Parent Directory
| 
Revision Log
Revision 1.9 - (view) (download)
| 1 : | Skal | 1.1 | ;/**************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - optimized SSIM routines - | ||
| 5 : | ; * | ||
| 6 : | ; * Copyright(C) 2006 Johannes Reinhardt <johannes.reinhardt@gmx.de> | ||
| 7 : | ; * | ||
| 8 : | ; * This program is free software; you can redistribute it and/or modify it | ||
| 9 : | ; * under the terms of the GNU General Public License as published by | ||
| 10 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
| 11 : | ; * (at your option) any later version. | ||
| 12 : | ; * | ||
| 13 : | ; * This program 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 | ||
| 16 : | ; * GNU General Public License for more details. | ||
| 17 : | ; * | ||
| 18 : | ; * You should have received a copy of the GNU General Public License | ||
| 19 : | ; * along with this program; if not, write to the Free Software | ||
| 20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 21 : | ; * | ||
| 22 : | ; * | ||
| 23 : | ; ***************************************************************************/ | ||
| 24 : | |||
| 25 : | BITS 32 | ||
| 26 : | |||
| 27 : | Isibaar | 1.8 | %include "nasm.inc" |
| 28 : | Skal | 1.1 | |
| 29 : | Skal | 1.2 | %macro ACC_ROW 2 |
| 30 : | Isibaar | 1.8 | movq %1,[ TMP0] |
| 31 : | movq %2,[TMP0+TMP1] | ||
| 32 : | Skal | 1.1 | psadbw %1,mm0 |
| 33 : | Skal | 1.2 | psadbw %2,mm0 |
| 34 : | Isibaar | 1.8 | lea TMP0, [TMP0+2*TMP1] |
| 35 : | Skal | 1.2 | paddw %1, %2 |
| 36 : | Skal | 1.1 | %endmacro |
| 37 : | |||
| 38 : | %macro CONSIM_1x8_SSE2 0 | ||
| 39 : | Isibaar | 1.8 | movdqu xmm0,[TMP0] |
| 40 : | movdqu xmm1,[TMP1] | ||
| 41 : | Skal | 1.1 | |
| 42 : | ;unpack to words | ||
| 43 : | punpcklbw xmm0,xmm2 | ||
| 44 : | punpcklbw xmm1,xmm2 | ||
| 45 : | |||
| 46 : | Skal | 1.3 | movaps xmm3,xmm0 |
| 47 : | movaps xmm4,xmm1 | ||
| 48 : | |||
| 49 : | pmaddwd xmm0,xmm0;orig | ||
| 50 : | pmaddwd xmm1,xmm1;comp | ||
| 51 : | pmaddwd xmm3,xmm4;corr | ||
| 52 : | |||
| 53 : | paddd xmm5,xmm0 | ||
| 54 : | paddd xmm6,xmm1 | ||
| 55 : | paddd xmm7,xmm3 | ||
| 56 : | Skal | 1.1 | %endmacro |
| 57 : | |||
| 58 : | Skal | 1.3 | %macro CONSIM_1x8_MMX 0 |
| 59 : | Isibaar | 1.8 | movq mm0,[TMP0];orig |
| 60 : | movq mm1,[TMP1];comp | ||
| 61 : | Skal | 1.1 | |
| 62 : | Skal | 1.3 | ;unpack low half of qw to words |
| 63 : | punpcklbw mm0,mm2 | ||
| 64 : | punpcklbw mm1,mm2 | ||
| 65 : | |||
| 66 : | movq mm3,mm0 | ||
| 67 : | pmaddwd mm3,mm0 | ||
| 68 : | paddd mm5,mm3; | ||
| 69 : | |||
| 70 : | movq mm4,mm1 | ||
| 71 : | pmaddwd mm4,mm1 | ||
| 72 : | paddd mm6,mm4; | ||
| 73 : | Skal | 1.1 | |
| 74 : | pmaddwd mm1,mm0 | ||
| 75 : | Skal | 1.3 | paddd mm7,mm1 |
| 76 : | |||
| 77 : | Isibaar | 1.8 | movq mm0,[TMP0];orig |
| 78 : | movq mm1,[TMP1];comp | ||
| 79 : | Skal | 1.1 | |
| 80 : | Skal | 1.3 | ;unpack high half of qw to words |
| 81 : | punpckhbw mm0,mm2 | ||
| 82 : | punpckhbw mm1,mm2 | ||
| 83 : | |||
| 84 : | movq mm3,mm0 | ||
| 85 : | pmaddwd mm3,mm0 | ||
| 86 : | paddd mm5,mm3; | ||
| 87 : | |||
| 88 : | movq mm4,mm1 | ||
| 89 : | pmaddwd mm4,mm1 | ||
| 90 : | paddd mm6,mm4; | ||
| 91 : | Skal | 1.1 | |
| 92 : | pmaddwd mm1,mm0 | ||
| 93 : | Skal | 1.3 | paddd mm7,mm1 |
| 94 : | Skal | 1.1 | %endmacro |
| 95 : | |||
| 96 : | Skal | 1.3 | %macro CONSIM_WRITEOUT 3 |
| 97 : | Isibaar | 1.9 | mov eax,prm4d;lumo |
| 98 : | mul eax; lumo^2 | ||
| 99 : | add eax, 32 | ||
| 100 : | shr eax, 6; 64*lum0^2 | ||
| 101 : | Isibaar | 1.8 | movd TMP0d,%1 |
| 102 : | Isibaar | 1.9 | sub TMP0d, eax |
| 103 : | Isibaar | 1.8 | |
| 104 : | mov TMP1,prm6; pdevo | ||
| 105 : | Isibaar | 1.9 | mov dword [TMP1],TMP0d |
| 106 : | Isibaar | 1.8 | |
| 107 : | mov eax,prm5d ;lumc | ||
| 108 : | Isibaar | 1.9 | mul eax; lumc^2 |
| 109 : | add eax, 32 | ||
| 110 : | shr eax, 6; 64*lumc^2 | ||
| 111 : | Isibaar | 1.8 | movd TMP0d,%2 |
| 112 : | Isibaar | 1.9 | sub TMP0d, eax |
| 113 : | Isibaar | 1.8 | |
| 114 : | mov TMP1,prm7; pdevc | ||
| 115 : | Isibaar | 1.9 | mov dword [TMP1],TMP0d |
| 116 : | Isibaar | 1.8 | |
| 117 : | Isibaar | 1.9 | mov eax,prm4d;lumo |
| 118 : | Isibaar | 1.8 | mul prm5d; lumo*lumc, should fit in _EAX |
| 119 : | Isibaar | 1.9 | add eax, 32 |
| 120 : | shr eax, 6; 64*lumo*lumc | ||
| 121 : | Isibaar | 1.8 | movd TMP0d,%3 |
| 122 : | Isibaar | 1.9 | sub TMP0d, eax |
| 123 : | Skal | 1.3 | |
| 124 : | Isibaar | 1.8 | mov TMP1,prm8; pcorr |
| 125 : | Isibaar | 1.9 | mov dword [TMP1],TMP0d |
| 126 : | Skal | 1.3 | %endmacro |
| 127 : | Skal | 1.1 | |
| 128 : | |||
| 129 : | Isibaar | 1.8 | SECTION .rotext align=SECTION_ALIGN |
| 130 : | Skal | 1.1 | |
| 131 : | cglobal lum_8x8_mmx | ||
| 132 : | cglobal consim_sse2 | ||
| 133 : | cglobal consim_mmx | ||
| 134 : | |||
| 135 : | ;int lum_8x8_c(uint8_t* ptr, uint32_t stride) | ||
| 136 : | |||
| 137 : | Isibaar | 1.8 | ALIGN SECTION_ALIGN |
| 138 : | Skal | 1.1 | lum_8x8_mmx: |
| 139 : | Isibaar | 1.8 | mov TMP0, prm1 ;ptr |
| 140 : | mov TMP1, prm2 ;stride | ||
| 141 : | Skal | 1.1 | |
| 142 : | pxor mm0,mm0 | ||
| 143 : | |||
| 144 : | Skal | 1.2 | ACC_ROW mm1, mm2 |
| 145 : | Skal | 1.1 | |
| 146 : | Skal | 1.2 | ACC_ROW mm3, mm4 |
| 147 : | |||
| 148 : | ACC_ROW mm5, mm6 | ||
| 149 : | |||
| 150 : | ACC_ROW mm7, mm4 | ||
| 151 : | Skal | 1.1 | |
| 152 : | paddw mm1, mm3 | ||
| 153 : | paddw mm5, mm7 | ||
| 154 : | paddw mm1, mm5 | ||
| 155 : | |||
| 156 : | movd eax,mm1 | ||
| 157 : | ret | ||
| 158 : | Isibaar | 1.7 | ENDFUNC |
| 159 : | Skal | 1.1 | |
| 160 : | Isibaar | 1.8 | ALIGN SECTION_ALIGN |
| 161 : | Skal | 1.3 | consim_sse2: |
| 162 : | Isibaar | 1.8 | mov TMP0,prm1 ;ptro |
| 163 : | mov TMP1,prm2 ;ptrc | ||
| 164 : | mov _EAX, prm3 ;stride | ||
| 165 : | Skal | 1.1 | |
| 166 : | Skal | 1.3 | pxor xmm2,xmm2;null vektor |
| 167 : | pxor xmm5,xmm5;devo | ||
| 168 : | pxor xmm6,xmm6;devc | ||
| 169 : | pxor xmm7,xmm7;corr | ||
| 170 : | |||
| 171 : | ;broadcast lumo/c | ||
| 172 : | punpcklbw xmm6,xmm6 | ||
| 173 : | punpcklwd xmm6,xmm6 | ||
| 174 : | pshufd xmm6,xmm6,00000000b;or shufps | ||
| 175 : | punpcklbw xmm7,xmm7 | ||
| 176 : | punpcklwd xmm7,xmm7 | ||
| 177 : | pshufd xmm7,xmm7,00000000b | ||
| 178 : | |||
| 179 : | CONSIM_1x8_SSE2 | ||
| 180 : | Isibaar | 1.8 | add TMP0,_EAX |
| 181 : | add TMP1,_EAX | ||
| 182 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 183 : | Isibaar | 1.8 | add TMP0,_EAX |
| 184 : | add TMP1,_EAX | ||
| 185 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 186 : | Isibaar | 1.8 | add TMP0,_EAX |
| 187 : | add TMP1,_EAX | ||
| 188 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 189 : | Isibaar | 1.8 | add TMP0,_EAX |
| 190 : | add TMP1,_EAX | ||
| 191 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 192 : | Isibaar | 1.8 | add TMP0,_EAX |
| 193 : | add TMP1,_EAX | ||
| 194 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 195 : | Isibaar | 1.8 | add TMP0,_EAX |
| 196 : | add TMP1,_EAX | ||
| 197 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 198 : | Isibaar | 1.8 | add TMP0,_EAX |
| 199 : | add TMP1,_EAX | ||
| 200 : | Skal | 1.3 | CONSIM_1x8_SSE2 |
| 201 : | Skal | 1.1 | |
| 202 : | Skal | 1.3 | ;accumulate xmm5-7 |
| 203 : | chl | 1.5 | pshufd xmm0, xmm5, 0x0E |
| 204 : | Skal | 1.3 | paddd xmm5, xmm0 |
| 205 : | chl | 1.5 | pshufd xmm0, xmm5, 0x01 |
| 206 : | Skal | 1.3 | paddd xmm5, xmm0 |
| 207 : | |||
| 208 : | chl | 1.5 | pshufd xmm1, xmm6, 0x0E |
| 209 : | Skal | 1.3 | paddd xmm6, xmm1 |
| 210 : | chl | 1.5 | pshufd xmm1, xmm6, 0x01 |
| 211 : | Skal | 1.3 | paddd xmm6, xmm1 |
| 212 : | |||
| 213 : | chl | 1.5 | pshufd xmm2, xmm7, 0x0E |
| 214 : | Skal | 1.3 | paddd xmm7, xmm2 |
| 215 : | chl | 1.5 | pshufd xmm2, xmm7, 0x01 |
| 216 : | Skal | 1.3 | paddd xmm7, xmm2 |
| 217 : | Skal | 1.1 | |
| 218 : | Skal | 1.3 | CONSIM_WRITEOUT xmm5,xmm6,xmm7 |
| 219 : | Skal | 1.1 | ret |
| 220 : | Isibaar | 1.7 | ENDFUNC |
| 221 : | Skal | 1.1 | |
| 222 : | Skal | 1.3 | |
| 223 : | |||
| 224 : | |||
| 225 : | |||
| 226 : | Isibaar | 1.8 | ALIGN SECTION_ALIGN |
| 227 : | Skal | 1.3 | consim_mmx: |
| 228 : | Isibaar | 1.8 | mov TMP0,prm1 ;ptro |
| 229 : | mov TMP1,prm2 ;ptrc | ||
| 230 : | mov _EAX,prm3;stride | ||
| 231 : | Skal | 1.3 | pxor mm2,mm2;null |
| 232 : | pxor mm5,mm5;devo | ||
| 233 : | pxor mm6,mm6;devc | ||
| 234 : | pxor mm7,mm7;corr | ||
| 235 : | Skal | 1.1 | |
| 236 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 237 : | Isibaar | 1.8 | add TMP0,_EAX |
| 238 : | add TMP1,_EAX | ||
| 239 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 240 : | Isibaar | 1.8 | add TMP0,_EAX |
| 241 : | add TMP1,_EAX | ||
| 242 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 243 : | Isibaar | 1.8 | add TMP0,_EAX |
| 244 : | add TMP1,_EAX | ||
| 245 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 246 : | Isibaar | 1.8 | add TMP0,_EAX |
| 247 : | add TMP1,_EAX | ||
| 248 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 249 : | Isibaar | 1.8 | add TMP0,_EAX |
| 250 : | add TMP1,_EAX | ||
| 251 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 252 : | Isibaar | 1.8 | add TMP0,_EAX |
| 253 : | add TMP1,_EAX | ||
| 254 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 255 : | Isibaar | 1.8 | add TMP0,_EAX |
| 256 : | add TMP1,_EAX | ||
| 257 : | Skal | 1.3 | CONSIM_1x8_MMX |
| 258 : | |||
| 259 : | movq mm0,mm5 | ||
| 260 : | psrlq mm0,32 | ||
| 261 : | paddd mm5,mm0 | ||
| 262 : | movq mm1,mm6 | ||
| 263 : | psrlq mm1,32 | ||
| 264 : | paddd mm6,mm1 | ||
| 265 : | movq mm2,mm7 | ||
| 266 : | psrlq mm2,32 | ||
| 267 : | paddd mm7,mm2 | ||
| 268 : | Skal | 1.1 | |
| 269 : | Skal | 1.3 | CONSIM_WRITEOUT mm5,mm6,mm7 |
| 270 : | Skal | 1.1 | ret |
| 271 : | Isibaar | 1.7 | ENDFUNC |
| 272 : | Isibaar | 1.6 | |
| 273 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 274 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 275 : | %endif |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |