Merge commit 'b4987f72197e0c62cf2633bf835a9c32d2a445ae'
[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 FFmpeg.
7  *
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.
12  *
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.
17  *
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
21  */
22
23 #include <stdint.h>
24
25 #include "libavutil/internal.h"
26 #include "libavutil/x86/asm.h"
27 #include "libavcodec/mpegvideo.h"
28
29 #undef MMREG_WIDTH
30 #undef MM
31 #undef MOVQ
32 #undef SPREADW
33 #undef PMAXW
34 #undef PMAX
35 #undef SAVE_SIGN
36 #undef RESTORE_SIGN
37
38 #if COMPILE_TEMPLATE_SSE2
39 #define MMREG_WIDTH "16"
40 #define MM "%%xmm"
41 #define MOVQ "movdqa"
42 #define SPREADW(a) \
43             "pshuflw $0, "a", "a"       \n\t"\
44             "punpcklwd "a", "a"         \n\t"
45 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
46 #define PMAX(a,b) \
47             "movhlps "a", "b"           \n\t"\
48             PMAXW(b, a)\
49             "pshuflw $0x0E, "a", "b"    \n\t"\
50             PMAXW(b, a)\
51             "pshuflw $0x01, "a", "b"    \n\t"\
52             PMAXW(b, a)
53 #else
54 #define MMREG_WIDTH "8"
55 #define MM "%%mm"
56 #define MOVQ "movq"
57 #if COMPILE_TEMPLATE_MMXEXT
58 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
59 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
60 #define PMAX(a,b) \
61             "pshufw $0x0E, "a", "b"     \n\t"\
62             PMAXW(b, a)\
63             "pshufw $0x01, "a", "b"     \n\t"\
64             PMAXW(b, a)
65 #else
66 #define SPREADW(a) \
67             "punpcklwd "a", "a"         \n\t"\
68             "punpcklwd "a", "a"         \n\t"
69 #define PMAXW(a,b) \
70             "psubusw "a", "b"           \n\t"\
71             "paddw "a", "b"             \n\t"
72 #define PMAX(a,b)  \
73             "movq "a", "b"              \n\t"\
74             "psrlq $32, "a"             \n\t"\
75             PMAXW(b, a)\
76             "movq "a", "b"              \n\t"\
77             "psrlq $16, "a"             \n\t"\
78             PMAXW(b, a)
79
80 #endif
81 #endif
82
83 #if COMPILE_TEMPLATE_SSSE3
84 #define SAVE_SIGN(a,b) \
85             "movdqa "b", "a"            \n\t"\
86             "pabsw  "b", "b"            \n\t"
87 #define RESTORE_SIGN(a,b) \
88             "psignw "a", "b"            \n\t"
89 #else
90 #define SAVE_SIGN(a,b) \
91             "pxor "a", "a"              \n\t"\
92             "pcmpgtw "b", "a"           \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
93             "pxor "a", "b"              \n\t"\
94             "psubw "a", "b"             \n\t" /* ABS(block[i]) */
95 #define RESTORE_SIGN(a,b) \
96             "pxor "a", "b"              \n\t"\
97             "psubw "a", "b"             \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
98 #endif
99
100 static int RENAME(dct_quantize)(MpegEncContext *s,
101                             int16_t *block, int n,
102                             int qscale, int *overflow)
103 {
104     x86_reg last_non_zero_p1;
105     int level=0, q; //=0 is because gcc says uninitialized ...
106     const uint16_t *qmat, *bias;
107     LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
108
109     av_assert2((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
110
111     //s->fdct (block);
112     RENAME_FDCT(ff_fdct)(block); // cannot be anything else ...
113
114     if(s->dct_error_sum)
115         s->denoise_dct(s, block);
116
117     if (s->mb_intra) {
118         int dummy;
119         if (n < 4){
120             q = s->y_dc_scale;
121             bias = s->q_intra_matrix16[qscale][1];
122             qmat = s->q_intra_matrix16[qscale][0];
123         }else{
124             q = s->c_dc_scale;
125             bias = s->q_chroma_intra_matrix16[qscale][1];
126             qmat = s->q_chroma_intra_matrix16[qscale][0];
127         }
128         /* note: block[0] is assumed to be positive */
129         if (!s->h263_aic) {
130         __asm__ volatile (
131                 "mul %%ecx                \n\t"
132                 : "=d" (level), "=a"(dummy)
133                 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
134         );
135         } else
136             /* For AIC we skip quant/dequant of INTRADC */
137             level = (block[0] + 4)>>3;
138
139         block[0]=0; //avoid fake overflow
140 //        temp_block[0] = (block[0] + (q >> 1)) / q;
141         last_non_zero_p1 = 1;
142     } else {
143         last_non_zero_p1 = 0;
144         bias = s->q_inter_matrix16[qscale][1];
145         qmat = s->q_inter_matrix16[qscale][0];
146     }
147
148     if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
149
150         __asm__ volatile(
151             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
152             SPREADW(MM"3")
153             "pxor "MM"7, "MM"7                  \n\t" // 0
154             "pxor "MM"4, "MM"4                  \n\t" // 0
155             MOVQ" (%2), "MM"5                   \n\t" // qmat[0]
156             "pxor "MM"6, "MM"6                  \n\t"
157             "psubw (%3), "MM"6                  \n\t" // -bias[0]
158             "mov $-128, %%"REG_a"               \n\t"
159             ".p2align 4                         \n\t"
160             "1:                                 \n\t"
161             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
162             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
163             "psubusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
164             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
165             "por "MM"0, "MM"4                   \n\t"
166             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
167             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
168             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
169             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
170             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
171             "pandn "MM"1, "MM"0                 \n\t"
172             PMAXW(MM"0", MM"3")
173             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
174             " js 1b                             \n\t"
175             PMAX(MM"3", MM"0")
176             "movd "MM"3, %%"REG_a"              \n\t"
177             "movzbl %%al, %%eax                 \n\t" // last_non_zero_p1
178             : "+a" (last_non_zero_p1)
179             : "r" (block+64), "r" (qmat), "r" (bias),
180               "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
181               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
182                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
183         );
184     }else{ // FMT_H263
185         __asm__ volatile(
186             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
187             SPREADW(MM"3")
188             "pxor "MM"7, "MM"7                  \n\t" // 0
189             "pxor "MM"4, "MM"4                  \n\t" // 0
190             "mov $-128, %%"REG_a"               \n\t"
191             ".p2align 4                         \n\t"
192             "1:                                 \n\t"
193             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
194             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
195             MOVQ" (%3, %%"REG_a"), "MM"6        \n\t" // bias[0]
196             "paddusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
197             MOVQ" (%2, %%"REG_a"), "MM"5        \n\t" // qmat[i]
198             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
199             "por "MM"0, "MM"4                   \n\t"
200             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
201             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
202             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
203             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
204             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
205             "pandn "MM"1, "MM"0                 \n\t"
206             PMAXW(MM"0", MM"3")
207             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
208             " js 1b                             \n\t"
209             PMAX(MM"3", MM"0")
210             "movd "MM"3, %%"REG_a"              \n\t"
211             "movzbl %%al, %%eax                 \n\t" // last_non_zero_p1
212             : "+a" (last_non_zero_p1)
213             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
214               "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
215               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
216                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
217         );
218     }
219     __asm__ volatile(
220         "movd %1, "MM"1                     \n\t" // max_qcoeff
221         SPREADW(MM"1")
222         "psubusw "MM"1, "MM"4               \n\t"
223         "packuswb "MM"4, "MM"4              \n\t"
224 #if COMPILE_TEMPLATE_SSE2
225         "packsswb "MM"4, "MM"4              \n\t"
226 #endif
227         "movd "MM"4, %0                     \n\t" // *overflow
228         : "=g" (*overflow)
229         : "g" (s->max_qcoeff)
230     );
231
232     if(s->mb_intra) block[0]= level;
233     else            block[0]= temp_block[0];
234
235     if (s->idsp.perm_type == FF_IDCT_PERM_SIMPLE) {
236         if(last_non_zero_p1 <= 1) goto end;
237         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
238         block[0x20] = temp_block[0x10];
239         if(last_non_zero_p1 <= 4) goto end;
240         block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
241         block[0x09] = temp_block[0x03];
242         if(last_non_zero_p1 <= 7) goto end;
243         block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
244         block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
245         if(last_non_zero_p1 <= 11) goto end;
246         block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
247         block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
248         block[0x0C] = temp_block[0x05];
249         if(last_non_zero_p1 <= 16) goto end;
250         block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
251         block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
252         block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
253         block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
254         if(last_non_zero_p1 <= 24) goto end;
255         block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
256         block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
257         block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
258         block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
259         if(last_non_zero_p1 <= 32) goto end;
260         block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
261         block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
262         block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
263         block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
264         if(last_non_zero_p1 <= 40) goto end;
265         block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
266         block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
267         block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
268         block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
269         if(last_non_zero_p1 <= 48) goto end;
270         block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
271         block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
272         block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
273         block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
274         if(last_non_zero_p1 <= 56) goto end;
275         block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
276         block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
277         block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
278         block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
279     }else if(s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2){
280         if(last_non_zero_p1 <= 1) goto end;
281         block[0x04] = temp_block[0x01];
282         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
283         if(last_non_zero_p1 <= 4) goto end;
284         block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
285         block[0x05] = temp_block[0x03];
286         if(last_non_zero_p1 <= 7) goto end;
287         block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
288         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
289         if(last_non_zero_p1 <= 11) goto end;
290         block[0x1C] = temp_block[0x19];
291         block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
292         block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
293         if(last_non_zero_p1 <= 16) goto end;
294         block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
295         block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
296         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
297         block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
298         if(last_non_zero_p1 <= 24) goto end;
299         block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
300         block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
301         block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
302         block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
303         if(last_non_zero_p1 <= 32) goto end;
304         block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
305         block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
306         block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
307         block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
308         if(last_non_zero_p1 <= 40) goto end;
309         block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
310         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
311         block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
312         block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
313         if(last_non_zero_p1 <= 48) goto end;
314         block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
315         block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
316             block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
317         block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
318         if(last_non_zero_p1 <= 56) goto end;
319         block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
320         block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
321         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
322         block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
323     }else{
324         if(last_non_zero_p1 <= 1) goto end;
325         block[0x01] = temp_block[0x01];
326         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
327         if(last_non_zero_p1 <= 4) goto end;
328         block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
329         block[0x03] = temp_block[0x03];
330         if(last_non_zero_p1 <= 7) goto end;
331         block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
332         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
333         if(last_non_zero_p1 <= 11) goto end;
334         block[0x19] = temp_block[0x19];
335         block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
336         block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
337         if(last_non_zero_p1 <= 16) goto end;
338         block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
339         block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
340         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
341         block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
342         if(last_non_zero_p1 <= 24) goto end;
343         block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
344         block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
345         block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
346         block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
347         if(last_non_zero_p1 <= 32) goto end;
348         block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
349         block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
350         block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
351         block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
352         if(last_non_zero_p1 <= 40) goto end;
353         block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
354         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
355         block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
356         block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
357         if(last_non_zero_p1 <= 48) goto end;
358         block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
359         block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
360         block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
361         block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
362         if(last_non_zero_p1 <= 56) goto end;
363         block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
364         block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
365         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
366         block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
367     }
368     end:
369     return last_non_zero_p1 - 1;
370 }