hevc: SSE2 and SSSE3 loop filters
[ffmpeg.git] / libavcodec / x86 / mpegvideoenc_template.c
1 /*
2  * MPEG video MMX templates
3  *
4  * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This file is part of Libav.
7  *
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.
12  *
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.
17  *
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
21  */
22
23 #include <stdint.h>
24
25 #include "libavutil/internal.h"
26 #include "libavutil/x86/asm.h"
27 #include "libavcodec/mpegvideo.h"
28 #include "fdct.h"
29
30 #undef MMREG_WIDTH
31 #undef MM
32 #undef MOVQ
33 #undef SPREADW
34 #undef PMAXW
35 #undef PMAX
36 #undef SAVE_SIGN
37 #undef RESTORE_SIGN
38
39 #if COMPILE_TEMPLATE_SSE2
40 #define MMREG_WIDTH "16"
41 #define MM "%%xmm"
42 #define MOVQ "movdqa"
43 #define SPREADW(a) \
44             "pshuflw $0, "a", "a"       \n\t"\
45             "punpcklwd "a", "a"         \n\t"
46 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
47 #define PMAX(a,b) \
48             "movhlps "a", "b"           \n\t"\
49             PMAXW(b, a)\
50             "pshuflw $0x0E, "a", "b"    \n\t"\
51             PMAXW(b, a)\
52             "pshuflw $0x01, "a", "b"    \n\t"\
53             PMAXW(b, a)
54 #else
55 #define MMREG_WIDTH "8"
56 #define MM "%%mm"
57 #define MOVQ "movq"
58 #if COMPILE_TEMPLATE_MMXEXT
59 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
60 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
61 #define PMAX(a,b) \
62             "pshufw $0x0E, "a", "b"     \n\t"\
63             PMAXW(b, a)\
64             "pshufw $0x01, "a", "b"     \n\t"\
65             PMAXW(b, a)
66 #else
67 #define SPREADW(a) \
68             "punpcklwd "a", "a"         \n\t"\
69             "punpcklwd "a", "a"         \n\t"
70 #define PMAXW(a,b) \
71             "psubusw "a", "b"           \n\t"\
72             "paddw "a", "b"             \n\t"
73 #define PMAX(a,b)  \
74             "movq "a", "b"              \n\t"\
75             "psrlq $32, "a"             \n\t"\
76             PMAXW(b, a)\
77             "movq "a", "b"              \n\t"\
78             "psrlq $16, "a"             \n\t"\
79             PMAXW(b, a)
80
81 #endif
82 #endif
83
84 #if COMPILE_TEMPLATE_SSSE3
85 #define SAVE_SIGN(a,b) \
86             "movdqa "b", "a"            \n\t"\
87             "pabsw  "b", "b"            \n\t"
88 #define RESTORE_SIGN(a,b) \
89             "psignw "a", "b"            \n\t"
90 #else
91 #define SAVE_SIGN(a,b) \
92             "pxor "a", "a"              \n\t"\
93             "pcmpgtw "b", "a"           \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
94             "pxor "a", "b"              \n\t"\
95             "psubw "a", "b"             \n\t" /* ABS(block[i]) */
96 #define RESTORE_SIGN(a,b) \
97             "pxor "a", "b"              \n\t"\
98             "psubw "a", "b"             \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
99 #endif
100
101 static int RENAME(dct_quantize)(MpegEncContext *s,
102                             int16_t *block, int n,
103                             int qscale, int *overflow)
104 {
105     x86_reg last_non_zero_p1;
106     int level=0, q; //=0 is because gcc says uninitialized ...
107     const uint16_t *qmat, *bias;
108     LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
109
110     assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
111
112     //s->fdct (block);
113     RENAME_FDCT(ff_fdct)(block); // cannot be anything else ...
114
115     if(s->dct_error_sum)
116         s->denoise_dct(s, block);
117
118     if (s->mb_intra) {
119         int dummy;
120         if (n < 4)
121             q = s->y_dc_scale;
122         else
123             q = s->c_dc_scale;
124         /* note: block[0] is assumed to be positive */
125         if (!s->h263_aic) {
126         __asm__ volatile (
127                 "mul %%ecx                \n\t"
128                 : "=d" (level), "=a"(dummy)
129                 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
130         );
131         } else
132             /* For AIC we skip quant/dequant of INTRADC */
133             level = (block[0] + 4)>>3;
134
135         block[0]=0; //avoid fake overflow
136 //        temp_block[0] = (block[0] + (q >> 1)) / q;
137         last_non_zero_p1 = 1;
138         bias = s->q_intra_matrix16[qscale][1];
139         qmat = s->q_intra_matrix16[qscale][0];
140     } else {
141         last_non_zero_p1 = 0;
142         bias = s->q_inter_matrix16[qscale][1];
143         qmat = s->q_inter_matrix16[qscale][0];
144     }
145
146     if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
147
148         __asm__ volatile(
149             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
150             SPREADW(MM"3")
151             "pxor "MM"7, "MM"7                  \n\t" // 0
152             "pxor "MM"4, "MM"4                  \n\t" // 0
153             MOVQ" (%2), "MM"5                   \n\t" // qmat[0]
154             "pxor "MM"6, "MM"6                  \n\t"
155             "psubw (%3), "MM"6                  \n\t" // -bias[0]
156             "mov $-128, %%"REG_a"               \n\t"
157             ".p2align 4                         \n\t"
158             "1:                                 \n\t"
159             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
160             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
161             "psubusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
162             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
163             "por "MM"0, "MM"4                   \n\t"
164             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
165             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
166             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
167             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
168             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
169             "pandn "MM"1, "MM"0                 \n\t"
170             PMAXW(MM"0", MM"3")
171             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
172             " js 1b                             \n\t"
173             PMAX(MM"3", MM"0")
174             "movd "MM"3, %%"REG_a"              \n\t"
175             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
176             : "+a" (last_non_zero_p1)
177             : "r" (block+64), "r" (qmat), "r" (bias),
178               "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
179               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
180                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
181         );
182     }else{ // FMT_H263
183         __asm__ volatile(
184             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
185             SPREADW(MM"3")
186             "pxor "MM"7, "MM"7                  \n\t" // 0
187             "pxor "MM"4, "MM"4                  \n\t" // 0
188             "mov $-128, %%"REG_a"               \n\t"
189             ".p2align 4                         \n\t"
190             "1:                                 \n\t"
191             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
192             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
193             MOVQ" (%3, %%"REG_a"), "MM"6        \n\t" // bias[0]
194             "paddusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
195             MOVQ" (%2, %%"REG_a"), "MM"5        \n\t" // qmat[i]
196             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
197             "por "MM"0, "MM"4                   \n\t"
198             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
199             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
200             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
201             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
202             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
203             "pandn "MM"1, "MM"0                 \n\t"
204             PMAXW(MM"0", MM"3")
205             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
206             " js 1b                             \n\t"
207             PMAX(MM"3", MM"0")
208             "movd "MM"3, %%"REG_a"              \n\t"
209             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
210             : "+a" (last_non_zero_p1)
211             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
212               "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
213               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
214                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
215         );
216     }
217     __asm__ volatile(
218         "movd %1, "MM"1                     \n\t" // max_qcoeff
219         SPREADW(MM"1")
220         "psubusw "MM"1, "MM"4               \n\t"
221         "packuswb "MM"4, "MM"4              \n\t"
222 #if COMPILE_TEMPLATE_SSE2
223         "packuswb "MM"4, "MM"4              \n\t"
224 #endif
225         "movd "MM"4, %0                     \n\t" // *overflow
226         : "=g" (*overflow)
227         : "g" (s->max_qcoeff)
228     );
229
230     if(s->mb_intra) block[0]= level;
231     else            block[0]= temp_block[0];
232
233     if (s->idsp.perm_type == FF_IDCT_PERM_SIMPLE) {
234         if(last_non_zero_p1 <= 1) goto end;
235         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
236         block[0x20] = temp_block[0x10];
237         if(last_non_zero_p1 <= 4) goto end;
238         block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
239         block[0x09] = temp_block[0x03];
240         if(last_non_zero_p1 <= 7) goto end;
241         block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
242         block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
243         if(last_non_zero_p1 <= 11) goto end;
244         block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
245         block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
246         block[0x0C] = temp_block[0x05];
247         if(last_non_zero_p1 <= 16) goto end;
248         block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
249         block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
250         block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
251         block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
252         if(last_non_zero_p1 <= 24) goto end;
253         block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
254         block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
255         block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
256         block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
257         if(last_non_zero_p1 <= 32) goto end;
258         block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
259         block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
260         block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
261         block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
262         if(last_non_zero_p1 <= 40) goto end;
263         block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
264         block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
265         block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
266         block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
267         if(last_non_zero_p1 <= 48) goto end;
268         block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
269         block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
270         block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
271         block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
272         if(last_non_zero_p1 <= 56) goto end;
273         block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
274         block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
275         block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
276         block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
277     }else{
278         if(last_non_zero_p1 <= 1) goto end;
279         block[0x01] = temp_block[0x01];
280         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
281         if(last_non_zero_p1 <= 4) goto end;
282         block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
283         block[0x03] = temp_block[0x03];
284         if(last_non_zero_p1 <= 7) goto end;
285         block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
286         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
287         if(last_non_zero_p1 <= 11) goto end;
288         block[0x19] = temp_block[0x19];
289         block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
290         block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
291         if(last_non_zero_p1 <= 16) goto end;
292         block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
293         block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
294         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
295         block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
296         if(last_non_zero_p1 <= 24) goto end;
297         block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
298         block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
299         block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
300         block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
301         if(last_non_zero_p1 <= 32) goto end;
302         block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
303         block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
304         block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
305         block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
306         if(last_non_zero_p1 <= 40) goto end;
307         block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
308         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
309         block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
310         block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
311         if(last_non_zero_p1 <= 48) goto end;
312         block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
313         block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
314         block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
315         block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
316         if(last_non_zero_p1 <= 56) goto end;
317         block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
318         block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
319         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
320         block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
321     }
322     end:
323     return last_non_zero_p1 - 1;
324 }