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/attributes.h"
26#include "libavutil/cpu.h"
27#include "libavutil/mem_internal.h"
28#include "libavutil/ppc/cpu.h"
29#include "libavutil/ppc/util_altivec.h"
30
31#include "libavcodec/vp8dsp.h"
32
33#include "hpeldsp_altivec.h"
34
35#if HAVE_ALTIVEC
36#define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
37
38// h subpel filter uses msum to multiply+add 4 pixel taps at once
39static const vec_s8 h_subpel_filters_inner[7] =
40{
41    REPT4( -6, 123,  12,  -1),
42    REPT4(-11, 108,  36,  -8),
43    REPT4( -9,  93,  50,  -6),
44    REPT4(-16,  77,  77, -16),
45    REPT4( -6,  50,  93,  -9),
46    REPT4( -8,  36, 108, -11),
47    REPT4( -1,  12, 123,  -6),
48};
49
50// for 6tap filters, these are the outer two taps
51// The zeros mask off pixels 4-7 when filtering 0-3
52// and vice-versa
53static const vec_s8 h_subpel_filters_outer[3] =
54{
55    REPT4(0, 0, 2, 1),
56    REPT4(0, 0, 3, 3),
57    REPT4(0, 0, 1, 2),
58};
59
60#define LOAD_H_SUBPEL_FILTER(i) \
61    vec_s8 filter_inner  = h_subpel_filters_inner[i]; \
62    vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
63    vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
64
65#if HAVE_BIGENDIAN
66#define GET_PIXHL(offset)                   \
67    a = vec_ld((offset)-is6tap-1, src);     \
68    b = vec_ld((offset)-is6tap-1+15, src);  \
69    pixh  = vec_perm(a, b, permh##offset);  \
70    pixl  = vec_perm(a, b, perml##offset)
71
72#define GET_OUTER(offset) outer = vec_perm(a, b, perm_6tap##offset)
73#else
74#define GET_PIXHL(offset)                   \
75    a = vec_vsx_ld((offset)-is6tap-1, src); \
76    pixh  = vec_perm(a, a, perm_inner);     \
77    pixl  = vec_perm(a, a, vec_add(perm_inner, vec_splat_u8(4)))
78
79#define GET_OUTER(offset) outer = vec_perm(a, a, perm_outer)
80#endif
81
82#define FILTER_H(dstv, off) \
83    GET_PIXHL(off);                            \
84    filth = vec_msum(filter_inner, pixh, c64); \
85    filtl = vec_msum(filter_inner, pixl, c64); \
86\
87    if (is6tap) { \
88        GET_OUTER(off);                                \
89        filth = vec_msum(filter_outerh, outer, filth); \
90        filtl = vec_msum(filter_outerl, outer, filtl); \
91    } \
92    if (w == 4) \
93        filtl = filth; /* discard pixels 4-7 */ \
94    dstv = vec_packs(filth, filtl); \
95    dstv = vec_sra(dstv, c7)
96
97static av_always_inline
98void put_vp8_epel_h_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
99                                 uint8_t *src, ptrdiff_t src_stride,
100                                 int h, int mx, int w, int is6tap)
101{
102    LOAD_H_SUBPEL_FILTER(mx-1);
103#if HAVE_BIGENDIAN
104    vec_u8 align_vec0, align_vec8, permh0, permh8;
105    vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
106    vec_u8 b;
107#endif
108    vec_u8 filt, a, pixh, pixl, outer;
109    vec_s16 f16h, f16l;
110    vec_s32 filth, filtl;
111
112    vec_u8 perm_inner6 = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
113    vec_u8 perm_inner4 = { 0,1,2,3, 1,2,3,4, 2,3,4,5, 3,4,5,6 };
114    vec_u8 perm_inner  = is6tap ? perm_inner6 : perm_inner4;
115    vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
116    vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
117    vec_u16 c7  = vec_splat_u16(7);
118
119#if HAVE_BIGENDIAN
120    align_vec0 = vec_lvsl( -is6tap-1, src);
121    align_vec8 = vec_lvsl(8-is6tap-1, src);
122
123    permh0     = vec_perm(align_vec0, align_vec0, perm_inner);
124    permh8     = vec_perm(align_vec8, align_vec8, perm_inner);
125    perm_inner = vec_add(perm_inner, vec_splat_u8(4));
126    perml0     = vec_perm(align_vec0, align_vec0, perm_inner);
127    perml8     = vec_perm(align_vec8, align_vec8, perm_inner);
128    perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
129    perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
130#endif
131
132    while (h --> 0) {
133        FILTER_H(f16h, 0);
134
135        if (w == 16) {
136            FILTER_H(f16l, 8);
137            filt = vec_packsu(f16h, f16l);
138            vec_st(filt, 0, dst);
139        } else {
140            filt = vec_packsu(f16h, f16h);
141            vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
142            if (w == 8)
143                vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
144        }
145        src += src_stride;
146        dst += dst_stride;
147    }
148}
149
150// v subpel filter does a simple vertical multiply + add
151static const vec_u8 v_subpel_filters[7] =
152{
153    { 0,   6, 123,  12,   1,   0 },
154    { 2,  11, 108,  36,   8,   1 },
155    { 0,   9,  93,  50,   6,   0 },
156    { 3,  16,  77,  77,  16,   3 },
157    { 0,   6,  50,  93,   9,   0 },
158    { 1,   8,  36, 108,  11,   2 },
159    { 0,   1,  12, 123,   6,   0 },
160};
161
162#define LOAD_V_SUBPEL_FILTER(i) \
163    vec_u8 subpel_filter = v_subpel_filters[i]; \
164    vec_u8 f0 = vec_splat(subpel_filter, 0); \
165    vec_u8 f1 = vec_splat(subpel_filter, 1); \
166    vec_u8 f2 = vec_splat(subpel_filter, 2); \
167    vec_u8 f3 = vec_splat(subpel_filter, 3); \
168    vec_u8 f4 = vec_splat(subpel_filter, 4); \
169    vec_u8 f5 = vec_splat(subpel_filter, 5)
170
171#define FILTER_V(dstv, vec_mul) \
172    s1f = (vec_s16)vec_mul(s1, f1); \
173    s2f = (vec_s16)vec_mul(s2, f2); \
174    s3f = (vec_s16)vec_mul(s3, f3); \
175    s4f = (vec_s16)vec_mul(s4, f4); \
176    s2f = vec_subs(s2f, s1f); \
177    s3f = vec_subs(s3f, s4f); \
178    if (is6tap) { \
179        s0f = (vec_s16)vec_mul(s0, f0); \
180        s5f = (vec_s16)vec_mul(s5, f5); \
181        s2f = vec_adds(s2f, s0f); \
182        s3f = vec_adds(s3f, s5f); \
183    } \
184    dstv = vec_adds(s2f, s3f); \
185    dstv = vec_adds(dstv, c64); \
186    dstv = vec_sra(dstv, c7)
187
188#if HAVE_BIGENDIAN
189#define LOAD_HL(off, s, perm) load_with_perm_vec(off, s, perm)
190#else
191#define LOAD_HL(off, s, perm) vec_mergeh(vec_vsx_ld(off,s), vec_vsx_ld(off+8,s))
192#endif
193
194static av_always_inline
195void put_vp8_epel_v_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
196                                 uint8_t *src, ptrdiff_t src_stride,
197                                 int h, int my, int w, int is6tap)
198{
199    LOAD_V_SUBPEL_FILTER(my-1);
200    vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
201    vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
202    vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
203    vec_u16 c7  = vec_splat_u16(7);
204
205#if HAVE_BIGENDIAN
206    // we want pixels 0-7 to be in the even positions and 8-15 in the odd,
207    // so combine this permute with the alignment permute vector
208    align_vech = vec_lvsl(0, src);
209    align_vecl = vec_sld(align_vech, align_vech, 8);
210    if (w ==16)
211        perm_vec = vec_mergeh(align_vech, align_vecl);
212    else
213        perm_vec = vec_mergeh(align_vech, align_vech);
214#endif
215
216    if (is6tap)
217        s0 = LOAD_HL(-2*src_stride, src, perm_vec);
218    s1 = LOAD_HL(-1*src_stride, src, perm_vec);
219    s2 = LOAD_HL( 0*src_stride, src, perm_vec);
220    s3 = LOAD_HL( 1*src_stride, src, perm_vec);
221    if (is6tap)
222        s4 = LOAD_HL( 2*src_stride, src, perm_vec);
223
224    src += (2+is6tap)*src_stride;
225
226    while (h --> 0) {
227        if (is6tap)
228            s5 = LOAD_HL(0, src, perm_vec);
229        else
230            s4 = LOAD_HL(0, src, perm_vec);
231
232        FILTER_V(f16h, vec_mule);
233
234        if (w == 16) {
235            FILTER_V(f16l, vec_mulo);
236            filt = vec_packsu(f16h, f16l);
237            vec_st(filt, 0, dst);
238        } else {
239            filt = vec_packsu(f16h, f16h);
240            if (w == 4)
241                filt = (vec_u8)vec_splat((vec_u32)filt, 0);
242            else
243                vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
244            vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
245        }
246
247        if (is6tap)
248            s0 = s1;
249        s1 = s2;
250        s2 = s3;
251        s3 = s4;
252        if (is6tap)
253            s4 = s5;
254
255        dst += dst_stride;
256        src += src_stride;
257    }
258}
259
260#define EPEL_FUNCS(WIDTH, TAPS) \
261static av_noinline \
262void 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) \
263{ \
264    put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
265} \
266\
267static av_noinline \
268void 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) \
269{ \
270    put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
271}
272
273#define EPEL_HV(WIDTH, HTAPS, VTAPS) \
274static 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) \
275{ \
276    DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
277    if (VTAPS == 6) { \
278        put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,      src-2*sstride, sstride, h+5, mx, my); \
279        put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+2*16,      16,      h,   mx, my); \
280    } else { \
281        put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,      src-sstride, sstride, h+4, mx, my); \
282        put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+16,      16,      h,   mx, my); \
283    } \
284}
285
286EPEL_FUNCS(16,6)
287EPEL_FUNCS(8, 6)
288EPEL_FUNCS(8, 4)
289EPEL_FUNCS(4, 6)
290EPEL_FUNCS(4, 4)
291
292EPEL_HV(16, 6,6)
293EPEL_HV(8,  6,6)
294EPEL_HV(8,  4,6)
295EPEL_HV(8,  6,4)
296EPEL_HV(8,  4,4)
297EPEL_HV(4,  6,6)
298EPEL_HV(4,  4,6)
299EPEL_HV(4,  6,4)
300EPEL_HV(4,  4,4)
301
302static void put_vp8_pixels16_altivec(uint8_t *dst, ptrdiff_t dstride, uint8_t *src, ptrdiff_t sstride, int h, int mx, int my)
303{
304    register vector unsigned char perm;
305    int i;
306    register ptrdiff_t dstride2 = dstride << 1, sstride2 = sstride << 1;
307    register ptrdiff_t dstride3 = dstride2 + dstride, sstride3 = sstride + sstride2;
308    register ptrdiff_t dstride4 = dstride << 2, sstride4 = sstride << 2;
309
310#if HAVE_BIGENDIAN
311    perm = vec_lvsl(0, src);
312#endif
313// hand-unrolling the loop by 4 gains about 15%
314// mininum execution time goes from 74 to 60 cycles
315// it's faster than -funroll-loops, but using
316// -funroll-loops w/ this is bad - 74 cycles again.
317// all this is on a 7450, tuning for the 7450
318    for (i = 0; i < h; i += 4) {
319        vec_st(load_with_perm_vec(0, src, perm), 0, dst);
320        vec_st(load_with_perm_vec(sstride, src, perm), dstride, dst);
321        vec_st(load_with_perm_vec(sstride2, src, perm), dstride2, dst);
322        vec_st(load_with_perm_vec(sstride3, src, perm), dstride3, dst);
323        src += sstride4;
324        dst += dstride4;
325    }
326}
327
328#endif /* HAVE_ALTIVEC */
329
330
331av_cold void ff_vp78dsp_init_ppc(VP8DSPContext *c)
332{
333#if HAVE_ALTIVEC
334    if (!PPC_ALTIVEC(av_get_cpu_flags()))
335        return;
336
337    c->put_vp8_epel_pixels_tab[0][0][0] = put_vp8_pixels16_altivec;
338    c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
339    c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
340    c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
341
342    c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
343    c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
344    c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
345    c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
346
347    c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
348    c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
349    c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
350    c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
351
352    c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
353    c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
354    c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
355    c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
356
357    c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
358    c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
359    c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
360    c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
361#endif /* HAVE_ALTIVEC */
362}
363