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);
99 /* 1/2^n downscaling functions from imgconvert.c */
100 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
101 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
102 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
103 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
105 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
106 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
108 /* minimum alignment rules ;)
109 If you notice errors in the align stuff, need more alignment for some ASM code
110 for some CPU or need to use a function with less aligned data then send a mail
111 to the ffmpeg-devel mailing list, ...
113 !warning These alignments might not match reality, (missing attribute((align))
114 stuff somewhere possible).
115 I (Michael) did not check them, these are just the alignments which I think
116 could be reached easily ...
118 !future video codecs might need functions with less strict alignment
122 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
123 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
124 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
125 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
126 void clear_blocks_c(DCTELEM *blocks);
129 /* add and put pixel (decoding) */
130 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
131 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
132 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
133 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);
134 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
135 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
136 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset);
137 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset);
139 typedef void (*op_fill_func)(uint8_t *block/*align width (8 or 16)*/, uint8_t value, int line_size, int h);
141 #define DEF_OLD_QPEL(name)\
142 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
143 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
144 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
146 DEF_OLD_QPEL(qpel16_mc11_old_c)
147 DEF_OLD_QPEL(qpel16_mc31_old_c)
148 DEF_OLD_QPEL(qpel16_mc12_old_c)
149 DEF_OLD_QPEL(qpel16_mc32_old_c)
150 DEF_OLD_QPEL(qpel16_mc13_old_c)
151 DEF_OLD_QPEL(qpel16_mc33_old_c)
152 DEF_OLD_QPEL(qpel8_mc11_old_c)
153 DEF_OLD_QPEL(qpel8_mc31_old_c)
154 DEF_OLD_QPEL(qpel8_mc12_old_c)
155 DEF_OLD_QPEL(qpel8_mc32_old_c)
156 DEF_OLD_QPEL(qpel8_mc13_old_c)
157 DEF_OLD_QPEL(qpel8_mc33_old_c)
159 #define CALL_2X_PIXELS(a, b, n)\
160 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
161 b(block , pixels , line_size, h);\
162 b(block+n, pixels+n, line_size, h);\
165 /* motion estimation */
166 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
167 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
168 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))*/;
172 typedef struct slice_buffer_s slice_buffer;
177 typedef struct ScanTable{
178 const uint8_t *scantable;
179 uint8_t permutated[64];
180 uint8_t raster_end[64];
182 /** Used by dct_quantize_altivec to find last-non-zero */
183 DECLARE_ALIGNED(16, uint8_t, inverse)[64];
187 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
189 void ff_emulated_edge_mc(uint8_t *buf, uint8_t *src, int linesize,
190 int block_w, int block_h,
191 int src_x, int src_y, int w, int h);
196 typedef struct DSPContext {
197 /* pixel ops : interface with DCT */
198 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
199 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
200 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
201 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
202 void (*put_pixels_nonclamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
203 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
204 void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
205 void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
206 int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
208 * translational global motion compensation.
210 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
212 * global motion compensation.
214 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
215 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
216 void (*clear_block)(DCTELEM *block/*align 16*/);
217 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
218 int (*pix_sum)(uint8_t * pix, int line_size);
219 int (*pix_norm1)(uint8_t * pix, int line_size);
220 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
222 me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
224 me_cmp_func hadamard8_diff[6];
225 me_cmp_func dct_sad[6];
226 me_cmp_func quant_psnr[6];
234 me_cmp_func dct_max[6];
235 me_cmp_func dct264_sad[6];
237 me_cmp_func me_pre_cmp[6];
238 me_cmp_func me_cmp[6];
239 me_cmp_func me_sub_cmp[6];
240 me_cmp_func mb_cmp[6];
241 me_cmp_func ildct_cmp[6]; //only width 16 used
242 me_cmp_func frame_skip_cmp[6]; //only width 8 used
244 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
248 * Halfpel motion compensation with rounding (a+b+1)>>1.
249 * this is an array[4][4] of motion compensation functions for 4
250 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
251 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
252 * @param block destination where the result is stored
253 * @param pixels source
254 * @param line_size number of bytes in a horizontal line of block
257 op_pixels_func put_pixels_tab[4][4];
260 * Halfpel motion compensation with rounding (a+b+1)>>1.
261 * This is an array[4][4] of motion compensation functions for 4
262 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
263 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
264 * @param block destination into which the result is averaged (a+b+1)>>1
265 * @param pixels source
266 * @param line_size number of bytes in a horizontal line of block
269 op_pixels_func avg_pixels_tab[4][4];
272 * Halfpel motion compensation with no rounding (a+b)>>1.
273 * this is an array[2][4] of motion compensation functions for 2
274 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
275 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
276 * @param block destination where the result is stored
277 * @param pixels source
278 * @param line_size number of bytes in a horizontal line of block
281 op_pixels_func put_no_rnd_pixels_tab[4][4];
284 * Halfpel motion compensation with no rounding (a+b)>>1.
285 * this is an array[2][4] of motion compensation functions for 2
286 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
287 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
288 * @param block destination into which the result is averaged (a+b)>>1
289 * @param pixels source
290 * @param line_size number of bytes in a horizontal line of block
293 op_pixels_func avg_no_rnd_pixels_tab[4][4];
295 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);
298 * Thirdpel motion compensation with rounding (a+b+1)>>1.
299 * this is an array[12] of motion compensation functions for the 9 thirdpe
301 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
302 * @param block destination where the result is stored
303 * @param pixels source
304 * @param line_size number of bytes in a horizontal line of block
307 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
308 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
310 qpel_mc_func put_qpel_pixels_tab[2][16];
311 qpel_mc_func avg_qpel_pixels_tab[2][16];
312 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
313 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
314 qpel_mc_func put_mspel_pixels_tab[8];
319 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
320 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
321 /* This is really one func used in VC-1 decoding */
322 h264_chroma_mc_func put_no_rnd_vc1_chroma_pixels_tab[3];
323 h264_chroma_mc_func avg_no_rnd_vc1_chroma_pixels_tab[3];
325 qpel_mc_func put_h264_qpel_pixels_tab[4][16];
326 qpel_mc_func avg_h264_qpel_pixels_tab[4][16];
328 qpel_mc_func put_2tap_qpel_pixels_tab[4][16];
329 qpel_mc_func avg_2tap_qpel_pixels_tab[4][16];
331 h264_weight_func weight_h264_pixels_tab[10];
332 h264_biweight_func biweight_h264_pixels_tab[10];
335 qpel_mc_func put_cavs_qpel_pixels_tab[2][16];
336 qpel_mc_func avg_cavs_qpel_pixels_tab[2][16];
337 void (*cavs_filter_lv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
338 void (*cavs_filter_lh)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
339 void (*cavs_filter_cv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
340 void (*cavs_filter_ch)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
341 void (*cavs_idct8_add)(uint8_t *dst, DCTELEM *block, int stride);
343 me_cmp_func pix_abs[2][4];
345 /* huffyuv specific */
346 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
347 void (*add_bytes_l2)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/, int w);
348 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
350 * subtract huffyuv's variant of median prediction
351 * note, this might read from src1[-1], src2[-1]
353 void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
354 void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
355 int (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
356 void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
357 /* this might write to dst[w] */
358 void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
359 void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
361 void (*h264_v_loop_filter_luma)(uint8_t *pix/*align 16*/, int stride, int alpha, int beta, int8_t *tc0);
362 void (*h264_h_loop_filter_luma)(uint8_t *pix/*align 4 */, int stride, int alpha, int beta, int8_t *tc0);
363 /* v/h_loop_filter_luma_intra: align 16 */
364 void (*h264_v_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
365 void (*h264_h_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
366 void (*h264_v_loop_filter_chroma)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta, int8_t *tc0);
367 void (*h264_h_loop_filter_chroma)(uint8_t *pix/*align 4*/, int stride, int alpha, int beta, int8_t *tc0);
368 void (*h264_v_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
369 void (*h264_h_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
370 // h264_loop_filter_strength: simd only. the C version is inlined in h264.c
371 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],
372 int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field);
374 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
375 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
377 void (*h261_loop_filter)(uint8_t *src, int stride);
379 void (*x8_v_loop_filter)(uint8_t *src, int stride, int qscale);
380 void (*x8_h_loop_filter)(uint8_t *src, int stride, int qscale);
382 void (*vp3_v_loop_filter)(uint8_t *src, int stride, int *bounding_values);
383 void (*vp3_h_loop_filter)(uint8_t *src, int stride, int *bounding_values);
385 void (*vp6_filter_diag4)(uint8_t *dst, uint8_t *src, int stride,
386 const int16_t *h_weights,const int16_t *v_weights);
388 /* assume len is a multiple of 4, and arrays are 16-byte aligned */
389 void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
390 void (*ac3_downmix)(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
391 /* no alignment needed */
392 void (*lpc_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
393 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
394 void (*vector_fmul)(float *dst, const float *src, int len);
395 void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
396 /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
397 void (*vector_fmul_add)(float *dst, const float *src0, const float *src1, const float *src2, int len);
398 /* assume len is a multiple of 4, and arrays are 16-byte aligned */
399 void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len);
400 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
401 void (*int32_to_float_fmul_scalar)(float *dst, const int *src, float mul, int len);
402 void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
404 * Multiply a vector of floats by a scalar float. Source and
405 * destination vectors must overlap exactly or not at all.
406 * @param dst result vector, 16-byte aligned
407 * @param src input vector, 16-byte aligned
408 * @param mul scalar value
409 * @param len length of vector, multiple of 4
411 void (*vector_fmul_scalar)(float *dst, const float *src, float mul,
414 * Multiply a vector of floats by concatenated short vectors of
415 * floats and by a scalar float. Source and destination vectors
416 * must overlap exactly or not at all.
417 * [0]: short vectors of length 2, 8-byte aligned
418 * [1]: short vectors of length 4, 16-byte aligned
419 * @param dst output vector, 16-byte aligned
420 * @param src input vector, 16-byte aligned
421 * @param sv array of pointers to short vectors
422 * @param mul scalar value
423 * @param len number of elements in src and dst, multiple of 4
425 void (*vector_fmul_sv_scalar[2])(float *dst, const float *src,
426 const float **sv, float mul, int len);
428 * Multiply short vectors of floats by a scalar float, store
429 * concatenated result.
430 * [0]: short vectors of length 2, 8-byte aligned
431 * [1]: short vectors of length 4, 16-byte aligned
432 * @param dst output vector, 16-byte aligned
433 * @param sv array of pointers to short vectors
434 * @param mul scalar value
435 * @param len number of output elements, multiple of 4
437 void (*sv_fmul_scalar[2])(float *dst, const float **sv,
440 * Calculate the scalar product of two vectors of floats.
441 * @param v1 first vector, 16-byte aligned
442 * @param v2 second vector, 16-byte aligned
443 * @param len length of vectors, multiple of 4
445 float (*scalarproduct_float)(const float *v1, const float *v2, int len);
447 * Calculate the sum and difference of two vectors of floats.
448 * @param v1 first input vector, sum output, 16-byte aligned
449 * @param v2 second input vector, difference output, 16-byte aligned
450 * @param len length of vectors, multiple of 4
452 void (*butterflies_float)(float *restrict v1, float *restrict v2, int len);
454 /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
455 * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */
456 void (*float_to_int16)(int16_t *dst, const float *src, long len);
457 void (*float_to_int16_interleave)(int16_t *dst, const float **src, long len, int channels);
460 void (*fdct)(DCTELEM *block/* align 16*/);
461 void (*fdct248)(DCTELEM *block/* align 16*/);
464 void (*idct)(DCTELEM *block/* align 16*/);
467 * block -> idct -> clip to unsigned 8 bit -> dest.
468 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
469 * @param line_size size in bytes of a horizontal line of dest
471 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
474 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
475 * @param line_size size in bytes of a horizontal line of dest
477 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
480 * idct input permutation.
481 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
483 * this permutation must be performed before the idct_put/add, note, normally this can be merged
484 * with the zigzag/alternate scan<br>
485 * an example to avoid confusion:
486 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
487 * - (x -> referece dct -> reference idct -> x)
488 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
489 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
491 uint8_t idct_permutation[64];
492 int idct_permutation_type;
493 #define FF_NO_IDCT_PERM 1
494 #define FF_LIBMPEG2_IDCT_PERM 2
495 #define FF_SIMPLE_IDCT_PERM 3
496 #define FF_TRANSPOSE_IDCT_PERM 4
497 #define FF_PARTTRANS_IDCT_PERM 5
498 #define FF_SSE2_IDCT_PERM 6
500 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
501 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
502 #define BASIS_SHIFT 16
503 #define RECON_SHIFT 6
505 void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w);
506 #define EDGE_WIDTH 16
509 /* NOTE!!! if you implement any of h264_idct8_add, h264_idct8_add4 then you must implement all of them
510 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
511 The reason for above, is that no 2 out of one list may use a different permutation.
513 void (*h264_idct_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
514 void (*h264_idct8_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
515 void (*h264_idct_dc_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
516 void (*h264_idct8_dc_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
517 void (*h264_dct)(DCTELEM block[4][4]);
518 void (*h264_idct_add16)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
519 void (*h264_idct8_add4)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
520 void (*h264_idct_add8)(uint8_t **dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
521 void (*h264_idct_add16intra)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
524 void (*vertical_compose97i)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
525 void (*horizontal_compose97i)(IDWTELEM *b, int width);
526 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);
528 void (*prefetch)(void *mem, int stride, int h);
530 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
532 /* mlp/truehd functions */
533 void (*mlp_filter_channel)(int32_t *state, const int32_t *coeff,
534 int firorder, int iirorder,
535 unsigned int filter_shift, int32_t mask, int blocksize,
536 int32_t *sample_buffer);
539 void (*vc1_inv_trans_8x8)(DCTELEM *b);
540 void (*vc1_inv_trans_8x4)(uint8_t *dest, int line_size, DCTELEM *block);
541 void (*vc1_inv_trans_4x8)(uint8_t *dest, int line_size, DCTELEM *block);
542 void (*vc1_inv_trans_4x4)(uint8_t *dest, int line_size, DCTELEM *block);
543 void (*vc1_inv_trans_8x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
544 void (*vc1_inv_trans_8x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
545 void (*vc1_inv_trans_4x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
546 void (*vc1_inv_trans_4x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
547 void (*vc1_v_overlap)(uint8_t* src, int stride);
548 void (*vc1_h_overlap)(uint8_t* src, int stride);
549 void (*vc1_v_loop_filter4)(uint8_t *src, int stride, int pq);
550 void (*vc1_h_loop_filter4)(uint8_t *src, int stride, int pq);
551 void (*vc1_v_loop_filter8)(uint8_t *src, int stride, int pq);
552 void (*vc1_h_loop_filter8)(uint8_t *src, int stride, int pq);
553 void (*vc1_v_loop_filter16)(uint8_t *src, int stride, int pq);
554 void (*vc1_h_loop_filter16)(uint8_t *src, int stride, int pq);
555 /* put 8x8 block with bicubic interpolation and quarterpel precision
556 * last argument is actually round value instead of height
558 op_pixels_func put_vc1_mspel_pixels_tab[16];
559 op_pixels_func avg_vc1_mspel_pixels_tab[16];
561 /* intrax8 functions */
562 void (*x8_spatial_compensation[12])(uint8_t *src , uint8_t *dst, int linesize);
563 void (*x8_setup_spatial_compensation)(uint8_t *src, uint8_t *dst, int linesize,
564 int * range, int * sum, int edges);
567 * Calculate scalar product of two vectors.
568 * @param len length of vectors, should be multiple of 16
569 * @param shift number of bits to discard from product
571 int32_t (*scalarproduct_int16)(int16_t *v1, int16_t *v2/*align 16*/, int len, int shift);
574 * Calculate scalar product of v1 and v2,
575 * and v1[i] += v3[i] * mul
576 * @param len length of vectors, should be multiple of 16
578 int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, int16_t *v2, int16_t *v3, int len, int mul);
581 qpel_mc_func put_rv30_tpel_pixels_tab[4][16];
582 qpel_mc_func avg_rv30_tpel_pixels_tab[4][16];
585 qpel_mc_func put_rv40_qpel_pixels_tab[4][16];
586 qpel_mc_func avg_rv40_qpel_pixels_tab[4][16];
587 h264_chroma_mc_func put_rv40_chroma_pixels_tab[3];
588 h264_chroma_mc_func avg_rv40_chroma_pixels_tab[3];
591 op_fill_func fill_block_tab[2];
592 void (*scale_block)(const uint8_t src[64]/*align 8*/, uint8_t *dst/*align 8*/, int linesize);
595 void dsputil_static_init(void);
596 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
598 int ff_check_alignment(void);
601 * permute block according to permuatation.
602 * @param last last non zero element in scantable order
604 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
606 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
608 #define BYTE_VEC32(c) ((c)*0x01010101UL)
610 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
612 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
615 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
617 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
620 static inline int get_penalty_factor(int lambda, int lambda2, int type){
624 return lambda>>FF_LAMBDA_SHIFT;
626 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
628 return (4*lambda)>>(FF_LAMBDA_SHIFT);
630 return (2*lambda)>>(FF_LAMBDA_SHIFT);
633 return (2*lambda)>>FF_LAMBDA_SHIFT;
638 return lambda2>>FF_LAMBDA_SHIFT;
646 * this must be called between any dsp function and float/double code.
647 * for example sin(); dsp->idct_put(); emms_c(); cos()
651 /* should be defined by architectures supporting
652 one or more MultiMedia extension */
653 int mm_support(void);
656 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
657 void dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
658 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
659 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
660 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
661 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
662 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
663 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
664 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
670 static inline void emms(void)
672 __asm__ volatile ("emms;":::"memory");
678 if (mm_flags & FF_MM_MMX)\
685 # define STRIDE_ALIGN 16
690 #define STRIDE_ALIGN 16
694 #define STRIDE_ALIGN 16
699 #define mm_support() 0
704 # define STRIDE_ALIGN 8
707 #define LOCAL_ALIGNED(a, t, v, s, ...) \
708 uint8_t la_##v[sizeof(t s __VA_ARGS__) + (a)]; \
709 t (*v) __VA_ARGS__ = (void *)FFALIGN((uintptr_t)la_##v, a)
711 #if HAVE_LOCAL_ALIGNED_8
712 # define LOCAL_ALIGNED_8(t, v, s, ...) DECLARE_ALIGNED(8, t, v) s __VA_ARGS__
714 # define LOCAL_ALIGNED_8(t, v, s, ...) LOCAL_ALIGNED(8, t, v, s, __VA_ARGS__)
717 #if HAVE_LOCAL_ALIGNED_16
718 # define LOCAL_ALIGNED_16(t, v, s, ...) DECLARE_ALIGNED(16, t, v) s __VA_ARGS__
720 # define LOCAL_ALIGNED_16(t, v, s, ...) LOCAL_ALIGNED(16, t, v, s, __VA_ARGS__)
724 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
725 int orig_linesize[3], int coded_linesize,
726 AVCodecContext *avctx);
728 /* FFT computation */
730 /* NOTE: soon integer code will be added, so you must use the
732 typedef float FFTSample;
734 typedef struct FFTComplex {
738 typedef struct FFTContext {
743 FFTComplex *exptab1; /* only used by SSE code */
745 int mdct_size; /* size of MDCT (i.e. number of input data * 2) */
746 int mdct_bits; /* n = 2^nbits */
747 /* pre/post rotation tables */
750 void (*fft_permute)(struct FFTContext *s, FFTComplex *z);
751 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
752 void (*imdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
753 void (*imdct_half)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
754 void (*mdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
757 #define FF_MDCT_PERM_NONE 0
758 #define FF_MDCT_PERM_INTERLEAVE 1
761 #if CONFIG_HARDCODED_TABLES
762 #define COSTABLE_CONST const
763 #define SINTABLE_CONST const
764 #define SINETABLE_CONST const
766 #define COSTABLE_CONST
767 #define SINTABLE_CONST
768 #define SINETABLE_CONST
771 #define COSTABLE(size) \
772 COSTABLE_CONST DECLARE_ALIGNED(16, FFTSample, ff_cos_##size)[size/2]
773 #define SINTABLE(size) \
774 SINTABLE_CONST DECLARE_ALIGNED(16, FFTSample, ff_sin_##size)[size/2]
775 #define SINETABLE(size) \
776 SINETABLE_CONST DECLARE_ALIGNED(16, float, ff_sine_##size)[size]
780 extern COSTABLE(128);
781 extern COSTABLE(256);
782 extern COSTABLE(512);
783 extern COSTABLE(1024);
784 extern COSTABLE(2048);
785 extern COSTABLE(4096);
786 extern COSTABLE(8192);
787 extern COSTABLE(16384);
788 extern COSTABLE(32768);
789 extern COSTABLE(65536);
790 extern COSTABLE_CONST FFTSample* const ff_cos_tabs[17];
793 * Initializes the cosine table in ff_cos_tabs[index]
794 * \param index index in ff_cos_tabs array of the table to initialize
796 void ff_init_ff_cos_tabs(int index);
801 extern SINTABLE(128);
802 extern SINTABLE(256);
803 extern SINTABLE(512);
804 extern SINTABLE(1024);
805 extern SINTABLE(2048);
806 extern SINTABLE(4096);
807 extern SINTABLE(8192);
808 extern SINTABLE(16384);
809 extern SINTABLE(32768);
810 extern SINTABLE(65536);
813 * Sets up a complex FFT.
814 * @param nbits log2 of the length of the input array
815 * @param inverse if 0 perform the forward transform, if 1 perform the inverse
817 int ff_fft_init(FFTContext *s, int nbits, int inverse);
818 void ff_fft_permute_c(FFTContext *s, FFTComplex *z);
819 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
821 void ff_fft_init_altivec(FFTContext *s);
822 void ff_fft_init_mmx(FFTContext *s);
823 void ff_fft_init_arm(FFTContext *s);
826 * Do the permutation needed BEFORE calling ff_fft_calc().
828 static inline void ff_fft_permute(FFTContext *s, FFTComplex *z)
830 s->fft_permute(s, z);
833 * Do a complex FFT with the parameters defined in ff_fft_init(). The
834 * input data must be permuted before. No 1.0/sqrt(n) normalization is done.
836 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
840 void ff_fft_end(FFTContext *s);
842 /* MDCT computation */
844 static inline void ff_imdct_calc(FFTContext *s, FFTSample *output, const FFTSample *input)
846 s->imdct_calc(s, output, input);
848 static inline void ff_imdct_half(FFTContext *s, FFTSample *output, const FFTSample *input)
850 s->imdct_half(s, output, input);
853 static inline void ff_mdct_calc(FFTContext *s, FFTSample *output,
854 const FFTSample *input)
856 s->mdct_calc(s, output, input);
860 * Generate a Kaiser-Bessel Derived Window.
861 * @param window pointer to half window
862 * @param alpha determines window shape
863 * @param n size of half window
865 void ff_kbd_window_init(float *window, float alpha, int n);
868 * Generate a sine window.
869 * @param window pointer to half window
870 * @param n size of half window
872 void ff_sine_window_init(float *window, int n);
874 * initialize the specified entry of ff_sine_windows
876 void ff_init_ff_sine_windows(int index);
877 extern SINETABLE( 32);
878 extern SINETABLE( 64);
879 extern SINETABLE( 128);
880 extern SINETABLE( 256);
881 extern SINETABLE( 512);
882 extern SINETABLE(1024);
883 extern SINETABLE(2048);
884 extern SINETABLE(4096);
885 extern SINETABLE_CONST float * const ff_sine_windows[13];
887 int ff_mdct_init(FFTContext *s, int nbits, int inverse, double scale);
888 void ff_imdct_calc_c(FFTContext *s, FFTSample *output, const FFTSample *input);
889 void ff_imdct_half_c(FFTContext *s, FFTSample *output, const FFTSample *input);
890 void ff_mdct_calc_c(FFTContext *s, FFTSample *output, const FFTSample *input);
891 void ff_mdct_end(FFTContext *s);
893 /* Real Discrete Fourier Transform */
895 enum RDFTransformType {
907 /* pre/post rotation tables */
908 const FFTSample *tcos;
909 SINTABLE_CONST FFTSample *tsin;
914 * Sets up a real FFT.
915 * @param nbits log2 of the length of the input array
916 * @param trans the type of transform
918 int ff_rdft_init(RDFTContext *s, int nbits, enum RDFTransformType trans);
919 void ff_rdft_calc(RDFTContext *s, FFTSample *data);
920 void ff_rdft_end(RDFTContext *s);
922 /* Discrete Cosine Transform */
934 * Sets up (Inverse)DCT.
935 * @param nbits log2 of the length of the input array
936 * @param inverse >0 forward transform, <0 inverse transform
938 int ff_dct_init(DCTContext *s, int nbits, int inverse);
939 void ff_dct_calc(DCTContext *s, FFTSample *data);
940 void ff_dct_end (DCTContext *s);
942 #define WRAPPER8_16(name8, name16)\
943 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
944 return name8(s, dst , src , stride, h)\
945 +name8(s, dst+8 , src+8 , stride, h);\
948 #define WRAPPER8_16_SQ(name8, name16)\
949 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
951 score +=name8(s, dst , src , stride, 8);\
952 score +=name8(s, dst+8 , src+8 , stride, 8);\
956 score +=name8(s, dst , src , stride, 8);\
957 score +=name8(s, dst+8 , src+8 , stride, 8);\
963 static inline void copy_block2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
968 AV_WN16(dst , AV_RN16(src ));
974 static inline void copy_block4(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
979 AV_WN32(dst , AV_RN32(src ));
985 static inline void copy_block8(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
990 AV_WN32(dst , AV_RN32(src ));
991 AV_WN32(dst+4 , AV_RN32(src+4 ));
997 static inline void copy_block9(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
1002 AV_WN32(dst , AV_RN32(src ));
1003 AV_WN32(dst+4 , AV_RN32(src+4 ));
1010 static inline void copy_block16(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
1015 AV_WN32(dst , AV_RN32(src ));
1016 AV_WN32(dst+4 , AV_RN32(src+4 ));
1017 AV_WN32(dst+8 , AV_RN32(src+8 ));
1018 AV_WN32(dst+12, AV_RN32(src+12));
1024 static inline void copy_block17(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
1029 AV_WN32(dst , AV_RN32(src ));
1030 AV_WN32(dst+4 , AV_RN32(src+4 ));
1031 AV_WN32(dst+8 , AV_RN32(src+8 ));
1032 AV_WN32(dst+12, AV_RN32(src+12));
1039 #endif /* AVCODEC_DSPUTIL_H */