2 * MPEG video MMX templates
4 * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 #define SPREADW(a) "pshufw $0, " #a ", " #a " \n\t"
24 #define PMAXW(a,b) "pmaxsw " #a ", " #b " \n\t"
28 "punpcklwd " #a ", " #a " \n\t"\
29 "punpcklwd " #a ", " #a " \n\t"
31 "psubusw " #a ", " #b " \n\t"\
32 "paddw " #a ", " #b " \n\t"
35 static int RENAME(dct_quantize)(MpegEncContext *s,
36 DCTELEM *block, int n,
37 int qscale, int *overflow)
39 int level=0, last_non_zero_p1, q; //=0 is cuz gcc says uninitalized ...
40 const UINT16 *qmat, *bias;
41 static __align8 INT16 temp_block[64];
51 /* note: block[0] is assumed to be positive */
55 "xorl %%edx, %%edx \n\t"
57 : "=d" (level), "=a"(dummy)
58 : "a" (block[0] + (q >> 1)), "c" (inverse[q])
62 "xorl %%edx, %%edx \n\t"
64 "movzwl %%ax, %%eax \n\t"
66 : "a" (block[0] + (q >> 1)), "c" (q)
71 /* For AIC we skip quant/dequant of INTRADC */
74 block[0]=0; //avoid fake overflow
75 // temp_block[0] = (block[0] + (q >> 1)) / q;
77 bias = s->q_intra_matrix16_bias[qscale];
78 qmat = s->q_intra_matrix16[qscale];
81 bias = s->q_inter_matrix16_bias[qscale];
82 qmat = s->q_inter_matrix16[qscale];
85 if(s->out_format == FMT_H263 && s->mpeg_quant==0){
88 "movd %%eax, %%mm3 \n\t" // last_non_zero_p1
90 "pxor %%mm7, %%mm7 \n\t" // 0
91 "pxor %%mm4, %%mm4 \n\t" // 0
92 "movq (%2), %%mm5 \n\t" // qmat[0]
93 "pxor %%mm6, %%mm6 \n\t"
94 "psubw (%3), %%mm6 \n\t" // -bias[0]
95 "movl $-128, %%eax \n\t"
98 "pxor %%mm1, %%mm1 \n\t" // 0
99 "movq (%1, %%eax), %%mm0 \n\t" // block[i]
100 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
101 "pxor %%mm1, %%mm0 \n\t"
102 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
103 "psubusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
104 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
105 "por %%mm0, %%mm4 \n\t"
106 "pxor %%mm1, %%mm0 \n\t"
107 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
108 "movq %%mm0, (%5, %%eax) \n\t"
109 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
110 "movq (%4, %%eax), %%mm1 \n\t"
111 "movq %%mm7, (%1, %%eax) \n\t" // 0
112 "pandn %%mm1, %%mm0 \n\t"
114 "addl $8, %%eax \n\t"
116 "movq %%mm3, %%mm0 \n\t"
117 "psrlq $32, %%mm3 \n\t"
119 "movq %%mm3, %%mm0 \n\t"
120 "psrlq $16, %%mm3 \n\t"
122 "movd %%mm3, %%eax \n\t"
123 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
124 : "+a" (last_non_zero_p1)
125 : "r" (block+64), "r" (qmat), "r" (bias),
126 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
128 // note the asm is split cuz gcc doesnt like that many operands ...
130 "movd %1, %%mm1 \n\t" // max_qcoeff
132 "psubusw %%mm1, %%mm4 \n\t"
133 "packuswb %%mm4, %%mm4 \n\t"
134 "movd %%mm4, %0 \n\t" // *overflow
136 : "g" (s->max_qcoeff)
140 "movd %%eax, %%mm3 \n\t" // last_non_zero_p1
142 "pxor %%mm7, %%mm7 \n\t" // 0
143 "pxor %%mm4, %%mm4 \n\t" // 0
144 "movl $-128, %%eax \n\t"
147 "pxor %%mm1, %%mm1 \n\t" // 0
148 "movq (%1, %%eax), %%mm0 \n\t" // block[i]
149 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
150 "pxor %%mm1, %%mm0 \n\t"
151 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
152 "movq (%3, %%eax), %%mm6 \n\t" // bias[0]
153 "paddusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
154 "movq (%2, %%eax), %%mm5 \n\t" // qmat[i]
155 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
156 "por %%mm0, %%mm4 \n\t"
157 "pxor %%mm1, %%mm0 \n\t"
158 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
159 "movq %%mm0, (%5, %%eax) \n\t"
160 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
161 "movq (%4, %%eax), %%mm1 \n\t"
162 "movq %%mm7, (%1, %%eax) \n\t" // 0
163 "pandn %%mm1, %%mm0 \n\t"
165 "addl $8, %%eax \n\t"
167 "movq %%mm3, %%mm0 \n\t"
168 "psrlq $32, %%mm3 \n\t"
170 "movq %%mm3, %%mm0 \n\t"
171 "psrlq $16, %%mm3 \n\t"
173 "movd %%mm3, %%eax \n\t"
174 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
175 : "+a" (last_non_zero_p1)
176 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
177 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
179 // note the asm is split cuz gcc doesnt like that many operands ...
181 "movd %1, %%mm1 \n\t" // max_qcoeff
183 "psubusw %%mm1, %%mm4 \n\t"
184 "packuswb %%mm4, %%mm4 \n\t"
185 "movd %%mm4, %0 \n\t" // *overflow
187 : "g" (s->max_qcoeff)
191 if(s->mb_intra) temp_block[0]= level; //FIXME move afer permute
193 // last_non_zero_p1=64;
194 /* permute for IDCT */
196 "movl %0, %%eax \n\t"
198 "movl %%esp, " MANGLE(esp_temp) "\n\t"
200 "movzbl (%1, %%eax), %%ebx \n\t"
201 "movzbl 1(%1, %%eax), %%ebp \n\t"
202 "movw (%2, %%ebx, 2), %%cx \n\t"
203 "movw (%2, %%ebp, 2), %%sp \n\t"
204 "movzbl " MANGLE(permutation) "(%%ebx), %%ebx\n\t"
205 "movzbl " MANGLE(permutation) "(%%ebp), %%ebp\n\t"
206 "movw %%cx, (%3, %%ebx, 2) \n\t"
207 "movw %%sp, (%3, %%ebp, 2) \n\t"
208 "addl $2, %%eax \n\t"
210 "movl " MANGLE(esp_temp) ", %%esp\n\t"
213 : "g" (-last_non_zero_p1), "d" (zigzag_direct_noperm+last_non_zero_p1), "S" (temp_block), "D" (block)
214 : "%eax", "%ebx", "%ecx"
217 for(i=0; i<last_non_zero_p1; i++)
219 int j= zigzag_direct_noperm[i];
220 block[block_permute_op(j)]= temp_block[j];
223 //block_permute(block);
225 return last_non_zero_p1 - 1;