2 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
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
21 #include "libavutil/cpu.h"
22 #include "libavutil/x86_cpu.h"
23 #include "libavcodec/h264dsp.h"
24 #include "dsputil_mmx.h"
26 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_3_1 ) = 0x0103010301030103ULL;
28 /***********************************/
31 void ff_h264_idct_add_mmx (uint8_t *dst, int16_t *block, int stride);
32 void ff_h264_idct8_add_mmx (uint8_t *dst, int16_t *block, int stride);
33 void ff_h264_idct8_add_sse2 (uint8_t *dst, int16_t *block, int stride);
34 void ff_h264_idct_dc_add_mmx2 (uint8_t *dst, int16_t *block, int stride);
35 void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride);
37 void ff_h264_idct_add16_mmx (uint8_t *dst, const int *block_offset,
38 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
39 void ff_h264_idct8_add4_mmx (uint8_t *dst, const int *block_offset,
40 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
41 void ff_h264_idct_add16_mmx2 (uint8_t *dst, const int *block_offset,
42 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
43 void ff_h264_idct_add16intra_mmx (uint8_t *dst, const int *block_offset,
44 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
45 void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset,
46 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
47 void ff_h264_idct8_add4_mmx2 (uint8_t *dst, const int *block_offset,
48 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
49 void ff_h264_idct8_add4_sse2 (uint8_t *dst, const int *block_offset,
50 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
51 void ff_h264_idct_add8_mmx (uint8_t **dest, const int *block_offset,
52 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
53 void ff_h264_idct_add8_mmx2 (uint8_t **dest, const int *block_offset,
54 DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
56 void ff_h264_idct_add16_sse2 (uint8_t *dst, const int *block_offset, DCTELEM *block,
57 int stride, const uint8_t nnzc[6*8]);
58 void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block,
59 int stride, const uint8_t nnzc[6*8]);
60 void ff_h264_idct_add8_sse2 (uint8_t **dest, const int *block_offset, DCTELEM *block,
61 int stride, const uint8_t nnzc[6*8]);
62 void ff_h264_luma_dc_dequant_idct_mmx (DCTELEM *output, DCTELEM *input, int qmul);
63 void ff_h264_luma_dc_dequant_idct_sse2(DCTELEM *output, DCTELEM *input, int qmul);
65 /***********************************/
68 #define h264_loop_filter_strength_iteration_mmx2(bS, nz, ref, mv, bidir, edges, step, mask_mv, dir, d_idx, mask_dir) \
72 for( b_idx=0; b_idx<edges; b_idx+=step ) { \
75 "pxor %%mm0, %%mm0 \n\t" \
78 if(!(mask_mv & b_idx)) { \
81 "movd %a3(%0,%2), %%mm2 \n" \
82 "punpckldq %a4(%0,%2), %%mm2 \n" /* { ref0[bn], ref1[bn] } */ \
83 "pshufw $0x44, 12(%0,%2), %%mm0 \n" /* { ref0[b], ref0[b] } */ \
84 "pshufw $0x44, 52(%0,%2), %%mm1 \n" /* { ref1[b], ref1[b] } */ \
85 "pshufw $0x4E, %%mm2, %%mm3 \n" \
86 "psubb %%mm2, %%mm0 \n" /* { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] } */ \
87 "psubb %%mm3, %%mm1 \n" /* { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] } */ \
89 "por %%mm1, %%mm0 \n" \
90 "movq %a5(%1,%2,4), %%mm1 \n" \
91 "movq %a6(%1,%2,4), %%mm2 \n" \
92 "movq %%mm1, %%mm3 \n" \
93 "movq %%mm2, %%mm4 \n" \
94 "psubw 48(%1,%2,4), %%mm1 \n" \
95 "psubw 56(%1,%2,4), %%mm2 \n" \
96 "psubw 208(%1,%2,4), %%mm3 \n" \
97 "psubw 216(%1,%2,4), %%mm4 \n" \
98 "packsswb %%mm2, %%mm1 \n" \
99 "packsswb %%mm4, %%mm3 \n" \
100 "paddb %%mm6, %%mm1 \n" \
101 "paddb %%mm6, %%mm3 \n" \
102 "psubusb %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
103 "psubusb %%mm5, %%mm3 \n" \
104 "packsswb %%mm3, %%mm1 \n" \
106 "por %%mm1, %%mm0 \n" \
107 "movq %a7(%1,%2,4), %%mm1 \n" \
108 "movq %a8(%1,%2,4), %%mm2 \n" \
109 "movq %%mm1, %%mm3 \n" \
110 "movq %%mm2, %%mm4 \n" \
111 "psubw 48(%1,%2,4), %%mm1 \n" \
112 "psubw 56(%1,%2,4), %%mm2 \n" \
113 "psubw 208(%1,%2,4), %%mm3 \n" \
114 "psubw 216(%1,%2,4), %%mm4 \n" \
115 "packsswb %%mm2, %%mm1 \n" \
116 "packsswb %%mm4, %%mm3 \n" \
117 "paddb %%mm6, %%mm1 \n" \
118 "paddb %%mm6, %%mm3 \n" \
119 "psubusb %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
120 "psubusb %%mm5, %%mm3 \n" \
121 "packsswb %%mm3, %%mm1 \n" \
123 "pshufw $0x4E, %%mm1, %%mm1 \n" \
124 "por %%mm1, %%mm0 \n" \
125 "pshufw $0x4E, %%mm0, %%mm1 \n" \
126 "pminub %%mm1, %%mm0 \n" \
139 "movd 12(%0,%2), %%mm0 \n" \
140 "psubb %a3(%0,%2), %%mm0 \n" /* ref[b] != ref[bn] */ \
141 "movq 48(%1,%2,4), %%mm1 \n" \
142 "movq 56(%1,%2,4), %%mm2 \n" \
143 "psubw %a4(%1,%2,4), %%mm1 \n" \
144 "psubw %a5(%1,%2,4), %%mm2 \n" \
145 "packsswb %%mm2, %%mm1 \n" \
146 "paddb %%mm6, %%mm1 \n" \
147 "psubusb %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
148 "packsswb %%mm1, %%mm1 \n" \
149 "por %%mm1, %%mm0 \n" \
160 "movd 12(%0,%1), %%mm1 \n" \
161 "por %a2(%0,%1), %%mm1 \n" /* nnz[b] || nnz[bn] */ \
167 "pminub %%mm7, %%mm1 \n" \
168 "pminub %%mm7, %%mm0 \n" \
169 "psllw $1, %%mm1 \n" \
170 "pxor %%mm2, %%mm2 \n" \
171 "pmaxub %%mm0, %%mm1 \n" \
172 "punpcklbw %%mm2, %%mm1 \n" \
173 "movq %%mm1, %a1(%0,%2) \n" \
182 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
183 int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
187 ::"m"(ff_pb_1), "m"(ff_pb_3)
195 "movq %%mm6, %%mm5 \n"
196 "paddb %%mm5, %%mm5 \n"
199 // could do a special case for dir==0 && edges==1, but it only reduces the
200 // average filter time by 1.2%
203 h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir, edges, step, mask_mv1, 1, -8, 0);
204 h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir, 32, 8, mask_mv0, 0, -1, -1);
207 "movq (%0), %%mm0 \n\t"
208 "movq 8(%0), %%mm1 \n\t"
209 "movq 16(%0), %%mm2 \n\t"
210 "movq 24(%0), %%mm3 \n\t"
211 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
212 "movq %%mm0, (%0) \n\t"
213 "movq %%mm3, 8(%0) \n\t"
214 "movq %%mm4, 16(%0) \n\t"
215 "movq %%mm2, 24(%0) \n\t"
221 #define LF_FUNC(DIR, TYPE, DEPTH, OPT) \
222 void ff_deblock_ ## DIR ## _ ## TYPE ## _ ## DEPTH ## _ ## OPT (uint8_t *pix, int stride, \
223 int alpha, int beta, int8_t *tc0);
224 #define LF_IFUNC(DIR, TYPE, DEPTH, OPT) \
225 void ff_deblock_ ## DIR ## _ ## TYPE ## _ ## DEPTH ## _ ## OPT (uint8_t *pix, int stride, \
226 int alpha, int beta);
228 #define LF_FUNCS(type, depth)\
229 LF_FUNC (h, chroma, depth, mmxext)\
230 LF_IFUNC(h, chroma_intra, depth, mmxext)\
231 LF_FUNC (v, chroma, depth, mmxext)\
232 LF_IFUNC(v, chroma_intra, depth, mmxext)\
233 LF_FUNC (h, luma, depth, mmxext)\
234 LF_IFUNC(h, luma_intra, depth, mmxext)\
235 LF_FUNC (h, luma, depth, sse2)\
236 LF_IFUNC(h, luma_intra, depth, sse2)\
237 LF_FUNC (v, luma, depth, sse2)\
238 LF_IFUNC(v, luma_intra, depth, sse2)\
239 LF_FUNC (h, chroma, depth, sse2)\
240 LF_IFUNC(h, chroma_intra, depth, sse2)\
241 LF_FUNC (v, chroma, depth, sse2)\
242 LF_IFUNC(v, chroma_intra, depth, sse2)\
243 LF_FUNC (h, luma, depth, avx)\
244 LF_IFUNC(h, luma_intra, depth, avx)\
245 LF_FUNC (v, luma, depth, avx)\
246 LF_IFUNC(v, luma_intra, depth, avx)\
247 LF_FUNC (h, chroma, depth, avx)\
248 LF_IFUNC(h, chroma_intra, depth, avx)\
249 LF_FUNC (v, chroma, depth, avx)\
250 LF_IFUNC(v, chroma_intra, depth, avx)
252 LF_FUNCS( uint8_t, 8)
253 LF_FUNCS(uint16_t, 10)
256 LF_FUNC (v8, luma, 8, mmxext)
257 static void ff_deblock_v_luma_8_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
259 if((tc0[0] & tc0[1]) >= 0)
260 ff_deblock_v8_luma_8_mmxext(pix+0, stride, alpha, beta, tc0);
261 if((tc0[2] & tc0[3]) >= 0)
262 ff_deblock_v8_luma_8_mmxext(pix+8, stride, alpha, beta, tc0+2);
264 LF_IFUNC(v8, luma_intra, 8, mmxext)
265 static void ff_deblock_v_luma_intra_8_mmxext(uint8_t *pix, int stride, int alpha, int beta)
267 ff_deblock_v8_luma_intra_8_mmxext(pix+0, stride, alpha, beta);
268 ff_deblock_v8_luma_intra_8_mmxext(pix+8, stride, alpha, beta);
270 #endif /* ARCH_X86_32 */
272 LF_FUNC (v, luma, 10, mmxext)
273 LF_IFUNC(v, luma_intra, 10, mmxext)
275 /***********************************/
276 /* weighted prediction */
278 #define H264_WEIGHT(W, H, OPT) \
279 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
280 int stride, int log2_denom, int weight, int offset);
282 #define H264_BIWEIGHT(W, H, OPT) \
283 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
284 uint8_t *src, int stride, int log2_denom, int weightd, \
285 int weights, int offset);
287 #define H264_BIWEIGHT_MMX(W,H) \
288 H264_WEIGHT (W, H, mmx2) \
289 H264_BIWEIGHT(W, H, mmx2)
291 #define H264_BIWEIGHT_MMX_SSE(W,H) \
292 H264_BIWEIGHT_MMX(W, H) \
293 H264_WEIGHT (W, H, sse2) \
294 H264_BIWEIGHT (W, H, sse2) \
295 H264_BIWEIGHT (W, H, ssse3)
297 H264_BIWEIGHT_MMX_SSE(16, 16)
298 H264_BIWEIGHT_MMX_SSE(16, 8)
299 H264_BIWEIGHT_MMX_SSE( 8, 16)
300 H264_BIWEIGHT_MMX_SSE( 8, 8)
301 H264_BIWEIGHT_MMX_SSE( 8, 4)
302 H264_BIWEIGHT_MMX ( 4, 8)
303 H264_BIWEIGHT_MMX ( 4, 4)
304 H264_BIWEIGHT_MMX ( 4, 2)
306 void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth)
308 int mm_flags = av_get_cpu_flags();
310 if (bit_depth == 8) {
311 if (mm_flags & AV_CPU_FLAG_MMX2) {
312 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
315 if (mm_flags & AV_CPU_FLAG_MMX) {
317 c->h264_idct_add= ff_h264_idct_add_mmx;
318 c->h264_idct8_dc_add=
319 c->h264_idct8_add= ff_h264_idct8_add_mmx;
321 c->h264_idct_add16 = ff_h264_idct_add16_mmx;
322 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx;
323 c->h264_idct_add8 = ff_h264_idct_add8_mmx;
324 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx;
325 c->h264_luma_dc_dequant_idct= ff_h264_luma_dc_dequant_idct_mmx;
327 if (mm_flags & AV_CPU_FLAG_MMX2) {
328 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
329 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
330 c->h264_idct_add16 = ff_h264_idct_add16_mmx2;
331 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx2;
332 c->h264_idct_add8 = ff_h264_idct_add8_mmx2;
333 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2;
335 c->h264_v_loop_filter_chroma= ff_deblock_v_chroma_8_mmxext;
336 c->h264_h_loop_filter_chroma= ff_deblock_h_chroma_8_mmxext;
337 c->h264_v_loop_filter_chroma_intra= ff_deblock_v_chroma_intra_8_mmxext;
338 c->h264_h_loop_filter_chroma_intra= ff_deblock_h_chroma_intra_8_mmxext;
340 c->h264_v_loop_filter_luma= ff_deblock_v_luma_8_mmxext;
341 c->h264_h_loop_filter_luma= ff_deblock_h_luma_8_mmxext;
342 c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_mmxext;
343 c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_mmxext;
345 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
346 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
347 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
348 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
349 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
350 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
351 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
352 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
354 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
355 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
356 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
357 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
358 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
359 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
360 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
361 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
363 if (mm_flags&AV_CPU_FLAG_SSE2) {
364 c->h264_idct8_add = ff_h264_idct8_add_sse2;
365 c->h264_idct8_add4= ff_h264_idct8_add4_sse2;
366 c->h264_luma_dc_dequant_idct= ff_h264_luma_dc_dequant_idct_sse2;
368 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2;
369 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2;
370 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2;
371 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2;
372 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2;
374 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2;
375 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2;
376 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2;
377 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2;
378 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2;
380 #if HAVE_ALIGNED_STACK
381 c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_sse2;
382 c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_sse2;
383 c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_sse2;
384 c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_sse2;
387 c->h264_idct_add16 = ff_h264_idct_add16_sse2;
388 c->h264_idct_add8 = ff_h264_idct_add8_sse2;
389 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2;
391 if (mm_flags&AV_CPU_FLAG_SSSE3) {
392 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3;
393 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3;
394 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3;
395 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3;
396 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3;
398 if (mm_flags&AV_CPU_FLAG_AVX) {
399 #if HAVE_ALIGNED_STACK
400 c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_avx;
401 c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_avx;
402 c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_avx;
403 c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_avx;
409 } else if (bit_depth == 10) {
411 if (mm_flags & AV_CPU_FLAG_MMX) {
412 if (mm_flags & AV_CPU_FLAG_MMX2) {
414 c->h264_v_loop_filter_chroma= ff_deblock_v_chroma_10_mmxext;
415 c->h264_v_loop_filter_chroma_intra= ff_deblock_v_chroma_intra_10_mmxext;
416 c->h264_v_loop_filter_luma= ff_deblock_v_luma_10_mmxext;
417 c->h264_h_loop_filter_luma= ff_deblock_h_luma_10_mmxext;
418 c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_mmxext;
419 c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_mmxext;
421 if (mm_flags&AV_CPU_FLAG_SSE2) {
422 c->h264_v_loop_filter_chroma= ff_deblock_v_chroma_10_sse2;
423 c->h264_v_loop_filter_chroma_intra= ff_deblock_v_chroma_intra_10_sse2;
424 #if HAVE_ALIGNED_STACK
425 c->h264_v_loop_filter_luma = ff_deblock_v_luma_10_sse2;
426 c->h264_h_loop_filter_luma = ff_deblock_h_luma_10_sse2;
427 c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_sse2;
428 c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_sse2;
431 if (mm_flags&AV_CPU_FLAG_AVX) {
432 c->h264_v_loop_filter_chroma= ff_deblock_v_chroma_10_avx;
433 c->h264_v_loop_filter_chroma_intra= ff_deblock_v_chroma_intra_10_avx;
434 #if HAVE_ALIGNED_STACK
435 c->h264_v_loop_filter_luma = ff_deblock_v_luma_10_avx;
436 c->h264_h_loop_filter_luma = ff_deblock_h_luma_10_avx;
437 c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_avx;
438 c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_avx;