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 #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
RENAME(dct_quantize)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 av_assert2((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 bias = s->q_intra_matrix16[qscale][1];
123 qmat = s->q_intra_matrix16[qscale][0];
124 }else{
125 q = s->c_dc_scale;
126 bias = s->q_chroma_intra_matrix16[qscale][1];
127 qmat = s->q_chroma_intra_matrix16[qscale][0];
128 }
129 /* note: block[0] is assumed to be positive */
130 if (!s->h263_aic) {
131 __asm__ volatile (
132 "mul %%ecx \n\t"
133 : "=d" (level), "=a"(dummy)
134 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
135 );
136 } else
137 /* For AIC we skip quant/dequant of INTRADC */
138 level = (block[0] + 4)>>3;
139
140 block[0]=0; //avoid fake overflow
141 // temp_block[0] = (block[0] + (q >> 1)) / q;
142 last_non_zero_p1 = 1;
143 } else {
144 last_non_zero_p1 = 0;
145 bias = s->q_inter_matrix16[qscale][1];
146 qmat = s->q_inter_matrix16[qscale][0];
147 }
148
149 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
150
151 __asm__ volatile(
152 "movd %%"REG_a", "MM"3 \n\t" // last_non_zero_p1
153 SPREADW(MM"3")
154 "pxor "MM"7, "MM"7 \n\t" // 0
155 "pxor "MM"4, "MM"4 \n\t" // 0
156 MOVQ" (%2), "MM"5 \n\t" // qmat[0]
157 "pxor "MM"6, "MM"6 \n\t"
158 "psubw (%3), "MM"6 \n\t" // -bias[0]
159 "mov $-128, %%"REG_a" \n\t"
160 ".p2align 4 \n\t"
161 "1: \n\t"
162 MOVQ" (%1, %%"REG_a"), "MM"0 \n\t" // block[i]
163 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
164 "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
165 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
166 "por "MM"0, "MM"4 \n\t"
167 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
168 MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
169 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
170 MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
171 MOVQ" "MM"7, (%1, %%"REG_a") \n\t" // 0
172 "pandn "MM"1, "MM"0 \n\t"
173 PMAXW(MM"0", MM"3")
174 "add $"MMREG_WIDTH", %%"REG_a" \n\t"
175 " js 1b \n\t"
176 PMAX(MM"3", MM"0")
177 "movd "MM"3, %%"REG_a" \n\t"
178 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
179 : "+a" (last_non_zero_p1)
180 : "r" (block+64), "r" (qmat), "r" (bias),
181 "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
182 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
183 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
184 );
185 }else{ // FMT_H263
186 __asm__ volatile(
187 "movd %%"REG_a", "MM"3 \n\t" // last_non_zero_p1
188 SPREADW(MM"3")
189 "pxor "MM"7, "MM"7 \n\t" // 0
190 "pxor "MM"4, "MM"4 \n\t" // 0
191 "mov $-128, %%"REG_a" \n\t"
192 ".p2align 4 \n\t"
193 "1: \n\t"
194 MOVQ" (%1, %%"REG_a"), "MM"0 \n\t" // block[i]
195 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
196 MOVQ" (%3, %%"REG_a"), "MM"6 \n\t" // bias[0]
197 "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
198 MOVQ" (%2, %%"REG_a"), "MM"5 \n\t" // qmat[i]
199 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
200 "por "MM"0, "MM"4 \n\t"
201 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
202 MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
203 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
204 MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
205 MOVQ" "MM"7, (%1, %%"REG_a") \n\t" // 0
206 "pandn "MM"1, "MM"0 \n\t"
207 PMAXW(MM"0", MM"3")
208 "add $"MMREG_WIDTH", %%"REG_a" \n\t"
209 " js 1b \n\t"
210 PMAX(MM"3", MM"0")
211 "movd "MM"3, %%"REG_a" \n\t"
212 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
213 : "+a" (last_non_zero_p1)
214 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
215 "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
216 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
217 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
218 );
219 }
220 __asm__ volatile(
221 "movd %1, "MM"1 \n\t" // max_qcoeff
222 SPREADW(MM"1")
223 "psubusw "MM"1, "MM"4 \n\t"
224 "packuswb "MM"4, "MM"4 \n\t"
225 #if COMPILE_TEMPLATE_SSE2
226 "packsswb "MM"4, "MM"4 \n\t"
227 #endif
228 "movd "MM"4, %0 \n\t" // *overflow
229 : "=g" (*overflow)
230 : "g" (s->max_qcoeff)
231 );
232
233 if(s->mb_intra) block[0]= level;
234 else block[0]= temp_block[0];
235
236 if (s->idsp.perm_type == FF_IDCT_PERM_SIMPLE) {
237 if(last_non_zero_p1 <= 1) goto end;
238 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
239 block[0x20] = temp_block[0x10];
240 if(last_non_zero_p1 <= 4) goto end;
241 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
242 block[0x09] = temp_block[0x03];
243 if(last_non_zero_p1 <= 7) goto end;
244 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
245 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
246 if(last_non_zero_p1 <= 11) goto end;
247 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
248 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
249 block[0x0C] = temp_block[0x05];
250 if(last_non_zero_p1 <= 16) goto end;
251 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
252 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
253 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
254 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
255 if(last_non_zero_p1 <= 24) goto end;
256 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
257 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
258 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
259 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
260 if(last_non_zero_p1 <= 32) goto end;
261 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
262 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
263 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
264 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
265 if(last_non_zero_p1 <= 40) goto end;
266 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
267 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
268 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
269 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
270 if(last_non_zero_p1 <= 48) goto end;
271 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
272 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
273 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
274 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
275 if(last_non_zero_p1 <= 56) goto end;
276 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
277 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
278 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
279 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
280 }else if(s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2){
281 if(last_non_zero_p1 <= 1) goto end;
282 block[0x04] = temp_block[0x01];
283 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
284 if(last_non_zero_p1 <= 4) goto end;
285 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
286 block[0x05] = temp_block[0x03];
287 if(last_non_zero_p1 <= 7) goto end;
288 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
289 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
290 if(last_non_zero_p1 <= 11) goto end;
291 block[0x1C] = temp_block[0x19];
292 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
293 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
294 if(last_non_zero_p1 <= 16) goto end;
295 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
296 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
297 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
298 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
299 if(last_non_zero_p1 <= 24) goto end;
300 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
301 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
302 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
303 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
304 if(last_non_zero_p1 <= 32) goto end;
305 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
306 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
307 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
308 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
309 if(last_non_zero_p1 <= 40) goto end;
310 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
311 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
312 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
313 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
314 if(last_non_zero_p1 <= 48) goto end;
315 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
316 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
317 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
318 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
319 if(last_non_zero_p1 <= 56) goto end;
320 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
321 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
322 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
323 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
324 }else{
325 if(last_non_zero_p1 <= 1) goto end;
326 block[0x01] = temp_block[0x01];
327 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
328 if(last_non_zero_p1 <= 4) goto end;
329 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
330 block[0x03] = temp_block[0x03];
331 if(last_non_zero_p1 <= 7) goto end;
332 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
333 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
334 if(last_non_zero_p1 <= 11) goto end;
335 block[0x19] = temp_block[0x19];
336 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
337 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
338 if(last_non_zero_p1 <= 16) goto end;
339 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
340 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
341 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
342 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
343 if(last_non_zero_p1 <= 24) goto end;
344 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
345 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
346 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
347 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
348 if(last_non_zero_p1 <= 32) goto end;
349 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
350 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
351 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
352 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
353 if(last_non_zero_p1 <= 40) goto end;
354 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
355 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
356 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
357 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
358 if(last_non_zero_p1 <= 48) goto end;
359 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
360 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
361 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
362 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
363 if(last_non_zero_p1 <= 56) goto end;
364 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
365 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
366 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
367 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
368 }
369 end:
370 return last_non_zero_p1 - 1;
371 }
372