FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
h264_altivec.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20 
21 #include "libavutil/attributes.h"
22 #include "libavutil/cpu.h"
23 #include "libavutil/intreadwrite.h"
26 #include "libavcodec/h264data.h"
27 #include "libavcodec/h264dsp.h"
28 
29 /****************************************************************************
30  * IDCT transform:
31  ****************************************************************************/
32 
33 #define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3) \
34  /* 1st stage */ \
35  vz0 = vec_add(vb0,vb2); /* temp[0] = Y[0] + Y[2] */ \
36  vz1 = vec_sub(vb0,vb2); /* temp[1] = Y[0] - Y[2] */ \
37  vz2 = vec_sra(vb1,vec_splat_u16(1)); \
38  vz2 = vec_sub(vz2,vb3); /* temp[2] = Y[1].1/2 - Y[3] */ \
39  vz3 = vec_sra(vb3,vec_splat_u16(1)); \
40  vz3 = vec_add(vb1,vz3); /* temp[3] = Y[1] + Y[3].1/2 */ \
41  /* 2nd stage: output */ \
42  va0 = vec_add(vz0,vz3); /* x[0] = temp[0] + temp[3] */ \
43  va1 = vec_add(vz1,vz2); /* x[1] = temp[1] + temp[2] */ \
44  va2 = vec_sub(vz1,vz2); /* x[2] = temp[1] - temp[2] */ \
45  va3 = vec_sub(vz0,vz3) /* x[3] = temp[0] - temp[3] */
46 
47 #define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \
48  b0 = vec_mergeh( a0, a0 ); \
49  b1 = vec_mergeh( a1, a0 ); \
50  b2 = vec_mergeh( a2, a0 ); \
51  b3 = vec_mergeh( a3, a0 ); \
52  a0 = vec_mergeh( b0, b2 ); \
53  a1 = vec_mergel( b0, b2 ); \
54  a2 = vec_mergeh( b1, b3 ); \
55  a3 = vec_mergel( b1, b3 ); \
56  b0 = vec_mergeh( a0, a2 ); \
57  b1 = vec_mergel( a0, a2 ); \
58  b2 = vec_mergeh( a1, a3 ); \
59  b3 = vec_mergel( a1, a3 )
60 
61 #define VEC_LOAD_U8_ADD_S16_STORE_U8(va) \
62  vdst_orig = vec_ld(0, dst); \
63  vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask); \
64  vdst_ss = (vec_s16) vec_mergeh(zero_u8v, vdst); \
65  va = vec_add(va, vdst_ss); \
66  va_u8 = vec_packsu(va, zero_s16v); \
67  va_u32 = vec_splat((vec_u32)va_u8, 0); \
68  vec_ste(va_u32, element, (uint32_t*)dst);
69 
70 static void ff_h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride)
71 {
72  vec_s16 va0, va1, va2, va3;
73  vec_s16 vz0, vz1, vz2, vz3;
74  vec_s16 vtmp0, vtmp1, vtmp2, vtmp3;
75  vec_u8 va_u8;
76  vec_u32 va_u32;
77  vec_s16 vdst_ss;
78  const vec_u16 v6us = vec_splat_u16(6);
79  vec_u8 vdst, vdst_orig;
80  vec_u8 vdst_mask = vec_lvsl(0, dst);
81  int element = ((unsigned long)dst & 0xf) >> 2;
82  LOAD_ZERO;
83 
84  block[0] += 32; /* add 32 as a DC-level for rounding */
85 
86  vtmp0 = vec_ld(0,block);
87  vtmp1 = vec_sld(vtmp0, vtmp0, 8);
88  vtmp2 = vec_ld(16,block);
89  vtmp3 = vec_sld(vtmp2, vtmp2, 8);
90  memset(block, 0, 16 * sizeof(int16_t));
91 
92  VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
93  VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3);
94  VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
95 
96  va0 = vec_sra(va0,v6us);
97  va1 = vec_sra(va1,v6us);
98  va2 = vec_sra(va2,v6us);
99  va3 = vec_sra(va3,v6us);
100 
102  dst += stride;
104  dst += stride;
106  dst += stride;
108 }
109 
110 #define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7, d0, d1, d2, d3, d4, d5, d6, d7) {\
111  /* a0 = SRC(0) + SRC(4); */ \
112  vec_s16 a0v = vec_add(s0, s4); \
113  /* a2 = SRC(0) - SRC(4); */ \
114  vec_s16 a2v = vec_sub(s0, s4); \
115  /* a4 = (SRC(2)>>1) - SRC(6); */ \
116  vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6); \
117  /* a6 = (SRC(6)>>1) + SRC(2); */ \
118  vec_s16 a6v = vec_add(vec_sra(s6, onev), s2); \
119  /* b0 = a0 + a6; */ \
120  vec_s16 b0v = vec_add(a0v, a6v); \
121  /* b2 = a2 + a4; */ \
122  vec_s16 b2v = vec_add(a2v, a4v); \
123  /* b4 = a2 - a4; */ \
124  vec_s16 b4v = vec_sub(a2v, a4v); \
125  /* b6 = a0 - a6; */ \
126  vec_s16 b6v = vec_sub(a0v, a6v); \
127  /* a1 = SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \
128  /* a1 = (SRC(5)-SRC(3)) - (SRC(7) + (SRC(7)>>1)); */ \
129  vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \
130  /* a3 = SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \
131  /* a3 = (SRC(7)+SRC(1)) - (SRC(3) + (SRC(3)>>1)); */ \
132  vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\
133  /* a5 = SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \
134  /* a5 = (SRC(7)-SRC(1)) + SRC(5) + (SRC(5)>>1); */ \
135  vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\
136  /* a7 = SRC(5)+SRC(3) + SRC(1) + (SRC(1)>>1); */ \
137  vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\
138  /* b1 = (a7>>2) + a1; */ \
139  vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \
140  /* b3 = a3 + (a5>>2); */ \
141  vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \
142  /* b5 = (a3>>2) - a5; */ \
143  vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \
144  /* b7 = a7 - (a1>>2); */ \
145  vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \
146  /* DST(0, b0 + b7); */ \
147  d0 = vec_add(b0v, b7v); \
148  /* DST(1, b2 + b5); */ \
149  d1 = vec_add(b2v, b5v); \
150  /* DST(2, b4 + b3); */ \
151  d2 = vec_add(b4v, b3v); \
152  /* DST(3, b6 + b1); */ \
153  d3 = vec_add(b6v, b1v); \
154  /* DST(4, b6 - b1); */ \
155  d4 = vec_sub(b6v, b1v); \
156  /* DST(5, b4 - b3); */ \
157  d5 = vec_sub(b4v, b3v); \
158  /* DST(6, b2 - b5); */ \
159  d6 = vec_sub(b2v, b5v); \
160  /* DST(7, b0 - b7); */ \
161  d7 = vec_sub(b0v, b7v); \
162 }
163 
164 #define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \
165  /* unaligned load */ \
166  vec_u8 hv = vec_ld( 0, dest ); \
167  vec_u8 lv = vec_ld( 7, dest ); \
168  vec_u8 dstv = vec_perm( hv, lv, (vec_u8)perm_ldv ); \
169  vec_s16 idct_sh6 = vec_sra(idctv, sixv); \
170  vec_u16 dst16 = (vec_u16)vec_mergeh(zero_u8v, dstv); \
171  vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16); \
172  vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum); \
173  vec_u8 edgehv; \
174  /* unaligned store */ \
175  vec_u8 bodyv = vec_perm( idstsum8, idstsum8, perm_stv );\
176  vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv ); \
177  lv = vec_sel( lv, bodyv, edgelv ); \
178  vec_st( lv, 7, dest ); \
179  hv = vec_ld( 0, dest ); \
180  edgehv = vec_perm( zero_u8v, sel, perm_stv ); \
181  hv = vec_sel( hv, bodyv, edgehv ); \
182  vec_st( hv, 0, dest ); \
183  }
184 
185 static void ff_h264_idct8_add_altivec( uint8_t *dst, int16_t *dct, int stride ) {
186  vec_s16 s0, s1, s2, s3, s4, s5, s6, s7;
187  vec_s16 d0, d1, d2, d3, d4, d5, d6, d7;
188  vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7;
189 
190  vec_u8 perm_ldv = vec_lvsl(0, dst);
191  vec_u8 perm_stv = vec_lvsr(8, dst);
192 
193  const vec_u16 onev = vec_splat_u16(1);
194  const vec_u16 twov = vec_splat_u16(2);
195  const vec_u16 sixv = vec_splat_u16(6);
196 
197  const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1};
198  LOAD_ZERO;
199 
200  dct[0] += 32; // rounding for the >>6 at the end
201 
202  s0 = vec_ld(0x00, (int16_t*)dct);
203  s1 = vec_ld(0x10, (int16_t*)dct);
204  s2 = vec_ld(0x20, (int16_t*)dct);
205  s3 = vec_ld(0x30, (int16_t*)dct);
206  s4 = vec_ld(0x40, (int16_t*)dct);
207  s5 = vec_ld(0x50, (int16_t*)dct);
208  s6 = vec_ld(0x60, (int16_t*)dct);
209  s7 = vec_ld(0x70, (int16_t*)dct);
210  memset(dct, 0, 64 * sizeof(int16_t));
211 
212  IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,
213  d0, d1, d2, d3, d4, d5, d6, d7);
214 
215  TRANSPOSE8( d0, d1, d2, d3, d4, d5, d6, d7 );
216 
217  IDCT8_1D_ALTIVEC(d0, d1, d2, d3, d4, d5, d6, d7,
218  idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7);
219 
220  ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel);
221  ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel);
222  ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel);
223  ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel);
224  ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel);
225  ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel);
226  ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel);
227  ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel);
228 }
229 
231 {
232  vec_s16 dc16;
233  vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
234  LOAD_ZERO;
235  DECLARE_ALIGNED(16, int, dc);
236  int i;
237 
238  dc = (block[0] + 32) >> 6;
239  block[0] = 0;
240  dc16 = vec_splat((vec_s16) vec_lde(0, &dc), 1);
241 
242  if (size == 4)
243  dc16 = vec_sld(dc16, zero_s16v, 8);
244  dcplus = vec_packsu(dc16, zero_s16v);
245  dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
246 
247  aligner = vec_lvsr(0, dst);
248  dcplus = vec_perm(dcplus, dcplus, aligner);
249  dcminus = vec_perm(dcminus, dcminus, aligner);
250 
251  for (i = 0; i < size; i += 4) {
252  v0 = vec_ld(0, dst+0*stride);
253  v1 = vec_ld(0, dst+1*stride);
254  v2 = vec_ld(0, dst+2*stride);
255  v3 = vec_ld(0, dst+3*stride);
256 
257  v0 = vec_adds(v0, dcplus);
258  v1 = vec_adds(v1, dcplus);
259  v2 = vec_adds(v2, dcplus);
260  v3 = vec_adds(v3, dcplus);
261 
262  v0 = vec_subs(v0, dcminus);
263  v1 = vec_subs(v1, dcminus);
264  v2 = vec_subs(v2, dcminus);
265  v3 = vec_subs(v3, dcminus);
266 
267  vec_st(v0, 0, dst+0*stride);
268  vec_st(v1, 0, dst+1*stride);
269  vec_st(v2, 0, dst+2*stride);
270  vec_st(v3, 0, dst+3*stride);
271 
272  dst += 4*stride;
273  }
274 }
275 
276 static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
277 {
278  h264_idct_dc_add_internal(dst, block, stride, 4);
279 }
280 
282 {
283  h264_idct_dc_add_internal(dst, block, stride, 8);
284 }
285 
286 static void ff_h264_idct_add16_altivec(uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
287  int i;
288  for(i=0; i<16; i++){
289  int nnz = nnzc[ scan8[i] ];
290  if(nnz){
291  if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
292  else ff_h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
293  }
294  }
295 }
296 
297 static void ff_h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
298  int i;
299  for(i=0; i<16; i++){
300  if(nnzc[ scan8[i] ]) ff_h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
301  else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
302  }
303 }
304 
305 static void ff_h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
306  int i;
307  for(i=0; i<16; i+=4){
308  int nnz = nnzc[ scan8[i] ];
309  if(nnz){
310  if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
311  else ff_h264_idct8_add_altivec (dst + block_offset[i], block + i*16, stride);
312  }
313  }
314 }
315 
316 static void ff_h264_idct_add8_altivec(uint8_t **dest, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
317  int i, j;
318  for (j = 1; j < 3; j++) {
319  for(i = j * 16; i < j * 16 + 4; i++){
320  if(nnzc[ scan8[i] ])
321  ff_h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
322  else if(block[i*16])
323  h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
324  }
325  }
326 }
327 
328 #define transpose4x16(r0, r1, r2, r3) { \
329  register vec_u8 r4; \
330  register vec_u8 r5; \
331  register vec_u8 r6; \
332  register vec_u8 r7; \
333  \
334  r4 = vec_mergeh(r0, r2); /*0, 2 set 0*/ \
335  r5 = vec_mergel(r0, r2); /*0, 2 set 1*/ \
336  r6 = vec_mergeh(r1, r3); /*1, 3 set 0*/ \
337  r7 = vec_mergel(r1, r3); /*1, 3 set 1*/ \
338  \
339  r0 = vec_mergeh(r4, r6); /*all set 0*/ \
340  r1 = vec_mergel(r4, r6); /*all set 1*/ \
341  r2 = vec_mergeh(r5, r7); /*all set 2*/ \
342  r3 = vec_mergel(r5, r7); /*all set 3*/ \
343 }
344 
345 static inline void write16x4(uint8_t *dst, int dst_stride,
346  register vec_u8 r0, register vec_u8 r1,
347  register vec_u8 r2, register vec_u8 r3) {
348  DECLARE_ALIGNED(16, unsigned char, result)[64];
349  uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
350  int int_dst_stride = dst_stride/4;
351 
352  vec_st(r0, 0, result);
353  vec_st(r1, 16, result);
354  vec_st(r2, 32, result);
355  vec_st(r3, 48, result);
356  /* FIXME: there has to be a better way!!!! */
357  *dst_int = *src_int;
358  *(dst_int+ int_dst_stride) = *(src_int + 1);
359  *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
360  *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
361  *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
362  *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
363  *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
364  *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
365  *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
366  *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
367  *(dst_int+10*int_dst_stride) = *(src_int + 10);
368  *(dst_int+11*int_dst_stride) = *(src_int + 11);
369  *(dst_int+12*int_dst_stride) = *(src_int + 12);
370  *(dst_int+13*int_dst_stride) = *(src_int + 13);
371  *(dst_int+14*int_dst_stride) = *(src_int + 14);
372  *(dst_int+15*int_dst_stride) = *(src_int + 15);
373 }
374 
375 /** @brief performs a 6x16 transpose of data in src, and stores it to dst
376  @todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
377  out of unaligned_load() */
378 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
379  register vec_u8 r0 = unaligned_load(0, src); \
380  register vec_u8 r1 = unaligned_load( src_stride, src); \
381  register vec_u8 r2 = unaligned_load(2* src_stride, src); \
382  register vec_u8 r3 = unaligned_load(3* src_stride, src); \
383  register vec_u8 r4 = unaligned_load(4* src_stride, src); \
384  register vec_u8 r5 = unaligned_load(5* src_stride, src); \
385  register vec_u8 r6 = unaligned_load(6* src_stride, src); \
386  register vec_u8 r7 = unaligned_load(7* src_stride, src); \
387  register vec_u8 r14 = unaligned_load(14*src_stride, src); \
388  register vec_u8 r15 = unaligned_load(15*src_stride, src); \
389  \
390  r8 = unaligned_load( 8*src_stride, src); \
391  r9 = unaligned_load( 9*src_stride, src); \
392  r10 = unaligned_load(10*src_stride, src); \
393  r11 = unaligned_load(11*src_stride, src); \
394  r12 = unaligned_load(12*src_stride, src); \
395  r13 = unaligned_load(13*src_stride, src); \
396  \
397  /*Merge first pairs*/ \
398  r0 = vec_mergeh(r0, r8); /*0, 8*/ \
399  r1 = vec_mergeh(r1, r9); /*1, 9*/ \
400  r2 = vec_mergeh(r2, r10); /*2,10*/ \
401  r3 = vec_mergeh(r3, r11); /*3,11*/ \
402  r4 = vec_mergeh(r4, r12); /*4,12*/ \
403  r5 = vec_mergeh(r5, r13); /*5,13*/ \
404  r6 = vec_mergeh(r6, r14); /*6,14*/ \
405  r7 = vec_mergeh(r7, r15); /*7,15*/ \
406  \
407  /*Merge second pairs*/ \
408  r8 = vec_mergeh(r0, r4); /*0,4, 8,12 set 0*/ \
409  r9 = vec_mergel(r0, r4); /*0,4, 8,12 set 1*/ \
410  r10 = vec_mergeh(r1, r5); /*1,5, 9,13 set 0*/ \
411  r11 = vec_mergel(r1, r5); /*1,5, 9,13 set 1*/ \
412  r12 = vec_mergeh(r2, r6); /*2,6,10,14 set 0*/ \
413  r13 = vec_mergel(r2, r6); /*2,6,10,14 set 1*/ \
414  r14 = vec_mergeh(r3, r7); /*3,7,11,15 set 0*/ \
415  r15 = vec_mergel(r3, r7); /*3,7,11,15 set 1*/ \
416  \
417  /*Third merge*/ \
418  r0 = vec_mergeh(r8, r12); /*0,2,4,6,8,10,12,14 set 0*/ \
419  r1 = vec_mergel(r8, r12); /*0,2,4,6,8,10,12,14 set 1*/ \
420  r2 = vec_mergeh(r9, r13); /*0,2,4,6,8,10,12,14 set 2*/ \
421  r4 = vec_mergeh(r10, r14); /*1,3,5,7,9,11,13,15 set 0*/ \
422  r5 = vec_mergel(r10, r14); /*1,3,5,7,9,11,13,15 set 1*/ \
423  r6 = vec_mergeh(r11, r15); /*1,3,5,7,9,11,13,15 set 2*/ \
424  /* Don't need to compute 3 and 7*/ \
425  \
426  /*Final merge*/ \
427  r8 = vec_mergeh(r0, r4); /*all set 0*/ \
428  r9 = vec_mergel(r0, r4); /*all set 1*/ \
429  r10 = vec_mergeh(r1, r5); /*all set 2*/ \
430  r11 = vec_mergel(r1, r5); /*all set 3*/ \
431  r12 = vec_mergeh(r2, r6); /*all set 4*/ \
432  r13 = vec_mergel(r2, r6); /*all set 5*/ \
433  /* Don't need to compute 14 and 15*/ \
434  \
435 }
436 
437 // out: o = |x-y| < a
438 static inline vec_u8 diff_lt_altivec ( register vec_u8 x,
439  register vec_u8 y,
440  register vec_u8 a) {
441 
442  register vec_u8 diff = vec_subs(x, y);
443  register vec_u8 diffneg = vec_subs(y, x);
444  register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */
445  o = (vec_u8)vec_cmplt(o, a);
446  return o;
447 }
448 
449 static inline vec_u8 h264_deblock_mask ( register vec_u8 p0,
450  register vec_u8 p1,
451  register vec_u8 q0,
452  register vec_u8 q1,
453  register vec_u8 alpha,
454  register vec_u8 beta) {
455 
456  register vec_u8 mask;
457  register vec_u8 tempmask;
458 
459  mask = diff_lt_altivec(p0, q0, alpha);
460  tempmask = diff_lt_altivec(p1, p0, beta);
461  mask = vec_and(mask, tempmask);
462  tempmask = diff_lt_altivec(q1, q0, beta);
463  mask = vec_and(mask, tempmask);
464 
465  return mask;
466 }
467 
468 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
469 static inline vec_u8 h264_deblock_q1(register vec_u8 p0,
470  register vec_u8 p1,
471  register vec_u8 p2,
472  register vec_u8 q0,
473  register vec_u8 tc0) {
474 
475  register vec_u8 average = vec_avg(p0, q0);
476  register vec_u8 temp;
477  register vec_u8 uncliped;
478  register vec_u8 ones;
479  register vec_u8 max;
480  register vec_u8 min;
481  register vec_u8 newp1;
482 
483  temp = vec_xor(average, p2);
484  average = vec_avg(average, p2); /*avg(p2, avg(p0, q0)) */
485  ones = vec_splat_u8(1);
486  temp = vec_and(temp, ones); /*(p2^avg(p0, q0)) & 1 */
487  uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
488  max = vec_adds(p1, tc0);
489  min = vec_subs(p1, tc0);
490  newp1 = vec_max(min, uncliped);
491  newp1 = vec_min(max, newp1);
492  return newp1;
493 }
494 
495 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) { \
496  \
497  const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4)); \
498  \
499  register vec_u8 pq0bit = vec_xor(p0,q0); \
500  register vec_u8 q1minus; \
501  register vec_u8 p0minus; \
502  register vec_u8 stage1; \
503  register vec_u8 stage2; \
504  register vec_u8 vec160; \
505  register vec_u8 delta; \
506  register vec_u8 deltaneg; \
507  \
508  q1minus = vec_nor(q1, q1); /* 255 - q1 */ \
509  stage1 = vec_avg(p1, q1minus); /* (p1 - q1 + 256)>>1 */ \
510  stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */ \
511  p0minus = vec_nor(p0, p0); /* 255 - p0 */ \
512  stage1 = vec_avg(q0, p0minus); /* (q0 - p0 + 256)>>1 */ \
513  pq0bit = vec_and(pq0bit, vec_splat_u8(1)); \
514  stage2 = vec_avg(stage2, pq0bit); /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \
515  stage2 = vec_adds(stage2, stage1); /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
516  vec160 = vec_ld(0, &A0v); \
517  deltaneg = vec_subs(vec160, stage2); /* -d */ \
518  delta = vec_subs(stage2, vec160); /* d */ \
519  deltaneg = vec_min(tc0masked, deltaneg); \
520  delta = vec_min(tc0masked, delta); \
521  p0 = vec_subs(p0, deltaneg); \
522  q0 = vec_subs(q0, delta); \
523  p0 = vec_adds(p0, delta); \
524  q0 = vec_adds(q0, deltaneg); \
525 }
526 
527 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) { \
528  DECLARE_ALIGNED(16, unsigned char, temp)[16]; \
529  register vec_u8 alphavec; \
530  register vec_u8 betavec; \
531  register vec_u8 mask; \
532  register vec_u8 p1mask; \
533  register vec_u8 q1mask; \
534  register vector signed char tc0vec; \
535  register vec_u8 finaltc0; \
536  register vec_u8 tc0masked; \
537  register vec_u8 newp1; \
538  register vec_u8 newq1; \
539  \
540  temp[0] = alpha; \
541  temp[1] = beta; \
542  alphavec = vec_ld(0, temp); \
543  betavec = vec_splat(alphavec, 0x1); \
544  alphavec = vec_splat(alphavec, 0x0); \
545  mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */ \
546  \
547  AV_COPY32(temp, tc0); \
548  tc0vec = vec_ld(0, (signed char*)temp); \
549  tc0vec = vec_mergeh(tc0vec, tc0vec); \
550  tc0vec = vec_mergeh(tc0vec, tc0vec); \
551  mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1))); /* if tc0[i] >= 0 */ \
552  finaltc0 = vec_and((vec_u8)tc0vec, mask); /* tc = tc0 */ \
553  \
554  p1mask = diff_lt_altivec(p2, p0, betavec); \
555  p1mask = vec_and(p1mask, mask); /* if ( |p2 - p0| < beta) */ \
556  tc0masked = vec_and(p1mask, (vec_u8)tc0vec); \
557  finaltc0 = vec_sub(finaltc0, p1mask); /* tc++ */ \
558  newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked); \
559  /*end if*/ \
560  \
561  q1mask = diff_lt_altivec(q2, q0, betavec); \
562  q1mask = vec_and(q1mask, mask); /* if ( |q2 - q0| < beta ) */\
563  tc0masked = vec_and(q1mask, (vec_u8)tc0vec); \
564  finaltc0 = vec_sub(finaltc0, q1mask); /* tc++ */ \
565  newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked); \
566  /*end if*/ \
567  \
568  h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0); \
569  p1 = newp1; \
570  q1 = newq1; \
571 }
572 
573 static void h264_v_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
574 
575  if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
576  register vec_u8 p2 = vec_ld(-3*stride, pix);
577  register vec_u8 p1 = vec_ld(-2*stride, pix);
578  register vec_u8 p0 = vec_ld(-1*stride, pix);
579  register vec_u8 q0 = vec_ld(0, pix);
580  register vec_u8 q1 = vec_ld(stride, pix);
581  register vec_u8 q2 = vec_ld(2*stride, pix);
582  h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
583  vec_st(p1, -2*stride, pix);
584  vec_st(p0, -1*stride, pix);
585  vec_st(q0, 0, pix);
586  vec_st(q1, stride, pix);
587  }
588 }
589 
590 static void h264_h_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
591 
592  register vec_u8 line0, line1, line2, line3, line4, line5;
593  if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
594  return;
595  readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
596  h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
597  transpose4x16(line1, line2, line3, line4);
598  write16x4(pix-2, stride, line1, line2, line3, line4);
599 }
600 
601 static av_always_inline
603  int log2_denom, int weight, int offset, int w)
604 {
605  int y, aligned;
606  vec_u8 vblock;
607  vec_s16 vtemp, vweight, voffset, v0, v1;
608  vec_u16 vlog2_denom;
609  DECLARE_ALIGNED(16, int32_t, temp)[4];
610  LOAD_ZERO;
611 
612  offset <<= log2_denom;
613  if(log2_denom) offset += 1<<(log2_denom-1);
614  temp[0] = log2_denom;
615  temp[1] = weight;
616  temp[2] = offset;
617 
618  vtemp = (vec_s16)vec_ld(0, temp);
619  vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
620  vweight = vec_splat(vtemp, 3);
621  voffset = vec_splat(vtemp, 5);
622  aligned = !((unsigned long)block & 0xf);
623 
624  for (y = 0; y < height; y++) {
625  vblock = vec_ld(0, block);
626 
627  v0 = (vec_s16)vec_mergeh(zero_u8v, vblock);
628  v1 = (vec_s16)vec_mergel(zero_u8v, vblock);
629 
630  if (w == 16 || aligned) {
631  v0 = vec_mladd(v0, vweight, zero_s16v);
632  v0 = vec_adds(v0, voffset);
633  v0 = vec_sra(v0, vlog2_denom);
634  }
635  if (w == 16 || !aligned) {
636  v1 = vec_mladd(v1, vweight, zero_s16v);
637  v1 = vec_adds(v1, voffset);
638  v1 = vec_sra(v1, vlog2_denom);
639  }
640  vblock = vec_packsu(v0, v1);
641  vec_st(vblock, 0, block);
642 
643  block += stride;
644  }
645 }
646 
647 static av_always_inline
649  int log2_denom, int weightd, int weights, int offset, int w)
650 {
651  int y, dst_aligned, src_aligned;
652  vec_u8 vsrc, vdst;
653  vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
654  vec_u16 vlog2_denom;
655  DECLARE_ALIGNED(16, int32_t, temp)[4];
656  LOAD_ZERO;
657 
658  offset = ((offset + 1) | 1) << log2_denom;
659  temp[0] = log2_denom+1;
660  temp[1] = weights;
661  temp[2] = weightd;
662  temp[3] = offset;
663 
664  vtemp = (vec_s16)vec_ld(0, temp);
665  vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
666  vweights = vec_splat(vtemp, 3);
667  vweightd = vec_splat(vtemp, 5);
668  voffset = vec_splat(vtemp, 7);
669  dst_aligned = !((unsigned long)dst & 0xf);
670  src_aligned = !((unsigned long)src & 0xf);
671 
672  for (y = 0; y < height; y++) {
673  vdst = vec_ld(0, dst);
674  vsrc = vec_ld(0, src);
675 
676  v0 = (vec_s16)vec_mergeh(zero_u8v, vdst);
677  v1 = (vec_s16)vec_mergel(zero_u8v, vdst);
678  v2 = (vec_s16)vec_mergeh(zero_u8v, vsrc);
679  v3 = (vec_s16)vec_mergel(zero_u8v, vsrc);
680 
681  if (w == 8) {
682  if (src_aligned)
683  v3 = v2;
684  else
685  v2 = v3;
686  }
687 
688  if (w == 16 || dst_aligned) {
689  v0 = vec_mladd(v0, vweightd, zero_s16v);
690  v2 = vec_mladd(v2, vweights, zero_s16v);
691 
692  v0 = vec_adds(v0, voffset);
693  v0 = vec_adds(v0, v2);
694  v0 = vec_sra(v0, vlog2_denom);
695  }
696  if (w == 16 || !dst_aligned) {
697  v1 = vec_mladd(v1, vweightd, zero_s16v);
698  v3 = vec_mladd(v3, vweights, zero_s16v);
699 
700  v1 = vec_adds(v1, voffset);
701  v1 = vec_adds(v1, v3);
702  v1 = vec_sra(v1, vlog2_denom);
703  }
704  vdst = vec_packsu(v0, v1);
705  vec_st(vdst, 0, dst);
706 
707  dst += stride;
708  src += stride;
709  }
710 }
711 
712 #define H264_WEIGHT(W) \
713 static void ff_weight_h264_pixels ## W ## _altivec(uint8_t *block, int stride, int height, \
714  int log2_denom, int weight, int offset){ \
715  weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \
716 }\
717 static void ff_biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, int stride, int height, \
718  int log2_denom, int weightd, int weights, int offset){ \
719  biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \
720 }
721 
722 H264_WEIGHT(16)
723 H264_WEIGHT( 8)
724 
725 av_cold void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth,
726  const int chroma_format_idc)
727 {
729  if (bit_depth == 8) {
730  c->h264_idct_add = ff_h264_idct_add_altivec;
731  if (chroma_format_idc == 1)
732  c->h264_idct_add8 = ff_h264_idct_add8_altivec;
733  c->h264_idct_add16 = ff_h264_idct_add16_altivec;
734  c->h264_idct_add16intra = ff_h264_idct_add16intra_altivec;
735  c->h264_idct_dc_add= h264_idct_dc_add_altivec;
736  c->h264_idct8_dc_add = ff_h264_idct8_dc_add_altivec;
737  c->h264_idct8_add = ff_h264_idct8_add_altivec;
738  c->h264_idct8_add4 = ff_h264_idct8_add4_altivec;
739  c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec;
740  c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec;
741 
742  c->weight_h264_pixels_tab[0] = ff_weight_h264_pixels16_altivec;
743  c->weight_h264_pixels_tab[1] = ff_weight_h264_pixels8_altivec;
744  c->biweight_h264_pixels_tab[0] = ff_biweight_h264_pixels16_altivec;
745  c->biweight_h264_pixels_tab[1] = ff_biweight_h264_pixels8_altivec;
746  }
747  }
748 }