4a17ff3da78328e2c794d518059521a9df58e954
[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
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     assert((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         else
122             q = s->c_dc_scale;
123         /* note: block[0] is assumed to be positive */
124         if (!s->h263_aic) {
125         __asm__ volatile (
126                 "mul %%ecx                \n\t"
127                 : "=d" (level), "=a"(dummy)
128                 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
129         );
130         } else
131             /* For AIC we skip quant/dequant of INTRADC */
132             level = (block[0] + 4)>>3;
133
134         block[0]=0; //avoid fake overflow
135 //        temp_block[0] = (block[0] + (q >> 1)) / q;
136         last_non_zero_p1 = 1;
137         bias = s->q_intra_matrix16[qscale][1];
138         qmat = s->q_intra_matrix16[qscale][0];
139     } else {
140         last_non_zero_p1 = 0;
141         bias = s->q_inter_matrix16[qscale][1];
142         qmat = s->q_inter_matrix16[qscale][0];
143     }
144
145     if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
146
147         __asm__ volatile(
148             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
149             SPREADW(MM"3")
150             "pxor "MM"7, "MM"7                  \n\t" // 0
151             "pxor "MM"4, "MM"4                  \n\t" // 0
152             MOVQ" (%2), "MM"5                   \n\t" // qmat[0]
153             "pxor "MM"6, "MM"6                  \n\t"
154             "psubw (%3), "MM"6                  \n\t" // -bias[0]
155             "mov $-128, %%"REG_a"               \n\t"
156             ".p2align 4                         \n\t"
157             "1:                                 \n\t"
158             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
159             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
160             "psubusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
161             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
162             "por "MM"0, "MM"4                   \n\t"
163             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
164             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
165             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
166             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
167             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
168             "pandn "MM"1, "MM"0                 \n\t"
169             PMAXW(MM"0", MM"3")
170             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
171             " js 1b                             \n\t"
172             PMAX(MM"3", MM"0")
173             "movd "MM"3, %%"REG_a"              \n\t"
174             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
175             : "+a" (last_non_zero_p1)
176             : "r" (block+64), "r" (qmat), "r" (bias),
177               "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
178               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
179                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
180         );
181     }else{ // FMT_H263
182         __asm__ volatile(
183             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
184             SPREADW(MM"3")
185             "pxor "MM"7, "MM"7                  \n\t" // 0
186             "pxor "MM"4, "MM"4                  \n\t" // 0
187             "mov $-128, %%"REG_a"               \n\t"
188             ".p2align 4                         \n\t"
189             "1:                                 \n\t"
190             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
191             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
192             MOVQ" (%3, %%"REG_a"), "MM"6        \n\t" // bias[0]
193             "paddusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
194             MOVQ" (%2, %%"REG_a"), "MM"5        \n\t" // qmat[i]
195             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
196             "por "MM"0, "MM"4                   \n\t"
197             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
198             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
199             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
200             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
201             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
202             "pandn "MM"1, "MM"0                 \n\t"
203             PMAXW(MM"0", MM"3")
204             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
205             " js 1b                             \n\t"
206             PMAX(MM"3", MM"0")
207             "movd "MM"3, %%"REG_a"              \n\t"
208             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
209             : "+a" (last_non_zero_p1)
210             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
211               "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
212               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
213                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
214         );
215     }
216     __asm__ volatile(
217         "movd %1, "MM"1                     \n\t" // max_qcoeff
218         SPREADW(MM"1")
219         "psubusw "MM"1, "MM"4               \n\t"
220         "packuswb "MM"4, "MM"4              \n\t"
221 #if COMPILE_TEMPLATE_SSE2
222         "packuswb "MM"4, "MM"4              \n\t"
223 #endif
224         "movd "MM"4, %0                     \n\t" // *overflow
225         : "=g" (*overflow)
226         : "g" (s->max_qcoeff)
227     );
228
229     if(s->mb_intra) block[0]= level;
230     else            block[0]= temp_block[0];
231
232     if (s->idsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM) {
233         if(last_non_zero_p1 <= 1) goto end;
234         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
235         block[0x20] = temp_block[0x10];
236         if(last_non_zero_p1 <= 4) goto end;
237         block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
238         block[0x09] = temp_block[0x03];
239         if(last_non_zero_p1 <= 7) goto end;
240         block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
241         block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
242         if(last_non_zero_p1 <= 11) goto end;
243         block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
244         block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
245         block[0x0C] = temp_block[0x05];
246         if(last_non_zero_p1 <= 16) goto end;
247         block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
248         block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
249         block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
250         block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
251         if(last_non_zero_p1 <= 24) goto end;
252         block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
253         block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
254         block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
255         block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
256         if(last_non_zero_p1 <= 32) goto end;
257         block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
258         block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
259         block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
260         block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
261         if(last_non_zero_p1 <= 40) goto end;
262         block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
263         block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
264         block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
265         block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
266         if(last_non_zero_p1 <= 48) goto end;
267         block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
268         block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
269         block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
270         block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
271         if(last_non_zero_p1 <= 56) goto end;
272         block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
273         block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
274         block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
275         block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
276     }else{
277         if(last_non_zero_p1 <= 1) goto end;
278         block[0x01] = temp_block[0x01];
279         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
280         if(last_non_zero_p1 <= 4) goto end;
281         block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
282         block[0x03] = temp_block[0x03];
283         if(last_non_zero_p1 <= 7) goto end;
284         block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
285         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
286         if(last_non_zero_p1 <= 11) goto end;
287         block[0x19] = temp_block[0x19];
288         block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
289         block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
290         if(last_non_zero_p1 <= 16) goto end;
291         block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
292         block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
293         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
294         block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
295         if(last_non_zero_p1 <= 24) goto end;
296         block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
297         block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
298         block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
299         block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
300         if(last_non_zero_p1 <= 32) goto end;
301         block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
302         block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
303         block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
304         block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
305         if(last_non_zero_p1 <= 40) goto end;
306         block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
307         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
308         block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
309         block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
310         if(last_non_zero_p1 <= 48) goto end;
311         block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
312         block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
313         block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
314         block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
315         if(last_non_zero_p1 <= 56) goto end;
316         block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
317         block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
318         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
319         block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
320     }
321     end:
322     return last_non_zero_p1 - 1;
323 }