1 /*
2  * VP8 compatible video decoder
3  *
4  * Copyright (C) 2010 David Conrad
5  *
6  * This file is part of FFmpeg.
7  *
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.
12  *
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.
17  *
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
21  */
22 
23 #include "config.h"
24 
25 #include "libavutil/cpu.h"
26 #include "libavutil/mem.h"
27 #include "libavutil/ppc/cpu.h"
28 #include "libavutil/ppc/util_altivec.h"
29 
30 #include "libavcodec/vp8dsp.h"
31 
32 #include "hpeldsp_altivec.h"
33 
34 #if HAVE_ALTIVEC
35 #define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
36 
37 // h subpel filter uses msum to multiply+add 4 pixel taps at once
38 static const vec_s8 h_subpel_filters_inner[7] =
39 {
40     REPT4( -6, 123,  12,  -1),
41     REPT4(-11, 108,  36,  -8),
42     REPT4( -9,  93,  50,  -6),
43     REPT4(-16,  77,  77, -16),
44     REPT4( -6,  50,  93,  -9),
45     REPT4( -8,  36, 108, -11),
46     REPT4( -1,  12, 123,  -6),
47 };
48 
49 // for 6tap filters, these are the outer two taps
50 // The zeros mask off pixels 4-7 when filtering 0-3
51 // and vice-versa
52 static const vec_s8 h_subpel_filters_outer[3] =
53 {
54     REPT4(0, 0, 2, 1),
55     REPT4(0, 0, 3, 3),
56     REPT4(0, 0, 1, 2),
57 };
58 
59 #define LOAD_H_SUBPEL_FILTER(i) \
60     vec_s8 filter_inner  = h_subpel_filters_inner[i]; \
61     vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
62     vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
63 
64 #if HAVE_BIGENDIAN
65 #define GET_PIXHL(offset)                   \
66     a = vec_ld((offset)-is6tap-1, src);     \
67     b = vec_ld((offset)-is6tap-1+15, src);  \
68     pixh  = vec_perm(a, b, permh##offset);  \
69     pixl  = vec_perm(a, b, perml##offset)
70 
71 #define GET_OUTER(offset) outer = vec_perm(a, b, perm_6tap##offset)
72 #else
73 #define GET_PIXHL(offset)                   \
74     a = vec_vsx_ld((offset)-is6tap-1, src); \
75     pixh  = vec_perm(a, a, perm_inner);     \
76     pixl  = vec_perm(a, a, vec_add(perm_inner, vec_splat_u8(4)))
77 
78 #define GET_OUTER(offset) outer = vec_perm(a, a, perm_outer)
79 #endif
80 
81 #define FILTER_H(dstv, off) \
82     GET_PIXHL(off);                            \
83     filth = vec_msum(filter_inner, pixh, c64); \
84     filtl = vec_msum(filter_inner, pixl, c64); \
85 \
86     if (is6tap) { \
87         GET_OUTER(off);                                \
88         filth = vec_msum(filter_outerh, outer, filth); \
89         filtl = vec_msum(filter_outerl, outer, filtl); \
90     } \
91     if (w == 4) \
92         filtl = filth; /* discard pixels 4-7 */ \
93     dstv = vec_packs(filth, filtl); \
94     dstv = vec_sra(dstv, c7)
95 
96 static av_always_inline
put_vp8_epel_h_altivec_core(uint8_t * dst,ptrdiff_t dst_stride,uint8_t * src,ptrdiff_t src_stride,int h,int mx,int w,int is6tap)97 void put_vp8_epel_h_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
98                                  uint8_t *src, ptrdiff_t src_stride,
99                                  int h, int mx, int w, int is6tap)
100 {
101     LOAD_H_SUBPEL_FILTER(mx-1);
102 #if HAVE_BIGENDIAN
103     vec_u8 align_vec0, align_vec8, permh0, permh8;
104     vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
105     vec_u8 b;
106 #endif
107     vec_u8 filt, a, pixh, pixl, outer;
108     vec_s16 f16h, f16l;
109     vec_s32 filth, filtl;
110 
111     vec_u8 perm_inner6 = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
112     vec_u8 perm_inner4 = { 0,1,2,3, 1,2,3,4, 2,3,4,5, 3,4,5,6 };
113     vec_u8 perm_inner  = is6tap ? perm_inner6 : perm_inner4;
114     vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
115     vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
116     vec_u16 c7  = vec_splat_u16(7);
117 
118 #if HAVE_BIGENDIAN
119     align_vec0 = vec_lvsl( -is6tap-1, src);
120     align_vec8 = vec_lvsl(8-is6tap-1, src);
121 
122     permh0     = vec_perm(align_vec0, align_vec0, perm_inner);
123     permh8     = vec_perm(align_vec8, align_vec8, perm_inner);
124     perm_inner = vec_add(perm_inner, vec_splat_u8(4));
125     perml0     = vec_perm(align_vec0, align_vec0, perm_inner);
126     perml8     = vec_perm(align_vec8, align_vec8, perm_inner);
127     perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
128     perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
129 #endif
130 
131     while (h --> 0) {
132         FILTER_H(f16h, 0);
133 
134         if (w == 16) {
135             FILTER_H(f16l, 8);
136             filt = vec_packsu(f16h, f16l);
137             vec_st(filt, 0, dst);
138         } else {
139             filt = vec_packsu(f16h, f16h);
140             vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
141             if (w == 8)
142                 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
143         }
144         src += src_stride;
145         dst += dst_stride;
146     }
147 }
148 
149 // v subpel filter does a simple vertical multiply + add
150 static const vec_u8 v_subpel_filters[7] =
151 {
152     { 0,   6, 123,  12,   1,   0 },
153     { 2,  11, 108,  36,   8,   1 },
154     { 0,   9,  93,  50,   6,   0 },
155     { 3,  16,  77,  77,  16,   3 },
156     { 0,   6,  50,  93,   9,   0 },
157     { 1,   8,  36, 108,  11,   2 },
158     { 0,   1,  12, 123,   6,   0 },
159 };
160 
161 #define LOAD_V_SUBPEL_FILTER(i) \
162     vec_u8 subpel_filter = v_subpel_filters[i]; \
163     vec_u8 f0 = vec_splat(subpel_filter, 0); \
164     vec_u8 f1 = vec_splat(subpel_filter, 1); \
165     vec_u8 f2 = vec_splat(subpel_filter, 2); \
166     vec_u8 f3 = vec_splat(subpel_filter, 3); \
167     vec_u8 f4 = vec_splat(subpel_filter, 4); \
168     vec_u8 f5 = vec_splat(subpel_filter, 5)
169 
170 #define FILTER_V(dstv, vec_mul) \
171     s1f = (vec_s16)vec_mul(s1, f1); \
172     s2f = (vec_s16)vec_mul(s2, f2); \
173     s3f = (vec_s16)vec_mul(s3, f3); \
174     s4f = (vec_s16)vec_mul(s4, f4); \
175     s2f = vec_subs(s2f, s1f); \
176     s3f = vec_subs(s3f, s4f); \
177     if (is6tap) { \
178         s0f = (vec_s16)vec_mul(s0, f0); \
179         s5f = (vec_s16)vec_mul(s5, f5); \
180         s2f = vec_adds(s2f, s0f); \
181         s3f = vec_adds(s3f, s5f); \
182     } \
183     dstv = vec_adds(s2f, s3f); \
184     dstv = vec_adds(dstv, c64); \
185     dstv = vec_sra(dstv, c7)
186 
187 #if HAVE_BIGENDIAN
188 #define LOAD_HL(off, s, perm) load_with_perm_vec(off, s, perm)
189 #else
190 #define LOAD_HL(off, s, perm) vec_mergeh(vec_vsx_ld(off,s), vec_vsx_ld(off+8,s))
191 #endif
192 
193 static av_always_inline
put_vp8_epel_v_altivec_core(uint8_t * dst,ptrdiff_t dst_stride,uint8_t * src,ptrdiff_t src_stride,int h,int my,int w,int is6tap)194 void put_vp8_epel_v_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
195                                  uint8_t *src, ptrdiff_t src_stride,
196                                  int h, int my, int w, int is6tap)
197 {
198     LOAD_V_SUBPEL_FILTER(my-1);
199     vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
200     vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
201     vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
202     vec_u16 c7  = vec_splat_u16(7);
203 
204 #if HAVE_BIGENDIAN
205     // we want pixels 0-7 to be in the even positions and 8-15 in the odd,
206     // so combine this permute with the alignment permute vector
207     align_vech = vec_lvsl(0, src);
208     align_vecl = vec_sld(align_vech, align_vech, 8);
209     if (w ==16)
210         perm_vec = vec_mergeh(align_vech, align_vecl);
211     else
212         perm_vec = vec_mergeh(align_vech, align_vech);
213 #endif
214 
215     if (is6tap)
216         s0 = LOAD_HL(-2*src_stride, src, perm_vec);
217     s1 = LOAD_HL(-1*src_stride, src, perm_vec);
218     s2 = LOAD_HL( 0*src_stride, src, perm_vec);
219     s3 = LOAD_HL( 1*src_stride, src, perm_vec);
220     if (is6tap)
221         s4 = LOAD_HL( 2*src_stride, src, perm_vec);
222 
223     src += (2+is6tap)*src_stride;
224 
225     while (h --> 0) {
226         if (is6tap)
227             s5 = LOAD_HL(0, src, perm_vec);
228         else
229             s4 = LOAD_HL(0, src, perm_vec);
230 
231         FILTER_V(f16h, vec_mule);
232 
233         if (w == 16) {
234             FILTER_V(f16l, vec_mulo);
235             filt = vec_packsu(f16h, f16l);
236             vec_st(filt, 0, dst);
237         } else {
238             filt = vec_packsu(f16h, f16h);
239             if (w == 4)
240                 filt = (vec_u8)vec_splat((vec_u32)filt, 0);
241             else
242                 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
243             vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
244         }
245 
246         if (is6tap)
247             s0 = s1;
248         s1 = s2;
249         s2 = s3;
250         s3 = s4;
251         if (is6tap)
252             s4 = s5;
253 
254         dst += dst_stride;
255         src += src_stride;
256     }
257 }
258 
259 #define EPEL_FUNCS(WIDTH, TAPS) \
260 static av_noinline \
261 void put_vp8_epel ## WIDTH ## _h ## TAPS ## _altivec(uint8_t *dst, ptrdiff_t dst_stride, uint8_t *src, ptrdiff_t src_stride, int h, int mx, int my) \
262 { \
263     put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
264 } \
265 \
266 static av_noinline \
267 void put_vp8_epel ## WIDTH ## _v ## TAPS ## _altivec(uint8_t *dst, ptrdiff_t dst_stride, uint8_t *src, ptrdiff_t src_stride, int h, int mx, int my) \
268 { \
269     put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
270 }
271 
272 #define EPEL_HV(WIDTH, HTAPS, VTAPS) \
273 static void put_vp8_epel ## WIDTH ## _h ## HTAPS ## v ## VTAPS ## _altivec(uint8_t *dst, ptrdiff_t dstride, uint8_t *src, ptrdiff_t sstride, int h, int mx, int my) \
274 { \
275     DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
276     if (VTAPS == 6) { \
277         put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,      src-2*sstride, sstride, h+5, mx, my); \
278         put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+2*16,      16,      h,   mx, my); \
279     } else { \
280         put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,      src-sstride, sstride, h+4, mx, my); \
281         put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+16,      16,      h,   mx, my); \
282     } \
283 }
284 
285 EPEL_FUNCS(16,6)
286 EPEL_FUNCS(8, 6)
287 EPEL_FUNCS(8, 4)
288 EPEL_FUNCS(4, 6)
289 EPEL_FUNCS(4, 4)
290 
291 EPEL_HV(16, 6,6)
292 EPEL_HV(8,  6,6)
293 EPEL_HV(8,  4,6)
294 EPEL_HV(8,  6,4)
295 EPEL_HV(8,  4,4)
296 EPEL_HV(4,  6,6)
297 EPEL_HV(4,  4,6)
298 EPEL_HV(4,  6,4)
299 EPEL_HV(4,  4,4)
300 
put_vp8_pixels16_altivec(uint8_t * dst,ptrdiff_t dstride,uint8_t * src,ptrdiff_t sstride,int h,int mx,int my)301 static void put_vp8_pixels16_altivec(uint8_t *dst, ptrdiff_t dstride, uint8_t *src, ptrdiff_t sstride, int h, int mx, int my)
302 {
303     register vector unsigned char perm;
304     int i;
305     register ptrdiff_t dstride2 = dstride << 1, sstride2 = sstride << 1;
306     register ptrdiff_t dstride3 = dstride2 + dstride, sstride3 = sstride + sstride2;
307     register ptrdiff_t dstride4 = dstride << 2, sstride4 = sstride << 2;
308 
309 #if HAVE_BIGENDIAN
310     perm = vec_lvsl(0, src);
311 #endif
312 // hand-unrolling the loop by 4 gains about 15%
313 // mininum execution time goes from 74 to 60 cycles
314 // it's faster than -funroll-loops, but using
315 // -funroll-loops w/ this is bad - 74 cycles again.
316 // all this is on a 7450, tuning for the 7450
317     for (i = 0; i < h; i += 4) {
318         vec_st(load_with_perm_vec(0, src, perm), 0, dst);
319         vec_st(load_with_perm_vec(sstride, src, perm), dstride, dst);
320         vec_st(load_with_perm_vec(sstride2, src, perm), dstride2, dst);
321         vec_st(load_with_perm_vec(sstride3, src, perm), dstride3, dst);
322         src += sstride4;
323         dst += dstride4;
324     }
325 }
326 
327 #endif /* HAVE_ALTIVEC */
328 
329 
ff_vp78dsp_init_ppc(VP8DSPContext * c)330 av_cold void ff_vp78dsp_init_ppc(VP8DSPContext *c)
331 {
332 #if HAVE_ALTIVEC
333     if (!PPC_ALTIVEC(av_get_cpu_flags()))
334         return;
335 
336     c->put_vp8_epel_pixels_tab[0][0][0] = put_vp8_pixels16_altivec;
337     c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
338     c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
339     c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
340 
341     c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
342     c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
343     c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
344     c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
345 
346     c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
347     c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
348     c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
349     c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
350 
351     c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
352     c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
353     c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
354     c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
355 
356     c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
357     c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
358     c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
359     c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
360 #endif /* HAVE_ALTIVEC */
361 }
362