1 /*
2  * Copyright (C) 2002-2012 Michael Niedermayer
3  * Copyright (C) 2012 Ronald S. Bultje
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 #include "config.h"
23 
24 #include "libavutil/attributes.h"
25 #include "libavutil/avassert.h"
26 #include "libavutil/common.h"
27 #include "libavutil/cpu.h"
28 #include "libavutil/mem.h"
29 #include "libavutil/x86/asm.h"
30 #include "libavutil/x86/cpu.h"
31 #include "libavcodec/videodsp.h"
32 
33 #if HAVE_YASM
34 typedef void emu_edge_vfix_func(uint8_t *dst, x86_reg dst_stride,
35                                 const uint8_t *src, x86_reg src_stride,
36                                 x86_reg start_y, x86_reg end_y, x86_reg bh);
37 typedef void emu_edge_vvar_func(uint8_t *dst, x86_reg dst_stride,
38                                 const uint8_t *src, x86_reg src_stride,
39                                 x86_reg start_y, x86_reg end_y, x86_reg bh,
40                                 x86_reg w);
41 
42 extern emu_edge_vfix_func ff_emu_edge_vfix1_mmx;
43 extern emu_edge_vfix_func ff_emu_edge_vfix2_mmx;
44 extern emu_edge_vfix_func ff_emu_edge_vfix3_mmx;
45 extern emu_edge_vfix_func ff_emu_edge_vfix4_mmx;
46 extern emu_edge_vfix_func ff_emu_edge_vfix5_mmx;
47 extern emu_edge_vfix_func ff_emu_edge_vfix6_mmx;
48 extern emu_edge_vfix_func ff_emu_edge_vfix7_mmx;
49 extern emu_edge_vfix_func ff_emu_edge_vfix8_mmx;
50 extern emu_edge_vfix_func ff_emu_edge_vfix9_mmx;
51 extern emu_edge_vfix_func ff_emu_edge_vfix10_mmx;
52 extern emu_edge_vfix_func ff_emu_edge_vfix11_mmx;
53 extern emu_edge_vfix_func ff_emu_edge_vfix12_mmx;
54 extern emu_edge_vfix_func ff_emu_edge_vfix13_mmx;
55 extern emu_edge_vfix_func ff_emu_edge_vfix14_mmx;
56 extern emu_edge_vfix_func ff_emu_edge_vfix15_mmx;
57 extern emu_edge_vfix_func ff_emu_edge_vfix16_mmx;
58 extern emu_edge_vfix_func ff_emu_edge_vfix17_mmx;
59 extern emu_edge_vfix_func ff_emu_edge_vfix18_mmx;
60 extern emu_edge_vfix_func ff_emu_edge_vfix19_mmx;
61 extern emu_edge_vfix_func ff_emu_edge_vfix20_mmx;
62 extern emu_edge_vfix_func ff_emu_edge_vfix21_mmx;
63 extern emu_edge_vfix_func ff_emu_edge_vfix22_mmx;
64 #if ARCH_X86_32
65 static emu_edge_vfix_func * const vfixtbl_mmx[22] = {
66     &ff_emu_edge_vfix1_mmx,  &ff_emu_edge_vfix2_mmx,  &ff_emu_edge_vfix3_mmx,
67     &ff_emu_edge_vfix4_mmx,  &ff_emu_edge_vfix5_mmx,  &ff_emu_edge_vfix6_mmx,
68     &ff_emu_edge_vfix7_mmx,  &ff_emu_edge_vfix8_mmx,  &ff_emu_edge_vfix9_mmx,
69     &ff_emu_edge_vfix10_mmx, &ff_emu_edge_vfix11_mmx, &ff_emu_edge_vfix12_mmx,
70     &ff_emu_edge_vfix13_mmx, &ff_emu_edge_vfix14_mmx, &ff_emu_edge_vfix15_mmx,
71     &ff_emu_edge_vfix16_mmx, &ff_emu_edge_vfix17_mmx, &ff_emu_edge_vfix18_mmx,
72     &ff_emu_edge_vfix19_mmx, &ff_emu_edge_vfix20_mmx, &ff_emu_edge_vfix21_mmx,
73     &ff_emu_edge_vfix22_mmx
74 };
75 #endif
76 extern emu_edge_vvar_func ff_emu_edge_vvar_mmx;
77 extern emu_edge_vfix_func ff_emu_edge_vfix16_sse;
78 extern emu_edge_vfix_func ff_emu_edge_vfix17_sse;
79 extern emu_edge_vfix_func ff_emu_edge_vfix18_sse;
80 extern emu_edge_vfix_func ff_emu_edge_vfix19_sse;
81 extern emu_edge_vfix_func ff_emu_edge_vfix20_sse;
82 extern emu_edge_vfix_func ff_emu_edge_vfix21_sse;
83 extern emu_edge_vfix_func ff_emu_edge_vfix22_sse;
84 static emu_edge_vfix_func * const vfixtbl_sse[22] = {
85     ff_emu_edge_vfix1_mmx,  ff_emu_edge_vfix2_mmx,  ff_emu_edge_vfix3_mmx,
86     ff_emu_edge_vfix4_mmx,  ff_emu_edge_vfix5_mmx,  ff_emu_edge_vfix6_mmx,
87     ff_emu_edge_vfix7_mmx,  ff_emu_edge_vfix8_mmx,  ff_emu_edge_vfix9_mmx,
88     ff_emu_edge_vfix10_mmx, ff_emu_edge_vfix11_mmx, ff_emu_edge_vfix12_mmx,
89     ff_emu_edge_vfix13_mmx, ff_emu_edge_vfix14_mmx, ff_emu_edge_vfix15_mmx,
90     ff_emu_edge_vfix16_sse, ff_emu_edge_vfix17_sse, ff_emu_edge_vfix18_sse,
91     ff_emu_edge_vfix19_sse, ff_emu_edge_vfix20_sse, ff_emu_edge_vfix21_sse,
92     ff_emu_edge_vfix22_sse
93 };
94 extern emu_edge_vvar_func ff_emu_edge_vvar_sse;
95 
96 typedef void emu_edge_hfix_func(uint8_t *dst, x86_reg dst_stride,
97                                 x86_reg start_x, x86_reg bh);
98 typedef void emu_edge_hvar_func(uint8_t *dst, x86_reg dst_stride,
99                                 x86_reg start_x, x86_reg n_words, x86_reg bh);
100 
101 extern emu_edge_hfix_func ff_emu_edge_hfix2_mmx;
102 extern emu_edge_hfix_func ff_emu_edge_hfix4_mmx;
103 extern emu_edge_hfix_func ff_emu_edge_hfix6_mmx;
104 extern emu_edge_hfix_func ff_emu_edge_hfix8_mmx;
105 extern emu_edge_hfix_func ff_emu_edge_hfix10_mmx;
106 extern emu_edge_hfix_func ff_emu_edge_hfix12_mmx;
107 extern emu_edge_hfix_func ff_emu_edge_hfix14_mmx;
108 extern emu_edge_hfix_func ff_emu_edge_hfix16_mmx;
109 extern emu_edge_hfix_func ff_emu_edge_hfix18_mmx;
110 extern emu_edge_hfix_func ff_emu_edge_hfix20_mmx;
111 extern emu_edge_hfix_func ff_emu_edge_hfix22_mmx;
112 #if ARCH_X86_32
113 static emu_edge_hfix_func * const hfixtbl_mmx[11] = {
114     ff_emu_edge_hfix2_mmx,  ff_emu_edge_hfix4_mmx,  ff_emu_edge_hfix6_mmx,
115     ff_emu_edge_hfix8_mmx,  ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx,
116     ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_mmx, ff_emu_edge_hfix18_mmx,
117     ff_emu_edge_hfix20_mmx, ff_emu_edge_hfix22_mmx
118 };
119 #endif
120 extern emu_edge_hvar_func ff_emu_edge_hvar_mmx;
121 extern emu_edge_hfix_func ff_emu_edge_hfix16_sse2;
122 extern emu_edge_hfix_func ff_emu_edge_hfix18_sse2;
123 extern emu_edge_hfix_func ff_emu_edge_hfix20_sse2;
124 extern emu_edge_hfix_func ff_emu_edge_hfix22_sse2;
125 static emu_edge_hfix_func * const hfixtbl_sse2[11] = {
126     ff_emu_edge_hfix2_mmx,  ff_emu_edge_hfix4_mmx,  ff_emu_edge_hfix6_mmx,
127     ff_emu_edge_hfix8_mmx,  ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx,
128     ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_sse2, ff_emu_edge_hfix18_sse2,
129     ff_emu_edge_hfix20_sse2, ff_emu_edge_hfix22_sse2
130 };
131 extern emu_edge_hvar_func ff_emu_edge_hvar_sse2;
132 
emulated_edge_mc(uint8_t * dst,const uint8_t * src,ptrdiff_t dst_stride,ptrdiff_t src_stride,x86_reg block_w,x86_reg block_h,x86_reg src_x,x86_reg src_y,x86_reg w,x86_reg h,emu_edge_vfix_func ** vfix_tbl,emu_edge_vvar_func * v_extend_var,emu_edge_hfix_func ** hfix_tbl,emu_edge_hvar_func * h_extend_var)133 static av_always_inline void emulated_edge_mc(uint8_t *dst, const uint8_t *src,
134                                               ptrdiff_t dst_stride,
135                                               ptrdiff_t src_stride,
136                                               x86_reg block_w, x86_reg block_h,
137                                               x86_reg src_x, x86_reg src_y,
138                                               x86_reg w, x86_reg h,
139                                               emu_edge_vfix_func **vfix_tbl,
140                                               emu_edge_vvar_func *v_extend_var,
141                                               emu_edge_hfix_func **hfix_tbl,
142                                               emu_edge_hvar_func *h_extend_var)
143 {
144     x86_reg start_y, start_x, end_y, end_x, src_y_add = 0, p;
145 
146     if (!w || !h)
147         return;
148 
149     if (src_y >= h) {
150         src -= src_y*src_stride;
151         src_y_add = h - 1;
152         src_y     = h - 1;
153     } else if (src_y <= -block_h) {
154         src -= src_y*src_stride;
155         src_y_add = 1 - block_h;
156         src_y     = 1 - block_h;
157     }
158     if (src_x >= w) {
159         src   += w - 1 - src_x;
160         src_x  = w - 1;
161     } else if (src_x <= -block_w) {
162         src   += 1 - block_w - src_x;
163         src_x  = 1 - block_w;
164     }
165 
166     start_y = FFMAX(0, -src_y);
167     start_x = FFMAX(0, -src_x);
168     end_y   = FFMIN(block_h, h-src_y);
169     end_x   = FFMIN(block_w, w-src_x);
170     av_assert2(start_x < end_x && block_w > 0);
171     av_assert2(start_y < end_y && block_h > 0);
172 
173     // fill in the to-be-copied part plus all above/below
174     src += (src_y_add + start_y) * src_stride + start_x;
175     w = end_x - start_x;
176     if (w <= 22) {
177         vfix_tbl[w - 1](dst + start_x, dst_stride, src, src_stride,
178                         start_y, end_y, block_h);
179     } else {
180         v_extend_var(dst + start_x, dst_stride, src, src_stride,
181                      start_y, end_y, block_h, w);
182     }
183 
184     // fill left
185     if (start_x) {
186         if (start_x <= 22) {
187             hfix_tbl[(start_x - 1) >> 1](dst, dst_stride, start_x, block_h);
188         } else {
189             h_extend_var(dst, dst_stride,
190                          start_x, (start_x + 1) >> 1, block_h);
191         }
192     }
193 
194     // fill right
195     p = block_w - end_x;
196     if (p) {
197         if (p <= 22) {
198             hfix_tbl[(p - 1) >> 1](dst + end_x - (p & 1), dst_stride,
199                                    -!(p & 1), block_h);
200         } else {
201             h_extend_var(dst + end_x - (p & 1), dst_stride,
202                          -!(p & 1), (p + 1) >> 1, block_h);
203         }
204     }
205 }
206 
207 #if ARCH_X86_32
emulated_edge_mc_mmx(uint8_t * buf,const uint8_t * src,ptrdiff_t buf_stride,ptrdiff_t src_stride,int block_w,int block_h,int src_x,int src_y,int w,int h)208 static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
209                                              ptrdiff_t buf_stride,
210                                              ptrdiff_t src_stride,
211                                              int block_w, int block_h,
212                                              int src_x, int src_y, int w, int h)
213 {
214     emulated_edge_mc(buf, src, buf_stride, src_stride, block_w, block_h,
215                      src_x, src_y, w, h, (emu_edge_vfix_func **)vfixtbl_mmx, &ff_emu_edge_vvar_mmx,
216                      (emu_edge_hfix_func **)hfixtbl_mmx, &ff_emu_edge_hvar_mmx);
217 }
218 
emulated_edge_mc_sse(uint8_t * buf,const uint8_t * src,ptrdiff_t buf_stride,ptrdiff_t src_stride,int block_w,int block_h,int src_x,int src_y,int w,int h)219 static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
220                                              ptrdiff_t buf_stride,
221                                              ptrdiff_t src_stride,
222                                              int block_w, int block_h,
223                                              int src_x, int src_y, int w, int h)
224 {
225     emulated_edge_mc(buf, src, buf_stride, src_stride, block_w, block_h,
226                      src_x, src_y, w, h, (emu_edge_vfix_func **)vfixtbl_sse, &ff_emu_edge_vvar_sse,
227                      (emu_edge_hfix_func **)hfixtbl_mmx, &ff_emu_edge_hvar_mmx);
228 }
229 #endif
230 
emulated_edge_mc_sse2(uint8_t * buf,const uint8_t * src,ptrdiff_t buf_stride,ptrdiff_t src_stride,int block_w,int block_h,int src_x,int src_y,int w,int h)231 static av_noinline void emulated_edge_mc_sse2(uint8_t *buf, const uint8_t *src,
232                                               ptrdiff_t buf_stride,
233                                               ptrdiff_t src_stride,
234                                               int block_w, int block_h,
235                                               int src_x, int src_y, int w,
236                                               int h)
237 {
238     emulated_edge_mc(buf, src, buf_stride, src_stride, block_w, block_h,
239                      src_x, src_y, w, h, (emu_edge_vfix_func **)vfixtbl_sse, &ff_emu_edge_vvar_sse,
240                      (emu_edge_hfix_func **)hfixtbl_sse2, &ff_emu_edge_hvar_sse2);
241 }
242 #endif /* HAVE_YASM */
243 
244 void ff_prefetch_mmxext(uint8_t *buf, ptrdiff_t stride, int h);
245 void ff_prefetch_3dnow(uint8_t *buf, ptrdiff_t stride, int h);
246 
ff_videodsp_init_x86(VideoDSPContext * ctx,int bpc)247 av_cold void ff_videodsp_init_x86(VideoDSPContext *ctx, int bpc)
248 {
249 #if HAVE_YASM
250     int cpu_flags = av_get_cpu_flags();
251 
252 #if ARCH_X86_32
253     if (EXTERNAL_MMX(cpu_flags) && bpc <= 8) {
254         ctx->emulated_edge_mc = emulated_edge_mc_mmx;
255     }
256     if (EXTERNAL_AMD3DNOW(cpu_flags)) {
257         ctx->prefetch = ff_prefetch_3dnow;
258     }
259 #endif /* ARCH_X86_32 */
260     if (EXTERNAL_MMXEXT(cpu_flags)) {
261         ctx->prefetch = ff_prefetch_mmxext;
262     }
263 #if ARCH_X86_32
264     if (EXTERNAL_SSE(cpu_flags) && bpc <= 8) {
265         ctx->emulated_edge_mc = emulated_edge_mc_sse;
266     }
267 #endif /* ARCH_X86_32 */
268     if (EXTERNAL_SSE2(cpu_flags) && bpc <= 8) {
269         ctx->emulated_edge_mc = emulated_edge_mc_sse2;
270     }
271 #endif /* HAVE_YASM */
272 }
273