2 * VP9 SIMD optimizations
4 * Copyright (c) 2013 Ronald S. Bultje <rsbultje@gmail.com>
6 * This file is part of Libav.
8 * Libav 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 * Libav 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 Libav; 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/internal.h"
26 #include "libavutil/mem.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavcodec/vp9.h"
32 #define fpel_func(avg, sz, opt) \
33 void ff_vp9_ ## avg ## sz ## _ ## opt(uint8_t *dst, const uint8_t *src, \
34 ptrdiff_t dst_stride, \
35 ptrdiff_t src_stride, \
36 int h, int mx, int my)
38 fpel_func(put, 4, mmx);
39 fpel_func(put, 8, mmx);
40 fpel_func(put, 16, sse);
41 fpel_func(put, 32, sse);
42 fpel_func(put, 64, sse);
43 fpel_func(avg, 4, mmxext);
44 fpel_func(avg, 8, mmxext);
45 fpel_func(avg, 16, sse2);
46 fpel_func(avg, 32, sse2);
47 fpel_func(avg, 64, sse2);
48 fpel_func(put, 32, avx);
49 fpel_func(put, 64, avx);
50 fpel_func(avg, 32, avx2);
51 fpel_func(avg, 64, avx2);
54 #define mc_func(avg, sz, dir, opt, type, f_sz) \
56 ff_vp9_ ## avg ## _8tap_1d_ ## dir ## _ ## sz ## _ ## opt(uint8_t *dst, \
58 ptrdiff_t dst_stride, \
59 ptrdiff_t src_stride, \
61 const type (*filter)[f_sz])
63 #define mc_funcs(sz, opt, type, f_sz) \
64 mc_func(put, sz, h, opt, type, f_sz); \
65 mc_func(avg, sz, h, opt, type, f_sz); \
66 mc_func(put, sz, v, opt, type, f_sz); \
67 mc_func(avg, sz, v, opt, type, f_sz)
69 mc_funcs(4, mmxext, int16_t, 8);
70 mc_funcs(8, sse2, int16_t, 8);
71 mc_funcs(4, ssse3, int8_t, 32);
72 mc_funcs(8, ssse3, int8_t, 32);
74 mc_funcs(16, ssse3, int8_t, 32);
75 mc_funcs(32, avx2, int8_t, 32);
81 #define mc_rep_func(avg, sz, hsz, dir, opt, type, f_sz) \
82 static av_always_inline void \
83 ff_vp9_ ## avg ## _8tap_1d_ ## dir ## _ ## sz ## _ ## opt(uint8_t *dst, \
85 ptrdiff_t dst_stride, \
86 ptrdiff_t src_stride, \
88 const type (*filter)[f_sz]) \
90 ff_vp9_ ## avg ## _8tap_1d_ ## dir ## _ ## hsz ## _ ## opt(dst, src, \
95 ff_vp9_ ## avg ## _8tap_1d_ ## dir ## _ ## hsz ## _ ## opt(dst + hsz, \
102 #define mc_rep_funcs(sz, hsz, opt, type, f_sz) \
103 mc_rep_func(put, sz, hsz, h, opt, type, f_sz); \
104 mc_rep_func(avg, sz, hsz, h, opt, type, f_sz); \
105 mc_rep_func(put, sz, hsz, v, opt, type, f_sz); \
106 mc_rep_func(avg, sz, hsz, v, opt, type, f_sz)
108 mc_rep_funcs(16, 8, sse2, int16_t, 8);
110 mc_rep_funcs(16, 8, ssse3, int8_t, 32);
112 mc_rep_funcs(32, 16, sse2, int16_t, 8);
113 mc_rep_funcs(32, 16, ssse3, int8_t, 32);
114 mc_rep_funcs(64, 32, sse2, int16_t, 8);
115 mc_rep_funcs(64, 32, ssse3, int8_t, 32);
116 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
117 mc_rep_funcs(64, 32, avx2, int8_t, 32);
123 extern const int8_t ff_filters_ssse3[3][15][4][32];
124 extern const int16_t ff_filters_sse2[3][15][8][8];
126 #define filter_8tap_2d_fn(op, sz, f, f_opt, fname, align, opt) \
128 op ## _8tap_ ## fname ## _ ## sz ## hv_ ## opt(uint8_t *dst, \
129 const uint8_t *src, \
130 ptrdiff_t dst_stride, \
131 ptrdiff_t src_stride, \
132 int h, int mx, int my) \
134 LOCAL_ALIGNED_ ## align(uint8_t, temp, [71 * 64]); \
135 ff_vp9_put_8tap_1d_h_ ## sz ## _ ## opt(temp, src - 3 * src_stride, \
138 ff_filters_ ## f_opt[f][mx - 1]); \
139 ff_vp9_ ## op ## _8tap_1d_v_ ## sz ## _ ## opt(dst, temp + 3 * 64, \
142 ff_filters_ ## f_opt[f][my - 1]); \
145 #define filters_8tap_2d_fn(op, sz, align, opt, f_opt) \
146 filter_8tap_2d_fn(op, sz, FILTER_8TAP_REGULAR, f_opt, regular, align, opt) \
147 filter_8tap_2d_fn(op, sz, FILTER_8TAP_SHARP, f_opt, sharp, align, opt) \
148 filter_8tap_2d_fn(op, sz, FILTER_8TAP_SMOOTH, f_opt, smooth, align, opt)
150 #define filters_8tap_2d_fn2(op, align, opt4, opt8, f_opt) \
151 filters_8tap_2d_fn(op, 64, align, opt8, f_opt) \
152 filters_8tap_2d_fn(op, 32, align, opt8, f_opt) \
153 filters_8tap_2d_fn(op, 16, align, opt8, f_opt) \
154 filters_8tap_2d_fn(op, 8, align, opt8, f_opt) \
155 filters_8tap_2d_fn(op, 4, align, opt4, f_opt)
158 filters_8tap_2d_fn2(put, 16, mmxext, sse2, sse2)
159 filters_8tap_2d_fn2(avg, 16, mmxext, sse2, sse2)
160 filters_8tap_2d_fn2(put, 16, ssse3, ssse3, ssse3)
161 filters_8tap_2d_fn2(avg, 16, ssse3, ssse3, ssse3)
162 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
163 filters_8tap_2d_fn(put, 64, 32, avx2, ssse3)
164 filters_8tap_2d_fn(put, 32, 32, avx2, ssse3)
165 filters_8tap_2d_fn(avg, 64, 32, avx2, ssse3)
166 filters_8tap_2d_fn(avg, 32, 32, avx2, ssse3)
169 #undef filters_8tap_2d_fn2
170 #undef filters_8tap_2d_fn
171 #undef filter_8tap_2d_fn
173 #define filter_8tap_1d_fn(op, sz, f, f_opt, fname, dir, dvar, opt) \
175 op ## _8tap_ ## fname ## _ ## sz ## dir ## _ ## opt(uint8_t *dst, \
176 const uint8_t *src, \
177 ptrdiff_t dst_stride, \
178 ptrdiff_t src_stride, \
182 ff_vp9_ ## op ## _8tap_1d_ ## dir ## _ ## sz ## _ ## opt(dst, src, \
185 ff_filters_ ## f_opt[f][dvar - 1]); \
188 #define filters_8tap_1d_fn(op, sz, dir, dvar, opt, f_opt) \
189 filter_8tap_1d_fn(op, sz, FILTER_8TAP_REGULAR, f_opt, regular, dir, dvar, opt) \
190 filter_8tap_1d_fn(op, sz, FILTER_8TAP_SHARP, f_opt, sharp, dir, dvar, opt) \
191 filter_8tap_1d_fn(op, sz, FILTER_8TAP_SMOOTH, f_opt, smooth, dir, dvar, opt)
193 #define filters_8tap_1d_fn2(op, sz, opt, f_opt) \
194 filters_8tap_1d_fn(op, sz, h, mx, opt, f_opt) \
195 filters_8tap_1d_fn(op, sz, v, my, opt, f_opt)
197 #define filters_8tap_1d_fn3(op, opt4, opt8, f_opt) \
198 filters_8tap_1d_fn2(op, 64, opt8, f_opt) \
199 filters_8tap_1d_fn2(op, 32, opt8, f_opt) \
200 filters_8tap_1d_fn2(op, 16, opt8, f_opt) \
201 filters_8tap_1d_fn2(op, 8, opt8, f_opt) \
202 filters_8tap_1d_fn2(op, 4, opt4, f_opt)
204 filters_8tap_1d_fn3(put, mmxext, sse2, sse2)
205 filters_8tap_1d_fn3(avg, mmxext, sse2, sse2)
206 filters_8tap_1d_fn3(put, ssse3, ssse3, ssse3)
207 filters_8tap_1d_fn3(avg, ssse3, ssse3, ssse3)
208 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
209 filters_8tap_1d_fn2(put, 64, avx2, ssse3)
210 filters_8tap_1d_fn2(put, 32, avx2, ssse3)
211 filters_8tap_1d_fn2(avg, 64, avx2, ssse3)
212 filters_8tap_1d_fn2(avg, 32, avx2, ssse3)
215 #undef filters_8tap_1d_fn
216 #undef filters_8tap_1d_fn2
217 #undef filters_8tap_1d_fn3
218 #undef filter_8tap_1d_fn
220 #define lpf_funcs(size1, size2, opt) \
221 void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
222 int E, int I, int H); \
223 void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
226 lpf_funcs(16, 16, sse2);
227 lpf_funcs(16, 16, ssse3);
228 lpf_funcs(16, 16, avx);
229 lpf_funcs(44, 16, sse2);
230 lpf_funcs(44, 16, ssse3);
231 lpf_funcs(44, 16, avx);
232 lpf_funcs(84, 16, sse2);
233 lpf_funcs(84, 16, ssse3);
234 lpf_funcs(84, 16, avx);
235 lpf_funcs(48, 16, sse2);
236 lpf_funcs(48, 16, ssse3);
237 lpf_funcs(48, 16, avx);
238 lpf_funcs(88, 16, sse2);
239 lpf_funcs(88, 16, ssse3);
240 lpf_funcs(88, 16, avx);
244 #endif /* HAVE_YASM */
246 av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
249 int cpu_flags = av_get_cpu_flags();
251 #define init_fpel(idx1, idx2, sz, type, opt) \
252 dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][0][0] = \
253 dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][0][0] = \
254 dsp->mc[idx1][FILTER_8TAP_SHARP ][idx2][0][0] = \
255 dsp->mc[idx1][FILTER_BILINEAR ][idx2][0][0] = ff_vp9_ ## type ## sz ## _ ## opt
258 #define init_subpel1(idx1, idx2, idxh, idxv, sz, dir, type, opt) \
259 dsp->mc[idx1][FILTER_8TAP_SMOOTH][idx2][idxh][idxv] = type ## _8tap_smooth_ ## sz ## dir ## _ ## opt; \
260 dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][idxh][idxv] = type ## _8tap_regular_ ## sz ## dir ## _ ## opt; \
261 dsp->mc[idx1][FILTER_8TAP_SHARP][idx2][idxh][idxv] = type ## _8tap_sharp_ ## sz ## dir ## _ ## opt
263 #define init_subpel2(idx1, idx2, sz, type, opt) \
264 init_subpel1(idx1, idx2, 1, 1, sz, hv, type, opt); \
265 init_subpel1(idx1, idx2, 0, 1, sz, v, type, opt); \
266 init_subpel1(idx1, idx2, 1, 0, sz, h, type, opt)
268 #define init_subpel3_32_64(idx, type, opt) \
269 init_subpel2(0, idx, 64, type, opt); \
270 init_subpel2(1, idx, 32, type, opt)
272 #define init_subpel3_8to64(idx, type, opt) \
273 init_subpel3_32_64(idx, type, opt); \
274 init_subpel2(2, idx, 16, type, opt); \
275 init_subpel2(3, idx, 8, type, opt)
277 #define init_subpel3(idx, type, opt) \
278 init_subpel3_8to64(idx, type, opt); \
279 init_subpel2(4, idx, 4, type, opt)
281 #define init_lpf(opt) do { \
282 dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_##opt; \
283 dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_##opt; \
284 dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_##opt; \
285 dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_##opt; \
286 dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_##opt; \
287 dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_##opt; \
288 dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_##opt; \
289 dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_##opt; \
290 dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_##opt; \
291 dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_##opt; \
294 if (EXTERNAL_MMX(cpu_flags)) {
295 init_fpel(4, 0, 4, put, mmx);
296 init_fpel(3, 0, 8, put, mmx);
299 if (EXTERNAL_MMXEXT(cpu_flags)) {
300 init_subpel2(4, 0, 4, put, mmxext);
301 init_subpel2(4, 1, 4, avg, mmxext);
302 init_fpel(4, 1, 4, avg, mmxext);
303 init_fpel(3, 1, 8, avg, mmxext);
306 if (EXTERNAL_SSE(cpu_flags)) {
307 init_fpel(2, 0, 16, put, sse);
308 init_fpel(1, 0, 32, put, sse);
309 init_fpel(0, 0, 64, put, sse);
312 if (EXTERNAL_SSE2(cpu_flags)) {
313 init_subpel3_8to64(0, put, sse2);
314 init_subpel3_8to64(1, avg, sse2);
315 init_fpel(2, 1, 16, avg, sse2);
316 init_fpel(1, 1, 32, avg, sse2);
317 init_fpel(0, 1, 64, avg, sse2);
321 if (EXTERNAL_SSSE3(cpu_flags)) {
322 init_subpel3(0, put, ssse3);
323 init_subpel3(1, avg, ssse3);
327 if (EXTERNAL_AVX(cpu_flags)) {
328 init_fpel(1, 0, 32, put, avx);
329 init_fpel(0, 0, 64, put, avx);
333 if (EXTERNAL_AVX2(cpu_flags)) {
334 init_fpel(1, 1, 32, avg, avx2);
335 init_fpel(0, 1, 64, avg, avx2);
337 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
338 init_subpel3_32_64(0, put, avx2);
339 init_subpel3_32_64(1, avg, avx2);
340 #endif /* ARCH_X86_64 && HAVE_AVX2_EXTERNAL */
348 #endif /* HAVE_YASM */