2 * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
4 * This file is part of Libav.
6 * Libav 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.
11 * Libav 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.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with Libav; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include "libavutil/attributes.h"
27 #include "libavutil/cpu.h"
28 #include "libavutil/intreadwrite.h"
29 #include "libavutil/mem.h"
30 #include "libavutil/ppc/cpu.h"
31 #include "libavutil/ppc/types_altivec.h"
32 #include "libavutil/ppc/util_altivec.h"
34 #include "libavcodec/h264.h"
35 #include "libavcodec/h264dsp.h"
37 #if HAVE_ALTIVEC && HAVE_BIGENDIAN
39 /****************************************************************************
41 ****************************************************************************/
43 #define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3) \
45 vz0 = vec_add(vb0,vb2); /* temp[0] = Y[0] + Y[2] */ \
46 vz1 = vec_sub(vb0,vb2); /* temp[1] = Y[0] - Y[2] */ \
47 vz2 = vec_sra(vb1,vec_splat_u16(1)); \
48 vz2 = vec_sub(vz2,vb3); /* temp[2] = Y[1].1/2 - Y[3] */ \
49 vz3 = vec_sra(vb3,vec_splat_u16(1)); \
50 vz3 = vec_add(vb1,vz3); /* temp[3] = Y[1] + Y[3].1/2 */ \
51 /* 2nd stage: output */ \
52 va0 = vec_add(vz0,vz3); /* x[0] = temp[0] + temp[3] */ \
53 va1 = vec_add(vz1,vz2); /* x[1] = temp[1] + temp[2] */ \
54 va2 = vec_sub(vz1,vz2); /* x[2] = temp[1] - temp[2] */ \
55 va3 = vec_sub(vz0,vz3) /* x[3] = temp[0] - temp[3] */
57 #define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \
58 b0 = vec_mergeh( a0, a0 ); \
59 b1 = vec_mergeh( a1, a0 ); \
60 b2 = vec_mergeh( a2, a0 ); \
61 b3 = vec_mergeh( a3, a0 ); \
62 a0 = vec_mergeh( b0, b2 ); \
63 a1 = vec_mergel( b0, b2 ); \
64 a2 = vec_mergeh( b1, b3 ); \
65 a3 = vec_mergel( b1, b3 ); \
66 b0 = vec_mergeh( a0, a2 ); \
67 b1 = vec_mergel( a0, a2 ); \
68 b2 = vec_mergeh( a1, a3 ); \
69 b3 = vec_mergel( a1, a3 )
71 #define VEC_LOAD_U8_ADD_S16_STORE_U8(va) \
72 vdst_orig = vec_ld(0, dst); \
73 vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask); \
74 vdst_ss = (vec_s16) vec_mergeh(zero_u8v, vdst); \
75 va = vec_add(va, vdst_ss); \
76 va_u8 = vec_packsu(va, zero_s16v); \
77 va_u32 = vec_splat((vec_u32)va_u8, 0); \
78 vec_ste(va_u32, element, (uint32_t*)dst);
80 static void h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride)
82 vec_s16 va0, va1, va2, va3;
83 vec_s16 vz0, vz1, vz2, vz3;
84 vec_s16 vtmp0, vtmp1, vtmp2, vtmp3;
88 const vec_u16 v6us = vec_splat_u16(6);
89 vec_u8 vdst, vdst_orig;
90 vec_u8 vdst_mask = vec_lvsl(0, dst);
91 int element = ((unsigned long)dst & 0xf) >> 2;
94 block[0] += 32; /* add 32 as a DC-level for rounding */
96 vtmp0 = vec_ld(0,block);
97 vtmp1 = vec_sld(vtmp0, vtmp0, 8);
98 vtmp2 = vec_ld(16,block);
99 vtmp3 = vec_sld(vtmp2, vtmp2, 8);
100 memset(block, 0, 16 * sizeof(int16_t));
102 VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
103 VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3);
104 VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
106 va0 = vec_sra(va0,v6us);
107 va1 = vec_sra(va1,v6us);
108 va2 = vec_sra(va2,v6us);
109 va3 = vec_sra(va3,v6us);
111 VEC_LOAD_U8_ADD_S16_STORE_U8(va0);
113 VEC_LOAD_U8_ADD_S16_STORE_U8(va1);
115 VEC_LOAD_U8_ADD_S16_STORE_U8(va2);
117 VEC_LOAD_U8_ADD_S16_STORE_U8(va3);
120 #define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7, d0, d1, d2, d3, d4, d5, d6, d7) {\
121 /* a0 = SRC(0) + SRC(4); */ \
122 vec_s16 a0v = vec_add(s0, s4); \
123 /* a2 = SRC(0) - SRC(4); */ \
124 vec_s16 a2v = vec_sub(s0, s4); \
125 /* a4 = (SRC(2)>>1) - SRC(6); */ \
126 vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6); \
127 /* a6 = (SRC(6)>>1) + SRC(2); */ \
128 vec_s16 a6v = vec_add(vec_sra(s6, onev), s2); \
129 /* b0 = a0 + a6; */ \
130 vec_s16 b0v = vec_add(a0v, a6v); \
131 /* b2 = a2 + a4; */ \
132 vec_s16 b2v = vec_add(a2v, a4v); \
133 /* b4 = a2 - a4; */ \
134 vec_s16 b4v = vec_sub(a2v, a4v); \
135 /* b6 = a0 - a6; */ \
136 vec_s16 b6v = vec_sub(a0v, a6v); \
137 /* a1 = SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \
138 /* a1 = (SRC(5)-SRC(3)) - (SRC(7) + (SRC(7)>>1)); */ \
139 vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \
140 /* a3 = SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \
141 /* a3 = (SRC(7)+SRC(1)) - (SRC(3) + (SRC(3)>>1)); */ \
142 vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\
143 /* a5 = SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \
144 /* a5 = (SRC(7)-SRC(1)) + SRC(5) + (SRC(5)>>1); */ \
145 vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\
146 /* a7 = SRC(5)+SRC(3) + SRC(1) + (SRC(1)>>1); */ \
147 vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\
148 /* b1 = (a7>>2) + a1; */ \
149 vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \
150 /* b3 = a3 + (a5>>2); */ \
151 vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \
152 /* b5 = (a3>>2) - a5; */ \
153 vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \
154 /* b7 = a7 - (a1>>2); */ \
155 vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \
156 /* DST(0, b0 + b7); */ \
157 d0 = vec_add(b0v, b7v); \
158 /* DST(1, b2 + b5); */ \
159 d1 = vec_add(b2v, b5v); \
160 /* DST(2, b4 + b3); */ \
161 d2 = vec_add(b4v, b3v); \
162 /* DST(3, b6 + b1); */ \
163 d3 = vec_add(b6v, b1v); \
164 /* DST(4, b6 - b1); */ \
165 d4 = vec_sub(b6v, b1v); \
166 /* DST(5, b4 - b3); */ \
167 d5 = vec_sub(b4v, b3v); \
168 /* DST(6, b2 - b5); */ \
169 d6 = vec_sub(b2v, b5v); \
170 /* DST(7, b0 - b7); */ \
171 d7 = vec_sub(b0v, b7v); \
174 #define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \
175 /* unaligned load */ \
176 vec_u8 hv = vec_ld( 0, dest ); \
177 vec_u8 lv = vec_ld( 7, dest ); \
178 vec_u8 dstv = vec_perm( hv, lv, (vec_u8)perm_ldv ); \
179 vec_s16 idct_sh6 = vec_sra(idctv, sixv); \
180 vec_u16 dst16 = (vec_u16)vec_mergeh(zero_u8v, dstv); \
181 vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16); \
182 vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum); \
184 /* unaligned store */ \
185 vec_u8 bodyv = vec_perm( idstsum8, idstsum8, perm_stv );\
186 vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv ); \
187 lv = vec_sel( lv, bodyv, edgelv ); \
188 vec_st( lv, 7, dest ); \
189 hv = vec_ld( 0, dest ); \
190 edgehv = vec_perm( zero_u8v, sel, perm_stv ); \
191 hv = vec_sel( hv, bodyv, edgehv ); \
192 vec_st( hv, 0, dest ); \
195 static void h264_idct8_add_altivec(uint8_t *dst, int16_t *dct, int stride)
197 vec_s16 s0, s1, s2, s3, s4, s5, s6, s7;
198 vec_s16 d0, d1, d2, d3, d4, d5, d6, d7;
199 vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7;
201 vec_u8 perm_ldv = vec_lvsl(0, dst);
202 vec_u8 perm_stv = vec_lvsr(8, dst);
204 const vec_u16 onev = vec_splat_u16(1);
205 const vec_u16 twov = vec_splat_u16(2);
206 const vec_u16 sixv = vec_splat_u16(6);
208 const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1};
211 dct[0] += 32; // rounding for the >>6 at the end
213 s0 = vec_ld(0x00, (int16_t*)dct);
214 s1 = vec_ld(0x10, (int16_t*)dct);
215 s2 = vec_ld(0x20, (int16_t*)dct);
216 s3 = vec_ld(0x30, (int16_t*)dct);
217 s4 = vec_ld(0x40, (int16_t*)dct);
218 s5 = vec_ld(0x50, (int16_t*)dct);
219 s6 = vec_ld(0x60, (int16_t*)dct);
220 s7 = vec_ld(0x70, (int16_t*)dct);
221 memset(dct, 0, 64 * sizeof(int16_t));
223 IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,
224 d0, d1, d2, d3, d4, d5, d6, d7);
226 TRANSPOSE8( d0, d1, d2, d3, d4, d5, d6, d7 );
228 IDCT8_1D_ALTIVEC(d0, d1, d2, d3, d4, d5, d6, d7,
229 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7);
231 ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel);
232 ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel);
233 ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel);
234 ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel);
235 ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel);
236 ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel);
237 ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel);
238 ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel);
241 static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
244 vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
246 DECLARE_ALIGNED(16, int, dc);
249 dc = (block[0] + 32) >> 6;
251 dc16 = vec_splat((vec_s16) vec_lde(0, &dc), 1);
254 dc16 = vec_sld(dc16, zero_s16v, 8);
255 dcplus = vec_packsu(dc16, zero_s16v);
256 dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
258 aligner = vec_lvsr(0, dst);
259 dcplus = vec_perm(dcplus, dcplus, aligner);
260 dcminus = vec_perm(dcminus, dcminus, aligner);
262 for (i = 0; i < size; i += 4) {
263 v0 = vec_ld(0, dst+0*stride);
264 v1 = vec_ld(0, dst+1*stride);
265 v2 = vec_ld(0, dst+2*stride);
266 v3 = vec_ld(0, dst+3*stride);
268 v0 = vec_adds(v0, dcplus);
269 v1 = vec_adds(v1, dcplus);
270 v2 = vec_adds(v2, dcplus);
271 v3 = vec_adds(v3, dcplus);
273 v0 = vec_subs(v0, dcminus);
274 v1 = vec_subs(v1, dcminus);
275 v2 = vec_subs(v2, dcminus);
276 v3 = vec_subs(v3, dcminus);
278 vec_st(v0, 0, dst+0*stride);
279 vec_st(v1, 0, dst+1*stride);
280 vec_st(v2, 0, dst+2*stride);
281 vec_st(v3, 0, dst+3*stride);
287 static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
289 h264_idct_dc_add_internal(dst, block, stride, 4);
292 static void h264_idct8_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
294 h264_idct_dc_add_internal(dst, block, stride, 8);
297 static void h264_idct_add16_altivec(uint8_t *dst, const int *block_offset,
298 int16_t *block, int stride,
299 const uint8_t nnzc[15 * 8])
303 int nnz = nnzc[ scan8[i] ];
305 if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
306 else h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
311 static void h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset,
312 int16_t *block, int stride,
313 const uint8_t nnzc[15 * 8])
317 if(nnzc[ scan8[i] ]) h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
318 else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
322 static void h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset,
323 int16_t *block, int stride,
324 const uint8_t nnzc[15 * 8])
327 for(i=0; i<16; i+=4){
328 int nnz = nnzc[ scan8[i] ];
330 if(nnz==1 && block[i*16]) h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
331 else h264_idct8_add_altivec(dst + block_offset[i], block + i*16, stride);
336 static void h264_idct_add8_altivec(uint8_t **dest, const int *block_offset,
337 int16_t *block, int stride,
338 const uint8_t nnzc[15 * 8])
341 for (j = 1; j < 3; j++) {
342 for(i = j * 16; i < j * 16 + 4; i++){
344 h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
346 h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
351 #define transpose4x16(r0, r1, r2, r3) { \
352 register vec_u8 r4; \
353 register vec_u8 r5; \
354 register vec_u8 r6; \
355 register vec_u8 r7; \
357 r4 = vec_mergeh(r0, r2); /*0, 2 set 0*/ \
358 r5 = vec_mergel(r0, r2); /*0, 2 set 1*/ \
359 r6 = vec_mergeh(r1, r3); /*1, 3 set 0*/ \
360 r7 = vec_mergel(r1, r3); /*1, 3 set 1*/ \
362 r0 = vec_mergeh(r4, r6); /*all set 0*/ \
363 r1 = vec_mergel(r4, r6); /*all set 1*/ \
364 r2 = vec_mergeh(r5, r7); /*all set 2*/ \
365 r3 = vec_mergel(r5, r7); /*all set 3*/ \
368 static inline void write16x4(uint8_t *dst, int dst_stride,
369 register vec_u8 r0, register vec_u8 r1,
370 register vec_u8 r2, register vec_u8 r3) {
371 DECLARE_ALIGNED(16, unsigned char, result)[64];
372 uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
373 int int_dst_stride = dst_stride/4;
375 vec_st(r0, 0, result);
376 vec_st(r1, 16, result);
377 vec_st(r2, 32, result);
378 vec_st(r3, 48, result);
379 /* FIXME: there has to be a better way!!!! */
381 *(dst_int+ int_dst_stride) = *(src_int + 1);
382 *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
383 *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
384 *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
385 *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
386 *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
387 *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
388 *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
389 *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
390 *(dst_int+10*int_dst_stride) = *(src_int + 10);
391 *(dst_int+11*int_dst_stride) = *(src_int + 11);
392 *(dst_int+12*int_dst_stride) = *(src_int + 12);
393 *(dst_int+13*int_dst_stride) = *(src_int + 13);
394 *(dst_int+14*int_dst_stride) = *(src_int + 14);
395 *(dst_int+15*int_dst_stride) = *(src_int + 15);
398 /** @brief performs a 6x16 transpose of data in src, and stores it to dst
399 @todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
400 out of unaligned_load() */
401 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
402 register vec_u8 r0 = unaligned_load(0, src); \
403 register vec_u8 r1 = unaligned_load( src_stride, src); \
404 register vec_u8 r2 = unaligned_load(2* src_stride, src); \
405 register vec_u8 r3 = unaligned_load(3* src_stride, src); \
406 register vec_u8 r4 = unaligned_load(4* src_stride, src); \
407 register vec_u8 r5 = unaligned_load(5* src_stride, src); \
408 register vec_u8 r6 = unaligned_load(6* src_stride, src); \
409 register vec_u8 r7 = unaligned_load(7* src_stride, src); \
410 register vec_u8 r14 = unaligned_load(14*src_stride, src); \
411 register vec_u8 r15 = unaligned_load(15*src_stride, src); \
413 r8 = unaligned_load( 8*src_stride, src); \
414 r9 = unaligned_load( 9*src_stride, src); \
415 r10 = unaligned_load(10*src_stride, src); \
416 r11 = unaligned_load(11*src_stride, src); \
417 r12 = unaligned_load(12*src_stride, src); \
418 r13 = unaligned_load(13*src_stride, src); \
420 /*Merge first pairs*/ \
421 r0 = vec_mergeh(r0, r8); /*0, 8*/ \
422 r1 = vec_mergeh(r1, r9); /*1, 9*/ \
423 r2 = vec_mergeh(r2, r10); /*2,10*/ \
424 r3 = vec_mergeh(r3, r11); /*3,11*/ \
425 r4 = vec_mergeh(r4, r12); /*4,12*/ \
426 r5 = vec_mergeh(r5, r13); /*5,13*/ \
427 r6 = vec_mergeh(r6, r14); /*6,14*/ \
428 r7 = vec_mergeh(r7, r15); /*7,15*/ \
430 /*Merge second pairs*/ \
431 r8 = vec_mergeh(r0, r4); /*0,4, 8,12 set 0*/ \
432 r9 = vec_mergel(r0, r4); /*0,4, 8,12 set 1*/ \
433 r10 = vec_mergeh(r1, r5); /*1,5, 9,13 set 0*/ \
434 r11 = vec_mergel(r1, r5); /*1,5, 9,13 set 1*/ \
435 r12 = vec_mergeh(r2, r6); /*2,6,10,14 set 0*/ \
436 r13 = vec_mergel(r2, r6); /*2,6,10,14 set 1*/ \
437 r14 = vec_mergeh(r3, r7); /*3,7,11,15 set 0*/ \
438 r15 = vec_mergel(r3, r7); /*3,7,11,15 set 1*/ \
441 r0 = vec_mergeh(r8, r12); /*0,2,4,6,8,10,12,14 set 0*/ \
442 r1 = vec_mergel(r8, r12); /*0,2,4,6,8,10,12,14 set 1*/ \
443 r2 = vec_mergeh(r9, r13); /*0,2,4,6,8,10,12,14 set 2*/ \
444 r4 = vec_mergeh(r10, r14); /*1,3,5,7,9,11,13,15 set 0*/ \
445 r5 = vec_mergel(r10, r14); /*1,3,5,7,9,11,13,15 set 1*/ \
446 r6 = vec_mergeh(r11, r15); /*1,3,5,7,9,11,13,15 set 2*/ \
447 /* Don't need to compute 3 and 7*/ \
450 r8 = vec_mergeh(r0, r4); /*all set 0*/ \
451 r9 = vec_mergel(r0, r4); /*all set 1*/ \
452 r10 = vec_mergeh(r1, r5); /*all set 2*/ \
453 r11 = vec_mergel(r1, r5); /*all set 3*/ \
454 r12 = vec_mergeh(r2, r6); /*all set 4*/ \
455 r13 = vec_mergel(r2, r6); /*all set 5*/ \
456 /* Don't need to compute 14 and 15*/ \
460 // out: o = |x-y| < a
461 static inline vec_u8 diff_lt_altivec ( register vec_u8 x,
465 register vec_u8 diff = vec_subs(x, y);
466 register vec_u8 diffneg = vec_subs(y, x);
467 register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */
468 o = (vec_u8)vec_cmplt(o, a);
472 static inline vec_u8 h264_deblock_mask ( register vec_u8 p0,
476 register vec_u8 alpha,
477 register vec_u8 beta) {
479 register vec_u8 mask;
480 register vec_u8 tempmask;
482 mask = diff_lt_altivec(p0, q0, alpha);
483 tempmask = diff_lt_altivec(p1, p0, beta);
484 mask = vec_and(mask, tempmask);
485 tempmask = diff_lt_altivec(q1, q0, beta);
486 mask = vec_and(mask, tempmask);
491 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
492 static inline vec_u8 h264_deblock_q1(register vec_u8 p0,
496 register vec_u8 tc0) {
498 register vec_u8 average = vec_avg(p0, q0);
499 register vec_u8 temp;
500 register vec_u8 uncliped;
501 register vec_u8 ones;
504 register vec_u8 newp1;
506 temp = vec_xor(average, p2);
507 average = vec_avg(average, p2); /*avg(p2, avg(p0, q0)) */
508 ones = vec_splat_u8(1);
509 temp = vec_and(temp, ones); /*(p2^avg(p0, q0)) & 1 */
510 uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
511 max = vec_adds(p1, tc0);
512 min = vec_subs(p1, tc0);
513 newp1 = vec_max(min, uncliped);
514 newp1 = vec_min(max, newp1);
518 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) { \
520 const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4)); \
522 register vec_u8 pq0bit = vec_xor(p0,q0); \
523 register vec_u8 q1minus; \
524 register vec_u8 p0minus; \
525 register vec_u8 stage1; \
526 register vec_u8 stage2; \
527 register vec_u8 vec160; \
528 register vec_u8 delta; \
529 register vec_u8 deltaneg; \
531 q1minus = vec_nor(q1, q1); /* 255 - q1 */ \
532 stage1 = vec_avg(p1, q1minus); /* (p1 - q1 + 256)>>1 */ \
533 stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */ \
534 p0minus = vec_nor(p0, p0); /* 255 - p0 */ \
535 stage1 = vec_avg(q0, p0minus); /* (q0 - p0 + 256)>>1 */ \
536 pq0bit = vec_and(pq0bit, vec_splat_u8(1)); \
537 stage2 = vec_avg(stage2, pq0bit); /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \
538 stage2 = vec_adds(stage2, stage1); /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
539 vec160 = vec_ld(0, &A0v); \
540 deltaneg = vec_subs(vec160, stage2); /* -d */ \
541 delta = vec_subs(stage2, vec160); /* d */ \
542 deltaneg = vec_min(tc0masked, deltaneg); \
543 delta = vec_min(tc0masked, delta); \
544 p0 = vec_subs(p0, deltaneg); \
545 q0 = vec_subs(q0, delta); \
546 p0 = vec_adds(p0, delta); \
547 q0 = vec_adds(q0, deltaneg); \
550 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) { \
551 DECLARE_ALIGNED(16, unsigned char, temp)[16]; \
552 register vec_u8 alphavec; \
553 register vec_u8 betavec; \
554 register vec_u8 mask; \
555 register vec_u8 p1mask; \
556 register vec_u8 q1mask; \
557 register vector signed char tc0vec; \
558 register vec_u8 finaltc0; \
559 register vec_u8 tc0masked; \
560 register vec_u8 newp1; \
561 register vec_u8 newq1; \
565 alphavec = vec_ld(0, temp); \
566 betavec = vec_splat(alphavec, 0x1); \
567 alphavec = vec_splat(alphavec, 0x0); \
568 mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */ \
570 AV_COPY32(temp, tc0); \
571 tc0vec = vec_ld(0, (signed char*)temp); \
572 tc0vec = vec_mergeh(tc0vec, tc0vec); \
573 tc0vec = vec_mergeh(tc0vec, tc0vec); \
574 mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1))); /* if tc0[i] >= 0 */ \
575 finaltc0 = vec_and((vec_u8)tc0vec, mask); /* tc = tc0 */ \
577 p1mask = diff_lt_altivec(p2, p0, betavec); \
578 p1mask = vec_and(p1mask, mask); /* if ( |p2 - p0| < beta) */ \
579 tc0masked = vec_and(p1mask, (vec_u8)tc0vec); \
580 finaltc0 = vec_sub(finaltc0, p1mask); /* tc++ */ \
581 newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked); \
584 q1mask = diff_lt_altivec(q2, q0, betavec); \
585 q1mask = vec_and(q1mask, mask); /* if ( |q2 - q0| < beta ) */\
586 tc0masked = vec_and(q1mask, (vec_u8)tc0vec); \
587 finaltc0 = vec_sub(finaltc0, q1mask); /* tc++ */ \
588 newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked); \
591 h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0); \
596 static void h264_v_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
598 if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
599 register vec_u8 p2 = vec_ld(-3*stride, pix);
600 register vec_u8 p1 = vec_ld(-2*stride, pix);
601 register vec_u8 p0 = vec_ld(-1*stride, pix);
602 register vec_u8 q0 = vec_ld(0, pix);
603 register vec_u8 q1 = vec_ld(stride, pix);
604 register vec_u8 q2 = vec_ld(2*stride, pix);
605 h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
606 vec_st(p1, -2*stride, pix);
607 vec_st(p0, -1*stride, pix);
609 vec_st(q1, stride, pix);
613 static void h264_h_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
615 register vec_u8 line0, line1, line2, line3, line4, line5;
616 if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
618 readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
619 h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
620 transpose4x16(line1, line2, line3, line4);
621 write16x4(pix-2, stride, line1, line2, line3, line4);
624 static av_always_inline
625 void weight_h264_W_altivec(uint8_t *block, int stride, int height,
626 int log2_denom, int weight, int offset, int w)
630 vec_s16 vtemp, vweight, voffset, v0, v1;
632 DECLARE_ALIGNED(16, int32_t, temp)[4];
635 offset <<= log2_denom;
636 if(log2_denom) offset += 1<<(log2_denom-1);
637 temp[0] = log2_denom;
641 vtemp = (vec_s16)vec_ld(0, temp);
642 vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
643 vweight = vec_splat(vtemp, 3);
644 voffset = vec_splat(vtemp, 5);
645 aligned = !((unsigned long)block & 0xf);
647 for (y = 0; y < height; y++) {
648 vblock = vec_ld(0, block);
650 v0 = (vec_s16)vec_mergeh(zero_u8v, vblock);
651 v1 = (vec_s16)vec_mergel(zero_u8v, vblock);
653 if (w == 16 || aligned) {
654 v0 = vec_mladd(v0, vweight, zero_s16v);
655 v0 = vec_adds(v0, voffset);
656 v0 = vec_sra(v0, vlog2_denom);
658 if (w == 16 || !aligned) {
659 v1 = vec_mladd(v1, vweight, zero_s16v);
660 v1 = vec_adds(v1, voffset);
661 v1 = vec_sra(v1, vlog2_denom);
663 vblock = vec_packsu(v0, v1);
664 vec_st(vblock, 0, block);
670 static av_always_inline
671 void biweight_h264_W_altivec(uint8_t *dst, uint8_t *src, int stride, int height,
672 int log2_denom, int weightd, int weights, int offset, int w)
674 int y, dst_aligned, src_aligned;
676 vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
678 DECLARE_ALIGNED(16, int32_t, temp)[4];
681 offset = ((offset + 1) | 1) << log2_denom;
682 temp[0] = log2_denom+1;
687 vtemp = (vec_s16)vec_ld(0, temp);
688 vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
689 vweights = vec_splat(vtemp, 3);
690 vweightd = vec_splat(vtemp, 5);
691 voffset = vec_splat(vtemp, 7);
692 dst_aligned = !((unsigned long)dst & 0xf);
693 src_aligned = !((unsigned long)src & 0xf);
695 for (y = 0; y < height; y++) {
696 vdst = vec_ld(0, dst);
697 vsrc = vec_ld(0, src);
699 v0 = (vec_s16)vec_mergeh(zero_u8v, vdst);
700 v1 = (vec_s16)vec_mergel(zero_u8v, vdst);
701 v2 = (vec_s16)vec_mergeh(zero_u8v, vsrc);
702 v3 = (vec_s16)vec_mergel(zero_u8v, vsrc);
711 if (w == 16 || dst_aligned) {
712 v0 = vec_mladd(v0, vweightd, zero_s16v);
713 v2 = vec_mladd(v2, vweights, zero_s16v);
715 v0 = vec_adds(v0, voffset);
716 v0 = vec_adds(v0, v2);
717 v0 = vec_sra(v0, vlog2_denom);
719 if (w == 16 || !dst_aligned) {
720 v1 = vec_mladd(v1, vweightd, zero_s16v);
721 v3 = vec_mladd(v3, vweights, zero_s16v);
723 v1 = vec_adds(v1, voffset);
724 v1 = vec_adds(v1, v3);
725 v1 = vec_sra(v1, vlog2_denom);
727 vdst = vec_packsu(v0, v1);
728 vec_st(vdst, 0, dst);
735 #define H264_WEIGHT(W) \
736 static void weight_h264_pixels ## W ## _altivec(uint8_t *block, int stride, int height, \
737 int log2_denom, int weight, int offset) \
739 weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \
741 static void biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, int stride, int height, \
742 int log2_denom, int weightd, int weights, int offset) \
744 biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \
749 #endif /* HAVE_ALTIVEC */
751 av_cold void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth,
752 const int chroma_format_idc)
754 #if HAVE_ALTIVEC && HAVE_BIGENDIAN
755 if (!PPC_ALTIVEC(av_get_cpu_flags()))
758 if (bit_depth == 8) {
759 c->h264_idct_add = h264_idct_add_altivec;
760 if (chroma_format_idc <= 1)
761 c->h264_idct_add8 = h264_idct_add8_altivec;
762 c->h264_idct_add16 = h264_idct_add16_altivec;
763 c->h264_idct_add16intra = h264_idct_add16intra_altivec;
764 c->h264_idct_dc_add= h264_idct_dc_add_altivec;
765 c->h264_idct8_dc_add = h264_idct8_dc_add_altivec;
766 c->h264_idct8_add = h264_idct8_add_altivec;
767 c->h264_idct8_add4 = h264_idct8_add4_altivec;
768 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec;
769 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec;
771 c->weight_h264_pixels_tab[0] = weight_h264_pixels16_altivec;
772 c->weight_h264_pixels_tab[1] = weight_h264_pixels8_altivec;
773 c->biweight_h264_pixels_tab[0] = biweight_h264_pixels16_altivec;
774 c->biweight_h264_pixels_tab[1] = biweight_h264_pixels8_altivec;
776 #endif /* HAVE_ALTIVEC */