libavcodec/ppc/vp8dsp_altivec.c

   1 /*
   2  * VP8 compatible video decoder
   3  *
   4  * Copyright (C) 2010 David Conrad
   5  *
   6  * This file is part of Libav.
   7  *
   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.
  12  *
  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.
  17  *
  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
  21  */
  22
  23 #include "libavutil/cpu.h"
  24 #include "libavutil/mem.h"
  25 #include "libavutil/ppc/types_altivec.h"
  26 #include "libavutil/ppc/util_altivec.h"
  27 #include "libavcodec/vp8dsp.h"
  28 #include "dsputil_altivec.h"
  29
  30 #define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
  31
  32 // h subpel filter uses msum to multiply+add 4 pixel taps at once
  33 static const vec_s8 h_subpel_filters_inner[7] =
  34 {
  35     REPT4( -6, 123,  12,  -1),
  36     REPT4(-11, 108,  36,  -8),
  37     REPT4( -9,  93,  50,  -6),
  38     REPT4(-16,  77,  77, -16),
  39     REPT4( -6,  50,  93,  -9),
  40     REPT4( -8,  36, 108, -11),
  41     REPT4( -1,  12, 123,  -6),
  42 };
  43
  44 // for 6tap filters, these are the outer two taps
  45 // The zeros mask off pixels 4-7 when filtering 0-3
  46 // and vice-versa
  47 static const vec_s8 h_subpel_filters_outer[3] =
  48 {
  49     REPT4(0, 0, 2, 1),
  50     REPT4(0, 0, 3, 3),
  51     REPT4(0, 0, 1, 2),
  52 };
  53
  54 #define LOAD_H_SUBPEL_FILTER(i) \
  55     vec_s8 filter_inner  = h_subpel_filters_inner[i]; \
  56     vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
  57     vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
  58
  59 #define FILTER_H(dstv, off) \
  60     a = vec_ld((off)-is6tap-1,    src); \
  61     b = vec_ld((off)-is6tap-1+15, src); \
  62 \
  63     pixh  = vec_perm(a, b, permh##off); \
  64     pixl  = vec_perm(a, b, perml##off); \
  65     filth = vec_msum(filter_inner, pixh, c64); \
  66     filtl = vec_msum(filter_inner, pixl, c64); \
  67 \
  68     if (is6tap) { \
  69         outer = vec_perm(a, b, perm_6tap##off); \
  70         filth = vec_msum(filter_outerh, outer, filth); \
  71         filtl = vec_msum(filter_outerl, outer, filtl); \
  72     } \
  73     if (w == 4) \
  74         filtl = filth; /* discard pixels 4-7 */ \
  75     dstv = vec_packs(filth, filtl); \
  76     dstv = vec_sra(dstv, c7)
  77
  78 static av_always_inline
  79 void put_vp8_epel_h_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
  80                                  uint8_t *src, ptrdiff_t src_stride,
  81                                  int h, int mx, int w, int is6tap)
  82 {
  83     LOAD_H_SUBPEL_FILTER(mx-1);
  84     vec_u8 align_vec0, align_vec8, permh0, permh8, filt;
  85     vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
  86     vec_u8 a, b, pixh, pixl, outer;
  87     vec_s16 f16h, f16l;
  88     vec_s32 filth, filtl;
  89
  90     vec_u8 perm_inner6 = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
  91     vec_u8 perm_inner4 = { 0,1,2,3, 1,2,3,4, 2,3,4,5, 3,4,5,6 };
  92     vec_u8 perm_inner  = is6tap ? perm_inner6 : perm_inner4;
  93     vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
  94     vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
  95     vec_u16 c7  = vec_splat_u16(7);
  96
  97     align_vec0 = vec_lvsl( -is6tap-1, src);
  98     align_vec8 = vec_lvsl(8-is6tap-1, src);
  99
 100     permh0     = vec_perm(align_vec0, align_vec0, perm_inner);
 101     permh8     = vec_perm(align_vec8, align_vec8, perm_inner);
 102     perm_inner = vec_add(perm_inner, vec_splat_u8(4));
 103     perml0     = vec_perm(align_vec0, align_vec0, perm_inner);
 104     perml8     = vec_perm(align_vec8, align_vec8, perm_inner);
 105     perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
 106     perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
 107
 108     while (h --> 0) {
 109         FILTER_H(f16h, 0);
 110
 111         if (w == 16) {
 112             FILTER_H(f16l, 8);
 113             filt = vec_packsu(f16h, f16l);
 114             vec_st(filt, 0, dst);
 115         } else {
 116             filt = vec_packsu(f16h, f16h);
 117             vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
 118             if (w == 8)
 119                 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
 120         }
 121         src += src_stride;
 122         dst += dst_stride;
 123     }
 124 }
 125
 126 // v subpel filter does a simple vertical multiply + add
 127 static const vec_u8 v_subpel_filters[7] =
 128 {
 129     { 0,   6, 123,  12,   1,   0 },
 130     { 2,  11, 108,  36,   8,   1 },
 131     { 0,   9,  93,  50,   6,   0 },
 132     { 3,  16,  77,  77,  16,   3 },
 133     { 0,   6,  50,  93,   9,   0 },
 134     { 1,   8,  36, 108,  11,   2 },
 135     { 0,   1,  12, 123,   6,   0 },
 136 };
 137
 138 #define LOAD_V_SUBPEL_FILTER(i) \
 139     vec_u8 subpel_filter = v_subpel_filters[i]; \
 140     vec_u8 f0 = vec_splat(subpel_filter, 0); \
 141     vec_u8 f1 = vec_splat(subpel_filter, 1); \
 142     vec_u8 f2 = vec_splat(subpel_filter, 2); \
 143     vec_u8 f3 = vec_splat(subpel_filter, 3); \
 144     vec_u8 f4 = vec_splat(subpel_filter, 4); \
 145     vec_u8 f5 = vec_splat(subpel_filter, 5)
 146
 147 #define FILTER_V(dstv, vec_mul) \
 148     s1f = (vec_s16)vec_mul(s1, f1); \
 149     s2f = (vec_s16)vec_mul(s2, f2); \
 150     s3f = (vec_s16)vec_mul(s3, f3); \
 151     s4f = (vec_s16)vec_mul(s4, f4); \
 152     s2f = vec_subs(s2f, s1f); \
 153     s3f = vec_subs(s3f, s4f); \
 154     if (is6tap) { \
 155         s0f = (vec_s16)vec_mul(s0, f0); \
 156         s5f = (vec_s16)vec_mul(s5, f5); \
 157         s2f = vec_adds(s2f, s0f); \
 158         s3f = vec_adds(s3f, s5f); \
 159     } \
 160     dstv = vec_adds(s2f, s3f); \
 161     dstv = vec_adds(dstv, c64); \
 162     dstv = vec_sra(dstv, c7)
 163
 164 static av_always_inline
 165 void put_vp8_epel_v_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
 166                                  uint8_t *src, ptrdiff_t src_stride,
 167                                  int h, int my, int w, int is6tap)
 168 {
 169     LOAD_V_SUBPEL_FILTER(my-1);
 170     vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
 171     vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
 172     vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
 173     vec_u16 c7  = vec_splat_u16(7);
 174
 175     // we want pixels 0-7 to be in the even positions and 8-15 in the odd,
 176     // so combine this permute with the alignment permute vector
 177     align_vech = vec_lvsl(0, src);
 178     align_vecl = vec_sld(align_vech, align_vech, 8);
 179     if (w ==16)
 180         perm_vec = vec_mergeh(align_vech, align_vecl);
 181     else
 182         perm_vec = vec_mergeh(align_vech, align_vech);
 183
 184     if (is6tap)
 185         s0 = load_with_perm_vec(-2*src_stride, src, perm_vec);
 186     s1 = load_with_perm_vec(-1*src_stride, src, perm_vec);
 187     s2 = load_with_perm_vec( 0*src_stride, src, perm_vec);
 188     s3 = load_with_perm_vec( 1*src_stride, src, perm_vec);
 189     if (is6tap)
 190         s4 = load_with_perm_vec( 2*src_stride, src, perm_vec);
 191
 192     src += (2+is6tap)*src_stride;
 193
 194     while (h --> 0) {
 195         if (is6tap)
 196             s5 = load_with_perm_vec(0, src, perm_vec);
 197         else
 198             s4 = load_with_perm_vec(0, src, perm_vec);
 199
 200         FILTER_V(f16h, vec_mule);
 201
 202         if (w == 16) {
 203             FILTER_V(f16l, vec_mulo);
 204             filt = vec_packsu(f16h, f16l);
 205             vec_st(filt, 0, dst);
 206         } else {
 207             filt = vec_packsu(f16h, f16h);
 208             if (w == 4)
 209                 filt = (vec_u8)vec_splat((vec_u32)filt, 0);
 210             else
 211                 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
 212             vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
 213         }
 214
 215         if (is6tap)
 216             s0 = s1;
 217         s1 = s2;
 218         s2 = s3;
 219         s3 = s4;
 220         if (is6tap)
 221             s4 = s5;
 222
 223         dst += dst_stride;
 224         src += src_stride;
 225     }
 226 }
 227
 228 #define EPEL_FUNCS(WIDTH, TAPS) \
 229 static av_noinline \
 230 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) \
 231 { \
 232     put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
 233 } \
 234 \
 235 static av_noinline \
 236 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) \
 237 { \
 238     put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
 239 }
 240
 241 #define EPEL_HV(WIDTH, HTAPS, VTAPS) \
 242 static void put_vp8_epel ## WIDTH ## _h ## HTAPS ## v ## VTAPS ## _altivec(uint8_t *dst, ptrdiff_t stride, uint8_t *src, ptrdiff_t s, int h, int mx, int my) \
 243 { \
 244     DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
 245     if (VTAPS == 6) { \
 246         put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,     src-2*stride, stride, h+5, mx, my); \
 247         put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, stride, tmp+2*16,     16,     h,   mx, my); \
 248     } else { \
 249         put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,     src-stride, stride, h+4, mx, my); \
 250         put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, stride, tmp+16,     16,     h,   mx, my); \
 251     } \
 252 }
 253
 254 EPEL_FUNCS(16,6)
 255 EPEL_FUNCS(8, 6)
 256 EPEL_FUNCS(8, 4)
 257 EPEL_FUNCS(4, 6)
 258 EPEL_FUNCS(4, 4)
 259
 260 EPEL_HV(16, 6,6)
 261 EPEL_HV(8,  6,6)
 262 EPEL_HV(8,  4,6)
 263 EPEL_HV(8,  6,4)
 264 EPEL_HV(8,  4,4)
 265 EPEL_HV(4,  6,6)
 266 EPEL_HV(4,  4,6)
 267 EPEL_HV(4,  6,4)
 268 EPEL_HV(4,  4,4)
 269
 270 static void put_vp8_pixels16_altivec(uint8_t *dst, ptrdiff_t stride, uint8_t *src, ptrdiff_t s, int h, int mx, int my)
 271 {
 272     ff_put_pixels16_altivec(dst, src, stride, h);
 273 }
 274
 275 av_cold void ff_vp8dsp_init_altivec(VP8DSPContext *c)
 276 {
 277     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
 278         return;
 279
 280     c->put_vp8_epel_pixels_tab[0][0][0] = put_vp8_pixels16_altivec;
 281     c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
 282     c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
 283     c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
 284
 285     c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
 286     c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
 287     c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
 288     c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
 289
 290     c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
 291     c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
 292     c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
 293     c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
 294
 295     c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
 296     c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
 297     c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
 298     c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
 299
 300     c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
 301     c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
 302     c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
 303     c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
 304 }