Merge remote-tracking branch 'qatar/master'
[ffmpeg.git] / libavcodec / alpha / dsputil_alpha.c
1 /*
2  * Alpha optimized DSP utils
3  * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 #include "libavcodec/dsputil.h"
23 #include "dsputil_alpha.h"
24 #include "asm.h"
25
26 void (*put_pixels_clamped_axp_p)(const DCTELEM *block, uint8_t *pixels,
27                                  int line_size);
28 void (*add_pixels_clamped_axp_p)(const DCTELEM *block, uint8_t *pixels,
29                                  int line_size);
30
31 #if 0
32 /* These functions were the base for the optimized assembler routines,
33    and remain here for documentation purposes.  */
34 static void put_pixels_clamped_mvi(const DCTELEM *block, uint8_t *pixels,
35                                    int line_size)
36 {
37     int i = 8;
38     uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
39
40     do {
41         uint64_t shorts0, shorts1;
42
43         shorts0 = ldq(block);
44         shorts0 = maxsw4(shorts0, 0);
45         shorts0 = minsw4(shorts0, clampmask);
46         stl(pkwb(shorts0), pixels);
47
48         shorts1 = ldq(block + 4);
49         shorts1 = maxsw4(shorts1, 0);
50         shorts1 = minsw4(shorts1, clampmask);
51         stl(pkwb(shorts1), pixels + 4);
52
53         pixels += line_size;
54         block += 8;
55     } while (--i);
56 }
57
58 void add_pixels_clamped_mvi(const DCTELEM *block, uint8_t *pixels,
59                             int line_size)
60 {
61     int h = 8;
62     /* Keep this function a leaf function by generating the constants
63        manually (mainly for the hack value ;-).  */
64     uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
65     uint64_t signmask  = zap(-1, 0x33);
66     signmask ^= signmask >> 1;  /* 0x8000800080008000 */
67
68     do {
69         uint64_t shorts0, pix0, signs0;
70         uint64_t shorts1, pix1, signs1;
71
72         shorts0 = ldq(block);
73         shorts1 = ldq(block + 4);
74
75         pix0    = unpkbw(ldl(pixels));
76         /* Signed subword add (MMX paddw).  */
77         signs0  = shorts0 & signmask;
78         shorts0 &= ~signmask;
79         shorts0 += pix0;
80         shorts0 ^= signs0;
81         /* Clamp. */
82         shorts0 = maxsw4(shorts0, 0);
83         shorts0 = minsw4(shorts0, clampmask);
84
85         /* Next 4.  */
86         pix1    = unpkbw(ldl(pixels + 4));
87         signs1  = shorts1 & signmask;
88         shorts1 &= ~signmask;
89         shorts1 += pix1;
90         shorts1 ^= signs1;
91         shorts1 = maxsw4(shorts1, 0);
92         shorts1 = minsw4(shorts1, clampmask);
93
94         stl(pkwb(shorts0), pixels);
95         stl(pkwb(shorts1), pixels + 4);
96
97         pixels += line_size;
98         block += 8;
99     } while (--h);
100 }
101 #endif
102
103 static void clear_blocks_axp(DCTELEM *blocks) {
104     uint64_t *p = (uint64_t *) blocks;
105     int n = sizeof(DCTELEM) * 6 * 64;
106
107     do {
108         p[0] = 0;
109         p[1] = 0;
110         p[2] = 0;
111         p[3] = 0;
112         p[4] = 0;
113         p[5] = 0;
114         p[6] = 0;
115         p[7] = 0;
116         p += 8;
117         n -= 8 * 8;
118     } while (n);
119 }
120
121 static inline uint64_t avg2_no_rnd(uint64_t a, uint64_t b)
122 {
123     return (a & b) + (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
124 }
125
126 static inline uint64_t avg2(uint64_t a, uint64_t b)
127 {
128     return (a | b) - (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
129 }
130
131 #if 0
132 /* The XY2 routines basically utilize this scheme, but reuse parts in
133    each iteration.  */
134 static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4)
135 {
136     uint64_t r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
137                 + ((l2 & ~BYTE_VEC(0x03)) >> 2)
138                 + ((l3 & ~BYTE_VEC(0x03)) >> 2)
139                 + ((l4 & ~BYTE_VEC(0x03)) >> 2);
140     uint64_t r2 = ((  (l1 & BYTE_VEC(0x03))
141                     + (l2 & BYTE_VEC(0x03))
142                     + (l3 & BYTE_VEC(0x03))
143                     + (l4 & BYTE_VEC(0x03))
144                     + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
145     return r1 + r2;
146 }
147 #endif
148
149 #define OP(LOAD, STORE)                         \
150     do {                                        \
151         STORE(LOAD(pixels), block);             \
152         pixels += line_size;                    \
153         block += line_size;                     \
154     } while (--h)
155
156 #define OP_X2(LOAD, STORE)                                      \
157     do {                                                        \
158         uint64_t pix1, pix2;                                    \
159                                                                 \
160         pix1 = LOAD(pixels);                                    \
161         pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56);        \
162         STORE(AVG2(pix1, pix2), block);                         \
163         pixels += line_size;                                    \
164         block += line_size;                                     \
165     } while (--h)
166
167 #define OP_Y2(LOAD, STORE)                      \
168     do {                                        \
169         uint64_t pix = LOAD(pixels);            \
170         do {                                    \
171             uint64_t next_pix;                  \
172                                                 \
173             pixels += line_size;                \
174             next_pix = LOAD(pixels);            \
175             STORE(AVG2(pix, next_pix), block);  \
176             block += line_size;                 \
177             pix = next_pix;                     \
178         } while (--h);                          \
179     } while (0)
180
181 #define OP_XY2(LOAD, STORE)                                                 \
182     do {                                                                    \
183         uint64_t pix1 = LOAD(pixels);                                       \
184         uint64_t pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56);           \
185         uint64_t pix_l = (pix1 & BYTE_VEC(0x03))                            \
186                        + (pix2 & BYTE_VEC(0x03));                           \
187         uint64_t pix_h = ((pix1 & ~BYTE_VEC(0x03)) >> 2)                    \
188                        + ((pix2 & ~BYTE_VEC(0x03)) >> 2);                   \
189                                                                             \
190         do {                                                                \
191             uint64_t npix1, npix2;                                          \
192             uint64_t npix_l, npix_h;                                        \
193             uint64_t avg;                                                   \
194                                                                             \
195             pixels += line_size;                                            \
196             npix1 = LOAD(pixels);                                           \
197             npix2 = npix1 >> 8 | ((uint64_t) pixels[8] << 56);              \
198             npix_l = (npix1 & BYTE_VEC(0x03))                               \
199                    + (npix2 & BYTE_VEC(0x03));                              \
200             npix_h = ((npix1 & ~BYTE_VEC(0x03)) >> 2)                       \
201                    + ((npix2 & ~BYTE_VEC(0x03)) >> 2);                      \
202             avg = (((pix_l + npix_l + AVG4_ROUNDER) >> 2) & BYTE_VEC(0x03)) \
203                 + pix_h + npix_h;                                           \
204             STORE(avg, block);                                              \
205                                                                             \
206             block += line_size;                                             \
207             pix_l = npix_l;                                                 \
208             pix_h = npix_h;                                                 \
209         } while (--h);                                                      \
210     } while (0)
211
212 #define MAKE_OP(OPNAME, SUFF, OPKIND, STORE)                                \
213 static void OPNAME ## _pixels ## SUFF ## _axp                               \
214         (uint8_t *restrict block, const uint8_t *restrict pixels,           \
215          int line_size, int h)                                              \
216 {                                                                           \
217     if ((size_t) pixels & 0x7) {                                            \
218         OPKIND(uldq, STORE);                                                \
219     } else {                                                                \
220         OPKIND(ldq, STORE);                                                 \
221     }                                                                       \
222 }                                                                           \
223                                                                             \
224 static void OPNAME ## _pixels16 ## SUFF ## _axp                             \
225         (uint8_t *restrict block, const uint8_t *restrict pixels,           \
226          int line_size, int h)                                              \
227 {                                                                           \
228     OPNAME ## _pixels ## SUFF ## _axp(block,     pixels,     line_size, h); \
229     OPNAME ## _pixels ## SUFF ## _axp(block + 8, pixels + 8, line_size, h); \
230 }
231
232 #define PIXOP(OPNAME, STORE)                    \
233     MAKE_OP(OPNAME, ,     OP,     STORE)        \
234     MAKE_OP(OPNAME, _x2,  OP_X2,  STORE)        \
235     MAKE_OP(OPNAME, _y2,  OP_Y2,  STORE)        \
236     MAKE_OP(OPNAME, _xy2, OP_XY2, STORE)
237
238 /* Rounding primitives.  */
239 #define AVG2 avg2
240 #define AVG4 avg4
241 #define AVG4_ROUNDER BYTE_VEC(0x02)
242 #define STORE(l, b) stq(l, b)
243 PIXOP(put, STORE);
244
245 #undef STORE
246 #define STORE(l, b) stq(AVG2(l, ldq(b)), b);
247 PIXOP(avg, STORE);
248
249 /* Not rounding primitives.  */
250 #undef AVG2
251 #undef AVG4
252 #undef AVG4_ROUNDER
253 #undef STORE
254 #define AVG2 avg2_no_rnd
255 #define AVG4 avg4_no_rnd
256 #define AVG4_ROUNDER BYTE_VEC(0x01)
257 #define STORE(l, b) stq(l, b)
258 PIXOP(put_no_rnd, STORE);
259
260 #undef STORE
261 #define STORE(l, b) stq(AVG2(l, ldq(b)), b);
262 PIXOP(avg_no_rnd, STORE);
263
264 static void put_pixels16_axp_asm(uint8_t *block, const uint8_t *pixels,
265                                  int line_size, int h)
266 {
267     put_pixels_axp_asm(block,     pixels,     line_size, h);
268     put_pixels_axp_asm(block + 8, pixels + 8, line_size, h);
269 }
270
271 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx)
272 {
273     const int high_bit_depth = avctx->codec_id == CODEC_ID_H264 && avctx->bits_per_raw_sample > 8;
274
275     if (!high_bit_depth) {
276     c->put_pixels_tab[0][0] = put_pixels16_axp_asm;
277     c->put_pixels_tab[0][1] = put_pixels16_x2_axp;
278     c->put_pixels_tab[0][2] = put_pixels16_y2_axp;
279     c->put_pixels_tab[0][3] = put_pixels16_xy2_axp;
280
281     c->put_no_rnd_pixels_tab[0][0] = put_pixels16_axp_asm;
282     c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_axp;
283     c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_axp;
284     c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_axp;
285
286     c->avg_pixels_tab[0][0] = avg_pixels16_axp;
287     c->avg_pixels_tab[0][1] = avg_pixels16_x2_axp;
288     c->avg_pixels_tab[0][2] = avg_pixels16_y2_axp;
289     c->avg_pixels_tab[0][3] = avg_pixels16_xy2_axp;
290
291     c->avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_axp;
292     c->avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_axp;
293     c->avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_axp;
294     c->avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_axp;
295
296     c->put_pixels_tab[1][0] = put_pixels_axp_asm;
297     c->put_pixels_tab[1][1] = put_pixels_x2_axp;
298     c->put_pixels_tab[1][2] = put_pixels_y2_axp;
299     c->put_pixels_tab[1][3] = put_pixels_xy2_axp;
300
301     c->put_no_rnd_pixels_tab[1][0] = put_pixels_axp_asm;
302     c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels_x2_axp;
303     c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels_y2_axp;
304     c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels_xy2_axp;
305
306     c->avg_pixels_tab[1][0] = avg_pixels_axp;
307     c->avg_pixels_tab[1][1] = avg_pixels_x2_axp;
308     c->avg_pixels_tab[1][2] = avg_pixels_y2_axp;
309     c->avg_pixels_tab[1][3] = avg_pixels_xy2_axp;
310
311     c->avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels_axp;
312     c->avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels_x2_axp;
313     c->avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels_y2_axp;
314     c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels_xy2_axp;
315
316     c->clear_blocks = clear_blocks_axp;
317     }
318
319     /* amask clears all bits that correspond to present features.  */
320     if (amask(AMASK_MVI) == 0) {
321         c->put_pixels_clamped = put_pixels_clamped_mvi_asm;
322         c->add_pixels_clamped = add_pixels_clamped_mvi_asm;
323
324         c->get_pixels       = get_pixels_mvi;
325         c->diff_pixels      = diff_pixels_mvi;
326         c->sad[0]           = pix_abs16x16_mvi_asm;
327         c->sad[1]           = pix_abs8x8_mvi;
328         c->pix_abs[0][0]    = pix_abs16x16_mvi_asm;
329         c->pix_abs[1][0]    = pix_abs8x8_mvi;
330         c->pix_abs[0][1]    = pix_abs16x16_x2_mvi;
331         c->pix_abs[0][2]    = pix_abs16x16_y2_mvi;
332         c->pix_abs[0][3]    = pix_abs16x16_xy2_mvi;
333     }
334
335     put_pixels_clamped_axp_p = c->put_pixels_clamped;
336     add_pixels_clamped_axp_p = c->add_pixels_clamped;
337
338     if (!avctx->lowres && avctx->bits_per_raw_sample <= 8 &&
339         (avctx->idct_algo == FF_IDCT_AUTO ||
340          avctx->idct_algo == FF_IDCT_SIMPLEALPHA)) {
341         c->idct_put = ff_simple_idct_put_axp;
342         c->idct_add = ff_simple_idct_add_axp;
343         c->idct =     ff_simple_idct_axp;
344     }
345 }