FFmpeg
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 decl_fpel_func(put, 4, , mmx);
30 decl_fpel_func(put, 8, , mmx);
31 decl_fpel_func(put, 16, , sse);
32 decl_fpel_func(put, 32, , sse);
33 decl_fpel_func(put, 64, , sse);
34 decl_fpel_func(avg, 4, _8, mmxext);
35 decl_fpel_func(avg, 8, _8, mmxext);
36 decl_fpel_func(avg, 16, _8, sse2);
37 decl_fpel_func(avg, 32, _8, sse2);
38 decl_fpel_func(avg, 64, _8, sse2);
39 decl_fpel_func(put, 32, , avx);
40 decl_fpel_func(put, 64, , avx);
41 decl_fpel_func(avg, 32, _8, avx2);
42 decl_fpel_func(avg, 64, _8, avx2);
43 
44 decl_mc_funcs(8, sse2, int16_t, 8, 8);
45 decl_mc_funcs(4, ssse3, int8_t, 32, 8);
46 decl_mc_funcs(8, ssse3, int8_t, 32, 8);
47 #if ARCH_X86_64
48 decl_mc_funcs(16, ssse3, int8_t, 32, 8);
49 decl_mc_funcs(32, avx2, int8_t, 32, 8);
50 decl_subpel_asm( 4, 8, avx512icl);
51 decl_subpel_asm( 8, 8, avx512icl);
52 decl_subpel_asm(16, 8, avx512icl);
53 decl_subpel_asm(32, 8, avx512icl);
54 decl_subpel_asm(64, 8, avx512icl);
55 #endif
56 
57 mc_rep_funcs(16, 8, 8, sse2, int16_t, 8, 8)
58 #if ARCH_X86_32
59 mc_rep_funcs(16, 8, 8, ssse3, int8_t, 32, 8)
60 #endif
61 mc_rep_funcs(32, 16, 16, sse2, int16_t, 8, 8)
62 mc_rep_funcs(32, 16, 16, ssse3, int8_t, 32, 8)
63 mc_rep_funcs(64, 32, 32, sse2, int16_t, 8, 8)
64 mc_rep_funcs(64, 32, 32, ssse3, int8_t, 32, 8)
65 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
66 mc_rep_funcs(64, 32, 32, avx2, int8_t, 32, 8)
67 #endif
68 
69 extern const int8_t ff_filters_ssse3[3][15][4][32];
70 extern const int16_t ff_filters_sse2[3][15][8][8];
71 
72 filters_8tap_2d_fn2(put, 16, 8, 1, sse2, sse2)
73 filters_8tap_2d_fn2(avg, 16, 8, 1, sse2, sse2)
74 filters_8tap_2d_fn3(put, 16, 8, 1, ssse3, ssse3)
75 filters_8tap_2d_fn3(avg, 16, 8, 1, ssse3, ssse3)
76 
77 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
78 filters_8tap_2d_fn(put, 64, 32, 8, 1, avx2, ssse3)
79 filters_8tap_2d_fn(put, 32, 32, 8, 1, avx2, ssse3)
80 filters_8tap_2d_fn(avg, 64, 32, 8, 1, avx2, ssse3)
81 filters_8tap_2d_fn(avg, 32, 32, 8, 1, avx2, ssse3)
82 #endif
83 
84 filters_8tap_1d_fn3(put, 8, sse2, sse2)
85 filters_8tap_1d_fn3(avg, 8, sse2, sse2)
86 filters_8tap_1d_fn4(put, 8, ssse3, ssse3)
87 filters_8tap_1d_fn4(avg, 8, ssse3, ssse3)
88 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
89 filters_8tap_1d_fn2(put, 64, 8, avx2, ssse3)
90 filters_8tap_1d_fn2(put, 32, 8, avx2, ssse3)
91 filters_8tap_1d_fn2(avg, 64, 8, avx2, ssse3)
92 filters_8tap_1d_fn2(avg, 32, 8, avx2, ssse3)
93 #endif
94 
95 #define itxfm_func(typea, typeb, size, opt) \
96 void ff_vp9_##typea##_##typeb##_##size##x##size##_add_##opt(uint8_t *dst, ptrdiff_t stride, \
97  int16_t *block, int eob)
98 #define itxfm_funcs(size, opt) \
99 itxfm_func(idct, idct, size, opt); \
100 itxfm_func(iadst, idct, size, opt); \
101 itxfm_func(idct, iadst, size, opt); \
102 itxfm_func(iadst, iadst, size, opt)
103 
104 itxfm_func(idct, iadst, 4, sse2);
105 itxfm_func(iadst, idct, 4, sse2);
106 itxfm_func(iadst, iadst, 4, sse2);
107 itxfm_funcs(4, ssse3);
108 itxfm_funcs(4, avx2);
109 itxfm_funcs(8, sse2);
110 itxfm_funcs(8, ssse3);
111 itxfm_funcs(8, avx2);
112 itxfm_funcs(16, sse2);
113 itxfm_funcs(16, ssse3);
114 itxfm_func(idct, idct, 32, sse2);
115 itxfm_func(idct, idct, 32, ssse3);
116 itxfm_func(iwht, iwht, 4, mmx);
117 itxfm_func(iwht, iwht, 4, avx2);
118 itxfm_funcs(16, avx2);
119 itxfm_funcs(16, avx512icl);
120 itxfm_func(idct, idct, 32, avx2);
121 itxfm_func(idct, idct, 32, avx512icl);
122 
123 #undef itxfm_func
124 #undef itxfm_funcs
125 
126 #define lpf_funcs(size1, size2, opt) \
127 void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
128  int E, int I, int H); \
129 void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
130  int E, int I, int H)
131 
132 lpf_funcs(4, 8, mmxext);
133 lpf_funcs(8, 8, mmxext);
134 lpf_funcs(16, 16, sse2);
135 lpf_funcs(16, 16, ssse3);
136 lpf_funcs(16, 16, avx);
137 lpf_funcs(44, 16, sse2);
138 lpf_funcs(44, 16, ssse3);
139 lpf_funcs(44, 16, avx);
140 lpf_funcs(84, 16, sse2);
141 lpf_funcs(84, 16, ssse3);
142 lpf_funcs(84, 16, avx);
143 lpf_funcs(48, 16, sse2);
144 lpf_funcs(48, 16, ssse3);
145 lpf_funcs(48, 16, avx);
146 lpf_funcs(88, 16, sse2);
147 lpf_funcs(88, 16, ssse3);
148 lpf_funcs(88, 16, avx);
149 
150 #undef lpf_funcs
151 
152 #define ipred_func(size, type, opt) \
153 void ff_vp9_ipred_##type##_##size##x##size##_##opt(uint8_t *dst, ptrdiff_t stride, \
154  const uint8_t *l, const uint8_t *a)
155 
156 ipred_func(4, hd, mmxext);
157 ipred_func(4, vl, mmxext);
158 ipred_func(8, v, mmx);
159 
160 #define ipred_dc_funcs(size, opt) \
161 ipred_func(size, dc, opt); \
162 ipred_func(size, dc_left, opt); \
163 ipred_func(size, dc_top, opt)
164 
165 #define ipred_dir_tm_funcs(size, opt) \
166 ipred_func(size, tm, opt); \
167 ipred_func(size, dl, opt); \
168 ipred_func(size, dr, opt); \
169 ipred_func(size, hd, opt); \
170 ipred_func(size, hu, opt); \
171 ipred_func(size, vl, opt); \
172 ipred_func(size, vr, opt)
173 
174 ipred_func(16, v, sse);
175 ipred_func(32, v, sse);
176 
177 ipred_dc_funcs(16, sse2);
178 ipred_dc_funcs(32, sse2);
179 
180 #define ipred_dir_tm_h_funcs(size, opt) \
181 ipred_dir_tm_funcs(size, opt); \
182 ipred_func(size, h, opt)
183 
184 ipred_dir_tm_h_funcs(8, sse2);
185 ipred_dir_tm_h_funcs(16, sse2);
186 ipred_dir_tm_h_funcs(32, sse2);
187 
188 ipred_func(4, h, sse2);
189 
190 #define ipred_all_funcs(size, opt) \
191 ipred_dc_funcs(size, opt); \
192 ipred_dir_tm_h_funcs(size, opt)
193 
194 // FIXME hd/vl_4x4_ssse3 does not exist
195 ipred_all_funcs(4, ssse3);
196 ipred_all_funcs(8, ssse3);
197 ipred_all_funcs(16, ssse3);
198 ipred_all_funcs(32, ssse3);
199 
200 ipred_dir_tm_h_funcs(8, avx);
201 ipred_dir_tm_h_funcs(16, avx);
202 ipred_dir_tm_h_funcs(32, avx);
203 
204 ipred_all_funcs(32, avx2);
205 ipred_func(32, v, avx2);
206 
207 #undef ipred_func
208 #undef ipred_dir_tm_h_funcs
209 #undef ipred_dir_tm_funcs
210 #undef ipred_dc_funcs
211 
212 av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
213 {
214  int cpu_flags;
215 
216  if (bpp == 10) {
217  ff_vp9dsp_init_10bpp_x86(dsp, bitexact);
218  return;
219  } else if (bpp == 12) {
220  ff_vp9dsp_init_12bpp_x86(dsp, bitexact);
221  return;
222  }
223 
224  cpu_flags = av_get_cpu_flags();
225 
226 #define init_lpf(opt) do { \
227  dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_##opt; \
228  dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_##opt; \
229  dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_##opt; \
230  dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_##opt; \
231  dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_##opt; \
232  dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_##opt; \
233  dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_##opt; \
234  dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_##opt; \
235  dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_##opt; \
236  dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_##opt; \
237 } while (0)
238 
239 #define init_ipred(sz, opt, t, e) \
240  dsp->intra_pred[TX_##sz##X##sz][e##_PRED] = ff_vp9_ipred_##t##_##sz##x##sz##_##opt
241 
242 #define ff_vp9_ipred_hd_4x4_ssse3 ff_vp9_ipred_hd_4x4_mmxext
243 #define ff_vp9_ipred_vl_4x4_ssse3 ff_vp9_ipred_vl_4x4_mmxext
244 #define init_dir_tm_ipred(sz, opt) do { \
245  init_ipred(sz, opt, dl, DIAG_DOWN_LEFT); \
246  init_ipred(sz, opt, dr, DIAG_DOWN_RIGHT); \
247  init_ipred(sz, opt, hd, HOR_DOWN); \
248  init_ipred(sz, opt, vl, VERT_LEFT); \
249  init_ipred(sz, opt, hu, HOR_UP); \
250  init_ipred(sz, opt, tm, TM_VP8); \
251  init_ipred(sz, opt, vr, VERT_RIGHT); \
252 } while (0)
253 #define init_dir_tm_h_ipred(sz, opt) do { \
254  init_dir_tm_ipred(sz, opt); \
255  init_ipred(sz, opt, h, HOR); \
256 } while (0)
257 #define init_dc_ipred(sz, opt) do { \
258  init_ipred(sz, opt, dc, DC); \
259  init_ipred(sz, opt, dc_left, LEFT_DC); \
260  init_ipred(sz, opt, dc_top, TOP_DC); \
261 } while (0)
262 #define init_all_ipred(sz, opt) do { \
263  init_dc_ipred(sz, opt); \
264  init_dir_tm_h_ipred(sz, opt); \
265 } while (0)
266 
267  if (EXTERNAL_MMX(cpu_flags)) {
268  init_fpel_func(4, 0, 4, put, , mmx);
269  init_fpel_func(3, 0, 8, put, , mmx);
270  if (!bitexact) {
271  dsp->itxfm_add[4 /* lossless */][DCT_DCT] =
272  dsp->itxfm_add[4 /* lossless */][ADST_DCT] =
273  dsp->itxfm_add[4 /* lossless */][DCT_ADST] =
274  dsp->itxfm_add[4 /* lossless */][ADST_ADST] = ff_vp9_iwht_iwht_4x4_add_mmx;
275  }
276  init_ipred(8, mmx, v, VERT);
277  }
278 
279  if (EXTERNAL_MMXEXT(cpu_flags)) {
280  dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_mmxext;
281  dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_mmxext;
282  dsp->loop_filter_8[1][0] = ff_vp9_loop_filter_h_8_8_mmxext;
283  dsp->loop_filter_8[1][1] = ff_vp9_loop_filter_v_8_8_mmxext;
284  init_fpel_func(4, 1, 4, avg, _8, mmxext);
285  init_fpel_func(3, 1, 8, avg, _8, mmxext);
286  dsp->intra_pred[TX_4X4][HOR_DOWN_PRED] = ff_vp9_ipred_hd_4x4_mmxext;
287  dsp->intra_pred[TX_4X4][VERT_LEFT_PRED] = ff_vp9_ipred_vl_4x4_mmxext;
288  }
289 
290  if (EXTERNAL_SSE(cpu_flags)) {
291  init_fpel_func(2, 0, 16, put, , sse);
292  init_fpel_func(1, 0, 32, put, , sse);
293  init_fpel_func(0, 0, 64, put, , sse);
294  init_ipred(16, sse, v, VERT);
295  init_ipred(32, sse, v, VERT);
296  }
297 
298  if (EXTERNAL_SSE2(cpu_flags)) {
299  init_subpel3_8to64(0, put, 8, sse2);
300  init_subpel3_8to64(1, avg, 8, sse2);
301  init_fpel_func(2, 1, 16, avg, _8, sse2);
302  init_fpel_func(1, 1, 32, avg, _8, sse2);
303  init_fpel_func(0, 1, 64, avg, _8, sse2);
304  init_lpf(sse2);
305  dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_sse2;
306  dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_sse2;
307  dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_sse2;
308  dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_sse2;
309  dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_sse2;
310  dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_sse2;
311  dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_sse2;
312  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_sse2;
313  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_sse2;
314  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_sse2;
315  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_sse2;
316  dsp->itxfm_add[TX_32X32][ADST_ADST] =
317  dsp->itxfm_add[TX_32X32][ADST_DCT] =
318  dsp->itxfm_add[TX_32X32][DCT_ADST] =
319  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_sse2;
320  init_dc_ipred(16, sse2);
321  init_dc_ipred(32, sse2);
322  init_dir_tm_h_ipred(8, sse2);
323  init_dir_tm_h_ipred(16, sse2);
324  init_dir_tm_h_ipred(32, sse2);
325  init_ipred(4, sse2, h, HOR);
326  }
327 
328  if (EXTERNAL_SSSE3(cpu_flags)) {
329  init_subpel3(0, put, 8, ssse3);
330  init_subpel3(1, avg, 8, ssse3);
331  dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_ssse3;
332  dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_ssse3;
333  dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_ssse3;
334  dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_ssse3;
335  dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_ssse3;
336  dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_ssse3;
337  dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_ssse3;
338  dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_ssse3;
339  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_ssse3;
340  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_ssse3;
341  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_ssse3;
342  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_ssse3;
343  dsp->itxfm_add[TX_32X32][ADST_ADST] =
344  dsp->itxfm_add[TX_32X32][ADST_DCT] =
345  dsp->itxfm_add[TX_32X32][DCT_ADST] =
346  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_ssse3;
347  init_lpf(ssse3);
348  init_all_ipred(4, ssse3);
349  init_all_ipred(8, ssse3);
350  init_all_ipred(16, ssse3);
351  init_all_ipred(32, ssse3);
352  }
353 
354  if (EXTERNAL_AVX(cpu_flags)) {
355  init_lpf(avx);
356  init_dir_tm_h_ipred(8, avx);
357  init_dir_tm_h_ipred(16, avx);
358  init_dir_tm_h_ipred(32, avx);
359  }
360  if (EXTERNAL_AVX_FAST(cpu_flags)) {
361  init_fpel_func(1, 0, 32, put, , avx);
362  init_fpel_func(0, 0, 64, put, , avx);
363  }
364 
365  if (EXTERNAL_AVX2_FAST(cpu_flags)) {
366  init_fpel_func(1, 1, 32, avg, _8, avx2);
367  init_fpel_func(0, 1, 64, avg, _8, avx2);
368  if (ARCH_X86_64) {
369 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
370  dsp->itxfm_add[4 /* lossless */][DCT_DCT] =
371  dsp->itxfm_add[4 /* lossless */][ADST_DCT] =
372  dsp->itxfm_add[4 /* lossless */][DCT_ADST] =
373  dsp->itxfm_add[4 /* lossless */][ADST_ADST] = ff_vp9_iwht_iwht_4x4_add_avx2;
374  dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_avx2;
375  dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_avx2;
376  dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_avx2;
377  dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_avx2;
378  dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_avx2;
379  dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_avx2;
380  dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_avx2;
381  dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_avx2;
382  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx2;
383  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx2;
384  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx2;
385  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx2;
386  dsp->itxfm_add[TX_32X32][ADST_ADST] =
387  dsp->itxfm_add[TX_32X32][ADST_DCT] =
388  dsp->itxfm_add[TX_32X32][DCT_ADST] =
389  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx2;
390  init_subpel3_32_64(0, put, 8, avx2);
391  init_subpel3_32_64(1, avg, 8, avx2);
392 #endif
393  }
394  init_all_ipred(32, avx2);
395  init_ipred(32, avx2, v, VERT);
396  }
397 
398 #if ARCH_X86_64
399  if (EXTERNAL_AVX512ICL(cpu_flags)) {
400  dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx512icl;
401  dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx512icl;
402  dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx512icl;
403  dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx512icl;
404  dsp->itxfm_add[TX_32X32][ADST_ADST] =
405  dsp->itxfm_add[TX_32X32][ADST_DCT] =
406  dsp->itxfm_add[TX_32X32][DCT_ADST] =
407  dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx512icl;
408  init_subpel_asm(4, 4, 8, avx512icl);
409  init_subpel_asm(3, 8, 8, avx512icl);
410  init_subpel_asm(2, 16, 8, avx512icl);
411  init_subpel_asm(1, 32, 8, avx512icl);
412  init_subpel_asm(0, 64, 8, avx512icl);
413  }
414 #endif
415 
416 #undef init_fpel
417 #undef init_subpel1
418 #undef init_subpel2
419 #undef init_subpel3
420 }
filters_8tap_2d_fn3
#define filters_8tap_2d_fn3(op, align, bpp, bytes, opt, f_opt)
Definition: vp9dsp_init.h:145
mc_rep_funcs
mc_rep_funcs(mc_rep_funcs(16, mc_rep_funcs(8, mc_rep_funcs(8, mc_rep_funcs(sse2, mc_rep_funcs(int16_t, mc_rep_funcs(8, mc_rep_funcs(8)
Definition: vp9dsp_init.c:57
cpu.h
ipred_all_funcs
#define ipred_all_funcs(size, opt)
EXTERNAL_AVX2_FAST
#define EXTERNAL_AVX2_FAST(flags)
Definition: cpu.h:72
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:2775
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
VERT_LEFT_PRED
@ VERT_LEFT_PRED
Definition: vp9.h:53
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
init_all_ipred
#define init_all_ipred(sz, opt)
init_lpf
#define init_lpf(opt)
init_dir_tm_h_ipred
#define init_dir_tm_h_ipred(sz, opt)
filters_8tap_2d_fn2
#define filters_8tap_2d_fn2(op, align, bpp, bytes, opt8, f_opt)
Definition: vp9dsp_init.h:139
idct
static void idct(int16_t block[64])
Definition: 4xm.c:167
av_cold
#define av_cold
Definition: attributes.h:106
ff_vp9dsp_init_10bpp_x86
void ff_vp9dsp_init_10bpp_x86(VP9DSPContext *dsp, int bitexact)
EXTERNAL_SSE
#define EXTERNAL_SSE(flags)
Definition: cpu.h:51
EXTERNAL_AVX_FAST
#define EXTERNAL_AVX_FAST(flags)
Definition: cpu.h:64
init_ipred
#define init_ipred(sz, opt, t, e)
DCT_ADST
@ DCT_ADST
Definition: vp9.h:39
init_subpel3_32_64
#define init_subpel3_32_64(idx, type, bpp, opt)
Definition: vp9dsp_init.h:168
TX_8X8
@ TX_8X8
Definition: vp9.h:29
ipred_dc_funcs
#define ipred_dc_funcs(size, opt)
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
VP9DSPContext::intra_pred
void(* intra_pred[N_TXFM_SIZES][N_INTRA_PRED_MODES])(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, const uint8_t *top)
Definition: vp9dsp.h:52
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
decl_fpel_func
decl_fpel_func(put, 4,, mmx)
avg
#define avg(a, b, c, d)
Definition: colorspacedsp_template.c:28
itxfm_funcs
#define itxfm_funcs(size, opt)
TX_4X4
@ TX_4X4
Definition: vp9.h:28
decl_subpel_asm
#define decl_subpel_asm(sz, bpp, opt)
Definition: vp9dsp_init.h:194
attributes.h
EXTERNAL_SSE2
#define EXTERNAL_SSE2(flags)
Definition: cpu.h:52
filters_8tap_2d_fn
#define filters_8tap_2d_fn(op, sz, align, bpp, bytes, opt, f_opt)
Definition: vp9dsp_init.h:134
lpf_funcs
#define lpf_funcs(size1, size2, opt)
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
itxfm_func
#define itxfm_func(typea, typeb, size, opt)
ipred_dir_tm_h_funcs
#define ipred_dir_tm_h_funcs(size, opt)
EXTERNAL_AVX
#define EXTERNAL_AVX(flags)
Definition: cpu.h:63
init_dc_ipred
#define init_dc_ipred(sz, opt)
EXTERNAL_AVX512ICL
#define EXTERNAL_AVX512ICL(flags)
Definition: cpu.h:76
init_subpel3_8to64
#define init_subpel3_8to64(idx, type, bpp, opt)
Definition: vp9dsp_init.h:172
decl_mc_funcs
decl_mc_funcs(8, sse2, int16_t, 8, 8)
init_fpel_func
#define init_fpel_func(idx1, idx2, sz, type, bpp, opt)
Definition: vp9dsp_init.h:149
filters_8tap_1d_fn4
#define filters_8tap_1d_fn4(op, bpp, opt, f_opt)
Definition: vp9dsp_init.h:116
ipred_func
#define ipred_func(size, type, opt)
filters_8tap_1d_fn3
#define filters_8tap_1d_fn3(op, bpp, opt8, f_opt)
Definition: vp9dsp_init.h:110
HOR_DOWN_PRED
@ HOR_DOWN_PRED
Definition: vp9.h:52
TX_32X32
@ TX_32X32
Definition: vp9.h:31
vp9dsp_init.h
ff_vp9dsp_init_x86
void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
ADST_DCT
@ ADST_DCT
Definition: vp9.h:40
h
h
Definition: vp9dsp_template.c:2070
EXTERNAL_SSSE3
#define EXTERNAL_SSSE3(flags)
Definition: cpu.h:58
EXTERNAL_MMX
#define EXTERNAL_MMX(flags)
Definition: cpu.h:49
EXTERNAL_MMXEXT
#define EXTERNAL_MMXEXT(flags)
Definition: cpu.h:50
init_subpel_asm
#define init_subpel_asm(idx, sz, bpp, opt)
Definition: vp9dsp_init.h:207
Generated on Sun Dec 21 2025 19:22:25 for FFmpeg by   doxygen 1.8.17