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_add_add_c(float *dst
, const float *src0
, const float *src1
,
70 const float *src2
, int src3
, int blocksize
, int step
);
71 void ff_vector_fmul_window_c(float *dst
, const float *src0
, const float *src1
,
72 const float *win
, float add_bias
, int len
);
73 void ff_float_to_int16_c(int16_t *dst
, const float *src
, long len
);
74 void ff_float_to_int16_interleave_c(int16_t *dst
, const float **src
, long len
, int channels
);
77 extern const uint8_t ff_alternate_horizontal_scan
[64];
78 extern const uint8_t ff_alternate_vertical_scan
[64];
79 extern const uint8_t ff_zigzag_direct
[64];
80 extern const uint8_t ff_zigzag248_direct
[64];
82 /* pixel operations */
83 #define MAX_NEG_CROP 1024
86 extern uint32_t ff_squareTbl
[512];
87 extern uint8_t ff_cropTbl
[256 + 2 * MAX_NEG_CROP
];
89 /* VP3 DSP functions */
90 void ff_vp3_idct_c(DCTELEM
*block
/* align 16*/);
91 void ff_vp3_idct_put_c(uint8_t *dest
/*align 8*/, int line_size
, DCTELEM
*block
/*align 16*/);
92 void ff_vp3_idct_add_c(uint8_t *dest
/*align 8*/, int line_size
, DCTELEM
*block
/*align 16*/);
94 void ff_vp3_v_loop_filter_c(uint8_t *src
, int stride
, int *bounding_values
);
95 void ff_vp3_h_loop_filter_c(uint8_t *src
, int stride
, int *bounding_values
);
97 /* VP6 DSP functions */
98 void ff_vp6_filter_diag4_c(uint8_t *dst
, uint8_t *src
, int stride
,
99 const int16_t *h_weights
, const int16_t *v_weights
);
101 /* 1/2^n downscaling functions from imgconvert.c */
102 void ff_img_copy_plane(uint8_t *dst
, int dst_wrap
, const uint8_t *src
, int src_wrap
, int width
, int height
);
103 void ff_shrink22(uint8_t *dst
, int dst_wrap
, const uint8_t *src
, int src_wrap
, int width
, int height
);
104 void ff_shrink44(uint8_t *dst
, int dst_wrap
, const uint8_t *src
, int src_wrap
, int width
, int height
);
105 void ff_shrink88(uint8_t *dst
, int dst_wrap
, const uint8_t *src
, int src_wrap
, int width
, int height
);
107 void ff_gmc_c(uint8_t *dst
, uint8_t *src
, int stride
, int h
, int ox
, int oy
,
108 int dxx
, int dxy
, int dyx
, int dyy
, int shift
, int r
, int width
, int height
);
110 /* minimum alignment rules ;)
111 If you notice errors in the align stuff, need more alignment for some ASM code
112 for some CPU or need to use a function with less aligned data then send a mail
113 to the ffmpeg-devel mailing list, ...
115 !warning These alignments might not match reality, (missing attribute((align))
116 stuff somewhere possible).
117 I (Michael) did not check them, these are just the alignments which I think
118 could be reached easily ...
120 !future video codecs might need functions with less strict alignment
124 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
125 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
126 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
127 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
128 void clear_blocks_c(DCTELEM *blocks);
131 /* add and put pixel (decoding) */
132 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
133 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
134 typedef void (*op_pixels_func
)(uint8_t *block
/*align width (8 or 16)*/, const uint8_t *pixels
/*align 1*/, int line_size
, int h
);
135 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
);
136 typedef void (*qpel_mc_func
)(uint8_t *dst
/*align width (8 or 16)*/, uint8_t *src
/*align 1*/, int stride
);
137 typedef void (*h264_chroma_mc_func
)(uint8_t *dst
/*align 8*/, uint8_t *src
/*align 1*/, int srcStride
, int h
, int x
, int y
);
138 typedef void (*h264_weight_func
)(uint8_t *block
, int stride
, int log2_denom
, int weight
, int offset
);
139 typedef void (*h264_biweight_func
)(uint8_t *dst
, uint8_t *src
, int stride
, int log2_denom
, int weightd
, int weights
, int offset
);
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 (*add_pixels_clamped
)(const DCTELEM
*block
/*align 16*/, uint8_t *pixels
/*align 8*/, int line_size
);
203 void (*add_pixels8
)(uint8_t *pixels
, DCTELEM
*block
, int line_size
);
204 void (*add_pixels4
)(uint8_t *pixels
, DCTELEM
*block
, int line_size
);
205 int (*sum_abs_dctelem
)(DCTELEM
*block
/*align 16*/);
207 * translational global motion compensation.
209 void (*gmc1
)(uint8_t *dst
/*align 8*/, uint8_t *src
/*align 1*/, int srcStride
, int h
, int x16
, int y16
, int rounder
);
211 * global motion compensation.
213 void (*gmc
)(uint8_t *dst
/*align 8*/, uint8_t *src
/*align 1*/, int stride
, int h
, int ox
, int oy
,
214 int dxx
, int dxy
, int dyx
, int dyy
, int shift
, int r
, int width
, int height
);
215 void (*clear_block
)(DCTELEM
*block
/*align 16*/);
216 void (*clear_blocks
)(DCTELEM
*blocks
/*align 16*/);
217 int (*pix_sum
)(uint8_t * pix
, int line_size
);
218 int (*pix_norm1
)(uint8_t * pix
, int line_size
);
219 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
221 me_cmp_func sad
[6]; /* identical to pix_absAxA except additional void * */
223 me_cmp_func hadamard8_diff
[6];
224 me_cmp_func dct_sad
[6];
225 me_cmp_func quant_psnr
[6];
233 me_cmp_func dct_max
[6];
234 me_cmp_func dct264_sad
[6];
236 me_cmp_func me_pre_cmp
[6];
237 me_cmp_func me_cmp
[6];
238 me_cmp_func me_sub_cmp
[6];
239 me_cmp_func mb_cmp
[6];
240 me_cmp_func ildct_cmp
[6]; //only width 16 used
241 me_cmp_func frame_skip_cmp
[6]; //only width 8 used
243 int (*ssd_int8_vs_int16
)(const int8_t *pix1
, const int16_t *pix2
,
247 * Halfpel motion compensation with rounding (a+b+1)>>1.
248 * this is an array[4][4] of motion compensation functions for 4
249 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
250 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
251 * @param block destination where the result is stored
252 * @param pixels source
253 * @param line_size number of bytes in a horizontal line of block
256 op_pixels_func put_pixels_tab
[4][4];
259 * Halfpel motion compensation with rounding (a+b+1)>>1.
260 * This is an array[4][4] of motion compensation functions for 4
261 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
262 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
263 * @param block destination into which the result is averaged (a+b+1)>>1
264 * @param pixels source
265 * @param line_size number of bytes in a horizontal line of block
268 op_pixels_func avg_pixels_tab
[4][4];
271 * Halfpel motion compensation with no rounding (a+b)>>1.
272 * this is an array[2][4] of motion compensation functions for 2
273 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
274 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
275 * @param block destination where the result is stored
276 * @param pixels source
277 * @param line_size number of bytes in a horizontal line of block
280 op_pixels_func put_no_rnd_pixels_tab
[4][4];
283 * Halfpel motion compensation with no rounding (a+b)>>1.
284 * this is an array[2][4] of motion compensation functions for 2
285 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
286 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
287 * @param block destination into which the result is averaged (a+b)>>1
288 * @param pixels source
289 * @param line_size number of bytes in a horizontal line of block
292 op_pixels_func avg_no_rnd_pixels_tab
[4][4];
294 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
);
297 * Thirdpel motion compensation with rounding (a+b+1)>>1.
298 * this is an array[12] of motion compensation functions for the 9 thirdpe
300 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
301 * @param block destination where the result is stored
302 * @param pixels source
303 * @param line_size number of bytes in a horizontal line of block
306 tpel_mc_func put_tpel_pixels_tab
[11]; //FIXME individual func ptr per width?
307 tpel_mc_func avg_tpel_pixels_tab
[11]; //FIXME individual func ptr per width?
309 qpel_mc_func put_qpel_pixels_tab
[2][16];
310 qpel_mc_func avg_qpel_pixels_tab
[2][16];
311 qpel_mc_func put_no_rnd_qpel_pixels_tab
[2][16];
312 qpel_mc_func avg_no_rnd_qpel_pixels_tab
[2][16];
313 qpel_mc_func put_mspel_pixels_tab
[8];
318 h264_chroma_mc_func put_h264_chroma_pixels_tab
[3];
319 h264_chroma_mc_func avg_h264_chroma_pixels_tab
[3];
320 /* This is really one func used in VC-1 decoding */
321 h264_chroma_mc_func put_no_rnd_vc1_chroma_pixels_tab
[3];
322 h264_chroma_mc_func avg_no_rnd_vc1_chroma_pixels_tab
[3];
324 qpel_mc_func put_h264_qpel_pixels_tab
[4][16];
325 qpel_mc_func avg_h264_qpel_pixels_tab
[4][16];
327 qpel_mc_func put_2tap_qpel_pixels_tab
[4][16];
328 qpel_mc_func avg_2tap_qpel_pixels_tab
[4][16];
330 h264_weight_func weight_h264_pixels_tab
[10];
331 h264_biweight_func biweight_h264_pixels_tab
[10];
334 qpel_mc_func put_cavs_qpel_pixels_tab
[2][16];
335 qpel_mc_func avg_cavs_qpel_pixels_tab
[2][16];
336 void (*cavs_filter_lv
)(uint8_t *pix
, int stride
, int alpha
, int beta
, int tc
, int bs1
, int bs2
);
337 void (*cavs_filter_lh
)(uint8_t *pix
, int stride
, int alpha
, int beta
, int tc
, int bs1
, int bs2
);
338 void (*cavs_filter_cv
)(uint8_t *pix
, int stride
, int alpha
, int beta
, int tc
, int bs1
, int bs2
);
339 void (*cavs_filter_ch
)(uint8_t *pix
, int stride
, int alpha
, int beta
, int tc
, int bs1
, int bs2
);
340 void (*cavs_idct8_add
)(uint8_t *dst
, DCTELEM
*block
, int stride
);
342 me_cmp_func pix_abs
[2][4];
344 /* huffyuv specific */
345 void (*add_bytes
)(uint8_t *dst
/*align 16*/, uint8_t *src
/*align 16*/, int w
);
346 void (*add_bytes_l2
)(uint8_t *dst
/*align 16*/, uint8_t *src1
/*align 16*/, uint8_t *src2
/*align 16*/, int w
);
347 void (*diff_bytes
)(uint8_t *dst
/*align 16*/, uint8_t *src1
/*align 16*/, uint8_t *src2
/*align 1*/,int w
);
349 * subtract huffyuv's variant of median prediction
350 * note, this might read from src1[-1], src2[-1]
352 void (*sub_hfyu_median_prediction
)(uint8_t *dst
, uint8_t *src1
, uint8_t *src2
, int w
, int *left
, int *left_top
);
353 void (*add_hfyu_median_prediction
)(uint8_t *dst
, uint8_t *top
, uint8_t *diff
, int w
, int *left
, int *left_top
);
354 /* this might write to dst[w] */
355 void (*add_png_paeth_prediction
)(uint8_t *dst
, uint8_t *src
, uint8_t *top
, int w
, int bpp
);
356 void (*bswap_buf
)(uint32_t *dst
, const uint32_t *src
, int w
);
358 void (*h264_v_loop_filter_luma
)(uint8_t *pix
/*align 16*/, int stride
, int alpha
, int beta
, int8_t *tc0
);
359 void (*h264_h_loop_filter_luma
)(uint8_t *pix
/*align 4 */, int stride
, int alpha
, int beta
, int8_t *tc0
);
360 /* v/h_loop_filter_luma_intra: align 16 */
361 void (*h264_v_loop_filter_luma_intra
)(uint8_t *pix
, int stride
, int alpha
, int beta
);
362 void (*h264_h_loop_filter_luma_intra
)(uint8_t *pix
, int stride
, int alpha
, int beta
);
363 void (*h264_v_loop_filter_chroma
)(uint8_t *pix
/*align 8*/, int stride
, int alpha
, int beta
, int8_t *tc0
);
364 void (*h264_h_loop_filter_chroma
)(uint8_t *pix
/*align 4*/, int stride
, int alpha
, int beta
, int8_t *tc0
);
365 void (*h264_v_loop_filter_chroma_intra
)(uint8_t *pix
/*align 8*/, int stride
, int alpha
, int beta
);
366 void (*h264_h_loop_filter_chroma_intra
)(uint8_t *pix
/*align 8*/, int stride
, int alpha
, int beta
);
367 // h264_loop_filter_strength: simd only. the C version is inlined in h264.c
368 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],
369 int bidir
, int edges
, int step
, int mask_mv0
, int mask_mv1
, int field
);
371 void (*h263_v_loop_filter
)(uint8_t *src
, int stride
, int qscale
);
372 void (*h263_h_loop_filter
)(uint8_t *src
, int stride
, int qscale
);
374 void (*h261_loop_filter
)(uint8_t *src
, int stride
);
376 void (*x8_v_loop_filter
)(uint8_t *src
, int stride
, int qscale
);
377 void (*x8_h_loop_filter
)(uint8_t *src
, int stride
, int qscale
);
379 void (*vp3_v_loop_filter
)(uint8_t *src
, int stride
, int *bounding_values
);
380 void (*vp3_h_loop_filter
)(uint8_t *src
, int stride
, int *bounding_values
);
382 void (*vp6_filter_diag4
)(uint8_t *dst
, uint8_t *src
, int stride
,
383 const int16_t *h_weights
,const int16_t *v_weights
);
385 /* assume len is a multiple of 4, and arrays are 16-byte aligned */
386 void (*vorbis_inverse_coupling
)(float *mag
, float *ang
, int blocksize
);
387 void (*ac3_downmix
)(float (*samples
)[256], float (*matrix
)[2], int out_ch
, int in_ch
, int len
);
388 /* no alignment needed */
389 void (*flac_compute_autocorr
)(const int32_t *data
, int len
, int lag
, double *autoc
);
390 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
391 void (*vector_fmul
)(float *dst
, const float *src
, int len
);
392 void (*vector_fmul_reverse
)(float *dst
, const float *src0
, const float *src1
, int len
);
393 /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
394 void (*vector_fmul_add_add
)(float *dst
, const float *src0
, const float *src1
, const float *src2
, int src3
, int len
, int step
);
395 /* assume len is a multiple of 4, and arrays are 16-byte aligned */
396 void (*vector_fmul_window
)(float *dst
, const float *src0
, const float *src1
, const float *win
, float add_bias
, int len
);
397 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
398 void (*int32_to_float_fmul_scalar
)(float *dst
, const int *src
, float mul
, int len
);
399 void (*vector_clipf
)(float *dst
/* align 16 */, const float *src
/* align 16 */, float min
, float max
, int len
/* align 16 */);
401 * Multiply a vector of floats by a scalar float. Source and
402 * destination vectors must overlap exactly or not at all.
403 * @param dst result vector, 16-byte aligned
404 * @param src input vector, 16-byte aligned
405 * @param mul scalar value
406 * @param len length of vector, multiple of 4
408 void (*vector_fmul_scalar
)(float *dst
, const float *src
, float mul
,
411 * Multiply a vector of floats by concatenated short vectors of
412 * floats and by a scalar float. Source and destination vectors
413 * must overlap exactly or not at all.
414 * [0]: short vectors of length 2, 8-byte aligned
415 * [1]: short vectors of length 4, 16-byte aligned
416 * @param dst output vector, 16-byte aligned
417 * @param src input vector, 16-byte aligned
418 * @param sv array of pointers to short vectors
419 * @param mul scalar value
420 * @param len number of elements in src and dst, multiple of 4
422 void (*vector_fmul_sv_scalar
[2])(float *dst
, const float *src
,
423 const float **sv
, float mul
, int len
);
425 * Multiply short vectors of floats by a scalar float, store
426 * concatenated result.
427 * [0]: short vectors of length 2, 8-byte aligned
428 * [1]: short vectors of length 4, 16-byte aligned
429 * @param dst output vector, 16-byte aligned
430 * @param sv array of pointers to short vectors
431 * @param mul scalar value
432 * @param len number of output elements, multiple of 4
434 void (*sv_fmul_scalar
[2])(float *dst
, const float **sv
,
437 * Calculate the scalar product of two vectors of floats.
438 * @param v1 first vector, 16-byte aligned
439 * @param v2 second vector, 16-byte aligned
440 * @param len length of vectors, multiple of 4
442 float (*scalarproduct_float
)(const float *v1
, const float *v2
, int len
);
444 * Calculate the sum and difference of two vectors of floats.
445 * @param v1 first input vector, sum output, 16-byte aligned
446 * @param v2 second input vector, difference output, 16-byte aligned
447 * @param len length of vectors, multiple of 4
449 void (*butterflies_float
)(float *restrict v1
, float *restrict v2
, int len
);
451 /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
452 * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */
453 void (*float_to_int16
)(int16_t *dst
, const float *src
, long len
);
454 void (*float_to_int16_interleave
)(int16_t *dst
, const float **src
, long len
, int channels
);
457 void (*fdct
)(DCTELEM
*block
/* align 16*/);
458 void (*fdct248
)(DCTELEM
*block
/* align 16*/);
461 void (*idct
)(DCTELEM
*block
/* align 16*/);
464 * block -> idct -> clip to unsigned 8 bit -> dest.
465 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
466 * @param line_size size in bytes of a horizontal line of dest
468 void (*idct_put
)(uint8_t *dest
/*align 8*/, int line_size
, DCTELEM
*block
/*align 16*/);
471 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
472 * @param line_size size in bytes of a horizontal line of dest
474 void (*idct_add
)(uint8_t *dest
/*align 8*/, int line_size
, DCTELEM
*block
/*align 16*/);
477 * idct input permutation.
478 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
480 * this permutation must be performed before the idct_put/add, note, normally this can be merged
481 * with the zigzag/alternate scan<br>
482 * an example to avoid confusion:
483 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
484 * - (x -> referece dct -> reference idct -> x)
485 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
486 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
488 uint8_t idct_permutation
[64];
489 int idct_permutation_type
;
490 #define FF_NO_IDCT_PERM 1
491 #define FF_LIBMPEG2_IDCT_PERM 2
492 #define FF_SIMPLE_IDCT_PERM 3
493 #define FF_TRANSPOSE_IDCT_PERM 4
494 #define FF_PARTTRANS_IDCT_PERM 5
495 #define FF_SSE2_IDCT_PERM 6
497 int (*try_8x8basis
)(int16_t rem
[64], int16_t weight
[64], int16_t basis
[64], int scale
);
498 void (*add_8x8basis
)(int16_t rem
[64], int16_t basis
[64], int scale
);
499 #define BASIS_SHIFT 16
500 #define RECON_SHIFT 6
502 void (*draw_edges
)(uint8_t *buf
, int wrap
, int width
, int height
, int w
);
503 #define EDGE_WIDTH 16
506 /* NOTE!!! if you implement any of h264_idct8_add, h264_idct8_add4 then you must implement all of them
507 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
508 The reason for above, is that no 2 out of one list may use a different permutation.
510 void (*h264_idct_add
)(uint8_t *dst
/*align 4*/, DCTELEM
*block
/*align 16*/, int stride
);
511 void (*h264_idct8_add
)(uint8_t *dst
/*align 8*/, DCTELEM
*block
/*align 16*/, int stride
);
512 void (*h264_idct_dc_add
)(uint8_t *dst
/*align 4*/, DCTELEM
*block
/*align 16*/, int stride
);
513 void (*h264_idct8_dc_add
)(uint8_t *dst
/*align 8*/, DCTELEM
*block
/*align 16*/, int stride
);
514 void (*h264_dct
)(DCTELEM block
[4][4]);
515 void (*h264_idct_add16
)(uint8_t *dst
/*align 16*/, const int *blockoffset
, DCTELEM
*block
/*align 16*/, int stride
, const uint8_t nnzc
[6*8]);
516 void (*h264_idct8_add4
)(uint8_t *dst
/*align 16*/, const int *blockoffset
, DCTELEM
*block
/*align 16*/, int stride
, const uint8_t nnzc
[6*8]);
517 void (*h264_idct_add8
)(uint8_t **dst
/*align 16*/, const int *blockoffset
, DCTELEM
*block
/*align 16*/, int stride
, const uint8_t nnzc
[6*8]);
518 void (*h264_idct_add16intra
)(uint8_t *dst
/*align 16*/, const int *blockoffset
, DCTELEM
*block
/*align 16*/, int stride
, const uint8_t nnzc
[6*8]);
521 void (*vertical_compose97i
)(IDWTELEM
*b0
, IDWTELEM
*b1
, IDWTELEM
*b2
, IDWTELEM
*b3
, IDWTELEM
*b4
, IDWTELEM
*b5
, int width
);
522 void (*horizontal_compose97i
)(IDWTELEM
*b
, int width
);
523 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
);
525 void (*prefetch
)(void *mem
, int stride
, int h
);
527 void (*shrink
[4])(uint8_t *dst
, int dst_wrap
, const uint8_t *src
, int src_wrap
, int width
, int height
);
529 /* mlp/truehd functions */
530 void (*mlp_filter_channel
)(int32_t *state
, const int32_t *coeff
,
531 int firorder
, int iirorder
,
532 unsigned int filter_shift
, int32_t mask
, int blocksize
,
533 int32_t *sample_buffer
);
536 void (*vc1_inv_trans_8x8
)(DCTELEM
*b
);
537 void (*vc1_inv_trans_8x4
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
538 void (*vc1_inv_trans_4x8
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
539 void (*vc1_inv_trans_4x4
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
540 void (*vc1_inv_trans_8x8_dc
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
541 void (*vc1_inv_trans_8x4_dc
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
542 void (*vc1_inv_trans_4x8_dc
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
543 void (*vc1_inv_trans_4x4_dc
)(uint8_t *dest
, int line_size
, DCTELEM
*block
);
544 void (*vc1_v_overlap
)(uint8_t* src
, int stride
);
545 void (*vc1_h_overlap
)(uint8_t* src
, int stride
);
546 void (*vc1_v_loop_filter4
)(uint8_t *src
, int stride
, int pq
);
547 void (*vc1_h_loop_filter4
)(uint8_t *src
, int stride
, int pq
);
548 void (*vc1_v_loop_filter8
)(uint8_t *src
, int stride
, int pq
);
549 void (*vc1_h_loop_filter8
)(uint8_t *src
, int stride
, int pq
);
550 void (*vc1_v_loop_filter16
)(uint8_t *src
, int stride
, int pq
);
551 void (*vc1_h_loop_filter16
)(uint8_t *src
, int stride
, int pq
);
552 /* put 8x8 block with bicubic interpolation and quarterpel precision
553 * last argument is actually round value instead of height
555 op_pixels_func put_vc1_mspel_pixels_tab
[16];
556 op_pixels_func avg_vc1_mspel_pixels_tab
[16];
558 /* intrax8 functions */
559 void (*x8_spatial_compensation
[12])(uint8_t *src
, uint8_t *dst
, int linesize
);
560 void (*x8_setup_spatial_compensation
)(uint8_t *src
, uint8_t *dst
, int linesize
,
561 int * range
, int * sum
, int edges
);
565 * Add contents of the second vector to the first one.
566 * @param len length of vectors, should be multiple of 16
568 void (*add_int16
)(int16_t *v1
/*align 16*/, int16_t *v2
, int len
);
570 * Add contents of the second vector to the first one.
571 * @param len length of vectors, should be multiple of 16
573 void (*sub_int16
)(int16_t *v1
/*align 16*/, int16_t *v2
, int len
);
575 * Calculate scalar product of two vectors.
576 * @param len length of vectors, should be multiple of 16
577 * @param shift number of bits to discard from product
579 int32_t (*scalarproduct_int16
)(int16_t *v1
, int16_t *v2
/*align 16*/, int len
, int shift
);
582 qpel_mc_func put_rv30_tpel_pixels_tab
[4][16];
583 qpel_mc_func avg_rv30_tpel_pixels_tab
[4][16];
586 qpel_mc_func put_rv40_qpel_pixels_tab
[4][16];
587 qpel_mc_func avg_rv40_qpel_pixels_tab
[4][16];
588 h264_chroma_mc_func put_rv40_chroma_pixels_tab
[3];
589 h264_chroma_mc_func avg_rv40_chroma_pixels_tab
[3];
592 void dsputil_static_init(void);
593 void dsputil_init(DSPContext
* p
, AVCodecContext
*avctx
);
595 int ff_check_alignment(void);
598 * permute block according to permuatation.
599 * @param last last non zero element in scantable order
601 void ff_block_permute(DCTELEM
*block
, uint8_t *permutation
, const uint8_t *scantable
, int last
);
603 void ff_set_cmp(DSPContext
* c
, me_cmp_func
*cmp
, int type
);
605 #define BYTE_VEC32(c) ((c)*0x01010101UL)
607 static inline uint32_t rnd_avg32(uint32_t a
, uint32_t b
)
609 return (a
| b
) - (((a
^ b
) & ~BYTE_VEC32(0x01)) >> 1);
612 static inline uint32_t no_rnd_avg32(uint32_t a
, uint32_t b
)
614 return (a
& b
) + (((a
^ b
) & ~BYTE_VEC32(0x01)) >> 1);
617 static inline int get_penalty_factor(int lambda
, int lambda2
, int type
){
621 return lambda
>>FF_LAMBDA_SHIFT
;
623 return (3*lambda
)>>(FF_LAMBDA_SHIFT
+1);
625 return (4*lambda
)>>(FF_LAMBDA_SHIFT
);
627 return (2*lambda
)>>(FF_LAMBDA_SHIFT
);
630 return (2*lambda
)>>FF_LAMBDA_SHIFT
;
635 return lambda2
>>FF_LAMBDA_SHIFT
;
643 * this must be called between any dsp function and float/double code.
644 * for example sin(); dsp->idct_put(); emms_c(); cos()
648 /* should be defined by architectures supporting
649 one or more MultiMedia extension */
650 int mm_support(void);
653 void dsputil_init_alpha(DSPContext
* c
, AVCodecContext
*avctx
);
654 void dsputil_init_arm(DSPContext
* c
, AVCodecContext
*avctx
);
655 void dsputil_init_bfin(DSPContext
* c
, AVCodecContext
*avctx
);
656 void dsputil_init_mlib(DSPContext
* c
, AVCodecContext
*avctx
);
657 void dsputil_init_mmi(DSPContext
* c
, AVCodecContext
*avctx
);
658 void dsputil_init_mmx(DSPContext
* c
, AVCodecContext
*avctx
);
659 void dsputil_init_ppc(DSPContext
* c
, AVCodecContext
*avctx
);
660 void dsputil_init_sh4(DSPContext
* c
, AVCodecContext
*avctx
);
661 void dsputil_init_vis(DSPContext
* c
, AVCodecContext
*avctx
);
663 #define DECLARE_ALIGNED_16(t, v) DECLARE_ALIGNED(16, t, v)
664 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v)
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
708 void get_psnr(uint8_t *orig_image
[3], uint8_t *coded_image
[3],
709 int orig_linesize
[3], int coded_linesize
,
710 AVCodecContext
*avctx
);
712 /* FFT computation */
714 /* NOTE: soon integer code will be added, so you must use the
716 typedef float FFTSample
;
718 typedef struct FFTComplex
{
722 typedef struct FFTContext
{
727 FFTComplex
*exptab1
; /* only used by SSE code */
729 int mdct_size
; /* size of MDCT (i.e. number of input data * 2) */
730 int mdct_bits
; /* n = 2^nbits */
731 /* pre/post rotation tables */
734 void (*fft_permute
)(struct FFTContext
*s
, FFTComplex
*z
);
735 void (*fft_calc
)(struct FFTContext
*s
, FFTComplex
*z
);
736 void (*imdct_calc
)(struct FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
);
737 void (*imdct_half
)(struct FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
);
738 void (*mdct_calc
)(struct FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
);
741 #define FF_MDCT_PERM_NONE 0
742 #define FF_MDCT_PERM_INTERLEAVE 1
745 extern FFTSample
* const ff_cos_tabs
[13];
748 * Sets up a complex FFT.
749 * @param nbits log2 of the length of the input array
750 * @param inverse if 0 perform the forward transform, if 1 perform the inverse
752 int ff_fft_init(FFTContext
*s
, int nbits
, int inverse
);
753 void ff_fft_permute_c(FFTContext
*s
, FFTComplex
*z
);
754 void ff_fft_calc_c(FFTContext
*s
, FFTComplex
*z
);
756 void ff_fft_init_altivec(FFTContext
*s
);
757 void ff_fft_init_mmx(FFTContext
*s
);
758 void ff_fft_init_arm(FFTContext
*s
);
761 * Do the permutation needed BEFORE calling ff_fft_calc().
763 static inline void ff_fft_permute(FFTContext
*s
, FFTComplex
*z
)
765 s
->fft_permute(s
, z
);
768 * Do a complex FFT with the parameters defined in ff_fft_init(). The
769 * input data must be permuted before. No 1.0/sqrt(n) normalization is done.
771 static inline void ff_fft_calc(FFTContext
*s
, FFTComplex
*z
)
775 void ff_fft_end(FFTContext
*s
);
777 /* MDCT computation */
779 static inline void ff_imdct_calc(FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
)
781 s
->imdct_calc(s
, output
, input
);
783 static inline void ff_imdct_half(FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
)
785 s
->imdct_half(s
, output
, input
);
788 static inline void ff_mdct_calc(FFTContext
*s
, FFTSample
*output
,
789 const FFTSample
*input
)
791 s
->mdct_calc(s
, output
, input
);
795 * Generate a Kaiser-Bessel Derived Window.
796 * @param window pointer to half window
797 * @param alpha determines window shape
798 * @param n size of half window
800 void ff_kbd_window_init(float *window
, float alpha
, int n
);
803 * Generate a sine window.
804 * @param window pointer to half window
805 * @param n size of half window
807 void ff_sine_window_init(float *window
, int n
);
808 extern float ff_sine_32
[ 32];
809 extern float ff_sine_64
[ 64];
810 extern float ff_sine_128
[ 128];
811 extern float ff_sine_256
[ 256];
812 extern float ff_sine_512
[ 512];
813 extern float ff_sine_1024
[1024];
814 extern float ff_sine_2048
[2048];
815 extern float ff_sine_4096
[4096];
816 extern float * const ff_sine_windows
[13];
818 int ff_mdct_init(FFTContext
*s
, int nbits
, int inverse
, double scale
);
819 void ff_imdct_calc_c(FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
);
820 void ff_imdct_half_c(FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
);
821 void ff_mdct_calc_c(FFTContext
*s
, FFTSample
*output
, const FFTSample
*input
);
822 void ff_mdct_end(FFTContext
*s
);
824 /* Real Discrete Fourier Transform */
826 enum RDFTransformType
{
838 /* pre/post rotation tables */
845 * Sets up a real FFT.
846 * @param nbits log2 of the length of the input array
847 * @param trans the type of transform
849 int ff_rdft_init(RDFTContext
*s
, int nbits
, enum RDFTransformType trans
);
850 void ff_rdft_calc(RDFTContext
*s
, FFTSample
*data
);
851 void ff_rdft_end(RDFTContext
*s
);
853 #define WRAPPER8_16(name8, name16)\
854 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
855 return name8(s, dst , src , stride, h)\
856 +name8(s, dst+8 , src+8 , stride, h);\
859 #define WRAPPER8_16_SQ(name8, name16)\
860 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
862 score +=name8(s, dst , src , stride, 8);\
863 score +=name8(s, dst+8 , src+8 , stride, 8);\
867 score +=name8(s, dst , src , stride, 8);\
868 score +=name8(s, dst+8 , src+8 , stride, 8);\
874 static inline void copy_block2(uint8_t *dst
, const uint8_t *src
, int dstStride
, int srcStride
, int h
)
879 AV_WN16(dst
, AV_RN16(src
));
885 static inline void copy_block4(uint8_t *dst
, const uint8_t *src
, int dstStride
, int srcStride
, int h
)
890 AV_WN32(dst
, AV_RN32(src
));
896 static inline void copy_block8(uint8_t *dst
, const uint8_t *src
, int dstStride
, int srcStride
, int h
)
901 AV_WN32(dst
, AV_RN32(src
));
902 AV_WN32(dst
+4 , AV_RN32(src
+4 ));
908 static inline void copy_block9(uint8_t *dst
, const uint8_t *src
, int dstStride
, int srcStride
, int h
)
913 AV_WN32(dst
, AV_RN32(src
));
914 AV_WN32(dst
+4 , AV_RN32(src
+4 ));
921 static inline void copy_block16(uint8_t *dst
, const uint8_t *src
, int dstStride
, int srcStride
, int h
)
926 AV_WN32(dst
, AV_RN32(src
));
927 AV_WN32(dst
+4 , AV_RN32(src
+4 ));
928 AV_WN32(dst
+8 , AV_RN32(src
+8 ));
929 AV_WN32(dst
+12, AV_RN32(src
+12));
935 static inline void copy_block17(uint8_t *dst
, const uint8_t *src
, int dstStride
, int srcStride
, int h
)
940 AV_WN32(dst
, AV_RN32(src
));
941 AV_WN32(dst
+4 , AV_RN32(src
+4 ));
942 AV_WN32(dst
+8 , AV_RN32(src
+8 ));
943 AV_WN32(dst
+12, AV_RN32(src
+12));
950 #endif /* AVCODEC_DSPUTIL_H */