vp8dsp_altivec.c

   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 #include "libavutil/cpu.h"
  25 #include "libavutil/mem.h"
  26 #include "libavutil/ppc/cpu.h"
  27 #include "libavutil/ppc/types_altivec.h"
  28 #include "libavutil/ppc/util_altivec.h"
  29 #include "libavcodec/vp8dsp.h"
  30 #include "hpeldsp_altivec.h"
  31
  32 #if HAVE_ALTIVEC
  33 #define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
  34
  35 // h subpel filter uses msum to multiply+add 4 pixel taps at once
  36 static const vec_s8 h_subpel_filters_inner[7] =
  37 {
  38     REPT4( -6, 123,  12,  -1),
  39     REPT4(-11, 108,  36,  -8),
  40     REPT4( -9,  93,  50,  -6),
  41     REPT4(-16,  77,  77, -16),
  42     REPT4( -6,  50,  93,  -9),
  43     REPT4( -8,  36, 108, -11),
  44     REPT4( -1,  12, 123,  -6),
  45 };
  46
  47 // for 6tap filters, these are the outer two taps
  48 // The zeros mask off pixels 4-7 when filtering 0-3
  49 // and vice-versa
  50 static const vec_s8 h_subpel_filters_outer[3] =
  51 {
  52     REPT4(0, 0, 2, 1),
  53     REPT4(0, 0, 3, 3),
  54     REPT4(0, 0, 1, 2),
  55 };
  56
  57 #define LOAD_H_SUBPEL_FILTER(i) \
  58     vec_s8 filter_inner  = h_subpel_filters_inner[i]; \
  59     vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
  60     vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
  61
  62 #define FILTER_H(dstv, off) \
  63     a = vec_ld((off)-is6tap-1,    src); \
  64     b = vec_ld((off)-is6tap-1+15, src); \
  65 \
  66     pixh  = vec_perm(a, b, permh##off); \
  67     pixl  = vec_perm(a, b, perml##off); \
  68     filth = vec_msum(filter_inner, pixh, c64); \
  69     filtl = vec_msum(filter_inner, pixl, c64); \
  70 \
  71     if (is6tap) { \
  72         outer = vec_perm(a, b, perm_6tap##off); \
  73         filth = vec_msum(filter_outerh, outer, filth); \
  74         filtl = vec_msum(filter_outerl, outer, filtl); \
  75     } \
  76     if (w == 4) \
  77         filtl = filth; /* discard pixels 4-7 */ \
  78     dstv = vec_packs(filth, filtl); \
  79     dstv = vec_sra(dstv, c7)
  80
  81 static av_always_inline
  82 void put_vp8_epel_h_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
  83                                  uint8_t *src, ptrdiff_t src_stride,
  84                                  int h, int mx, int w, int is6tap)
  85 {
  86     LOAD_H_SUBPEL_FILTER(mx-1);
  87     vec_u8 align_vec0, align_vec8, permh0, permh8, filt;
  88     vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
  89     vec_u8 a, b, pixh, pixl, outer;
  90     vec_s16 f16h, f16l;
  91     vec_s32 filth, filtl;
  92
  93     vec_u8 perm_inner6 = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
  94     vec_u8 perm_inner4 = { 0,1,2,3, 1,2,3,4, 2,3,4,5, 3,4,5,6 };
  95     vec_u8 perm_inner  = is6tap ? perm_inner6 : perm_inner4;
  96     vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
  97     vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
  98     vec_u16 c7  = vec_splat_u16(7);
  99
 100     align_vec0 = vec_lvsl( -is6tap-1, src);
 101     align_vec8 = vec_lvsl(8-is6tap-1, src);
 102
 103     permh0     = vec_perm(align_vec0, align_vec0, perm_inner);
 104     permh8     = vec_perm(align_vec8, align_vec8, perm_inner);
 105     perm_inner = vec_add(perm_inner, vec_splat_u8(4));
 106     perml0     = vec_perm(align_vec0, align_vec0, perm_inner);
 107     perml8     = vec_perm(align_vec8, align_vec8, perm_inner);
 108     perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
 109     perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
 110
 111     while (h --> 0) {
 112         FILTER_H(f16h, 0);
 113
 114         if (w == 16) {
 115             FILTER_H(f16l, 8);
 116             filt = vec_packsu(f16h, f16l);
 117             vec_st(filt, 0, dst);
 118         } else {
 119             filt = vec_packsu(f16h, f16h);
 120             vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
 121             if (w == 8)
 122                 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
 123         }
 124         src += src_stride;
 125         dst += dst_stride;
 126     }
 127 }
 128
 129 // v subpel filter does a simple vertical multiply + add
 130 static const vec_u8 v_subpel_filters[7] =
 131 {
 132     { 0,   6, 123,  12,   1,   0 },
 133     { 2,  11, 108,  36,   8,   1 },
 134     { 0,   9,  93,  50,   6,   0 },
 135     { 3,  16,  77,  77,  16,   3 },
 136     { 0,   6,  50,  93,   9,   0 },
 137     { 1,   8,  36, 108,  11,   2 },
 138     { 0,   1,  12, 123,   6,   0 },
 139 };
 140
 141 #define LOAD_V_SUBPEL_FILTER(i) \
 142     vec_u8 subpel_filter = v_subpel_filters[i]; \
 143     vec_u8 f0 = vec_splat(subpel_filter, 0); \
 144     vec_u8 f1 = vec_splat(subpel_filter, 1); \
 145     vec_u8 f2 = vec_splat(subpel_filter, 2); \
 146     vec_u8 f3 = vec_splat(subpel_filter, 3); \
 147     vec_u8 f4 = vec_splat(subpel_filter, 4); \
 148     vec_u8 f5 = vec_splat(subpel_filter, 5)
 149
 150 #define FILTER_V(dstv, vec_mul) \
 151     s1f = (vec_s16)vec_mul(s1, f1); \
 152     s2f = (vec_s16)vec_mul(s2, f2); \
 153     s3f = (vec_s16)vec_mul(s3, f3); \
 154     s4f = (vec_s16)vec_mul(s4, f4); \
 155     s2f = vec_subs(s2f, s1f); \
 156     s3f = vec_subs(s3f, s4f); \
 157     if (is6tap) { \
 158         s0f = (vec_s16)vec_mul(s0, f0); \
 159         s5f = (vec_s16)vec_mul(s5, f5); \
 160         s2f = vec_adds(s2f, s0f); \
 161         s3f = vec_adds(s3f, s5f); \
 162     } \
 163     dstv = vec_adds(s2f, s3f); \
 164     dstv = vec_adds(dstv, c64); \
 165     dstv = vec_sra(dstv, c7)
 166
 167 static av_always_inline
 168 void put_vp8_epel_v_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
 169                                  uint8_t *src, ptrdiff_t src_stride,
 170                                  int h, int my, int w, int is6tap)
 171 {
 172     LOAD_V_SUBPEL_FILTER(my-1);
 173     vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
 174     vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
 175     vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
 176     vec_u16 c7  = vec_splat_u16(7);
 177
 178     // we want pixels 0-7 to be in the even positions and 8-15 in the odd,
 179     // so combine this permute with the alignment permute vector
 180     align_vech = vec_lvsl(0, src);
 181     align_vecl = vec_sld(align_vech, align_vech, 8);
 182     if (w ==16)
 183         perm_vec = vec_mergeh(align_vech, align_vecl);
 184     else
 185         perm_vec = vec_mergeh(align_vech, align_vech);
 186
 187     if (is6tap)
 188         s0 = load_with_perm_vec(-2*src_stride, src, perm_vec);
 189     s1 = load_with_perm_vec(-1*src_stride, src, perm_vec);
 190     s2 = load_with_perm_vec( 0*src_stride, src, perm_vec);
 191     s3 = load_with_perm_vec( 1*src_stride, src, perm_vec);
 192     if (is6tap)
 193         s4 = load_with_perm_vec( 2*src_stride, src, perm_vec);
 194
 195     src += (2+is6tap)*src_stride;
 196
 197     while (h --> 0) {
 198         if (is6tap)
 199             s5 = load_with_perm_vec(0, src, perm_vec);
 200         else
 201             s4 = load_with_perm_vec(0, src, perm_vec);
 202
 203         FILTER_V(f16h, vec_mule);
 204
 205         if (w == 16) {
 206             FILTER_V(f16l, vec_mulo);
 207             filt = vec_packsu(f16h, f16l);
 208             vec_st(filt, 0, dst);
 209         } else {
 210             filt = vec_packsu(f16h, f16h);
 211             if (w == 4)
 212                 filt = (vec_u8)vec_splat((vec_u32)filt, 0);
 213             else
 214                 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
 215             vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
 216         }
 217
 218         if (is6tap)
 219             s0 = s1;
 220         s1 = s2;
 221         s2 = s3;
 222         s3 = s4;
 223         if (is6tap)
 224             s4 = s5;
 225
 226         dst += dst_stride;
 227         src += src_stride;
 228     }
 229 }
 230
 231 #define EPEL_FUNCS(WIDTH, TAPS) \
 232 static av_noinline \
 233 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) \
 234 { \
 235     put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
 236 } \
 237 \
 238 static av_noinline \
 239 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) \
 240 { \
 241     put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
 242 }
 243
 244 #define EPEL_HV(WIDTH, HTAPS, VTAPS) \
 245 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) \
 246 { \
 247     DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
 248     if (VTAPS == 6) { \
 249         put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,      src-2*sstride, sstride, h+5, mx, my); \
 250         put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+2*16,      16,      h,   mx, my); \
 251     } else { \
 252         put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16,      src-sstride, sstride, h+4, mx, my); \
 253         put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+16,      16,      h,   mx, my); \
 254     } \
 255 }
 256
 257 EPEL_FUNCS(16,6)
 258 EPEL_FUNCS(8, 6)
 259 EPEL_FUNCS(8, 4)
 260 EPEL_FUNCS(4, 6)
 261 EPEL_FUNCS(4, 4)
 262
 263 EPEL_HV(16, 6,6)
 264 EPEL_HV(8,  6,6)
 265 EPEL_HV(8,  4,6)
 266 EPEL_HV(8,  6,4)
 267 EPEL_HV(8,  4,4)
 268 EPEL_HV(4,  6,6)
 269 EPEL_HV(4,  4,6)
 270 EPEL_HV(4,  6,4)
 271 EPEL_HV(4,  4,4)
 272
 273 static void put_vp8_pixels16_altivec(uint8_t *dst, ptrdiff_t dstride, uint8_t *src, ptrdiff_t sstride, int h, int mx, int my)
 274 {
 275     register vector unsigned char pixelsv1, pixelsv2;
 276     register vector unsigned char pixelsv1B, pixelsv2B;
 277     register vector unsigned char pixelsv1C, pixelsv2C;
 278     register vector unsigned char pixelsv1D, pixelsv2D;
 279
 280     register vector unsigned char perm = vec_lvsl(0, src);
 281     int i;
 282     register ptrdiff_t dstride2 = dstride << 1, sstride2 = sstride << 1;
 283     register ptrdiff_t dstride3 = dstride2 + dstride, sstride3 = sstride + sstride2;
 284     register ptrdiff_t dstride4 = dstride << 2, sstride4 = sstride << 2;
 285
 286 // hand-unrolling the loop by 4 gains about 15%
 287 // mininum execution time goes from 74 to 60 cycles
 288 // it's faster than -funroll-loops, but using
 289 // -funroll-loops w/ this is bad - 74 cycles again.
 290 // all this is on a 7450, tuning for the 7450
 291     for (i = 0; i < h; i += 4) {
 292         pixelsv1  = vec_ld( 0, src);
 293         pixelsv2  = vec_ld(15, src);
 294         pixelsv1B = vec_ld(sstride, src);
 295         pixelsv2B = vec_ld(15 + sstride, src);
 296         pixelsv1C = vec_ld(sstride2, src);
 297         pixelsv2C = vec_ld(15 + sstride2, src);
 298         pixelsv1D = vec_ld(sstride3, src);
 299         pixelsv2D = vec_ld(15 + sstride3, src);
 300         vec_st(vec_perm(pixelsv1, pixelsv2, perm),
 301                0, (unsigned char*)dst);
 302         vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
 303                dstride, (unsigned char*)dst);
 304         vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
 305                dstride2, (unsigned char*)dst);
 306         vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
 307                dstride3, (unsigned char*)dst);
 308         src += sstride4;
 309         dst += dstride4;
 310     }
 311 }
 312
 313 #endif /* HAVE_ALTIVEC */
 314
 315
 316 av_cold void ff_vp78dsp_init_ppc(VP8DSPContext *c)
 317 {
 318 #if HAVE_ALTIVEC
 319     if (!PPC_ALTIVEC(av_get_cpu_flags()))
 320         return;
 321
 322     c->put_vp8_epel_pixels_tab[0][0][0] = put_vp8_pixels16_altivec;
 323     c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
 324     c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
 325     c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
 326
 327     c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
 328     c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
 329     c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
 330     c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
 331
 332     c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
 333     c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
 334     c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
 335     c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
 336
 337     c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
 338     c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
 339     c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
 340     c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
 341
 342     c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
 343     c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
 344     c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
 345     c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
 346 #endif /* HAVE_ALTIVEC */
 347 }