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/asm.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavcodec/vp9dsp.h"
32 #define fpel_func(avg, sz, opt) \
33 void ff_vp9_##avg##sz##_##opt(uint8_t *dst, ptrdiff_t dst_stride, \
34 const uint8_t *src, ptrdiff_t src_stride, \
35 int h, int mx, int my)
36 fpel_func(put
, 4, mmx
);
37 fpel_func(put
, 8, mmx
);
38 fpel_func(put
, 16, sse
);
39 fpel_func(put
, 32, sse
);
40 fpel_func(put
, 64, sse
);
41 fpel_func(avg
, 4, mmxext
);
42 fpel_func(avg
, 8, mmxext
);
43 fpel_func(avg
, 16, sse2
);
44 fpel_func(avg
, 32, sse2
);
45 fpel_func(avg
, 64, sse2
);
48 #define mc_func(avg, sz, dir, opt) \
49 void ff_vp9_##avg##_8tap_1d_##dir##_##sz##_##opt(uint8_t *dst, ptrdiff_t dst_stride, \
50 const uint8_t *src, ptrdiff_t src_stride, \
51 int h, const int8_t (*filter)[16])
52 #define mc_funcs(sz) \
53 mc_func(put, sz, h, ssse3); \
54 mc_func(avg, sz, h, ssse3); \
55 mc_func(put, sz, v, ssse3); \
56 mc_func(avg, sz, v, ssse3)
67 #define mc_rep_func(avg, sz, hsz, dir, opt) \
68 static av_always_inline void \
69 ff_vp9_##avg##_8tap_1d_##dir##_##sz##_##opt(uint8_t *dst, ptrdiff_t dst_stride, \
70 const uint8_t *src, ptrdiff_t src_stride, \
71 int h, const int8_t (*filter)[16]) \
73 ff_vp9_##avg##_8tap_1d_##dir##_##hsz##_##opt(dst, dst_stride, src, \
74 src_stride, h, filter); \
75 ff_vp9_##avg##_8tap_1d_##dir##_##hsz##_##opt(dst + hsz, dst_stride, src + hsz, \
76 src_stride, h, filter); \
79 #define mc_rep_funcs(sz, hsz) \
80 mc_rep_func(put, sz, hsz, h, ssse3); \
81 mc_rep_func(avg, sz, hsz, h, ssse3); \
82 mc_rep_func(put, sz, hsz, v, ssse3); \
83 mc_rep_func(avg, sz, hsz, v, ssse3)
94 extern const int8_t ff_filters_ssse3
[3][15][4][16];
96 #define filter_8tap_2d_fn(op, sz, f, fname) \
97 static void op##_8tap_##fname##_##sz##hv_ssse3(uint8_t *dst, ptrdiff_t dst_stride, \
98 const uint8_t *src, ptrdiff_t src_stride, \
99 int h, int mx, int my) \
101 LOCAL_ALIGNED_16(uint8_t, temp, [71 * 64]); \
102 ff_vp9_put_8tap_1d_h_##sz##_ssse3(temp, 64, src - 3 * src_stride, src_stride, \
103 h + 7, ff_filters_ssse3[f][mx - 1]); \
104 ff_vp9_##op##_8tap_1d_v_##sz##_ssse3(dst, dst_stride, temp + 3 * 64, 64, \
105 h, ff_filters_ssse3[f][my - 1]); \
108 #define filters_8tap_2d_fn(op, sz) \
109 filter_8tap_2d_fn(op, sz, FILTER_8TAP_REGULAR, regular) \
110 filter_8tap_2d_fn(op, sz, FILTER_8TAP_SHARP, sharp) \
111 filter_8tap_2d_fn(op, sz, FILTER_8TAP_SMOOTH, smooth)
113 #define filters_8tap_2d_fn2(op) \
114 filters_8tap_2d_fn(op, 64) \
115 filters_8tap_2d_fn(op, 32) \
116 filters_8tap_2d_fn(op, 16) \
117 filters_8tap_2d_fn(op, 8) \
118 filters_8tap_2d_fn(op, 4)
120 filters_8tap_2d_fn2(put
)
121 filters_8tap_2d_fn2(avg
)
123 #undef filters_8tap_2d_fn2
124 #undef filters_8tap_2d_fn
125 #undef filter_8tap_2d_fn
127 #define filter_8tap_1d_fn(op, sz, f, fname, dir, dvar) \
128 static void op##_8tap_##fname##_##sz##dir##_ssse3(uint8_t *dst, ptrdiff_t dst_stride, \
129 const uint8_t *src, ptrdiff_t src_stride, \
130 int h, int mx, int my) \
132 ff_vp9_##op##_8tap_1d_##dir##_##sz##_ssse3(dst, dst_stride, src, src_stride, \
133 h, ff_filters_ssse3[f][dvar - 1]); \
136 #define filters_8tap_1d_fn(op, sz, dir, dvar) \
137 filter_8tap_1d_fn(op, sz, FILTER_8TAP_REGULAR, regular, dir, dvar) \
138 filter_8tap_1d_fn(op, sz, FILTER_8TAP_SHARP, sharp, dir, dvar) \
139 filter_8tap_1d_fn(op, sz, FILTER_8TAP_SMOOTH, smooth, dir, dvar)
141 #define filters_8tap_1d_fn2(op, sz) \
142 filters_8tap_1d_fn(op, sz, h, mx) \
143 filters_8tap_1d_fn(op, sz, v, my)
145 #define filters_8tap_1d_fn3(op) \
146 filters_8tap_1d_fn2(op, 64) \
147 filters_8tap_1d_fn2(op, 32) \
148 filters_8tap_1d_fn2(op, 16) \
149 filters_8tap_1d_fn2(op, 8) \
150 filters_8tap_1d_fn2(op, 4)
152 filters_8tap_1d_fn3(put
)
153 filters_8tap_1d_fn3(avg
)
155 #undef filters_8tap_1d_fn
156 #undef filters_8tap_1d_fn2
157 #undef filters_8tap_1d_fn3
158 #undef filter_8tap_1d_fn
160 #define itxfm_func(typea, typeb, size, opt) \
161 void ff_vp9_##typea##_##typeb##_##size##x##size##_add_##opt(uint8_t *dst, ptrdiff_t stride, \
162 int16_t *block, int eob)
163 #define itxfm_funcs(size, opt) \
164 itxfm_func(idct, idct, size, opt); \
165 itxfm_func(iadst, idct, size, opt); \
166 itxfm_func(idct, iadst, size, opt); \
167 itxfm_func(iadst, iadst, size, opt)
169 itxfm_funcs(4, ssse3
);
170 itxfm_funcs(8, ssse3
);
172 itxfm_funcs(16, ssse3
);
173 itxfm_funcs(16, avx
);
174 itxfm_func(idct
, idct
, 32, ssse3
);
175 itxfm_func(idct
, idct
, 32, avx
);
176 itxfm_func(iwht
, iwht
, 4, mmx
);
181 #define lpf_funcs(size1, size2, opt) \
182 void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
183 int E, int I, int H); \
184 void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
187 lpf_funcs(16, 16, sse2
);
188 lpf_funcs(16, 16, ssse3
);
189 lpf_funcs(16, 16, avx
);
190 lpf_funcs(44, 16, sse2
);
191 lpf_funcs(44, 16, ssse3
);
192 lpf_funcs(44, 16, avx
);
193 lpf_funcs(84, 16, sse2
);
194 lpf_funcs(84, 16, ssse3
);
195 lpf_funcs(84, 16, avx
);
196 lpf_funcs(48, 16, sse2
);
197 lpf_funcs(48, 16, ssse3
);
198 lpf_funcs(48, 16, avx
);
199 lpf_funcs(88, 16, sse2
);
200 lpf_funcs(88, 16, ssse3
);
201 lpf_funcs(88, 16, avx
);
205 #define ipred_func(size, type, opt) \
206 void ff_vp9_ipred_##type##_##size##x##size##_##opt(uint8_t *dst, ptrdiff_t stride, \
207 const uint8_t *l, const uint8_t *a)
209 #define ipred_funcs(type, opt) \
210 ipred_func(4, type, opt); \
211 ipred_func(8, type, opt); \
212 ipred_func(16, type, opt); \
213 ipred_func(32, type, opt)
215 ipred_funcs(dc
, ssse3
);
216 ipred_funcs(dc_left
, ssse3
);
217 ipred_funcs(dc_top
, ssse3
);
221 ipred_func(8, v
, mmx
);
222 ipred_func(16, v
, sse2
);
223 ipred_func(32, v
, sse2
);
225 #define ipred_func_set(size, type, opt1, opt2) \
226 ipred_func(size, type, opt1); \
227 ipred_func(size, type, opt2)
229 #define ipred_funcs(type, opt1, opt2) \
230 ipred_func(4, type, opt1); \
231 ipred_func_set(8, type, opt1, opt2); \
232 ipred_func_set(16, type, opt1, opt2); \
233 ipred_func_set(32, type, opt1, opt2)
235 ipred_funcs(h
, ssse3
, avx
);
236 ipred_funcs(tm
, ssse3
, avx
);
237 ipred_funcs(dl
, ssse3
, avx
);
238 ipred_funcs(dr
, ssse3
, avx
);
239 ipred_funcs(hu
, ssse3
, avx
);
240 ipred_funcs(hd
, ssse3
, avx
);
241 ipred_funcs(vl
, ssse3
, avx
);
242 ipred_funcs(vr
, ssse3
, avx
);
244 ipred_func(32, dc
, avx2
);
245 ipred_func(32, dc_left
, avx2
);
246 ipred_func(32, dc_top
, avx2
);
247 ipred_func(32, v
, avx2
);
248 ipred_func(32, h
, avx2
);
249 ipred_func(32, tm
, avx2
);
252 #undef ipred_func_set
255 #endif /* HAVE_YASM */
257 av_cold
void ff_vp9dsp_init_x86(VP9DSPContext
*dsp
)
260 int cpu_flags
= av_get_cpu_flags();
262 #define init_fpel(idx1, idx2, sz, type, opt) \
263 dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][0][0] = \
264 dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][0][0] = \
265 dsp->mc[idx1][FILTER_8TAP_SHARP ][idx2][0][0] = \
266 dsp->mc[idx1][FILTER_BILINEAR ][idx2][0][0] = ff_vp9_##type##sz##_##opt
268 #define init_subpel1(idx1, idx2, idxh, idxv, sz, dir, type, opt) \
269 dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][idxh][idxv] = type##_8tap_smooth_##sz##dir##_##opt; \
270 dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][idxh][idxv] = type##_8tap_regular_##sz##dir##_##opt; \
271 dsp->mc[idx1][FILTER_8TAP_SHARP ][idx2][idxh][idxv] = type##_8tap_sharp_##sz##dir##_##opt
273 #define init_subpel2(idx, idxh, idxv, dir, type, opt) \
274 init_subpel1(0, idx, idxh, idxv, 64, dir, type, opt); \
275 init_subpel1(1, idx, idxh, idxv, 32, dir, type, opt); \
276 init_subpel1(2, idx, idxh, idxv, 16, dir, type, opt); \
277 init_subpel1(3, idx, idxh, idxv, 8, dir, type, opt); \
278 init_subpel1(4, idx, idxh, idxv, 4, dir, type, opt)
280 #define init_subpel3(idx, type, opt) \
281 init_subpel2(idx, 1, 1, hv, type, opt); \
282 init_subpel2(idx, 0, 1, v, type, opt); \
283 init_subpel2(idx, 1, 0, h, type, opt)
285 #define init_lpf(opt) do { \
287 dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_##opt; \
288 dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_##opt; \
289 dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_##opt; \
290 dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_##opt; \
291 dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_##opt; \
292 dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_##opt; \
293 dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_##opt; \
294 dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_##opt; \
295 dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_##opt; \
296 dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_##opt; \
300 #define init_ipred(tx, sz, opt) do { \
301 dsp->intra_pred[tx][HOR_PRED] = ff_vp9_ipred_h_##sz##x##sz##_##opt; \
302 dsp->intra_pred[tx][DIAG_DOWN_LEFT_PRED] = ff_vp9_ipred_dl_##sz##x##sz##_##opt; \
303 dsp->intra_pred[tx][DIAG_DOWN_RIGHT_PRED] = ff_vp9_ipred_dr_##sz##x##sz##_##opt; \
304 dsp->intra_pred[tx][HOR_DOWN_PRED] = ff_vp9_ipred_hd_##sz##x##sz##_##opt; \
305 dsp->intra_pred[tx][VERT_LEFT_PRED] = ff_vp9_ipred_vl_##sz##x##sz##_##opt; \
306 dsp->intra_pred[tx][HOR_UP_PRED] = ff_vp9_ipred_hu_##sz##x##sz##_##opt; \
307 if (ARCH_X86_64 || tx != TX_32X32) { \
308 dsp->intra_pred[tx][VERT_RIGHT_PRED] = ff_vp9_ipred_vr_##sz##x##sz##_##opt; \
309 dsp->intra_pred[tx][TM_VP8_PRED] = ff_vp9_ipred_tm_##sz##x##sz##_##opt; \
312 #define init_dc_ipred(tx, sz, opt) do { \
313 init_ipred(tx, sz, opt); \
314 dsp->intra_pred[tx][DC_PRED] = ff_vp9_ipred_dc_##sz##x##sz##_##opt; \
315 dsp->intra_pred[tx][LEFT_DC_PRED] = ff_vp9_ipred_dc_left_##sz##x##sz##_##opt; \
316 dsp->intra_pred[tx][TOP_DC_PRED] = ff_vp9_ipred_dc_top_##sz##x##sz##_##opt; \
319 if (EXTERNAL_MMX(cpu_flags
)) {
320 init_fpel(4, 0, 4, put
, mmx
);
321 init_fpel(3, 0, 8, put
, mmx
);
322 dsp
->itxfm_add
[4 /* lossless */][DCT_DCT
] =
323 dsp
->itxfm_add
[4 /* lossless */][ADST_DCT
] =
324 dsp
->itxfm_add
[4 /* lossless */][DCT_ADST
] =
325 dsp
->itxfm_add
[4 /* lossless */][ADST_ADST
] = ff_vp9_iwht_iwht_4x4_add_mmx
;
326 dsp
->intra_pred
[TX_8X8
][VERT_PRED
] = ff_vp9_ipred_v_8x8_mmx
;
329 if (EXTERNAL_MMXEXT(cpu_flags
)) {
330 init_fpel(4, 1, 4, avg
, mmxext
);
331 init_fpel(3, 1, 8, avg
, mmxext
);
334 if (EXTERNAL_SSE(cpu_flags
)) {
335 init_fpel(2, 0, 16, put
, sse
);
336 init_fpel(1, 0, 32, put
, sse
);
337 init_fpel(0, 0, 64, put
, sse
);
340 if (EXTERNAL_SSE2(cpu_flags
)) {
341 init_fpel(2, 1, 16, avg
, sse2
);
342 init_fpel(1, 1, 32, avg
, sse2
);
343 init_fpel(0, 1, 64, avg
, sse2
);
345 dsp
->intra_pred
[TX_16X16
][VERT_PRED
] = ff_vp9_ipred_v_16x16_sse2
;
346 dsp
->intra_pred
[TX_32X32
][VERT_PRED
] = ff_vp9_ipred_v_32x32_sse2
;
349 if (EXTERNAL_SSSE3(cpu_flags
)) {
350 init_subpel3(0, put
, ssse3
);
351 init_subpel3(1, avg
, ssse3
);
352 dsp
->itxfm_add
[TX_4X4
][DCT_DCT
] = ff_vp9_idct_idct_4x4_add_ssse3
;
353 dsp
->itxfm_add
[TX_4X4
][ADST_DCT
] = ff_vp9_idct_iadst_4x4_add_ssse3
;
354 dsp
->itxfm_add
[TX_4X4
][DCT_ADST
] = ff_vp9_iadst_idct_4x4_add_ssse3
;
355 dsp
->itxfm_add
[TX_4X4
][ADST_ADST
] = ff_vp9_iadst_iadst_4x4_add_ssse3
;
357 dsp
->itxfm_add
[TX_8X8
][DCT_DCT
] = ff_vp9_idct_idct_8x8_add_ssse3
;
358 dsp
->itxfm_add
[TX_8X8
][ADST_DCT
] = ff_vp9_idct_iadst_8x8_add_ssse3
;
359 dsp
->itxfm_add
[TX_8X8
][DCT_ADST
] = ff_vp9_iadst_idct_8x8_add_ssse3
;
360 dsp
->itxfm_add
[TX_8X8
][ADST_ADST
] = ff_vp9_iadst_iadst_8x8_add_ssse3
;
361 dsp
->itxfm_add
[TX_16X16
][DCT_DCT
] = ff_vp9_idct_idct_16x16_add_ssse3
;
362 dsp
->itxfm_add
[TX_16X16
][ADST_DCT
] = ff_vp9_idct_iadst_16x16_add_ssse3
;
363 dsp
->itxfm_add
[TX_16X16
][DCT_ADST
] = ff_vp9_iadst_idct_16x16_add_ssse3
;
364 dsp
->itxfm_add
[TX_16X16
][ADST_ADST
] = ff_vp9_iadst_iadst_16x16_add_ssse3
;
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_ssse3
;
371 init_dc_ipred(TX_4X4
, 4, ssse3
);
372 init_dc_ipred(TX_8X8
, 8, ssse3
);
373 init_dc_ipred(TX_16X16
, 16, ssse3
);
374 init_dc_ipred(TX_32X32
, 32, ssse3
);
377 if (EXTERNAL_AVX(cpu_flags
)) {
379 dsp
->itxfm_add
[TX_8X8
][DCT_DCT
] = ff_vp9_idct_idct_8x8_add_avx
;
380 dsp
->itxfm_add
[TX_8X8
][ADST_DCT
] = ff_vp9_idct_iadst_8x8_add_avx
;
381 dsp
->itxfm_add
[TX_8X8
][DCT_ADST
] = ff_vp9_iadst_idct_8x8_add_avx
;
382 dsp
->itxfm_add
[TX_8X8
][ADST_ADST
] = ff_vp9_iadst_iadst_8x8_add_avx
;
383 dsp
->itxfm_add
[TX_16X16
][DCT_DCT
] = ff_vp9_idct_idct_16x16_add_avx
;
384 dsp
->itxfm_add
[TX_16X16
][ADST_DCT
] = ff_vp9_idct_iadst_16x16_add_avx
;
385 dsp
->itxfm_add
[TX_16X16
][DCT_ADST
] = ff_vp9_iadst_idct_16x16_add_avx
;
386 dsp
->itxfm_add
[TX_16X16
][ADST_ADST
] = ff_vp9_iadst_iadst_16x16_add_avx
;
387 dsp
->itxfm_add
[TX_32X32
][ADST_ADST
] =
388 dsp
->itxfm_add
[TX_32X32
][ADST_DCT
] =
389 dsp
->itxfm_add
[TX_32X32
][DCT_ADST
] =
390 dsp
->itxfm_add
[TX_32X32
][DCT_DCT
] = ff_vp9_idct_idct_32x32_add_avx
;
393 init_ipred(TX_8X8
, 8, avx
);
394 init_ipred(TX_16X16
, 16, avx
);
395 init_ipred(TX_32X32
, 32, avx
);
398 if (EXTERNAL_AVX2(cpu_flags
)) {
399 dsp
->intra_pred
[TX_32X32
][DC_PRED
] = ff_vp9_ipred_dc_32x32_avx2
;
400 dsp
->intra_pred
[TX_32X32
][LEFT_DC_PRED
] = ff_vp9_ipred_dc_left_32x32_avx2
;
401 dsp
->intra_pred
[TX_32X32
][TOP_DC_PRED
] = ff_vp9_ipred_dc_top_32x32_avx2
;
402 dsp
->intra_pred
[TX_32X32
][VERT_PRED
] = ff_vp9_ipred_v_32x32_avx2
;
403 dsp
->intra_pred
[TX_32X32
][HOR_PRED
] = ff_vp9_ipred_h_32x32_avx2
;
404 dsp
->intra_pred
[TX_32X32
][TM_VP8_PRED
] = ff_vp9_ipred_tm_32x32_avx2
;
412 #endif /* HAVE_YASM */