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
36 extern const uint8_t ff_alternate_horizontal_scan[64];
37 extern const uint8_t ff_alternate_vertical_scan[64];
39 extern uint32_t ff_square_tab[512];
41 void ff_put_pixels8x8_c(uint8_t *dst, uint8_t *src, ptrdiff_t stride);
42 void ff_avg_pixels8x8_c(uint8_t *dst, uint8_t *src, ptrdiff_t stride);
43 void ff_put_pixels16x16_c(uint8_t *dst, uint8_t *src, ptrdiff_t stride);
44 void ff_avg_pixels16x16_c(uint8_t *dst, uint8_t *src, ptrdiff_t stride);
46 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
47 int dxx, int dxy, int dyx, int dyy, int shift, int r,
48 int width, int height);
50 /* minimum alignment rules ;)
51 * If you notice errors in the align stuff, need more alignment for some ASM code
52 * for some CPU or need to use a function with less aligned data then send a mail
53 * to the libav-devel mailing list, ...
55 * !warning These alignments might not match reality, (missing attribute((align))
56 * stuff somewhere possible).
57 * I (Michael) did not check them, these are just the alignments which I think
58 * could be reached easily ...
60 * !future video codecs might need functions with less strict alignment
63 /* add and put pixel (decoding)
64 * Block sizes for op_pixels_func are 8x4,8x8 16x8 16x16.
65 * h for op_pixels_func is limited to { width / 2, width },
66 * but never larger than 16 and never smaller than 4. */
67 typedef void (*qpel_mc_func)(uint8_t *dst /* align width (8 or 16) */,
68 uint8_t *src /* align 1 */, ptrdiff_t stride);
70 typedef void (*op_fill_func)(uint8_t *block /* align width (8 or 16) */,
71 uint8_t value, int line_size, int h);
73 #define DEF_OLD_QPEL(name) \
74 void ff_put_ ## name(uint8_t *dst /* align width (8 or 16) */, \
75 uint8_t *src /* align 1 */, ptrdiff_t stride); \
76 void ff_put_no_rnd_ ## name(uint8_t *dst /* align width (8 or 16) */, \
77 uint8_t *src /* align 1 */, ptrdiff_t stride); \
78 void ff_avg_ ## name(uint8_t *dst /* align width (8 or 16) */, \
79 uint8_t *src /* align 1 */, ptrdiff_t stride);
81 DEF_OLD_QPEL(qpel16_mc11_old_c)
82 DEF_OLD_QPEL(qpel16_mc31_old_c)
83 DEF_OLD_QPEL(qpel16_mc12_old_c)
84 DEF_OLD_QPEL(qpel16_mc32_old_c)
85 DEF_OLD_QPEL(qpel16_mc13_old_c)
86 DEF_OLD_QPEL(qpel16_mc33_old_c)
87 DEF_OLD_QPEL(qpel8_mc11_old_c)
88 DEF_OLD_QPEL(qpel8_mc31_old_c)
89 DEF_OLD_QPEL(qpel8_mc12_old_c)
90 DEF_OLD_QPEL(qpel8_mc32_old_c)
91 DEF_OLD_QPEL(qpel8_mc13_old_c)
92 DEF_OLD_QPEL(qpel8_mc33_old_c)
94 struct MpegEncContext;
96 * h is limited to { width / 2, width, 2 * width },
97 * but never larger than 16 and never smaller than 2.
98 * Although currently h < 4 is not used as functions with
99 * width < 8 are neither used nor implemented. */
100 typedef int (*me_cmp_func)(struct MpegEncContext *c,
101 uint8_t *blk1 /* align width (8 or 16) */,
102 uint8_t *blk2 /* align 1 */, int line_size, int h);
107 typedef struct ScanTable {
108 const uint8_t *scantable;
109 uint8_t permutated[64];
110 uint8_t raster_end[64];
113 void ff_init_scantable(uint8_t *permutation, ScanTable *st,
114 const uint8_t *src_scantable);
115 void ff_init_scantable_permutation(uint8_t *idct_permutation,
116 int idct_permutation_type);
121 typedef struct DSPContext {
122 /* pixel ops : interface with DCT */
123 void (*get_pixels)(int16_t *block /* align 16 */,
124 const uint8_t *pixels /* align 8 */,
126 void (*diff_pixels)(int16_t *block /* align 16 */,
127 const uint8_t *s1 /* align 8 */,
128 const uint8_t *s2 /* align 8 */,
130 void (*put_pixels_clamped)(const int16_t *block /* align 16 */,
131 uint8_t *pixels /* align 8 */,
133 void (*put_signed_pixels_clamped)(const int16_t *block /* align 16 */,
134 uint8_t *pixels /* align 8 */,
136 void (*add_pixels_clamped)(const int16_t *block /* align 16 */,
137 uint8_t *pixels /* align 8 */,
139 void (*add_pixels8)(uint8_t *pixels, int16_t *block, int line_size);
140 int (*sum_abs_dctelem)(int16_t *block /* align 16 */);
142 * translational global motion compensation.
144 void (*gmc1)(uint8_t *dst /* align 8 */, uint8_t *src /* align 1 */,
145 int srcStride, int h, int x16, int y16, int rounder);
147 * global motion compensation.
149 void (*gmc)(uint8_t *dst /* align 8 */, uint8_t *src /* align 1 */,
150 int stride, int h, int ox, int oy,
151 int dxx, int dxy, int dyx, int dyy,
152 int shift, int r, int width, int height);
153 void (*clear_block)(int16_t *block /* align 16 */);
154 void (*clear_blocks)(int16_t *blocks /* align 16 */);
155 int (*pix_sum)(uint8_t *pix, int line_size);
156 int (*pix_norm1)(uint8_t *pix, int line_size);
158 me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
160 me_cmp_func hadamard8_diff[6];
161 me_cmp_func dct_sad[6];
162 me_cmp_func quant_psnr[6];
168 me_cmp_func dct_max[6];
169 me_cmp_func dct264_sad[6];
171 me_cmp_func me_pre_cmp[6];
172 me_cmp_func me_cmp[6];
173 me_cmp_func me_sub_cmp[6];
174 me_cmp_func mb_cmp[6];
175 me_cmp_func ildct_cmp[6]; // only width 16 used
176 me_cmp_func frame_skip_cmp[6]; // only width 8 used
178 int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
181 qpel_mc_func put_qpel_pixels_tab[2][16];
182 qpel_mc_func avg_qpel_pixels_tab[2][16];
183 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
184 qpel_mc_func put_mspel_pixels_tab[8];
186 me_cmp_func pix_abs[2][4];
188 /* HuffYUV specific */
189 void (*add_bytes)(uint8_t *dst /* align 16 */,
190 uint8_t *src /* align 16 */,
192 void (*diff_bytes)(uint8_t *dst /* align 16 */,
193 uint8_t *src1 /* align 16 */,
194 uint8_t *src2 /* align 1 */,
197 * Subtract HuffYUV's variant of median prediction.
198 * Note, this might read from src1[-1], src2[-1].
200 void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1,
201 const uint8_t *src2, int w,
202 int *left, int *left_top);
203 void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top,
204 const uint8_t *diff, int w,
205 int *left, int *left_top);
206 int (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src,
208 void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src,
209 int w, int *red, int *green,
210 int *blue, int *alpha);
211 void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
212 void (*bswap16_buf)(uint16_t *dst, const uint16_t *src, int len);
214 /* assume len is a multiple of 8, and arrays are 16-byte aligned */
215 void (*vector_clipf)(float *dst /* align 16 */,
216 const float *src /* align 16 */,
217 float min, float max, int len /* align 16 */);
220 void (*fdct)(int16_t *block /* align 16 */);
221 void (*fdct248)(int16_t *block /* align 16 */);
224 void (*idct)(int16_t *block /* align 16 */);
227 * block -> idct -> clip to unsigned 8 bit -> dest.
228 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
229 * @param line_size size in bytes of a horizontal line of dest
231 void (*idct_put)(uint8_t *dest /* align 8 */,
232 int line_size, int16_t *block /* align 16 */);
235 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
236 * @param line_size size in bytes of a horizontal line of dest
238 void (*idct_add)(uint8_t *dest /* align 8 */,
239 int line_size, int16_t *block /* align 16 */);
242 * IDCT input permutation.
243 * Several optimized IDCTs need a permutated input (relative to the
244 * normal order of the reference IDCT).
245 * This permutation must be performed before the idct_put/add.
246 * Note, normally this can be merged with the zigzag/alternate scan<br>
247 * An example to avoid confusion:
248 * - (->decode coeffs -> zigzag reorder -> dequant -> reference IDCT -> ...)
249 * - (x -> reference DCT -> reference IDCT -> x)
250 * - (x -> reference DCT -> simple_mmx_perm = idct_permutation
251 * -> simple_idct_mmx -> x)
252 * - (-> decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant
253 * -> simple_idct_mmx -> ...)
255 uint8_t idct_permutation[64];
256 int idct_permutation_type;
257 #define FF_NO_IDCT_PERM 1
258 #define FF_LIBMPEG2_IDCT_PERM 2
259 #define FF_SIMPLE_IDCT_PERM 3
260 #define FF_TRANSPOSE_IDCT_PERM 4
261 #define FF_PARTTRANS_IDCT_PERM 5
262 #define FF_SSE2_IDCT_PERM 6
264 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64],
265 int16_t basis[64], int scale);
266 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
267 #define BASIS_SHIFT 16
268 #define RECON_SHIFT 6
270 void (*draw_edges)(uint8_t *buf, int wrap, int width, int height,
271 int w, int h, int sides);
272 #define EDGE_WIDTH 16
274 #define EDGE_BOTTOM 2
276 void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src,
277 int src_wrap, int width, int height);
280 * Calculate scalar product of two vectors.
281 * @param len length of vectors, should be multiple of 16
283 int32_t (*scalarproduct_int16)(const int16_t *v1,
284 const int16_t *v2 /* align 16 */, int len);
287 * Calculate scalar product of v1 and v2,
288 * and v1[i] += v3[i] * mul
289 * @param len length of vectors, should be multiple of 16
291 int32_t (*scalarproduct_and_madd_int16)(int16_t *v1 /* align 16 */,
297 * Clip each element in an array of int32_t to a given minimum and
299 * @param dst destination array
300 * constraints: 16-byte aligned
301 * @param src source array
302 * constraints: 16-byte aligned
303 * @param min minimum value
304 * constraints: must be in the range [-(1 << 24), 1 << 24]
305 * @param max maximum value
306 * constraints: must be in the range [-(1 << 24), 1 << 24]
307 * @param len number of elements in the array
308 * constraints: multiple of 32 greater than zero
310 void (*vector_clip_int32)(int32_t *dst, const int32_t *src, int32_t min,
311 int32_t max, unsigned int len);
313 op_fill_func fill_block_tab[2];
316 void ff_dsputil_static_init(void);
317 void ff_dsputil_init(DSPContext *p, AVCodecContext *avctx);
319 void ff_set_cmp(DSPContext *c, me_cmp_func *cmp, int type);
321 void ff_dsputil_init_arm(DSPContext *c, AVCodecContext *avctx,
322 unsigned high_bit_depth);
323 void ff_dsputil_init_bfin(DSPContext *c, AVCodecContext *avctx,
324 unsigned high_bit_depth);
325 void ff_dsputil_init_ppc(DSPContext *c, AVCodecContext *avctx,
326 unsigned high_bit_depth);
327 void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx,
328 unsigned high_bit_depth);
330 #endif /* AVCODEC_DSPUTIL_H */