
|
271 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16]; |
936 | 272 qpel_mc_func put_mspel_pixels_tab[8]; |
2967 | 273 |
1168 | 274 /** |
4751 | 275 * h264 Chroma MC |
1168 | 276 */ |
277 h264_chroma_mc_func put_h264_chroma_pixels_tab[3]; | |
3663 | 278 /* This is really one func used in VC-1 decoding */ |
279 h264_chroma_mc_func put_no_rnd_h264_chroma_pixels_tab[3]; | |
1168 | 280 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
|
281 |
3020
c75fb0747e74
use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents:
3010
diff
changeset
|
282 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
|
283 qpel_mc_func avg_h264_qpel_pixels_tab[4][16]; |
2967 | 284 |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3746
diff
changeset
|
285 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
|
286 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
|
287 |
2415 | 288 h264_weight_func weight_h264_pixels_tab[10]; |
289 h264_biweight_func biweight_h264_pixels_tab[10]; | |
2967 | 290 |
3395
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
291 /* AVS specific */ |
adccbf4a1040
CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents:
3279
diff
changeset
|
292 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
|
293 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
|
294 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
|
295 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
|
296 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
|
297 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
|
298 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
|
299 |
1708 | 300 me_cmp_func pix_abs[2][4]; |
2967 | 301 |
866 | 302 /* huffyuv specific */ |
303 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w); | |
936 | 304 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w); |
1527 | 305 /** |
306 * subtract huffyuv's variant of median prediction | |
307 * note, this might read from src1[-1], src2[-1] | |
308 */ | |
309 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top); | |
1273 | 310 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w); |
2633 | 311 |
2707
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
312 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
|
313 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
|
314 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
|
315 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
|
316 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
|
317 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
|
318 // 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
|
319 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
|
320 int bidir, int edges, int step, int mask_mv0, int mask_mv1); |
2967 | 321 |
1644 | 322 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale); |
323 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale); | |
324 | |
2045 | 325 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
|
326 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
327 /* 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
|
328 void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize); |
3574 | 329 /* 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
|
330 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
|
331 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
|
332 /* 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
|
333 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
|
334 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3555
diff
changeset
|
335 /* 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
|
336 * 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
|
337 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
|
338 |
1092 | 339 /* (I)DCT */ |
340 void (*fdct)(DCTELEM *block/* align 16*/); | |
1567 | 341 void (*fdct248)(DCTELEM *block/* align 16*/); |
2967 | 342 |
1324
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
343 /* IDCT really*/ |
7d328fd9d8a5
the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents:
1320
diff
changeset
|
344 void (*idct)(DCTELEM *block/* align 16*/); |
2967 | 345 |
1102 | 346 /** |
1104 | 347 * block -> idct -> clip to unsigned 8 bit -> dest. |
1102 | 348 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...) |
1111 | 349 * @param line_size size in bytes of a horizotal line of dest |
1102 | 350 */ |
1092 | 351 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
2967 | 352 |
1102 | 353 /** |
354 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest. | |
1111 | 355 * @param line_size size in bytes of a horizotal line of dest |
1102 | 356 */ |
1092 | 357 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/); |
2967 | 358 |
1102 | 359 /** |
1104 | 360 * idct input permutation. |
1241 | 361 * several optimized IDCTs need a permutated input (relative to the normal order of the reference |
362 * IDCT) | |
363 * this permutation must be performed before the idct_put/add, note, normally this can be merged | |
364 * with the zigzag/alternate scan<br> | |
1102 | 365 * an example to avoid confusion: |
366 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...) | |
367 * - (x -> referece dct -> reference idct -> x) | |
368 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x) | |
369 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...) | |
370 */ | |
1092 | 371 uint8_t idct_permutation[64]; |
372 int idct_permutation_type; | |
373 #define FF_NO_IDCT_PERM 1 | |
374 #define FF_LIBMPEG2_IDCT_PERM 2 | |
375 #define FF_SIMPLE_IDCT_PERM 3 | |
376 #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
|
377 #define FF_PARTTRANS_IDCT_PERM 5 |
1092 | 378 |
1784 | 379 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale); |
380 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale); | |
381 #define BASIS_SHIFT 16 | |
382 #define RECON_SHIFT 6 | |
2967 | 383 |
4268 | 384 /* h264 functions */ |
2272
cd43603c46f9
move h264 idct to its own file and call via function pointer in DspContext
michael
parents:
2259
diff
changeset
|
385 void (*h264_idct_add)(uint8_t *dst, DCTELEM *block, int stride); |
2755 | 386 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
|
387 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
|
388 void (*h264_idct8_dc_add)(uint8_t *dst, DCTELEM *block, int stride); |
4279 | 389 void (*h264_dct)(DCTELEM block[4][4]); |
3198
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
390 |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
391 /* snow wavelet */ |
6b9f0c4fbdbe
First part of a series of speed-enchancing patches.
gpoirier
parents:
3175
diff
changeset
|
392 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
|
393 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
|
394 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
|
395 |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3198
diff
changeset
|
396 void (*prefetch)(void *mem, int stride, int h); |
3245 | 397 |
398 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height); | |
3526 | 399 |
400 /* vc1 functions */ | |
401 void (*vc1_inv_trans_8x8)(DCTELEM *b); | |
402 void (*vc1_inv_trans_8x4)(DCTELEM *b, int n); | |
403 void (*vc1_inv_trans_4x8)(DCTELEM *b, int n); | |
404 void (*vc1_inv_trans_4x4)(DCTELEM *b, int n); | |
4239 | 405 void (*vc1_v_overlap)(uint8_t* src, int stride); |
406 void (*vc1_h_overlap)(uint8_t* src, int stride); | |
3526 | 407 /* put 8x8 block with bicubic interpolation and quarterpel precision |
408 * last argument is actually round value instead of height | |
409 */ | |
410 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
|
411 } DSPContext; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
412 |
4197 | 413 void dsputil_static_init(void); |
1092 | 414 void dsputil_init(DSPContext* p, AVCodecContext *avctx); |
0 | 415 |
4281
de525a2b41db
ff_check_alignment to warn the user about a missaligned stack
michael
parents:
4279
diff
changeset
|
416 int ff_check_alignment(void); |
de525a2b41db
ff_check_alignment to warn the user about a missaligned stack
michael
parents:
4279
diff
changeset
|
417 |
764 | 418 /** |
419 * permute block according to permuatation. | |
420 * @param last last non zero element in scantable order | |
421 */ | |
1064 | 422 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last); |
34 | 423 |
1729 | 424 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type); |
425 | |
2979 | 426 #define BYTE_VEC32(c) ((c)*0x01010101UL) |
1264 | 427 |
428 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b) | |
429 { | |
430 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
431 } | |
432 | |
433 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b) | |
434 { | |
435 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1); | |
436 } | |
437 | |
2184 | 438 static inline int get_penalty_factor(int lambda, int lambda2, int type){ |
439 switch(type&0xFF){ | |
440 default: | |
441 case FF_CMP_SAD: | |
442 return lambda>>FF_LAMBDA_SHIFT; | |
443 case FF_CMP_DCT: | |
444 return (3*lambda)>>(FF_LAMBDA_SHIFT+1); | |
445 case FF_CMP_W53: | |
446 return (4*lambda)>>(FF_LAMBDA_SHIFT); | |
447 case FF_CMP_W97: | |
448 return (2*lambda)>>(FF_LAMBDA_SHIFT); | |
449 case FF_CMP_SATD: | |
3010
533c6386eca9
8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents:
2979
diff
changeset
|
450 case FF_CMP_DCT264: |
2184 | 451 return (2*lambda)>>FF_LAMBDA_SHIFT; |
452 case FF_CMP_RD: | |
453 case FF_CMP_PSNR: | |
454 case FF_CMP_SSE: | |
455 case FF_CMP_NSSE: | |
456 return lambda2>>FF_LAMBDA_SHIFT; | |
457 case FF_CMP_BIT: | |
458 return 1; | |
459 } | |
460 } | |
461 | |
1102 | 462 /** |
1104 | 463 * Empty mmx state. |
1102 | 464 * this must be called between any dsp function and float/double code. |
465 * for example sin(); dsp->idct_put(); emms_c(); cos() | |
466 */ | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
467 #define emms_c() |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
468 |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
469 /* should be defined by architectures supporting |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
470 one or more MultiMedia extension */ |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
471 int mm_support(void); |
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
984
diff
changeset
|
472 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
473 #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
|
474 |
62 | 475 #if defined(HAVE_MMX) |
0 | 476 |
862 | 477 #undef emms_c |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
838
diff
changeset
|
478 |
0 | 479 #define MM_MMX 0x0001 /* standard MMX */ |
480 #define MM_3DNOW 0x0004 /* AMD 3DNOW */ | |
481 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */ | |
482 #define MM_SSE 0x0008 /* SSE functions */ | |
483 #define MM_SSE2 0x0010 /* PIV SSE2 functions */ | |
2388 | 484 #define MM_3DNOWEXT 0x0020 /* AMD 3DNowExt */ |
3279
647a677c00a4
Remove unused and unsupported Cyrix's "Extended MMX",
gpoirier
parents:
3248
diff
changeset
|
485 #define MM_SSE3 0x0040 /* Prescott SSE3 functions */ |
4311
f9325f2bc762
Add SSSE3 (Core2 aka Conroe/Merom/Woodcrester new instructions) detection
gpoirier
parents:
4281
diff
changeset
|
486 #define MM_SSSE3 0x0080 /* Conroe SSSE3 functions */ |
0 | 487 |
4197 | 488 extern int mm_flags; |
489 | |
1064 | 490 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
491 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
|
492 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size); |
0 | 493 |
494 static inline void emms(void) | |
495 { | |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
496 __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
|
497 } |
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
498 |
936 | 499 |
6
ec4642daa6fe
added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents:
2
diff
changeset
|
500 #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
|
501 {\ |
4197 | 502 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
|
503 emms();\ |
0 | 504 } |
505 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
506 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
3089 | 507 |
2324 | 508 #define STRIDE_ALIGN 8 |
0 | 509 |
1092 | 510 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx); |
511 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx); | |
1065 | 512 |
62 | 513 #elif defined(ARCH_ARMV4L) |
514 | |
515 /* 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
|
516 line optimizations */ |
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(4, t, v) |
2324 | 518 #define STRIDE_ALIGN 4 |
62 | 519 |
2776
930e56f92c57
IWMMXT configure support + runtime selection patch by (Gildas Bazin, gbazin : altern org)
michael
parents:
2763
diff
changeset
|
520 #define MM_IWMMXT 0x0100 /* XScale IWMMXT */ |
930e56f92c57
IWMMXT configure support + runtime selection patch by (Gildas Bazin, gbazin : altern org)
michael
parents:
2763
diff
changeset
|
521 |
4197 | 522 extern int mm_flags; |
523 | |
1092 | 524 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx); |
62 | 525 |
88 | 526 #elif defined(HAVE_MLIB) |
527 | |
528 /* 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
|
529 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 530 #define STRIDE_ALIGN 8 |
88 | 531 |
1092 | 532 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx); |
88 | 533 |
1959
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
534 #elif defined(ARCH_SPARC) |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
535 |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
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 |
1959
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
539 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx); |
55b7435c59b8
VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents:
1879
diff
changeset
|
540 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
541 #elif defined(ARCH_ALPHA) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
542 |
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 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
545 |
1092 | 546 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx); |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
190
diff
changeset
|
547 |
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
|
548 #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
|
549 |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
884
diff
changeset
|
550 #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
|
551 |
4197 | 552 extern int mm_flags; |
553 | |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
554 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN) |
1653 | 555 #define pixel altivec_pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
556 #include <altivec.h> |
1653 | 557 #undef pixel |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
558 #endif |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1008
diff
changeset
|
559 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
560 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v) |
2324 | 561 #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
|
562 |
1092 | 563 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx); |
623
92e99e506920
first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
612
diff
changeset
|
564 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
565 #elif defined(HAVE_MMI) |
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
566 |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
567 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v) |
2324 | 568 #define STRIDE_ALIGN 16 |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
569 |
1092 | 570 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx); |
689
efcbfbd18864
ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
687
diff
changeset
|
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 |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
577 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx); |
e8c3884f2c7e
sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents:
1241
diff
changeset
|
578 |
3728 | 579 #elif defined(ARCH_BFIN) |
580 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
581 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
3728 | 582 #define STRIDE_ALIGN 8 |
583 | |
584 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx); | |
585 | |
0 | 586 #else |
587 | |
4999
0c4bf6b7d1c6
use DECLARE_ALIGNED from mem.h (which is included through common.h)
alex
parents:
4988
diff
changeset
|
588 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v) |
2324 | 589 #define STRIDE_ALIGN 8 |
0 | 590 |
591 #endif | |
592 | |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
214
diff
changeset
|
593 /* PSNR */ |
1064 | 594 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
|
595 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
|
596 AVCodecContext *avctx); |
781 | 597 |
598 /* FFT computation */ | |
599 | |
600 /* NOTE: soon integer code will be added, so you must use the | |
601 FFTSample type */ | |
602 typedef float FFTSample; | |
603 | |
3555 | 604 struct MDCTContext; |
605 | |
781 | 606 typedef struct FFTComplex { |
607 FFTSample re, im; | |
608 } FFTComplex; | |
609 | |
610 typedef struct FFTContext { | |
611 int nbits; | |
612 int inverse; | |
613 uint16_t *revtab; | |
614 FFTComplex *exptab; | |
615 FFTComplex *exptab1; /* only used by SSE code */ | |
616 void (*fft_calc)(struct FFTContext *s, FFTComplex *z); | |
3555 | 617 void (*imdct_calc)(struct MDCTContext *s, FFTSample *output, |
618 const FFTSample *input, FFTSample *tmp); | |
781 | 619 } FFTContext; |
620 | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
621 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
|
622 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
|
623 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
|
624 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z); |
3175 | 625 void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z); |
626 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
|
627 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
|
628 |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
629 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z) |
781 | 630 { |
631 s->fft_calc(s, z); | |
632 } | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1866
diff
changeset
|
633 void ff_fft_end(FFTContext *s); |
781 | 634 |
635 /* MDCT computation */ | |
636 | |
637 typedef struct MDCTContext { | |
638 int n; /* size of MDCT (i.e. number of input data * 2) */ | |
639 int nbits; /* n = 2^nbits */ | |
640 /* pre/post rotation tables */ | |
641 FFTSample *tcos; | |
642 FFTSample *tsin; | |
643 FFTContext fft; | |
644 } MDCTContext; | |
645 | |
794 | 646 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
|
647 void ff_imdct_calc(MDCTContext *s, FFTSample *output, |
781 | 648 const FFTSample *input, FFTSample *tmp); |
3555 | 649 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output, |
650 const FFTSample *input, FFTSample *tmp); | |
3746 | 651 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, |
652 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
|
653 void ff_mdct_calc(MDCTContext *s, FFTSample *out, |
781 | 654 const FFTSample *input, FFTSample *tmp); |
794 | 655 void ff_mdct_end(MDCTContext *s); |
781 | 656 |
1708 | 657 #define WARPER8_16(name8, name16)\ |
658 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
659 return name8(s, dst , src , stride, h)\ | |
660 +name8(s, dst+8 , src+8 , stride, h);\ | |
661 } | |
662 | |
663 #define WARPER8_16_SQ(name8, name16)\ | |
664 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\ | |
665 int score=0;\ | |
666 score +=name8(s, dst , src , stride, 8);\ | |
667 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
668 if(h==16){\ | |
669 dst += 8*stride;\ | |
670 src += 8*stride;\ | |
671 score +=name8(s, dst , src , stride, 8);\ | |
672 score +=name8(s, dst+8 , src+8 , stride, 8);\ | |
673 }\ | |
674 return score;\ | |
936 | 675 } |
676 | |
4240 | 677 |
678 static inline void copy_block2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
679 { | |
680 int i; | |
681 for(i=0; i<h; i++) | |
682 { | |
683 ST16(dst , LD16(src )); | |
684 dst+=dstStride; | |
685 src+=srcStride; | |
686 } | |
687 } | |
688 | |
689 static inline void copy_block4(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
690 { | |
691 int i; | |
692 for(i=0; i<h; i++) | |
693 { | |
694 ST32(dst , LD32(src )); | |
695 dst+=dstStride; | |
696 src+=srcStride; | |
697 } | |
698 } | |
699 | |
700 static inline void copy_block8(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
701 { | |
702 int i; | |
703 for(i=0; i<h; i++) | |
704 { | |
705 ST32(dst , LD32(src )); | |
706 ST32(dst+4 , LD32(src+4 )); | |
707 dst+=dstStride; | |
708 src+=srcStride; | |
709 } | |
710 } | |
711 | |
712 static inline void copy_block9(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
713 { | |
714 int i; | |
715 for(i=0; i<h; i++) | |
716 { | |
717 ST32(dst , LD32(src )); | |
718 ST32(dst+4 , LD32(src+4 )); | |
719 dst[8]= src[8]; | |
720 dst+=dstStride; | |
721 src+=srcStride; | |
722 } | |
723 } | |
724 | |
725 static inline void copy_block16(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
726 { | |
727 int i; | |
728 for(i=0; i<h; i++) | |
729 { | |
730 ST32(dst , LD32(src )); | |
731 ST32(dst+4 , LD32(src+4 )); | |
732 ST32(dst+8 , LD32(src+8 )); | |
733 ST32(dst+12, LD32(src+12)); | |
734 dst+=dstStride; | |
735 src+=srcStride; | |
736 } | |
737 } | |
738 | |
739 static inline void copy_block17(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h) | |
740 { | |
741 int i; | |
742 for(i=0; i<h; i++) | |
743 { | |
744 ST32(dst , LD32(src )); | |
745 ST32(dst+4 , LD32(src+4 )); | |
746 ST32(dst+8 , LD32(src+8 )); | |
747 ST32(dst+12, LD32(src+12)); | |
748 dst[16]= src[16]; | |
749 dst+=dstStride; | |
750 src+=srcStride; | |
751 } | |
752 } | |
753 | |
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
|
754 #endif |