3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of Libav.
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 * note, many functions in here may use MMX which trashes the FPU state, it is
27 * absolutely necessary to call emms_c() between dsp & float/double code
30 #ifndef AVCODEC_DSPUTIL_H
31 #define AVCODEC_DSPUTIL_H
33 #include "libavutil/intreadwrite.h"
40 void ff_fdct_ifast(int16_t *data);
41 void ff_fdct_ifast248(int16_t *data);
42 void ff_jpeg_fdct_islow_8(int16_t *data);
43 void ff_jpeg_fdct_islow_10(int16_t *data);
44 void ff_fdct248_islow_8(int16_t *data);
45 void ff_fdct248_islow_10(int16_t *data);
47 void ff_j_rev_dct(int16_t *data);
49 void ff_fdct_mmx(int16_t *block);
50 void ff_fdct_mmxext(int16_t *block);
51 void ff_fdct_sse2(int16_t *block);
54 extern const uint8_t ff_alternate_horizontal_scan[64];
55 extern const uint8_t ff_alternate_vertical_scan[64];
56 extern const uint8_t ff_zigzag_direct[64];
57 extern const uint8_t ff_zigzag248_direct[64];
59 /* pixel operations */
60 #define MAX_NEG_CROP 1024
63 extern uint32_t ff_squareTbl[512];
64 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP];
66 #define PUTAVG_PIXELS(depth)\
67 void ff_put_pixels8x8_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
68 void ff_avg_pixels8x8_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
69 void ff_put_pixels16x16_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
70 void ff_avg_pixels16x16_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);
76 #define ff_put_pixels8x8_c ff_put_pixels8x8_8_c
77 #define ff_avg_pixels8x8_c ff_avg_pixels8x8_8_c
78 #define ff_put_pixels16x16_c ff_put_pixels16x16_8_c
79 #define ff_avg_pixels16x16_c ff_avg_pixels16x16_8_c
82 void ff_put_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride);
83 void ff_avg_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride);
84 void ff_put_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride);
85 void ff_avg_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride);
87 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
88 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
90 /* minimum alignment rules ;)
91 If you notice errors in the align stuff, need more alignment for some ASM code
92 for some CPU or need to use a function with less aligned data then send a mail
93 to the libav-devel mailing list, ...
95 !warning These alignments might not match reality, (missing attribute((align))
96 stuff somewhere possible).
97 I (Michael) did not check them, these are just the alignments which I think
98 could be reached easily ...
100 !future video codecs might need functions with less strict alignment
103 /* add and put pixel (decoding) */
104 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
105 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller than 4
106 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, ptrdiff_t line_size, int h);
107 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);
108 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
110 typedef void (*op_fill_func)(uint8_t *block/*align width (8 or 16)*/, uint8_t value, int line_size, int h);
112 #define DEF_OLD_QPEL(name)\
113 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
114 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
115 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
117 DEF_OLD_QPEL(qpel16_mc11_old_c)
118 DEF_OLD_QPEL(qpel16_mc31_old_c)
119 DEF_OLD_QPEL(qpel16_mc12_old_c)
120 DEF_OLD_QPEL(qpel16_mc32_old_c)
121 DEF_OLD_QPEL(qpel16_mc13_old_c)
122 DEF_OLD_QPEL(qpel16_mc33_old_c)
123 DEF_OLD_QPEL(qpel8_mc11_old_c)
124 DEF_OLD_QPEL(qpel8_mc31_old_c)
125 DEF_OLD_QPEL(qpel8_mc12_old_c)
126 DEF_OLD_QPEL(qpel8_mc32_old_c)
127 DEF_OLD_QPEL(qpel8_mc13_old_c)
128 DEF_OLD_QPEL(qpel8_mc33_old_c)
130 #define CALL_2X_PIXELS(a, b, n)\
131 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
132 b(block , pixels , line_size, h);\
133 b(block+n, pixels+n, line_size, h);\
136 /* motion estimation */
137 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller than 2
138 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
139 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))*/;
144 typedef struct ScanTable{
145 const uint8_t *scantable;
146 uint8_t permutated[64];
147 uint8_t raster_end[64];
150 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
151 void ff_init_scantable_permutation(uint8_t *idct_permutation,
152 int idct_permutation_type);
157 typedef struct DSPContext {
159 * Size of DCT coefficients.
163 /* pixel ops : interface with DCT */
164 void (*get_pixels)(int16_t *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
165 void (*diff_pixels)(int16_t *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
166 void (*put_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
167 void (*put_signed_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
168 void (*add_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
169 void (*add_pixels8)(uint8_t *pixels, int16_t *block, int line_size);
170 void (*add_pixels4)(uint8_t *pixels, int16_t *block, int line_size);
171 int (*sum_abs_dctelem)(int16_t *block/*align 16*/);
173 * translational global motion compensation.
175 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
177 * global motion compensation.
179 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
180 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
181 void (*clear_block)(int16_t *block/*align 16*/);
182 void (*clear_blocks)(int16_t *blocks/*align 16*/);
183 int (*pix_sum)(uint8_t * pix, int line_size);
184 int (*pix_norm1)(uint8_t * pix, int line_size);
185 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
187 me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
189 me_cmp_func hadamard8_diff[6];
190 me_cmp_func dct_sad[6];
191 me_cmp_func quant_psnr[6];
197 me_cmp_func dct_max[6];
198 me_cmp_func dct264_sad[6];
200 me_cmp_func me_pre_cmp[6];
201 me_cmp_func me_cmp[6];
202 me_cmp_func me_sub_cmp[6];
203 me_cmp_func mb_cmp[6];
204 me_cmp_func ildct_cmp[6]; //only width 16 used
205 me_cmp_func frame_skip_cmp[6]; //only width 8 used
207 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
211 * Halfpel motion compensation with rounding (a+b+1)>>1.
212 * this is an array[4][4] of motion compensation functions for 4
213 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
214 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
215 * @param block destination where the result is stored
216 * @param pixels source
217 * @param line_size number of bytes in a horizontal line of block
220 op_pixels_func put_pixels_tab[4][4];
223 * Halfpel motion compensation with rounding (a+b+1)>>1.
224 * This is an array[4][4] of motion compensation functions for 4
225 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
226 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
227 * @param block destination into which the result is averaged (a+b+1)>>1
228 * @param pixels source
229 * @param line_size number of bytes in a horizontal line of block
232 op_pixels_func avg_pixels_tab[4][4];
235 * Halfpel motion compensation with no rounding (a+b)>>1.
236 * this is an array[2][4] of motion compensation functions for 2
237 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
238 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
239 * @param block destination where the result is stored
240 * @param pixels source
241 * @param line_size number of bytes in a horizontal line of block
244 op_pixels_func put_no_rnd_pixels_tab[2][4];
247 * Halfpel motion compensation with no rounding (a+b)>>1.
248 * this is an array[4] of motion compensation functions for 1
249 * horizontal blocksize (16) and the 4 halfpel positions<br>
250 * *pixels_tab[0][ xhalfpel + 2*yhalfpel ]
251 * @param block destination into which the result is averaged (a+b)>>1
252 * @param pixels source
253 * @param line_size number of bytes in a horizontal line of block
256 op_pixels_func avg_no_rnd_pixels_tab[4];
259 * Thirdpel motion compensation with rounding (a+b+1)>>1.
260 * this is an array[12] of motion compensation functions for the 9 thirdpe
262 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
263 * @param block destination where the result is stored
264 * @param pixels source
265 * @param line_size number of bytes in a horizontal line of block
268 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
269 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
271 qpel_mc_func put_qpel_pixels_tab[2][16];
272 qpel_mc_func avg_qpel_pixels_tab[2][16];
273 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
274 qpel_mc_func put_mspel_pixels_tab[8];
276 me_cmp_func pix_abs[2][4];
278 /* huffyuv specific */
279 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
280 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
282 * subtract huffyuv's variant of median prediction
283 * note, this might read from src1[-1], src2[-1]
285 void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
286 void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
287 int (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
288 void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
289 void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
290 void (*bswap16_buf)(uint16_t *dst, const uint16_t *src, int len);
292 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
293 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
295 void (*h261_loop_filter)(uint8_t *src, int stride);
297 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
298 void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
301 void (*fdct)(int16_t *block/* align 16*/);
302 void (*fdct248)(int16_t *block/* align 16*/);
305 void (*idct)(int16_t *block/* align 16*/);
308 * block -> idct -> clip to unsigned 8 bit -> dest.
309 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
310 * @param line_size size in bytes of a horizontal line of dest
312 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, int16_t *block/*align 16*/);
315 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
316 * @param line_size size in bytes of a horizontal line of dest
318 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, int16_t *block/*align 16*/);
321 * idct input permutation.
322 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
324 * this permutation must be performed before the idct_put/add, note, normally this can be merged
325 * with the zigzag/alternate scan<br>
326 * an example to avoid confusion:
327 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
328 * - (x -> reference dct -> reference idct -> x)
329 * - (x -> reference dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
330 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
332 uint8_t idct_permutation[64];
333 int idct_permutation_type;
334 #define FF_NO_IDCT_PERM 1
335 #define FF_LIBMPEG2_IDCT_PERM 2
336 #define FF_SIMPLE_IDCT_PERM 3
337 #define FF_TRANSPOSE_IDCT_PERM 4
338 #define FF_PARTTRANS_IDCT_PERM 5
339 #define FF_SSE2_IDCT_PERM 6
341 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
342 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
343 #define BASIS_SHIFT 16
344 #define RECON_SHIFT 6
346 void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w, int h, int sides);
347 #define EDGE_WIDTH 16
349 #define EDGE_BOTTOM 2
351 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
354 * Calculate scalar product of two vectors.
355 * @param len length of vectors, should be multiple of 16
357 int32_t (*scalarproduct_int16)(const int16_t *v1, const int16_t *v2/*align 16*/, int len);
360 * Calculate scalar product of v1 and v2,
361 * and v1[i] += v3[i] * mul
362 * @param len length of vectors, should be multiple of 16
364 int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, const int16_t *v2, const int16_t *v3, int len, int mul);
367 * Apply symmetric window in 16-bit fixed-point.
368 * @param output destination array
369 * constraints: 16-byte aligned
370 * @param input source array
371 * constraints: 16-byte aligned
372 * @param window window array
373 * constraints: 16-byte aligned, at least len/2 elements
374 * @param len full window length
375 * constraints: multiple of ? greater than zero
377 void (*apply_window_int16)(int16_t *output, const int16_t *input,
378 const int16_t *window, unsigned int len);
381 * Clip each element in an array of int32_t to a given minimum and maximum value.
382 * @param dst destination array
383 * constraints: 16-byte aligned
384 * @param src source array
385 * constraints: 16-byte aligned
386 * @param min minimum value
387 * constraints: must be in the range [-(1 << 24), 1 << 24]
388 * @param max maximum value
389 * constraints: must be in the range [-(1 << 24), 1 << 24]
390 * @param len number of elements in the array
391 * constraints: multiple of 32 greater than zero
393 void (*vector_clip_int32)(int32_t *dst, const int32_t *src, int32_t min,
394 int32_t max, unsigned int len);
396 op_fill_func fill_block_tab[2];
399 void ff_dsputil_static_init(void);
400 void ff_dsputil_init(DSPContext* p, AVCodecContext *avctx);
402 int ff_check_alignment(void);
404 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
406 void ff_dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
407 void ff_dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
408 void ff_dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
409 void ff_dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
410 void ff_dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
411 void ff_dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
412 void ff_dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
414 #if (ARCH_ARM && HAVE_NEON) || ARCH_PPC || HAVE_MMX
415 # define STRIDE_ALIGN 16
417 # define STRIDE_ALIGN 8
420 #endif /* AVCODEC_DSPUTIL_H */