FFmpeg
•All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Modules Pages
vp9dsp_init.c
Go to the documentation of this file.
1 /*
2  * VP9 SIMD optimizations
3  *
4  * Copyright (c) 2013 Ronald S. Bultje <rsbultje gmail com>
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 "libavutil/attributes.h"
24 #include "libavutil/cpu.h"
25 #include "libavutil/x86/cpu.h"
26 #include "libavcodec/vp9dsp.h"
27 #include "libavcodec/x86/vp9dsp_init.h"
28 
29 #if HAVE_X86ASM
30 
31 decl_fpel_func(put, 4, , mmx);
32 decl_fpel_func(put, 8, , mmx);
33 decl_fpel_func(put, 16, , sse);
34 decl_fpel_func(put, 32, , sse);
35 decl_fpel_func(put, 64, , sse);
36 decl_fpel_func(avg, 4, _8, mmxext);
37 decl_fpel_func(avg, 8, _8, mmxext);
38 decl_fpel_func(avg, 16, _8, sse2);
39 decl_fpel_func(avg, 32, _8, sse2);
40 decl_fpel_func(avg, 64, _8, sse2);
41 decl_fpel_func(put, 32, , avx);
42 decl_fpel_func(put, 64, , avx);
43 decl_fpel_func(avg, 32, _8, avx2);
44 decl_fpel_func(avg, 64, _8, avx2);
45 
46 decl_mc_funcs(4, mmxext, int16_t, 8, 8);
47 decl_mc_funcs(8, sse2, int16_t, 8, 8);
48 decl_mc_funcs(4, ssse3, int8_t, 32, 8);
49 decl_mc_funcs(8, ssse3, int8_t, 32, 8);
50 #if ARCH_X86_64
51 decl_mc_funcs(16, ssse3, int8_t, 32, 8);
52 decl_mc_funcs(32, avx2, int8_t, 32, 8);
53 #endif
54 
55 mc_rep_funcs(16, 8, 8, sse2, int16_t, 8, 8)
56 #if ARCH_X86_32
57 mc_rep_funcs(16, 8, 8, ssse3, int8_t, 32, 8)
58 #endif
59 mc_rep_funcs(32, 16, 16, sse2, int16_t, 8, 8)
60 mc_rep_funcs(32, 16, 16, ssse3, int8_t, 32, 8)
61 mc_rep_funcs(64, 32, 32, sse2, int16_t, 8, 8)
62 mc_rep_funcs(64, 32, 32, ssse3, int8_t, 32, 8)
63 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
64 mc_rep_funcs(64, 32, 32, avx2, int8_t, 32, 8)
65 #endif
66 
67 extern const int8_t ff_filters_ssse3[3][15][4][32];
68 extern const int16_t ff_filters_sse2[3][15][8][8];
69 
70 filters_8tap_2d_fn2(put, 16, 8, 1, mmxext, sse2, sse2)
71 filters_8tap_2d_fn2(avg, 16, 8, 1, mmxext, sse2, sse2)
72 filters_8tap_2d_fn2(put, 16, 8, 1, ssse3, ssse3, ssse3)
73 filters_8tap_2d_fn2(avg, 16, 8, 1, ssse3, ssse3, ssse3)
74 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
75 filters_8tap_2d_fn(put, 64, 32, 8, 1, avx2, ssse3)
76 filters_8tap_2d_fn(put, 32, 32, 8, 1, avx2, ssse3)
77 filters_8tap_2d_fn(avg, 64, 32, 8, 1, avx2, ssse3)
78 filters_8tap_2d_fn(avg, 32, 32, 8, 1, avx2, ssse3)
79 #endif
80 
81 filters_8tap_1d_fn3(put, 8, mmxext, sse2, sse2)
82 filters_8tap_1d_fn3(avg, 8, mmxext, sse2, sse2)
83 filters_8tap_1d_fn3(put, 8, ssse3, ssse3, ssse3)
84 filters_8tap_1d_fn3(avg, 8, ssse3, ssse3, ssse3)
85 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
86 filters_8tap_1d_fn2(put, 64, 8, avx2, ssse3)
87 filters_8tap_1d_fn2(put, 32, 8, avx2, ssse3)
88 filters_8tap_1d_fn2(avg, 64, 8, avx2, ssse3)
89 filters_8tap_1d_fn2(avg, 32, 8, avx2, ssse3)
90 #endif
91 
92 #define itxfm_func(typea, typeb, size, opt) \
93 void ff_vp9_##typea##_##typeb##_##size##x##size##_add_##opt(uint8_t *dst, ptrdiff_t stride, \
94  int16_t *block, int eob)
95 #define itxfm_funcs(size, opt) \
96 itxfm_func(idct, idct, size, opt); \
97 itxfm_func(iadst, idct, size, opt); \
98 itxfm_func(idct, iadst, size, opt); \
99 itxfm_func(iadst, iadst, size, opt)
100 
101 itxfm_func(idct, idct, 4, mmxext);
102 itxfm_func(idct, iadst, 4, sse2);
103 itxfm_func(iadst, idct, 4, sse2);
104 itxfm_func(iadst, iadst, 4, sse2);
105 itxfm_funcs(4, ssse3);
106 itxfm_funcs(8, sse2);
107 itxfm_funcs(8, ssse3);
108 itxfm_funcs(8, avx);
109 itxfm_funcs(16, sse2);
110 itxfm_funcs(16, ssse3);
111 itxfm_funcs(16, avx);
112 itxfm_func(idct, idct, 32, sse2);
113 itxfm_func(idct, idct, 32, ssse3);
114 itxfm_func(idct, idct, 32, avx);
115 itxfm_func(iwht, iwht, 4, mmx);
116 itxfm_funcs(16, avx2);
117 itxfm_funcs(16, avx512icl);
118 itxfm_func(idct, idct, 32, avx2);
119 itxfm_func(idct, idct, 32, avx512icl);
120 
121 #undef itxfm_func
122 #undef itxfm_funcs
123 
124 #define lpf_funcs(size1, size2, opt) \
125 void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
126  int E, int I, int H); \
127 void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
128  int E, int I, int H)
129 
130 lpf_funcs(4, 8, mmxext);
131 lpf_funcs(8, 8, mmxext);
132 lpf_funcs(16, 16, sse2);
133 lpf_funcs(16, 16, ssse3);
134 lpf_funcs(16, 16, avx);
135 lpf_funcs(44, 16, sse2);
136 lpf_funcs(44, 16, ssse3);
137 lpf_funcs(44, 16, avx);
138 lpf_funcs(84, 16, sse2);
139 lpf_funcs(84, 16, ssse3);
140 lpf_funcs(84, 16, avx);
141 lpf_funcs(48, 16, sse2);
142 lpf_funcs(48, 16, ssse3);
143 lpf_funcs(48, 16, avx);
144 lpf_funcs(88, 16, sse2);
145 lpf_funcs(88, 16, ssse3);
146 lpf_funcs(88, 16, avx);
147 
148 #undef lpf_funcs
149 
150 #define ipred_func(size, type, opt) \
151 void ff_vp9_ipred_##type##_##size##x##size##_##opt(uint8_t *dst, ptrdiff_t stride, \
152  const uint8_t *l, const uint8_t *a)
153 
154 ipred_func(8, v, mmx);
155 
156 #define ipred_dc_funcs(size, opt) \
157 ipred_func(size, dc, opt); \
158 ipred_func(size, dc_left, opt); \
159 ipred_func(size, dc_top, opt)
160 
161 ipred_dc_funcs(4, mmxext);
162 ipred_dc_funcs(8, mmxext);
163 
164 #define ipred_dir_tm_funcs(size, opt) \
165 ipred_func(size, tm, opt); \
166 ipred_func(size, dl, opt); \
167 ipred_func(size, dr, opt); \
168 ipred_func(size, hd, opt); \
169 ipred_func(size, hu, opt); \
170 ipred_func(size, vl, opt); \
171 ipred_func(size, vr, opt)
172 
173 ipred_dir_tm_funcs(4, mmxext);
174 
175 ipred_func(16, v, sse);
176 ipred_func(32, v, sse);
177 
178 ipred_dc_funcs(16, sse2);
179 ipred_dc_funcs(32, sse2);
180 
181 #define ipred_dir_tm_h_funcs(size, opt) \
182 ipred_dir_tm_funcs(size, opt); \
183 ipred_func(size, h, opt)
184 
185 ipred_dir_tm_h_funcs(8, sse2);
186 ipred_dir_tm_h_funcs(16, sse2);
187 ipred_dir_tm_h_funcs(32, sse2);
188 
189 ipred_func(4, h, sse2);
190 
191 #define ipred_all_funcs(size, opt) \
192 ipred_dc_funcs(size, opt); \
193 ipred_dir_tm_h_funcs(size, opt)
194 
195 // FIXME hd/vl_4x4_ssse3 does not exist
196 ipred_all_funcs(4, ssse3);
197 ipred_all_funcs(8, ssse3);
198 ipred_all_funcs(16, ssse3);
199 ipred_all_funcs(32, ssse3);
200 
201 ipred_dir_tm_h_funcs(8, avx);
202 ipred_dir_tm_h_funcs(16, avx);
203 ipred_dir_tm_h_funcs(32, avx);
204 
205 ipred_func(32, v, avx);
206 
207 ipred_dc_funcs(32, avx2);
208 ipred_func(32, h, avx2);
209 ipred_func(32, tm, avx2);
210 
211 #undef ipred_func
212 #undef ipred_dir_tm_h_funcs
213 #undef ipred_dir_tm_funcs
214 #undef ipred_dc_funcs
215 
216 #endif /* HAVE_X86ASM */
217 
218 av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
219 {
220 #if HAVE_X86ASM
221  int cpu_flags;
222 
223  if (bpp == 10) {
224  ff_vp9dsp_init_10bpp_x86(dsp, bitexact);
225  return;
226  } else if (bpp == 12) {
227  ff_vp9dsp_init_12bpp_x86(dsp, bitexact);
228  return;
229  }
230 
231  cpu_flags = av_get_cpu_flags();
232 
233 #define init_lpf(opt) do { \
234  dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_##opt; \
235  dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_##opt; \
236  dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_##opt; \
237  dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_##opt; \
238  dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_##opt; \
239  dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_##opt; \
240  dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_##opt; \
241  dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_##opt; \
242  dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_##opt; \
243  dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_##opt; \
244 } while (0)
245 
246 #define init_ipred(sz, opt, t, e) \
247  dsp->intra_pred[TX_##sz##X##sz][e##_PRED] = ff_vp9_ipred_##t##_##sz##x##sz##_##opt
248 
249 #define ff_vp9_ipred_hd_4x4_ssse3 ff_vp9_ipred_hd_4x4_mmxext
250 #define ff_vp9_ipred_vl_4x4_ssse3 ff_vp9_ipred_vl_4x4_mmxext
251 #define init_dir_tm_ipred(sz, opt) do { \
252  init_ipred(sz, opt, dl, DIAG_DOWN_LEFT); \
253  init_ipred(sz, opt, dr, DIAG_DOWN_RIGHT); \
254  init_ipred(sz, opt, hd, HOR_DOWN); \
255  init_ipred(sz, opt, vl, VERT_LEFT); \
256  init_ipred(sz, opt, hu, HOR_UP); \
257  init_ipred(sz, opt, tm, TM_VP8); \
258  init_ipred(sz, opt, vr, VERT_RIGHT); \
259 } while (0)
260 #define init_dir_tm_h_ipred(sz, opt) do { \
261  init_dir_tm_ipred(sz, opt); \
262  init_ipred(sz, opt, h, HOR); \
263 } while (0)
264 #define init_dc_ipred(sz, opt) do { \
265  init_ipred(sz, opt, dc, DC); \
266  init_ipred(sz, opt, dc_left, LEFT_DC); \
267  init_ipred(sz, opt, dc_top, TOP_DC); \
268 } while (0)
269 #define init_all_ipred(sz, opt) do { \
270  init_dc_ipred(sz, opt); \
271  init_dir_tm_h_ipred(sz, opt); \
272 } while (0)
273 
274  if (EXTERNAL_MMX(cpu_flags)) {
275  init_fpel_func(4, 0, 4, put, , mmx);
276  init_fpel_func(3, 0, 8, put, , mmx);
277  if (!bitexact) {
278  dsp->itxfm_add[4 /* lossless */][DCT_DCT] =
279  dsp->itxfm_add[4 /* lossless */][ADST_DCT] =
280  dsp->itxfm_add[4 /* lossless */][DCT_ADST] =
281  dsp->itxfm_add[4 /* lossless */][ADST_ADST] = ff_vp9_iwht_iwht_4x4_add_mmx;
282  }
283  init_ipred(8, mmx, v, VERT);
284  }
285 
286  if (EXTERNAL_MMXEXT(cpu_flags)) {
287  dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_mmxext;
288  dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_mmxext;
289  dsp->loop_filter_8[1][0] = ff_vp9_loop_filter_h_8_8_mmxext;
290  dsp->loop_filter_8[1][1] = ff_vp9_loop_filter_v_8_8_mmxext;
291  init_subpel2(4, 0, 4, put, 8, mmxext);
292  init_subpel2(4, 1, 4, avg, 8, mmxext);
293  init_fpel_func(4, 1, 4, avg, _8, mmxext);
294  init_fpel_func(3, 1, 8, avg, _8, mmxext);
295  dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_mmxext;
296  init_dc_ipred(4, mmxext);
297  init_dc_ipred(8, mmxext);
298  init_dir_tm_ipred(4, mmxext);
299  }
300 
301  if (EXTERNAL_SSE(cpu_flags)) {
302  init_fpel_func(2, 0, 16, put, , sse);
303  init_fpel_func(1, 0, 32, put, , sse);
304  init_fpel_func(0, 0, 64, put, , sse);
305  init_ipred(16, sse, v, VERT);
306  init_ipred(32, sse, v, VERT);
307  }
308 
309  if (EXTERNAL_SSE2(cpu_flags)) {
310  init_subpel3_8to64(0, put, 8, sse2);
311  init_subpel3_8to64(1, avg, 8, sse2);
312  init_fpel_func(2, 1, 16, avg, _8, sse2);
313  init_fpel_func(1, 1, 32, avg, _8, sse2);
314  init_fpel_func(0, 1, 64, avg, _8, sse2);
315  init_lpf(sse2);
316  dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_sse2;
317  dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_sse2;
318  dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_sse2;
319  dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_sse2;
320  dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_sse2;
321  dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_sse2;
322  dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_sse2;
323  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_sse2;
324  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_sse2;
325  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_sse2;
326  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_sse2;
327  dsp->itxfm_add[TX_32X32][ADST_ADST] =
328  dsp->itxfm_add[TX_32X32][ADST_DCT] =
329  dsp->itxfm_add[TX_32X32][DCT_ADST] =
330  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_sse2;
331  init_dc_ipred(16, sse2);
332  init_dc_ipred(32, sse2);
333  init_dir_tm_h_ipred(8, sse2);
334  init_dir_tm_h_ipred(16, sse2);
335  init_dir_tm_h_ipred(32, sse2);
336  init_ipred(4, sse2, h, HOR);
337  }
338 
339  if (EXTERNAL_SSSE3(cpu_flags)) {
340  init_subpel3(0, put, 8, ssse3);
341  init_subpel3(1, avg, 8, ssse3);
342  dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_ssse3;
343  dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_ssse3;
344  dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_ssse3;
345  dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_ssse3;
346  dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_ssse3;
347  dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_ssse3;
348  dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_ssse3;
349  dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_ssse3;
350  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_ssse3;
351  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_ssse3;
352  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_ssse3;
353  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_ssse3;
354  dsp->itxfm_add[TX_32X32][ADST_ADST] =
355  dsp->itxfm_add[TX_32X32][ADST_DCT] =
356  dsp->itxfm_add[TX_32X32][DCT_ADST] =
357  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_ssse3;
358  init_lpf(ssse3);
359  init_all_ipred(4, ssse3);
360  init_all_ipred(8, ssse3);
361  init_all_ipred(16, ssse3);
362  init_all_ipred(32, ssse3);
363  }
364 
365  if (EXTERNAL_AVX(cpu_flags)) {
366  dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_avx;
367  dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_avx;
368  dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_avx;
369  dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_avx;
370  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx;
371  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx;
372  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx;
373  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx;
374  dsp->itxfm_add[TX_32X32][ADST_ADST] =
375  dsp->itxfm_add[TX_32X32][ADST_DCT] =
376  dsp->itxfm_add[TX_32X32][DCT_ADST] =
377  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx;
378  init_lpf(avx);
379  init_dir_tm_h_ipred(8, avx);
380  init_dir_tm_h_ipred(16, avx);
381  init_dir_tm_h_ipred(32, avx);
382  }
383  if (EXTERNAL_AVX_FAST(cpu_flags)) {
384  init_fpel_func(1, 0, 32, put, , avx);
385  init_fpel_func(0, 0, 64, put, , avx);
386  init_ipred(32, avx, v, VERT);
387  }
388 
389  if (EXTERNAL_AVX2_FAST(cpu_flags)) {
390  init_fpel_func(1, 1, 32, avg, _8, avx2);
391  init_fpel_func(0, 1, 64, avg, _8, avx2);
392  if (ARCH_X86_64) {
393 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
394  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx2;
395  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx2;
396  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx2;
397  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx2;
398  dsp->itxfm_add[TX_32X32][ADST_ADST] =
399  dsp->itxfm_add[TX_32X32][ADST_DCT] =
400  dsp->itxfm_add[TX_32X32][DCT_ADST] =
401  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx2;
402  init_subpel3_32_64(0, put, 8, avx2);
403  init_subpel3_32_64(1, avg, 8, avx2);
404 #endif
405  }
406  init_dc_ipred(32, avx2);
407  init_ipred(32, avx2, h, HOR);
408  init_ipred(32, avx2, tm, TM_VP8);
409  }
410 
411 #if ARCH_X86_64
412  if (EXTERNAL_AVX512ICL(cpu_flags)) {
413  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx512icl;
414  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx512icl;
415  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx512icl;
416  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx512icl;
417  dsp->itxfm_add[TX_32X32][ADST_ADST] =
418  dsp->itxfm_add[TX_32X32][ADST_DCT] =
419  dsp->itxfm_add[TX_32X32][DCT_ADST] =
420  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx512icl;
421  }
422 #endif
423 
424 #undef init_fpel
425 #undef init_subpel1
426 #undef init_subpel2
427 #undef init_subpel3
428 
429 #endif /* HAVE_X86ASM */
430 }
cpu.h
EXTERNAL_AVX2_FAST
#define EXTERNAL_AVX2_FAST(flags)
Definition: cpu.h:79
sse
static int sse(const MPVEncContext *const s, const uint8_t *src1, const uint8_t *src2, int w, int h, int stride)
Definition: mpegvideo_enc.c:2779
VP9DSPContext::loop_filter_8
void(* loop_filter_8[3][2])(uint8_t *dst, ptrdiff_t stride, int mb_lim, int lim, int hev_thr)
Definition: vp9dsp.h:81
VP9DSPContext
Definition: vp9dsp.h:40
av_get_cpu_flags
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:109
cpu_flags
static atomic_int cpu_flags
Definition: cpu.c:56
filters_8tap_1d_fn3
#define filters_8tap_1d_fn3(op, bpp, opt4, opt8, f_opt)
Definition: vp9dsp_init.h:110
idct
static void idct(int16_t block[64])
Definition: 4xm.c:167
av_cold
#define av_cold
Definition: attributes.h:90
decl_mc_funcs
#define decl_mc_funcs(sz, bpp)
Definition: vp9dsp_init_16bpp_aarch64_template.c:60
ff_vp9dsp_init_10bpp_x86
void ff_vp9dsp_init_10bpp_x86(VP9DSPContext *dsp, int bitexact)
EXTERNAL_SSE
#define EXTERNAL_SSE(flags)
Definition: cpu.h:58
EXTERNAL_AVX_FAST
#define EXTERNAL_AVX_FAST(flags)
Definition: cpu.h:71
DCT_ADST
@ DCT_ADST
Definition: vp9.h:39
ff_vp9dsp_init_x86
av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
Definition: vp9dsp_init.c:218
init_subpel3_32_64
#define init_subpel3_32_64(idx, type, bpp, opt)
Definition: vp9dsp_init.h:162
TX_8X8
@ TX_8X8
Definition: vp9.h:29
TX_16X16
@ TX_16X16
Definition: vp9.h:30
ff_vp9dsp_init_12bpp_x86
void ff_vp9dsp_init_12bpp_x86(VP9DSPContext *dsp, int bitexact)
VP9DSPContext::itxfm_add
void(* itxfm_add[N_TXFM_SIZES+1][N_TXFM_TYPES])(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob)
Definition: vp9dsp.h:71
filters_8tap_2d_fn2
#define filters_8tap_2d_fn2(op, align, bpp, bytes, opt4, opt8, f_opt)
Definition: vp9dsp_init.h:136
vp9dsp.h
DCT_DCT
@ DCT_DCT
Definition: vp9.h:38
init_subpel3
#define init_subpel3(idx, type)
Definition: vp9dsp_init_loongarch.c:42
cpu.h
avg
#define avg(a, b, c, d)
Definition: colorspacedsp_template.c:28
TX_4X4
@ TX_4X4
Definition: vp9.h:28
attributes.h
EXTERNAL_SSE2
#define EXTERNAL_SSE2(flags)
Definition: cpu.h:59
decl_fpel_func
#define decl_fpel_func(avg, sz, bpp, opt)
Definition: vp9dsp_init.h:34
filters_8tap_2d_fn
#define filters_8tap_2d_fn(op, sz, align, bpp, bytes, opt, f_opt)
Definition: vp9dsp_init.h:131
filters_8tap_1d_fn2
#define filters_8tap_1d_fn2(op, sz, bpp, opt, f_opt)
Definition: vp9dsp_init.h:106
ADST_ADST
@ ADST_ADST
Definition: vp9.h:41
EXTERNAL_AVX
#define EXTERNAL_AVX(flags)
Definition: cpu.h:70
EXTERNAL_AVX512ICL
#define EXTERNAL_AVX512ICL(flags)
Definition: cpu.h:83
init_subpel3_8to64
#define init_subpel3_8to64(idx, type, bpp, opt)
Definition: vp9dsp_init.h:166
init_fpel_func
#define init_fpel_func(idx1, idx2, sz, type, bpp, opt)
Definition: vp9dsp_init.h:143
mc_rep_funcs
#define mc_rep_funcs(name, bitd, step, W, opt)
Definition: h2656dsp.c:53
TX_32X32
@ TX_32X32
Definition: vp9.h:31
vp9dsp_init.h
ADST_DCT
@ ADST_DCT
Definition: vp9.h:40
h
h
Definition: vp9dsp_template.c:2070
EXTERNAL_SSSE3
#define EXTERNAL_SSSE3(flags)
Definition: cpu.h:65
EXTERNAL_MMX
#define EXTERNAL_MMX(flags)
Definition: cpu.h:56
EXTERNAL_MMXEXT
#define EXTERNAL_MMXEXT(flags)
Definition: cpu.h:57
init_subpel2
#define init_subpel2(idx, idxh, idxv, dir, type)
Definition: vp9dsp_init_loongarch.c:35
Generated on Sat Jun 28 2025 19:22:44 for FFmpeg by   doxygen 1.8.17