1/*
2 * BlackFin DSPUTILS
3 *
4 * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
5 * Copyright (c) 2006 Michael Benjamin <michael.benjamin@analog.com>
6 *
7 * This file is part of FFmpeg.
8 *
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24#include <unistd.h>
25#include "libavcodec/avcodec.h"
26#include "libavcodec/dsputil.h"
27#include "dsputil_bfin.h"
28
29int off;
30
31
32void ff_bfin_idct (DCTELEM *block) attribute_l1_text;
33void ff_bfin_fdct (DCTELEM *block) attribute_l1_text;
34void ff_bfin_vp3_idct (DCTELEM *block);
35void ff_bfin_vp3_idct_put (uint8_t *dest, int line_size, DCTELEM *block);
36void ff_bfin_vp3_idct_add (uint8_t *dest, int line_size, DCTELEM *block);
37void ff_bfin_add_pixels_clamped (DCTELEM *block, uint8_t *dest, int line_size) attribute_l1_text;
38void ff_bfin_put_pixels_clamped (DCTELEM *block, uint8_t *dest, int line_size) attribute_l1_text;
39void ff_bfin_diff_pixels (DCTELEM *block, uint8_t *s1, uint8_t *s2, int stride)  attribute_l1_text;
40void ff_bfin_get_pixels  (DCTELEM *restrict block, const uint8_t *pixels, int line_size) attribute_l1_text;
41int  ff_bfin_pix_norm1  (uint8_t * pix, int line_size) attribute_l1_text;
42int  ff_bfin_z_sad8x8   (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h) attribute_l1_text;
43int  ff_bfin_z_sad16x16 (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h) attribute_l1_text;
44
45void ff_bfin_z_put_pixels16_xy2     (uint8_t *block, const uint8_t *s0, int dest_size, int line_size, int h) attribute_l1_text;
46void ff_bfin_z_put_pixels8_xy2      (uint8_t *block, const uint8_t *s0, int dest_size, int line_size, int h) attribute_l1_text;
47void ff_bfin_put_pixels16_xy2_nornd (uint8_t *block, const uint8_t *s0, int line_size, int h) attribute_l1_text;
48void ff_bfin_put_pixels8_xy2_nornd  (uint8_t *block, const uint8_t *s0, int line_size, int h) attribute_l1_text;
49
50
51int  ff_bfin_pix_sum (uint8_t *p, int stride) attribute_l1_text;
52
53void ff_bfin_put_pixels8uc        (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int dest_size, int line_size, int h) attribute_l1_text;
54void ff_bfin_put_pixels16uc       (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int dest_size, int line_size, int h) attribute_l1_text;
55void ff_bfin_put_pixels8uc_nornd  (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int line_size, int h) attribute_l1_text;
56void ff_bfin_put_pixels16uc_nornd (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int line_size, int h) attribute_l1_text;
57
58int ff_bfin_sse4  (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) attribute_l1_text;
59int ff_bfin_sse8  (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) attribute_l1_text;
60int ff_bfin_sse16 (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) attribute_l1_text;
61
62
63static void bfin_idct_add (uint8_t *dest, int line_size, DCTELEM *block)
64{
65    ff_bfin_idct (block);
66    ff_bfin_add_pixels_clamped (block, dest, line_size);
67}
68
69static void bfin_idct_put (uint8_t *dest, int line_size, DCTELEM *block)
70{
71    ff_bfin_idct (block);
72    ff_bfin_put_pixels_clamped (block, dest, line_size);
73}
74
75
76static void bfin_clear_blocks (DCTELEM *blocks)
77{
78    // This is just a simple memset.
79    //
80    __asm__("P0=192; "
81        "I0=%0;  "
82        "R0=0;   "
83        "LSETUP(clear_blocks_blkfn_lab,clear_blocks_blkfn_lab)LC0=P0;"
84        "clear_blocks_blkfn_lab:"
85        "[I0++]=R0;"
86        ::"a" (blocks):"P0","I0","R0");
87}
88
89
90
91static void bfin_put_pixels8 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
92{
93    ff_bfin_put_pixels8uc (block, pixels, pixels, line_size, line_size, h);
94}
95
96static void bfin_put_pixels8_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
97{
98    ff_bfin_put_pixels8uc (block, pixels, pixels+1, line_size, line_size, h);
99}
100
101static void bfin_put_pixels8_y2 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
102{
103    ff_bfin_put_pixels8uc (block, pixels, pixels+line_size, line_size, line_size, h);
104}
105
106static void bfin_put_pixels8_xy2 (uint8_t *block, const uint8_t *s0, int line_size, int h)
107{
108    ff_bfin_z_put_pixels8_xy2 (block,s0,line_size, line_size, h);
109}
110
111static void bfin_put_pixels16 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
112{
113    ff_bfin_put_pixels16uc (block, pixels, pixels, line_size, line_size, h);
114}
115
116static void bfin_put_pixels16_x2 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
117{
118    ff_bfin_put_pixels16uc (block, pixels, pixels+1, line_size, line_size, h);
119}
120
121static void bfin_put_pixels16_y2 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
122{
123    ff_bfin_put_pixels16uc (block, pixels, pixels+line_size, line_size, line_size, h);
124}
125
126static void bfin_put_pixels16_xy2 (uint8_t *block, const uint8_t *s0, int line_size, int h)
127{
128    ff_bfin_z_put_pixels16_xy2 (block,s0,line_size, line_size, h);
129}
130
131void bfin_put_pixels8_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
132{
133    ff_bfin_put_pixels8uc_nornd (block, pixels, pixels, line_size, h);
134}
135
136static void bfin_put_pixels8_x2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
137{
138    ff_bfin_put_pixels8uc_nornd (block, pixels, pixels+1, line_size, h);
139}
140
141static void bfin_put_pixels8_y2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
142{
143    ff_bfin_put_pixels8uc_nornd (block, pixels, pixels+line_size, line_size, h);
144}
145
146
147void bfin_put_pixels16_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
148{
149    ff_bfin_put_pixels16uc_nornd (block, pixels, pixels, line_size, h);
150}
151
152static void bfin_put_pixels16_x2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
153{
154    ff_bfin_put_pixels16uc_nornd (block, pixels, pixels+1, line_size, h);
155}
156
157static void bfin_put_pixels16_y2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
158{
159    ff_bfin_put_pixels16uc_nornd (block, pixels, pixels+line_size, line_size, h);
160}
161
162static int bfin_pix_abs16 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
163{
164    return ff_bfin_z_sad16x16 (blk1,blk2,line_size,line_size,h);
165}
166
167static int bfin_vsad_intra16 (void *c, uint8_t *blk1, uint8_t *dummy, int stride, int h) {
168    return ff_bfin_z_sad16x16 (blk1,blk1+stride,stride<<1,stride<<1,h);
169}
170
171static int bfin_vsad (void *c, uint8_t *blk1, uint8_t *blk2, int stride, int h) {
172    return ff_bfin_z_sad16x16 (blk1,blk1+stride,stride<<1,stride<<1,h)
173        + ff_bfin_z_sad16x16 (blk2,blk2+stride,stride<<1,stride<<1,h);
174}
175
176static uint8_t vtmp_blk[256] attribute_l1_data_b;
177
178static int bfin_pix_abs16_x2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
179{
180    ff_bfin_put_pixels16uc (vtmp_blk, blk2, blk2+1, 16, line_size, h);
181    return ff_bfin_z_sad16x16 (blk1, vtmp_blk, line_size, 16, h);
182}
183
184static int bfin_pix_abs16_y2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
185{
186    ff_bfin_put_pixels16uc (vtmp_blk, blk2, blk2+line_size, 16, line_size, h);
187    return ff_bfin_z_sad16x16 (blk1, vtmp_blk, line_size, 16, h);
188}
189
190static int bfin_pix_abs16_xy2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
191{
192    ff_bfin_z_put_pixels16_xy2 (vtmp_blk, blk2, 16, line_size, h);
193    return ff_bfin_z_sad16x16 (blk1, vtmp_blk, line_size, 16, h);
194}
195
196static int bfin_pix_abs8 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
197{
198    return ff_bfin_z_sad8x8 (blk1,blk2,line_size,line_size, h);
199}
200
201static int bfin_pix_abs8_x2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
202{
203    ff_bfin_put_pixels8uc (vtmp_blk, blk2, blk2+1, 8, line_size, h);
204    return ff_bfin_z_sad8x8 (blk1, vtmp_blk, line_size, 8, h);
205}
206
207static int bfin_pix_abs8_y2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
208{
209    ff_bfin_put_pixels8uc (vtmp_blk, blk2, blk2+line_size, 8, line_size, h);
210    return ff_bfin_z_sad8x8 (blk1, vtmp_blk, line_size, 8, h);
211}
212
213static int bfin_pix_abs8_xy2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
214{
215    ff_bfin_z_put_pixels8_xy2 (vtmp_blk, blk2, 8, line_size, h);
216    return ff_bfin_z_sad8x8 (blk1, vtmp_blk, line_size, 8, h);
217}
218
219
220/*
221  decoder optimization
222  start on 2/11 100 frames of 352x240@25 compiled with no optimization -g debugging
223  9.824s ~ 2.44x off
224  6.360s ~ 1.58x off with -O2
225  5.740s ~ 1.43x off with idcts
226
227  2.64s    2/20 same sman.mp4 decode only
228
229*/
230
231void dsputil_init_bfin( DSPContext* c, AVCodecContext *avctx )
232{
233    c->get_pixels         = ff_bfin_get_pixels;
234    c->diff_pixels        = ff_bfin_diff_pixels;
235    c->put_pixels_clamped = ff_bfin_put_pixels_clamped;
236    c->add_pixels_clamped = ff_bfin_add_pixels_clamped;
237
238    c->clear_blocks       = bfin_clear_blocks;
239    c->pix_sum            = ff_bfin_pix_sum;
240    c->pix_norm1          = ff_bfin_pix_norm1;
241
242    c->sad[0]             = bfin_pix_abs16;
243    c->sad[1]             = bfin_pix_abs8;
244
245    c->vsad[0]            = bfin_vsad;
246    c->vsad[4]            = bfin_vsad_intra16;
247
248    /* TODO [0] 16  [1] 8 */
249    c->pix_abs[0][0] = bfin_pix_abs16;
250    c->pix_abs[0][1] = bfin_pix_abs16_x2;
251    c->pix_abs[0][2] = bfin_pix_abs16_y2;
252    c->pix_abs[0][3] = bfin_pix_abs16_xy2;
253
254    c->pix_abs[1][0] = bfin_pix_abs8;
255    c->pix_abs[1][1] = bfin_pix_abs8_x2;
256    c->pix_abs[1][2] = bfin_pix_abs8_y2;
257    c->pix_abs[1][3] = bfin_pix_abs8_xy2;
258
259
260    c->sse[0] = ff_bfin_sse16;
261    c->sse[1] = ff_bfin_sse8;
262    c->sse[2] = ff_bfin_sse4;
263
264
265    /**
266     * Halfpel motion compensation with rounding (a+b+1)>>1.
267     * This is an array[4][4] of motion compensation functions for 4
268     * horizontal blocksizes (8,16) and the 4 halfpel positions
269     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
270     * @param block destination where the result is stored
271     * @param pixels source
272     * @param line_size number of bytes in a horizontal line of block
273     * @param h height
274     */
275
276    c->put_pixels_tab[0][0] = bfin_put_pixels16;
277    c->put_pixels_tab[0][1] = bfin_put_pixels16_x2;
278    c->put_pixels_tab[0][2] = bfin_put_pixels16_y2;
279    c->put_pixels_tab[0][3] = bfin_put_pixels16_xy2;
280
281    c->put_pixels_tab[1][0] = bfin_put_pixels8;
282    c->put_pixels_tab[1][1] = bfin_put_pixels8_x2;
283    c->put_pixels_tab[1][2] = bfin_put_pixels8_y2;
284    c->put_pixels_tab[1][3] = bfin_put_pixels8_xy2;
285
286    c->put_no_rnd_pixels_tab[1][0] = bfin_put_pixels8_nornd;
287    c->put_no_rnd_pixels_tab[1][1] = bfin_put_pixels8_x2_nornd;
288    c->put_no_rnd_pixels_tab[1][2] = bfin_put_pixels8_y2_nornd;
289    c->put_no_rnd_pixels_tab[1][3] = ff_bfin_put_pixels8_xy2_nornd;
290
291    c->put_no_rnd_pixels_tab[0][0] = bfin_put_pixels16_nornd;
292    c->put_no_rnd_pixels_tab[0][1] = bfin_put_pixels16_x2_nornd;
293    c->put_no_rnd_pixels_tab[0][2] = bfin_put_pixels16_y2_nornd;
294    c->put_no_rnd_pixels_tab[0][3] = ff_bfin_put_pixels16_xy2_nornd;
295
296    c->idct_permutation_type = FF_NO_IDCT_PERM;
297    c->fdct                  = ff_bfin_fdct;
298    if (avctx->idct_algo==FF_IDCT_VP3) {
299        c->idct               = ff_bfin_vp3_idct;
300        c->idct_add           = ff_bfin_vp3_idct_add;
301        c->idct_put           = ff_bfin_vp3_idct_put;
302    } else {
303        c->idct               = ff_bfin_idct;
304        c->idct_add           = bfin_idct_add;
305        c->idct_put           = bfin_idct_put;
306    }
307}
308
309
310
311