3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of FFmpeg.
8 * FFmpeg 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 * FFmpeg 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 FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * @file libavcodec/dsputil.h
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"
39 typedef short DCTELEM;
41 typedef short IDWTELEM;
43 void fdct_ifast (DCTELEM *data);
44 void fdct_ifast248 (DCTELEM *data);
45 void ff_jpeg_fdct_islow (DCTELEM *data);
46 void ff_fdct248_islow (DCTELEM *data);
48 void j_rev_dct (DCTELEM *data);
49 void j_rev_dct4 (DCTELEM *data);
50 void j_rev_dct2 (DCTELEM *data);
51 void j_rev_dct1 (DCTELEM *data);
52 void ff_wmv2_idct_c(DCTELEM *data);
54 void ff_fdct_mmx(DCTELEM *block);
55 void ff_fdct_mmx2(DCTELEM *block);
56 void ff_fdct_sse2(DCTELEM *block);
58 void ff_h264_idct8_add_c(uint8_t *dst, DCTELEM *block, int stride);
59 void ff_h264_idct_add_c(uint8_t *dst, DCTELEM *block, int stride);
60 void ff_h264_idct8_dc_add_c(uint8_t *dst, DCTELEM *block, int stride);
61 void ff_h264_idct_dc_add_c(uint8_t *dst, DCTELEM *block, int stride);
62 void ff_h264_lowres_idct_add_c(uint8_t *dst, int stride, DCTELEM *block);
63 void ff_h264_lowres_idct_put_c(uint8_t *dst, int stride, DCTELEM *block);
64 void ff_h264_idct_add16_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
65 void ff_h264_idct_add16intra_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
66 void ff_h264_idct8_add4_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
67 void ff_h264_idct_add8_c(uint8_t **dest, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
69 void ff_vector_fmul_window_c(float *dst, const float *src0, const float *src1,
70 const float *win, float add_bias, int len);
71 void ff_float_to_int16_c(int16_t *dst, const float *src, long len);
72 void ff_float_to_int16_interleave_c(int16_t *dst, const float **src, long len, int channels);
75 extern const uint8_t ff_alternate_horizontal_scan[64];
76 extern const uint8_t ff_alternate_vertical_scan[64];
77 extern const uint8_t ff_zigzag_direct[64];
78 extern const uint8_t ff_zigzag248_direct[64];
80 /* pixel operations */
81 #define MAX_NEG_CROP 1024
84 extern uint32_t ff_squareTbl[512];
85 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP];
87 /* VP3 DSP functions */
88 void ff_vp3_idct_c(DCTELEM *block/* align 16*/);
89 void ff_vp3_idct_put_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
90 void ff_vp3_idct_add_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
92 void ff_vp3_v_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
93 void ff_vp3_h_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
95 /* VP6 DSP functions */
96 void ff_vp6_filter_diag4_c(uint8_t *dst, uint8_t *src, int stride,
97 const int16_t *h_weights, const int16_t *v_weights);
100 void ff_put_vc1_mspel_mc00_c(uint8_t *dst, const uint8_t *src, int stride, int rnd);
101 void ff_avg_vc1_mspel_mc00_c(uint8_t *dst, const uint8_t *src, int stride, int rnd);
103 /* 1/2^n downscaling functions from imgconvert.c */
104 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
105 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
106 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
107 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
109 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
110 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
112 /* minimum alignment rules ;)
113 If you notice errors in the align stuff, need more alignment for some ASM code
114 for some CPU or need to use a function with less aligned data then send a mail
115 to the ffmpeg-devel mailing list, ...
117 !warning These alignments might not match reality, (missing attribute((align))
118 stuff somewhere possible).
119 I (Michael) did not check them, these are just the alignments which I think
120 could be reached easily ...
122 !future video codecs might need functions with less strict alignment
126 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
127 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
128 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
129 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
130 void clear_blocks_c(DCTELEM *blocks);
133 /* add and put pixel (decoding) */
134 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
135 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
136 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
137 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);
138 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
139 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
140 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset);
141 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset);
143 typedef void (*op_fill_func)(uint8_t *block/*align width (8 or 16)*/, uint8_t value, int line_size, int h);
145 #define DEF_OLD_QPEL(name)\
146 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
147 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
148 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
150 DEF_OLD_QPEL(qpel16_mc11_old_c)
151 DEF_OLD_QPEL(qpel16_mc31_old_c)
152 DEF_OLD_QPEL(qpel16_mc12_old_c)
153 DEF_OLD_QPEL(qpel16_mc32_old_c)
154 DEF_OLD_QPEL(qpel16_mc13_old_c)
155 DEF_OLD_QPEL(qpel16_mc33_old_c)
156 DEF_OLD_QPEL(qpel8_mc11_old_c)
157 DEF_OLD_QPEL(qpel8_mc31_old_c)
158 DEF_OLD_QPEL(qpel8_mc12_old_c)
159 DEF_OLD_QPEL(qpel8_mc32_old_c)
160 DEF_OLD_QPEL(qpel8_mc13_old_c)
161 DEF_OLD_QPEL(qpel8_mc33_old_c)
163 #define CALL_2X_PIXELS(a, b, n)\
164 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
165 b(block , pixels , line_size, h);\
166 b(block+n, pixels+n, line_size, h);\
169 /* motion estimation */
170 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
171 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
172 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))*/;
176 typedef struct slice_buffer_s slice_buffer;
181 typedef struct ScanTable{
182 const uint8_t *scantable;
183 uint8_t permutated[64];
184 uint8_t raster_end[64];
186 /** Used by dct_quantize_altivec to find last-non-zero */
187 DECLARE_ALIGNED(16, uint8_t, inverse)[64];
191 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
193 void ff_emulated_edge_mc(uint8_t *buf, uint8_t *src, int linesize,
194 int block_w, int block_h,
195 int src_x, int src_y, int w, int h);
200 typedef struct DSPContext {
201 /* pixel ops : interface with DCT */
202 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
203 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
204 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
205 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
206 void (*put_pixels_nonclamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
207 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
208 void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
209 void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
210 int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
212 * translational global motion compensation.
214 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
216 * global motion compensation.
218 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
219 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
220 void (*clear_block)(DCTELEM *block/*align 16*/);
221 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
222 int (*pix_sum)(uint8_t * pix, int line_size);
223 int (*pix_norm1)(uint8_t * pix, int line_size);
224 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
226 me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
228 me_cmp_func hadamard8_diff[6];
229 me_cmp_func dct_sad[6];
230 me_cmp_func quant_psnr[6];
238 me_cmp_func dct_max[6];
239 me_cmp_func dct264_sad[6];
241 me_cmp_func me_pre_cmp[6];
242 me_cmp_func me_cmp[6];
243 me_cmp_func me_sub_cmp[6];
244 me_cmp_func mb_cmp[6];
245 me_cmp_func ildct_cmp[6]; //only width 16 used
246 me_cmp_func frame_skip_cmp[6]; //only width 8 used
248 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
252 * Halfpel motion compensation with rounding (a+b+1)>>1.
253 * this is an array[4][4] of motion compensation functions for 4
254 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
255 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
256 * @param block destination where the result is stored
257 * @param pixels source
258 * @param line_size number of bytes in a horizontal line of block
261 op_pixels_func put_pixels_tab[4][4];
264 * Halfpel motion compensation with rounding (a+b+1)>>1.
265 * This is an array[4][4] of motion compensation functions for 4
266 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
267 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
268 * @param block destination into which the result is averaged (a+b+1)>>1
269 * @param pixels source
270 * @param line_size number of bytes in a horizontal line of block
273 op_pixels_func avg_pixels_tab[4][4];
276 * Halfpel motion compensation with no rounding (a+b)>>1.
277 * this is an array[2][4] of motion compensation functions for 2
278 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
279 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
280 * @param block destination where the result is stored
281 * @param pixels source
282 * @param line_size number of bytes in a horizontal line of block
285 op_pixels_func put_no_rnd_pixels_tab[4][4];
288 * Halfpel motion compensation with no rounding (a+b)>>1.
289 * this is an array[2][4] of motion compensation functions for 2
290 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
291 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
292 * @param block destination into which the result is averaged (a+b)>>1
293 * @param pixels source
294 * @param line_size number of bytes in a horizontal line of block
297 op_pixels_func avg_no_rnd_pixels_tab[4][4];
299 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);
302 * Thirdpel motion compensation with rounding (a+b+1)>>1.
303 * this is an array[12] of motion compensation functions for the 9 thirdpe
305 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
306 * @param block destination where the result is stored
307 * @param pixels source
308 * @param line_size number of bytes in a horizontal line of block
311 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
312 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
314 qpel_mc_func put_qpel_pixels_tab[2][16];
315 qpel_mc_func avg_qpel_pixels_tab[2][16];
316 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
317 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
318 qpel_mc_func put_mspel_pixels_tab[8];
323 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
324 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
325 /* This is really one func used in VC-1 decoding */
326 h264_chroma_mc_func put_no_rnd_vc1_chroma_pixels_tab[3];
327 h264_chroma_mc_func avg_no_rnd_vc1_chroma_pixels_tab[3];
329 qpel_mc_func put_h264_qpel_pixels_tab[4][16];
330 qpel_mc_func avg_h264_qpel_pixels_tab[4][16];
332 qpel_mc_func put_2tap_qpel_pixels_tab[4][16];
333 qpel_mc_func avg_2tap_qpel_pixels_tab[4][16];
335 h264_weight_func weight_h264_pixels_tab[10];
336 h264_biweight_func biweight_h264_pixels_tab[10];
339 qpel_mc_func put_cavs_qpel_pixels_tab[2][16];
340 qpel_mc_func avg_cavs_qpel_pixels_tab[2][16];
341 void (*cavs_filter_lv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
342 void (*cavs_filter_lh)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
343 void (*cavs_filter_cv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
344 void (*cavs_filter_ch)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
345 void (*cavs_idct8_add)(uint8_t *dst, DCTELEM *block, int stride);
347 me_cmp_func pix_abs[2][4];
349 /* huffyuv specific */
350 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
351 void (*add_bytes_l2)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/, int w);
352 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
354 * subtract huffyuv's variant of median prediction
355 * note, this might read from src1[-1], src2[-1]
357 void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
358 void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
359 int (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
360 void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
361 /* this might write to dst[w] */
362 void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
363 void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
365 void (*h264_v_loop_filter_luma)(uint8_t *pix/*align 16*/, int stride, int alpha, int beta, int8_t *tc0);
366 void (*h264_h_loop_filter_luma)(uint8_t *pix/*align 4 */, int stride, int alpha, int beta, int8_t *tc0);
367 /* v/h_loop_filter_luma_intra: align 16 */
368 void (*h264_v_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
369 void (*h264_h_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
370 void (*h264_v_loop_filter_chroma)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta, int8_t *tc0);
371 void (*h264_h_loop_filter_chroma)(uint8_t *pix/*align 4*/, int stride, int alpha, int beta, int8_t *tc0);
372 void (*h264_v_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
373 void (*h264_h_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
374 // h264_loop_filter_strength: simd only. the C version is inlined in h264.c
375 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],
376 int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field);
378 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
379 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
381 void (*h261_loop_filter)(uint8_t *src, int stride);
383 void (*x8_v_loop_filter)(uint8_t *src, int stride, int qscale);
384 void (*x8_h_loop_filter)(uint8_t *src, int stride, int qscale);
386 void (*vp3_v_loop_filter)(uint8_t *src, int stride, int *bounding_values);
387 void (*vp3_h_loop_filter)(uint8_t *src, int stride, int *bounding_values);
389 void (*vp6_filter_diag4)(uint8_t *dst, uint8_t *src, int stride,
390 const int16_t *h_weights,const int16_t *v_weights);
392 /* assume len is a multiple of 4, and arrays are 16-byte aligned */
393 void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
394 void (*ac3_downmix)(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
395 /* no alignment needed */
396 void (*lpc_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
397 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
398 void (*vector_fmul)(float *dst, const float *src, int len);
399 void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
400 /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
401 void (*vector_fmul_add)(float *dst, const float *src0, const float *src1, const float *src2, int len);
402 /* assume len is a multiple of 4, and arrays are 16-byte aligned */
403 void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len);
404 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
405 void (*int32_to_float_fmul_scalar)(float *dst, const int *src, float mul, int len);
406 void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
408 * Multiply a vector of floats by a scalar float. Source and
409 * destination vectors must overlap exactly or not at all.
410 * @param dst result vector, 16-byte aligned
411 * @param src input vector, 16-byte aligned
412 * @param mul scalar value
413 * @param len length of vector, multiple of 4
415 void (*vector_fmul_scalar)(float *dst, const float *src, float mul,
418 * Multiply a vector of floats by concatenated short vectors of
419 * floats and by a scalar float. Source and destination vectors
420 * must overlap exactly or not at all.
421 * [0]: short vectors of length 2, 8-byte aligned
422 * [1]: short vectors of length 4, 16-byte aligned
423 * @param dst output vector, 16-byte aligned
424 * @param src input vector, 16-byte aligned
425 * @param sv array of pointers to short vectors
426 * @param mul scalar value
427 * @param len number of elements in src and dst, multiple of 4
429 void (*vector_fmul_sv_scalar[2])(float *dst, const float *src,
430 const float **sv, float mul, int len);
432 * Multiply short vectors of floats by a scalar float, store
433 * concatenated result.
434 * [0]: short vectors of length 2, 8-byte aligned
435 * [1]: short vectors of length 4, 16-byte aligned
436 * @param dst output vector, 16-byte aligned
437 * @param sv array of pointers to short vectors
438 * @param mul scalar value
439 * @param len number of output elements, multiple of 4
441 void (*sv_fmul_scalar[2])(float *dst, const float **sv,
444 * Calculate the scalar product of two vectors of floats.
445 * @param v1 first vector, 16-byte aligned
446 * @param v2 second vector, 16-byte aligned
447 * @param len length of vectors, multiple of 4
449 float (*scalarproduct_float)(const float *v1, const float *v2, int len);
451 * Calculate the sum and difference of two vectors of floats.
452 * @param v1 first input vector, sum output, 16-byte aligned
453 * @param v2 second input vector, difference output, 16-byte aligned
454 * @param len length of vectors, multiple of 4
456 void (*butterflies_float)(float *restrict v1, float *restrict v2, int len);
458 /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
459 * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */
460 void (*float_to_int16)(int16_t *dst, const float *src, long len);
461 void (*float_to_int16_interleave)(int16_t *dst, const float **src, long len, int channels);
464 void (*fdct)(DCTELEM *block/* align 16*/);
465 void (*fdct248)(DCTELEM *block/* align 16*/);
468 void (*idct)(DCTELEM *block/* align 16*/);
471 * block -> idct -> clip to unsigned 8 bit -> dest.
472 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
473 * @param line_size size in bytes of a horizontal line of dest
475 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
478 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
479 * @param line_size size in bytes of a horizontal line of dest
481 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
484 * idct input permutation.
485 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
487 * this permutation must be performed before the idct_put/add, note, normally this can be merged
488 * with the zigzag/alternate scan<br>
489 * an example to avoid confusion:
490 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
491 * - (x -> referece dct -> reference idct -> x)
492 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
493 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
495 uint8_t idct_permutation[64];
496 int idct_permutation_type;
497 #define FF_NO_IDCT_PERM 1
498 #define FF_LIBMPEG2_IDCT_PERM 2
499 #define FF_SIMPLE_IDCT_PERM 3
500 #define FF_TRANSPOSE_IDCT_PERM 4
501 #define FF_PARTTRANS_IDCT_PERM 5
502 #define FF_SSE2_IDCT_PERM 6
504 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
505 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
506 #define BASIS_SHIFT 16
507 #define RECON_SHIFT 6
509 void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w);
510 #define EDGE_WIDTH 16
513 /* NOTE!!! if you implement any of h264_idct8_add, h264_idct8_add4 then you must implement all of them
514 NOTE!!! if you implement any of h264_idct_add, h264_idct_add16, h264_idct_add16intra, h264_idct_add8 then you must implement all of them
515 The reason for above, is that no 2 out of one list may use a different permutation.
517 void (*h264_idct_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
518 void (*h264_idct8_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
519 void (*h264_idct_dc_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
520 void (*h264_idct8_dc_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
521 void (*h264_dct)(DCTELEM block[4][4]);
522 void (*h264_idct_add16)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
523 void (*h264_idct8_add4)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
524 void (*h264_idct_add8)(uint8_t **dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
525 void (*h264_idct_add16intra)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
528 void (*vertical_compose97i)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
529 void (*horizontal_compose97i)(IDWTELEM *b, int width);
530 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);
532 void (*prefetch)(void *mem, int stride, int h);
534 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
536 /* mlp/truehd functions */
537 void (*mlp_filter_channel)(int32_t *state, const int32_t *coeff,
538 int firorder, int iirorder,
539 unsigned int filter_shift, int32_t mask, int blocksize,
540 int32_t *sample_buffer);
543 void (*vc1_inv_trans_8x8)(DCTELEM *b);
544 void (*vc1_inv_trans_8x4)(uint8_t *dest, int line_size, DCTELEM *block);
545 void (*vc1_inv_trans_4x8)(uint8_t *dest, int line_size, DCTELEM *block);
546 void (*vc1_inv_trans_4x4)(uint8_t *dest, int line_size, DCTELEM *block);
547 void (*vc1_inv_trans_8x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
548 void (*vc1_inv_trans_8x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
549 void (*vc1_inv_trans_4x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
550 void (*vc1_inv_trans_4x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
551 void (*vc1_v_overlap)(uint8_t* src, int stride);
552 void (*vc1_h_overlap)(uint8_t* src, int stride);
553 void (*vc1_v_loop_filter4)(uint8_t *src, int stride, int pq);
554 void (*vc1_h_loop_filter4)(uint8_t *src, int stride, int pq);
555 void (*vc1_v_loop_filter8)(uint8_t *src, int stride, int pq);
556 void (*vc1_h_loop_filter8)(uint8_t *src, int stride, int pq);
557 void (*vc1_v_loop_filter16)(uint8_t *src, int stride, int pq);
558 void (*vc1_h_loop_filter16)(uint8_t *src, int stride, int pq);
559 /* put 8x8 block with bicubic interpolation and quarterpel precision
560 * last argument is actually round value instead of height
562 op_pixels_func put_vc1_mspel_pixels_tab[16];
563 op_pixels_func avg_vc1_mspel_pixels_tab[16];
565 /* intrax8 functions */
566 void (*x8_spatial_compensation[12])(uint8_t *src , uint8_t *dst, int linesize);
567 void (*x8_setup_spatial_compensation)(uint8_t *src, uint8_t *dst, int linesize,
568 int * range, int * sum, int edges);
571 * Calculate scalar product of two vectors.
572 * @param len length of vectors, should be multiple of 16
573 * @param shift number of bits to discard from product
575 int32_t (*scalarproduct_int16)(int16_t *v1, int16_t *v2/*align 16*/, int len, int shift);
578 * Calculate scalar product of v1 and v2,
579 * and v1[i] += v3[i] * mul
580 * @param len length of vectors, should be multiple of 16
582 int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, int16_t *v2, int16_t *v3, int len, int mul);
585 qpel_mc_func put_rv30_tpel_pixels_tab[4][16];
586 qpel_mc_func avg_rv30_tpel_pixels_tab[4][16];
589 qpel_mc_func put_rv40_qpel_pixels_tab[4][16];
590 qpel_mc_func avg_rv40_qpel_pixels_tab[4][16];
591 h264_chroma_mc_func put_rv40_chroma_pixels_tab[3];
592 h264_chroma_mc_func avg_rv40_chroma_pixels_tab[3];
595 op_fill_func fill_block_tab[2];
596 void (*scale_block)(const uint8_t src[64]/*align 8*/, uint8_t *dst/*align 8*/, int linesize);
599 void dsputil_static_init(void);
600 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
602 int ff_check_alignment(void);
605 * permute block according to permuatation.
606 * @param last last non zero element in scantable order
608 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
610 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
612 #define BYTE_VEC32(c) ((c)*0x01010101UL)
614 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
616 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
619 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
621 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
624 static inline int get_penalty_factor(int lambda, int lambda2, int type){
628 return lambda>>FF_LAMBDA_SHIFT;
630 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
632 return (4*lambda)>>(FF_LAMBDA_SHIFT);
634 return (2*lambda)>>(FF_LAMBDA_SHIFT);
637 return (2*lambda)>>FF_LAMBDA_SHIFT;
642 return lambda2>>FF_LAMBDA_SHIFT;
650 * this must be called between any dsp function and float/double code.
651 * for example sin(); dsp->idct_put(); emms_c(); cos()
655 /* should be defined by architectures supporting
656 one or more MultiMedia extension */
657 int mm_support(void);
660 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
661 void dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
662 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
663 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
664 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
665 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
666 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
667 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
668 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
670 void ff_cavsdsp_init(DSPContext* c, AVCodecContext *avctx);
671 void ff_rv30dsp_init(DSPContext* c, AVCodecContext *avctx);
672 void ff_rv40dsp_init(DSPContext* c, AVCodecContext *avctx);
673 void ff_vc1dsp_init(DSPContext* c, AVCodecContext *avctx);
674 void ff_intrax8dsp_init(DSPContext* c, AVCodecContext *avctx);
675 void ff_mlp_init(DSPContext* c, AVCodecContext *avctx);
676 void ff_mlp_init_x86(DSPContext* c, AVCodecContext *avctx);
682 static inline void emms(void)
684 __asm__ volatile ("emms;":::"memory");
690 if (mm_flags & FF_MM_MMX)\
697 # define STRIDE_ALIGN 16
702 #define STRIDE_ALIGN 16
706 #define STRIDE_ALIGN 16
711 #define mm_support() 0
716 # define STRIDE_ALIGN 8
719 #define LOCAL_ALIGNED(a, t, v, s, ...) \
720 uint8_t la_##v[sizeof(t s __VA_ARGS__) + (a)]; \
721 t (*v) __VA_ARGS__ = (void *)FFALIGN((uintptr_t)la_##v, a)
723 #if HAVE_LOCAL_ALIGNED_8
724 # define LOCAL_ALIGNED_8(t, v, s, ...) DECLARE_ALIGNED(8, t, v) s __VA_ARGS__
726 # define LOCAL_ALIGNED_8(t, v, s, ...) LOCAL_ALIGNED(8, t, v, s, __VA_ARGS__)
729 #if HAVE_LOCAL_ALIGNED_16
730 # define LOCAL_ALIGNED_16(t, v, s, ...) DECLARE_ALIGNED(16, t, v) s __VA_ARGS__
732 # define LOCAL_ALIGNED_16(t, v, s, ...) LOCAL_ALIGNED(16, t, v, s, __VA_ARGS__)
736 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
737 int orig_linesize[3], int coded_linesize,
738 AVCodecContext *avctx);
740 #define WRAPPER8_16(name8, name16)\
741 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
742 return name8(s, dst , src , stride, h)\
743 +name8(s, dst+8 , src+8 , stride, h);\
746 #define WRAPPER8_16_SQ(name8, name16)\
747 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
749 score +=name8(s, dst , src , stride, 8);\
750 score +=name8(s, dst+8 , src+8 , stride, 8);\
754 score +=name8(s, dst , src , stride, 8);\
755 score +=name8(s, dst+8 , src+8 , stride, 8);\
761 static inline void copy_block2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
766 AV_WN16(dst , AV_RN16(src ));
772 static inline void copy_block4(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
777 AV_WN32(dst , AV_RN32(src ));
783 static inline void copy_block8(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
788 AV_WN32(dst , AV_RN32(src ));
789 AV_WN32(dst+4 , AV_RN32(src+4 ));
795 static inline void copy_block9(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
800 AV_WN32(dst , AV_RN32(src ));
801 AV_WN32(dst+4 , AV_RN32(src+4 ));
808 static inline void copy_block16(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
813 AV_WN32(dst , AV_RN32(src ));
814 AV_WN32(dst+4 , AV_RN32(src+4 ));
815 AV_WN32(dst+8 , AV_RN32(src+8 ));
816 AV_WN32(dst+12, AV_RN32(src+12));
822 static inline void copy_block17(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
827 AV_WN32(dst , AV_RN32(src ));
828 AV_WN32(dst+4 , AV_RN32(src+4 ));
829 AV_WN32(dst+8 , AV_RN32(src+8 ));
830 AV_WN32(dst+12, AV_RN32(src+12));
837 #endif /* AVCODEC_DSPUTIL_H */