3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of Libav.
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * Libav is distributed in the hope that it will be useful,
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.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
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
30 #ifndef AVCODEC_DSPUTIL_H
31 #define AVCODEC_DSPUTIL_H
33 #include "libavutil/intreadwrite.h"
40 extern const uint8_t ff_alternate_horizontal_scan[64];
41 extern const uint8_t ff_alternate_vertical_scan[64];
42 extern const uint8_t ff_zigzag_direct[64];
43 extern const uint8_t ff_zigzag248_direct[64];
45 /* pixel operations */
46 #define MAX_NEG_CROP 1024
49 extern uint32_t ff_squareTbl[512];
50 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP];
52 #define PUTAVG_PIXELS(depth)\
53 void ff_put_pixels8x8_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
54 void ff_avg_pixels8x8_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
55 void ff_put_pixels16x16_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
56 void ff_avg_pixels16x16_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);
62 #define ff_put_pixels8x8_c ff_put_pixels8x8_8_c
63 #define ff_avg_pixels8x8_c ff_avg_pixels8x8_8_c
64 #define ff_put_pixels16x16_c ff_put_pixels16x16_8_c
65 #define ff_avg_pixels16x16_c ff_avg_pixels16x16_8_c
68 void ff_put_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride);
69 void ff_avg_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride);
70 void ff_put_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride);
71 void ff_avg_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride);
73 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
74 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
76 /* minimum alignment rules ;)
77 If you notice errors in the align stuff, need more alignment for some ASM code
78 for some CPU or need to use a function with less aligned data then send a mail
79 to the libav-devel mailing list, ...
81 !warning These alignments might not match reality, (missing attribute((align))
82 stuff somewhere possible).
83 I (Michael) did not check them, these are just the alignments which I think
84 could be reached easily ...
86 !future video codecs might need functions with less strict alignment
89 /* add and put pixel (decoding) */
90 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
91 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller than 4
92 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, ptrdiff_t line_size, int h);
93 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);
94 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
96 typedef void (*op_fill_func)(uint8_t *block/*align width (8 or 16)*/, uint8_t value, int line_size, int h);
98 #define DEF_OLD_QPEL(name)\
99 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
100 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
101 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
103 DEF_OLD_QPEL(qpel16_mc11_old_c)
104 DEF_OLD_QPEL(qpel16_mc31_old_c)
105 DEF_OLD_QPEL(qpel16_mc12_old_c)
106 DEF_OLD_QPEL(qpel16_mc32_old_c)
107 DEF_OLD_QPEL(qpel16_mc13_old_c)
108 DEF_OLD_QPEL(qpel16_mc33_old_c)
109 DEF_OLD_QPEL(qpel8_mc11_old_c)
110 DEF_OLD_QPEL(qpel8_mc31_old_c)
111 DEF_OLD_QPEL(qpel8_mc12_old_c)
112 DEF_OLD_QPEL(qpel8_mc32_old_c)
113 DEF_OLD_QPEL(qpel8_mc13_old_c)
114 DEF_OLD_QPEL(qpel8_mc33_old_c)
116 #define CALL_2X_PIXELS(a, b, n)\
117 static void a(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){\
118 b(block , pixels , line_size, h);\
119 b(block+n, pixels+n, line_size, h);\
122 /* motion estimation */
123 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller than 2
124 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
125 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))*/;
130 typedef struct ScanTable{
131 const uint8_t *scantable;
132 uint8_t permutated[64];
133 uint8_t raster_end[64];
136 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
137 void ff_init_scantable_permutation(uint8_t *idct_permutation,
138 int idct_permutation_type);
143 typedef struct DSPContext {
145 * Size of DCT coefficients.
149 /* pixel ops : interface with DCT */
150 void (*get_pixels)(int16_t *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
151 void (*diff_pixels)(int16_t *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
152 void (*put_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
153 void (*put_signed_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
154 void (*add_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
155 void (*add_pixels8)(uint8_t *pixels, int16_t *block, int line_size);
156 int (*sum_abs_dctelem)(int16_t *block/*align 16*/);
158 * translational global motion compensation.
160 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
162 * global motion compensation.
164 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
165 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
166 void (*clear_block)(int16_t *block/*align 16*/);
167 void (*clear_blocks)(int16_t *blocks/*align 16*/);
168 int (*pix_sum)(uint8_t * pix, int line_size);
169 int (*pix_norm1)(uint8_t * pix, int line_size);
170 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
172 me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
174 me_cmp_func hadamard8_diff[6];
175 me_cmp_func dct_sad[6];
176 me_cmp_func quant_psnr[6];
182 me_cmp_func dct_max[6];
183 me_cmp_func dct264_sad[6];
185 me_cmp_func me_pre_cmp[6];
186 me_cmp_func me_cmp[6];
187 me_cmp_func me_sub_cmp[6];
188 me_cmp_func mb_cmp[6];
189 me_cmp_func ildct_cmp[6]; //only width 16 used
190 me_cmp_func frame_skip_cmp[6]; //only width 8 used
192 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
196 * Halfpel motion compensation with rounding (a+b+1)>>1.
197 * this is an array[4][4] of motion compensation functions for 4
198 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
199 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
200 * @param block destination where the result is stored
201 * @param pixels source
202 * @param line_size number of bytes in a horizontal line of block
205 op_pixels_func put_pixels_tab[4][4];
208 * Halfpel motion compensation with rounding (a+b+1)>>1.
209 * This is an array[4][4] of motion compensation functions for 4
210 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
211 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
212 * @param block destination into which the result is averaged (a+b+1)>>1
213 * @param pixels source
214 * @param line_size number of bytes in a horizontal line of block
217 op_pixels_func avg_pixels_tab[4][4];
220 * Halfpel motion compensation with no rounding (a+b)>>1.
221 * this is an array[2][4] of motion compensation functions for 2
222 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
223 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
224 * @param block destination where the result is stored
225 * @param pixels source
226 * @param line_size number of bytes in a horizontal line of block
229 op_pixels_func put_no_rnd_pixels_tab[2][4];
232 * Halfpel motion compensation with no rounding (a+b)>>1.
233 * this is an array[4] of motion compensation functions for 1
234 * horizontal blocksize (16) and the 4 halfpel positions<br>
235 * *pixels_tab[0][ xhalfpel + 2*yhalfpel ]
236 * @param block destination into which the result is averaged (a+b)>>1
237 * @param pixels source
238 * @param line_size number of bytes in a horizontal line of block
241 op_pixels_func avg_no_rnd_pixels_tab[4];
244 * Thirdpel motion compensation with rounding (a+b+1)>>1.
245 * this is an array[12] of motion compensation functions for the 9 thirdpe
247 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
248 * @param block destination where the result is stored
249 * @param pixels source
250 * @param line_size number of bytes in a horizontal line of block
253 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
254 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
256 qpel_mc_func put_qpel_pixels_tab[2][16];
257 qpel_mc_func avg_qpel_pixels_tab[2][16];
258 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
259 qpel_mc_func put_mspel_pixels_tab[8];
261 me_cmp_func pix_abs[2][4];
263 /* huffyuv specific */
264 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
265 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
267 * subtract huffyuv's variant of median prediction
268 * note, this might read from src1[-1], src2[-1]
270 void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
271 void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
272 int (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
273 void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
274 void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
275 void (*bswap16_buf)(uint16_t *dst, const uint16_t *src, int len);
277 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
278 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
280 void (*h261_loop_filter)(uint8_t *src, int stride);
282 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
283 void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
286 void (*fdct)(int16_t *block/* align 16*/);
287 void (*fdct248)(int16_t *block/* align 16*/);
290 void (*idct)(int16_t *block/* align 16*/);
293 * block -> idct -> clip to unsigned 8 bit -> dest.
294 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
295 * @param line_size size in bytes of a horizontal line of dest
297 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, int16_t *block/*align 16*/);
300 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
301 * @param line_size size in bytes of a horizontal line of dest
303 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, int16_t *block/*align 16*/);
306 * idct input permutation.
307 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
309 * this permutation must be performed before the idct_put/add, note, normally this can be merged
310 * with the zigzag/alternate scan<br>
311 * an example to avoid confusion:
312 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
313 * - (x -> reference dct -> reference idct -> x)
314 * - (x -> reference dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
315 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
317 uint8_t idct_permutation[64];
318 int idct_permutation_type;
319 #define FF_NO_IDCT_PERM 1
320 #define FF_LIBMPEG2_IDCT_PERM 2
321 #define FF_SIMPLE_IDCT_PERM 3
322 #define FF_TRANSPOSE_IDCT_PERM 4
323 #define FF_PARTTRANS_IDCT_PERM 5
324 #define FF_SSE2_IDCT_PERM 6
326 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
327 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
328 #define BASIS_SHIFT 16
329 #define RECON_SHIFT 6
331 void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w, int h, int sides);
332 #define EDGE_WIDTH 16
334 #define EDGE_BOTTOM 2
336 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
339 * Calculate scalar product of two vectors.
340 * @param len length of vectors, should be multiple of 16
342 int32_t (*scalarproduct_int16)(const int16_t *v1, const int16_t *v2/*align 16*/, int len);
345 * Calculate scalar product of v1 and v2,
346 * and v1[i] += v3[i] * mul
347 * @param len length of vectors, should be multiple of 16
349 int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, const int16_t *v2, const int16_t *v3, int len, int mul);
352 * Apply symmetric window in 16-bit fixed-point.
353 * @param output destination array
354 * constraints: 16-byte aligned
355 * @param input source array
356 * constraints: 16-byte aligned
357 * @param window window array
358 * constraints: 16-byte aligned, at least len/2 elements
359 * @param len full window length
360 * constraints: multiple of ? greater than zero
362 void (*apply_window_int16)(int16_t *output, const int16_t *input,
363 const int16_t *window, unsigned int len);
366 * Clip each element in an array of int32_t to a given minimum and maximum value.
367 * @param dst destination array
368 * constraints: 16-byte aligned
369 * @param src source array
370 * constraints: 16-byte aligned
371 * @param min minimum value
372 * constraints: must be in the range [-(1 << 24), 1 << 24]
373 * @param max maximum value
374 * constraints: must be in the range [-(1 << 24), 1 << 24]
375 * @param len number of elements in the array
376 * constraints: multiple of 32 greater than zero
378 void (*vector_clip_int32)(int32_t *dst, const int32_t *src, int32_t min,
379 int32_t max, unsigned int len);
381 op_fill_func fill_block_tab[2];
384 void ff_dsputil_static_init(void);
385 void ff_dsputil_init(DSPContext* p, AVCodecContext *avctx);
387 int ff_check_alignment(void);
389 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
391 void ff_dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
392 void ff_dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
393 void ff_dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
394 void ff_dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
395 void ff_dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
396 void ff_dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
397 void ff_dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
399 #endif /* AVCODEC_DSPUTIL_H */