2 * VP9 SIMD optimizations
4 * Copyright (c) 2013 Ronald S. Bultje <rsbultje gmail com>
6 * This file is part of FFmpeg.
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.
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.
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
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"
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);
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);
52 decl_mc_funcs(16, ssse3, int8_t, 32, 8);
53 decl_mc_funcs(32, avx2, int8_t, 32, 8);
56 mc_rep_funcs(16, 8, 8, sse2, int16_t, 8, 8)
58 mc_rep_funcs(16, 8, 8, ssse3, int8_t, 32, 8)
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)
68 extern const int8_t ff_filters_ssse3[3][15][4][32];
69 extern const int16_t ff_filters_sse2[3][15][8][8];
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)
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)
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)
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);
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);
121 #define lpf_funcs(size1, size2, opt) \
122 void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
123 int E, int I, int H); \
124 void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
127 lpf_funcs(16, 16, sse2);
128 lpf_funcs(16, 16, ssse3);
129 lpf_funcs(16, 16, avx);
130 lpf_funcs(44, 16, sse2);
131 lpf_funcs(44, 16, ssse3);
132 lpf_funcs(44, 16, avx);
133 lpf_funcs(84, 16, sse2);
134 lpf_funcs(84, 16, ssse3);
135 lpf_funcs(84, 16, avx);
136 lpf_funcs(48, 16, sse2);
137 lpf_funcs(48, 16, ssse3);
138 lpf_funcs(48, 16, avx);
139 lpf_funcs(88, 16, sse2);
140 lpf_funcs(88, 16, ssse3);
141 lpf_funcs(88, 16, avx);
145 #define ipred_func(size, type, opt) \
146 void ff_vp9_ipred_##type##_##size##x##size##_##opt(uint8_t *dst, ptrdiff_t stride, \
147 const uint8_t *l, const uint8_t *a)
149 ipred_func(8, v, mmx);
151 #define ipred_dc_funcs(size, opt) \
152 ipred_func(size, dc, opt); \
153 ipred_func(size, dc_left, opt); \
154 ipred_func(size, dc_top, opt)
156 ipred_dc_funcs(4, mmxext);
157 ipred_dc_funcs(8, mmxext);
159 #define ipred_dir_tm_funcs(size, opt) \
160 ipred_func(size, tm, opt); \
161 ipred_func(size, dl, opt); \
162 ipred_func(size, dr, opt); \
163 ipred_func(size, hd, opt); \
164 ipred_func(size, hu, opt); \
165 ipred_func(size, vl, opt); \
166 ipred_func(size, vr, opt)
168 ipred_dir_tm_funcs(4, mmxext);
170 ipred_func(16, v, sse);
171 ipred_func(32, v, sse);
173 ipred_dc_funcs(16, sse2);
174 ipred_dc_funcs(32, sse2);
176 #define ipred_dir_tm_h_funcs(size, opt) \
177 ipred_dir_tm_funcs(size, opt); \
178 ipred_func(size, h, opt)
180 ipred_dir_tm_h_funcs(8, sse2);
181 ipred_dir_tm_h_funcs(16, sse2);
182 ipred_dir_tm_h_funcs(32, sse2);
184 ipred_func(4, h, sse2);
186 #define ipred_all_funcs(size, opt) \
187 ipred_dc_funcs(size, opt); \
188 ipred_dir_tm_h_funcs(size, opt)
190 // FIXME hd/vl_4x4_ssse3 does not exist
191 ipred_all_funcs(4, ssse3);
192 ipred_all_funcs(8, ssse3);
193 ipred_all_funcs(16, ssse3);
194 ipred_all_funcs(32, ssse3);
196 ipred_dir_tm_h_funcs(8, avx);
197 ipred_dir_tm_h_funcs(16, avx);
198 ipred_dir_tm_h_funcs(32, avx);
200 ipred_func(32, v, avx);
202 ipred_dc_funcs(32, avx2);
203 ipred_func(32, h, avx2);
204 ipred_func(32, tm, avx2);
207 #undef ipred_dir_tm_h_funcs
208 #undef ipred_dir_tm_funcs
209 #undef ipred_dc_funcs
211 #endif /* HAVE_YASM */
213 av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
219 ff_vp9dsp_init_10bpp_x86(dsp, bitexact);
221 } else if (bpp == 12) {
222 ff_vp9dsp_init_12bpp_x86(dsp, bitexact);
226 cpu_flags = av_get_cpu_flags();
228 #define init_lpf(opt) do { \
229 dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_##opt; \
230 dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_##opt; \
231 dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_##opt; \
232 dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_##opt; \
233 dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_##opt; \
234 dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_##opt; \
235 dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_##opt; \
236 dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_##opt; \
237 dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_##opt; \
238 dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_##opt; \
241 #define init_ipred(sz, opt, t, e) \
242 dsp->intra_pred[TX_##sz##X##sz][e##_PRED] = ff_vp9_ipred_##t##_##sz##x##sz##_##opt
244 #define ff_vp9_ipred_hd_4x4_ssse3 ff_vp9_ipred_hd_4x4_mmxext
245 #define ff_vp9_ipred_vl_4x4_ssse3 ff_vp9_ipred_vl_4x4_mmxext
246 #define init_dir_tm_ipred(sz, opt) do { \
247 init_ipred(sz, opt, dl, DIAG_DOWN_LEFT); \
248 init_ipred(sz, opt, dr, DIAG_DOWN_RIGHT); \
249 init_ipred(sz, opt, hd, HOR_DOWN); \
250 init_ipred(sz, opt, vl, VERT_LEFT); \
251 init_ipred(sz, opt, hu, HOR_UP); \
252 init_ipred(sz, opt, tm, TM_VP8); \
253 init_ipred(sz, opt, vr, VERT_RIGHT); \
255 #define init_dir_tm_h_ipred(sz, opt) do { \
256 init_dir_tm_ipred(sz, opt); \
257 init_ipred(sz, opt, h, HOR); \
259 #define init_dc_ipred(sz, opt) do { \
260 init_ipred(sz, opt, dc, DC); \
261 init_ipred(sz, opt, dc_left, LEFT_DC); \
262 init_ipred(sz, opt, dc_top, TOP_DC); \
264 #define init_all_ipred(sz, opt) do { \
265 init_dc_ipred(sz, opt); \
266 init_dir_tm_h_ipred(sz, opt); \
269 if (EXTERNAL_MMX(cpu_flags)) {
270 init_fpel_func(4, 0, 4, put, , mmx);
271 init_fpel_func(3, 0, 8, put, , mmx);
273 dsp->itxfm_add[4 /* lossless */][DCT_DCT] =
274 dsp->itxfm_add[4 /* lossless */][ADST_DCT] =
275 dsp->itxfm_add[4 /* lossless */][DCT_ADST] =
276 dsp->itxfm_add[4 /* lossless */][ADST_ADST] = ff_vp9_iwht_iwht_4x4_add_mmx;
278 init_ipred(8, mmx, v, VERT);
281 if (EXTERNAL_MMXEXT(cpu_flags)) {
282 init_subpel2(4, 0, 4, put, 8, mmxext);
283 init_subpel2(4, 1, 4, avg, 8, mmxext);
284 init_fpel_func(4, 1, 4, avg, _8, mmxext);
285 init_fpel_func(3, 1, 8, avg, _8, mmxext);
286 dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_mmxext;
287 init_dc_ipred(4, mmxext);
288 init_dc_ipred(8, mmxext);
289 init_dir_tm_ipred(4, mmxext);
292 if (EXTERNAL_SSE(cpu_flags)) {
293 init_fpel_func(2, 0, 16, put, , sse);
294 init_fpel_func(1, 0, 32, put, , sse);
295 init_fpel_func(0, 0, 64, put, , sse);
296 init_ipred(16, sse, v, VERT);
297 init_ipred(32, sse, v, VERT);
300 if (EXTERNAL_SSE2(cpu_flags)) {
301 init_subpel3_8to64(0, put, 8, sse2);
302 init_subpel3_8to64(1, avg, 8, sse2);
303 init_fpel_func(2, 1, 16, avg, _8, sse2);
304 init_fpel_func(1, 1, 32, avg, _8, sse2);
305 init_fpel_func(0, 1, 64, avg, _8, sse2);
307 dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_sse2;
308 dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_sse2;
309 dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_sse2;
310 dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_sse2;
311 dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_sse2;
312 dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_sse2;
313 dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_sse2;
314 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_sse2;
315 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_sse2;
316 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_sse2;
317 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_sse2;
318 dsp->itxfm_add[TX_32X32][ADST_ADST] =
319 dsp->itxfm_add[TX_32X32][ADST_DCT] =
320 dsp->itxfm_add[TX_32X32][DCT_ADST] =
321 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_sse2;
322 init_dc_ipred(16, sse2);
323 init_dc_ipred(32, sse2);
324 init_dir_tm_h_ipred(8, sse2);
325 init_dir_tm_h_ipred(16, sse2);
326 init_dir_tm_h_ipred(32, sse2);
327 init_ipred(4, sse2, h, HOR);
330 if (EXTERNAL_SSSE3(cpu_flags)) {
331 init_subpel3(0, put, 8, ssse3);
332 init_subpel3(1, avg, 8, ssse3);
333 dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_ssse3;
334 dsp->itxfm_add[TX_4X4][ADST_DCT] = ff_vp9_idct_iadst_4x4_add_ssse3;
335 dsp->itxfm_add[TX_4X4][DCT_ADST] = ff_vp9_iadst_idct_4x4_add_ssse3;
336 dsp->itxfm_add[TX_4X4][ADST_ADST] = ff_vp9_iadst_iadst_4x4_add_ssse3;
337 dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_ssse3;
338 dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_ssse3;
339 dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_ssse3;
340 dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_ssse3;
341 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_ssse3;
342 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_ssse3;
343 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_ssse3;
344 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_ssse3;
345 dsp->itxfm_add[TX_32X32][ADST_ADST] =
346 dsp->itxfm_add[TX_32X32][ADST_DCT] =
347 dsp->itxfm_add[TX_32X32][DCT_ADST] =
348 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_ssse3;
350 init_all_ipred(4, ssse3);
351 init_all_ipred(8, ssse3);
352 init_all_ipred(16, ssse3);
353 init_all_ipred(32, ssse3);
356 if (EXTERNAL_AVX(cpu_flags)) {
357 dsp->itxfm_add[TX_8X8][DCT_DCT] = ff_vp9_idct_idct_8x8_add_avx;
358 dsp->itxfm_add[TX_8X8][ADST_DCT] = ff_vp9_idct_iadst_8x8_add_avx;
359 dsp->itxfm_add[TX_8X8][DCT_ADST] = ff_vp9_iadst_idct_8x8_add_avx;
360 dsp->itxfm_add[TX_8X8][ADST_ADST] = ff_vp9_iadst_iadst_8x8_add_avx;
361 dsp->itxfm_add[TX_16X16][DCT_DCT] = ff_vp9_idct_idct_16x16_add_avx;
362 dsp->itxfm_add[TX_16X16][ADST_DCT] = ff_vp9_idct_iadst_16x16_add_avx;
363 dsp->itxfm_add[TX_16X16][DCT_ADST] = ff_vp9_iadst_idct_16x16_add_avx;
364 dsp->itxfm_add[TX_16X16][ADST_ADST] = ff_vp9_iadst_iadst_16x16_add_avx;
365 dsp->itxfm_add[TX_32X32][ADST_ADST] =
366 dsp->itxfm_add[TX_32X32][ADST_DCT] =
367 dsp->itxfm_add[TX_32X32][DCT_ADST] =
368 dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx;
370 init_dir_tm_h_ipred(8, avx);
371 init_dir_tm_h_ipred(16, avx);
372 init_dir_tm_h_ipred(32, avx);
374 if (EXTERNAL_AVX_FAST(cpu_flags)) {
375 init_fpel_func(1, 0, 32, put, , avx);
376 init_fpel_func(0, 0, 64, put, , avx);
377 init_ipred(32, avx, v, VERT);
380 if (EXTERNAL_AVX2(cpu_flags)) {
381 init_fpel_func(1, 1, 32, avg, _8, avx2);
382 init_fpel_func(0, 1, 64, avg, _8, avx2);
384 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
385 init_subpel3_32_64(0, put, 8, avx2);
386 init_subpel3_32_64(1, avg, 8, avx2);
389 init_dc_ipred(32, avx2);
390 init_ipred(32, avx2, h, HOR);
391 init_ipred(32, avx2, tm, TM_VP8);
399 #endif /* HAVE_YASM */