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/mem.h"
26 #include "libavutil/x86/cpu.h"
27 #include "libavcodec/vp9dsp.h"
28 #include "libavcodec/x86/vp9dsp_init.h"
29
30 #if HAVE_X86ASM
31
32 decl_fpel_func(put, 4, , mmx);
33 decl_fpel_func(put, 8, , mmx);
34 decl_fpel_func(put, 16, , sse);
35 decl_fpel_func(put, 32, , sse);
36 decl_fpel_func(put, 64, , sse);
37 decl_fpel_func(avg, 4, _8, mmxext);
38 decl_fpel_func(avg, 8, _8, mmxext);
39 decl_fpel_func(avg, 16, _8, sse2);
40 decl_fpel_func(avg, 32, _8, sse2);
41 decl_fpel_func(avg, 64, _8, sse2);
42 decl_fpel_func(put, 32, , avx);
43 decl_fpel_func(put, 64, , avx);
44 decl_fpel_func(avg, 32, _8, avx2);
45 decl_fpel_func(avg, 64, _8, avx2);
46
47 decl_mc_funcs(4, mmxext, int16_t, 8, 8);
48 decl_mc_funcs(8, sse2, int16_t, 8, 8);
49 decl_mc_funcs(4, ssse3, int8_t, 32, 8);
50 decl_mc_funcs(8, ssse3, int8_t, 32, 8);
51 #if ARCH_X86_64
52 decl_mc_funcs(16, ssse3, int8_t, 32, 8);
53 decl_mc_funcs(32, avx2, int8_t, 32, 8);
54 #endif
55
56 mc_rep_funcs(16, 8, 8, sse2, int16_t, 8, 8)
57 #if ARCH_X86_32
58 mc_rep_funcs(16, 8, 8, ssse3, int8_t, 32, 8)
59 #endif
60 mc_rep_funcs(32, 16, 16, sse2, int16_t, 8, 8)
61 mc_rep_funcs(32, 16, 16, ssse3, int8_t, 32, 8)
62 mc_rep_funcs(64, 32, 32, sse2, int16_t, 8, 8)
63 mc_rep_funcs(64, 32, 32, ssse3, int8_t, 32, 8)
64 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
65 mc_rep_funcs(64, 32, 32, avx2, int8_t, 32, 8)
66 #endif
67
68 extern const int8_t ff_filters_ssse3[3][15][4][32];
69 extern const int16_t ff_filters_sse2[3][15][8][8];
70
71 filters_8tap_2d_fn2(put, 16, 8, 1, mmxext, sse2, sse2)
72 filters_8tap_2d_fn2(avg, 16, 8, 1, mmxext, sse2, sse2)
73 filters_8tap_2d_fn2(put, 16, 8, 1, ssse3, ssse3, ssse3)
74 filters_8tap_2d_fn2(avg, 16, 8, 1, ssse3, ssse3, ssse3)
75 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
76 filters_8tap_2d_fn(put, 64, 32, 8, 1, avx2, ssse3)
77 filters_8tap_2d_fn(put, 32, 32, 8, 1, avx2, ssse3)
78 filters_8tap_2d_fn(avg, 64, 32, 8, 1, avx2, ssse3)
79 filters_8tap_2d_fn(avg, 32, 32, 8, 1, avx2, ssse3)
80 #endif
81
82 filters_8tap_1d_fn3(put, 8, mmxext, sse2, sse2)
83 filters_8tap_1d_fn3(avg, 8, mmxext, sse2, sse2)
84 filters_8tap_1d_fn3(put, 8, ssse3, ssse3, ssse3)
85 filters_8tap_1d_fn3(avg, 8, ssse3, ssse3, ssse3)
86 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
87 filters_8tap_1d_fn2(put, 64, 8, avx2, ssse3)
88 filters_8tap_1d_fn2(put, 32, 8, avx2, ssse3)
89 filters_8tap_1d_fn2(avg, 64, 8, avx2, ssse3)
90 filters_8tap_1d_fn2(avg, 32, 8, avx2, ssse3)
91 #endif
92
93 #define itxfm_func(typea, typeb, size, opt) \
94 void ff_vp9_##typea##_##typeb##_##size##x##size##_add_##opt(uint8_t *dst, ptrdiff_t stride, \
95 int16_t *block, int eob)
96 #define itxfm_funcs(size, opt) \
97 itxfm_func(idct, idct, size, opt); \
98 itxfm_func(iadst, idct, size, opt); \
99 itxfm_func(idct, iadst, size, opt); \
100 itxfm_func(iadst, iadst, size, opt)
101
102 itxfm_func(idct, idct, 4, mmxext);
103 itxfm_func(idct, iadst, 4, sse2);
104 itxfm_func(iadst, idct, 4, sse2);
105 itxfm_func(iadst, iadst, 4, sse2);
106 itxfm_funcs(4, ssse3);
107 itxfm_funcs(8, sse2);
108 itxfm_funcs(8, ssse3);
109 itxfm_funcs(8, avx);
110 itxfm_funcs(16, sse2);
111 itxfm_funcs(16, ssse3);
112 itxfm_funcs(16, avx);
113 itxfm_func(idct, idct, 32, sse2);
114 itxfm_func(idct, idct, 32, ssse3);
115 itxfm_func(idct, idct, 32, avx);
116 itxfm_func(iwht, iwht, 4, mmx);
117 itxfm_funcs(16, avx2);
118 itxfm_func(idct, idct, 32, avx2);
119
120 #undef itxfm_func
121 #undef itxfm_funcs
122
123 #define lpf_funcs(size1, size2, opt) \
124 void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
125 int E, int I, int H); \
126 void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
127 int E, int I, int H)
128
129 lpf_funcs(4, 8, mmxext);
130 lpf_funcs(8, 8, mmxext);
131 lpf_funcs(16, 16, sse2);
132 lpf_funcs(16, 16, ssse3);
133 lpf_funcs(16, 16, avx);
134 lpf_funcs(44, 16, sse2);
135 lpf_funcs(44, 16, ssse3);
136 lpf_funcs(44, 16, avx);
137 lpf_funcs(84, 16, sse2);
138 lpf_funcs(84, 16, ssse3);
139 lpf_funcs(84, 16, avx);
140 lpf_funcs(48, 16, sse2);
141 lpf_funcs(48, 16, ssse3);
142 lpf_funcs(48, 16, avx);
143 lpf_funcs(88, 16, sse2);
144 lpf_funcs(88, 16, ssse3);
145 lpf_funcs(88, 16, avx);
146
147 #undef lpf_funcs
148
149 #define ipred_func(size, type, opt) \
150 void ff_vp9_ipred_##type##_##size##x##size##_##opt(uint8_t *dst, ptrdiff_t stride, \
151 const uint8_t *l, const uint8_t *a)
152
153 ipred_func(8, v, mmx);
154
155 #define ipred_dc_funcs(size, opt) \
156 ipred_func(size, dc, opt); \
157 ipred_func(size, dc_left, opt); \
158 ipred_func(size, dc_top, opt)
159
160 ipred_dc_funcs(4, mmxext);
161 ipred_dc_funcs(8, mmxext);
162
163 #define ipred_dir_tm_funcs(size, opt) \
164 ipred_func(size, tm, opt); \
165 ipred_func(size, dl, opt); \
166 ipred_func(size, dr, opt); \
167 ipred_func(size, hd, opt); \
168 ipred_func(size, hu, opt); \
169 ipred_func(size, vl, opt); \
170 ipred_func(size, vr, opt)
171
172 ipred_dir_tm_funcs(4, mmxext);
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_func(32, v, avx);
205
206 ipred_dc_funcs(32, avx2);
207 ipred_func(32, h, avx2);
208 ipred_func(32, tm, avx2);
209
210 #undef ipred_func
211 #undef ipred_dir_tm_h_funcs
212 #undef ipred_dir_tm_funcs
213 #undef ipred_dc_funcs
214
215 #endif /* HAVE_X86ASM */
216
ff_vp9dsp_init_x86(VP9DSPContext * dsp,int bpp,int bitexact)217 av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
218 {
219 #if HAVE_X86ASM
220 int cpu_flags;
221
222 if (bpp == 10) {
223 ff_vp9dsp_init_10bpp_x86(dsp, bitexact);
224 return;
225 } else if (bpp == 12) {
226 ff_vp9dsp_init_12bpp_x86(dsp, bitexact);
227 return;
228 }
229
230 cpu_flags = av_get_cpu_flags();
231
232 #define init_lpf(opt) do { \
233 dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_##opt; \
234 dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_##opt; \
235 dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_##opt; \
236 dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_##opt; \
237 dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_##opt; \
238 dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_##opt; \
239 dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_##opt; \
240 dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_##opt; \
241 dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_##opt; \
242 dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_##opt; \
243 } while (0)
244
245 #define init_ipred(sz, opt, t, e) \
246 dsp->intra_pred[TX_##sz##X##sz][e##_PRED] = ff_vp9_ipred_##t##_##sz##x##sz##_##opt
247
248 #define ff_vp9_ipred_hd_4x4_ssse3 ff_vp9_ipred_hd_4x4_mmxext
249 #define ff_vp9_ipred_vl_4x4_ssse3 ff_vp9_ipred_vl_4x4_mmxext
250 #define init_dir_tm_ipred(sz, opt) do { \
251 init_ipred(sz, opt, dl, DIAG_DOWN_LEFT); \
252 init_ipred(sz, opt, dr, DIAG_DOWN_RIGHT); \
253 init_ipred(sz, opt, hd, HOR_DOWN); \
254 init_ipred(sz, opt, vl, VERT_LEFT); \
255 init_ipred(sz, opt, hu, HOR_UP); \
256 init_ipred(sz, opt, tm, TM_VP8); \
257 init_ipred(sz, opt, vr, VERT_RIGHT); \
258 } while (0)
259 #define init_dir_tm_h_ipred(sz, opt) do { \
260 init_dir_tm_ipred(sz, opt); \
261 init_ipred(sz, opt, h, HOR); \
262 } while (0)
263 #define init_dc_ipred(sz, opt) do { \
264 init_ipred(sz, opt, dc, DC); \
265 init_ipred(sz, opt, dc_left, LEFT_DC); \
266 init_ipred(sz, opt, dc_top, TOP_DC); \
267 } while (0)
268 #define init_all_ipred(sz, opt) do { \
269 init_dc_ipred(sz, opt); \
270 init_dir_tm_h_ipred(sz, opt); \
271 } while (0)
272
273 if (EXTERNAL_MMX(cpu_flags)) {
274 init_fpel_func(4, 0, 4, put, , mmx);
275 init_fpel_func(3, 0, 8, put, , mmx);
276 if (!bitexact) {
277 dsp->itxfm_add[4 /* lossless */][DCT_DCT] =
278 dsp->itxfm_add[4 /* lossless */][ADST_DCT] =
279 dsp->itxfm_add[4 /* lossless */][DCT_ADST] =
280 dsp->itxfm_add[4 /* lossless */][ADST_ADST] = ff_vp9_iwht_iwht_4x4_add_mmx;
281 }
282 init_ipred(8, mmx, v, VERT);
283 }
284
285 if (EXTERNAL_MMXEXT(cpu_flags)) {
286 dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_mmxext;
287 dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_mmxext;
288 dsp->loop_filter_8[1][0] = ff_vp9_loop_filter_h_8_8_mmxext;
289 dsp->loop_filter_8[1][1] = ff_vp9_loop_filter_v_8_8_mmxext;
290 init_subpel2(4, 0, 4, put, 8, mmxext);
291 init_subpel2(4, 1, 4, avg, 8, mmxext);
292 init_fpel_func(4, 1, 4, avg, _8, mmxext);
293 init_fpel_func(3, 1, 8, avg, _8, mmxext);
294 dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_mmxext;
295 init_dc_ipred(4, mmxext);
296 init_dc_ipred(8, mmxext);
297 init_dir_tm_ipred(4, mmxext);
298 }
299
300 if (EXTERNAL_SSE(cpu_flags)) {
301 init_fpel_func(2, 0, 16, put, , sse);
302 init_fpel_func(1, 0, 32, put, , sse);
303 init_fpel_func(0, 0, 64, put, , sse);
304 init_ipred(16, sse, v, VERT);
305 init_ipred(32, sse, v, VERT);
306 }
307
308 if (EXTERNAL_SSE2(cpu_flags)) {
309 init_subpel3_8to64(0, put, 8, sse2);
310 init_subpel3_8to64(1, avg, 8, sse2);
311 init_fpel_func(2, 1, 16, avg, _8, sse2);
312 init_fpel_func(1, 1, 32, avg, _8, sse2);
313 init_fpel_func(0, 1, 64, avg, _8, sse2);
314 init_lpf(sse2);
315 dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_sse2;
316 dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_sse2;
317 dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_sse2;
318 dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_sse2;
319 dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_sse2;
320 dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_sse2;
321 dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_sse2;
322 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_sse2;
323 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_sse2;
324 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_sse2;
325 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_sse2;
326 dsp->itxfm_add[TX_32X32][ADST_ADST] =
327 dsp->itxfm_add[TX_32X32][ADST_DCT] =
328 dsp->itxfm_add[TX_32X32][DCT_ADST] =
329 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_sse2;
330 init_dc_ipred(16, sse2);
331 init_dc_ipred(32, sse2);
332 init_dir_tm_h_ipred(8, sse2);
333 init_dir_tm_h_ipred(16, sse2);
334 init_dir_tm_h_ipred(32, sse2);
335 init_ipred(4, sse2, h, HOR);
336 }
337
338 if (EXTERNAL_SSSE3(cpu_flags)) {
339 init_subpel3(0, put, 8, ssse3);
340 init_subpel3(1, avg, 8, ssse3);
341 dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_ssse3;
342 dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_ssse3;
343 dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_ssse3;
344 dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_ssse3;
345 dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_ssse3;
346 dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_ssse3;
347 dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_ssse3;
348 dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_ssse3;
349 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_ssse3;
350 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_ssse3;
351 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_ssse3;
352 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_ssse3;
353 dsp->itxfm_add[TX_32X32][ADST_ADST] =
354 dsp->itxfm_add[TX_32X32][ADST_DCT] =
355 dsp->itxfm_add[TX_32X32][DCT_ADST] =
356 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_ssse3;
357 init_lpf(ssse3);
358 init_all_ipred(4, ssse3);
359 init_all_ipred(8, ssse3);
360 init_all_ipred(16, ssse3);
361 init_all_ipred(32, ssse3);
362 }
363
364 if (EXTERNAL_AVX(cpu_flags)) {
365 dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_avx;
366 dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_avx;
367 dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_avx;
368 dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_avx;
369 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx;
370 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx;
371 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx;
372 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx;
373 dsp->itxfm_add[TX_32X32][ADST_ADST] =
374 dsp->itxfm_add[TX_32X32][ADST_DCT] =
375 dsp->itxfm_add[TX_32X32][DCT_ADST] =
376 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx;
377 init_lpf(avx);
378 init_dir_tm_h_ipred(8, avx);
379 init_dir_tm_h_ipred(16, avx);
380 init_dir_tm_h_ipred(32, avx);
381 }
382 if (EXTERNAL_AVX_FAST(cpu_flags)) {
383 init_fpel_func(1, 0, 32, put, , avx);
384 init_fpel_func(0, 0, 64, put, , avx);
385 init_ipred(32, avx, v, VERT);
386 }
387
388 if (EXTERNAL_AVX2_FAST(cpu_flags)) {
389 init_fpel_func(1, 1, 32, avg, _8, avx2);
390 init_fpel_func(0, 1, 64, avg, _8, avx2);
391 if (ARCH_X86_64) {
392 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
393 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx2;
394 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx2;
395 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx2;
396 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx2;
397 dsp->itxfm_add[TX_32X32][ADST_ADST] =
398 dsp->itxfm_add[TX_32X32][ADST_DCT] =
399 dsp->itxfm_add[TX_32X32][DCT_ADST] =
400 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx2;
401 init_subpel3_32_64(0, put, 8, avx2);
402 init_subpel3_32_64(1, avg, 8, avx2);
403 #endif
404 }
405 init_dc_ipred(32, avx2);
406 init_ipred(32, avx2, h, HOR);
407 init_ipred(32, avx2, tm, TM_VP8);
408 }
409
410 #undef init_fpel
411 #undef init_subpel1
412 #undef init_subpel2
413 #undef init_subpel3
414
415 #endif /* HAVE_X86ASM */
416 }
417